session svm: non blocking mq

Avoid synchronizing producers and the consumer. Instead, only use mutex
or spinlock (if eventfds are configured) to synchronize producers.

Type: improvement

Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ie2aafbdc2e07fced5d5e46ee2df6b30a186faa2f

2 files changed, 181 insertions, 74 deletions

diff --git a/src/svm/message_queue.c b/src/svm/message_queue.c
index fdf9293b18c..b423826cb83 100644
--- a/src/svm/message_queue.c
+++ b/src/svm/message_queue.c
@@ -163,21 +163,36 @@ svm_msg_q_attach (svm_msg_q_t *mq, void *smq_base)
       offset = sizeof (*ring) + ring->nitems * ring->elsize;
       ring = (void *) ((u8 *) ring + offset);
     }
+  clib_spinlock_init (&mq->q.lock);
 }
 
 void
 svm_msg_q_free (svm_msg_q_t * mq)
 {
   clib_mem_free (mq->q.shr);
+  clib_spinlock_free (&mq->q.lock);
   clib_mem_free (mq);
 }
 
 static void
-svm_msg_q_send_signal (svm_msg_q_t *mq)
+svm_msg_q_send_signal (svm_msg_q_t *mq, u8 is_consumer)
 {
   if (mq->q.evtfd == -1)
     {
+      if (is_consumer)
+	{
+	  int rv = pthread_mutex_lock (&mq->q.shr->mutex);
+	  if (PREDICT_FALSE (rv == EOWNERDEAD))
+	    {
+	      rv = pthread_mutex_consistent (&mq->q.shr->mutex);
+	      return;
+	    }
+	}
+
       (void) pthread_cond_broadcast (&mq->q.shr->condvar);
+
+      if (is_consumer)
+	pthread_mutex_unlock (&mq->q.shr->mutex);
     }
   else
     {
@@ -232,7 +247,7 @@ svm_msg_q_lock_and_alloc_msg_w_ring (svm_msg_q_t * mq, u32 ring_index,
       svm_msg_q_lock (mq);
       while (svm_msg_q_is_full (mq)
 	     || svm_msg_q_ring_is_full (mq, ring_index))
-	svm_msg_q_wait (mq);
+	svm_msg_q_wait (mq, SVM_MQ_WAIT_FULL);
       *msg = svm_msg_q_alloc_msg_w_ring (mq, ring_index);
     }
   return 0;
@@ -253,7 +268,7 @@ svm_msg_q_alloc_msg (svm_msg_q_t * mq, u32 nbytes)
     msg.ring_index = ring - mq->rings;
     msg.elt_index = sr->tail;
     sr->tail = (sr->tail + 1) % ring->nitems;
-    clib_atomic_fetch_add_rel (&sr->cursize, 1);
+    clib_atomic_fetch_add_relax (&sr->cursize, 1);
     break;
   }
   return msg;
@@ -271,7 +286,7 @@ svm_msg_q_free_msg (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
 {
   svm_msg_q_ring_shared_t *sr;
   svm_msg_q_ring_t *ring;
-  int need_signal;
+  u32 need_signal;
 
   ASSERT (vec_len (mq->rings) > msg->ring_index);
   ring = svm_msg_q_ring_inline (mq, msg->ring_index);
@@ -282,16 +297,17 @@ svm_msg_q_free_msg (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
     }
   else
     {
-      clib_warning ("message out of order");
+      clib_warning ("message out of order: elt %u head %u ring %u",
+		    msg->elt_index, sr->head, msg->ring_index);
       /* for now, expect messages to be processed in order */
       ASSERT (0);
     }
 
