Impact: use deeper function tracing depth safely
Some tests showed that function return tracing needed a more deeper depth
of function calls. But it could be unsafe to store these return addresses
to the stack.
So these arrays will now be allocated dynamically into task_struct of current
only when the tracer is activated.
Typical scheme when tracer is activated:
- allocate a return stack for each task in global list.
- fork: allocate the return stack for the newly created task
- exit: free return stack of current
- idle init: same as fork
I chose a default depth of 50. I don't have overruns anymore.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: help to find the better depth of trace
We decided to arbitrary define the depth of function return trace as
"20". Perhaps this is not enough. To help finding an optimal depth, we
measure now the overrun: the number of functions that have been missed
for the current thread. By default this is not displayed, we have to
do set a particular flag on the return tracer: echo overrun >
/debug/tracing/trace_options And the overrun will be printed on the
right.
As the trace shows below, the current 20 depth is not enough.
update_wall_time+0x37f/0x8c0 -> update_xtime_cache (345 ns) (Overruns: 2838)
update_wall_time+0x384/0x8c0 -> clocksource_get_next (1141 ns) (Overruns: 2838)
do_timer+0x23/0x100 -> update_wall_time (3882 ns) (Overruns: 2838)
tick_do_update_jiffies64+0xbf/0x160 -> do_timer (5339 ns) (Overruns: 2838)
tick_sched_timer+0x6a/0xf0 -> tick_do_update_jiffies64 (7209 ns) (Overruns: 2838)
vgacon_set_cursor_size+0x98/0x120 -> native_io_delay (2613 ns) (Overruns: 274)
vgacon_cursor+0x16e/0x1d0 -> vgacon_set_cursor_size (33151 ns) (Overruns: 274)
set_cursor+0x5f/0x80 -> vgacon_cursor (36432 ns) (Overruns: 274)
con_flush_chars+0x34/0x40 -> set_cursor (38790 ns) (Overruns: 274)
release_console_sem+0x1ec/0x230 -> up (721 ns) (Overruns: 274)
release_console_sem+0x225/0x230 -> wake_up_klogd (316 ns) (Overruns: 274)
con_flush_chars+0x39/0x40 -> release_console_sem (2996 ns) (Overruns: 274)
con_write+0x22/0x30 -> con_flush_chars (46067 ns) (Overruns: 274)
n_tty_write+0x1cc/0x360 -> con_write (292670 ns) (Overruns: 274)
smp_apic_timer_interrupt+0x2a/0x90 -> native_apic_mem_write (330 ns) (Overruns: 274)
irq_enter+0x17/0x70 -> idle_cpu (413 ns) (Overruns: 274)
smp_apic_timer_interrupt+0x2f/0x90 -> irq_enter (1525 ns) (Overruns: 274)
ktime_get_ts+0x40/0x70 -> getnstimeofday (465 ns) (Overruns: 274)
ktime_get_ts+0x60/0x70 -> set_normalized_timespec (436 ns) (Overruns: 274)
ktime_get+0x16/0x30 -> ktime_get_ts (2501 ns) (Overruns: 274)
hrtimer_interrupt+0x77/0x1a0 -> ktime_get (3439 ns) (Overruns: 274)
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds the support for dynamic tracing on the function return tracer.
The whole difference with normal dynamic function tracing is that we don't need
to hook on a particular callback. The only pro that we want is to nop or set
dynamically the calls to ftrace_caller (which is ftrace_return_caller here).
Some security checks ensure that we are not trying to launch dynamic tracing for
return tracing while normal function tracing is already running.
An example of trace with getnstimeofday set as a filter:
ktime_get_ts+0x22/0x50 -> getnstimeofday (2283 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1396 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1382 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1825 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1426 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1464 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1524 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1382 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1382 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1434 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1464 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1502 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1404 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1397 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1051 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1314 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1344 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1163 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1390 ns)
ktime_get_ts+0x22/0x50 -> getnstimeofday (1374 ns)
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: fix possible race condition in ftrace function return tracer
This fixes a possible race condition if index incrementation
is not immediately flushed in memory.
Thanks for Andi Kleen and Steven Rostedt for pointing out this issue
and give me this solution.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: allow archs more flexibility on dynamic ftrace implementations
Dynamic ftrace has largly been developed on x86. Since x86 does not
have the same limitations as other architectures, the ftrace interaction
between the generic code and the architecture specific code was not
flexible enough to handle some of the issues that other architectures
have.
Most notably, module trampolines. Due to the limited branch distance
that archs make in calling kernel core code from modules, the module
load code must create a trampoline to jump to what will make the
larger jump into core kernel code.
The problem arises when this happens to a call to mcount. Ftrace checks
all code before modifying it and makes sure the current code is what
it expects. Right now, there is not enough information to handle modifying
module trampolines.
