Re: [PATCH v5 01/50] perf comm: Use regular mutex

[Ian Rogers]

On Wed, Dec 6, 2023 at 4:05 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Sat, Dec 2, 2023 at 3:55 PM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> >
> > On Thu, Nov 30, 2023 at 10:28 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> > >
> > > On Wed, Nov 29, 2023 at 4:56 PM Namhyung Kim <namhyung@kernelorg> wrote:
> > > >
> > > > On Mon, Nov 27, 2023 at 2:09 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> > > > >
> > > > > The rwsem is only after used for writing so switch to a mutex that has
> > > > > better error checking.
> > > > >
> > > > > Fixes: 7a8f349e9d14 ("perf rwsem: Add debug mode that uses a mutex")
> > > >
> > > > I think we talked about fixing this separately, no?
> > >
> > > Sorry, I'm unclear on an action to do. Currently changing the
> > > RWS_ERRORCHECK in tools/perf/util/rwsem.h will break the build without
> > > this change.
> >
> > Can it be like this?
> >
> > #ifdef RWS_ERRORCHECK
> > #define RWSEM_INITIALIZER  { .lock = PTHREAD_MUTEX_INITIALIZER, }
> > #else
> > #define RWSEM_INITIALIZER  { .lock = PTHREAD_RWLOCK_INITIALIZER, }
> > #endif
> >
> > >
> > > > > Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> > > > > ---
> > > > >  tools/perf/util/comm.c | 10 +++++-----
> > > > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > > > >
> > > > > diff --git a/tools/perf/util/comm.c b/tools/perf/util/comm.c
> > > > > index afb8d4fd2644..4ae7bc2aa9a6 100644
> > > > > --- a/tools/perf/util/comm.c
> > > > > +++ b/tools/perf/util/comm.c
> > > > > @@ -17,7 +17,7 @@ struct comm_str {
> > > > >
> > > > >  /* Should perhaps be moved to struct machine */
> > > > >  static struct rb_root comm_str_root;
> > > > > -static struct rw_semaphore comm_str_lock = {.lock = PTHREAD_RWLOCK_INITIALIZER,};
> > > > > +static struct mutex comm_str_lock = {.lock = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP,};
> > > >
> > > > IIUC it has a problem with musl libc.  Actually I think it's better to
> > > > hide the field and the pthread initializer under some macro like
> > > > MUTEX_INITIALIZER or DEFINE_MUTEX() like in the kernel.
> > >
> > > Will there be enough use to justify this? I think ideally we'd not be
> > > having global locks needing global initializers as we run into
> > > problems like we see in metrics needing to mix counting and sampling.
> >
> > I don't know but there might be a reason to use global locks.
> > Then we need to support the initialization and it'd be better
> > to make it easier to handle internal changes like this.
>
> Right. So you are suggesting I make a macro for initialization but
> when this change is applied it will remove the only user of the macro.
> The macro would clearly be redundant which is why I didn't do a
> separate fix for that before doing this change - to use a mutex as the
> rwsem here is only ever used as a write lock. If we're looking to
> improve rwsem I don't think adding unused macros is the best thing,
> for example, we could remove references to perf_singlethreaded which
> is an idea that has had its day.

Note, RWS_ERRORCHECK being enabled still is broken without this.

Thanks,
Ian

> Thanks,
> Ian
>
> > Thanks,
> > Namhyung