-  need_signal = sr->cursize == ring->nitems;
-  clib_atomic_fetch_sub_rel (&sr->cursize, 1);
+  need_signal = clib_atomic_load_relax_n (&sr->cursize) == ring->nitems;
+  clib_atomic_fetch_sub_relax (&sr->cursize, 1);
 
   if (PREDICT_FALSE (need_signal))
-    svm_msg_q_send_signal (mq);
+    svm_msg_q_send_signal (mq, 1 /* is consumer */);
 }
 
 static int
@@ -331,7 +347,7 @@ svm_msg_q_add_raw (svm_msg_q_t *mq, u8 *elem)
 
   sz = clib_atomic_fetch_add_rel (&sq->cursize, 1);
   if (!sz)
-    svm_msg_q_send_signal (mq);
+    svm_msg_q_send_signal (mq, 0 /* is consumer */);
 }
 
 int
@@ -355,7 +371,7 @@ svm_msg_q_add (svm_msg_q_t * mq, svm_msg_q_msg_t * msg, int nowait)
       if (nowait)
 	return (-2);
       while (svm_msg_q_is_full (mq))
-	svm_msg_q_wait (mq);
+	svm_msg_q_wait (mq, SVM_MQ_WAIT_FULL);
     }
 
   svm_msg_q_add_raw (mq, (u8 *) msg);
@@ -373,8 +389,8 @@ svm_msg_q_add_and_unlock (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
   svm_msg_q_unlock (mq);
 }
 
-static int
-svm_msg_q_sub_raw (svm_msg_q_t *mq, u8 *elem)
+int
+svm_msg_q_sub_raw (svm_msg_q_t *mq, svm_msg_q_msg_t *elem)
 {
   svm_msg_q_shared_queue_t *sq = mq->q.shr;
   i8 *headp;
@@ -387,69 +403,76 @@ svm_msg_q_sub_raw (svm_msg_q_t *mq, u8 *elem)
 
   sq->head = (sq->head + 1) % sq->maxsize;
 
-  sz = clib_atomic_fetch_sub_rel (&sq->cursize, 1);
+  sz = clib_atomic_fetch_sub_relax (&sq->cursize, 1);
   if (PREDICT_FALSE (sz == sq->maxsize))
-    svm_msg_q_send_signal (mq);
+    svm_msg_q_send_signal (mq, 1 /* is consumer */);
 
   return 0;
 }
 
 int
-svm_msg_q_sub (svm_msg_q_t * mq, svm_msg_q_msg_t * msg,
-	       svm_q_conditional_wait_t cond, u32 time)
+svm_msg_q_sub_raw_batch (svm_msg_q_t *mq, svm_msg_q_msg_t *msg_buf, u32 n_msgs)
 {
-  int rc = 0;
+  svm_msg_q_shared_queue_t *sq = mq->q.shr;
+  u32 sz, to_deq;
+  i8 *headp;
 
-  if (cond == SVM_Q_NOWAIT)
+  sz = svm_msg_q_size (mq);
+  ASSERT (sz);
+  to_deq = clib_min (sz, n_msgs);
+
+  headp = (i8 *) (&sq->data[0] + sq->elsize * sq->head);
+
+  if (sq->head + to_deq < sq->maxsize)
     {
-      /* zero on success */
-      if (svm_msg_q_try_lock (mq))
-	{
-	  return (-1);
-	}
+      clib_memcpy_fast (msg_buf, headp, sq->elsize * to_deq);
+      sq->head += to_deq;
     }
   else
-    svm_msg_q_lock (mq);
+    {
+      u32 first_batch = sq->maxsize - sq->head;
+      clib_memcpy_fast (msg_buf, headp, sq->elsize * first_batch);
+      clib_memcpy_fast (msg_buf + first_batch, sq->data,
+			sq->elsize * (to_deq - first_batch));
+      sq->head = (sq->head + to_deq) % sq->maxsize;
+    }
+
+  clib_atomic_fetch_sub_relax (&sq->cursize, to_deq);
+  if (PREDICT_FALSE (sz == sq->maxsize))
+    svm_msg_q_send_signal (mq, 1 /* is consumer */);
+
+  return to_deq;
+}
+
+int
+svm_msg_q_sub (svm_msg_q_t *mq, svm_msg_q_msg_t *msg,
+	       svm_q_conditional_wait_t cond, u32 time)
+{
+  int rc = 0;
 
