sfc: Enable RX DMA scattering where possible

Enable RX DMA scattering iff an RX buffer large enough for the current
MTU will not fit into a single page and the NIC supports DMA
scattering for kernel-mode RX queues.

On Falcon and Siena, the RX_USR_BUF_SIZE field is used as the DMA
limit for both all RX queues with scatter enabled.  Set it to 1824,
matching what Onload uses now.

Maintain a statistic for frames truncated due to lack of descriptors
(rx_nodesc_trunc).  This is distinct from rx_frm_trunc which may be
incremented when scattering is disabled and implies an over-length
frame.

Whenever an MTU change causes scattering to be turned on or off,
update filters that point to the PF queues, but leave others
unchanged, as VF drivers assume scattering is off.

Add n_frags parameters to various functions, and make them iterate:
- efx_rx_packet()
- efx_recycle_rx_buffers()
- efx_rx_mk_skb()
- efx_rx_deliver()

Make efx_handle_rx_event() responsible for updating
efx_rx_queue::removed_count.

Change the RX pipeline state to a starting ring index and number of
fragments, and make __efx_rx_packet() responsible for clearing it.

Based on earlier versions by David Riddoch and Jon Cooper.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>

drivers/net/ethernet/sfc/efx.c

@@ -88,8 +88,6 @@ const char *const efx_reset_type_names[] = {
 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
 };
 
-#define EFX_MAX_MTU (9 * 1024)
-
 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
  * queued onto this work queue. This is not a per-nic work queue, because
  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
@@ -627,9 +625,11 @@ static int efx_probe_channels(struct efx_nic *efx)
  */
 static void efx_start_datapath(struct efx_nic *efx)
 {
+	bool old_rx_scatter = efx->rx_scatter;
 	struct efx_tx_queue *tx_queue;
 	struct efx_rx_queue *rx_queue;
 	struct efx_channel *channel;
+	size_t rx_buf_len;
 
 	/* Calculate the rx buffer allocation parameters required to
 	 * support the current MTU, including padding for header
@@ -638,8 +638,32 @@ static void efx_start_datapath(struct efx_nic *efx)
 	efx->rx_dma_len = (efx->type->rx_buffer_hash_size +
 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
 			   efx->type->rx_buffer_padding);
-	efx->rx_buffer_order = get_order(sizeof(struct efx_rx_page_state) +
+	rx_buf_len = (sizeof(struct efx_rx_page_state) +
 		      EFX_PAGE_IP_ALIGN + efx->rx_dma_len);
+	if (rx_buf_len <= PAGE_SIZE) {
+		efx->rx_scatter = false;
+		efx->rx_buffer_order = 0;
+		if (rx_buf_len <= PAGE_SIZE / 2)
+			efx->rx_buffer_truesize = PAGE_SIZE / 2;
+		else
+			efx->rx_buffer_truesize = PAGE_SIZE;
+	} else if (efx->type->can_rx_scatter) {
+		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
+			     EFX_PAGE_IP_ALIGN + EFX_RX_USR_BUF_SIZE >
+			     PAGE_SIZE / 2);
+		efx->rx_scatter = true;
+		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
+		efx->rx_buffer_order = 0;
+		efx->rx_buffer_truesize = PAGE_SIZE / 2;
+	} else {
+		efx->rx_scatter = false;
+		efx->rx_buffer_order = get_order(rx_buf_len);
+		efx->rx_buffer_truesize = PAGE_SIZE << efx->rx_buffer_order;
+	}
+
+	/* RX filters also have scatter-enabled flags */
+	if (efx->rx_scatter != old_rx_scatter)
+		efx_filter_update_rx_scatter(efx);
 
 	/* We must keep at least one descriptor in a TX ring empty.
 	 * We could avoid this when the queue size does not exactly
@@ -661,7 +685,7 @@ static void efx_start_datapath(struct efx_nic *efx)
 			efx_nic_generate_fill_event(rx_queue);
 		}
 
