[PATCH 0/4] perf ring-buffer fixes

[PATCH 0/4] perf ring-buffer fixes

[Peter Zijlstra]

This is basically a split of Yabin's last patch.

[PATCH 1/4] perf/ring_buffer: Fix exposing a temporarily decreased data_head.

[Peter Zijlstra]

In perf_output_put_handle(), an IRQ/NMI can happen in below location and
write records to the same ring buffer:
	...
	local_dec_and_test(&rb->nest)
	...                          <-- an IRQ/NMI can happen here
	rb->user_page->data_head = head;
	...

In this case, a value A is written to data_head in the IRQ, then a value
B is written to data_head after the IRQ. And A > B. As a result,
data_head is temporarily decreased from A to B. And a reader may see
data_head < data_tail if it read the buffer frequently enough, which
creates unexpected behaviors.

This can be fixed by moving dec(&rb->nest) to after updating data_head,
which prevents the IRQ/NMI above from updating data_head.

Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
Signed-off-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20190516184010.167903-1-yabinc@google.com
---
 kernel/events/ring_buffer.c |   24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -51,11 +51,18 @@ static void perf_output_put_handle(struc
 	head = local_read(&rb->head);
 
 	/*
-	 * IRQ/NMI can happen here, which means we can miss a head update.
+	 * IRQ/NMI can happen here and advance @rb->head, causing our
+	 * load above to be stale.
 	 */
 
-	if (!local_dec_and_test(&rb->nest))
+	/*
+	 * If this isn't the outermost nesting, we don't have to update
+	 * @rb->user_page->data_head.
+	 */
+	if (local_read(&rb->nest) > 1) {
+		local_dec(&rb->nest);
 		goto out;
+	}
 
 	/*
 	 * Since the mmap() consumer (userspace) can run on a different CPU:
@@ -87,9 +94,18 @@ static void perf_output_put_handle(struc
 	rb->user_page->data_head = head;
 
 	/*
-	 * Now check if we missed an update -- rely on previous implied
-	 * compiler barriers to force a re-read.
+	 * We must publish the head before decrementing the nest count,
+	 * otherwise an IRQ/NMI can publish a more recent head value and our
+	 * write will (temporarily) publish a stale value.
+	 */
+	barrier();
+	local_set(&rb->nest, 0);
+
+	/*
+	 * Ensure we decrement @rb->nest before we validate the @rb->head.
+	 * Otherwise we cannot be sure we caught the 'last' nested update.
 	 */
+	barrier();
 	if (unlikely(head != local_read(&rb->head))) {
 		local_inc(&rb->nest);
 		goto again;

Re: [PATCH 1/4] perf/ring_buffer: Fix exposing a temporarily decreased data_head.

[Ingo Molnar]

* Peter Zijlstra <peterz@infradead.org> wrote:

> In perf_output_put_handle(), an IRQ/NMI can happen in below location and
> write records to the same ring buffer:
> 	...
> 	local_dec_and_test(&rb->nest)
> 	...                          <-- an IRQ/NMI can happen here
> 	rb->user_page->data_head = head;
> 	...
> 
> In this case, a value A is written to data_head in the IRQ, then a value
> B is written to data_head after the IRQ. And A > B. As a result,
> data_head is temporarily decreased from A to B. And a reader may see
> data_head < data_tail if it read the buffer frequently enough, which
> creates unexpected behaviors.
> 
> This can be fixed by moving dec(&rb->nest) to after updating data_head,
> which prevents the IRQ/NMI above from updating data_head.
> 
> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Jiri Olsa <jolsa@redhat.com>
> Cc: Namhyung Kim <namhyung@kernel.org>
> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
> Signed-off-by: Yabin Cui <yabinc@google.com>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Link: https://lkml.kernel.org/r/20190516184010.167903-1-yabinc@google.com

So these are missing a bunch of:

  From: Yabin Cui <yabinc@google.com>

lines, right?

Thanks,

	Ingo

Re: [PATCH 1/4] perf/ring_buffer: Fix exposing a temporarily decreased data_head.

