Use rte_mempool private data for storing vlib_buffer_t

Change-Id: If3fc88a35bc0b736376113a39667caea42802ea1
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 80 insertions, 53 deletions

diff --git a/vlib/vlib/buffer.h b/vlib/vlib/buffer.h
index 6322481b696..9c148ef2f21 100644
--- a/vlib/vlib/buffer.h
+++ b/vlib/vlib/buffer.h
@@ -45,7 +45,16 @@
 #include <vppinfra/serialize.h>
 #include <vppinfra/vector.h>
 #include <vlib/error.h>		/* for vlib_error_t */
+
+#if DPDK > 0
+#include <rte_config.h>
+#define VLIB_BUFFER_DATA_SIZE		(2048)
+#define VLIB_BUFFER_PRE_DATA_SIZE	RTE_PKTMBUF_HEADROOM
+#else
 #include <vlib/config.h>        /* for __PRE_DATA_SIZE */
+#define VLIB_BUFFER_DATA_SIZE		(512)
+#define VLIB_BUFFER_PRE_DATA_SIZE	__PRE_DATA_SIZE
+#endif
 
 #ifdef CLIB_HAVE_VEC128
 typedef u8x16 vlib_copy_unit_t;
@@ -62,6 +71,7 @@ typedef uword vlib_copy_unit_t;
     
 /* VLIB buffer representation. */
 typedef struct {
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline0);
   /* Offset within data[] that we are currently processing.
      If negative current header points into predata area. */ 
   i16 current_data;  /**< signed offset in data[], pre_data[]  
@@ -124,23 +134,25 @@ typedef struct {
   u32 opaque[8]; /**< Opaque data used by sub-graphs for their own purposes. 
                     See .../vnet/vnet/buffer.h
                  */
-  /***** end of first cache line */
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline1);
 
   u32 opaque2[16];  /**< More opaque data, in its own cache line */
 
   /***** end of second cache line */
-  u8 pre_data [__PRE_DATA_SIZE]; /**< Space for inserting data 
-                                     before buffer start.  
-                                     Packet rewrite string will be
-                                     rewritten backwards and may extend 
-                                     back before buffer->data[0].
-                                     Must come directly before packet data.
-                                 */
-
-#define VLIB_BUFFER_PRE_DATA_SIZE (ARRAY_LEN (((vlib_buffer_t *)0)->pre_data))
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline2);
+  u8 pre_data [VLIB_BUFFER_PRE_DATA_SIZE]; /**< Space for inserting data
+                                               before buffer start.
+                                               Packet rewrite string will be
+                                               rewritten backwards and may extend
+                                               back before buffer->data[0].
+                                               Must come directly before packet data.
+                                            */
+
   u8 data[0]; /**< Packet data. Hardware DMA here */
 } vlib_buffer_t;  /* Must be a multiple of 64B. */
 
+#define VLIB_BUFFER_HDR_SIZE  (sizeof(vlib_buffer_t) - VLIB_BUFFER_PRE_DATA_SIZE)
+
 /** \brief Prefetch buffer metadata.
     The first 64 bytes of buffer contains most header information
 
@@ -283,13 +295,7 @@ typedef struct {
      initializing static data for each packet generated. */
   vlib_buffer_free_list_t * buffer_free_list_pool;
 #define VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX (0)
-
-#if DPDK == 1
-/* must be same as dpdk buffer size */
-#define VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES (2048)
-#else
-#define VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES (512)
-#endif
+#define VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES VLIB_BUFFER_DATA_SIZE
 
   /* Hash table mapping buffer size (rounded to next unit of
      sizeof (vlib_buffer_t)) to free list index. */
@@ -357,6 +363,11 @@ serialize_vlib_buffer_n_bytes (serialize_main_t * m)
   return sm->tx.n_total_data_bytes + s->current_buffer_index + vec_len (s->overflow_buffer);
 }
 
+#if DPDK > 0
+#define rte_mbuf_from_vlib_buffer(x) (((struct rte_mbuf *)x) - 1)
+#define vlib_buffer_from_rte_mbuf(x) ((vlib_buffer_t *)(x+1))
+#endif
+
 /*
  */
 
diff --git a/vlib/vlib/buffer_funcs.h b/vlib/vlib/buffer_funcs.h
index eea417a94f0..f7bdb12d680 100644
--- a/vlib/vlib/buffer_funcs.h
+++ b/vlib/vlib/buffer_funcs.h
@@ -254,7 +254,7 @@ u8 * vlib_validate_buffers (vlib_main_t * vm,
 
 clib_error_t *
 vlib_buffer_pool_create(vlib_main_t * vm, unsigned num_mbufs,
-                        unsigned mbuf_size, unsigned socket_id);
+                        unsigned socket_id);
 
