1 files changed, 410 insertions, 383 deletions

diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index bb579a6128c8..e73e30bac10e 100644
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -16,6 +16,7 @@
 #include <linux/udp.h>
 #include <linux/prefetch.h>
 #include <linux/moduleparam.h>
+#include <linux/iommu.h>
 #include <net/ip.h>
 #include <net/checksum.h>
 #include "net_driver.h"
@@ -24,85 +25,39 @@
 #include "selftest.h"
 #include "workarounds.h"
 
-/* Number of RX descriptors pushed at once. */
-#define EFX_RX_BATCH  8
+/* Preferred number of descriptors to fill at once */
+#define EFX_RX_PREFERRED_BATCH 8U
 
-/* Maximum size of a buffer sharing a page */
-#define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
+/* Number of RX buffers to recycle pages for.  When creating the RX page recycle
+ * ring, this number is divided by the number of buffers per page to calculate
+ * the number of pages to store in the RX page recycle ring.
+ */
+#define EFX_RECYCLE_RING_SIZE_IOMMU 4096
+#define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
 
 /* Size of buffer allocated for skb header area. */
 #define EFX_SKB_HEADERS  64u
 
-/*
- * rx_alloc_method - RX buffer allocation method
- *
- * This driver supports two methods for allocating and using RX buffers:
- * each RX buffer may be backed by an skb or by an order-n page.
- *
- * When GRO is in use then the second method has a lower overhead,
- * since we don't have to allocate then free skbs on reassembled frames.
- *
- * Values:
- *   - RX_ALLOC_METHOD_AUTO = 0
- *   - RX_ALLOC_METHOD_SKB  = 1
- *   - RX_ALLOC_METHOD_PAGE = 2
- *
- * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
- * controlled by the parameters below.
- *
- *   - Since pushing and popping descriptors are separated by the rx_queue
- *     size, so the watermarks should be ~rxd_size.
- *   - The performance win by using page-based allocation for GRO is less
- *     than the performance hit of using page-based allocation of non-GRO,
- *     so the watermarks should reflect this.
- *
- * Per channel we maintain a single variable, updated by each channel:
- *
- *   rx_alloc_level += (gro_performed ? RX_ALLOC_FACTOR_GRO :
- *                      RX_ALLOC_FACTOR_SKB)
- * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
- * limits the hysteresis), and update the allocation strategy:
- *
- *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_GRO ?
- *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
- */
-static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
-
-#define RX_ALLOC_LEVEL_GRO 0x2000
-#define RX_ALLOC_LEVEL_MAX 0x3000
-#define RX_ALLOC_FACTOR_GRO 1
-#define RX_ALLOC_FACTOR_SKB (-2)
-
 /* This is the percentage fill level below which new RX descriptors
  * will be added to the RX descriptor ring.
  */
 static unsigned int rx_refill_threshold;
 
+/* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
+#define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
+				      EFX_RX_USR_BUF_SIZE)
+
 /*
  * RX maximum head room required.
  *
- * This must be at least 1 to prevent overflow and at least 2 to allow
- * pipelined receives.
+ * This must be at least 1 to prevent overflow, plus one packet-worth
+ * to allow pipelined receives.
  */
-#define EFX_RXD_HEAD_ROOM 2
+#define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
 
-/* Offset of ethernet header within page */
-static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
-					     struct efx_rx_buffer *buf)
+static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
 {
-	return buf->page_offset + efx->type->rx_buffer_hash_size;
-}
-static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
-{
-	return PAGE_SIZE << efx->rx_buffer_order;
-}
-
-static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
-{
-	if (buf->flags & EFX_RX_BUF_PAGE)
-		return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
-	else
-		return (u8 *)buf->u.skb->data + efx->type->rx_buffer_hash_size;
+	return page_address(buf->page) + buf->page_offset;
 }
 
 static inline u32 efx_rx_buf_hash(const u8 *eh)
@@ -119,66 +74,81 @@ static inline u32 efx_rx_buf_hash(const u8 *eh)
 #endif
 }
 
