svm_fifo rework to avoid contention on cursize

Problems Addressed:
- Contention of cursize by producer and consumer.
- Reduce the no of modulo operations.

Changes:
- Synchronization between producer and consumer changed from cursize
  to head and tail indexes
  Implications: reduces the usable size of fifo by 1.
- Using weaker memory ordering C++11 atomics to access head and tail
  based on producer and consumer role.
- Head and tail indexes are unsigned 32 bit integers. Additions and
  subtraction on them are implicit 32 bit Modulo operation.
- Adding weaker memory ordering variants of max_enq, max_deq, is_empty
  and is_full Using them appropriately in all places.

Perfomance improvement (iperf3 via Hoststack):

iperf3 Server: Marvell ThunderX2(AArch64) - iperf3 Client: Skylake(x86)
   ~6%(256 rxd/txd) - ~11%(2048 rxd/txd)

Change-Id: I1d484e000e437430fdd5a819657d1c6b62443018
Signed-off-by: Sirshak Das <sirshak.das@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>

3 files changed, 393 insertions, 244 deletions

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 7aecbb81619..918ad77956e 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -1,5 +1,9 @@
 /*
  * Copyright (c) 2016-2019 Cisco and/or its affiliates.
+ * Copyright (c) 2019 Arm Limited
+ * Copyright (c) 2010-2017 Intel Corporation and/or its affiliates.
+ * Copyright (c) 2007-2009 Kip Macy kmacy@freebsd.org
+ * Inspired from DPDK rte_ring.h (SPSC only) (derived from freebsd bufring.h).
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at:
@@ -17,37 +21,37 @@
 #include <vppinfra/cpu.h>
 
 static inline u8
-position_lt (svm_fifo_t * f, u32 a, u32 b)
+position_lt (svm_fifo_t * f, u32 a, u32 b, u32 tail)
 {
-  return (ooo_segment_distance_from_tail (f, a)
-	  < ooo_segment_distance_from_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a, tail)
+	  < ooo_segment_distance_from_tail (f, b, tail));
 }
 
 static inline u8
-position_leq (svm_fifo_t * f, u32 a, u32 b)
+position_leq (svm_fifo_t * f, u32 a, u32 b, u32 tail)
 {
-  return (ooo_segment_distance_from_tail (f, a)
-	  <= ooo_segment_distance_from_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a, tail)
+	  <= ooo_segment_distance_from_tail (f, b, tail));
 }
 
 static inline u8
-position_gt (svm_fifo_t * f, u32 a, u32 b)
+position_gt (svm_fifo_t * f, u32 a, u32 b, u32 tail)
 {
-  return (ooo_segment_distance_from_tail (f, a)
-	  > ooo_segment_distance_from_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a, tail)
+	  > ooo_segment_distance_from_tail (f, b, tail));
 }
 
 static inline u32
-position_diff (svm_fifo_t * f, u32 posa, u32 posb)
+position_diff (svm_fifo_t * f, u32 posa, u32 posb, u32 tail)
 {
-  return ooo_segment_distance_from_tail (f, posa)
-    - ooo_segment_distance_from_tail (f, posb);
+  return ooo_segment_distance_from_tail (f, posa, tail)
+    - ooo_segment_distance_from_tail (f, posb, tail);
 }
 
 static inline u32
 ooo_segment_end_pos (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return (s->start + s->length) % f->nitems;
+  return s->start + s->length;
 }
 
 #ifndef CLIB_MARCH_VARIANT
@@ -57,9 +61,9 @@ format_ooo_segment (u8 * s, va_list * args)
 {
   svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
   ooo_segment_t *seg = va_arg (*args, ooo_segment_t *);
-  u32 normalized_start = (seg->start + f->nitems - f->tail) % f->nitems;
+  u32 normalized_start = (seg->start + f->nitems - f->tail) % f->size;
   s = format (s, "[%u, %u], len %u, next %d, prev %d", normalized_start,
-	      (normalized_start + seg->length) % f->nitems, seg->length,
+	      (normalized_start + seg->length) % f->size, seg->length,
 	      seg->next, seg->prev);
   return s;
 }
@@ -108,7 +112,7 @@ svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose)
   trace_len = 0;
 #endif
 
-  dummy_fifo = svm_fifo_create (f->nitems);
+  dummy_fifo = svm_fifo_create (f->size);
   clib_memset (f->data, 0xFF, f->nitems);
 
   vec_validate (data, f->nitems);
@@ -122,8 +126,7 @@ svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose)
 	{
 	  if (verbose)
 	    s = format (s, "adding [%u, %u]:", trace[i].offset,
-			(trace[i].offset +
-			 trace[i].len) % dummy_fifo->nitems);
+			(trace[i].offset + trace[i].len) % dummy_fifo->size);
 	  svm_fifo_enqueue_with_offset (dummy_fifo, trace[i].offset,
 					trace[i].len, &data[offset]);
 	}
@@ -179,9 +182,10 @@ format_svm_fifo (u8 * s, va_list * args)
 
   indent = format_get_indent (s);
   s = format (s, "cursize %u nitems %u has_event %d\n",
-	      f->cursize, f->nitems, f->has_event);
-  s = format (s, "%Uhead %d tail %d segment manager %u\n", format_white_space,
-	      indent, f->head, f->tail, f->segment_manager);
+	      svm_fifo_max_dequeue (f), f->nitems, f->has_event);
+  s = format (s, "%Uhead %u tail %u segment manager %u\n", format_white_space,
+	      indent, (f->head % f->size), (f->tail % f->size),
+	      f->segment_manager);
 