-		WARN_ON(channel->rx_pkt != NULL);
+		WARN_ON(channel->rx_pkt_n_frags);
 	}
 
 	if (netif_device_present(efx->net_dev))

drivers/net/ethernet/sfc/efx.h

@@ -39,16 +39,14 @@ extern void efx_init_rx_queue(struct efx_rx_queue *rx_queue);
 extern void efx_fini_rx_queue(struct efx_rx_queue *rx_queue);
 extern void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue);
 extern void efx_rx_slow_fill(unsigned long context);
-extern void __efx_rx_packet(struct efx_channel *channel,
-			    struct efx_rx_buffer *rx_buf);
-extern void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
+extern void __efx_rx_packet(struct efx_channel *channel);
+extern void efx_rx_packet(struct efx_rx_queue *rx_queue,
+			  unsigned int index, unsigned int n_frags,
 			  unsigned int len, u16 flags);
 static inline void efx_rx_flush_packet(struct efx_channel *channel)
 {
-	if (channel->rx_pkt) {
-		__efx_rx_packet(channel, channel->rx_pkt);
-		channel->rx_pkt = NULL;
-	}
+	if (channel->rx_pkt_n_frags)
+		__efx_rx_packet(channel);
 }
 extern void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue);
 
@@ -73,6 +71,7 @@ extern void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue);
 extern int efx_probe_filters(struct efx_nic *efx);
 extern void efx_restore_filters(struct efx_nic *efx);
 extern void efx_remove_filters(struct efx_nic *efx);
+extern void efx_filter_update_rx_scatter(struct efx_nic *efx);
 extern s32 efx_filter_insert_filter(struct efx_nic *efx,
 				    struct efx_filter_spec *spec,
 				    bool replace);

drivers/net/ethernet/sfc/ethtool.c

@@ -154,6 +154,7 @@ static const struct efx_ethtool_stat efx_ethtool_stats[] = {
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_tcp_udp_chksum_err),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_mcast_mismatch),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_frm_trunc),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_nodesc_trunc),
 };
 
 /* Number of ethtool statistics */
@@ -978,7 +979,8 @@ static int efx_ethtool_set_class_rule(struct efx_nic *efx,
 	     rule->m_ext.data[1]))
 		return -EINVAL;
 
-	efx_filter_init_rx(&spec, EFX_FILTER_PRI_MANUAL, 0,
+	efx_filter_init_rx(&spec, EFX_FILTER_PRI_MANUAL,
+			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 			   (rule->ring_cookie == RX_CLS_FLOW_DISC) ?
 			   0xfff : rule->ring_cookie);
 

drivers/net/ethernet/sfc/falcon.c

@@ -1546,10 +1546,6 @@ static int falcon_probe_nic(struct efx_nic *efx)
 
 static void falcon_init_rx_cfg(struct efx_nic *efx)
 {
-	/* Prior to Siena the RX DMA engine will split each frame at
-	 * intervals of RX_USR_BUF_SIZE (32-byte units). We set it to
-	 * be so large that that never happens. */
-	const unsigned huge_buf_size = (3 * 4096) >> 5;
 	/* RX control FIFO thresholds (32 entries) */
 	const unsigned ctrl_xon_thr = 20;
 	const unsigned ctrl_xoff_thr = 25;
@@ -1557,10 +1553,15 @@ static void falcon_init_rx_cfg(struct efx_nic *efx)
 
