2016-11-30 10:51:22

by Jiri Olsa

[permalink] [raw]
Subject: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question

hi,
I'm trying to find out some documentation background for this part of uncore code:

---
static int uncore_pmu_event_init(struct perf_event *event)
{
...
if (event->attr.config == UNCORE_FIXED_EVENT) {
/* no fixed counter */
if (!pmu->type->fixed_ctl)
return -EINVAL;
/*
* if there is only one fixed counter, only the first pmu
* can access the fixed counter
*/
if (pmu->type->single_fixed && pmu->pmu_idx > 0)
return -EINVAL;
...
---

that for some uncore types (those with pmu->type->single_fixed) only
the first pmu (code_id == 0) will allow to touch the clocktick event

other cores boxes will not allow to open clocktick event, eventhough
it's announced via /sys/../events/..

I'm probably missing some HW logic of specific boxes that would explain
that, but I can't find it.

thanks for info,
jirka


2016-11-30 14:45:27

by Liang, Kan

[permalink] [raw]
Subject: RE: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question



> hi,
> I'm trying to find out some documentation background for this part of
> uncore code:
>
> ---
> static int uncore_pmu_event_init(struct perf_event *event) {
> ...
> if (event->attr.config == UNCORE_FIXED_EVENT) {
> /* no fixed counter */
> if (!pmu->type->fixed_ctl)
> return -EINVAL;
> /*
> * if there is only one fixed counter, only the first pmu
> * can access the fixed counter
> */
> if (pmu->type->single_fixed && pmu->pmu_idx > 0)
> return -EINVAL;
> ...
> ---
>
> that for some uncore types (those with pmu->type->single_fixed) only the
> first pmu (code_id == 0) will allow to touch the clocktick event
>
> other cores boxes will not allow to open clocktick event, eventhough it's
> announced via /sys/../events/..
>
> I'm probably missing some HW logic of specific boxes that would explain
> that, but I can't find it.

The client uncore has a standalone clocktick fixed counter. It doesn't belong
to any boxes, which is different from server uncore.

But client and server uncore share the same uncore_pmu_event_init.
So it forces that only the first box can access the fixed counter.

Maybe we should create a clocktick box for client uncore to fix it.

You can find the fixed counter information from 18.11.6 in latest
SDM (Order Number: 325384-060US).

There should be a Skylake client uncore document published somewhere.
But I cannot find it from Google. Let me ask around.

Thanks,
Kan

2016-11-30 15:27:56

by Liang, Kan

[permalink] [raw]
Subject: RE: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question


>
> > hi,
> > I'm trying to find out some documentation background for this part of
> > uncore code:
> >
> > ---
> > static int uncore_pmu_event_init(struct perf_event *event) {
> > ...
> > if (event->attr.config == UNCORE_FIXED_EVENT) {
> > /* no fixed counter */
> > if (!pmu->type->fixed_ctl)
> > return -EINVAL;
> > /*
> > * if there is only one fixed counter, only the first pmu
> > * can access the fixed counter
> > */
> > if (pmu->type->single_fixed && pmu->pmu_idx > 0)
> > return -EINVAL;
> > ...
> > ---
> >
> > that for some uncore types (those with pmu->type->single_fixed) only
> > the first pmu (code_id == 0) will allow to touch the clocktick event
> >
> > other cores boxes will not allow to open clocktick event, eventhough
> > it's announced via /sys/../events/..
> >
> > I'm probably missing some HW logic of specific boxes that would
> > explain that, but I can't find it.
>
> The client uncore has a standalone clocktick fixed counter. It doesn't belong
> to any boxes, which is different from server uncore.
>
> But client and server uncore share the same uncore_pmu_event_init.
> So it forces that only the first box can access the fixed counter.
>
> Maybe we should create a clocktick box for client uncore to fix it.
>
> You can find the fixed counter information from 18.11.6 in latest SDM
> (Order Number: 325384-060US).
>
> There should be a Skylake client uncore document published somewhere.
> But I cannot find it from Google. Let me ask around.

