Pull request: sfc-next 2013-09-21

Pull request: sfc-next 2013-09-21

[Ben Hutchings]

The following changes since commit e7d33bb5ea82922e6ddcfc6b28a630b1a4ced071:

  lockref: add ability to mark lockrefs "dead" (2013-09-07 15:49:18 -0700)

are available in the git repository at:

  git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next.git for-davem

for you to fetch changes up to 48419e3f080704470ffa1350311c6d5640eb4da6:

  sfc: Add static tracepoints to datapath (2013-09-20 01:18:51 +0100)

1. Some cleanup from Fengguang Wu and his kbuild robot.
2. Support for ethtool register dump on EF10.
3. Change soft-TSO to take advantage of firmware assistance on EF10.
4. Support for PIO TX buffers and descriptors on EF10, enabled on
architectures that support write-combining mappings.
5. Accelerated RFS support for TCP/IPv6 and UDP/IPv6 on EF10.
6. Optional static tracepoints to support testing of various offload
features.

Ben.

----------------------------------------------------------------
Ben Hutchings (8):
      sfc: Add EF10 registers to register dump
      sfc: Fold tso_get_head_fragment() into tso_start()
      sfc: Implement firmware-assisted TSO for EF10
      sfc: Allocate and link PIO buffers; map them with write-combining
      sfc: Separate out queue-empty check from efx_nic_may_push_tx_desc()
      sfc: Introduce inline functions to simplify TX insertion
      sfc: Support ARFS for IPv6 flows
      sfc: Add static tracepoints to datapath

Fengguang Wu (2):
      sfc: efx_ethtool_get_ts_info() can be static
      sfc: efx_ef10_filter_update_rx_scatter() can be static

Jon Cooper (1):
      sfc: Use TX PIO for sufficiently small packets

 MAINTAINERS                           |   1 +
 drivers/net/ethernet/sfc/Kconfig      |  14 ++
 drivers/net/ethernet/sfc/ef10.c       | 319 ++++++++++++++++++++++++-
 drivers/net/ethernet/sfc/ef10_regs.h  |   1 +
 drivers/net/ethernet/sfc/efx.c        |   6 +
 drivers/net/ethernet/sfc/efx.h        |   1 +
 drivers/net/ethernet/sfc/ethtool.c    |   5 +-
 drivers/net/ethernet/sfc/io.h         |   5 +
 drivers/net/ethernet/sfc/net_driver.h |   7 +
 drivers/net/ethernet/sfc/nic.c        |  73 +++---
 drivers/net/ethernet/sfc/nic.h        |  46 +++-
 drivers/net/ethernet/sfc/ptp.c        |   3 +
 drivers/net/ethernet/sfc/rx.c         |  92 +++++---
 drivers/net/ethernet/sfc/tx.c         | 422 +++++++++++++++++++++++++++-------
 include/trace/events/sfc.h            | 160 +++++++++++++
 15 files changed, 996 insertions(+), 159 deletions(-)
 create mode 100644 include/trace/events/sfc.h

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 01/11] sfc: efx_ethtool_get_ts_info() can be static

[Ben Hutchings]

From: Fengguang Wu <fengguang.wu@intel.com>

Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/ethtool.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/sfc/ethtool.c b/drivers/net/ethernet/sfc/ethtool.c
index 5b471cf..c8dc407 100644
--- a/drivers/net/ethernet/sfc/ethtool.c
+++ b/drivers/net/ethernet/sfc/ethtool.c
@@ -1035,8 +1035,8 @@ static int efx_ethtool_set_rxfh_indir(struct net_device *net_dev,
 	return 0;
 }
 
-int efx_ethtool_get_ts_info(struct net_device *net_dev,
-			    struct ethtool_ts_info *ts_info)
+static int efx_ethtool_get_ts_info(struct net_device *net_dev,
+				   struct ethtool_ts_info *ts_info)
 {
 	struct efx_nic *efx = netdev_priv(net_dev);
 


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 02/11] sfc: efx_ef10_filter_update_rx_scatter() can be static

[Ben Hutchings]

From: Fengguang Wu <fengguang.wu@intel.com>

Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/ef10.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/sfc/ef10.c b/drivers/net/ethernet/sfc/ef10.c
index 5f42313..6956bd2 100644
--- a/drivers/net/ethernet/sfc/ef10.c
+++ b/drivers/net/ethernet/sfc/ef10.c
@@ -2121,7 +2121,7 @@ out_unlock:
 	return rc;
 }
 
-void efx_ef10_filter_update_rx_scatter(struct efx_nic *efx)
+static void efx_ef10_filter_update_rx_scatter(struct efx_nic *efx)
 {
 	/* no need to do anything here on EF10 */
 }


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 03/11] sfc: Add EF10 registers to register dump

[Ben Hutchings]

There are very few readable registers, but we may as well report them.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/nic.c | 73 ++++++++++++++++++++++++------------------
 1 file changed, 42 insertions(+), 31 deletions(-)

diff --git a/drivers/net/ethernet/sfc/nic.c b/drivers/net/ethernet/sfc/nic.c
index e7dbd2d..c75009b 100644
--- a/drivers/net/ethernet/sfc/nic.c
+++ b/drivers/net/ethernet/sfc/nic.c
@@ -19,6 +19,7 @@
 #include "bitfield.h"
 #include "efx.h"
 #include "nic.h"
+#include "ef10_regs.h"
 #include "farch_regs.h"
 #include "io.h"
 #include "workarounds.h"
@@ -166,26 +167,30 @@ void efx_nic_fini_interrupt(struct efx_nic *efx)
 
 /* Register dump */
 
-#define REGISTER_REVISION_A	1
-#define REGISTER_REVISION_B	2
-#define REGISTER_REVISION_C	3
-#define REGISTER_REVISION_Z	3	/* latest revision */
+#define REGISTER_REVISION_FA	1
+#define REGISTER_REVISION_FB	2
+#define REGISTER_REVISION_FC	3
+#define REGISTER_REVISION_FZ	3	/* last Falcon arch revision */
+#define REGISTER_REVISION_ED	4
+#define REGISTER_REVISION_EZ	4	/* latest EF10 revision */
 
 struct efx_nic_reg {
 	u32 offset:24;
-	u32 min_revision:2, max_revision:2;
+	u32 min_revision:3, max_revision:3;
 };
 
-#define REGISTER(name, min_rev, max_rev) {				\
-	FR_ ## min_rev ## max_rev ## _ ## name,				\
-	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
+#define REGISTER(name, arch, min_rev, max_rev) {			\
+	arch ## R_ ## min_rev ## max_rev ## _ ## name,			\
+	REGISTER_REVISION_ ## arch ## min_rev,				\
+	REGISTER_REVISION_ ## arch ## max_rev				\
 }
-#define REGISTER_AA(name) REGISTER(name, A, A)
-#define REGISTER_AB(name) REGISTER(name, A, B)
-#define REGISTER_AZ(name) REGISTER(name, A, Z)
-#define REGISTER_BB(name) REGISTER(name, B, B)
-#define REGISTER_BZ(name) REGISTER(name, B, Z)
-#define REGISTER_CZ(name) REGISTER(name, C, Z)
+#define REGISTER_AA(name) REGISTER(name, F, A, A)
+#define REGISTER_AB(name) REGISTER(name, F, A, B)
+#define REGISTER_AZ(name) REGISTER(name, F, A, Z)
+#define REGISTER_BB(name) REGISTER(name, F, B, B)
+#define REGISTER_BZ(name) REGISTER(name, F, B, Z)
+#define REGISTER_CZ(name) REGISTER(name, F, C, Z)
+#define REGISTER_DZ(name) REGISTER(name, E, D, Z)
 
 static const struct efx_nic_reg efx_nic_regs[] = {
 	REGISTER_AZ(ADR_REGION),
@@ -292,37 +297,42 @@ static const struct efx_nic_reg efx_nic_regs[] = {
 	REGISTER_AB(XX_TXDRV_CTL),
 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
 	/* XX_CORE_STAT is partly RC */
+	REGISTER_DZ(BIU_HW_REV_ID),
+	REGISTER_DZ(MC_DB_LWRD),
+	REGISTER_DZ(MC_DB_HWRD),
 };
 
 struct efx_nic_reg_table {
 	u32 offset:24;
-	u32 min_revision:2, max_revision:2;
+	u32 min_revision:3, max_revision:3;
 	u32 step:6, rows:21;
 };
 
-#define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
+#define REGISTER_TABLE_DIMENSIONS(_, offset, arch, min_rev, max_rev, step, rows) { \
 	offset,								\
-	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
+	REGISTER_REVISION_ ## arch ## min_rev,				\
+	REGISTER_REVISION_ ## arch ## max_rev,				\
 	step, rows							\
 }
-#define REGISTER_TABLE(name, min_rev, max_rev)				\
+#define REGISTER_TABLE(name, arch, min_rev, max_rev)			\
 	REGISTER_TABLE_DIMENSIONS(					\
-		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
-		min_rev, max_rev,					\
-		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
-		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
-#define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
-#define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
-#define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
-#define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
+		name, arch ## R_ ## min_rev ## max_rev ## _ ## name,	\
+		arch, min_rev, max_rev,					\
+		arch ## R_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
+		arch ## R_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
+#define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, F, A, A)
+#define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, F, A, Z)
+#define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, F, B, B)
+#define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, F, B, Z)
 #define REGISTER_TABLE_BB_CZ(name)					\
-	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
+	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, F, B, B,	\
 				  FR_BZ_ ## name ## _STEP,		\
 				  FR_BB_ ## name ## _ROWS),		\
-	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
+	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, F, C, Z,	\
 				  FR_BZ_ ## name ## _STEP,		\
 				  FR_CZ_ ## name ## _ROWS)
-#define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
+#define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, F, C, Z)
+#define REGISTER_TABLE_DZ(name) REGISTER_TABLE(name, E, D, Z)
 
 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
 	/* DRIVER is not used */
@@ -340,9 +350,9 @@ static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
 	 * 1K entries allows for some expansion of queue count and
 	 * size before we need to change the version. */
 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
-				  A, A, 8, 1024),
+				  F, A, A, 8, 1024),
 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
-				  B, Z, 8, 1024),
+				  F, B, Z, 8, 1024),
 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
@@ -353,6 +363,7 @@ static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
 	/* MSIX_PBA_TABLE is not mapped */
 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
+	REGISTER_TABLE_DZ(BIU_MC_SFT_STATUS),
 };
 
 size_t efx_nic_get_regs_len(struct efx_nic *efx)