-  if (PREDICT_FALSE (svm_msg_q_is_empty (mq)))
+  if (svm_msg_q_is_empty (mq))
     {
       if (cond == SVM_Q_NOWAIT)
 	{
-	  svm_msg_q_unlock (mq);
 	  return (-2);
 	}
       else if (cond == SVM_Q_TIMEDWAIT)
 	{
-	  while (svm_msg_q_is_empty (mq) && rc == 0)
-	    rc = svm_msg_q_timedwait (mq, time);
-
-	  if (rc == ETIMEDOUT)
-	    {
-	      svm_msg_q_unlock (mq);
-	      return ETIMEDOUT;
-	    }
+	  if ((rc = svm_msg_q_timedwait (mq, time)))
+	    return rc;
 	}
       else
 	{
-	  while (svm_msg_q_is_empty (mq))
-	    svm_msg_q_wait (mq);
+	  svm_msg_q_wait (mq, SVM_MQ_WAIT_EMPTY);
 	}
     }
 
-  svm_msg_q_sub_raw (mq, (u8 *) msg);
-
-  svm_msg_q_unlock (mq);
+  svm_msg_q_sub_raw (mq, msg);
 
   return 0;
 }
 
 void
-svm_msg_q_sub_w_lock (svm_msg_q_t *mq, svm_msg_q_msg_t *msg)
-{
-  svm_msg_q_sub_raw (mq, (u8 *) msg);
-}
-
-void
 svm_msg_q_set_eventfd (svm_msg_q_t *mq, int fd)
 {
   mq->q.evtfd = fd;
@@ -465,29 +488,46 @@ svm_msg_q_alloc_eventfd (svm_msg_q_t *mq)
   return 0;
 }
 
-void
-svm_msg_q_wait (svm_msg_q_t *mq)
+int
+svm_msg_q_wait (svm_msg_q_t *mq, svm_msg_q_wait_type_t type)
 {
+  u8 (*fn) (svm_msg_q_t *);
+  int rv;
+
+  fn = (type == SVM_MQ_WAIT_EMPTY) ? svm_msg_q_is_empty : svm_msg_q_is_full;
+
   if (mq->q.evtfd == -1)
     {
-      pthread_cond_wait (&mq->q.shr->condvar, &mq->q.shr->mutex);
+      rv = pthread_mutex_lock (&mq->q.shr->mutex);
+      if (PREDICT_FALSE (rv == EOWNERDEAD))
+	{
+	  rv = pthread_mutex_consistent (&mq->q.shr->mutex);
+	  return rv;
+	}
+
+      while (fn (mq))
+	pthread_cond_wait (&mq->q.shr->condvar, &mq->q.shr->mutex);
+
+      pthread_mutex_unlock (&mq->q.shr->mutex);
     }
   else
     {
       u64 buf;
-      int rv;
 
-      svm_msg_q_unlock (mq);
-      while ((rv = read (mq->q.evtfd, &buf, sizeof (buf))) < 0)
+      while (fn (mq))
 	{
-	  if (errno != EAGAIN)
+	  while ((rv = read (mq->q.evtfd, &buf, sizeof (buf))) < 0)
 	    {
-	      clib_unix_warning ("read error");
-	      return;
+	      if (errno != EAGAIN)
+		{
+		  clib_unix_warning ("read error");
+		  return rv;
+		}
 	    }
 	}
-      svm_msg_q_lock (mq);
     }
+
+  return 0;
 }
 
