Skip to contentSkip to conversation

pull request: sfc-next-2.6 2010-12-10

PATCH New Ben Hutchings 2 participants 12 messages Dec 10, 2010

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

The following changes since commit defb3519a64141608725e2dac5a5aa9a3c644bae:

net: Abstract away all dst_entry metrics accesses. (2010-12-09 10:46:36 -0800)

are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next-2.6.git for-davem

These changes include the 'TX push' feature I mentioned at netconf. When an TX queue is empty this feature allows the driver to write the next DMA descriptor directly to the NIC, reducing latency by about 0.5 us (dependent on the system).

Ben.

Ben Hutchings (10):
sfc: Reorder struct efx_nic to separate fields by volatility
sfc: Use ACCESS_ONCE when copying efx_tx_queue::read_count
sfc: Expand/correct comments on collector behaviour and function usage
sfc: Remove redundant memory barriers between MMIOs
sfc: Add compile-time checks for correctness of paged register writes
sfc: Remove locking from implementation of efx_writeo_paged()
sfc: Use TX push whenever adding descriptors to an empty queue
sfc: Log start and end of ethtool self-test at INFO level
sfc: Remove filter table IDs from filter functions
sfc: Generalise filter spec initialisation

 drivers/net/sfc/efx.h        |    5 +-
 drivers/net/sfc/ethtool.c    |   99 ++++++++---------
 drivers/net/sfc/filter.c     |  252 ++++++++++++++++++++++++++++++++++--------
 drivers/net/sfc/filter.h     |  149 ++++++-------------------
 drivers/net/sfc/io.h         |  153 +++++++++++++++-----------
 drivers/net/sfc/net_driver.h |   57 +++++++---
 drivers/net/sfc/nic.c        |   42 +++++++-
 drivers/net/sfc/tx.c         |   17 +++-
 8 files changed, 468 insertions(+), 306 deletions(-)
Signature · 3 lines
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#1 lore raw

PATCH 1/10 sfc: Reorder struct efx_nic to separate fields by volatility

Place the regularly updated fields (locks, MAC stats, etc.) on a separate cache-line from fields which are mostly constant. This should reduce cache misses for access to the latter on the data path.

---
 drivers/net/sfc/net_driver.h |   41 +++++++++++++++++++++++------------------
 1 files changed, 23 insertions(+), 18 deletions(-)
drivers/net/sfc/net_driver.h +23 -18
index 0d19fbf..60d6371 100644
@@ -625,10 +625,8 @@ struct efx_filter_state;
  *	Work items do not hold and must not acquire RTNL.
  * @workqueue_name: Name of workqueue
  * @reset_work: Scheduled reset workitem
- * @monitor_work: Hardware monitor workitem
  * @membase_phys: Memory BAR value as physical address
  * @membase: Memory BAR value
- * @biu_lock: BIU (bus interface unit) lock
  * @interrupt_mode: Interrupt mode
  * @irq_rx_adaptive: Adaptive IRQ moderation enabled for RX event queues
  * @irq_rx_moderation: IRQ moderation time for RX event queues
@@ -652,14 +650,9 @@ struct efx_filter_state;
  * @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
- * @last_irq_cpu: Last CPU to handle interrupt.
- *	This register is written with the SMP processor ID whenever an
- *	interrupt is handled.  It is used by efx_nic_test_interrupt()
- *	to verify that an interrupt has occurred.
  * @irq_zero_count: Number of legacy IRQs seen with queue flags == 0
  * @fatal_irq_level: IRQ level (bit number) used for serious errors
  * @mtd_list: List of MTDs attached to the NIC
- * @n_rx_nodesc_drop_cnt: RX no descriptor drop count
  * @nic_data: Hardware dependant state
  * @mac_lock: MAC access lock. Protects @port_enabled, @phy_mode,
  *	@port_inhibited, efx_monitor() and efx_reconfigure_port()
@@ -672,11 +665,7 @@ struct efx_filter_state;
  * @port_initialized: Port initialized?
  * @net_dev: Operating system network device. Consider holding the rtnl lock
  * @rx_checksum_enabled: RX checksumming enabled
- * @mac_stats: MAC statistics. These include all statistics the MACs
- *	can provide.  Generic code converts these into a standard
- *	&struct net_device_stats.
  * @stats_buffer: DMA buffer for statistics
- * @stats_lock: Statistics update lock. Serialises statistics fetches
  * @mac_op: MAC interface
  * @phy_type: PHY type
  * @phy_op: PHY interface
@@ -694,10 +683,23 @@ struct efx_filter_state;
  * @loopback_mode: Loopback status
  * @loopback_modes: Supported loopback mode bitmask
  * @loopback_selftest: Offline self-test private state
+ * @monitor_work: Hardware monitor workitem
+ * @biu_lock: BIU (bus interface unit) lock
+ * @last_irq_cpu: Last CPU to handle interrupt.
+ *	This register is written with the SMP processor ID whenever an
+ *	interrupt is handled.  It is used by efx_nic_test_interrupt()
+ *	to verify that an interrupt has occurred.
+ * @n_rx_nodesc_drop_cnt: RX no descriptor drop count
+ * @mac_stats: MAC statistics. These include all statistics the MACs
+ *	can provide.  Generic code converts these into a standard
+ *	&struct net_device_stats.
+ * @stats_lock: Statistics update lock. Serialises statistics fetches
  *
  * This is stored in the private area of the &struct net_device.
  */
 struct efx_nic {
+	/* The following fields should be written very rarely */
+
 	char name[IFNAMSIZ];
 	struct pci_dev *pci_dev;
 	const struct efx_nic_type *type;
@@ -705,10 +707,9 @@ struct efx_nic {
 	struct workqueue_struct *workqueue;
 	char workqueue_name[16];
 	struct work_struct reset_work;
-	struct delayed_work monitor_work;
 	resource_size_t membase_phys;
 	void __iomem *membase;
-	spinlock_t biu_lock;
+
 	enum efx_int_mode interrupt_mode;
 	bool irq_rx_adaptive;
 	unsigned int irq_rx_moderation;
@@ -735,7 +736,6 @@ struct efx_nic {
 	unsigned long int_error_expire;
 
