svm: allow mq attachments at random offsets

Type: feature

Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ic373cd2c11272da539eb4b0db27227f36f2f9688

6 files changed, 198 insertions, 70 deletions

diff --git a/src/svm/fifo_segment.c b/src/svm/fifo_segment.c
index 636f223cde1..0c3a79a613b 100644
--- a/src/svm/fifo_segment.c
+++ b/src/svm/fifo_segment.c
@@ -764,9 +764,15 @@ void
 fifo_segment_cleanup (fifo_segment_t *fs)
 {
   int slice_index;
+  svm_msg_q_t *mq = 0;
 
   for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
     clib_mem_bulk_destroy (fs->slices[slice_index].fifos);
+
+  vec_foreach (fs->mqs, mq)
+    vec_free (mq->rings);
+
+  vec_free (fs->mqs);
 }
 
 /**
@@ -944,6 +950,69 @@ fifo_segment_attach_fifo (fifo_segment_t * fs, svm_fifo_t * f,
     }
 }
 
+svm_msg_q_t *
+fifo_segment_msg_q_alloc (fifo_segment_t *fs, u32 mq_index,
+			  svm_msg_q_cfg_t *cfg)
+{
+  fifo_segment_header_t *fsh = fs->h;
+  svm_msg_q_shared_t *smq;
+  svm_msg_q_t *mq;
+  void *base;
+  u32 size;
+
+  if (!fs->mqs)
+    {
+      u32 n_mqs = clib_max (fs->h->n_mqs, 1);
+      vec_validate (fs->mqs, n_mqs - 1);
+    }
+
+  size = svm_msg_q_size_to_alloc (cfg);
+  base = fsh_alloc_aligned (fsh, size, 8);
+  fsh->n_reserved_bytes += size;
+
+  smq = svm_msg_q_init (base, cfg);
+  mq = vec_elt_at_index (fs->mqs, mq_index);
+  svm_msg_q_attach (mq, smq);
+
+  return mq;
+}
+
+svm_msg_q_t *
+fifo_segment_msg_q_attach (fifo_segment_t *fs, uword offset, u32 mq_index)
+{
+  svm_msg_q_t *mq;
+
+  if (!fs->mqs)
+    {
+      u32 n_mqs = clib_max (fs->h->n_mqs, 1);
+      vec_validate (fs->mqs, n_mqs - 1);
+    }
+
+  mq = vec_elt_at_index (fs->mqs, mq_index);
+
+  if (!mq->q)
+    {
+      svm_msg_q_shared_t *smq;
+      smq = (svm_msg_q_shared_t *) ((u8 *) fs->h + offset);
+      svm_msg_q_attach (mq, smq);
+    }
+
+  ASSERT (fifo_segment_msg_q_offset (fs, mq_index) == offset);
+
+  return mq;
+}
+
+uword
+fifo_segment_msg_q_offset (fifo_segment_t *fs, u32 mq_index)
+{
+  svm_msg_q_t *mq = vec_elt_at_index (fs->mqs, mq_index);
+
+  if (mq->q == 0)
+    return ~0ULL;
+
+  return (uword) ((u8 *) mq->q - (u8 *) fs->h) - sizeof (svm_msg_q_shared_t);
+}
+
 int
 fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 slice_index,
 				 u32 batch_size)
diff --git a/src/svm/fifo_segment.h b/src/svm/fifo_segment.h
index 006ffc40f63..195869a25dd 100644
--- a/src/svm/fifo_segment.h
+++ b/src/svm/fifo_segment.h
@@ -17,6 +17,7 @@
 
 #include <svm/ssvm.h>
 #include <svm/fifo_types.h>
+#include <svm/message_queue.h>
 #include <svm/svm_fifo.h>
 
 typedef enum
@@ -70,6 +71,7 @@ typedef struct
   uword max_byte_index;
   u8 n_slices;			/**< number of fifo segment slices */
   fifo_slice_private_t *slices; /**< private slice information */
+  svm_msg_q_t *mqs;		/**< private vec of attached mqs */
 } fifo_segment_t;
 
 typedef struct
@@ -130,6 +132,37 @@ void fifo_segment_attach_fifo (fifo_segment_t * fs, svm_fifo_t * f,
 			       u32 slice_index);
 
