So I recently moved from a dual-xeon box to a zen 2 based threadripper
workstation.
My usual incantation for measuring profiles for compile time isn't working:
$ perf record -e cycles:pp --freq=128 --call-graph lbr -- make LLVM=1 -j$(nproc)
Error:
Invalid event (cycles:pp) in per-thread mode, enable system wide with '-a'.
I've already set /proc/sys/kernel/perf_event_paranoid and
/proc/sys/kernel/kptr_restrict to 0.
I remember hearing rumblings about issues with zen 2, LBR, vs zen 3.
Is this a known issue, or am I holding it wrong?
--
Thanks,
~Nick Desaulniers
On 30-Sep-22 9:56 AM, Namhyung Kim wrote:
> Hello,
>
> On Thu, Sep 29, 2022 at 8:49 PM Ian Rogers <[email protected]> wrote:
>>
>> On Thu, Sep 29, 2022 at 3:10 PM Slade Watkins <[email protected]> wrote:
>>>
>>> Hey Nick,
>>>
>>>> On Sep 29, 2022, at 5:54 PM, Nick Desaulniers <[email protected]> wrote:
>>>>
>>>> I remember hearing rumblings about issues with zen 2, LBR, vs zen 3.
>>>> Is this a known issue, or am I holding it wrong?
>>>
>>> Hm… I also remember this. I have a Zen 2 based system that I can do testing on, so I will do so when I’m able.
>>>
>>> If I discover something of note, I’ll get back to you.
>>>
>>> Cheers,
>>> -srw
>>>
>>
>> LBR isn't yet supported for Zen but is coming:
>> https://lore.kernel.org/lkml/166155216401.401.5809694678609694438.tip-bot2@tip-bot2/
>> I'd recommend frame-pointers.
>>
>> +Ravi who may be able to say if there are any issues with the precise
>> sampling on AMD.
>
> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> since it has no task context. Ravi is working on it..
Right.
https://lore.kernel.org/lkml/[email protected]
Thanks,
Ravi
+ Stephane, Kees, Sandipan, Bill, ClangBuiltLinux mailing list, Yonghong
https://www.spinics.net/lists/linux-perf-users/msg23103.html
starts the thread, for context.
On Thu, Sep 29, 2022 at 9:32 PM Ravi Bangoria <[email protected]> wrote:
>
> On 30-Sep-22 9:56 AM, Namhyung Kim wrote:
> > Hello,
> >
> > On Thu, Sep 29, 2022 at 8:49 PM Ian Rogers <[email protected]> wrote:
> >>
> >> On Thu, Sep 29, 2022 at 3:10 PM Slade Watkins <[email protected]> wrote:
> >>>
> >>> Hey Nick,
> >>>
> >>>> On Sep 29, 2022, at 5:54 PM, Nick Desaulniers <[email protected]> wrote:
> >>>>
> >>>> I remember hearing rumblings about issues with zen 2, LBR, vs zen 3.
> >>>> Is this a known issue, or am I holding it wrong?
> >>>
> >>> Hm… I also remember this. I have a Zen 2 based system that I can do testing on, so I will do so when I’m able.
> >>>
> >>> If I discover something of note, I’ll get back to you.
> >>>
> >>> Cheers,
> >>> -srw
> >>>
> >>
> >> LBR isn't yet supported for Zen but is coming:
> >> https://lore.kernel.org/lkml/166155216401.401.5809694678609694438.tip-bot2@tip-bot2/
> >> I'd recommend frame-pointers.
Having to recompile is less than ideal for my workflow. I have added a note to
https://github.com/ClangBuiltLinux/profiling/tree/main/perf#errors
Please let me know how I might improve the documentation.
> >>
> >> +Ravi who may be able to say if there are any issues with the precise
> >> sampling on AMD.
> >
> > Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> > since it has no task context. Ravi is working on it..
>
> Right.
> https://lore.kernel.org/lkml/[email protected]
Cool, thanks for working on this Ravi.
I'm not sure yet whether I may replace the kernel on my corporate
provided workstation, so I'm not sure yet I can help test that patch.
Can you confirm that
$ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
works with just that patch applied? Or is there more work required?
What is the status of that patch?
For context, we had difficulty upstreaming support for instrumentation
based profile guided optimizations in the Linux kernel.
https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
We'd like to be able to use either instrumentation or sampling to
optimize our builds. The major barrier to sample based approaches are
architecture / micro architecture issues with sample based profile
data collection, and bitrot of data processing utilities.
https://github.com/google/autofdo/issues/144
--
Thanks,
~Nick Desaulniers
On Wed, Oct 5, 2022 at 2:56 PM Nick Desaulniers <[email protected]> wrote:
>
> + Stephane, Kees, Sandipan, Bill, ClangBuiltLinux mailing list, Yonghong
> https://www.spinics.net/lists/linux-perf-users/msg23103.html
> starts the thread, for context.
