2023-09-01 07:00:24

by Jijie Shao

[permalink] [raw]
Subject: [PATCH V2 drivers/perf: hisi:] drivers/perf: hisi: fix set wrong filter mode for running events issue

From: Hao Chen <[email protected]>

hns3_pmu_select_filter_mode() includes A series of mode judgments such
as global mode ,function mode, function-queue mode, port mode, port-tc
mode.

For a special scenario:
command use parameter
perf stat -a -e hns3_pmu_sicl_0/bdf=0x3700,config=0x3,queue=0x0,
and hns3_pmu_is_enabled_func_mode() has a judgement as below:
if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_FUNC))

filter_support of event 0x3 hasn't set bit for func mode, so it can't
enter func-mode branch, and continue to func-queue mode judgement, port
judgement, port-tc mode, then enter port-tc mode.

It's not up to expectations, it shouldn't enter any modes but
return -ENOENT.

port-tc mode parameter show as below:
perf stat -a -e hns3_pmu_sicl_0/config=0x00001,port=0x0,tc=0x1

port-tc mode should use bdf parameter as 0, so, add judgement of
bdf parameter to fix it.

Signed-off-by: Hao Chen <[email protected]>
Signed-off-by: Jijie Shao <[email protected]>
---
changeLog:
v2: add more details in log message suggested by Will
v1 link: https://lore.kernel.org/all/[email protected]/
---
drivers/perf/hisilicon/hns3_pmu.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/perf/hisilicon/hns3_pmu.c b/drivers/perf/hisilicon/hns3_pmu.c
index e0457d84af6b..2aa9cb045705 100644
--- a/drivers/perf/hisilicon/hns3_pmu.c
+++ b/drivers/perf/hisilicon/hns3_pmu.c
@@ -998,12 +998,13 @@ static bool
hns3_pmu_is_enabled_port_tc_mode(struct perf_event *event,
struct hns3_pmu_event_attr *pmu_event)
{
+ u16 bdf = hns3_pmu_get_bdf(event);
u8 tc_id = hns3_pmu_get_tc(event);

if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_PORT_TC))
return false;

- return tc_id != HNS3_PMU_FILTER_ALL_TC;
+ return (tc_id != HNS3_PMU_FILTER_ALL_TC) && (!bdf);
}

static bool
--
2.30.0



2023-09-03 13:15:16

by Jonathan Cameron

[permalink] [raw]
Subject: Re: [PATCH V2 drivers/perf: hisi:] drivers/perf: hisi: fix set wrong filter mode for running events issue

On Fri, 1 Sep 2023 11:50:27 +0800
Jijie Shao <[email protected]> wrote:

> From: Hao Chen <[email protected]>

Mention which hisi pmu this is in the patch description (hns)

>
> hns3_pmu_select_filter_mode() includes A series of mode judgments such

includes a series

> as global mode ,function mode, function-queue mode, port mode, port-tc
> mode.
>
> For a special scenario:
> command use parameter
> perf stat -a -e hns3_pmu_sicl_0/bdf=0x3700,config=0x3,queue=0x0,
> and hns3_pmu_is_enabled_func_mode() has a judgement as below:
> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_FUNC))
>
> filter_support of event 0x3 hasn't set bit for func mode, so it can't
> enter func-mode branch, and continue to func-queue mode judgement, port
> judgement, port-tc mode, then enter port-tc mode.
>
> It's not up to expectations, it shouldn't enter any modes but
> return -ENOENT.
>
> port-tc mode parameter show as below:
> perf stat -a -e hns3_pmu_sicl_0/config=0x00001,port=0x0,tc=0x1
>
> port-tc mode should use bdf parameter as 0, so, add judgement of
> bdf parameter to fix it.

I don't follow the description here. As far as I can see from the code
the change just checks that BDF is not set before allowing a port based
filter.

>
> Signed-off-by: Hao Chen <[email protected]>
> Signed-off-by: Jijie Shao <[email protected]>
> ---
> changeLog:
> v2: add more details in log message suggested by Will
> v1 link: https://lore.kernel.org/all/[email protected]/
> ---
> drivers/perf/hisilicon/hns3_pmu.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/perf/hisilicon/hns3_pmu.c b/drivers/perf/hisilicon/hns3_pmu.c
> index e0457d84af6b..2aa9cb045705 100644
> --- a/drivers/perf/hisilicon/hns3_pmu.c
> +++ b/drivers/perf/hisilicon/hns3_pmu.c
> @@ -998,12 +998,13 @@ static bool
> hns3_pmu_is_enabled_port_tc_mode(struct perf_event *event,
> struct hns3_pmu_event_attr *pmu_event)
> {
> + u16 bdf = hns3_pmu_get_bdf(event);
> u8 tc_id = hns3_pmu_get_tc(event);
>
> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_PORT_TC))
> return false;
>
> - return tc_id != HNS3_PMU_FILTER_ALL_TC;
> + return (tc_id != HNS3_PMU_FILTER_ALL_TC) && (!bdf);

No need for brackets on !bdf

> }
>
> static bool

2023-09-04 12:36:31

by chenhao (EZ)

[permalink] [raw]
Subject: Re: [PATCH V2 drivers/perf: hisi:] drivers/perf: hisi: fix set wrong filter mode for running events issue



