Horizontal (nSessions) scaling draft

- Data structure preallocation.
- Input state machine fixes for mid-stream 3-way handshake retries.
- Batch connections in the builtin_client
- Multiple private fifo segment support
- Fix elog simultaneous event type registration
- Fix sacks when segment hole is added after highest sacked
- Add "accepting" session state for sessions pending accept
- Add ssvm non-recursive locking
- Estimate RTT for syn-ack
- Don't init fifo pointers. We're using relative offsets for ooo
  segments
- CLI to dump individual session

Change-Id: Ie0598563fd246537bafba4feed7985478ea1d415
Signed-off-by: Dave Barach <dbarach@cisco.com>
Signed-off-by: Florin Coras <fcoras@cisco.com>

6 files changed, 158 insertions, 59 deletions

diff --git a/src/svm/ssvm.h b/src/svm/ssvm.h
index bccfc164bef..8466e15579b 100644
--- a/src/svm/ssvm.h
+++ b/src/svm/ssvm.h
@@ -102,6 +102,15 @@ ssvm_lock (ssvm_shared_header_t * h, u32 my_pid, u32 tag)
 }
 
 always_inline void
+ssvm_lock_non_recursive (ssvm_shared_header_t * h, u32 tag)
+{
+  while (__sync_lock_test_and_set (&h->lock, 1))
+    ;
+
+  h->tag = tag;
+}
+
+always_inline void
 ssvm_unlock (ssvm_shared_header_t * h)
 {
   if (--h->recursion_count == 0)
@@ -113,6 +122,14 @@ ssvm_unlock (ssvm_shared_header_t * h)
     }
 }
 
+always_inline void
+ssvm_unlock_non_recursive (ssvm_shared_header_t * h)
+{
+  h->tag = 0;
+  CLIB_MEMORY_BARRIER ();
+  h->lock = 0;
+}
+
 static inline void *
 ssvm_push_heap (ssvm_shared_header_t * sh)
 {
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index aed5d6a7420..da60fee51ec 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -19,29 +19,29 @@
 static inline u8
 position_lt (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  < ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  < ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u8
 position_leq (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  <= ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  <= ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u8
 position_gt (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  > ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  > ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u32
 position_diff (svm_fifo_t * f, u32 posa, u32 posb)
 {
-  return ooo_segment_distance_to_tail (f, posa)
-    - ooo_segment_distance_to_tail (f, posb);
+  return ooo_segment_distance_from_tail (f, posa)
+    - ooo_segment_distance_from_tail (f, posb);
 }
 
 static inline u32
@@ -113,7 +113,7 @@ svm_fifo_create (u32 data_size_in_bytes)
   if (f == 0)
     return 0;
 
-  memset (f, 0, sizeof (*f) + data_size_in_bytes);
+  memset (f, 0, sizeof (*f));
   f->nitems = data_size_in_bytes;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
 
@@ -204,7 +204,19 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
     {
       s = prev;
       s_end_pos = ooo_segment_end_pos (f, s);
-      goto merge;
+
+      /* Check head and tail now since segment may be wider at both ends so
+       * merge tests lower won't work */
+      if (position_lt (f, normalized_position, s->start))
+	{
+	  s->start = normalized_position;
+	  s->length = position_diff (f, s_end_pos, s->start);
+	}
+      if (position_gt (f, normalized_end_position, s_end_pos))
+	{
+	  s->length = position_diff (f, normalized_end_position, s->start);
+	}
+      goto check_tail;
     }
 
   s_index = s - f->ooo_segments;
@@ -257,8 +269,6 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
    * Merge needed
    */
 
-merge:
-
   /* Merge at head */
   if (position_lt (f, normalized_position, s->start))
     {
@@ -278,6 +288,7 @@ merge:
       goto done;
     }
 
