From: NeilBrown Subject: mm: reclaim mustn't enter FS for SWP_FS_OPS swap-space If swap-out is using filesystem operations (SWP_FS_OPS), then it is not safe to enter the FS for reclaim. So only down-grade the requirement for swap pages to __GFP_IO after checking that SWP_FS_OPS are not being used. This makes the calculation of "may_enter_fs" slightly more complex, so move it into a separate function. with that done, there is little value in maintaining the bool variable any more. So replace the may_enter_fs variable with a may_enter_fs() function. This removes any risk for the variable becoming out-of-date. Link: https://lkml.kernel.org/r/164859778124.29473.16176717935781721855.stgit@noble.brown Signed-off-by: NeilBrown Reviewed-by: Christoph Hellwig Tested-by: David Howells Tested-by: Geert Uytterhoeven Cc: Hugh Dickins Cc: Mel Gorman Cc: Trond Myklebust Signed-off-by: Andrew Morton --- mm/swap.h | 8 ++++++++ mm/vmscan.c | 29 ++++++++++++++++++++--------- 2 files changed, 28 insertions(+), 9 deletions(-) --- a/mm/swap.h~mm-reclaim-mustnt-enter-fs-for-swp_fs_ops-swap-space +++ a/mm/swap.h @@ -50,6 +50,10 @@ struct page *swap_cluster_readahead(swp_ struct page *swapin_readahead(swp_entry_t entry, gfp_t flag, struct vm_fault *vmf); +static inline unsigned int page_swap_flags(struct page *page) +{ + return page_swap_info(page)->flags; +} #else /* CONFIG_SWAP */ static inline int swap_readpage(struct page *page, bool do_poll) { @@ -129,5 +133,9 @@ static inline void clear_shadow_from_swa { } +static inline unsigned int page_swap_flags(struct page *page) +{ + return 0; +} #endif /* CONFIG_SWAP */ #endif /* _MM_SWAP_H */ --- a/mm/vmscan.c~mm-reclaim-mustnt-enter-fs-for-swp_fs_ops-swap-space +++ a/mm/vmscan.c @@ -1502,6 +1502,22 @@ static unsigned int demote_page_list(str return nr_succeeded; } +static bool may_enter_fs(struct page *page, gfp_t gfp_mask) +{ + if (gfp_mask & __GFP_FS) + return true; + if (!PageSwapCache(page) || !(gfp_mask & __GFP_IO)) + return false; + /* + * We can "enter_fs" for swap-cache with only __GFP_IO + * providing this isn't SWP_FS_OPS. + * ->flags can be updated non-atomicially (scan_swap_map_slots), + * but that will never affect SWP_FS_OPS, so the data_race + * is safe. + */ + return !data_race(page_swap_flags(page) & SWP_FS_OPS); +} + /* * shrink_page_list() returns the number of reclaimed pages */ @@ -1528,7 +1544,7 @@ retry: struct page *page; struct folio *folio; enum page_references references = PAGEREF_RECLAIM; - bool dirty, writeback, may_enter_fs; + bool dirty, writeback; unsigned int nr_pages; cond_resched(); @@ -1553,9 +1569,6 @@ retry: if (!sc->may_unmap && page_mapped(page)) goto keep_locked; - may_enter_fs = (sc->gfp_mask & __GFP_FS) || - (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); - /* * The number of dirty pages determines if a node is marked * reclaim_congested. kswapd will stall and start writing @@ -1598,7 +1611,7 @@ retry: * not to fs). In this case mark the page for immediate * reclaim and continue scanning. * - * Require may_enter_fs because we would wait on fs, which + * Require may_enter_fs() because we would wait on fs, which * may not have submitted IO yet. And the loop driver might * enter reclaim, and deadlock if it waits on a page for * which it is needed to do the write (loop masks off @@ -1630,7 +1643,7 @@ retry: /* Case 2 above */ } else if (writeback_throttling_sane(sc) || - !PageReclaim(page) || !may_enter_fs) { + !PageReclaim(page) || !may_enter_fs(page, sc->gfp_mask)) { /* * This is slightly racy - end_page_writeback() * might have just cleared PageReclaim, then @@ -1720,8 +1733,6 @@ retry: goto activate_locked_split; } - may_enter_fs = true; - /* Adding to swap updated mapping */ mapping = page_mapping(page); } @@ -1792,7 +1803,7 @@ retry: if (references == PAGEREF_RECLAIM_CLEAN) goto keep_locked; - if (!may_enter_fs) + if (!may_enter_fs(page, sc->gfp_mask)) goto keep_locked; if (!sc->may_writepage) goto keep_locked; _