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 04/11] sfc: Fold tso_get_head_fragment() into tso_start()

[Ben Hutchings]

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/tx.c | 62 +++++++++++++++++++++----------------------
 1 file changed, 31 insertions(+), 31 deletions(-)

diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 2ac91c5..9025356 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -822,11 +822,17 @@ static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 

 /* Parse the SKB header and initialise state. */
-static void tso_start(struct tso_state *st, const struct sk_buff *skb)
+static int tso_start(struct tso_state *st, struct efx_nic *efx,
+		     const struct sk_buff *skb)
 {
+	unsigned int header_len, in_len;
+
 	st->ip_off = skb_network_header(skb) - skb->data;
 	st->tcp_off = skb_transport_header(skb) - skb->data;
-	st->header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
+	header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
+	in_len = skb_headlen(skb) - header_len;
+	st->header_len = header_len;
+	st->in_len = in_len;
 	if (st->protocol == htons(ETH_P_IP)) {
 		st->ip_base_len = st->header_len - st->ip_off;
 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
@@ -840,9 +846,24 @@ static void tso_start(struct tso_state *st, const struct sk_buff *skb)
 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
 
-	st->out_len = skb->len - st->header_len;
-	st->unmap_len = 0;
-	st->dma_flags = 0;
+	st->out_len = skb->len - header_len;
+
+	if (likely(in_len == 0)) {
+		st->unmap_len = 0;
+		st->dma_flags = 0;
+		return 0;
+	}
+
+	st->unmap_addr = dma_map_single(&efx->pci_dev->dev,
+					skb->data + header_len, in_len,
+					DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr)))
+		return -ENOMEM;
+
+	st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
+	st->unmap_len = in_len;
+	st->dma_addr = st->unmap_addr;
+	return 0;
 }
 
 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
@@ -860,24 +881,6 @@ static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
 	return -ENOMEM;
 }
 
-static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
-				 const struct sk_buff *skb)
-{
-	int hl = st->header_len;
-	int len = skb_headlen(skb) - hl;
-
-	st->unmap_addr = dma_map_single(&efx->pci_dev->dev, skb->data + hl,
-					len, DMA_TO_DEVICE);
-	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
-		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
-		st->unmap_len = len;
-		st->in_len = len;
-		st->dma_addr = st->unmap_addr;
-		return 0;
-	}
-	return -ENOMEM;
-}
-
 
 /**
  * tso_fill_packet_with_fragment - form descriptors for the current fragment
@@ -1023,12 +1026,11 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
 
 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
 
-	tso_start(&state, skb);
+	rc = tso_start(&state, efx, skb);
+	if (rc)
+		goto mem_err;
 
-	/* Assume that skb header area contains exactly the headers, and
-	 * all payload is in the frag list.
-	 */
-	if (skb_headlen(skb) == state.header_len) {
+	if (likely(state.in_len == 0)) {
 		/* Grab the first payload fragment. */
 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
 		frag_i = 0;
@@ -1037,9 +1039,7 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
 		if (rc)
 			goto mem_err;
 	} else {
-		rc = tso_get_head_fragment(&state, efx, skb);
-		if (rc)
-			goto mem_err;
+		/* Payload starts in the header area. */
 		frag_i = -1;
 	}
 


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 05/11] sfc: Implement firmware-assisted TSO for EF10

[Ben Hutchings]

Segmentation remains in the driver, but we generate option descriptors
describing the required packet editing rather than making our own
copies.

Reduce tso_state::ipv4_id to 16 bits, so it doesn't overflow into the
TCP_FLAGS field of the option descriptor.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/tx.c | 173 +++++++++++++++++++++++++++++-------------
 1 file changed, 121 insertions(+), 52 deletions(-)

diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 9025356..a23ba0d 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -21,6 +21,7 @@
 #include "efx.h"
 #include "nic.h"
 #include "workarounds.h"
+#include "ef10_regs.h"
 
 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 			       struct efx_tx_buffer *buffer,
@@ -83,8 +84,10 @@ unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
 	 */
 	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
 
-	/* Possibly one more per segment for the alignment workaround */
-	if (EFX_WORKAROUND_5391(efx))
+	/* Possibly one more per segment for the alignment workaround,
+	 * or for option descriptors
+	 */
+	if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
 		max_descs += EFX_TSO_MAX_SEGS;
 
 	/* Possibly more for PCIe page boundaries within input fragments */
@@ -628,6 +631,9 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
  * @tcp_off: Offset of TCP header
  * @header_len: Number of bytes of header
  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
+ * @header_dma_addr: Header DMA address, when using option descriptors
+ * @header_unmap_len: Header DMA mapped length, or 0 if not using option
+ *	descriptors
  *
  * The state used during segmentation.  It is put into this data structure
  * just to make it easy to pass into inline functions.
@@ -636,7 +642,7 @@ struct tso_state {
 	/* Output position */
 	unsigned out_len;
 	unsigned seqnum;
-	unsigned ipv4_id;
+	u16 ipv4_id;
 	unsigned packet_space;
 
 	/* Input position */
@@ -651,6 +657,8 @@ struct tso_state {
 	unsigned int tcp_off;
 	unsigned header_len;
 	unsigned int ip_base_len;
+	dma_addr_t header_dma_addr;
+	unsigned int header_unmap_len;
 };
 

@@ -825,7 +833,10 @@ static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 static int tso_start(struct tso_state *st, struct efx_nic *efx,
 		     const struct sk_buff *skb)
 {
+	bool use_options = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
+	struct device *dma_dev = &efx->pci_dev->dev;
 	unsigned int header_len, in_len;
+	dma_addr_t dma_addr;
 
 	st->ip_off = skb_network_header(skb) - skb->data;
 	st->tcp_off = skb_transport_header(skb) - skb->data;
@@ -848,22 +859,32 @@ static int tso_start(struct tso_state *st, struct efx_nic *efx,
 
 	st->out_len = skb->len - header_len;
 
-	if (likely(in_len == 0)) {
-		st->unmap_len = 0;
+	if (!use_options) {
+		st->header_unmap_len = 0;
+
+		if (likely(in_len == 0)) {
+			st->dma_flags = 0;
+			st->unmap_len = 0;
+			return 0;
+		}
+
+		dma_addr = dma_map_single(dma_dev, skb->data + header_len,
+					  in_len, DMA_TO_DEVICE);
+		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
+		st->dma_addr = dma_addr;
+		st->unmap_addr = dma_addr;
+		st->unmap_len = in_len;
+	} else {
+		dma_addr = dma_map_single(dma_dev, skb->data,
+					  skb_headlen(skb), DMA_TO_DEVICE);
+		st->header_dma_addr = dma_addr;
+		st->header_unmap_len = skb_headlen(skb);
 		st->dma_flags = 0;
-		return 0;
+		st->dma_addr = dma_addr + header_len;
+		st->unmap_len = 0;
 	}
 
-	st->unmap_addr = dma_map_single(&efx->pci_dev->dev,
-					skb->data + header_len, in_len,
-					DMA_TO_DEVICE);
-	if (unlikely(dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr)))
-		return -ENOMEM;
-
-	st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
-	st->unmap_len = in_len;
-	st->dma_addr = st->unmap_addr;
-	return 0;
+	return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
 }
 
 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
@@ -948,54 +969,97 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 {
 	struct efx_tx_buffer *buffer =
 		&tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
-	struct tcphdr *tsoh_th;
-	unsigned ip_length;
-	u8 *header;
-	int rc;
-
-	/* Allocate and insert a DMA-mapped header buffer. */
-	header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
-	if (!header)
-		return -ENOMEM;
+	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
+	u8 tcp_flags_clear;
 
-	tsoh_th = (struct tcphdr *)(header + st->tcp_off);
-
-	/* Copy and update the headers. */
-	memcpy(header, skb->data, st->header_len);
-
-	tsoh_th->seq = htonl(st->seqnum);
-	st->seqnum += skb_shinfo(skb)->gso_size;
-	if (st->out_len > skb_shinfo(skb)->gso_size) {
-		/* This packet will not finish the TSO burst. */
+	if (!is_last) {
 		st->packet_space = skb_shinfo(skb)->gso_size;
-		tsoh_th->fin = 0;
-		tsoh_th->psh = 0;
+		tcp_flags_clear = 0x09; /* mask out FIN and PSH */
 	} else {
-		/* This packet will be the last in the TSO burst. */
 		st->packet_space = st->out_len;
-		tsoh_th->fin = tcp_hdr(skb)->fin;
-		tsoh_th->psh = tcp_hdr(skb)->psh;
+		tcp_flags_clear = 0x00;
 	}
-	ip_length = st->ip_base_len + st->packet_space;
 
-	if (st->protocol == htons(ETH_P_IP)) {
-		struct iphdr *tsoh_iph = (struct iphdr *)(header + st->ip_off);
+	if (!st->header_unmap_len) {
+		/* Allocate and insert a DMA-mapped header buffer. */
+		struct tcphdr *tsoh_th;
+		unsigned ip_length;
+		u8 *header;
+		int rc;
 
-		tsoh_iph->tot_len = htons(ip_length);
+		header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
+		if (!header)
+			return -ENOMEM;
 
-		/* Linux leaves suitable gaps in the IP ID space for us to fill. */
-		tsoh_iph->id = htons(st->ipv4_id);
-		st->ipv4_id++;
+		tsoh_th = (struct tcphdr *)(header + st->tcp_off);
+
+		/* Copy and update the headers. */
+		memcpy(header, skb->data, st->header_len);
+
+		tsoh_th->seq = htonl(st->seqnum);
+		((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
+
+		ip_length = st->ip_base_len + st->packet_space;
+
+		if (st->protocol == htons(ETH_P_IP)) {
+			struct iphdr *tsoh_iph =
+				(struct iphdr *)(header + st->ip_off);
+
+			tsoh_iph->tot_len = htons(ip_length);
+			tsoh_iph->id = htons(st->ipv4_id);
+		} else {
+			struct ipv6hdr *tsoh_iph =
+				(struct ipv6hdr *)(header + st->ip_off);
+
+			tsoh_iph->payload_len = htons(ip_length);
+		}
+
+		rc = efx_tso_put_header(tx_queue, buffer, header);
+		if (unlikely(rc))
+			return rc;
 	} else {
-		struct ipv6hdr *tsoh_iph =
-			(struct ipv6hdr *)(header + st->ip_off);
+		/* Send the original headers with a TSO option descriptor
+		 * in front
+		 */
+		u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
+
+		buffer->flags = EFX_TX_BUF_OPTION;
+		buffer->len = 0;
+		buffer->unmap_len = 0;
+		EFX_POPULATE_QWORD_5(buffer->option,
+				     ESF_DZ_TX_DESC_IS_OPT, 1,
+				     ESF_DZ_TX_OPTION_TYPE,
+				     ESE_DZ_TX_OPTION_DESC_TSO,
+				     ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
+				     ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
+				     ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
+		++tx_queue->insert_count;
 
