Message ID | 20230613215346.1022773-7-peterx@redhat.com (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | mm/gup: Unify hugetlb, speed up thp | expand |
On Tue, Jun 13, 2023 at 05:53:45PM -0400, Peter Xu wrote: > + if (page_increm > 1) > + WARN_ON_ONCE( > + try_grab_folio(compound_head(page), You don't need to call compound_head() here; try_grab_folio() works on tail pages just fine. > + page_increm - 1, > + foll_flags) == NULL); > + > + for (j = 0; j < page_increm; j++) { > + subpage = nth_page(page, j); > + pages[i+j] = subpage; > + flush_anon_page(vma, subpage, start + j * PAGE_SIZE); > + flush_dcache_page(subpage); You're better off calling flush_dcache_folio() right at the end.
On Wed, Jun 14, 2023 at 03:58:34PM +0100, Matthew Wilcox wrote: > On Tue, Jun 13, 2023 at 05:53:45PM -0400, Peter Xu wrote: > > + if (page_increm > 1) > > + WARN_ON_ONCE( > > + try_grab_folio(compound_head(page), > > You don't need to call compound_head() here; try_grab_folio() works > on tail pages just fine. I did it with caution because two things I'm not sure: either is_pci_p2pdma_page() or is_longterm_pinnable_page() inside, both calls is_zone_device_page() on the page*. But I just noticed try_grab_folio() is also used in gup_pte_range() where the thp can be pte mapped, so I assume we at least need that to handle tail page well. Do we perhaps need the compound_head() in try_grab_folio() as a separate patch? Or maybe I was wrong on is_zone_device_page()? > > > + page_increm - 1, > > + foll_flags) == NULL); > > + > > + for (j = 0; j < page_increm; j++) { > > + subpage = nth_page(page, j); > > + pages[i+j] = subpage; > > + flush_anon_page(vma, subpage, start + j * PAGE_SIZE); > > + flush_dcache_page(subpage); > > You're better off calling flush_dcache_folio() right at the end. Will do. Thanks,
On Wed, Jun 14, 2023 at 11:19:48AM -0400, Peter Xu wrote: > > > + for (j = 0; j < page_increm; j++) { > > > + subpage = nth_page(page, j); > > > + pages[i+j] = subpage; > > > + flush_anon_page(vma, subpage, start + j * PAGE_SIZE); > > > + flush_dcache_page(subpage); > > > > You're better off calling flush_dcache_folio() right at the end. > > Will do. Ah when I start to modify it I noticed it's a two-sided sword: we'll then also do flush dcache over the whole folio even if we gup one page. We'll start to get benefit only if some arch at least starts to impl flush_dcache_folio() (which seems to be none, right now..), and we'll already start to lose on amplifying the flush when gup on partial folio. Perhaps I still keep it as-is which will still be accurate, always faster than old code, and definitely not regress in any form?
On Tue, Jun 13, 2023 at 05:53:45PM -0400, Peter Xu wrote: > The acceleration of THP was done with ctx.page_mask, however it'll be > ignored if **pages is non-NULL. > > The old optimization was introduced in 2013 in 240aadeedc4a ("mm: > accelerate mm_populate() treatment of THP pages"). It didn't explain why > we can't optimize the **pages non-NULL case. It's possible that at that > time the major goal was for mm_populate() which should be enough back then. > > Optimize thp for all cases, by properly looping over each subpage, doing > cache flushes, and boost refcounts / pincounts where needed in one go. > > This can be verified using gup_test below: > > # chrt -f 1 ./gup_test -m 512 -t -L -n 1024 -r 10 > > Before: 13992.50 ( +-8.75%) > After: 378.50 (+-69.62%) > > Signed-off-by: Peter Xu <peterx@redhat.com> > --- > mm/gup.c | 36 +++++++++++++++++++++++++++++------- > 1 file changed, 29 insertions(+), 7 deletions(-) > > diff --git a/mm/gup.c b/mm/gup.c > index a2d1b3c4b104..cdabc8ea783b 100644 > --- a/mm/gup.c > +++ b/mm/gup.c > @@ -1210,16 +1210,38 @@ static long __get_user_pages(struct mm_struct *mm, > goto out; > } > next_page: > - if (pages) { > - pages[i] = page; > - flush_anon_page(vma, page, start); > - flush_dcache_page(page); > - ctx.page_mask = 