   if (verbose > 1)
     s = format (s, "%Uvpp session %d thread %d app session %d thread %d\n",
@@ -215,7 +219,12 @@ svm_fifo_create (u32 data_size_in_bytes)
     return 0;
 
   clib_memset (f, 0, sizeof (*f));
-  f->nitems = data_size_in_bytes;
+  f->size = rounded_data_size;
+  /*
+   * usable size of the fifo set to rounded_data_size - 1
+   * to differentiate between free fifo and empty fifo.
+   */
+  f->nitems = f->size - 1;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
   f->ct_session_index = SVM_FIFO_INVALID_SESSION_INDEX;
   f->segment_index = SVM_FIFO_INVALID_INDEX;
@@ -281,21 +290,21 @@ ooo_segment_del (svm_fifo_t * f, u32 index)
  * adjacent segments and removing overlapping ones.
  */
 static void
-ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
+ooo_segment_add (svm_fifo_t * f, u32 offset, u32 head, u32 tail, u32 length)
 {
   ooo_segment_t *s, *new_s, *prev, *next, *it;
   u32 new_index, s_end_pos, s_index;
-  u32 normalized_position, normalized_end_position;
+  u32 offset_pos, offset_end_pos;
 
-  ASSERT (offset + length <= ooo_segment_distance_from_tail (f, f->head));
-  normalized_position = (f->tail + offset) % f->nitems;
-  normalized_end_position = (f->tail + offset + length) % f->nitems;
+  ASSERT (offset + length <= ooo_segment_distance_from_tail (f, head, tail));
+  offset_pos = tail + offset;
+  offset_end_pos = tail + offset + length;
 
   f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
   if (f->ooos_list_head == OOO_SEGMENT_INVALID_INDEX)
     {
-      s = ooo_segment_new (f, normalized_position, length);
+      s = ooo_segment_new (f, offset_pos, length);
       f->ooos_list_head = s - f->ooo_segments;
       f->ooos_newest = f->ooos_list_head;
       return;
@@ -304,20 +313,20 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   /* Find first segment that starts after new segment */
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
   while (s->next != OOO_SEGMENT_INVALID_INDEX
-	 && position_lt (f, s->start, normalized_position))
+	 && position_lt (f, s->start, offset_pos, tail))
     s = pool_elt_at_index (f->ooo_segments, s->next);
 
   /* If we have a previous and we overlap it, use it as starting point */
   prev = ooo_segment_get_prev (f, s);
   if (prev
-      && position_leq (f, normalized_position, ooo_segment_end_pos (f, prev)))
+      && position_leq (f, offset_pos, ooo_segment_end_pos (f, prev), tail))
     {
       s = prev;
       s_end_pos = ooo_segment_end_pos (f, s);
 
-      /* Since we have previous, normalized start position cannot be smaller
+      /* Since we have previous, offset start position cannot be smaller
        * than prev->start. Check tail */
-      ASSERT (position_lt (f, s->start, normalized_position));
+      ASSERT (position_lt (f, s->start, offset_pos, tail));
       goto check_tail;
     }
 
@@ -325,9 +334,9 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   s_end_pos = ooo_segment_end_pos (f, s);
 
   /* No overlap, add before current segment */
-  if (position_lt (f, normalized_end_position, s->start))
+  if (position_lt (f, offset_end_pos, s->start, tail))
     {
-      new_s = ooo_segment_new (f, normalized_position, length);
+      new_s = ooo_segment_new (f, offset_pos, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
@@ -350,9 +359,9 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       return;
     }
   /* No overlap, add after current segment */
-  else if (position_gt (f, normalized_position, s_end_pos))
+  else if (position_gt (f, offset_pos, s_end_pos, tail))
     {
-      new_s = ooo_segment_new (f, normalized_position, length);
+      new_s = ooo_segment_new (f, offset_pos, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
@@ -373,24 +382,24 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
    */
 
   /* Merge at head */
-  if (position_lt (f, normalized_position, s->start))
+  if (position_lt (f, offset_pos, s->start, tail))
     {
-      s->start = normalized_position;
-      s->length = position_diff (f, s_end_pos, s->start);
+      s->start = offset_pos;
+      s->length = position_diff (f, s_end_pos, s->start, tail);
       f->ooos_newest = s - f->ooo_segments;
     }
 
 check_tail:
 
   /* Overlapping tail */
-  if (position_gt (f, normalized_end_position, s_end_pos))
+  if (position_gt (f, offset_end_pos, s_end_pos, tail))
     {
-      s->length = position_diff (f, normalized_end_position, s->start);
+      s->length = position_diff (f, offset_end_pos, s->start, tail);
 
       /* Remove the completely overlapped segments in the tail */
       it = ooo_segment_next (f, s);
       while (it && position_leq (f, ooo_segment_end_pos (f, it),
-				 normalized_end_position))
+				 offset_end_pos, tail))
 	{
 	  next = ooo_segment_next (f, it);
 	  ooo_segment_del (f, it - f->ooo_segments);
@@ -398,10 +407,10 @@ check_tail:
 	}
 
