2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Implement CPU time clocks for the POSIX clock interface.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include <linux/sched.h>
|
|
|
|
|
#include <linux/posix-timers.h>
|
|
|
|
|
#include <linux/errno.h>
|
2008-05-01 18:34:31 +07:00
|
|
|
#include <linux/math64.h>
|
|
|
|
|
#include <asm/uaccess.h>
|
2008-09-12 23:54:39 +07:00
|
|
|
#include <linux/kernel_stat.h>
|
2009-08-10 09:52:30 +07:00
|
|
|
#include <trace/events/timer.h>
|
2005-04-17 05:20:36 +07:00
|
|
|
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
/*
|
2010-03-12 05:04:37 +07:00
|
|
|
* Called after updating RLIMIT_CPU to run cpu timer and update
|
|
|
|
|
* tsk->signal->cputime_expires expiration cache if necessary. Needs
|
|
|
|
|
* siglock protection since other code may update expiration cache as
|
|
|
|
|
* well.
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
*/
|
2009-08-28 19:05:12 +07:00
|
|
|
void update_rlimit_cpu(struct task_struct *task, unsigned long rlim_new)
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
{
|
2009-07-29 17:15:26 +07:00
|
|
|
cputime_t cputime = secs_to_cputime(rlim_new);
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
|
2009-08-28 19:05:12 +07:00
|
|
|
spin_lock_irq(&task->sighand->siglock);
|
|
|
|
|
set_process_cpu_timer(task, CPUCLOCK_PROF, &cputime, NULL);
|
|
|
|
|
spin_unlock_irq(&task->sighand->siglock);
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
}
|
|
|
|
|
|
2006-01-10 11:52:27 +07:00
|
|
|
static int check_clock(const clockid_t which_clock)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
int error = 0;
|
|
|
|
|
struct task_struct *p;
|
|
|
|
|
const pid_t pid = CPUCLOCK_PID(which_clock);
|
|
|
|
|
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) >= CPUCLOCK_MAX)
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
|
|
if (pid == 0)
|
|
|
|
|
return 0;
|
|
|
|
|
|
2010-11-03 23:52:56 +07:00
|
|
|
rcu_read_lock();
|
2008-02-08 19:21:52 +07:00
|
|
|
p = find_task_by_vpid(pid);
|
2007-10-19 13:40:18 +07:00
|
|
|
if (!p || !(CPUCLOCK_PERTHREAD(which_clock) ?
|
2010-11-03 23:52:56 +07:00
|
|
|
same_thread_group(p, current) : has_group_leader_pid(p))) {
|
2005-04-17 05:20:36 +07:00
|
|
|
error = -EINVAL;
|
|
|
|
|
}
|
2010-11-03 23:52:56 +07:00
|
|
|
rcu_read_unlock();
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline union cpu_time_count
|
2006-01-10 11:52:27 +07:00
|
|
|
timespec_to_sample(const clockid_t which_clock, const struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
union cpu_time_count ret;
|
|
|
|
|
ret.sched = 0; /* high half always zero when .cpu used */
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
|
2005-11-29 04:43:55 +07:00
|
|
|
ret.sched = (unsigned long long)tp->tv_sec * NSEC_PER_SEC + tp->tv_nsec;
|
2005-04-17 05:20:36 +07:00
|
|
|
} else {
|
|
|
|
|
ret.cpu = timespec_to_cputime(tp);
|
|
|
|
|
}
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
2006-01-10 11:52:27 +07:00
|
|
|
static void sample_to_timespec(const clockid_t which_clock,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count cpu,
|
|
|
|
|
struct timespec *tp)
|
|
|
|
|
{
|
2008-05-01 18:34:31 +07:00
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED)
|
|
|
|
|
*tp = ns_to_timespec(cpu.sched);
|
|
|
|
|
else
|
2005-04-17 05:20:36 +07:00
|
|
|
cputime_to_timespec(cpu.cpu, tp);
|
|
|
|
|
}
|
|
|
|
|
|
2006-01-10 11:52:27 +07:00
|
|
|
static inline int cpu_time_before(const clockid_t which_clock,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count now,
|
|
|
|
|
union cpu_time_count then)
|
|
|
|
|
{
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
|
|
|
|
|
return now.sched < then.sched;
|
|
|
|
|
} else {
|
|
|
|
|
return cputime_lt(now.cpu, then.cpu);
|
|
|
|
|
}
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static inline void cpu_time_add(const clockid_t which_clock,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count *acc,
|
|
|
|
|
union cpu_time_count val)
|
|
|
|
|
{
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
|
|
|
|
|
acc->sched += val.sched;
|
|
|
|
|
} else {
|
|
|
|
|
acc->cpu = cputime_add(acc->cpu, val.cpu);
|
|
|
|
|
}
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static inline union cpu_time_count cpu_time_sub(const clockid_t which_clock,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count a,
|
|
|
|
|
union cpu_time_count b)
|
|
|
|
|
{
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
|
|
|
|
|
a.sched -= b.sched;
|
|
|
|
|
} else {
|
|
|
|
|
a.cpu = cputime_sub(a.cpu, b.cpu);
|
|
|
|
|
}
|
|
|
|
|
return a;
|
|
|
|
|
}
|
|
|
|
|
|
[PATCH] posix-cpu-timers: prevent signal delivery starvation
The integer divisions in the timer accounting code can round the result
down to 0. Adding 0 is without effect and the signal delivery stops.
Clamp the division result to minimum 1 to avoid this.
Problem was reported by Seongbae Park <spark@google.com>, who provided
also an inital patch.
Roland sayeth:
I have had some more time to think about the problem, and to reproduce it
using Toyo's test case. For the record, if my understanding of the problem
is correct, this happens only in one very particular case. First, the
expiry time has to be so soon that in cputime_t units (usually 1s/HZ ticks)
it's < nthreads so the division yields zero. Second, it only affects each
thread that is so new that its CPU time accumulation is zero so now+0 is
still zero and ->it_*_expires winds up staying zero. For the VIRT and PROF
clocks when cputime_t is tick granularity (or the SCHED clock on
configurations where sched_clock's value only advances on clock ticks), this
is not hard to arrange with new threads starting up and blocking before they
accumulate a whole tick of CPU time. That's what happens in Toyo's test
case.
Note that in general it is fine for that division to round down to zero,
and set each thread's expiry time to its "now" time. The problem only
arises with thread's whose "now" value is still zero, so that now+0 winds up
0 and is interpreted as "not set" instead of ">= now". So it would be a
sufficient and more precise fix to just use max(ticks, 1) inside the loop
when setting each it_*_expires value.
But, it does no harm to round the division up to one and always advance
every thread's expiry time. If the thread didn't already fire timers for
the expiry time of "now", there is no expectation that it will do so before
the next tick anyway. So I followed Thomas's patch in lifting the max out
of the loops.
This patch also covers the reload cases, which are harder to write a test
for (and I didn't try). I've tested it with Toyo's case and it fixes that.
[toyoa@mvista.com: fix: min_t -> max_t]
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Roland McGrath <roland@redhat.com>
Cc: Daniel Walker <dwalker@mvista.com>
Cc: Toyo Abe <toyoa@mvista.com>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Roman Zippel <zippel@linux-m68k.org>
Cc: Seongbae Park <spark@google.com>
Cc: Peter Mattis <pmattis@google.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Martin Bligh <mbligh@google.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-10-17 14:09:39 +07:00
|
|
|
/*
|
|
|
|
|
* Divide and limit the result to res >= 1
|
|
|
|
|
*
|
|
|
|
|
* This is necessary to prevent signal delivery starvation, when the result of
|
|
|
|
|
* the division would be rounded down to 0.
|
|
|
|
|
*/
|
|
|
|
|
static inline cputime_t cputime_div_non_zero(cputime_t time, unsigned long div)
|
|
|
|
|
{
|
|
|
|
|
cputime_t res = cputime_div(time, div);
|
|
|
|
|
|
|
|
|
|
return max_t(cputime_t, res, 1);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Update expiry time from increment, and increase overrun count,
|
|
|
|
|
* given the current clock sample.
|
|
|
|
|
*/
|
2005-10-26 23:26:53 +07:00
|
|
|
static void bump_cpu_timer(struct k_itimer *timer,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count now)
|
|
|
|
|
{
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
if (timer->it.cpu.incr.sched == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (CPUCLOCK_WHICH(timer->it_clock) == CPUCLOCK_SCHED) {
|
|
|
|
|
unsigned long long delta, incr;
|
|
|
|
|
|
|
|
|
|
if (now.sched < timer->it.cpu.expires.sched)
|
|
|
|
|
return;
|
|
|
|
|
incr = timer->it.cpu.incr.sched;
|
|
|
|
|
delta = now.sched + incr - timer->it.cpu.expires.sched;
|
|
|
|
|
/* Don't use (incr*2 < delta), incr*2 might overflow. */
|
|
|
|
|
for (i = 0; incr < delta - incr; i++)
|
|
|
|
|
incr = incr << 1;
|
|
|
|
|
for (; i >= 0; incr >>= 1, i--) {
|
2005-10-26 23:26:53 +07:00
|
|
|
if (delta < incr)
|
2005-04-17 05:20:36 +07:00
|
|
|
continue;
|
|
|
|
|
timer->it.cpu.expires.sched += incr;
|
|
|
|
|
timer->it_overrun += 1 << i;
|
|
|
|
|
delta -= incr;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
cputime_t delta, incr;
|
|
|
|
|
|
|
|
|
|
if (cputime_lt(now.cpu, timer->it.cpu.expires.cpu))
|
|
|
|
|
return;
|
|
|
|
|
incr = timer->it.cpu.incr.cpu;
|
|
|
|
|
delta = cputime_sub(cputime_add(now.cpu, incr),
|
|
|
|
|
timer->it.cpu.expires.cpu);
|
|
|
|
|
/* Don't use (incr*2 < delta), incr*2 might overflow. */
|
|
|
|
|
for (i = 0; cputime_lt(incr, cputime_sub(delta, incr)); i++)
|
|
|
|
|
incr = cputime_add(incr, incr);
|
|
|
|
|
for (; i >= 0; incr = cputime_halve(incr), i--) {
|
2005-10-26 23:26:53 +07:00
|
|
|
if (cputime_lt(delta, incr))
|
2005-04-17 05:20:36 +07:00
|
|
|
continue;
|
|
|
|
|
timer->it.cpu.expires.cpu =
|
|
|
|
|
cputime_add(timer->it.cpu.expires.cpu, incr);
|
|
|
|
|
timer->it_overrun += 1 << i;
|
|
|
|
|
delta = cputime_sub(delta, incr);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline cputime_t prof_ticks(struct task_struct *p)
|
|
|
|
|
{
|
|
|
|
|
return cputime_add(p->utime, p->stime);
|
|
|
|
|
}
|
|
|
|
|
static inline cputime_t virt_ticks(struct task_struct *p)
|
|
|
|
|
{
|
|
|
|
|
return p->utime;
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static int
|
|
|
|
|
posix_cpu_clock_getres(const clockid_t which_clock, struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
int error = check_clock(which_clock);
|
|
|
|
|
if (!error) {
|
|
|
|
|
tp->tv_sec = 0;
|
|
|
|
|
tp->tv_nsec = ((NSEC_PER_SEC + HZ - 1) / HZ);
|
|
|
|
|
if (CPUCLOCK_WHICH(which_clock) == CPUCLOCK_SCHED) {
|
|
|
|
|
/*
|
|
|
|
|
* If sched_clock is using a cycle counter, we
|
|
|
|
|
* don't have any idea of its true resolution
|
|
|
|
|
* exported, but it is much more than 1s/HZ.
