The bandwidth throttle code dequeues a group when it runs out of quota, and
re-queues it once the period rolls over and the quota gets refreshed.
Sadly it failed to take the hierarchy into consideration. Share more of the
enqueue/dequeue code with regular task opterations.
Also, some operations like sched_setscheduler() can dequeue/enqueue tasks that
are in throttled runqueues, we should not inadvertly re-enqueue empty runqueues
so check for that.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Daniel K. <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Don't re-set the entity's runqueue to the wrong rq after we've set it
to the right one.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Daniel K. <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Fix this warning, which appears with !CONFIG_SMP:
kernel/sched.c:1216: warning: `init_hrtick' defined but not used
Signed-off-by: Rabin Vincent <rabin@rab.in>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Fix error checking routine to catch an error which occurs in first
__register_*probe().
Signed-off-by: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Jim Keniston <jkenisto@us.ibm.com>
Cc: David Miller <davem@davemloft.net>
Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
(overflow means weight >= 2^32 here, because inv_weigh = 2^32/weight)
A weight of a cfs_rq is the sum of weights of which entities
are queued on this cfs_rq, so it will overflow when there are
too many entities.
Although, overflow occurs very rarely, but it break fairness when
it occurs. 64-bits systems have more memory than 32-bit systems
and 64-bit systems can create more process usually, so overflow may
occur more frequently.
This patch guarantees fairness when overflow happens on 64-bit systems.
Thanks to the optimization of compiler, it changes nothing on 32-bit.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
I found a bug which can be reproduced by this way:(linux-2.6.26-rc5, x86-64)
(use 2^32, 2^33, ...., 2^63 as shares value)
# mkdir /dev/cpuctl
# mount -t cgroup -o cpu cpuctl /dev/cpuctl
# cd /dev/cpuctl
# mkdir sub
# echo 0x8000000000000000 > sub/cpu.shares
# echo $$ > sub/tasks
oops here! divide by zero.
This is because do_div() expects the 2th parameter to be 32 bits,
but unsigned long is 64 bits in x86_64.
Peter Zijstra pointed it out that the sane thing to do is limit the
shares value to something smaller instead of using an even more
expensive divide.
Also, I found another bug about "the shares value is too large":
pid1 and pid2 are set affinity to cpu#0
pid1 is attached to cg1 and pid2 is attached to cg2
if cg1/cpu.shares = 1024 cg2/cpu.shares = 2000000000
then pid2 got 100% usage of cpu, and pid1 0%
if cg1/cpu.shares = 1024 cg2/cpu.shares = 20000000000
then pid2 got 0% usage of cpu, and pid1 100%
And a weight of a cfs_rq is the sum of weights of which entities
are queued on this cfs_rq, so the shares value should be limited
to a smaller value.
I think that (1UL << 18) is a good limited value:
1) it's not too large, we can create a lot of group before overflow
2) it's several times the weight value for nice=-19 (not too small)
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
schedule() has the special "TASK_INTERRUPTIBLE && signal_pending()" case,
this allows us to do
current->state = TASK_INTERRUPTIBLE;
schedule();
without fear to sleep with pending signal.
However, the code like
current->state = TASK_KILLABLE;
schedule();
is not right, schedule() doesn't take TASK_WAKEKILL into account. This means
that mutex_lock_killable(), wait_for_completion_killable(), down_killable(),
schedule_timeout_killable() can miss SIGKILL (and btw the second SIGKILL has
no effect).
Introduce the new helper, signal_pending_state(), and change schedule() to
use it. Hopefully it will have more users, that is why the task's state is
passed separately.
Note this "__TASK_STOPPED | __TASK_TRACED" check in signal_pending_state().
This is needed to preserve the current behaviour (ptrace_notify). I hope
this check will be removed soon, but this (afaics good) change needs the
separate discussion.
