On Tue, Dec 15, 2015 at 09:52:33AM +0100, Michal Hocko wrote: > On Mon 14-12-15 14:04:56, Kirill A. Shutemov wrote: > > On Wed, Dec 09, 2015 at 02:02:05PM +0100, Michal Hocko wrote: > > > Hi Kirill, > > > > [ sorry for late reply, just back from vacation. ] > > > > > while looking at the issue reported by Minchan [1] I have noticed that > > > there is nothing to prevent from "isolating" a tail page from LRU because > > > isolate_lru_page checks PageLRU which is > > > PAGEFLAG(LRU, lru, PF_HEAD) > > > so it is checked on the head page rather than the given page directly > > > but the rest of the operation is done on the given (tail) page. > > > > Looks like most (all?) callers already exclude PTE-mapped THP already one > > way or another. > > I can see e.g. do_move_page_to_node_array not doing a similar thing. It > isolates and then migrates potentially a tail page. No, it doesn't. follow_page(FOLL_SPLIT) would split THP pages. > I haven't looked closer whether there is other hand break on the way > though. The point I was trying to make is that this is really _subtle_. > We are changing something else than we operate later on. > > > Probably, VM_BUG_ON_PAGE(PageTail(page), page) in isolate_lru_page() would > > be appropriate. > > > > > This is really subtle because this expects that every caller of this > > > function checks for the tail page otherwise we would clobber statistics > > > and who knows what else (I haven't checked that in detail) as the page > > > cannot be on the LRU list and the operation makes sense only on the head > > > page. > > > > > > Would it make more sense to make PageLRU PF_ANY? That would return > > > false for PageLRU on any tail page and so it would be ignored by > > > isolate_lru_page. > > > > I don't think this is right way to go. What we put on LRU is compound > > page, not 4k subpages. PageLRU() should return true if the compound page > > is on LRU regardless if you ask for head or tail page. > > Hmm, but then we should operate on the head page because that is what > PageLRU operated on, no? head page is what linked into LRU, but not nessesary the way we obtain the page to check. If we check PageLRU(pte_page(*pte)) it should produce the right result. > > False-negatives PageLRU() can be as bad as bug Minchan reported, but > > perhaps more silent. > > -- > Michal Hocko > SUSE Labs -- Kirill A. Shutemov -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>