From: Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> Subject: mm/filemap.c: fix a data race in filemap_fault() struct file_ra_state ra.mmap_miss could be accessed concurrently during page faults as noticed by KCSAN, BUG: KCSAN: data-race in filemap_fault / filemap_map_pages write to 0xffff9b1700a2c1b4 of 4 bytes by task 3292 on cpu 30: filemap_fault+0x920/0xfc0 do_sync_mmap_readahead at mm/filemap.c:2384 (inlined by) filemap_fault at mm/filemap.c:2486 __xfs_filemap_fault+0x112/0x3e0 [xfs] xfs_filemap_fault+0x74/0x90 [xfs] __do_fault+0x9e/0x220 do_fault+0x4a0/0x920 __handle_mm_fault+0xc69/0xd00 handle_mm_fault+0xfc/0x2f0 do_page_fault+0x263/0x6f9 page_fault+0x34/0x40 read to 0xffff9b1700a2c1b4 of 4 bytes by task 3313 on cpu 32: filemap_map_pages+0xc2e/0xd80 filemap_map_pages at mm/filemap.c:2625 do_fault+0x3da/0x920 __handle_mm_fault+0xc69/0xd00 handle_mm_fault+0xfc/0x2f0 do_page_fault+0x263/0x6f9 page_fault+0x34/0x40 Reported by Kernel Concurrency Sanitizer on: CPU: 32 PID: 3313 Comm: systemd-udevd Tainted: G W L 5.5.0-next-20200210+ #1 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 ra.mmap_miss is used to contribute the readahead decisions, a data race could be undesirable. Both the read and write is only under non-exclusive mmap_sem, two concurrent writers could even underflow the counter. Fix the underflow by writing to a local variable before committing a final store to ra.mmap_miss given a small inaccuracy of the counter should be acceptable. Link: http://lkml.kernel.org/r/20200211030134.1847-1-cai@xxxxxx Signed-off-by: Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> Signed-off-by: Qian Cai <cai@xxxxxx> Tested-by: Qian Cai <cai@xxxxxx> Reviewed-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx> Cc: Marco Elver <elver@xxxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- mm/filemap.c | 20 +++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-) --- a/mm/filemap.c~mm-filemap-fix-a-data-race-in-filemap_fault +++ a/mm/filemap.c @@ -2468,6 +2468,7 @@ static struct file *do_sync_mmap_readahe struct address_space *mapping = file->f_mapping; struct file *fpin = NULL; pgoff_t offset = vmf->pgoff; + unsigned int mmap_miss; /* If we don't want any read-ahead, don't bother */ if (vmf->vma->vm_flags & VM_RAND_READ) @@ -2483,14 +2484,15 @@ static struct file *do_sync_mmap_readahe } /* Avoid banging the cache line if not needed */ - if (ra->mmap_miss < MMAP_LOTSAMISS * 10) - ra->mmap_miss++; + mmap_miss = READ_ONCE(ra->mmap_miss); + if (mmap_miss < MMAP_LOTSAMISS * 10) + WRITE_ONCE(ra->mmap_miss, ++mmap_miss); /* * Do we miss much more than hit in this file? If so, * stop bothering with read-ahead. It will only hurt. */ - if (ra->mmap_miss > MMAP_LOTSAMISS) + if (mmap_miss > MMAP_LOTSAMISS) return fpin; /* @@ -2516,13 +2518,15 @@ static struct file *do_async_mmap_readah struct file_ra_state *ra = &file->f_ra; struct address_space *mapping = file->f_mapping; struct file *fpin = NULL; + unsigned int mmap_miss; pgoff_t offset = vmf->pgoff; /* If we don't want any read-ahead, don't bother */ if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages) return fpin; - if (ra->mmap_miss > 0) - ra->mmap_miss--; + mmap_miss = READ_ONCE(ra->mmap_miss); + if (mmap_miss) + WRITE_ONCE(ra->mmap_miss, --mmap_miss); if (PageReadahead(page)) { fpin = maybe_unlock_mmap_for_io(vmf, fpin); page_cache_async_readahead(mapping, ra, file, @@ -2688,6 +2692,7 @@ void filemap_map_pages(struct vm_fault * unsigned long max_idx; XA_STATE(xas, &mapping->i_pages, start_pgoff); struct page *page; + unsigned int mmap_miss = READ_ONCE(file->f_ra.mmap_miss); rcu_read_lock(); xas_for_each(&xas, page, end_pgoff) { @@ -2724,8 +2729,8 @@ void filemap_map_pages(struct vm_fault * if (page->index >= max_idx) goto unlock; - if (file->f_ra.mmap_miss > 0) - file->f_ra.mmap_miss--; + if (mmap_miss > 0) + mmap_miss--; vmf->address += (xas.xa_index - last_pgoff) << PAGE_SHIFT; if (vmf->pte) @@ -2745,6 +2750,7 @@ next: break; } rcu_read_unlock(); + WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss); } EXPORT_SYMBOL(filemap_map_pages); _