[Peter Zijlstra]

On Fri, May 17, 2019 at 03:05:09PM +0200, Ingo Molnar wrote:
> 
> * Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > In perf_output_put_handle(), an IRQ/NMI can happen in below location and
> > write records to the same ring buffer:
> > 	...
> > 	local_dec_and_test(&rb->nest)
> > 	...                          <-- an IRQ/NMI can happen here
> > 	rb->user_page->data_head = head;
> > 	...
> > 
> > In this case, a value A is written to data_head in the IRQ, then a value
> > B is written to data_head after the IRQ. And A > B. As a result,
> > data_head is temporarily decreased from A to B. And a reader may see
> > data_head < data_tail if it read the buffer frequently enough, which
> > creates unexpected behaviors.
> > 
> > This can be fixed by moving dec(&rb->nest) to after updating data_head,
> > which prevents the IRQ/NMI above from updating data_head.
> > 
> > Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> > Cc: Ingo Molnar <mingo@redhat.com>
> > Cc: Jiri Olsa <jolsa@redhat.com>
> > Cc: Namhyung Kim <namhyung@kernel.org>
> > Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> > Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
> > Signed-off-by: Yabin Cui <yabinc@google.com>
> > Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > Link: https://lkml.kernel.org/r/20190516184010.167903-1-yabinc@google.com
> 
> So these are missing a bunch of:
> 
>   From: Yabin Cui <yabinc@google.com>
> 
> lines, right?
> 

The first. certainly, the rest, while inspired by his patch, is more
complete than what he did.

Re: [PATCH 1/4] perf/ring_buffer: Fix exposing a temporarily decreased data_head.

[Ingo Molnar]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Fri, May 17, 2019 at 03:05:09PM +0200, Ingo Molnar wrote:
> > 
> > * Peter Zijlstra <peterz@infradead.org> wrote:
> > 
> > > In perf_output_put_handle(), an IRQ/NMI can happen in below location and
> > > write records to the same ring buffer:
> > > 	...
> > > 	local_dec_and_test(&rb->nest)
> > > 	...                          <-- an IRQ/NMI can happen here
> > > 	rb->user_page->data_head = head;
> > > 	...
> > > 
> > > In this case, a value A is written to data_head in the IRQ, then a value
> > > B is written to data_head after the IRQ. And A > B. As a result,
> > > data_head is temporarily decreased from A to B. And a reader may see
> > > data_head < data_tail if it read the buffer frequently enough, which
> > > creates unexpected behaviors.
> > > 
> > > This can be fixed by moving dec(&rb->nest) to after updating data_head,
> > > which prevents the IRQ/NMI above from updating data_head.
> > > 
> > > Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> > > Cc: Ingo Molnar <mingo@redhat.com>
> > > Cc: Jiri Olsa <jolsa@redhat.com>
> > > Cc: Namhyung Kim <namhyung@kernel.org>
> > > Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> > > Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
> > > Signed-off-by: Yabin Cui <yabinc@google.com>
> > > Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > > Link: https://lkml.kernel.org/r/20190516184010.167903-1-yabinc@google.com
> > 
> > So these are missing a bunch of:
> > 
> >   From: Yabin Cui <yabinc@google.com>
> > 
> > lines, right?
> > 
> 
> The first. certainly, the rest, while inspired by his patch, is more
> complete than what he did.

Oh, indeed - I adjusted the first patch only.

Thanks,

	Ingo

[tip:perf/urgent] perf/ring_buffer: Fix exposing a temporarily decreased data_head

[tip-bot for Yabin Cui]

Commit-ID:  1b038c6e05ff70a1e66e3e571c2e6106bdb75f53
Gitweb:     https://git.kernel.org/tip/1b038c6e05ff70a1e66e3e571c2e6106bdb75f53
Author:     Yabin Cui <yabinc@google.com>
AuthorDate: Fri, 17 May 2019 13:52:31 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 24 May 2019 09:00:10 +0200

perf/ring_buffer: Fix exposing a temporarily decreased data_head

In perf_output_put_handle(), an IRQ/NMI can happen in below location and
write records to the same ring buffer:

	...
	local_dec_and_test(&rb->nest)
	...                          <-- an IRQ/NMI can happen here
	rb->user_page->data_head = head;
	...

