buffers: major cleanup and improvements

This patch introduces following changes:
- deprecated free lists which are not used and not compatible
  with external buffer managers (i.e. DPDK)
- introduces native support for per-numa buffer pools
- significantly improves performance of buffer alloc and free

Change-Id: I4a8e723ae47056717afd6cac0efe87cb731b5be7
Signed-off-by: Damjan Marion <damarion@cisco.com>

1 files changed, 271 insertions, 479 deletions

diff --git a/src/plugins/dpdk/buffer.c b/src/plugins/dpdk/buffer.c
index aed194074cc..73310eff95a 100644
--- a/src/plugins/dpdk/buffer.c
+++ b/src/plugins/dpdk/buffer.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 Cisco and/or its affiliates.
+ * Copyright (c) 2017-2019 Cisco and/or its affiliates.
  * 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:
@@ -12,586 +12,378 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-/*
- * buffer.c: allocate/free network buffers.
- *
- * Copyright (c) 2008 Eliot Dresselhaus
- *
- * Permission is hereby granted, free of charge, to any person obtaining
- * a copy of this software and associated documentation files (the
- * "Software"), to deal in the Software without restriction, including
- * without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sublicense, and/or sell copies of the Software, and to
- * permit persons to whom the Software is furnished to do so, subject to
- * the following conditions:
- *
- * The above copyright notice and this permission notice shall be
- * included in all copies or substantial portions of the Software.
- *
- *  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
- *  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- *  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
- *  NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
- *  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
- *  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
- *  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- */
-
-/**
- * @file
- *
- * Allocate/free network buffers.
- */
 
 #include <unistd.h>
+#include <errno.h>
 
 #include <rte_config.h>
-
-#include <rte_common.h>
-#include <rte_log.h>
-#include <rte_memory.h>
-#include <rte_memzone.h>
-#include <rte_tailq.h>
-#include <rte_eal.h>
-#include <rte_per_lcore.h>
-#include <rte_launch.h>
-#include <rte_atomic.h>
-#include <rte_cycles.h>
-#include <rte_prefetch.h>
-#include <rte_lcore.h>
-#include <rte_per_lcore.h>
-#include <rte_branch_prediction.h>
-#include <rte_interrupts.h>
-#include <rte_vfio.h>
-#include <rte_random.h>
-#include <rte_debug.h>
-#include <rte_ether.h>
-#include <rte_ethdev.h>
-#include <rte_ring.h>
-#include <rte_mempool.h>
 #include <rte_mbuf.h>
-#include <rte_version.h>
+#include <rte_ethdev.h>
+#include <rte_vfio.h>
 
 #include <vlib/vlib.h>
-#include <vlib/unix/unix.h>
-#include <vnet/vnet.h>
-#include <dpdk/device/dpdk.h>
-#include <dpdk/device/dpdk_priv.h>
 #include <dpdk/buffer.h>
 
 STATIC_ASSERT (VLIB_BUFFER_PRE_DATA_SIZE == RTE_PKTMBUF_HEADROOM,
 	       "VLIB_BUFFER_PRE_DATA_SIZE must be equal to RTE_PKTMBUF_HEADROOM");
 
-typedef struct
-{
-  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
-  struct rte_mbuf **mbuf_alloc_list;
-} dpdk_buffer_per_thread_data;
 
 typedef struct
 {
-  int vfio_container_fd;
-  dpdk_buffer_per_thread_data *ptd;
-} dpdk_buffer_main_t;
-
-dpdk_buffer_main_t dpdk_buffer_main;
-
-static_always_inline void
-dpdk_rte_pktmbuf_free (vlib_main_t * vm, u32 thread_index, vlib_buffer_t * b,
-		       int maybe_next)
-{
-  struct rte_mbuf *mb;
-  u32 next, flags;
+  /* must be first */
+  struct rte_pktmbuf_pool_private mbp_priv;
+  u8 buffer_pool_index;
+} dpdk_mempool_private_t;
 
-next:
-  flags = b->flags;
-  next = b->next_buffer;
-  mb = rte_mbuf_from_vlib_buffer (b);
-
-  if (PREDICT_FALSE (b->n_add_refs))
-    {
-      rte_mbuf_refcnt_update (mb, b->n_add_refs);
-      b->n_add_refs = 0;
-    }
-
-  if ((mb = rte_pktmbuf_prefree_seg (mb)))
-    rte_mempool_put (mb->pool, mb);
-
-  if (maybe_next && (flags & VLIB_BUFFER_NEXT_PRESENT))
-    {
-      b = vlib_get_buffer (vm, next);
-      goto next;
-    }
-}
+#ifndef CLIB_MARCH_VARIANT
+struct rte_mempool **dpdk_mempool_by_buffer_pool_index = 0;
+struct rte_mempool **dpdk_no_cache_mempool_by_buffer_pool_index = 0;
 