       /* If partial overlap with last, merge */
-      if (it && position_leq (f, it->start, normalized_end_position))
+      if (it && position_leq (f, it->start, offset_end_pos, tail))
 	{
 	  s->length = position_diff (f, ooo_segment_end_pos (f, it),
-				     s->start);
+				     s->start, tail);
 	  ooo_segment_del (f, it - f->ooo_segments);
 	}
       f->ooos_newest = s - f->ooo_segments;
@@ -413,14 +422,14 @@ check_tail:
  * advanced. Returns the number of bytes added to tail.
  */
 static int
-ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
+ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued, u32 * tail)
 {
   ooo_segment_t *s;
   u32 index, bytes = 0;
   i32 diff;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
-  diff = ooo_segment_distance_to_tail (f, s->start);
+  diff = ooo_segment_distance_to_tail (f, s->start, *tail);
 
   ASSERT (diff != n_bytes_enqueued);
 
@@ -436,8 +445,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
       if (s->length > diff)
 	{
 	  bytes = s->length - diff;
-	  f->tail += bytes;
-	  f->tail %= f->nitems;
+	  *tail = *tail + bytes;
 	  ooo_segment_del (f, index);
 	  break;
 	}
@@ -446,7 +454,7 @@ 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 = ooo_segment_distance_to_tail (f, s->start);
+	  diff = ooo_segment_distance_to_tail (f, s->start, *tail);
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -465,64 +473,63 @@ CLIB_MARCH_FN (svm_fifo_enqueue_nowait, int, svm_fifo_t * f, u32 max_bytes,
 	       const u8 * copy_from_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems;
+  u32 tail, head, free_count, tail_idx;
+
+  f_load_head_tail_prod (f, &head, &tail);
+
+  /* free space in fifo can only increase during enqueue: SPSC */
+  free_count = f_free_count (f, head, tail);
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
   f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
-  if (PREDICT_FALSE (cursize == f->nitems))
+  if (PREDICT_FALSE (free_count == 0))
     return SVM_FIFO_FULL;
 
-  nitems = f->nitems;
+  /* number of bytes we're going to copy */
+  total_copy_bytes = free_count < max_bytes ? free_count : max_bytes;
 
-  /* Number of bytes we're going to copy */
-  total_copy_bytes = (nitems - cursize) < max_bytes ?
-    (nitems - cursize) : max_bytes;
+  tail_idx = tail % f->size;
 
   if (PREDICT_TRUE (copy_from_here != 0))
     {
-      /* Number of bytes in first copy segment */
-      first_copy_bytes = ((nitems - f->tail) < total_copy_bytes)
-	? (nitems - f->tail) : total_copy_bytes;
-
-      clib_memcpy_fast (&f->data[f->tail], copy_from_here, first_copy_bytes);
-      f->tail += first_copy_bytes;
-      f->tail = (f->tail == nitems) ? 0 : f->tail;
-
-      /* Number of bytes in second copy segment, if any */
-      second_copy_bytes = total_copy_bytes - first_copy_bytes;
-      if (second_copy_bytes)
+      first_copy_bytes = f->size - tail_idx;
+      if (first_copy_bytes < total_copy_bytes)
 	{
-	  clib_memcpy_fast (&f->data[f->tail],
+	  clib_memcpy_fast (&f->data[tail_idx], copy_from_here,
+			    first_copy_bytes);
+	  /* number of bytes in second copy segment */
+	  second_copy_bytes = total_copy_bytes - first_copy_bytes;
+	  /* wrap around */
+	  clib_memcpy_fast (&f->data[0],
 			    copy_from_here + first_copy_bytes,
 			    second_copy_bytes);
-	  f->tail += second_copy_bytes;
-	  f->tail = (f->tail == nitems) ? 0 : f->tail;
 	}
+      else
+	{
+	  clib_memcpy_fast (&f->data[tail_idx], copy_from_here,
+			    total_copy_bytes);
+	}
+      tail += total_copy_bytes;
     }
   else
     {
       ASSERT (0);
-
       /* Account for a zero-copy enqueue done elsewhere */
-      ASSERT (max_bytes <= (nitems - cursize));
-      f->tail += max_bytes;
-      f->tail = f->tail % nitems;
-      total_copy_bytes = max_bytes;
+      tail += max_bytes;
     }
 
-  svm_fifo_trace_add (f, f->head, total_copy_bytes, 2);
+  svm_fifo_trace_add (f, head, total_copy_bytes, 2);
 
-  /* Any out-of-order segments to collect? */
+  /* collect out-of-order segments */
   if (PREDICT_FALSE (f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX))
-    total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
+    total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes, &tail);
 
-  /* Atomically increase the queue length */
-  ASSERT (cursize + total_copy_bytes <= nitems);
-  clib_atomic_fetch_add_rel (&f->cursize, total_copy_bytes);
+  ASSERT (total_copy_bytes <= free_count);
 
-  return (total_copy_bytes);
+  /* store-rel: producer owned index (paired with load-acq in consumer) */
+  clib_atomic_store_rel_n (&f->tail, tail);
+
+  return total_copy_bytes;
 }
 
 #ifndef CLIB_MARCH_VARIANT
@@ -546,50 +553,53 @@ CLIB_MARCH_FN (svm_fifo_enqueue_with_offset, int, svm_fifo_t * f,
 	       u32 offset, u32 required_bytes, u8 * copy_from_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems, normalized_offset;
+  u32 tail_offset;
+  u32 tail, head, free_count, tail_offset_idx;
 