-		tsoh_iph->payload_len = htons(ip_length);
+		/* We mapped the headers in tso_start().  Unmap them
+		 * when the last segment is completed.
+		 */
+		buffer = &tx_queue->buffer[tx_queue->insert_count &
+					   tx_queue->ptr_mask];
+		buffer->dma_addr = st->header_dma_addr;
+		buffer->len = st->header_len;
+		if (is_last) {
+			buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
+			buffer->unmap_len = st->header_unmap_len;
+			/* Ensure we only unmap them once in case of a
+			 * later DMA mapping error and rollback
+			 */
+			st->header_unmap_len = 0;
+		} else {
+			buffer->flags = EFX_TX_BUF_CONT;
+			buffer->unmap_len = 0;
+		}
+		++tx_queue->insert_count;
 	}
 
-	rc = efx_tso_put_header(tx_queue, buffer, header);
-	if (unlikely(rc))
-		return rc;
+	st->seqnum += skb_shinfo(skb)->gso_size;
+
+	/* Linux leaves suitable gaps in the IP ID space for us to fill. */
+	++st->ipv4_id;
 
 	++tx_queue->tso_packets;
 
@@ -1091,6 +1155,11 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
 				       state.unmap_len, DMA_TO_DEVICE);
 	}
 
+	/* Free the header DMA mapping, if using option descriptors */
+	if (state.header_unmap_len)
+		dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
+				 state.header_unmap_len, DMA_TO_DEVICE);
+
 	efx_enqueue_unwind(tx_queue);
 	return NETDEV_TX_OK;
 }


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 06/11] sfc: Allocate and link PIO buffers; map them with write-combining

[Ben Hutchings]

Try to allocate a segment of PIO buffer to each TX channel.  If
allocation fails, log an error but continue.

PIO buffers must be mapped separately from the NIC registers, with
write-combining enabled.  Where the host page size is 4K, we could
potentially map each VI's registers and PIO buffer separately.
However, this would add significant complexity, and we also need to
support architectures such as POWER which have a greater page size.
So make a single contiguous write-combining mapping after the
uncacheable mapping, aligned to the host page size, and link PIO
buffers there.  Where necessary, allocate additional VIs within
the write-combining mapping purely for access to PIO buffers.

Link all TX buffers to TX queues and the additional VIs in
efx_ef10_dimension_resources() and in efx_ef10_init_nic() after
an MC reboot.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/ef10.c       | 317 +++++++++++++++++++++++++++++++++-
 drivers/net/ethernet/sfc/efx.h        |   1 +
 drivers/net/ethernet/sfc/io.h         |   5 +
 drivers/net/ethernet/sfc/net_driver.h |   5 +
 drivers/net/ethernet/sfc/nic.h        |  18 ++
 drivers/net/ethernet/sfc/tx.c         |  10 ++
 6 files changed, 353 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/sfc/ef10.c b/drivers/net/ethernet/sfc/ef10.c
index 6956bd2..9165c35 100644
--- a/drivers/net/ethernet/sfc/ef10.c
+++ b/drivers/net/ethernet/sfc/ef10.c
@@ -284,6 +284,181 @@ static int efx_ef10_free_vis(struct efx_nic *efx)
 	return rc;
 }
 
+#ifdef EFX_USE_PIO
+
+static void efx_ef10_free_piobufs(struct efx_nic *efx)
+{
+	struct efx_ef10_nic_data *nic_data = efx->nic_data;
+	MCDI_DECLARE_BUF(inbuf, MC_CMD_FREE_PIOBUF_IN_LEN);
+	unsigned int i;
+	int rc;
+
+	BUILD_BUG_ON(MC_CMD_FREE_PIOBUF_OUT_LEN != 0);
+
+	for (i = 0; i < nic_data->n_piobufs; i++) {
+		MCDI_SET_DWORD(inbuf, FREE_PIOBUF_IN_PIOBUF_HANDLE,
+			       nic_data->piobuf_handle[i]);
+		rc = efx_mcdi_rpc(efx, MC_CMD_FREE_PIOBUF, inbuf, sizeof(inbuf),
+				  NULL, 0, NULL);
+		WARN_ON(rc);
+	}
+
+	nic_data->n_piobufs = 0;
+}
+
+static int efx_ef10_alloc_piobufs(struct efx_nic *efx, unsigned int n)
+{
+	struct efx_ef10_nic_data *nic_data = efx->nic_data;
+	MCDI_DECLARE_BUF(outbuf, MC_CMD_ALLOC_PIOBUF_OUT_LEN);
+	unsigned int i;
+	size_t outlen;
+	int rc = 0;
+
+	BUILD_BUG_ON(MC_CMD_ALLOC_PIOBUF_IN_LEN != 0);
+
+	for (i = 0; i < n; i++) {
+		rc = efx_mcdi_rpc(efx, MC_CMD_ALLOC_PIOBUF, NULL, 0,
+				  outbuf, sizeof(outbuf), &outlen);
+		if (rc)
+			break;
+		if (outlen < MC_CMD_ALLOC_PIOBUF_OUT_LEN) {
+			rc = -EIO;
+			break;
+		}
+		nic_data->piobuf_handle[i] =
+			MCDI_DWORD(outbuf, ALLOC_PIOBUF_OUT_PIOBUF_HANDLE);
+		netif_dbg(efx, probe, efx->net_dev,
+			  "allocated PIO buffer %u handle %x\n", i,
+			  nic_data->piobuf_handle[i]);
+	}
+
+	nic_data->n_piobufs = i;
+	if (rc)
+		efx_ef10_free_piobufs(efx);
+	return rc;
+}
+
+static int efx_ef10_link_piobufs(struct efx_nic *efx)
+{
+	struct efx_ef10_nic_data *nic_data = efx->nic_data;
+	MCDI_DECLARE_BUF(inbuf,
+			 max(MC_CMD_LINK_PIOBUF_IN_LEN,
+			     MC_CMD_UNLINK_PIOBUF_IN_LEN));
+	struct efx_channel *channel;
+	struct efx_tx_queue *tx_queue;
+	unsigned int offset, index;
+	int rc;
+
+	BUILD_BUG_ON(MC_CMD_LINK_PIOBUF_OUT_LEN != 0);
+	BUILD_BUG_ON(MC_CMD_UNLINK_PIOBUF_OUT_LEN != 0);
+
+	/* Link a buffer to each VI in the write-combining mapping */
+	for (index = 0; index < nic_data->n_piobufs; ++index) {
+		MCDI_SET_DWORD(inbuf, LINK_PIOBUF_IN_PIOBUF_HANDLE,
+			       nic_data->piobuf_handle[index]);
+		MCDI_SET_DWORD(inbuf, LINK_PIOBUF_IN_TXQ_INSTANCE,
+			       nic_data->pio_write_vi_base + index);
+		rc = efx_mcdi_rpc(efx, MC_CMD_LINK_PIOBUF,
+				  inbuf, MC_CMD_LINK_PIOBUF_IN_LEN,
+				  NULL, 0, NULL);
+		if (rc) {
+			netif_err(efx, drv, efx->net_dev,
+				  "failed to link VI %u to PIO buffer %u (%d)\n",
+				  nic_data->pio_write_vi_base + index, index,
+				  rc);
+			goto fail;
+		}
+		netif_dbg(efx, probe, efx->net_dev,
+			  "linked VI %u to PIO buffer %u\n",
+			  nic_data->pio_write_vi_base + index, index);
+	}
+
+	/* Link a buffer to each TX queue */
+	efx_for_each_channel(channel, efx) {
+		efx_for_each_channel_tx_queue(tx_queue, channel) {
+			/* We assign the PIO buffers to queues in
+			 * reverse order to allow for the following
+			 * special case.
+			 */
+			offset = ((efx->tx_channel_offset + efx->n_tx_channels -
+				   tx_queue->channel->channel - 1) *
+				  efx_piobuf_size);
+			index = offset / ER_DZ_TX_PIOBUF_SIZE;
+			offset = offset % ER_DZ_TX_PIOBUF_SIZE;
+
+			/* When the host page size is 4K, the first
+			 * host page in the WC mapping may be within
+			 * the same VI page as the last TX queue.  We
+			 * can only link one buffer to each VI.
+			 */
+			if (tx_queue->queue == nic_data->pio_write_vi_base) {
+				BUG_ON(index != 0);
+				rc = 0;
+			} else {
+				MCDI_SET_DWORD(inbuf,
+					       LINK_PIOBUF_IN_PIOBUF_HANDLE,
+					       nic_data->piobuf_handle[index]);
+				MCDI_SET_DWORD(inbuf,
+					       LINK_PIOBUF_IN_TXQ_INSTANCE,
+					       tx_queue->queue);
+				rc = efx_mcdi_rpc(efx, MC_CMD_LINK_PIOBUF,
+						  inbuf, MC_CMD_LINK_PIOBUF_IN_LEN,
+						  NULL, 0, NULL);
+			}
+
+			if (rc) {
+				/* This is non-fatal; the TX path just
+				 * won't use PIO for this queue
+				 */
+				netif_err(efx, drv, efx->net_dev,
+					  "failed to link VI %u to PIO buffer %u (%d)\n",
+					  tx_queue->queue, index, rc);
+				tx_queue->piobuf = NULL;
+			} else {
+				tx_queue->piobuf =
+					nic_data->pio_write_base +
+					index * EFX_VI_PAGE_SIZE + offset;
+				tx_queue->piobuf_offset = offset;
+				netif_dbg(efx, probe, efx->net_dev,
+					  "linked VI %u to PIO buffer %u offset %x addr %p\n",
+					  tx_queue->queue, index,
+					  tx_queue->piobuf_offset,
+					  tx_queue->piobuf);
+			}
+		}
+	}
+
+	return 0;
+
+fail:
+	while (index--) {
+		MCDI_SET_DWORD(inbuf, UNLINK_PIOBUF_IN_TXQ_INSTANCE,
+			       nic_data->pio_write_vi_base + index);
+		efx_mcdi_rpc(efx, MC_CMD_UNLINK_PIOBUF,
+			     inbuf, MC_CMD_UNLINK_PIOBUF_IN_LEN,
+			     NULL, 0, NULL);
+	}
+	return rc;
+}
+
+#else /* !EFX_USE_PIO */
+
+static int efx_ef10_alloc_piobufs(struct efx_nic *efx, unsigned int n)
+{
+	return n == 0 ? 0 : -ENOBUFS;
+}
+
+static int efx_ef10_link_piobufs(struct efx_nic *efx)
+{
+	return 0;
+}
+
+static void efx_ef10_free_piobufs(struct efx_nic *efx)
+{
+}
+
+#endif /* EFX_USE_PIO */
+
 static void efx_ef10_remove(struct efx_nic *efx)
 {
 	struct efx_ef10_nic_data *nic_data = efx->nic_data;
@@ -294,9 +469,15 @@ static void efx_ef10_remove(struct efx_nic *efx)
 	/* This needs to be after efx_ptp_remove_channel() with no filters */
 	efx_ef10_rx_free_indir_table(efx);
 