 	struct efx_buffer irq_status;
-	volatile signed int last_irq_cpu;
 	unsigned irq_zero_count;
 	unsigned fatal_irq_level;
 
@@ -743,8 +743,6 @@ struct efx_nic {
 	struct list_head mtd_list;
 #endif
 
-	unsigned n_rx_nodesc_drop_cnt;
-
 	void *nic_data;
 
 	struct mutex mac_lock;
@@ -756,9 +754,7 @@ struct efx_nic {
 	struct net_device *net_dev;
 	bool rx_checksum_enabled;
 
-	struct efx_mac_stats mac_stats;
 	struct efx_buffer stats_buffer;
-	spinlock_t stats_lock;
 
 	struct efx_mac_operations *mac_op;
 
@@ -784,6 +780,15 @@ struct efx_nic {
 	void *loopback_selftest;
 
 	struct efx_filter_state *filter_state;
+
+	/* The following fields may be written more often */
+
+	struct delayed_work monitor_work ____cacheline_aligned_in_smp;
+	spinlock_t biu_lock;
+	volatile signed int last_irq_cpu;
+	unsigned n_rx_nodesc_drop_cnt;
+	struct efx_mac_stats mac_stats;
+	spinlock_t stats_lock;
 };
 
 static inline int efx_dev_registered(struct efx_nic *efx)
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#2 lore raw

PATCH 2/10 sfc: Use ACCESS_ONCE when copying efx_tx_queue::read_count

In reply to Ben Hutchings
---
 drivers/net/sfc/tx.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)
drivers/net/sfc/tx.c +2 -3
index 03194f7..fef2235 100644
@@ -240,8 +240,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 				 * of read_count. */
 				smp_mb();
 				tx_queue->old_read_count =
-					*(volatile unsigned *)
-					&tx_queue->read_count;
+					ACCESS_ONCE(tx_queue->read_count);
 				fill_level = (tx_queue->insert_count
 					      - tx_queue->old_read_count);
 				q_space = efx->txq_entries - 1 - fill_level;
@@ -764,7 +763,7 @@ static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 			 * stopped from the access of read_count. */
 			smp_mb();
 			tx_queue->old_read_count =
-				*(volatile unsigned *)&tx_queue->read_count;
+				ACCESS_ONCE(tx_queue->read_count);
 			fill_level = (tx_queue->insert_count
 				      - tx_queue->old_read_count);
 			q_space = efx->txq_entries - 1 - fill_level;
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#3 lore raw

PATCH 3/10 sfc: Expand/correct comments on collector behaviour and function usage

In reply to Ben Hutchings

Document exactly which registers and functions have special behaviour, and why races on writes to descriptor pointers are safe.

