Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 5:07 PM, Peter Zijlstra wrote:

On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:

This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto reload.
Both period_left and event->count are carefully recalculated for auto
reload.


How does prev_count make sense when we've already reloaded a bunch of
times?


Same as non-auto-reload, it's the 'left' (unfinished) period from last time.
The period for the first record should always be the 'left' period no 
matter on which case.
For auto-reload, it doesn't need to increase the prev_count with the 
reload. Because for later records, the period should be exactly the same 
as the reload value.


To calculate the event->count,
For auto-reload, the event->count = prev_count + (reload times - 1) * 
reload value + gap between PMI trigger and PMI handler.


For non-auto-reload, the event->count = prev_count + gap between PMI 
trigger and PMI handler.


The 'prev_count' is same for both auto-reload and non-auto-reload.

The gap is a little bit tricky for auto-reload. Because it starts from 
-reload_value. But for non-auto-reload, it starts from 0.

"delta += (reload_val << shift);" is used to correct it.




For x86_perf_event_set_period(), there is nothing special needed for auto
reload. The period is fixed. The period_left from x86_perf_event_update() is
already handled.


Hurm.. I see. But rather than make an ever bigger trainwreck of things,
I'd rather you just write a special purpose intel_pmu_save_and_restart()
just for AUTO_RELOAD.


OK. I will do it in V2.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 5:07 PM, Peter Zijlstra wrote:

On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:

This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto reload.
Both period_left and event->count are carefully recalculated for auto
reload.


How does prev_count make sense when we've already reloaded a bunch of
times?


Same as non-auto-reload, it's the 'left' (unfinished) period from last time.
The period for the first record should always be the 'left' period no 
matter on which case.
For auto-reload, it doesn't need to increase the prev_count with the 
reload. Because for later records, the period should be exactly the same 
as the reload value.


To calculate the event->count,
For auto-reload, the event->count = prev_count + (reload times - 1) * 
reload value + gap between PMI trigger and PMI handler.


For non-auto-reload, the event->count = prev_count + gap between PMI 
trigger and PMI handler.


The 'prev_count' is same for both auto-reload and non-auto-reload.

The gap is a little bit tricky for auto-reload. Because it starts from 
-reload_value. But for non-auto-reload, it starts from 0.

"delta += (reload_val << shift);" is used to correct it.




For x86_perf_event_set_period(), there is nothing special needed for auto
reload. The period is fixed. The period_left from x86_perf_event_update() is
already handled.


Hurm.. I see. But rather than make an ever bigger trainwreck of things,
I'd rather you just write a special purpose intel_pmu_save_and_restart()
just for AUTO_RELOAD.


OK. I will do it in V2.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Andi Kleen]

On Tue, Dec 19, 2017 at 11:07:09PM +0100, Peter Zijlstra wrote:
> On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:
> > > This all looks very wrong... In auto reload we should never call
> > > intel_pmu_save_and_restore() in the first place I think.
> > > 
> > > Things like x86_perf_event_update() and x86_perf_event_set_period()
> > > simply _cannot_ do the right thing when we auto reload the counter.
> > > 
> > 
> > I think it should be OK to call it in first place.
> > For x86_perf_event_update(), the reload_times will tell if it's auto reload.
> > Both period_left and event->count are carefully recalculated for auto
> > reload.
> 
> How does prev_count make sense when we've already reloaded a bunch of
> times?

We can figure it out how often there was a reload based on the PEBS index.

-Andi

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Andi Kleen]

On Tue, Dec 19, 2017 at 11:07:09PM +0100, Peter Zijlstra wrote:
> On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:
> > > This all looks very wrong... In auto reload we should never call
> > > intel_pmu_save_and_restore() in the first place I think.
> > > 
> > > Things like x86_perf_event_update() and x86_perf_event_set_period()
> > > simply _cannot_ do the right thing when we auto reload the counter.
> > > 
> > 
> > I think it should be OK to call it in first place.
> > For x86_perf_event_update(), the reload_times will tell if it's auto reload.
> > Both period_left and event->count are carefully recalculated for auto
> > reload.
> 
> How does prev_count make sense when we've already reloaded a bunch of
> times?

We can figure it out how often there was a reload based on the PEBS index.

