> On 14 Oct 2016, at 16:17, Ted Unangst <[email protected]> wrote:
>
> David Gwynne wrote:
>> this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
>> can be used on some very common nics that have annoying constraints
>> on their rx descriptors.
>>
>> this in turn simplifies the code in those drivers and lets them
>> always operate on ETHER_ALIGN boundaries.
>>
>> the pool is cheap, pages will only be allocated in it if something
>> asks for them, and it keeps this complexity out of the drivers.
>
> pool is efficient, but it can't work miracles. on archs like amd64, this is
> going to only pack 15 clusters in a 32k "page" instead of 16.
it is only about 70 bytes per cluster though. i could raise the size by that to
avoid slack space if you want, or we can let the page colouring have some fun
here.
>
> if i understand correctly, currently em is using 4k clusters on strict
> alignment? i think it makes a lot of sense to add a 2k2 pool because that
> means using less memory. but not strict alignment shouldn't be forced to use
> more memory. that's a regression.
you could also argue that 2k clusters waste ~500 bytes in the most common case
since the vast majority of large packets are still only 1514 bytes. we could
pack 10 1600 byte clusters into 16k instead of the 8 we do now.
more seriously though, im happy to burn a bit more memory to keep the code
simple, and overall our memory consumption by network cards is much lower than
other platforms because of the rx ring moderation we do.
dlg
>
>
>>
>> ok?
>>
>> Index: net/if.h
>> ===================================================================
>> RCS file: /cvs/src/sys/net/if.h,v
>> retrieving revision 1.179
>> diff -u -p -r1.179 if.h
>> --- net/if.h 4 Sep 2016 15:10:59 -0000 1.179
>> +++ net/if.h 14 Oct 2016 03:46:22 -0000
>> @@ -68,7 +68,7 @@ struct if_clonereq {
>> char *ifcr_buffer; /* buffer for cloner names */
>> };
>>
>> -#define MCLPOOLS 7 /* number of cluster pools */
>> +#define MCLPOOLS 8 /* number of cluster pools */
>>
>> struct if_rxring {
>> int rxr_adjusted;
>> Index: kern/uipc_mbuf.c
>> ===================================================================
>> RCS file: /cvs/src/sys/kern/uipc_mbuf.c,v
>> retrieving revision 1.233
>> diff -u -p -r1.233 uipc_mbuf.c
>> --- kern/uipc_mbuf.c 10 Oct 2016 00:41:17 -0000 1.233
>> +++ kern/uipc_mbuf.c 14 Oct 2016 03:46:22 -0000
>> @@ -107,6 +107,7 @@ struct pool mtagpool;
>> /* mbuf cluster pools */
>> u_int mclsizes[MCLPOOLS] = {
>> MCLBYTES, /* must be at slot 0 */
>> + MCLBYTES + 2, /* ETHER_ALIGNED 2k mbufs */
>> 4 * 1024,
>> 8 * 1024,
>> 9 * 1024,
>> @@ -142,6 +143,7 @@ void
>> mbinit(void)
>> {
>> int i;
>> + unsigned int lowbits;
>>
>> #if DIAGNOSTIC
>> if (mclsizes[0] != MCLBYTES)
>> @@ -158,9 +160,15 @@ mbinit(void)
>> IPL_NET, 0, "mtagpl", NULL);
>>
>> for (i = 0; i < nitems(mclsizes); i++) {
>> - snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
>> - mclsizes[i] >> 10);
>> - pool_init(&mclpools[i], mclsizes[i], 0, IPL_NET, 0,
>> + lowbits = mclsizes[i] & ((1 << 10) - 1);
>> + if (lowbits) {