|
|
|
|
|
*/
|
|
|
|
|
tp->tv_nsec = 1;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static int
|
|
|
|
|
posix_cpu_clock_set(const clockid_t which_clock, const struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* You can never reset a CPU clock, but we check for other errors
|
|
|
|
|
* in the call before failing with EPERM.
|
|
|
|
|
*/
|
|
|
|
|
int error = check_clock(which_clock);
|
|
|
|
|
if (error == 0) {
|
|
|
|
|
error = -EPERM;
|
|
|
|
|
}
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Sample a per-thread clock for the given task.
|
|
|
|
|
*/
|
2006-01-10 11:52:27 +07:00
|
|
|
static int cpu_clock_sample(const clockid_t which_clock, struct task_struct *p,
|
2005-04-17 05:20:36 +07:00
|
|
|
union cpu_time_count *cpu)
|
|
|
|
|
{
|
|
|
|
|
switch (CPUCLOCK_WHICH(which_clock)) {
|
|
|
|
|
default:
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
case CPUCLOCK_PROF:
|
|
|
|
|
cpu->cpu = prof_ticks(p);
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_VIRT:
|
|
|
|
|
cpu->cpu = virt_ticks(p);
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_SCHED:
|
2009-03-31 14:56:03 +07:00
|
|
|
cpu->sched = task_sched_runtime(p);
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2009-02-05 18:24:16 +07:00
|
|
|
void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
|
|
|
|
|
{
|
2010-06-11 06:09:56 +07:00
|
|
|
struct signal_struct *sig = tsk->signal;
|
2009-02-05 18:24:16 +07:00
|
|
|
struct task_struct *t;
|
|
|
|
|
|
2010-06-11 06:09:56 +07:00
|
|
|
times->utime = sig->utime;
|
|
|
|
|
times->stime = sig->stime;
|
|
|
|
|
times->sum_exec_runtime = sig->sum_sched_runtime;
|
2009-02-05 18:24:16 +07:00
|
|
|
|
|
|
|
|
rcu_read_lock();
|
2010-06-11 06:09:56 +07:00
|
|
|
/* make sure we can trust tsk->thread_group list */
|
|
|
|
|
if (!likely(pid_alive(tsk)))
|
2009-02-05 18:24:16 +07:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
t = tsk;
|
|
|
|
|
do {
|
|
|
|
|
times->utime = cputime_add(times->utime, t->utime);
|
|
|
|
|
times->stime = cputime_add(times->stime, t->stime);
|
|
|
|
|
times->sum_exec_runtime += t->se.sum_exec_runtime;
|
2010-06-11 06:09:56 +07:00
|
|
|
} while_each_thread(tsk, t);
|
2009-02-05 18:24:16 +07:00
|
|
|
out:
|
|
|
|
|
rcu_read_unlock();
|
|
|
|
|
}
|
|
|
|
|
|
2009-02-11 17:30:27 +07:00
|
|
|
static void update_gt_cputime(struct task_cputime *a, struct task_cputime *b)
|
|
|
|
|
{
|
|
|
|
|
if (cputime_gt(b->utime, a->utime))
|
|
|
|
|
a->utime = b->utime;
|
|
|
|
|
|
|
|
|
|
if (cputime_gt(b->stime, a->stime))
|
|
|
|
|
a->stime = b->stime;
|
|
|
|
|
|
|
|
|
|
if (b->sum_exec_runtime > a->sum_exec_runtime)
|
|
|
|
|
a->sum_exec_runtime = b->sum_exec_runtime;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void thread_group_cputimer(struct task_struct *tsk, struct task_cputime *times)
|
|
|
|
|
{
|
|
|
|
|
struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
|
|
|
|
|
struct task_cputime sum;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&cputimer->lock, flags);
|
|
|
|
|
if (!cputimer->running) {
|
|
|
|
|
cputimer->running = 1;
|
|
|
|
|
/*
|
|
|
|
|
* The POSIX timer interface allows for absolute time expiry
|
|
|
|
|
* values through the TIMER_ABSTIME flag, therefore we have
|
|
|
|
|
* to synchronize the timer to the clock every time we start
|
|
|
|
|
* it.
|
|
|
|
|
*/
|
|
|
|
|
thread_group_cputime(tsk, &sum);
|
|
|
|
|
update_gt_cputime(&cputimer->cputime, &sum);
|
|
|
|
|
}
|
|
|
|
|
*times = cputimer->cputime;
|
|
|
|
|
spin_unlock_irqrestore(&cputimer->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Sample a process (thread group) clock for the given group_leader task.
|
|
|
|
|
* Must be called with tasklist_lock held for reading.
|
|
|
|
|
*/
|
2008-09-12 23:54:39 +07:00
|
|
|
static int cpu_clock_sample_group(const clockid_t which_clock,
|
|
|
|
|
struct task_struct *p,
|
|
|
|
|
union cpu_time_count *cpu)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
struct task_cputime cputime;
|
|
|
|
|
|
2008-11-24 21:46:31 +07:00
|
|
|
switch (CPUCLOCK_WHICH(which_clock)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
default:
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
case CPUCLOCK_PROF:
|
2009-03-31 14:56:03 +07:00
|
|
|
thread_group_cputime(p, &cputime);
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
cpu->cpu = cputime_add(cputime.utime, cputime.stime);
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_VIRT:
|
2009-03-31 14:56:03 +07:00
|
|
|
thread_group_cputime(p, &cputime);
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
cpu->cpu = cputime.utime;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_SCHED:
|
2009-03-31 14:56:03 +07:00
|
|
|
cpu->sched = thread_group_sched_runtime(p);
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static int posix_cpu_clock_get(const clockid_t which_clock, struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
const pid_t pid = CPUCLOCK_PID(which_clock);
|
|
|
|
|
int error = -EINVAL;
|
|
|
|
|
union cpu_time_count rtn;
|
|
|
|
|
|
|
|
|
|
if (pid == 0) {
|
|
|
|
|
/*
|
|
|
|
|
* Special case constant value for our own clocks.
|
|
|
|
|
* We don't have to do any lookup to find ourselves.
|
|
|
|
|
*/
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(which_clock)) {
|
|
|
|
|
/*
|
|
|
|
|
* Sampling just ourselves we can do with no locking.
|
|
|
|
|
*/
|
|
|
|
|
error = cpu_clock_sample(which_clock,
|
|
|
|
|
current, &rtn);
|
|
|
|
|
} else {
|
|
|
|
|
read_lock(&tasklist_lock);
|
|
|
|
|
error = cpu_clock_sample_group(which_clock,
|
|
|
|
|
current, &rtn);
|
|
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/*
|
|
|
|
|
* Find the given PID, and validate that the caller
|
|
|
|
|
* should be able to see it.
|
|
|
|
|
*/
|
|
|
|
|
struct task_struct *p;
|
2007-02-16 16:28:22 +07:00
|
|
|
rcu_read_lock();
|
2008-02-08 19:21:52 +07:00
|
|
|
p = find_task_by_vpid(pid);
|
2005-04-17 05:20:36 +07:00
|
|
|
if (p) {
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(which_clock)) {
|
2007-10-19 13:40:18 +07:00
|
|
|
if (same_thread_group(p, current)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
error = cpu_clock_sample(which_clock,
|
|
|
|
|
p, &rtn);
|
|
|
|
|
}
|
2007-02-16 16:28:22 +07:00
|
|
|
} else {
|
|
|
|
|
read_lock(&tasklist_lock);
|
2010-05-27 04:43:13 +07:00
|
|
|
if (thread_group_leader(p) && p->sighand) {
|
2007-02-16 16:28:22 +07:00
|
|
|
error =
|
|
|
|
|
cpu_clock_sample_group(which_clock,
|
|
|
|
|
p, &rtn);
|
|
|
|
|
}
|
|
|
|
|
read_unlock(&tasklist_lock);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
}
|
2007-02-16 16:28:22 +07:00
|
|
|
rcu_read_unlock();
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (error)
|
|
|
|
|
return error;
|
|
|
|
|
sample_to_timespec(which_clock, rtn, tp);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Validate the clockid_t for a new CPU-clock timer, and initialize the timer.
|
2009-11-18 05:14:13 +07:00
|
|
|
* This is called from sys_timer_create() and do_cpu_nanosleep() with the
|
|
|
|
|
* new timer already all-zeros initialized.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
2011-02-01 20:52:12 +07:00
|
|
|
static int posix_cpu_timer_create(struct k_itimer *new_timer)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
int ret = 0;
|
|
|
|
|
const pid_t pid = CPUCLOCK_PID(new_timer->it_clock);
|
|
|
|
|
struct task_struct *p;
|
|
|
|
|
|
|
|
|
|
if (CPUCLOCK_WHICH(new_timer->it_clock) >= CPUCLOCK_MAX)
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
|
|
INIT_LIST_HEAD(&new_timer->it.cpu.entry);
|
|
|
|
|
|
2010-11-03 23:52:56 +07:00
|
|
|
rcu_read_lock();
|
2005-04-17 05:20:36 +07:00
|
|
|
if (CPUCLOCK_PERTHREAD(new_timer->it_clock)) {
|
|
|
|
|
if (pid == 0) {
|
|
|
|
|
p = current;
|
|
|
|
|
} else {
|
2008-02-08 19:21:52 +07:00
|
|
|
p = find_task_by_vpid(pid);
|
2007-10-19 13:40:18 +07:00
|
|
|
if (p && !same_thread_group(p, current))
|
2005-04-17 05:20:36 +07:00
|
|
|
p = NULL;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
if (pid == 0) {
|
|
|
|
|
p = current->group_leader;
|
|
|
|
|
} else {
|
2008-02-08 19:21:52 +07:00
|
|
|
p = find_task_by_vpid(pid);
|
2010-11-03 23:52:56 +07:00
|
|
|
if (p && !has_group_leader_pid(p))
|
2005-04-17 05:20:36 +07:00
|
|
|
p = NULL;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
new_timer->it.cpu.task = p;
|
|
|
|
|
if (p) {
|
|
|
|
|
get_task_struct(p);
|
|
|
|
|
} else {
|
|
|
|
|
ret = -EINVAL;
|
|
|
|
|
}
|
2010-11-03 23:52:56 +07:00
|
|
|
rcu_read_unlock();
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Clean up a CPU-clock timer that is about to be destroyed.
|
|
|
|
|
* This is called from timer deletion with the timer already locked.
|
|
|
|
|
* If we return TIMER_RETRY, it's necessary to release the timer's lock
|
|
|
|
|
* and try again. (This happens when the timer is in the middle of firing.)
|
|
|
|
|
*/
|
2011-02-01 20:52:12 +07:00
|
|
|
static int posix_cpu_timer_del(struct k_itimer *timer)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
struct task_struct *p = timer->it.cpu.task;
|
2005-10-23 23:25:39 +07:00
|
|
|
int ret = 0;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2005-10-23 23:25:39 +07:00
|
|
|
if (likely(p != NULL)) {
|
2005-10-22 05:36:00 +07:00
|
|
|
read_lock(&tasklist_lock);
|
2010-05-27 04:43:13 +07:00
|
|
|
if (unlikely(p->sighand == NULL)) {
|
2005-10-22 05:36:00 +07:00
|
|
|
/*
|
|
|
|
|
* We raced with the reaping of the task.
|
|
|
|
|
* The deletion should have cleared us off the list.
