On Tue, Mar 12, 2019 at 07:09:50PM +0800, Qu Wenruo wrote: > > > On 2019/3/12 下午7:07, Nikolay Borisov wrote: > > > > > > On 12.03.19 г. 11:10 ч., Qu Wenruo wrote: > >> [BUG] > >> When reading a file from a fuzzed image, kernel can panic like: > >> BTRFS warning (device loop0): csum failed root 5 ino 270 off 0 csum 0x98f94189 expected csum 0x00000000 mirror 1 > >> assertion failed: !memcmp_extent_buffer(b, &disk_key, offsetof(struct btrfs_leaf, items[0].key), sizeof(disk_key)), file: fs/btrfs/ctree.c, line: 2544 > >> ------------[ cut here ]------------ > >> kernel BUG at fs/btrfs/ctree.h:3500! > >> invalid opcode: 0000 [#1] PREEMPT SMP NOPTI > >> RIP: 0010:btrfs_search_slot.cold.24+0x61/0x63 [btrfs] > >> Call Trace: > >> btrfs_lookup_csum+0x52/0x150 [btrfs] > >> __btrfs_lookup_bio_sums+0x209/0x640 [btrfs] > >> btrfs_submit_bio_hook+0x103/0x170 [btrfs] > >> submit_one_bio+0x59/0x80 [btrfs] > >> extent_read_full_page+0x58/0x80 [btrfs] > >> generic_file_read_iter+0x2f6/0x9d0 > >> __vfs_read+0x14d/0x1a0 > >> vfs_read+0x8d/0x140 > >> ksys_read+0x52/0xc0 > >> do_syscall_64+0x60/0x210 > >> entry_SYSCALL_64_after_hwframe+0x49/0xbe > >> > >> [CAUSE] > >> The fuzzed image has a corrupted leaf whose first key doesn't match with its parent: > >> checksum tree key (CSUM_TREE ROOT_ITEM 0) > >> node 29741056 level 1 items 14 free 107 generation 19 owner CSUM_TREE > >> fs uuid 3381d111-94a3-4ac7-8f39-611bbbdab7e6 > >> chunk uuid 9af1c3c7-2af5-488b-8553-530bd515f14c > >> ... > >> key (EXTENT_CSUM EXTENT_CSUM 79691776) block 29761536 gen 19 > >> > >> leaf 29761536 items 1 free space 1726 generation 19 owner CSUM_TREE > >> leaf 29761536 flags 0x1(WRITTEN) backref revision 1 > >> fs uuid 3381d111-94a3-4ac7-8f39-611bbbdab7e6 > >> chunk uuid 9af1c3c7-2af5-488b-8553-530bd515f14c > >> item 0 key (EXTENT_CSUM EXTENT_CSUM 8798638964736) itemoff 1751 itemsize 2244 > >> range start 8798638964736 end 8798641262592 length 2297856 > >> > >> When reading above tree block, we have extent_buffer->refs = 2 in the > >> context: > >> - initial one from __alloc_extent_buffer() > >> alloc_extent_buffer() > >> |- __alloc_extent_buffer() > >> |- atomic_set(&eb->refs, 1) > >> > >> - one being added to fs_info->buffer_radix > >> alloc_extent_buffer() > >> |- check_buffer_tree_ref() > >> |- atomic_inc(&eb->refs) > >> > >> So even we call free_extent_buffer() in read_tree_block or other similar > >> situation, we only decrease the refs by 1, it doesn't reach 0 and won't > >> be freed right now. > >> > >> The staled eb and its corrupted content will still be kept cached. > >> > >> Further more, we have several extra cases where we either don't do > >> first key check or the check is not proper for all callers: > >> - scrub > >> We just don't have first key in this context. > >> > >> - shared tree block > >> One tree block can be shared by several snapshot/subvolume trees. > >> In that case, the first key check for one subvolume doesn't apply to > >> another. > >> > >> So for above reasons, a corrupted extent buffer can sneak into the > >> buffer cache. > >> > >> [FIX] > >> Export verify_level_key() as btrfs_verify_level_key() and call it in > >> read_block_for_search() to fill the hole. > >> > >> Due to above described reasons, even we can free corrupted extent buffer > >> from cache, we still need the check in read_block_for_search(), for > >> scrub and shared tree blocks. > >> > >> Reported-by: Yoon Jungyeon <jungyeon@xxxxxxxxxx> > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202755 > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202757 > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202759 > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202761 > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202767 > >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=202769 > >> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx> > >> --- > >> changelog: > >> v2: > >> - Commit message update to show the reason why stale ebs are kept in > >> cache. > >> - Commit message update to show extra reasons why we still need the > >> check, mainly for scrub and shared tree blocks. > >> --- > >> fs/btrfs/ctree.c | 10 ++++++++++ > >> fs/btrfs/disk-io.c | 10 +++++----- > >> fs/btrfs/disk-io.h | 3 +++ > >> 3 files changed, 18 insertions(+), 5 deletions(-) > >> > >> diff --git a/fs/btrfs/ctree.c b/fs/btrfs/ctree.c > >> index 5a6c39b44c84..7672932aa5b4 100644 > >> --- a/fs/btrfs/ctree.c > >> +++ b/fs/btrfs/ctree.c > >> @@ -2401,6 +2401,16 @@ read_block_for_search(struct btrfs_root *root, struct btrfs_path *p, > >> if (tmp) { > >> /* first we do an atomic uptodate check */ > >> if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) { > >> + /* > >> + * Do extra check for first_key, eb can be stale due to > >> + * being cached, read from scrub, or have multiple > >> + * parents (shared tree blocks). > >> + */ > >> + if (btrfs_verify_level_key(fs_info, tmp, > >> + parent_level - 1, &first_key, gen)) { > >> + free_extent_buffer(tmp); > >> + return -EUCLEAN; > >> + } > > > > What about verify_parent_transid? Shouldn't it also be checked here? > > Previous btrfs_buffer_uptodate() call implies verify_parent_transid() > check, and that's why btrfs_buffer_uptodate() needs transid parameter. Agreed, though following btrfs_buffer_uptodate is not exactly straightforward as it can return 3 values depending on the atomic parameter (the 3rd one).
