perf/aux: Ensure aux_wakeup represents most recent wakeup index
authorWill Deacon <will.deacon@arm.com>
Wed, 16 Aug 2017 16:18:17 +0000 (17:18 +0100)
committerIngo Molnar <mingo@kernel.org>
Fri, 25 Aug 2017 09:04:16 +0000 (11:04 +0200)
The aux_watermark member of struct ring_buffer represents the period (in
terms of bytes) at which wakeup events should be generated when data is
written to the aux buffer in non-snapshot mode. On hardware that cannot
generate an interrupt when the aux_head reaches an arbitrary wakeup index
(such as ARM SPE), the aux_head sampled from handle->head in
perf_aux_output_{skip,end} may in fact be past the wakeup index. This
can lead to wakeup slowly falling behind the head. For example, consider
the case where hardware can only generate an interrupt on a page-boundary
and the aux buffer is initialised as follows:

  // Buffer size is 2 * PAGE_SIZE
  rb->aux_head = rb->aux_wakeup = 0
  rb->aux_watermark = PAGE_SIZE / 2

following the first perf_aux_output_begin call, the handle is
initialised with:

  handle->head = 0
  handle->size = 2 * PAGE_SIZE
  handle->wakeup = PAGE_SIZE / 2

and the hardware will be programmed to generate an interrupt at
PAGE_SIZE.

When the interrupt is raised, the hardware head will be at PAGE_SIZE,
so calling perf_aux_output_end(handle, PAGE_SIZE) puts the ring buffer
into the following state:

  rb->aux_head = PAGE_SIZE
  rb->aux_wakeup = PAGE_SIZE / 2
  rb->aux_watermark = PAGE_SIZE / 2

and then the next call to perf_aux_output_begin will result in:

  handle->head = handle->wakeup = PAGE_SIZE

for which the semantics are unclear and, for a smaller aux_watermark
(e.g. PAGE_SIZE / 4), then the wakeup would in fact be behind head at
this point.

This patch fixes the problem by rounding down the aux_head (as sampled
from the handle) to the nearest aux_watermark boundary when updating
rb->aux_wakeup, therefore taking into account any overruns by the
hardware.

Reported-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-arm-kernel@lists.infradead.org
Link: http://lkml.kernel.org/r/1502900297-21839-2-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/events/internal.h
kernel/events/ring_buffer.c

index 2941b868353ca6e159ad4344a0e6b65d77125273..5377c591c57a25f3d0dc989ede53f8e5299bf3c7 100644 (file)
@@ -40,7 +40,7 @@ struct ring_buffer {
        /* AUX area */
        long                            aux_head;
        local_t                         aux_nest;
-       long                            aux_wakeup;
+       long                            aux_wakeup;     /* last aux_watermark boundary crossed by aux_head */
        unsigned long                   aux_pgoff;
        int                             aux_nr_pages;
        int                             aux_overwrite;
index 25437fda56e31e0db028a4b4961fe3818e94f2cd..af71a84e12eea343c6047184599fedf7f5592e65 100644 (file)
@@ -453,7 +453,7 @@ void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
        rb->user_page->aux_head = rb->aux_head;
        if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
                wakeup = true;
-               rb->aux_wakeup += rb->aux_watermark;
+               rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
        }
 
        if (wakeup) {
@@ -486,7 +486,7 @@ int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
        rb->user_page->aux_head = rb->aux_head;
        if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
                perf_output_wakeup(handle);
-               rb->aux_wakeup += rb->aux_watermark;
+               rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
                handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
        }