session svm: per thread fifo segment slices

Type: refactor

Change-Id: I9be652e56cdb48b0aee3253f7ce8d9bed299d824
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dave@barachs.net>

3 files changed, 324 insertions, 205 deletions

diff --git a/src/svm/fifo_segment.c b/src/svm/fifo_segment.c
index 698a422127d..95b88836ddb 100644
--- a/src/svm/fifo_segment.c
+++ b/src/svm/fifo_segment.c
@@ -34,6 +34,30 @@ fs_free_space (fifo_segment_t * fs)
   return dlminfo.fordblks;
 }
 
+static inline void
+fsh_free_bytes_sub (fifo_segment_header_t * fsh, int size)
+{
+  clib_atomic_fetch_sub_rel (&fsh->n_free_bytes, size);
+}
+
+static inline uword
+fsh_n_free_bytes (fifo_segment_header_t * fsh)
+{
+  return clib_atomic_load_relax_n (&fsh->n_free_bytes);
+}
+
+static inline fifo_segment_slice_t *
+fsh_slice_get (fifo_segment_header_t * fsh, u32 slice_index)
+{
+  return &fsh->slices[slice_index];
+}
+
+static inline void
+fsh_active_fifos_update (fifo_segment_header_t * fsh, int inc)
+{
+  clib_atomic_fetch_add_rel (&fsh->n_active_fifos, inc);
+}
+
 /**
  * Initialize fifo segment shared header
  */
@@ -41,15 +65,45 @@ int
 fifo_segment_init (fifo_segment_t * fs)
 {
   fifo_segment_header_t *fsh;
+  fifo_segment_slice_t *fss;
   ssvm_shared_header_t *sh;
+  u32 max_chunk_sz;
+  uword max_fifo;
   void *oldheap;
+  int i;
 
   sh = fs->ssvm.sh;
   oldheap = ssvm_push_heap (sh);
 
-  fsh = clib_mem_alloc (sizeof (*fsh));
+  /*
+   * Manually align the fifo segment header to sizeof(uword) = 8 bytes.
+   * Long story made short: the "process-private" fifo segment
+   * is allocated from the main heap, not mmapped. dlmalloc
+   * only guarantees 4-byte alignment, and on aarch64
+   * the fsh can end up 4-byte but not 8-byte aligned.
+   * That eventually causes the atomic op in fifo_segment_update_free_bytes
+   * to backfire.
+   */
+  fsh = clib_mem_alloc_aligned (sizeof (*fsh), sizeof (uword));
   clib_memset (fsh, 0, sizeof (*fsh));
   fs->h = sh->opaque[0] = fsh;
+  fs->n_slices = clib_max (fs->n_slices, 1);
+
+  fsh->ssvm_sh = fs->ssvm.sh;
+  fsh->n_slices = fs->n_slices;
+  max_fifo = clib_min ((fs_free_space (fs) - 4096) / 2,
+		       FIFO_SEGMENT_MAX_FIFO_SIZE);
+  fsh->max_log2_chunk_size = max_log2 (max_fifo);
+
+  fsh->slices = clib_mem_alloc (sizeof (*fss) * fs->n_slices);
+  clib_memset (fsh->slices, 0, sizeof (*fss) * fs->n_slices);
+  max_chunk_sz = fsh->max_log2_chunk_size - FIFO_SEGMENT_MIN_LOG2_FIFO_SIZE;
+
+  for (i = 0; i < fs->n_slices; i++)
+    {
+      fss = fsh_slice_get (fsh, i);
+      vec_validate_init_empty (fss->free_chunks, max_chunk_sz, 0);
+    }
 
   ssvm_pop_heap (oldheap);
 
@@ -98,30 +152,30 @@ fifo_segment_create (fifo_segment_main_t * sm, fifo_segment_create_args_t * a)
 int
 fifo_segment_attach (fifo_segment_main_t * sm, fifo_segment_create_args_t * a)
 {
-  fifo_segment_t *s;
+  fifo_segment_t *fs;
   int rv;
 
-  pool_get_zero (sm->segments, s);
+  pool_get_zero (sm->segments, fs);
 
-  s->ssvm.ssvm_size = a->segment_size;
-  s->ssvm.my_pid = getpid ();
-  s->ssvm.name = format (0, "%s%c", a->segment_name, 0);
-  s->ssvm.requested_va = sm->next_baseva;
+  fs->ssvm.ssvm_size = a->segment_size;
+  fs->ssvm.my_pid = getpid ();
+  fs->ssvm.name = format (0, "%s%c", a->segment_name, 0);
+  fs->ssvm.requested_va = sm->next_baseva;
   if (a->segment_type == SSVM_SEGMENT_MEMFD)
-    s->ssvm.fd = a->memfd_fd;
+    fs->ssvm.fd = a->memfd_fd;
   else
-    s->ssvm.attach_timeout = sm->timeout_in_seconds;
+    fs->ssvm.attach_timeout = sm->timeout_in_seconds;
 
-  if ((rv = ssvm_slave_init (&s->ssvm, a->segment_type)))
+  if ((rv = ssvm_slave_init (&fs->ssvm, a->segment_type)))
     {
-      _vec_len (s) = vec_len (s) - 1;
+      _vec_len (fs) = vec_len (fs) - 1;
       return (rv);
     }
 
   /* Fish the segment header */
-  s->h = s->ssvm.sh->opaque[0];
+  fs->h = fs->ssvm.sh->opaque[0];
 
-  vec_add1 (a->new_segment_indices, s - sm->segments);
+  vec_add1 (a->new_segment_indices, fs - sm->segments);
   return (0);
 }
 
@@ -139,12 +193,6 @@ fifo_segment_index (fifo_segment_main_t * sm, fifo_segment_t * s)
   return s - sm->segments;
 }
 
