Skip to contentSkip to conversation

btrfs: use mount idmap for defrag permission check

PATCH In review Tao Cui 4 participants 6 messages Aug 13 – Aug 14

Subsystems (2)
lore
Share & tools
lore

Tao Cui

author
#0 lore raw

PATCH 1/1 btrfs: use mount idmap for defrag permission check

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

---
 fs/btrfs/ioctl.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/btrfs/ioctl.c +1 -1
index 72bc9d4f7708..607329dedd50 100644
@@ -2458,7 +2458,7 @@ static int btrfs_ioctl_defrag(struct file *file, void __user *argp)
 		 * running and allows defrag on files open in read-only mode.
 		 */
 		if (!capable(CAP_SYS_ADMIN) &&
-		    inode_permission(&nop_mnt_idmap, inode, MAY_WRITE)) {
+		    inode_permission(file_mnt_idmap(file), inode, MAY_WRITE)) {
 			ret = -EPERM;
 			goto out;
 		}
Signature · 1 line
2.43.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 8 lines · Tao Cui

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Quoted 7 lines

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

Fixes 4609e1f18e19 ("fs: port ->permission() to pass mnt_idmap")

Otherwise looks good to me.

Quoted 17 lines
---
  fs/btrfs/ioctl.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/btrfs/ioctl.c +1 -1
index 72bc9d4f7708..607329dedd50 100644
@@ -2458,7 +2458,7 @@ static int btrfs_ioctl_defrag(struct file *file, void __user *argp)
  		 * running and allows defrag on files open in read-only mode.
  		 */
  		if (!capable(CAP_SYS_ADMIN) &&
-		    inode_permission(&nop_mnt_idmap, inode, MAY_WRITE)) {
+		    inode_permission(file_mnt_idmap(file), inode, MAY_WRITE)) {
  			ret = -EPERM;
  			goto out;
  		}
#2 lore raw
Quoted 17 lines · Qu Wenruo
Quoted 14 lines · Tao Cui

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

Quoted 3 lines
Quoted 1 line
Fixes 4609e1f18e19 ("fs: port ->permission() to pass mnt_idmap")

This is a misattribution. Using nop_mnt_idmap there keeps the behavior the same as it was before the commit. Switching to the mount idmap opens up BTRFS_IOC_DEFRAG* to idmapped users when they weren't before, which is a separate policy decision that didn't belong in that change.

Saying that these ioctls should be allowed for idmapped users just because others are isn't a good justification. Why are these ioctls safe for idmapped users? Do they actually require this capability?

Thanks,
Seth

#3 lore raw
Quoted 18 lines · Seth Forshee
Quoted 17 lines · Qu Wenruo
Quoted 14 lines · Tao Cui

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

Quoted 8 lines
Quoted 3 lines
Quoted 1 line
Fixes 4609e1f18e19 ("fs: port ->permission() to pass mnt_idmap")

This is a misattribution. Using nop_mnt_idmap there keeps the behavior the same as it was before the commit. Switching to the mount idmap opens up BTRFS_IOC_DEFRAG* to idmapped users when they weren't before, which is a separate policy decision that didn't belong in that change.

OK, removed from for-next, and will not add the fixes tag.

Quoted 4 lines

Saying that these ioctls should be allowed for idmapped users just because others are isn't a good justification. Why are these ioctls safe for idmapped users? Do they actually require this capability?

I'll let the author to do the explanation.

Thanks,
Qu

Quoted 4 lines

Thanks,
Seth

author
#4 lore raw
In reply to Seth Forshee
Quoted 18 lines · Seth Forshee
Quoted 17 lines · Qu Wenruo
Quoted 14 lines · Tao Cui

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

Quoted 9 lines
Quoted 3 lines
Quoted 1 line
Fixes 4609e1f18e19 ("fs: port ->permission() to pass mnt_idmap")

This is a misattribution. Using nop_mnt_idmap there keeps the behavior the same as it was before the commit. Switching to the mount idmap opens up BTRFS_IOC_DEFRAG* to idmapped users when they weren't before, which is a separate policy decision that didn't belong in that change.