Here is the published document.
http://www.intel.com/content/www/us/en/processors/core/6th-gen-core-family-uncore-performance-monitoring-manual.html



>
> Thanks,
> Kan

2016-11-30 15:38:30

by Jiri Olsa

[permalink] [raw]
Subject: Re: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question

On Wed, Nov 30, 2016 at 03:27:52PM +0000, Liang, Kan wrote:
>
> >
> > > hi,
> > > I'm trying to find out some documentation background for this part of
> > > uncore code:
> > >
> > > ---
> > > static int uncore_pmu_event_init(struct perf_event *event) {
> > > ...
> > > if (event->attr.config == UNCORE_FIXED_EVENT) {
> > > /* no fixed counter */
> > > if (!pmu->type->fixed_ctl)
> > > return -EINVAL;
> > > /*
> > > * if there is only one fixed counter, only the first pmu
> > > * can access the fixed counter
> > > */
> > > if (pmu->type->single_fixed && pmu->pmu_idx > 0)
> > > return -EINVAL;
> > > ...
> > > ---
> > >
> > > that for some uncore types (those with pmu->type->single_fixed) only
> > > the first pmu (code_id == 0) will allow to touch the clocktick event
> > >
> > > other cores boxes will not allow to open clocktick event, eventhough
> > > it's announced via /sys/../events/..
> > >
> > > I'm probably missing some HW logic of specific boxes that would
> > > explain that, but I can't find it.
> >
> > The client uncore has a standalone clocktick fixed counter. It doesn't belong
> > to any boxes, which is different from server uncore.
> >
> > But client and server uncore share the same uncore_pmu_event_init.
> > So it forces that only the first box can access the fixed counter.
> >
> > Maybe we should create a clocktick box for client uncore to fix it.
> >
> > You can find the fixed counter information from 18.11.6 in latest SDM
> > (Order Number: 325384-060US).
> >
> > There should be a Skylake client uncore document published somewhere.
> > But I cannot find it from Google. Let me ask around.
>
> Here is the published document.
> http://www.intel.com/content/www/us/en/processors/core/6th-gen-core-family-uncore-performance-monitoring-manual.html
>

will check, thanks a lot!

jirka

2016-11-30 17:52:24

by Andi Kleen

[permalink] [raw]
Subject: Re: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question

"Liang, Kan" <[email protected]> writes:
>
> Maybe we should create a clocktick box for client uncore to fix it.

That would break all the user space code that uses the existing format.

-Andi

2016-12-01 16:34:51

by Liang, Kan

[permalink] [raw]
Subject: RE: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question


>
> "Liang, Kan" <[email protected]> writes:
> >
> > Maybe we should create a clocktick box for client uncore to fix it.
>
> That would break all the user space code that uses the existing format.
>
OK.
If so, I think we should remove the non-exist events for non-first boxes.

What do you think about the patch as below?

Andi? Jirka?

------
>From 844c12b5e6fc1e2c27b2f094c89098ea9405ea41 Mon Sep 17 00:00:00 2001
From: Kan Liang <[email protected]>
Date: Thu, 1 Dec 2016 03:29:29 -0500
Subject: [PATCH] perf/x86/intel/uncore: remove non-exist clockticks events

For client cbox, only the first box can access the clockticks event.
Other cboxes will not allow to open clocktick event, eventhough it's
announced via /sys/../events/..

The client uncore has a standalone clocktick fixed counter. It doesn't
belong to any boxes, which is different from server uncore. To make the
code compatible, the client forces that only the first box can access
the fixed counter.

The clocktick event should be removed from other boxes.
Currently, all the pmus of same type share the same attr_groups, which
include all the events.
no_fixed_attr_groups is introduced for other boxes, which remove the
non-exist events.