|
|
|
|
|
*/
|
|
|
|
|
BUG_ON(!list_empty(&timer->it.cpu.entry));
|
|
|
|
|
} else {
|
|
|
|
|
spin_lock(&p->sighand->siglock);
|
2005-10-23 23:25:39 +07:00
|
|
|
if (timer->it.cpu.firing)
|
|
|
|
|
ret = TIMER_RETRY;
|
|
|
|
|
else
|
|
|
|
|
list_del(&timer->it.cpu.entry);
|
2005-10-22 05:36:00 +07:00
|
|
|
spin_unlock(&p->sighand->siglock);
|
|
|
|
|
}
|
|
|
|
|
read_unlock(&tasklist_lock);
|
2005-10-23 23:25:39 +07:00
|
|
|
|
|
|
|
|
if (!ret)
|
|
|
|
|
put_task_struct(p);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
2005-10-23 23:25:39 +07:00
|
|
|
return ret;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Clean out CPU timers still ticking when a thread exited. The task
|
|
|
|
|
* pointer is cleared, and the expiry time is replaced with the residual
|
|
|
|
|
* time for later timer_gettime calls to return.
|
|
|
|
|
* This must be called with the siglock held.
|
|
|
|
|
*/
|
|
|
|
|
static void cleanup_timers(struct list_head *head,
|
|
|
|
|
cputime_t utime, cputime_t stime,
|
2007-07-09 23:51:58 +07:00
|
|
|
unsigned long long sum_exec_runtime)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
struct cpu_timer_list *timer, *next;
|
|
|
|
|
cputime_t ptime = cputime_add(utime, stime);
|
|
|
|
|
|
|
|
|
|
list_for_each_entry_safe(timer, next, head, entry) {
|
|
|
|
|
list_del_init(&timer->entry);
|
|
|
|
|
if (cputime_lt(timer->expires.cpu, ptime)) {
|
|
|
|
|
timer->expires.cpu = cputime_zero;
|
|
|
|
|
} else {
|
|
|
|
|
timer->expires.cpu = cputime_sub(timer->expires.cpu,
|
|
|
|
|
ptime);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++head;
|
|
|
|
|
list_for_each_entry_safe(timer, next, head, entry) {
|
|
|
|
|
list_del_init(&timer->entry);
|
|
|
|
|
if (cputime_lt(timer->expires.cpu, utime)) {
|
|
|
|
|
timer->expires.cpu = cputime_zero;
|
|
|
|
|
} else {
|
|
|
|
|
timer->expires.cpu = cputime_sub(timer->expires.cpu,
|
|
|
|
|
utime);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++head;
|
|
|
|
|
list_for_each_entry_safe(timer, next, head, entry) {
|
|
|
|
|
list_del_init(&timer->entry);
|
2007-07-09 23:51:58 +07:00
|
|
|
if (timer->expires.sched < sum_exec_runtime) {
|
2005-04-17 05:20:36 +07:00
|
|
|
timer->expires.sched = 0;
|
|
|
|
|
} else {
|
2007-07-09 23:51:58 +07:00
|
|
|
timer->expires.sched -= sum_exec_runtime;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* These are both called with the siglock held, when the current thread
|
|
|
|
|
* is being reaped. When the final (leader) thread in the group is reaped,
|
|
|
|
|
* posix_cpu_timers_exit_group will be called after posix_cpu_timers_exit.
|
|
|
|
|
*/
|
|
|
|
|
void posix_cpu_timers_exit(struct task_struct *tsk)
|
|
|
|
|
{
|
|
|
|
|
cleanup_timers(tsk->cpu_timers,
|
2007-07-09 23:51:58 +07:00
|
|
|
tsk->utime, tsk->stime, tsk->se.sum_exec_runtime);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
}
|
|
|
|
|
void posix_cpu_timers_exit_group(struct task_struct *tsk)
|
|
|
|
|
{
|
2009-08-07 06:03:30 +07:00
|
|
|
struct signal_struct *const sig = tsk->signal;
|
2005-10-24 17:36:28 +07:00
|
|
|
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
cleanup_timers(tsk->signal->cpu_timers,
|
2009-08-07 06:03:30 +07:00
|
|
|
cputime_add(tsk->utime, sig->utime),
|
|
|
|
|
cputime_add(tsk->stime, sig->stime),
|
|
|
|
|
tsk->se.sum_exec_runtime + sig->sum_sched_runtime);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void clear_dead_task(struct k_itimer *timer, union cpu_time_count now)
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* That's all for this thread or process.
|
|
|
|
|
* We leave our residual in expires to be reported.
|
|
|
|
|
*/
|
|
|
|
|
put_task_struct(timer->it.cpu.task);
|
|
|
|
|
timer->it.cpu.task = NULL;
|
|
|
|
|
timer->it.cpu.expires = cpu_time_sub(timer->it_clock,
|
|
|
|
|
timer->it.cpu.expires,
|
|
|
|
|
now);
|
|
|
|
|
}
|
|
|
|
|
|
2009-07-29 17:15:28 +07:00
|
|
|
static inline int expires_gt(cputime_t expires, cputime_t new_exp)
|
|
|
|
|
{
|
|
|
|
|
return cputime_eq(expires, cputime_zero) ||
|
|
|
|
|
cputime_gt(expires, new_exp);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Insert the timer on the appropriate list before any timers that
|
|
|
|
|
* expire later. This must be called with the tasklist_lock held
|
2010-03-12 05:04:42 +07:00
|
|
|
* for reading, interrupts disabled and p->sighand->siglock taken.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
2010-03-12 05:04:38 +07:00
|
|
|
static void arm_timer(struct k_itimer *timer)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
struct task_struct *p = timer->it.cpu.task;
|
|
|
|
|
struct list_head *head, *listpos;
|
2010-03-12 05:04:38 +07:00
|
|
|
struct task_cputime *cputime_expires;
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list *const nt = &timer->it.cpu;
|
|
|
|
|
struct cpu_timer_list *next;
|
|
|
|
|
|
2010-03-12 05:04:38 +07:00
|
|
|
if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
|
|
|
|
|
head = p->cpu_timers;
|
|
|
|
|
cputime_expires = &p->cputime_expires;
|
|
|
|
|
} else {
|
|
|
|
|
head = p->signal->cpu_timers;
|
|
|
|
|
cputime_expires = &p->signal->cputime_expires;
|
|
|
|
|
}
|
2005-04-17 05:20:36 +07:00
|
|
|
head += CPUCLOCK_WHICH(timer->it_clock);
|
|
|
|
|
|
|
|
|
|
listpos = head;
|
2010-03-12 05:04:38 +07:00
|
|
|
list_for_each_entry(next, head, entry) {
|
|
|
|
|
if (cpu_time_before(timer->it_clock, nt->expires, next->expires))
|
|
|
|
|
break;
|
|
|
|
|
listpos = &next->entry;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
list_add(&nt->entry, listpos);
|
|
|
|
|
|
|
|
|
|
if (listpos == head) {
|
2010-03-12 05:04:38 +07:00
|
|
|
union cpu_time_count *exp = &nt->expires;
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
2010-03-12 05:04:38 +07:00
|
|
|
* We are the new earliest-expiring POSIX 1.b timer, hence
|
|
|
|
|
* need to update expiration cache. Take into account that
|
|
|
|
|
* for process timers we share expiration cache with itimers
|
|
|
|
|
* and RLIMIT_CPU and for thread timers with RLIMIT_RTTIME.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
|
|
|
|
|
2010-03-12 05:04:38 +07:00
|
|
|
switch (CPUCLOCK_WHICH(timer->it_clock)) {
|
|
|
|
|
case CPUCLOCK_PROF:
|
|
|
|
|
if (expires_gt(cputime_expires->prof_exp, exp->cpu))
|
|
|
|
|
cputime_expires->prof_exp = exp->cpu;
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_VIRT:
|
|
|
|
|
if (expires_gt(cputime_expires->virt_exp, exp->cpu))
|
|
|
|
|
cputime_expires->virt_exp = exp->cpu;
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_SCHED:
|
|
|
|
|
if (cputime_expires->sched_exp == 0 ||
|
|
|
|
|
cputime_expires->sched_exp > exp->sched)
|
|
|
|
|
cputime_expires->sched_exp = exp->sched;
|
|
|
|
|
break;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The timer is locked, fire it and arrange for its reload.
|
|
|
|
|
*/
|
|
|
|
|
static void cpu_timer_fire(struct k_itimer *timer)
|
|
|
|
|
{
|
cpu-timers: Change SIGEV_NONE timer implementation
When user sets up a timer without associated signal and process does
not use any other cpu timers and does not exit, tsk->signal->cputimer
is enabled and running forever.
Avoid running the timer for no reason.
I used below program to check patch does not break current user space
visible behavior.
#include <sys/time.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
#include <assert.h>
void consume_cpu(void)
{
int i = 0;
int count = 0;
for(i=0; i<100000000; i++)
count++;
}
int main(void)
{
int i;
struct sigaction act;
struct sigevent evt = { };
timer_t tid;
struct itimerspec spec = { };
evt.sigev_notify = SIGEV_NONE;
assert(timer_create(CLOCK_PROCESS_CPUTIME_ID, &evt, &tid) == 0);
spec.it_value.tv_sec = 10;
assert(timer_settime(tid, 0, &spec, NULL) == 0);
for (i = 0; i < 30; i++) {
consume_cpu();
memset(&spec, 0, sizeof(spec));
assert(timer_gettime(tid, &spec) == 0);
printf("%lu.%09lu\n",
(unsigned long) spec.it_value.tv_sec,
(unsigned long) spec.it_value.tv_nsec);
}
assert(timer_delete(tid) == 0);
return 0;
}
Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-03-12 05:04:41 +07:00
|
|
|
if ((timer->it_sigev_notify & ~SIGEV_THREAD_ID) == SIGEV_NONE) {
|
|
|
|
|
/*
|
|
|
|
|
* User don't want any signal.
|
|
|
|
|
*/
|
|
|
|
|
timer->it.cpu.expires.sched = 0;
|
|
|
|
|
} else if (unlikely(timer->sigq == NULL)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* This a special case for clock_nanosleep,
|
|
|
|
|
* not a normal timer from sys_timer_create.
|
|
|
|
|
*/
|
|
|
|
|
wake_up_process(timer->it_process);
|
|
|
|
|
timer->it.cpu.expires.sched = 0;
|
|
|
|
|
} else if (timer->it.cpu.incr.sched == 0) {
|
|
|
|
|
/*
|
|
|
|
|
* One-shot timer. Clear it as soon as it's fired.
|
|
|
|
|
*/
|
|
|
|
|
posix_timer_event(timer, 0);
|
|
|
|
|
timer->it.cpu.expires.sched = 0;
|
|
|
|
|
} else if (posix_timer_event(timer, ++timer->it_requeue_pending)) {
|
|
|
|
|
/*
|
|
|
|
|
* The signal did not get queued because the signal
|
|
|
|
|
* was ignored, so we won't get any callback to
|
|
|
|
|
* reload the timer. But we need to keep it
|
|
|
|
|
* ticking in case the signal is deliverable next time.
|
|
|
|
|
*/
|
|
|
|
|
posix_cpu_timer_schedule(timer);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2009-02-12 21:00:52 +07:00
|
|
|
/*
|
|
|
|
|
* Sample a process (thread group) timer for the given group_leader task.
|
|
|
|
|
* Must be called with tasklist_lock held for reading.
|
|
|
|
|
*/
|
|
|
|
|
static int cpu_timer_sample_group(const clockid_t which_clock,
|
|
|
|
|
struct task_struct *p,
|
|
|
|
|
union cpu_time_count *cpu)
|
|
|
|
|
{
|
|
|
|
|
struct task_cputime cputime;
|
|
|
|
|
|
|
|
|
|
thread_group_cputimer(p, &cputime);
|
|
|
|
|
switch (CPUCLOCK_WHICH(which_clock)) {
|
|
|
|
|
default:
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
case CPUCLOCK_PROF:
|
|
|
|
|
cpu->cpu = cputime_add(cputime.utime, cputime.stime);
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_VIRT:
|
|
|
|
|
cpu->cpu = cputime.utime;
|
|
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_SCHED:
|
|
|
|
|
cpu->sched = cputime.sum_exec_runtime + task_delta_exec(p);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Guts of sys_timer_settime for CPU timers.
|
|
|
|
|
* This is called with the timer locked and interrupts disabled.
|
|
|
|
|
* If we return TIMER_RETRY, it's necessary to release the timer's lock
|
|
|
|
|
* and try again. (This happens when the timer is in the middle of firing.)
|
|
|
|
|
*/
|
2011-02-01 20:52:12 +07:00
|
|
|
static int posix_cpu_timer_set(struct k_itimer *timer, int flags,
|
|
|
|
|
struct itimerspec *new, struct itimerspec *old)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
struct task_struct *p = timer->it.cpu.task;
|
cpu-timers: Return correct previous timer reload value
According POSIX we need to correctly set old timer it_interval value when
user request that in timer_settime(). Tested using below program.