+	if (nic_data->wc_membase)
+		iounmap(nic_data->wc_membase);
+
 	rc = efx_ef10_free_vis(efx);
 	WARN_ON(rc != 0);
 
+	if (!nic_data->must_restore_piobufs)
+		efx_ef10_free_piobufs(efx);
+
 	efx_mcdi_fini(efx);
 	efx_nic_free_buffer(efx, &nic_data->mcdi_buf);
 	kfree(nic_data);
@@ -329,12 +510,126 @@ static int efx_ef10_alloc_vis(struct efx_nic *efx,
 	return 0;
 }
 
+/* Note that the failure path of this function does not free
+ * resources, as this will be done by efx_ef10_remove().
+ */
 static int efx_ef10_dimension_resources(struct efx_nic *efx)
 {
-	unsigned int n_vis =
-		max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
+	struct efx_ef10_nic_data *nic_data = efx->nic_data;
+	unsigned int uc_mem_map_size, wc_mem_map_size;
+	unsigned int min_vis, pio_write_vi_base, max_vis;
+	void __iomem *membase;
+	int rc;
+
+	min_vis = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
+
+#ifdef EFX_USE_PIO
+	/* Try to allocate PIO buffers if wanted and if the full
+	 * number of PIO buffers would be sufficient to allocate one
+	 * copy-buffer per TX channel.  Failure is non-fatal, as there
+	 * are only a small number of PIO buffers shared between all
+	 * functions of the controller.
+	 */
+	if (efx_piobuf_size != 0 &&
+	    ER_DZ_TX_PIOBUF_SIZE / efx_piobuf_size * EF10_TX_PIOBUF_COUNT >=
+	    efx->n_tx_channels) {
+		unsigned int n_piobufs =
+			DIV_ROUND_UP(efx->n_tx_channels,
+				     ER_DZ_TX_PIOBUF_SIZE / efx_piobuf_size);
+
+		rc = efx_ef10_alloc_piobufs(efx, n_piobufs);
+		if (rc)
+			netif_err(efx, probe, efx->net_dev,
+				  "failed to allocate PIO buffers (%d)\n", rc);
+		else
+			netif_dbg(efx, probe, efx->net_dev,
+				  "allocated %u PIO buffers\n", n_piobufs);
+	}
+#else
+	nic_data->n_piobufs = 0;
+#endif
+
+	/* PIO buffers should be mapped with write-combining enabled,
+	 * and we want to make single UC and WC mappings rather than
+	 * several of each (in fact that's the only option if host
+	 * page size is >4K).  So we may allocate some extra VIs just
+	 * for writing PIO buffers through.
+	 */
+	uc_mem_map_size = PAGE_ALIGN((min_vis - 1) * EFX_VI_PAGE_SIZE +
+				     ER_DZ_TX_PIOBUF);
+	if (nic_data->n_piobufs) {
+		pio_write_vi_base = uc_mem_map_size / EFX_VI_PAGE_SIZE;
+		wc_mem_map_size = (PAGE_ALIGN((pio_write_vi_base +
+					       nic_data->n_piobufs) *
+					      EFX_VI_PAGE_SIZE) -
+				   uc_mem_map_size);
+		max_vis = pio_write_vi_base + nic_data->n_piobufs;
+	} else {
+		pio_write_vi_base = 0;
+		wc_mem_map_size = 0;
+		max_vis = min_vis;
+	}
+
+	/* In case the last attached driver failed to free VIs, do it now */
+	rc = efx_ef10_free_vis(efx);
+	if (rc != 0)
+		return rc;
+
+	rc = efx_ef10_alloc_vis(efx, min_vis, max_vis);
+	if (rc != 0)
+		return rc;
 
-	return efx_ef10_alloc_vis(efx, n_vis, n_vis);
+	/* If we didn't get enough VIs to map all the PIO buffers, free the
+	 * PIO buffers
+	 */
+	if (nic_data->n_piobufs &&
+	    nic_data->n_allocated_vis <
+	    pio_write_vi_base + nic_data->n_piobufs) {
+		netif_dbg(efx, probe, efx->net_dev,
+			  "%u VIs are not sufficient to map %u PIO buffers\n",
+			  nic_data->n_allocated_vis, nic_data->n_piobufs);
+		efx_ef10_free_piobufs(efx);
+	}
+
+	/* Shrink the original UC mapping of the memory BAR */
+	membase = ioremap_nocache(efx->membase_phys, uc_mem_map_size);
+	if (!membase) {
+		netif_err(efx, probe, efx->net_dev,
+			  "could not shrink memory BAR to %x\n",
+			  uc_mem_map_size);
+		return -ENOMEM;
+	}
+	iounmap(efx->membase);
+	efx->membase = membase;
+
+	/* Set up the WC mapping if needed */
+	if (wc_mem_map_size) {
+		nic_data->wc_membase = ioremap_wc(efx->membase_phys +
+						  uc_mem_map_size,
+						  wc_mem_map_size);
+		if (!nic_data->wc_membase) {
+			netif_err(efx, probe, efx->net_dev,
+				  "could not allocate WC mapping of size %x\n",
+				  wc_mem_map_size);
+			return -ENOMEM;
+		}
+		nic_data->pio_write_vi_base = pio_write_vi_base;
+		nic_data->pio_write_base =
+			nic_data->wc_membase +
+			(pio_write_vi_base * EFX_VI_PAGE_SIZE + ER_DZ_TX_PIOBUF -
+			 uc_mem_map_size);
+
+		rc = efx_ef10_link_piobufs(efx);
+		if (rc)
+			efx_ef10_free_piobufs(efx);
+	}
+
+	netif_dbg(efx, probe, efx->net_dev,
+		  "memory BAR at %pa (virtual %p+%x UC, %p+%x WC)\n",
+		  &efx->membase_phys, efx->membase, uc_mem_map_size,
+		  nic_data->wc_membase, wc_mem_map_size);
+
+	return 0;
 }
 
 static int efx_ef10_init_nic(struct efx_nic *efx)
@@ -351,6 +646,21 @@ static int efx_ef10_init_nic(struct efx_nic *efx)
 		nic_data->must_realloc_vis = false;
 	}
 
+	if (nic_data->must_restore_piobufs && nic_data->n_piobufs) {
+		rc = efx_ef10_alloc_piobufs(efx, nic_data->n_piobufs);
+		if (rc == 0) {
+			rc = efx_ef10_link_piobufs(efx);
+			if (rc)
+				efx_ef10_free_piobufs(efx);
+		}
+
+		/* Log an error on failure, but this is non-fatal */
+		if (rc)
+			netif_err(efx, drv, efx->net_dev,
+				  "failed to restore PIO buffers (%d)\n", rc);
+		nic_data->must_restore_piobufs = false;
+	}
+
 	efx_ef10_rx_push_indir_table(efx);
 	return 0;
 }
@@ -708,6 +1018,7 @@ static int efx_ef10_mcdi_poll_reboot(struct efx_nic *efx)
 	/* All our allocations have been reset */
 	nic_data->must_realloc_vis = true;
 	nic_data->must_restore_filters = true;
+	nic_data->must_restore_piobufs = true;
 	nic_data->rx_rss_context = EFX_EF10_RSS_CONTEXT_INVALID;
 
 	return -EIO;
diff --git a/drivers/net/ethernet/sfc/efx.h b/drivers/net/ethernet/sfc/efx.h
index 34d00f5..31d0128 100644
--- a/drivers/net/ethernet/sfc/efx.h
+++ b/drivers/net/ethernet/sfc/efx.h
@@ -30,6 +30,7 @@ efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb);
 extern void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index);
 extern int efx_setup_tc(struct net_device *net_dev, u8 num_tc);
 extern unsigned int efx_tx_max_skb_descs(struct efx_nic *efx);
+extern unsigned int efx_piobuf_size;
 
 /* RX */
 extern void efx_rx_config_page_split(struct efx_nic *efx);
diff --git a/drivers/net/ethernet/sfc/io.h b/drivers/net/ethernet/sfc/io.h
index 96ce507..4d3f119 100644
--- a/drivers/net/ethernet/sfc/io.h
+++ b/drivers/net/ethernet/sfc/io.h
@@ -66,6 +66,11 @@
 #define EFX_USE_QWORD_IO 1
 #endif
 
