Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 22 Nov 2008 08:05:05 +0000 (UTC)
From:      Kip Macy <kmacy@FreeBSD.org>
To:        src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org
Subject:   svn commit: r185165 - head/sys/dev/cxgb
Message-ID:  <200811220805.mAM855Qs010505@svn.freebsd.org>

next in thread | raw e-mail | index | archive | help
Author: kmacy
Date: Sat Nov 22 08:05:05 2008
New Revision: 185165
URL: http://svn.freebsd.org/changeset/base/185165

Log:
  - enable multiple transmit queues
  - invert sense of hw.cxgb.singleq tunable to hw.cxgb.multiq
  - don't wake up transmitting thread by default
  - add per tx queue ifaltq to handle ALTQ
  - remove several unused functions in cxgb_multiq.c
  - add several sysctls: multiq_tx_enable, coalesce_tx_enable,
    and wakeup_tx_thread
  - this obsoletes the hw.cxgb.snd_queue_len as ifq is replaced
    by a buf_ring

Modified:
  head/sys/dev/cxgb/cxgb_adapter.h
  head/sys/dev/cxgb/cxgb_config.h
  head/sys/dev/cxgb/cxgb_main.c
  head/sys/dev/cxgb/cxgb_multiq.c
  head/sys/dev/cxgb/cxgb_sge.c

Modified: head/sys/dev/cxgb/cxgb_adapter.h
==============================================================================
--- head/sys/dev/cxgb/cxgb_adapter.h	Sat Nov 22 07:35:45 2008	(r185164)
+++ head/sys/dev/cxgb/cxgb_adapter.h	Sat Nov 22 08:05:05 2008	(r185165)
@@ -260,7 +260,9 @@ struct sge_txq {
 	 */
 	struct mbuf_head cleanq;	
 	struct buf_ring *txq_mr;
+	struct ifaltq	*txq_ifq;
 	struct mbuf     *immpkt;
+
 	uint32_t        txq_drops;
 	uint32_t        txq_skipped;
 	uint32_t        txq_coalesced;
@@ -604,7 +606,7 @@ static inline int offload_running(adapte
 }
 
 int cxgb_pcpu_enqueue_packet(struct ifnet *ifp, struct mbuf *m);
-int cxgb_pcpu_start(struct ifnet *ifp, struct mbuf *m);
+int cxgb_pcpu_transmit(struct ifnet *ifp, struct mbuf *m);
 void cxgb_pcpu_shutdown_threads(struct adapter *sc);
 void cxgb_pcpu_startup_threads(struct adapter *sc);
 

Modified: head/sys/dev/cxgb/cxgb_config.h
==============================================================================
--- head/sys/dev/cxgb/cxgb_config.h	Sat Nov 22 07:35:45 2008	(r185164)
+++ head/sys/dev/cxgb/cxgb_config.h	Sat Nov 22 08:05:05 2008	(r185165)
@@ -32,6 +32,7 @@ $FreeBSD$
 #ifndef _CXGB_CONFIG_H_
 #define _CXGB_CONFIG_H_
 
-#define CONFIG_CHELSIO_T3_CORE
+#define	CONFIG_CHELSIO_T3_CORE
+#define	IFNET_MULTIQUEUE
 
 #endif

Modified: head/sys/dev/cxgb/cxgb_main.c
==============================================================================
--- head/sys/dev/cxgb/cxgb_main.c	Sat Nov 22 07:35:45 2008	(r185164)
+++ head/sys/dev/cxgb/cxgb_main.c	Sat Nov 22 08:05:05 2008	(r185165)
@@ -206,17 +206,17 @@ SYSCTL_UINT(_hw_cxgb, OID_AUTO, ofld_dis
 
 /*
  * The driver uses an auto-queue algorithm by default.
- * To disable it and force a single queue-set per port, use singleq = 1.
+ * To disable it and force a single queue-set per port, use multiq = 0
  */
-static int singleq = 0;
-TUNABLE_INT("hw.cxgb.singleq", &singleq);
-SYSCTL_UINT(_hw_cxgb, OID_AUTO, singleq, CTLFLAG_RDTUN, &singleq, 0,
-    "use a single queue-set per port");
-
+static int multiq = 1;
+TUNABLE_INT("hw.cxgb.multiq", &multiq);
+SYSCTL_UINT(_hw_cxgb, OID_AUTO, multiq, CTLFLAG_RDTUN, &multiq, 0,
+    "use min(ncpus/ports, 8) queue-sets per port");
 