#include <sys/time.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <unistd.h>
#include <assert.h>
int main(void)
{
struct sigaction act;
struct sigevent evt = { };
timer_t tid;
struct itimerspec spec, u_spec, k_spec;
evt.sigev_notify = SIGEV_SIGNAL;
evt.sigev_signo = SIGPROF;
assert(timer_create(CLOCK_PROCESS_CPUTIME_ID, &evt, &tid) == 0);
spec.it_value.tv_sec = 1;
spec.it_value.tv_nsec = 2;
spec.it_interval.tv_sec = 3;
spec.it_interval.tv_nsec = 4;
u_spec = spec;
assert(timer_settime(tid, 0, &spec, NULL) == 0);
spec.it_value.tv_sec = 5;
spec.it_value.tv_nsec = 6;
spec.it_interval.tv_sec = 7;
spec.it_interval.tv_nsec = 8;
assert(timer_settime(tid, 0, &spec, &k_spec) == 0);
#define PRT(val) printf(#val ":\t%d/%d\n", (int) u_spec.val, (int) k_spec.val)
PRT(it_value.tv_sec);
PRT(it_value.tv_nsec);
PRT(it_interval.tv_sec);
PRT(it_interval.tv_nsec);
return 0;
}
Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-03-12 05:04:39 +07:00
|
|
|
union cpu_time_count old_expires, new_expires, old_incr, val;
|
2005-04-17 05:20:36 +07:00
|
|
|
int ret;
|
|
|
|
|
|
|
|
|
|
if (unlikely(p == NULL)) {
|
|
|
|
|
/*
|
|
|
|
|
* Timer refers to a dead task's clock.
|
|
|
|
|
*/
|
|
|
|
|
return -ESRCH;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
new_expires = timespec_to_sample(timer->it_clock, &new->it_value);
|
|
|
|
|
|
|
|
|
|
read_lock(&tasklist_lock);
|
|
|
|
|
/*
|
|
|
|
|
* We need the tasklist_lock to protect against reaping that
|
2010-05-27 04:43:13 +07:00
|
|
|
* clears p->sighand. If p has just been reaped, we can no
|
2005-04-17 05:20:36 +07:00
|
|
|
* longer get any information about it at all.
|
|
|
|
|
*/
|
2010-05-27 04:43:13 +07:00
|
|
|
if (unlikely(p->sighand == NULL)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
put_task_struct(p);
|
|
|
|
|
timer->it.cpu.task = NULL;
|
|
|
|
|
return -ESRCH;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Disarm any old timer after extracting its expiry time.
|
|
|
|
|
*/
|
|
|
|
|
BUG_ON(!irqs_disabled());
|
2005-10-24 21:29:58 +07:00
|
|
|
|
|
|
|
|
ret = 0;
|
cpu-timers: Return correct previous timer reload value
According POSIX we need to correctly set old timer it_interval value when
user request that in timer_settime(). Tested using below program.
#include <sys/time.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <unistd.h>
#include <assert.h>
int main(void)
{
struct sigaction act;
struct sigevent evt = { };
timer_t tid;
struct itimerspec spec, u_spec, k_spec;
evt.sigev_notify = SIGEV_SIGNAL;
evt.sigev_signo = SIGPROF;
assert(timer_create(CLOCK_PROCESS_CPUTIME_ID, &evt, &tid) == 0);
spec.it_value.tv_sec = 1;
spec.it_value.tv_nsec = 2;
spec.it_interval.tv_sec = 3;
spec.it_interval.tv_nsec = 4;
u_spec = spec;
assert(timer_settime(tid, 0, &spec, NULL) == 0);
spec.it_value.tv_sec = 5;
spec.it_value.tv_nsec = 6;
spec.it_interval.tv_sec = 7;
spec.it_interval.tv_nsec = 8;
assert(timer_settime(tid, 0, &spec, &k_spec) == 0);
#define PRT(val) printf(#val ":\t%d/%d\n", (int) u_spec.val, (int) k_spec.val)
PRT(it_value.tv_sec);
PRT(it_value.tv_nsec);
PRT(it_interval.tv_sec);
PRT(it_interval.tv_nsec);
return 0;
}
Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-03-12 05:04:39 +07:00
|
|
|
old_incr = timer->it.cpu.incr;
|
2005-04-17 05:20:36 +07:00
|
|
|
spin_lock(&p->sighand->siglock);
|
|
|
|
|
old_expires = timer->it.cpu.expires;
|
2005-10-24 21:29:58 +07:00
|
|
|
if (unlikely(timer->it.cpu.firing)) {
|
|
|
|
|
timer->it.cpu.firing = -1;
|
|
|
|
|
ret = TIMER_RETRY;
|
|
|
|
|
} else
|
|
|
|
|
list_del_init(&timer->it.cpu.entry);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We need to sample the current value to convert the new
|
|
|
|
|
* value from to relative and absolute, and to convert the
|
|
|
|
|
* old value from absolute to relative. To set a process
|
|
|
|
|
* timer, we need a sample to balance the thread expiry
|
|
|
|
|
* times (in arm_timer). With an absolute time, we must
|
|
|
|
|
* check if it's already passed. In short, we need a sample.
|
|
|
|
|
*/
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
|
|
|
|
|
cpu_clock_sample(timer->it_clock, p, &val);
|
|
|
|
|
} else {
|
2009-02-12 21:00:52 +07:00
|
|
|
cpu_timer_sample_group(timer->it_clock, p, &val);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (old) {
|
|
|
|
|
if (old_expires.sched == 0) {
|
|
|
|
|
old->it_value.tv_sec = 0;
|
|
|
|
|
old->it_value.tv_nsec = 0;
|
|
|
|
|
} else {
|
|
|
|
|
/*
|
|
|
|
|
* Update the timer in case it has
|
|
|
|
|
* overrun already. If it has,
|
|
|
|
|
* we'll report it as having overrun
|
|
|
|
|
* and with the next reloaded timer
|
|
|
|
|
* already ticking, though we are
|
|
|
|
|
* swallowing that pending
|
|
|
|
|
* notification here to install the
|
|
|
|
|
* new setting.
|
|
|
|
|
*/
|
|
|
|
|
bump_cpu_timer(timer, val);
|
|
|
|
|
if (cpu_time_before(timer->it_clock, val,
|
|
|
|
|
timer->it.cpu.expires)) {
|
|
|
|
|
old_expires = cpu_time_sub(
|
|
|
|
|
timer->it_clock,
|
|
|
|
|
timer->it.cpu.expires, val);
|
|
|
|
|
sample_to_timespec(timer->it_clock,
|
|
|
|
|
old_expires,
|
|
|
|
|
&old->it_value);
|
|
|
|
|
} else {
|
|
|
|
|
old->it_value.tv_nsec = 1;
|
|
|
|
|
old->it_value.tv_sec = 0;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2005-10-24 21:29:58 +07:00
|
|
|
if (unlikely(ret)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* We are colliding with the timer actually firing.
|
|
|
|
|
* Punt after filling in the timer's old value, and
|
|
|
|
|
* disable this firing since we are already reporting
|
|
|
|
|
* it as an overrun (thanks to bump_cpu_timer above).
|
|
|
|
|
*/
|
2010-03-12 05:04:42 +07:00
|
|
|
spin_unlock(&p->sighand->siglock);
|
2005-04-17 05:20:36 +07:00
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) {
|
|
|
|
|
cpu_time_add(timer->it_clock, &new_expires, val);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Install the new expiry time (or zero).
|
|
|
|
|
* For a timer with no notification action, we don't actually
|
|
|
|
|
* arm the timer (we'll just fake it for timer_gettime).
|
|
|
|
|
*/
|
|
|
|
|
timer->it.cpu.expires = new_expires;
|
|
|
|
|
if (new_expires.sched != 0 &&
|
|
|
|
|
cpu_time_before(timer->it_clock, val, new_expires)) {
|
2010-03-12 05:04:38 +07:00
|
|
|
arm_timer(timer);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
2010-03-12 05:04:42 +07:00
|
|
|
spin_unlock(&p->sighand->siglock);
|
2005-04-17 05:20:36 +07:00
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Install the new reload setting, and
|
|
|
|
|
* set up the signal and overrun bookkeeping.
|
|
|
|
|
*/
|
|
|
|
|
timer->it.cpu.incr = timespec_to_sample(timer->it_clock,
|
|
|
|
|
&new->it_interval);
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This acts as a modification timestamp for the timer,
|
|
|
|
|
* so any automatic reload attempt will punt on seeing
|
|
|
|
|
* that we have reset the timer manually.
|
|
|
|
|
*/
|
|
|
|
|
timer->it_requeue_pending = (timer->it_requeue_pending + 2) &
|
|
|
|
|
~REQUEUE_PENDING;
|
|
|
|
|
timer->it_overrun_last = 0;
|
|
|
|
|
timer->it_overrun = -1;
|
|
|
|
|
|
|
|
|
|
if (new_expires.sched != 0 &&
|
|
|
|
|
!cpu_time_before(timer->it_clock, val, new_expires)) {
|
|
|
|
|
/*
|
|
|
|
|
* The designated time already passed, so we notify
|
|
|
|
|
* immediately, even if the thread never runs to
|
|
|
|
|
* accumulate more time on this clock.
|
|
|
|
|
*/
|
|
|
|
|
cpu_timer_fire(timer);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ret = 0;
|
|
|
|
|
out:
|
|
|
|
|
if (old) {
|
|
|
|
|
sample_to_timespec(timer->it_clock,
|
cpu-timers: Return correct previous timer reload value
According POSIX we need to correctly set old timer it_interval value when
user request that in timer_settime(). Tested using below program.