This patch changes the API between generic dynamic ftrace code and
the arch dependent code. There is now two functions for modifying code:
ftrace_make_nop(mod, rec, addr) - convert the code at rec->ip into
a nop, where the original text is calling addr. (mod is the
module struct if called by module init)
ftrace_make_caller(rec, addr) - convert the code rec->ip that should
be a nop into a caller to addr.
The record "rec" now has a new field called "arch" where the architecture
can add any special attributes to each call site record.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Optimize a bit the function return tracer
This patch changes the calling convention of prepare_ftrace_return to
pass its arguments by register. This will optimize it a bit and
prepare it to support dynamic tracing.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: remove spinlocks and irq disabling in function return tracer.
I've tried to figure out all of the race condition that could happen
when the tracer pushes or pops a return address trace to/from the
current thread_info.
Theory:
_ One thread can only execute on one cpu at a time. So this code
doesn't need to be SMP-safe. Just drop the spinlock.
_ The only race could happen between the current thread and an
interrupt. If an interrupt is raised, it will increase the index of
the return stack storage and then execute until the end of the
tracing to finally free the index it used. We don't need to disable
irqs.
This is theorical. In practice, I've tested it with a two-core SMP and
had no problem at all. Perhaps -tip testing could confirm it.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
fix:
arch/x86/kernel/ftrace.c: In function 'ftrace_return_to_handler':
arch/x86/kernel/ftrace.c:112: error: implicit declaration of function 'cpu_clock'
cpu_clock() is implicitly included via a number of ways, but its real
location is sched.h. (Build failure is triggerable if enough other
kernel components are turned off.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
fix:
arch/x86/kernel/ftrace.c: Assembler messages:
arch/x86/kernel/ftrace.c:140: Error: missing ')'
arch/x86/kernel/ftrace.c:140: Error: junk `(%ebp))' after expression
arch/x86/kernel/ftrace.c:141: Error: missing ')'
arch/x86/kernel/ftrace.c:141: Error: junk `(%ebp))' after expression
the [parent_replaced] is used in an =rm fashion, so that constraint
is correct in isolation - but [parent_old] aliases register %0 and uses
it in an addressing mode that is only valid with registers - so change
the constraint from =rm to =r.
This fixes the build failure.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: add infrastructure for function-return tracing
Add low level support for ftrace return tracing.
This plug-in stores return addresses on the thread_info structure of
the current task.
The index of the current return address is initialized when the task
is the first one (init) and when a process forks (the child). It is
not needed when a task does a sys_execve because after this syscall,
it still needs to return on the kernel functions it called.
Note that the code of return_to_handler has been suggested by Steven
Rostedt as almost all of the ideas of improvements in this V3.
For purpose of security, arch/x86/kernel/process_32.c is not traced
because __switch_to() changes the current task during its execution.
That could cause inconsistency in the stored return address of this
function even if I didn't have any crash after testing with tracing on
this function enabled.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
This patch cleans up the NMI safe code for dynamic ftrace as suggested
by Andrew Morton.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: add more debug info to /debugfs/tracing/dyn_ftrace_total_info
This patch adds dynamic ftrace NMI update statistics to the
/debugfs/tracing/dyn_ftrace_total_info stat file.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: fix crashes that can occur in NMI handlers, if their code is modified
Modifying code is something that needs special care. On SMP boxes,
if code that is being modified is also being executed on another CPU,
that CPU will have undefined results.
The dynamic ftrace uses kstop_machine to make the system act like a
uniprocessor system. But this does not address NMIs, that can still
run on other CPUs.
One approach to handle this is to make all code that are used by NMIs
not be traced. But NMIs can call notifiers that spread throughout the
kernel and this will be very hard to maintain, and the chance of missing
a function is very high.
The approach that this patch takes is to have the NMIs modify the code
if the modification is taking place. The way this works is that just
writing to code executing on another CPU is not harmful if what is
written is the same as what exists.
Two buffers are used: an IP buffer and a "code" buffer.
The steps that the patcher takes are:
1) Put in the instruction pointer into the IP buffer
and the new code into the "code" buffer.
2) Set a flag that says we are modifying code
3) Wait for any running NMIs to finish.
4) Write the code
5) clear the flag.
6) Wait for any running NMIs to finish.
If an NMI is executed, it will also write the pending code.
Multiple writes are OK, because what is being written is the same.
Then the patcher must wait for all running NMIs to finish before
going to the next line that must be patched.
This is basically the RCU approach to code modification.
Thanks to Ingo Molnar for suggesting the idea, and to Arjan van de Ven
for his guidence on what is safe and what is not.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: avoid section mismatch warning, clean up
The dynamic ftrace determines which nop is safe to use at start up.
