GitLab

This patch is a fix specific to the 3.19 - 4.4 kernels. The 4.5 kernel
inadvertently fixed this bug differently (db3cbfff), but is not
a stable candidate due it being a complicated re-write of the entire
feature.

This patch fixes a potential timing bug with nvme's asynchronous queue
deletion, which causes an allocated request to be accidentally released
due to the ordering of the shared completion context among the sq/cq
pair. The completion context saves the request that issued the queue
deletion. If the submission side deletion happens to reset the active
request, the completion side will release the wrong request tag back into
the pool of available tags. This means the driver will create multiple
commands with the same tag, corrupting the queue context.

The error is observable in the kernel logs like:

  "nvme XX:YY:ZZ completed id XX twice on qid:0"

In this particular case, this message occurs because the queue is
corrupted.

The following timing sequence demonstrates the error:

  CPU A                                 CPU B
  -----------------------               -----------------------------
  nvme_irq
   nvme_process_cq
    async_completion
     queue_kthread_work  ----------->   nvme_del_sq_work_handler
                                         nvme_delete_cq
                                          adapter_async_del_queue
                                           nvme_submit_admin_async_cmd
                                            cmdinfo->req = req;

     blk_mq_free_request(cmdinfo->req); <-- wrong request!!!

This patch fixes the bug by releasing the request in the completion side
prior to waking the submission thread, such that that thread can't muck
with the shared completion context.

Fixes: a4aea562

 ("NVMe: Convert to blk-mq")
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>