Re: [RFC] perf_events: support for uncore a.k.a. nest units

[stephane eranian]

Seems to me that struct pmu is a shared resource across all CPUs.
I don't understand why scheduling on one CPU would have to impact
all the other CPUs, unless I am missing something here.


On Wed, Apr 21, 2010 at 10:08 AM, Lin Ming <ming.m.lin@xxxxxxxxx> wrote:
> On Tue, 2010-04-20 at 20:03 +0800, Peter Zijlstra wrote:
>> On Tue, 2010-04-20 at 19:55 +0800, Lin Ming wrote:
>>
>> > > One thing not on that list, which should happen first I guess, is to
>> > > remove hw_perf_group_sched_in(). The idea is to add some sort of
>> > > transactional API to the struct pmu, so that we can delay the
>> > > schedulability check until commit time (and roll back when it fails).
>> > >
>> > > Something as simple as:
>> > >
>> > > Â struct pmu {
>> > > Â Â void start_txn(struct pmu *);
>> > > Â Â void commit_txn(struct pmu *);
>> > >
>> > > Â Â ,,,
>> > > Â };
>> >
>> > Could you please explain a bit more?
>> >
>> > Does it mean that "start_txn" perform the schedule events stuff
>> > and "commit_txn" perform the assign events stuff?
>> >
>> > Does "commit time" mean the actual activation in hw_perf_enable?
>>
>> No, the idea behind hw_perf_group_sched_in() is to not perform
>> schedulability tests on each event in the group, but to add the group as
>> a whole and then perform one test.
>>
>> Of course, when that test fails, you'll have to roll-back the whole
>> group again.
>>
>> So start_txn (or a better name) would simply toggle a flag in the pmu
>> implementation that will make pmu::enable() not perform the
>> schedulablilty test.
>>
>> Then commit_txn() will perform the schedulability test (so note the
>> method has to have a !void return value, my mistake in the earlier
>> email).
>>
>> This will allow us to use the regular
>> kernel/perf_event.c::group_sched_in() and all the rollback code.
>> Currently each hw_perf_group_sched_in() implementation duplicates all
>> the rolllback code (with various bugs).
>>
>>
>>
>> We must get rid of all weak hw_perf_*() functions before we can properly
>> consider multiple struct pmu implementations.
>>
>
> Thanks for the clear explanation.
>
> Does below patch show what you mean?
>
> I only touch the x86 arch code now.
>
> ---
> Âarch/x86/kernel/cpu/perf_event.c | Â161 +++++++++++--------------------------
> Âinclude/linux/perf_event.h    |  10 ++-
> Âkernel/perf_event.c       Â|  28 +++----
> Â3 files changed, 67 insertions(+), 132 deletions(-)
>
> diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
> index 626154a..62aa9a1 100644
> --- a/arch/x86/kernel/cpu/perf_event.c
> +++ b/arch/x86/kernel/cpu/perf_event.c
> @@ -944,6 +944,9 @@ static int x86_pmu_enable(struct perf_event *event)
> Â Â Â Âif (n < 0)
> Â Â Â Â Â Â Â Âreturn n;
>
> + Â Â Â if (!(event->pmu->flag & PERF_EVENT_TRAN_STARTED))
> + Â Â Â Â Â Â Â goto out;
> +
> Â Â Â Âret = x86_pmu.schedule_events(cpuc, n, assign);
> Â Â Â Âif (ret)
> Â Â Â Â Â Â Â Âreturn ret;
> @@ -953,6 +956,7 @@ static int x86_pmu_enable(struct perf_event *event)