-  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
+  f_load_head_tail_prod (f, &head, &tail);
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
-  nitems = f->nitems;
+  /* free space in fifo can only increase during enqueue: SPSC */
+  free_count = f_free_count (f, head, tail);
 
-  ASSERT (required_bytes < nitems);
+  /* will this request fit? */
+  if ((required_bytes + offset) > free_count)
+    return -1;
 
-  normalized_offset = (f->tail + offset) % nitems;
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
-  /* Will this request fit? */
-  if ((required_bytes + offset) > (nitems - cursize))
-    return -1;
+  ASSERT (required_bytes < f->nitems);
+
+  tail_offset = tail + offset;
+  tail_offset_idx = tail_offset % f->size;
 
   svm_fifo_trace_add (f, offset, required_bytes, 1);
 
-  ooo_segment_add (f, offset, required_bytes);
+  ooo_segment_add (f, offset, head, tail, required_bytes);
 
-  /* Number of bytes we're going to copy */
+  /* number of bytes we're going to copy */
   total_copy_bytes = required_bytes;
 
-  /* Number of bytes in first copy segment */
-  first_copy_bytes = ((nitems - normalized_offset) < total_copy_bytes)
-    ? (nitems - normalized_offset) : total_copy_bytes;
-
-  clib_memcpy_fast (&f->data[normalized_offset], copy_from_here,
-		    first_copy_bytes);
+  /* number of bytes in first copy segment */
+  first_copy_bytes = f->size - tail_offset_idx;
 
-  /* Number of bytes in second copy segment, if any */
-  second_copy_bytes = total_copy_bytes - first_copy_bytes;
-  if (second_copy_bytes)
+  if (first_copy_bytes < total_copy_bytes)
     {
-      normalized_offset += first_copy_bytes;
-      normalized_offset %= nitems;
+      clib_memcpy_fast (&f->data[tail_offset_idx], copy_from_here,
+			first_copy_bytes);
 
-      ASSERT (normalized_offset == 0);
-
-      clib_memcpy_fast (&f->data[normalized_offset],
+      /* number of bytes in second copy segment */
+      second_copy_bytes = total_copy_bytes - first_copy_bytes;
+      /* wrap around */
+      clib_memcpy_fast (&f->data[0],
 			copy_from_here + first_copy_bytes, second_copy_bytes);
     }
+  else
+    {
+      clib_memcpy_fast (&f->data[tail_offset_idx], copy_from_here,
+			total_copy_bytes);
+    }
 
-  return (0);
+  return 0;
 }
 
 #ifndef CLIB_MARCH_VARIANT
@@ -607,13 +617,17 @@ void
 svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len)
 {
   u32 first_chunk;
-  first_chunk = f->nitems - f->head;
+  u32 head, tail, head_idx;
+
+  f_load_head_tail_cons (f, &head, &tail);
+  head_idx = head % f->size;
+  first_chunk = f->size - (head_idx);
   ASSERT (len <= f->nitems);
   if (len <= first_chunk)
-    clib_memcpy_fast (&f->data[f->head], data, len);
+    clib_memcpy_fast (&f->data[head_idx], data, len);
   else
     {
-      clib_memcpy_fast (&f->data[f->head], data, first_chunk);
+      clib_memcpy_fast (&f->data[head_idx], data, first_chunk);
       clib_memcpy_fast (&f->data[0], data + first_chunk, len - first_chunk);
     }
 }
@@ -623,53 +637,52 @@ CLIB_MARCH_FN (svm_fifo_dequeue_nowait, int, svm_fifo_t * f, u32 max_bytes,
 	       u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems;
+  u32 tail, head, cursize, head_idx;
+
+  f_load_head_tail_cons (f, &head, &tail);
+
+  /* current size of fifo can only increase during dequeue: SPSC */
+  cursize = f_cursize (f, head, tail);
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
   if (PREDICT_FALSE (cursize == 0))
     return -2;			/* nothing in the fifo */
 
-  nitems = f->nitems;
-
-  /* Number of bytes we're going to copy */
+  /* number of bytes we're going to copy */
   total_copy_bytes = (cursize < max_bytes) ? cursize : max_bytes;
 
+  head_idx = head % f->size;
+
   if (PREDICT_TRUE (copy_here != 0))
     {
-      /* Number of bytes in first copy segment */
-      first_copy_bytes = ((nitems - f->head) < total_copy_bytes)
-	? (nitems - f->head) : total_copy_bytes;
-      clib_memcpy_fast (copy_here, &f->data[f->head], first_copy_bytes);
-      f->head += first_copy_bytes;
-      f->head = (f->head == nitems) ? 0 : f->head;
-
-      /* Number of bytes in second copy segment, if any */
-      second_copy_bytes = total_copy_bytes - first_copy_bytes;
-      if (second_copy_bytes)
+      /* number of bytes in first copy segment */
+      first_copy_bytes = f->size - head_idx;
+      if (first_copy_bytes < total_copy_bytes)
 	{
+	  clib_memcpy_fast (copy_here, &f->data[head_idx], first_copy_bytes);
+	  /* number of bytes in second copy segment */
+	  second_copy_bytes = total_copy_bytes - first_copy_bytes;
+	  /* wrap around */
 	  clib_memcpy_fast (copy_here + first_copy_bytes,
-			    &f->data[f->head], second_copy_bytes);
-	  f->head += second_copy_bytes;
-	  f->head = (f->head == nitems) ? 0 : f->head;
+			    &f->data[0], second_copy_bytes);
+	}
+      else
+	{
+	  clib_memcpy_fast (copy_here, &f->data[head_idx], total_copy_bytes);
 	}
+      head += total_copy_bytes;
     }
   else
     {
       ASSERT (0);
       /* Account for a zero-copy dequeue done elsewhere */
-      ASSERT (max_bytes <= cursize);
-      f->head += max_bytes;
-      f->head = f->head % nitems;
-      cursize -= max_bytes;
-      total_copy_bytes = max_bytes;
+      head += max_bytes;
     }
 