-void *
-svm_fifo_segment_heap (fifo_segment_t * seg)
-{
-  return seg->ssvm.sh->heap;
-}
-
 fifo_segment_t *
 fifo_segment_get_segment (fifo_segment_main_t * sm, u32 segment_index)
 {
@@ -169,41 +217,41 @@ fifo_segment_main_init (fifo_segment_main_t * sm, u64 baseva,
 static inline u32
 fs_freelist_for_size (u32 size)
 {
-  return max_log2 (size) - max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE);
+  return max_log2 (size) - FIFO_SEGMENT_MIN_LOG2_FIFO_SIZE;
 }
 
 static inline u32
 fs_freelist_index_to_size (u32 fl_index)
 {
-  return 1 << (fl_index + max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE));
+  return 1 << (fl_index + FIFO_SEGMENT_MIN_LOG2_FIFO_SIZE);
 }
 
 static inline int
-fs_chunk_size_is_valid (u32 size)
+fs_chunk_size_is_valid (fifo_segment_header_t * fsh, u32 size)
 {
   /*
    * 4K minimum. It's not likely that anything good will happen
    * with a smaller FIFO.
    */
   return size >= FIFO_SEGMENT_MIN_FIFO_SIZE
-    && size <= FIFO_SEGMENT_MAX_FIFO_SIZE;
+    && size <= (1 << fsh->max_log2_chunk_size);
 }
 
 static svm_fifo_t *
-fs_try_alloc_fifo_freelist (fifo_segment_t * fs, u32 fl_index, u32 data_bytes)
+fs_try_alloc_fifo_freelist (fifo_segment_slice_t * fss,
+			    u32 fl_index, u32 data_bytes)
 {
-  fifo_segment_header_t *fsh = fs->h;
   svm_fifo_chunk_t *c;
   svm_fifo_t *f;
 
-  f = fsh->free_fifos;
-  c = fsh->free_chunks[fl_index];
+  f = fss->free_fifos;
+  c = fss->free_chunks[fl_index];
 
   if (!f || !c)
     return 0;
 
-  fsh->free_fifos = f->next;
-  fsh->free_chunks[fl_index] = c->next;
+  fss->free_fifos = f->next;
+  fss->free_chunks[fl_index] = c->next;
   c->next = c;
   c->start_byte = 0;
   c->length = data_bytes;
@@ -211,44 +259,44 @@ fs_try_alloc_fifo_freelist (fifo_segment_t * fs, u32 fl_index, u32 data_bytes)
   f->start_chunk = c;
   f->end_chunk = c;
 
-  fsh->n_fl_chunk_bytes -= fs_freelist_index_to_size (fl_index);
+  fss->n_fl_chunk_bytes -= fs_freelist_index_to_size (fl_index);
   return f;
 }
 
 static svm_fifo_t *
-fs_try_alloc_fifo_freelist_multi_chunk (fifo_segment_t * fs, u32 data_bytes)
+fs_try_alloc_fifo_freelist_multi_chunk (fifo_segment_header_t * fsh,
+					fifo_segment_slice_t * fss,
+					u32 data_bytes)
 {
   svm_fifo_chunk_t *c, *first = 0, *last = 0;
-  fifo_segment_header_t *fsh = fs->h;
   u32 fl_index, fl_size, n_alloc = 0;
   svm_fifo_t *f;
 
-  f = fsh->free_fifos;
+  f = fss->free_fifos;
   if (!f)
     {
-      void *oldheap = ssvm_push_heap (fs->ssvm.sh);
+      void *oldheap = ssvm_push_heap (fsh->ssvm_sh);
       f = clib_mem_alloc_aligned (sizeof (*f), CLIB_CACHE_LINE_BYTES);
       ssvm_pop_heap (oldheap);
       if (!f)
 	return 0;
       memset (f, 0, sizeof (*f));
-      fsh->n_free_bytes -= sizeof (*f);
+      fsh_free_bytes_sub (fsh, sizeof (*f));
     }
   else
     {
-      fsh->free_fifos = f->next;
+      fss->free_fifos = f->next;
     }
 
   fl_index = fs_freelist_for_size (data_bytes) - 1;
-  vec_validate_init_empty (fsh->free_chunks, fl_index, 0);
   fl_size = fs_freelist_index_to_size (fl_index);
 
   while (data_bytes)
     {
-      c = fsh->free_chunks[fl_index];
+      c = fss->free_chunks[fl_index];
       if (c)
 	{
-	  fsh->free_chunks[fl_index] = c->next;
+	  fss->free_chunks[fl_index] = c->next;
 	  if (!last)
 	    last = c;
 	  c->next = first;
@@ -267,14 +315,15 @@ fs_try_alloc_fifo_freelist_multi_chunk (fifo_segment_t * fs, u32 data_bytes)
   f->start_chunk = first;
   f->end_chunk = last;
   last->next = first;
-  fsh->n_fl_chunk_bytes -= n_alloc;
+  fss->n_fl_chunk_bytes -= n_alloc;
   return f;
 }
 
 static int
-fs_try_alloc_fifo_batch (fifo_segment_t * fs, u32 fl_index, u32 batch_size)
+fs_try_alloc_fifo_batch (fifo_segment_header_t * fsh,
+			 fifo_segment_slice_t * fss,
+			 u32 fl_index, u32 batch_size)
 {
-  fifo_segment_header_t *fsh = fs->h;
   u32 size, hdrs, rounded_data_size;
   svm_fifo_chunk_t *c;
   svm_fifo_t *f;
@@ -282,12 +331,11 @@ fs_try_alloc_fifo_batch (fifo_segment_t * fs, u32 fl_index, u32 batch_size)
   u8 *fmem;
   int i;
 