When it finds a safe nop for patching, it sets a pointer called ftrace_nop
to point to the code. All call sites are then patched to this nop.
Later, when tracing is turned on, this ftrace_nop variable is again used
to compare the location to make sure it is a nop before we update it to
an mcount call. If this fails just once, a warning is printed and ftrace
is disabled.
Rakib Mullick noted that the code that sets up the nop is a .init section
where as the nop itself is in the .text section. This is needed because
the nop is used later on after boot up. The problem is that the test of the
nop jumps back to the setup code and causes a "section mismatch" warning.
Rakib first recommended to convert the nop to .init.text, but as stated
above, this would fail since that text is used later.
The real solution is to extend Rabik's patch, and to make the ftrace_nop
into an array, and just save the code from the assembly to this array.
Now the section can stay as an init section, and we have a nop to use
later on.
Reported-by: Rakib Mullick <rakib.mullick@gmail.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The entire file of ftrace.c in the arch code needs to be marked
as notrace. It is much cleaner to do this from the Makefile with
CFLAGS_REMOVE_ftrace.o.
[ powerpc already had this in its Makefile. ]
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The arch dependent function ftrace_mcount_set was only used by the daemon
start up code. This patch removes it.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Andrew Morton suggested using the proper API for reading and writing
kernel areas that might fault.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add comments to explain what is happening in the x86 arch ftrace code.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Have the ftrace_modify_code return error values:
-EFAULT on error of reading the address
-EINVAL if what is read does not match what it expected
-EPERM if the write fails to update after a successful match.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In ftrace, logic is defined in the WARN_ON_ONCE, which can become a
nop with some configs. This patch fixes it.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Commit 4c3dc21b136f8cb4b72afee16c3ba7e961656c0b in tip introduced the
5-byte NOP ftrace_test_p6nop:
jmp . + 5
.byte 0x00, 0x00, 0x00
This is not friendly to disassemblers because an odd number of 0x00s
ends in the middle of an instruction boundary. This changes the 0x00s
to 1-byte NOPs (0x90).
Signed-off-by: Anders Kaseorg <andersk@mit.edu>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
With latest -tip I get this bug:
[ 49.439988] in_atomic():0, irqs_disabled():1
[ 49.440118] INFO: lockdep is turned off.
[ 49.440118] Pid: 2814, comm: modprobe Tainted: G W 2.6.27-rc7 #4
[ 49.440118] [<c01215e1>] __might_sleep+0xe1/0x120
[ 49.440118] [<c01148ea>] ftrace_modify_code+0x2a/0xd0
[ 49.440118] [<c01148a2>] ? ftrace_test_p6nop+0x0/0xa
[ 49.440118] [<c016e80e>] __ftrace_update_code+0xfe/0x2f0
[ 49.440118] [<c01148a2>] ? ftrace_test_p6nop+0x0/0xa
[ 49.440118] [<c016f190>] ftrace_convert_nops+0x50/0x80
[ 49.440118] [<c016f1d6>] ftrace_init_module+0x16/0x20
[ 49.440118] [<c015498b>] load_module+0x185b/0x1d30
[ 49.440118] [<c01767a0>] ? find_get_page+0x0/0xf0
[ 49.440118] [<c02463c0>] ? sprintf+0x0/0x30
[ 49.440118] [<c034e012>] ? mutex_lock_interruptible_nested+0x1f2/0x350
[ 49.440118] [<c0154eb3>] sys_init_module+0x53/0x1b0
[ 49.440118] [<c0352340>] ? do_page_fault+0x0/0x740
[ 49.440118] [<c0104012>] syscall_call+0x7/0xb
[ 49.440118] =======================
It is because ftrace_modify_code() calls copy_to_user and
copy_from_user.
These functions have been inserted after guessing that there
couldn't be any race condition but copy_[to/from]_user might
sleep and __ftrace_update_code is called with local_irq_saved.
These function have been inserted since this commit:
d5e92e8978fd2574e415dc2792c5eb592978243d:
"ftrace: x86 use copy from user function"
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Could just as easily change the three casts to cast to the correct
type...this patch changes the type of ftrace_nop instead.
Supresses sparse warnings:
arch/x86/kernel/ftrace.c:157:14: warning: incorrect type in assignment (different signedness)
arch/x86/kernel/ftrace.c:157:14: expected long *static [toplevel] ftrace_nop
arch/x86/kernel/ftrace.c:157:14: got unsigned long *<noident>
arch/x86/kernel/ftrace.c:161:14: warning: incorrect type in assignment (different signedness)
arch/x86/kernel/ftrace.c:161:14: expected long *static [toplevel] ftrace_nop
arch/x86/kernel/ftrace.c:161:14: got unsigned long *<noident>
arch/x86/kernel/ftrace.c:165:14: warning: incorrect type in assignment (different signedness)
arch/x86/kernel/ftrace.c:165:14: expected long *static [toplevel] ftrace_nop
arch/x86/kernel/ftrace.c:165:14: got unsigned long *<noident>
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The modification of code is performed either by kstop_machine, before
SMP starts, or on module code before the module is executed. There is
no reason to do the modifications from assembly. The copy to and from
user functions are sufficient and produces cleaner and easier to read
code.
