On Mon 29-05-23 16:01:48, Baokun Li wrote: > Syzbot reported a BUG_ON: > ================================================================== > EXT4-fs (loop0): mounted filesystem without journal. Quota mode: none. > EXT4-fs error (device loop0): ext4_mb_generate_buddy:1098: group 0, block > bitmap and bg descriptor inconsistent: 25 vs 150994969 free clusters > ------------[ cut here ]------------ > kernel BUG at fs/ext4/ext4_jbd2.c:53! > invalid opcode: 0000 [#1] PREEMPT SMP KASAN > CPU: 1 PID: 494 Comm: syz-executor.0 6.1.0-rc7-syzkaller-ga4412fdd49dc #0 > RIP: 0010:__ext4_journal_stop+0x1b3/0x1c0 > [...] > Call Trace: > ext4_write_inline_data_end+0xa39/0xdf0 > ext4_da_write_end+0x1e2/0x950 > generic_perform_write+0x401/0x5f0 > ext4_buffered_write_iter+0x35f/0x640 > ext4_file_write_iter+0x198/0x1cd0 > vfs_write+0x8b5/0xef0 > [...] > ================================================================== > > The above BUG_ON is triggered by the following race: > > cpu1 cpu2 > ________________________|________________________ > ksys_write > vfs_write > new_sync_write > ext4_file_write_iter > ext4_buffered_write_iter > generic_perform_write > ext4_da_write_begin > do_fault > do_page_mkwrite > ext4_page_mkwrite > ext4_convert_inline_data > ext4_convert_inline_data_nolock > ext4_destroy_inline_data_nolock > //clear EXT4_STATE_MAY_INLINE_DATA > ext4_map_blocks --> return error > ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) > ext4_block_write_begin > ext4_restore_inline_data > // set EXT4_STATE_MAY_INLINE_DATA > ext4_da_write_end > ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) > ext4_write_inline_data_end > handle=NULL > ext4_journal_stop(handle) > __ext4_journal_stop > ext4_put_nojournal(handle) > ref_cnt = (unsigned long)handle > BUG_ON(ref_cnt == 0) ---> BUG_ON > > The root cause of this problem is that the ext4_convert_inline_data() in > ext4_page_mkwrite() does not grab i_rwsem, so it may race with > ext4_buffered_write_iter() and cause the write_begin() and write_end() > functions to be inconsistent and trigger BUG_ON. > > To solve the above issue, we cannot add inode_lock directly to > ext4_page_mkwrite(), because this function is a hot path and frequent calls > to inode_lock will cause performance degradation for multi-threaded reads > and writes. Hence, we move ext4_convert_inline_data() to ext4_file_mmap(), > and only when inline_data is enabled and mmap a file in shared write mode, > we hold the lock to convert, which can reduce the impact on performance. > > Reported-by: Jun Nie <jun.nie@xxxxxxxxxx> > Closes: https://lore.kernel.org/lkml/63903521.5040307@xxxxxxxxxx/t/ > Reported-by: syzbot+a158d886ca08a3fecca4@xxxxxxxxxxxxxxxxxxxxxxxxx > Closes: https://syzkaller.appspot.com/bug?id=899b37f20ce4072bcdfecfe1647b39602e956e36 > Fixes: 7b4cc9787fe3 ("ext4: evict inline data when writing to memory map") > CC: stable@xxxxxxxxxxxxxxx # 4.12+ > Signed-off-by: Baokun Li <libaokun1@xxxxxxxxxx> Thanks for the patch! The problem with i_rwsem in ext4_page_mkwrite() is not so much about performance as about lock ordering. In ext4_page_mkwrite() we are called with mmap_sem held and so we cannot acquire i_rwsem because it ranks about it. > --- > fs/ext4/file.c | 24 +++++++++++++++++++++++- > fs/ext4/inode.c | 4 ---- > 2 files changed, 23 insertions(+), 5 deletions(-) > > diff --git a/fs/ext4/file.c b/fs/ext4/file.c > index d101b3b0c7da..7a04376c33f2 100644 > --- a/fs/ext4/file.c > +++ b/fs/ext4/file.c > @@ -795,7 +795,8 @@ static const struct vm_operations_struct ext4_file_vm_ops = { > static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma) > { > struct inode *inode = file->f_mapping->host; > - struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); > + struct super_block *sb = inode->i_sb; > + struct ext4_sb_info *sbi = EXT4_SB(sb); > struct dax_device *dax_dev = sbi->s_daxdev; > > if (unlikely(ext4_forced_shutdown(sbi))) > @@ -808,6 +809,27 @@ static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma) > if (!daxdev_mapping_supported(vma, dax_dev)) > return -EOPNOTSUPP; > > + /* > + * Writing via mmap has no logic to handle inline data, so we > + * need to call ext4_convert_inline_data() to convert the inode > + * to normal format before doing so, otherwise a BUG_ON will be > + * triggered in ext4_writepages() due to the > + * EXT4_STATE_MAY_INLINE_DATA flag. Moreover, we need to grab > + * i_rwsem during conversion, since clearing and setting the > + * inline data flag may race with ext4_buffered_write_iter() > + * to trigger a BUG_ON. > + */ > + if (ext4_has_feature_inline_data(sb) && > + vma->vm_flags & VM_SHARED && vma->vm_flags & VM_WRITE) { Sadly this does not work because we can mmap(2) the file read-only and then use mprotect(2) to make file writeable. But we can test for VM_MAYWRITE which gets set when mapping can be made writeable (basically anytime when the file descriptor itself is writeable). Otherwise the patch looks good. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