 	efx_reado(efx, &reg, FR_AZ_RX_CFG);
 	if (efx_nic_rev(efx) <= EFX_REV_FALCON_A1) {
-		/* Data FIFO size is 5.5K */
+		/* Data FIFO size is 5.5K.  The RX DMA engine only
+		 * supports scattering for user-mode queues, but will
+		 * split DMA writes at intervals of RX_USR_BUF_SIZE
+		 * (32-byte units) even for kernel-mode queues.  We
+		 * set it to be so large that that never happens.
+		 */
 		EFX_SET_OWORD_FIELD(reg, FRF_AA_RX_DESC_PUSH_EN, 0);
 		EFX_SET_OWORD_FIELD(reg, FRF_AA_RX_USR_BUF_SIZE,
-				    huge_buf_size);
+				    (3 * 4096) >> 5);
 		EFX_SET_OWORD_FIELD(reg, FRF_AA_RX_XON_MAC_TH, 512 >> 8);
 		EFX_SET_OWORD_FIELD(reg, FRF_AA_RX_XOFF_MAC_TH, 2048 >> 8);
 		EFX_SET_OWORD_FIELD(reg, FRF_AA_RX_XON_TX_TH, ctrl_xon_thr);
@@ -1569,7 +1570,7 @@ static void falcon_init_rx_cfg(struct efx_nic *efx)
 		/* Data FIFO size is 80K; register fields moved */
 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_RX_DESC_PUSH_EN, 0);
 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_RX_USR_BUF_SIZE,
-				    huge_buf_size);
+				    EFX_RX_USR_BUF_SIZE >> 5);
 		/* Send XON and XOFF at ~3 * max MTU away from empty/full */
 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_RX_XON_MAC_TH, 27648 >> 8);
 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_RX_XOFF_MAC_TH, 54272 >> 8);
@@ -1815,6 +1816,7 @@ const struct efx_nic_type falcon_a1_nic_type = {
 	.evq_rptr_tbl_base = FR_AA_EVQ_RPTR_KER,
 	.max_dma_mask = DMA_BIT_MASK(FSF_AZ_TX_KER_BUF_ADDR_WIDTH),
 	.rx_buffer_padding = 0x24,
+	.can_rx_scatter = false,
 	.max_interrupt_mode = EFX_INT_MODE_MSI,
 	.phys_addr_channels = 4,
 	.timer_period_max =  1 << FRF_AB_TC_TIMER_VAL_WIDTH,
@@ -1865,6 +1867,7 @@ const struct efx_nic_type falcon_b0_nic_type = {
 	.max_dma_mask = DMA_BIT_MASK(FSF_AZ_TX_KER_BUF_ADDR_WIDTH),
 	.rx_buffer_hash_size = 0x10,
 	.rx_buffer_padding = 0,
+	.can_rx_scatter = true,
 	.max_interrupt_mode = EFX_INT_MODE_MSIX,
 	.phys_addr_channels = 32, /* Hardware limit is 64, but the legacy
 				   * interrupt handler only supports 32

drivers/net/ethernet/sfc/filter.c

@@ -172,6 +172,25 @@ static void efx_filter_push_rx_config(struct efx_nic *efx)
 			filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
 			!!(table->spec[EFX_FILTER_INDEX_MC_DEF].flags &
 			   EFX_FILTER_FLAG_RX_RSS));
+
+		/* There is a single bit to enable RX scatter for all
+		 * unmatched packets.  Only set it if scatter is
+		 * enabled in both filter specs.
+		 */
+		EFX_SET_OWORD_FIELD(
+			filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
+			!!(table->spec[EFX_FILTER_INDEX_UC_DEF].flags &
+			   table->spec[EFX_FILTER_INDEX_MC_DEF].flags &
+			   EFX_FILTER_FLAG_RX_SCATTER));
+	} else if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
+		/* We don't expose 'default' filters because unmatched
+		 * packets always go to the queue number found in the
+		 * RSS table.  But we still need to set the RX scatter
+		 * bit here.
+		 */
+		EFX_SET_OWORD_FIELD(
+			filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
+			efx->rx_scatter);
 	}
 
 	efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
@@ -413,13 +432,18 @@ static void efx_filter_reset_rx_def(struct efx_nic *efx, unsigned filter_idx)
 	struct efx_filter_state *state = efx->filter_state;
 	struct efx_filter_table *table = &state->table[EFX_FILTER_TABLE_RX_DEF];
 	struct efx_filter_spec *spec = &table->spec[filter_idx];
+	enum efx_filter_flags flags = 0;
 
