From c4ed9b0fb10132a9683362ae90caf4c64a29062c Mon Sep 17 00:00:00 2001 From: Shaun Tancheff Date: Wed, 6 May 2020 03:19:48 -0500 Subject: [PATCH 1/1] LU-13476 llite: Fix lock ordering in pagevec_dirty In vvp_set_pagevec_dirty lock order between i_pages and lock_page_memcg was inverted with the expectation that no other users would conflict. However in vvp_page_completion_write the call to test_clear_page_writeback does expect to be able to lock_page_memcg then lock i_pages which appears to conflict with the original analysis. The reported case shows as RCU stalls with vvp_set_pagevec_dirty blocked attempting to lock i_pages. Fixes: a7299cb012f ("LU-9920 vvp: dirty pages with pagevec") HPE-bug-id: LUS-8798 Signed-off-by: Shaun Tancheff Change-Id: I47c2107ddbef4a76325928e982abfc0ea666f39b Reviewed-on: https://review.whamcloud.com/38317 Tested-by: jenkins Tested-by: Maloo Reviewed-by: Wang Shilong Reviewed-by: Patrick Farrell Reviewed-by: Oleg Drokin --- lustre/llite/vvp_io.c | 34 +++++++++++++++++++++------------- 1 file changed, 21 insertions(+), 13 deletions(-) diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index 94df8b5..5fb6e13 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -988,21 +988,33 @@ void vvp_set_pagevec_dirty(struct pagevec *pvec) struct page *page = pvec->pages[0]; struct address_space *mapping = page->mapping; unsigned long flags; + unsigned long skip_pages = 0; int count = pagevec_count(pvec); int dirtied = 0; - int i = 0; + int i; ENTRY; - /* From set_page_dirty */ - for (i = 0; i < count; i++) - ClearPageReclaim(pvec->pages[i]); - + BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG); LASSERTF(page->mapping, "mapping must be set. page %p, page->private (cl_page) %p\n", page, (void *) page->private); - /* Rest of code derived from __set_page_dirty_nobuffers */ + for (i = 0; i < count; i++) { + page = pvec->pages[i]; + + ClearPageReclaim(page); + + lock_page_memcg(page); + if (TestSetPageDirty(page)) { + /* page is already dirty .. no extra work needed + * set a flag for the i'th page to be skipped + */ + unlock_page_memcg(page); + skip_pages |= (1 << i); + } + } + ll_xa_lock_irqsave(&mapping->i_pages, flags); /* Notes on differences with __set_page_dirty_nobuffers: @@ -1013,17 +1025,13 @@ void vvp_set_pagevec_dirty(struct pagevec *pvec) * 3. No mapping is impossible. (Race w/truncate mentioned in * dirty_nobuffers should be impossible because we hold the page lock.) * 4. All mappings are the same because i/o is only to one file. - * 5. We invert the lock order on lock_page_memcg(page) and the mapping - * xa_lock, but this is the only function that should use that pair of - * locks and it can't race because Lustre locks pages throughout i/o. */ for (i = 0; i < count; i++) { page = pvec->pages[i]; - lock_page_memcg(page); - if (TestSetPageDirty(page)) { - unlock_page_memcg(page); + /* if the i'th page was unlocked above, skip it here */ + if ((skip_pages >> i) & 1) continue; - } + LASSERTF(page->mapping == mapping, "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n", page, page->mapping, mapping); -- 1.8.3.1