From: Catherine Sullivan csully@google.com
[ Upstream commit f81781835f0adfae8d701545386030d223efcd6f ]
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: f5cedc84a30d ("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 | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c index 77e79d2939ba..6ea0975d74a1 100644 --- a/drivers/net/ethernet/google/gve/gve_main.c +++ b/drivers/net/ethernet/google/gve/gve_main.c @@ -121,7 +121,7 @@ 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)