-  vec_validate_init_empty (fsh->free_chunks, fl_index, 0);
   rounded_data_size = fs_freelist_index_to_size (fl_index);
   hdrs = sizeof (*f) + sizeof (*c);
   size = (hdrs + rounded_data_size) * batch_size;
 
-  oldheap = ssvm_push_heap (fs->ssvm.sh);
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
   fmem = clib_mem_alloc_aligned_at_offset (size, CLIB_CACHE_LINE_BYTES,
 					   0 /* align_offset */ ,
 					   0 /* os_out_of_memory */ );
@@ -302,18 +350,18 @@ fs_try_alloc_fifo_batch (fifo_segment_t * fs, u32 fl_index, u32 batch_size)
     {
       f = (svm_fifo_t *) fmem;
       memset (f, 0, sizeof (*f));
-      f->next = fsh->free_fifos;
-      fsh->free_fifos = f;
+      f->next = fss->free_fifos;
+      fss->free_fifos = f;
       c = (svm_fifo_chunk_t *) (fmem + sizeof (*f));
       c->start_byte = 0;
       c->length = rounded_data_size;
-      c->next = fsh->free_chunks[fl_index];
-      fsh->free_chunks[fl_index] = c;
+      c->next = fss->free_chunks[fl_index];
+      fss->free_chunks[fl_index] = c;
       fmem += hdrs + rounded_data_size;
     }
 
-  fsh->n_fl_chunk_bytes += batch_size * rounded_data_size;
-  fsh->n_free_bytes -= size;
+  fss->n_fl_chunk_bytes += batch_size * rounded_data_size;
+  fsh_free_bytes_sub (fsh, size);
 
   return 0;
 }
@@ -328,45 +376,46 @@ fs_try_alloc_fifo_batch (fifo_segment_t * fs, u32 fl_index, u32 batch_size)
  * - grab multiple fifo chunks from freelists
  */
 static svm_fifo_t *
-fs_try_alloc_fifo (fifo_segment_t * fs, u32 data_bytes)
+fs_try_alloc_fifo (fifo_segment_header_t * fsh, fifo_segment_slice_t * fss,
+		   u32 data_bytes)
 {
-  fifo_segment_header_t *fsh = fs->h;
   u32 fifo_sz, fl_index;
   svm_fifo_t *f = 0;
+  uword n_free_bytes;
 
   fl_index = fs_freelist_for_size (data_bytes);
-  vec_validate_init_empty (fsh->free_chunks, fl_index, 0);
   fifo_sz = sizeof (svm_fifo_t) + sizeof (svm_fifo_chunk_t);
   fifo_sz += 1 << max_log2 (data_bytes);
 
-  if (fsh->free_fifos && fsh->free_chunks[fl_index])
+  if (fss->free_fifos && fss->free_chunks[fl_index])
     {
-      f = fs_try_alloc_fifo_freelist (fs, fl_index, data_bytes);
+      f = fs_try_alloc_fifo_freelist (fss, fl_index, data_bytes);
       if (f)
 	goto done;
     }
-  if (fifo_sz * FIFO_SEGMENT_ALLOC_BATCH_SIZE < fsh->n_free_bytes)
+  n_free_bytes = fsh_n_free_bytes (fsh);
+  if (fifo_sz * FIFO_SEGMENT_ALLOC_BATCH_SIZE < n_free_bytes)
     {
-      if (fs_try_alloc_fifo_batch (fs, fl_index,
+      if (fs_try_alloc_fifo_batch (fsh, fss, fl_index,
 				   FIFO_SEGMENT_ALLOC_BATCH_SIZE))
 	goto done;
 
-      f = fs_try_alloc_fifo_freelist (fs, fl_index, data_bytes);
+      f = fs_try_alloc_fifo_freelist (fss, fl_index, data_bytes);
       goto done;
     }
-  if (fifo_sz <= fsh->n_free_bytes)
+  if (fifo_sz <= n_free_bytes)
     {
-      void *oldheap = ssvm_push_heap (fs->ssvm.sh);
+      void *oldheap = ssvm_push_heap (fsh->ssvm_sh);
       f = svm_fifo_create (data_bytes);
       ssvm_pop_heap (oldheap);
       if (f)
 	{
-	  fsh->n_free_bytes -= fifo_sz;
+	  fsh_free_bytes_sub (fsh, fifo_sz);
 	  goto done;
 	}
     }
-  if (data_bytes <= fsh->n_fl_chunk_bytes)
-    f = fs_try_alloc_fifo_freelist_multi_chunk (fs, data_bytes);
+  if (data_bytes <= fss->n_fl_chunk_bytes)
+    f = fs_try_alloc_fifo_freelist_multi_chunk (fsh, fss, data_bytes);
 
 done:
 
@@ -377,32 +426,29 @@ done:
  * Allocate fifo in fifo segment
  */
 svm_fifo_t *
-fifo_segment_alloc_fifo (fifo_segment_t * fs, u32 data_bytes,
-			 fifo_segment_ftype_t ftype)
+fifo_segment_alloc_fifo_w_slice (fifo_segment_t * fs, u32 slice_index,
+				 u32 data_bytes, fifo_segment_ftype_t ftype)
 {
-  fifo_segment_header_t *fsh;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
   svm_fifo_t *f = 0;
 
-  if (!fs_chunk_size_is_valid (data_bytes))
-    {
-      clib_warning ("fifo size out of range %d", data_bytes);
-      return 0;
-    }
-
-  fsh = fs->h;
-  ssvm_lock_non_recursive (fs->ssvm.sh, 1);
+  ASSERT (slice_index < fs->n_slices);
 
-  f = fs_try_alloc_fifo (fs, data_bytes);
+  fss = fsh_slice_get (fsh, slice_index);
+  f = fs_try_alloc_fifo (fsh, fss, data_bytes);
   if (!f)
     goto done;
 