-/**
- * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
- *
- * @rx_queue:		Efx RX queue
- *
- * 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_buffers_skb(struct efx_rx_queue *rx_queue)
+static inline struct efx_rx_buffer *
+efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
+{
+	if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
+		return efx_rx_buffer(rx_queue, 0);
+	else
+		return rx_buf + 1;
+}
+
+static inline void efx_sync_rx_buffer(struct efx_nic *efx,
+				      struct efx_rx_buffer *rx_buf,
+				      unsigned int len)
+{
+	dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
+				DMA_FROM_DEVICE);
+}
+
+void efx_rx_config_page_split(struct efx_nic *efx)
+{
+	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + EFX_PAGE_IP_ALIGN,
+				      L1_CACHE_BYTES);
+	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
+		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
+		 efx->rx_page_buf_step);
+	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
+		efx->rx_bufs_per_page;
+	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
+					       efx->rx_bufs_per_page);
+}
+
+/* Check the RX page recycle ring for a page that can be reused. */
+static struct page *efx_reuse_page(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;
-	struct sk_buff *skb;
-	int skb_len = efx->rx_buffer_len;
-	unsigned index, count;
+	struct page *page;
+	struct efx_rx_page_state *state;
+	unsigned index;
 
-	for (count = 0; count < EFX_RX_BATCH; ++count) {
-		index = rx_queue->added_count & rx_queue->ptr_mask;
-		rx_buf = efx_rx_buffer(rx_queue, index);
-
-		rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
-		if (unlikely(!skb))
-			return -ENOMEM;
-
-		/* Adjust the SKB for padding */
-		skb_reserve(skb, NET_IP_ALIGN);
-		rx_buf->len = skb_len - NET_IP_ALIGN;
-		rx_buf->flags = 0;
-
-		rx_buf->dma_addr = dma_map_single(&efx->pci_dev->dev,
-						  skb->data, rx_buf->len,
-						  DMA_FROM_DEVICE);
-		if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
-					       rx_buf->dma_addr))) {
-			dev_kfree_skb_any(skb);
-			rx_buf->u.skb = NULL;
-			return -EIO;
-		}
+	index = rx_queue->page_remove & rx_queue->page_ptr_mask;
+	page = rx_queue->page_ring[index];
+	if (page == NULL)
+		return NULL;
+
+	rx_queue->page_ring[index] = NULL;
+	/* page_remove cannot exceed page_add. */
+	if (rx_queue->page_remove != rx_queue->page_add)
+		++rx_queue->page_remove;
 
-		++rx_queue->added_count;
-		++rx_queue->alloc_skb_count;
+	/* If page_count is 1 then we hold the only reference to this page. */
+	if (page_count(page) == 1) {
+		++rx_queue->page_recycle_count;
+		return page;
+	} else {
+		state = page_address(page);
+		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
+			       PAGE_SIZE << efx->rx_buffer_order,
+			       DMA_FROM_DEVICE);
+		put_page(page);
+		++rx_queue->page_recycle_failed;
 	}
 
-	return 0;
+	return NULL;
 }
 
 /**
- * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
+ * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
  *
  * @rx_queue:		Efx RX queue
  *
- * 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.
+ * This allocates a batch of pages, 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 used for multiple buffers,
+ * then the page will either be inserted fully, or not at all.
  */
-static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
+static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue)
 {
 	struct efx_nic *efx = rx_queue->efx;
 	struct efx_rx_buffer *rx_buf;
@@ -188,150 +158,140 @@ static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
 	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 = dma_map_page(&efx->pci_dev->dev, page, 0,
-					efx_rx_buf_size(efx),
-					DMA_FROM_DEVICE);
-		if (unlikely(dma_mapping_error(&efx->pci_dev->dev, dma_addr))) {
-			__free_pages(page, efx->rx_buffer_order);
-			return -EIO;
+	count = 0;
+	do {
+		page = efx_reuse_page(rx_queue);
+		if (page == NULL) {
+			page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
+					   efx->rx_buffer_order);
+			if (unlikely(page == NULL))
+				return -ENOMEM;
+			dma_addr =
+				dma_map_page(&efx->pci_dev->dev, page, 0,
+					     PAGE_SIZE << efx->rx_buffer_order,
+					     DMA_FROM_DEVICE);
+			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
+						       dma_addr))) {
+				__free_pages(page, efx->rx_buffer_order);
+				return -EIO;
+			}
+			state = page_address(page);
+			state->dma_addr = dma_addr;
+		} else {
+			state = page_address(page);
+			dma_addr = state->dma_addr;
 		}
-		state = page_address(page);
-		state->refcnt = 0;
-		state->dma_addr = dma_addr;
 
 		dma_addr += sizeof(struct efx_rx_page_state);
 		page_offset = sizeof(struct efx_rx_page_state);
 