+check_tail:
   /* The new segment's tail may cover multiple smaller ones */
   if (position_gt (f, normalized_end_position, s_end_pos))
     {
@@ -296,7 +307,8 @@ merge:
       /* If partial overlap with last, merge */
       if (it && position_leq (f, it->start, normalized_end_position))
 	{
-	  s->length = ooo_segment_end_pos (f, it) - s->start;
+	  s->length =
+	    position_diff (f, ooo_segment_end_pos (f, it), s->start);
 	  ooo_segment_del (f, it - f->ooo_segments);
 	}
     }
@@ -319,9 +331,9 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   i32 diff;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
+  diff = ooo_segment_distance_to_tail (f, s->start);
 
-  diff = (f->tail >= s->start) ?
-    f->tail - s->start : f->nitems + f->tail - s->start;
+  ASSERT (diff != n_bytes_enqueued);
 
   if (diff > n_bytes_enqueued)
     return 0;
@@ -345,8 +357,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
       if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  diff = (f->tail >= s->start) ?
-	    f->tail - s->start : f->nitems + f->tail - s->start;
+	  diff = ooo_segment_distance_to_tail (f, s->start);
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -357,6 +368,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 	}
     }
 
+  ASSERT (bytes >= 0 && bytes <= f->nitems);
   return bytes;
 }
 
@@ -401,6 +413,8 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
     }
   else
     {
+      ASSERT (0);
+
       /* Account for a zero-copy enqueue done elsewhere */
       ASSERT (max_bytes <= (nitems - cursize));
       f->tail += max_bytes;
@@ -413,6 +427,7 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
     total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
 
   /* Atomically increase the queue length */
+  ASSERT (cursize + total_copy_bytes <= nitems);
   __sync_fetch_and_add (&f->cursize, total_copy_bytes);
 
   return (total_copy_bytes);
@@ -475,6 +490,8 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   cursize = svm_fifo_max_dequeue (f);
   nitems = f->nitems;
 
+  ASSERT (required_bytes < nitems);
+
   normalized_offset = (f->tail + offset) % nitems;
 
   /* Will this request fit? */
@@ -557,6 +574,7 @@ svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
     }
   else
     {
+      ASSERT (0);
       /* Account for a zero-copy dequeue done elsewhere */
       ASSERT (max_bytes <= cursize);
       f->head += max_bytes;
@@ -565,6 +583,8 @@ svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
       total_copy_bytes = max_bytes;
     }
 
+  ASSERT (f->head <= nitems);
+  ASSERT (cursize >= total_copy_bytes);
   __sync_fetch_and_sub (&f->cursize, total_copy_bytes);
 
   return (total_copy_bytes);
@@ -702,6 +722,8 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
       f->head = (f->head == nitems) ? 0 : f->head;
     }
 
+  ASSERT (f->head <= nitems);
+  ASSERT (cursize >= total_drop_bytes);
   __sync_fetch_and_sub (&f->cursize, total_drop_bytes);
 
   return total_drop_bytes;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index f32ef41da7d..fe21de476d7 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -133,25 +133,31 @@ svm_fifo_newest_ooo_segment (svm_fifo_t * f)
 }
 
 always_inline u32
-ooo_segment_distance_to_tail (svm_fifo_t * f, u32 a)
+ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
 {
   /* Ambiguous. Assumption is that ooo segments don't touch tail */
-  if (a == f->tail && f->tail == f->head)
+  if (PREDICT_FALSE (pos == f->tail && f->tail == f->head))
     return f->nitems;
 
-  return ((f->nitems + a - f->tail) % f->nitems);
+  return (((f->nitems + pos) - f->tail) % f->nitems);
+}
+
+always_inline u32
+ooo_segment_distance_to_tail (svm_fifo_t * f, u32 pos)
+{
+  return (((f->nitems + f->tail) - pos) % f->nitems);
 }
 
 always_inline u32
 ooo_segment_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ooo_segment_distance_to_tail (f, s->start);
+  return ooo_segment_distance_from_tail (f, s->start);
 }
 
 always_inline u32
 ooo_segment_end_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ooo_segment_distance_to_tail (f, s->start) + s->length;
+  return ooo_segment_distance_from_tail (f, s->start) + s->length;
 }
 
 always_inline u32
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index c4ac235252c..69d4ecb9387 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -35,6 +35,11 @@ preallocate_fifo_pairs (svm_fifo_segment_header_t * fsh,
   rx_fifo_size = (sizeof (*f) + a->rx_fifo_size) * a->preallocated_fifo_pairs;
   tx_fifo_size = (sizeof (*f) + a->tx_fifo_size) * a->preallocated_fifo_pairs;
 