+/* PIO is a win only if write-combining is possible */
+#ifdef ARCH_HAS_IOREMAP_WC
+#define EFX_USE_PIO 1
+#endif
+
 #ifdef EFX_USE_QWORD_IO
 static inline void _efx_writeq(struct efx_nic *efx, __le64 value,
 				  unsigned int reg)
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
index b172ed1..6febecc 100644
--- a/drivers/net/ethernet/sfc/net_driver.h
+++ b/drivers/net/ethernet/sfc/net_driver.h
@@ -182,6 +182,9 @@ struct efx_tx_buffer {
  * @tsoh_page: Array of pages of TSO header buffers
  * @txd: The hardware descriptor ring
  * @ptr_mask: The size of the ring minus 1.
+ * @piobuf: PIO buffer region for this TX queue (shared with its partner).
+ *	Size of the region is efx_piobuf_size.
+ * @piobuf_offset: Buffer offset to be specified in PIO descriptors
  * @initialised: Has hardware queue been initialised?
  * @read_count: Current read pointer.
  *	This is the number of buffers that have been removed from both rings.
@@ -223,6 +226,8 @@ struct efx_tx_queue {
 	struct efx_buffer *tsoh_page;
 	struct efx_special_buffer txd;
 	unsigned int ptr_mask;
+	void __iomem *piobuf;
+	unsigned int piobuf_offset;
 	bool initialised;
 
 	/* Members used mainly on the completion path */
diff --git a/drivers/net/ethernet/sfc/nic.h b/drivers/net/ethernet/sfc/nic.h
index 4b1e188..a6e4336 100644
--- a/drivers/net/ethernet/sfc/nic.h
+++ b/drivers/net/ethernet/sfc/nic.h
@@ -389,6 +389,12 @@ enum {
 	EF10_STAT_COUNT
 };
 
+/* Maximum number of TX PIO buffers we may allocate to a function.
+ * This matches the total number of buffers on each SFC9100-family
+ * controller.
+ */
+#define EF10_TX_PIOBUF_COUNT 16
+
 /**
  * struct efx_ef10_nic_data - EF10 architecture NIC state
  * @mcdi_buf: DMA buffer for MCDI
@@ -397,6 +403,13 @@ enum {
  * @n_allocated_vis: Number of VIs allocated to this function
  * @must_realloc_vis: Flag: VIs have yet to be reallocated after MC reboot
  * @must_restore_filters: Flag: filters have yet to be restored after MC reboot
+ * @n_piobufs: Number of PIO buffers allocated to this function
+ * @wc_membase: Base address of write-combining mapping of the memory BAR
+ * @pio_write_base: Base address for writing PIO buffers
+ * @pio_write_vi_base: Relative VI number for @pio_write_base
+ * @piobuf_handle: Handle of each PIO buffer allocated
+ * @must_restore_piobufs: Flag: PIO buffers have yet to be restored after MC
+ *	reboot
  * @rx_rss_context: Firmware handle for our RSS context
  * @stats: Hardware statistics
  * @workaround_35388: Flag: firmware supports workaround for bug 35388
@@ -410,6 +423,11 @@ struct efx_ef10_nic_data {
 	unsigned int n_allocated_vis;
 	bool must_realloc_vis;
 	bool must_restore_filters;
+	unsigned int n_piobufs;
+	void __iomem *wc_membase, *pio_write_base;
+	unsigned int pio_write_vi_base;
+	unsigned int piobuf_handle[EF10_TX_PIOBUF_COUNT];
+	bool must_restore_piobufs;
 	u32 rx_rss_context;
 	u64 stats[EF10_STAT_COUNT];
 	bool workaround_35388;
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index a23ba0d..8754347 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -17,12 +17,22 @@
 #include <net/ipv6.h>
 #include <linux/if_ether.h>
 #include <linux/highmem.h>
+#include <linux/cache.h>
 #include "net_driver.h"
 #include "efx.h"
+#include "io.h"
 #include "nic.h"
 #include "workarounds.h"
 #include "ef10_regs.h"
 
+#ifdef EFX_USE_PIO
+
+#define EFX_PIOBUF_SIZE_MAX ER_DZ_TX_PIOBUF_SIZE
+#define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES)
+unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF;
+
+#endif /* EFX_USE_PIO */
+
 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 			       struct efx_tx_buffer *buffer,
 			       unsigned int *pkts_compl,


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 07/11] sfc: Separate out queue-empty check from efx_nic_may_push_tx_desc()

[Ben Hutchings]

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/nic.h | 28 ++++++++++++++++++++++------
 1 file changed, 22 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/sfc/nic.h b/drivers/net/ethernet/sfc/nic.h
index a6e4336..53f8cf6 100644
--- a/drivers/net/ethernet/sfc/nic.h
+++ b/drivers/net/ethernet/sfc/nic.h
@@ -71,6 +71,26 @@ efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
 	return ((efx_qword_t *) (tx_queue->txd.buf.addr)) + index;
 }
 
+/* Report whether the NIC considers this TX queue empty, given the
+ * write_count used for the last doorbell push.  May return false
+ * negative.
+ */
+static inline bool __efx_nic_tx_is_empty(struct efx_tx_queue *tx_queue,
+					 unsigned int write_count)
+{
+	unsigned int empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
+
+	if (empty_read_count == 0)
+		return false;
+
+	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
+}
+
+static inline bool efx_nic_tx_is_empty(struct efx_tx_queue *tx_queue)
+{
+	return __efx_nic_tx_is_empty(tx_queue, tx_queue->write_count);
+}
+
 /* Decide whether to push a TX descriptor to the NIC vs merely writing
  * the doorbell.  This can reduce latency when we are adding a single
  * descriptor to an empty queue, but is otherwise pointless.  Further,
@@ -80,14 +100,10 @@ efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
 static inline bool efx_nic_may_push_tx_desc(struct efx_tx_queue *tx_queue,
 					    unsigned int write_count)
 {
-	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
-
-	if (empty_read_count == 0)
-		return false;
+	bool was_empty = __efx_nic_tx_is_empty(tx_queue, write_count);
 
 	tx_queue->empty_read_count = 0;
-	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
-		&& tx_queue->write_count - write_count == 1;
+	return was_empty && tx_queue->write_count - write_count == 1;
 }
 
 /* Returns a pointer to the specified descriptor in the RX descriptor queue */


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 08/11] sfc: Introduce inline functions to simplify TX insertion

[Ben Hutchings]

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/tx.c | 53 +++++++++++++++++++++++++++----------------
 1 file changed, 33 insertions(+), 20 deletions(-)

diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 8754347..6cecda3 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -33,6 +33,31 @@ unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF;
 
 #endif /* EFX_USE_PIO */
 
+static inline unsigned int
+efx_tx_queue_get_insert_index(const struct efx_tx_queue *tx_queue)
+{
+	return tx_queue->insert_count & tx_queue->ptr_mask;
+}
+
+static inline struct efx_tx_buffer *
+__efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue)
+{
+	return &tx_queue->buffer[efx_tx_queue_get_insert_index(tx_queue)];
+}
+
+static inline struct efx_tx_buffer *
+efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue)
+{
+	struct efx_tx_buffer *buffer =
+		__efx_tx_queue_get_insert_buffer(tx_queue);
+
+	EFX_BUG_ON_PARANOID(buffer->len);
+	EFX_BUG_ON_PARANOID(buffer->flags);
+	EFX_BUG_ON_PARANOID(buffer->unmap_len);
+
+	return buffer;
+}
+
 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 			       struct efx_tx_buffer *buffer,
 			       unsigned int *pkts_compl,
@@ -180,7 +205,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 	struct device *dma_dev = &efx->pci_dev->dev;
 	struct efx_tx_buffer *buffer;
 	skb_frag_t *fragment;
-	unsigned int len, unmap_len = 0, insert_ptr;
+	unsigned int len, unmap_len = 0;
 	dma_addr_t dma_addr, unmap_addr = 0;
 	unsigned int dma_len;
 	unsigned short dma_flags;
@@ -221,11 +246,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 
 		/* Add to TX queue, splitting across DMA boundaries */
 		do {
-			insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
-			buffer = &tx_queue->buffer[insert_ptr];
-			EFX_BUG_ON_PARANOID(buffer->flags);
-			EFX_BUG_ON_PARANOID(buffer->len);
-			EFX_BUG_ON_PARANOID(buffer->unmap_len);
+			buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 
 			dma_len = efx_max_tx_len(efx, dma_addr);
 			if (likely(dma_len >= len))
@@ -283,8 +304,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 	while (tx_queue->insert_count != tx_queue->write_count) {
 		unsigned int pkts_compl = 0, bytes_compl = 0;
 		--tx_queue->insert_count;
-		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
-		buffer = &tx_queue->buffer[insert_ptr];
+		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
 	}
 
@@ -755,23 +775,18 @@ static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 {
 	struct efx_tx_buffer *buffer;
 	struct efx_nic *efx = tx_queue->efx;
-	unsigned dma_len, insert_ptr;
+	unsigned dma_len;
 
 	EFX_BUG_ON_PARANOID(len <= 0);
 
 	while (1) {
-		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
-		buffer = &tx_queue->buffer[insert_ptr];
+		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 		++tx_queue->insert_count;
 
 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
 				    tx_queue->read_count >=
 				    efx->txq_entries);
 
-		EFX_BUG_ON_PARANOID(buffer->len);
-		EFX_BUG_ON_PARANOID(buffer->unmap_len);
-		EFX_BUG_ON_PARANOID(buffer->flags);
-
 		buffer->dma_addr = dma_addr;
 
 		dma_len = efx_max_tx_len(efx, dma_addr);
@@ -832,8 +847,7 @@ static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 	/* Work backwards until we hit the original insert pointer value */
 	while (tx_queue->insert_count != tx_queue->write_count) {
 		--tx_queue->insert_count;
-		buffer = &tx_queue->buffer[tx_queue->insert_count &
-					   tx_queue->ptr_mask];
+		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
 	}
 }
@@ -978,7 +992,7 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 				struct tso_state *st)
 {
 	struct efx_tx_buffer *buffer =
-		&tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
+		efx_tx_queue_get_insert_buffer(tx_queue);
 	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
 	u8 tcp_flags_clear;
 
@@ -1048,8 +1062,7 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 		/* We mapped the headers in tso_start().  Unmap them
 		 * when the last segment is completed.
 		 */
-		buffer = &tx_queue->buffer[tx_queue->insert_count &
-					   tx_queue->ptr_mask];
+		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 		buffer->dma_addr = st->header_dma_addr;
 		buffer->len = st->header_len;
 		if (is_last) {


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 09/11] sfc: Use TX PIO for sufficiently small packets

[Ben Hutchings]

From: Jon Cooper <jcooper@solarflare.com>

Sufficiently small linear packets can be copied into the PIO buffer
with a single call to memcpy_toio().  Non-linear packets require an
intermediate cache-line-sized buffer.

