Messages in this thread | | | Date | Wed, 1 Jul 2009 15:45:03 +0200 | From | Frederic Weisbecker <> | Subject | Re: [PATCH][RFC] Adding information of counts processes acquired how many spinlocks to schedstat |
| |
On Wed, Jul 01, 2009 at 09:31:39AM +0200, Ingo Molnar wrote: > > * Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp> wrote: > > > Hi, > > > > I wrote a test patch which add information of counts processes > > acquired how many spinlocks to schedstat. After applied this > > patch, /proc/<PID>/sched will change like this, > > > > init (1, #threads: 1) > > --------------------------------------------------------- > > se.exec_start : 482130.851458 > > se.vruntime : 26883.107980 > > se.sum_exec_runtime : 2316.651816 > > se.avg_overlap : 0.480053 > > se.avg_wakeup : 14.999993 > > .... > > se.nr_wakeups_passive : 1 > > se.nr_wakeups_idle : 0 > > se.nr_acquired_spinlock : 74483 > > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > > Looks potentially useful - but it would be nice and go one step > further and add lock acquire stats as a software-counter. > > Perfcounters is a feature of the latest upstream kernel, there's a > (still very small) Wiki page about it at: > > http://perf.wiki.kernel.org > > With perfcounters we can instrument various software properties of > the kernel as well, for example the number of page-faults in the > system per second: > > $ perf stat -a -e page-faults sleep 1 > > Performance counter stats for 'sleep 1': > > 294387 page-faults > > 1.022318527 seconds time elapsed > > Now, it would be nice to have a lock-acquire software-counter as > well, which would output things like: > > $ perf stat -a -e lock-acquires sleep 1 > > Performance counter stats for 'sleep 1': > > 294387 lock-acquires > > 1.022318527 seconds time elapsed > > Furthermore, beyond plain counts, doing this would also allow the > profiling of lock acquire places: perf record -e lock-acquires and > perf report would work fine. > > It is really easy to add a new sw counter, check how it is done for > the pagefault counter(s), see the uses of PERF_COUNT_SW_PAGE_FAULTS > in the following files: > > $ git grep -l PERF_COUNT_SW_PAGE_FAULTS > > arch/powerpc/mm/fault.c > arch/x86/mm/fault.c > include/linux/perf_counter.h > kernel/perf_counter.c > tools/perf/builtin-stat.c > tools/perf/design.txt
Indeed, the raw number of lock acquired may be useful for a perf profiling especially in the case of profile comparison.
But IMHO, this information is too much orphan and lonesome. We would gain a lot if this information is provided per lock. Another useful info would be the rate of the time spent in a contended state for a given lock.
Which makes me think it may be better to use the existing ftrace lock events as softwares counters for that, which takes into account the following events:
- lock_acquire - lock_release - lock_contended - lock_acquired
And these events are per lock.
Now the missing piece is the sampling count for events...
| |