> Â Â Â Â */
> Â Â Â Âmemcpy(cpuc->assign, assign, n*sizeof(int));
>
> +out:
> Â Â Â Âcpuc->n_events = n;
> Â Â Â Âcpuc->n_added += n - n0;
>
> @@ -1210,119 +1214,6 @@ x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
> Â Â Â Âreturn &unconstrained;
> Â}
>
> -static int x86_event_sched_in(struct perf_event *event,
> - Â Â Â Â Â Â Â Â Â Â Â Â struct perf_cpu_context *cpuctx)
> -{
> - Â Â Â int ret = 0;
> -
> - Â Â Â event->state = PERF_EVENT_STATE_ACTIVE;
> - Â Â Â event->oncpu = smp_processor_id();
> - Â Â Â event->tstamp_running += event->ctx->time - event->tstamp_stopped;
> -
> - Â Â Â if (!is_x86_event(event))
> - Â Â Â Â Â Â Â ret = event->pmu->enable(event);
> -
> - Â Â Â if (!ret && !is_software_event(event))
> - Â Â Â Â Â Â Â cpuctx->active_oncpu++;
> -
> - Â Â Â if (!ret && event->attr.exclusive)
> - Â Â Â Â Â Â Â cpuctx->exclusive = 1;
> -
> - Â Â Â return ret;
> -}
> -
> -static void x86_event_sched_out(struct perf_event *event,
> - Â Â Â Â Â Â Â Â Â Â Â Â Â struct perf_cpu_context *cpuctx)
> -{
> - Â Â Â event->state = PERF_EVENT_STATE_INACTIVE;
> - Â Â Â event->oncpu = -1;
> -
> - Â Â Â if (!is_x86_event(event))
> - Â Â Â Â Â Â Â event->pmu->disable(event);
> -
> - Â Â Â event->tstamp_running -= event->ctx->time - event->tstamp_stopped;
> -
> - Â Â Â if (!is_software_event(event))
> - Â Â Â Â Â Â Â cpuctx->active_oncpu--;
> -
> - Â Â Â if (event->attr.exclusive || !cpuctx->active_oncpu)
> - Â Â Â Â Â Â Â cpuctx->exclusive = 0;
> -}
> -
> -/*
> - * Called to enable a whole group of events.
> - * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
> - * Assumes the caller has disabled interrupts and has
> - * frozen the PMU with hw_perf_save_disable.
> - *
> - * called with PMU disabled. If successful and return value 1,
> - * then guaranteed to call perf_enable() and hw_perf_enable()
> - */
> -int hw_perf_group_sched_in(struct perf_event *leader,
> - Â Â Â Â Â Â Âstruct perf_cpu_context *cpuctx,
> - Â Â Â Â Â Â Âstruct perf_event_context *ctx)
> -{
> - Â Â Â struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
> - Â Â Â struct perf_event *sub;
> - Â Â Â int assign[X86_PMC_IDX_MAX];
> - Â Â Â int n0, n1, ret;
> -
> - Â Â Â if (!x86_pmu_initialized())
> - Â Â Â Â Â Â Â return 0;
> -
> - Â Â Â /* n0 = total number of events */
> - Â Â Â n0 = collect_events(cpuc, leader, true);
> - Â Â Â if (n0 < 0)
> - Â Â Â Â Â Â Â return n0;
> -
> - Â Â Â ret = x86_pmu.schedule_events(cpuc, n0, assign);
> - Â Â Â if (ret)
> - Â Â Â Â Â Â Â return ret;
> -
> - Â Â Â ret = x86_event_sched_in(leader, cpuctx);
> - Â Â Â if (ret)
> - Â Â Â Â Â Â Â return ret;
> -
> - Â Â Â n1 = 1;
> - Â Â Â list_for_each_entry(sub, &leader->sibling_list, group_entry) {
> - Â Â Â Â Â Â Â if (sub->state > PERF_EVENT_STATE_OFF) {
> - Â Â Â Â Â Â Â Â Â Â Â ret = x86_event_sched_in(sub, cpuctx);
> - Â Â Â Â Â Â Â Â Â Â Â if (ret)
> - Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â goto undo;
> - Â Â Â Â Â Â Â Â Â Â Â ++n1;
> - Â Â Â Â Â Â Â }
> - Â Â Â }
> - Â Â Â /*
> - Â Â Â Â* copy new assignment, now we know it is possible