-/* Make sure free list has at least given number of free buffers. */
-uword
-CLIB_MULTIARCH_FN (dpdk_buffer_fill_free_list) (vlib_main_t * vm,
-						vlib_buffer_free_list_t * fl,
-						uword min_free_buffers)
+clib_error_t *
+dpdk_buffer_pool_init (vlib_main_t * vm, vlib_buffer_pool_t * bp)
 {
-  dpdk_main_t *dm = &dpdk_main;
-  dpdk_buffer_main_t *dbm = &dpdk_buffer_main;
-  struct rte_mbuf **mb;
-  uword n_left, first;
-  word n_alloc;
-  unsigned socket_id = rte_socket_id ();
-  u32 thread_index = vlib_get_thread_index ();
-  dpdk_buffer_per_thread_data *d = vec_elt_at_index (dbm->ptd, thread_index);
-  struct rte_mempool *rmp = dm->pktmbuf_pools[socket_id];
-  dpdk_mempool_private_t *privp = rte_mempool_get_priv (rmp);
-  vlib_buffer_t bt;
+  struct rte_mempool *mp, *nmp;
+  dpdk_mempool_private_t priv;
+  enum rte_iova_mode iova_mode;
   u32 *bi;
+  u8 *name = 0;
 
-  /* Too early? */
-  if (PREDICT_FALSE (rmp == 0))
-    return 0;
-
-  /* Already have enough free buffers on free list? */
-  n_alloc = min_free_buffers - vec_len (fl->buffers);
-  if (n_alloc <= 0)
-    return min_free_buffers;
-
-  /* Always allocate round number of buffers. */
-  n_alloc = round_pow2 (n_alloc, CLIB_CACHE_LINE_BYTES / sizeof (u32));
+  u32 elt_size =
+    sizeof (struct rte_mbuf) + sizeof (vlib_buffer_t) + bp->data_size;
 
-  /* Always allocate new buffers in reasonably large sized chunks. */
-  n_alloc = clib_max (n_alloc, fl->min_n_buffers_each_alloc);
-
-  vec_validate_aligned (d->mbuf_alloc_list, n_alloc - 1,
+  /* create empty mempools */
+  vec_validate_aligned (dpdk_mempool_by_buffer_pool_index, bp->index,
+			CLIB_CACHE_LINE_BYTES);
+  vec_validate_aligned (dpdk_no_cache_mempool_by_buffer_pool_index, bp->index,
 			CLIB_CACHE_LINE_BYTES);
 
-  if (rte_mempool_get_bulk (rmp, (void *) d->mbuf_alloc_list, n_alloc) < 0)
-    return 0;
+  /* normal mempool */
+  name = format (name, "vpp pool %u%c", bp->index, 0);
+  mp = rte_mempool_create_empty ((char *) name, vec_len (bp->buffers),
+				 elt_size, 512, sizeof (priv),
+				 bp->numa_node, 0);
+  vec_reset_length (name);
 
-  clib_memset (&bt, 0, sizeof (vlib_buffer_t));
-  bt.buffer_pool_index = privp->buffer_pool_index;
+  /* non-cached mempool */
+  name = format (name, "vpp pool %u (no cache)%c", bp->index, 0);
+  nmp = rte_mempool_create_empty ((char *) name, vec_len (bp->buffers),
+				  elt_size, 0, sizeof (priv),
+				  bp->numa_node, 0);
+  vec_free (name);
 
-  _vec_len (d->mbuf_alloc_list) = n_alloc;
+  dpdk_mempool_by_buffer_pool_index[bp->index] = mp;
+  dpdk_no_cache_mempool_by_buffer_pool_index[bp->index] = nmp;
 
-  first = vec_len (fl->buffers);
-  vec_resize_aligned (fl->buffers, n_alloc, CLIB_CACHE_LINE_BYTES);
+  rte_mempool_set_ops_byname (mp, "vpp", NULL);
+  rte_mempool_set_ops_byname (nmp, "vpp-no-cache", NULL);
 
-  n_left = n_alloc;
-  mb = d->mbuf_alloc_list;
-  bi = fl->buffers + first;
+  /* Call the mempool priv initializer */
+  priv.mbp_priv.mbuf_data_room_size = VLIB_BUFFER_PRE_DATA_SIZE +
+    VLIB_BUFFER_DATA_SIZE;
+  priv.mbp_priv.mbuf_priv_size = VLIB_BUFFER_HDR_SIZE;
+  priv.buffer_pool_index = bp->index;
+  rte_pktmbuf_pool_init (mp, &priv);
+  rte_pktmbuf_pool_init (nmp, &priv);
 
