bnx2x: uses build_skb() in receive path
authorEric Dumazet <eric.dumazet@gmail.com>
Mon, 14 Nov 2011 06:05:34 +0000 (06:05 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 14 Nov 2011 19:13:30 +0000 (14:13 -0500)
bnx2x uses following formula to compute its rx_buf_sz :

dev->mtu + 2*L1_CACHE_BYTES + 14 + 8 + 8 + 2

Then core network adds NET_SKB_PAD and SKB_DATA_ALIGN(sizeof(struct
skb_shared_info))

Final allocated size for skb head on x86_64 (L1_CACHE_BYTES = 64,
MTU=1500) : 2112 bytes : SLUB/SLAB round this to 4096 bytes.

Since skb truesize is then bigger than SK_MEM_QUANTUM, we have lot of
false sharing because of mem_reclaim in UDP stack.

One possible way to half truesize is to reduce the need by 64 bytes
(2112 -> 2048 bytes)

Instead of allocating a full cache line at the end of packet for
alignment, we can use the fact that skb_shared_info sits at the end of
skb->head, and we can use this room, if we convert bnx2x to new
build_skb() infrastructure.

skb_shared_info will be initialized after hardware finished its
transfert, so we can eventually overwrite the final padding.

Using build_skb() also reduces cache line misses in the driver, since we
use cache hot skb instead of cold ones. Number of in-flight sk_buff
structures is lower, they are recycled while still hot.

Performance results :

(820.000 pps on a rx UDP monothread benchmark, instead of 720.000 pps)

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Eilon Greenstein <eilong@broadcom.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
CC: Tom Herbert <therbert@google.com>
CC: Jamal Hadi Salim <hadi@mojatatu.com>
CC: Stephen Hemminger <shemminger@vyatta.com>
CC: Thomas Graf <tgraf@infradead.org>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Acked-by: Eilon Greenstein <eilong@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/broadcom/bnx2x/bnx2x.h
drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c
drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.h
drivers/net/ethernet/broadcom/bnx2x/bnx2x_ethtool.c
drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c

index 4b90b518d6a6d3eb854de5300aa68539f916cbe0..0f7b7a463eba0d4e6e548ff0b9fe1f2a81bd39cf 100644 (file)
@@ -293,8 +293,13 @@ enum {
 #define FCOE_TXQ_IDX(bp)       (MAX_ETH_TXQ_IDX(bp))
 
 /* fast path */
+/*
+ * This driver uses new build_skb() API :
+ * RX ring buffer contains pointer to kmalloc() data only,
+ * skb are built only after Hardware filled the frame.
+ */
 struct sw_rx_bd {
-       struct sk_buff  *skb;
+       u8              *data;
        DEFINE_DMA_UNMAP_ADDR(mapping);
 };
 
@@ -424,8 +429,8 @@ union host_hc_status_block {
 
 struct bnx2x_agg_info {
        /*
-        * First aggregation buffer is an skb, the following - are pages.
-        * We will preallocate the skbs for each aggregation when
+        * First aggregation buffer is a data buffer, the following - are pages.
+        * We will preallocate the data buffer for each aggregation when
         * we open the interface and will replace the BD at the consumer
         * with this one when we receive the TPA_START CQE in order to
         * keep the Rx BD ring consistent.
@@ -439,6 +444,7 @@ struct bnx2x_agg_info {
        u16                     parsing_flags;
        u16                     vlan_tag;
        u16                     len_on_bd;
+       u32                     rxhash;
 };
 
 #define Q_STATS_OFFSET32(stat_name) \
@@ -1187,10 +1193,20 @@ struct bnx2x {
 #define ETH_MAX_JUMBO_PACKET_SIZE      9600
 