 int
@@ -495,11 +535,32 @@ svm_msg_q_timedwait (svm_msg_q_t *mq, double timeout)
 {
   if (mq->q.evtfd == -1)
     {
+      svm_msg_q_shared_queue_t *sq = mq->q.shr;
       struct timespec ts;
+      u32 sz;
+      int rv;
+
+      rv = pthread_mutex_lock (&sq->mutex);
+      if (PREDICT_FALSE (rv == EOWNERDEAD))
+	{
+	  rv = pthread_mutex_consistent (&sq->mutex);
+	  return rv;
+	}
+
+      /* check if we're still in a signalable state after grabbing lock */
+      sz = svm_msg_q_size (mq);
+      if (sz != 0 && sz != sq->maxsize)
+	{
+	  pthread_mutex_unlock (&sq->mutex);
+	  return 0;
+	}
+
       ts.tv_sec = unix_time_now () + (u32) timeout;
       ts.tv_nsec = (timeout - (u32) timeout) * 1e9;
-      return pthread_cond_timedwait (&mq->q.shr->condvar, &mq->q.shr->mutex,
-				     &ts);
+      rv = pthread_cond_timedwait (&sq->condvar, &sq->mutex, &ts);
+
+      pthread_mutex_unlock (&sq->mutex);
+      return rv;
     }
   else
     {
@@ -512,11 +573,9 @@ svm_msg_q_timedwait (svm_msg_q_t *mq, double timeout)
       setsockopt (mq->q.evtfd, SOL_SOCKET, SO_RCVTIMEO, (const char *) &tv,
 		  sizeof tv);
 
-      svm_msg_q_unlock (mq);
       rv = read (mq->q.evtfd, &buf, sizeof (buf));
       if (rv < 0)
 	clib_warning ("read %u", errno);
-      svm_msg_q_lock (mq);
 
       return rv < 0 ? errno : 0;
     }
diff --git a/src/svm/message_queue.h b/src/svm/message_queue.h
index 7716c6724d8..1ef773d9f0a 100644
--- a/src/svm/message_queue.h
+++ b/src/svm/message_queue.h
@@ -22,6 +22,7 @@
 
 #include <vppinfra/clib.h>
 #include <vppinfra/error.h>
+#include <vppinfra/lock.h>
 #include <svm/queue.h>
 
 typedef struct svm_msg_q_shr_queue_
@@ -41,6 +42,7 @@ typedef struct svm_msg_q_queue_
 {
   svm_msg_q_shared_queue_t *shr; /**< pointer to shared queue */
   int evtfd;			 /**< producer/consumer eventfd */
+  clib_spinlock_t lock;		 /**< private lock for multi-producer */
 } svm_msg_q_queue_t;
 
 typedef struct svm_msg_q_ring_shared_
@@ -99,6 +101,13 @@ typedef union
 } svm_msg_q_msg_t;
 
 #define SVM_MQ_INVALID_MSG { .as_u64 = ~0 }
+
+typedef enum svm_msg_q_wait_type_
+{
+  SVM_MQ_WAIT_EMPTY,
+  SVM_MQ_WAIT_FULL
+} svm_msg_q_wait_type_t;
+
 /**
  * Allocate message queue
  *
@@ -206,6 +215,7 @@ void svm_msg_q_add_and_unlock (svm_msg_q_t * mq, svm_msg_q_msg_t * msg);
  * Consumer dequeue one message from queue
  *
  * This returns the message pointing to the data in the message rings.
+ * Should only be used in single consumer scenarios as no locks are grabbed.
  * The consumer is expected to call @ref svm_msg_q_free_msg once it
  * finishes processing/copies the message data.
  *
@@ -219,18 +229,34 @@ int svm_msg_q_sub (svm_msg_q_t * mq, svm_msg_q_msg_t * msg,
 		   svm_q_conditional_wait_t cond, u32 time);
 