+  f->slice_index = slice_index;
+
   /* (re)initialize the fifo, as in svm_fifo_create */
   svm_fifo_init (f, data_bytes);
 
   /* Initialize chunks and rbtree for multi-chunk fifos */
   if (f->start_chunk->next != f->start_chunk)
     {
-      void *oldheap = ssvm_push_heap (fs->ssvm.sh);
+      void *oldheap = ssvm_push_heap (fsh->ssvm_sh);
       svm_fifo_init_chunks (f);
       ssvm_pop_heap (oldheap);
     }
@@ -413,18 +459,17 @@ fifo_segment_alloc_fifo (fifo_segment_t * fs, u32 data_bytes,
    * only one. */
   if (ftype == FIFO_SEGMENT_RX_FIFO)
     {
-      if (fsh->fifos)
+      if (fss->fifos)
 	{
-	  fsh->fifos->prev = f;
-	  f->next = fsh->fifos;
+	  fss->fifos->prev = f;
+	  f->next = fss->fifos;
 	}
-      fsh->fifos = f;
+      fss->fifos = f;
       f->flags |= SVM_FIFO_F_LL_TRACKED;
     }
-  fsh->n_active_fifos++;
+  fsh_active_fifos_update (fsh, 1);
 
 done:
-  ssvm_unlock_non_recursive (fs->ssvm.sh);
   return (f);
 }
 
@@ -434,9 +479,9 @@ done:
 void
 fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
 {
+  fifo_segment_header_t *fsh = fs->h;
   svm_fifo_chunk_t *cur, *next;
-  fifo_segment_header_t *fsh;
-  ssvm_shared_header_t *sh;
+  fifo_segment_slice_t *fss;
   void *oldheap;
   int fl_index;
 
@@ -445,10 +490,7 @@ fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
   if (--f->refcnt > 0)
     return;
 
-  sh = fs->ssvm.sh;
-  fsh = fs->h;
-
-  ssvm_lock_non_recursive (sh, 2);
+  fss = fsh_slice_get (fsh, f->slice_index);
 
   /* Remove from active list. Only rx fifos are tracked */
   if (f->flags & SVM_FIFO_F_LL_TRACKED)
@@ -456,16 +498,16 @@ fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
       if (f->prev)
 	f->prev->next = f->next;
       else
-	fsh->fifos = f->next;
+	fss->fifos = f->next;
       if (f->next)
 	f->next->prev = f->prev;
       f->flags &= ~SVM_FIFO_F_LL_TRACKED;
     }
 
   /* Add to free list */
-  f->next = fsh->free_fifos;
+  f->next = fss->free_fifos;
   f->prev = 0;
-  fsh->free_fifos = f;
+  fss->free_fifos = f;
 
   /* Free fifo chunks */
   cur = f->start_chunk;
@@ -473,10 +515,10 @@ fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
     {
       next = cur->next;
       fl_index = fs_freelist_for_size (cur->length);
-      ASSERT (fl_index < vec_len (fsh->free_chunks));
-      cur->next = fsh->free_chunks[fl_index];
-      fsh->free_chunks[fl_index] = cur;
-      fsh->n_fl_chunk_bytes += fs_freelist_index_to_size (fl_index);
+      ASSERT (fl_index < vec_len (fss->free_chunks));
+      cur->next = fss->free_chunks[fl_index];
+      fss->free_chunks[fl_index] = cur;
+      fss->n_fl_chunk_bytes += fs_freelist_index_to_size (fl_index);
       cur = next;
     }
   while (cur != f->start_chunk);
@@ -484,7 +526,7 @@ fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
   f->start_chunk = f->end_chunk = f->new_chunks = 0;
   f->head_chunk = f->tail_chunk = f->ooo_enq = f->ooo_deq = 0;
 
-  oldheap = ssvm_push_heap (sh);
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
   svm_fifo_free_chunk_lookup (f);
   ssvm_pop_heap (oldheap);
 
@@ -497,23 +539,25 @@ fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f)
       f->master_thread_index = ~0;
     }
 
-  fsh->n_active_fifos--;
-  ssvm_unlock_non_recursive (sh);
+  fsh_active_fifos_update (fsh, -1);
 }
 
 int
-fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 batch_size)
+fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 slice_index,
+				 u32 batch_size)
 {
   fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
   svm_fifo_t *f;
   void *oldheap;
   u32 size;
   u8 *fmem;
   int i;
 
+  fss = fsh_slice_get (fsh, slice_index);
   size = (sizeof (*f)) * batch_size;
 
-  oldheap = ssvm_push_heap (fs->ssvm.sh);
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
   fmem = clib_mem_alloc_aligned_at_offset (size, CLIB_CACHE_LINE_BYTES,
 					   0 /* align_offset */ ,
 					   0 /* os_out_of_memory */ );
@@ -528,39 +572,39 @@ fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 batch_size)
     {
       f = (svm_fifo_t *) fmem;
       memset (f, 0, sizeof (*f));
-      f->next = fsh->free_fifos;
-      fsh->free_fifos = f;
+      f->next = fss->free_fifos;
+      fss->free_fifos = f;
       fmem += sizeof (*f);
     }
 
-  fsh->n_free_bytes -= size;
+  fsh_free_bytes_sub (fsh, size);
 
   return 0;
 }
 
 int
-fifo_segment_prealloc_fifo_chunks (fifo_segment_t * fs, u32 chunk_size,
-				   u32 batch_size)
+fifo_segment_prealloc_fifo_chunks (fifo_segment_t * fs, u32 slice_index,
+				   u32 chunk_size, u32 batch_size)
 {
-  fifo_segment_header_t *fsh = fs->h;
   u32 size, rounded_data_size, fl_index;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
   svm_fifo_chunk_t *c;
   void *oldheap;
   u8 *cmem;
   int i;
 