In this case, a value A is written to data_head in the IRQ, then a value
B is written to data_head after the IRQ. And A > B. As a result,
data_head is temporarily decreased from A to B. And a reader may see
data_head < data_tail if it read the buffer frequently enough, which
creates unexpected behaviors.

This can be fixed by moving dec(&rb->nest) to after updating data_head,
which prevents the IRQ/NMI above from updating data_head.

[ Split up by peterz. ]

Signed-off-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: mark.rutland@arm.com
Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
Link: http://lkml.kernel.org/r/20190517115418.224478157@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/ring_buffer.c | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 674b35383491..009467a60578 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -51,11 +51,18 @@ again:
 	head = local_read(&rb->head);
 
 	/*
-	 * IRQ/NMI can happen here, which means we can miss a head update.
+	 * IRQ/NMI can happen here and advance @rb->head, causing our
+	 * load above to be stale.
 	 */
 
-	if (!local_dec_and_test(&rb->nest))
+	/*
+	 * If this isn't the outermost nesting, we don't have to update
+	 * @rb->user_page->data_head.
+	 */
+	if (local_read(&rb->nest) > 1) {
+		local_dec(&rb->nest);
 		goto out;
+	}
 
 	/*
 	 * Since the mmap() consumer (userspace) can run on a different CPU:
@@ -87,9 +94,18 @@ again:
 	rb->user_page->data_head = head;
 
 	/*
-	 * Now check if we missed an update -- rely on previous implied
-	 * compiler barriers to force a re-read.
+	 * We must publish the head before decrementing the nest count,
+	 * otherwise an IRQ/NMI can publish a more recent head value and our
+	 * write will (temporarily) publish a stale value.
+	 */
+	barrier();
+	local_set(&rb->nest, 0);
+
+	/*
+	 * Ensure we decrement @rb->nest before we validate the @rb->head.
+	 * Otherwise we cannot be sure we caught the 'last' nested update.
 	 */
+	barrier();
 	if (unlikely(head != local_read(&rb->head))) {
 		local_inc(&rb->nest);
 		goto again;

[PATCH 2/4] perf/ring_buffer: Add ordering to rb->nest increment

[Peter Zijlstra]

Similar to how decrementing rb->next too early can cause data_head to
(temporarily) be observed to go backward, so too can this happen when
we increment too late.

This barrier() ensures the rb->head load happens after the increment,
both the one in the 'goto again' path, as the one from
perf_output_get_handle() -- albeit very unlikely to matter for the
latter.

Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
Suggested-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/events/ring_buffer.c |    9 +++++++++
 1 file changed, 9 insertions(+)

--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -48,6 +48,15 @@ static void perf_output_put_handle(struc
 	unsigned long head;
 
 again:
+	/*
+	 * In order to avoid publishing a head value that goes backwards,
+	 * we must ensure the load of @rb->head happens after we've
+	 * incremented @rb->nest.
+	 *
+	 * Otherwise we can observe a @rb->head value before one published
+	 * by an IRQ/NMI happening between the load and the increment.
+	 */
+	barrier();
 	head = local_read(&rb->head);
 
 	/*

[tip:perf/urgent] perf/ring_buffer: Add ordering to rb->nest increment

[tip-bot for Peter Zijlstra]

Commit-ID:  3f9fbe9bd86c534eba2faf5d840fd44c6049f50e
Gitweb:     https://git.kernel.org/tip/3f9fbe9bd86c534eba2faf5d840fd44c6049f50e
Author:     Peter Zijlstra <peterz@infradead.org>
AuthorDate: Fri, 17 May 2019 13:52:32 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 24 May 2019 09:00:10 +0200

perf/ring_buffer: Add ordering to rb->nest increment

Similar to how decrementing rb->next too early can cause data_head to
(temporarily) be observed to go backward, so too can this happen when
we increment too late.

