[External] Re: Current status of RISC-V init_cache_level()

[Conor Dooley]

On Fri, Jan 19, 2024 at 05:24:48PM +0800, yunhui cui wrote:
> Hi Conor,
> 
> On Fri, Jan 19, 2024 at 4:59 PM yunhui cui <cuiyunhui at bytedance.com> wrote:
> >
> > Hi Conor，
> >
> > On Fri, Jan 19, 2024 at 4:22 PM Conor Dooley <conor.dooley at microchip.com> wrote:
> > >
> > > On Fri, Jan 19, 2024 at 10:32:38AM +0800, yunhui cui wrote:
> > > > On Thu, Jan 18, 2024 at 9:25 PM Conor Dooley <conor at kernel.org> wrote:
> > > > > On Thu, Jan 18, 2024 at 07:40:42PM +0800, yunhui cui wrote:
> > > > >
> > > > > Firstly, please don't send me off-list mails about such things
> > > > > and instead, please reply to the relevant threads on lkml.
> > > > >
> > > > > > There is no cache subdirectory in /sys/devices/system/cpu/cpu0/, so
> > > > > > lscpu cannot see the cache information. I found that the reason is
> > > > > > that init_cache_level() is not implemented on RISC-V.
> > > > >
> > > > > What version of the kernel are you using? I had a brief check to make
> > > > > sure something had not gone awry recently and I could see them on my
> > > > > system. What do the cpu nodes in your DT look like, assuming you are
> > > > > on a DT system?
> > > >
> > > > The results of top commit using linux-next on qemu, It should not
> > > > matter if it is combined with DT, because the following function flow
> > > > directly fails.
> > >
> > > Oh no, it totally does matter what the DT looks like. The default DT in
> > > QEMU's virt machine does not populate any cache properties. On a system
> > > where the DT populates these values, init_of_cache_level() will ensure
> > > that the correct sysfs entries are set up.
> > >
> > > > cacheinfo_sysfs_init()...init_cache_level()
> > > > int __weak init_cache_level(unsigned int cpu)
> > > > {
> > > > return -ENOENT;
> > > > }
> > > > Is it necessary to implement an init_cache_level() on RISC-V like other arches?
> > >
> > > In order to implement that function, we would need some mechanism for
> > > finding that information. Where do you suggest we get it from, if it is not
> > > provided to us from DT?
> >
> > It's not that we don't get cache information from DT. What I want to
> > express is that we need to implement init_cache_level() on RISC-V,
> > otherwise cacheinfo_cpu_online() will return from
> > detect_cache_attributes() and will not execute cache_add_dev(), so it
> > will not be in/ sys/devices/system/cpu/cpux/ creates a "cache" node.
> 
> What do you think? If so, I will post a patch to fix it.

By all means if you have a patch that you think solves a problem, post
it. Be sure to CC Pierre and Sudeep if you do. Also explain exactly the
situation in which you encountered it and how to reproduce.

I briefly tested a hotunplug and replug on my setup here and the cache
information was populated correctly on replug, so there must be some
specific situation in which this is hit that I am not aware of as the
hotplug side of things I am not super au fait with.

Cheers,
Conor.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20240119/89b25052/attachment.sig>