 /**
+ * Allocate message queue on segment
+ *
+ * @param fs		fifo segment for mq
+ * @param mq_index	index in private mqs vector to use to attach
+ * @param cfg		configuration for mq
+ * @return		attached message queue
+ */
+svm_msg_q_t *fifo_segment_msg_q_alloc (fifo_segment_t *fs, u32 mq_index,
+				       svm_msg_q_cfg_t *cfg);
+
+/**
+ *  Attach message queue at fifo segment offset
+ *
+ *  @param fs		fifo segment for mq
+ *  @param offset	offset for shared mq on the segment
+ *  @param mq_index	index in private mqs vector to use to attach
+ *  @return		attached message queue
+ */
+svm_msg_q_t *fifo_segment_msg_q_attach (fifo_segment_t *fs, uword offset,
+					u32 mq_index);
+
+/**
+ * Message queue offset on segment
+ *
+ * @param fs		fifo segment for mq
+ * @param mq_index	index of mq in private mqs vector
+ * @return		offset of the shared mq the private mq is attached to
+ */
+uword fifo_segment_msg_q_offset (fifo_segment_t *fs, u32 mq_index);
+
+/**
  * Try to preallocate fifo headers
  *
  * Tries to preallocate fifo headers and adds them to freelist.
diff --git a/src/svm/fifo_types.h b/src/svm/fifo_types.h
index 85e67bb039c..bfd1a41a17b 100644
--- a/src/svm/fifo_types.h
+++ b/src/svm/fifo_types.h
@@ -146,6 +146,7 @@ struct fifo_segment_header_
   u8 high_watermark;			/**< Memory pressure watermark high */
   u8 low_watermark;			/**< Memory pressure watermark low */
   u8 pct_first_alloc;			/**< Pct of fifo size to alloc */
+  u8 n_mqs;				/**< Num mqs for mqs segment */
   CLIB_CACHE_LINE_ALIGN_MARK (allocator);
   uword byte_index;
   uword max_byte_index;
diff --git a/src/svm/message_queue.c b/src/svm/message_queue.c
index e586841cdb1..0ebce702430 100644
--- a/src/svm/message_queue.c
+++ b/src/svm/message_queue.c
@@ -34,45 +34,34 @@ static inline void *
 svm_msg_q_ring_data (svm_msg_q_ring_t * ring, u32 elt_index)
 {
   ASSERT (elt_index < ring->nitems);
-  return (ring->data + elt_index * ring->elsize);
+  return (ring->shr->data + elt_index * ring->elsize);
 }
 
-svm_msg_q_t *
+svm_msg_q_shared_t *
 svm_msg_q_init (void *base, svm_msg_q_cfg_t *cfg)
 {
-  svm_msg_q_ring_t *ring;
-  vec_header_t *vh;
-  svm_msg_q_t *mq;
-  u8 *rings_ptr;
-  u32 q_sz;
+  svm_msg_q_ring_shared_t *ring;
+  svm_msg_q_shared_t *smq;
+  u32 q_sz, offset;
   int i;
 
   q_sz = sizeof (svm_queue_t) + cfg->q_nitems * sizeof (svm_msg_q_msg_t);
 
-  mq = (svm_msg_q_t *) base;
-  mq->q = svm_queue_init (base + sizeof (svm_msg_q_t), cfg->q_nitems,
-			  sizeof (svm_msg_q_msg_t));
-  mq->q->consumer_pid = cfg->consumer_pid;
-  vh = (vec_header_t *) ((u8 *) mq->q + q_sz);
-  vh->len = cfg->n_rings;
-  mq->rings = (svm_msg_q_ring_t *) (vh + 1);
-  rings_ptr = (u8 *) mq->rings + sizeof (svm_msg_q_ring_t) * cfg->n_rings;
+  smq = (svm_msg_q_shared_t *) base;
+  svm_queue_init (&smq->q, cfg->q_nitems, sizeof (svm_msg_q_msg_t));
+  smq->q->consumer_pid = cfg->consumer_pid;
+  smq->n_rings = cfg->n_rings;
+  ring = (void *) ((u8 *) smq->q + q_sz);
   for (i = 0; i < cfg->n_rings; i++)
     {
-      ring = &mq->rings[i];
       ring->elsize = cfg->ring_cfgs[i].elsize;
       ring->nitems = cfg->ring_cfgs[i].nitems;
       ring->cursize = ring->head = ring->tail = 0;
-      if (cfg->ring_cfgs[i].data)
-	ring->data = cfg->ring_cfgs[i].data;
-      else
-	{
-	  ring->data = rings_ptr;
-	  rings_ptr += (uword) ring->nitems * ring->elsize;
-	}
+      offset = sizeof (*ring) + ring->nitems * ring->elsize;
+      ring = (void *) ((u8 *) ring + offset);
     }
 
-  return mq;
+  return smq;
 }
 
 uword
@@ -80,12 +69,12 @@ svm_msg_q_size_to_alloc (svm_msg_q_cfg_t *cfg)
 {
   svm_msg_q_ring_cfg_t *ring_cfg;
   uword rings_sz = 0, mq_sz;
-  u32 vec_sz, q_sz;
+  u32 q_sz;
   int i;
 
   ASSERT (cfg);
 
-  vec_sz = vec_header_bytes (0) + sizeof (svm_msg_q_ring_t) * cfg->n_rings;
+  rings_sz = sizeof (svm_msg_q_ring_shared_t) * cfg->n_rings;
   for (i = 0; i < cfg->n_rings; i++)
     {
       if (cfg->ring_cfgs[i].data)
@@ -95,33 +84,18 @@ svm_msg_q_size_to_alloc (svm_msg_q_cfg_t *cfg)
     }
 
   q_sz = sizeof (svm_queue_t) + cfg->q_nitems * sizeof (svm_msg_q_msg_t);
-  mq_sz = sizeof (svm_msg_q_t) + vec_sz + rings_sz + q_sz;
+  mq_sz = sizeof (svm_msg_q_shared_t) + q_sz + rings_sz;
 
   return mq_sz;
 }
 
-svm_msg_q_t *
+svm_msg_q_shared_t *
 svm_msg_q_alloc (svm_msg_q_cfg_t *cfg)
 {
-  svm_msg_q_ring_cfg_t *ring_cfg;
-  uword rings_sz = 0, mq_sz;
-  u32 vec_sz, q_sz;
+  uword mq_sz;
   u8 *base;
-  int i;
-
-  ASSERT (cfg);
 
-  vec_sz = vec_header_bytes (0) + sizeof (svm_msg_q_ring_t) * cfg->n_rings;
-  for (i = 0; i < cfg->n_rings; i++)
-    {
-      if (cfg->ring_cfgs[i].data)
-	continue;
-      ring_cfg = &cfg->ring_cfgs[i];
-      rings_sz += (uword) ring_cfg->nitems * ring_cfg->elsize;
-    }
-
-  q_sz = sizeof (svm_queue_t) + cfg->q_nitems * sizeof (svm_msg_q_msg_t);
-  mq_sz = sizeof (svm_msg_q_t) + vec_sz + rings_sz + q_sz;
+  mq_sz = svm_msg_q_size_to_alloc (cfg);
   base = clib_mem_alloc_aligned (mq_sz, CLIB_CACHE_LINE_BYTES);
   if (!base)
     return 0;
@@ -130,6 +104,29 @@ svm_msg_q_alloc (svm_msg_q_cfg_t *cfg)
 }
 
 void
+svm_msg_q_attach (svm_msg_q_t *mq, void *smq_base)
+{
+  svm_msg_q_ring_shared_t *ring;
+  svm_msg_q_shared_t *smq;
+  u32 i, n_rings, q_sz, offset;
+
+  smq = (svm_msg_q_shared_t *) smq_base;
+  mq->q = smq->q;
+  n_rings = smq->n_rings;
+  vec_validate (mq->rings, n_rings - 1);
+  q_sz = sizeof (svm_queue_t) + mq->q->maxsize * sizeof (svm_msg_q_msg_t);
+  ring = (void *) ((u8 *) smq->q + q_sz);
+  for (i = 0; i < n_rings; i++)
+    {
+      mq->rings[i].nitems = ring->nitems;
+      mq->rings[i].elsize = ring->elsize;
+      mq->rings[i].shr = ring;
+      offset = sizeof (*ring) + ring->nitems * ring->elsize;
+      ring = (void *) ((u8 *) ring + offset);
+    }
+}
+
+void
 svm_msg_q_free (svm_msg_q_t * mq)
 {
   svm_queue_free (mq->q);
@@ -139,14 +136,18 @@ svm_msg_q_free (svm_msg_q_t * mq)
 svm_msg_q_msg_t
 svm_msg_q_alloc_msg_w_ring (svm_msg_q_t * mq, u32 ring_index)
 {
+  svm_msg_q_ring_shared_t *sr;
+  svm_msg_q_ring_t *ring;
   svm_msg_q_msg_t msg;
-  svm_msg_q_ring_t *ring = svm_msg_q_ring_inline (mq, ring_index);
 
