nvmf/vfio-user: fix SQ race condition

The host driver should do a wmb() before it updates the SQ tail doorbell
to ensure that any writes to the SQ are guaranteed to be visible if a
doorbell update is visible (a store-release) - and indeed, this is what
the Linux NVMe driver does.

Therefore, we require a rmb() after we read the tail doorbell in order
to synchronise properly with the host driver (we need a load-acquire),
and guarantee that the updates to the SQ are visisble to us.

Signed-off-by: Andreas Economides <andreas.economides@nutanix.com>
Change-Id: I57eb1e1f0dfb1091a8f10f40f8ee0e2604d9268c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9514
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This commit is contained in:
Andreas Economides 2021-09-15 16:59:03 +00:00 committed by Tomasz Zawadzki
parent 49911c47fb
commit 93db90f7f3

View File

@ -2746,7 +2746,16 @@ nvmf_vfio_user_qpair_poll(struct nvmf_vfio_user_qpair *qpair)
ctrlr = qpair->ctrlr;
/* Load-Acquire. */
new_tail = *tdbl(ctrlr, &qpair->sq);
/*
* Ensure that changes to the queue are visible to us.
* The host driver should write the queue first, do a wmb(), and then
* update the SQ tail doorbell (their Store-Release).
*/
spdk_rmb();
if (sq_head(qpair) == new_tail) {
return 0;
}