sfc: Support only two rx buffers per page
authorSteve Hodgson <shodgson@solarflare.com>
Tue, 1 Jun 2010 11:33:17 +0000 (11:33 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 2 Jun 2010 09:21:08 +0000 (02:21 -0700)
- Pull the loop handling into efx_init_rx_buffers_(skb|page)
- Remove rx_queue->buf_page, and associated clean up code
- Remove unmap_addr, since unmap_addr is trivially calculable

This will allow us to recycle discarded buffers directly
from efx_rx_packet(), since will never be in the middle of
splitting a page.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/sfc/net_driver.h
drivers/net/sfc/rx.c

index 45398039dee675b2450e5c45679c61b53ce7fd8a..59c8ecc39aee713211757597b59a4314404b2113 100644 (file)
@@ -222,7 +222,6 @@ struct efx_tx_queue {
  *     If both this and skb are %NULL, the buffer slot is currently free.
  * @data: Pointer to ethernet header
  * @len: Buffer length, in bytes.
- * @unmap_addr: DMA address to unmap
  */
 struct efx_rx_buffer {
        dma_addr_t dma_addr;
@@ -230,7 +229,6 @@ struct efx_rx_buffer {
        struct page *page;
        char *data;
        unsigned int len;
-       dma_addr_t unmap_addr;
 };
 
 /**
@@ -257,11 +255,6 @@ struct efx_rx_buffer {
  * @alloc_page_count: RX allocation strategy counter.
  * @alloc_skb_count: RX allocation strategy counter.
  * @slow_fill: Timer used to defer efx_nic_generate_fill_event().
- * @buf_page: Page for next RX buffer.
- *     We can use a single page for multiple RX buffers. This tracks
- *     the remaining space in the allocation.
- * @buf_dma_addr: Page's DMA address.
- * @buf_data: Page's host address.
  * @flushed: Use when handling queue flushing
  */
 struct efx_rx_queue {
@@ -284,9 +277,6 @@ struct efx_rx_queue {
        struct timer_list slow_fill;
        unsigned int slow_fill_count;
 
-       struct page *buf_page;
-       dma_addr_t buf_dma_addr;
-       char *buf_data;
        enum efx_flush_state flushed;
 };
 
index bf1e55e7869e770251a7215b0611ea22b8904d87..615a1fcd66441f5e9b5396229cf238ebbbd37b88 100644 (file)
@@ -98,155 +98,132 @@ static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
        return PAGE_SIZE << efx->rx_buffer_order;
 }
 
-
 /**
- * efx_init_rx_buffer_skb - create new RX buffer using skb-based allocation
+ * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
  *
  * @rx_queue:          Efx RX queue
- * @rx_buf:            RX buffer structure to populate
  *
- * This allocates memory for a new receive buffer, maps it for DMA,
- * and populates a struct efx_rx_buffer with the relevant
- * information.  Return a negative error code or 0 on success.
+ * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
+ * struct efx_rx_buffer for each one. Return a negative error code or 0
+ * on success. May fail having only inserted fewer than EFX_RX_BATCH
+ * buffers.
  */
-static int efx_init_rx_buffer_skb(struct efx_rx_queue *rx_queue,
-                                 struct efx_rx_buffer *rx_buf)
+static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
 {
        struct efx_nic *efx = rx_queue->efx;
        struct net_device *net_dev = efx->net_dev;
+       struct efx_rx_buffer *rx_buf;
        int skb_len = efx->rx_buffer_len;
+       unsigned index, count;
 
-       rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
-       if (unlikely(!rx_buf->skb))
-               return -ENOMEM;
+       for (count = 0; count < EFX_RX_BATCH; ++count) {
+               index = rx_queue->added_count & EFX_RXQ_MASK;
+               rx_buf = efx_rx_buffer(rx_queue, index);
 
-       /* Adjust the SKB for padding and checksum */
-       skb_reserve(rx_buf->skb, NET_IP_ALIGN);
-       rx_buf->len = skb_len - NET_IP_ALIGN;
-       rx_buf->data = (char *)rx_buf->skb->data;
-       rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
+               rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
+               if (unlikely(!rx_buf->skb))
+                       return -ENOMEM;
+               rx_buf->page = NULL;
 
-       rx_buf->dma_addr = pci_map_single(efx->pci_dev,
-                                         rx_buf->data, rx_buf->len,
-                                         PCI_DMA_FROMDEVICE);
+               /* Adjust the SKB for padding and checksum */
+               skb_reserve(rx_buf->skb, NET_IP_ALIGN);
+               rx_buf->len = skb_len - NET_IP_ALIGN;
+               rx_buf->data = (char *)rx_buf->skb->data;
+               rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+               rx_buf->dma_addr = pci_map_single(efx->pci_dev,
+                                                 rx_buf->data, rx_buf->len,
+                                                 PCI_DMA_FROMDEVICE);
+               if (unlikely(pci_dma_mapping_error(efx->pci_dev,
+                                                  rx_buf->dma_addr))) {
+                       dev_kfree_skb_any(rx_buf->skb);
+                       rx_buf->skb = NULL;
+                       return -EIO;
+               }
 
