eth: fbnic: ring the doorbell if a burst ends in a drop

fbnic_tx_map() skips the doorbell write, and the completion request, for every packet handed to it with xmit_more set, counting on the packet which ends the burst to publish them all. When that packet is dropped instead - skb_put_padto(), skb_cow_head() or a DMA mapping failure - nothing rings. The descriptors of the preceding packets stay invisible to the HW until the next transmit on that queue, which for a burst-then-idle workload may never come.

Remember the meta descriptor of the last packet left without a doorbell and flush it from the error paths. The completion request has to be set on that descriptor rather than simply writing the tail, otherwise the HW would transmit the packets but never report a head, and the ring would fill up and stall for good.

This is very similar to Joe's recent series of fixes for bnxt. Not seen in real life, reproduced under QEMU with failure injection.

---
Reproduced under the QEMU fbnic model by failing the DMA mapping of the last segment of a software-GSO burst: 21 descriptors (7 packets) stayed in the ring with head stuck at 24 and tail at 45, 10094 bytes stuck in BQL, and QEMU saw none of the 7 frames.

---
 drivers/net/ethernet/meta/fbnic/fbnic_txrx.h |  9 ++++-
 drivers/net/ethernet/meta/fbnic/fbnic_txrx.c | 42 +++++++++++++++-----
 2 files changed, 40 insertions(+), 11 deletions(-)

index e03c9d2c38dc..f5899446dcc5 100644
@@ -128,9 +128,14 @@ struct fbnic_ring {
 		/* Rx BDQs only */
 		struct page_pool *page_pool;
 
-		/* Deferred_head is used to cache the head for TWQ1 if
+		/* TWQ0 only, index of the meta descriptor of the last packet
+		 * placed in the ring without ringing the doorbell, -1 if the
+		 * doorbell is in sync with the tail.
+		 */
+		s32 deferred_meta;
+
+		/* TCQ only, used to cache the head for TWQ1 if
 		 * an attempt is made to clean TWQ1 with zero napi_budget.
-		 * We do not use it for any other ring.
 		 */
 		s32 deferred_head;
 	};

2.55.0