mlx4: allow order-0 memory allocations in RX path
authorEric Dumazet <eric.dumazet@gmail.com>
Sun, 23 Jun 2013 15:17:56 +0000 (08:17 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 25 Jun 2013 23:18:31 +0000 (16:18 -0700)
Signed-off-by: Eric Dumazet <edumazet@google.com>
mlx4 exclusively uses order-2 allocations in RX path, which are
likely to fail under memory pressure.

We therefore drop frames more than needed.

This patch tries order-3, order-2, order-1 and finally order-0
allocations to keep good performance, yet allow allocations if/when
memory gets fragmented.

By using larger pages, and avoiding unnecessary get_page()/put_page()
on compound pages, this patch improves performance as well, lowering
false sharing on struct page.

Also use GFP_KERNEL allocations in initialization path, as allocating 12
MB (390 order-3 pages) can easily fail with GFP_ATOMIC.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Amir Vadai <amirv@mellanox.com>
Acked-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx4/en_rx.c
drivers/net/ethernet/mellanox/mlx4/mlx4_en.h

index 9c57581b021c78e92ef2877292866b46999fe726..76997b93fdfebda261f26621b27b19e1f9e95feb 100644 (file)
 
 #include "mlx4_en.h"
 
+static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
+                           struct mlx4_en_rx_alloc *page_alloc,
+                           const struct mlx4_en_frag_info *frag_info,
+                           gfp_t _gfp)
+{
+       int order;
+       struct page *page;
+       dma_addr_t dma;
+
+       for (order = MLX4_EN_ALLOC_PREFER_ORDER; ;) {
+               gfp_t gfp = _gfp;
+
+               if (order)
+                       gfp |= __GFP_COMP | __GFP_NOWARN;
+               page = alloc_pages(gfp, order);
+               if (likely(page))
+                       break;
+               if (--order < 0 ||
+                   ((PAGE_SIZE << order) < frag_info->frag_size))
+                       return -ENOMEM;
+       }
+       dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
+                          PCI_DMA_FROMDEVICE);
+       if (dma_mapping_error(priv->ddev, dma)) {
+               put_page(page);
+               return -ENOMEM;
+       }
+       page_alloc->size = PAGE_SIZE << order;
+       page_alloc->page = page;
+       page_alloc->dma = dma;
+       page_alloc->offset = frag_info->frag_align;
+       /* Not doing get_page() for each frag is a big win
+        * on asymetric workloads.
+        */
+       atomic_set(&page->_count, page_alloc->size / frag_info->frag_stride);
+       return 0;
+}
+
 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
                               struct mlx4_en_rx_desc *rx_desc,
                               struct mlx4_en_rx_alloc *frags,
-                              struct mlx4_en_rx_alloc *ring_alloc)
+                              struct mlx4_en_rx_alloc *ring_alloc,
+                              gfp_t gfp)
 {
        struct mlx4_en_rx_alloc page_alloc[MLX4_EN_MAX_RX_FRAGS];
-       struct mlx4_en_frag_info *frag_info;
+       const struct mlx4_en_frag_info *frag_info;
        struct page *page;
        dma_addr_t dma;
        int i;
 
        for (i = 0; i < priv->num_frags; i++) {
                frag_info = &priv->frag_info[i];
-               if (ring_alloc[i].offset == frag_info->last_offset) {
-                       page = alloc_pages(GFP_ATOMIC | __GFP_COMP,
-                                       MLX4_EN_ALLOC_ORDER);
-                       if (!page)
-                               goto out;
-                       dma = dma_map_page(priv->ddev, page, 0,
-                               MLX4_EN_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
-                       if (dma_mapping_error(priv->ddev, dma)) {
-                               put_page(page);
-                               goto out;
-                       }
-                       page_alloc[i].page = page;
-                       page_alloc[i].dma = dma;
-                       page_alloc[i].offset = frag_info->frag_align;
-               } else {
-                       page_alloc[i].page = ring_alloc[i].page;
-                       get_page(ring_alloc[i].page);
-                       page_alloc[i].dma = ring_alloc[i].dma;
-                       page_alloc[i].offset = ring_alloc[i].offset +
-                                               frag_info->frag_stride;
-               }
+               page_alloc[i] = ring_alloc[i];
+               page_alloc[i].offset += frag_info->frag_stride;
+               if (page_alloc[i].offset + frag_info->frag_stride <= ring_alloc[i].size)
+                       continue;
+               if (mlx4_alloc_pages(priv, &page_alloc[i], frag_info, gfp))
+                       goto out;
        }
 
        for (i = 0; i < priv->num_frags; i++) {
@@ -88,14 +112,16 @@ static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
 
        return 0;
 