-  ASSERT (f->head <= nitems);
   ASSERT (cursize >= total_copy_bytes);
-  clib_atomic_fetch_sub_rel (&f->cursize, total_copy_bytes);
+  /* store-rel: consumer owned index (paired with load-acq in producer) */
+  clib_atomic_store_rel_n (&f->head, head);
 
-  return (total_copy_bytes);
+  return total_copy_bytes;
 }
 
 #ifndef CLIB_MARCH_VARIANT
@@ -686,36 +699,42 @@ CLIB_MARCH_FN (svm_fifo_peek, int, svm_fifo_t * f, u32 relative_offset,
 	       u32 max_bytes, u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems, real_head;
+  u32 tail, head, cursize;
+  u32 relative_head_idx;
+
+  f_load_head_tail_cons (f, &head, &tail);
+
+  /* current size of fifo can only increase during peek: SPSC */
+  cursize = f_cursize (f, head, tail);
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
   if (PREDICT_FALSE (cursize < relative_offset))
     return -2;			/* nothing in the fifo */
 
-  nitems = f->nitems;
-  real_head = f->head + relative_offset;
-  real_head = real_head >= nitems ? real_head - nitems : real_head;
+  relative_head_idx = (head + relative_offset) % f->size;
 
-  /* Number of bytes we're going to copy */
-  total_copy_bytes = (cursize - relative_offset < max_bytes) ?
+  /* number of bytes we're going to copy */
+  total_copy_bytes = ((cursize - relative_offset) < max_bytes) ?
     cursize - relative_offset : max_bytes;
 
   if (PREDICT_TRUE (copy_here != 0))
     {
-      /* Number of bytes in first copy segment */
-      first_copy_bytes =
-	((nitems - real_head) < total_copy_bytes) ?
-	(nitems - real_head) : total_copy_bytes;
-      clib_memcpy_fast (copy_here, &f->data[real_head], first_copy_bytes);
-
-      /* Number of bytes in second copy segment, if any */
-      second_copy_bytes = total_copy_bytes - first_copy_bytes;
-      if (second_copy_bytes)
+      /* number of bytes in first copy segment */
+      first_copy_bytes = f->size - relative_head_idx;
+      if (first_copy_bytes < total_copy_bytes)
 	{
+	  clib_memcpy_fast (copy_here, &f->data[relative_head_idx],
+			    first_copy_bytes);
+
+	  /* number of bytes in second copy segment */
+	  second_copy_bytes = total_copy_bytes - first_copy_bytes;
 	  clib_memcpy_fast (copy_here + first_copy_bytes, &f->data[0],
 			    second_copy_bytes);
 	}
+      else
+	{
+	  clib_memcpy_fast (copy_here, &f->data[relative_head_idx],
+			    total_copy_bytes);
+	}
     }
   return total_copy_bytes;
 }
@@ -733,39 +752,28 @@ svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
 int
 svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
 {
-  u32 total_drop_bytes, first_drop_bytes, second_drop_bytes;
-  u32 cursize, nitems;
+  u32 total_drop_bytes;
+  u32 tail, head, cursize;
+
+  f_load_head_tail_cons (f, &head, &tail);
+
+  /* number of bytes we're going to drop */
+  cursize = f_cursize (f, head, tail);
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
   if (PREDICT_FALSE (cursize == 0))
     return -2;			/* nothing in the fifo */
 
-  nitems = f->nitems;
+  svm_fifo_trace_add (f, tail, total_drop_bytes, 3);
 
-  /* Number of bytes we're going to drop */
+  /* number of bytes we're going to drop */
   total_drop_bytes = (cursize < max_bytes) ? cursize : max_bytes;
 
-  svm_fifo_trace_add (f, f->tail, total_drop_bytes, 3);
-
-  /* Number of bytes in first copy segment */
-  first_drop_bytes =
-    ((nitems - f->head) < total_drop_bytes) ?
-    (nitems - f->head) : total_drop_bytes;
-  f->head += first_drop_bytes;
-  f->head = (f->head == nitems) ? 0 : f->head;
-
-  /* Number of bytes in second drop segment, if any */
-  second_drop_bytes = total_drop_bytes - first_drop_bytes;
-  if (second_drop_bytes)
-    {
-      f->head += second_drop_bytes;
-      f->head = (f->head == nitems) ? 0 : f->head;
-    }
+  /* move head */
+  head += total_drop_bytes;
 