---
 drivers/net/sfc/io.h |   98 +++++++++++++++++++++++++-------------------------
 1 files changed, 49 insertions(+), 49 deletions(-)
drivers/net/sfc/io.h +49 -49
index 85a99fe..0764e84 100644
@@ -22,28 +22,39 @@
  *
  * Notes on locking strategy:
  *
- * Most NIC registers require 16-byte (or 8-byte, for SRAM) atomic writes
- * which necessitates locking.
- * Under normal operation few writes to NIC registers are made and these
- * registers (EVQ_RPTR_REG, RX_DESC_UPD_REG and TX_DESC_UPD_REG) are special
- * cased to allow 4-byte (hence lockless) accesses.
+ * Most CSRs are 128-bit (oword) and therefore cannot be read or
+ * written atomically.  Access from the host is buffered by the Bus
+ * Interface Unit (BIU).  Whenever the host reads from the lowest
+ * address of such a register, or from the address of a different such
+ * register, the BIU latches the register's value.  Subsequent reads
+ * from higher addresses of the same register will read the latched
+ * value.  Whenever the host writes part of such a register, the BIU
+ * collects the written value and does not write to the underlying
+ * register until all 4 dwords have been written.  A similar buffering
+ * scheme applies to host access to the NIC's 64-bit SRAM.
  *
- * It *is* safe to write to these 4-byte registers in the middle of an
- * access to an 8-byte or 16-byte register.  We therefore use a
- * spinlock to protect accesses to the larger registers, but no locks
- * for the 4-byte registers.
+ * Access to different CSRs and 64-bit SRAM words must be serialised,
+ * since interleaved access can result in lost writes or lost
+ * information from read-to-clear fields.  We use efx_nic::biu_lock
+ * for this.  (We could use separate locks for read and write, but
+ * this is not normally a performance bottleneck.)
  *
- * A write barrier is needed to ensure that DW3 is written after DW0/1/2
- * due to the way the 16byte registers are "collected" in the BIU.
+ * The DMA descriptor pointers (RX_DESC_UPD and TX_DESC_UPD) are
+ * 128-bit but are special-cased in the BIU to avoid the need for
+ * locking in the host:
  *
- * We also lock when carrying out reads, to ensure consistency of the
- * data (made possible since the BIU reads all 128 bits into a cache).
- * Reads are very rare, so this isn't a significant performance
- * impact.  (Most data transferred from NIC to host is DMAed directly
- * into host memory).
- *
- * I/O BAR access uses locks for both reads and writes (but is only provided
- * for testing purposes).
+ * - They are write-only.
+ * - The semantics of writing to these registers are such that
+ *   replacing the low 96 bits with zero does not affect functionality.
+ * - If the host writes to the last dword address of such a register
+ *   (i.e. the high 32 bits) the underlying register will always be
+ *   written.  If the collector does not hold values for the low 96
+ *   bits of the register, they will be written as zero.  Writing to
+ *   the last qword does not have this effect and must not be done.
+ * - If the host writes to the address of any other part of such a
+ *   register while the collector already holds values for some other
+ *   register, the write is discarded and the collector maintains its
+ *   current state.
  */
 
 #if BITS_PER_LONG == 64
@@ -72,7 +83,7 @@ static inline __le32 _efx_readd(struct efx_nic *efx, unsigned int reg)
 	return (__force __le32)__raw_readl(efx->membase + reg);
 }
 
-/* Writes to a normal 16-byte Efx register, locking as appropriate. */
+/* Write a normal 128-bit CSR, locking as appropriate. */
 static inline void efx_writeo(struct efx_nic *efx, efx_oword_t *value,
 			      unsigned int reg)
 {
@@ -98,8 +109,7 @@ static inline void efx_writeo(struct efx_nic *efx, efx_oword_t *value,
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
 }
 
-/* Write an 8-byte NIC SRAM entry through the supplied mapping,
- * locking as appropriate. */
+/* Write 64-bit SRAM through the supplied mapping, locking as appropriate. */
 static inline void efx_sram_writeq(struct efx_nic *efx, void __iomem *membase,
 				   efx_qword_t *value, unsigned int index)
 {
@@ -122,29 +132,19 @@ static inline void efx_sram_writeq(struct efx_nic *efx, void __iomem *membase,
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
 }
 
-/* Write dword to NIC register that allows partial writes
- *
- * Some registers (EVQ_RPTR_REG, RX_DESC_UPD_REG and
- * TX_DESC_UPD_REG) can be written to as a single dword.  This allows
- * for lockless writes.
- */
+/* Write a 32-bit CSR or the last dword of a special 128-bit CSR */
 static inline void efx_writed(struct efx_nic *efx, efx_dword_t *value,
 			      unsigned int reg)
 {
 	netif_vdbg(efx, hw, efx->net_dev,
-		   "writing partial register %x with "EFX_DWORD_FMT"\n",
+		   "writing register %x with "EFX_DWORD_FMT"\n",
 		   reg, EFX_DWORD_VAL(*value));
 
 	/* No lock required */
 	_efx_writed(efx, value->u32[0], reg);
 }
 
-/* Read from a NIC register
- *
- * This reads an entire 16-byte register in one go, locking as
- * appropriate.  It is essential to read the first dword first, as this
- * prompts the NIC to load the current value into the shadow register.
- */
+/* Read a 128-bit CSR, locking as appropriate. */
 static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 			     unsigned int reg)
 {
@@ -163,8 +163,7 @@ static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 		   EFX_OWORD_VAL(*value));
 }
 
-/* Read an 8-byte SRAM entry through supplied mapping,
- * locking as appropriate. */
+/* Read 64-bit SRAM through the supplied mapping, locking as appropriate. */
 static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 				  efx_qword_t *value, unsigned int index)
 {
@@ -186,7 +185,7 @@ static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 		   addr, EFX_QWORD_VAL(*value));
 }
 