Signed-off-by: Kan Liang <[email protected]>
---
arch/x86/events/intel/uncore.c | 70 ++++++++++++++++++++++++++++++++----------
arch/x86/events/intel/uncore.h | 1 +
2 files changed, 55 insertions(+), 16 deletions(-)

diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index dbaaf7dc..6d3606e 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -739,6 +739,9 @@ static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
pmu->pmu.attr_groups = pmu->type->attr_groups;
}

+ if (pmu->type->single_fixed && pmu->pmu_idx)
+ pmu->pmu.attr_groups = pmu->type->no_fixed_attr_groups;
+
if (pmu->type->num_boxes == 1) {
if (strlen(pmu->type->name) > 0)
sprintf(pmu->name, "uncore_%s", pmu->type->name);
@@ -811,6 +814,8 @@ static void uncore_type_exit(struct intel_uncore_type *type)
}
kfree(type->events_group);
type->events_group = NULL;
+ kfree(type->no_fixed_attr_groups[2]);
+ type->no_fixed_attr_groups[2] = NULL;
}

static void uncore_types_exit(struct intel_uncore_type **types)
@@ -819,13 +824,45 @@ static void uncore_types_exit(struct intel_uncore_type **types)
uncore_type_exit(*types);
}

-static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
+static struct attribute_group *
+uncore_alloc_event_group(struct intel_uncore_type *type,
+ bool single_fixed)
{
- struct intel_uncore_pmu *pmus;
struct attribute_group *attr_group;
struct attribute **attrs;
+ char fixed_event_name[11];
+ int i, j, index = 0;
+
+ if (single_fixed)
+ snprintf(fixed_event_name, 11, "event=0x%x", UNCORE_FIXED_EVENT);
+
+ for (i = 0; type->event_descs[i].attr.attr.name; i++)
+ ;
+
+ attr_group = kzalloc(sizeof(struct attribute *) * (i + 1) +
+ sizeof(*attr_group), GFP_KERNEL);
+ if (!attr_group)
+ return NULL;
+
+ attrs = (struct attribute **)(attr_group + 1);
+ attr_group->name = "events";
+ attr_group->attrs = attrs;
+
+ for (j = 0; j < i; j++) {
+ if (single_fixed && !strncmp(type->event_descs[j].config, fixed_event_name, 10))
+ continue;
+
+ attrs[index++] = &type->event_descs[j].attr.attr;
+ }
+
+ return attr_group;
+}
+
+static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
+{
+ struct intel_uncore_pmu *pmus;
size_t size;
- int i, j;
+ int i;

pmus = kzalloc(sizeof(*pmus) * type->num_boxes, GFP_KERNEL);
if (!pmus)
@@ -848,24 +885,25 @@ static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
0, type->num_counters, 0, 0);

if (type->event_descs) {
- for (i = 0; type->event_descs[i].attr.attr.name; i++);
-
- attr_group = kzalloc(sizeof(struct attribute *) * (i + 1) +
- sizeof(*attr_group), GFP_KERNEL);
- if (!attr_group)
+ type->events_group = uncore_alloc_event_group(type, false);
+ if (!type->events_group)
return -ENOMEM;
+ if (type->single_fixed) {
+ type->no_fixed_attr_groups[2] = uncore_alloc_event_group(type, true);
+ if (!type->no_fixed_attr_groups[2]) {
+ kfree(type->events_group);
+ return -ENOMEM;
+ }
+ }
+ }

- attrs = (struct attribute **)(attr_group + 1);
- attr_group->name = "events";
- attr_group->attrs = attrs;
-
- for (j = 0; j < i; j++)
- attrs[j] = &type->event_descs[j].attr.attr;
+ type->pmu_group = &uncore_pmu_attr_group;

- type->events_group = attr_group;
+ if (type->single_fixed) {
+ type->no_fixed_attr_groups[0] = type->pmu_group;
+ type->no_fixed_attr_groups[1] = type->format_group;
}

