On Tue, Oct 11, 2022 at 11:38:37AM -0700, Tadeusz Struk wrote: > On 9/30/22 13:25, Tadeusz Struk wrote: > > Syzbot reported an issue with ext4 extents. The reproducer creates > > a corrupted ext4 fs image in memory, and mounts it as a loop device. > > It invokes the ext4_cache_extents() and ext4_find_extent(), which > > eventually triggers a BUG() in ext4_es_end() causing a kernel crash. > > It triggers on mainline, and every kernel version back to v4.14. > > Add a call ext4_ext_check_inode() in ext4_find_extent() to prevent > > the crash. > > > > To: "Theodore Ts'o"<tytso@xxxxxxx> > > Cc: "Andreas Dilger"<adilger.kernel@xxxxxxxxx> > > Cc:<linux-ext4@xxxxxxxxxxxxxxx> > > Cc:<linux-kernel@xxxxxxxxxxxxxxx> > > Cc:<stable@xxxxxxxxxxxxxxx> > > > > Link:https://syzkaller.appspot.com/bug?id=641e7a4b900015c5d7a729d6cc1fba7a928a88f9 > > Reported-by:syzbot+a22dc4b0744ac658ed9b@xxxxxxxxxxxxxxxxxxxxxxxxx > > Signed-off-by: Tadeusz Struk<tadeusz.struk@xxxxxxxxxx> > > Any comments/feedback on this one? Hi, I can't replicate the problem using the syzkaller repro: root@kvm-xfstests:/vdc# /vtmp/repro [ 14.840406] loop0: detected capacity change from 0 to 1051 [ 14.840965] EXT4-fs (loop0): ext4_check_descriptors: Checksum for group 0 failed (60935!=0) [ 14.841468] EXT4-fs error (device loop0): ext4_ext_check_inode:520: inode #2: comm repro: pblk 0 bad header/extent: invalid magic - magic 9fe4, entries 42, max 0(0), depth 0(0) [ 14.842188] EXT4-fs (loop0): get root inode failed [ 14.842401] EXT4-fs (loop0): mount failed And by inspection, if there is a corrupted inode, it should have been caught much sooner, in ext4_iget(): } else if (!ext4_has_inline_data(inode)) { /* validate the block references in the inode */ if (!(EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) || (S_ISLNK(inode->i_mode) && !ext4_inode_is_fast_symlink(inode)))) { if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) ret = ext4_ext_check_inode(inode); else ret = ext4_ind_check_inode(inode); } } if (ret) goto bad_inode; ... and this check has been around for quite some time. It's much more efficient to check for a corrupted inode data structure at the time when the inode is read into memory, as opposed to every single time we try to lookup a logical->physical block map, in ext4_find_extent(). If you can reproduce a failure on a modern kernel, please let me know, but it looks like syzkaller had only reported it on androd-54, android-5-10, linux-4.14, and linux-4.19 in any case. Cheers, - Ted