        /* Max supported alignment is 256 (8 shift) */
-#define BNX2X_RX_ALIGN_SHIFT           ((L1_CACHE_SHIFT < 8) ? \
-                                        L1_CACHE_SHIFT : 8)
-       /* FW use 2 Cache lines Alignment for start packet and size  */
-#define BNX2X_FW_RX_ALIGN              (2 << BNX2X_RX_ALIGN_SHIFT)
+#define BNX2X_RX_ALIGN_SHIFT           min(8, L1_CACHE_SHIFT)
+
+       /* FW uses 2 Cache lines Alignment for start packet and size
+        *
+        * We assume skb_build() uses sizeof(struct skb_shared_info) bytes
+        * at the end of skb->data, to avoid wasting a full cache line.
+        * This reduces memory use (skb->truesize).
+        */
+#define BNX2X_FW_RX_ALIGN_START        (1UL << BNX2X_RX_ALIGN_SHIFT)
+
+#define BNX2X_FW_RX_ALIGN_END                                  \
+       max(1UL << BNX2X_RX_ALIGN_SHIFT,                        \
+           SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
+
 #define BNX2X_PXP_DRAM_ALIGN           (BNX2X_RX_ALIGN_SHIFT - 5)
 
        struct host_sp_status_block *def_status_blk;
index 13dad9230dbc8901c6a1b21cf1ebb55f7c4d7d4f..0d60b9e633ad16e7113f3b282767e99549c51ae1 100644 (file)
@@ -294,8 +294,21 @@ static void bnx2x_update_sge_prod(struct bnx2x_fastpath *fp,
           fp->last_max_sge, fp->rx_sge_prod);
 }
 
+/* Set Toeplitz hash value in the skb using the value from the
+ * CQE (calculated by HW).
+ */
+static u32 bnx2x_get_rxhash(const struct bnx2x *bp,
+                           const struct eth_fast_path_rx_cqe *cqe)
+{
+       /* Set Toeplitz hash from CQE */
+       if ((bp->dev->features & NETIF_F_RXHASH) &&
+           (cqe->status_flags & ETH_FAST_PATH_RX_CQE_RSS_HASH_FLG))
+               return le32_to_cpu(cqe->rss_hash_result);
+       return 0;
+}
+
 static void bnx2x_tpa_start(struct bnx2x_fastpath *fp, u16 queue,
-                           struct sk_buff *skb, u16 cons, u16 prod,
+                           u16 cons, u16 prod,
                            struct eth_fast_path_rx_cqe *cqe)
 {
        struct bnx2x *bp = fp->bp;
@@ -310,9 +323,9 @@ static void bnx2x_tpa_start(struct bnx2x_fastpath *fp, u16 queue,
        if (tpa_info->tpa_state != BNX2X_TPA_STOP)
                BNX2X_ERR("start of bin not in stop [%d]\n", queue);
 
-       /* Try to map an empty skb from the aggregation info  */
+       /* Try to map an empty data buffer from the aggregation info  */
        mapping = dma_map_single(&bp->pdev->dev,
-                                first_buf->skb->data,
+                                first_buf->data + NET_SKB_PAD,
                                 fp->rx_buf_size, DMA_FROM_DEVICE);
        /*
         *  ...if it fails - move the skb from the consumer to the producer
@@ -322,15 +335,15 @@ static void bnx2x_tpa_start(struct bnx2x_fastpath *fp, u16 queue,
 
        if (unlikely(dma_mapping_error(&bp->pdev->dev, mapping))) {
                /* Move the BD from the consumer to the producer */
-               bnx2x_reuse_rx_skb(fp, cons, prod);
+               bnx2x_reuse_rx_data(fp, cons, prod);
                tpa_info->tpa_state = BNX2X_TPA_ERROR;
                return;
        }
 
-       /* move empty skb from pool to prod */
-       prod_rx_buf->skb = first_buf->skb;
+       /* move empty data from pool to prod */
+       prod_rx_buf->data = first_buf->data;
        dma_unmap_addr_set(prod_rx_buf, mapping, mapping);
-       /* point prod_bd to new skb */
+       /* point prod_bd to new data */
        prod_bd->addr_hi = cpu_to_le32(U64_HI(mapping));
        prod_bd->addr_lo = cpu_to_le32(U64_LO(mapping));
 