>
> On Thu, Sep 29, 2022 at 9:32 PM Ravi Bangoria <[email protected]> wrote:
> >
> > On 30-Sep-22 9:56 AM, Namhyung Kim wrote:
> > > Hello,
> > >
> > > On Thu, Sep 29, 2022 at 8:49 PM Ian Rogers <[email protected]> wrote:
> > >>
> > >> On Thu, Sep 29, 2022 at 3:10 PM Slade Watkins <[email protected]> wrote:
> > >>>
> > >>> Hey Nick,
> > >>>
> > >>>> On Sep 29, 2022, at 5:54 PM, Nick Desaulniers <[email protected]> wrote:
> > >>>>
> > >>>> I remember hearing rumblings about issues with zen 2, LBR, vs zen 3.
> > >>>> Is this a known issue, or am I holding it wrong?
> > >>>
> > >>> Hm… I also remember this. I have a Zen 2 based system that I can do testing on, so I will do so when I’m able.
> > >>>
> > >>> If I discover something of note, I’ll get back to you.
> > >>>
> > >>> Cheers,
> > >>> -srw
> > >>>
> > >>
> > >> LBR isn't yet supported for Zen but is coming:
> > >> https://lore.kernel.org/lkml/166155216401.401.5809694678609694438.tip-bot2@tip-bot2/
> > >> I'd recommend frame-pointers.
>
> Having to recompile is less than ideal for my workflow. I have added a note to
> https://github.com/ClangBuiltLinux/profiling/tree/main/perf#errors
> Please let me know how I might improve the documentation.
>
> > >>
> > >> +Ravi who may be able to say if there are any issues with the precise
> > >> sampling on AMD.
> > >
> > > Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> > > since it has no task context. Ravi is working on it..
> >
> > Right.
> > https://lore.kernel.org/lkml/[email protected]
>
> Cool, thanks for working on this Ravi.
>
> I'm not sure yet whether I may replace the kernel on my corporate
> provided workstation, so I'm not sure yet I can help test that patch.
>
> Can you confirm that
> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>
> works with just that patch applied? Or is there more work required?
> What is the status of that patch?
>
> For context, we had difficulty upstreaming support for instrumentation
> based profile guided optimizations in the Linux kernel.
> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
> We'd like to be able to use either instrumentation or sampling to
> optimize our builds. The major barrier to sample based approaches are
> architecture / micro architecture issues with sample based profile
> data collection, and bitrot of data processing utilities.
> https://github.com/google/autofdo/issues/144
On existing AMD Zen2, Zen3 the following cmdline:
$ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
does not work. I see two reasons:
1. cycles:pp is likely converted into IBS op in cycle mode.
Current kernels do not support IBS in per-thread mode.
This is purely a kernel limitation
2. call-graph lbr is not supported on AMD because they do
not have LBR and therefore no LBR callstack mode
The best way to get what you want here today on AMD Zen2 and Zen3:
$ perf record -e cycles --freq=128 -g -- <command to profile>
On AMD Zen3, there is a precursor to LBR with Branch Sampling (BRS),
and you can use it to sample taken branches but not for callstacks. I
mention the cmdline here for reference:
$ perf record -e cpu/branch-brs/ -c 1000037 -b -- <command to profile>
Note that AMD Zen3 BRS is enough to get the autoFDO usage of an
LBR working as per the cmdline above.
Hope this helps.
+cc: PeterZ
>>>>> +Ravi who may be able to say if there are any issues with the precise
>>>>> sampling on AMD.
>>>>
>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
>>>> since it has no task context. Ravi is working on it..
>>>
>>> Right.
>>> https://lore.kernel.org/lkml/[email protected]
>>
>> Cool, thanks for working on this Ravi.
>>
>> I'm not sure yet whether I may replace the kernel on my corporate
>> provided workstation, so I'm not sure yet I can help test that patch.
>>
>> Can you confirm that
>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>
>> works with just that patch applied? Or is there more work required?
>> What is the status of that patch?
>>
>> For context, we had difficulty upstreaming support for instrumentation
>> based profile guided optimizations in the Linux kernel.
>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
>> We'd like to be able to use either instrumentation or sampling to
>> optimize our builds. The major barrier to sample based approaches are
>> architecture / micro architecture issues with sample based profile
>> data collection, and bitrot of data processing utilities.
>> https://github.com/google/autofdo/issues/144
>
> On existing AMD Zen2, Zen3 the following cmdline:
> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>
> does not work. I see two reasons:
>
> 1. cycles:pp is likely converted into IBS op in cycle mode.
