On Sun, Jan 25, 2015 at 03:56:33PM +0000, Mark Rutland wrote:
> On Sun, Jan 25, 2015 at 04:34:28AM +0000, Fengguang Wu wrote:
> > Greetings,
>
> Hi Fengguang,
>
> Thanks very much for the report.
>
> > 0day kernel testing robot got the below dmesg and the first bad commit is
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git perf/core
> >
> > commit d26bb7f73a2881f2412c340a27438b185f0cc3dc
> > Author: Mark Rutland <mark.rutl...@arm.com>
> > AuthorDate: Wed Jan 7 15:01:54 2015 +0000
> > Commit: Peter Zijlstra <pet...@infradead.org>
> > CommitDate: Fri Jan 23 15:17:56 2015 +0100
> >
> > perf: decouple unthrottling and rotating
>
> [...]
>
> What seems to be happening is:
>
> * An event is created in the context of task A on CPU0. As the task's hw
> context is empty of events, we call perf_pmu_ctx_activate. This adds
> the cpuctx of the relevant HW PMU to the active_ctx_list. Note that we
> checked the task's ctx for emptiness, then added the PMU's hw context.
>
> * An event is created (as a result of a clone()) in the context of task
> B on CPU0, and we do the same thing, finding the ctx empty of events
> we add the HW PMU's cpuctx to the active_ctx_list. As it's already in
> there, the WARN_ON(!list_empty(&cpuctx->active_ctx_list)) explodes.
>
> So I guess what we actually want to do is to turn the active_ctx list
> into a list of perf_event_contexts rather than perf_event_cpu_contexts,
> and add/remove from the list when a context is scheduled in/out (or
> updated empty<->nonempty). That way we remove chances for erroneous
> add/remove, and we don't need to treat task and CPU contexts separately
> in perf_event_task_tick.
>
> I managed to get the reproducer running on a box at home, so I'll spin a
> potential fix against that for a while, and send that out if I don't see
> explosions.
The below patch is a fix up to the broken patch above. I've given it
some testing with Fengguang's reproducer and some custom tests, and as
far as I can tell everything is working as expected.
With Fengguang's reproducer I see some failures in netlink, but I
believe that's an unrelated issue that trinity is tickling.
Peter, does the below patch look OK to you?
Thanks,
Mark.
---->8----
>From c88d36594f956666ea1936abde7a05414bf56032 Mon Sep 17 00:00:00 2001
From: Mark Rutland <mark.rutl...@arm.com>
Date: Mon, 26 Jan 2015 18:01:41 +0000
Subject: [PATCH] Perf: fix broken rotation/unthrolling fix
Commit d26bb7f73a2881f2 ("perf: decouple unthrottling and rotating")
bodged its attempted cleanup of the rotation_list usage. The PMU cpu
contexts were added to the active_ctx_list whenever an event was added
to any context, and hence could be added repeatedly, corrupting the list.
Additionally, as contexts were added to the active_ctx_list whevener the
first event was added to them in list_add_event, they could be accounted
for on the wrong CPU (if they were not active at the time). Thus a
context could be accounted as active on one CPU while it ran on another.
This patch fixes the issue by updating the active_ctx_list when events
are scheduled in/out, ensuring that the active_ctx_list stays in sync
with the HW state. By maintaining a list of perf_event_contexts rather
than perf_event_cpu_contexts, we don't need separate paths to handle the
cpu and task contexts, which makes the code a little simpler.
Signed-off-by: Mark Rutland <mark.rutl...@arm.com>
Reported-by: Fengguang Wu <fengguang...@intel.com>
Cc: Will Deacon <will.dea...@arm.com>
Cc: Paul Mackerras <pau...@samba.org>
Cc: Ingo Molnar <mi...@redhat.com>
Cc: Arnaldo Carvalho de Melo <a...@kernel.org>
---
include/linux/perf_event.h | 2 +-
kernel/events/core.c | 51 ++++++++++++++++------------------------------
2 files changed, 19 insertions(+), 34 deletions(-)
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index b21e2a1..4eb5f9c 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -489,6 +489,7 @@ struct perf_event_context {
*/
struct mutex mutex;
+ struct list_head active_ctx_list;
struct list_head pinned_groups;
struct list_head flexible_groups;
struct list_head event_list;
@@ -539,7 +540,6 @@ struct perf_cpu_context {
int exclusive;
struct hrtimer hrtimer;
ktime_t hrtimer_interval;
- struct list_head active_ctx_list;
struct pmu *unique_pmu;
struct perf_cgroup *cgrp;
};
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 1d02140..05a8da5 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -875,30 +875,29 @@ void perf_pmu_enable(struct pmu *pmu)
static DEFINE_PER_CPU(struct list_head, active_ctx_list);
/*
- * perf_pmu_ctx_activate(), perf_pmu_ctx_deactivate(), and
- * perf_even_task_tick() are fully serialized because they're strictly cpu
- * affine and perf_pmu_ctx{activate,deactivate} are called with IRQs disabled,
- * while perf_event_task_tick is called from IRQ context.
+ * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
+ * perf_event_task_tick() are fully serialized because they're strictly cpu
+ * affine and perf_event_ctx{activate,deactivate} are called with IRQs
+ * disabled, while perf_event_task_tick is called from IRQ context.
*/
-static void perf_pmu_ctx_activate(struct pmu *pmu)
+static void perf_event_ctx_activate(struct perf_event_context *ctx)
{
- struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
struct list_head *head = this_cpu_ptr(&active_ctx_list);
WARN_ON(!irqs_disabled());
- WARN_ON(!list_empty(&cpuctx->active_ctx_list));
+ WARN_ON(!list_empty(&ctx->active_ctx_list));
- list_add(&cpuctx->active_ctx_list, head);
+ list_add(&ctx->active_ctx_list, head);
}
-static void perf_pmu_ctx_deactivate(struct pmu *pmu)
+static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
{
- struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
-
WARN_ON(!irqs_disabled());
- list_del_init(&cpuctx->active_ctx_list);
+ WARN_ON(list_empty(&ctx->active_ctx_list));
+
+ list_del_init(&ctx->active_ctx_list);
}
static void get_ctx(struct perf_event_context *ctx)
@@ -1243,8 +1242,6 @@ list_add_event(struct perf_event *event, struct
perf_event_context *ctx)
ctx->nr_cgroups++;
list_add_rcu(&event->event_entry, &ctx->event_list);
- if (!ctx->nr_events)
- perf_pmu_ctx_activate(ctx->pmu);
ctx->nr_events++;
if (event->attr.inherit_stat)
ctx->nr_stat++;
@@ -1415,8 +1412,6 @@ list_del_event(struct perf_event *event, struct
perf_event_context *ctx)
ctx->nr_stat--;
list_del_rcu(&event->event_entry);
- if (!ctx->nr_events)
- perf_pmu_ctx_deactivate(ctx->pmu);
if (event->group_leader == event)
list_del_init(&event->group_entry);
@@ -1570,7 +1565,8 @@ event_sched_out(struct perf_event *event,
if (!is_software_event(event))
cpuctx->active_oncpu--;
- ctx->nr_active--;
+ if (!--ctx->nr_active)
+ perf_event_ctx_deactivate(ctx);
if (event->attr.freq && event->attr.sample_freq)
ctx->nr_freq--;
if (event->attr.exclusive || !cpuctx->active_oncpu)
@@ -1894,7 +1890,8 @@ event_sched_in(struct perf_event *event,
if (!is_software_event(event))
cpuctx->active_oncpu++;
- ctx->nr_active++;
+ if (!ctx->nr_active++)
+ perf_event_ctx_activate(ctx);
if (event->attr.freq && event->attr.sample_freq)
ctx->nr_freq++;
@@ -3035,11 +3032,6 @@ static void rotate_ctx(struct perf_event_context *ctx)
list_rotate_left(&ctx->flexible_groups);
}
-/*
- * perf_pmu_ctx_activate() and perf_rotate_context() are fully serialized
- * because they're strictly cpu affine and rotate_start is called with IRQs
- * disabled, while rotate_context is called from IRQ context.
- */
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
{
struct perf_event_context *ctx = NULL;
@@ -3093,8 +3085,7 @@ bool perf_event_can_stop_tick(void)
void perf_event_task_tick(void)
{
struct list_head *head = this_cpu_ptr(&active_ctx_list);
- struct perf_cpu_context *cpuctx, *tmp;
- struct perf_event_context *ctx;
+ struct perf_event_context *ctx, *tmp;
int throttled;
WARN_ON(!irqs_disabled());
@@ -3102,14 +3093,8 @@ void perf_event_task_tick(void)
__this_cpu_inc(perf_throttled_seq);
throttled = __this_cpu_xchg(perf_throttled_count, 0);
- list_for_each_entry_safe(cpuctx, tmp, head, active_ctx_list) {
- ctx = &cpuctx->ctx;
+ list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
perf_adjust_freq_unthr_context(ctx, throttled);
-
- ctx = cpuctx->task_ctx;
- if (ctx)
- perf_adjust_freq_unthr_context(ctx, throttled);
- }
}
static int event_enable_on_exec(struct perf_event *event,
@@ -3268,6 +3253,7 @@ static void __perf_event_init_context(struct
perf_event_context *ctx)
{
raw_spin_lock_init(&ctx->lock);
mutex_init(&ctx->mutex);
+ INIT_LIST_HEAD(&ctx->active_ctx_list);
INIT_LIST_HEAD(&ctx->pinned_groups);
INIT_LIST_HEAD(&ctx->flexible_groups);
INIT_LIST_HEAD(&ctx->event_list);
@@ -6980,7 +6966,6 @@ skip_type:
__perf_cpu_hrtimer_init(cpuctx, cpu);
- INIT_LIST_HEAD(&cpuctx->active_ctx_list);
cpuctx->unique_pmu = pmu;
}
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
