On Mon, May 23, 2022 at 10:16:58PM -0700, Minchan Kim wrote: > On Mon, May 23, 2022 at 07:55:25PM -0700, John Hubbard wrote: > > On 5/23/22 09:33, Minchan Kim wrote: > > ... > > > > So then: > > > > > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > > > index 0e42038382c1..b404f87e2682 100644 > > > > +++ b/mm/page_alloc.c > > > > @@ -482,7 +482,12 @@ unsigned long __get_pfnblock_flags_mask(const struct page *page, > > > > word_bitidx = bitidx / BITS_PER_LONG; > > > > bitidx &= (BITS_PER_LONG-1); > > > > > > > > - word = bitmap[word_bitidx]; > > > > + /* > > > > + * This races, without locks, with set_pageblock_migratetype(). Ensure > > > set_pfnblock_flags_mask would be better? > > > > + * a consistent (non-tearing) read of the memory array, so that results, > > > > > > Thanks for proceeding and suggestion, John. > > > > > > IIUC, the load tearing wouldn't be an issue since [1] fixed the issue. > > > > Did it? [1] fixed something, but I'm not sure we can claim that that > > code is now safe against tearing in all possible cases, especially given > > the recent discussion here. Specifically, having this code do a read, > > then follow that up with calculations, seems correct. Anything else is > > The load tearing you are trying to explain in the comment would be > solved by [1] since the bits will always align on a word and accessing > word size based on word aligned address is always atomic so there is > no load tearing problem IIUC. That is not technically true. It is exactly the sort of thing READ_ONCE is intended to guard against. > Instead of the tearing problem, what we are trying to solve with > READ_ONCE is to prevent refetching when the function would be > inlined in the future. It is the same problem, who is to say it doesn't refetch while doing the maths? Jason