-  ASSERT (f->head <= nitems);
   ASSERT (cursize >= total_drop_bytes);
-  clib_atomic_fetch_sub_rel (&f->cursize, total_drop_bytes);
+  /* store-rel: consumer owned index (paired with load-acq in producer) */
+  clib_atomic_store_rel_n (&f->head, head);
 
   return total_drop_bytes;
 }
@@ -773,32 +781,38 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
 void
 svm_fifo_dequeue_drop_all (svm_fifo_t * f)
 {
-  f->head = f->tail;
-  clib_atomic_fetch_sub_rel (&f->cursize, f->cursize);
+  /* consumer foreign index */
+  u32 tail = clib_atomic_load_acq_n (&f->tail);
+  /* store-rel: consumer owned index (paired with load-acq in producer) */
+  clib_atomic_store_rel_n (&f->head, tail);
 }
 
 int
 svm_fifo_segments (svm_fifo_t * f, svm_fifo_segment_t * fs)
 {
-  u32 cursize, nitems;
+  u32 cursize, head, tail, head_idx;
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
-  if (PREDICT_FALSE (cursize == 0))
-    return -2;
+  f_load_head_tail_cons (f, &head, &tail);
+
+  /* consumer function, cursize can only increase while we're working */
+  cursize = f_cursize (f, head, tail);
 
-  nitems = f->nitems;
+  if (PREDICT_FALSE (cursize == 0))
+    return -2;			/* nothing in the fifo */
 
-  fs[0].len = ((nitems - f->head) < cursize) ? (nitems - f->head) : cursize;
-  fs[0].data = f->data + f->head;
+  head_idx = head % f->size;
 
-  if (fs[0].len < cursize)
+  if (tail < head)
     {
+      fs[0].len = f->size - head_idx;
+      fs[0].data = f->data + head_idx;
       fs[1].len = cursize - fs[0].len;
       fs[1].data = f->data;
     }
   else
     {
+      fs[0].len = cursize;
+      fs[0].data = f->data + head_idx;
       fs[1].len = 0;
       fs[1].data = 0;
     }
@@ -808,20 +822,28 @@ svm_fifo_segments (svm_fifo_t * f, svm_fifo_segment_t * fs)
 void
 svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_segment_t * fs)
 {
-  u32 total_drop_bytes;
+  u32 head, head_idx;
 
-  ASSERT (fs[0].data == f->data + f->head);
-  if (fs[1].len)
-    {
-      f->head = fs[1].len;
-      total_drop_bytes = fs[0].len + fs[1].len;
-    }
-  else
-    {
-      f->head = (f->head + fs[0].len) % f->nitems;
-      total_drop_bytes = fs[0].len;
-    }
-  clib_atomic_fetch_sub_rel (&f->cursize, total_drop_bytes);
+  /* consumer owned index */
+  head = f->head;
+  head_idx = head % f->size;
+
+  ASSERT (fs[0].data == f->data + head_idx);
+  head += fs[0].len + fs[1].len;
+  /* store-rel: consumer owned index (paired with load-acq in producer) */
+  clib_atomic_store_rel_n (&f->head, head);
+}
+
+/* Assumption: no prod and cons are accessing either dest or src fifo */
+void
+svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf)
+{
+  u32 head, tail;
+  clib_memcpy_fast (df->data, sf->data, sf->size);
+
+  f_load_head_tail_all_acq (sf, &head, &tail);
+  clib_atomic_store_rel_n (&df->head, head);
+  clib_atomic_store_rel_n (&df->tail, tail);
 }
 
 u32
@@ -842,7 +864,8 @@ svm_fifo_first_ooo_segment (svm_fifo_t * f)
 void
 svm_fifo_init_pointers (svm_fifo_t * f, u32 pointer)
 {
-  f->head = f->tail = pointer % f->nitems;
+  clib_atomic_store_rel_n (&f->head, pointer);
+  clib_atomic_store_rel_n (&f->tail, pointer);
 }
 
 void
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 4a119341d8b..94a05938d26 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -1,5 +1,9 @@
 /*
  * Copyright (c) 2016-2019 Cisco and/or its affiliates.
+ * Copyright (c) 2019 Arm Limited
+ * Copyright (c) 2010-2017 Intel Corporation and/or its affiliates.
+ * Copyright (c) 2007-2009 Kip Macy kmacy@freebsd.org
+ * Inspired from DPDK rte_ring.h (SPSC only) (derived from freebsd bufring.h).
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at:
@@ -59,8 +63,8 @@ typedef struct
 typedef struct _svm_fifo
 {
   CLIB_CACHE_LINE_ALIGN_MARK (shared_first);
-  volatile u32 cursize;		/**< current fifo size */
-  u32 nitems;
+  u32 size;			/**< size of the fifo(must be power of 2) */
+  u32 nitems;			/**< usable size(size-1) */
 
     CLIB_CACHE_LINE_ALIGN_MARK (shared_second);
   volatile u32 has_event;	/**< non-zero if deq event exists */
@@ -125,28 +129,141 @@ typedef struct svm_fifo_segment_
 u8 *svm_fifo_dump_trace (u8 * s, svm_fifo_t * f);
 u8 *svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose);
 