 /**
- * Consumer dequeue one message from queue with mutex held
+ * Consumer dequeue one message from queue
  *
- * Returns the message pointing to the data in the message rings under the
- * assumption that the message queue lock is already held. The consumer is
- * expected to call @ref svm_msg_q_free_msg once it finishes
+ * Returns the message pointing to the data in the message rings. Should only
+ * be used in single consumer scenarios as no locks are grabbed. The consumer
+ * is expected to call @ref svm_msg_q_free_msg once it finishes
  * processing/copies the message data.
  *
  * @param mq		message queue
  * @param msg		pointer to structure where message is to be received
  * @return		success status
  */
-void svm_msg_q_sub_w_lock (svm_msg_q_t * mq, svm_msg_q_msg_t * msg);
+int svm_msg_q_sub_raw (svm_msg_q_t *mq, svm_msg_q_msg_t *elem);
+
+/**
+ * Consumer dequeue multiple messages from queue
+ *
+ * Returns the message pointing to the data in the message rings. Should only
+ * be used in single consumer scenarios as no locks are grabbed. The consumer
+ * is expected to call @ref svm_msg_q_free_msg once it finishes
+ * processing/copies the message data.
+ *
+ * @param mq		message queue
+ * @param msg_buf	pointer to array of messages to received
+ * @param n_msgs	lengt of msg_buf array
+ * @return		number of messages dequeued
+ */
+int svm_msg_q_sub_raw_batch (svm_msg_q_t *mq, svm_msg_q_msg_t *msg_buf,
+			     u32 n_msgs);
 
 /**
  * Get data for message in queue
@@ -321,10 +347,17 @@ svm_msg_q_msg_is_invalid (svm_msg_q_msg_t * msg)
 static inline int
 svm_msg_q_try_lock (svm_msg_q_t * mq)
 {
-  int rv = pthread_mutex_trylock (&mq->q.shr->mutex);
-  if (PREDICT_FALSE (rv == EOWNERDEAD))
-    rv = pthread_mutex_consistent (&mq->q.shr->mutex);
-  return rv;
+  if (mq->q.evtfd == -1)
+    {
+      int rv = pthread_mutex_trylock (&mq->q.shr->mutex);
+      if (PREDICT_FALSE (rv == EOWNERDEAD))
+	rv = pthread_mutex_consistent (&mq->q.shr->mutex);
+      return rv;
+    }
+  else
+    {
+      return !clib_spinlock_trylock (&mq->q.lock);
+    }
 }
 
 /**
@@ -333,10 +366,18 @@ svm_msg_q_try_lock (svm_msg_q_t * mq)
 static inline int
 svm_msg_q_lock (svm_msg_q_t * mq)
 {
-  int rv = pthread_mutex_lock (&mq->q.shr->mutex);
-  if (PREDICT_FALSE (rv == EOWNERDEAD))
-    rv = pthread_mutex_consistent (&mq->q.shr->mutex);
-  return rv;
+  if (mq->q.evtfd == -1)
+    {
+      int rv = pthread_mutex_lock (&mq->q.shr->mutex);
+      if (PREDICT_FALSE (rv == EOWNERDEAD))
+	rv = pthread_mutex_consistent (&mq->q.shr->mutex);
+      return rv;
+    }
+  else
+    {
+      clib_spinlock_lock (&mq->q.lock);
+      return 0;
+    }
 }
 
 /**
@@ -345,7 +386,14 @@ svm_msg_q_lock (svm_msg_q_t * mq)
 static inline void
 svm_msg_q_unlock (svm_msg_q_t * mq)
 {
-  pthread_mutex_unlock (&mq->q.shr->mutex);
+  if (mq->q.evtfd == -1)
+    {
+      pthread_mutex_unlock (&mq->q.shr->mutex);
+    }
+  else
+    {
+      clib_spinlock_unlock (&mq->q.lock);
+    }
 }
 
 /**
@@ -354,7 +402,7 @@ svm_msg_q_unlock (svm_msg_q_t * mq)
  * Must be called with mutex held. The queue only works non-blocking
  * with eventfds, so handle blocking calls as an exception here.
  */
-void svm_msg_q_wait (svm_msg_q_t *mq);
+int svm_msg_q_wait (svm_msg_q_t *mq, svm_msg_q_wait_type_t type);
 
 /**
  * Timed wait for message queue event