@@ -344,6 +357,7 @@ static void bnx2x_tpa_start(struct bnx2x_fastpath *fp, u16 queue,
        tpa_info->tpa_state = BNX2X_TPA_START;
        tpa_info->len_on_bd = le16_to_cpu(cqe->len_on_bd);
        tpa_info->placement_offset = cqe->placement_offset;
+       tpa_info->rxhash = bnx2x_get_rxhash(bp, cqe);
 
 #ifdef BNX2X_STOP_ON_ERROR
        fp->tpa_queue_used |= (1 << queue);
@@ -471,11 +485,12 @@ static void bnx2x_tpa_stop(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 {
        struct bnx2x_agg_info *tpa_info = &fp->tpa_info[queue];
        struct sw_rx_bd *rx_buf = &tpa_info->first_buf;
-       u8 pad = tpa_info->placement_offset;
+       u32 pad = tpa_info->placement_offset;
        u16 len = tpa_info->len_on_bd;
-       struct sk_buff *skb = rx_buf->skb;
+       struct sk_buff *skb = NULL;
+       u8 *data = rx_buf->data;
        /* alloc new skb */
-       struct sk_buff *new_skb;
+       u8 *new_data;
        u8 old_tpa_state = tpa_info->tpa_state;
 
        tpa_info->tpa_state = BNX2X_TPA_STOP;
@@ -486,18 +501,18 @@ static void bnx2x_tpa_stop(struct bnx2x *bp, struct bnx2x_fastpath *fp,
        if (old_tpa_state == BNX2X_TPA_ERROR)
                goto drop;
 
-       /* Try to allocate the new skb */
-       new_skb = netdev_alloc_skb(bp->dev, fp->rx_buf_size);
+       /* Try to allocate the new data */
+       new_data = kmalloc(fp->rx_buf_size + NET_SKB_PAD, GFP_ATOMIC);
 
        /* Unmap skb in the pool anyway, as we are going to change
           pool entry status to BNX2X_TPA_STOP even if new skb allocation
           fails. */
        dma_unmap_single(&bp->pdev->dev, dma_unmap_addr(rx_buf, mapping),
                         fp->rx_buf_size, DMA_FROM_DEVICE);
+       if (likely(new_data))
+               skb = build_skb(data);
 
-       if (likely(new_skb)) {
-               prefetch(skb);
-               prefetch(((char *)(skb)) + L1_CACHE_BYTES);
+       if (likely(skb)) {
 
 #ifdef BNX2X_STOP_ON_ERROR
                if (pad + len > fp->rx_buf_size) {
@@ -509,8 +524,9 @@ static void bnx2x_tpa_stop(struct bnx2x *bp, struct bnx2x_fastpath *fp,
                }
 #endif
 
-               skb_reserve(skb, pad);
+               skb_reserve(skb, pad + NET_SKB_PAD);
                skb_put(skb, len);
+               skb->rxhash = tpa_info->rxhash;
 
                skb->protocol = eth_type_trans(skb, bp->dev);
                skb->ip_summed = CHECKSUM_UNNECESSARY;
@@ -526,8 +542,8 @@ static void bnx2x_tpa_stop(struct bnx2x *bp, struct bnx2x_fastpath *fp,
                }
 
 
-               /* put new skb in bin */
-               rx_buf->skb = new_skb;
+               /* put new data in bin */
+               rx_buf->data = new_data;
 
                return;
        }
@@ -539,19 +555,6 @@ drop:
        fp->eth_q_stats.rx_skb_alloc_failed++;
 }
 
-/* Set Toeplitz hash value in the skb using the value from the
- * CQE (calculated by HW).
- */
-static inline void bnx2x_set_skb_rxhash(struct bnx2x *bp, union eth_rx_cqe *cqe,
-                                       struct sk_buff *skb)
-{
-       /* Set Toeplitz hash from CQE */
-       if ((bp->dev->features & NETIF_F_RXHASH) &&
-           (cqe->fast_path_cqe.status_flags &
-            ETH_FAST_PATH_RX_CQE_RSS_HASH_FLG))
-               skb->rxhash =
-               le32_to_cpu(cqe->fast_path_cqe.rss_hash_result);
-}
 
 int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
 {
@@ -594,6 +597,7 @@ int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
                u8 cqe_fp_flags;
                enum eth_rx_cqe_type cqe_fp_type;
                u16 len, pad;
+               u8 *data;
 
 #ifdef BNX2X_STOP_ON_ERROR
                if (unlikely(bp->panic))
@@ -604,13 +608,6 @@ int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
                bd_prod = RX_BD(bd_prod);
                bd_cons = RX_BD(bd_cons);
 
-               /* Prefetch the page containing the BD descriptor
-                  at producer's index. It will be needed when new skb is
-                  allocated */
-               prefetch((void *)(PAGE_ALIGN((unsigned long)
-                                            (&fp->rx_desc_ring[bd_prod])) -
-                                 PAGE_SIZE + 1));
-
                cqe = &fp->rx_comp_ring[comp_ring_cons];
                cqe_fp = &cqe->fast_path_cqe;
                cqe_fp_flags = cqe_fp->type_error_flags;
@@ -626,125 +623,110 @@ int bnx2x_rx_int(struct bnx2x_fastpath *fp, int budget)
                if (unlikely(CQE_TYPE_SLOW(cqe_fp_type))) {
                        bnx2x_sp_event(fp, cqe);
                        goto next_cqe;
+               }
+               rx_buf = &fp->rx_buf_ring[bd_cons];
+               data = rx_buf->data;
 
-               /* this is an rx packet */
-               } else {
-                       rx_buf = &fp->rx_buf_ring[bd_cons];
-                       skb = rx_buf->skb;
-                       prefetch(skb);
-
-                       if (!CQE_TYPE_FAST(cqe_fp_type)) {
+               if (!CQE_TYPE_FAST(cqe_fp_type)) {
 #ifdef BNX2X_STOP_ON_ERROR
-                               /* sanity check */
-                               if (fp->disable_tpa &&
-                                   (CQE_TYPE_START(cqe_fp_type) ||
-                                    CQE_TYPE_STOP(cqe_fp_type)))
-                                       BNX2X_ERR("START/STOP packet while "
-                                                 "disable_tpa type %x\n",
-                                                 CQE_TYPE(cqe_fp_type));
+                       /* sanity check */
+                       if (fp->disable_tpa &&
+                           (CQE_TYPE_START(cqe_fp_type) ||
+                            CQE_TYPE_STOP(cqe_fp_type)))
+                               BNX2X_ERR("START/STOP packet while "
+                                         "disable_tpa type %x\n",
+                                         CQE_TYPE(cqe_fp_type));
 #endif
 
-                               if (CQE_TYPE_START(cqe_fp_type)) {
-                                       u16 queue = cqe_fp->queue_index;
-                                       DP(NETIF_MSG_RX_STATUS,
-                                          "calling tpa_start on queue %d\n",
-                                          queue);
-
-                                       bnx2x_tpa_start(fp, queue, skb,
-                                                       bd_cons, bd_prod,
-                                                       cqe_fp);
-
-                                       /* Set Toeplitz hash for LRO skb */
-                                       bnx2x_set_skb_rxhash(bp, cqe, skb);
-
-                                       goto next_rx;
-
-                               } else {
-                                       u16 queue =
-                                               cqe->end_agg_cqe.queue_index;
-                                       DP(NETIF_MSG_RX_STATUS,
-                                          "calling tpa_stop on queue %d\n",
-                                          queue);
+                       if (CQE_TYPE_START(cqe_fp_type)) {
+                               u16 queue = cqe_fp->queue_index;
+                               DP(NETIF_MSG_RX_STATUS,
+                                  "calling tpa_start on queue %d\n",
+                                  queue);
 
-                                       bnx2x_tpa_stop(bp, fp, queue,
-                                                      &cqe->end_agg_cqe,
-                                                      comp_ring_cons);
+                               bnx2x_tpa_start(fp, queue,
+                                               bd_cons, bd_prod,
+                                               cqe_fp);
+                               goto next_rx;
+                       } else {
+                               u16 queue =
+                                       cqe->end_agg_cqe.queue_index;
+                               DP(NETIF_MSG_RX_STATUS,
+                                  "calling tpa_stop on queue %d\n",
+                                  queue);
+
+                               bnx2x_tpa_stop(bp, fp, queue,
+                                              &cqe->end_agg_cqe,
+                                              comp_ring_cons);
 #ifdef BNX2X_STOP_ON_ERROR
-                                       if (bp->panic)
-                                               return 0;
+                               if (bp->panic)
+                                       return 0;
 #endif
 
-                                       bnx2x_update_sge_prod(fp, cqe_fp);
-                                       goto next_cqe;
-                               }
+                               bnx2x_update_sge_prod(fp, cqe_fp);
+                               goto next_cqe;
                        }
-                       /* non TPA */
-                       len = le16_to_cpu(cqe_fp->pkt_len);
-                       pad = cqe_fp->placement_offset;
-                       dma_sync_single_for_cpu(&bp->pdev->dev,
+               }
+               /* non TPA */
+               len = le16_to_cpu(cqe_fp->pkt_len);
+               pad = cqe_fp->placement_offset;
+               dma_sync_single_for_cpu(&bp->pdev->dev,
                                        dma_unmap_addr(rx_buf, mapping),
-                                                      pad + RX_COPY_THRESH,
-                                                      DMA_FROM_DEVICE);
-                       prefetch(((char *)(skb)) + L1_CACHE_BYTES);
+                                       pad + RX_COPY_THRESH,
+                                       DMA_FROM_DEVICE);
+               pad += NET_SKB_PAD;
+               prefetch(data + pad); /* speedup eth_type_trans() */
+               /* is this an error packet? */
+               if (unlikely(cqe_fp_flags & ETH_RX_ERROR_FALGS)) {
+                       DP(NETIF_MSG_RX_ERR,
+                          "ERROR  flags %x  rx packet %u\n",
+                          cqe_fp_flags, sw_comp_cons);
+                       fp->eth_q_stats.rx_err_discard_pkt++;
+                       goto reuse_rx;
+               }
 
-                       /* is this an error packet? */
-                       if (unlikely(cqe_fp_flags & ETH_RX_ERROR_FALGS)) {
+               /* Since we don't have a jumbo ring
+                * copy small packets if mtu > 1500
+                */
+               if ((bp->dev->mtu > ETH_MAX_PACKET_SIZE) &&
+                   (len <= RX_COPY_THRESH)) {
+                       skb = netdev_alloc_skb_ip_align(bp->dev, len);
+                       if (skb == NULL) {
                                DP(NETIF_MSG_RX_ERR,
-                                  "ERROR  flags %x  rx packet %u\n",
-                                  cqe_fp_flags, sw_comp_cons);
-                               fp->eth_q_stats.rx_err_discard_pkt++;
+                                  "ERROR  packet dropped because of alloc failure\n");
+                               fp->eth_q_stats.rx_skb_alloc_failed++;
                                goto reuse_rx;
                        }
-
-                       /* Since we don't have a jumbo ring
-                        * copy small packets if mtu > 1500
-                        */
-                       if ((bp->dev->mtu > ETH_MAX_PACKET_SIZE) &&
-                           (len <= RX_COPY_THRESH)) {
-                               struct sk_buff *new_skb;
-
-                               new_skb = netdev_alloc_skb(bp->dev, len + pad);
-                               if (new_skb == NULL) {
-                                       DP(NETIF_MSG_RX_ERR,
-                                          "ERROR  packet dropped "
-                                          "because of alloc failure\n");
-                                       fp->eth_q_stats.rx_skb_alloc_failed++;
-                                       goto reuse_rx;
-                               }
-
-                               /* aligned copy */
-                               skb_copy_from_linear_data_offset(skb, pad,
-                                                   new_skb->data + pad, len);
-                               skb_reserve(new_skb, pad);
-                               skb_put(new_skb, len);
-
-                               bnx2x_reuse_rx_skb(fp, bd_cons, bd_prod);
-
-                               skb = new_skb;
-
-                       } else
-                       if (likely(bnx2x_alloc_rx_skb(bp, fp, bd_prod) == 0)) {
+                       memcpy(skb->data, data + pad, len);
+                       bnx2x_reuse_rx_data(fp, bd_cons, bd_prod);
+               } else {
+                       if (likely(bnx2x_alloc_rx_data(bp, fp, bd_prod) == 0)) {
                                dma_unmap_single(&bp->pdev->dev,
-                                       dma_unmap_addr(rx_buf, mapping),
+                                                dma_unmap_addr(rx_buf, mapping),
                                                 fp->rx_buf_size,
                                                 DMA_FROM_DEVICE);
+                               skb = build_skb(data);
+                               if (unlikely(!skb)) {
+                                       kfree(data);
+                                       fp->eth_q_stats.rx_skb_alloc_failed++;
+                                       goto next_rx;
+                               }
                                skb_reserve(skb, pad);
-                               skb_put(skb, len);
-
                        } else {
                                DP(NETIF_MSG_RX_ERR,
                                   "ERROR  packet dropped because "
                                   "of alloc failure\n");
                                fp->eth_q_stats.rx_skb_alloc_failed++;
 reuse_rx:
-                               bnx2x_reuse_rx_skb(fp, bd_cons, bd_prod);
+                               bnx2x_reuse_rx_data(fp, bd_cons, bd_prod);
                                goto next_rx;
                        }
 
