On Wed, Nov 04, 2020 at 12:11:16PM -0500, Liang, Kan wrote:
> On 10/13/2020 12:34 PM, Peter Zijlstra wrote:
> > @@ -7037,13 +7057,20 @@ static u64 __perf_get_page_size(struct m
> > return 0;
> > }
> > + page = pte_page(*pte);
> > + if (PageHuge(page)) {
> > + u64 size = page_size(compound_head(page));
> > + pte_unmap(pte);
> > + return size;
> > + }
> > +
>
> The PageHuge() check for PTE crashes my machine when I did page size test.
> (Sorry, I didn't find the issue earlier. I just found some time to re-run
> the page size test.)
>
> It seems we don't need the check for PTE here. The size should be always
> PAGE_SIZE, no? After I remove the check, everything looks good.
That's the thing, an architecture could have non-page-table aligned
huge-pages. For example using 4 consecutive 4k pages to create 16k
pages. In that case the above code would trigger and find a 16k compound
page with HUGETLB_PAGE_DTOR (assuming it was created through hugetlbfs).
What is this page size test; I'd like to reproduce.