-	split:
-		index = rx_queue->added_count & rx_queue->ptr_mask;
-		rx_buf = efx_rx_buffer(rx_queue, index);
-		rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
-		rx_buf->u.page = page;
-		rx_buf->page_offset = page_offset + EFX_PAGE_IP_ALIGN;
-		rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
-		rx_buf->flags = EFX_RX_BUF_PAGE;
-		++rx_queue->added_count;
-		++rx_queue->alloc_page_count;
-		++state->refcnt;
-
-		if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
-			/* Use the second half of the page */
+		do {
+			index = rx_queue->added_count & rx_queue->ptr_mask;
+			rx_buf = efx_rx_buffer(rx_queue, index);
+			rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
+			rx_buf->page = page;
+			rx_buf->page_offset = page_offset + EFX_PAGE_IP_ALIGN;
+			rx_buf->len = efx->rx_dma_len;
+			rx_buf->flags = 0;
+			++rx_queue->added_count;
 			get_page(page);
-			dma_addr += (PAGE_SIZE >> 1);
-			page_offset += (PAGE_SIZE >> 1);
-			++count;
-			goto split;
-		}
-	}
+			dma_addr += efx->rx_page_buf_step;
+			page_offset += efx->rx_page_buf_step;
+		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
+
+		rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
+	} while (++count < efx->rx_pages_per_batch);
 
 	return 0;
 }
 
+/* Unmap a DMA-mapped page.  This function is only called for the final RX
+ * buffer in a page.
+ */
 static void efx_unmap_rx_buffer(struct efx_nic *efx,
-				struct efx_rx_buffer *rx_buf,
-				unsigned int used_len)
+				struct efx_rx_buffer *rx_buf)
 {
-	if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
-		struct efx_rx_page_state *state;
-
-		state = page_address(rx_buf->u.page);
-		if (--state->refcnt == 0) {
-			dma_unmap_page(&efx->pci_dev->dev,
-				       state->dma_addr,
-				       efx_rx_buf_size(efx),
-				       DMA_FROM_DEVICE);
-		} else if (used_len) {
-			dma_sync_single_for_cpu(&efx->pci_dev->dev,
-						rx_buf->dma_addr, used_len,
-						DMA_FROM_DEVICE);
-		}
-	} else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
-		dma_unmap_single(&efx->pci_dev->dev, rx_buf->dma_addr,
-				 rx_buf->len, DMA_FROM_DEVICE);
+	struct page *page = rx_buf->page;
+
+	if (page) {
+		struct efx_rx_page_state *state = page_address(page);
+		dma_unmap_page(&efx->pci_dev->dev,
+			       state->dma_addr,
+			       PAGE_SIZE << efx->rx_buffer_order,
+			       DMA_FROM_DEVICE);
 	}
 }
 
-static void efx_free_rx_buffer(struct efx_nic *efx,
-			       struct efx_rx_buffer *rx_buf)
+static void efx_free_rx_buffer(struct efx_rx_buffer *rx_buf)
 {
-	if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
-		__free_pages(rx_buf->u.page, efx->rx_buffer_order);
-		rx_buf->u.page = NULL;
-	} else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
-		dev_kfree_skb_any(rx_buf->u.skb);
-		rx_buf->u.skb = NULL;
+	if (rx_buf->page) {
+		put_page(rx_buf->page);
+		rx_buf->page = NULL;
 	}
 }
 
-static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
-			       struct efx_rx_buffer *rx_buf)
+/* Attempt to recycle the page if there is an RX recycle ring; the page can
+ * only be added if this is the final RX buffer, to prevent pages being used in
+ * the descriptor ring and appearing in the recycle ring simultaneously.
+ */
+static void efx_recycle_rx_page(struct efx_channel *channel,
+				struct efx_rx_buffer *rx_buf)
 {
-	efx_unmap_rx_buffer(rx_queue->efx, rx_buf, 0);
-	efx_free_rx_buffer(rx_queue->efx, rx_buf);
-}
+	struct page *page = rx_buf->page;
+	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
+	struct efx_nic *efx = rx_queue->efx;
+	unsigned index;
 
-/* Attempt to resurrect the other receive buffer that used to share this page,
- * which had previously been passed up to the kernel and freed. */
-static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
-				    struct efx_rx_buffer *rx_buf)
-{
-	struct efx_rx_page_state *state = page_address(rx_buf->u.page);
-	struct efx_rx_buffer *new_buf;
-	unsigned fill_level, index;
-
-	/* +1 because efx_rx_packet() incremented removed_count. +1 because
-	 * we'd like to insert an additional descriptor whilst leaving
-	 * EFX_RXD_HEAD_ROOM for the non-recycle path */
-	fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
-	if (unlikely(fill_level > rx_queue->max_fill)) {
-		/* We could place "state" on a list, and drain the list in
-		 * efx_fast_push_rx_descriptors(). For now, this will do. */
+	/* Only recycle the page after processing the final buffer. */
+	if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
 		return;
-	}
 