 /*
- * The driver uses an auto-queue algorithm by default.
- * To disable it and force a single queue-set per port, use singleq = 1.
+ * By default the driver will not update the firmware unless
+ * it was compiled against a newer version
+ * 
  */
 static int force_fw_update = 0;
 TUNABLE_INT("hw.cxgb.force_fw_update", &force_fw_update);
@@ -228,15 +228,6 @@ TUNABLE_INT("hw.cxgb.use_16k_clusters", 
 SYSCTL_UINT(_hw_cxgb, OID_AUTO, use_16k_clusters, CTLFLAG_RDTUN,
     &cxgb_use_16k_clusters, 0, "use 16kB clusters for the jumbo queue ");
 
-/*
- * Tune the size of the output queue.
- */
-int cxgb_snd_queue_len = IFQ_MAXLEN;
-TUNABLE_INT("hw.cxgb.snd_queue_len", &cxgb_snd_queue_len);
-SYSCTL_UINT(_hw_cxgb, OID_AUTO, snd_queue_len, CTLFLAG_RDTUN,
-    &cxgb_snd_queue_len, 0, "send queue size ");
-
-
 enum {
 	MAX_TXQ_ENTRIES      = 16384,
 	MAX_CTRL_TXQ_ENTRIES = 1024,
@@ -368,8 +359,8 @@ cxgb_controller_probe(device_t dev)
 		ports = "ports";
 
 	snprintf(buf, sizeof(buf), "%s %sNIC, rev: %d nports: %d %s",
-		 ai->desc, is_offload(sc) ? "R" : "",
-		 sc->params.rev, nports, ports);
+	    ai->desc, is_offload(sc) ? "R" : "",
+	    sc->params.rev, nports, ports);
 	device_set_desc_copy(dev, buf);
 	return (BUS_PROBE_DEFAULT);
 }
@@ -415,8 +406,6 @@ cxgb_controller_attach(device_t dev)
 	int msi_needed, reg;
 #endif
 	int must_load = 0;
-	char buf[80];
-
 	sc = device_get_softc(dev);
 	sc->dev = dev;
 	sc->msi_count = 0;
@@ -537,7 +526,7 @@ cxgb_controller_attach(device_t dev)
 		sc->cxgb_intr = t3b_intr;
 	}
 
-	if ((sc->flags & USING_MSIX) && !singleq)
+	if ((sc->flags & USING_MSIX) && multiq)
 		port_qsets = min((SGE_QSETS/(sc)->params.nports), mp_ncpus);
 	
 	/* Create a private taskqueue thread for handling driver events */
@@ -629,11 +618,6 @@ cxgb_controller_attach(device_t dev)
 	    G_FW_VERSION_MAJOR(vers), G_FW_VERSION_MINOR(vers),
 	    G_FW_VERSION_MICRO(vers));
 
-	snprintf(buf, sizeof(buf), "%s\t E/C: %s S/N: %s", 
-		 ai->desc,
-		 sc->params.vpd.ec, sc->params.vpd.sn);
-	device_set_desc_copy(dev, buf);
-
 	device_printf(sc->dev, "Firmware Version %s\n", &sc->fw_version[0]);
 	callout_reset(&sc->cxgb_tick_ch, CXGB_TICKS(sc), cxgb_tick, sc);
 	t3_add_attach_sysctls(sc);
@@ -843,15 +827,18 @@ cxgb_setup_msix(adapter_t *sc, int msix_
 				device_printf(sc->dev, "Cannot set up "
 				    "interrupt for message %d\n", rid);
 				return (EINVAL);
+				
 			}
+#if 0			
 #ifdef IFNET_MULTIQUEUE			
-			if (singleq == 0) {
+			if (multiq) {
 				int vector = rman_get_start(sc->msix_irq_res[k]);
 				if (bootverbose)
 					device_printf(sc->dev, "binding vector=%d to cpu=%d\n", vector, k % mp_ncpus);
 				intr_bind(vector, k % mp_ncpus);
 			}
-#endif			
+#endif
+#endif
 		}
 	}
 