+                       skb_put(skb, len);
                        skb->protocol = eth_type_trans(skb, bp->dev);
 
                        /* Set Toeplitz hash for a none-LRO skb */
-                       bnx2x_set_skb_rxhash(bp, cqe, skb);
+                       skb->rxhash = bnx2x_get_rxhash(bp, cqe_fp);
 
                        skb_checksum_none_assert(skb);
 
@@ -767,7 +749,7 @@ reuse_rx:
 
 
 next_rx:
-               rx_buf->skb = NULL;
+               rx_buf->data = NULL;
 
                bd_cons = NEXT_RX_IDX(bd_cons);
                bd_prod = NEXT_RX_IDX(bd_prod);
@@ -1013,9 +995,9 @@ void bnx2x_init_rx_rings(struct bnx2x *bp)
                                struct sw_rx_bd *first_buf =
                                        &tpa_info->first_buf;
 
-                               first_buf->skb = netdev_alloc_skb(bp->dev,
-                                                      fp->rx_buf_size);
-                               if (!first_buf->skb) {
+                               first_buf->data = kmalloc(fp->rx_buf_size + NET_SKB_PAD,
+                                                         GFP_ATOMIC);
+                               if (!first_buf->data) {
                                        BNX2X_ERR("Failed to allocate TPA "
                                                  "skb pool for queue[%d] - "
                                                  "disabling TPA on this "
@@ -1118,16 +1100,16 @@ static void bnx2x_free_rx_bds(struct bnx2x_fastpath *fp)
 
        for (i = 0; i < NUM_RX_BD; i++) {
                struct sw_rx_bd *rx_buf = &fp->rx_buf_ring[i];
-               struct sk_buff *skb = rx_buf->skb;
+               u8 *data = rx_buf->data;
 
-               if (skb == NULL)
+               if (data == NULL)
                        continue;
                dma_unmap_single(&bp->pdev->dev,
                                 dma_unmap_addr(rx_buf, mapping),
                                 fp->rx_buf_size, DMA_FROM_DEVICE);
 
-               rx_buf->skb = NULL;
-               dev_kfree_skb(skb);
+               rx_buf->data = NULL;
+               kfree(data);
        }
 }
 
@@ -1509,6 +1491,7 @@ static inline void bnx2x_set_rx_buf_size(struct bnx2x *bp)
 
        for_each_queue(bp, i) {
                struct bnx2x_fastpath *fp = &bp->fp[i];
+               u32 mtu;
 