> Current kernels do not support IBS in per-thread mode.
> This is purely a kernel limitation
Right, it's purely a kernel limitation. And below simple patch on top
of event-context rewrite patch[1] should be sufficient to make cycles:pp
working in per-process mode on AMD Zen.
---
diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
index c251bc44c088..de01b5d27e40 100644
--- a/arch/x86/events/amd/ibs.c
+++ b/arch/x86/events/amd/ibs.c
@@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
static struct perf_ibs perf_ibs_op = {
.pmu = {
- .task_ctx_nr = perf_invalid_context,
+ .task_ctx_nr = perf_hw_context,
.event_init = perf_ibs_init,
.add = perf_ibs_add,
---
[1]: https://lore.kernel.org/lkml/[email protected]
On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
>
> +cc: PeterZ
>
> >>>>> +Ravi who may be able to say if there are any issues with the precise
> >>>>> sampling on AMD.
> >>>>
> >>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> >>>> since it has no task context. Ravi is working on it..
> >>>
> >>> Right.
> >>> https://lore.kernel.org/lkml/[email protected]
> >>
> >> Cool, thanks for working on this Ravi.
> >>
> >> I'm not sure yet whether I may replace the kernel on my corporate
> >> provided workstation, so I'm not sure yet I can help test that patch.
> >>
> >> Can you confirm that
> >> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> >>
> >> works with just that patch applied? Or is there more work required?
> >> What is the status of that patch?
> >>
> >> For context, we had difficulty upstreaming support for instrumentation
> >> based profile guided optimizations in the Linux kernel.
> >> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
> >> We'd like to be able to use either instrumentation or sampling to
> >> optimize our builds. The major barrier to sample based approaches are
> >> architecture / micro architecture issues with sample based profile
> >> data collection, and bitrot of data processing utilities.
> >> https://github.com/google/autofdo/issues/144
> >
> > On existing AMD Zen2, Zen3 the following cmdline:
> > $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> >
> > does not work. I see two reasons:
> >
> > 1. cycles:pp is likely converted into IBS op in cycle mode.
> > Current kernels do not support IBS in per-thread mode.
> > This is purely a kernel limitation
>
> Right, it's purely a kernel limitation. And below simple patch on top
> of event-context rewrite patch[1] should be sufficient to make cycles:pp
> working in per-process mode on AMD Zen.
>
> ---
> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
> index c251bc44c088..de01b5d27e40 100644
> --- a/arch/x86/events/amd/ibs.c
> +++ b/arch/x86/events/amd/ibs.c
> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
>
> static struct perf_ibs perf_ibs_op = {
> .pmu = {
> - .task_ctx_nr = perf_invalid_context,
> + .task_ctx_nr = perf_hw_context,
>
> .event_init = perf_ibs_init,
> .add = perf_ibs_add,
> ---
>
> [1]: https://lore.kernel.org/lkml/[email protected]
Hi Ravi,
I didn't see the above diff in
https://lore.kernel.org/lkml/[email protected]/
Was there another distinct patch you were going to send for the above?
--
Thanks,
~Nick Desaulniers
On Wed, Oct 5, 2022 at 3:50 PM Stephane Eranian <[email protected]> wrote:
>
> On Wed, Oct 5, 2022 at 2:56 PM Nick Desaulniers <[email protected]> wrote:
> >
> > + Stephane, Kees, Sandipan, Bill, ClangBuiltLinux mailing list, Yonghong
> > https://www.spinics.net/lists/linux-perf-users/msg23103.html
> > starts the thread, for context.
> >
> > On Thu, Sep 29, 2022 at 9:32 PM Ravi Bangoria <[email protected]> wrote:
> > >
> > > On 30-Sep-22 9:56 AM, Namhyung Kim wrote:
> > > > Hello,
> > > >
> > > > On Thu, Sep 29, 2022 at 8:49 PM Ian Rogers <[email protected]> wrote:
> > > >>
> > > >> On Thu, Sep 29, 2022 at 3:10 PM Slade Watkins <[email protected]> wrote:
> > > >>>
> > > >>> Hey Nick,
> > > >>>
> > > >>>> On Sep 29, 2022, at 5:54 PM, Nick Desaulniers <[email protected]> wrote:
> > > >>>>
> > > >>>> I remember hearing rumblings about issues with zen 2, LBR, vs zen 3.
> > > >>>> Is this a known issue, or am I holding it wrong?
> > > >>>
> > > >>> Hm… I also remember this. I have a Zen 2 based system that I can do testing on, so I will do so when I’m able.
> > > >>>
> > > >>> If I discover something of note, I’ll get back to you.