@@ -941,16 +928,11 @@ cxgb_port_attach(device_t dev)
 	ifp->if_ioctl = cxgb_ioctl;
 	ifp->if_start = cxgb_start;
 
-#if 0	
-#ifdef IFNET_MULTIQUEUE
-	ifp->if_flags |= IFF_MULTIQ;
-	ifp->if_mq_start = cxgb_pcpu_start;
-#endif
-#endif	
+
 	ifp->if_timer = 0;	/* Disable ifnet watchdog */
 	ifp->if_watchdog = NULL;
 
-	ifp->if_snd.ifq_drv_maxlen = cxgb_snd_queue_len;
+	ifp->if_snd.ifq_drv_maxlen = IFQ_MAXLEN;
 	IFQ_SET_MAXLEN(&ifp->if_snd, ifp->if_snd.ifq_drv_maxlen);
 	IFQ_SET_READY(&ifp->if_snd);
 
@@ -968,6 +950,8 @@ cxgb_port_attach(device_t dev)
 	}
 
 	ether_ifattach(ifp, p->hw_addr);
+
+	ifp->if_transmit = cxgb_pcpu_transmit;
 	/*
 	 * Only default to jumbo frames on 10GigE
 	 */

Modified: head/sys/dev/cxgb/cxgb_multiq.c
==============================================================================
--- head/sys/dev/cxgb/cxgb_multiq.c	Sat Nov 22 07:35:45 2008	(r185164)
+++ head/sys/dev/cxgb/cxgb_multiq.c	Sat Nov 22 08:05:05 2008	(r185165)
@@ -27,8 +27,6 @@ POSSIBILITY OF SUCH DAMAGE.
 
 ***************************************************************************/
 
-#define DEBUG_BUFRING
-
 #include <sys/cdefs.h>
 __FBSDID("$FreeBSD$");
 
@@ -89,12 +87,11 @@ __FBSDID("$FreeBSD$");
 #include <sys/mvec.h>
 
 extern int txq_fills;
-extern struct sysctl_oid_list sysctl__hw_cxgb_children;
-static int cxgb_pcpu_tx_coalesce = 0;
-TUNABLE_INT("hw.cxgb.tx_coalesce", &cxgb_pcpu_tx_coalesce);
-SYSCTL_UINT(_hw_cxgb, OID_AUTO, tx_coalesce, CTLFLAG_RDTUN, &cxgb_pcpu_tx_coalesce, 0,
-    "coalesce small packets into a single work request");
+int multiq_tx_enable = 1;
+int coalesce_tx_enable = 0;
+int wakeup_tx_thread = 0;
 
+extern struct sysctl_oid_list sysctl__hw_cxgb_children;
 static int sleep_ticks = 1;
 TUNABLE_INT("hw.cxgb.sleep_ticks", &sleep_ticks);
 SYSCTL_UINT(_hw_cxgb, OID_AUTO, sleep_ticks, CTLFLAG_RDTUN, &sleep_ticks, 0,
@@ -106,12 +103,12 @@ SYSCTL_UINT(_hw_cxgb, OID_AUTO, txq_mr_s
     "size of per-queue mbuf ring");
 
 
-static inline int32_t cxgb_pcpu_calc_cookie(struct ifnet *ifp, struct mbuf *immpkt);
 static void cxgb_pcpu_start_proc(void *arg);
+static int cxgb_tx(struct sge_qset *qs, uint32_t txmax);
+
 #ifdef IFNET_MULTIQUEUE
-static int cxgb_pcpu_cookie_to_qidx(struct port_info *, uint32_t cookie);
+static int cxgb_pcpu_cookie_to_qidx(struct port_info *pi, uint32_t cookie);
 #endif
-static int cxgb_tx(struct sge_qset *qs, uint32_t txmax);
 
 static inline int
 cxgb_pcpu_enqueue_packet_(struct sge_qset *qs, struct mbuf *m)
@@ -119,9 +116,6 @@ cxgb_pcpu_enqueue_packet_(struct sge_qse
 	struct sge_txq *txq;
 	int err = 0;
 
-#ifndef IFNET_MULTIQUEUE
-	panic("not expecting enqueue without multiqueue");
-#endif	
 	KASSERT(m != NULL, ("null mbuf"));
 	KASSERT(m->m_type == MT_DATA, ("bad mbuf type %d", m->m_type));
 	if (qs->qs_flags & QS_EXITING) {
@@ -134,9 +128,9 @@ cxgb_pcpu_enqueue_packet_(struct sge_qse
 		txq->txq_drops++;
 		m_freem(m);
 	}
-	if ((qs->txq[TXQ_ETH].flags & TXQ_TRANSMITTING) == 0)
+	if (wakeup_tx_thread && ((txq->flags & TXQ_TRANSMITTING) == 0))
 		wakeup(qs);
-
+	
 	return (err);
 }
 
@@ -149,29 +143,34 @@ cxgb_pcpu_enqueue_packet(struct ifnet *i
 #ifdef IFNET_MULTIQUEUE
 	int32_t calc_cookie;
 
-	calc_cookie = m->m_pkthdr.rss_hash;
+	calc_cookie = m->m_pkthdr.flowid;
 	qidx = cxgb_pcpu_cookie_to_qidx(pi, calc_cookie);
 #else
 	qidx = 0;
 #endif	    
 	qs = &pi->adapter->sge.qs[qidx];
-	err = cxgb_pcpu_enqueue_packet_(qs, m);
+	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
+		IFQ_ENQUEUE(qs->txq[0].txq_ifq, m, err);
+	} else {
+		err = cxgb_pcpu_enqueue_packet_(qs, m);
+	}
 	return (err);
 }
 
 static int
 cxgb_dequeue_packet(struct sge_txq *txq, struct mbuf **m_vec)
 {
-	struct mbuf *m;
+	struct mbuf *m, *m0;
 	struct sge_qset *qs;
 	int count, size, coalesced;
 	struct adapter *sc;
+
 #ifndef IFNET_MULTIQUEUE
 	struct port_info *pi = txq->port;
 
+	mtx_assert(&txq->lock, MA_OWNED);
 	if (txq->immpkt != NULL)
 		panic("immediate packet set");
-	mtx_assert(&txq->lock, MA_OWNED);
 
 	IFQ_DRV_DEQUEUE(&pi->ifp->if_snd, m);
 	if (m == NULL)
@@ -180,7 +179,16 @@ cxgb_dequeue_packet(struct sge_txq *txq,
 	m_vec[0] = m;
 	return (1);
 #endif
-
+	if (ALTQ_IS_ENABLED(txq->txq_ifq)) {
+		IFQ_DRV_DEQUEUE(txq->txq_ifq, m);
+		if (m == NULL)
+			return (0);
+	
+		m_vec[0] = m;
+		return (1);		
+	}
+	
+	mtx_assert(&txq->lock, MA_OWNED);
 	coalesced = count = size = 0;
 	qs = txq_to_qset(txq, TXQ_ETH);
 	if (qs->qs_flags & QS_EXITING)
@@ -199,9 +207,10 @@ cxgb_dequeue_packet(struct sge_txq *txq,
 		return (0);
 
 	count = 1;
+
 	m_vec[0] = m;
 	if (m->m_pkthdr.tso_segsz > 0 || m->m_pkthdr.len > TX_WR_SIZE_MAX ||
-	    m->m_next != NULL || (cxgb_pcpu_tx_coalesce == 0)) {
+	    m->m_next != NULL || (coalesce_tx_enable == 0)) {
 		return (count);
 	}
 
@@ -213,7 +222,11 @@ cxgb_dequeue_packet(struct sge_txq *txq,
 		    size + m->m_pkthdr.len > TX_WR_SIZE_MAX || m->m_next != NULL)
 			break;
 
-		buf_ring_dequeue_sc(txq->txq_mr);
+		m0 = buf_ring_dequeue_sc(txq->txq_mr);
+#ifdef DEBUG_BUFRING
+		if (m0 != m)
+			panic("peek and dequeue don't match");
+#endif		
 		size += m->m_pkthdr.len;
 		m_vec[count++] = m;
 
@@ -227,134 +240,6 @@ cxgb_dequeue_packet(struct sge_txq *txq,
 	return (count);
 }
 
-static int32_t
-cxgb_pcpu_get_cookie(struct ifnet *ifp, struct in6_addr *lip, uint16_t lport, struct in6_addr *rip, uint16_t rport, int ipv6)
-{
-	uint32_t base;
-	uint8_t buf[36];
-	int count;
-	int32_t cookie;
-
-	critical_enter();
-	/* 
-	 * Can definitely bypass bcopy XXX
-	 */
-	if (ipv6 == 0) {
-		count = 12;
-		bcopy(rip, &buf[0], 4);
-		bcopy(lip, &buf[4], 4);
-		bcopy(&rport, &buf[8], 2);
-		bcopy(&lport, &buf[10], 2);
-	} else {
-		count = 36;
-		bcopy(rip, &buf[0], 16);
-		bcopy(lip, &buf[16], 16);
-		bcopy(&rport, &buf[32], 2);
-		bcopy(&lport, &buf[34], 2);
-	}
-	
-	base = 0xffffffff;
-	base = update_crc32(base, buf, count);
-	base = sctp_csum_finalize(base);
-
-	/*
-	 * Indirection table is 128 bits
-	 * -> cookie indexes into indirection table which maps connection to queue
-	 * -> RSS map maps queue to CPU
-	 */
-	cookie = (base & (RSS_TABLE_SIZE-1));
-	critical_exit();
-	
-	return (cookie);
-}
-
-static int32_t
-cxgb_pcpu_calc_cookie(struct ifnet *ifp, struct mbuf *immpkt)
-{
-	struct in6_addr lip, rip;
-	uint16_t lport, rport;
-	struct ether_header *eh;
-	int32_t cookie;
-	struct ip *ip;
-	struct ip6_hdr *ip6;
-	struct tcphdr *th;
-	struct udphdr *uh;
-	struct sctphdr *sh;
-	uint8_t *next, proto;
-	int etype;
-
-	if (immpkt == NULL)
-		return -1;
-
-#if 1	
-	/*
-	 * XXX perf test
-	 */
-	return (0);
-#endif	
-	rport = lport = 0;
-	cookie = -1;
-	next = NULL;
-	eh = mtod(immpkt, struct ether_header *);
-	etype = ntohs(eh->ether_type);
-
-	switch (etype) {
-	case ETHERTYPE_IP:
-		ip = (struct ip *)(eh + 1);
-		next = (uint8_t *)(ip + 1);
-		bcopy(&ip->ip_src, &lip, 4);
-		bcopy(&ip->ip_dst, &rip, 4);
-		proto = ip->ip_p;
-		break;
-	case ETHERTYPE_IPV6:
-		ip6 = (struct ip6_hdr *)(eh + 1);
-		next = (uint8_t *)(ip6 + 1);
-		bcopy(&ip6->ip6_src, &lip, sizeof(struct in6_addr));
-		bcopy(&ip6->ip6_dst, &rip, sizeof(struct in6_addr));
-		if (ip6->ip6_nxt == IPPROTO_HOPOPTS) {
-			struct ip6_hbh *hbh;
-
-			hbh = (struct ip6_hbh *)(ip6 + 1);
-			proto = hbh->ip6h_nxt;
-		} else 
-			proto = ip6->ip6_nxt;
-		break;
-	case ETHERTYPE_ARP:
-	default:
-		/*
-		 * Default to queue zero
-		 */
-		proto = cookie = 0;
-	}
-	if (proto) {
-		switch (proto) {
-		case IPPROTO_TCP:
-			th = (struct tcphdr *)next;
-			lport = th->th_sport;
-			rport = th->th_dport;
-			break;
-		case IPPROTO_UDP:
-			uh = (struct udphdr *)next;
-			lport = uh->uh_sport;
-			rport = uh->uh_dport;
-			break;
-		case IPPROTO_SCTP:
-			sh = (struct sctphdr *)next;
-			lport = sh->src_port;
-			rport = sh->dest_port;
-			break;
-		default:
-			/* nothing to do */
-			break;
-		}
-	}
-	
-	if (cookie) 		
-		cookie = cxgb_pcpu_get_cookie(ifp, &lip, lport, &rip, rport, (etype == ETHERTYPE_IPV6));
-
-	return (cookie);
-}
-
 static void
 cxgb_pcpu_free(struct sge_qset *qs)
 {
@@ -377,10 +262,6 @@ cxgb_pcpu_reclaim_tx(struct sge_txq *txq
 	int reclaimable;
 	struct sge_qset *qs = txq_to_qset(txq, TXQ_ETH);
 
