On 12/17/2013 01:31 AM, Andrew Morton wrote:
On Mon, 16 Dec 2013 11:14:15 +0100 Vlastimil Babka <vba...@suse.cz> wrote:Commit 7225522bb ("mm: munlock: batch non-THP page isolation and munlock+putback using pagevec" introduced __munlock_pagevec() to speed up munlock by holding lru_lock over multiple isolated pages. Pages that fail to be isolated are put_back() immediately, also within the lock. This can lead to deadlock when __munlock_pagevec() becomes the holder of the last page pin and put_back() leads to __page_cache_release() which also locks lru_lock. The deadlock has been observed by Sasha Levin using trinity. This patch avoids the deadlock by deferring put_back() operations until lru_lock is released. Another pagevec (which is also used by later phases of the function is reused to gather the pages for put_back() operation. ...Thanks for fixing this one. I'll cross it off the rather large list of recent MM regressions :(
Well I made this one in the first place :/
--- a/mm/mlock.c +++ b/mm/mlock.c @@ -295,10 +295,12 @@ static void __munlock_pagevec(struct pagevec *pvec, struct zone *zone) { int i; int nr = pagevec_count(pvec); - int delta_munlocked = -nr; + int delta_munlocked; struct pagevec pvec_putback; int pgrescued = 0; + pagevec_init(&pvec_putback, 0); + /* Phase 1: page isolation */ spin_lock_irq(&zone->lru_lock); for (i = 0; i < nr; i++) { @@ -327,16 +329,22 @@ skip_munlock: /* * We won't be munlocking this page in the next phase * but we still need to release the follow_page_mask() - * pin. + * pin. We cannot do it under lru_lock however. If it's + * the last pin, __page_cache_release would deadlock. */ + pagevec_add(&pvec_putback, pvec->pages[i]); pvec->pages[i] = NULL; - put_page(page); - delta_munlocked++; } } + delta_munlocked = -nr + pagevec_count(&pvec_putback); __mod_zone_page_state(zone, NR_MLOCK, delta_munlocked); spin_unlock_irq(&zone->lru_lock); + /* Now we can release pins of pages that we are not munlocking */ + for (i = 0; i < pagevec_count(&pvec_putback); i++) { + put_page(pvec_putback.pages[i]); + } +We could just do --- a/mm/mlock.c~mm-munlock-fix-deadlock-in-__munlock_pagevec-fix +++ a/mm/mlock.c @@ -341,12 +341,9 @@ skip_munlock: spin_unlock_irq(&zone->lru_lock); /* Now we can release pins of pages that we are not munlocking */ - for (i = 0; i < pagevec_count(&pvec_putback); i++) { - put_page(pvec_putback.pages[i]); - } + pagevec_release(&pvec_putback); /* Phase 2: page munlock */ - pagevec_init(&pvec_putback, 0); for (i = 0; i < nr; i++) { struct page *page = pvec->pages[i];
Yeah that looks nicer.
The lru_add_drain() is unnecessary overhead here. What do you think?
I would expect these isolation failures to be sufficiently rare so that it doesn't matter. Especially in process exit path which was the original target of my munlock work. But I don't have any numbers and my mmtests benchmark for munlock is most likely too simple to trigger this. But even once per pagevec the drain shouldn't hurt I guess...
-- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