+  if (0)
+    clib_warning ("rx_fifo_size %u (%d mb), tx_fifo_size %u (%d mb)",
+		  rx_fifo_size, rx_fifo_size >> 20,
+		  tx_fifo_size, tx_fifo_size >> 20);
+
   /* Allocate rx fifo space. May fail. */
   rx_fifo_space = clib_mem_alloc_aligned_at_offset
     (rx_fifo_size, CLIB_CACHE_LINE_BYTES, 0 /* align_offset */ ,
@@ -129,7 +134,7 @@ svm_fifo_segment_create (svm_fifo_segment_create_args_t * a)
   ssvm_pop_heap (oldheap);
 
   sh->ready = 1;
-  a->new_segment_index = s - sm->segments;
+  vec_add1 (a->new_segment_indices, s - sm->segments);
   return (0);
 }
 
@@ -141,35 +146,81 @@ svm_fifo_segment_create_process_private (svm_fifo_segment_create_args_t * a)
   svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
   ssvm_shared_header_t *sh;
   svm_fifo_segment_header_t *fsh;
+  void *oldheap;
+  u8 **heaps = 0;
+  mheap_t *heap_header;
+  int segment_count = 1;
+  int i;
 
-  /* Allocate a fresh segment */
-  pool_get (sm->segments, s);
-  memset (s, 0, sizeof (*s));
-
-  s->ssvm.ssvm_size = ~0;
-  s->ssvm.i_am_master = 1;
-  s->ssvm.my_pid = getpid ();
-  s->ssvm.name = (u8 *) a->segment_name;
-  s->ssvm.requested_va = ~0;
-
-  /* Allocate a [sic] shared memory header, in process memory... */
-  sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
-  s->ssvm.sh = sh;
+  if (a->private_segment_count && a->private_segment_size)
+    {
+      void *mem;
+      u8 *heap;
+      u32 pagesize = clib_mem_get_page_size ();
+      u32 rnd_size;
 
-  memset (sh, 0, sizeof (*sh));
-  sh->heap = clib_mem_get_heap ();
+      for (i = 0; i < a->private_segment_count; i++)
+	{
+	  rnd_size = (a->private_segment_size + (pagesize - 1)) & ~pagesize;
+
+	  mem = mmap (0, rnd_size, PROT_READ | PROT_WRITE,
+		      MAP_PRIVATE | MAP_ANONYMOUS,
+		      -1 /* fd */ , 0 /* offset */ );
+
+	  if (mem == MAP_FAILED)
+	    {
+	      clib_unix_warning ("mmap");
+	      return -1;
+	    }
+	  heap = mheap_alloc (mem, rnd_size);
+	  heap_header = mheap_header (heap);
+	  heap_header->flags |= MHEAP_FLAG_THREAD_SAFE;
+	  vec_add1 (heaps, heap);
+	}
+      segment_count = a->private_segment_count;
+    }
 
-  /* Set up svm_fifo_segment shared header */
-  fsh = clib_mem_alloc (sizeof (*fsh));
-  memset (fsh, 0, sizeof (*fsh));
-  sh->opaque[0] = fsh;
-  s->h = fsh;
-  fsh->segment_name = format (0, "%s%c", a->segment_name, 0);
+  /* Spread preallocated fifo pairs across segments */
+  a->preallocated_fifo_pairs /= segment_count;
 