                /* Always use a mini-jumbo MTU for the FCoE L2 ring */
                if (IS_FCOE_IDX(i))
@@ -1518,13 +1501,15 @@ static inline void bnx2x_set_rx_buf_size(struct bnx2x *bp)
                         * IP_HEADER_ALIGNMENT_PADDING to prevent a buffer
                         * overrun attack.
                         */
-                       fp->rx_buf_size =
-                               BNX2X_FCOE_MINI_JUMBO_MTU + ETH_OVREHEAD +
-                               BNX2X_FW_RX_ALIGN + IP_HEADER_ALIGNMENT_PADDING;
+                       mtu = BNX2X_FCOE_MINI_JUMBO_MTU;
                else
-                       fp->rx_buf_size =
-                               bp->dev->mtu + ETH_OVREHEAD +
-                               BNX2X_FW_RX_ALIGN + IP_HEADER_ALIGNMENT_PADDING;
+                       mtu = bp->dev->mtu;
+               fp->rx_buf_size = BNX2X_FW_RX_ALIGN_START +
+                                 IP_HEADER_ALIGNMENT_PADDING +
+                                 ETH_OVREHEAD +
+                                 mtu +
+                                 BNX2X_FW_RX_ALIGN_END;
+               /* Note : rx_buf_size doesnt take into account NET_SKB_PAD */
        }
 }
 
index 260226d099168e12e23266614bc859d510c565d9..41eb17e7720f2ecbbe5525c6df02611a02d72504 100644 (file)
@@ -910,26 +910,27 @@ static inline int bnx2x_alloc_rx_sge(struct bnx2x *bp,
        return 0;
 }
 
-static inline int bnx2x_alloc_rx_skb(struct bnx2x *bp,
-                                    struct bnx2x_fastpath *fp, u16 index)
+static inline int bnx2x_alloc_rx_data(struct bnx2x *bp,
+                                     struct bnx2x_fastpath *fp, u16 index)
 {
-       struct sk_buff *skb;
+       u8 *data;
        struct sw_rx_bd *rx_buf = &fp->rx_buf_ring[index];
        struct eth_rx_bd *rx_bd = &fp->rx_desc_ring[index];
        dma_addr_t mapping;
 
-       skb = netdev_alloc_skb(bp->dev, fp->rx_buf_size);
-       if (unlikely(skb == NULL))
+       data = kmalloc(fp->rx_buf_size + NET_SKB_PAD, GFP_ATOMIC);
+       if (unlikely(data == NULL))
                return -ENOMEM;
 
-       mapping = dma_map_single(&bp->pdev->dev, skb->data, fp->rx_buf_size,
+       mapping = dma_map_single(&bp->pdev->dev, data + NET_SKB_PAD,
+                                fp->rx_buf_size,
                                 DMA_FROM_DEVICE);
        if (unlikely(dma_mapping_error(&bp->pdev->dev, mapping))) {
-               dev_kfree_skb_any(skb);
+               kfree(data);
                return -ENOMEM;
        }
 
-       rx_buf->skb = skb;
+       rx_buf->data = data;
        dma_unmap_addr_set(rx_buf, mapping, mapping);
 
        rx_bd->addr_hi = cpu_to_le32(U64_HI(mapping));
@@ -938,12 +939,12 @@ static inline int bnx2x_alloc_rx_skb(struct bnx2x *bp,
        return 0;
 }
 
-/* note that we are not allocating a new skb,
+/* note that we are not allocating a new buffer,
  * we are just moving one from cons to prod
  * we are not creating a new mapping,
  * so there is no need to check for dma_mapping_error().
  */
-static inline void bnx2x_reuse_rx_skb(struct bnx2x_fastpath *fp,
+static inline void bnx2x_reuse_rx_data(struct bnx2x_fastpath *fp,
                                      u16 cons, u16 prod)
 {
        struct sw_rx_bd *cons_rx_buf = &fp->rx_buf_ring[cons];
@@ -953,7 +954,7 @@ static inline void bnx2x_reuse_rx_skb(struct bnx2x_fastpath *fp,
 
        dma_unmap_addr_set(prod_rx_buf, mapping,
                           dma_unmap_addr(cons_rx_buf, mapping));
-       prod_rx_buf->skb = cons_rx_buf->skb;
+       prod_rx_buf->data = cons_rx_buf->data;
        *prod_bd = *cons_bd;
 }
 
