[email protected] writes:
> From: Mohammad Shahid <[email protected]> > > vpa_dtl_event_init acquires dtl_access_lock when initializing the > first vpa_dtl PMU event. > > If vpa_dtl_mem_alloc() fails,the function returns -ENOMEM without > releasing the semaphore. > > Release the semaphore when buffer allocation fails and no other > PMU events are active. > > Signed-off-by: Mohammad Shahid <[email protected]> > --- > arch/powerpc/perf/vpa-dtl.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/arch/powerpc/perf/vpa-dtl.c b/arch/powerpc/perf/vpa-dtl.c > index 3e3d65b6c796..20eafeaffa6b 100644 > --- a/arch/powerpc/perf/vpa-dtl.c > +++ b/arch/powerpc/perf/vpa-dtl.c > @@ -396,6 +396,8 @@ static int vpa_dtl_event_init(struct perf_event *event) > > /* Allocate dtl buffer memory */ > if (vpa_dtl_mem_alloc(event->cpu)) { > + if (dtl_global_refc == 0) > + up_write(&dtl_access_lock); > spin_unlock(&dtl_global_lock); > return -ENOMEM; > } I looked at this code and I am unable to wrap my head around it. We have dtl_access_lock and dtl_global_lock. Then we have vpa_dtl, which is a per-cpu variable anyway. So I am not sure why we need both locks held while doing vpa_dtl_mem_alloc(). What exactly are they protecting? Can we move the allocation out of the lock and just assign dtl->buf later within the lock? In one other place I see dtl->buf = NULL without any lock. Also, do we need a spinlock to protect dtl_global_refc? Would an atomic variable not work here? The above fix is not wrong per se, based on today's code, but it would be better if someone who owns this code takes another look and fixes its design while we are at it. Another motivation for doing this work now is to bring in clang's context analysis for arch/powerpc/perf/, which can statically check that such locking rules are followed, provided there are proper annotations for every lock and the data it protects. -ritesh