-  ASSERT (n_left % 8 == 0);
+  iova_mode = rte_eal_iova_mode ();
 
-  while (n_left >= 8)
+  /* populate mempool object buffer header */
+  vec_foreach (bi, bp->buffers)
+  {
+    struct rte_mempool_objhdr *hdr;
+    vlib_buffer_t *b = vlib_get_buffer (vm, *bi);
+    struct rte_mbuf *mb = rte_mbuf_from_vlib_buffer (b);
+    hdr = (struct rte_mempool_objhdr *) RTE_PTR_SUB (mb, sizeof (*hdr));
+    hdr->mp = mp;
+    hdr->iova = (iova_mode == RTE_IOVA_VA) ?
+      pointer_to_uword (mb) : vlib_physmem_get_pa (vm, mb);
+    STAILQ_INSERT_TAIL (&mp->elt_list, hdr, next);
+    STAILQ_INSERT_TAIL (&nmp->elt_list, hdr, next);
+    mp->populated_size++;
+    nmp->populated_size++;
+  }
+
+  /* call the object initializers */
+  rte_mempool_obj_iter (mp, rte_pktmbuf_init, 0);
+
+  /* map DMA pages if at least one physical device exists */
+  if (rte_eth_dev_count_avail ())
     {
-      if (PREDICT_FALSE (n_left < 24))
-	goto no_prefetch;
-
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[16]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[17]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[18]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[19]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[20]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[21]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[22]), STORE);
-      vlib_prefetch_buffer_header (vlib_buffer_from_rte_mbuf (mb[23]), STORE);
-
-    no_prefetch:
-      vlib_get_buffer_indices_with_offset (vm, (void **) mb, bi, 8,
-					   sizeof (struct rte_mbuf));
+      uword i;
+      size_t page_sz;
+      vlib_physmem_map_t *pm;
 
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[0]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[1]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[2]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[3]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[4]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[5]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[6]), &bt);
-      vlib_buffer_copy_template (vlib_buffer_from_rte_mbuf (mb[7]), &bt);
-
-      n_left -= 8;
-      mb += 8;
-      bi += 8;
-    }
+      pm = vlib_physmem_get_map (vm, bp->physmem_map_index);
+      page_sz = 1ULL << pm->log2_page_size;
 
-  if (fl->buffer_init_function)
-    fl->buffer_init_function (vm, fl, fl->buffers + first, n_alloc);
+      for (i = 0; i < pm->n_pages; i++)
+	{
+	  char *va = ((char *) pm->base) + i * page_sz;
+	  uword pa = (iova_mode == RTE_IOVA_VA) ?
+	    pointer_to_uword (va) : pm->page_table[i];
 
-  fl->n_alloc += n_alloc;
+	  if (rte_vfio_dma_map (pointer_to_uword (va), pa, page_sz))
+	    break;
+	}
+    }
 
-  return n_alloc;
+  return 0;
 }
 
-static_always_inline void
-dpdk_prefetch_buffer (vlib_buffer_t * b)
+static int
+dpdk_ops_vpp_alloc (struct rte_mempool *mp)
 {
-  struct rte_mbuf *mb;
-  mb = rte_mbuf_from_vlib_buffer (b);
-  CLIB_PREFETCH (mb, 2 * CLIB_CACHE_LINE_BYTES, LOAD);
-  CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, LOAD);
+  clib_warning ("");
+  return 0;
 }
 
-static_always_inline void
-recycle_or_free (vlib_main_t * vm, vlib_buffer_main_t * bm, u32 bi,
-		 vlib_buffer_t * b)
+static void
+dpdk_ops_vpp_free (struct rte_mempool *mp)
 {
-  u32 thread_index = vlib_get_thread_index ();
-
-  dpdk_rte_pktmbuf_free (vm, thread_index, b, 1);
+  clib_warning ("");
 }
 
+#endif
+
 static_always_inline void
-vlib_buffer_free_inline (vlib_main_t * vm,
-			 u32 * buffers, u32 n_buffers, u32 follow_buffer_next)
+dpdk_ops_vpp_enqueue_one (vlib_buffer_t * bt, void *obj)
 {
-  vlib_buffer_main_t *bm = vm->buffer_main;
-  vlib_buffer_t *bufp[n_buffers], **b = bufp;
-  u32 thread_index = vm->thread_index;
-  int i = 0;
-  u32 simple_mask = VLIB_BUFFER_NEXT_PRESENT;
-  u32 n_left, *bi;
-  u32 (*cb) (vlib_main_t * vm, u32 * buffers, u32 n_buffers,
-	     u32 follow_buffer_next);
-
-  cb = bm->buffer_free_callback;
+  /* Only non-replicated packets (b->ref_count == 1) expected */
 
