Re: [PATCH V9 1/4] perf/core: Add PERF_SAMPLE_DATA_PAGE_SIZE

[Peter Zijlstra]

On Tue, Oct 13, 2020 at 10:57:41AM -0400, Liang, Kan wrote:
> 
> 
> On 10/12/2020 4:48 AM, Will Deacon wrote:
> > On Sat, Oct 10, 2020 at 12:28:39AM +1100, Michael Ellerman wrote:
> > > Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
> > > > Patch 4 makes it all far worse by exposing it to pretty much everybody.
> > > > 
> > > > Now, I think we can fix at least the user mappings with the below delta,
> > > > but if archs are using non-page-table MMU sizes we'll need arch helpers.
> > > > 
> > > > ARM64 is in that last boat.
> > > 
> > > I think we probably need it to be weak so we can provide our own
> > > version.
> > 
> > I guess the same thing applies to us, but I can't really tell how accurate
> > this stuff needs to be for userspace. If it's trying to use the page-table
> > configuration to infer properties of the TLB, that's never going to be
> > reliable because the TLB can choose both to split and coalesce entries
> > as long as software can't tell.
> > 
> 
> Hi Peter,
> 
> It looks like everybody wants a __weak function. If so, I guess we should
> drop the generic code in this patch. For X86, we have existing functions to
> retrieve the page level and the page size. I think we don't need the generic
> code either.
> https://lkml.kernel.org/r/1549648509-12704-2-git-send-email-kan.liang@xxxxxxxxxxxxxxx/
> 
> Should I send a V10 patch to drop the generic code and implement an X86
> specific perf_get_page_size()? Will, Michael, and others can implement their
> version later separately.

Nah, that generic function, should work for 90% of all archs, it's just
a few oddballs that need something else.

Also, if we add that hugetlb exception, we'll even get the usermap for
those oddballs right.

I'll take this version after the merge window, I'll add __weak for the
oddballs and also add the hugetlb userspace thing on top.