 	/* If there's only one channel then disable RSS for non VF
 	 * traffic, thereby allowing VFs to use RSS when the PF can't.
 	 */
-	efx_filter_init_rx(spec, EFX_FILTER_PRI_MANUAL,
-			   efx->n_rx_channels > 1 ? EFX_FILTER_FLAG_RX_RSS : 0,
-			   0);
+	if (efx->n_rx_channels > 1)
+		flags |= EFX_FILTER_FLAG_RX_RSS;
+
+	if (efx->rx_scatter)
+		flags |= EFX_FILTER_FLAG_RX_SCATTER;
+
+	efx_filter_init_rx(spec, EFX_FILTER_PRI_MANUAL, flags, 0);
 	spec->type = EFX_FILTER_UC_DEF + filter_idx;
 	table->used_bitmap[0] |= 1 << filter_idx;
 }
@@ -1101,6 +1125,50 @@ void efx_remove_filters(struct efx_nic *efx)
 	kfree(state);
 }
 
+/* Update scatter enable flags for filters pointing to our own RX queues */
+void efx_filter_update_rx_scatter(struct efx_nic *efx)
+{
+	struct efx_filter_state *state = efx->filter_state;
+	enum efx_filter_table_id table_id;
+	struct efx_filter_table *table;
+	efx_oword_t filter;
+	unsigned int filter_idx;
+
+	spin_lock_bh(&state->lock);
+
+	for (table_id = EFX_FILTER_TABLE_RX_IP;
+	     table_id <= EFX_FILTER_TABLE_RX_DEF;
+	     table_id++) {
+		table = &state->table[table_id];
+
+		for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
+			if (!test_bit(filter_idx, table->used_bitmap) ||
+			    table->spec[filter_idx].dmaq_id >=
+			    efx->n_rx_channels)
+				continue;
+
+			if (efx->rx_scatter)
+				table->spec[filter_idx].flags |=
+					EFX_FILTER_FLAG_RX_SCATTER;
+			else
+				table->spec[filter_idx].flags &=
+					~EFX_FILTER_FLAG_RX_SCATTER;
+
+			if (table_id == EFX_FILTER_TABLE_RX_DEF)
+				/* Pushed by efx_filter_push_rx_config() */
+				continue;
+
+			efx_filter_build(&filter, &table->spec[filter_idx]);
+			efx_writeo(efx, &filter,
+				   table->offset + table->step * filter_idx);
+		}
+	}
+
+	efx_filter_push_rx_config(efx);
+
+	spin_unlock_bh(&state->lock);
+}
+
 #ifdef CONFIG_RFS_ACCEL
 
 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,

drivers/net/ethernet/sfc/net_driver.h

@@ -69,6 +69,12 @@
 #define EFX_TXQ_TYPES		4
 #define EFX_MAX_TX_QUEUES	(EFX_TXQ_TYPES * EFX_MAX_CHANNELS)
 
+/* Maximum possible MTU the driver supports */
+#define EFX_MAX_MTU (9 * 1024)
+
+/* Size of an RX scatter buffer.  Small enough to pack 2 into a 4K page. */
+#define EFX_RX_USR_BUF_SIZE 1824
+
 /* Forward declare Precision Time Protocol (PTP) support structure. */
 struct efx_ptp_data;
 