-
 out:
        while (i--) {
                frag_info = &priv->frag_info[i];
-               if (ring_alloc[i].offset == frag_info->last_offset)
+               if (page_alloc[i].page != ring_alloc[i].page) {
                        dma_unmap_page(priv->ddev, page_alloc[i].dma,
-                               MLX4_EN_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
-               put_page(page_alloc[i].page);
+                               page_alloc[i].size, PCI_DMA_FROMDEVICE);
+                       page = page_alloc[i].page;
+                       atomic_set(&page->_count, 1);
+                       put_page(page);
+               }
        }
        return -ENOMEM;
 }
@@ -104,12 +130,12 @@ static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
                              struct mlx4_en_rx_alloc *frags,
                              int i)
 {
-       struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
+       const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
 
-       if (frags[i].offset == frag_info->last_offset) {
-               dma_unmap_page(priv->ddev, frags[i].dma, MLX4_EN_ALLOC_SIZE,
+       if (frags[i].offset + frag_info->frag_stride > frags[i].size)
+               dma_unmap_page(priv->ddev, frags[i].dma, frags[i].size,
                                         PCI_DMA_FROMDEVICE);
-       }
+
        if (frags[i].page)
                put_page(frags[i].page);
 }
@@ -117,35 +143,28 @@ static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
                                  struct mlx4_en_rx_ring *ring)
 {
-       struct mlx4_en_rx_alloc *page_alloc;
        int i;
+       struct mlx4_en_rx_alloc *page_alloc;
 
        for (i = 0; i < priv->num_frags; i++) {
-               page_alloc = &ring->page_alloc[i];
-               page_alloc->page = alloc_pages(GFP_ATOMIC | __GFP_COMP,
-                                              MLX4_EN_ALLOC_ORDER);
-               if (!page_alloc->page)
-                       goto out;
+               const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
 
-               page_alloc->dma = dma_map_page(priv->ddev, page_alloc->page, 0,
-                                       MLX4_EN_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
-               if (dma_mapping_error(priv->ddev, page_alloc->dma)) {
-                       put_page(page_alloc->page);
-                       page_alloc->page = NULL;
+               if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
+                                    frag_info, GFP_KERNEL))
                        goto out;
-               }
-               page_alloc->offset = priv->frag_info[i].frag_align;
-               en_dbg(DRV, priv, "Initialized allocator:%d with page:%p\n",
-                      i, page_alloc->page);
        }
        return 0;
 
 out:
        while (i--) {
+               struct page *page;
+
                page_alloc = &ring->page_alloc[i];
                dma_unmap_page(priv->ddev, page_alloc->dma,
-                               MLX4_EN_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
-               put_page(page_alloc->page);
+                              page_alloc->size, PCI_DMA_FROMDEVICE);
+               page = page_alloc->page;
+               atomic_set(&page->_count, 1);
+               put_page(page);
                page_alloc->page = NULL;
        }
        return -ENOMEM;
@@ -158,13 +177,18 @@ static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
        int i;
 
        for (i = 0; i < priv->num_frags; i++) {
+               const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
+
                page_alloc = &ring->page_alloc[i];
                en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
                       i, page_count(page_alloc->page));
 
                dma_unmap_page(priv->ddev, page_alloc->dma,
-                               MLX4_EN_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
-               put_page(page_alloc->page);
+                               page_alloc->size, PCI_DMA_FROMDEVICE);
+               while (page_alloc->offset + frag_info->frag_stride < page_alloc->size) {
+                       put_page(page_alloc->page);
+                       page_alloc->offset += frag_info->frag_stride;
+               }
                page_alloc->page = NULL;
        }
 }
@@ -195,13 +219,14 @@ static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
 }
 
 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
-                                  struct mlx4_en_rx_ring *ring, int index)
+                                  struct mlx4_en_rx_ring *ring, int index,
+                                  gfp_t gfp)
 {
        struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
        struct mlx4_en_rx_alloc *frags = ring->rx_info +
                                        (index << priv->log_rx_info);
 
-       return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc);
+       return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
 }
 
 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
@@ -235,7 +260,8 @@ static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
                        ring = &priv->rx_ring[ring_ind];
 
                        if (mlx4_en_prepare_rx_desc(priv, ring,
-                                                   ring->actual_size)) {
+                                                   ring->actual_size,
+                                                   GFP_KERNEL)) {
                                if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
                                        en_err(priv, "Failed to allocate "
                                                     "enough rx buffers\n");
@@ -450,11 +476,11 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
                                        DMA_FROM_DEVICE);
 
