perf: register pmu implementations
in [Kernel]
|
Prev: [PATCH] procfs: Do not release pid_ns->proc_mnt too early
Next: perf/perf_events: misleading number of samples due to mmap()
From: Frederic Weisbecker on 18 Jun 2010 01:00 On Wed, Jun 16, 2010 at 07:48:00PM +0200, Peter Zijlstra wrote: > > > + > > > +struct pmu *perf_init_event(struct perf_event *event) > > > +{ > > > + struct pmu *pmu; > > > + int idx; > > > + > > > + idx = srcu_read_lock(&pmus_srcu); > > > + list_for_each_entry_rcu(pmu, &pmus, entry) { > > > + int ret = pmu->event_init(event); > > > + if (!ret) > > > + break; > > > + if (ret != -ENOENT) { > > > + pmu = ERR_PTR(ret); > > > + break; > > > } > > > - pmu = &perf_ops_generic; > > > - break; > > > } > > > + srcu_read_unlock(&pmus_srcu, idx); > > > > > > > > This could use a simple mutex instead of a spinlock + srcu_sync on > > writer and srcu on reader. > > Right, that spinlock needs to be a mutex for sure, a later patch adds an > allocation under it. > > But even with a mutex we need srcu_sync in there to sync against the > readers. > > > That doesn't matter much that said. What I don't understand is > > why we need to synchronize the writers. Walking the list with > > list_*_rcu() looks justified once we support boot events, but > > until then... > > Well, the typical unregister user would be a module, if you unregister > and then dealloc the struct pmu by unloading the module a reader might > still see a reference to it if you don't srcu_sync it. Ok, I see what you mean. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Peter Zijlstra on 18 Jun 2010 04:50 On Fri, 2010-06-18 at 09:31 +1000, Paul Mackerras wrote: > > > @@ -1011,7 +1001,7 @@ static int hw_perf_cache_event(u64 confi > > return 0; > > } > > > > -struct pmu *hw_perf_event_init(struct perf_event *event) > > +static int power_pmu_event_init(struct perf_event *event) > > How does power_pmu_event_init ever get called now? I don't see any > other references to it in the patch. Should struct pmu have a > reference to it? Uhm yeah.. looks like that went pear-shaped. Fixed it. --- Index: linux-2.6/arch/powerpc/kernel/perf_event.c =================================================================== --- linux-2.6.orig/arch/powerpc/kernel/perf_event.c +++ linux-2.6/arch/powerpc/kernel/perf_event.c @@ -1125,6 +1125,7 @@ static int power_pmu_event_init(struct p } struct pmu power_pmu = { + .event_init = power_pmu_event_init, .enable = power_pmu_enable, .disable = power_pmu_disable, .read = power_pmu_read, -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Frederic Weisbecker on 28 Jun 2010 09:30 On Thu, Jun 24, 2010 at 04:28:09PM +0200, Peter Zijlstra wrote: > + if (bp->attr.type != PERF_TYPE_BREAKPOINT) > + return -ENOENT; > + > + err = register_perf_hw_breakpoint(bp); > + if (err) > + return err; > + > + bp->destroy = bp_perf_event_destroy; Seems it would make sense to also have destroy in the pmu, it's the same along every events in the same class right? But this can be for later. > +static LIST_HEAD(pmus); > +static DEFINE_MUTEX(pmus_lock); > +static struct srcu_struct pmus_srcu; > + > +int perf_pmu_register(struct pmu *pmu) > +{ > + mutex_lock(&pmus_lock); > + list_add_rcu(&pmu->entry, &pmus); > + mutex_unlock(&pmus_lock); > + > + return 0; > +} > + > +void perf_pmu_unregister(struct pmu *pmu) > +{ > + mutex_lock(&pmus_lock); > + list_del_rcu(&pmu->entry); > + mutex_unlock(&pmus_lock); > > - atomic_inc(&perf_swevent_enabled[event_id]); > - event->destroy = sw_perf_event_destroy; > + synchronize_srcu(&pmus_srcu); > +} > + > +struct pmu *perf_init_event(struct perf_event *event) > +{ > + struct pmu *pmu = NULL; > + int idx; > + > + idx = srcu_read_lock(&pmus_srcu); > + list_for_each_entry_rcu(pmu, &pmus, entry) { > + int ret = pmu->event_init(event); > + if (!ret) > + break; > + if (ret != -ENOENT) { > + pmu = ERR_PTR(ret); > + break; > } > - pmu = &perf_ops_generic; > - break; > } > + srcu_read_unlock(&pmus_srcu, idx); > > return pmu; > } I'm still not sure why all this locking is needed. We don't even support pmus in modules. Is there something coming soon that will use this? I remember something about KVM. And who will have to use srcu? It seems the event fastpath would be concerned, right? Will that have an impact on the performances? > @@ -5743,15 +5742,15 @@ perf_cpu_notify(struct notifier_block *s > { > unsigned int cpu = (long)hcpu; > > - switch (action) { > + switch (action & ~CPU_TASKS_FROZEN) { > > case CPU_UP_PREPARE: > - case CPU_UP_PREPARE_FROZEN: > + case CPU_DOWN_FAILED: > perf_event_init_cpu(cpu); > break; > > + case CPU_UP_CANCELED: > case CPU_DOWN_PREPARE: > - case CPU_DOWN_PREPARE_FROZEN: > perf_event_exit_cpu(cpu); > break; That doesn't seem to be related to this patch initial topic. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Peter Zijlstra on 28 Jun 2010 11:20 On Mon, 2010-06-28 at 15:21 +0200, Frederic Weisbecker wrote: > On Thu, Jun 24, 2010 at 04:28:09PM +0200, Peter Zijlstra wrote: > > + if (bp->attr.type != PERF_TYPE_BREAKPOINT) > > + return -ENOENT; > > + > > + err = register_perf_hw_breakpoint(bp); > > + if (err) > > + return err; > > + > > + bp->destroy = bp_perf_event_destroy; > Seems it would make sense to also have destroy in the pmu, it's the same > along every events in the same class right? > > But this can be for later. Ah, indeed. > > +static LIST_HEAD(pmus); > > +static DEFINE_MUTEX(pmus_lock); > > +static struct srcu_struct pmus_srcu; > > + > > +int perf_pmu_register(struct pmu *pmu) > > +{ > > + mutex_lock(&pmus_lock); > > + list_add_rcu(&pmu->entry, &pmus); > > + mutex_unlock(&pmus_lock); > > + > > + return 0; > > +} > > + > > +void perf_pmu_unregister(struct pmu *pmu) > > +{ > > + mutex_lock(&pmus_lock); > > + list_del_rcu(&pmu->entry); > > + mutex_unlock(&pmus_lock); > > > > - atomic_inc(&perf_swevent_enabled[event_id]); > > - event->destroy = sw_perf_event_destroy; > > + synchronize_srcu(&pmus_srcu); > > +} > > + > > +struct pmu *perf_init_event(struct perf_event *event) > > +{ > > + struct pmu *pmu = NULL; > > + int idx; > > + > > + idx = srcu_read_lock(&pmus_srcu); > > + list_for_each_entry_rcu(pmu, &pmus, entry) { > > + int ret = pmu->event_init(event); > > + if (!ret) > > + break; > > + if (ret != -ENOENT) { > > + pmu = ERR_PTR(ret); > > + break; > > } > > - pmu = &perf_ops_generic; > > - break; > > } > > + srcu_read_unlock(&pmus_srcu, idx); > > > > return pmu; > > } > > > > I'm still not sure why all this locking is needed. We don't even > support pmus in modules. > > Is there something coming soon that will use this? > I remember something about KVM. Possibly, not sure. We could put the unregister thing in a later patch, but I wanted to make sure it was sanely possibly and its only a few lines of code. > And who will have to use srcu? It seems the event fastpath would > be concerned, right? Will that have an impact on the performances? Only event creation like above (perf_init_event) will have to use SRCU, so not really a hot path. > > @@ -5743,15 +5742,15 @@ perf_cpu_notify(struct notifier_block *s > > { > > unsigned int cpu = (long)hcpu; > > > > - switch (action) { > > + switch (action & ~CPU_TASKS_FROZEN) { > > > > case CPU_UP_PREPARE: > > - case CPU_UP_PREPARE_FROZEN: > > + case CPU_DOWN_FAILED: > > perf_event_init_cpu(cpu); > > break; > > > > + case CPU_UP_CANCELED: > > case CPU_DOWN_PREPARE: > > - case CPU_DOWN_PREPARE_FROZEN: > > perf_event_exit_cpu(cpu); > > break; > > > > That doesn't seem to be related to this patch initial topic. Ah indeed, that needs to go live in its own patch. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Frederic Weisbecker on 28 Jun 2010 11:30
On Mon, Jun 28, 2010 at 05:16:37PM +0200, Peter Zijlstra wrote: > > > } > > > + srcu_read_unlock(&pmus_srcu, idx); > > > > > > return pmu; > > > } > > > > > > > > I'm still not sure why all this locking is needed. We don't even > > support pmus in modules. > > > > Is there something coming soon that will use this? > > I remember something about KVM. > > Possibly, not sure. We could put the unregister thing in a later patch, > but I wanted to make sure it was sanely possibly and its only a few > lines of code. Ok. > > And who will have to use srcu? It seems the event fastpath would > > be concerned, right? Will that have an impact on the performances? > > Only event creation like above (perf_init_event) will have to use SRCU, > so not really a hot path. Ah I see. The event itself is synchronized against the fast-path using rcu. And then pmus themselves would be synchronized against events. Right that makes sense. But then why RCU (or SRCU, whatever)? I mean parent event creation is quite rare. And child events won't need to be synchronized as far as the parent keeps a reference to the pmu. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |