[src/trunk]: src/sys/arch/x86/pci vmx(4) support if_transmit and Tx multiqueu...

[knakahara <knakahara%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/6bef2a95b500
branches:  trunk
changeset: 962525:6bef2a95b500
user:      knakahara <knakahara%NetBSD.org@localhost>
date:      Wed Jul 24 10:13:14 2019 +0000

description:
vmx(4) support if_transmit and Tx multiqueue (1/2)

Implemented Tx processing only. Fix Tx interrupt handler later.

diffstat:

 sys/arch/x86/pci/if_vmx.c |  105 +++++++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 103 insertions(+), 2 deletions(-)

diffs (174 lines):

diff -r 6d54b22dd7bf -r 6bef2a95b500 sys/arch/x86/pci/if_vmx.c
--- a/sys/arch/x86/pci/if_vmx.c Wed Jul 24 10:05:57 2019 +0000
+++ b/sys/arch/x86/pci/if_vmx.c Wed Jul 24 10:13:14 2019 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: if_vmx.c,v 1.37 2019/07/23 04:50:05 knakahara Exp $    */
+/*     $NetBSD: if_vmx.c,v 1.38 2019/07/24 10:13:14 knakahara Exp $    */
 /*     $OpenBSD: if_vmx.c,v 1.16 2014/01/22 06:04:17 brad Exp $        */
 
 /*
@@ -19,7 +19,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: if_vmx.c,v 1.37 2019/07/23 04:50:05 knakahara Exp $");
+__KERNEL_RCSID(0, "$NetBSD: if_vmx.c,v 1.38 2019/07/24 10:13:14 knakahara Exp $");
 
 #include <sys/param.h>
 #include <sys/cpu.h>
@@ -29,6 +29,7 @@
 #include <sys/device.h>
 #include <sys/mbuf.h>
 #include <sys/sockio.h>
+#include <sys/pcq.h>
 
 #include <net/bpf.h>
 #include <net/if.h>
@@ -174,6 +175,7 @@
        int vxtxq_id;
        int vxtxq_intr_idx;
        int vxtxq_watchdog;
+       pcq_t *vxtxq_interq;
        struct vmxnet3_txring vxtxq_cmd_ring;
        struct vmxnet3_comp_ring vxtxq_comp_ring;
        struct vmxnet3_txq_stats vxtxq_stats;
@@ -371,6 +373,8 @@
 int vmxnet3_txq_encap(struct vmxnet3_txqueue *, struct mbuf **);
 void vmxnet3_start_locked(struct ifnet *);
 void vmxnet3_start(struct ifnet *);
+void vmxnet3_transmit_locked(struct ifnet *, struct vmxnet3_txqueue *);
+int vmxnet3_transmit(struct ifnet *, struct mbuf *);
 
 void vmxnet3_set_rxfilter(struct vmxnet3_softc *);
 int vmxnet3_ioctl(struct ifnet *, u_long, void *);
@@ -1061,6 +1065,8 @@
 
        txq->vxtxq_comp_ring.vxcr_ndesc = sc->vmx_ntxdescs;
 
+       txq->vxtxq_interq = pcq_create(sc->vmx_ntxdescs, KM_SLEEP);
+
        return (0);
 }
 
@@ -1134,12 +1140,17 @@
 vmxnet3_destroy_txq(struct vmxnet3_txqueue *txq)
 {
        struct vmxnet3_txring *txr;
+       struct mbuf *m;
 
        txr = &txq->vxtxq_cmd_ring;
 
        txq->vxtxq_sc = NULL;
        txq->vxtxq_id = -1;
 
+       while ((m = pcq_get(txq->vxtxq_interq)) != NULL)
+               m_freem(m);
+       pcq_destroy(txq->vxtxq_interq);
+
        if (txr->vxtxr_txbuf != NULL) {
                kmem_free(txr->vxtxr_txbuf,
                    txr->vxtxr_ndesc * sizeof(struct vmxnet3_txbuf));
@@ -1704,6 +1715,7 @@
        ifp->if_extflags = IFEF_MPSAFE;
        ifp->if_ioctl = vmxnet3_ioctl;
        ifp->if_start = vmxnet3_start;
+       ifp->if_transmit = vmxnet3_transmit;
        ifp->if_watchdog = NULL;
        ifp->if_init = vmxnet3_init;
        ifp->if_stop = vmxnet3_stop;
@@ -2805,6 +2817,95 @@
        VMXNET3_TXQ_UNLOCK(txq);
 }
 
+static int
+vmxnet3_select_txqueue(struct ifnet *ifp, struct mbuf *m __unused)
+{
+       struct vmxnet3_softc *sc;
+       u_int cpuid;
+
+       sc = ifp->if_softc;
+       cpuid = cpu_index(curcpu());
+       /*
+        * Furure work
+        * We should select txqueue to even up the load even if ncpu is
+        * different from sc->vmx_ntxqueues. Currently, the load is not
+        * even, that is, when ncpu is six and ntxqueues is four, the load
+        * of vmx_txq[0] and vmx_txq[1] is higher than vmx_txq[2] and
+        * vmx_txq[3] because CPU#4 always uses vmx_txq[0] and CPU#5 always
+        * uses vmx_txq[1].
+        * Furthermore, we should not use random value to select txqueue to
+        * avoid reordering. We should use flow information of mbuf.
+        */
+       return cpuid % sc->vmx_ntxqueues;
+}
+
+void
+vmxnet3_transmit_locked(struct ifnet *ifp, struct vmxnet3_txqueue *txq)
+{
+       struct vmxnet3_softc *sc;
+       struct vmxnet3_txring *txr;
+       struct mbuf *m_head;
+       int tx;
+
+       sc = ifp->if_softc;
+       txr = &txq->vxtxq_cmd_ring;
+       tx = 0;
+
+       VMXNET3_TXQ_LOCK_ASSERT(txq);
+
+       if ((ifp->if_flags & IFF_RUNNING) == 0 ||
+           sc->vmx_link_active == 0)
+               return;
+
+       for (;;) {
+               m_head = pcq_peek(txq->vxtxq_interq);
+               if (m_head == NULL)
+                       break;
+
+               if (vmxnet3_txring_avail(txr) < VMXNET3_TX_MAXSEGS)
+                       break;
+
+               m_head = pcq_get(txq->vxtxq_interq);
+               if (m_head == NULL)
+                       break;
+
+               if (vmxnet3_txq_encap(txq, &m_head) != 0) {
+                       if (m_head != NULL)
+                               m_freem(m_head);
+                       break;
+               }
+
+               tx++;
+               bpf_mtap(ifp, m_head, BPF_D_OUT);
+       }
+
+       if (tx > 0)
+               txq->vxtxq_watchdog = VMXNET3_WATCHDOG_TIMEOUT;
+}
+
+int
+vmxnet3_transmit(struct ifnet *ifp, struct mbuf *m)
+{
+       struct vmxnet3_softc *sc;
+       struct vmxnet3_txqueue *txq;
+       int qid;
+
+       qid = vmxnet3_select_txqueue(ifp, m);
+       sc = ifp->if_softc;
+       txq = &sc->vmx_txq[qid];
+
+       if (__predict_false(!pcq_put(txq->vxtxq_interq, m))) {
+               m_freem(m);
+               return ENOBUFS;
+       }
+
+       VMXNET3_TXQ_LOCK(txq);
+       vmxnet3_transmit_locked(ifp, txq);
+       VMXNET3_TXQ_UNLOCK(txq);
+
+       return 0;
+}
+
 void
 vmxnet3_set_rxfilter(struct vmxnet3_softc *sc)
 {