sfc: Generalise packet hash lookup to support EF10 RX prefix

EF10 uses an entirely different RX prefix format from Falcon-arch.
Extend struct efx_nic_type to describe this.

[bwh: Also replace the magic numbers used for the Falcon-arch RX prefix]
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

drivers/net/ethernet/sfc/efx.c

@@ -573,7 +573,7 @@ static void efx_start_datapath(struct efx_nic *efx)
 	 * support the current MTU, including padding for header
 	 * alignment and overruns.
 	 */
-	efx->rx_dma_len = (efx->type->rx_buffer_hash_size +
+	efx->rx_dma_len = (efx->rx_prefix_size +
 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
 			   efx->type->rx_buffer_padding);
 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
@@ -2456,6 +2456,9 @@ static int efx_init_struct(struct efx_nic *efx,
 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
 
 	efx->net_dev = net_dev;
+	efx->rx_prefix_size = efx->type->rx_prefix_size;
+	efx->rx_packet_hash_offset =
+		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
 	spin_lock_init(&efx->stats_lock);
 	mutex_init(&efx->mac_lock);
 	efx->phy_op = &efx_dummy_phy_operations;

drivers/net/ethernet/sfc/falcon.c

@@ -2827,7 +2827,8 @@ const struct efx_nic_type falcon_b0_nic_type = {
 	.evq_ptr_tbl_base = FR_BZ_EVQ_PTR_TBL,
 	.evq_rptr_tbl_base = FR_BZ_EVQ_RPTR,
 	.max_dma_mask = DMA_BIT_MASK(FSF_AZ_TX_KER_BUF_ADDR_WIDTH),
-	.rx_buffer_hash_size = 0x10,
+	.rx_prefix_size = FS_BZ_RX_PREFIX_SIZE,
+	.rx_hash_offset = FS_BZ_RX_PREFIX_HASH_OFST,
 	.rx_buffer_padding = 0,
 	.can_rx_scatter = true,
 	.max_interrupt_mode = EFX_INT_MODE_MSIX,

drivers/net/ethernet/sfc/farch_regs.h

@@ -2925,4 +2925,8 @@
 #define FSF_AZ_DRV_GEN_EV_MAGIC_LBN 0
 #define FSF_AZ_DRV_GEN_EV_MAGIC_WIDTH 32
 
+/* RX packet prefix */
+#define FS_BZ_RX_PREFIX_HASH_OFST 12
+#define FS_BZ_RX_PREFIX_SIZE 16
+
 #endif /* EFX_FARCH_REGS_H */

drivers/net/ethernet/sfc/net_driver.h

@@ -663,6 +663,9 @@ struct vfdi_status;
  * @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_prefix_size: Size of RX prefix before packet data
+ * @rx_packet_hash_offset: Offset of RX flow hash from start of packet data
+ *	(valid only if @rx_prefix_size != 0; always negative)
  * @rx_hash_key: Toeplitz hash key for RSS
  * @rx_indir_table: Indirection table for RSS
  * @rx_scatter: Scatter mode enabled for receives
@@ -793,6 +796,8 @@ struct efx_nic {
 	unsigned int rx_page_buf_step;
 	unsigned int rx_bufs_per_page;
 	unsigned int rx_pages_per_batch;
+	unsigned int rx_prefix_size;
+	int rx_packet_hash_offset;
 	u8 rx_hash_key[40];
 	u32 rx_indir_table[128];
 	bool rx_scatter;
@@ -1009,7 +1014,8 @@ struct efx_mtd_partition {
  * @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 packet
+ * @rx_prefix_size: Size of RX prefix before packet data
+ * @rx_hash_offset: Offset of RX flow hash within prefix
  * @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
@@ -1126,7 +1132,8 @@ struct efx_nic_type {
 	unsigned int evq_ptr_tbl_base;
 	unsigned int evq_rptr_tbl_base;
 	u64 max_dma_mask;
-	unsigned int rx_buffer_hash_size;
+	unsigned int rx_prefix_size;
+	unsigned int rx_hash_offset;
 	unsigned int rx_buffer_padding;
 	bool can_rx_scatter;
 	unsigned int max_interrupt_mode;

drivers/net/ethernet/sfc/rx.c

@@ -61,13 +61,12 @@ static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
 	return page_address(buf->page) + buf->page_offset;
 }
 
-static inline u32 efx_rx_buf_hash(const u8 *eh)
+static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
 {
-	/* The ethernet header is always directly after any hash. */
-#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
-	return __le32_to_cpup((const __le32 *)(eh - 4));
+#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
+	return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
 #else
-	const u8 *data = eh - 4;
+	const u8 *data = eh + efx->rx_packet_hash_offset;
 	return (u32)data[0]	  |
 	       (u32)data[1] << 8  |
 	       (u32)data[2] << 16 |
@@ -439,7 +438,7 @@ efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 	}
 
 	if (efx->net_dev->features & NETIF_F_RXHASH)
-		skb->rxhash = efx_rx_buf_hash(eh);
+		skb->rxhash = efx_rx_buf_hash(efx, eh);
 	skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 			  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 
@@ -568,8 +567,8 @@ 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 -= efx->type->rx_buffer_hash_size;
+	rx_buf->page_offset += efx->rx_prefix_size;
+	rx_buf->len -= efx->rx_prefix_size;
 
 	if (n_frags > 1) {
 		/* Release/sync DMA mapping for additional fragments.

drivers/net/ethernet/sfc/siena.c

@@ -965,7 +965,8 @@ const struct efx_nic_type siena_a0_nic_type = {
 	.evq_ptr_tbl_base = FR_BZ_EVQ_PTR_TBL,
 	.evq_rptr_tbl_base = FR_BZ_EVQ_RPTR,
 	.max_dma_mask = DMA_BIT_MASK(FSF_AZ_TX_KER_BUF_ADDR_WIDTH),
-	.rx_buffer_hash_size = 0x10,
+	.rx_prefix_size = FS_BZ_RX_PREFIX_SIZE,
+	.rx_hash_offset = FS_BZ_RX_PREFIX_HASH_OFST,
 	.rx_buffer_padding = 0,
 	.can_rx_scatter = true,
 	.max_interrupt_mode = EFX_INT_MODE_MSIX,