- type->pmu_group = &uncore_pmu_attr_group;
return 0;
}

diff --git a/arch/x86/events/intel/uncore.h b/arch/x86/events/intel/uncore.h
index ad986c1..23bf3ff 100644
--- a/arch/x86/events/intel/uncore.h
+++ b/arch/x86/events/intel/uncore.h
@@ -59,6 +59,7 @@ struct intel_uncore_type {
struct intel_uncore_ops *ops;
struct uncore_event_desc *event_descs;
const struct attribute_group *attr_groups[4];
+ const struct attribute_group *no_fixed_attr_groups[4];
struct pmu *pmu; /* for custom pmu ops */
};

--
2.5.5

2016-12-01 16:38:06

by Peter Zijlstra

[permalink] [raw]
Subject: Re: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question

On Thu, Dec 01, 2016 at 04:34:43PM +0000, Liang, Kan wrote:
>
> >
> > "Liang, Kan" <[email protected]> writes:
> > >
> > > Maybe we should create a clocktick box for client uncore to fix it.
> >
> > That would break all the user space code that uses the existing format.
> >
> OK.
> If so, I think we should remove the non-exist events for non-first boxes.
>
> What do you think about the patch as below?
>
> Andi? Jirka?

Urgh, more lines for ugly :-(

I really would prefer to move the thing to its own PMU.

2016-12-12 14:49:08

by Liang, Kan

[permalink] [raw]
Subject: RE: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question



> I really would prefer to move the thing to its own PMU.

The patch as below creates a new PMU to fix the issue.

Jirka, could you please try the patch on your machine?


Thanks,
Kan
-------
>From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17 00:00:00 2001
From: Kan Liang <[email protected]>
Date: Mon, 12 Dec 2016 09:03:35 -0500
Subject: [PATCH] perf/x86/intel/uncore: fix nonexistent clockticks event for
client uncore

The clockticks event can only be used by the first Cbox pmu. Other
Cboxes don't allow to open clockticks event, eventhough it's announced
via /sys/../events/..

For client uncore, there is only one clocktick fixed counter. Current
kernel code forces that only the first box can access the fixed counter
in uncore_pmu_event_init. But it doesn't take care of the the
attr_groups. All the pmus of same type share the same attr_groups. If
the clockticks event is set for the first box, user can also observe the
event in other boxes.

The clocktick fixed counter is a standalone counter. It should be
removed from the Cbox PMUs. A new type of PMU is added which only
supports fixed counter events.

User observable changes with the patch.
clockticks event is removed from Cbox. It will return unsupported, if
uncore_cbox_0/clockticks/ is accessed. User may need to change their
script to use uncore_clock/clockticks/ to instead.

Signed-off-by: Kan Liang <[email protected]>
---
arch/x86/events/intel/uncore.c | 13 +++++++------
arch/x86/events/intel/uncore_snb.c | 38 +++++++++++++++++++++++++++-----------
2 files changed, 34 insertions(+), 17 deletions(-)

diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index dbaaf7dc..03afeca 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -652,6 +652,10 @@ static int uncore_pmu_event_init(struct perf_event *event)
if (hwc->sample_period)
return -EINVAL;

+ /* Single fixed PMU only has fixed event */
+ if (pmu->type->single_fixed && (event->attr.config != UNCORE_FIXED_EVENT))
+ return -EINVAL;
+
/*
* Place all uncore events for a particular physical package
* onto a single cpu
@@ -675,12 +679,9 @@ static int uncore_pmu_event_init(struct perf_event *event)
/* no fixed counter */
if (!pmu->type->fixed_ctl)
return -EINVAL;
- /*
- * if there is only one fixed counter, only the first pmu
- * can access the fixed counter
- */
- if (pmu->type->single_fixed && pmu->pmu_idx > 0)
- return -EINVAL;
+
+ if (pmu->type->single_fixed)
+ event->hw.idx = UNCORE_PMC_IDX_FIXED;

