fs: btrfs: fix btrfs_ioctl_space_info() slot_count TOCTOU which can lead to info-leak
PATCH 1/1 fs: btrfs: fix btrfs_ioctl_space_info() slot_count TOCTOU which can lead to info-leak
From: Yochai Eisenrich <[email protected]>
btrfs_ioctl_space_info() has a TOCTOU race between two passes over the block group RAID type lists. The first pass counts entries to determine the allocation size, then the second pass fills the buffer. The groups_sem rwlock is released between passes, allowing concurrent block group removal to reduce the entry count.
When the second pass fills fewer entries than the first pass counted, copy_to_user() copies the full alloc_size bytes including trailing uninitialized kmalloc bytes to userspace.
Fix by copying only total_spaces entries (the actually-filled count from the second pass) instead of alloc_size bytes, and switch to kzalloc so any future copy size mismatch cannot leak heap data.
--- fs/btrfs/ioctl.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
fs/btrfs/ioctl.c +3 -2
@@ -2897,7 +2897,7 @@ static long btrfs_ioctl_space_info(struct btrfs_fs_info *fs_info,return -ENOMEM ; space_args.total_spaces =0 ; - dest =kmalloc (alloc_size,GFP_KERNEL ); + dest =kzalloc (alloc_size,GFP_KERNEL );if (!dest)return -ENOMEM ; dest_orig = dest;
@@ -2953,7 +2953,8 @@ static long btrfs_ioctl_space_info(struct btrfs_fs_info *fs_info, user_dest = (struct btrfs_ioctl_space_info __user *) (arg +sizeof (struct btrfs_ioctl_space_args )); -if (copy_to_user (user_dest, dest_orig, alloc_size)) +if (copy_to_user (user_dest, dest_orig, + space_args.total_spaces *sizeof (*dest_orig)))return -EFAULT ; out:
--
2.53.0
Quoted 11 lines · Yochai Eisenrich
From: Yochai Eisenrich <[email protected]>
btrfs_ioctl_space_info()has a TOCTOU race between two passes over the block group RAID type lists. The first pass counts entries to determine the allocation size, then the second pass fills the buffer. Thegroups_semrwlock is released between passes, allowing concurrent block group removal to reduce the entry count.When the second pass fills fewer entries than the first pass counted,
copy_to_user()copies the fullalloc_sizebytes including trailing uninitialized kmalloc bytes to userspace.
This sounds correct.
Quoted 3 lines
Fix by copying only
total_spacesentries (the actually-filled count from the second pass) instead ofalloc_sizebytes, and switch to kzalloc so any future copy size mismatch cannot leak heap data.
Trying to hit this race looks very hard though, reducing number of block group types is quite rare.
The change to kzalloc looks like best fix, for all ioctls that are exposed to userspace. Copying the exact number makes sense. The other case (copying too much) has been fixed in 51788b1bdd0d68 ("btrfs:
prevent heap corruption in btrfs_ioctl_space_info()").
Quoted 4 lines
Fixes 7fde62bffb57 ("Btrfs: buffer results in the space_info ioctl")Signed-off-by Yochai Eisenrich <[email protected]>
Added to for-next, thanks.
Quoted 14 lines · David Sterba
Quoted 11 lines · Yochai Eisenrich
From: Yochai Eisenrich <[email protected]>
btrfs_ioctl_space_info()has a TOCTOU race between two passes over the block group RAID type lists. The first pass counts entries to determine the allocation size, then the second pass fills the buffer. Thegroups_semrwlock is released between passes, allowing concurrent block group removal to reduce the entry count.When the second pass fills fewer entries than the first pass counted,
copy_to_user()copies the fullalloc_sizebytes including trailing uninitialized kmalloc bytes to userspace.
Quoted 8 lines
This sounds correct.
Quoted 3 lines
Fix by copying only
total_spacesentries (the actually-filled count from the second pass) instead ofalloc_sizebytes, and switch to kzalloc so any future copy size mismatch cannot leak heap data.Trying to hit this race looks very hard though, reducing number of block group types is quite rare.
I agree that this may not be your btrfs typical behavior, but I wouldn't have raised the issue if I wasn't able to prove it. I can send the PoC code your way if you're interested - it leaks kernel data. A malicious user can utilize a fresh btrfs disk over e.g. zram to trigger the issue.
