Skip to contentSkip to conversation

perf report fractal output issue

DISCUSSION Anton Blanchard 2 participants 4 messages Oct 14, 2009 – Oct 20, 2009

lore
Share & tools
lore
#0 lore raw

Hi,

I noticed a problem where I wasn't always getting callchains:

# perf report
    98.47%  lock1_processes  [kernel]                [k] .lock_kernel

0.91% lock1_processes [kernel] [k] .fput

But a flat output confirmed the data is there:

# perf report -g flat
    98.47%  lock1_processes  [kernel]                             [k] .lock_kernel
            98.47%
                .lock_kernel
                .__posix_lock_file
                (nil)
                .fcntl_setlk
                .sys_fcntl
                syscall_exit
                0xfffb7d04844
                .testcase
                .testcase
                .affinitize
                .new_task
                .main
                0xfffb7c4933c
                0xfffb7c4955c
                (nil)

It looks like callchain__fprintf_graph() only prints the callchain if we have at least one rb entry at the top. In my case all the associated backtraces are identical so I assume we dont have to do any splitting.

Anton

#1 lore raw
Quoted 32 lines · Anton Blanchard

Hi,

I noticed a problem where I wasn't always getting callchains:

# perf report
    98.47%  lock1_processes  [kernel]                [k] .lock_kernel

0.91% lock1_processes [kernel] [k] .fput

But a flat output confirmed the data is there:

# perf report -g flat
    98.47%  lock1_processes  [kernel]                             [k] .lock_kernel
            98.47%
                .lock_kernel
                .__posix_lock_file
                (nil)
                .fcntl_setlk
                .sys_fcntl
                syscall_exit
                0xfffb7d04844
                .testcase
                .testcase
                .affinitize
                .new_task
                .main
                0xfffb7c4933c
                0xfffb7c4955c
                (nil)
Quoted 5 lines

It looks like callchain__fprintf_graph() only prints the callchain if we have at least one rb entry at the top. In my case all the associated backtraces are identical so I assume we dont have to do any splitting.

Anton

Hmm, could you please send me a perf.data where you can see that? Does that only happen on powerpc?

Thanks.

#2 lore raw

Hi,

Quoted 6 lines
Quoted 3 lines

It looks like callchain__fprintf_graph() only prints the callchain if we have at least one rb entry at the top. In my case all the associated backtraces are identical so I assume we dont have to do any splitting.

Hmm, could you please send me a perf.data where you can see that? Does that only happen on powerpc?

I can replicate this on x86 by taking a perf.data and truncating it after the first sample+backtrace:

# wget http://samba.org/~anton/perf_truncated.test
# perf report -i perf_truncated.test -g fractal
...
   100.00%     perf  [kernel]       [k] native_write_msr_safe
# perf report -i perf_truncated.test -g flat
...
   100.00%     perf  [kernel]       [k] native_write_msr_safe
           100.00%
                native_write_msr_safe
                intel_pmu_enable_all
                hw_perf_enable
                perf_enable
                __perf_counter_enable
                smp_call_function_single
                task_oncpu_function_call
                perf_counter_enable
                perf_counter_for_each_child
                perf_ioctl
                vfs_ioctl
                do_vfs_ioctl
                sys_ioctl
                system_call_fastpath
                0x7f00c1857537
#3 lore raw
Quoted 33 lines · Anton Blanchard

Hi,

Quoted 6 lines
Quoted 3 lines

It looks like callchain__fprintf_graph() only prints the callchain if we have at least one rb entry at the top. In my case all the associated backtraces are identical so I assume we dont have to do any splitting.

Hmm, could you please send me a perf.data where you can see that? Does that only happen on powerpc?

I can replicate this on x86 by taking a perf.data and truncating it after the first sample+backtrace:

# wget http://samba.org/~anton/perf_truncated.test
# perf report -i perf_truncated.test -g fractal
...
   100.00%     perf  [kernel]       [k] native_write_msr_safe
# perf report -i perf_truncated.test -g flat
...
   100.00%     perf  [kernel]       [k] native_write_msr_safe
           100.00%
                native_write_msr_safe
                intel_pmu_enable_all
                hw_perf_enable
                perf_enable
                __perf_counter_enable
                smp_call_function_single
                task_oncpu_function_call
Quoted 8 lines
                perf_counter_enable
                perf_counter_for_each_child
                perf_ioctl
                vfs_ioctl
                do_vfs_ioctl
                sys_ioctl
                system_call_fastpath
                0x7f00c1857537

Thanks, I can reproduce it with your file.
Actually it was there from the beginning but we have missed it:
the root of the callchain is never printed :-s

Say you have:

hist_entry

    |
    --------- f1
              f2
              |
              -------- f3
              |        f4
              |
              ---------f5
                       f6

Actually you never see that, instead you have:

hist_entry

    |
    --------- f3
    |         f4
    |
    --------- f5
              f6

Thanks a lot for this report!
I'm currently fixing it.

I just hope the fix won't be too big for .32 ....