-/* Read dword from register that allows partial writes (sic) */
+/* Read a 32-bit CSR or SRAM */
 static inline void efx_readd(struct efx_nic *efx, efx_dword_t *value,
 				unsigned int reg)
 {
@@ -196,28 +195,28 @@ static inline void efx_readd(struct efx_nic *efx, efx_dword_t *value,
 		   reg, EFX_DWORD_VAL(*value));
 }
 
-/* Write to a register forming part of a table */
+/* Write a 128-bit CSR forming part of a table */
 static inline void efx_writeo_table(struct efx_nic *efx, efx_oword_t *value,
 				      unsigned int reg, unsigned int index)
 {
 	efx_writeo(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Read to a register forming part of a table */
+/* Read a 128-bit CSR forming part of a table */
 static inline void efx_reado_table(struct efx_nic *efx, efx_oword_t *value,
 				     unsigned int reg, unsigned int index)
 {
 	efx_reado(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Write to a dword register forming part of a table */
+/* Write a 32-bit CSR forming part of a table, or 32-bit SRAM */
 static inline void efx_writed_table(struct efx_nic *efx, efx_dword_t *value,
 				       unsigned int reg, unsigned int index)
 {
 	efx_writed(efx, value, reg + index * sizeof(efx_oword_t));
 }
 
-/* Read from a dword register forming part of a table */
+/* Read a 32-bit CSR forming part of a table, or 32-bit SRAM */
 static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 				   unsigned int reg, unsigned int index)
 {
@@ -231,25 +230,26 @@ static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 #define EFX_PAGED_REG(page, reg) \
 	((page) * EFX_PAGE_BLOCK_SIZE + (reg))
 
-/* As for efx_writeo(), but for a page-mapped register. */
+/* Write the whole of RX_DESC_UPD or TX_DESC_UPD */
 static inline void efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
 				   unsigned int reg, unsigned int page)
 {
 	efx_writeo(efx, value, EFX_PAGED_REG(page, reg));
 }
 
-/* As for efx_writed(), but for a page-mapped register. */
+/* Write a page-mapped 32-bit CSR (EVQ_RPTR or the high bits of
+ * RX_DESC_UPD or TX_DESC_UPD)
+ */
 static inline void efx_writed_page(struct efx_nic *efx, efx_dword_t *value,
 				   unsigned int reg, unsigned int page)
 {
 	efx_writed(efx, value, EFX_PAGED_REG(page, reg));
 }
 
-/* Write dword to page-mapped register with an extra lock.
- *
- * As for efx_writed_page(), but for a register that suffers from
- * SFC bug 3181. Take out a lock so the BIU collector cannot be
- * confused. */
+/* Write TIMER_COMMAND.  This is a page-mapped 32-bit CSR, but a bug
+ * in the BIU means that writes to TIMER_COMMAND[0] invalidate the
+ * collector register.
+ */
 static inline void efx_writed_page_locked(struct efx_nic *efx,
 					  efx_dword_t *value,
 					  unsigned int reg,
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#4 lore raw

PATCH 4/10 sfc: Remove redundant memory barriers between MMIOs

In reply to Ben Hutchings
---
 drivers/net/sfc/io.h |    5 -----
 1 files changed, 0 insertions(+), 5 deletions(-)
drivers/net/sfc/io.h +0 -5
index 0764e84..896b31e 100644
@@ -96,13 +96,11 @@ static inline void efx_writeo(struct efx_nic *efx, efx_oword_t *value,
 	spin_lock_irqsave(&efx->biu_lock, flags);
 #ifdef EFX_USE_QWORD_IO
 	_efx_writeq(efx, value->u64[0], reg + 0);
-	wmb();
 	_efx_writeq(efx, value->u64[1], reg + 8);
 #else
 	_efx_writed(efx, value->u32[0], reg + 0);
 	_efx_writed(efx, value->u32[1], reg + 4);
 	_efx_writed(efx, value->u32[2], reg + 8);
-	wmb();
 	_efx_writed(efx, value->u32[3], reg + 12);
 #endif
 	mmiowb();
@@ -125,7 +123,6 @@ static inline void efx_sram_writeq(struct efx_nic *efx, void __iomem *membase,
 	__raw_writeq((__force u64)value->u64[0], membase + addr);
 #else
 	__raw_writel((__force u32)value->u32[0], membase + addr);
-	wmb();
 	__raw_writel((__force u32)value->u32[1], membase + addr + 4);
 #endif
 	mmiowb();
@@ -152,7 +149,6 @@ static inline void efx_reado(struct efx_nic *efx, efx_oword_t *value,
 
 	spin_lock_irqsave(&efx->biu_lock, flags);
 	value->u32[0] = _efx_readd(efx, reg + 0);
-	rmb();
 	value->u32[1] = _efx_readd(efx, reg + 4);
 	value->u32[2] = _efx_readd(efx, reg + 8);
 	value->u32[3] = _efx_readd(efx, reg + 12);
@@ -175,7 +171,6 @@ static inline void efx_sram_readq(struct efx_nic *efx, void __iomem *membase,
 	value->u64[0] = (__force __le64)__raw_readq(membase + addr);
 #else
 	value->u32[0] = (__force __le32)__raw_readl(membase + addr);
-	rmb();
 	value->u32[1] = (__force __le32)__raw_readl(membase + addr + 4);
 #endif
 	spin_unlock_irqrestore(&efx->biu_lock, flags);
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#5 lore raw

