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.

Quoted 22 lines

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;
        ...

---

Quoted 8 lines

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

Quoted 39 lines

Quoted 22 lines

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;
        ...

---

Quoted 8 lines

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.

Quoted 5 lines

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

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

Quoted 2 lines · Kan

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.

From 844c12b5e6fc1e2c27b2f094c89098ea9405ea41 Mon Sep 17 00:00:00 2001

Quoted 6 lines

Quoted 2 lines · Kan

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

Quoted 2 lines

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?

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

From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17 00:00:00 2001

Quoted 3 lines

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]>

Quoted 6 lines

Quoted 3 lines

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

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

Quoted 7 lines

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

Thanks,
Kan
-------
From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17

00:00:00 2001

Quoted 25 lines

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.

Quoted 1 line

Signed-off-by Kan Liang <[email protected]>

seems ok

Tested-by Jiri Olsa <[email protected]>

Quoted 41 lines · Kan

Quoted 6 lines

Quoted 3 lines

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

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

Quoted 7 lines

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

Thanks,
Kan
-------
From 2de8b2eda6b54734e08a608b5fc8c367b94369d3 Mon Sep 17

00:00:00 2001

Quoted 25 lines

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.

Quoted 6 lines

Quoted 1 line

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