Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6.git] / drivers / net / vmxnet3 / vmxnet3_drv.c
index 2c14736..fabcded 100644 (file)
@@ -573,7 +573,7 @@ vmxnet3_rq_alloc_rx_buf(struct vmxnet3_rx_queue *rq, u32 ring_idx,
        struct vmxnet3_cmd_ring *ring = &rq->rx_ring[ring_idx];
        u32 val;
 
-       while (num_allocated < num_to_alloc) {
+       while (num_allocated <= num_to_alloc) {
                struct vmxnet3_rx_buf_info *rbi;
                union Vmxnet3_GenericDesc *gd;
 
@@ -619,9 +619,15 @@ vmxnet3_rq_alloc_rx_buf(struct vmxnet3_rx_queue *rq, u32 ring_idx,
 
                BUG_ON(rbi->dma_addr == 0);
                gd->rxd.addr = cpu_to_le64(rbi->dma_addr);
-               gd->dword[2] = cpu_to_le32((ring->gen << VMXNET3_RXD_GEN_SHIFT)
+               gd->dword[2] = cpu_to_le32((!ring->gen << VMXNET3_RXD_GEN_SHIFT)
                                           | val | rbi->len);
 
+               /* Fill the last buffer but dont mark it ready, or else the
+                * device will think that the queue is full */
+               if (num_allocated == num_to_alloc)
+                       break;
+
+               gd->dword[2] |= cpu_to_le32(ring->gen << VMXNET3_RXD_GEN_SHIFT);
                num_allocated++;
                vmxnet3_cmd_ring_adv_next2fill(ring);
        }
@@ -1138,6 +1144,7 @@ vmxnet3_rq_rx_complete(struct vmxnet3_rx_queue *rq,
                VMXNET3_REG_RXPROD, VMXNET3_REG_RXPROD2
        };
        u32 num_rxd = 0;
+       bool skip_page_frags = false;
        struct Vmxnet3_RxCompDesc *rcd;
        struct vmxnet3_rx_ctx *ctx = &rq->rx_ctx;
 #ifdef __BIG_ENDIAN_BITFIELD
@@ -1148,11 +1155,12 @@ vmxnet3_rq_rx_complete(struct vmxnet3_rx_queue *rq,
                          &rxComp);
        while (rcd->gen == rq->comp_ring.gen) {
                struct vmxnet3_rx_buf_info *rbi;
-               struct sk_buff *skb;
+               struct sk_buff *skb, *new_skb = NULL;
+               struct page *new_page = NULL;
                int num_to_alloc;
                struct Vmxnet3_RxDesc *rxd;
                u32 idx, ring_idx;
-
+               struct vmxnet3_cmd_ring *ring = NULL;
                if (num_rxd >= quota) {
                        /* we may stop even before we see the EOP desc of
                         * the current pkt
@@ -1163,6 +1171,7 @@ vmxnet3_rq_rx_complete(struct vmxnet3_rx_queue *rq,
                BUG_ON(rcd->rqID != rq->qid && rcd->rqID != rq->qid2);
                idx = rcd->rxdIdx;
                ring_idx = rcd->rqID < adapter->num_rx_queues ? 0 : 1;
+               ring = rq->rx_ring + ring_idx;
                vmxnet3_getRxDesc(rxd, &rq->rx_ring[ring_idx].base[idx].rxd,
                                  &rxCmdDesc);
                rbi = rq->buf_info[ring_idx] + idx;
@@ -1191,37 +1200,80 @@ vmxnet3_rq_rx_complete(struct vmxnet3_rx_queue *rq,
                                goto rcd_done;
                        }
 
+                       skip_page_frags = false;
                        ctx->skb = rbi->skb;
-                       rbi->skb = NULL;
+                       new_skb = dev_alloc_skb(rbi->len + NET_IP_ALIGN);
+                       if (new_skb == NULL) {
+                               /* Skb allocation failed, do not handover this
+                                * skb to stack. Reuse it. Drop the existing pkt
+                                */
+                               rq->stats.rx_buf_alloc_failure++;
+                               ctx->skb = NULL;
+                               rq->stats.drop_total++;
+                               skip_page_frags = true;
+                               goto rcd_done;
+                       }
 
                        pci_unmap_single(adapter->pdev, rbi->dma_addr, rbi->len,
                                         PCI_DMA_FROMDEVICE);
 
                        skb_put(ctx->skb, rcd->len);
+
+                       /* Immediate refill */
+                       new_skb->dev = adapter->netdev;
+                       skb_reserve(new_skb, NET_IP_ALIGN);
+                       rbi->skb = new_skb;
+                       rbi->dma_addr = pci_map_single(adapter->pdev,
+                                       rbi->skb->data, rbi->len,
+                                       PCI_DMA_FROMDEVICE);
+                       rxd->addr = cpu_to_le64(rbi->dma_addr);
+                       rxd->len = rbi->len;
+
                } else {
-                       BUG_ON(ctx->skb == NULL);
+                       BUG_ON(ctx->skb == NULL && !skip_page_frags);
+
                        /* non SOP buffer must be type 1 in most cases */
-                       if (rbi->buf_type == VMXNET3_RX_BUF_PAGE) {
-                               BUG_ON(rxd->btype != VMXNET3_RXD_BTYPE_BODY);
+                       BUG_ON(rbi->buf_type != VMXNET3_RX_BUF_PAGE);
+                       BUG_ON(rxd->btype != VMXNET3_RXD_BTYPE_BODY);
 