The fast path is "(state & (INTERRUPTIBLE | WAKEKILL)) + signal_pending(p)",
basically the same that schedule() does now. However, this patch of course
bloats schedule().
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/chrisw/lsm-2.6:
capabilities: remain source compatible with 32-bit raw legacy capability support.
LSM: remove stale web site from MAINTAINERS
Adding a nonexistent cpu to a cpuset will be omitted quietly. It should
return -EINVAL.
Example: (real_nr_cpus <= 4 < NR_CPUS or cpu#4 was just offline)
# cat cpus
0-1
# /bin/echo 4 > cpus
# /bin/echo $?
0
# cat cpus
#
The same occurs when add a nonexistent mem.
This patch will fix this bug.
And when *buf == "", the check is unneeded.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Acked-by: Paul Jackson <pj@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jwessel/linux-2.6-kgdb:
kgdbts: Use HW breakpoints with CONFIG_DEBUG_RODATA
kgdb: use common ascii helpers and put_unaligned_be32 helper
Source code out there hard-codes a notion of what the
_LINUX_CAPABILITY_VERSION #define means in terms of the semantics of the
raw capability system calls capget() and capset(). Its unfortunate, but
true.
Since the confusing header file has been in a released kernel, there is
software that is erroneously using 64-bit capabilities with the semantics
of 32-bit compatibilities. These recently compiled programs may suffer
corruption of their memory when sys_getcap() overwrites more memory than
they are coded to expect, and the raising of added capabilities when using
sys_capset().
As such, this patch does a number of things to clean up the situation
for all. It
1. forces the _LINUX_CAPABILITY_VERSION define to always retain its
legacy value.
2. adopts a new #define strategy for the kernel's internal
implementation of the preferred magic.
3. deprecates v2 capability magic in favor of a new (v3) magic
number. The functionality of v3 is entirely equivalent to v2,
the only difference being that the v2 magic causes the kernel
to log a "deprecated" warning so the admin can find applications
that may be using v2 inappropriately.
[User space code continues to be encouraged to use the libcap API which
protects the application from details like this. libcap-2.10 is the first
to support v3 capabilities.]
Fixes issue reported in https://bugzilla.redhat.com/show_bug.cgi?id=447518.
Thanks to Bojan Smojver for the report.
[akpm@linux-foundation.org: s/depreciate/deprecate/g]
[akpm@linux-foundation.org: be robust about put_user size]
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Andrew G. Morgan <morgan@kernel.org>
Cc: Serge E. Hallyn <serue@us.ibm.com>
Cc: Bojan Smojver <bojan@rexursive.com>
Cc: stable@kernel.org
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Prevent short-running wakers of short-running threads from overloading a single
cpu via wakeup affinity, and wire up disconnected debug option.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Make sched_clock_cpu() return 0 before it has been initialized and avoid
corrupting its state due to doing so.
This fixes the weird printk timestamp jump reported.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Yanmin Zhang reported:
Comparing with 2.6.25, volanoMark has big regression with kernel 2.6.26-rc1.
It's about 50% on my 8-core stoakley, 16-core tigerton, and Itanium Montecito.
With bisect, I located the following patch:
| 18d95a2832 is first bad commit
| commit 18d95a2832
| Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
| Date: Sat Apr 19 19:45:00 2008 +0200
|
| sched: fair-group: SMP-nice for group scheduling
Revert it so that we get v2.6.25 behavior.
Bisected-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The Coverity checker spotted a memleak introduced by commit
39106dcf85 (cpumask: use new cpus_scnprintf
function).
It seems the kfree() got lost between v2 and v3 of this patch...
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Cc: Mike Travis <travis@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Splice isn't always incrementing the ppos correctly, which broke
relay splice.
Signed-off-by: Tom Zanussi <zanussi@comcast.net>
Tested-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Based on Roland's patch. This approach was suggested by Austin Clements
from the very beginning, and then by Linus.
