Re: [PATCH V3] perf vendor events riscv: add T-HEAD C9xx JSON file

[Namhyung Kim]

Hello,

On Fri, Jun 23, 2023 at 6:33 PM Inochi Amaoto <inochiama@xxxxxxxxxxx> wrote:
>
> Hi, Namhyung
>
> Since there is a fault in T-HEAD documentation, I am not sure whether
> the perf events are correct. As a result, I suggest temporarily suppressing
> this patch until I extract the correct events from its vendor perf driver.
> I will prepare a new V4 patch once it is finished.

Sure, thanks for letting me know.

>
> As for c9xx wildcard, the T-HEAD provides a `MCPUID` vendor CSR to allow
> its CKLINK to get the detail CPU info. The format of this CSR are:
>
> ------------------------------------------------
> |31   28|27  26|25    22|21   18|17    8|7    0|
> | index | WLRL | family | class | model | WLRL |
> ------------------------------------------------
>
> And for C9xx series (only index 0000 is vaild for us, as `MCPUID` also
> provides other index).
>
> | 0000 | xx | 0100 | class | xxxxxxxxxx | xxxxxxxx |
>
> The class codes are:
>
> C910: 0011
> c906: 0100
>
> The CSR is a M-mode only CSR, so now I'm exploring a clean way to
> integrate this CSR into the kernel. Any advice?

I don't know about the details.  Is this CSR available from user space?
If not, you could add it somewhere in the sysfs.

Thanks,
Namhyung


> > Hello,
> >
> > On Thu, May 18, 2023 at 2:50 AM Inochi Amaoto <inochiama@xxxxxxxxxxx> wrote:
> > >
> > >> licheerv # perf record
> > >> [  432.015618] watchdog: BUG: soft lockup - CPU#0 stuck for 26s!
> > >> [perf:117]
> > >> [  460.015617] watchdog: BUG: soft lockup - CPU#0 stuck for 52s!
> > >> [perf:117]
> > >> [  488.015616] watchdog: BUG: soft lockup - CPU#0 stuck for 78s!
> > >> [perf:117]
> > >> [  516.015617] watchdog: BUG: soft lockup - CPU#0 stuck for 104s!
> > >> [perf:117]
> > >>
> > >> But that's not related to your patch anyway.
> > >
> > > Same issue on c920, but it did not always occur.
> > > Like a sbi issue for T-HEAD cpus.
> > >
> > >> I am strongly against using "c9xx" wildcard, i would prefer declaring
> > >> them separate (especially taking in mind that c920 is c910 with vector
> > >> - AFAIK), but that's up to Arnaldo to decide.
> > >
> > > AFAIK, there is no reliable way to distinguish c906 and c910 cores. And
> > > the events of c910 and c920 are the same (according to the draft document
> > > of the c920).
> > >
> > > Anyway, I agree to let Arnaldo decide.
> > >
> > >> Tested-by: Nikita Shubin <n.shubin@xxxxxxxxx>
> >
> > I'm collecting patches on behalf of Arnaldo this time.
> > It seems this patch was not picked up for a long time.
> >
> > I think we can make changes for the c9xx wildcard later
> > if needed.  I'll process it in the current form.
> >
> > Thanks,
> > Namhyung
> >