Hi,
I notice a regression report on Bugzilla [1]. Quoting from it:
Since commit a44be64, remounting a read-only ext4 filesystem to become read-write fails when quotas are enabled. The mount syscall returns -EROFS and outputs the following in dmesg:
EXT4-fs warning (device loop0): ext4_enable_quotas:7028: Failed to enable quota tracking (type=0, err=-30, ino=3). Please run e2fsckRoot cause
The problem can be traced back to the changes introduced in commit a44be64. It appears that the issue arises because the SB_RDONLY bit of the s_flags field is now only cleared after executing the ext4_enable_quotas function. However, the vfs_setup_quota_inode function, called by ext4_enable_quotas, checks whether this bit is set (fs/quota/dquot.c:2331):
if (IS_RDONLY(inode)) return -EROFS;This condition therefore always triggers the -EROFS fail condition.
Steps to Reproduce
The bug can be reproduced by executing the following script on a current mainline kernel with defconfig:
#!/bin/bash set -ex truncate -s 1G /tmp/img mkfs.ext4 /tmp/img tune2fs -Q usrquota,grpquota,prjquota /tmp/img losetup /dev/loop0 /tmp/img mount -o ro /dev/loop0 /mnt mount -o remount,rw /mntExecuting the script results in the following output:
+ truncate -s 1G /tmp/img + mkfs.ext4 /tmp/img mke2fs 1.47.0 (5-Feb-2023) Discarding device blocks: done Creating filesystem with 262144 4k blocks and 65536 inodes Filesystem UUID: b96a3da2-043f-11ee-b6f0-47c69db05231 Superblock backups stored on blocks: 32768, 98304, 163840, 229376 Allocating group tables: done Writing inode tables: done Creating journal (8192 blocks): done Writing superblocks and filesystem accounting information: done + tune2fs -Q usrquota,grpquota,prjquota /tmp/img tune2fs 1.47.0 (5-Feb-2023) + losetup /dev/loop0 /tmp/img [ 6.766763] loop0: detected capacity change from 0 to 2097152 + mount -o ro /dev/loop0 /mnt [ 6.791561] EXT4-fs (loop0): mounted filesystem b96a3da2-043f-11ee-b6f0-47c69db05231 ro with ordered data mode. Quota mode: journalled. + mount -o remount,rw /mnt [ 6.805546] EXT4-fs warning (device loop0): ext4_enable_quotas:7028: Failed to enable quota tracking (type=0, err=-30, ino=3). Please run e2fsck to fix. mount: /mnt: cannot remount /dev/loop0 read-write, is write-protected. dmesg(1) may have more information after failed mount system call.
See Bugzilla for the full thread.
Ted, it looks like this regression is caused by your ext4_xattr_block_set() fix to earlier syzbot report. Would you like to take a look on it?
Anyway, I'm adding it to regzbot:
#regzbot introduced: a44be64bbecb15 https://bugzilla.kernel.org/show_bug.cgi?id=217529 #regzbot title: Remounting ext4 filesystem from ro to rw fails when quotas are enabled
Thanks.
[1]: https://bugzilla.kernel.org/show_bug.cgi?id=217529
On Wed, Jun 07, 2023 at 12:51:26PM +0700, Bagas Sanjaya wrote:
I notice a regression report on Bugzilla [1]. Quoting from it:
Since commit a44be64, remounting a read-only ext4 filesystem to become read-write fails when quotas are enabled. The mount syscall returns -EROFS and outputs the following in dmesg:
Yeah, and I think all we can do is revert this commit:
ext4: don't clear SB_RDONLY when remounting r/w until quota is re-enabled
I think I saw some messages go by about this getting queued for the stable kernel; if so, could it please be dropped?
The problem can be traced back to the changes introduced in commit a44be64. It appears that the issue arises because the SB_RDONLY bit of the s_flags field is now only cleared after executing the ext4_enable_quotas function. However, the vfs_setup_quota_inode function, called by ext4_enable_quotas, checks whether this bit is set (fs/quota/dquot.c:2331):
The problem that we're trying to solve is the malicious syzbot reproducer is in one thread, twiddling the file system state from r/o to rw and back. In another thread, it's attempt to create files and directories. And occasionally, we're tripping this warning:
WARN_ON_ONCE(dquot_initialize_needed(inode));
That's because we're racing with the quota getting initialized, and the moment we clear the SB_RDONLY bit the thread which is trying to create a directory or file will proceed with the operation --- even though the quota subsystem hasn't been initialized yet. That's why the patch attempted to move the clearing the SB_RDONLY bit ahead of reiniitalization of the quota subsystem.
Since this is screwing up the ability to remount the file system rw, we need to revert this commit, at which point, we'll be able to trigger this warning again.
So how do we fix the warning? Well, we could just drop the WARN_ON_ONCE(); the downside is that when this race gets hit, the quota operations to allocate the block and inode will silently become a no-op, which means the quota will get out of sync with reality.
Alternatively, we could add a call to the beginning to ext4_xattr_block_Set():
if (dquot_initialize_needed(inode)) reutrn -EROFS;
... on the theory that the only time we should hit this is when there is a quota setup racing with inode creation, and it's better that we just let the mkdir or open with O_CREAT fail than to succeed, and allocate blocks before the quota subsystem has been initialized. I'm not sure how safe this would be on older quota setups (pre-ext4 quota feature), since I suspect the race window is a quite a bit bigger if I understand correctly how things worked with the legacy quota support.