> > > >>>
> > > >>> Cheers,
> > > >>> -srw
> > > >>>
> > > >>
> > > >> LBR isn't yet supported for Zen but is coming:
> > > >> https://lore.kernel.org/lkml/166155216401.401.5809694678609694438.tip-bot2@tip-bot2/
> > > >> I'd recommend frame-pointers.
> >
> > Having to recompile is less than ideal for my workflow. I have added a note to
> > https://github.com/ClangBuiltLinux/profiling/tree/main/perf#errors
> > Please let me know how I might improve the documentation.
> >
> > > >>
> > > >> +Ravi who may be able to say if there are any issues with the precise
> > > >> sampling on AMD.
> > > >
> > > > Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> > > > since it has no task context. Ravi is working on it..
> > >
> > > Right.
> > > https://lore.kernel.org/lkml/[email protected]
> >
> > Cool, thanks for working on this Ravi.
> >
> > I'm not sure yet whether I may replace the kernel on my corporate
> > provided workstation, so I'm not sure yet I can help test that patch.
> >
> > Can you confirm that
> > $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> >
> > works with just that patch applied? Or is there more work required?
> > What is the status of that patch?
> >
> > For context, we had difficulty upstreaming support for instrumentation
> > based profile guided optimizations in the Linux kernel.
> > https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
> > We'd like to be able to use either instrumentation or sampling to
> > optimize our builds. The major barrier to sample based approaches are
> > architecture / micro architecture issues with sample based profile
> > data collection, and bitrot of data processing utilities.
> > https://github.com/google/autofdo/issues/144
>
> On existing AMD Zen2, Zen3 the following cmdline:
> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>
> does not work. I see two reasons:
>
> 1. cycles:pp is likely converted into IBS op in cycle mode.
> Current kernels do not support IBS in per-thread mode.
> This is purely a kernel limitation
Sounds like Ravi has a diff that might work?
https://lore.kernel.org/linux-perf-users/[email protected]/
>
> 2. call-graph lbr is not supported on AMD because they do
> not have LBR and therefore no LBR callstack mode
>
> The best way to get what you want here today on AMD Zen2 and Zen3:
>
> $ perf record -e cycles --freq=128 -g -- <command to profile>
So without recompiling to explicitly re-enable frame pointers, code
I'm looking to profile is built with -gmlt, so I get symbols, but I
can't observe callers.
Sounds like I'd need to rebuild with -fno-omit-frame-pointers which is
more painful than my prior LBR based workflow.
>
> On AMD Zen3, there is a precursor to LBR with Branch Sampling (BRS),
> and you can use it to sample taken branches but not for callstacks. I
> mention the cmdline here for reference:
>
> $ perf record -e cpu/branch-brs/ -c 1000037 -b -- <command to profile>
>
> Note that AMD Zen3 BRS is enough to get the autoFDO usage of an
> LBR working as per the cmdline above.
Interesting, but I'm stuck with a Zen2 box for a couple years now
(corporate workstation). This pretty much blows up all of the
profiling work I was doing, and any hope I had of contributing towards
building the kernel with AutoFDO profiles until this works on my box
with the kernel it has.
>
> Hope this helps.
--
Thanks,
~Nick Desaulniers
On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
> On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
>>
>> +cc: PeterZ
>>
>>>>>>> +Ravi who may be able to say if there are any issues with the precise
>>>>>>> sampling on AMD.
>>>>>>
>>>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
>>>>>> since it has no task context. Ravi is working on it..
>>>>>
>>>>> Right.
>>>>> https://lore.kernel.org/lkml/[email protected]
>>>>
>>>> Cool, thanks for working on this Ravi.
>>>>
>>>> I'm not sure yet whether I may replace the kernel on my corporate
>>>> provided workstation, so I'm not sure yet I can help test that patch.
>>>>
>>>> Can you confirm that
>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>>
>>>> works with just that patch applied? Or is there more work required?
>>>> What is the status of that patch?
>>>>
>>>> For context, we had difficulty upstreaming support for instrumentation
>>>> based profile guided optimizations in the Linux kernel.
>>>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
>>>> We'd like to be able to use either instrumentation or sampling to
>>>> optimize our builds. The major barrier to sample based approaches are
>>>> architecture / micro architecture issues with sample based profile
>>>> data collection, and bitrot of data processing utilities.
>>>> https://github.com/google/autofdo/issues/144
>>>
>>> On existing AMD Zen2, Zen3 the following cmdline:
>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>
>>> does not work. I see two reasons:
>>>
>>> 1. cycles:pp is likely converted into IBS op in cycle mode.
>>> Current kernels do not support IBS in per-thread mode.
>>> This is purely a kernel limitation
>>
>> Right, it's purely a kernel limitation. And below simple patch on top
>> of event-context rewrite patch[1] should be sufficient to make cycles:pp
>> working in per-process mode on AMD Zen.