@@ -212,7 +218,8 @@ struct efx_tx_queue {
  *	If completed: offset in @page of Ethernet header.
  * @len: If pending: length for DMA descriptor.
  *	If completed: received length, excluding hash prefix.
- * @flags: Flags for buffer and packet state.
+ * @flags: Flags for buffer and packet state.  These are only set on the
+ *	first buffer of a scattered packet.
  */
 struct efx_rx_buffer {
 	dma_addr_t dma_addr;
@@ -256,6 +263,7 @@ struct efx_rx_page_state {
  * @added_count: Number of buffers added to the receive queue.
  * @notified_count: Number of buffers given to NIC (<= @added_count).
  * @removed_count: Number of buffers removed from the receive queue.
+ * @scatter_n: Number of buffers used by current packet
  * @max_fill: RX descriptor maximum fill level (<= ring size)
  * @fast_fill_trigger: RX descriptor fill level that will trigger a fast fill
  *	(<= @max_fill)
@@ -276,6 +284,7 @@ struct efx_rx_queue {
 	unsigned int added_count;
 	unsigned int notified_count;
 	unsigned int removed_count;
+	unsigned int scatter_n;
 	unsigned int max_fill;
 	unsigned int fast_fill_trigger;
 	unsigned int min_fill;
@@ -335,6 +344,12 @@ enum efx_rx_alloc_method {
  * @n_rx_frm_trunc: Count of RX_FRM_TRUNC errors
  * @n_rx_overlength: Count of RX_OVERLENGTH errors
  * @n_skbuff_leaks: Count of skbuffs leaked due to RX overrun
+ * @n_rx_nodesc_trunc: Number of RX packets truncated and then dropped due to
+ *	lack of descriptors
+ * @rx_pkt_n_frags: Number of fragments in next packet to be delivered by
+ *	__efx_rx_packet(), or zero if there is none
+ * @rx_pkt_index: Ring index of first buffer for next packet to be delivered
+ *	by __efx_rx_packet(), if @rx_pkt_n_frags != 0
  * @rx_queue: RX queue for this channel
  * @tx_queue: TX queues for this channel
  */
@@ -366,11 +381,10 @@ struct efx_channel {
 	unsigned n_rx_frm_trunc;
 	unsigned n_rx_overlength;
 	unsigned n_skbuff_leaks;
+	unsigned int n_rx_nodesc_trunc;
 
-	/* Used to pipeline received packets in order to optimise memory
-	 * access with prefetches.
-	 */
-	struct efx_rx_buffer *rx_pkt;
+	unsigned int rx_pkt_n_frags;
+	unsigned int rx_pkt_index;
 
 	struct efx_rx_queue rx_queue;
 	struct efx_tx_queue tx_queue[EFX_TXQ_TYPES];
@@ -672,8 +686,11 @@ struct vfdi_status;
  * @n_tx_channels: Number of channels used for TX
  * @rx_dma_len: Current maximum RX DMA length
  * @rx_buffer_order: Order (log2) of number of pages for each RX buffer
+ * @rx_buffer_truesize: Amortised allocation size of an RX buffer,
+ *	for use in sk_buff::truesize
  * @rx_hash_key: Toeplitz hash key for RSS
  * @rx_indir_table: Indirection table for RSS
+ * @rx_scatter: Scatter mode enabled for receives
  * @int_error_count: Number of internal errors seen recently
  * @int_error_expire: Time at which error count will be expired
  * @irq_status: Interrupt status buffer
@@ -788,8 +805,10 @@ struct efx_nic {
 	unsigned n_tx_channels;
 	unsigned int rx_dma_len;
 	unsigned int rx_buffer_order;
+	unsigned int rx_buffer_truesize;
 	u8 rx_hash_key[40];
 	u32 rx_indir_table[128];
+	bool rx_scatter;
 
 	unsigned int_error_count;
 	unsigned long int_error_expire;
@@ -920,8 +939,9 @@ static inline unsigned int efx_port_num(struct efx_nic *efx)
  * @evq_ptr_tbl_base: Event queue pointer table base address
  * @evq_rptr_tbl_base: Event queue read-pointer table base address
  * @max_dma_mask: Maximum possible DMA mask
- * @rx_buffer_hash_size: Size of hash at start of RX buffer
- * @rx_buffer_padding: Size of padding at end of RX buffer
+ * @rx_buffer_hash_size: Size of hash at start of RX packet
+ * @rx_buffer_padding: Size of padding at end of RX packet
+ * @can_rx_scatter: NIC is able to scatter packet to multiple buffers
  * @max_interrupt_mode: Highest capability interrupt mode supported
  *	from &enum efx_init_mode.
  * @phys_addr_channels: Number of channels with physically addressed
@@ -969,6 +989,7 @@ struct efx_nic_type {
 	u64 max_dma_mask;
 	unsigned int rx_buffer_hash_size;
 	unsigned int rx_buffer_padding;
+	bool can_rx_scatter;
 	unsigned int max_interrupt_mode;
 	unsigned int phys_addr_channels;
 	unsigned int timer_period_max;