 /** \brief Allocate buffers into supplied array
 
@@ -425,11 +425,9 @@ vlib_buffer_chain_init(vlib_buffer_t *first)
   first->flags &= ~VLIB_BUFFER_NEXT_PRESENT;
   first->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
 #if DPDK == 1
-  (((struct rte_mbuf *) first) - 1)->nb_segs = 1;
-  (((struct rte_mbuf *) first) - 1)->next = 0;
-  (((struct rte_mbuf *) first) - 1)->pkt_len = 0;
-  (((struct rte_mbuf *) first) - 1)->data_len = 0;
-  (((struct rte_mbuf *) first) - 1)->data_off = RTE_PKTMBUF_HEADROOM + first->current_data;
+  struct rte_mbuf * mb = rte_mbuf_from_vlib_buffer(first);
+  rte_pktmbuf_reset(mb);
+  mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + first->current_data;
 #endif
 }
 
@@ -446,11 +444,17 @@ vlib_buffer_chain_buffer(vlib_main_t *vm,
   next_buffer->current_length = 0;
   next_buffer->flags &= ~VLIB_BUFFER_NEXT_PRESENT;
 #if DPDK == 1
-  (((struct rte_mbuf *) first) - 1)->nb_segs++;
-  (((struct rte_mbuf *) last) - 1)->next = (((struct rte_mbuf *) next_buffer) - 1);
-  (((struct rte_mbuf *) next_buffer) - 1)->data_len = 0;
-  (((struct rte_mbuf *) next_buffer) - 1)->data_off = RTE_PKTMBUF_HEADROOM + next_buffer->current_data;
-  (((struct rte_mbuf *) next_buffer) - 1)->next = 0;
+  struct rte_mbuf * mb;
+  mb = rte_mbuf_from_vlib_buffer(first);
+  mb->nb_segs++;
+
+  mb = rte_mbuf_from_vlib_buffer(last);
+  mb->next = rte_mbuf_from_vlib_buffer(next_buffer);
+
+  mb = rte_mbuf_from_vlib_buffer(next_buffer);
+  mb->data_len = 0;
+  mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + next_buffer->current_data;
+  mb->next = 0;
 #endif
   return next_buffer;
 }
@@ -468,8 +472,10 @@ vlib_buffer_chain_increase_length(vlib_buffer_t *first,
   if (first != last)
     first->total_length_not_including_first_buffer += len;
 #if DPDK == 1
-  (((struct rte_mbuf *) first) - 1)->pkt_len += len;
-  (((struct rte_mbuf *) last) - 1)->data_len += len;
+  struct rte_mbuf * mb_first = rte_mbuf_from_vlib_buffer(first);
+  struct rte_mbuf * mb_last = rte_mbuf_from_vlib_buffer(last);
+  mb_first->pkt_len += len;
+  mb_last->data_len += len;
 #endif
 }
 
@@ -589,6 +595,11 @@ vlib_buffer_init_for_free_list (vlib_buffer_t * _dst,
   vlib_buffer_union_t * dst = (vlib_buffer_union_t *) _dst;
   vlib_buffer_union_t * src = (vlib_buffer_union_t *) &fl->buffer_init_template;
 
+  /* Make sure vlib_buffer_t is cacheline aligned and sized */
+  ASSERT(STRUCT_OFFSET_OF(vlib_buffer_t, cacheline0) == 0);
+  ASSERT(STRUCT_OFFSET_OF(vlib_buffer_t, cacheline1) == CLIB_CACHE_LINE_BYTES);
+  ASSERT(STRUCT_OFFSET_OF(vlib_buffer_t, cacheline2) == CLIB_CACHE_LINE_BYTES * 2);
+
   /* Make sure buffer template is sane. */
   ASSERT (fl->index == fl->buffer_init_template.free_list_index);
 
diff --git a/vlib/vlib/dpdk_buffer.c b/vlib/vlib/dpdk_buffer.c
index 04a6447d633..db1fde18c77 100644
--- a/vlib/vlib/dpdk_buffer.c
+++ b/vlib/vlib/dpdk_buffer.c
@@ -66,12 +66,10 @@
 
 #include <vlib/vlib.h>
 
-phys_addr_t          __attribute__ ((weak)) rte_mem_virt2phy();
-int                  __attribute__ ((weak)) rte_eal_has_hugepages();
-unsigned             __attribute__ ((weak)) rte_socket_id();
-struct rte_mempool * __attribute__ ((weak)) rte_mempool_create();
-void                 __attribute__ ((weak)) rte_pktmbuf_init();
-void                 __attribute__ ((weak)) rte_pktmbuf_pool_init();
+#pragma weak rte_mem_virt2phy
+#pragma weak rte_eal_has_hugepages
+#pragma weak rte_socket_id
+#pragma weak rte_pktmbuf_pool_create
 
 uword vlib_buffer_length_in_chain_slow_path (vlib_main_t * vm, vlib_buffer_t * b_first)
 {
@@ -400,13 +398,13 @@ del_free_list (vlib_main_t * vm, vlib_buffer_free_list_t * f)
 
   for (i = 0; i < vec_len (f->unaligned_buffers); i++) {
       b = vlib_get_buffer (vm, f->unaligned_buffers[i]);
-      mb = ((struct rte_mbuf *)b)-1;
+      mb = rte_mbuf_from_vlib_buffer(b);
       ASSERT(rte_mbuf_refcnt_read(mb) == 1);
       rte_pktmbuf_free (mb);
   }
   for (i = 0; i < vec_len (f->aligned_buffers); i++) {
       b = vlib_get_buffer (vm, f->aligned_buffers[i]);
-      mb = ((struct rte_mbuf *)b)-1;
+      mb = rte_mbuf_from_vlib_buffer(b);
       ASSERT(rte_mbuf_refcnt_read(mb) == 1);
       rte_pktmbuf_free (mb);
   }
@@ -487,7 +485,7 @@ fill_free_list (vlib_main_t * vm,
       mb->data_off = RTE_PKTMBUF_HEADROOM;
       mb->nb_segs = 1;
 