-  if (PREDICT_FALSE (cb != 0))
-    n_buffers = (*cb) (vm, buffers, n_buffers, follow_buffer_next);
+  struct rte_mbuf *mb = obj;
+  vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb);
+  ASSERT (b->ref_count == 1);
+  ASSERT (b->buffer_pool_index == bt->buffer_pool_index);
+  vlib_buffer_copy_template (b, bt);
+}
 
-  if (!n_buffers)
-    return;
+int
+CLIB_MULTIARCH_FN (dpdk_ops_vpp_enqueue) (struct rte_mempool * mp,
+					  void *const *obj_table, unsigned n)
+{
+  const int batch_size = 32;
+  vlib_main_t *vm = vlib_get_main ();
+  vlib_buffer_t bt;
+  dpdk_mempool_private_t *privp = rte_mempool_get_priv (mp);
+  u8 buffer_pool_index = privp->buffer_pool_index;
+  vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index);
+  u32 bufs[batch_size];
+  u32 n_left = n;
+  void *const *obj = obj_table;
 
-  n_left = n_buffers;
-  bi = buffers;
-  b = bufp;
-  vlib_get_buffers (vm, bi, b, n_buffers);
+  vlib_buffer_copy_template (&bt, &bp->buffer_template);
 
   while (n_left >= 4)
     {
-      u32 or_flags;
-      vlib_buffer_t **p;
-
-      if (n_left < 16)
-	goto no_prefetch;
-
-      p = b + 12;
-      dpdk_prefetch_buffer (p[0]);
-      dpdk_prefetch_buffer (p[1]);
-      dpdk_prefetch_buffer (p[2]);
-      dpdk_prefetch_buffer (p[3]);
-    no_prefetch:
-
-      for (i = 0; i < 4; i++)
-	VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[i]);
-
-      or_flags = b[0]->flags | b[1]->flags | b[2]->flags | b[3]->flags;
-
-      if (or_flags & simple_mask)
-	{
-	  recycle_or_free (vm, bm, bi[0], b[0]);
-	  recycle_or_free (vm, bm, bi[1], b[1]);
-	  recycle_or_free (vm, bm, bi[2], b[2]);
-	  recycle_or_free (vm, bm, bi[3], b[3]);
-	}
-      else
-	{
-	  dpdk_rte_pktmbuf_free (vm, thread_index, b[0], 0);
-	  dpdk_rte_pktmbuf_free (vm, thread_index, b[1], 0);
-	  dpdk_rte_pktmbuf_free (vm, thread_index, b[2], 0);
-	  dpdk_rte_pktmbuf_free (vm, thread_index, b[3], 0);
-	}
-      bi += 4;
-      b += 4;
+      dpdk_ops_vpp_enqueue_one (&bt, obj[0]);
+      dpdk_ops_vpp_enqueue_one (&bt, obj[1]);
+      dpdk_ops_vpp_enqueue_one (&bt, obj[2]);
+      dpdk_ops_vpp_enqueue_one (&bt, obj[3]);
+      obj += 4;
       n_left -= 4;
     }
+
   while (n_left)
     {
-      VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[0]);
-      recycle_or_free (vm, bm, bi[0], b[0]);
-      bi += 1;
-      b += 1;
+      dpdk_ops_vpp_enqueue_one (&bt, obj[0]);
+      obj += 1;
       n_left -= 1;
     }
-}
-
-void
-CLIB_MULTIARCH_FN (dpdk_buffer_free) (vlib_main_t * vm, u32 * buffers,
-				      u32 n_buffers)
-{
-  vlib_buffer_free_inline (vm, buffers, n_buffers,	/* follow_buffer_next */
-			   1);
-}
-
-void
-CLIB_MULTIARCH_FN (dpdk_buffer_free_no_next) (vlib_main_t * vm, u32 * buffers,
-					      u32 n_buffers)
-{
-  vlib_buffer_free_inline (vm, buffers, n_buffers,	/* follow_buffer_next */
-			   0);
-}
 
-#ifndef CLIB_MARCH_VARIANT
-clib_error_t *
-dpdk_pool_create (vlib_main_t * vm, u8 * pool_name, u32 elt_size,
-		  u32 num_elts, u32 pool_priv_size, u16 cache_size, u8 numa,
-		  struct rte_mempool **_mp, u32 * map_index)
-{
-  struct rte_mempool *mp;
-  enum rte_iova_mode iova_mode;
-  dpdk_mempool_private_t priv;
-  vlib_physmem_map_t *pm;
-  clib_error_t *error = 0;
-  size_t min_chunk_size, align;
-  int map_dma = 1;
-  u32 size;
-  i32 ret;
-  uword i;
-
-  mp = rte_mempool_create_empty ((char *) pool_name, num_elts, elt_size,
-				 512, pool_priv_size, numa, 0);
-  if (!mp)
-    return clib_error_return (0, "failed to create %s", pool_name);
-
-  rte_mempool_set_ops_byname (mp, RTE_MBUF_DEFAULT_MEMPOOL_OPS, NULL);
-
-  size = rte_mempool_op_calc_mem_size_default (mp, num_elts, 21,
-					       &min_chunk_size, &align);
-
-  if ((error = vlib_physmem_shared_map_create (vm, (char *) pool_name, size,
-					       0, numa, map_index)))
+  while (n >= batch_size)
     {
-      rte_mempool_free (mp);
-      return error;
+      vlib_get_buffer_indices_with_offset (vm, (void **) obj_table, bufs,
+					   batch_size,
+					   sizeof (struct rte_mbuf));
+      vlib_buffer_pool_put (vm, buffer_pool_index, bufs, batch_size);
+      n -= batch_size;
+      obj_table += batch_size;
     }
-  pm = vlib_physmem_get_map (vm, *map_index);
-
-  /* Call the mempool priv initializer */
-  priv.mbp_priv.mbuf_data_room_size = VLIB_BUFFER_PRE_DATA_SIZE +
-    VLIB_BUFFER_DATA_SIZE;
-  priv.mbp_priv.mbuf_priv_size = VLIB_BUFFER_HDR_SIZE;
-  rte_pktmbuf_pool_init (mp, &priv);
 
-  if (rte_eth_dev_count_avail () == 0)
-    map_dma = 0;
-
-  iova_mode = rte_eal_iova_mode ();
-  for (i = 0; i < pm->n_pages; i++)
+  if (n)
     {
-      size_t page_sz = 1ULL << pm->log2_page_size;
-      char *va = ((char *) pm->base) + i * page_sz;
-      uword pa = iova_mode == RTE_IOVA_VA ?
-	pointer_to_uword (va) : pm->page_table[i];
-      ret = rte_mempool_populate_iova (mp, va, pa, page_sz, 0, 0);
-      if (ret < 0)
-	{
-	  rte_mempool_free (mp);
-	  return clib_error_return (0, "failed to populate %s", pool_name);
-	}
-      /* -1 likely means there is no PCI devices assigned to vfio
-         container or noiommu mode is used  so we stop trying */
-      if (map_dma && rte_vfio_dma_map (pointer_to_uword (va), pa, page_sz))
-	map_dma = 0;
+      vlib_get_buffer_indices_with_offset (vm, (void **) obj_table, bufs,
+					   n, sizeof (struct rte_mbuf));
+      vlib_buffer_pool_put (vm, buffer_pool_index, bufs, batch_size);
     }
 
-  _mp[0] = mp;
-
   return 0;
 }
 
-clib_error_t *
-dpdk_buffer_pool_create (vlib_main_t * vm, unsigned num_mbufs,
-			 unsigned socket_id)
-{
-  dpdk_main_t *dm = &dpdk_main;
-  struct rte_mempool *rmp;
-  clib_error_t *error = 0;
-  u8 *pool_name;
-  u32 elt_size, i;
-  u32 map_index;
+CLIB_MARCH_FN_REGISTRATION (dpdk_ops_vpp_enqueue);
 
-  vec_validate_aligned (dm->pktmbuf_pools, socket_id, CLIB_CACHE_LINE_BYTES);
-
-  /* pool already exists, nothing to do */
-  if (dm->pktmbuf_pools[socket_id])
-    return 0;
-
-  pool_name = format (0, "dpdk_mbuf_pool_socket%u%c", socket_id, 0);
-
-  elt_size = sizeof (struct rte_mbuf) +
-    VLIB_BUFFER_HDR_SIZE /* priv size */  +
-    VLIB_BUFFER_PRE_DATA_SIZE + VLIB_BUFFER_DATA_SIZE;	/*data room size */
-
-  error = dpdk_pool_create (vm, pool_name, elt_size, num_mbufs,
-			    sizeof (dpdk_mempool_private_t), 512, socket_id,
-			    &rmp, &map_index);
-
-  vec_free (pool_name);
+static_always_inline void
+dpdk_ops_vpp_enqueue_no_cache_one (vlib_main_t * vm, struct rte_mempool *old,
+				   struct rte_mempool *new, void *obj,
+				   vlib_buffer_t * bt)
+{
+  struct rte_mbuf *mb = obj;
+  vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb);
 
