Messages in this thread | | | Date | Tue, 3 Mar 2015 15:35:24 +0200 | From | "Kirill A. Shutemov" <> | Subject | Re: [PATCHv3 03/24] mm: avoid PG_locked on tail pages |
| |
On Tue, Mar 03, 2015 at 06:51:11PM +0530, Aneesh Kumar K.V wrote: > Rik van Riel <riel@redhat.com> writes: > > > -----BEGIN PGP SIGNED MESSAGE----- > > Hash: SHA1 > > > > On 02/12/2015 02:55 PM, Rik van Riel wrote: > >> On 02/12/2015 11:18 AM, Kirill A. Shutemov wrote: > > > >>> @@ -490,6 +493,7 @@ extern int > >>> wait_on_page_bit_killable_timeout(struct page *page, > >> > >>> static inline int wait_on_page_locked_killable(struct page *page) > >>> { + page = compound_head(page); if (PageLocked(page)) return > >>> wait_on_page_bit_killable(page, PG_locked); return 0; @@ -510,6 > >>> +514,7 @@ static inline void wake_up_page(struct page *page, int > >>> bit) */ static inline void wait_on_page_locked(struct page *page) > >>> { + page = compound_head(page); if (PageLocked(page)) > >>> wait_on_page_bit(page, PG_locked); } > >> > >> These are all atomic operations. > >> > >> This may be a stupid question with the answer lurking somewhere in > >> the other patches, but how do you ensure you operate on the right > >> page lock during a THP collapse or split? > > > > Kirill answered that question on IRC. > > > > The VM takes a refcount on a page before attempting to take a page > > lock, which prevents the THP code from doing anything with the > > page. In other words, while we have a refcount on the page, we > > will dereference the same page lock. > > Can we explain this more ? Don't we allow a thp split to happen even if > we have page refcount ?.
The patchset changes this. Have you read the cover letter?
-- Kirill A. Shutemov
| |