On Fri, 22 Nov 2019, Johannes Weiner wrote: > > But that leaves me with one more worry: compaction. We locked out > charge moving now, so between that and knowing that the page is alive, > we have page->mem_cgroup stable. But compaction doesn't know whether > the page is alive - it comes from a pfn and finds out using PageLRU. > > In the current code, pgdat->lru_lock remains the same before and after > the page is charged to a cgroup, so once compaction has that locked > and it observes PageLRU, it can go ahead and isolate the page. > > But lruvec->lru_lock changes during charging, and then compaction may > hold the wrong lock during isolation: > > compaction: generic_file_buffered_read: > > page_cache_alloc() > > !PageBuddy() > > lock_page_lruvec(page) > lruvec = mem_cgroup_page_lruvec() > spin_lock(&lruvec->lru_lock) > if lruvec != mem_cgroup_page_lruvec() > goto again > > add_to_page_cache_lru() > mem_cgroup_commit_charge() > page->mem_cgroup = foo > lru_cache_add() > __pagevec_lru_add() > SetPageLRU() > > if PageLRU(page): > __isolate_lru_page() > > I don't see what prevents the lruvec from changing under compaction, > neither in your patches nor in Hugh's. Maybe I'm missing something? Speaking for my patches only: I'm humbled, I think you have caught me, I cannot find any argument against the race you suggest here. The race with mem_cgroup_move_account(), which Konstantin pointed out in 2012's https://lore.kernel.org/lkml/4F433418.3010401@xxxxxxxxxx/ but I later misunderstood, and came to think I needed no patch against, until this week coming to perceive the same race in isolate_lru_page(): that one is easily and satisfactorily fixed by holding lruvec lock in mem_cgroup_move_account() - embarrassing, but not too serious. Your race here (again, lruvec lock taken then PageLRU observed, but page->mem_cgroup changed in between) really questions my whole scheme: I am not going to propose a solution now, I'll have to go back and recheck my assumptions all over. Certainly isolate_migratepage_block() has a harder job than any other, but I need to re-review it all. Maybe we got it right back in the days of PageCgroupUsed, and then I paid too little attention when rebasing to your welcome simplifications. I don't think any of us want to bring back PageCgroupUsed! And maybe we could get it right by always holding lruvec lock in commit_charge(), lrucare or not; but that's a much hotter path, and not a change I'd expect anyone to embrace. I'll go away and re-examine it all; probably start by verifying that your race actually happens in practice, though we never observed it. Heavy-hearted thanks, Hannes! Hugh