-  preallocate_fifo_pairs (fsh, a);
+  /* Allocate segments */
+  for (i = 0; i < segment_count; i++)
+    {
+      pool_get (sm->segments, s);
+      memset (s, 0, sizeof (*s));
+
+      s->ssvm.ssvm_size = ~0;
+      s->ssvm.i_am_master = 1;
+      s->ssvm.my_pid = getpid ();
+      s->ssvm.name = (u8 *) a->segment_name;
+      s->ssvm.requested_va = ~0;
+
+      /* Allocate a [sic] shared memory header, in process memory... */
+      sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
+      s->ssvm.sh = sh;
+
+      memset (sh, 0, sizeof (*sh));
+      sh->heap = a->private_segment_count ? heaps[i] : clib_mem_get_heap ();
+
+      /* Set up svm_fifo_segment shared header */
+      fsh = clib_mem_alloc (sizeof (*fsh));
+      memset (fsh, 0, sizeof (*fsh));
+      sh->opaque[0] = fsh;
+      s->h = fsh;
+      fsh->segment_name = format (0, "%s%c", a->segment_name, 0);
+
+      if (a->private_segment_count)
+	{
+	  oldheap = clib_mem_get_heap ();
+	  clib_mem_set_heap (sh->heap);
+	  preallocate_fifo_pairs (fsh, a);
+	  clib_mem_set_heap (oldheap);
+	}
 
-  sh->ready = 1;
-  a->new_segment_index = s - sm->segments;
+      sh->ready = 1;
+      vec_add1 (a->new_segment_indices, s - sm->segments);
+    }
+  vec_free (heaps);
   return (0);
 }
 
@@ -205,7 +256,7 @@ svm_fifo_segment_attach (svm_fifo_segment_create_args_t * a)
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
   s->h = fsh;
 
-  a->new_segment_index = s - sm->segments;
+  vec_add1 (a->new_segment_indices, s - sm->segments);
   return (0);
 }
 
@@ -230,7 +281,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
-  ssvm_lock (sh, 1, 0);
+  ssvm_lock_non_recursive (sh, 1);
   oldheap = ssvm_push_heap (sh);
 
   switch (list_index)
@@ -261,7 +312,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
   if (PREDICT_FALSE (f == 0))
     {
       ssvm_pop_heap (oldheap);
-      ssvm_unlock (sh);
+      ssvm_unlock_non_recursive (sh);
       return (0);
     }
 
@@ -281,7 +332,7 @@ found:
     }
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock (sh);
+  ssvm_unlock_non_recursive (sh);
   return (f);
 }
 
@@ -293,10 +344,11 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
   svm_fifo_segment_header_t *fsh;
   void *oldheap;
 
+
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
-  ssvm_lock (sh, 1, 0);
+  ssvm_lock_non_recursive (sh, 2);
   oldheap = ssvm_push_heap (sh);
 
   switch (list_index)
@@ -325,7 +377,7 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
     }
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock (sh);
+  ssvm_unlock_non_recursive (sh);
 }
 
 void
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index 31e14db50f6..a7a3f469bab 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -57,10 +57,12 @@ typedef struct
 {
   char *segment_name;
   u32 segment_size;
-  u32 new_segment_index;
+  u32 *new_segment_indices;
   u32 rx_fifo_size;
   u32 tx_fifo_size;
   u32 preallocated_fifo_pairs;
+  u32 private_segment_count;
+  u32 private_segment_size;
 } svm_fifo_segment_create_args_t;
 
 static inline svm_fifo_segment_private_t *
diff --git a/src/svm/test_svm_fifo1.c b/src/svm/test_svm_fifo1.c
index 63b4a9b7ba5..63d7584532e 100644
--- a/src/svm/test_svm_fifo1.c
+++ b/src/svm/test_svm_fifo1.c
@@ -39,7 +39,7 @@ hello_world (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 4096, FIFO_SEGMENT_RX_FREELIST);
 
@@ -92,7 +92,7 @@ master (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 4096, FIFO_SEGMENT_RX_FREELIST);
 
@@ -128,7 +128,7 @@ mempig (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   for (i = 0; i < 1000; i++)
     {
@@ -186,7 +186,7 @@ offset (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 200 << 10, FIFO_SEGMENT_RX_FREELIST);
 
@@ -246,7 +246,7 @@ slave (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_attach returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
   sh = sp->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];