-	++state->refcnt;
-	get_page(rx_buf->u.page);
+	index = rx_queue->page_add & rx_queue->page_ptr_mask;
+	if (rx_queue->page_ring[index] == NULL) {
+		unsigned read_index = rx_queue->page_remove &
+			rx_queue->page_ptr_mask;
 
-	index = rx_queue->added_count & rx_queue->ptr_mask;
-	new_buf = efx_rx_buffer(rx_queue, index);
-	new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
-	new_buf->u.page = rx_buf->u.page;
-	new_buf->len = rx_buf->len;
-	new_buf->flags = EFX_RX_BUF_PAGE;
-	++rx_queue->added_count;
+		/* The next slot in the recycle ring is available, but
+		 * increment page_remove if the read pointer currently
+		 * points here.
+		 */
+		if (read_index == index)
+			++rx_queue->page_remove;
+		rx_queue->page_ring[index] = page;
+		++rx_queue->page_add;
+		return;
+	}
+	++rx_queue->page_recycle_full;
+	efx_unmap_rx_buffer(efx, rx_buf);
+	put_page(rx_buf->page);
 }
 
-/* Recycle the given rx buffer directly back into the rx_queue. There is
- * always room to add this buffer, because we've just popped a buffer. */
-static void efx_recycle_rx_buffer(struct efx_channel *channel,
-				  struct efx_rx_buffer *rx_buf)
+static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
+			       struct efx_rx_buffer *rx_buf)
 {
-	struct efx_nic *efx = channel->efx;
-	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
-	struct efx_rx_buffer *new_buf;
-	unsigned index;
-
-	rx_buf->flags &= EFX_RX_BUF_PAGE;
-
-	if ((rx_buf->flags & EFX_RX_BUF_PAGE) &&
-	    efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
-	    page_count(rx_buf->u.page) == 1)
-		efx_resurrect_rx_buffer(rx_queue, rx_buf);
+	/* Release the page reference we hold for the buffer. */
+	if (rx_buf->page)
+		put_page(rx_buf->page);
+
+	/* If this is the last buffer in a page, unmap and free it. */
+	if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
+		efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
+		efx_free_rx_buffer(rx_buf);
+	}
+	rx_buf->page = NULL;
+}
 
-	index = rx_queue->added_count & rx_queue->ptr_mask;
-	new_buf = efx_rx_buffer(rx_queue, index);
+/* Recycle the pages that are used by buffers that have just been received. */
+static void efx_recycle_rx_buffers(struct efx_channel *channel,
+				   struct efx_rx_buffer *rx_buf,
+				   unsigned int n_frags)
+{
+	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 
-	memcpy(new_buf, rx_buf, sizeof(*new_buf));
-	rx_buf->u.page = NULL;
-	++rx_queue->added_count;
+	do {
+		efx_recycle_rx_page(channel, rx_buf);
+		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
+	} while (--n_frags);
 }
 
 /**
@@ -348,8 +308,8 @@ static void efx_recycle_rx_buffer(struct efx_channel *channel,
  */
 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 {
-	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
-	unsigned fill_level;
+	struct efx_nic *efx = rx_queue->efx;
+	unsigned int fill_level, batch_size;
 	int space, rc = 0;
 
 	/* Calculate current fill level, and exit if we don't need to fill */
@@ -364,28 +324,26 @@ void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 			rx_queue->min_fill = fill_level;
 	}
 
+	batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 	space = rx_queue->max_fill - fill_level;
-	EFX_BUG_ON_PARANOID(space < EFX_RX_BATCH);
+	EFX_BUG_ON_PARANOID(space < batch_size);
 
 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 		   "RX queue %d fast-filling descriptor ring from"
-		   " level %d to level %d using %s allocation\n",
+		   " level %d to level %d\n",
 		   efx_rx_queue_index(rx_queue), fill_level,
-		   rx_queue->max_fill,
-		   channel->rx_alloc_push_pages ? "page" : "skb");
+		   rx_queue->max_fill);
+
 
 	do {
-		if (channel->rx_alloc_push_pages)
-			rc = efx_init_rx_buffers_page(rx_queue);
-		else
-			rc = efx_init_rx_buffers_skb(rx_queue);
+		rc = efx_init_rx_buffers(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);
+	} while ((space -= batch_size) >= batch_size);
 