>>
>> ---
>> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
>> index c251bc44c088..de01b5d27e40 100644
>> --- a/arch/x86/events/amd/ibs.c
>> +++ b/arch/x86/events/amd/ibs.c
>> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
>>
>> static struct perf_ibs perf_ibs_op = {
>> .pmu = {
>> - .task_ctx_nr = perf_invalid_context,
>> + .task_ctx_nr = perf_hw_context,
>>
>> .event_init = perf_ibs_init,
>> .add = perf_ibs_add,
>> ---
>>
>> [1]: https://lore.kernel.org/lkml/[email protected]
>
> Hi Ravi,
> I didn't see the above diff in
> https://lore.kernel.org/lkml/[email protected]/
> Was there another distinct patch you were going to send for the above?
Yes Nick. I was planning to send it once the rewrite stuff goes in.
Thanks,
Ravi
On 12-Oct-22 9:36 AM, Ravi Bangoria wrote:
> On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
>> On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
>>>
>>> +cc: PeterZ
>>>
>>>>>>>> +Ravi who may be able to say if there are any issues with the precise
>>>>>>>> sampling on AMD.
>>>>>>>
>>>>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
>>>>>>> since it has no task context. Ravi is working on it..
>>>>>>
>>>>>> Right.
>>>>>> https://lore.kernel.org/lkml/[email protected]
>>>>>
>>>>> Cool, thanks for working on this Ravi.
>>>>>
>>>>> I'm not sure yet whether I may replace the kernel on my corporate
>>>>> provided workstation, so I'm not sure yet I can help test that patch.
>>>>>
>>>>> Can you confirm that
>>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>>>
>>>>> works with just that patch applied? Or is there more work required?
>>>>> What is the status of that patch?
>>>>>
>>>>> For context, we had difficulty upstreaming support for instrumentation
>>>>> based profile guided optimizations in the Linux kernel.
>>>>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
>>>>> We'd like to be able to use either instrumentation or sampling to
>>>>> optimize our builds. The major barrier to sample based approaches are
>>>>> architecture / micro architecture issues with sample based profile
>>>>> data collection, and bitrot of data processing utilities.
>>>>> https://github.com/google/autofdo/issues/144
>>>>
>>>> On existing AMD Zen2, Zen3 the following cmdline:
>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>>
>>>> does not work. I see two reasons:
>>>>
>>>> 1. cycles:pp is likely converted into IBS op in cycle mode.
>>>> Current kernels do not support IBS in per-thread mode.
>>>> This is purely a kernel limitation
>>>
>>> Right, it's purely a kernel limitation. And below simple patch on top
>>> of event-context rewrite patch[1] should be sufficient to make cycles:pp
>>> working in per-process mode on AMD Zen.
>>>
>>> ---
>>> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
>>> index c251bc44c088..de01b5d27e40 100644
>>> --- a/arch/x86/events/amd/ibs.c
>>> +++ b/arch/x86/events/amd/ibs.c
>>> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
>>>
>>> static struct perf_ibs perf_ibs_op = {
>>> .pmu = {
>>> - .task_ctx_nr = perf_invalid_context,
>>> + .task_ctx_nr = perf_hw_context,
>>>
>>> .event_init = perf_ibs_init,
>>> .add = perf_ibs_add,
>>> ---
>>>
>>> [1]: https://lore.kernel.org/lkml/[email protected]
>>
>> Hi Ravi,
>> I didn't see the above diff in
>> https://lore.kernel.org/lkml/[email protected]/
>> Was there another distinct patch you were going to send for the above?
>
> Yes Nick. I was planning to send it once the rewrite stuff goes in.
Hi Nick,
Since you have practical use case, would it be possible to run your workflow
with perf rewrite and IBS patches applied? It will help us in finding/fixing
more bugs and upstreaming these changes.
Thanks,
Ravi
On Tue, Oct 11, 2022 at 10:05 PM Ravi Bangoria <[email protected]> wrote:
>
> On 12-Oct-22 9:36 AM, Ravi Bangoria wrote:
> > On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
> >> On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
> >>>
> >>> +cc: PeterZ
> >>>
> >>>>>>>> +Ravi who may be able to say if there are any issues with the precise
> >>>>>>>> sampling on AMD.
> >>>>>>>
> >>>>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> >>>>>>> since it has no task context. Ravi is working on it..
> >>>>>>
> >>>>>> Right.
> >>>>>> https://lore.kernel.org/lkml/[email protected]
> >>>>>
> >>>>> Cool, thanks for working on this Ravi.
