Introduce deferred interrupt processing capability

Hyper-V interrupts seem to be sometimes delivered before the message
becomes available on the channel ring.  This is reproducible on hvs(4)
under load.  This change is modelled on the workaround found in the
Linux driver.

2 files changed, 84 insertions, 0 deletions

diff --git a/sys/dev/pv/hyperv.c b/sys/dev/pv/hyperv.c
index 2b6864dcb72..d62ae7d0c4d 100644
--- a/sys/dev/pv/hyperv.c
+++ b/sys/dev/pv/hyperv.c
@@ -104,6 +104,8 @@ struct hv_channel *
 	hv_channel_lookup(struct hv_softc *, uint32_t);
 int	hv_channel_ring_create(struct hv_channel *, uint32_t);
 void	hv_channel_ring_destroy(struct hv_channel *);
+void	hv_channel_pause(struct hv_channel *);
+uint	hv_channel_unpause(struct hv_channel *);
 extern void hv_attach_icdevs(struct hv_softc *);
 int	hv_attach_devices(struct hv_softc *);
 
@@ -1225,6 +1227,51 @@ hv_channel_setevent(struct hv_softc *sc, struct hv_channel *ch)
 		hv_intr_signal(sc, &ch->ch_monprm);
 }
 
+void
+hv_channel_intr(void *arg)
+{
+	struct hv_channel *ch = arg;
+	extern int ticks;
+	int start = ticks;
+
+	do {
+		ch->ch_handler(ch->ch_ctx);
+
+		if (hv_channel_unpause(ch) == 0)
+			return;
+
+		hv_channel_pause(ch);
+
+#if (defined(__amd64__) || defined(__i386__))
+		__asm volatile("pause": : : "memory");
+#endif
+	} while (ticks < start + 1);
+
+	hv_channel_schedule(ch);
+}
+
+int
+hv_channel_setdeferred(struct hv_channel *ch, const char *name)
+{
+	ch->ch_taskq = taskq_create(name, 1, IPL_NET, TASKQ_MPSAFE);
+	if (ch->ch_taskq == NULL)
+		return (-1);
+	task_set(&ch->ch_task, hv_channel_intr, ch);
+	return (0);
+}
+
+void
+hv_channel_schedule(struct hv_channel *ch)
+{
+	if (ch->ch_handler) {
+		if (!cold && (ch->ch_flags & CHF_BATCHED)) {
+			hv_channel_pause(ch);
+			task_add(ch->ch_taskq, &ch->ch_task);
+		} else
+			ch->ch_handler(ch->ch_ctx);
+	}
+}
+
 static inline void
 hv_ring_put(struct hv_ring_data *wrd, uint8_t *data, uint32_t datalen)
 {
@@ -1518,6 +1565,39 @@ hv_channel_recv(struct hv_channel *ch, void *data, uint32_t datalen,
 	return (rv);
 }
 
+static inline void
+hv_ring_mask(struct hv_ring_data *rd)
+{
+	virtio_membar_sync();
+	rd->rd_ring->br_imask = 1;
+	virtio_membar_sync();
+}
+
+static inline void
+hv_ring_unmask(struct hv_ring_data *rd)
+{
+	virtio_membar_sync();
+	rd->rd_ring->br_imask = 0;
+	virtio_membar_sync();
+}
+
+void
+hv_channel_pause(struct hv_channel *ch)
+{
+	hv_ring_mask(&ch->ch_rrd);
+}
+
+uint
+hv_channel_unpause(struct hv_channel *ch)
+{
+	uint32_t avail;
+
+	hv_ring_unmask(&ch->ch_rrd);
+	hv_ring_avail(&ch->ch_rrd, NULL, &avail);
+
+	return (avail);
+}
+
 /* How many PFNs can be referenced by the header */
 #define HV_NPFNHDR	((VMBUS_MSG_DSIZE_MAX -	\
 	  sizeof(struct vmbus_chanmsg_gpadl_conn)) / sizeof(uint64_t))
diff --git a/sys/dev/pv/hypervvar.h b/sys/dev/pv/hypervvar.h
index 3305a4459d6..9cd8036095d 100644
--- a/sys/dev/pv/hypervvar.h
+++ b/sys/dev/pv/hypervvar.h
@@ -80,6 +80,8 @@ struct hv_channel {
 	void				(*ch_handler)(void *);
 	void				 *ch_ctx;
 	struct evcount			  ch_evcnt;
+	struct taskq			*ch_taskq;
+	struct task			 ch_task;
 
 	uint32_t			 ch_flags;
 #define  CHF_BATCHED			  0x0001
@@ -207,6 +209,8 @@ void	hv_handle_free(struct hv_channel *, uint32_t);
 int	hv_channel_open(struct hv_channel *, size_t, void *, size_t,
 	    void (*)(void *), void *);
 int	hv_channel_close(struct hv_channel *);
+int	hv_channel_setdeferred(struct hv_channel *, const char *);
+void	hv_channel_schedule(struct hv_channel *);
 void	hv_evcount_attach(struct hv_channel *, const char *);
 int	hv_channel_send(struct hv_channel *, void *, uint32_t, uint64_t,
 	    int, uint32_t);