 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 		   "RX queue %d fast-filled descriptor ring "
@@ -408,7 +366,7 @@ void efx_rx_slow_fill(unsigned long context)
 
 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 				     struct efx_rx_buffer *rx_buf,
-				     int len, bool *leak_packet)
+				     int len)
 {
 	struct efx_nic *efx = rx_queue->efx;
 	unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
@@ -428,11 +386,6 @@ static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 				  "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 				  efx_rx_queue_index(rx_queue), len, max_len,
 				  efx->type->rx_buffer_padding);
-		/* If this buffer was skb-allocated, then the meta
-		 * data at the end of the skb will be trashed. So
-		 * we have no choice but to leak the fragment.
-		 */
-		*leak_packet = !(rx_buf->flags & EFX_RX_BUF_PAGE);
 		efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 	} else {
 		if (net_ratelimit())
@@ -448,212 +401,238 @@ static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 /* Pass a received packet up through GRO.  GRO can handle pages
  * regardless of checksum state and skbs with a good checksum.
  */
-static void efx_rx_packet_gro(struct efx_channel *channel,
-			      struct efx_rx_buffer *rx_buf,
-			      const u8 *eh)
+static void
+efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
+		  unsigned int n_frags, u8 *eh)
 {
 	struct napi_struct *napi = &channel->napi_str;
 	gro_result_t gro_result;
+	struct efx_nic *efx = channel->efx;
+	struct sk_buff *skb;
 
-	if (rx_buf->flags & EFX_RX_BUF_PAGE) {
-		struct efx_nic *efx = channel->efx;
-		struct page *page = rx_buf->u.page;
-		struct sk_buff *skb;
+	skb = napi_get_frags(napi);
+	if (unlikely(!skb)) {
+		while (n_frags--) {
+			put_page(rx_buf->page);
+			rx_buf->page = NULL;
+			rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
+		}
+		return;
+	}
 
-		rx_buf->u.page = NULL;
+	if (efx->net_dev->features & NETIF_F_RXHASH)
+		skb->rxhash = efx_rx_buf_hash(eh);
+	skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
+			  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
+
+	for (;;) {
+		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
+				   rx_buf->page, rx_buf->page_offset,
+				   rx_buf->len);
+		rx_buf->page = NULL;
+		skb->len += rx_buf->len;
+		if (skb_shinfo(skb)->nr_frags == n_frags)
+			break;
+
+		rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
+	}
 
-		skb = napi_get_frags(napi);
-		if (!skb) {
-			put_page(page);
-			return;
-		}
+	skb->data_len = skb->len;
+	skb->truesize += n_frags * efx->rx_buffer_truesize;
+
+	skb_record_rx_queue(skb, channel->rx_queue.core_index);
+
+	gro_result = napi_gro_frags(napi);
+	if (gro_result != GRO_DROP)
+		channel->irq_mod_score += 2;
+}
 
-		if (efx->net_dev->features & NETIF_F_RXHASH)
-			skb->rxhash = efx_rx_buf_hash(eh);
+/* Allocate and construct an SKB around page fragments */
+static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
+				     struct efx_rx_buffer *rx_buf,
+				     unsigned int n_frags,
+				     u8 *eh, int hdr_len)
+{
+	struct efx_nic *efx = channel->efx;
+	struct sk_buff *skb;
 
-		skb_fill_page_desc(skb, 0, page,
-				   efx_rx_buf_offset(efx, rx_buf), rx_buf->len);
+	/* Allocate an SKB to store the headers */
+	skb = netdev_alloc_skb(efx->net_dev, hdr_len + EFX_PAGE_SKB_ALIGN);
+	if (unlikely(skb == NULL))
+		return NULL;
 
-		skb->len = rx_buf->len;
-		skb->data_len = rx_buf->len;
-		skb->truesize += rx_buf->len;
-		skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
-				  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
+	EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 
-		skb_record_rx_queue(skb, channel->rx_queue.core_index);
+	skb_reserve(skb, EFX_PAGE_SKB_ALIGN);
+	memcpy(__skb_put(skb, hdr_len), eh, hdr_len);
 
-		gro_result = napi_gro_frags(napi);
-	} else {
-		struct sk_buff *skb = rx_buf->u.skb;
+	/* Append the remaining page(s) onto the frag list */
+	if (rx_buf->len > hdr_len) {
+		rx_buf->page_offset += hdr_len;
+		rx_buf->len -= hdr_len;
 