-       if (unlikely(pci_dma_mapping_error(efx->pci_dev, rx_buf->dma_addr))) {
-               dev_kfree_skb_any(rx_buf->skb);
-               rx_buf->skb = NULL;
-               return -EIO;
+               ++rx_queue->added_count;
+               ++rx_queue->alloc_skb_count;
        }
 
        return 0;
 }
 
 /**
- * efx_init_rx_buffer_page - create new RX buffer using page-based allocation
+ * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
  *
  * @rx_queue:          Efx RX queue
- * @rx_buf:            RX buffer structure to populate
  *
- * This allocates memory for a new receive buffer, maps it for DMA,
- * and populates a struct efx_rx_buffer with the relevant
- * information.  Return a negative error code or 0 on success.
+ * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
+ * and populates struct efx_rx_buffers for each one. Return a negative error
+ * code or 0 on success. If a single page can be split between two buffers,
+ * then the page will either be inserted fully, or not at at all.
  */
-static int efx_init_rx_buffer_page(struct efx_rx_queue *rx_queue,
-                                  struct efx_rx_buffer *rx_buf)
+static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
 {
        struct efx_nic *efx = rx_queue->efx;
-       int bytes, space, offset;
-
-       bytes = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
-
-       /* If there is space left in the previously allocated page,
-        * then use it. Otherwise allocate a new one */
-       rx_buf->page = rx_queue->buf_page;
-       if (rx_buf->page == NULL) {
-               dma_addr_t dma_addr;
-
-               rx_buf->page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
-                                          efx->rx_buffer_order);
-               if (unlikely(rx_buf->page == NULL))
+       struct efx_rx_buffer *rx_buf;
+       struct page *page;
+       char *page_addr;
+       dma_addr_t dma_addr;
+       unsigned index, count;
+
+       /* We can split a page between two buffers */
+       BUILD_BUG_ON(EFX_RX_BATCH & 1);
+
+       for (count = 0; count < EFX_RX_BATCH; ++count) {
+               page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
+                                  efx->rx_buffer_order);
+               if (unlikely(page == NULL))
                        return -ENOMEM;
-
-               dma_addr = pci_map_page(efx->pci_dev, rx_buf->page,
-                                       0, efx_rx_buf_size(efx),
+               dma_addr = pci_map_page(efx->pci_dev, page, 0,
+                                       efx_rx_buf_size(efx),
                                        PCI_DMA_FROMDEVICE);
-
                if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
-                       __free_pages(rx_buf->page, efx->rx_buffer_order);
-                       rx_buf->page = NULL;
+                       __free_pages(page, efx->rx_buffer_order);
                        return -EIO;
                }
-
-               rx_queue->buf_page = rx_buf->page;
-               rx_queue->buf_dma_addr = dma_addr;
-               rx_queue->buf_data = (page_address(rx_buf->page) +
-                                     EFX_PAGE_IP_ALIGN);
-       }
-
-       rx_buf->len = bytes;
-       rx_buf->data = rx_queue->buf_data;
-       offset = efx_rx_buf_offset(rx_buf);
-       rx_buf->dma_addr = rx_queue->buf_dma_addr + offset;
-
-       /* Try to pack multiple buffers per page */
-       if (efx->rx_buffer_order == 0) {
-               /* The next buffer starts on the next 512 byte boundary */
-               rx_queue->buf_data += ((bytes + 0x1ff) & ~0x1ff);
-               offset += ((bytes + 0x1ff) & ~0x1ff);
-
-               space = efx_rx_buf_size(efx) - offset;
-               if (space >= bytes) {
-                       /* Refs dropped on kernel releasing each skb */
-                       get_page(rx_queue->buf_page);
-                       goto out;
+               EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1));
+               page_addr = page_address(page) + EFX_PAGE_IP_ALIGN;
+               dma_addr += EFX_PAGE_IP_ALIGN;
+
+       split:
+               index = rx_queue->added_count & EFX_RXQ_MASK;
+               rx_buf = efx_rx_buffer(rx_queue, index);
+               rx_buf->dma_addr = dma_addr;
+               rx_buf->skb = NULL;
+               rx_buf->page = page;
+               rx_buf->data = page_addr;
+               rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
+               ++rx_queue->added_count;
+               ++rx_queue->alloc_page_count;
+
+               if ((~count & 1) && (efx->rx_buffer_len < (PAGE_SIZE >> 1))) {
+                       /* Use the second half of the page */
+                       get_page(page);
+                       dma_addr += (PAGE_SIZE >> 1);
+                       page_addr += (PAGE_SIZE >> 1);
+                       ++count;
+                       goto split;
                }
        }
 
