[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openbsd-tech
Subject:    Re: mcx(4): sync only received length on RX
From:       David Gwynne <david () gwynne ! id ! au>
Date:       2021-05-31 22:20:43
Message-ID: 59087CDB-3480-4708-82E4-358DB4ACD29E () gwynne ! id ! au
[Download RAW message or body]



> On 1 Jun 2021, at 04:15, Patrick Wildt <patrick@blueri.se> wrote:
> 
> Hi,
> 
> mcx(4) seems to sync the whole mapsize on processing a received packet.
> As far as I know, we usually only sync the actual size that we have
> received.  Noticed this when doing bounce buffer tests, seeing that
> it copied a lot more data than is necessary.
> 
> That's because the RX buffer size is maximum supported MTU, which is
> about 9500 bytes or so.  For small packets, or regular 1500 bytes,
> this adds overhead.
> 
> This change should not change anything for ARM machines that have a
> cache coherent PCIe bus or x86.
> 
> ok?

ok.

> 
> Patrick
> 
> diff --git a/sys/dev/pci/if_mcx.c b/sys/dev/pci/if_mcx.c
> index 38437e54897..065855d46d3 100644
> --- a/sys/dev/pci/if_mcx.c
> +++ b/sys/dev/pci/if_mcx.c
> @@ -6800,20 +6800,20 @@ mcx_process_rx(struct mcx_softc *sc, struct mcx_rx *rx,
> {
> 	struct mcx_slot *ms;
> 	struct mbuf *m;
> -	uint32_t flags;
> +	uint32_t flags, len;
> 	int slot;
> 
> +	len = bemtoh32(&cqe->cq_byte_cnt);
> 	slot = betoh16(cqe->cq_wqe_count) % (1 << MCX_LOG_RQ_SIZE);
> 
> 	ms = &rx->rx_slots[slot];
> -	bus_dmamap_sync(sc->sc_dmat, ms->ms_map, 0, ms->ms_map->dm_mapsize,
> -	    BUS_DMASYNC_POSTREAD);
> +	bus_dmamap_sync(sc->sc_dmat, ms->ms_map, 0, len, BUS_DMASYNC_POSTREAD);
> 	bus_dmamap_unload(sc->sc_dmat, ms->ms_map);
> 
> 	m = ms->ms_m;
> 	ms->ms_m = NULL;
> 
> -	m->m_pkthdr.len = m->m_len = bemtoh32(&cqe->cq_byte_cnt);
> +	m->m_pkthdr.len = m->m_len = len;
> 
> 	if (cqe->cq_rx_hash_type) {
> 		m->m_pkthdr.ph_flowid = betoh32(cqe->cq_rx_hash);
> 

[prev in list] [next in list] [prev in thread] [next in thread]