-		EFX_BUG_ON_PARANOID(!(rx_buf->flags & EFX_RX_PKT_CSUMMED));
-		rx_buf->u.skb = NULL;
-		skb->ip_summed = CHECKSUM_UNNECESSARY;
+		for (;;) {
+			skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
+					   rx_buf->page, rx_buf->page_offset,
+					   rx_buf->len);
+			rx_buf->page = NULL;
+			skb->len += rx_buf->len;
+			skb->data_len += rx_buf->len;
+			if (skb_shinfo(skb)->nr_frags == n_frags)
+				break;
 
-		gro_result = napi_gro_receive(napi, skb);
+			rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
+		}
+	} else {
+		__free_pages(rx_buf->page, efx->rx_buffer_order);
+		rx_buf->page = NULL;
+		n_frags = 0;
 	}
 
-	if (gro_result == GRO_NORMAL) {
-		channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
-	} else if (gro_result != GRO_DROP) {
-		channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
-		channel->irq_mod_score += 2;
-	}
+	skb->truesize += n_frags * efx->rx_buffer_truesize;
+
+	/* Move past the ethernet header */
+	skb->protocol = eth_type_trans(skb, efx->net_dev);
+
+	return skb;
 }
 
 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
-		   unsigned int len, u16 flags)
+		   unsigned int n_frags, unsigned int len, u16 flags)
 {
 	struct efx_nic *efx = rx_queue->efx;
 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 	struct efx_rx_buffer *rx_buf;
-	bool leak_packet = false;
 
 	rx_buf = efx_rx_buffer(rx_queue, index);
 	rx_buf->flags |= flags;
 
-	/* This allows the refill path to post another buffer.
-	 * EFX_RXD_HEAD_ROOM ensures that the slot we are using
-	 * isn't overwritten yet.
-	 */
-	rx_queue->removed_count++;
-
-	/* Validate the length encoded in the event vs the descriptor pushed */
-	efx_rx_packet__check_len(rx_queue, rx_buf, len, &leak_packet);
+	/* Validate the number of fragments and completed length */
+	if (n_frags == 1) {
+		efx_rx_packet__check_len(rx_queue, rx_buf, len);
+	} else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
+		   unlikely(len <= (n_frags - 1) * EFX_RX_USR_BUF_SIZE) ||
+		   unlikely(len > n_frags * EFX_RX_USR_BUF_SIZE) ||
+		   unlikely(!efx->rx_scatter)) {
+		/* If this isn't an explicit discard request, either
+		 * the hardware or the driver is broken.
+		 */
+		WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
+		rx_buf->flags |= EFX_RX_PKT_DISCARD;
+	}
 
 	netif_vdbg(efx, rx_status, efx->net_dev,
-		   "RX queue %d received id %x at %llx+%x %s%s\n",
+		   "RX queue %d received ids %x-%x len %d %s%s\n",
 		   efx_rx_queue_index(rx_queue), index,
-		   (unsigned long long)rx_buf->dma_addr, len,
+		   (index + n_frags - 1) & rx_queue->ptr_mask, len,
 		   (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 		   (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 
-	/* Discard packet, if instructed to do so */
+	/* Discard packet, if instructed to do so.  Process the
+	 * previous receive first.
+	 */
 	if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
-		if (unlikely(leak_packet))
-			channel->n_skbuff_leaks++;
-		else
-			efx_recycle_rx_buffer(channel, rx_buf);
-
-		/* Don't hold off the previous receive */
-		rx_buf = NULL;
-		goto out;
+		efx_rx_flush_packet(channel);
+		put_page(rx_buf->page);
+		efx_recycle_rx_buffers(channel, rx_buf, n_frags);
+		return;
 	}
 