PATCH 5/10 sfc: Add compile-time checks for correctness of paged register writes

In reply to Ben Hutchings
---
 drivers/net/sfc/io.h |   31 +++++++++++++++++++++++--------
 1 files changed, 23 insertions(+), 8 deletions(-)
drivers/net/sfc/io.h +23 -8
index 896b31e..7f3adf2 100644
@@ -226,29 +226,40 @@ static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 	((page) * EFX_PAGE_BLOCK_SIZE + (reg))
 
 /* Write the whole of RX_DESC_UPD or TX_DESC_UPD */
-static inline void efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
-				   unsigned int reg, unsigned int page)
+static inline void _efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
+				    unsigned int reg, unsigned int page)
 {
 	efx_writeo(efx, value, EFX_PAGED_REG(page, reg));
 }
+#define efx_writeo_page(efx, value, reg, page)				\
+	_efx_writeo_page(efx, value,					\
+			 reg +						\
+			 BUILD_BUG_ON_ZERO((reg) != 0x830 && (reg) != 0xa10), \
+			 page)
 
 /* Write a page-mapped 32-bit CSR (EVQ_RPTR or the high bits of
  * RX_DESC_UPD or TX_DESC_UPD)
  */
-static inline void efx_writed_page(struct efx_nic *efx, efx_dword_t *value,
-				   unsigned int reg, unsigned int page)
+static inline void _efx_writed_page(struct efx_nic *efx, efx_dword_t *value,
+				    unsigned int reg, unsigned int page)
 {
 	efx_writed(efx, value, EFX_PAGED_REG(page, reg));
 }
+#define efx_writed_page(efx, value, reg, page)				\
+	_efx_writed_page(efx, value,					\
+			 reg +						\
+			 BUILD_BUG_ON_ZERO((reg) != 0x400 && (reg) != 0x83c \
+					   && (reg) != 0xa1c),		\
+			 page)
 
 /* Write TIMER_COMMAND.  This is a page-mapped 32-bit CSR, but a bug
  * in the BIU means that writes to TIMER_COMMAND[0] invalidate the
  * collector register.
  */
-static inline void efx_writed_page_locked(struct efx_nic *efx,
-					  efx_dword_t *value,
-					  unsigned int reg,
-					  unsigned int page)
+static inline void _efx_writed_page_locked(struct efx_nic *efx,
+					   efx_dword_t *value,
+					   unsigned int reg,
+					   unsigned int page)
 {
 	unsigned long flags __attribute__ ((unused));
 
@@ -260,5 +271,9 @@ static inline void efx_writed_page_locked(struct efx_nic *efx,
 		efx_writed(efx, value, EFX_PAGED_REG(page, reg));
 	}
 }
+#define efx_writed_page_locked(efx, value, reg, page)			\
+	_efx_writed_page_locked(efx, value,				\
+				reg + BUILD_BUG_ON_ZERO((reg) != 0x420), \
+				page)
 
 #endif /* EFX_IO_H */
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#6 lore raw

PATCH 6/10 sfc: Remove locking from implementation of efx_writeo_paged()

In reply to Ben Hutchings

It is not necessary to serialise writes to the paged 128-bit registers. However, if we don't then we must always write the last dword separately, not as part of a qword write.

---
 drivers/net/sfc/io.h |   15 ++++++++++++++-
 1 files changed, 14 insertions(+), 1 deletions(-)
drivers/net/sfc/io.h +14 -1
index 7f3adf2..6da4ae2 100644
@@ -229,7 +229,20 @@ static inline void efx_readd_table(struct efx_nic *efx, efx_dword_t *value,
 static inline void _efx_writeo_page(struct efx_nic *efx, efx_oword_t *value,
 				    unsigned int reg, unsigned int page)
 {
-	efx_writeo(efx, value, EFX_PAGED_REG(page, reg));
+	reg = EFX_PAGED_REG(page, reg);
+
+	netif_vdbg(efx, hw, efx->net_dev,
+		   "writing register %x with " EFX_OWORD_FMT "\n", reg,
+		   EFX_OWORD_VAL(*value));
+
+#ifdef EFX_USE_QWORD_IO
+	_efx_writeq(efx, value->u64[0], reg + 0);
+#else
+	_efx_writed(efx, value->u32[0], reg + 0);
+	_efx_writed(efx, value->u32[1], reg + 4);
+#endif
+	_efx_writed(efx, value->u32[2], reg + 8);
+	_efx_writed(efx, value->u32[3], reg + 12);
 }
 #define efx_writeo_page(efx, value, reg, page)				\
 	_efx_writeo_page(efx, value,					\
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#7 lore raw