As Austin pointed out, the execing task can be killed by SI_TIMER signal
because exec flushes the signal handlers, but doesn't discard the pending
signals generated by posix timers. Perhaps not a bug, but people find this
surprising. See http://bugzilla.kernel.org/show_bug.cgi?id=10460
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Austin Clements <amdragon+kernelbugzilla@mit.edu>
Cc: Roland McGrath <roland@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently sigqueue_free() removes sigqueue from list, but doesn't cancel the
pending signal. This is not consistent, the task should either receive the
"full" signal along with siginfo_t, or it shouldn't receive the signal at all.
Change sigqueue_free() to clear SIGQUEUE_PREALLOC but leave sigqueue on list
if it is queued.
This is a user-visible change. If the signal is blocked, it stays queued
after sys_timer_delete() until unblocked with the "stale" si_code/si_value,
and of course it is still counted wrt RLIMIT_SIGPENDING which also limits
the number of posix timers.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Austin Clements <amdragon+kernelbugzilla@mit.edu>
Cc: Roland McGrath <roland@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a slight change in the namespace cgroup subsystem api.
The change is that previously when cgroup_clone() was called (currently
only from the unshare path in ns_proxy cgroup, you'd get a new group named
"node_$pid" whereas now you'll get a group named after just your pid.)
The only users who would notice it are those who are using the ns_proxy
cgroup subsystem to auto-create cgroups when namespaces are unshared -
something of an experimental feature, which I think really needs more
complete container/namespace support in order to be useful. I suspect the
only users are Cedric and Serge, or maybe a few others on
containers@lists.linux-foundation.org. And in fact it would only be
noticed by the users who make the assumption about how the name is
generated, rather than getting it from the /proc/<pid>/cgroups file for
the process in question.
Whether the change is actually needed or not I'm fairly agnostic on, but I
guess it is more elegant to just use the pid as the new group name rather
than adding a fairly arbitrary "node_" prefix on the front.
[menage@google.com: provided changelog]
Signed-off-by: Cedric Le Goater <clg@fr.ibm.com>
Cc: "Paul Menage" <menage@google.com>
Cc: "Serge E. Hallyn" <serue@us.ibm.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If none of the switch cases match, the PR_SET_PDEATHSIG and
PR_SET_DUMPABLE cases of the switch statement will never write to local
variable `error'.
Signed-off-by: Shi Weihua <shiwh@cn.fujitsu.com>
Cc: Andrew G. Morgan <morgan@kernel.org>
Acked-by: "Serge E. Hallyn" <serue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
__exit_signal() does flush_sigqueue(tsk->pending) outside of ->siglock.
This can race with another thread doing sigqueue_free(), we can free the
same SIGQUEUE_PREALLOC sigqueue twice or corrupt the pending->list.
Note that even sys_exit_group() can trigger this race, not only
sys_timer_delete().
Move the callsite of flush_sigqueue(tsk->pending) under ->siglock.
This patch doesn't touch flush_sigqueue(->shared_pending) below, it is
called when there are no other threads which can play with signals, and
sigqueue_free() can't be used outside of our thread group.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On kvm I have seen some rare hangs in stop_machine when I used more guest
cpus than hosts cpus. e.g. 32 guest cpus on 1 host cpu triggered the
hang quite often. I could also reproduce the problem on a 4 way z/VM host with
a 64 way guest.
It turned out that the guest was consuming all available cpus mostly for
spinning on scheduler locks like rq->lock. This is expected as the threads are
calling yield all the time.
The problem is now, that the host scheduling decisings together with the guest
scheduling decisions and spinlocks not being fair managed to create an
interesting scenario similar to a live lock. (Sometimes the hang resolved
itself after some minutes)
Changing stop_machine to yield the cpu to the hypervisor when yielding inside
the guest fixed the problem for me. While I am not completely happy with this
patch, I think it causes no harm and it really improves the situation for me.