-  ASSERT (ring->cursize < ring->nitems);
+  ring = svm_msg_q_ring_inline (mq, ring_index);
+  sr = ring->shr;
+
+  ASSERT (sr->cursize < ring->nitems);
   msg.ring_index = ring - mq->rings;
-  msg.elt_index = ring->tail;
-  ring->tail = (ring->tail + 1) % ring->nitems;
-  clib_atomic_fetch_add (&ring->cursize, 1);
+  msg.elt_index = sr->tail;
+  sr->tail = (sr->tail + 1) % ring->nitems;
+  clib_atomic_fetch_add (&sr->cursize, 1);
   return msg;
 }
 
@@ -181,16 +182,18 @@ svm_msg_q_msg_t
 svm_msg_q_alloc_msg (svm_msg_q_t * mq, u32 nbytes)
 {
   svm_msg_q_msg_t msg = {.as_u64 = ~0 };
+  svm_msg_q_ring_shared_t *sr;
   svm_msg_q_ring_t *ring;
 
   vec_foreach (ring, mq->rings)
   {
-    if (ring->elsize < nbytes || ring->cursize == ring->nitems)
+    sr = ring->shr;
+    if (ring->elsize < nbytes || sr->cursize == ring->nitems)
       continue;
     msg.ring_index = ring - mq->rings;
-    msg.elt_index = ring->tail;
-    ring->tail = (ring->tail + 1) % ring->nitems;
-    clib_atomic_fetch_add (&ring->cursize, 1);
+    msg.elt_index = sr->tail;
+    sr->tail = (sr->tail + 1) % ring->nitems;
+    clib_atomic_fetch_add (&sr->cursize, 1);
     break;
   }
   return msg;
@@ -206,14 +209,16 @@ svm_msg_q_msg_data (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
 void
 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;
 
   ASSERT (vec_len (mq->rings) > msg->ring_index);
-  ring = &mq->rings[msg->ring_index];
-  if (msg->elt_index == ring->head)
+  ring = svm_msg_q_ring_inline (mq, msg->ring_index);
+  sr = ring->shr;
+  if (msg->elt_index == sr->head)
     {
-      ring->head = (ring->head + 1) % ring->nitems;
+      sr->head = (sr->head + 1) % ring->nitems;
     }
   else
     {
@@ -222,8 +227,8 @@ svm_msg_q_free_msg (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
       ASSERT (0);
     }
 
-  need_signal = ring->cursize == ring->nitems;
-  clib_atomic_fetch_sub (&ring->cursize, 1);
+  need_signal = sr->cursize == ring->nitems;
+  clib_atomic_fetch_sub (&sr->cursize, 1);
 
   if (PREDICT_FALSE (need_signal))
     svm_queue_send_signal (mq->q, 0);
@@ -233,17 +238,20 @@ static int
 svm_msq_q_msg_is_valid (svm_msg_q_t * mq, svm_msg_q_msg_t * msg)
 {
   u32 dist1, dist2, tail, head;
+  svm_msg_q_ring_shared_t *sr;
   svm_msg_q_ring_t *ring;
 
   if (vec_len (mq->rings) <= msg->ring_index)
     return 0;
-  ring = &mq->rings[msg->ring_index];
-  tail = ring->tail;
-  head = ring->head;
+
+  ring = svm_msg_q_ring_inline (mq, msg->ring_index);
+  sr = ring->shr;
+  tail = sr->tail;
+  head = sr->head;
 