This barrier() ensures the rb->head load happens after the increment,
both the one in the 'goto again' path, as the one from
perf_output_get_handle() -- albeit very unlikely to matter for the
latter.

Suggested-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: acme@kernel.org
Cc: mark.rutland@arm.com
Cc: namhyung@kernel.org
Fixes: ef60777c9abd ("perf: Optimize the perf_output() path by removing IRQ-disables")
Link: http://lkml.kernel.org/r/20190517115418.309516009@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/ring_buffer.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 009467a60578..4b5f8d932400 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -48,6 +48,15 @@ static void perf_output_put_handle(struct perf_output_handle *handle)
 	unsigned long head;
 
 again:
+	/*
+	 * In order to avoid publishing a head value that goes backwards,
+	 * we must ensure the load of @rb->head happens after we've
+	 * incremented @rb->nest.
+	 *
+	 * Otherwise we can observe a @rb->head value before one published
+	 * by an IRQ/NMI happening between the load and the increment.
+	 */
+	barrier();
 	head = local_read(&rb->head);
 
 	/*

[PATCH 3/4] perf/ring-buffer: Always use {READ,WRITE}_ONCE() for rb->user_page data

[Peter Zijlstra]

We must use {READ,WRITE}_ONCE() on rb->user_page data such that
concurrent usage will see whole values. A few key sites were missing
this.

Fixes:7b732a750477 ("perf_counter: new output ABI - part 1")
Suggested-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/events/ring_buffer.c |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -100,7 +100,7 @@ static void perf_output_put_handle(struc
 	 * See perf_output_begin().
 	 */
 	smp_wmb(); /* B, matches C */
-	rb->user_page->data_head = head;
+	WRITE_ONCE(rb->user_page->data_head, head);
 
 	/*
 	 * We must publish the head before decrementing the nest count,
@@ -496,7 +496,7 @@ void perf_aux_output_end(struct perf_out
 		perf_event_aux_event(handle->event, aux_head, size,
 				     handle->aux_flags);
 
-	rb->user_page->aux_head = rb->aux_head;
+	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
 	if (rb_need_aux_wakeup(rb))
 		wakeup = true;
 
@@ -528,7 +528,7 @@ int perf_aux_output_skip(struct perf_out
 
 	rb->aux_head += size;
 
-	rb->user_page->aux_head = rb->aux_head;
+	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
 	if (rb_need_aux_wakeup(rb)) {
 		perf_output_wakeup(handle);
 		handle->wakeup = rb->aux_wakeup + rb->aux_watermark;

[tip:perf/urgent] perf/ring-buffer: Always use {READ,WRITE}_ONCE() for rb->user_page data

[tip-bot for Peter Zijlstra]

Commit-ID:  4d839dd9e4356bbacf3eb0ab13a549b83b008c21
Gitweb:     https://git.kernel.org/tip/4d839dd9e4356bbacf3eb0ab13a549b83b008c21
Author:     Peter Zijlstra <peterz@infradead.org>
AuthorDate: Fri, 17 May 2019 13:52:33 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 24 May 2019 09:00:11 +0200

perf/ring-buffer: Always use {READ,WRITE}_ONCE() for rb->user_page data

We must use {READ,WRITE}_ONCE() on rb->user_page data such that
concurrent usage will see whole values. A few key sites were missing
this.

Suggested-by: Yabin Cui <yabinc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: acme@kernel.org
Cc: mark.rutland@arm.com
Cc: namhyung@kernel.org
Fixes: 7b732a750477 ("perf_counter: new output ABI - part 1")
Link: http://lkml.kernel.org/r/20190517115418.394192145@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/ring_buffer.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 4b5f8d932400..7a0c73e4b3eb 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -100,7 +100,7 @@ again:
 	 * See perf_output_begin().
 	 */
 	smp_wmb(); /* B, matches C */
-	rb->user_page->data_head = head;
+	WRITE_ONCE(rb->user_page->data_head, head);
 