-  if (!error)
+  if (clib_atomic_sub_fetch (&b->ref_count, 1) == 0)
     {
-      /* call the object initializers */
-      rte_mempool_obj_iter (rmp, rte_pktmbuf_init, 0);
-
-      dpdk_mempool_private_t *privp = rte_mempool_get_priv (rmp);
-      privp->buffer_pool_index =
-	vlib_buffer_register_physmem_map (vm, map_index);
+      u32 bi = vlib_get_buffer_index (vm, b);
+      mb->pool = new;
+      vlib_buffer_copy_template (b, bt);
+      vlib_buffer_pool_put (vm, bt->buffer_pool_index, &bi, 1);
+      return;
+    }
+}
 
-      dm->pktmbuf_pools[socket_id] = rmp;
+int
+CLIB_MULTIARCH_FN (dpdk_ops_vpp_enqueue_no_cache) (struct rte_mempool * cmp,
+						   void *const *obj_table,
+						   unsigned n)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  vlib_buffer_t bt;
+  dpdk_mempool_private_t *privp = rte_mempool_get_priv (cmp);
+  struct rte_mempool *mp;
+  mp = dpdk_mempool_by_buffer_pool_index[privp->buffer_pool_index];
+  u8 buffer_pool_index = privp->buffer_pool_index;
+  vlib_buffer_pool_t *bp = vlib_get_buffer_pool (vm, buffer_pool_index);
+  vlib_buffer_copy_template (&bt, &bp->buffer_template);
 
-      return 0;
+  while (n >= 4)
+    {
+      dpdk_ops_vpp_enqueue_no_cache_one (vm, cmp, mp, obj_table[0], &bt);
+      dpdk_ops_vpp_enqueue_no_cache_one (vm, cmp, mp, obj_table[1], &bt);
+      dpdk_ops_vpp_enqueue_no_cache_one (vm, cmp, mp, obj_table[2], &bt);
+      dpdk_ops_vpp_enqueue_no_cache_one (vm, cmp, mp, obj_table[3], &bt);
+      obj_table += 4;
+      n -= 4;
     }
 
-  clib_error_report (error);
-
-  /* no usable pool for this socket, try to use pool from another one */
-  for (i = 0; i < vec_len (dm->pktmbuf_pools); i++)
+  while (n)
     {
-      if (dm->pktmbuf_pools[i])
-	{
-	  clib_warning ("WARNING: Failed to allocate mempool for CPU socket "
-			"%u. Threads running on socket %u will use socket %u "
-			"mempool.", socket_id, socket_id, i);
-	  dm->pktmbuf_pools[socket_id] = dm->pktmbuf_pools[i];
-	  return 0;
-	}
+      dpdk_ops_vpp_enqueue_no_cache_one (vm, cmp, mp, obj_table[0], &bt);
+      obj_table += 1;
+      n -= 1;
     }
 
-  return clib_error_return (0, "failed to allocate mempool on socket %u",
-			    socket_id);
+  return 0;
 }
 
-#if CLIB_DEBUG > 0
+CLIB_MARCH_FN_REGISTRATION (dpdk_ops_vpp_enqueue_no_cache);
 
-u32 *vlib_buffer_state_validation_lock;
-uword *vlib_buffer_state_validation_hash;
-void *vlib_buffer_state_heap;
-
-static clib_error_t *
-buffer_state_validation_init (vlib_main_t * vm)
+int
+CLIB_MULTIARCH_FN (dpdk_ops_vpp_dequeue) (struct rte_mempool * mp,
+					  void **obj_table, unsigned n)
 {
-  void *oldheap;
-
-  vlib_buffer_state_heap =
-    mheap_alloc_with_lock (0, 10 << 20, 0 /* locked */ );
-  oldheap = clib_mem_set_heap (vlib_buffer_state_heap);
+  const int batch_size = 32;
+  vlib_main_t *vm = vlib_get_main ();
+  u32 bufs[batch_size], total = 0, n_alloc = 0;
+  dpdk_mempool_private_t *privp = rte_mempool_get_priv (mp);
+  u8 buffer_pool_index = privp->buffer_pool_index;
+  void **obj = obj_table;
+
+  while (n >= batch_size)
+    {
+      n_alloc = vlib_buffer_alloc_from_pool (vm, bufs, batch_size,
+					     buffer_pool_index);
+      if (n_alloc != batch_size)
+	goto alloc_fail;
+
+      vlib_get_buffers_with_offset (vm, bufs, obj, batch_size,
+				    -(i32) sizeof (struct rte_mbuf));
+      total += batch_size;
+      obj += batch_size;
+      n -= batch_size;
+    }
 