PATCH 7/10 sfc: Use TX push whenever adding descriptors to an empty queue

In reply to Ben Hutchings

Whenever we add DMA descriptors to a TX ring and update the ring pointer, the TX DMA engine must first read the new DMA descriptors and then start reading packet data. However, all released Solarflare 10G controllers have a 'TX push' feature that allows us to reduce latency by writing the first new DMA descriptor along with the pointer update. This is only useful when the queue is empty. The hardware should ignore the pushed descriptor if the queue is not empty, but this check is buggy, so we must do it in software.

In order to tell whether a TX queue is empty, we need to compare the previous transmission count (write_count) and completion count (read_count). However, if we do that every time we update the ring pointer then read_count may ping-pong between the caches of two CPUs running the transmission and completion paths for the queue. Therefore, we split the check for an empty queue between the completion path and the transmission path:

- Add an empty_read_count field representing a point at which the completion path saw the TX queue as empty.
- Add an old_write_count field for use on the completion path.
- On the completion path, whenever read_count reaches or passes old_write_count the TX queue may be empty. We then read write_count, set empty_read_count if read_count == write_count, and update old_write_count.
- On the transmission path, we read empty_read_count. If it's set, we compare it with the value of write_count before the current set of descriptors was added. If they match, the queue really is empty and we can use TX push.

---
 drivers/net/sfc/net_driver.h |   16 ++++++++++++++++
 drivers/net/sfc/nic.c        |   42 ++++++++++++++++++++++++++++++++++++++++--
 drivers/net/sfc/tx.c         |   12 ++++++++++++
 3 files changed, 68 insertions(+), 2 deletions(-)
drivers/net/sfc/net_driver.h +16 -0
index 60d6371..270e217 100644
@@ -142,6 +142,12 @@ struct efx_tx_buffer {
  * @flushed: Used when handling queue flushing
  * @read_count: Current read pointer.
  *	This is the number of buffers that have been removed from both rings.
+ * @old_write_count: The value of @write_count when last checked.
+ *	This is here for performance reasons.  The xmit path will
+ *	only get the up-to-date value of @write_count if this
+ *	variable indicates that the queue is empty.  This is to
+ *	avoid cache-line ping-pong between the xmit path and the
+ *	completion path.
  * @stopped: Stopped count.
  *	Set if this TX queue is currently stopping its port.
  * @insert_count: Current insert pointer
@@ -163,6 +169,10 @@ struct efx_tx_buffer {
  * @tso_long_headers: Number of packets with headers too long for standard
  *	blocks
  * @tso_packets: Number of packets via the TSO xmit path
+ * @pushes: Number of times the TX push 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.
  */
 struct efx_tx_queue {
 	/* Members which don't change on the fast path */
@@ -177,6 +187,7 @@ struct efx_tx_queue {
 
 	/* Members used mainly on the completion path */
 	unsigned int read_count ____cacheline_aligned_in_smp;
+	unsigned int old_write_count;
 	int stopped;
 
 	/* Members used only on the xmit path */
@@ -187,6 +198,11 @@ struct efx_tx_queue {
 	unsigned int tso_bursts;
 	unsigned int tso_long_headers;
 	unsigned int tso_packets;
+	unsigned int pushes;
+
+	/* Members shared between paths and sometimes updated */
+	unsigned int empty_read_count ____cacheline_aligned_in_smp;
+#define EFX_EMPTY_COUNT_VALID 0x80000000
 };
 
 /**
drivers/net/sfc/nic.c +40 -2

66 lines not on this page · raw message

drivers/net/sfc/tx.c +12 -0
index fef2235..bdb92b4 100644
@@ -428,6 +428,16 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 			__netif_tx_unlock(queue);
 		}
 	}
+
+	/* Check whether the hardware queue is now empty */
+	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
+		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
+		if (tx_queue->read_count == tx_queue->old_write_count) {
+			smp_mb();
+			tx_queue->empty_read_count =
+				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
+		}
+	}
 }
 
 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
@@ -473,8 +483,10 @@ void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 
 	tx_queue->insert_count = 0;
 	tx_queue->write_count = 0;
+	tx_queue->old_write_count = 0;
 	tx_queue->read_count = 0;
 	tx_queue->old_read_count = 0;
+	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
 	BUG_ON(tx_queue->stopped);
 
 	/* Set up TX descriptor ring */
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#8 lore raw

PATCH 8/10 sfc: Log start and end of ethtool self-test at INFO level

In reply to Ben Hutchings

Add message at start of self-test and increase log level of message at end of self-test, so that any other messages produced during the test are clearly associated with it.

