It may happen that the work to destroy a queue
(for example nvmet_tcp_release_queue_work()) is started while
an auth-send or auth-receive command is still completing.
nvmet_sq_destroy() will block, waiting for all the references
to the sq to be dropped, the last reference is then
dropped when nvmet_req_complete() is called.
When this happens, both nvmet_sq_destroy() and
nvmet_execute_auth_send()/_receive() will free the dhchap pointers by
calling nvmet_auth_sq_free().
Since there isn't any lock, the two threads may race against each other,
causing double frees and memory corruptions, as reported by KASAN.
Reproduced by stress blktests nvme/041 nvme/042 nvme/043
nvme nvme2: qid 0: authenticated with hash hmac(sha512) dhgroup ffdhe4096
==================================================================
BUG: KASAN: double-free in kfree+0xec/0x4b0
Call Trace:
<TASK>
kfree+0xec/0x4b0
nvmet_auth_sq_free+0xe1/0x160 [nvmet]
nvmet_execute_auth_send+0x482/0x16d0 [nvmet]
process_one_work+0x8e5/0x1510
Allocated by task 191846:
__kasan_kmalloc+0x81/0xa0
nvmet_auth_ctrl_sesskey+0xf6/0x380 [nvmet]
nvmet_auth_reply+0x119/0x990 [nvmet]
Freed by task 143270:
kfree+0xec/0x4b0
nvmet_auth_sq_free+0xe1/0x160 [nvmet]
process_one_work+0x8e5/0x1510
Fix this bug by calling nvmet_req_complete() only after freeing the
pointers, so we will prevent the race by holding the sq reference.
V2: remove redundant code
Fixes: db1312dd95 ("nvmet: implement basic In-Band Authentication")
Signed-off-by: Maurizio Lombardi <mlombard@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Keith Busch <kbusch@kernel.org>
f965b281fd