 	/*
 	 * We must publish the head before decrementing the nest count,
@@ -496,7 +496,7 @@ void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
 		perf_event_aux_event(handle->event, aux_head, size,
 				     handle->aux_flags);
 
-	rb->user_page->aux_head = rb->aux_head;
+	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
 	if (rb_need_aux_wakeup(rb))
 		wakeup = true;
 
@@ -528,7 +528,7 @@ int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
 
 	rb->aux_head += size;
 
-	rb->user_page->aux_head = rb->aux_head;
+	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
 	if (rb_need_aux_wakeup(rb)) {
 		perf_output_wakeup(handle);
 		handle->wakeup = rb->aux_wakeup + rb->aux_watermark;

[PATCH 4/4] perf/ring-buffer: Use regular variables for nesting

[Peter Zijlstra]

While the IRQ/NMI will nest, the nest-count will be invariant over the
actual exception, since it will decrement equal to increment.

This means we can -- carefully -- use a regular variable since the
typical LOAD-STORE race doesn't exist (similar to preempt_count).

This optimizes the ring-buffer for all LOAD-STORE architectures, since
they need to use atomic ops to implement local_t.

Suggested-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/events/internal.h    |    4 ++--
 kernel/events/ring_buffer.c |   41 ++++++++++++++++++++++++++---------------
 2 files changed, 28 insertions(+), 17 deletions(-)

--- a/kernel/events/internal.h
+++ b/kernel/events/internal.h
@@ -24,7 +24,7 @@ struct ring_buffer {
 	atomic_t			poll;		/* POLL_ for wakeups */
 
 	local_t				head;		/* write position    */
-	local_t				nest;		/* nested writers    */
+	unsigned int			nest;		/* nested writers    */
 	local_t				events;		/* event limit       */
 	local_t				wakeup;		/* wakeup stamp      */
 	local_t				lost;		/* nr records lost   */
@@ -41,7 +41,7 @@ struct ring_buffer {
 
 	/* AUX area */
 	long				aux_head;
-	local_t				aux_nest;
+	unsigned int			aux_nest;
 	long				aux_wakeup;	/* last aux_watermark boundary crossed by aux_head */
 	unsigned long			aux_pgoff;
 	int				aux_nr_pages;
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -38,7 +38,12 @@ static void perf_output_get_handle(struc
 	struct ring_buffer *rb = handle->rb;
 
 	preempt_disable();
-	local_inc(&rb->nest);
+
+	/*
+	 * Avoid an explicit LOAD/STORE such that architectures with memops
+	 * can use them.
+	 */
+	(*(volatile unsigned int *)&rb->nest)++;
 	handle->wakeup = local_read(&rb->wakeup);
 }
 
@@ -46,6 +51,17 @@ static void perf_output_put_handle(struc
 {
 	struct ring_buffer *rb = handle->rb;
 	unsigned long head;
+	unsigned int nest;
+
+	/*
+	 * If this isn't the outermost nesting, we don't have to update
+	 * @rb->user_page->data_head.
+	 */
+	nest = READ_ONCE(rb->nest);
+	if (nest > 1) {
+		WRITE_ONCE(rb->nest, nest - 1);
+		goto out;
+	}
 
 again:
 	/*
@@ -65,15 +81,6 @@ static void perf_output_put_handle(struc
 	 */
 
 	/*
-	 * If this isn't the outermost nesting, we don't have to update
-	 * @rb->user_page->data_head.
-	 */
-	if (local_read(&rb->nest) > 1) {
-		local_dec(&rb->nest);
-		goto out;
-	}
-
-	/*
 	 * Since the mmap() consumer (userspace) can run on a different CPU:
 	 *
 	 *   kernel				user
@@ -108,7 +115,7 @@ static void perf_output_put_handle(struc
 	 * write will (temporarily) publish a stale value.
 	 */
 	barrier();
-	local_set(&rb->nest, 0);
+	WRITE_ONCE(rb->nest, 0);
 