0; > - } > - > page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); > if (page_increm > nr_pages) > page_increm = nr_pages; > + > + if (pages) { > + struct page *subpage; > + unsigned int j; > + > + /* > + * This must be a large folio (and doesn't need to > + * be the whole folio; it can be part of it), do > + * the refcount work for all the subpages too. > + * Since we already hold refcount on the head page, > + * it should never fail. > + * > + * NOTE: here the page may not be the head page > + * e.g. when start addr is not thp-size aligned. > + */ > + if (page_increm > 1) > + WARN_ON_ONCE( > + try_grab_folio(compound_head(page), > + page_increm - 1, > + foll_flags) == NULL); I'm not sure this should be warning but otherwise ignoring this returning NULL? This feels like a case that could come up in realtiy, e.g. folio_ref_try_add_rcu() fails, or !folio_is_longterm_pinnable(). Side-note: I _hate_ the semantics of GUP such that try_grab_folio() (invoked, other than for huge page cases, by the GUP-fast logic) will explicitly fail if neither FOLL_GET or FOLL_PIN are specified, differentiating it from try_grab_page() in this respect. This is a side-note and not relevant here, as all callers to __get_user_pages() either explicitly set FOLL_GET if not set by user (in __get_user_pages_locked()) or don't set pages (e.g. in faultin_vma_page_range()) > + > + for (j = 0; j < page_increm; j++) { > + subpage = nth_page(page, j); > + pages[i+j] = subpage; > + flush_anon_page(vma, subpage, start + j * PAGE_SIZE); > + flush_dcache_page(subpage); > + } > + } > + > i += page_increm; > start += page_increm * PAGE_SIZE; > nr_pages -= page_increm; > -- > 2.40.1 >
On Sat, Jun 17, 2023 at 09:27:22PM +0100, Lorenzo Stoakes wrote: > On Tue, Jun 13, 2023 at 05:53:45PM -0400, Peter Xu wrote: > > The acceleration of THP was done with ctx.page_mask, however it'll be > > ignored if **pages is non-NULL. > > > > The old optimization was introduced in 2013 in 240aadeedc4a ("mm: > > accelerate mm_populate() treatment of THP pages"). It didn't explain why > > we can't optimize the **pages non-NULL case. It's possible that at that > > time the major goal was for mm_populate() which should be enough back then. > > > > Optimize thp for all cases, by properly looping over each subpage, doing > > cache flushes, and boost refcounts / pincounts where needed in one go. > > > > This can be verified using gup_test below: > > > > # chrt -f 1 ./gup_test -m 512 -t -L -n 1024 -r 10 > > > > Before: 13992.50 ( +-8.75%) > > After: 378.50 (+-69.62%) > > > > Signed-off-by: Peter Xu <peterx@redhat.com> > > --- > > mm/gup.c | 36 +++++++++++++++++++++++++++++------- > > 1 file changed, 29 insertions(+), 7 deletions(-) > > > > diff --git a/mm/gup.c b/mm/gup.c > > index a2d1b3c4b104..cdabc8ea783b 100644 > > --- a/mm/gup.c > > +++ b/mm/gup.c > > @@ -1210,16 +1210,38 @@ static long __get_user_pages(struct mm_struct *mm, > > goto out; > > } > > next_page: > > - if (pages) { > > - pages[i] = page; > > - flush_anon_page(vma, page, start); > > - flush_dcache_page(page); > > - ctx.page_mask = 0; > > - } > > - > > page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); > > if (page_increm > nr_pages) > > page_increm = nr_pages; > > + > > + if (pages) { > > + struct page *subpage; > > + unsigned int j; > > + > > + /* > > + * This must be a large folio (and doesn't need to > > + * be the whole folio; it can be part of it), do > > + * the refcount work for all the subpages too. > > + * Since we already hold refcount on the head page, > > + * it should never fail. > > + * > > + * NOTE: here the page may not be the head page > > + * e.g. when start addr is not thp-size aligned. > > + */ > > + if (page_increm > 1) > > + WARN_ON_ONCE( > > + try_grab_folio(compound_head(page), > > + page_increm - 1, > > + foll_flags) == NULL); > > I'm not sure this should be warning but otherwise ignoring this returning > NULL? This feels like a case that could come up in realtiy, > e.g. folio_ref_try_add_rcu() fails, or !folio_is_longterm_pinnable(). Note that we hold already at least 1 refcount on the folio (also mentioned in the comment above this chunk of code), so both folio_ref_try_add_rcu() and folio_is_longterm_pinnable() should already have been called on the same folio and passed. If it will fail it should have already, afaict. I still don't see how that would trigger if the refcount won't overflow. Here what I can do is still guard this try_grab_folio() and fail the GUP if for any reason it failed. Perhaps then it means I'll also keep that one untouched in hugetlb_follow_page_mask() too. But I suppose keeping the WARN_ON_ONCE() seems still proper. Thanks,
On Mon, Jun 19, 2023 at 03:37:30PM -0400, Peter Xu wrote: > Here what I can do is still guard this try_grab_folio() and fail the GUP if > for any reason it failed. Perhaps then it means I'll also keep that one > untouched in hugetlb_follow_page_mask() too. But I suppose keeping the > WARN_ON_ONCE() seems still proper. Here's the outcome that I plan to post in the new version, taking care of try_grab_folio() failures even if it happens, meanwhile remove the compound_head() redundancy on the page. __get_user_pages(): ... ===8<=== /* * This must be a large folio (and doesn't need to * be the whole folio; it can be part of it), do * the refcount work for all the subpages too. * * NOTE: here the page may not be the head page * e.g. when start addr is not thp-size aligned. * try_grab_folio() should have taken care of tail * pages. */ if (page_increm > 1) { struct folio *folio; /* * Since we already hold refcount on the * large folio, this should never fail. */ folio = try_grab_folio(page, page_increm - 1, foll_flags); if (WARN_ON_ONCE(!folio)) { /* * Release the 1st page ref if the * folio is problematic, fail hard. */ gup_put_folio(page_folio(page), 1, foll_flags); ret = -EFAULT; goto out; } } ===8<=== Thanks,
diff --git a/mm/gup.c b/mm/gup.c index a2d1b3c4b104..cdabc8ea783b 100644 --- a/mm/gup.c +++ b/mm/gup.c @@ -1210,16 +1210,38 @@ static long __get_user_pages(struct mm_struct *mm, goto out; } next_page: - if (pages) { - pages[i] = page; - flush_anon_page(vma, page, start); - flush_dcache_page(page); - ctx.page_mask = 0; - } - page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); if (page_increm > nr_pages) page_increm = nr_pages; + + if (pages) { + struct page *subpage; + unsigned int j; + + /* + * This must be a large folio (and doesn't need to + * be the whole folio; it can be part of it), do + * the refcount work for all the subpages too. + * Since we already hold refcount on the head page, + * it should never fail. + * + * NOTE: here the page may not be the head page + * e.g. when start addr is not thp-size aligned. + */ + if (page_increm > 1) + WARN_ON_ONCE( + try_grab_folio(compound_head(page), + page_increm - 1, + foll_flags) == NULL); + + for (j = 0; j < page_increm; j++) { + subpage = nth_page(page, j); + pages[i+j] = subpage; + flush_anon_page(vma, subpage, start + j * PAGE_SIZE); + flush_dcache_page(subpage); + } + } + i += page_increm; start += page_increm * PAGE_SIZE; nr_pages -= page_increm;
The acceleration of THP was done with ctx.page_mask, however it'll be ignored if **pages is non-NULL. The old optimization was introduced in 2013 in 240aadeedc4a ("mm: accelerate mm_populate() treatment of THP pages"). It didn't explain why we can't optimize the **pages non-NULL case. It's possible that at that time the major goal was for mm_populate() which should be enough back then. Optimize thp for all cases, by properly looping over each subpage, doing cache flushes, and boost refcounts / pincounts where needed in one go. This can be verified using gup_test below: # chrt -f 1 ./gup_test -m 512 -t -L -n 1024 -r 10 Before: 13992.50 ( +-8.75%) After: 378.50 (+-69.62%) Signed-off-by: Peter Xu <peterx@redhat.com> --- mm/gup.c | 36 +++++++++++++++++++++++++++++------- 1 file changed, 29 insertions(+), 7 deletions(-)