On 07/02/2018 02:53 AM, Jan Kara wrote: > On Sun 01-07-18 17:56:53, john.hubb...@gmail.com wrote: >> From: John Hubbard <jhubb...@nvidia.com> >> > ... > >> @@ -904,12 +907,24 @@ static inline void get_page(struct page *page) >> */ >> VM_BUG_ON_PAGE(page_ref_count(page) <= 0, page); >> page_ref_inc(page); >> + >> + if (unlikely(PageDmaPinned(page))) >> + __get_page_for_pinned_dma(page); >> } >> >> static inline void put_page(struct page *page) >> { >> page = compound_head(page); >> >> + /* Because the page->dma_pinned_* fields are unioned with >> + * page->lru, there is no way to do classical refcount-style >> + * decrement-and-test-for-zero. Instead, PageDmaPinned(page) must >> + * be checked, in order to safely check if we are allowed to decrement >> + * page->dma_pinned_count at all. >> + */ >> + if (unlikely(PageDmaPinned(page))) >> + __put_page_for_pinned_dma(page); >> + > > These two are just wrong. You cannot make any page reference for > PageDmaPinned() account against a pin count. First, it is just conceptually > wrong as these references need not be long term pins, second, you can > easily race like: > > Pinner Random process > get_page(page) > pin_page_for_dma() > put_page(page) > -> oops, page gets unpinned too early >
I'll drop this approach, without mentioning any of the locking that is hiding in there, since that was probably breaking other rules anyway. :) Thanks for your patience in reviewing this. > So you really have to create counterpart to get_user_pages() - like > put_user_page() or whatever... It is inconvenient to have to modify all GUP > users but I don't see a way around that. OK, there will be a long-ish pause, while I go visit all the gup sites. I count about 88 callers, which is not nearly as crazy as my first casual grep showed, but still quite a chunk, since I have to track down where each one does its put_page call(s). It's definitely worth the effort, though. These pins just plain need some special handling in order to get everything correct. thanks, -- John Hubbard NVIDIA