+/* internal function */
+static inline void
+f_load_head_tail_cons (svm_fifo_t * f, u32 * head, u32 * tail)
+{
+  /* load-relaxed: consumer owned index */
+  *head = f->head;
+  /* load-acq: consumer foreign index (paired with store-rel in producer) */
+  *tail = clib_atomic_load_acq_n (&f->tail);
+}
+
+/* internal function */
+static inline void
+f_load_head_tail_prod (svm_fifo_t * f, u32 * head, u32 * tail)
+{
+  /* load relaxed: producer owned index */
+  *tail = f->tail;
+  /* load-acq: producer foreign index (paired with store-rel in consumer) */
+  *head = clib_atomic_load_acq_n (&f->head);
+}
+
+/* producer consumer role independent */
+/* internal function */
+static inline void
+f_load_head_tail_all_acq (svm_fifo_t * f, u32 * head, u32 * tail)
+{
+  /* load-acq : consumer foreign index (paired with store-rel) */
+  *tail = clib_atomic_load_acq_n (&f->tail);
+  /* load-acq : producer foriegn index (paired with store-rel) */
+  *head = clib_atomic_load_acq_n (&f->head);
+}
+
+/* internal function */
+static inline u32
+f_free_count (svm_fifo_t * f, u32 head, u32 tail)
+{
+  return (f->nitems + head - tail);
+}
+
+/* internal function */
+static inline u32
+f_cursize (svm_fifo_t * f, u32 head, u32 tail)
+{
+  return (f->nitems - f_free_count (f, head, tail));
+}
+
+/* used by consumer */
+static inline u32
+svm_fifo_max_dequeue_cons (svm_fifo_t * f)
+{
+  u32 tail, head;
+  f_load_head_tail_cons (f, &head, &tail);
+  return f_cursize (f, head, tail);
+}
+
+/* used by producer*/
+static inline u32
+svm_fifo_max_dequeue_prod (svm_fifo_t * f)
+{
+  u32 tail, head;
+  f_load_head_tail_prod (f, &head, &tail);
+  return f_cursize (f, head, tail);
+}
+
+/* use producer or consumer specific functions for perfomance.
+ * svm_fifo_max_dequeue_cons (svm_fifo_t *f)
+ * svm_fifo_max_dequeue_prod (svm_fifo_t *f)
+ */
 static inline u32
 svm_fifo_max_dequeue (svm_fifo_t * f)
 {
-  return clib_atomic_load_acq_n (&f->cursize);
+  u32 tail, head;
+  f_load_head_tail_all_acq (f, &head, &tail);
+  return f_cursize (f, head, tail);
 }
 
+/* used by producer */
+static inline int
+svm_fifo_is_full_prod (svm_fifo_t * f)
+{
+  return (svm_fifo_max_dequeue_prod (f) == f->nitems);
+}
+
+/* use producer or consumer specific functions for perfomance.
+ * svm_fifo_is_full_prod (svm_fifo_t * f)
+ * add cons version if needed
+ */
 static inline int
 svm_fifo_is_full (svm_fifo_t * f)
 {
-  return (clib_atomic_load_acq_n (&f->cursize) == f->nitems);
+  return (svm_fifo_max_dequeue (f) == f->nitems);
+}
+
+/* used by consumer */
+static inline int
+svm_fifo_is_empty_cons (svm_fifo_t * f)
+{
+  return (svm_fifo_max_dequeue_cons (f) == 0);
 }
 
+/* used by producer */
+static inline int
+svm_fifo_is_empty_prod (svm_fifo_t * f)
+{
+  return (svm_fifo_max_dequeue_prod (f) == 0);
+}
+
+/* use producer or consumer specific functions for perfomance.
+ * svm_fifo_is_empty_cons (svm_fifo_t * f)
+ * svm_fifo_is_empty_prod (svm_fifo_t * f)
+ */
 static inline int
 svm_fifo_is_empty (svm_fifo_t * f)
 {
-  return (clib_atomic_load_acq_n (&f->cursize) == 0);
+  return (svm_fifo_max_dequeue (f) == 0);
 }
 
+/* used by producer*/
+static inline u32
+svm_fifo_max_enqueue_prod (svm_fifo_t * f)
+{
+  u32 head, tail;
+  f_load_head_tail_prod (f, &head, &tail);
+  return f_free_count (f, head, tail);
+}
+
+/* use producer or consumer specfic functions for perfomance.
+ * svm_fifo_max_enqueue_prod (svm_fifo_t *f)
+ * add consumer specific version if needed.
+ */
 static inline u32
 svm_fifo_max_enqueue (svm_fifo_t * f)
 {
-  return f->nitems - svm_fifo_max_dequeue (f);
+  u32 head, tail;
+  f_load_head_tail_all_acq (f, &head, &tail);
+  return f_free_count (f, head, tail);
 }
 
 static inline int
@@ -164,7 +281,7 @@ svm_fifo_has_ooo_data (svm_fifo_t * f)
 /**
  * Sets fifo event flag.
  *
- * Also acts as a release barrier.
+ * Also acts as a release ordering.
  *
  * @return 1 if flag was not set.
  */
@@ -202,6 +319,7 @@ void svm_fifo_dequeue_drop_all (svm_fifo_t * f);
 int svm_fifo_segments (svm_fifo_t * f, svm_fifo_segment_t * fs);
 void svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_segment_t * fs);
 void svm_fifo_init_pointers (svm_fifo_t * f, u32 pointer);
+void svm_fifo_clone (svm_fifo_t * df, svm_fifo_t * sf);
 void svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len);
 void svm_fifo_add_subscriber (svm_fifo_t * f, u8 subscriber);
 void svm_fifo_del_subscriber (svm_fifo_t * f, u8 subscriber);
@@ -213,7 +331,12 @@ format_function_t format_svm_fifo;
 always_inline u32
 svm_fifo_max_read_chunk (svm_fifo_t * f)
 {
-  return ((f->tail > f->head) ? (f->tail - f->head) : (f->nitems - f->head));
+  u32 head, tail;
+  u32 head_idx, tail_idx;
+  f_load_head_tail_cons (f, &head, &tail);
+  head_idx = head % f->size;
+  tail_idx = tail % f->size;
+  return tail_idx > head_idx ? (tail_idx - head_idx) : (f->size - head_idx);
 }
 