Agreed, I'll resend it as a behavior change with no Fixes tag.

Quoted 4 lines

Saying that these ioctls should be allowed for idmapped users just because others are isn't a good justification. Why are these ioctls safe for idmapped users? Do they actually require this capability?

The check isn't a privilege gate. 616d374efa23 added it to test "whether the file could have been opened rw", and the caller can already do that on the idmapped mount -- open O_RDWR, write(), fallocate(). Only the regular-file branch changes, and defrag only reorganizes the caller's own extents; the S_IFDIR branch (whole-subvolume defrag) keeps its capable(CAP_SYS_ADMIN).

Nothing is weakened: !capable(CAP_SYS_ADMIN) stays exactly as is.

may_dedupe_file() in fs/remap_range.c already gates FIDEDUPERANGE the same way -- capable(CAP_SYS_ADMIN), then i_uid_into_vfsuid( file_mnt_idmap(file)) / inode_permission(idmap, MAY_WRITE) -- and dedupe is stronger, since it can rewrite one file's extents from another file's contents. Idmapped users can already dedupe on btrfs.

Thanks,
Tao

Quoted 2 lines

Thanks,
Seth

#5 lore raw
Quoted 42 lines · Tao Cui
Quoted 18 lines · Seth Forshee
Quoted 17 lines · Qu Wenruo
Quoted 14 lines · Tao Cui

From: Tao Cui <[email protected]>

btrfs_ioctl_defrag() checks MAY_WRITE with nop_mnt_idmap, which skips the mount idmap. On an idmapped mount the owner comparison then uses the caller's fsuid against the raw on-disk uid, dropping the mapping. Every other permission/owner check in btrfs ioctl uses file_mnt_idmap(file) (e.g. :1152, :1310, :1946); this one missed it.

Switch to file_mnt_idmap(file). It equals nop_mnt_idmap on a normal mount, and the check stays behind !capable(CAP_SYS_ADMIN), so only unprivileged callers on idmapped btrfs change. The RO-fd note in the comment above is about the file descriptor, not this inode check, and is unaffected.

Quoted 9 lines
Quoted 3 lines
Quoted 1 line
Fixes 4609e1f18e19 ("fs: port ->permission() to pass mnt_idmap")

This is a misattribution. Using nop_mnt_idmap there keeps the behavior the same as it was before the commit. Switching to the mount idmap opens up BTRFS_IOC_DEFRAG* to idmapped users when they weren't before, which is a separate policy decision that didn't belong in that change.

Agreed, I'll resend it as a behavior change with no Fixes tag.

Quoted 4 lines

Saying that these ioctls should be allowed for idmapped users just because others are isn't a good justification. Why are these ioctls safe for idmapped users? Do they actually require this capability?

The check isn't a privilege gate. 616d374efa23 added it to test "whether the file could have been opened rw", and the caller can already do that on the idmapped mount -- open O_RDWR, write(), fallocate(). Only the regular-file branch changes, and defrag only reorganizes the caller's own extents; the S_IFDIR branch (whole-subvolume defrag) keeps its capable(CAP_SYS_ADMIN).

Quoted 7 lines

Nothing is weakened: !capable(CAP_SYS_ADMIN) stays exactly as is.

may_dedupe_file() in fs/remap_range.c already gates FIDEDUPERANGE the same way -- capable(CAP_SYS_ADMIN), then i_uid_into_vfsuid( file_mnt_idmap(file)) / inode_permission(idmap, MAY_WRITE) -- and dedupe is stronger, since it can rewrite one file's extents from another file's contents. Idmapped users can already dedupe on btrfs.

It seems perfectly reasonable to relax the check. I just think the justification in the commit message should be updated to explain why these specific ioctls should be allowed for idmapped users, similar to pervious commits allowing idmapped use of the other ioctls.

Thanks,
Seth