> >>>>>
> >>>>> I'm not sure yet whether I may replace the kernel on my corporate
> >>>>> provided workstation, so I'm not sure yet I can help test that patch.
> >>>>>
> >>>>> Can you confirm that
> >>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> >>>>>
> >>>>> works with just that patch applied? Or is there more work required?
> >>>>> What is the status of that patch?
> >>>>>
> >>>>> For context, we had difficulty upstreaming support for instrumentation
> >>>>> based profile guided optimizations in the Linux kernel.
> >>>>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
> >>>>> We'd like to be able to use either instrumentation or sampling to
> >>>>> optimize our builds. The major barrier to sample based approaches are
> >>>>> architecture / micro architecture issues with sample based profile
> >>>>> data collection, and bitrot of data processing utilities.
> >>>>> https://github.com/google/autofdo/issues/144
> >>>>
> >>>> On existing AMD Zen2, Zen3 the following cmdline:
> >>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> >>>>
> >>>> does not work. I see two reasons:
> >>>>
> >>>> 1. cycles:pp is likely converted into IBS op in cycle mode.
> >>>> Current kernels do not support IBS in per-thread mode.
> >>>> This is purely a kernel limitation
> >>>
> >>> Right, it's purely a kernel limitation. And below simple patch on top
> >>> of event-context rewrite patch[1] should be sufficient to make cycles:pp
> >>> working in per-process mode on AMD Zen.
> >>>
> >>> ---
> >>> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
> >>> index c251bc44c088..de01b5d27e40 100644
> >>> --- a/arch/x86/events/amd/ibs.c
> >>> +++ b/arch/x86/events/amd/ibs.c
> >>> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
> >>>
> >>> static struct perf_ibs perf_ibs_op = {
> >>> .pmu = {
> >>> - .task_ctx_nr = perf_invalid_context,
> >>> + .task_ctx_nr = perf_hw_context,
> >>>
> >>> .event_init = perf_ibs_init,
> >>> .add = perf_ibs_add,
> >>> ---
> >>>
> >>> [1]: https://lore.kernel.org/lkml/[email protected]
> >>
> >> Hi Ravi,
> >> I didn't see the above diff in
> >> https://lore.kernel.org/lkml/[email protected]/
> >> Was there another distinct patch you were going to send for the above?
> >
> > Yes Nick. I was planning to send it once the rewrite stuff goes in.
>
> Hi Nick,
>
> Since you have practical use case, would it be possible to run your workflow
> with perf rewrite and IBS patches applied? It will help us in finding/fixing
> more bugs and upstreaming these changes.
Hi Ravi,
Sorry, I'm not able to load a custom kernel image on my employer
provided workstation, and I never got approval to expense hardware for
testing this otherwise.
Was there ever any update on this? I'm on 6.1.25 now and still cant run
$ perf record -e cycles:pp --call-graph lbr <any command to profile>
$ cat /proc/cpuinfo
...
model name : AMD Ryzen Threadripper PRO 3995WX 64-Cores
...
--
Thanks,
~Nick Desaulniers
Hi Nick,
On Fri, Jun 23, 2023 at 9:23 AM Nick Desaulniers
<[email protected]> wrote:
>
> On Tue, Oct 11, 2022 at 10:05 PM Ravi Bangoria <[email protected]> wrote:
> >
> > On 12-Oct-22 9:36 AM, Ravi Bangoria wrote:
> > > On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
> > >> On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
> > >>>
> > >>> +cc: PeterZ
> > >>>
> > >>>>>>>> +Ravi who may be able to say if there are any issues with the precise
> > >>>>>>>> sampling on AMD.
> > >>>>>>>
> > >>>>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
> > >>>>>>> since it has no task context. Ravi is working on it..
> > >>>>>>
> > >>>>>> Right.
> > >>>>>> https://lore.kernel.org/lkml/[email protected]
> > >>>>>
> > >>>>> Cool, thanks for working on this Ravi.
> > >>>>>
> > >>>>> I'm not sure yet whether I may replace the kernel on my corporate
> > >>>>> provided workstation, so I'm not sure yet I can help test that patch.
> > >>>>>
> > >>>>> Can you confirm that
> > >>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> > >>>>>
> > >>>>> works with just that patch applied? Or is there more work required?
> > >>>>> What is the status of that patch?
> > >>>>>
> > >>>>> For context, we had difficulty upstreaming support for instrumentation
> > >>>>> based profile guided optimizations in the Linux kernel.
> > >>>>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
> > >>>>> We'd like to be able to use either instrumentation or sampling to
> > >>>>> optimize our builds. The major barrier to sample based approaches are
> > >>>>> architecture / micro architecture issues with sample based profile
> > >>>>> data collection, and bitrot of data processing utilities.