[bwh: I wrote the first version of this, but Jon did the hard work to
 handle non-linear packets.]
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/ef10_regs.h  |   1 +
 drivers/net/ethernet/sfc/ethtool.c    |   1 +
 drivers/net/ethernet/sfc/net_driver.h |   2 +
 drivers/net/ethernet/sfc/tx.c         | 151 ++++++++++++++++++++++++++++++++++
 4 files changed, 155 insertions(+)

diff --git a/drivers/net/ethernet/sfc/ef10_regs.h b/drivers/net/ethernet/sfc/ef10_regs.h
index b3f4e37..207ac9a 100644
--- a/drivers/net/ethernet/sfc/ef10_regs.h
+++ b/drivers/net/ethernet/sfc/ef10_regs.h
@@ -315,6 +315,7 @@
 #define	ESF_DZ_TX_PIO_TYPE_WIDTH 1
 #define	ESF_DZ_TX_PIO_OPT_LBN 60
 #define	ESF_DZ_TX_PIO_OPT_WIDTH 3
+#define	ESE_DZ_TX_OPTION_DESC_PIO 1
 #define	ESF_DZ_TX_PIO_CONT_LBN 59
 #define	ESF_DZ_TX_PIO_CONT_WIDTH 1
 #define	ESF_DZ_TX_PIO_BYTE_CNT_LBN 32
diff --git a/drivers/net/ethernet/sfc/ethtool.c b/drivers/net/ethernet/sfc/ethtool.c
index c8dc407..1f529fa 100644
--- a/drivers/net/ethernet/sfc/ethtool.c
+++ b/drivers/net/ethernet/sfc/ethtool.c
@@ -70,6 +70,7 @@ static const struct efx_sw_stat_desc efx_sw_stat_desc[] = {
 	EFX_ETHTOOL_UINT_TXQ_STAT(tso_long_headers),
 	EFX_ETHTOOL_UINT_TXQ_STAT(tso_packets),
 	EFX_ETHTOOL_UINT_TXQ_STAT(pushes),
+	EFX_ETHTOOL_UINT_TXQ_STAT(pio_packets),
 	EFX_ETHTOOL_ATOMIC_NIC_ERROR_STAT(rx_reset),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_tobe_disc),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_ip_hdr_chksum_err),
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
index 6febecc..aac22a1 100644
--- a/drivers/net/ethernet/sfc/net_driver.h
+++ b/drivers/net/ethernet/sfc/net_driver.h
@@ -212,6 +212,7 @@ struct efx_tx_buffer {
  *	blocks
  * @tso_packets: Number of packets via the TSO xmit path
  * @pushes: Number of times the TX push feature has been used
+ * @pio_packets: Number of times the TX PIO feature has been used
  * @empty_read_count: If the completion path has seen the queue as empty
  *	and the transmission path has not yet checked this, the value of
  *	@read_count bitwise-added to %EFX_EMPTY_COUNT_VALID; otherwise 0.
@@ -243,6 +244,7 @@ struct efx_tx_queue {
 	unsigned int tso_long_headers;
 	unsigned int tso_packets;
 	unsigned int pushes;
+	unsigned int pio_packets;
 
 	/* Members shared between paths and sometimes updated */
 	unsigned int empty_read_count ____cacheline_aligned_in_smp;
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 6cecda3..282692c 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -183,6 +183,145 @@ static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
 	}
 }
 
+#ifdef EFX_USE_PIO
+
+struct efx_short_copy_buffer {
+	int used;
+	u8 buf[L1_CACHE_BYTES];
+};
+
+/* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
+ * Advances piobuf pointer. Leaves additional data in the copy buffer.
+ */
+static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
+				    u8 *data, int len,
+				    struct efx_short_copy_buffer *copy_buf)
+{
+	int block_len = len & ~(sizeof(copy_buf->buf) - 1);
+
+	memcpy_toio(*piobuf, data, block_len);
+	*piobuf += block_len;
+	len -= block_len;
+
+	if (len) {
+		data += block_len;
+		BUG_ON(copy_buf->used);
+		BUG_ON(len > sizeof(copy_buf->buf));
+		memcpy(copy_buf->buf, data, len);
+		copy_buf->used = len;
+	}
+}
+
+/* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
+ * Advances piobuf pointer. Leaves additional data in the copy buffer.
+ */
+static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
+				       u8 *data, int len,
+				       struct efx_short_copy_buffer *copy_buf)
+{
+	if (copy_buf->used) {
+		/* if the copy buffer is partially full, fill it up and write */
+		int copy_to_buf =
+			min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
+
+		memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
+		copy_buf->used += copy_to_buf;
+
+		/* if we didn't fill it up then we're done for now */
+		if (copy_buf->used < sizeof(copy_buf->buf))
+			return;
+
+		memcpy_toio(*piobuf, copy_buf->buf, sizeof(copy_buf->buf));
+		*piobuf += sizeof(copy_buf->buf);
+		data += copy_to_buf;
+		len -= copy_to_buf;
+		copy_buf->used = 0;
+	}
+
+	efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
+}
+
+static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
+				  struct efx_short_copy_buffer *copy_buf)
+{
+	/* if there's anything in it, write the whole buffer, including junk */
+	if (copy_buf->used)
+		memcpy_toio(piobuf, copy_buf->buf, sizeof(copy_buf->buf));
+}
+
+/* Traverse skb structure and copy fragments in to PIO buffer.
+ * Advances piobuf pointer.
+ */
+static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
+				     u8 __iomem **piobuf,
+				     struct efx_short_copy_buffer *copy_buf)
+{
+	int i;
+
+	efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
+				copy_buf);
+
+	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
+		skb_frag_t *f = &skb_shinfo(skb)->frags[i];
+		u8 *vaddr;
+
+		vaddr = kmap_atomic(skb_frag_page(f));
+
+		efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset,
+					   skb_frag_size(f), copy_buf);
+		kunmap_atomic(vaddr);
+	}
+
+	EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list);
+}
+
+static struct efx_tx_buffer *
+efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
+{
+	struct efx_tx_buffer *buffer =
+		efx_tx_queue_get_insert_buffer(tx_queue);
+	u8 __iomem *piobuf = tx_queue->piobuf;
+
+	/* Copy to PIO buffer. Ensure the writes are padded to the end
+	 * of a cache line, as this is required for write-combining to be
+	 * effective on at least x86.
+	 */
+
+	if (skb_shinfo(skb)->nr_frags) {
+		/* The size of the copy buffer will ensure all writes
+		 * are the size of a cache line.
+		 */
+		struct efx_short_copy_buffer copy_buf;
+
+		copy_buf.used = 0;
+
+		efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
+					 &piobuf, &copy_buf);
+		efx_flush_copy_buffer(tx_queue->efx, piobuf, &copy_buf);
+	} else {
+		/* Pad the write to the size of a cache line.
+		 * We can do this because we know the skb_shared_info sruct is
+		 * after the source, and the destination buffer is big enough.
+		 */
+		BUILD_BUG_ON(L1_CACHE_BYTES >
+			     SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
+		memcpy_toio(tx_queue->piobuf, skb->data,
+			    ALIGN(skb->len, L1_CACHE_BYTES));
+	}
+
+	EFX_POPULATE_QWORD_5(buffer->option,
+			     ESF_DZ_TX_DESC_IS_OPT, 1,
+			     ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
+			     ESF_DZ_TX_PIO_CONT, 0,
+			     ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
+			     ESF_DZ_TX_PIO_BUF_ADDR,
+			     tx_queue->piobuf_offset);
+	++tx_queue->pio_packets;
+	++tx_queue->insert_count;
+	return buffer;
+}
+#endif /* EFX_USE_PIO */
+
 /*
  * Add a socket buffer to a TX queue
  *
@@ -227,6 +366,17 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 			return NETDEV_TX_OK;
 	}
 
+	/* Consider using PIO for short packets */
+#ifdef EFX_USE_PIO
+	if (skb->len <= efx_piobuf_size && tx_queue->piobuf &&
+	    efx_nic_tx_is_empty(tx_queue) &&
+	    efx_nic_tx_is_empty(efx_tx_queue_partner(tx_queue))) {
+		buffer = efx_enqueue_skb_pio(tx_queue, skb);
+		dma_flags = EFX_TX_BUF_OPTION;
+		goto finish_packet;
+	}
+#endif
+
 	/* Map for DMA.  Use dma_map_single rather than dma_map_page
 	 * since this is more efficient on machines with sparse
 	 * memory.
@@ -279,6 +429,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 	}
 
 	/* Transfer ownership of the skb to the final buffer */
+finish_packet:
 	buffer->skb = skb;
 	buffer->flags = EFX_TX_BUF_SKB | dma_flags;
 


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 10/11] sfc: Support ARFS for IPv6 flows

[Ben Hutchings]

Extend efx_filter_rfs() to map TCP/IPv6 and UDP/IPv6 flows into
efx_filter_spec.  These are only supported on EF10; on Falcon and
Siena they will be rejected by efx_farch_filter_from_gen_spec().

