Re: [PATCHSET 00/12] perf tools: Apply percent-limit to callchains

[Namhyung Kim]

On Tue, Jan 26, 2016 at 03:41:35PM +0100, Jiri Olsa wrote:
> On Tue, Jan 26, 2016 at 11:10:25PM +0900, Namhyung Kim wrote:
> > On Tue, Jan 26, 2016 at 02:27:26PM +0100, Jiri Olsa wrote:
> > > On Tue, Jan 26, 2016 at 09:51:59PM +0900, Namhyung Kim wrote:
> > > > On Tue, Jan 26, 2016 at 01:14:47PM +0100, Jiri Olsa wrote:
> > > > > On Sun, Jan 24, 2016 at 10:53:23PM +0900, Namhyung Kim wrote:
> > > > > > Hello,
> > > > > > 
> > > > > > This patchset tries to implement percent limit to callchains which was
> > > > > > requested by Andi Kleen.  For some reason, limiting callchains by
> > > > > > (overhead) percentage didn't work well.  This patch fixes it and make
> > > > > > --percent-limit also works for callchains as well as hist entries.
> > > > > > 
> > > > > > This is available on 'perf/callchain-limit-v1' branch in my tree:
> > > > > > 
> > > > > >   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> > > > > > 
> > > > > > Any comments are welcome,
> > > > > > 
> > > > > > Thanks,
> > > > > > Namhyung
> > > > > > 
> > > > > > 
> > > > > > Namhyung Kim (12):
> > > > > >   perf report: Apply --percent-limit to callchains also
> > > > > >   perf report: Apply callchain percent limit on --stdio
> > > > > >   perf report: Get rid of hist_entry__callchain_fprintf()
> > > > > >   perf report: Fix percent calculation on --stdio
> > > > > >   perf report: Hide output pipe for percent-limited callchains on stdio
> > > > > >   perf hists browser: Fix dump to show correct callchain style
> > > > > >   perf hists browser: Fix callchain_node__count_rows()
> > > > > >   perf hists browser: Apply callchain percent limit
> > > > > >   perf hists browser: Fix callchain counting when press ENTER key
> > > > > >   perf hists browser: Fix counting callchains when expand/collapse all
> > > > > >   perf hists browser: Update percent base for fractal callchain mode
> > > > > >   perf report: Fix callchain percent limit on --gtk
> > > > > 
> > > > > is 0.5 the default or one has to use the --percent-limit 0.5
> > > > > for the limit to be effective?
> > > > 
> > > > Yes, it's effective now.  I also think we need to change the default
> > > > limit of 0.5.  It was set for 'fractal' mode initially AFAIK so its
> > > > percentage is relative to each node.  In this case 0.5% of limit makes
> > > > sense because it'll be a very small (absolute) value.
> > > > 
> > > > But With 'graph' mode (now default), there're many entries under 0.5
> > > > overhead and they silently won't show callchains anymore.  Actually I
> > > > was confused by it when working with this patchset.
> > > > 
> > > > What about 0.005% for the new default?
> > > > 
> > > > 
> > > > > 
> > > > > without the option I'm getting empty callchains that are below 0.5
> > > > > but only in TUI mode (attached).. --stdio shows them all unfolded
> > > > 
> > > > It should not show them all.  But I found that I missed a check for
> > > > a stdio case.  Could you please test below?
> > > 
> > > did not help, it's still there.. same output as before
> > 
> > Hmm.. strange, could you show me the (part of) stdio output?
> > 
> 
> yea, that one changed as well.. no callchains now, attached
> 
> 
> jirka
> 
> 
> ---
> [jolsa@krava perf]$ ./perf report --stdio
> 
> ...
> 
>     46.69%    46.69%  ls       [kernel.vmlinux]  [k] intel_bts_enable_local      
>             |
>             ---0x1000
>                __statfs
>                entry_SYSCALL_64_fastpath
>                sys_statfs
>                SYSC_statfs
>                user_statfs
>                user_path_at_empty
>                filename_lookup
>                path_lookupat
>                link_path_walk
>                inode_permission
>                __inode_permission
>                kernfs_iop_permission
>                kernfs_refresh_inode
>                security_inode_notifysecctx
>                selinux_inode_notifysecctx
>                selinux_inode_setsecurity
>                security_context_to_sid
>                security_context_to_sid_core
>                string_to_context_struct
>                hashtab_search
>                apic_timer_interrupt
>                smp_apic_timer_interrupt
>                local_apic_timer_interrupt
>                hrtimer_interrupt
>                __hrtimer_run_queues
>                tick_sched_timer
>                tick_sched_handle.isra.17
>                update_process_times
>                scheduler_tick
>                perf_event_task_tick
>                perf_pmu_enable.part.87
>                x86_pmu_enable
>                intel_pmu_enable_all
>                intel_bts_enable_local
> 
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] perf_pmu_enable.part.87     
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] perf_event_context_sched_in 
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] perf_event_exec             
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] setup_new_exec              
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] load_elf_binary             
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] search_binary_handler       
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] do_execveat_common.isra.33  
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] sys_execve                  
>      0.08%     0.00%  perf     [kernel.vmlinux]  [k] return_from_execve          
>      0.08%     0.00%  perf     [unknown]         [k] 0x00007f2175b35e07          
>      0.04%     0.00%  perf     [kernel.vmlinux]  [k] perf_event_nmi_handler      
>      0.04%     0.00%  perf     [kernel.vmlinux]  [k] nmi_handle                  
>      0.04%     0.00%  perf     [kernel.vmlinux]  [k] default_do_nmi              
>      0.04%     0.00%  perf     [kernel.vmlinux]  [k] do_nmi                      
>      0.04%     0.00%  perf     [kernel.vmlinux]  [k] end_repeat_nmi              
>      0.04%     0.04%  perf     [kernel.vmlinux]  [k] x86_pmu_enable              
>      0.04%     0.04%  perf     [kernel.vmlinux]  [k] native_apic_mem_write       
> 

What's the problem?  Now by default callchains under 0.5% (absolute)
will not be shown.  I think this is intended output, and we need to
consider changing the default percent limit.

Thanks,
Namhyung