drivers/net/ethernet/sfc/nic.c

@@ -591,12 +591,22 @@ void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
 	struct efx_nic *efx = rx_queue->efx;
 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
 	bool iscsi_digest_en = is_b0;
+	bool jumbo_en;
+
+	/* For kernel-mode queues in Falcon A1, the JUMBO flag enables
+	 * DMA to continue after a PCIe page boundary (and scattering
+	 * is not possible).  In Falcon B0 and Siena, it enables
+	 * scatter.
+	 */
+	jumbo_en = !is_b0 || efx->rx_scatter;
 
 	netif_dbg(efx, hw, efx->net_dev,
 		  "RX queue %d ring in special buffers %d-%d\n",
 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
 
+	rx_queue->scatter_n = 0;
+
 	/* Pin RX descriptor ring */
 	efx_init_special_buffer(efx, &rx_queue->rxd);
 
@@ -613,8 +623,7 @@ void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
 			      FRF_AZ_RX_DESCQ_SIZE,
 			      __ffs(rx_queue->rxd.entries),
 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
-			      /* For >=B0 this is scatter so disable */
-			      FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
+			      FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
 			      FRF_AZ_RX_DESCQ_EN, 1);
 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
 			 efx_rx_queue_index(rx_queue));
@@ -968,13 +977,24 @@ static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
 		EFX_RX_PKT_DISCARD : 0;
 }
 
-/* Handle receive events that are not in-order. */
-static void
+/* Handle receive events that are not in-order. Return true if this
+ * can be handled as a partial packet discard, false if it's more
+ * serious.
+ */
+static bool
 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
 {
+	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 	struct efx_nic *efx = rx_queue->efx;
 	unsigned expected, dropped;
 
+	if (rx_queue->scatter_n &&
+	    index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
+		      rx_queue->ptr_mask)) {
+		++channel->n_rx_nodesc_trunc;
+		return true;
+	}
+
 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
 	dropped = (index - expected) & rx_queue->ptr_mask;
 	netif_info(efx, rx_err, efx->net_dev,
@@ -983,6 +1003,7 @@ efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
 
 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
+	return false;
 }
 
 /* Handle a packet received event
@@ -998,7 +1019,7 @@ efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
 	unsigned expected_ptr;
-	bool rx_ev_pkt_ok;
+	bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
 	u16 flags;
 	struct efx_rx_queue *rx_queue;
 	struct efx_nic *efx = channel->efx;
@@ -1006,21 +1027,56 @@ efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
 		return;
 
-	/* Basic packet information */
-	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
-	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
-	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
-	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
-	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
+	rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
+	rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
 		channel->channel);
 
 	rx_queue = efx_channel_get_rx_queue(channel);
 
 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
-	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
-	if (unlikely(rx_ev_desc_ptr != expected_ptr))
-		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
+	expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
+			rx_queue->ptr_mask);
+
+	/* Check for partial drops and other errors */
+	if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
+	    unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
+		if (rx_ev_desc_ptr != expected_ptr &&
+		    !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
+			return;
+
+		/* Discard all pending fragments */
+		if (rx_queue->scatter_n) {
+			efx_rx_packet(
+				rx_queue,
+				rx_queue->removed_count & rx_queue->ptr_mask,
+				rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
+			rx_queue->removed_count += rx_queue->scatter_n;
+			rx_queue->scatter_n = 0;
+		}
+
+		/* Return if there is no new fragment */
+		if (rx_ev_desc_ptr != expected_ptr)
+			return;
+
+		/* Discard new fragment if not SOP */
+		if (!rx_ev_sop) {
+			efx_rx_packet(
+				rx_queue,
+				rx_queue->removed_count & rx_queue->ptr_mask,
+				1, 0, EFX_RX_PKT_DISCARD);
+			++rx_queue->removed_count;
+			return;
+		}
+	}
+
+	++rx_queue->scatter_n;
+	if (rx_ev_cont)
+		return;
+
+	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
+	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
+	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
 
 	if (likely(rx_ev_pkt_ok)) {
 		/* If packet is marked as OK and packet type is TCP/IP or
@@ -1048,7 +1104,11 @@ efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
 	channel->irq_mod_score += 2;
 
