On Tue, 21 Feb 2012, Konstantin Khlebnikov wrote: > > > > Be particularly careful in compaction's isolate_migratepages() and > > vmscan's lumpy handling in isolate_lru_pages(): those approach the > > page by its physical location, and so can encounter pages which > > would not be found by any logical lookup. For those cases we have > > to change __isolate_lru_page() slightly: it must leave ClearPageLRU > > to the caller, because compaction and lumpy cannot safely interfere > > with a page until they have first isolated it and then locked lruvec. > > Yeah, this is most complicated part. Yes, I found myself leaving this patch until last when commenting. And was not entirely convinced by what I then said of move_account(). Indeed, I wondered if it might be improved and simplified by taking lruvec locks itself, in the manner that commit_charge lrucare ends up doing. > I found one race here, see below. Thank you! > > @@ -386,10 +364,24 @@ static isolate_migrate_t isolate_migrate > > continue; > > > > page_relock_lruvec(page,&lruvec); > > Here race with mem_cgroup_move_account() we hold lock for old lruvec, > while move_account() recharge page and put page back into other lruvec. > Thus we see PageLRU(), but below we isolate page from wrong lruvec. I expect you'll prove right on that, but I'm going to think about it more later. > > In my patch-set this is fixed with __wait_lru_unlock() [ spin_unlock_wait() ] > in mem_cgroup_move_account() Right now, after finishing mail, I want to concentrate on getting your series working under my swapping load. It's possible that I screwed up rediffing it to my slightly later base (though the only parts that appeared to need fixing up were as expected, near update_isolated_counts and move_active_pages_to_lru); but if so I'd expect that to show up differently. At present, although it runs fine with cgroup_disable=memory, with memcg two machines very soon hit that BUG at include/linux/mm_inline.h:41! when the lru_size or pages_count wraps around; on another it hit that precisely when I stopped the test. In all cases it's in release_pages from free_pages_and_swap_cache from tlb_flush_mmu from tlb_finish_mmu from exit_mmap from mmput from exit_mm from do_exit (but different processes exiting). Hugh -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>