#include <sys/time.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <unistd.h>
#include <assert.h>
int main(void)
{
struct sigaction act;
struct sigevent evt = { };
timer_t tid;
struct itimerspec spec, u_spec, k_spec;
evt.sigev_notify = SIGEV_SIGNAL;
evt.sigev_signo = SIGPROF;
assert(timer_create(CLOCK_PROCESS_CPUTIME_ID, &evt, &tid) == 0);
spec.it_value.tv_sec = 1;
spec.it_value.tv_nsec = 2;
spec.it_interval.tv_sec = 3;
spec.it_interval.tv_nsec = 4;
u_spec = spec;
assert(timer_settime(tid, 0, &spec, NULL) == 0);
spec.it_value.tv_sec = 5;
spec.it_value.tv_nsec = 6;
spec.it_interval.tv_sec = 7;
spec.it_interval.tv_nsec = 8;
assert(timer_settime(tid, 0, &spec, &k_spec) == 0);
#define PRT(val) printf(#val ":\t%d/%d\n", (int) u_spec.val, (int) k_spec.val)
PRT(it_value.tv_sec);
PRT(it_value.tv_nsec);
PRT(it_interval.tv_sec);
PRT(it_interval.tv_nsec);
return 0;
}
Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-03-12 05:04:39 +07:00
|
|
|
old_incr, &old->it_interval);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static void posix_cpu_timer_get(struct k_itimer *timer, struct itimerspec *itp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
union cpu_time_count now;
|
|
|
|
|
struct task_struct *p = timer->it.cpu.task;
|
|
|
|
|
int clear_dead;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Easy part: convert the reload time.
|
|
|
|
|
*/
|
|
|
|
|
sample_to_timespec(timer->it_clock,
|
|
|
|
|
timer->it.cpu.incr, &itp->it_interval);
|
|
|
|
|
|
|
|
|
|
if (timer->it.cpu.expires.sched == 0) { /* Timer not armed at all. */
|
|
|
|
|
itp->it_value.tv_sec = itp->it_value.tv_nsec = 0;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (unlikely(p == NULL)) {
|
|
|
|
|
/*
|
|
|
|
|
* This task already died and the timer will never fire.
|
|
|
|
|
* In this case, expires is actually the dead value.
|
|
|
|
|
*/
|
|
|
|
|
dead:
|
|
|
|
|
sample_to_timespec(timer->it_clock, timer->it.cpu.expires,
|
|
|
|
|
&itp->it_value);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Sample the clock to take the difference with the expiry time.
|
|
|
|
|
*/
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
|
|
|
|
|
cpu_clock_sample(timer->it_clock, p, &now);
|
|
|
|
|
clear_dead = p->exit_state;
|
|
|
|
|
} else {
|
|
|
|
|
read_lock(&tasklist_lock);
|
2010-05-27 04:43:13 +07:00
|
|
|
if (unlikely(p->sighand == NULL)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* The process has been reaped.
|
|
|
|
|
* We can't even collect a sample any more.
|
|
|
|
|
* Call the timer disarmed, nothing else to do.
|
|
|
|
|
*/
|
|
|
|
|
put_task_struct(p);
|
|
|
|
|
timer->it.cpu.task = NULL;
|
|
|
|
|
timer->it.cpu.expires.sched = 0;
|
|
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
goto dead;
|
|
|
|
|
} else {
|
2009-02-12 21:00:52 +07:00
|
|
|
cpu_timer_sample_group(timer->it_clock, p, &now);
|
2005-04-17 05:20:36 +07:00
|
|
|
clear_dead = (unlikely(p->exit_state) &&
|
|
|
|
|
thread_group_empty(p));
|
|
|
|
|
}
|
|
|
|
|
read_unlock(&tasklist_lock);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (unlikely(clear_dead)) {
|
|
|
|
|
/*
|
|
|
|
|
* We've noticed that the thread is dead, but
|
|
|
|
|
* not yet reaped. Take this opportunity to
|
|
|
|
|
* drop our task ref.
|
|
|
|
|
*/
|
|
|
|
|
clear_dead_task(timer, now);
|
|
|
|
|
goto dead;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (cpu_time_before(timer->it_clock, now, timer->it.cpu.expires)) {
|
|
|
|
|
sample_to_timespec(timer->it_clock,
|
|
|
|
|
cpu_time_sub(timer->it_clock,
|
|
|
|
|
timer->it.cpu.expires, now),
|
|
|
|
|
&itp->it_value);
|
|
|
|
|
} else {
|
|
|
|
|
/*
|
|
|
|
|
* The timer should have expired already, but the firing
|
|
|
|
|
* hasn't taken place yet. Say it's just about to expire.
|
|
|
|
|
*/
|
|
|
|
|
itp->it_value.tv_nsec = 1;
|
|
|
|
|
itp->it_value.tv_sec = 0;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check for any per-thread CPU timers that have fired and move them off
|
|
|
|
|
* the tsk->cpu_timers[N] list onto the firing list. Here we update the
|
|
|
|
|
* tsk->it_*_expires values to reflect the remaining thread CPU timers.
|
|
|
|
|
*/
|
|
|
|
|
static void check_thread_timers(struct task_struct *tsk,
|
|
|
|
|
struct list_head *firing)
|
|
|
|
|
{
|
2005-10-24 00:02:50 +07:00
|
|
|
int maxfire;
|
2005-04-17 05:20:36 +07:00
|
|
|
struct list_head *timers = tsk->cpu_timers;
|
2008-01-26 03:08:27 +07:00
|
|
|
struct signal_struct *const sig = tsk->signal;
|
2010-03-06 04:42:53 +07:00
|
|
|
unsigned long soft;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.prof_exp = cputime_zero;
|
2005-04-17 05:20:36 +07:00
|
|
|
while (!list_empty(timers)) {
|
Introduce a handy list_first_entry macro
There are many places in the kernel where the construction like
foo = list_entry(head->next, struct foo_struct, list);
are used.
The code might look more descriptive and neat if using the macro
list_first_entry(head, type, member) \
list_entry((head)->next, type, member)
Here is the macro itself and the examples of its usage in the generic code.
If it will turn out to be useful, I can prepare the set of patches to
inject in into arch-specific code, drivers, networking, etc.
Signed-off-by: Pavel Emelianov <xemul@openvz.org>
Signed-off-by: Kirill Korotaev <dev@openvz.org>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: John McCutchan <ttb@tentacle.dhs.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Ram Pai <linuxram@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-08 14:30:19 +07:00
|
|
|
struct cpu_timer_list *t = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2005-10-24 00:02:50 +07:00
|
|
|
if (!--maxfire || cputime_lt(prof_ticks(tsk), t->expires.cpu)) {
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.prof_exp = t->expires.cpu;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
t->firing = 1;
|
|
|
|
|
list_move_tail(&t->entry, firing);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++timers;
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.virt_exp = cputime_zero;
|
2005-04-17 05:20:36 +07:00
|
|
|
while (!list_empty(timers)) {
|
Introduce a handy list_first_entry macro
There are many places in the kernel where the construction like
foo = list_entry(head->next, struct foo_struct, list);
are used.
The code might look more descriptive and neat if using the macro
list_first_entry(head, type, member) \
list_entry((head)->next, type, member)
Here is the macro itself and the examples of its usage in the generic code.
If it will turn out to be useful, I can prepare the set of patches to
inject in into arch-specific code, drivers, networking, etc.
Signed-off-by: Pavel Emelianov <xemul@openvz.org>
Signed-off-by: Kirill Korotaev <dev@openvz.org>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: John McCutchan <ttb@tentacle.dhs.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Ram Pai <linuxram@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-08 14:30:19 +07:00
|
|
|
struct cpu_timer_list *t = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2005-10-24 00:02:50 +07:00
|
|
|
if (!--maxfire || cputime_lt(virt_ticks(tsk), t->expires.cpu)) {
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.virt_exp = t->expires.cpu;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
t->firing = 1;
|
|
|
|
|
list_move_tail(&t->entry, firing);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++timers;
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.sched_exp = 0;
|
2005-04-17 05:20:36 +07:00
|
|
|
while (!list_empty(timers)) {
|
Introduce a handy list_first_entry macro
There are many places in the kernel where the construction like
foo = list_entry(head->next, struct foo_struct, list);
are used.
The code might look more descriptive and neat if using the macro
list_first_entry(head, type, member) \
list_entry((head)->next, type, member)
Here is the macro itself and the examples of its usage in the generic code.
If it will turn out to be useful, I can prepare the set of patches to
inject in into arch-specific code, drivers, networking, etc.
Signed-off-by: Pavel Emelianov <xemul@openvz.org>
Signed-off-by: Kirill Korotaev <dev@openvz.org>
Cc: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: John McCutchan <ttb@tentacle.dhs.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Ram Pai <linuxram@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-08 14:30:19 +07:00
|
|
|
struct cpu_timer_list *t = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2007-07-09 23:51:58 +07:00
|
|
|
if (!--maxfire || tsk->se.sum_exec_runtime < t->expires.sched) {
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->cputime_expires.sched_exp = t->expires.sched;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
t->firing = 1;
|
|
|
|
|
list_move_tail(&t->entry, firing);
|
|
|
|
|
}
|
2008-01-26 03:08:27 +07:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check for the special case thread timers.
|
|
|
|
|
*/
|
2010-03-06 04:42:54 +07:00
|
|
|
soft = ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_cur);
|
2010-03-06 04:42:53 +07:00
|
|
|
if (soft != RLIM_INFINITY) {
|
2010-03-06 04:42:54 +07:00
|
|
|
unsigned long hard =
|
|
|
|
|
ACCESS_ONCE(sig->rlim[RLIMIT_RTTIME].rlim_max);
|
2008-01-26 03:08:27 +07:00
|
|
|
|
2008-01-26 03:08:32 +07:00
|
|
|
if (hard != RLIM_INFINITY &&
|
|
|
|
|
tsk->rt.timeout > DIV_ROUND_UP(hard, USEC_PER_SEC/HZ)) {
|
2008-01-26 03:08:27 +07:00
|
|
|
/*
|
|
|
|
|
* At the hard limit, we just die.
|
|
|
|
|
* No need to calculate anything else now.
|
|
|
|
|
*/
|
|
|
|
|
__group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2010-03-06 04:42:53 +07:00
|
|
|
if (tsk->rt.timeout > DIV_ROUND_UP(soft, USEC_PER_SEC/HZ)) {
|
2008-01-26 03:08:27 +07:00
|
|
|
/*
|
|
|
|
|
* At the soft limit, send a SIGXCPU every second.
|
|
|
|
|
*/
|
2010-03-06 04:42:53 +07:00
|
|
|
if (soft < hard) {
|
|
|
|
|
soft += USEC_PER_SEC;
|
|
|
|
|
sig->rlim[RLIMIT_RTTIME].rlim_cur = soft;
|
2008-01-26 03:08:27 +07:00
|
|
|
}
|
2008-05-16 09:42:49 +07:00
|
|
|
printk(KERN_INFO
|
|
|
|
|
"RT Watchdog Timeout: %s[%d]\n",
|
|
|
|
|
tsk->comm, task_pid_nr(tsk));
|
2008-01-26 03:08:27 +07:00
|
|
|
__group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
|
|
|
|
|
}
|
|
|
|
|
}
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
2010-03-12 05:04:31 +07:00
|
|
|
static void stop_process_timers(struct signal_struct *sig)
|
2009-02-10 22:37:31 +07:00
|
|
|
{
|
2010-03-12 05:04:31 +07:00
|
|
|
struct thread_group_cputimer *cputimer = &sig->cputimer;
|
2009-02-10 22:37:31 +07:00
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
|
|
spin_lock_irqsave(&cputimer->lock, flags);
|
|
|
|
|
cputimer->running = 0;
|
|
|
|
|
spin_unlock_irqrestore(&cputimer->lock, flags);
|
|
|
|
|
}
|
|
|
|
|
|
2009-07-29 17:15:27 +07:00
|
|
|
static u32 onecputick;
|
|
|
|
|
|
2009-07-29 17:15:26 +07:00
|
|
|
static void check_cpu_itimer(struct task_struct *tsk, struct cpu_itimer *it,
|
|
|
|
|
cputime_t *expires, cputime_t cur_time, int signo)
|
|
|
|
|
{
|
|
|
|
|
if (cputime_eq(it->expires, cputime_zero))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (cputime_ge(cur_time, it->expires)) {
|
2009-07-29 17:15:27 +07:00
|
|
|
if (!cputime_eq(it->incr, cputime_zero)) {
|
|
|
|
|
it->expires = cputime_add(it->expires, it->incr);
|
|
|
|
|
it->error += it->incr_error;
|
|
|
|
|
if (it->error >= onecputick) {
|
|
|
|
|
it->expires = cputime_sub(it->expires,
|
2009-07-29 17:15:29 +07:00
|
|
|
cputime_one_jiffy);
|
2009-07-29 17:15:27 +07:00
|
|
|
it->error -= onecputick;
|
|
|
|
|
}
|
2009-08-10 09:52:30 +07:00
|
|
|
} else {
|
2009-07-29 17:15:27 +07:00
|
|
|
it->expires = cputime_zero;
|
2009-08-10 09:52:30 +07:00
|
|
|
}
|
2009-07-29 17:15:26 +07:00
|
|
|
|
2009-08-10 09:52:30 +07:00
|
|
|
trace_itimer_expire(signo == SIGPROF ?
