dm: avoid leaking the caller's thread keyring via the table device file
PATCH 1/1 dm: avoid leaking the caller's thread keyring via the table device file
The refactoring in commit a28d893eb327 ("md: port block device access to file") accidentally causes the caller's thread keyring to be kept alive long beyond the caller's lifetime.
As a result, "cryptsetup luksSuspend" silently fails to wipe the LUKS volume key from memory.
In detail: "cryptsetup luksOpen" uses its supposedly ephemeral thread keyring to pass the volume key to the kernel. dm-crypt's crypt_set_keyring_key() copies the key material into its own crypt_config structure and then drops its own reference to the key in the keyring with key_put().
With this fix, restoring pre-v6.9 behavior, the copy in the thread keyring is then promptly garbage collected, such that exactly one copy of the volume key remains. This single copy is correctly wiped from memory on "cryptsetup luksSuspend".
Without this fix, the thread keyring and the volume key in it remains. This second copy is only freed on "luksClose". "luksSuspend" neither knows about this copy nor has any way to remove it, so the key remains recoverable from RAM after a suspend that is documented to have wiped it.
This fix should not introduce new security problems, as the code is anyway gated by CAP_SYS_ADMIN. The device-mapper core, not the calling task, is the legitimate owner of this long-lived file.
---
This is only my second kernel patch; I don't have a complete grasp of the potential long-range interactions: This changes the credentials used to open the backing device for *all* table devices, not just dm-crypt's. I roughly followed the similar pattern in drivers/block/nbd.c, where scoped_with_kernel_creds is used for dealing with the socket. Please exercise extra caution. Happy to revise.
drivers/md/dm.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)
drivers/md/dm.c +10 -1
@@ -735,7 +735,16 @@ static struct table_device *open_table_device(struct mapped_device *md,return ERR_PTR (-ENOMEM );refcount_set (&td->count,1 ); - bdev_file =bdev_file_open_by_dev (dev, mode, _dm_claim_ptr,NULL ); +/* +* Open the backing device with kernel rather than caller +* credentials. Otherwise the caller's credentials would be +* pinned in bdev_file->f_cred until the table device is closed. +* That would keep the caller's thread keyring alive long beyond the +* lifetime of the caller, breaking userspace expectation (e.g. +* cryptsetup(8) leaking the LUKS volume key). +*/ +scoped_with_kernel_creds () + bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);if (IS_ERR (bdev_file)) { r =PTR_ERR (bdev_file); goto out_free_td;
Signature · 1 line
2.43.0
Hi,
Quoted 3 lines · Ingo Blechschmidt
The refactoring in commit
a28d893eb327("md: port block device access to file") accidentally causes the caller's thread keyring to be kept alive long beyond the caller's lifetime.
For the record for anyone tracking the issue interdependently.
I have tested the patch and can ack it fixes the issue for block devices opened in DM table constructors. But that's not, unfortunately, the root cause of the issue. The keys put in caller's thread keyring can be pinned via many kernel APIs taking their own reference on current_cred() when a process calls into them. That's something I missed when I added the feature in libcryptsetup.
For example: while I was testing the patch I got hit by the issue again when I had test LUKS2 image in a file. The cryptsetup creates loop device in before it opens dm-crypt mapping over it. And voila the thread keyring got pinned again, via loop device.
So to fix it properly, I have opened MR also on the cryptsetup to manually release the keyring while releasing a crypt device context: https://gitlab.com/cryptsetup/cryptsetup/-/merge_requests/937