-  if (!fs_chunk_size_is_valid (chunk_size))
+  if (!fs_chunk_size_is_valid (fsh, chunk_size))
     {
       clib_warning ("chunk size out of range %d", chunk_size);
       return -1;
     }
 
   fl_index = fs_freelist_for_size (chunk_size);
-  vec_validate_init_empty (fsh->free_chunks, fl_index, 0);
   rounded_data_size = fs_freelist_index_to_size (fl_index);
   size = (sizeof (*c) + rounded_data_size) * batch_size;
 
-  oldheap = ssvm_push_heap (fs->ssvm.sh);
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
   cmem = clib_mem_alloc_aligned_at_offset (size, CLIB_CACHE_LINE_BYTES,
 					   0 /* align_offset */ ,
 					   0 /* os_out_of_memory */ );
@@ -570,19 +614,21 @@ fifo_segment_prealloc_fifo_chunks (fifo_segment_t * fs, u32 chunk_size,
   if (cmem == 0)
     return -1;
 
+  fss = fsh_slice_get (fsh, slice_index);
+
   /* Carve fifo + chunk space */
   for (i = 0; i < batch_size; i++)
     {
       c = (svm_fifo_chunk_t *) cmem;
       c->start_byte = 0;
       c->length = rounded_data_size;
-      c->next = fsh->free_chunks[fl_index];
-      fsh->free_chunks[fl_index] = c;
+      c->next = fss->free_chunks[fl_index];
+      fss->free_chunks[fl_index] = c;
       cmem += sizeof (*c) + rounded_data_size;
     }
 
-  fsh->n_fl_chunk_bytes += batch_size * rounded_data_size;
-  fsh->n_free_bytes -= size;
+  fss->n_fl_chunk_bytes += batch_size * rounded_data_size;
+  fsh_free_bytes_sub (fsh, size);
 
   return 0;
 }
@@ -596,21 +642,23 @@ fifo_segment_preallocate_fifo_pairs (fifo_segment_t * fs,
 				     u32 * n_fifo_pairs)
 {
   u32 rx_rounded_data_size, tx_rounded_data_size, pair_size, pairs_to_alloc;
-  int rx_fl_index, tx_fl_index;
+  fifo_segment_header_t *fsh = fs->h;
+  int rx_fl_index, tx_fl_index, i;
+  fifo_segment_slice_t *fss;
+  u32 hdrs, pairs_per_slice;
   uword space_available;
-  u32 hdrs;
 
   /* Parameter check */
   if (rx_fifo_size == 0 || tx_fifo_size == 0 || *n_fifo_pairs == 0)
     return;
 
-  if (!fs_chunk_size_is_valid (rx_fifo_size))
+  if (!fs_chunk_size_is_valid (fsh, rx_fifo_size))
     {
       clib_warning ("rx fifo_size out of range %d", rx_fifo_size);
       return;
     }
 
-  if (!fs_chunk_size_is_valid (tx_fifo_size))
+  if (!fs_chunk_size_is_valid (fsh, tx_fifo_size))
     {
       clib_warning ("tx fifo_size out of range %d", tx_fifo_size);
       return;
@@ -628,42 +676,39 @@ fifo_segment_preallocate_fifo_pairs (fifo_segment_t * fs,
   space_available = fs_free_space (fs);
   pairs_to_alloc = space_available / pair_size;
   pairs_to_alloc = clib_min (pairs_to_alloc, *n_fifo_pairs);
+  pairs_per_slice = pairs_to_alloc / fs->n_slices;
 
-  if (!pairs_to_alloc)
+  if (!pairs_per_slice)
     return;
 
-  if (fs_try_alloc_fifo_batch (fs, rx_fl_index, pairs_to_alloc))
-    clib_warning ("rx prealloc failed: pairs %u", pairs_to_alloc);
-  if (fs_try_alloc_fifo_batch (fs, tx_fl_index, pairs_to_alloc))
-    clib_warning ("tx prealloc failed: pairs %u", pairs_to_alloc);
+  for (i = 0; i < fs->n_slices; i++)
+    {
+      fss = fsh_slice_get (fsh, i);
+      if (fs_try_alloc_fifo_batch (fsh, fss, rx_fl_index, pairs_to_alloc))
+	clib_warning ("rx prealloc failed: pairs %u", pairs_to_alloc);
+      if (fs_try_alloc_fifo_batch (fsh, fss, tx_fl_index, pairs_to_alloc))
+	clib_warning ("tx prealloc failed: pairs %u", pairs_to_alloc);
+    }
 
   /* Account for the pairs allocated */
-  *n_fifo_pairs -= pairs_to_alloc;
+  *n_fifo_pairs -= pairs_per_slice * fs->n_slices;
 }
 
 int
 fifo_segment_grow_fifo (fifo_segment_t * fs, svm_fifo_t * f, u32 chunk_size)
 {
-  ssvm_shared_header_t *sh;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
   svm_fifo_chunk_t *c;
   void *oldheap;
   int fl_index;
 
-  if (!fs_chunk_size_is_valid (chunk_size))
-    {
-      clib_warning ("chunk size out of range %d", chunk_size);
-      return -1;
-    }
-
   fl_index = fs_freelist_for_size (chunk_size);
+  fss = fsh_slice_get (fsh, f->slice_index);
 
-  sh = fs->ssvm.sh;
-  ssvm_lock_non_recursive (sh, 1);
-
-  vec_validate_init_empty (fs->h->free_chunks, fl_index, 0);
-  c = fs->h->free_chunks[fl_index];
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
 