-Andi

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Peter Zijlstra]

On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:
> > This all looks very wrong... In auto reload we should never call
> > intel_pmu_save_and_restore() in the first place I think.
> > 
> > Things like x86_perf_event_update() and x86_perf_event_set_period()
> > simply _cannot_ do the right thing when we auto reload the counter.
> > 
> 
> I think it should be OK to call it in first place.
> For x86_perf_event_update(), the reload_times will tell if it's auto reload.
> Both period_left and event->count are carefully recalculated for auto
> reload.

How does prev_count make sense when we've already reloaded a bunch of
times?

> For x86_perf_event_set_period(), there is nothing special needed for auto
> reload. The period is fixed. The period_left from x86_perf_event_update() is
> already handled.

Hurm.. I see. But rather than make an ever bigger trainwreck of things,
I'd rather you just write a special purpose intel_pmu_save_and_restart()
just for AUTO_RELOAD.

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Peter Zijlstra]

On Tue, Dec 19, 2017 at 03:08:58PM -0500, Liang, Kan wrote:
> > This all looks very wrong... In auto reload we should never call
> > intel_pmu_save_and_restore() in the first place I think.
> > 
> > Things like x86_perf_event_update() and x86_perf_event_set_period()
> > simply _cannot_ do the right thing when we auto reload the counter.
> > 
> 
> I think it should be OK to call it in first place.
> For x86_perf_event_update(), the reload_times will tell if it's auto reload.
> Both period_left and event->count are carefully recalculated for auto
> reload.

How does prev_count make sense when we've already reloaded a bunch of
times?

> For x86_perf_event_set_period(), there is nothing special needed for auto
> reload. The period is fixed. The period_left from x86_perf_event_update() is
> already handled.

Hurm.. I see. But rather than make an ever bigger trainwreck of things,
I'd rather you just write a special purpose intel_pmu_save_and_restart()
just for AUTO_RELOAD.

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 3:08 PM, Liang, Kan wrote:



On 12/19/2017 1:58 PM, Peter Zijlstra wrote:
On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com 
wrote:

  arch/x86/events/core.c | 14 ++
  arch/x86/events/intel/ds.c |  8 +++-
  2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
   * of the count.
   */
  delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+    /*
+ * Take auto-reload into account
+ * For the auto-reload before the last time, it went through the
+ * whole period (reload_val) every time.
+ * Just simply add period * times to the event.
+ *
+ * For the last load, the elapsed delta (event-)time need to be
+ * corrected by adding the period. Because the start point is 
-period.

+ */
+    if (reload_times > 0) {
+    delta += (reload_val << shift);
+    local64_add(reload_val * (reload_times - 1), >count);
+    }
  delta >>= shift;
  local64_add(delta, >count);
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct 
perf_event *event,

 void *base, void *top,
 int bit, int count)
  {
+    struct hw_perf_event *hwc = >hw;
  struct perf_sample_data data;
  struct pt_regs regs;
  void *at = get_next_pebs_record_by_bit(base, top, bit);
-    if (!intel_pmu_save_and_restart(event, 0, 0) &&
+    /*
+ * Now, auto-reload is only enabled in fixed period mode.
+ * The reload value is always hwc->sample_period.
+ * May need to change it, if auto-reload is enabled in freq mode 
later.

+ */
+    if (!intel_pmu_save_and_restart(event, hwc->sample_period, count 
- 1) &&

  !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
  return;


This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto 
reload. Both period_left and event->count are carefully recalculated for 
auto reload.


Think a bit more. You are right. We cannot rely on count to tell us if 
it's auto reload.

The count could also be 1 if auto reload is enabled.

I will fix it in V2.

Thanks,
Kan

For x86_perf_event_set_period(), there is nothing special needed for 
auto reload. The period is fixed. The period_left from 
x86_perf_event_update() is already handled.



BTW: It should be 'count' not 'count - 1' which pass to
intel_pmu_save_and_restart(). I just found the issue. I will fix it in
V2 with other improvements if there are any.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 3:08 PM, Liang, Kan wrote:



On 12/19/2017 1:58 PM, Peter Zijlstra wrote:
On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com 
wrote:

  arch/x86/events/core.c | 14 ++
  arch/x86/events/intel/ds.c |  8 +++-
  2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
   * of the count.
   */
  delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+    /*
+ * Take auto-reload into account
+ * For the auto-reload before the last time, it went through the
+ * whole period (reload_val) every time.
+ * Just simply add period * times to the event.
+ *
+ * For the last load, the elapsed delta (event-)time need to be
+ * corrected by adding the period. Because the start point is 
-period.