/* fixed counters have event field hardcoded to zero */
hwc->config = 0ULL;
diff --git a/arch/x86/events/intel/uncore_snb.c b/arch/x86/events/intel/uncore_snb.c
index a3dcc12..1b037ea 100644
--- a/arch/x86/events/intel/uncore_snb.c
+++ b/arch/x86/events/intel/uncore_snb.c
@@ -117,7 +117,7 @@ static void snb_uncore_msr_exit_box(struct intel_uncore_box *box)
}

static struct uncore_event_desc snb_uncore_events[] = {
- INTEL_UNCORE_EVENT_DESC(clockticks, "event=0xff,umask=0x00"),
+ INTEL_UNCORE_EVENT_DESC(clockticks, "event=0xff"),
{ /* end: all zeroes */ },
};

@@ -155,17 +155,12 @@ static struct intel_uncore_type snb_uncore_cbox = {
.num_counters = 2,
.num_boxes = 4,
.perf_ctr_bits = 44,
- .fixed_ctr_bits = 48,
.perf_ctr = SNB_UNC_CBO_0_PER_CTR0,
.event_ctl = SNB_UNC_CBO_0_PERFEVTSEL0,
- .fixed_ctr = SNB_UNC_FIXED_CTR,
- .fixed_ctl = SNB_UNC_FIXED_CTR_CTRL,
- .single_fixed = 1,
.event_mask = SNB_UNC_RAW_EVENT_MASK,
.msr_offset = SNB_UNC_CBO_MSR_OFFSET,
.ops = &snb_uncore_msr_ops,
.format_group = &snb_uncore_format_group,
- .event_descs = snb_uncore_events,
};

static struct intel_uncore_type snb_uncore_arb = {
@@ -182,9 +177,34 @@ static struct intel_uncore_type snb_uncore_arb = {
.format_group = &snb_uncore_format_group,
};

+static struct attribute *snb_uncore_clock_formats_attr[] = {
+ &format_attr_event.attr,
+ NULL,
+};
+
+static struct attribute_group snb_uncore_clock_format_group = {
+ .name = "format",
+ .attrs = snb_uncore_clock_formats_attr,
+};
+
+static struct intel_uncore_type snb_uncore_clockbox = {
+ .name = "clock",
+ .num_counters = 1,
+ .num_boxes = 1,
+ .fixed_ctr_bits = 48,
+ .fixed_ctr = SNB_UNC_FIXED_CTR,
+ .fixed_ctl = SNB_UNC_FIXED_CTR_CTRL,
+ .single_fixed = 1,
+ .event_mask = SNB_UNC_CTL_EV_SEL_MASK,
+ .format_group = &snb_uncore_clock_format_group,
+ .ops = &snb_uncore_msr_ops,
+ .event_descs = snb_uncore_events,
+};
+
static struct intel_uncore_type *snb_msr_uncores[] = {
&snb_uncore_cbox,
&snb_uncore_arb,
+ &snb_uncore_clockbox,
NULL,
};

@@ -229,22 +249,18 @@ static struct intel_uncore_type skl_uncore_cbox = {
.num_counters = 4,
.num_boxes = 5,
.perf_ctr_bits = 44,
- .fixed_ctr_bits = 48,
.perf_ctr = SNB_UNC_CBO_0_PER_CTR0,
.event_ctl = SNB_UNC_CBO_0_PERFEVTSEL0,
- .fixed_ctr = SNB_UNC_FIXED_CTR,
- .fixed_ctl = SNB_UNC_FIXED_CTR_CTRL,
- .single_fixed = 1,
.event_mask = SNB_UNC_RAW_EVENT_MASK,
.msr_offset = SNB_UNC_CBO_MSR_OFFSET,
.ops = &skl_uncore_msr_ops,
.format_group = &snb_uncore_format_group,
- .event_descs = snb_uncore_events,
};

static struct intel_uncore_type *skl_msr_uncores[] = {
&skl_uncore_cbox,
&snb_uncore_arb,
+ &snb_uncore_clockbox,
NULL,
};