-  oldheap = ssvm_push_heap (sh);
+  c = fss->free_chunks[fl_index];
 
   if (!c)
     {
@@ -671,50 +716,47 @@ fifo_segment_grow_fifo (fifo_segment_t * fs, svm_fifo_t * f, u32 chunk_size)
       if (!c)
 	{
 	  ssvm_pop_heap (oldheap);
-	  ssvm_unlock_non_recursive (sh);
 	  return -1;
 	}
-      fs->h->n_free_bytes -= chunk_size + sizeof (*c);
+      fsh_free_bytes_sub (fsh, chunk_size + sizeof (*c));
     }
   else
     {
-      fs->h->free_chunks[fl_index] = c->next;
+      fss->free_chunks[fl_index] = c->next;
       c->next = 0;
-      fs->h->n_fl_chunk_bytes -= fs_freelist_index_to_size (fl_index);
+      fss->n_fl_chunk_bytes -= fs_freelist_index_to_size (fl_index);
     }
 
   svm_fifo_add_chunk (f, c);
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock_non_recursive (sh);
   return 0;
 }
 
 int
 fifo_segment_collect_fifo_chunks (fifo_segment_t * fs, svm_fifo_t * f)
 {
+  fifo_segment_header_t *fsh = fs->h;
   svm_fifo_chunk_t *cur, *next;
-  ssvm_shared_header_t *sh;
+  fifo_segment_slice_t *fss;
   void *oldheap;
   int fl_index;
 
-  sh = fs->ssvm.sh;
-  ssvm_lock_non_recursive (sh, 1);
-
-  oldheap = ssvm_push_heap (sh);
+  oldheap = ssvm_push_heap (fsh->ssvm_sh);
   cur = svm_fifo_collect_chunks (f);
 
+  fss = fsh_slice_get (fsh, f->slice_index);
+
   while (cur)
     {
       next = cur->next;
       fl_index = fs_freelist_for_size (cur->length);
-      cur->next = fs->h->free_chunks[fl_index];
-      fs->h->free_chunks[fl_index] = cur;
+      cur->next = fss->free_chunks[fl_index];
+      fss->free_chunks[fl_index] = cur;
       cur = next;
     }
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock_non_recursive (sh);
 
   return 0;
 }
@@ -725,17 +767,16 @@ fifo_segment_collect_fifo_chunks (fifo_segment_t * fs, svm_fifo_t * f)
 u32
 fifo_segment_num_fifos (fifo_segment_t * fs)
 {
-  return fs->h->n_active_fifos;
+  return clib_atomic_load_relax_n (&fs->h->n_active_fifos);
 }
 
-u32
-fifo_segment_num_free_fifos (fifo_segment_t * fs)
+static u32
+fs_slice_num_free_fifos (fifo_segment_slice_t * fss)
 {
-  fifo_segment_header_t *fsh = fs->h;
   svm_fifo_t *f;
   u32 count = 0;
 
-  f = fsh->free_fifos;
+  f = fss->free_fifos;
   if (f == 0)
     return 0;
 
@@ -748,21 +789,34 @@ fifo_segment_num_free_fifos (fifo_segment_t * fs)
 }
 
 u32
-fifo_segment_num_free_chunks (fifo_segment_t * fs, u32 size)
+fifo_segment_num_free_fifos (fifo_segment_t * fs)
+{
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
+  int slice_index;
+  u32 count = 0;
+
+  for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
+    {
+      fss = fsh_slice_get (fsh, slice_index);
+      count += fs_slice_num_free_fifos (fss);
+    }
+  return count;
+}
+
+static u32
+fs_slice_num_free_chunks (fifo_segment_slice_t * fss, u32 size)
 {
   u32 count = 0, rounded_size, fl_index;
-  fifo_segment_header_t *fsh;
   svm_fifo_chunk_t *c;
   int i;
 
-  fsh = fs->h;
-
   /* Count all free chunks? */
   if (size == ~0)
     {
-      for (i = 0; i < vec_len (fsh->free_chunks); i++)
+      for (i = 0; i < vec_len (fss->free_chunks); i++)
 	{
-	  c = fsh->free_chunks[i];
+	  c = fss->free_chunks[i];
 	  if (c == 0)
 	    continue;
 
@@ -778,10 +832,10 @@ fifo_segment_num_free_chunks (fifo_segment_t * fs, u32 size)
   rounded_size = (1 << (max_log2 (size)));
   fl_index = fs_freelist_for_size (rounded_size);
 
-  if (fl_index >= vec_len (fsh->free_chunks))
+  if (fl_index >= vec_len (fss->free_chunks))
     return 0;
 
-  c = fsh->free_chunks[fl_index];
+  c = fss->free_chunks[fl_index];
   if (c == 0)
     return 0;
 
@@ -793,34 +847,76 @@ fifo_segment_num_free_chunks (fifo_segment_t * fs, u32 size)
   return count;
 }
 
+u32
+fifo_segment_num_free_chunks (fifo_segment_t * fs, u32 size)
+{
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
+  int slice_index;
+  u32 count = 0;
+
+  for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
+    {
+      fss = fsh_slice_get (fsh, slice_index);
+      count += fs_slice_num_free_chunks (fss, size);
+    }
+  return count;
+}
+
 void
 fifo_segment_update_free_bytes (fifo_segment_t * fs)
 {
-  fs->h->n_free_bytes = fs_free_space (fs);
+  fifo_segment_header_t *fsh = fs->h;
+  clib_atomic_store_rel_n (&fsh->n_free_bytes, fs_free_space (fs));
 }
 
 u32
 fifo_segment_free_bytes (fifo_segment_t * fs)
 {
-  return fs->h->n_free_bytes;
+  return fsh_n_free_bytes (fs->h);
 }
 