+ */
+    if (reload_times > 0) {
+    delta += (reload_val << shift);
+    local64_add(reload_val * (reload_times - 1), >count);
+    }
  delta >>= shift;
  local64_add(delta, >count);
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct 
perf_event *event,

 void *base, void *top,
 int bit, int count)
  {
+    struct hw_perf_event *hwc = >hw;
  struct perf_sample_data data;
  struct pt_regs regs;
  void *at = get_next_pebs_record_by_bit(base, top, bit);
-    if (!intel_pmu_save_and_restart(event, 0, 0) &&
+    /*
+ * Now, auto-reload is only enabled in fixed period mode.
+ * The reload value is always hwc->sample_period.
+ * May need to change it, if auto-reload is enabled in freq mode 
later.

+ */
+    if (!intel_pmu_save_and_restart(event, hwc->sample_period, count 
- 1) &&

  !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
  return;


This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto 
reload. Both period_left and event->count are carefully recalculated for 
auto reload.


Think a bit more. You are right. We cannot rely on count to tell us if 
it's auto reload.

The count could also be 1 if auto reload is enabled.

I will fix it in V2.

Thanks,
Kan

For x86_perf_event_set_period(), there is nothing special needed for 
auto reload. The period is fixed. The period_left from 
x86_perf_event_update() is already handled.



BTW: It should be 'count' not 'count - 1' which pass to
intel_pmu_save_and_restart(). I just found the issue. I will fix it in
V2 with other improvements if there are any.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 1:58 PM, Peter Zijlstra wrote:

On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com wrote:

  arch/x86/events/core.c | 14 ++
  arch/x86/events/intel/ds.c |  8 +++-
  2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
 * of the count.
 */
delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+   /*
+* Take auto-reload into account
+* For the auto-reload before the last time, it went through the
+* whole period (reload_val) every time.
+* Just simply add period * times to the event.
+*
+* For the last load, the elapsed delta (event-)time need to be
+* corrected by adding the period. Because the start point is -period.
+*/
+   if (reload_times > 0) {
+   delta += (reload_val << shift);
+   local64_add(reload_val * (reload_times - 1), >count);
+   }
delta >>= shift;
  
  	local64_add(delta, >count);

diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
*event,
   void *base, void *top,
   int bit, int count)
  {
+   struct hw_perf_event *hwc = >hw;
struct perf_sample_data data;
struct pt_regs regs;
void *at = get_next_pebs_record_by_bit(base, top, bit);
  
-	if (!intel_pmu_save_and_restart(event, 0, 0) &&

+   /*
+* Now, auto-reload is only enabled in fixed period mode.
+* The reload value is always hwc->sample_period.
+* May need to change it, if auto-reload is enabled in freq mode later.
+*/
+   if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
return;
  


This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto 
reload. Both period_left and event->count are carefully recalculated for 
auto reload.
For x86_perf_event_set_period(), there is nothing special needed for 
auto reload. The period is fixed. The period_left from 
x86_perf_event_update() is already handled.



BTW: It should be 'count' not 'count - 1' which pass to
intel_pmu_save_and_restart(). I just found the issue. I will fix it in
V2 with other improvements if there are any.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 1:58 PM, Peter Zijlstra wrote:

On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com wrote:

  arch/x86/events/core.c | 14 ++
  arch/x86/events/intel/ds.c |  8 +++-
  2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
 * of the count.
 */
delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+   /*
+* Take auto-reload into account
+* For the auto-reload before the last time, it went through the
+* whole period (reload_val) every time.
+* Just simply add period * times to the event.
+*
+* For the last load, the elapsed delta (event-)time need to be
+* corrected by adding the period. Because the start point is -period.
+*/
+   if (reload_times > 0) {
+   delta += (reload_val << shift);
+   local64_add(reload_val * (reload_times - 1), >count);
+   }
delta >>= shift;
  
  	local64_add(delta, >count);

diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
*event,
   void *base, void *top,
   int bit, int count)
  {
+   struct hw_perf_event *hwc = >hw;
struct perf_sample_data data;
struct pt_regs regs;
void *at = get_next_pebs_record_by_bit(base, top, bit);
  
-	if (!intel_pmu_save_and_restart(event, 0, 0) &&

+   /*
+* Now, auto-reload is only enabled in fixed period mode.
+* The reload value is always hwc->sample_period.
+* May need to change it, if auto-reload is enabled in freq mode later.
+*/
+   if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
return;
  


This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.