> > >>>>> https://github.com/google/autofdo/issues/144
> > >>>>
> > >>>> On existing AMD Zen2, Zen3 the following cmdline:
> > >>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
> > >>>>
> > >>>> does not work. I see two reasons:
> > >>>>
> > >>>> 1. cycles:pp is likely converted into IBS op in cycle mode.
> > >>>> Current kernels do not support IBS in per-thread mode.
> > >>>> This is purely a kernel limitation
> > >>>
> > >>> Right, it's purely a kernel limitation. And below simple patch on top
> > >>> of event-context rewrite patch[1] should be sufficient to make cycles:pp
> > >>> working in per-process mode on AMD Zen.
> > >>>
> > >>> ---
> > >>> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
> > >>> index c251bc44c088..de01b5d27e40 100644
> > >>> --- a/arch/x86/events/amd/ibs.c
> > >>> +++ b/arch/x86/events/amd/ibs.c
> > >>> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
> > >>>
> > >>> static struct perf_ibs perf_ibs_op = {
> > >>> .pmu = {
> > >>> - .task_ctx_nr = perf_invalid_context,
> > >>> + .task_ctx_nr = perf_hw_context,
> > >>>
> > >>> .event_init = perf_ibs_init,
> > >>> .add = perf_ibs_add,
> > >>> ---
> > >>>
> > >>> [1]: https://lore.kernel.org/lkml/[email protected]
> > >>
> > >> Hi Ravi,
> > >> I didn't see the above diff in
> > >> https://lore.kernel.org/lkml/[email protected]/
> > >> Was there another distinct patch you were going to send for the above?
> > >
> > > Yes Nick. I was planning to send it once the rewrite stuff goes in.
> >
> > Hi Nick,
> >
> > Since you have practical use case, would it be possible to run your workflow
> > with perf rewrite and IBS patches applied? It will help us in finding/fixing
> > more bugs and upstreaming these changes.
>
> Hi Ravi,
> Sorry, I'm not able to load a custom kernel image on my employer
> provided workstation, and I never got approval to expense hardware for
> testing this otherwise.
>
> Was there ever any update on this? I'm on 6.1.25 now and still cant run
> $ perf record -e cycles:pp --call-graph lbr <any command to profile>
> $ cat /proc/cpuinfo
> ...
> model name : AMD Ryzen Threadripper PRO 3995WX 64-Cores
> ...
The commit 30093056f7b2 ("perf/amd/ibs: Make IBS a core pmu") in v6.2.
$ git name-rev --tags --refs=v[2-6].* 30093056f7b2
30093056f7b2 v6.2-rc1~176^2~16
https://git.kernel.org/torvalds/c/30093056f7b2
Thanks,
Namhyung
Hi Nick,
On 23-Jun-23 9:53 PM, Nick Desaulniers wrote:
> On Tue, Oct 11, 2022 at 10:05 PM Ravi Bangoria <[email protected]> wrote:
>>
>> On 12-Oct-22 9:36 AM, Ravi Bangoria wrote:
>>> On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
>>>> On Thu, Oct 6, 2022 at 8:56 PM Ravi Bangoria <[email protected]> wrote:
>>>>>
>>>>> +cc: PeterZ
>>>>>
>>>>>>>>>> +Ravi who may be able to say if there are any issues with the precise
>>>>>>>>>> sampling on AMD.
>>>>>>>>>
>>>>>>>>> Afaik cvcles:pp will use IBS but it doesn't support per-task profiling
>>>>>>>>> since it has no task context. Ravi is working on it..
>>>>>>>>
>>>>>>>> Right.
>>>>>>>> https://lore.kernel.org/lkml/[email protected]
>>>>>>>
>>>>>>> Cool, thanks for working on this Ravi.
>>>>>>>
>>>>>>> I'm not sure yet whether I may replace the kernel on my corporate
>>>>>>> provided workstation, so I'm not sure yet I can help test that patch.
>>>>>>>
>>>>>>> Can you confirm that
>>>>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>>>>>
>>>>>>> works with just that patch applied? Or is there more work required?
>>>>>>> What is the status of that patch?
>>>>>>>
>>>>>>> For context, we had difficulty upstreaming support for instrumentation
>>>>>>> based profile guided optimizations in the Linux kernel.
>>>>>>> https://lore.kernel.org/lkml/CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com/
>>>>>>> We'd like to be able to use either instrumentation or sampling to
>>>>>>> optimize our builds. The major barrier to sample based approaches are
>>>>>>> architecture / micro architecture issues with sample based profile
>>>>>>> data collection, and bitrot of data processing utilities.