-	/* Release and/or sync DMA mapping - assumes all RX buffers
-	 * consumed in-order per RX queue
+	if (n_frags == 1)
+		rx_buf->len = len;
+
+	/* Release and/or sync the DMA mapping - assumes all RX buffers
+	 * consumed in-order per RX queue.
 	 */
-	efx_unmap_rx_buffer(efx, rx_buf, len);
+	efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 
 	/* Prefetch nice and early so data will (hopefully) be in cache by
 	 * the time we look at it.
 	 */
-	prefetch(efx_rx_buf_eh(efx, rx_buf));
+	prefetch(efx_rx_buf_va(rx_buf));
+
+	rx_buf->page_offset += efx->type->rx_buffer_hash_size;
+	rx_buf->len -= efx->type->rx_buffer_hash_size;
+
+	if (n_frags > 1) {
+		/* Release/sync DMA mapping for additional fragments.
+		 * Fix length for last fragment.
+		 */
+		unsigned int tail_frags = n_frags - 1;
+
+		for (;;) {
+			rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
+			if (--tail_frags == 0)
+				break;
+			efx_sync_rx_buffer(efx, rx_buf, EFX_RX_USR_BUF_SIZE);
+		}
+		rx_buf->len = len - (n_frags - 1) * EFX_RX_USR_BUF_SIZE;
+		efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
+	}
+
+	/* All fragments have been DMA-synced, so recycle buffers and pages. */
+	rx_buf = efx_rx_buffer(rx_queue, index);
+	efx_recycle_rx_buffers(channel, rx_buf, n_frags);
 
 	/* Pipeline receives so that we give time for packet headers to be
 	 * prefetched into cache.
 	 */
-	rx_buf->len = len - efx->type->rx_buffer_hash_size;
-out:
-	if (channel->rx_pkt)
-		__efx_rx_packet(channel, channel->rx_pkt);
-	channel->rx_pkt = rx_buf;
+	efx_rx_flush_packet(channel);
+	channel->rx_pkt_n_frags = n_frags;
+	channel->rx_pkt_index = index;
 }
 
-static void efx_rx_deliver(struct efx_channel *channel,
-			   struct efx_rx_buffer *rx_buf)
+static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
+			   struct efx_rx_buffer *rx_buf,
+			   unsigned int n_frags)
 {
 	struct sk_buff *skb;
+	u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 
-	/* We now own the SKB */
-	skb = rx_buf->u.skb;
-	rx_buf->u.skb = NULL;
+	skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
+	if (unlikely(skb == NULL)) {
+		efx_free_rx_buffer(rx_buf);
+		return;
+	}
+	skb_record_rx_queue(skb, channel->rx_queue.core_index);
 
 	/* Set the SKB flags */
 	skb_checksum_none_assert(skb);
 
-	/* Record the rx_queue */
-	skb_record_rx_queue(skb, channel->rx_queue.core_index);
-
-	/* Pass the packet up */
 	if (channel->type->receive_skb)
-		channel->type->receive_skb(channel, skb);
-	else
-		netif_receive_skb(skb);
+		if (channel->type->receive_skb(channel, skb))
+			return;
 
-	/* Update allocation strategy method */
-	channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
+	/* Pass the packet up */
+	netif_receive_skb(skb);
 }
 
 /* Handle a received packet.  Second half: Touches packet payload. */
-void __efx_rx_packet(struct efx_channel *channel, struct efx_rx_buffer *rx_buf)
+void __efx_rx_packet(struct efx_channel *channel)
 {
 	struct efx_nic *efx = channel->efx;
-	u8 *eh = efx_rx_buf_eh(efx, rx_buf);
+	struct efx_rx_buffer *rx_buf =
+		efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
+	u8 *eh = efx_rx_buf_va(rx_buf);
 
 	/* If we're in loopback test, then pass the packet directly to the
 	 * loopback layer, and free the rx_buf here
 	 */
 	if (unlikely(efx->loopback_selftest)) {
 		efx_loopback_rx_packet(efx, eh, rx_buf->len);
-		efx_free_rx_buffer(efx, rx_buf);
-		return;
-	}
-
-	if (!(rx_buf->flags & EFX_RX_BUF_PAGE)) {
-		struct sk_buff *skb = rx_buf->u.skb;
-
-		prefetch(skb_shinfo(skb));
-
-		skb_reserve(skb, efx->type->rx_buffer_hash_size);
-		skb_put(skb, rx_buf->len);
-
-		if (efx->net_dev->features & NETIF_F_RXHASH)
-			skb->rxhash = efx_rx_buf_hash(eh);
-
-		/* Move past the ethernet header. rx_buf->data still points
-		 * at the ethernet header */
-		skb->protocol = eth_type_trans(skb, efx->net_dev);
-
-		skb_record_rx_queue(skb, channel->rx_queue.core_index);
+		efx_free_rx_buffer(rx_buf);
+		goto out;
 	}
 