-u32
+uword
 fifo_segment_fl_chunk_bytes (fifo_segment_t * fs)
 {
-  return fs->h->n_fl_chunk_bytes;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
+  uword n_bytes = 0;
+  int slice_index;
+
+  for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
+    {
+      fss = fsh_slice_get (fsh, slice_index);
+      n_bytes += fss->n_fl_chunk_bytes;
+    }
+
+  return n_bytes;
 }
 
 u8
 fifo_segment_has_fifos (fifo_segment_t * fs)
 {
-  return fs->h->fifos != 0;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
+  int slice_index;
+
+  for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
+    {
+      fss = fsh_slice_get (fsh, slice_index);
+      if (fss->fifos)
+	return 1;
+    }
+  return 0;
 }
 
 svm_fifo_t *
-fifo_segment_get_fifo_list (fifo_segment_t * fs)
+fifo_segment_get_slice_fifo_list (fifo_segment_t * fs, u32 slice_index)
 {
-  return fs->h->fifos;
+  fifo_segment_header_t *fsh = fs->h;
+  fifo_segment_slice_t *fss;
+
+  fss = fsh_slice_get (fsh, slice_index);
+  return fss->fifos;
 }
 
 u8 *
@@ -850,10 +946,12 @@ format_fifo_segment (u8 * s, va_list * args)
   u32 count, indent, active_fifos, free_fifos, fifo_hdr = 0, chunk_size;
   fifo_segment_t *fs = va_arg (*args, fifo_segment_t *);
   int verbose __attribute__ ((unused)) = va_arg (*args, int);
-  u32 est_chunk_bytes, est_free_seg_bytes;
+  u32 est_chunk_bytes, est_free_seg_bytes, free_chunks;
+  uword chunk_bytes = 0, free_seg_bytes;
   fifo_segment_header_t *fsh;
+  fifo_segment_slice_t *fss;
   svm_fifo_chunk_t *c;
-  u64 chunk_bytes = 0;
+  u32 slice_index;
   char *address;
   size_t size;
   int i;
@@ -861,9 +959,9 @@ format_fifo_segment (u8 * s, va_list * args)
   indent = format_get_indent (s) + 2;
 #if USE_DLMALLOC == 0
   s = format (s, "%U segment heap: %U\n", format_white_space, indent,
-	      format_mheap, svm_fifo_segment_heap (fs), verbose);
+	      format_mheap, fsh->ssvm_sh->heap, verbose);
   s = format (s, "%U segment has %u active fifos\n",
-	      format_white_space, indent, fifo_segment_num_fifos (fs));
+	      format_white_space, indent, fifo_segment_num_fifos (fsh));
 #endif
 
   if (fs == 0)
@@ -873,7 +971,6 @@ format_fifo_segment (u8 * s, va_list * args)
       return s;
     }
 
-  fsh = fs->h;
   fifo_segment_info (fs, &address, &size);
   active_fifos = fifo_segment_num_fifos (fs);
   free_fifos = fifo_segment_num_free_fifos (fs);
@@ -885,39 +982,47 @@ format_fifo_segment (u8 * s, va_list * args)
   if (!verbose)
     return s;
 
-  if (fsh->free_chunks)
+  fsh = fs->h;
+
+  free_chunks = fifo_segment_num_free_chunks (fs, ~0);
+  if (free_chunks)
     s = format (s, "\n\n%UFree chunks by size:\n", format_white_space,
 		indent + 2);
   else
     s = format (s, "\n");
 
-  for (i = 0; i < vec_len (fsh->free_chunks); i++)
+  for (slice_index = 0; slice_index < fs->n_slices; slice_index++)
     {
-      c = fsh->free_chunks[i];
-      if (c == 0)
-	continue;
-      count = 0;
-      while (c)
+      fss = fsh_slice_get (fsh, slice_index);
+      for (i = 0; i < vec_len (fss->free_chunks); i++)
 	{
-	  c = c->next;
-	  count++;
-	}
+	  c = fss->free_chunks[i];
+	  if (c == 0)
+	    continue;
+	  count = 0;
+	  while (c)
+	    {
+	      c = c->next;
+	      count++;
+	    }
 
-      chunk_size = fs_freelist_index_to_size (i);
-      s = format (s, "%U%-5u kB: %u\n", format_white_space, indent + 2,
-		  chunk_size >> 10, count);
+	  chunk_size = fs_freelist_index_to_size (i);
+	  s = format (s, "%U%-5u kB: %u\n", format_white_space, indent + 2,
+		      chunk_size >> 10, count);
 
-      chunk_bytes += count * chunk_size;
+	  chunk_bytes += count * chunk_size;
+	}
     }
 
   fifo_hdr = free_fifos * sizeof (svm_fifo_t);
   est_chunk_bytes = fifo_segment_fl_chunk_bytes (fs);
-  est_free_seg_bytes = fsh->n_free_bytes;
+  est_free_seg_bytes = fifo_segment_free_bytes (fs);
   fifo_segment_update_free_bytes (fs);
+  free_seg_bytes = fifo_segment_free_bytes (fs);
 
   s = format (s, "\n%Useg free bytes: %U (%u) estimated: %U (%u)\n",
 	      format_white_space, indent + 2, format_memory_size,
-	      fsh->n_free_bytes, fsh->n_free_bytes, format_memory_size,
+	      free_seg_bytes, free_seg_bytes, format_memory_size,
 	      est_free_seg_bytes, est_free_seg_bytes);
   s = format (s, "%Uchunk free bytes: %U (%lu) estimated: %U (%u)\n",
 	      format_white_space, indent + 2, format_memory_size, chunk_bytes,
diff --git a/src/svm/fifo_segment.h b/src/svm/fifo_segment.h
index 7922274cecf..ff36f45c4f2 100644
--- a/src/svm/fifo_segment.h
+++ b/src/svm/fifo_segment.h
@@ -26,8 +26,9 @@ typedef enum
   FIFO_SEGMENT_N_FTYPES
 } fifo_segment_ftype_t;
 