 	/* Handle received packet */
-	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
+	efx_rx_packet(rx_queue,
+		      rx_queue->removed_count & rx_queue->ptr_mask,
+		      rx_queue->scatter_n, rx_ev_byte_cnt, flags);
+	rx_queue->removed_count += rx_queue->scatter_n;
+	rx_queue->scatter_n = 0;
 }
 
 /* If this flush done event corresponds to a &struct efx_tx_queue, then

drivers/net/ethernet/sfc/rx.c

@@ -39,13 +39,17 @@
  */
 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)
 
 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
 {
@@ -66,6 +70,15 @@ static inline u32 efx_rx_buf_hash(const u8 *eh)
 #endif
 }
 
+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;
+}
+
 /**
  * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
  *
@@ -199,16 +212,19 @@ static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
 	++rx_queue->added_count;
 }
 
-/* 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)
+/* Recycle buffers directly back into the rx_queue. There is always
+ * room to add these buffer, because we've just popped them.
+ */
+static void efx_recycle_rx_buffers(struct efx_channel *channel,
+				   struct efx_rx_buffer *rx_buf,
+				   unsigned int n_frags)
 {
 	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;
 
+	do {
 		rx_buf->flags = 0;
 
 		if (efx->rx_dma_len <= EFX_RX_HALF_PAGE &&
@@ -221,6 +237,9 @@ static void efx_recycle_rx_buffer(struct efx_channel *channel,
 		memcpy(new_buf, rx_buf, sizeof(*new_buf));
 		rx_buf->page = NULL;
 		++rx_queue->added_count;
+
+		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
+	} while (--n_frags);
 }
 
 /**
@@ -328,46 +347,56 @@ 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 page *page = rx_buf->page;
 	struct sk_buff *skb;
 
-	rx_buf->page = NULL;
-
 	skb = napi_get_frags(napi);
-	if (!skb) {
-		put_page(page);
+	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;
 	}
 
 	if (efx->net_dev->features & NETIF_F_RXHASH)
 		skb->rxhash = efx_rx_buf_hash(eh);
-
-	skb_fill_page_desc(skb, 0, page, rx_buf->page_offset, rx_buf->len);
-
-	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);
 
+	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->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;
 }
 
-/* Allocate and construct an SKB around a struct page.*/
+/* 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;
@@ -381,25 +410,32 @@ static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 	EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 
 	skb_reserve(skb, EFX_PAGE_SKB_ALIGN);
+	memcpy(__skb_put(skb, hdr_len), eh, hdr_len);
 
-	skb->len = rx_buf->len;
-	skb->truesize = rx_buf->len + sizeof(struct sk_buff);
-	memcpy(skb->data, eh, hdr_len);
-	skb->tail += hdr_len;
-
-	/* Append the remaining page onto the frag list */
+	/* Append the remaining page(s) onto the frag list */
 	if (rx_buf->len > hdr_len) {
-		skb->data_len = skb->len - hdr_len;
-		skb_fill_page_desc(skb, 0, rx_buf->page,
-				   rx_buf->page_offset + hdr_len,
-				   skb->data_len);
+		rx_buf->page_offset += hdr_len;
+		rx_buf->len -= hdr_len;
+
+		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;
+
+			rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
+		}
 	} else {
 		__free_pages(rx_buf->page, efx->rx_buffer_order);
-		skb->data_len = 0;
+		rx_buf->page = NULL;
+		n_frags = 0;
 	}
 
-	/* Ownership has transferred from the rx_buf to skb */
-	rx_buf->page = NULL;
+	skb->truesize += n_frags * efx->rx_buffer_truesize;
 
 	/* Move past the ethernet header */
 	skb->protocol = eth_type_trans(skb, efx->net_dev);
@@ -408,7 +444,7 @@ static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 }
 
 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);
@@ -417,35 +453,43 @@ void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 	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 */
+	/* 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)) {
-		efx_recycle_rx_buffer(channel, rx_buf);
-
-		/* Don't hold off the previous receive */
-		rx_buf = NULL;
-		goto out;
+		efx_rx_flush_packet(channel);
+		efx_recycle_rx_buffers(channel, rx_buf, n_frags);
+		return;
 	}
 
+	if (n_frags == 1)
+		rx_buf->len = len;
+
 	/* Release and/or sync DMA mapping - assumes all RX buffers
 	 * consumed in-order per RX queue
 	 */
-	efx_unmap_rx_buffer(efx, rx_buf, len);
+	efx_unmap_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.
@@ -453,23 +497,40 @@ void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 	prefetch(efx_rx_buf_va(rx_buf));
 
 	rx_buf->page_offset += efx->type->rx_buffer_hash_size;
-	rx_buf->len = len - 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_unmap_rx_buffer(efx, rx_buf, EFX_RX_USR_BUF_SIZE);
+		}
+		rx_buf->len = len - (n_frags - 1) * EFX_RX_USR_BUF_SIZE;
+		efx_unmap_rx_buffer(efx, rx_buf, rx_buf->len);
+	}
 
 	/* Pipeline receives so that we give time for packet headers to be
 	 * prefetched into cache.
 	 */
-out:
 	efx_rx_flush_packet(channel);
-	channel->rx_pkt = rx_buf;
+	channel->rx_pkt_n_frags = n_frags;
+	channel->rx_pkt_index = index;
 }
 
 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
-			   struct efx_rx_buffer *rx_buf)
+			   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);
 
-	skb = efx_rx_mk_skb(channel, rx_buf, eh, hdr_len);
+	skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 	if (unlikely(skb == NULL)) {
 		efx_free_rx_buffer(channel->efx, rx_buf);
 		return;
@@ -488,9 +549,11 @@ static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 }
 
 /* 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;
+	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
@@ -499,16 +562,18 @@ void __efx_rx_packet(struct efx_channel *channel, struct efx_rx_buffer *rx_buf)
 	if (unlikely(efx->loopback_selftest)) {
 		efx_loopback_rx_packet(efx, eh, rx_buf->len);
 		efx_free_rx_buffer(efx, rx_buf);
-		return;
+		goto out;
 	}
 
 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 
 	if (!channel->type->receive_skb)
-		efx_rx_packet_gro(channel, rx_buf, eh);
+		efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 	else
-		efx_rx_deliver(channel, eh, rx_buf);
+		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)

drivers/net/ethernet/sfc/siena.c

@@ -414,6 +414,8 @@ static int siena_init_nic(struct efx_nic *efx)
 	EFX_SET_OWORD_FIELD(temp, FRF_BZ_RX_HASH_INSRT_HDR, 1);
 	EFX_SET_OWORD_FIELD(temp, FRF_BZ_RX_HASH_ALG, 1);
 	EFX_SET_OWORD_FIELD(temp, FRF_BZ_RX_IP_HASH, 1);
+	EFX_SET_OWORD_FIELD(temp, FRF_BZ_RX_USR_BUF_SIZE,
+			    EFX_RX_USR_BUF_SIZE >> 5);
 	efx_writeo(efx, &temp, FR_AZ_RX_CFG);
 
 	/* Set hash key for IPv4 */
@@ -718,6 +720,7 @@ const struct efx_nic_type siena_a0_nic_type = {
 	.max_dma_mask = DMA_BIT_MASK(FSF_AZ_TX_KER_BUF_ADDR_WIDTH),
 	.rx_buffer_hash_size = 0x10,
 	.rx_buffer_padding = 0,
+	.can_rx_scatter = true,
 	.max_interrupt_mode = EFX_INT_MODE_MSIX,
 	.phys_addr_channels = 32, /* Hardware limit is 64, but the legacy
 				   * interrupt handler only supports 32