---
 drivers/net/sfc/ethtool.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)
drivers/net/sfc/ethtool.c +6 -3
index aae756b..d51a6b1 100644
@@ -582,6 +582,9 @@ static void efx_ethtool_self_test(struct net_device *net_dev,
 		goto fail1;
 	}
 
+	netif_info(efx, drv, efx->net_dev, "starting %sline testing\n",
+		   (test->flags & ETH_TEST_FL_OFFLINE) ? "off" : "on");
+
 	/* We need rx buffers and interrupts. */
 	already_up = (efx->net_dev->flags & IFF_UP);
 	if (!already_up) {
@@ -600,9 +603,9 @@ static void efx_ethtool_self_test(struct net_device *net_dev,
 	if (!already_up)
 		dev_close(efx->net_dev);
 
-	netif_dbg(efx, drv, efx->net_dev, "%s %sline self-tests\n",
-		  rc == 0 ? "passed" : "failed",
-		  (test->flags & ETH_TEST_FL_OFFLINE) ? "off" : "on");
+	netif_info(efx, drv, efx->net_dev, "%s %sline self-tests\n",
+		   rc == 0 ? "passed" : "failed",
+		   (test->flags & ETH_TEST_FL_OFFLINE) ? "off" : "on");
 
  fail2:
  fail1:
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#9 lore raw

PATCH 9/10 sfc: Remove filter table IDs from filter functions

In reply to Ben Hutchings

The separation between filter tables is largely an internal detail and it may be removed in future hardware. To prepare for that:

- Merge table ID with filter index to make an opaque filter ID
- Wrap efx_filter_table_clear() with a function that clears filters from both RX tables, which is all that the current caller requires

---
 drivers/net/sfc/efx.h     |    5 ++---
 drivers/net/sfc/ethtool.c |    8 ++------
 drivers/net/sfc/filter.c  |   40 ++++++++++++++++++++++++++++++----------
 drivers/net/sfc/filter.h  |    6 ------
 4 files changed, 34 insertions(+), 25 deletions(-)
drivers/net/sfc/efx.h +2 -3
index 10a1bf4..003fdb3 100644
@@ -74,9 +74,8 @@ extern int efx_filter_insert_filter(struct efx_nic *efx,
 				    bool replace);
 extern int efx_filter_remove_filter(struct efx_nic *efx,
 				    struct efx_filter_spec *spec);
-extern void efx_filter_table_clear(struct efx_nic *efx,
-				   enum efx_filter_table_id table_id,
-				   enum efx_filter_priority priority);
+extern void efx_filter_clear_rx(struct efx_nic *efx,
+				enum efx_filter_priority priority);
 
 /* Channels */
 extern void efx_process_channel_now(struct efx_channel *channel);
drivers/net/sfc/ethtool.c +2 -6
index d51a6b1..0f46c1a 100644
@@ -558,12 +558,8 @@ static int efx_ethtool_set_flags(struct net_device *net_dev, u32 data)
 	if (rc)
 		return rc;
 
-	if (!(data & ETH_FLAG_NTUPLE)) {
-		efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_IP,
-				       EFX_FILTER_PRI_MANUAL);
-		efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_MAC,
-				       EFX_FILTER_PRI_MANUAL);
-	}
+	if (!(data & ETH_FLAG_NTUPLE))
+		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
 
 	return 0;
 }
drivers/net/sfc/filter.c +30 -10
index 44500b5..e96e6e8 100644
@@ -26,6 +26,12 @@
  */
 #define FILTER_CTL_SRCH_MAX 200
 
+enum efx_filter_table_id {
+	EFX_FILTER_TABLE_RX_IP = 0,
+	EFX_FILTER_TABLE_RX_MAC,
+	EFX_FILTER_TABLE_COUNT,
+};
+
 struct efx_filter_table {
 	u32		offset;		/* address of table relative to BAR */
 	unsigned	size;		/* number of entries */
@@ -206,6 +212,14 @@ found:
 	return filter_idx;
 }
 
+/* Construct/deconstruct external filter IDs */
+
+static inline int
+efx_filter_make_id(enum efx_filter_table_id table_id, unsigned index)
+{
+	return table_id << 16 | index;
+}
+
 /**
  * efx_filter_insert_filter - add or replace a filter
  * @efx: NIC in which to insert the filter
@@ -213,7 +227,7 @@ found:
  * @replace: Flag for whether the specified filter may replace a filter
  *	with an identical match expression and equal or lower priority
  *
- * On success, return the filter index within its table.
+ * On success, return the filter ID.
  * On failure, return a negative error code.
  */
 int efx_filter_insert_filter(struct efx_nic *efx, struct efx_filter_spec *spec,
@@ -273,6 +287,7 @@ int efx_filter_insert_filter(struct efx_nic *efx, struct efx_filter_spec *spec,
 	netif_vdbg(efx, hw, efx->net_dev,
 		   "%s: filter type %d index %d rxq %u set",
 		   __func__, spec->type, filter_idx, spec->dmaq_id);
+	rc = efx_filter_make_id(table_id, filter_idx);
 