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 drivers/net/ethernet/sfc/rx.c | 89 +++++++++++++++++++++++++++++--------------
 1 file changed, 61 insertions(+), 28 deletions(-)

diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index 4a59672..20c46f3 100644
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -818,44 +818,70 @@ int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 	struct efx_nic *efx = netdev_priv(net_dev);
 	struct efx_channel *channel;
 	struct efx_filter_spec spec;
-	const struct iphdr *ip;
 	const __be16 *ports;
+	__be16 ether_type;
 	int nhoff;
 	int rc;
 
-	nhoff = skb_network_offset(skb);
+	/* The core RPS/RFS code has already parsed and validated
+	 * VLAN, IP and transport headers.  We assume they are in the
+	 * header area.
+	 */
 
 	if (skb->protocol == htons(ETH_P_8021Q)) {
-		EFX_BUG_ON_PARANOID(skb_headlen(skb) <
-				    nhoff + sizeof(struct vlan_hdr));
-		if (((const struct vlan_hdr *)skb->data + nhoff)->
-		    h_vlan_encapsulated_proto != htons(ETH_P_IP))
-			return -EPROTONOSUPPORT;
+		const struct vlan_hdr *vh =
+			(const struct vlan_hdr *)skb->data;
 
-		/* This is IP over 802.1q VLAN.  We can't filter on the
-		 * IP 5-tuple and the vlan together, so just strip the
-		 * vlan header and filter on the IP part.
+		/* We can't filter on the IP 5-tuple and the vlan
+		 * together, so just strip the vlan header and filter
+		 * on the IP part.
 		 */
-		nhoff += sizeof(struct vlan_hdr);
-	} else if (skb->protocol != htons(ETH_P_IP)) {
-		return -EPROTONOSUPPORT;
+		EFX_BUG_ON_PARANOID(skb_headlen(skb) < sizeof(*vh));
+		ether_type = vh->h_vlan_encapsulated_proto;
+		nhoff = sizeof(struct vlan_hdr);
+	} else {
+		ether_type = skb->protocol;
+		nhoff = 0;
 	}
 
-	/* RFS must validate the IP header length before calling us */
-	EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
-	ip = (const struct iphdr *)(skb->data + nhoff);
-	if (ip_is_fragment(ip))
+	if (ether_type != htons(ETH_P_IP) && ether_type != htons(ETH_P_IPV6))
 		return -EPROTONOSUPPORT;
-	EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
-	ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
 
 	efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 			   rxq_index);
-	rc = efx_filter_set_ipv4_full(&spec, ip->protocol,
-				      ip->daddr, ports[1], ip->saddr, ports[0]);
-	if (rc)
-		return rc;
+	spec.match_flags =
+		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
+		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
+		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
+	spec.ether_type = ether_type;
+
+	if (ether_type == htons(ETH_P_IP)) {
+		const struct iphdr *ip =
+			(const struct iphdr *)(skb->data + nhoff);
+
+		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
+		if (ip_is_fragment(ip))
+			return -EPROTONOSUPPORT;
+		spec.ip_proto = ip->protocol;
+		spec.rem_host[0] = ip->saddr;
+		spec.loc_host[0] = ip->daddr;
+		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
+		ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
+	} else {
+		const struct ipv6hdr *ip6 =
+			(const struct ipv6hdr *)(skb->data + nhoff);
+
+		EFX_BUG_ON_PARANOID(skb_headlen(skb) <
+				    nhoff + sizeof(*ip6) + 4);
+		spec.ip_proto = ip6->nexthdr;
+		memcpy(spec.rem_host, &ip6->saddr, sizeof(ip6->saddr));
+		memcpy(spec.loc_host, &ip6->daddr, sizeof(ip6->daddr));
+		ports = (const __be16 *)(ip6 + 1);
+	}
+
+	spec.rem_port = ports[0];
+	spec.loc_port = ports[1];
 
 	rc = efx->type->filter_rfs_insert(efx, &spec);
 	if (rc < 0)
@@ -866,11 +892,18 @@ int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 	channel = efx_get_channel(efx, skb_get_rx_queue(skb));
 	++channel->rfs_filters_added;
 
-	netif_info(efx, rx_status, efx->net_dev,
-		   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
-		   (ip->protocol == IPPROTO_TCP) ? "TCP" : "UDP",
-		   &ip->saddr, ntohs(ports[0]), &ip->daddr, ntohs(ports[1]),
-		   rxq_index, flow_id, rc);
+	if (ether_type == htons(ETH_P_IP))
+		netif_info(efx, rx_status, efx->net_dev,
+			   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
+			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
+			   spec.rem_host, ntohs(ports[0]), spec.loc_host,
+			   ntohs(ports[1]), rxq_index, flow_id, rc);
+	else
+		netif_info(efx, rx_status, efx->net_dev,
+			   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
+			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
+			   spec.rem_host, ntohs(ports[0]), spec.loc_host,
+			   ntohs(ports[1]), rxq_index, flow_id, rc);
 
 	return rc;
 }


-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

[PATCH net-next 11/11] sfc: Add static tracepoints to datapath

[Ben Hutchings]

These tracepoints support the driver-specific datapath feature tests
we're running internally, though they might be useful for other
purposes.  The skb fields are chosen to cover driver features
implemented now or likely to be added soon.

(Includes a bug fix from Edward Cree.)

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
---
 MAINTAINERS                      |   1 +
 drivers/net/ethernet/sfc/Kconfig |  14 ++++
 drivers/net/ethernet/sfc/efx.c   |   6 ++
 drivers/net/ethernet/sfc/ptp.c   |   3 +
 drivers/net/ethernet/sfc/rx.c    |   3 +
 drivers/net/ethernet/sfc/tx.c    |   3 +
 include/trace/events/sfc.h       | 160 +++++++++++++++++++++++++++++++++++++++
 7 files changed, 190 insertions(+)
 create mode 100644 include/trace/events/sfc.h

diff --git a/MAINTAINERS b/MAINTAINERS
index b6b29c3..663c6e1 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -7424,6 +7424,7 @@ M:	Ben Hutchings <bhutchings@solarflare.com>
 L:	netdev@vger.kernel.org
 S:	Supported
 F:	drivers/net/ethernet/sfc/
+F:	include/trace/events/sfc.h
 
 SGI GRU DRIVER
 M:	Dimitri Sivanich <sivanich@sgi.com>
diff --git a/drivers/net/ethernet/sfc/Kconfig b/drivers/net/ethernet/sfc/Kconfig
index 8b71525..15cab77 100644
--- a/drivers/net/ethernet/sfc/Kconfig
+++ b/drivers/net/ethernet/sfc/Kconfig
@@ -36,3 +36,17 @@ config SFC_SRIOV
 	  This enables support for the SFC9000 I/O Virtualization
 	  features, allowing accelerated network performance in
 	  virtualized environments.
+config SFC_TRACING
+	bool "Solarflare datapath tracing"
+	depends on SFC
+	select GENERIC_TRACER
+	default n
+	---help---
+	  Enable tracing of packet RX and TX. If you say yes, sfc will
+	  register with the tracing framework and generate an event
+	  for each skb transferred, which can be dumped using ftrace.
+	  You may need to increase the ftrace ringbuffer size. See the
+	  ftrace documentation for more information.
+
+	  When tracing is not enabled, this option still has some
+	  overhead.
diff --git a/drivers/net/ethernet/sfc/efx.c b/drivers/net/ethernet/sfc/efx.c
index 07c9bc4..97c2fda 100644
--- a/drivers/net/ethernet/sfc/efx.c
+++ b/drivers/net/ethernet/sfc/efx.c
@@ -30,6 +30,12 @@
 #include "mcdi.h"
 #include "workarounds.h"
 
+/* sparse doesn't like the tracepoint definitions */
+#ifndef __CHECKER__
+#define CREATE_TRACE_POINTS
+#include <trace/events/sfc.h>
+#endif
+
 /**************************************************************************
  *
  * Type name strings
diff --git a/drivers/net/ethernet/sfc/ptp.c b/drivers/net/ethernet/sfc/ptp.c
index 03acf57..209c428 100644
--- a/drivers/net/ethernet/sfc/ptp.c
+++ b/drivers/net/ethernet/sfc/ptp.c
@@ -48,6 +48,7 @@
 #include "io.h"
 #include "farch_regs.h"
 #include "nic.h"
+#include <trace/events/sfc.h>
 
 /* Maximum number of events expected to make up a PTP event */
 #define	MAX_EVENT_FRAGS			3
@@ -343,6 +344,7 @@ static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
 
 	while ((skb = skb_dequeue(q))) {
 		local_bh_disable();
+		trace_sfc_receive(skb, false);
 		netif_receive_skb(skb);
 		local_bh_enable();
 	}
@@ -722,6 +724,7 @@ static bool efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
 {
 	local_bh_disable();
+	trace_sfc_receive(skb, false);
 	netif_receive_skb(skb);
 	local_bh_enable();
 }
diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index 20c46f3..079b08f 100644
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -25,6 +25,7 @@
 #include "nic.h"
 #include "selftest.h"
 #include "workarounds.h"
+#include <trace/events/sfc.h>
 
 /* Preferred number of descriptors to fill at once */
 #define EFX_RX_PREFERRED_BATCH 8U
@@ -459,6 +460,7 @@ efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 
 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
 
+	trace_sfc_receive(skb, true);
 	gro_result = napi_gro_frags(napi);
 	if (gro_result != GRO_DROP)
 		channel->irq_mod_score += 2;
@@ -623,6 +625,7 @@ static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 			return;
 
 	/* Pass the packet up */
+	trace_sfc_receive(skb, false);
 	netif_receive_skb(skb);
 }
 
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 282692c..bc35f44 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -24,6 +24,7 @@
 #include "nic.h"
 #include "workarounds.h"
 #include "ef10_regs.h"
+#include <trace/events/sfc.h>
 