   dist1 = ((ring->nitems + msg->elt_index) - head) % ring->nitems;
   if (tail == head)
-    dist2 = (ring->cursize == 0) ? 0 : ring->nitems;
+    dist2 = (sr->cursize == 0) ? 0 : ring->nitems;
   else
     dist2 = ((ring->nitems + tail) - head) % ring->nitems;
   return (dist1 < dist2);
@@ -316,7 +324,7 @@ format_svm_msg_q (u8 * s, va_list * args)
   s = format (s, " [Q:%d/%d]", mq->q->cursize, mq->q->maxsize);
   for (u32 i = 0; i < vec_len (mq->rings); i++)
     {
-      s = format (s, " [R%d:%d/%d]", i, mq->rings[i].cursize,
+      s = format (s, " [R%d:%d/%d]", i, mq->rings[i].shr->cursize,
 		  mq->rings[i].nitems);
     }
   return s;
diff --git a/src/svm/message_queue.h b/src/svm/message_queue.h
index 50f79fb0c19..4b314b837e9 100644
--- a/src/svm/message_queue.h
+++ b/src/svm/message_queue.h
@@ -24,16 +24,30 @@
 #include <vppinfra/error.h>
 #include <svm/queue.h>
 
-typedef struct svm_msg_q_ring_
+typedef struct svm_msg_q_ring_shared_
 {
   volatile u32 cursize;			/**< current size of the ring */
   u32 nitems;				/**< max size of the ring */
   volatile u32 head;			/**< current head (for dequeue) */
   volatile u32 tail;			/**< current tail (for enqueue) */
   u32 elsize;				/**< size of an element */
-  u8 *data;				/**< chunk of memory for msg data */
+  u8 data[0];				/**< chunk of memory for msg data */
+} svm_msg_q_ring_shared_t;
+
+typedef struct svm_msg_q_ring_
+{
+  u32 nitems;			/**< max size of the ring */
+  u32 elsize;			/**< size of an element */
+  svm_msg_q_ring_shared_t *shr; /**< ring in shared memory */
 } __clib_packed svm_msg_q_ring_t;
 
+typedef struct svm_msg_q_shared_
+{
+  u32 n_rings;	    /**< number of rings after q */
+  u32 pad;	    /**< 8 byte alignment for q */
+  svm_queue_t q[0]; /**< queue for exchanging messages */
+} __clib_packed svm_msg_q_shared_t;
+
 typedef struct svm_msg_q_
 {
   svm_queue_t *q;			/**< queue for exchanging messages */
@@ -77,10 +91,12 @@ typedef union
  * 			ring configs
  * @return		message queue
  */
-svm_msg_q_t *svm_msg_q_alloc (svm_msg_q_cfg_t * cfg);
-svm_msg_q_t *svm_msg_q_init (void *base, svm_msg_q_cfg_t *cfg);
+svm_msg_q_shared_t *svm_msg_q_alloc (svm_msg_q_cfg_t *cfg);
+svm_msg_q_shared_t *svm_msg_q_init (void *base, svm_msg_q_cfg_t *cfg);
 uword svm_msg_q_size_to_alloc (svm_msg_q_cfg_t *cfg);
 
+void svm_msg_q_attach (svm_msg_q_t *mq, void *smq_base);
+
 /**
  * Free message queue
  *
@@ -267,8 +283,8 @@ svm_msg_q_is_full (svm_msg_q_t * mq)
 static inline u8
 svm_msg_q_ring_is_full (svm_msg_q_t * mq, u32 ring_index)
 {
-  ASSERT (ring_index < vec_len (mq->rings));
-  return (mq->rings[ring_index].cursize == mq->rings[ring_index].nitems);
+  svm_msg_q_ring_t *ring = vec_elt_at_index (mq->rings, ring_index);
+  return (ring->shr->cursize >= ring->nitems);
 }
 
 /**
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index f79f37a4687..14729710aac 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -398,6 +398,7 @@ svm_fifo_init (svm_fifo_t * f, u32 size)
     {
       c->start_byte = prev->start_byte + prev->length;
       c->enq_rb_index = c->deq_rb_index = RBTREE_TNIL_INDEX;
+      ASSERT (c->length >= 1 << FS_MIN_LOG2_CHUNK_SZ);
       prev = c;
       c = f_cptr (f, c->next);
     }