Skip to contentSkip to conversation

ftrace: add tracepoint for timer event

PATCH In review Xiao Guangrong 2 participants 3 messages May 22, 2009 – May 27, 2009

lore
Share & tools
lore
#0 lore raw

cover ftrace: add tracepoint for timer event

Hi,

We can know timer's whole lifecycle as init/start/expire/cancel by these tracepoints.

Following tracepoints are introduced as ingo's suggestion in

	http://marc.info/?l=linux-kernel&m=123791899529128&w=2

for timer:
trace_timer_init()
trace_timer_start()
trace_timer_expire()
trace_timer_cancel()

for hrtimer:
trace_hrtimer_init()
trace_hrtimer_start()
trace_hrtimer_expire()
trace_hrtimer_cancel()

for itimer:
itimer_start()
itimer_expire()
itimer_cancel()

Example ftrace output:
for timer:

           <...>-2998  [000] 63501.542376: timer_init: timer=e0b374e0
           <...>-2998  [000] 63501.542424: timer_start: timer=e0b374e0 func=test_timerfuc expires=4294941565 cpu=0
	  <idle>-0     [000] 63514.508219: timer_expire: timer=e0b374e0 func=test_timerfuc
          <idle>-0     [000] 63514.508222: timer_cancel: timer=e0b374e0 func=test_timerfuc

for hrtimer:

           <...>-3628  [001] 64228.304772: hrtimer_init: timer=e0b404cc clockid=CLOCK_REALTIME mode=HRTIMER_MODE_ABS
           <...>-3628  [001] 64228.304793: hrtimer_start: timer=e0b404cc func=test_hrtime expires=1242920654000000000 ns softexpires=1242920654000000000 ns
     ksoftirqd/1-7     [001] 64228.304858: hrtimer_expire: timer=e0b404cc func=test_hrtime
     ksoftirqd/1-7     [001] 64228.304860: hrtimer_cancel: timer=e0b404cc func=test_hrtime

for itimer:

           <...>-4183  [001] 66533.730163: itimer_start: task=main which=0 it_value=13000000000 it_interval=0
          <idle>-0     [001] 66546.698154: itimer_expire: task=main which=0
           <...>-4183  [000] 66546.698533: itimer_cancel: task=main which=0

We already have debugobject in timer to init/activate/deactivate/free, but it can't be covered function of there tracepoints, because:
1: We can't get timer's lifecycle information in userspace by debugobject,
it is necessary for system engineer to investigate system trouble caused by using timer.
2: We can't get information of whole lifecycle of timer by debugobject,
for example, deactivation of a timer.
3: There are many different tracing code in many kernel subsystem as blktrace,
debugobject, and tracepoint is designed as generic way to unify these tracing way.

include/trace/events/timer.h | 249 +++++++++++++++++++++++++++++++++++++++++++

 kernel/hrtimer.c             |    5
 kernel/itimer.c              |   17 ++
 kernel/posix-cpu-timers.c    |    3
 kernel/timer.c               |    9 +
 5 files changed, 279 insertions(+), 4 deletions(-)
#1 lore raw
Quoted 10 lines · Xiao Guangrong

We already have debugobject in timer to init/activate/deactivate/free, but it can't be covered function of there tracepoints, because:
1: We can't get timer's lifecycle information in userspace by debugobject,
it is necessary for system engineer to investigate system trouble caused by using timer.
2: We can't get information of whole lifecycle of timer by debugobject,
for example, deactivation of a timer.
3: There are many different tracing code in many kernel subsystem as
blktrace, debugobject, and tracepoint is designed as generic way to unify these tracing way.

No. You can not unify debugobject into tracepoints. debugobjects is a totally different beast. It's main purpose is to prevent undebugable system crashes which we have seen several times e.g: freeing of an active timer, reinitializing of an active timer ...

Dealing with these problems is not covered by tracepoints by any means. The trace point does not prevent the system crash which happens 2 seconds after the fact that an active timer is kfree'd, debugobject does and it points you to the exact place where the shit happens.

I'm not opposed to add tracepoints to the timer code at all. In fact I appreciate that, but your idea of substituting debugobjects with tracing is just plain wrong.

It's the other way round. tracing can reuse the existing debugobject hooks to insert trace points, but it can not replace the functionality at all.

Thanks,

tglx

#2 lore raw
Quoted 22 lines · Thomas Gleixner
Quoted 10 lines · Xiao Guangrong

We already have debugobject in timer to init/activate/deactivate/free, but it can't be covered function of there tracepoints, because:
1: We can't get timer's lifecycle information in userspace by debugobject,
it is necessary for system engineer to investigate system trouble caused by using timer.
2: We can't get information of whole lifecycle of timer by debugobject,
for example, deactivation of a timer.
3: There are many different tracing code in many kernel subsystem as
blktrace, debugobject, and tracepoint is designed as generic way to unify these tracing way.

No. You can not unify debugobject into tracepoints. debugobjects is a totally different beast. It's main purpose is to prevent undebugable system crashes which we have seen several times e.g: freeing of an active timer, reinitializing of an active timer ...

Dealing with these problems is not covered by tracepoints by any means. The trace point does not prevent the system crash which happens 2 seconds after the fact that an active timer is kfree'd, debugobject does and it points you to the exact place where the shit happens.

Quoted 8 lines

I'm not opposed to add tracepoints to the timer code at all. In fact I appreciate that, but your idea of substituting debugobjects with tracing is just plain wrong.

It's the other way round. tracing can reuse the existing debugobject hooks to insert trace points, but it can not replace the functionality at all.

Hello tglx:
Thanks for you review!

Totally agree.
Actually I do know the difference between debugobject and tracepoint. Sorry for making you misunderstand what I said for my pool English.

Thanks,
Xiao Guangrong

Quoted 3 lines

tglx