|
|
|
|
|
ITIMER_PROF : ITIMER_VIRTUAL,
|
|
|
|
|
tsk->signal->leader_pid, cur_time);
|
2009-07-29 17:15:26 +07:00
|
|
|
__group_send_sig_info(signo, SEND_SIG_PRIV, tsk);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!cputime_eq(it->expires, cputime_zero) &&
|
|
|
|
|
(cputime_eq(*expires, cputime_zero) ||
|
|
|
|
|
cputime_lt(it->expires, *expires))) {
|
|
|
|
|
*expires = it->expires;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-28 04:12:15 +07:00
|
|
|
/**
|
|
|
|
|
* task_cputime_zero - Check a task_cputime struct for all zero fields.
|
|
|
|
|
*
|
|
|
|
|
* @cputime: The struct to compare.
|
|
|
|
|
*
|
|
|
|
|
* Checks @cputime to see if all fields are zero. Returns true if all fields
|
|
|
|
|
* are zero, false if any field is nonzero.
|
|
|
|
|
*/
|
|
|
|
|
static inline int task_cputime_zero(const struct task_cputime *cputime)
|
|
|
|
|
{
|
|
|
|
|
if (cputime_eq(cputime->utime, cputime_zero) &&
|
|
|
|
|
cputime_eq(cputime->stime, cputime_zero) &&
|
|
|
|
|
cputime->sum_exec_runtime == 0)
|
|
|
|
|
return 1;
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* Check for any per-thread CPU timers that have fired and move them
|
|
|
|
|
* off the tsk->*_timers list onto the firing list. Per-thread timers
|
|
|
|
|
* have already been taken off.
|
|
|
|
|
*/
|
|
|
|
|
static void check_process_timers(struct task_struct *tsk,
|
|
|
|
|
struct list_head *firing)
|
|
|
|
|
{
|
2005-10-24 00:02:50 +07:00
|
|
|
int maxfire;
|
2005-04-17 05:20:36 +07:00
|
|
|
struct signal_struct *const sig = tsk->signal;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
cputime_t utime, ptime, virt_expires, prof_expires;
|
2007-07-09 23:51:58 +07:00
|
|
|
unsigned long long sum_sched_runtime, sched_expires;
|
2005-04-17 05:20:36 +07:00
|
|
|
struct list_head *timers = sig->cpu_timers;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
struct task_cputime cputime;
|
2010-03-06 04:42:53 +07:00
|
|
|
unsigned long soft;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Collect the current process totals.
|
|
|
|
|
*/
|
2009-02-05 18:24:16 +07:00
|
|
|
thread_group_cputimer(tsk, &cputime);
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
utime = cputime.utime;
|
|
|
|
|
ptime = cputime_add(utime, cputime.stime);
|
|
|
|
|
sum_sched_runtime = cputime.sum_exec_runtime;
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
2005-04-17 05:20:36 +07:00
|
|
|
prof_expires = cputime_zero;
|
|
|
|
|
while (!list_empty(timers)) {
|
2008-04-05 01:54:10 +07:00
|
|
|
struct cpu_timer_list *tl = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2008-04-05 01:54:10 +07:00
|
|
|
if (!--maxfire || cputime_lt(ptime, tl->expires.cpu)) {
|
|
|
|
|
prof_expires = tl->expires.cpu;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
2008-04-05 01:54:10 +07:00
|
|
|
tl->firing = 1;
|
|
|
|
|
list_move_tail(&tl->entry, firing);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++timers;
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
2005-04-17 05:20:36 +07:00
|
|
|
virt_expires = cputime_zero;
|
|
|
|
|
while (!list_empty(timers)) {
|
2008-04-05 01:54:10 +07:00
|
|
|
struct cpu_timer_list *tl = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2008-04-05 01:54:10 +07:00
|
|
|
if (!--maxfire || cputime_lt(utime, tl->expires.cpu)) {
|
|
|
|
|
virt_expires = tl->expires.cpu;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
2008-04-05 01:54:10 +07:00
|
|
|
tl->firing = 1;
|
|
|
|
|
list_move_tail(&tl->entry, firing);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
++timers;
|
2005-10-24 00:02:50 +07:00
|
|
|
maxfire = 20;
|
2005-04-17 05:20:36 +07:00
|
|
|
sched_expires = 0;
|
|
|
|
|
while (!list_empty(timers)) {
|
2008-04-05 01:54:10 +07:00
|
|
|
struct cpu_timer_list *tl = list_first_entry(timers,
|
2005-04-17 05:20:36 +07:00
|
|
|
struct cpu_timer_list,
|
|
|
|
|
entry);
|
2008-04-05 01:54:10 +07:00
|
|
|
if (!--maxfire || sum_sched_runtime < tl->expires.sched) {
|
|
|
|
|
sched_expires = tl->expires.sched;
|
2005-04-17 05:20:36 +07:00
|
|
|
break;
|
|
|
|
|
}
|
2008-04-05 01:54:10 +07:00
|
|
|
tl->firing = 1;
|
|
|
|
|
list_move_tail(&tl->entry, firing);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check for the special case process timers.
|
|
|
|
|
*/
|
2009-07-29 17:15:26 +07:00
|
|
|
check_cpu_itimer(tsk, &sig->it[CPUCLOCK_PROF], &prof_expires, ptime,
|
|
|
|
|
SIGPROF);
|
|
|
|
|
check_cpu_itimer(tsk, &sig->it[CPUCLOCK_VIRT], &virt_expires, utime,
|
|
|
|
|
SIGVTALRM);
|
2010-03-06 04:42:54 +07:00
|
|
|
soft = ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_cur);
|
2010-03-06 04:42:53 +07:00
|
|
|
if (soft != RLIM_INFINITY) {
|
2005-04-17 05:20:36 +07:00
|
|
|
unsigned long psecs = cputime_to_secs(ptime);
|
2010-03-06 04:42:54 +07:00
|
|
|
unsigned long hard =
|
|
|
|
|
ACCESS_ONCE(sig->rlim[RLIMIT_CPU].rlim_max);
|
2005-04-17 05:20:36 +07:00
|
|
|
cputime_t x;
|
2010-03-06 04:42:53 +07:00
|
|
|
if (psecs >= hard) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* At the hard limit, we just die.
|
|
|
|
|
* No need to calculate anything else now.
|
|
|
|
|
*/
|
|
|
|
|
__group_send_sig_info(SIGKILL, SEND_SIG_PRIV, tsk);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2010-03-06 04:42:53 +07:00
|
|
|
if (psecs >= soft) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* At the soft limit, send a SIGXCPU every second.
|
|
|
|
|
*/
|
|
|
|
|
__group_send_sig_info(SIGXCPU, SEND_SIG_PRIV, tsk);
|
2010-03-06 04:42:53 +07:00
|
|
|
if (soft < hard) {
|
|
|
|
|
soft++;
|
|
|
|
|
sig->rlim[RLIMIT_CPU].rlim_cur = soft;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
}
|
2010-03-06 04:42:53 +07:00
|
|
|
x = secs_to_cputime(soft);
|
2005-04-17 05:20:36 +07:00
|
|
|
if (cputime_eq(prof_expires, cputime_zero) ||
|
|
|
|
|
cputime_lt(x, prof_expires)) {
|
|
|
|
|
prof_expires = x;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-28 04:12:15 +07:00
|
|
|
sig->cputime_expires.prof_exp = prof_expires;
|
|
|
|
|
sig->cputime_expires.virt_exp = virt_expires;
|
|
|
|
|
sig->cputime_expires.sched_exp = sched_expires;
|
|
|
|
|
if (task_cputime_zero(&sig->cputime_expires))
|
|
|
|
|
stop_process_timers(sig);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This is called from the signal code (via do_schedule_next_timer)
|
|
|
|
|
* when the last timer signal was delivered and we have to reload the timer.
|
|
|
|
|
*/
|
|
|
|
|
void posix_cpu_timer_schedule(struct k_itimer *timer)
|
|
|
|
|
{
|
|
|
|
|
struct task_struct *p = timer->it.cpu.task;
|
|
|
|
|
union cpu_time_count now;
|
|
|
|
|
|
|
|
|
|
if (unlikely(p == NULL))
|
|
|
|
|
/*
|
|
|
|
|
* The task was cleaned up already, no future firings.
|
|
|
|
|
*/
|
2005-10-31 06:03:13 +07:00
|
|
|
goto out;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Fetch the current sample and update the timer's expiry time.
|
|
|
|
|
*/
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(timer->it_clock)) {
|
|
|
|
|
cpu_clock_sample(timer->it_clock, p, &now);
|
|
|
|
|
bump_cpu_timer(timer, now);
|
|
|
|
|
if (unlikely(p->exit_state)) {
|
|
|
|
|
clear_dead_task(timer, now);
|
2005-10-31 06:03:13 +07:00
|
|
|
goto out;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
read_lock(&tasklist_lock); /* arm_timer needs it. */
|
2010-03-12 05:04:42 +07:00
|
|
|
spin_lock(&p->sighand->siglock);
|
2005-04-17 05:20:36 +07:00
|
|
|
} else {
|
|
|
|
|
read_lock(&tasklist_lock);
|
2010-05-27 04:43:13 +07:00
|
|
|
if (unlikely(p->sighand == NULL)) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* The process has been reaped.
|
|
|
|
|
* We can't even collect a sample any more.
|
|
|
|
|
*/
|
|
|
|
|
put_task_struct(p);
|
|
|
|
|
timer->it.cpu.task = p = NULL;
|
|
|
|
|
timer->it.cpu.expires.sched = 0;
|
2005-10-31 06:03:13 +07:00
|
|
|
goto out_unlock;
|
2005-04-17 05:20:36 +07:00
|
|
|
} else if (unlikely(p->exit_state) && thread_group_empty(p)) {
|
|
|
|
|
/*
|
|
|
|
|
* We've noticed that the thread is dead, but
|
|
|
|
|
* not yet reaped. Take this opportunity to
|
|
|
|
|
* drop our task ref.
|
|
|
|
|
*/
|
|
|
|
|
clear_dead_task(timer, now);
|
2005-10-31 06:03:13 +07:00
|
|
|
goto out_unlock;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
2010-03-12 05:04:42 +07:00
|
|
|
spin_lock(&p->sighand->siglock);
|
2009-02-12 21:00:52 +07:00
|
|
|
cpu_timer_sample_group(timer->it_clock, p, &now);
|
2005-04-17 05:20:36 +07:00
|
|
|
bump_cpu_timer(timer, now);
|
|
|
|
|
/* Leave the tasklist_lock locked for the call below. */
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Now re-arm for the new expiry time.