-#ifdef notyet
-	KASSERT(qs->qs_cpuid == curcpu, ("cpu qset mismatch cpuid=%d curcpu=%d",
-			qs->qs_cpuid, curcpu));
-#endif
 	mtx_assert(&txq->lock, MA_OWNED);
 	
 	reclaimable = desc_reclaimable(txq);
@@ -426,7 +307,8 @@ cxgb_pcpu_start_(struct sge_qset *qs, st
 		initerr = ENXIO;
 	else if (immpkt) {
 
-		if (!buf_ring_empty(txq->txq_mr)) 
+		if (!buf_ring_empty(txq->txq_mr)
+		    || ALTQ_IS_ENABLED(&pi->ifp->if_snd)) 
 			initerr = cxgb_pcpu_enqueue_packet_(qs, immpkt);
 		else
 			txq->immpkt = immpkt;
@@ -491,7 +373,7 @@ cxgb_pcpu_start_(struct sge_qset *qs, st
 }
 
 int
-cxgb_pcpu_start(struct ifnet *ifp, struct mbuf *immpkt)
+cxgb_pcpu_transmit(struct ifnet *ifp, struct mbuf *immpkt)
 {
 	uint32_t cookie;
 	int err, qidx, locked, resid;
@@ -506,10 +388,10 @@ cxgb_pcpu_start(struct ifnet *ifp, struc
 	qidx = resid = err = cookie = locked = 0;
 
 #ifdef IFNET_MULTIQUEUE	
-	if (immpkt && (immpkt->m_pkthdr.rss_hash != 0)) {
-		cookie = immpkt->m_pkthdr.rss_hash;
+	if (immpkt && (immpkt->m_pkthdr.flowid != 0)) {
+		cookie = immpkt->m_pkthdr.flowid;
 		qidx = cxgb_pcpu_cookie_to_qidx(pi, cookie);
-		DPRINTF("hash=0x%x qidx=%d cpu=%d\n", immpkt->m_pkthdr.rss_hash, qidx, curcpu);
+		DPRINTF("hash=0x%x qidx=%d cpu=%d\n", immpkt->m_pkthdr.flowid, qidx, curcpu);
 		qs = &pi->adapter->sge.qs[qidx];
 	} else
 #endif		
@@ -519,7 +401,7 @@ cxgb_pcpu_start(struct ifnet *ifp, struc
 
 	if (((sc->tunq_coalesce == 0) ||
 		(buf_ring_count(txq->txq_mr) >= TX_WR_COUNT_MAX) ||
-		(cxgb_pcpu_tx_coalesce == 0)) && mtx_trylock(&txq->lock)) {
+		(coalesce_tx_enable == 0)) && mtx_trylock(&txq->lock)) {
 		if (cxgb_debug)
 			printf("doing immediate transmit\n");
 		
@@ -534,10 +416,6 @@ cxgb_pcpu_start(struct ifnet *ifp, struc
 			    sc->tunq_coalesce, buf_ring_count(txq->txq_mr), mtx_owned(&txq->lock));
 		err = cxgb_pcpu_enqueue_packet_(qs, immpkt);
 	}
-	
-	if (resid && (txq->flags & TXQ_TRANSMITTING) == 0)
-		wakeup(qs);
-
 	return ((err == ENOSPC) ? 0 : err);
 }
 
@@ -552,7 +430,7 @@ cxgb_start(struct ifnet *ifp)
 	if (IFQ_DRV_IS_EMPTY(&ifp->if_snd))
 		return;
 
-	cxgb_pcpu_start(ifp, NULL);
+	cxgb_pcpu_transmit(ifp, NULL);
 }
 
 static void
@@ -562,11 +440,8 @@ cxgb_pcpu_start_proc(void *arg)
 	struct thread *td;
 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
 	int idleticks, err = 0;
-#ifdef notyet	
-	struct adapter *sc = qs->port->adapter;
-#endif
-	td = curthread;
 
+	td = curthread;
 	sleep_ticks = max(hz/1000, 1);
 	qs->qs_flags |= QS_RUNNING;
 	thread_lock(td);
@@ -642,6 +517,9 @@ cxgb_pcpu_cookie_to_qidx(struct port_inf
 {
 	int qidx;
 	uint32_t tmp;
+
+	if (multiq_tx_enable == 0)
+		return (pi->first_qset);
 	
