diff --git a/sys/dev/pci/if_de.c b/sys/dev/pci/if_de.c
index 552fbc3..ff08180 100644
--- a/sys/dev/pci/if_de.c
+++ b/sys/dev/pci/if_de.c
@@ -205,10 +205,11 @@ void tulip_intr_handler(tulip_softc_t * const sc, int *progress_p);
 int tulip_intr_shared(void *arg);
 int tulip_intr_normal(void *arg);
 struct mbuf *tulip_mbuf_compress(struct mbuf *m);
-struct mbuf *tulip_txput(tulip_softc_t * const sc, struct mbuf *m, int);
+struct mbuf *tulip_txput(tulip_softc_t * const sc, struct mbuf *m);
 void tulip_txput_setup(tulip_softc_t * const sc);
 int tulip_ifioctl(struct ifnet * ifp, u_long cmd, caddr_t data);
 void tulip_ifstart(struct ifnet *ifp);
+void tulip_ifstart_one(struct ifnet *ifp);
 void tulip_ifwatchdog(struct ifnet *ifp);
 int tulip_busdma_allocmem(tulip_softc_t * const sc, size_t size,
     bus_dmamap_t *map_p, tulip_desc_t **desc_p);
@@ -284,7 +285,7 @@ tulip_txprobe(tulip_softc_t * const sc)
     sc->tulip_flags |= TULIP_TXPROBE_ACTIVE;
     TULIP_CSR_WRITE(sc, csr_command, sc->tulip_cmdmode);
     TULIP_CSR_WRITE(sc, csr_intr, sc->tulip_intrmask);
-    if ((m = tulip_txput(sc, m, 1)) != NULL)
+    if ((m = tulip_txput(sc, m)) != NULL)
 	m_freem(m);
     sc->tulip_probe.probe_txprobes++;
     return (1);
@@ -3808,7 +3809,7 @@ tulip_mbuf_compress(struct mbuf *m)
 }
 
 struct mbuf *
-tulip_txput(tulip_softc_t * const sc, struct mbuf *m, int notonqueue)
+tulip_txput(tulip_softc_t * const sc, struct mbuf *m)
 {
     TULIP_PERFSTART(txput)
     tulip_ringinfo_t * const ri = &sc->tulip_txinfo;
@@ -3817,11 +3818,6 @@ tulip_txput(tulip_softc_t * const sc, struct mbuf *m, int notonqueue)
     u_int32_t d_status;
     bus_dmamap_t map;
     int error;
-    struct ifnet *ifp = &sc->tulip_if;
-#ifdef DIAGNOSTIC
-    struct mbuf *ombuf = m;
-#endif
-    int compressed = 0;
 
 #if defined(TULIP_DEBUG)
     if ((sc->tulip_cmdmode & TULIP_CMD_TXRUN) == 0) {
@@ -3882,19 +3878,6 @@ tulip_txput(tulip_softc_t * const sc, struct mbuf *m, int notonqueue)
 	     * entries that we can use for one packet, so we have
 	     * to recopy it into one mbuf and then try again.
 	     */
-	    struct mbuf *tmp;
-	    if (!notonqueue) {
-#ifdef DIAGNOSTIC
-		if (IF_IS_EMPTY(&ifp->if_snd))
-			panic("%s: if_snd queue empty", ifp->if_xname);
-#endif
-		IFQ_DEQUEUE(&ifp->if_snd, tmp);
-#ifdef DIAGNOSTIC
-		if (tmp != ombuf)
-		    panic("tulip_txput: different mbuf dequeued!");
-#endif
-	    }
-	    compressed = 1;
 	    m = tulip_mbuf_compress(m);
 	    if (m == NULL) {
 #if defined(TULIP_DEBUG)
@@ -3965,19 +3948,6 @@ tulip_txput(tulip_softc_t * const sc, struct mbuf *m, int notonqueue)
      * The descriptors have been filled in.  Now get ready
      * to transmit.
      */
-    if (!compressed && !notonqueue) {
-	/* remove the mbuf from the queue */
-	struct mbuf *tmp;
-#ifdef DIAGNOSTIC
-	if (IF_IS_EMPTY(&ifp->if_snd))
-	    panic("%s: if_snd queue empty", ifp->if_xname);
-#endif
-	IFQ_DEQUEUE(&ifp->if_snd, tmp);
-#ifdef DIAGNOSTIC
-	if (tmp != ombuf)
-	    panic("tulip_txput: different mbuf dequeued!");
-#endif
-    }
 