Thanks to Benjamin Herrenschmidt for suggesting the idea.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Mathieu Desnoyers revealed a bug in the original code. The nop that is
used to relpace the mcount caller can be a two part nop. This runs the
risk where a process can be preempted after executing the first nop, but
before the second part of the nop.
The ftrace code calls kstop_machine to keep multiple CPUs from executing
code that is being modified, but it does not protect against a task preempting
in the middle of a two part nop.
If the above preemption happens and the tracer is enabled, after the
kstop_machine runs, all those nops will be calls to the trace function.
If the preempted process that was preempted between the two nops is executed
again, it will execute half of the call to the trace function, and this
might crash the system.
This patch instead uses what both the latest Intel and AMD spec suggests.
That is the P6_NOP5 sequence of "0x0f 0x1f 0x44 0x00 0x00".
Note, some older CPUs and QEMU might fault on this nop, so this nop
is executed with fault handling first. If it detects a fault, it will then
use the code "0x66 0x66 0x66 0x66 0x90". If that faults, it will then
default to a simple "jmp 1f; .byte 0x00 0x00 0x00; 1:". The jmp is
not optimal but will do if the first two can not be executed.
TODO: Examine the cpuid to determine the nop to use.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
x86 now sets up the mcount locations through the build and no longer
needs to record the ip when the function is executed. This patch changes
the initial mcount to simply return. There's no need to do any other work.
If the ftrace start up test fails, the original mcount will be what everything
will use, so having this as fast as possible is a good thing.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Record the address of the mcount call-site. Currently all archs except sparc64
record the address of the instruction following the mcount call-site. Some
general cleanups are entailed. Storing mcount addresses in rec->ip enables
looking them up in the kprobe hash table later on to check if they're kprobe'd.
Signed-off-by: Abhishek Sagar <sagar.abhishek@gmail.com>
Cc: davem@davemloft.net
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The fault label to jump to on fault of updating the code was misplaced
preventing the fault from being recorded.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch replaces the indirect call to the mcount function
pointer with a direct call that will be patched by the
dynamic ftrace routines.
On boot up, the mcount function calls the ftace_stub function.
When the dynamic ftrace code is initialized, the ftrace_stub
is replaced with a call to the ftrace_record_ip, which records
the instruction pointers of the locations that call it.
Later, the ftraced daemon will call kstop_machine and patch all
the locations to nops.
When a ftrace is enabled, the original calls to mcount will now
be set top call ftrace_caller, which will do a direct call
to the registered ftrace function. This direct call is also patched
when the function that should be called is updated.
All patching is performed by a kstop_machine routine to prevent any
type of race conditions that is associated with modifying code
on the fly.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch moves the memory management of the ftrace
records out of the arch code and into the generic code
making the arch code simpler.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch patches the call to mcount with nops instead
of a jmp over the mcount call.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch adds a feature to dynamically replace the ftrace code
with the jmps to allow a kernel with ftrace configured to run
as fast as it can without it configured.
The way this works, is on bootup (if ftrace is enabled), a ftrace
function is registered to record the instruction pointer of all
places that call the function.
Later, if there's still any code to patch, a kthread is awoken
(rate limited to at most once a second) that performs a stop_machine,
and replaces all the code that was called with a jmp over the call
to ftrace. It only replaces what was found the previous time. Typically
the system reaches equilibrium quickly after bootup and there's no code
patching needed at all.
e.g.
call ftrace /* 5 bytes */
is replaced with
jmp 3f /* jmp is 2 bytes and we jump 3 forward */
3:
When we want to enable ftrace for function tracing, the IP recording
is removed, and stop_machine is called again to replace all the locations
of that were recorded back to the call of ftrace. When it is disabled,
we replace the code back to the jmp.
Allocation is done by the kthread. If the ftrace recording function is
called, and we don't have any record slots available, then we simply
skip that call. Once a second a new page (if needed) is allocated for
recording new ftrace function calls. A large batch is allocated at
boot up to get most of the calls there.
Because we do this via stop_machine, we don't have to worry about another
CPU executing a ftrace call as we modify it. But we do need to worry
about NMI's so all functions that might be called via nmi must be
annotated with notrace_nmi. When this code is configured in, the NMI code
will not call notrace.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