On 2023/9/1 18:11, Jonathan Cameron wrote:
> On Fri, 1 Sep 2023 11:50:27 +0800
> Jijie Shao <[email protected]> wrote:
>
>> From: Hao Chen <[email protected]>
>
> Mention which hisi pmu this is in the patch description (hns)
>
>>
>> hns3_pmu_select_filter_mode() includes A series of mode judgments such
>
> includes a series
>
>> as global mode ,function mode, function-queue mode, port mode, port-tc
>> mode.
>>
>> For a special scenario:
>> command use parameter
>> perf stat -a -e hns3_pmu_sicl_0/bdf=0x3700,config=0x3,queue=0x0,
>> and hns3_pmu_is_enabled_func_mode() has a judgement as below:
>> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_FUNC))
>>
>> filter_support of event 0x3 hasn't set bit for func mode, so it can't
>> enter func-mode branch, and continue to func-queue mode judgement, port
>> judgement, port-tc mode, then enter port-tc mode.
>>
>> It's not up to expectations, it shouldn't enter any modes but
>> return -ENOENT.
>>
>> port-tc mode parameter show as below:
>> perf stat -a -e hns3_pmu_sicl_0/config=0x00001,port=0x0,tc=0x1
>>
>> port-tc mode should use bdf parameter as 0, so, add judgement of
>> bdf parameter to fix it.
>
> I don't follow the description here. As far as I can see from the code
> the change just checks that BDF is not set before allowing a port based
> filter.

Thanks for review.

Maybe I don't think about it completely.

For this patch, I only consider this scenario mentioned in patch and exactly filter out
port based mode, then enter port-tc based mode.

Actually port based mode also need this change for other scenarios, do you agree with it?

>
>>
>> Signed-off-by: Hao Chen <[email protected]>
>> Signed-off-by: Jijie Shao <[email protected]>
>> ---
>> changeLog:
>> v2: add more details in log message suggested by Will
>> v1 link: https://lore.kernel.org/all/[email protected]/
>> ---
>> drivers/perf/hisilicon/hns3_pmu.c | 3 ++-
>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/perf/hisilicon/hns3_pmu.c b/drivers/perf/hisilicon/hns3_pmu.c
>> index e0457d84af6b..2aa9cb045705 100644
>> --- a/drivers/perf/hisilicon/hns3_pmu.c
>> +++ b/drivers/perf/hisilicon/hns3_pmu.c
>> @@ -998,12 +998,13 @@ static bool
>> hns3_pmu_is_enabled_port_tc_mode(struct perf_event *event,
>> struct hns3_pmu_event_attr *pmu_event)
>> {
>> + u16 bdf = hns3_pmu_get_bdf(event);
>> u8 tc_id = hns3_pmu_get_tc(event);
>>
>> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_PORT_TC))
>> return false;
>>
>> - return tc_id != HNS3_PMU_FILTER_ALL_TC;
>> + return (tc_id != HNS3_PMU_FILTER_ALL_TC) && (!bdf);
>
> No need for brackets on !bdf
>
>> }
>>
>> static bool
>
> .
>

2023-09-06 10:06:42

by Yicong Yang

[permalink] [raw]
Subject: Re: [PATCH V2 drivers/perf: hisi:] drivers/perf: hisi: fix set wrong filter mode for running events issue

On 2023/9/1 11:50, Jijie Shao wrote:
> From: Hao Chen <[email protected]>
>
> hns3_pmu_select_filter_mode() includes A series of mode judgments such
> as global mode ,function mode, function-queue mode, port mode, port-tc
> mode.
>
> For a special scenario:
> command use parameter
> perf stat -a -e hns3_pmu_sicl_0/bdf=0x3700,config=0x3,queue=0x0,
> and hns3_pmu_is_enabled_func_mode() has a judgement as below:
> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_FUNC))
>
> filter_support of event 0x3 hasn't set bit for func mode, so it can't
> enter func-mode branch, and continue to func-queue mode judgement, port
> judgement, port-tc mode, then enter port-tc mode.
>
> It's not up to expectations, it shouldn't enter any modes but
> return -ENOENT.
>
> port-tc mode parameter show as below:
> perf stat -a -e hns3_pmu_sicl_0/config=0x00001,port=0x0,tc=0x1
>
> port-tc mode should use bdf parameter as 0, so, add judgement of
> bdf parameter to fix it.
>

Will the change here block the below unsupported case for event 0x3?

$ perf stat -a -e hns3_pmu_sicl_0/bdf=0x0,config=0x3,queue=0x0/

Otherwise you should handle this case.

> Signed-off-by: Hao Chen <[email protected]>
> Signed-off-by: Jijie Shao <[email protected]>
> ---
> changeLog:
> v2: add more details in log message suggested by Will
> v1 link: https://lore.kernel.org/all/[email protected]/
> ---
> drivers/perf/hisilicon/hns3_pmu.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/perf/hisilicon/hns3_pmu.c b/drivers/perf/hisilicon/hns3_pmu.c
> index e0457d84af6b..2aa9cb045705 100644
> --- a/drivers/perf/hisilicon/hns3_pmu.c
> +++ b/drivers/perf/hisilicon/hns3_pmu.c
> @@ -998,12 +998,13 @@ static bool
> hns3_pmu_is_enabled_port_tc_mode(struct perf_event *event,
> struct hns3_pmu_event_attr *pmu_event)
> {
> + u16 bdf = hns3_pmu_get_bdf(event);
> u8 tc_id = hns3_pmu_get_tc(event);
>
> if (!(pmu_event->filter_support & HNS3_PMU_FILTER_SUPPORT_PORT_TC))
> return false;
>
> - return tc_id != HNS3_PMU_FILTER_ALL_TC;
> + return (tc_id != HNS3_PMU_FILTER_ALL_TC) && (!bdf);
> }
>
> static bool
>