 	 /*
 	 * Will probably need to be changed for 4-port XXX
@@ -757,7 +635,7 @@ cxgb_tx(struct sge_qset *qs, uint32_t tx
 		check_pkt_coalesce(qs);
 		count = cxgb_dequeue_packet(txq, m_vec);
 		if (count == 0) {
-			err = ENOBUFS;
+			err = ENOSPC;
 			break;
 		}
 		ETHER_BPF_MTAP(ifp, m_vec[0]);
@@ -767,28 +645,10 @@ cxgb_tx(struct sge_qset *qs, uint32_t tx
 		txq->txq_enqueued += count;
 		m_vec[0] = NULL;
 	}
-#if 0 /* !MULTIQ */
-	if (__predict_false(err)) {
-		if (err == ENOMEM) {
-			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
-			IFQ_LOCK(&ifp->if_snd);
-			IFQ_DRV_PREPEND(&ifp->if_snd, m_vec[0]);
-			IFQ_UNLOCK(&ifp->if_snd);
-		}
-	}
-	else if ((err == 0) &&  (txq->size <= txq->in_use + TX_MAX_DESC) &&
-	    (ifp->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
-		setbit(&qs->txq_stopped, TXQ_ETH);
-		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
-		txq_fills++;
-		err = ENOSPC;
-	}
-#else
 	if ((err == 0) &&  (txq->size <= txq->in_use + TX_MAX_DESC)) {
-		err = ENOSPC;
+		err = ENOBUFS;
 		txq_fills++;
 		setbit(&qs->txq_stopped, TXQ_ETH);
-		ifp->if_drv_flags |= IFF_DRV_OACTIVE;
 	}
 	if (err == ENOMEM) {
 		int i;
@@ -799,7 +659,6 @@ cxgb_tx(struct sge_qset *qs, uint32_t tx
 		for (i = 0; i < count; i++)
 			m_freem(m_vec[i]);
 	}
-#endif
 	return (err);
 }
 

Modified: head/sys/dev/cxgb/cxgb_sge.c
==============================================================================
--- head/sys/dev/cxgb/cxgb_sge.c	Sat Nov 22 07:35:45 2008	(r185164)
+++ head/sys/dev/cxgb/cxgb_sge.c	Sat Nov 22 08:05:05 2008	(r185165)
@@ -26,8 +26,6 @@ ARISING IN ANY WAY OUT OF THE USE OF THI
 POSSIBILITY OF SUCH DAMAGE.
 
 ***************************************************************************/
-#define DEBUG_BUFRING
-
 
 #include <sys/cdefs.h>
 __FBSDID("$FreeBSD$");
@@ -86,9 +84,9 @@ extern int cxgb_pcpu_cache_enable;
 extern int nmbjumbo4;
 extern int nmbjumbo9;
 extern int nmbjumbo16;
-
-
-
+extern int multiq_tx_enable;
+extern int coalesce_tx_enable;
+extern int wakeup_tx_thread;
 
 #define USE_GTS 0
 
@@ -1275,7 +1273,6 @@ t3_encap(struct sge_qset *qs, struct mbu
 	KASSERT(txsd->mi.mi_base == NULL,
 	    ("overwriting valid entry mi_base==%p", txsd->mi.mi_base));
 	if (count > 1) {
-		panic("count > 1 not support in CVS\n");
 		if ((err = busdma_map_sg_vec(m, &m0, segs, count)))
 			return (err);
 		nsegs = count;
@@ -1286,7 +1283,7 @@ t3_encap(struct sge_qset *qs, struct mbu
 	} 
 	KASSERT(m0->m_pkthdr.len, ("empty packet nsegs=%d count=%d", nsegs, count));
 
-	if (!(m0->m_pkthdr.len <= PIO_LEN)) {
+	if ((m0->m_pkthdr.len > PIO_LEN) || (count > 1)) {
 		mi_collapse_mbuf(&txsd->mi, m0);
 		mi = &txsd->mi;
 	}
@@ -1718,9 +1715,15 @@ t3_free_qset(adapter_t *sc, struct sge_q
 	
 	t3_free_tx_desc_all(&q->txq[TXQ_ETH]);
 	