> - Â Â Â Â* will be used by hw_perf_enable()
> - Â Â Â Â*/
> - Â Â Â memcpy(cpuc->assign, assign, n0*sizeof(int));
> -
> - Â Â Â cpuc->n_events Â= n0;
> - Â Â Â cpuc->n_added Â+= n1;
> - Â Â Â ctx->nr_active += n1;
> -
> - Â Â Â /*
> - Â Â Â Â* 1 means successful and events are active
> - Â Â Â Â* This is not quite true because we defer
> - Â Â Â Â* actual activation until hw_perf_enable() but
> - Â Â Â Â* this way we* ensure caller won't try to enable
> - Â Â Â Â* individual events
> - Â Â Â Â*/
> - Â Â Â return 1;
> -undo:
> - Â Â Â x86_event_sched_out(leader, cpuctx);
> - Â Â Â n0 Â= 1;
> - Â Â Â list_for_each_entry(sub, &leader->sibling_list, group_entry) {
> - Â Â Â Â Â Â Â if (sub->state == PERF_EVENT_STATE_ACTIVE) {
> - Â Â Â Â Â Â Â Â Â Â Â x86_event_sched_out(sub, cpuctx);
> - Â Â Â Â Â Â Â Â Â Â Â if (++n0 == n1)
> - Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â break;
> - Â Â Â Â Â Â Â }
> - Â Â Â }
> - Â Â Â return ret;
> -}
> -
> Â#include "perf_event_amd.c"
> Â#include "perf_event_p6.c"
> Â#include "perf_event_p4.c"
> @@ -1454,6 +1345,47 @@ static inline void x86_pmu_read(struct perf_event *event)
> Â Â Â Âx86_perf_event_update(event);
> Â}
>
> +/*
> + * Set the flag to make pmu::enable() not perform the
> + * schedulablilty test.
> + */
> +static void x86_pmu_start_txn(struct pmu *pmu)
> +{
> + Â Â Â pmu->flag |= PERF_EVENT_TRAN_STARTED;
> +}
> +
> +static void x86_pmu_stop_txn(struct pmu *pmu)
> +{
> + Â Â Â pmu->flag &= ~PERF_EVENT_TRAN_STARTED;
> +}
> +
> +/*
> + * Return 0 if commit transaction success
> + */
> +static int x86_pmu_commit_txn(struct pmu *pmu)
> +{
> + Â Â Â struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
> + Â Â Â int assign[X86_PMC_IDX_MAX];
> + Â Â Â int n, ret;
> +
> + Â Â Â n = cpuc->n_events;
> +
> + Â Â Â if (!x86_pmu_initialized())
> + Â Â Â Â Â Â Â return -EAGAIN;
> +
> + Â Â Â ret = x86_pmu.schedule_events(cpuc, n, assign);
> + Â Â Â if (ret)
> + Â Â Â Â Â Â Â return ret;
> +
> + Â Â Â /*
> + Â Â Â Â* copy new assignment, now we know it is possible
> + Â Â Â Â* will be used by hw_perf_enable()
> + Â Â Â Â*/
> + Â Â Â memcpy(cpuc->assign, assign, n*sizeof(int));
> +
> + Â Â Â return 0;
> +}
> +
> Âstatic const struct pmu pmu = {
>    Â.enable     = x86_pmu_enable,
>    Â.disable    Â= x86_pmu_disable,
> @@ -1461,6 +1393,9 @@ static const struct pmu pmu = {
>    Â.stop      = x86_pmu_stop,
>    Â.read      = x86_pmu_read,
>    Â.unthrottle   = x86_pmu_unthrottle,
> +    .start_txn   Â= x86_pmu_start_txn,
> +    .stop_txn    = x86_pmu_stop_txn,
> +    .commit_txn   = x86_pmu_commit_txn,
> Â};
>
> Â/*
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index bf896d0..93aa8d8 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -524,6 +524,8 @@ struct hw_perf_event {
>
> Âstruct perf_event;
>
> +#define PERF_EVENT_TRAN_STARTED 1
> +
> Â/**
> Â* struct pmu - generic performance monitoring unit
> Â*/
> @@ -534,6 +536,11 @@ struct pmu {
> Â Â Â Âvoid (*stop) Â Â Â Â Â Â Â Â Â Â(struct perf_event *event);