 #ifdef EFX_USE_PIO
 
@@ -525,6 +526,8 @@ netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
 
 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
 
+	trace_sfc_transmit(skb, net_dev);
+
 	/* PTP "event" packet */
 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
diff --git a/include/trace/events/sfc.h b/include/trace/events/sfc.h
new file mode 100644
index 0000000..a158849
--- /dev/null
+++ b/include/trace/events/sfc.h
@@ -0,0 +1,160 @@
+/****************************************************************************
+ * Driver for Solarflare network controllers and boards
+ * Copyright 2013 Solarflare Communications Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published
+ * by the Free Software Foundation, incorporated herein by reference.
+ */
+
+/* sparse doesn't like the tracepoint definitions.  It will complain
+ * about the tracepoint functions being undefined now, but at least it
+ * won't give up after trying to parse this file.
+ */
+#ifndef __CHECKER__
+
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM sfc
+
+#if !defined(TRACE_EVENTS_SFC_H) || defined(TRACE_HEADER_MULTI_READ)
+#define TRACE_EVENTS_SFC_H
+
+#ifndef CONFIG_SFC_TRACING
+#undef TP_PROTO
+#define TP_PROTO(proto...) proto
+#undef TRACE_EVENT
+#define TRACE_EVENT(name, proto, ...)			\
+	static inline void trace_ ## name(proto) {}
+#else
+#include <linux/tracepoint.h>
+#endif
+
+TRACE_EVENT(sfc_receive,
+
+	TP_PROTO(const struct sk_buff *skb, bool gro),
+
+	TP_ARGS(skb, gro),
+
+	TP_STRUCT__entry(
+		__string(	dev_name,		skb->dev->name	)
+		__field(	unsigned int,		napi_id		)
+		__field(	u16,			queue_mapping	)
+		__field(	const void *,		skbaddr		)
+		__field(	bool,			gro		)
+		__field(	bool,			vlan_tagged	)
+		__field(	u16,			vlan_proto	)
+		__field(	u16,			vlan_tci	)
+		__field(	u16,			protocol	)
+		__field(	u8,			ip_summed	)
+		__field(	u32,			rxhash		)
+		__field(	bool,			l4_rxhash	)
+		__field(	unsigned int,		len		)
+		__field(	unsigned int,		data_len	)
+		__field(	unsigned int,		truesize	)
+		__field(	bool,			mac_header_valid)
+		__field(	int,			mac_header	)
+		__field(	unsigned char,		nr_frags	)
+		__field(	u16,			gso_size	)
+		__field(	u16,			gso_type	)
+	),
+
+	TP_fast_assign(
+		__assign_str(dev_name, skb->dev->name);
+#ifdef CONFIG_NET_LL_RX_POLL
+		__entry->napi_id = skb->napi_id;
+#else
+		__entry->napi_id = 0;
+#endif
+		__entry->queue_mapping = skb->queue_mapping;
+		__entry->skbaddr = skb;
+		__entry->gro = gro;
+		__entry->vlan_tagged = vlan_tx_tag_present(skb);
+		__entry->vlan_proto = ntohs(skb->vlan_proto);
+		__entry->vlan_tci = vlan_tx_tag_get(skb);
+		__entry->protocol = ntohs(skb->protocol);
+		__entry->ip_summed = skb->ip_summed;
+		__entry->rxhash = skb->rxhash;
+		__entry->l4_rxhash = skb->l4_rxhash;
+		__entry->len = skb->len;
+		__entry->data_len = skb->data_len;
+		__entry->truesize = skb->truesize;
+		__entry->mac_header_valid = skb_mac_header_was_set(skb);
+		__entry->mac_header = skb_mac_header(skb) - skb->data;
+		__entry->nr_frags = skb_shinfo(skb)->nr_frags;
+		__entry->gso_size = skb_shinfo(skb)->gso_size;
+		__entry->gso_type = skb_shinfo(skb)->gso_type;
+	),
+
+	TP_printk("dev_name=%s napi_id=%#x queue_mapping=%u skbaddr=%p gro=%d vlan_tagged=%d vlan_proto=0x%04x vlan_tci=0x%04x protocol=0x%04x ip_summed=%d rxhash=0x%08x l4_rxhash=%d len=%u data_len=%u truesize=%u mac_header_valid=%d mac_header=%d nr_frags=%d gso_size=%d gso_type=%#x",
+		  __get_str(dev_name), __entry->napi_id, __entry->queue_mapping,
+		  __entry->skbaddr, __entry->gro, __entry->vlan_tagged,
+		  __entry->vlan_proto, __entry->vlan_tci, __entry->protocol,
+		  __entry->ip_summed, __entry->rxhash, __entry->l4_rxhash,
+		  __entry->len, __entry->data_len, __entry->truesize,
+		  __entry->mac_header_valid, __entry->mac_header,
+		  __entry->nr_frags, __entry->gso_size, __entry->gso_type)
+);
+
+TRACE_EVENT(sfc_transmit,
+
+	TP_PROTO(const struct sk_buff *skb, const struct net_device *net_dev),
+
+	TP_ARGS(skb, net_dev),
+
+	TP_STRUCT__entry(
+		__string(	dev_name,		net_dev->name	)
+		__field(	u16,			queue_mapping	)
+		__field(	const void *,		skbaddr		)
+		__field(	bool,			vlan_tagged	)
+		__field(	u16,			vlan_proto	)
+		__field(	u16,			vlan_tci	)
+		__field(	u16,			protocol	)
+		__field(	u8,			ip_summed	)
+		__field(	unsigned int,		len		)
+		__field(	unsigned int,		data_len	)
+		__field(	int,			network_offset	)
+		__field(	bool,			transport_offset_valid)
+		__field(	int,			transport_offset)
+		__field(	u8,			tx_flags	)
+		__field(	u16,			gso_size	)
+		__field(	u16,			gso_segs	)
+		__field(	u16,			gso_type	)
+	),
+
+	TP_fast_assign(
+		__assign_str(dev_name, net_dev->name);
+		__entry->queue_mapping = skb->queue_mapping;
+		__entry->skbaddr = skb;
+		__entry->vlan_tagged = vlan_tx_tag_present(skb);
+		__entry->vlan_proto = ntohs(skb->vlan_proto);
+		__entry->vlan_tci = vlan_tx_tag_get(skb);
+		__entry->protocol = ntohs(skb->protocol);
+		__entry->ip_summed = skb->ip_summed;
+		__entry->len = skb->len;
+		__entry->data_len = skb->data_len;
+		__entry->network_offset = skb_network_offset(skb);
+		__entry->transport_offset_valid =
+			skb_transport_header_was_set(skb);
+		__entry->transport_offset = skb_transport_offset(skb);
+		__entry->tx_flags = skb_shinfo(skb)->tx_flags;
+		__entry->gso_size = skb_shinfo(skb)->gso_size;
+		__entry->gso_segs = skb_shinfo(skb)->gso_segs;
+		__entry->gso_type = skb_shinfo(skb)->gso_type;
+	),
+
+	TP_printk("dev_name=%s queue_mapping=%u skbaddr=%p vlan_tagged=%d vlan_proto=0x%04x vlan_tci=0x%04x protocol=0x%04x ip_summed=%d len=%u data_len=%u network_offset=%d transport_offset_valid=%d transport_offset=%d tx_flags=%d gso_size=%d gso_segs=%d gso_type=%#x",
+		  __get_str(dev_name), __entry->queue_mapping, __entry->skbaddr,
+		  __entry->vlan_tagged, __entry->vlan_proto, __entry->vlan_tci,
+		  __entry->protocol, __entry->ip_summed, __entry->len, __entry->data_len,
+		  __entry->network_offset, __entry->transport_offset_valid,
+		  __entry->transport_offset, __entry->tx_flags,
+		  __entry->gso_size, __entry->gso_segs, __entry->gso_type)
+);
+
+#endif /* TRACE_EVENTS_SFC_H */
+
+#ifdef CONFIG_SFC_TRACING
+#include <trace/define_trace.h>
+#endif
+
+#endif /* __CHECKER__ */

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: [PATCH net-next 11/11] sfc: Add static tracepoints to datapath

[David Miller]

From: Ben Hutchings <bhutchings@solarflare.com>
Date: Sat, 21 Sep 2013 19:36:32 +0100

> These tracepoints support the driver-specific datapath feature tests
> we're running internally, though they might be useful for other
> purposes.  The skb fields are chosen to cover driver features
> implemented now or likely to be added soon.
> 
> (Includes a bug fix from Edward Cree.)
> 
> Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>

Make this generic, rather than specific to this driver.

You're putting a tracepoint right before calls to the generic core
receive interfaces.  That makes no sense at all.

Re: [PATCH net-next 11/11] sfc: Add static tracepoints to datapath

[Ben Hutchings]

On Mon, 2013-09-23 at 02:08 -0400, David Miller wrote:
> From: Ben Hutchings <bhutchings@solarflare.com>
> Date: Sat, 21 Sep 2013 19:36:32 +0100
> 
> > These tracepoints support the driver-specific datapath feature tests
> > we're running internally, though they might be useful for other
> > purposes.  The skb fields are chosen to cover driver features
> > implemented now or likely to be added soon.
> > 
> > (Includes a bug fix from Edward Cree.)
> > 
> > Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
> 
> Make this generic, rather than specific to this driver.
> 
> You're putting a tracepoint right before calls to the generic core
> receive interfaces.  That makes no sense at all.

I can make these generic, but there are several reasons why I didn't try
to do so:

1. There already *are* static tracepoints there (net_dev_xmit,
netif_receive_skb, netif_rx).  Just not logging enough information for
driver testing - and not in quite the right places to do that
(net_dev_xmit is after ndo_start_xmit so skb may be freed;
netif_receive_skb doesn't know whether GRO was used).  So I assumed that
detailed trace events were not wanted.

2. Generic trace events would be subject to changes that could make them
unsuitable for Solarflare's tests.

3. In Solarflare's test framework the system under tests logs over the
network to the test controller, so that in case of a crash everything
leading up to it will be logged.  If all interfaces on the SUT are
traced then logging trace events over the 'house network' leads to an
infinite loop.  (But obviously this can be avoided with some filtering
on the SUT.)

4. Static tracepoints always have run-time overhead, so these should
probably be controlled by a kconfig option (as in this patch).  Changing
a kconfig option for the core is more disruptive than changing a kconfig
option for a module (reinstall+reboot vs rmmod+insmod).

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: [PATCH net-next 11/11] sfc: Add static tracepoints to datapath

[David Miller]

From: Ben Hutchings <bhutchings@solarflare.com>
Date: Mon, 23 Sep 2013 16:51:07 +0100

> 2. Generic trace events would be subject to changes that could make them
> unsuitable for Solarflare's tests.

This is not an argument for doing things privately in your driver and
duplicating things.

We should make a facility that's useful to Solarflare _AS WELL AS_
everyone else.

> 4. Static tracepoints always have run-time overhead, so these should
> probably be controlled by a kconfig option (as in this patch).

Don't we use static branches and other tricks to make them close to zero
overhead?

Re: [PATCH net-next 11/11] sfc: Add static tracepoints to datapath

[Ben Hutchings]

On Mon, 2013-09-23 at 14:17 -0400, David Miller wrote:
> From: Ben Hutchings <bhutchings@solarflare.com>
> Date: Mon, 23 Sep 2013 16:51:07 +0100
> 
> > 2. Generic trace events would be subject to changes that could make them
> > unsuitable for Solarflare's tests.
> 
> This is not an argument for doing things privately in your driver and
> duplicating things.
> 
> We should make a facility that's useful to Solarflare _AS WELL AS_
> everyone else.

OK, I'll try that.

> > 4. Static tracepoints always have run-time overhead, so these should
> > probably be controlled by a kconfig option (as in this patch).
> 
> Don't we use static branches and other tricks to make them close to zero
> overhead?

You're right, it is a static branch if CONFIG_JUMP_LABEL is enabled.
Some documentation has not been updated to reflect this.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.