     ml_enqueue(&sc->tulip_txq, m);
     m = NULL;
@@ -4214,12 +4184,6 @@ tulip_ifioctl(struct ifnet * ifp, u_long cmd, caddr_t data)
 }
 
 /*
- * the original dequeueing policy is dequeue-and-prepend if something
- * goes wrong.
- * the modification becomes a bit complicated since tulip_txput() might
- * copy and modify the mbuf passed.
- */
-/*
  * These routines gets called at device spl (from ether_output).
  */
 
@@ -4234,24 +4198,41 @@ tulip_ifstart(struct ifnet * const ifp)
 	if ((sc->tulip_flags & (TULIP_WANTSETUP|TULIP_TXPROBE_ACTIVE)) == TULIP_WANTSETUP)
 	    tulip_txput_setup(sc);
 
-	while (!IFQ_IS_EMPTY(&sc->tulip_if.if_snd)) {
-	    struct mbuf *m, *m0;
-	    IFQ_POLL(&sc->tulip_if.if_snd, m);
+	while (!IFQ_IS_EMPTY(&ifp->if_snd)) {
+	    struct mbuf *m;
+	    IFQ_DEQUEUE(&ifp->if_snd, m);
 	    if (m == NULL)
 		break;
-	    if ((m0 = tulip_txput(sc, m, 0)) != NULL) {
-		if (m0 != m)
-		    /* should not happen */
-		    printf("tulip_if_start: txput failed!\n");
+	    if ((m = tulip_txput(sc, m)) != NULL) {
+		IFQ_REQUEUE(&ifp->if_snd, m);
 		break;
 	    }
 	}
+
+	if (IFQ_IS_EMPTY(&ifp->if_snd))
+	    ifp->if_start = tulip_ifstart_one;
     }
 
     TULIP_PERFEND(ifstart);
 }
 
 void
+tulip_ifstart_one(struct ifnet * const ifp)
+{
+    TULIP_PERFSTART(ifstart_one)
+    tulip_softc_t * const sc = TULIP_IFP_TO_SOFTC(ifp);
+    struct mbuf *m;
+
+    if (sc->tulip_if.if_flags & IFF_RUNNING) {
+	IFQ_DEQUEUE(&ifp->if_snd, m);
+	if (m != NULL && (m = tulip_txput(sc, m)) != NULL)
+	    IFQ_REQUEUE(&ifp->if_snd, m);
+    }
+
+    TULIP_PERFEND(ifstart_one);
+}
+
+void
 tulip_ifwatchdog(struct ifnet *ifp)
 {
     TULIP_PERFSTART(ifwatchdog)
diff --git a/sys/dev/pci/if_devar.h b/sys/dev/pci/if_devar.h
index 33bd9f9..13a6852 100644
--- a/sys/dev/pci/if_devar.h
+++ b/sys/dev/pci/if_devar.h
@@ -570,6 +570,7 @@ struct _tulip_softc_t {
     struct tulip_perfstats {
 	u_quad_t perf_intr_cycles;
 	u_quad_t perf_ifstart_cycles;
+	u_quad_t perf_ifstart_one_cycles;
 	u_quad_t perf_ifioctl_cycles;
 	u_quad_t perf_ifwatchdog_cycles;
 	u_quad_t perf_timeout_cycles;
@@ -579,6 +580,7 @@ struct _tulip_softc_t {
 	u_quad_t perf_rxget_cycles;
 	unsigned perf_intr;
 	unsigned perf_ifstart;
+	unsigned perf_ifstart_one;
 	unsigned perf_ifioctl;
 	unsigned perf_ifwatchdog;
 	unsigned perf_timeout;