gve: Upgrade memory barrier in poll routine

[ Upstream commit f817818 ]

As currently written, if the driver checks for more work (via
gve_tx_poll or gve_rx_poll) before the device posts work and the
irq doorbell is not unmasked
(via iowrite32be(GVE_IRQ_ACK | GVE_IRQ_EVENT, ...)) before the device
attempts to raise an interrupt, an interrupt is lost and this could
potentially lead to the traffic being completely halted. For
example, if a tx queue has already been stopped, the driver won't get
the chance to complete work and egress will be halted.

We need a full memory barrier in the poll
routine to ensure that the irq doorbell is unmasked before the driver
checks for more work.

Fixes: f5cedc8 ("gve: Add transmit and receive support")
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: David Awogbemila <awogbemila@google.com>
Acked-by: Willem de Brujin <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/net/ethernet/google/gve/gve_main.c

@@ -180,7 +180,7 @@ static int gve_napi_poll(struct napi_struct *napi, int budget)
 	/* Double check we have no extra work.
 	 * Ensure unmask synchronizes with checking for work.
 	 */
-	dma_rmb();
+	mb();
 	if (block->tx)
 		reschedule |= gve_tx_poll(block, -1);
 	if (block->rx)