|
|
|
|
|
*/
|
2010-03-12 05:04:42 +07:00
|
|
|
BUG_ON(!irqs_disabled());
|
2010-03-12 05:04:38 +07:00
|
|
|
arm_timer(timer);
|
2010-03-12 05:04:42 +07:00
|
|
|
spin_unlock(&p->sighand->siglock);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2005-10-31 06:03:13 +07:00
|
|
|
out_unlock:
|
2005-04-17 05:20:36 +07:00
|
|
|
read_unlock(&tasklist_lock);
|
2005-10-31 06:03:13 +07:00
|
|
|
|
|
|
|
|
out:
|
|
|
|
|
timer->it_overrun_last = timer->it_overrun;
|
|
|
|
|
timer->it_overrun = -1;
|
|
|
|
|
++timer->it_requeue_pending;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
/**
|
|
|
|
|
* task_cputime_expired - Compare two task_cputime entities.
|
|
|
|
|
*
|
|
|
|
|
* @sample: The task_cputime structure to be checked for expiration.
|
|
|
|
|
* @expires: Expiration times, against which @sample will be checked.
|
|
|
|
|
*
|
|
|
|
|
* Checks @sample against @expires to see if any field of @sample has expired.
|
|
|
|
|
* Returns true if any field of the former is greater than the corresponding
|
|
|
|
|
* field of the latter if the latter field is set. Otherwise returns false.
|
|
|
|
|
*/
|
|
|
|
|
static inline int task_cputime_expired(const struct task_cputime *sample,
|
|
|
|
|
const struct task_cputime *expires)
|
|
|
|
|
{
|
|
|
|
|
if (!cputime_eq(expires->utime, cputime_zero) &&
|
|
|
|
|
cputime_ge(sample->utime, expires->utime))
|
|
|
|
|
return 1;
|
|
|
|
|
if (!cputime_eq(expires->stime, cputime_zero) &&
|
|
|
|
|
cputime_ge(cputime_add(sample->utime, sample->stime),
|
|
|
|
|
expires->stime))
|
|
|
|
|
return 1;
|
|
|
|
|
if (expires->sum_exec_runtime != 0 &&
|
|
|
|
|
sample->sum_exec_runtime >= expires->sum_exec_runtime)
|
|
|
|
|
return 1;
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* fastpath_timer_check - POSIX CPU timers fast path.
|
|
|
|
|
*
|
|
|
|
|
* @tsk: The task (thread) being checked.
|
|
|
|
|
*
|
2008-09-12 23:54:39 +07:00
|
|
|
* Check the task and thread group timers. If both are zero (there are no
|
|
|
|
|
* timers set) return false. Otherwise snapshot the task and thread group
|
|
|
|
|
* timers and compare them with the corresponding expiration times. Return
|
|
|
|
|
* true if a timer has expired, else return false.
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
*/
|
2008-09-12 23:54:39 +07:00
|
|
|
static inline int fastpath_timer_check(struct task_struct *tsk)
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
{
|
2008-11-17 21:39:47 +07:00
|
|
|
struct signal_struct *sig;
|
2008-09-12 23:54:39 +07:00
|
|
|
|
|
|
|
|
if (!task_cputime_zero(&tsk->cputime_expires)) {
|
|
|
|
|
struct task_cputime task_sample = {
|
|
|
|
|
.utime = tsk->utime,
|
|
|
|
|
.stime = tsk->stime,
|
|
|
|
|
.sum_exec_runtime = tsk->se.sum_exec_runtime
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
if (task_cputime_expired(&task_sample, &tsk->cputime_expires))
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
2008-11-17 21:39:47 +07:00
|
|
|
|
|
|
|
|
sig = tsk->signal;
|
2010-04-28 04:12:15 +07:00
|
|
|
if (sig->cputimer.running) {
|
2008-09-12 23:54:39 +07:00
|
|
|
struct task_cputime group_sample;
|
|
|
|
|
|
2010-06-12 01:04:46 +07:00
|
|
|
spin_lock(&sig->cputimer.lock);
|
|
|
|
|
group_sample = sig->cputimer.cputime;
|
|
|
|
|
spin_unlock(&sig->cputimer.lock);
|
|
|
|
|
|
2008-09-12 23:54:39 +07:00
|
|
|
if (task_cputime_expired(&group_sample, &sig->cputime_expires))
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
2009-03-24 02:34:11 +07:00
|
|
|
|
2010-03-12 05:04:37 +07:00
|
|
|
return 0;
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
}
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* This is called from the timer interrupt handler. The irq handler has
|
|
|
|
|
* already updated our counts. We need to check if any timers fire now.
|
|
|
|
|
* Interrupts are disabled.
|
|
|
|
|
*/
|
|
|
|
|
void run_posix_cpu_timers(struct task_struct *tsk)
|
|
|
|
|
{
|
|
|
|
|
LIST_HEAD(firing);
|
|
|
|
|
struct k_itimer *timer, *next;
|
2010-06-11 06:10:18 +07:00
|
|
|
unsigned long flags;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
BUG_ON(!irqs_disabled());
|
|
|
|
|
|
|
|
|
|
/*
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
* The fast path checks that there are no expired thread or thread
|
2008-09-12 23:54:39 +07:00
|
|
|
* group timers. If that's so, just return.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
2008-09-12 23:54:39 +07:00
|
|
|
if (!fastpath_timer_check(tsk))
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
return;
|
2008-09-14 22:11:46 +07:00
|
|
|
|
2010-06-11 06:10:18 +07:00
|
|
|
if (!lock_task_sighand(tsk, &flags))
|
|
|
|
|
return;
|
2008-09-12 23:54:39 +07:00
|
|
|
/*
|
|
|
|
|
* Here we take off tsk->signal->cpu_timers[N] and
|
|
|
|
|
* tsk->cpu_timers[N] all the timers that are firing, and
|
|
|
|
|
* put them on the firing list.
|
|
|
|
|
*/
|
|
|
|
|
check_thread_timers(tsk, &firing);
|
2010-04-28 04:12:15 +07:00
|
|
|
/*
|
|
|
|
|
* If there are any active process wide timers (POSIX 1.b, itimers,
|
|
|
|
|
* RLIMIT_CPU) cputimer must be running.
|
|
|
|
|
*/
|
|
|
|
|
if (tsk->signal->cputimer.running)
|
|
|
|
|
check_process_timers(tsk, &firing);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2008-09-12 23:54:39 +07:00
|
|
|
/*
|
|
|
|
|
* We must release these locks before taking any timer's lock.
|
|
|
|
|
* There is a potential race with timer deletion here, as the
|
|
|
|
|
* siglock now protects our private firing list. We have set
|
|
|
|
|
* the firing flag in each timer, so that a deletion attempt
|
|
|
|
|
* that gets the timer lock before we do will give it up and
|
|
|
|
|
* spin until we've taken care of that timer below.
|
|
|
|
|
*/
|
2010-06-11 06:10:18 +07:00
|
|
|
unlock_task_sighand(tsk, &flags);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Now that all the timers on our list have the firing flag,
|
|
|
|
|
* noone will touch their list entries but us. We'll take
|
|
|
|
|
* each timer's lock before clearing its firing flag, so no
|
|
|
|
|
* timer call will interfere.
|
|
|
|
|
*/
|
|
|
|
|
list_for_each_entry_safe(timer, next, &firing, it.cpu.entry) {
|
2009-04-30 06:14:32 +07:00
|
|
|
int cpu_firing;
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
spin_lock(&timer->it_lock);
|
|
|
|
|
list_del_init(&timer->it.cpu.entry);
|
2009-04-30 06:14:32 +07:00
|
|
|
cpu_firing = timer->it.cpu.firing;
|
2005-04-17 05:20:36 +07:00
|
|
|
timer->it.cpu.firing = 0;
|
|
|
|
|
/*
|
|
|
|
|
* The firing flag is -1 if we collided with a reset
|
|
|
|
|
* of the timer, which already reported this
|
|
|
|
|
* almost-firing as an overrun. So don't generate an event.
|
|
|
|
|
*/
|
2009-04-30 06:14:32 +07:00
|
|
|
if (likely(cpu_firing >= 0))
|
2005-04-17 05:20:36 +07:00
|
|
|
cpu_timer_fire(timer);
|
|
|
|
|
spin_unlock(&timer->it_lock);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2010-03-12 05:04:37 +07:00
|
|
|
* Set one of the process-wide special case CPU timers or RLIMIT_CPU.
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
* The tsk->sighand->siglock must be held by the caller.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
|
|
|
|
void set_process_cpu_timer(struct task_struct *tsk, unsigned int clock_idx,
|
|
|
|
|
cputime_t *newval, cputime_t *oldval)
|
|
|
|
|
{
|
|
|
|
|
union cpu_time_count now;
|
|
|
|
|
|
|
|
|
|
BUG_ON(clock_idx == CPUCLOCK_SCHED);
|
2009-02-05 18:24:16 +07:00
|
|
|
cpu_timer_sample_group(clock_idx, tsk, &now);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
if (oldval) {
|
2010-03-12 05:04:37 +07:00
|
|
|
/*
|
|
|
|
|
* We are setting itimer. The *oldval is absolute and we update
|
|
|
|
|
* it to be relative, *newval argument is relative and we update
|
|
|
|
|
* it to be absolute.
|
|
|
|
|
*/
|
2005-04-17 05:20:36 +07:00
|
|
|
if (!cputime_eq(*oldval, cputime_zero)) {
|
|
|
|
|
if (cputime_le(*oldval, now.cpu)) {
|
|
|
|
|
/* Just about to fire. */
|
2009-07-29 17:15:29 +07:00
|
|
|
*oldval = cputime_one_jiffy;
|
2005-04-17 05:20:36 +07:00
|
|
|
} else {
|
|
|
|
|
*oldval = cputime_sub(*oldval, now.cpu);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (cputime_eq(*newval, cputime_zero))
|
|
|
|
|
return;
|
|
|
|
|
*newval = cputime_add(*newval, now.cpu);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2010-03-12 05:04:37 +07:00
|
|
|
* Update expiration cache if we are the earliest timer, or eventually
|
|
|
|
|
* RLIMIT_CPU limit is earlier than prof_exp cpu timer expire.
|
2005-04-17 05:20:36 +07:00
|
|
|
*/
|
2010-03-12 05:04:37 +07:00
|
|
|
switch (clock_idx) {
|
|
|
|
|
case CPUCLOCK_PROF:
|
|
|
|
|
if (expires_gt(tsk->signal->cputime_expires.prof_exp, *newval))
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->signal->cputime_expires.prof_exp = *newval;
|
2010-03-12 05:04:37 +07:00
|
|
|
break;
|
|
|
|
|
case CPUCLOCK_VIRT:
|
|
|
|
|
if (expires_gt(tsk->signal->cputime_expires.virt_exp, *newval))
|
timers: fix itimer/many thread hang
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 23:54:39 +07:00
|
|
|
tsk->signal->cputime_expires.virt_exp = *newval;
|
2010-03-12 05:04:37 +07:00
|
|
|
break;
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2006-09-29 16:00:29 +07:00
|
|
|
static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
|
|
|
|
|
struct timespec *rqtp, struct itimerspec *it)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
struct k_itimer timer;
|
|
|
|
|
int error;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Set up a temporary timer and then wait for it to go off.