>> + snprintf(mclnames[i], sizeof(mclnames[0]),
>> + "mcl%dk%u", mclsizes[i] >> 10, lowbits);
>> + } else {
>> + snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
>> + mclsizes[i] >> 10);
>> + }
>> + pool_init(&mclpools[i], mclsizes[i], 64, IPL_NET, 0,
>> mclnames[i], NULL);
>> pool_set_constraints(&mclpools[i], &kp_dma_contig);
>> pool_setlowat(&mclpools[i], mcllowat);
>> Index: dev/pci/if_em.c
>> ===================================================================
>> RCS file: /cvs/src/sys/dev/pci/if_em.c,v
>> retrieving revision 1.331
>> diff -u -p -r1.331 if_em.c
>> --- dev/pci/if_em.c 13 Apr 2016 10:34:32 -0000 1.331
>> +++ dev/pci/if_em.c 14 Oct 2016 03:47:21 -0000
>> @@ -2450,9 +2450,7 @@ em_get_buf(struct em_softc *sc, int i)
>> return (ENOBUFS);
>> }
>> m->m_len = m->m_pkthdr.len = EM_MCLBYTES;
>> -#ifdef __STRICT_ALIGNMENT
>> m_adj(m, ETHER_ALIGN);
>> -#endif
>>
>> error = bus_dmamap_load_mbuf(sc->sc_dmat, pkt->pkt_map,
>> m, BUS_DMA_NOWAIT);
>> Index: dev/pci/if_ix.c
>> ===================================================================
>> RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
>> retrieving revision 1.132
>> diff -u -p -r1.132 if_ix.c
>> --- dev/pci/if_ix.c 13 Apr 2016 10:34:32 -0000 1.132
>> +++ dev/pci/if_ix.c 14 Oct 2016 03:47:21 -0000
>> @@ -616,11 +616,7 @@ ixgbe_init(void *arg)
>> ixgbe_initialize_transmit_units(sc);
>>
>> /* Use 2k clusters, even for jumbo frames */
>> -#ifdef __STRICT_ALIGNMENT
>> sc->rx_mbuf_sz = MCLBYTES + ETHER_ALIGN;
>> -#else
>> - sc->rx_mbuf_sz = MCLBYTES;
>> -#endif
>>
>> /* Prepare receive descriptors and buffers */
>> if (ixgbe_setup_receive_structures(sc)) {
>> @@ -2458,9 +2454,7 @@ ixgbe_get_buf(struct rx_ring *rxr, int i
>> return (ENOBUFS);
>>
>> mp->m_len = mp->m_pkthdr.len = sc->rx_mbuf_sz;
>> -#ifdef __STRICT_ALIGNMENT
>> m_adj(mp, ETHER_ALIGN);
>> -#endif
>>
>> error = bus_dmamap_load_mbuf(rxr->rxdma.dma_tag, rxbuf->map,
>> mp, BUS_DMA_NOWAIT);
>> @@ -2667,11 +2661,7 @@ ixgbe_initialize_receive_units(struct ix
>> hlreg |= IXGBE_HLREG0_JUMBOEN;
>> IXGBE_WRITE_REG(&sc->hw, IXGBE_HLREG0, hlreg);
>>
>> -#ifdef __STRICT_ALIGNMENT
>> bufsz = (sc->rx_mbuf_sz - ETHER_ALIGN) >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;
>> -#else
>> - bufsz = sc->rx_mbuf_sz >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;
>> -#endif
>>
>> for (i = 0; i < sc->num_queues; i++, rxr++) {
>> uint64_t rdba = rxr->rxdma.dma_map->dm_segs[0].ds_addr;
>> Index: dev/pci/if_em.h
>> ===================================================================
>> RCS file: /cvs/src/sys/dev/pci/if_em.h,v
>> retrieving revision 1.72
>> diff -u -p -r1.72 if_em.h
>> --- dev/pci/if_em.h 18 Feb 2016 14:24:39 -0000 1.72
>> +++ dev/pci/if_em.h 14 Oct 2016 03:48:32 -0000
>> @@ -263,11 +263,7 @@ typedef int boolean_t;
>> #define EM_RXBUFFER_8192 8192
>> #define EM_RXBUFFER_16384 16384
>>
>> -#ifdef __STRICT_ALIGNMENT
>> #define EM_MCLBYTES (EM_RXBUFFER_2048 + ETHER_ALIGN)
>> -#else
>> -#define EM_MCLBYTES EM_RXBUFFER_2048
>> -#endif
>>
>> #define EM_MAX_SCATTER 64
>> #define EM_TSO_SIZE 65535
>>
