https://bugzilla.kernel.org/show_bug.cgi?id=218820 --- Comment #1 from Theodore Tso (tytso@xxxxxxx) --- On Wed, May 08, 2024 at 01:33:56PM +0000, bugzilla-daemon@xxxxxxxxxx wrote: > > The following is the triggering script: > ``` > dd if=/dev/zero of=ext4-0.img bs=1M count=120 > mkfs.ext4 ext4-0.img > g++ -static reproduce.c > losetup /dev/loop0 ext4-0.img > mkdir /root/mnt > ./a.out > stat /root/mnt/a > ``` > > After run the script, you will get the following outputs: > ``` > File: /root/mnt/a > Size: 0 Blocks: 82 IO Block: 1024 regular empty > file > Device: 700h/1792d Inode: 12 Links: 1 > Access: (0755/-rwxr-xr-x) Uid: ( 0/ root) Gid: ( 0/ root) > Context: system_u:object_r:unlabeled_t:s0 > Access: 2024-05-08 11:47:48.000000000 +0000 > Modify: 2024-05-08 11:47:48.000000000 +0000 > Change: 2024-05-08 11:47:48.000000000 +0000 > Birth: - > ``` Thanks for the bug report. What the reproducer script is doing is opening a file for writing, and then remounting the file system to disable delayhed allocations. It then writes 40k to the file, and then truncates the file, and close it. The reproducer script leaves the file system mounted; if you unmount the file system, the kernel will issue a warning message: EXT4-fs (loop0): Inode 13 (0000000082f8ff6c): i_reserved_data_blocks (40) not cleared! ... and then if you examine the on-disk image, you'll see that the i_blocks field is correct. The root cause is that when the file is open, the address_operations which is instaniated is one which is designed for delayed allocation. So when 40k is written to the file, although we haven't allocated the space yet, we *will* allocate the space, so we update in-memory i_blocks to reflect the to-be-allocated disk space (although we don't update the on-disk i_blocks until the allocation actually takes place, so if we crash, the file system stays consistent). However, some of the *other* logic is done based on whether the delayed allocation flag is set in the struct super (as opposed to the address_operations in the struct inode at the time that file is opened). This includes what happens when we truncate the file, where in the nodelalloc case, the in-memory and on-disk i_blocks are in sync, and when the blocks are delallocated, the i_blocks field is dropped. But since the blocks weren't actually allocated because the file descriptor was in delalloc mode, there was nothing to delallocate, so the in-memory i_blocks stayed elevanted. There are a number ways of fixing this. The simplest is to simply not allow the delayed allocation mode to be changed on a remount. Since in the long term, once we fix some performance for some specialized use case which has caused people to want to disable delayed allocation, this is the one that probably makes the most sense. The two other approaches involve a lot more complexity. The first alternate approach is to iterate over all open files, and resolve any pending delayed allocations, and then update the address_operations to use ext4_aops instead of ext4_da_aops, while avoiding the races involving writes happening while we are trying to do the remount. Yelch. The second alternate approach is to treat the delayed allocation status on a per-inode basis, with a per-inode flag indicating whether delayed allocation is active, so that the truncate logic stays consistent with the ext4_da_ops active on file descriptors associated with the file. But this gets super messy since subsequent file opens on the inode also need to use the same delayed allocation mode until the last file descriptor associated with the inode is closed. Double Yelch. So preventing the "mount -o remount,nodelalloc" in the reproducer program is how we'll address this issue. - Ted -- You may reply to this email to add a comment. You are receiving this mail because: You are watching the assignee of the bug.