 	/*
 	 * Ensure we decrement @rb->nest before we validate the @rb->head.
@@ -116,7 +123,7 @@ static void perf_output_put_handle(struc
 	 */
 	barrier();
 	if (unlikely(head != local_read(&rb->head))) {
-		local_inc(&rb->nest);
+		WRITE_ONCE(rb->nest, 1);
 		goto again;
 	}
 
@@ -355,6 +362,7 @@ void *perf_aux_output_begin(struct perf_
 	struct perf_event *output_event = event;
 	unsigned long aux_head, aux_tail;
 	struct ring_buffer *rb;
+	unsigned int nest;
 
 	if (output_event->parent)
 		output_event = output_event->parent;
@@ -385,13 +393,16 @@ void *perf_aux_output_begin(struct perf_
 	if (!refcount_inc_not_zero(&rb->aux_refcount))
 		goto err;
 
+	nest = READ_ONCE(rb->aux_nest);
 	/*
 	 * Nesting is not supported for AUX area, make sure nested
 	 * writers are caught early
 	 */
-	if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
+	if (WARN_ON_ONCE(nest))
 		goto err_put;
 
+	WRITE_ONCE(rb->aux_nest, nest + 1);
+
 	aux_head = rb->aux_head;
 
 	handle->rb = rb;
@@ -419,7 +430,7 @@ void *perf_aux_output_begin(struct perf_
 		if (!handle->size) { /* A, matches D */
 			event->pending_disable = smp_processor_id();
 			perf_output_wakeup(handle);
-			local_set(&rb->aux_nest, 0);
+			WRITE_ONCE(rb->aux_nest, 0);
 			goto err_put;
 		}
 	}
@@ -508,7 +519,7 @@ void perf_aux_output_end(struct perf_out
 
 	handle->event = NULL;
 
-	local_set(&rb->aux_nest, 0);
+	WRITE_ONCE(rb->aux_nest, 0);
 	/* can't be last */
 	rb_free_aux(rb);
 	ring_buffer_put(rb);

[tip:perf/urgent] perf/ring-buffer: Use regular variables for nesting

[tip-bot for Peter Zijlstra]

Commit-ID:  5322ea58a06da2e69c5ef36a9b4d4b9255edd423
Gitweb:     https://git.kernel.org/tip/5322ea58a06da2e69c5ef36a9b4d4b9255edd423
Author:     Peter Zijlstra <peterz@infradead.org>
AuthorDate: Fri, 17 May 2019 13:52:34 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 24 May 2019 09:00:11 +0200

perf/ring-buffer: Use regular variables for nesting

While the IRQ/NMI will nest, the nest-count will be invariant over the
actual exception, since it will decrement equal to increment.

This means we can -- carefully -- use a regular variable since the
typical LOAD-STORE race doesn't exist (similar to preempt_count).

This optimizes the ring-buffer for all LOAD-STORE architectures, since
they need to use atomic ops to implement local_t.

Suggested-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: acme@kernel.org
Cc: mark.rutland@arm.com
Cc: namhyung@kernel.org
Cc: yabinc@google.com
Link: http://lkml.kernel.org/r/20190517115418.481392777@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/internal.h    |  4 ++--
 kernel/events/ring_buffer.c | 41 ++++++++++++++++++++++++++---------------
 2 files changed, 28 insertions(+), 17 deletions(-)

diff --git a/kernel/events/internal.h b/kernel/events/internal.h
index 79c47076700a..3aef4191798c 100644
--- a/kernel/events/internal.h
+++ b/kernel/events/internal.h
@@ -24,7 +24,7 @@ struct ring_buffer {
 	atomic_t			poll;		/* POLL_ for wakeups */
 
 	local_t				head;		/* write position    */
-	local_t				nest;		/* nested writers    */
+	unsigned int			nest;		/* nested writers    */
 	local_t				events;		/* event limit       */
 	local_t				wakeup;		/* wakeup stamp      */
 	local_t				lost;		/* nr records lost   */
@@ -41,7 +41,7 @@ struct ring_buffer {
 
