From: Haiyang Zhang <haiyangz@microsoft.com> Sent: Wednesday, December 8, 2021 12:14 PM

quoted

From: Tianyu Lan <redacted>
Sent: Tuesday, December 7, 2021 2:56 AM

[snip]

quoted

 static inline int netvsc_send_pkt(
 	struct hv_device *device,
 	struct hv_netvsc_packet *packet,

@@ -986,14 +1105,24 @@ static inline int netvsc_send_pkt(

 	trace_nvsp_send_pkt(ndev, out_channel, rpkt);

+	packet->dma_range = NULL;
 	if (packet->page_buf_cnt) {
 		if (packet->cp_partial)
 			pb += packet->rmsg_pgcnt;

+		ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb);
+		if (ret) {
+			ret = -EAGAIN;
+			goto exit;
+		}

Returning EAGAIN will let the upper network layer busy retry,
which may make things worse.
I suggest to return ENOSPC here like another place in this
function, which will just drop the packet, and let the network
protocol/app layer decide how to recover.

Thanks,
- Haiyang

I made the original suggestion to return -EAGAIN here.   A
DMA mapping failure should occur only if swiotlb bounce
buffer space is unavailable, which is a transient condition.
The existing code already stops the queue and returns
-EAGAIN when the ring buffer is full, which is also a transient
condition.  My sense is that the two conditions should be
handled the same way.  Or is there a reason why a ring
buffer full condition should stop the queue and retry, while
a mapping failure should drop the packet?

Michael