-                               if (rcd->len) {
-                                       pci_unmap_page(adapter->pdev,
-                                                      rbi->dma_addr, rbi->len,
-                                                      PCI_DMA_FROMDEVICE);
+                       /* If an sop buffer was dropped, skip all
+                        * following non-sop fragments. They will be reused.
+                        */
+                       if (skip_page_frags)
+                               goto rcd_done;
 
-                                       vmxnet3_append_frag(ctx->skb, rcd, rbi);
-                                       rbi->page = NULL;
-                               }
-                       } else {
-                               /*
-                                * The only time a non-SOP buffer is type 0 is
-                                * when it's EOP and error flag is raised, which
-                                * has already been handled.
+                       new_page = alloc_page(GFP_ATOMIC);
+                       if (unlikely(new_page == NULL)) {
+                               /* Replacement page frag could not be allocated.
+                                * Reuse this page. Drop the pkt and free the
+                                * skb which contained this page as a frag. Skip
+                                * processing all the following non-sop frags.
                                 */
-                               BUG_ON(true);
+                               rq->stats.rx_buf_alloc_failure++;
+                               dev_kfree_skb(ctx->skb);
+                               ctx->skb = NULL;
+                               skip_page_frags = true;
+                               goto rcd_done;
+                       }
+
+                       if (rcd->len) {
+                               pci_unmap_page(adapter->pdev,
+                                              rbi->dma_addr, rbi->len,
+                                              PCI_DMA_FROMDEVICE);
+
+                               vmxnet3_append_frag(ctx->skb, rcd, rbi);
                        }
+
+                       /* Immediate refill */
+                       rbi->page = new_page;
+                       rbi->dma_addr = pci_map_page(adapter->pdev, rbi->page,
+                                                    0, PAGE_SIZE,
+                                                    PCI_DMA_FROMDEVICE);
+                       rxd->addr = cpu_to_le64(rbi->dma_addr);
+                       rxd->len = rbi->len;
                }
 
+
                skb = ctx->skb;
                if (rcd->eop) {
                        skb->len += skb->data_len;
@@ -1243,26 +1295,27 @@ vmxnet3_rq_rx_complete(struct vmxnet3_rx_queue *rq,
                }
 
 rcd_done:
-               /* device may skip some rx descs */
-               rq->rx_ring[ring_idx].next2comp = idx;
-               VMXNET3_INC_RING_IDX_ONLY(rq->rx_ring[ring_idx].next2comp,
-                                         rq->rx_ring[ring_idx].size);
-
-               /* refill rx buffers frequently to avoid starving the h/w */
-               num_to_alloc = vmxnet3_cmd_ring_desc_avail(rq->rx_ring +
-                                                          ring_idx);
-               if (unlikely(num_to_alloc > VMXNET3_RX_ALLOC_THRESHOLD(rq,
-                                                       ring_idx, adapter))) {
-                       vmxnet3_rq_alloc_rx_buf(rq, ring_idx, num_to_alloc,
-                                               adapter);
-
-                       /* if needed, update the register */
-                       if (unlikely(rq->shared->updateRxProd)) {
-                               VMXNET3_WRITE_BAR0_REG(adapter,
-                                       rxprod_reg[ring_idx] + rq->qid * 8,
-                                       rq->rx_ring[ring_idx].next2fill);
-                               rq->uncommitted[ring_idx] = 0;
-                       }
+               /* device may have skipped some rx descs */
+               ring->next2comp = idx;
+               num_to_alloc = vmxnet3_cmd_ring_desc_avail(ring);
+               ring = rq->rx_ring + ring_idx;
+               while (num_to_alloc) {
+                       vmxnet3_getRxDesc(rxd, &ring->base[ring->next2fill].rxd,
+                                         &rxCmdDesc);
+                       BUG_ON(!rxd->addr);
+
+                       /* Recv desc is ready to be used by the device */
+                       rxd->gen = ring->gen;
+                       vmxnet3_cmd_ring_adv_next2fill(ring);
+                       num_to_alloc--;
+               }
+
+               /* if needed, update the register */
+               if (unlikely(rq->shared->updateRxProd)) {
+                       VMXNET3_WRITE_BAR0_REG(adapter,
+                               rxprod_reg[ring_idx] + rq->qid * 8,
+                               ring->next2fill);
+                       rq->uncommitted[ring_idx] = 0;
                }
 
                vmxnet3_comp_ring_adv_next2proc(&rq->comp_ring);