I used cpu_relax for yielding to the hypervisor, does that work on all
architectures?
p.s.: If you want to reproduce the problem, cpu hotplug and kprobes use
stop_machine_run and both triggered the problem after some retries.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
CC: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
All uses of list_for_each_rcu() can be profitably replaced by the
easier-to-use list_for_each_entry_rcu(). This patch makes this change
for the Audit system, in preparation for removing the list_for_each_rcu()
API entirely. This time with well-formed SOB.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Addresses http://bugzilla.kernel.org/show_bug.cgi?id=10663
Reporter: Daniel Marjamki <danielm77@spray.se>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Limit sysctl_nr_open - we don't want ->max_fds to exceed MAX_INT and
we don't want size calculation for ->fd[] to overflow.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Add a common hex array in hexdump.c so everyone can use it.
Add a common hi/lo helper to avoid the shifting masking that is
done to get the upper and lower nibbles of a byte value.
Pull the pack_hex_byte helper from kgdb as it is opencoded many
places in the tree that will be consolidated.
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Acked-by: Paul Mundt <lethal@linux-sh.org>
Cc: Jason Wessel <jason.wessel@windriver.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Return type of cpu_rt_runtime_write() should be int instead of ssize_t.
Signed-off-by: Mirco Tischler <mt-ml@gmx.de>
Acked-by: Paul Menage <menage@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It acts exactly like a regular 'cond_resched()', but will not get
optimized away when CONFIG_PREEMPT is set.
Normal kernel code is already preemptable in the presense of
CONFIG_PREEMPT, so cond_resched() is optimized away (see commit
02b67cc3ba "sched: do not do
cond_resched() when CONFIG_PREEMPT").
But when wanting to conditionally reschedule while holding a lock, you
need to use "cond_sched_lock(lock)", and the new function is the BKL
equivalent of that.
Also make fs/locks.c use it.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The generic semaphore rewrite had a huge performance regression on AIM7
(and potentially other BKL-heavy benchmarks) because the generic
semaphores had been rewritten to be simple to understand and fair. The
latter, in particular, turns a semaphore-based BKL implementation into a
mess of scheduling.
The attempt to fix the performance regression failed miserably (see the
previous commit 00b41ec261 'Revert
"semaphore: fix"'), and so for now the simple and sane approach is to
instead just go back to the old spinlock-based BKL implementation that
never had any issues like this.
This patch also has the advantage of being reported to fix the
regression completely according to Yanmin Zhang, unlike the semaphore
hack which still left a couple percentage point regression.
As a spinlock, the BKL obviously has the potential to be a latency
issue, but it's not really any different from any other spinlock in that
respect. We do want to get rid of the BKL asap, but that has been the
plan for several years.
These days, the biggest users are in the tty layer (open/release in
particular) and Alan holds out some hope:
"tty release is probably a few months away from getting cured - I'm
afraid it will almost certainly be the very last user of the BKL in
tty to get fixed as it depends on everything else being sanely locked."
so while we're not there yet, we do have a plan of action.
Tested-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Matthew Wilcox <matthew@wil.cx>
Cc: Alexander Viro <viro@ftp.linux.org.uk>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit bf726eab37, as it has
been reported to cause a regression with processes stuck in __down(),
apparently because some missing wakeup.
Quoth Sven Wegener:
"I'm currently investigating a regression that has showed up with my
last git pull yesterday. Bisecting the commits showed bf726e
"semaphore: fix" to be the culprit, reverting it fixed the issue.
Symptoms: During heavy filesystem usage (e.g. a kernel compile) I get
several compiler processes in uninterruptible sleep, blocking all i/o
on the filesystem. System is an Intel Core 2 Quad running a 64bit
kernel and userspace. Filesystem is xfs on top of lvm. See below for
the output of sysrq-w."
See
http://lkml.org/lkml/2008/5/10/45
for full report.
In the meantime, we can just fix the BKL performance regression by
reverting back to the good old BKL spinlock implementation instead,
since any sleeping lock will generally perform badly, especially if it
tries to be fair.