I think it should be OK to call it in first place.
For x86_perf_event_update(), the reload_times will tell if it's auto 
reload. Both period_left and event->count are carefully recalculated for 
auto reload.
For x86_perf_event_set_period(), there is nothing special needed for 
auto reload. The period is fixed. The period_left from 
x86_perf_event_update() is already handled.



BTW: It should be 'count' not 'count - 1' which pass to
intel_pmu_save_and_restart(). I just found the issue. I will fix it in
V2 with other improvements if there are any.

Thanks,
Kan

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Peter Zijlstra]

On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com wrote:
>  arch/x86/events/core.c | 14 ++
>  arch/x86/events/intel/ds.c |  8 +++-
>  2 files changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 35552ea..f74e21d 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
>* of the count.
>*/
>   delta = (new_raw_count << shift) - (prev_raw_count << shift);
> +
> + /*
> +  * Take auto-reload into account
> +  * For the auto-reload before the last time, it went through the
> +  * whole period (reload_val) every time.
> +  * Just simply add period * times to the event.
> +  *
> +  * For the last load, the elapsed delta (event-)time need to be
> +  * corrected by adding the period. Because the start point is -period.
> +  */
> + if (reload_times > 0) {
> + delta += (reload_val << shift);
> + local64_add(reload_val * (reload_times - 1), >count);
> + }
>   delta >>= shift;
>  
>   local64_add(delta, >count);
> diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
> index 0b693b7..f0f6026 100644
> --- a/arch/x86/events/intel/ds.c
> +++ b/arch/x86/events/intel/ds.c
> @@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
> *event,
>  void *base, void *top,
>  int bit, int count)
>  {
> + struct hw_perf_event *hwc = >hw;
>   struct perf_sample_data data;
>   struct pt_regs regs;
>   void *at = get_next_pebs_record_by_bit(base, top, bit);
>  
> - if (!intel_pmu_save_and_restart(event, 0, 0) &&
> + /*
> +  * Now, auto-reload is only enabled in fixed period mode.
> +  * The reload value is always hwc->sample_period.
> +  * May need to change it, if auto-reload is enabled in freq mode later.
> +  */
> + if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
>   !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
>   return;
>  

This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.

Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Peter Zijlstra]

On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.li...@linux.intel.com wrote:
>  arch/x86/events/core.c | 14 ++
>  arch/x86/events/intel/ds.c |  8 +++-
>  2 files changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 35552ea..f74e21d 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
>* of the count.
>*/
>   delta = (new_raw_count << shift) - (prev_raw_count << shift);
> +
> + /*
> +  * Take auto-reload into account
> +  * For the auto-reload before the last time, it went through the
> +  * whole period (reload_val) every time.
> +  * Just simply add period * times to the event.
> +  *
> +  * For the last load, the elapsed delta (event-)time need to be
> +  * corrected by adding the period. Because the start point is -period.
> +  */
> + if (reload_times > 0) {
> + delta += (reload_val << shift);
> + local64_add(reload_val * (reload_times - 1), >count);
> + }
>   delta >>= shift;
>  
>   local64_add(delta, >count);
> diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
> index 0b693b7..f0f6026 100644
> --- a/arch/x86/events/intel/ds.c
> +++ b/arch/x86/events/intel/ds.c
> @@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
> *event,
>  void *base, void *top,
>  int bit, int count)
>  {
> + struct hw_perf_event *hwc = >hw;
>   struct perf_sample_data data;
>   struct pt_regs regs;
>   void *at = get_next_pebs_record_by_bit(base, top, bit);
>  
> - if (!intel_pmu_save_and_restart(event, 0, 0) &&
> + /*
> +  * Now, auto-reload is only enabled in fixed period mode.
> +  * The reload value is always hwc->sample_period.
> +  * May need to change it, if auto-reload is enabled in freq mode later.
> +  */
> + if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
>   !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
>   return;
>  

This all looks very wrong... In auto reload we should never call
intel_pmu_save_and_restore() in the first place I think.

Things like x86_perf_event_update() and x86_perf_event_set_period()
simply _cannot_ do the right thing when we auto reload the counter.

[PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[kan . liang]

From: Kan Liang 

There is bug when mmap read event->count with large PEBS enabled.
Here is an example.
 #./read_count
 0x71f0
 0x122c0
 0x11c54
 0x10001257d
 0x2bdc5

There is auto-reload mechanism enabled for PEBS events in fixed period
mode. But the calculation of event->count does not take the auto-reload
values into account. Anyone who read the event->count will get wrong
result, e.g x86_pmu_read. Also, the calculation of hwc->period_left is
wrong either. It impacts the accuracy of period for the first record in
PEBS multiple records.

The issue is introduced with the auto-reload mechanism enabled by
commit 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism
when possible")

For the auto-reload before the last time, it went through the whole
period (reload value) every time. So period * times should be added into
the event->count.
For the last load, the elapsed delta (event-)time need to be corrected
by adding the period (reload value). Because the start point is -period.

Signed-off-by: Kan Liang 
---
 arch/x86/events/core.c | 14 ++
 arch/x86/events/intel/ds.c |  8 +++-
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
 * of the count.
 */
delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+   /*
+* Take auto-reload into account
+* For the auto-reload before the last time, it went through the
+* whole period (reload_val) every time.
+* Just simply add period * times to the event.
+*
+* For the last load, the elapsed delta (event-)time need to be
+* corrected by adding the period. Because the start point is -period.
+*/
+   if (reload_times > 0) {
+   delta += (reload_val << shift);
+   local64_add(reload_val * (reload_times - 1), >count);
+   }
delta >>= shift;
 
local64_add(delta, >count);
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
*event,
   void *base, void *top,
   int bit, int count)
 {
+   struct hw_perf_event *hwc = >hw;
struct perf_sample_data data;
struct pt_regs regs;
void *at = get_next_pebs_record_by_bit(base, top, bit);
 
-   if (!intel_pmu_save_and_restart(event, 0, 0) &&
+   /*
+* Now, auto-reload is only enabled in fixed period mode.
+* The reload value is always hwc->sample_period.
+* May need to change it, if auto-reload is enabled in freq mode later.
+*/
+   if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
return;
 
-- 
2.7.4

[PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[kan . liang]

From: Kan Liang 

There is bug when mmap read event->count with large PEBS enabled.
Here is an example.
 #./read_count
 0x71f0
 0x122c0
 0x11c54
 0x10001257d
 0x2bdc5

There is auto-reload mechanism enabled for PEBS events in fixed period
mode. But the calculation of event->count does not take the auto-reload
values into account. Anyone who read the event->count will get wrong
result, e.g x86_pmu_read. Also, the calculation of hwc->period_left is
wrong either. It impacts the accuracy of period for the first record in
PEBS multiple records.

The issue is introduced with the auto-reload mechanism enabled by
commit 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism
when possible")

For the auto-reload before the last time, it went through the whole
period (reload value) every time. So period * times should be added into
the event->count.
For the last load, the elapsed delta (event-)time need to be corrected
by adding the period (reload value). Because the start point is -period.

Signed-off-by: Kan Liang 
---
 arch/x86/events/core.c | 14 ++
 arch/x86/events/intel/ds.c |  8 +++-
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 35552ea..f74e21d 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
 * of the count.
 */
delta = (new_raw_count << shift) - (prev_raw_count << shift);
+
+   /*
+* Take auto-reload into account
+* For the auto-reload before the last time, it went through the
+* whole period (reload_val) every time.
+* Just simply add period * times to the event.
+*
+* For the last load, the elapsed delta (event-)time need to be
+* corrected by adding the period. Because the start point is -period.
+*/
+   if (reload_times > 0) {
+   delta += (reload_val << shift);
+   local64_add(reload_val * (reload_times - 1), >count);
+   }
delta >>= shift;
 
local64_add(delta, >count);
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 0b693b7..f0f6026 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct perf_event 
*event,
   void *base, void *top,
   int bit, int count)
 {
+   struct hw_perf_event *hwc = >hw;
struct perf_sample_data data;
struct pt_regs regs;
void *at = get_next_pebs_record_by_bit(base, top, bit);
 
-   if (!intel_pmu_save_and_restart(event, 0, 0) &&
+   /*
+* Now, auto-reload is only enabled in fixed period mode.
+* The reload value is always hwc->sample_period.
+* May need to change it, if auto-reload is enabled in freq mode later.
+*/
+   if (!intel_pmu_save_and_restart(event, hwc->sample_period, count - 1) &&
!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
return;
 
-- 
2.7.4