 /**
@@ -222,7 +345,12 @@ svm_fifo_max_read_chunk (svm_fifo_t * f)
 always_inline u32
 svm_fifo_max_write_chunk (svm_fifo_t * f)
 {
-  return ((f->tail >= f->head) ? (f->nitems - f->tail) : (f->head - f->tail));
+  u32 head, tail;
+  u32 head_idx, tail_idx;
+  f_load_head_tail_prod (f, &head, &tail);
+  head_idx = head % f->size;
+  tail_idx = tail % f->size;
+  return tail_idx >= head_idx ? (f->size - tail_idx) : (head_idx - tail_idx);
 }
 
 /**
@@ -233,27 +361,26 @@ svm_fifo_max_write_chunk (svm_fifo_t * f)
 always_inline void
 svm_fifo_enqueue_nocopy (svm_fifo_t * f, u32 bytes)
 {
-  ASSERT (bytes <= svm_fifo_max_enqueue (f));
-  f->tail = (f->tail + bytes) % f->nitems;
-  clib_atomic_fetch_add_rel (&f->cursize, bytes);
+  ASSERT (bytes <= svm_fifo_max_enqueue_prod (f));
+  /* load-relaxed: producer owned index */
+  u32 tail = f->tail;
+  tail += bytes;
+  /* store-rel: producer owned index (paired with load-acq in consumer) */
+  clib_atomic_store_rel_n (&f->tail, tail);
 }
 
 always_inline u8 *
 svm_fifo_head (svm_fifo_t * f)
 {
-  return (f->data + f->head);
+  /* load-relaxed: consumer owned index */
+  return (f->data + (f->head % f->size));
 }
 
 always_inline u8 *
 svm_fifo_tail (svm_fifo_t * f)
 {
-  return (f->data + f->tail);
-}
-
-always_inline u32
-svm_fifo_nitems (svm_fifo_t * f)
-{
-  return f->nitems;
+  /* load-relaxed: producer owned index */
+  return (f->data + (f->tail % f->size));
 }
 
 static inline void
@@ -293,8 +420,8 @@ svm_fifo_needs_tx_ntf (svm_fifo_t * f, u32 n_last_deq)
     return 1;
   else if (want_ntf & SVM_FIFO_WANT_TX_NOTIF_IF_FULL)
     {
-      u32 max_deq = svm_fifo_max_dequeue (f);
-      u32 nitems = svm_fifo_nitems (f);
+      u32 max_deq = svm_fifo_max_dequeue_cons (f);
+      u32 nitems = f->nitems;
       if (!f->has_tx_ntf && max_deq < nitems
 	  && max_deq + n_last_deq >= nitems)
 	return 1;
@@ -328,31 +455,25 @@ svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f)
 }
 
 always_inline u32
-ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
+ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos, u32 tail)
 {
-  /* Ambiguous. Assumption is that ooo segments don't touch tail */
-  if (PREDICT_FALSE (pos == f->tail && f->tail == f->head))
-    return f->nitems;
-
-  return (((f->nitems + pos) - f->tail) % f->nitems);
+  return ((pos - tail) % f->size);
 }
 
 always_inline u32
-ooo_segment_distance_to_tail (svm_fifo_t * f, u32 pos)
+ooo_segment_distance_to_tail (svm_fifo_t * f, u32 pos, u32 tail)
 {
-  return (((f->nitems + f->tail) - pos) % f->nitems);
+  return ((tail - pos) % f->size);
 }
 
 always_inline u32
-ooo_segment_offset (svm_fifo_t * f, ooo_segment_t * s)
+ooo_segment_offset_prod (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ooo_segment_distance_from_tail (f, s->start);
-}
+  u32 tail;
+  /* load-relaxed: producer owned index */
+  tail = f->tail;
 
-always_inline u32
-ooo_segment_end_offset (svm_fifo_t * f, ooo_segment_t * s)
-{
-  return ooo_segment_distance_from_tail (f, s->start) + s->length;
+  return ooo_segment_distance_from_tail (f, s->start, tail);
 }
 
 always_inline u32
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index e377160f6d8..94bf6a5e73e 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -381,7 +381,12 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * fs,
 	  fsh->free_fifos[freelist_index] = f->next;
 	  /* (re)initialize the fifo, as in svm_fifo_create */
 	  clib_memset (f, 0, sizeof (*f));
-	  f->nitems = data_size_in_bytes;
+	  f->size = (1 << (max_log2 (data_size_in_bytes)));
+	  /*
+	   * usable size of the fifo set to rounded_data_size - 1
+	   * to differentiate between free fifo and empty fifo.
+	   */
+	  f->nitems = f->size - 1;
 	  f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
 	  f->ct_session_index = SVM_FIFO_INVALID_SESSION_INDEX;
 	  f->refcnt = 1;