Reported-by: Sven Wegener <sven.wegener@stealer.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Linus found a logic bug: we ignore the version number in a module's
vermagic string if we have CONFIG_MODVERSIONS set, but modversions
also lets through a module with no __versions section for modprobe
--force (with tainting, but still).
We should only ignore the start of the vermagic string if the module
actually *has* crcs to check. Rather than (say) having an
entertaining hissy fit and creating a config option to work around the
buggy code.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We allow missing __versions sections, because modprobe --force strips
it. It makes less sense to allow sections where there's no version
for a specific symbol the module uses, so disallow that.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for-linus' of git://git.kernel.dk/linux-2.6-block:
Revert "relay: fix splice problem"
docbook: fix bio missing parameter
block: use unitialized_var() in bio_alloc_bioset()
block: avoid duplicate calls to get_part() in disk stat code
cfq-iosched: make io priorities inherit CPU scheduling class as well as nice
block: optimize generic_unplug_device()
block: get rid of likely/unlikely predictions in merge logic
vfs: splice remove_suid() cleanup
cfq-iosched: fix RCU race in the cfq io_context destructor handling
block: adjust tagging function queue bit locking
block: sysfs store function needs to grab queue_lock and use queue_flag_*()
Due to a merge conflict, the sched_relax_domain_level control file was marked
as being handled by cpuset_read/write_u64, but the code to handle it was
actually in cpuset_common_file_read/write.
Since the value being written/read is in fact a signed integer, it should be
treated as such; this patch adds cpuset_read/write_s64 functions, and uses
them to handle the sched_relax_domain_level file.
With this patch, the sched_relax_domain_level can be read and written, and the
correct contents seen/updated.
Signed-off-by: Paul Menage <menage@google.com>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The conversion between virtual and real time is as follows:
dvt = rw/w * dt <=> dt = w/rw * dvt
Since we want the fair sleeper granularity to be in real time, we actually
need to do:
dvt = - rw/w * l
This bug could be related to the regression reported by Yanmin Zhang:
| Comparing with kernel 2.6.25, sysbench+mysql(oltp, readonly) has lots
| of regressions with 2.6.26-rc1:
|
| 1) 8-core stoakley: 28%;
| 2) 16-core tigerton: 20%;
| 3) Itanium Montvale: 50%.
Reported-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Yanmin Zhang reported:
| Comparing with kernel 2.6.25, AIM7 (use tmpfs) has more th
| regression under 2.6.26-rc1 on my 8-core stoakley, 16-core tigerton,
| and Itanium Montecito. Bisect located the patch below:
|
| 64ac24e738 is first bad commit
| commit 64ac24e738
| Author: Matthew Wilcox <matthew@wil.cx>
| Date: Fri Mar 7 21:55:58 2008 -0500
|
| Generic semaphore implementation
|
| After I manually reverted the patch against 2.6.26-rc1 while fixing
| lots of conflicts/errors, aim7 regression became less than 2%.
i reproduced the AIM7 workload and can confirm Yanmin's findings that
-.26-rc1 regresses over .25 - by over 67% here.
Looking at the workload i found and fixed what i believe to be the real
bug causing the AIM7 regression: it was inefficient wakeup / scheduling
/ locking behavior of the new generic semaphore code, causing suboptimal
performance.