The final really hacky thing I could imagine is to hack dquot_initialize() to something like this:
int dquot_initialize(struct inode *inode) { ret = __dquot_initialize(inode, -1); if (ret) return ret; if (dquot_initialize_needed(inode)) { msleep(1000) return __dquot_initialize(inode, -1); } return 0; }
But I threw up a little in my mouth while writing it....
So I'm tempted to just remove the WARN_ON's, and just accept that if superuser wants to twiddle the read/only state of a file system with quota, at high rates, while something else is trying to create files/directories, most of which will fail while the file system is read-only, then the quota may gets out of sync, and... ¯_(ツ)_/¯
Since this is only something that crazy people or malicious Syzbot reproducers would try to do, I'm really having a hard time bringing myself to care. Especially since we have plenty of other places where we aren't doing the dquot_initialize_needed() check, so the opportunities for the quota to get out of sync already exist in other code paths.
Jan, what do you think?
- Ted
On Thu 08-06-23 00:40:56, Theodore Ts'o wrote:
On Wed, Jun 07, 2023 at 12:51:26PM +0700, Bagas Sanjaya wrote:
I notice a regression report on Bugzilla [1]. Quoting from it:
Since commit a44be64, remounting a read-only ext4 filesystem to become read-write fails when quotas are enabled. The mount syscall returns -EROFS and outputs the following in dmesg:
Yeah, and I think all we can do is revert this commit:
ext4: don't clear SB_RDONLY when remounting r/w until quota is re-enabledI think I saw some messages go by about this getting queued for the stable kernel; if so, could it please be dropped?
Yeah, for now I'd revert the fix. It needs more thought and it is not like it is fixing any serious vulnerability or so.
The problem can be traced back to the changes introduced in commit a44be64. It appears that the issue arises because the SB_RDONLY bit of the s_flags field is now only cleared after executing the ext4_enable_quotas function. However, the vfs_setup_quota_inode function, called by ext4_enable_quotas, checks whether this bit is set (fs/quota/dquot.c:2331):
The problem that we're trying to solve is the malicious syzbot reproducer is in one thread, twiddling the file system state from r/o to rw and back. In another thread, it's attempt to create files and directories. And occasionally, we're tripping this warning:
WARN_ON_ONCE(dquot_initialize_needed(inode));
That's because we're racing with the quota getting initialized, and the moment we clear the SB_RDONLY bit the thread which is trying to create a directory or file will proceed with the operation --- even though the quota subsystem hasn't been initialized yet. That's why the patch attempted to move the clearing the SB_RDONLY bit ahead of reiniitalization of the quota subsystem.
Since this is screwing up the ability to remount the file system rw, we need to revert this commit, at which point, we'll be able to trigger this warning again.
So how do we fix the warning? Well, we could just drop the WARN_ON_ONCE(); the downside is that when this race gets hit, the quota operations to allocate the block and inode will silently become a no-op, which means the quota will get out of sync with reality.
Yes, the warning triggering is actually a good thing because it shows us we have a race where quota can get out of sync with real space usage.
Alternatively, we could add a call to the beginning to ext4_xattr_block_Set():
if (dquot_initialize_needed(inode)) reutrn -EROFS;
... on the theory that the only time we should hit this is when there is a quota setup racing with inode creation, and it's better that we just let the mkdir or open with O_CREAT fail than to succeed, and allocate blocks before the quota subsystem has been initialized. I'm not sure how safe this would be on older quota setups (pre-ext4 quota feature), since I suspect the race window is a quite a bit bigger if I understand correctly how things worked with the legacy quota support.
No, xattr code is just a messenger here. There is a fundamental problem that once we clear SB_RDONLY in superblock flags, anybody can come up with any operation modifying the filesystem but until we are finished with dquot_resume(), this operation can escape proper quota accounting.
The final really hacky thing I could imagine is to hack dquot_initialize() to something like this:
int dquot_initialize(struct inode *inode) { ret = __dquot_initialize(inode, -1); if (ret) return ret; if (dquot_initialize_needed(inode)) { msleep(1000) return __dquot_initialize(inode, -1); } return 0; }
But I threw up a little in my mouth while writing it....
So I'm tempted to just remove the WARN_ON's, and just accept that if superuser wants to twiddle the read/only state of a file system with quota, at high rates, while something else is trying to create files/directories, most of which will fail while the file system is read-only, then the quota may gets out of sync, and... ¯_(ツ)_/¯
Since this is only something that crazy people or malicious Syzbot reproducers would try to do, I'm really having a hard time bringing myself to care. Especially since we have plenty of other places where we aren't doing the dquot_initialize_needed() check, so the opportunities for the quota to get out of sync already exist in other code paths.
Jan, what do you think?
Well, I think we should fix this properly. The problem is that currently changing read-only state is just flipping the flag but we need an intermediate state like "superblock is really RW but for userspace it is RO" so that we can prepare the filesystem for taking the writes - in fact we've already faced very similar problem with orphan replay but there we've worked around it by forbidding orphan replay on remount and playing a bit hacky games on mount.
I think I can see a way to solve this problem. I'll send patches.
Honza
linux-stable-mirror@lists.linaro.org
-
Bagas Sanjaya -
Jan Kara -
Theodore Ts'o