 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 
-	if (likely(rx_buf->flags & (EFX_RX_BUF_PAGE | EFX_RX_PKT_CSUMMED)) &&
-	    !channel->type->receive_skb)
-		efx_rx_packet_gro(channel, rx_buf, eh);
+	if (!channel->type->receive_skb)
+		efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 	else
-		efx_rx_deliver(channel, rx_buf);
-}
-
-void efx_rx_strategy(struct efx_channel *channel)
-{
-	enum efx_rx_alloc_method method = rx_alloc_method;
-
-	if (channel->type->receive_skb) {
-		channel->rx_alloc_push_pages = false;
-		return;
-	}
-
-	/* Only makes sense to use page based allocation if GRO is enabled */
-	if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
-		method = RX_ALLOC_METHOD_SKB;
-	} else if (method == RX_ALLOC_METHOD_AUTO) {
-		/* Constrain the rx_alloc_level */
-		if (channel->rx_alloc_level < 0)
-			channel->rx_alloc_level = 0;
-		else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
-			channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
-
-		/* Decide on the allocation method */
-		method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
-			  RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
-	}
-
-	/* Push the option */
-	channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
+		efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
+out:
+	channel->rx_pkt_n_frags = 0;
 }
 
 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
@@ -683,9 +662,32 @@ int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 		kfree(rx_queue->buffer);
 		rx_queue->buffer = NULL;
 	}
+
 	return rc;
 }
 
+static void efx_init_rx_recycle_ring(struct efx_nic *efx,
+				     struct efx_rx_queue *rx_queue)
+{
+	unsigned int bufs_in_recycle_ring, page_ring_size;
+
+	/* Set the RX recycle ring size */
+#ifdef CONFIG_PPC64
+	bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
+#else
+	if (efx->pci_dev->dev.iommu_group)
+		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
+	else
+		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
+#endif /* CONFIG_PPC64 */
+
+	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
+					    efx->rx_bufs_per_page);
+	rx_queue->page_ring = kcalloc(page_ring_size,
+				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
+	rx_queue->page_ptr_mask = page_ring_size - 1;
+}
+
 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 {
 	struct efx_nic *efx = rx_queue->efx;
@@ -699,10 +701,18 @@ void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 	rx_queue->notified_count = 0;
 	rx_queue->removed_count = 0;
 	rx_queue->min_fill = -1U;
+	efx_init_rx_recycle_ring(efx, rx_queue);
+
+	rx_queue->page_remove = 0;
+	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
+	rx_queue->page_recycle_count = 0;
+	rx_queue->page_recycle_failed = 0;
+	rx_queue->page_recycle_full = 0;
 
 	/* Initialise limit fields */
 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
-	max_trigger = max_fill - EFX_RX_BATCH;
+	max_trigger =
+		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 	if (rx_refill_threshold != 0) {
 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 		if (trigger > max_trigger)
@@ -722,6 +732,7 @@ void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 {
 	int i;
+	struct efx_nic *efx = rx_queue->efx;
 	struct efx_rx_buffer *rx_buf;
 
 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
@@ -733,13 +744,32 @@ void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 	del_timer_sync(&rx_queue->slow_fill);
 	efx_nic_fini_rx(rx_queue);
 
-	/* Release RX buffers NB start at index 0 not current HW ptr */
+	/* Release RX buffers from the current read ptr to the write ptr */
 	if (rx_queue->buffer) {
-		for (i = 0; i <= rx_queue->ptr_mask; i++) {
-			rx_buf = efx_rx_buffer(rx_queue, i);
+		for (i = rx_queue->removed_count; i < rx_queue->added_count;
+		     i++) {
+			unsigned index = i & rx_queue->ptr_mask;
+			rx_buf = efx_rx_buffer(rx_queue, index);
 			efx_fini_rx_buffer(rx_queue, rx_buf);
 		}
 	}
+
+	/* Unmap and release the pages in the recycle ring. Remove the ring. */
+	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
+		struct page *page = rx_queue->page_ring[i];
+		struct efx_rx_page_state *state;
+
+		if (page == NULL)
+			continue;
+
+		state = page_address(page);
+		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
+			       PAGE_SIZE << efx->rx_buffer_order,
+			       DMA_FROM_DEVICE);
+		put_page(page);
+	}
+	kfree(rx_queue->page_ring);
+	rx_queue->page_ring = NULL;
 }
 
 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
@@ -754,9 +784,6 @@ void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 }
 
 
-module_param(rx_alloc_method, int, 0644);
-MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
-
 module_param(rx_refill_threshold, uint, 0444);
 MODULE_PARM_DESC(rx_refill_threshold,
 		 "RX descriptor ring refill threshold (%)");