-  vlib_buffer_state_validation_hash = hash_create (0, sizeof (uword));
-  vec_validate_aligned (vlib_buffer_state_validation_lock, 0,
-			CLIB_CACHE_LINE_BYTES);
-  clib_mem_set_heap (oldheap);
-  return 0;
-}
+  if (n)
+    {
+      n_alloc = vlib_buffer_alloc_from_pool (vm, bufs, n, buffer_pool_index);
 
-VLIB_INIT_FUNCTION (buffer_state_validation_init);
-#endif
+      if (n_alloc != n)
+	goto alloc_fail;
 
-#if CLI_DEBUG
-struct dpdk_validate_buf_result
-{
-  u32 invalid;
-  u32 uninitialized;
-};
+      vlib_get_buffers_with_offset (vm, bufs, obj, n,
+				    -(i32) sizeof (struct rte_mbuf));
+    }
 
-#define DPDK_TRAJECTORY_POISON 31
+  return 0;
 
-static void
-dpdk_buffer_validate_trajectory (struct rte_mempool *mp, void *opaque,
-				 void *obj, unsigned obj_idx)
-{
-  vlib_buffer_t *b;
-  struct dpdk_validate_buf_result *counter = opaque;
-  b = vlib_buffer_from_rte_mbuf ((struct rte_mbuf *) obj);
-  if (b->pre_data[0] != 0)
+alloc_fail:
+  /* dpdk doesn't support partial alloc, so we need to return what we
+     already got */
+  if (n_alloc)
+    vlib_buffer_pool_put (vm, buffer_pool_index, bufs, n_alloc);
+  obj = obj_table;
+  while (total)
     {
-      if (b->pre_data[0] == DPDK_TRAJECTORY_POISON)
-	counter->uninitialized++;
-      else
-	counter->invalid++;
+      vlib_get_buffer_indices_with_offset (vm, obj, bufs, batch_size,
+					   sizeof (struct rte_mbuf));
+      vlib_buffer_pool_put (vm, buffer_pool_index, bufs, batch_size);
+
+      obj += batch_size;
+      total -= batch_size;
     }
+  return -ENOENT;
 }
 
-int
-dpdk_buffer_validate_trajectory_all (u32 * uninitialized)
+CLIB_MARCH_FN_REGISTRATION (dpdk_ops_vpp_dequeue);
+
+#ifndef CLIB_MARCH_VARIANT
+
+static int
+dpdk_ops_vpp_dequeue_no_cache (struct rte_mempool *mp, void **obj_table,
+			       unsigned n)
 {
-  dpdk_main_t *dm = &dpdk_main;
-  struct dpdk_validate_buf_result counter = { 0 };
-  int i;
-
-  for (i = 0; i < vec_len (dm->pktmbuf_pools); i++)
-    rte_mempool_obj_iter (dm->pktmbuf_pools[i],
-			  dpdk_buffer_validate_trajectory, &counter);
-  if (uninitialized)
-    *uninitialized = counter.uninitialized;
-  return counter.invalid;
+  clib_error ("bug");
+  return 0;
 }
 
-static void
-dpdk_buffer_poison_trajectory (struct rte_mempool *mp, void *opaque,
-			       void *obj, unsigned obj_idx)
+static unsigned
+dpdk_ops_vpp_get_count (const struct rte_mempool *mp)
 {
-  vlib_buffer_t *b;
-  b = vlib_buffer_from_rte_mbuf ((struct rte_mbuf *) obj);
-  b->pre_data[0] = DPDK_TRAJECTORY_POISON;
+  clib_warning ("");
+  return 0;
 }
 
-void
-dpdk_buffer_poison_trajectory_all (void)
+static unsigned
+dpdk_ops_vpp_get_count_no_cache (const struct rte_mempool *mp)
 {
-  dpdk_main_t *dm = &dpdk_main;
-  int i;
-
-  for (i = 0; i < vec_len (dm->pktmbuf_pools); i++)
-    rte_mempool_obj_iter (dm->pktmbuf_pools[i], dpdk_buffer_poison_trajectory,
-			  0);
+  dpdk_mempool_private_t *privp;
+  struct rte_mempool *cmp;
+  privp = rte_mempool_get_priv ((struct rte_mempool *) mp);
+  cmp = dpdk_no_cache_mempool_by_buffer_pool_index[privp->buffer_pool_index];
+  return dpdk_ops_vpp_get_count (cmp);
 }
-#endif
 
