nvme-pci: dma read memory barrier for completions

Control dependencies do not guarantee load order across the condition,
allowing a CPU to predict and speculate memory reads.

Commit 324b494c28 inlined verifying a new completion with its
handling. At least one architecture was observed to access the contents
out of order, resulting in the driver using stale data for the
completion.

Add a dma read barrier before reading the completion queue entry and
after the condition its contents depend on to ensure the read order is
determinsitic.

Reported-by: John Garry <john.garry@huawei.com>
Suggested-by: Will Deacon <will@kernel.org>
Signed-off-by: Keith Busch <kbusch@kernel.org>
Tested-by: John Garry <john.garry@huawei.com>
Acked-by: Will Deacon <will@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

drivers/nvme/host/pci.c

@@ -989,6 +989,11 @@ static inline int nvme_process_cq(struct nvme_queue *nvmeq)
 
 	while (nvme_cqe_pending(nvmeq)) {
 		found++;
+		/*
+		 * load-load control dependency between phase and the rest of
+		 * the cqe requires a full read memory barrier
+		 */
+		dma_rmb();
 		nvme_handle_cqe(nvmeq, nvmeq->cq_head);
 		nvme_update_cq_head(nvmeq);
 	}