original report: https://lore.kernel.org/all/CAKhLTr1UL3ePTpYjXOx2AJfNk8Ku2EdcEfu+CH1sf3Asr=B-Dw@xxxxxxxxxxxxxx/T/ When doing buffered writes with FGP_NOWAIT, under memory pressure, the system returned ENOMEM despite there was plenty of available memory. The user space used io_uring interface, which in turn submits I/O with FGP_NOWAIT (the fast path). retsnoop pointed to iomap_get_folio: 00:34:16.180612 -> 00:34:16.180651 TID/PID 253786/253721 (reactor-1/combined_tests): entry_SYSCALL_64_after_hwframe+0x76 do_syscall_64+0x82 __do_sys_io_uring_enter+0x265 io_submit_sqes+0x209 io_issue_sqe+0x5b io_write+0xdd xfs_file_buffered_write+0x84 iomap_file_buffered_write+0x1a6 32us [-ENOMEM] iomap_write_begin+0x408 iter=&{.inode=0xffff8c67aa031138,.len=4096,.flags=33,.iomap={.addr=0xffffffffffffffff,.length=4096,.type=1,.flags=3,.bdev=0x… pos=0 len=4096 foliop=0xffffb32c296b7b80 ! 4us [-ENOMEM] iomap_get_folio iter=&{.inode=0xffff8c67aa031138,.len=4096,.flags=33,.iomap={.addr=0xffffffffffffffff,.length=4096,.type=1,.flags=3,.bdev=0x… pos=0 len=4096 This is likely a regression caused by 66dabbb65d67 ("mm: return an ERR_PTR from __filemap_get_folio"), which performed the following changes: --- a/fs/iomap/buffered-io.c +++ b/fs/iomap/buffered-io.c @@ -468,19 +468,12 @@ EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate); struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos) { unsigned fgp = FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE | FGP_NOFS; - struct folio *folio; if (iter->flags & IOMAP_NOWAIT) fgp |= FGP_NOWAIT; - folio = __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT, + return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT, fgp, mapping_gfp_mask(iter->inode->i_mapping)); - if (folio) - return folio; - - if (iter->flags & IOMAP_NOWAIT) - return ERR_PTR(-EAGAIN); - return ERR_PTR(-ENOMEM); } Essentially, that patch is moving error picking decision to __filemap_get_folio, but it missed proper FGP_NOWAIT handling, so ENOMEM is being escaped to user space. Had it correctly returned -EAGAIN with NOWAIT, either io_uring or user space itself would be able to retry the request. It's not enough to patch io_uring since the iomap interface is the one responsible for it, and pwritev2(RWF_NOWAIT) and AIO interfaces must return the proper error too. The patch was tested with scylladb test suite (its original reproducer), and the tests all pass now when memory is pressured. Signed-off-by: Raphael S. Carvalho <raphaelsc@xxxxxxxxxxxx> --- mm/filemap.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/mm/filemap.c b/mm/filemap.c index 804d7365680c..b06bd6eedaf7 100644 --- a/mm/filemap.c +++ b/mm/filemap.c @@ -1986,8 +1986,15 @@ struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index, if (err == -EEXIST) goto repeat; - if (err) + if (err) { + /* + * Presumably ENOMEM, either from when allocating or + * adding folio (this one for xarray node) + */ + if (fgp_flags & FGP_NOWAIT) + err = -EAGAIN; return ERR_PTR(err); + } /* * filemap_add_folio locks the page, and for mmap * we expect an unlocked page. -- 2.48.1