-#define FIFO_SEGMENT_MIN_FIFO_SIZE 4096	/* 4kB min fifo size */
-#define FIFO_SEGMENT_MAX_FIFO_SIZE (2 << 30)	/* 2GB max fifo size */
+#define FIFO_SEGMENT_MIN_LOG2_FIFO_SIZE 12	/**< 4kB min fifo size */
+#define FIFO_SEGMENT_MIN_FIFO_SIZE 4096		/**< 4kB min fifo size */
+#define FIFO_SEGMENT_MAX_FIFO_SIZE (2 << 30)	/**< 2GB max fifo size */
 #define FIFO_SEGMENT_ALLOC_BATCH_SIZE 32	/* Allocation quantum */
 
 typedef enum fifo_segment_flags_
@@ -36,21 +37,30 @@ typedef enum fifo_segment_flags_
   FIFO_SEGMENT_F_WILL_DELETE = 1 << 1,
 } fifo_segment_flags_t;
 
-typedef struct
+typedef struct fifo_segment_slice_
 {
   svm_fifo_t *fifos;			/**< Linked list of active RX fifos */
   svm_fifo_t *free_fifos;		/**< Freelists by fifo size  */
   svm_fifo_chunk_t **free_chunks;	/**< Freelists by chunk size */
+  u32 n_fl_chunk_bytes;			/**< Chunk bytes on freelist */
+} fifo_segment_slice_t;
+
+typedef struct
+{
+  fifo_segment_slice_t *slices;		/** Fixed array of slices */
+  ssvm_shared_header_t *ssvm_sh;	/**< Pointer to fs ssvm shared hdr */
+  uword n_free_bytes;			/**< Segment free bytes */
   u32 n_active_fifos;			/**< Number of active fifos */
+  u32 max_log2_chunk_size;		/**< Max log2(chunk size) for fs */
   u8 flags;				/**< Segment flags */
-  u32 n_free_bytes;			/**< Bytes usable for new allocs */
-  u32 n_fl_chunk_bytes;			/**< Chunk bytes on freelist */
+  u8 n_slices;				/**< Number of slices */
 } fifo_segment_header_t;
 
 typedef struct
 {
   ssvm_private_t ssvm;		/**< ssvm segment data */
   fifo_segment_header_t *h;	/**< fifo segment data */
+  u8 n_slices;			/**< number of fifo segment slices */
 } fifo_segment_t;
 
 typedef struct
@@ -90,9 +100,10 @@ void fifo_segment_info (fifo_segment_t * seg, char **address, size_t * size);
  * @param ftype		fifo type @ref fifo_segment_ftype_t
  * @return		new fifo or 0 if alloc failed
  */
-svm_fifo_t *fifo_segment_alloc_fifo (fifo_segment_t * fs,
-				     u32 data_bytes,
-				     fifo_segment_ftype_t ftype);
+svm_fifo_t *fifo_segment_alloc_fifo_w_slice (fifo_segment_t * fs,
+					     u32 slice_index,
+					     u32 data_bytes,
+					     fifo_segment_ftype_t ftype);
 
 /**
  * Free fifo allocated in fifo segment
@@ -111,7 +122,8 @@ void fifo_segment_free_fifo (fifo_segment_t * fs, svm_fifo_t * f);
  * @param batch_size	number of chunks to be allocated
  * @return		0 on success, negative number otherwise
  */
-int fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 batch_size);
+int fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 slice_index,
+				     u32 batch_size);
 
 /**
  * Try to preallocate fifo chunks on segment
@@ -124,8 +136,8 @@ int fifo_segment_prealloc_fifo_hdrs (fifo_segment_t * fs, u32 batch_size);
  * @param batch_size	number of chunks to be allocated
  * @return		0 on success, negative number otherwise
  */
-int fifo_segment_prealloc_fifo_chunks (fifo_segment_t * fs, u32 chunk_size,
-				       u32 batch_size);
+int fifo_segment_prealloc_fifo_chunks (fifo_segment_t * fs, u32 slice_index,
+				       u32 chunk_size, u32 batch_size);
 /**
  * Pre-allocates fifo pairs in fifo segment
  *
@@ -192,9 +204,10 @@ void fifo_segment_update_free_bytes (fifo_segment_t * fs);
  * @param fs		fifo segment
  * @return		free bytes on chunk free lists
  */
-u32 fifo_segment_fl_chunk_bytes (fifo_segment_t * fs);
+uword fifo_segment_fl_chunk_bytes (fifo_segment_t * fs);
 u8 fifo_segment_has_fifos (fifo_segment_t * fs);
-svm_fifo_t *fifo_segment_get_fifo_list (fifo_segment_t * fs);
+svm_fifo_t *fifo_segment_get_slice_fifo_list (fifo_segment_t * fs,
+					      u32 slice_index);
 u32 fifo_segment_num_fifos (fifo_segment_t * fs);
 u32 fifo_segment_num_free_fifos (fifo_segment_t * fs);
 /**
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index ce4c53d9abe..0b7c70790c0 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -78,11 +78,12 @@ typedef struct _svm_fifo
   CLIB_CACHE_LINE_ALIGN_MARK (shared_first);
   u32 size;			/**< size of the fifo in bytes */
   u32 nitems;			/**< usable size (size-1) */
-  u8 flags;			/**< fifo flags */
   svm_fifo_chunk_t *start_chunk;/**< first chunk in fifo chunk list */
   svm_fifo_chunk_t *end_chunk;	/**< end chunk in fifo chunk list */
   svm_fifo_chunk_t *new_chunks;	/**< chunks yet to be added to list */
-  rb_tree_t chunk_lookup;
+  rb_tree_t chunk_lookup;	/**< rbtree for chunk lookup */
+  u8 flags;			/**< fifo flags */
+  u8 slice_index;		/**< segment slice for fifo */
 
     CLIB_CACHE_LINE_ALIGN_MARK (shared_second);
   volatile u32 has_event;	/**< non-zero if deq event exists */