--
2.4.3

2016-12-14 11:57:25

by Jiri Olsa

[permalink] [raw]
Subject: Re: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question

On Mon, Dec 12, 2016 at 02:49:03PM +0000, Liang, Kan wrote:
>
>
> > I really would prefer to move the thing to its own PMU.
>
> The patch as below creates a new PMU to fix the issue.
>
> Jirka, could you please try the patch on your machine?
>
>
> Thanks,
> Kan
> -------
> From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17 00:00:00 2001
> From: Kan Liang <[email protected]>
> Date: Mon, 12 Dec 2016 09:03:35 -0500
> Subject: [PATCH] perf/x86/intel/uncore: fix nonexistent clockticks event for
> client uncore
>
> The clockticks event can only be used by the first Cbox pmu. Other
> Cboxes don't allow to open clockticks event, eventhough it's announced
> via /sys/../events/..
>
> For client uncore, there is only one clocktick fixed counter. Current
> kernel code forces that only the first box can access the fixed counter
> in uncore_pmu_event_init. But it doesn't take care of the the
> attr_groups. All the pmus of same type share the same attr_groups. If
> the clockticks event is set for the first box, user can also observe the
> event in other boxes.
>
> The clocktick fixed counter is a standalone counter. It should be
> removed from the Cbox PMUs. A new type of PMU is added which only
> supports fixed counter events.
>
> User observable changes with the patch.
> clockticks event is removed from Cbox. It will return unsupported, if
> uncore_cbox_0/clockticks/ is accessed. User may need to change their
> script to use uncore_clock/clockticks/ to instead.
>
> Signed-off-by: Kan Liang <[email protected]>

seems ok

Tested-by: Jiri Olsa <[email protected]>

thanks,
jirka

2016-12-20 18:05:23

by Liang, Kan

[permalink] [raw]
Subject: RE: [RFC] perf/x86/intel/uncore: pmu->type->single_fixed question


> On Mon, Dec 12, 2016 at 02:49:03PM +0000, Liang, Kan wrote:
> >
> >
> > > I really would prefer to move the thing to its own PMU.
> >
> > The patch as below creates a new PMU to fix the issue.
> >
> > Jirka, could you please try the patch on your machine?
> >
> >
> > Thanks,
> > Kan
> > -------
> > From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17
> 00:00:00 2001
> > From: Kan Liang <[email protected]>
> > Date: Mon, 12 Dec 2016 09:03:35 -0500
> > Subject: [PATCH] perf/x86/intel/uncore: fix nonexistent clockticks
> > event for client uncore
> >
> > The clockticks event can only be used by the first Cbox pmu. Other
> > Cboxes don't allow to open clockticks event, eventhough it's announced
> > via /sys/../events/..
> >
> > For client uncore, there is only one clocktick fixed counter. Current
> > kernel code forces that only the first box can access the fixed
> > counter in uncore_pmu_event_init. But it doesn't take care of the the
> > attr_groups. All the pmus of same type share the same attr_groups. If
> > the clockticks event is set for the first box, user can also observe
> > the event in other boxes.
> >
> > The clocktick fixed counter is a standalone counter. It should be
> > removed from the Cbox PMUs. A new type of PMU is added which only
> > supports fixed counter events.
> >
> > User observable changes with the patch.
> > clockticks event is removed from Cbox. It will return unsupported, if
> > uncore_cbox_0/clockticks/ is accessed. User may need to change their
> > script to use uncore_clock/clockticks/ to instead.
> >
> > Signed-off-by: Kan Liang <[email protected]>
>
> seems ok
>
> Tested-by: Jiri Olsa <[email protected]>
>

Hi Peter and Ingo,

Any comments for the patch?
Do I need to re-send it?

Thanks,
Kan