|
|
|
|
|
*/
|
|
|
|
|
memset(&timer, 0, sizeof timer);
|
|
|
|
|
spin_lock_init(&timer.it_lock);
|
|
|
|
|
timer.it_clock = which_clock;
|
|
|
|
|
timer.it_overrun = -1;
|
|
|
|
|
error = posix_cpu_timer_create(&timer);
|
|
|
|
|
timer.it_process = current;
|
|
|
|
|
if (!error) {
|
|
|
|
|
static struct itimerspec zero_it;
|
2006-09-29 16:00:29 +07:00
|
|
|
|
|
|
|
|
memset(it, 0, sizeof *it);
|
|
|
|
|
it->it_value = *rqtp;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
|
|
|
|
spin_lock_irq(&timer.it_lock);
|
2006-09-29 16:00:29 +07:00
|
|
|
error = posix_cpu_timer_set(&timer, flags, it, NULL);
|
2005-04-17 05:20:36 +07:00
|
|
|
if (error) {
|
|
|
|
|
spin_unlock_irq(&timer.it_lock);
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
while (!signal_pending(current)) {
|
|
|
|
|
if (timer.it.cpu.expires.sched == 0) {
|
|
|
|
|
/*
|
|
|
|
|
* Our timer fired and was reset.
|
|
|
|
|
*/
|
|
|
|
|
spin_unlock_irq(&timer.it_lock);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Block until cpu_timer_fire (or a signal) wakes us.
|
|
|
|
|
*/
|
|
|
|
|
__set_current_state(TASK_INTERRUPTIBLE);
|
|
|
|
|
spin_unlock_irq(&timer.it_lock);
|
|
|
|
|
schedule();
|
|
|
|
|
spin_lock_irq(&timer.it_lock);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We were interrupted by a signal.
|
|
|
|
|
*/
|
|
|
|
|
sample_to_timespec(which_clock, timer.it.cpu.expires, rqtp);
|
2006-09-29 16:00:29 +07:00
|
|
|
posix_cpu_timer_set(&timer, 0, &zero_it, it);
|
2005-04-17 05:20:36 +07:00
|
|
|
spin_unlock_irq(&timer.it_lock);
|
|
|
|
|
|
2006-09-29 16:00:29 +07:00
|
|
|
if ((it->it_value.tv_sec | it->it_value.tv_nsec) == 0) {
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
|
|
|
|
* It actually did fire already.
|
|
|
|
|
*/
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2006-09-29 16:00:29 +07:00
|
|
|
error = -ERESTART_RESTARTBLOCK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static long posix_cpu_nsleep_restart(struct restart_block *restart_block);
|
|
|
|
|
|
|
|
|
|
static int posix_cpu_nsleep(const clockid_t which_clock, int flags,
|
|
|
|
|
struct timespec *rqtp, struct timespec __user *rmtp)
|
2006-09-29 16:00:29 +07:00
|
|
|
{
|
|
|
|
|
struct restart_block *restart_block =
|
2011-02-01 20:51:20 +07:00
|
|
|
¤t_thread_info()->restart_block;
|
2006-09-29 16:00:29 +07:00
|
|
|
struct itimerspec it;
|
|
|
|
|
int error;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Diagnose required errors first.
|
|
|
|
|
*/
|
|
|
|
|
if (CPUCLOCK_PERTHREAD(which_clock) &&
|
|
|
|
|
(CPUCLOCK_PID(which_clock) == 0 ||
|
|
|
|
|
CPUCLOCK_PID(which_clock) == current->pid))
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
|
|
error = do_cpu_nanosleep(which_clock, flags, rqtp, &it);
|
|
|
|
|
|
|
|
|
|
if (error == -ERESTART_RESTARTBLOCK) {
|
|
|
|
|
|
2011-02-01 20:51:20 +07:00
|
|
|
if (flags & TIMER_ABSTIME)
|
2006-09-29 16:00:29 +07:00
|
|
|
return -ERESTARTNOHAND;
|
2005-04-17 05:20:36 +07:00
|
|
|
/*
|
2011-02-01 20:51:20 +07:00
|
|
|
* Report back to the user the time still remaining.
|
|
|
|
|
*/
|
|
|
|
|
if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
|
2005-04-17 05:20:36 +07:00
|
|
|
return -EFAULT;
|
|
|
|
|
|
2006-09-29 16:00:28 +07:00
|
|
|
restart_block->fn = posix_cpu_nsleep_restart;
|
2011-02-01 20:51:20 +07:00
|
|
|
restart_block->nanosleep.index = which_clock;
|
|
|
|
|
restart_block->nanosleep.rmtp = rmtp;
|
|
|
|
|
restart_block->nanosleep.expires = timespec_to_ns(rqtp);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:52:12 +07:00
|
|
|
static long posix_cpu_nsleep_restart(struct restart_block *restart_block)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
2011-02-01 20:51:20 +07:00
|
|
|
clockid_t which_clock = restart_block->nanosleep.index;
|
2006-01-10 11:52:37 +07:00
|
|
|
struct timespec t;
|
2006-09-29 16:00:29 +07:00
|
|
|
struct itimerspec it;
|
|
|
|
|
int error;
|
2006-01-10 11:52:37 +07:00
|
|
|
|
2011-02-01 20:51:20 +07:00
|
|
|
t = ns_to_timespec(restart_block->nanosleep.expires);
|
2006-01-10 11:52:37 +07:00
|
|
|
|
2006-09-29 16:00:29 +07:00
|
|
|
error = do_cpu_nanosleep(which_clock, TIMER_ABSTIME, &t, &it);
|
|
|
|
|
|
|
|
|
|
if (error == -ERESTART_RESTARTBLOCK) {
|
2011-02-01 20:51:20 +07:00
|
|
|
struct timespec __user *rmtp = restart_block->nanosleep.rmtp;
|
2006-09-29 16:00:29 +07:00
|
|
|
/*
|
2011-02-01 20:51:20 +07:00
|
|
|
* Report back to the user the time still remaining.
|
|
|
|
|
*/
|
|
|
|
|
if (rmtp && copy_to_user(rmtp, &it.it_value, sizeof *rmtp))
|
2006-09-29 16:00:29 +07:00
|
|
|
return -EFAULT;
|
|
|
|
|
|
2011-02-01 20:51:20 +07:00
|
|
|
restart_block->nanosleep.expires = timespec_to_ns(&t);
|
2006-09-29 16:00:29 +07:00
|
|
|
}
|
|
|
|
|
return error;
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#define PROCESS_CLOCK MAKE_PROCESS_CPUCLOCK(0, CPUCLOCK_SCHED)
|
|
|
|
|
#define THREAD_CLOCK MAKE_THREAD_CPUCLOCK(0, CPUCLOCK_SCHED)
|
|
|
|
|
|
2006-01-10 11:52:27 +07:00
|
|
|
static int process_cpu_clock_getres(const clockid_t which_clock,
|
|
|
|
|
struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
return posix_cpu_clock_getres(PROCESS_CLOCK, tp);
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static int process_cpu_clock_get(const clockid_t which_clock,
|
|
|
|
|
struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
return posix_cpu_clock_get(PROCESS_CLOCK, tp);
|
|
|
|
|
}
|
|
|
|
|
static int process_cpu_timer_create(struct k_itimer *timer)
|
|
|
|
|
{
|
|
|
|
|
timer->it_clock = PROCESS_CLOCK;
|
|
|
|
|
return posix_cpu_timer_create(timer);
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static int process_cpu_nsleep(const clockid_t which_clock, int flags,
|
2006-01-10 11:52:37 +07:00
|
|
|
struct timespec *rqtp,
|
|
|
|
|
struct timespec __user *rmtp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
2006-01-10 11:52:37 +07:00
|
|
|
return posix_cpu_nsleep(PROCESS_CLOCK, flags, rqtp, rmtp);
|
2005-04-17 05:20:36 +07:00
|
|
|
}
|
2006-09-29 16:00:28 +07:00
|
|
|
static long process_cpu_nsleep_restart(struct restart_block *restart_block)
|
|
|
|
|
{
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static int thread_cpu_clock_getres(const clockid_t which_clock,
|
|
|
|
|
struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
return posix_cpu_clock_getres(THREAD_CLOCK, tp);
|
|
|
|
|
}
|
2006-01-10 11:52:27 +07:00
|
|
|
static int thread_cpu_clock_get(const clockid_t which_clock,
|
|
|
|
|
struct timespec *tp)
|
2005-04-17 05:20:36 +07:00
|
|
|
{
|
|
|
|
|
return posix_cpu_clock_get(THREAD_CLOCK, tp);
|
|
|
|
|
}
|
|
|
|
|
static int thread_cpu_timer_create(struct k_itimer *timer)
|
|
|
|
|
{
|
|
|
|
|
timer->it_clock = THREAD_CLOCK;
|
|
|
|
|
return posix_cpu_timer_create(timer);
|
|
|
|
|
}
|
|
|
|
|
|
2011-02-01 20:51:06 +07:00
|
|
|
struct k_clock clock_posix_cpu = {
|
|
|
|
|
.clock_getres = posix_cpu_clock_getres,
|
|
|
|
|
.clock_set = posix_cpu_clock_set,
|
|
|
|
|
.clock_get = posix_cpu_clock_get,
|
|
|
|
|
.timer_create = posix_cpu_timer_create,
|
|
|
|
|
.nsleep = posix_cpu_nsleep,
|
|
|
|
|
.nsleep_restart = posix_cpu_nsleep_restart,
|
|
|
|
|
.timer_set = posix_cpu_timer_set,
|
|
|
|
|
.timer_del = posix_cpu_timer_del,
|
|
|
|
|
.timer_get = posix_cpu_timer_get,
|
|
|
|
|
};
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
static __init int init_posix_cpu_timers(void)
|
|
|
|
|
{
|
|
|
|
|
struct k_clock process = {
|
2011-02-01 20:51:03 +07:00
|
|
|
.clock_getres = process_cpu_clock_getres,
|
|
|
|
|
.clock_get = process_cpu_clock_get,
|
|
|
|
|
.timer_create = process_cpu_timer_create,
|
|
|
|
|
.nsleep = process_cpu_nsleep,
|
|
|
|
|
.nsleep_restart = process_cpu_nsleep_restart,
|
2005-04-17 05:20:36 +07:00
|
|
|
};
|
|
|
|
|
struct k_clock thread = {
|
2011-02-01 20:51:03 +07:00
|
|
|
.clock_getres = thread_cpu_clock_getres,
|
|
|
|
|
.clock_get = thread_cpu_clock_get,
|
|
|
|
|
.timer_create = thread_cpu_timer_create,
|
2005-04-17 05:20:36 +07:00
|
|
|
};
|
2009-07-29 17:15:27 +07:00
|
|
|
struct timespec ts;
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2011-02-02 18:10:09 +07:00
|
|
|
posix_timers_register_clock(CLOCK_PROCESS_CPUTIME_ID, &process);
|
|
|
|
|
posix_timers_register_clock(CLOCK_THREAD_CPUTIME_ID, &thread);
|
2005-04-17 05:20:36 +07:00
|
|
|
|
2009-07-29 17:15:29 +07:00
|
|
|
cputime_to_timespec(cputime_one_jiffy, &ts);
|
2009-07-29 17:15:27 +07:00
|
|
|
onecputick = ts.tv_nsec;
|
|
|
|
|
WARN_ON(ts.tv_sec != 0);
|
|
|
|
|
|
2005-04-17 05:20:36 +07:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
__initcall(init_posix_cpu_timers);
|