Quoted 4 lines
The change to kzalloc looks like best fix, for all ioctls that are exposed to userspace. Copying the exact number makes sense. The other case (copying too much) has been fixed in
51788b1bdd0d68("btrfs:
prevent heap corruption inbtrfs_ioctl_space_info()").
Just to make sure we're on the same page:
1. Following the above, do you approve of the copy_to_user fix I suggested?
2. I think it makes sense to treat other ioctl kmallocs in a different patch, no?
Quoted 6 lines
Quoted 4 lines
Fixes 7fde62bffb57 ("Btrfs: buffer results in the space_info ioctl")Signed-off-by Yochai Eisenrich <[email protected]>Added to for-next, thanks.
Quoted 21 lines · Yochai E
Quoted 20 lines · David Sterba
Quoted 11 lines · Yochai Eisenrich
From: Yochai Eisenrich <[email protected]>
btrfs_ioctl_space_info()has a TOCTOU race between two passes over the block group RAID type lists. The first pass counts entries to determine the allocation size, then the second pass fills the buffer. Thegroups_semrwlock is released between passes, allowing concurrent block group removal to reduce the entry count.When the second pass fills fewer entries than the first pass counted,
copy_to_user()copies the fullalloc_sizebytes including trailing uninitialized kmalloc bytes to userspace.This sounds correct.
Quoted 3 lines
Fix by copying only
total_spacesentries (the actually-filled count from the second pass) instead ofalloc_sizebytes, and switch to kzalloc so any future copy size mismatch cannot leak heap data.
Quoted 7 lines
Quoted 2 lines
Trying to hit this race looks very hard though, reducing number of block group types is quite rare.
I agree that this may not be your btrfs typical behavior, but I wouldn't have raised the issue if I wasn't able to prove it. I can send the PoC code your way if you're interested - it leaks kernel data. A malicious user can utilize a fresh btrfs disk over e.g. zram to trigger the issue.
I believe you, the pattern for leaking data works, I'd apply the fix even without a PoC. Using zeroing allocations in ioctls should be used by default, this one was missing it.
For a user to trigger that I don't see how this can be well timed with just passively check the space info ioctl ("btrfs fi df") when the administrator runs the rebalance command and the block group number decreases. Like on an installed system it takes the defaults and stays like that forever.
Active user would have to be able to mkfs an image (trivial) and have it mounted (possible), which on itself opens a lot of other options. FS_USERNS_MOUNT is usually not enabled for that reason.
Quoted 8 lines
Quoted 4 lines
The change to kzalloc looks like best fix, for all ioctls that are exposed to userspace. Copying the exact number makes sense. The other case (copying too much) has been fixed in
51788b1bdd0d68("btrfs:
prevent heap corruption inbtrfs_ioctl_space_info()").Just to make sure we're on the same page:
1. Following the above, do you approve of thecopy_to_userfix I suggested?
Yeah it's ok. For normal case there's no difference, the user side will read up to args.total_spaces which matches the valid data, and ignore the empty slot. Attacker will find zeros there.
Quoted 2 lines
2. I think it makes sense to treat other ioctl kmallocs in a different patch, no?
Yes, one thing per patch. I was mentioning that as further work as it's good to look for the similar problems in other code.
Re: [PATCH] fs: btrfs: fix btrfs_ioctl_space_info() slot_count
In reply to Yochai EisenrichApplied and tested on v7.0-rc4 in a QEMU VM with btrfs as the root filesystem.
Verified that BTRFS_IOC_SPACE_INFO (via btrfs fi usage) returns correct results under normal operation and while running concurrent filesystem activity (file creation/deletion in a loop).
No kernel warnings or errors observed in dmesg.
It's hard to create a race condition between 2 passes so didn't test whether before-patch behaviour.
Re: [PATCH] fs: btrfs: fix btrfs_ioctl_space_info() slot_count
Quoted 8 lines · Teng Liu
Applied and tested on v7.0-rc4 in a QEMU VM with btrfs as the root filesystem.
Verified that
BTRFS_IOC_SPACE_INFO(via btrfs fi usage) returns correct results under normal operation and while running concurrent filesystem activity (file creation/deletion in a loop).
Quoted 6 lines
No kernel warnings or errors observed in dmesg.
It's hard to create a race condition between 2 passes so didn't test whether before-patch behaviour.
Tested-by Teng Liu <[email protected]>
If it's relevant, I can provide a reproducer.