On Mon, Feb 25, 2019 at 08:44:35AM +0100, Claudio Jeker wrote: > On Mon, Feb 25, 2019 at 10:49:16AM +1000, David Gwynne wrote: > > the mcl2k2 pool, aka the intel mbuf cluster pool, gets set up to allocate > > at least 2048 + 2 bytes, which gets rounded up by 64 bytes to 2112 > > bytes. this diff makes ix move the reception of packets to the end of > > the 2112 byte allocation so there's space left at the front of the mbuf. > > > > this in turn makes it more likely that an m_prepend at another point in > > the system will work without an extra mbuf allocation. eg, if you're > > bridging or routing between vlans and vlans on svlans somewhere else, > > this will be a bit faster with this diff. > > > > thoughts? ok? > > I think using m_align() here may be benefitial. Since it does exactly > that. Apart from that I have to agree, shifting the packet back makes a > lot of sense.
Like this? Index: if_ix.c =================================================================== RCS file: /cvs/src/sys/dev/pci/if_ix.c,v retrieving revision 1.153 diff -u -p -r1.153 if_ix.c --- if_ix.c 21 Feb 2019 03:16:47 -0000 1.153 +++ if_ix.c 25 Feb 2019 10:06:59 -0000 @@ -2389,8 +2395,8 @@ ixgbe_get_buf(struct rx_ring *rxr, int i if (!mp) return (ENOBUFS); + m_align(mp, sc->rx_mbuf_sz); mp->m_len = mp->m_pkthdr.len = sc->rx_mbuf_sz; - m_adj(mp, ETHER_ALIGN); error = bus_dmamap_load_mbuf(rxr->rxdma.dma_tag, rxbuf->map, mp, BUS_DMA_NOWAIT);