-       /* This is the final RX buffer for this page, so mark it for
-        * unmapping */
-       rx_queue->buf_page = NULL;
-       rx_buf->unmap_addr = rx_queue->buf_dma_addr;
-
- out:
        return 0;
 }
 
-/* This allocates memory for a new receive buffer, maps it for DMA,
- * and populates a struct efx_rx_buffer with the relevant
- * information.
- */
-static int efx_init_rx_buffer(struct efx_rx_queue *rx_queue,
-                             struct efx_rx_buffer *new_rx_buf)
-{
-       int rc = 0;
-
-       if (rx_queue->channel->rx_alloc_push_pages) {
-               new_rx_buf->skb = NULL;
-               rc = efx_init_rx_buffer_page(rx_queue, new_rx_buf);
-               rx_queue->alloc_page_count++;
-       } else {
-               new_rx_buf->page = NULL;
-               rc = efx_init_rx_buffer_skb(rx_queue, new_rx_buf);
-               rx_queue->alloc_skb_count++;
-       }
-
-       if (unlikely(rc < 0))
-               EFX_LOG_RL(rx_queue->efx, "%s RXQ[%d] =%d\n", __func__,
-                          rx_queue->queue, rc);
-       return rc;
-}
-
 static void efx_unmap_rx_buffer(struct efx_nic *efx,
                                struct efx_rx_buffer *rx_buf)
 {
        if (rx_buf->page) {
                EFX_BUG_ON_PARANOID(rx_buf->skb);
-               if (rx_buf->unmap_addr) {
-                       pci_unmap_page(efx->pci_dev, rx_buf->unmap_addr,
+
+               /* Unmap the buffer if there's only one buffer per page(s),
+                * or this is the second half of a two buffer page. */
+               if (efx->rx_buffer_order != 0 ||
+                   (efx_rx_buf_offset(rx_buf) & (PAGE_SIZE >> 1)) != 0) {
+                       pci_unmap_page(efx->pci_dev,
+                                      rx_buf->dma_addr & ~(PAGE_SIZE - 1),
                                       efx_rx_buf_size(efx),
                                       PCI_DMA_FROMDEVICE);
-                       rx_buf->unmap_addr = 0;
                }
        } else if (likely(rx_buf->skb)) {
                pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
@@ -286,9 +263,9 @@ static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
  */
 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 {
-       struct efx_rx_buffer *rx_buf;
-       unsigned fill_level, index;
-       int i, space, rc = 0;
+       struct efx_channel *channel = rx_queue->channel;
+       unsigned fill_level;
+       int space, rc = 0;
 
        /* Calculate current fill level, and exit if we don't need to fill */
        fill_level = (rx_queue->added_count - rx_queue->removed_count);
@@ -309,21 +286,18 @@ void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
        EFX_TRACE(rx_queue->efx, "RX queue %d fast-filling descriptor ring from"
                  " level %d to level %d using %s allocation\n",
                  rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
-                 rx_queue->channel->rx_alloc_push_pages ? "page" : "skb");
+                 channel->rx_alloc_push_pages ? "page" : "skb");
 
        do {
-               for (i = 0; i < EFX_RX_BATCH; ++i) {
-                       index = rx_queue->added_count & EFX_RXQ_MASK;
-                       rx_buf = efx_rx_buffer(rx_queue, index);
-                       rc = efx_init_rx_buffer(rx_queue, rx_buf);
-                       if (unlikely(rc)) {
-                               /* Ensure that we don't leave the rx queue
-                                * empty */
-                               if (rx_queue->added_count == rx_queue->removed_count)
-                                       efx_schedule_slow_fill(rx_queue);
-                               goto out;
-                       }
-                       ++rx_queue->added_count;
+               if (channel->rx_alloc_push_pages)
+                       rc = efx_init_rx_buffers_page(rx_queue);
+               else
+                       rc = efx_init_rx_buffers_skb(rx_queue);
+               if (unlikely(rc)) {
+                       /* Ensure that we don't leave the rx queue empty */
+                       if (rx_queue->added_count == rx_queue->removed_count)
+                               efx_schedule_slow_fill(rx_queue);
+                       goto out;
                }
        } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 
@@ -638,16 +612,6 @@ void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
                        efx_fini_rx_buffer(rx_queue, rx_buf);
                }
        }
-
-       /* For a page that is part-way through splitting into RX buffers */
-       if (rx_queue->buf_page != NULL) {
-               pci_unmap_page(rx_queue->efx->pci_dev, rx_queue->buf_dma_addr,
-                              efx_rx_buf_size(rx_queue->efx),
-                              PCI_DMA_FROMDEVICE);
-               __free_pages(rx_queue->buf_page,
-                            rx_queue->efx->rx_buffer_order);
-               rx_queue->buf_page = NULL;
-       }
 }
 
 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)