-      b = (vlib_buffer_t *)(mb+1);
+      b = vlib_buffer_from_rte_mbuf(mb);
       bi = vlib_get_buffer_index (vm, b);
       
       vec_add1_aligned (fl->aligned_buffers, bi, sizeof (vlib_copy_unit_t));
@@ -726,7 +724,7 @@ vlib_buffer_free_inline (vlib_main_t * vm,
         {
 	  if (PREDICT_TRUE (b->clone_count == 0))
 	    {
-	      mb = ((struct rte_mbuf *)b)-1;
+	      mb = rte_mbuf_from_vlib_buffer(b);
 	      ASSERT(rte_mbuf_refcnt_read(mb) == 1);
 	      rte_pktmbuf_free (mb);
 	    }
@@ -820,7 +818,7 @@ vlib_packet_template_get_packet (vlib_main_t * vm,
 
   /* Fix up mbuf header length fields */
   struct rte_mbuf * mb;
-  mb = ((struct rte_mbuf *)b) - 1;
+  mb = rte_mbuf_from_vlib_buffer(b);
   mb->data_len = b->current_length;
   mb->pkt_len = b->current_length;
 
@@ -916,22 +914,26 @@ vlib_buffer_chain_append_data_with_alloc(vlib_main_t *vm,
 void vlib_buffer_chain_validate (vlib_main_t * vm, vlib_buffer_t * b_first)
 {
   vlib_buffer_t *b = b_first, *prev = b_first;
-  struct rte_mbuf *mb_first = ((struct rte_mbuf *) b) - 1;
+  struct rte_mbuf *mb_prev, *mb, *mb_first;
 
-  mb_first->pkt_len = mb_first-> data_len = b_first->current_length;
+  mb_first = rte_mbuf_from_vlib_buffer(b_first);
+
+  mb_first->pkt_len = mb_first->data_len = b_first->current_length;
   while (b->flags & VLIB_BUFFER_NEXT_PRESENT) {
       b = vlib_get_buffer(vm, b->next_buffer);
+      mb = rte_mbuf_from_vlib_buffer(b);
+      mb_prev = rte_mbuf_from_vlib_buffer(prev);
       mb_first->nb_segs++;
       mb_first->pkt_len += b->current_length;
-      (((struct rte_mbuf *) prev) - 1)->next = (((struct rte_mbuf *) b) - 1);
-      (((struct rte_mbuf *) b) - 1)->data_len = b->current_length;
+      mb_prev->next = mb;
+      mb->data_len = b->current_length;
       prev = b;
   }
 }
 
 clib_error_t *
 vlib_buffer_pool_create(vlib_main_t * vm, unsigned num_mbufs,
-                        unsigned mbuf_size, unsigned socket_id)
+                        unsigned socket_id)
 {
   vlib_buffer_main_t * bm = vm->buffer_main;
   vlib_physmem_main_t * vpm = &vm->physmem_main;
@@ -939,7 +941,7 @@ vlib_buffer_pool_create(vlib_main_t * vm, unsigned num_mbufs,
   uword new_start, new_size;
   int i;
 
-  if (!rte_mempool_create)
+  if (!rte_pktmbuf_pool_create)
     return clib_error_return (0, "not linked with DPDK");
 
   vec_validate_aligned(bm->pktmbuf_pools, socket_id, CLIB_CACHE_LINE_BYTES);
@@ -949,12 +951,15 @@ vlib_buffer_pool_create(vlib_main_t * vm, unsigned num_mbufs,
     return 0;
 
   u8 * pool_name = format(0, "mbuf_pool_socket%u%c",socket_id, 0);
-  rmp = rte_mempool_create((char *) pool_name,
-                           num_mbufs, mbuf_size, 512,
-                           sizeof(struct rte_pktmbuf_pool_private),
-                           rte_pktmbuf_pool_init, NULL,
-                           rte_pktmbuf_init, NULL,
-                           socket_id, 0);
+
+  rmp = rte_pktmbuf_pool_create((char *) pool_name,         /* pool name */
+				 num_mbufs,                 /* number of mbufs */
+				 512,                       /* cache size */
+				 VLIB_BUFFER_HDR_SIZE,      /* priv size */
+				 VLIB_BUFFER_PRE_DATA_SIZE
+				 + VLIB_BUFFER_DATA_SIZE,   /* dataroom size */
+				 socket_id);                /* cpu socket */
+
   vec_free(pool_name);
 
   if (rmp)