The problem comes from the following code. The new semaphore code does
this on down():
spin_lock_irqsave(&sem->lock, flags);
if (likely(sem->count > 0))
sem->count--;
else
__down(sem);
spin_unlock_irqrestore(&sem->lock, flags);
and this on up():
spin_lock_irqsave(&sem->lock, flags);
if (likely(list_empty(&sem->wait_list)))
sem->count++;
else
__up(sem);
spin_unlock_irqrestore(&sem->lock, flags);
where __up() does:
list_del(&waiter->list);
waiter->up = 1;
wake_up_process(waiter->task);
and where __down() does this in essence:
list_add_tail(&waiter.list, &sem->wait_list);
waiter.task = task;
waiter.up = 0;
for (;;) {
[...]
spin_unlock_irq(&sem->lock);
timeout = schedule_timeout(timeout);
spin_lock_irq(&sem->lock);
if (waiter.up)
return 0;
}
the fastpath looks good and obvious, but note the following property of
the contended path: if there's a task on the ->wait_list, the up() of
the current owner will "pass over" ownership to that waiting task, in a
wake-one manner, via the waiter->up flag and by removing the waiter from
the wait list.
That is all and fine in principle, but as implemented in
kernel/semaphore.c it also creates a nasty, hidden source of contention!
The contention comes from the following property of the new semaphore
code: the new owner owns the semaphore exclusively, even if it is not
running yet.
So if the old owner, even if just a few instructions later, does a
down() [lock_kernel()] again, it will be blocked and will have to wait
on the new owner to eventually be scheduled (possibly on another CPU)!
Or if another task gets to lock_kernel() sooner than the "new owner"
scheduled, it will be blocked unnecessarily and for a very long time
when there are 2000 tasks running.
I.e. the implementation of the new semaphores code does wake-one and
lock ownership in a very restrictive way - it does not allow
opportunistic re-locking of the lock at all and keeps the scheduler from
picking task order intelligently.
This kind of scheduling, with 2000 AIM7 processes running, creates awful
cross-scheduling between those 2000 tasks, causes reduced parallelism, a
throttled runqueue length and a lot of idle time. With increasing number
of CPUs it causes an exponentially worse behavior in AIM7, as the chance
for a newly woken new-owner task to actually run anytime soon is less
and less likely.
Note that it takes just a tiny bit of contention for the 'new-semaphore
catastrophy' to happen: the wakeup latencies get added to whatever small
contention there is, and quickly snowball out of control!
I believe Yanmin's findings and numbers support this analysis too.
The best fix for this problem is to use the same scheduling logic that
the kernel/mutex.c code uses: keep the wake-one behavior (that is OK and
wanted because we do not want to over-schedule), but also allow
opportunistic locking of the lock even if a wakee is already "in
flight".
The patch below implements this new logic. With this patch applied the
AIM7 regression is largely fixed on my quad testbox:
# v2.6.25 vanilla:
..................
Tasks Jobs/Min JTI Real CPU Jobs/sec/task
2000 56096.4 91 207.5 789.7 0.4675
2000 55894.4 94 208.2 792.7 0.4658
# v2.6.26-rc1-166-gc0a1811 vanilla:
...................................
Tasks Jobs/Min JTI Real CPU Jobs/sec/task
2000 33230.6 83 350.3 784.5 0.2769
2000 31778.1 86 366.3 783.6 0.2648
# v2.6.26-rc1-166-gc0a1811 + semaphore-speedup:
...............................................
Tasks Jobs/Min JTI Real CPU Jobs/sec/task
2000 55707.1 92 209.0 795.6 0.4642
2000 55704.4 96 209.0 796.0 0.4642
i.e. a 67% speedup. We are now back to within 1% of the v2.6.25
performance levels and have zero idle time during the test, as expected.
Btw., interactivity also improved dramatically with the fix - for
example console-switching became almost instantaneous during this
workload (which after all is running 2000 tasks at once!), without the
patch it was stuck for a minute at times.
There's another nice side-effect of this speedup patch, the new generic
semaphore code got even smaller:
text data bss dec hex filename
1241 0 0 1241 4d9 semaphore.o.before
1207 0 0 1207 4b7 semaphore.o.after
(because the waiter.up complication got removed.)
Longer-term we should look into using the mutex code for the generic
semaphore code as well - but i's not easy due to legacies and it's
outside of the scope of v2.6.26 and outside the scope of this patch as
well.
Bisected-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
