Re: [PATCH v5 09/19] arch_topology: Use the last level cache information from the cacheinfo

From: Sudeep Holla
Date: Wed Jun 29 2022 - 15:44:55 EST


On Wed, Jun 29, 2022 at 07:25:41PM +0000, Conor.Dooley@xxxxxxxxxxxxx wrote:
>
>
> On 29/06/2022 20:12, Sudeep Holla wrote:
> > On Wed, Jun 29, 2022 at 06:56:29PM +0000, Conor.Dooley@xxxxxxxxxxxxx wrote:
> >> On 29/06/2022 19:47, Sudeep Holla wrote:
> >>> On Wed, Jun 29, 2022 at 06:18:25PM +0000, Conor.Dooley@xxxxxxxxxxxxx wrote:
> >>>> On 29/06/2022 18:49, Conor.Dooley@xxxxxxxxxxxxx wrote:
> >>>>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> >>>>>
> >>>>> On 27/06/2022 17:50, Sudeep Holla wrote:
> >>>>>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> >>>>>>
> >>>>>> The cacheinfo is now initialised early along with the CPU topology
> >>>>>> initialisation. Instead of relying on the LLC ID information parsed
> >>>>>> separately only with ACPI PPTT elsewhere, migrate to use the similar
> >>>>>> information from the cacheinfo.
> >>>>>>
> >>>>>> This is generic for both DT and ACPI systems. The ACPI LLC ID information
> >>>>>> parsed separately can now be removed from arch specific code.
> >>>>>
> >>>>> Hey Sudeep,
> >>>>> I bisected broken boot on PolarFire SoC to this patch in next-20220629 :/
> >>>>> I suspect the issue is a missing "next-level-cache" in the the dt:
> >>>>> arch/riscv/boot/dts/microchip/mpfs.dtsi
> >>>
> >>> Good that I included this in -next, I had not received any feedback from
> >>> RISC-V even after 5 iterations.
> >>
> >> I'll be honest, I saw the titles and CC list and made some incorrect
> >> assumptions as to whether looking at it was worthwhile! I am not at
> >> this all too long and what is/isn't important to look at often is not
> >> obvious to me.
> >
> > No worries, that's why I thought better to include in -next to get some
> > attention and I did get it this time, hurray! 😄
> >
> >> But hey, our CI boots -next every day for a reason ;)
> >>
> >
> > Good to know and that is really great. Anyways let me know if the diff I sent
> > helps. I strongly suspect that is the reason, but I may be wrong.
>
> Aye, I'll get back to you on that one in a moment or two
>

Sure, take your time.

> >
> >>> I also see this DTS is very odd. It also
> >>> states CPU0 doesn't have L1-D$ while the other 4 CPUs have L1-D$. Is that
> >>> a mistake or is it the reality ?
> >>
> >> AFAIK, reality. It's the same for the SiFive fu540 (with which this shares
> >> a core complex. See page 12:
> >> https://static.dev.sifive.com/FU540-C000-v1.0.pdf
> >>
> >>> Another breakage in userspace cacheinfo
> >>> sysfs entry of cpu0 has both I$ and D$.
> >>
> >> Could you clarify what this means please?
> >
> > Ignore me if the cpu0 really doesn't have L1-D$. However the userspace
> > sysfs cacheinfo is incomplete without linking L2, so it can be considered
> > as wrong info presented to the user.
>
> Yeah, I'll send a patch hooking up the L2.
> It wasn't in the initial fu540 dtsi so I guess it was added after the
> initial dts for my stuff was created based on that.
>

Thanks!

> >
> > Check /sys/devices/system/cpu/cpu<n>/cache/index<i>/*.
> > L2 won't be present there as the link with next-level-cache is missing.
> > So userspace can interpret this as absence of L2.
> >
>
> # cat /sys/devices/system/cpu/cpu0/cache/index0/
> coherency_line_size shared_cpu_list type
> level shared_cpu_map uevent
> number_of_sets size ways_of_associativity
> # ls /sys/devices/system/cpu/cpu0/cache/
> index0 index1 uevent
> # cat /sys/devices/system/cpu/cpu0/cache/index0/level
> 1
> # cat /sys/devices/system/cpu/cpu0/cache/index1/level
> 1
>
Ideally there must /sys/devices/system/cpu/cpu*/cache/index2/level
which reads 2 once you link it in the DT.

> cpu0 is /not/ the one with only instruction cache, that is not
> running Linux.

Ah, so there Linux runs only on cpu 1-4 ?

--
Regards,
Sudeep