> Â Â Â Âvoid (*read) Â Â Â Â Â Â Â Â Â Â(struct perf_event *event);
> Â Â Â Âvoid (*unthrottle) Â Â Â Â Â Â Â(struct perf_event *event);
> + Â Â Â void (*start_txn) Â Â Â Â Â Â Â (struct pmu *pmu);
> + Â Â Â void (*stop_txn) Â Â Â Â Â Â Â Â(struct pmu *pmu);
> + Â Â Â int (*commit_txn) Â Â Â Â Â Â Â (struct pmu *pmu);
> +
> + Â Â Â u8 flag;
> Â};
>
> Â/**
> @@ -799,9 +806,6 @@ extern void perf_disable(void);
> Âextern void perf_enable(void);
> Âextern int perf_event_task_disable(void);
> Âextern int perf_event_task_enable(void);
> -extern int hw_perf_group_sched_in(struct perf_event *group_leader,
> - Â Â Â Â Â Â Âstruct perf_cpu_context *cpuctx,
> - Â Â Â Â Â Â Âstruct perf_event_context *ctx);
> Âextern void perf_event_update_userpage(struct perf_event *event);
> Âextern int perf_event_release_kernel(struct perf_event *event);
> Âextern struct perf_event *
> diff --git a/kernel/perf_event.c b/kernel/perf_event.c
> index 07b7a43..4537676 100644
> --- a/kernel/perf_event.c
> +++ b/kernel/perf_event.c
> @@ -83,14 +83,6 @@ extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
> Âvoid __weak hw_perf_disable(void) Â Â Â Â Â Â Â{ barrier(); }
> Âvoid __weak hw_perf_enable(void) Â Â Â Â Â Â Â { barrier(); }
>
> -int __weak
> -hw_perf_group_sched_in(struct perf_event *group_leader,
> - Â Â Â Â Â Â Âstruct perf_cpu_context *cpuctx,
> - Â Â Â Â Â Â Âstruct perf_event_context *ctx)
> -{
> - Â Â Â return 0;
> -}
> -
> Âvoid __weak perf_event_print_debug(void) Â Â Â { }
>
> Âstatic DEFINE_PER_CPU(int, perf_disable_count);
> @@ -642,14 +634,13 @@ group_sched_in(struct perf_event *group_event,
> Â Â Â Â Â Â Â struct perf_event_context *ctx)
> Â{
> Â Â Â Âstruct perf_event *event, *partial_group;
> + Â Â Â struct pmu *pmu = (struct pmu *)group_event->pmu;
> Â Â Â Âint ret;
>
> Â Â Â Âif (group_event->state == PERF_EVENT_STATE_OFF)
> Â Â Â Â Â Â Â Âreturn 0;
>
> - Â Â Â ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
> - Â Â Â if (ret)
> - Â Â Â Â Â Â Â return ret < 0 ? ret : 0;
> + Â Â Â pmu->start_txn(pmu);
>
> Â Â Â Âif (event_sched_in(group_event, cpuctx, ctx))
> Â Â Â Â Â Â Â Âreturn -EAGAIN;
> @@ -664,16 +655,21 @@ group_sched_in(struct perf_event *group_event,
> Â Â Â Â Â Â Â Â}
> Â Â Â Â}
>
> - Â Â Â return 0;
> + Â Â Â ret = pmu->commit_txn(pmu);
> + Â Â Â if (!ret) {
> + Â Â Â Â Â Â Â pmu->stop_txn(pmu);
> + Â Â Â Â Â Â Â return 0;
> + Â Â Â }
>
> Âgroup_error:
> + Â Â Â pmu->stop_txn(pmu);
> +
> Â Â Â Â/*
> - Â Â Â Â* Groups can be scheduled in as one unit only, so undo any
> - Â Â Â Â* partial group before returning:
> + Â Â Â Â* Commit transaction fails, rollback
> + Â Â Â Â* Groups can be scheduled in as one unit only, so undo
> + Â Â Â Â* whole group before returning:
> Â Â Â Â */
> Â Â Â Âlist_for_each_entry(event, &group_event->sibling_list, group_entry) {
> - Â Â Â Â Â Â Â if (event == partial_group)
> - Â Â Â Â Â Â Â Â Â Â Â break;
> Â Â Â Â Â Â Â Âevent_sched_out(event, cpuctx, ctx);
> Â Â Â Â}
> Â Â Â Âevent_sched_out(group_event, cpuctx, ctx);
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/