-static clib_error_t *
-dpdk_buffer_init (vlib_main_t * vm)
+clib_error_t *
+dpdk_buffer_pools_create (vlib_main_t * vm)
 {
-  dpdk_buffer_main_t *dbm = &dpdk_buffer_main;
-  vlib_thread_main_t *tm = vlib_get_thread_main ();
-
-  vec_validate_aligned (dbm->ptd, tm->n_vlib_mains - 1,
-			CLIB_CACHE_LINE_BYTES);
-
-  dbm->vfio_container_fd = -1;
-
+  clib_error_t *err;
+  vlib_buffer_pool_t *bp;
+
+  struct rte_mempool_ops ops = { };
+
+  strncpy (ops.name, "vpp", 4);
+  ops.alloc = dpdk_ops_vpp_alloc;
+  ops.free = dpdk_ops_vpp_free;
+  ops.get_count = dpdk_ops_vpp_get_count;
+  ops.enqueue = CLIB_MARCH_FN_POINTER (dpdk_ops_vpp_enqueue);
+  ops.dequeue = CLIB_MARCH_FN_POINTER (dpdk_ops_vpp_dequeue);
+  rte_mempool_register_ops (&ops);
+
+  strncpy (ops.name, "vpp-no-cache", 13);
+  ops.get_count = dpdk_ops_vpp_get_count_no_cache;
+  ops.enqueue = CLIB_MARCH_FN_POINTER (dpdk_ops_vpp_enqueue_no_cache);
+  ops.dequeue = dpdk_ops_vpp_dequeue_no_cache;
+  rte_mempool_register_ops (&ops);
+
+  /* *INDENT-OFF* */
+  vec_foreach (bp, vm->buffer_main->buffer_pools)
+    if (bp->start && (err = dpdk_buffer_pool_init (vm, bp)))
+      return err;
+  /* *INDENT-ON* */
   return 0;
 }
 
-VLIB_INIT_FUNCTION (dpdk_buffer_init);
-
-/* *INDENT-OFF* */
-VLIB_BUFFER_REGISTER_CALLBACKS (dpdk, static) = {
-  .vlib_buffer_fill_free_list_cb = &dpdk_buffer_fill_free_list,
-  .vlib_buffer_free_cb = &dpdk_buffer_free,
-  .vlib_buffer_free_no_next_cb = &dpdk_buffer_free_no_next,
-};
-/* *INDENT-ON* */
-
-#if __x86_64__
-vlib_buffer_fill_free_list_cb_t __clib_weak dpdk_buffer_fill_free_list_avx512;
-vlib_buffer_fill_free_list_cb_t __clib_weak dpdk_buffer_fill_free_list_avx2;
-vlib_buffer_free_cb_t __clib_weak dpdk_buffer_free_avx512;
-vlib_buffer_free_cb_t __clib_weak dpdk_buffer_free_avx2;
-vlib_buffer_free_no_next_cb_t __clib_weak dpdk_buffer_free_no_next_avx512;
-vlib_buffer_free_no_next_cb_t __clib_weak dpdk_buffer_free_no_next_avx2;
-
-static void __clib_constructor
-dpdk_input_multiarch_select (void)
-{
-  vlib_buffer_callbacks_t *cb = &__dpdk_buffer_callbacks;
-  if (dpdk_buffer_fill_free_list_avx512 && clib_cpu_supports_avx512f ())
-    {
-      cb->vlib_buffer_fill_free_list_cb = dpdk_buffer_fill_free_list_avx512;
-      cb->vlib_buffer_free_cb = dpdk_buffer_free_avx512;
-      cb->vlib_buffer_free_no_next_cb = dpdk_buffer_free_no_next_avx512;
-    }
-  else if (dpdk_buffer_fill_free_list_avx2 && clib_cpu_supports_avx2 ())
-    {
-      cb->vlib_buffer_fill_free_list_cb = dpdk_buffer_fill_free_list_avx2;
-      cb->vlib_buffer_free_cb = dpdk_buffer_free_avx2;
-      cb->vlib_buffer_free_no_next_cb = dpdk_buffer_free_no_next_avx2;
-    }
-}
-#endif
+VLIB_BUFFER_SET_EXT_HDR_SIZE (sizeof (struct rte_mempool_objhdr) +
+			      sizeof (struct rte_mbuf));
+
 #endif
 
 /** @endcond */