 	/* AUX area */
 	long				aux_head;
-	local_t				aux_nest;
+	unsigned int			aux_nest;
 	long				aux_wakeup;	/* last aux_watermark boundary crossed by aux_head */
 	unsigned long			aux_pgoff;
 	int				aux_nr_pages;
diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 7a0c73e4b3eb..ffb59a4ef4ff 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -38,7 +38,12 @@ static void perf_output_get_handle(struct perf_output_handle *handle)
 	struct ring_buffer *rb = handle->rb;
 
 	preempt_disable();
-	local_inc(&rb->nest);
+
+	/*
+	 * Avoid an explicit LOAD/STORE such that architectures with memops
+	 * can use them.
+	 */
+	(*(volatile unsigned int *)&rb->nest)++;
 	handle->wakeup = local_read(&rb->wakeup);
 }
 
@@ -46,6 +51,17 @@ static void perf_output_put_handle(struct perf_output_handle *handle)
 {
 	struct ring_buffer *rb = handle->rb;
 	unsigned long head;
+	unsigned int nest;
+
+	/*
+	 * If this isn't the outermost nesting, we don't have to update
+	 * @rb->user_page->data_head.
+	 */
+	nest = READ_ONCE(rb->nest);
+	if (nest > 1) {
+		WRITE_ONCE(rb->nest, nest - 1);
+		goto out;
+	}
 
 again:
 	/*
@@ -64,15 +80,6 @@ again:
 	 * load above to be stale.
 	 */
 
-	/*
-	 * If this isn't the outermost nesting, we don't have to update
-	 * @rb->user_page->data_head.
-	 */
-	if (local_read(&rb->nest) > 1) {
-		local_dec(&rb->nest);
-		goto out;
-	}
-
 	/*
 	 * Since the mmap() consumer (userspace) can run on a different CPU:
 	 *
@@ -108,7 +115,7 @@ again:
 	 * write will (temporarily) publish a stale value.
 	 */
 	barrier();
-	local_set(&rb->nest, 0);
+	WRITE_ONCE(rb->nest, 0);
 
 	/*
 	 * Ensure we decrement @rb->nest before we validate the @rb->head.
@@ -116,7 +123,7 @@ again:
 	 */
 	barrier();
 	if (unlikely(head != local_read(&rb->head))) {
-		local_inc(&rb->nest);
+		WRITE_ONCE(rb->nest, 1);
 		goto again;
 	}
 
@@ -355,6 +362,7 @@ void *perf_aux_output_begin(struct perf_output_handle *handle,
 	struct perf_event *output_event = event;
 	unsigned long aux_head, aux_tail;
 	struct ring_buffer *rb;
+	unsigned int nest;
 
 	if (output_event->parent)
 		output_event = output_event->parent;
@@ -385,13 +393,16 @@ void *perf_aux_output_begin(struct perf_output_handle *handle,
 	if (!refcount_inc_not_zero(&rb->aux_refcount))
 		goto err;
 
+	nest = READ_ONCE(rb->aux_nest);
 	/*
 	 * Nesting is not supported for AUX area, make sure nested
 	 * writers are caught early
 	 */
-	if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
+	if (WARN_ON_ONCE(nest))
 		goto err_put;
 
+	WRITE_ONCE(rb->aux_nest, nest + 1);
+
 	aux_head = rb->aux_head;
 
 	handle->rb = rb;
@@ -419,7 +430,7 @@ void *perf_aux_output_begin(struct perf_output_handle *handle,
 		if (!handle->size) { /* A, matches D */
 			event->pending_disable = smp_processor_id();
 			perf_output_wakeup(handle);
-			local_set(&rb->aux_nest, 0);
+			WRITE_ONCE(rb->aux_nest, 0);
 			goto err_put;
 		}
 	}
@@ -508,7 +519,7 @@ void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
 
 	handle->event = NULL;
 
-	local_set(&rb->aux_nest, 0);
+	WRITE_ONCE(rb->aux_nest, 0);
 	/* can't be last */
 	rb_free_aux(rb);
 	ring_buffer_put(rb);