                /* Save page reference in skb */
-               get_page(frags[nr].page);
                __skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
                skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
                skb_frags_rx[nr].page_offset = frags[nr].offset;
                skb->truesize += frag_info->frag_stride;
+               frags[nr].page = NULL;
        }
        /* Adjust size of last fragment to match actual length */
        if (nr > 0)
@@ -547,7 +573,7 @@ static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
        int index = ring->prod & ring->size_mask;
 
        while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
-               if (mlx4_en_prepare_rx_desc(priv, ring, index))
+               if (mlx4_en_prepare_rx_desc(priv, ring, index, GFP_ATOMIC))
                        break;
                ring->prod++;
                index = ring->prod & ring->size_mask;
@@ -805,21 +831,7 @@ int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
        return done;
 }
 
-
-/* Calculate the last offset position that accommodates a full fragment
- * (assuming fagment size = stride-align) */
-static int mlx4_en_last_alloc_offset(struct mlx4_en_priv *priv, u16 stride, u16 align)
-{
-       u16 res = MLX4_EN_ALLOC_SIZE % stride;
-       u16 offset = MLX4_EN_ALLOC_SIZE - stride - res + align;
-
-       en_dbg(DRV, priv, "Calculated last offset for stride:%d align:%d "
-                           "res:%d offset:%d\n", stride, align, res, offset);
-       return offset;
-}
-
-
-static int frag_sizes[] = {
+static const int frag_sizes[] = {
        FRAG_SZ0,
        FRAG_SZ1,
        FRAG_SZ2,
@@ -847,9 +859,6 @@ void mlx4_en_calc_rx_buf(struct net_device *dev)
                        priv->frag_info[i].frag_stride =
                                ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
                }
-               priv->frag_info[i].last_offset = mlx4_en_last_alloc_offset(
-                                               priv, priv->frag_info[i].frag_stride,
-                                               priv->frag_info[i].frag_align);
                buf_size += priv->frag_info[i].frag_size;
                i++;
        }
@@ -861,13 +870,13 @@ void mlx4_en_calc_rx_buf(struct net_device *dev)
        en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
                  "num_frags:%d):\n", eff_mtu, priv->num_frags);
        for (i = 0; i < priv->num_frags; i++) {
-               en_dbg(DRV, priv, "  frag:%d - size:%d prefix:%d align:%d "
-                               "stride:%d last_offset:%d\n", i,
-                               priv->frag_info[i].frag_size,
-                               priv->frag_info[i].frag_prefix_size,
-                               priv->frag_info[i].frag_align,
-                               priv->frag_info[i].frag_stride,
-                               priv->frag_info[i].last_offset);
+               en_err(priv,
+                      "  frag:%d - size:%d prefix:%d align:%d stride:%d\n",
+                      i,
+                      priv->frag_info[i].frag_size,
+                      priv->frag_info[i].frag_prefix_size,
+                      priv->frag_info[i].frag_align,
+                      priv->frag_info[i].frag_stride);
        }
 }
 
index 57192a8f1d5e02f410e16be9382d7d2623b4128d..35fb60e2320c2a896e8b241c2c9dbbe575a2dd02 100644 (file)
@@ -96,7 +96,8 @@
 
 /* Use the maximum between 16384 and a single page */
 #define MLX4_EN_ALLOC_SIZE     PAGE_ALIGN(16384)
-#define MLX4_EN_ALLOC_ORDER    get_order(MLX4_EN_ALLOC_SIZE)
+
+#define MLX4_EN_ALLOC_PREFER_ORDER     PAGE_ALLOC_COSTLY_ORDER
 
 /* Receive fragment sizes; we use at most 3 fragments (for 9600 byte MTU
  * and 4K allocations) */
@@ -234,9 +235,10 @@ struct mlx4_en_tx_desc {
 #define MLX4_EN_CX3_HIGH_ID    0x1005
 
 struct mlx4_en_rx_alloc {
-       struct page *page;
-       dma_addr_t dma;
-       u16 offset;
+       struct page     *page;
+       dma_addr_t      dma;
+       u32             offset;
+       u32             size;
 };
 
 struct mlx4_en_tx_ring {
@@ -439,8 +441,6 @@ struct mlx4_en_frag_info {
        u16 frag_prefix_size;
        u16 frag_stride;
        u16 frag_align;
-       u16 last_offset;
-
 };
 
 #ifdef CONFIG_MLX4_EN_DCB