Messages in this thread | | | Date | Mon, 23 Mar 2009 16:01:12 +0900 | From | KAMEZAWA Hiroyuki <> | Subject | Re: [PATCH -tip] cpuacct: per-cgroup utime/stime statistics - v4 |
| |
On Mon, 23 Mar 2009 12:06:03 +0530 Bharata B Rao <bharata@linux.vnet.ibm.com> wrote:
> On Mon, Mar 23, 2009 at 01:55:28PM +0900, KAMEZAWA Hiroyuki wrote: > > On Mon, 23 Mar 2009 10:05:38 +0530 > > Bharata B Rao <bharata@linux.vnet.ibm.com> wrote: > > > > > Hi Ingo, > > > > > > Here is the v4 of the cpuacct statistics patch to obtain per-cgroup > > > system and user times with appropriate tags and complete changelog. > > > This applies against the latest -tip plus the cpuacct hierarchy fix v2 > > > which I posted just now. Could you please include this in -tip ? > > > > > > Changelog: > > > > > > v4 > > > - Remove comments in cpuacct_update_stats() which explained why rcu_read_lock() > > > was needed (as per Peter Zijlstra's review comments). > > > - Don't say that percpu_counter_read() is broken in Documentation/cpuacct.txt > > > as per KAMEZAWA Hiroyuki's review comments. > > > > > Broken -> isn't safe ? no difference ;) > > From your comment last time, I thought you meant calling it broken would > be harsh. > > > Can't this help you ? > > > > > > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> > > --- > > Index: mmotm-2.6.29-Mar21/include/linux/percpu_counter.h > > =================================================================== > > --- mmotm-2.6.29-Mar21.orig/include/linux/percpu_counter.h > > +++ mmotm-2.6.29-Mar21/include/linux/percpu_counter.h > > @@ -62,6 +62,16 @@ static inline s64 percpu_counter_read(st > > return fbc->count; > > } > > > > +static inline s64 percpu_counter_read_safe(struct percpu_counter *fbc) > > +{ > > + s64 ret; > > + > > + spin_lock(&fbc->lock); > > + ret = fbc->count; > > + spin_unlock(&fbc->lock); > > + return ret; > > +} > > + > > /* > > * It is possible for the percpu_counter_read() to return a small negative > > * number for some counter which should never be negative. > > @@ -114,6 +124,11 @@ static inline s64 percpu_counter_read(st > > return fbc->count; > > } > > > > +static inline s64 percpu_counter_read_safe(struct percpu_counter *fbc) > > +{ > > + return fbc->count; > > +} > > + > > static inline s64 percpu_counter_read_positive(struct percpu_counter *fbc) > > { > > return fbc->count; > > Yes it helps, but for the records, let me note that it makes the readside > ~20% slower. > > [I measured time spent by cpuacct_stats_show(), which is part of > the read routine for cpuacct.stat by using kretprobe and for 100 reads > (or 100 executions of cpuacct_stats_show()), I get the following numbers: > percpu_counter_read() - 3166367 ns > percpu_counter_read_safe() - 3793546 ns which is ~20% slower. ] > > I would prefer that this patch should be included in its current form > and we could separately fix percpu_counter_read() given that there > has been an attempt in the past to fix this (http://lkml.org/lkml/2008/8/27/303) > > Moreover, I don't know how much acceptable it is to work around the problem > in percpu_counter_read() by introducing another variant > percpu_counter_read_safe(). >
Hmm, ok, how about this way ?
Removet this.
+- It is possible to see slightly outdated values for stime and utime + due to the batch processing nature of percpu_counter.
Anyway, we get overhead of vfs_read(), interrupts, at el ;)
put into todo list.
+TODO: +- It is theoretically possible to see wrong values for stime and utime. + This is because percpu_counter_read() on 32bit systems isn't safe + against concurrent writes. needs to be fixed.
Thanks, -Kame
| |