@@ -1029,9 +1030,9 @@ static inline void bnx2x_free_tpa_pool(struct bnx2x *bp,
        for (i = 0; i < last; i++) {
                struct bnx2x_agg_info *tpa_info = &fp->tpa_info[i];
                struct sw_rx_bd *first_buf = &tpa_info->first_buf;
-               struct sk_buff *skb = first_buf->skb;
+               u8 *data = first_buf->data;
 
-               if (skb == NULL) {
+               if (data == NULL) {
                        DP(NETIF_MSG_IFDOWN, "tpa bin %d empty on free\n", i);
                        continue;
                }
@@ -1039,8 +1040,8 @@ static inline void bnx2x_free_tpa_pool(struct bnx2x *bp,
                        dma_unmap_single(&bp->pdev->dev,
                                         dma_unmap_addr(first_buf, mapping),
                                         fp->rx_buf_size, DMA_FROM_DEVICE);
-               dev_kfree_skb(skb);
-               first_buf->skb = NULL;
+               kfree(data);
+               first_buf->data = NULL;
        }
 }
 
@@ -1148,7 +1149,7 @@ static inline int bnx2x_alloc_rx_bds(struct bnx2x_fastpath *fp,
         * fp->eth_q_stats.rx_skb_alloc_failed = 0
         */
        for (i = 0; i < rx_ring_size; i++) {
-               if (bnx2x_alloc_rx_skb(bp, fp, ring_prod) < 0) {
+               if (bnx2x_alloc_rx_data(bp, fp, ring_prod) < 0) {
                        fp->eth_q_stats.rx_skb_alloc_failed++;
                        continue;
                }
index f6402fac26fd0d9a8b0ed99848f474853c45f937..ec318711f483b24b339f1fa7ea87bd098a9a1895 100644 (file)
@@ -1740,6 +1740,7 @@ static int bnx2x_run_loopback(struct bnx2x *bp, int loopback_mode)
        struct sw_rx_bd *rx_buf;
        u16 len;
        int rc = -ENODEV;
+       u8 *data;
 
        /* check the loopback mode */
        switch (loopback_mode) {
@@ -1865,10 +1866,9 @@ static int bnx2x_run_loopback(struct bnx2x *bp, int loopback_mode)
        dma_sync_single_for_cpu(&bp->pdev->dev,
                                   dma_unmap_addr(rx_buf, mapping),
                                   fp_rx->rx_buf_size, DMA_FROM_DEVICE);
-       skb = rx_buf->skb;
-       skb_reserve(skb, cqe->fast_path_cqe.placement_offset);
+       data = rx_buf->data + NET_SKB_PAD + cqe->fast_path_cqe.placement_offset;
        for (i = ETH_HLEN; i < pkt_size; i++)
-               if (*(skb->data + i) != (unsigned char) (i & 0xff))
+               if (*(data + i) != (unsigned char) (i & 0xff))
                        goto test_loopback_rx_exit;
 
        rc = 0;
index 33ff60d9fec8e72858baff7541c4a747f2fdac59..80fb9b53271167f08ce296ab8d2a706655d56ab4 100644 (file)
@@ -2789,8 +2789,8 @@ static void bnx2x_pf_rx_q_prep(struct bnx2x *bp,
        /* This should be a maximum number of data bytes that may be
         * placed on the BD (not including paddings).
         */
-       rxq_init->buf_sz = fp->rx_buf_size - BNX2X_FW_RX_ALIGN -
-               IP_HEADER_ALIGNMENT_PADDING;
+       rxq_init->buf_sz = fp->rx_buf_size - BNX2X_FW_RX_ALIGN_START -
+               BNX2X_FW_RX_ALIGN_END - IP_HEADER_ALIGNMENT_PADDING;
 
        rxq_init->cl_qzone_id = fp->cl_qzone_id;
        rxq_init->tpa_agg_sz = tpa_agg_size;