On Nov 16, Florian Maurer wrote: > On Mon, Nov 09, 2026 at 12:00:15PM +0100, Lorenzo Bianconi wrote: > > q->head and q->tail are not the same when we have refilled the queue > > with new buffers. Moreover, I think if mt76_npu_fill_rx_queue() fails > > [...] it is not correct to use q->head for dma_idx > > You're right, q->tail it is. In my tests I could not tell your version > and mine apart, so nothing speaks for head + 1. > > > Can you please check the updated patch below? > > Sorry for the delay. I tested it on the W1700K (kernel 6.18.55, mt76 > 467ae2e3) and it works: no oops, no rx stall. > > What I ran: your diff as posted, plus read-only counters that write > nothing to the descriptors. Load was 62 minutes of iperf3 from an > ath11k mesh peer at 113-259 Mbit/s, with four busy loops at nice -20 > on the W1700K for 3 s every 18 s. In that time: > > - 61.9 M frames on MT_RXQ_NPU1, 1.88 M on MT_RXQ_NPU0, of which 227781 > span two slots > - no oops, no page_pool warning > - no descriptor that changed between consume and refill > - no refill failure, no skb failure, no tx-free parse error > > Tested-by: Florian Maurer > > I found another issue of the NPU while testing, so this took me some > time as well: > In one minute, 226 multi-slot frames in a row did not end on the slot > with LAST. Slots of the rx ring go missing under this kind of load (I > suspect the NPU drops them), and counting slots by PKT_COUNT then > assembles the following frames one slot off, in your version and in > mine. I think tx-free events are lost the same way. > I will post it as a separate [BUG] report in the next days. > > Optional: in my tree I made the poll report the budget as used when > the skb allocation fails, so napi polls again instead of waiting for > an interrupt. It never triggered here. > > Please post your patch with the Tested-by. My 2/2 can be dropped. > > Thanks, > Florian ack, I will do. Thanks for testing. Regards, Lorenzo