>>>>>>> https://github.com/google/autofdo/issues/144
>>>>>>
>>>>>> On existing AMD Zen2, Zen3 the following cmdline:
>>>>>> $ perf record -e cycles:pp --freq=128 --call-graph lbr -- <command to profile>
>>>>>>
>>>>>> does not work. I see two reasons:
>>>>>>
>>>>>> 1. cycles:pp is likely converted into IBS op in cycle mode.
>>>>>> Current kernels do not support IBS in per-thread mode.
>>>>>> This is purely a kernel limitation
>>>>>
>>>>> Right, it's purely a kernel limitation. And below simple patch on top
>>>>> of event-context rewrite patch[1] should be sufficient to make cycles:pp
>>>>> working in per-process mode on AMD Zen.
>>>>>
>>>>> ---
>>>>> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
>>>>> index c251bc44c088..de01b5d27e40 100644
>>>>> --- a/arch/x86/events/amd/ibs.c
>>>>> +++ b/arch/x86/events/amd/ibs.c
>>>>> @@ -665,7 +665,7 @@ static struct perf_ibs perf_ibs_fetch = {
>>>>>
>>>>> static struct perf_ibs perf_ibs_op = {
>>>>> .pmu = {
>>>>> - .task_ctx_nr = perf_invalid_context,
>>>>> + .task_ctx_nr = perf_hw_context,
>>>>>
>>>>> .event_init = perf_ibs_init,
>>>>> .add = perf_ibs_add,
>>>>> ---
>>>>>
>>>>> [1]: https://lore.kernel.org/lkml/[email protected]
>>>>
>>>> Hi Ravi,
>>>> I didn't see the above diff in
>>>> https://lore.kernel.org/lkml/[email protected]/
>>>> Was there another distinct patch you were going to send for the above?
>>>
>>> Yes Nick. I was planning to send it once the rewrite stuff goes in.
>>
>> Hi Nick,
>>
>> Since you have practical use case, would it be possible to run your workflow
>> with perf rewrite and IBS patches applied? It will help us in finding/fixing
>> more bugs and upstreaming these changes.
>
> Hi Ravi,
> Sorry, I'm not able to load a custom kernel image on my employer
> provided workstation, and I never got approval to expense hardware for
> testing this otherwise.
>
> Was there ever any update on this? I'm on 6.1.25 now and still cant run
> $ perf record -e cycles:pp --call-graph lbr <any command to profile>
Per-process precise sampling on AMD platforms should work from 6.2-rc1
onward. However, --call-graph=lbr is not supported on AMD (hw limitation).
Thanks,
Ravi
On Sun, Jun 25, 2023 at 10:45 PM Ravi Bangoria <[email protected]> wrote:
>
> Hi Nick,
>
> On 23-Jun-23 9:53 PM, Nick Desaulniers wrote:
> >>> On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
> > Hi Ravi,
> > Sorry, I'm not able to load a custom kernel image on my employer
> > provided workstation, and I never got approval to expense hardware for
> > testing this otherwise.
> >
> > Was there ever any update on this? I'm on 6.1.25 now and still cant run
> > $ perf record -e cycles:pp --call-graph lbr <any command to profile>
>
> Per-process precise sampling on AMD platforms should work from 6.2-rc1
> onward.
Ok, I can wait for my employer to ship 6.2 on our workstations.
> However, --call-graph=lbr is not supported on AMD (hw limitation).
On any AMD uarches? Is there an equivalent? LBR encoding is compact
which makes working it much faster than DWARF or stack frame
unwinding.
--
Thanks,
~Nick Desaulniers
On 11-Jul-23 2:52 AM, Nick Desaulniers wrote:
> On Sun, Jun 25, 2023 at 10:45 PM Ravi Bangoria <[email protected]> wrote:
>>
>> Hi Nick,
>>
>> On 23-Jun-23 9:53 PM, Nick Desaulniers wrote:
>>>>> On 12-Oct-22 3:02 AM, Nick Desaulniers wrote:
>>> Hi Ravi,
>>> Sorry, I'm not able to load a custom kernel image on my employer
>>> provided workstation, and I never got approval to expense hardware for
>>> testing this otherwise.
>>>
>>> Was there ever any update on this? I'm on 6.1.25 now and still cant run
>>> $ perf record -e cycles:pp --call-graph lbr <any command to profile>
>>
>> Per-process precise sampling on AMD platforms should work from 6.2-rc1
>> onward.
>
> Ok, I can wait for my employer to ship 6.2 on our workstations.
>
>> However, --call-graph=lbr is not supported on AMD (hw limitation).
>
> On any AMD uarches? Is there an equivalent? LBR encoding is compact
> which makes working it much faster than DWARF or stack frame
> unwinding.
I understand that LBR call-stack is the fastest option but unfortunately
none of the current AMD uarch supports it.
Thanks,
Ravi