-	for (i = 0; i < SGE_TXQ_PER_SET; i++) 
+	for (i = 0; i < SGE_TXQ_PER_SET; i++) {
 		if (q->txq[i].txq_mr != NULL) 
 			buf_ring_free(q->txq[i].txq_mr, M_DEVBUF);
+		if (q->txq[i].txq_ifq != NULL) {
+			ifq_detach(q->txq[i].txq_ifq);
+			free(q->txq[i].txq_ifq, M_DEVBUF);
+		}
+	}
+	
 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
 		if (q->fl[i].desc) {
 			mtx_lock_spin(&sc->sge.reg_lock);
@@ -1882,19 +1885,16 @@ t3_free_tx_desc(struct sge_txq *q, int r
 				bus_dmamap_unload(q->entry_tag, txsd->map);
 				txsd->flags &= ~TX_SW_DESC_MAPPED;
 			}
-			m_freem_iovec(&txsd->mi);	
+			m_freem_iovec(&txsd->mi);
+#if 0
+			buf_ring_scan(&q->txq_mr, txsd->mi.mi_base, __FILE__, __LINE__);
+#endif
 			txsd->mi.mi_base = NULL;
 			/*
 			 * XXX check for cache hit rate here
 			 *
 			 */
 			q->port->ifp->if_opackets++;
-#if defined(DIAGNOSTIC) && 0
-			if (m_get_priority(txsd->m[0]) != cidx) 
-				printf("pri=%d cidx=%d\n",
-				    (int)m_get_priority(txsd->m[0]), cidx);
-#endif			
-
 		} else
 			q->txq_skipped++;
 		
@@ -2288,11 +2288,16 @@ t3_sge_alloc_qset(adapter_t *sc, u_int i
 			device_printf(sc->dev, "failed to allocate mbuf ring\n");
 			goto err;
 		}
+		if ((q->txq[i].txq_ifq =
+			malloc(sizeof(struct ifaltq), M_DEVBUF, M_NOWAIT|M_ZERO))
+		    == NULL) {
+			device_printf(sc->dev, "failed to allocate ifq\n");
+			goto err;
+		}
+		ifq_attach(q->txq[i].txq_ifq, pi->ifp);
 	}
-
 	init_qset_cntxt(q, id);
 	q->idx = id;
-	
 	if ((ret = alloc_ring(sc, p->fl_size, sizeof(struct rx_desc),
 		    sizeof(struct rx_sw_desc), &q->fl[0].phys_addr,
 		    &q->fl[0].desc, &q->fl[0].sdesc,
@@ -2880,7 +2885,7 @@ process_responses(adapter_t *adap, struc
 			eop = get_packet(adap, drop_thresh, qs, &rspq->rspq_mbuf, r);
 #endif
 #ifdef IFNET_MULTIQUEUE
-			rspq->rspq_mh.mh_head->m_pkthdr.rss_hash = rss_hash;
+			rspq->rspq_mh.mh_head->m_pkthdr.flowid = rss_hash;
 #endif			
 			ethpad = 2;
 		} else {
@@ -3365,6 +3370,18 @@ t3_add_attach_sysctls(adapter_t *sc)
 	    CTLFLAG_RW, &cxgb_pcpu_cache_enable,
 	    0, "#enable driver local pcpu caches");
 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, 
+	    "multiq_tx_enable",
+	    CTLFLAG_RW, &multiq_tx_enable,
+	    0, "enable transmit by multiple tx queues");
+	SYSCTL_ADD_INT(ctx, children, OID_AUTO, 
+	    "coalesce_tx_enable",
+	    CTLFLAG_RW, &coalesce_tx_enable,
+	    0, "coalesce small packets in work requests - WARNING ALPHA");
+	SYSCTL_ADD_INT(ctx, children, OID_AUTO, 
+	    "wakeup_tx_thread",
+	    CTLFLAG_RW, &wakeup_tx_thread,
+	    0, "wakeup tx thread if no transmitter running");
+	SYSCTL_ADD_INT(ctx, children, OID_AUTO, 
 	    "cache_alloc",
 	    CTLFLAG_RD, &cxgb_cached_allocations,
 	    0, "#times a cluster was allocated from cache");



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200811220805.mAM855Qs010505>