 out:
 	spin_unlock_bh(&state->lock);
@@ -340,15 +355,9 @@ out:
 	return rc;
 }
 
-/**
- * efx_filter_table_clear - remove filters from a table by priority
- * @efx: NIC from which to remove the filters
- * @table_id: Table from which to remove the filters
- * @priority: Maximum priority to remove
- */
-void efx_filter_table_clear(struct efx_nic *efx,
-			    enum efx_filter_table_id table_id,
-			    enum efx_filter_priority priority)
+static void efx_filter_table_clear(struct efx_nic *efx,
+				   enum efx_filter_table_id table_id,
+				   enum efx_filter_priority priority)
 {
 	struct efx_filter_state *state = efx->filter_state;
 	struct efx_filter_table *table = &state->table[table_id];
@@ -365,6 +374,17 @@ void efx_filter_table_clear(struct efx_nic *efx,
 	spin_unlock_bh(&state->lock);
 }
 
+/**
+ * efx_filter_clear_rx - remove RX filters by priority
+ * @efx: NIC from which to remove the filters
+ * @priority: Maximum priority to remove
+ */
+void efx_filter_clear_rx(struct efx_nic *efx, enum efx_filter_priority priority)
+{
+	efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_IP, priority);
+	efx_filter_table_clear(efx, EFX_FILTER_TABLE_RX_MAC, priority);
+}
+
 /* Restore filter stater after reset */
 void efx_restore_filters(struct efx_nic *efx)
 {
drivers/net/sfc/filter.h +0 -6
index a53319d..d11e4aa 100644
@@ -12,12 +12,6 @@
 
 #include <linux/types.h>
 
-enum efx_filter_table_id {
-	EFX_FILTER_TABLE_RX_IP = 0,
-	EFX_FILTER_TABLE_RX_MAC,
-	EFX_FILTER_TABLE_COUNT,
-};
-
 /**
  * enum efx_filter_type - type of hardware filter
  * @EFX_FILTER_RX_TCP_FULL: RX, matching TCP/IPv4 4-tuple
Signature · 8 lines
1.7.3.2



-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#10 lore raw

PATCH 10/10 sfc: Generalise filter spec initialisation

In reply to Ben Hutchings

Move search_depth arrays into per-table state.

Define initialisation function efx_filter_init_rx() which sets everything apart from the match fields.

Define efx_filter_set_{ipv4_local,ipv4_full,eth_local}() to set the match fields. This allows some simplification of callers and later support for additional protocols and more flexible matching using multiple calls to these functions.

---
 drivers/net/sfc/ethtool.c |   82 ++++++++----------
 drivers/net/sfc/filter.c  |  214 +++++++++++++++++++++++++++++++++++++--------
 drivers/net/sfc/filter.h  |  143 ++++++++----------------------
 3 files changed, 248 insertions(+), 191 deletions(-)
drivers/net/sfc/ethtool.c +36 -46
index 0f46c1a..5e50e57 100644
@@ -11,6 +11,7 @@
 #include <linux/netdevice.h>
 #include <linux/ethtool.h>
 #include <linux/rtnetlink.h>
+#include <linux/in.h>
 #include "net_driver.h"
 #include "workarounds.h"
 #include "selftest.h"
@@ -920,6 +921,7 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 	struct ethhdr *mac_entry = &ntuple->fs.h_u.ether_spec;
 	struct ethhdr *mac_mask = &ntuple->fs.m_u.ether_spec;
 	struct efx_filter_spec filter;
+	int rc;
 
 	/* Range-check action */
 	if (ntuple->fs.action < ETHTOOL_RXNTUPLE_ACTION_CLEAR ||
@@ -929,9 +931,16 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 	if (~ntuple->fs.data_mask)
 		return -EINVAL;
 
+	efx_filter_init_rx(&filter, EFX_FILTER_PRI_MANUAL, 0,
+			   (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_DROP) ?
+			   0xfff : ntuple->fs.action);
+
 	switch (ntuple->fs.flow_type) {
 	case TCP_V4_FLOW:
-	case UDP_V4_FLOW:
+	case UDP_V4_FLOW: {
+		u8 proto = (ntuple->fs.flow_type == TCP_V4_FLOW ?
+			    IPPROTO_TCP : IPPROTO_UDP);
+
 		/* Must match all of destination, */
 		if (ip_mask->ip4dst | ip_mask->pdst)
 			return -EINVAL;
@@ -943,7 +952,22 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 		/* and nothing else */
 		if ((u8)~ip_mask->tos | (u16)~ntuple->fs.vlan_tag_mask)
 			return -EINVAL;
+
+		if (!ip_mask->ip4src)
+			rc = efx_filter_set_ipv4_full(&filter, proto,
+						      ip_entry->ip4dst,
+						      ip_entry->pdst,
+						      ip_entry->ip4src,
+						      ip_entry->psrc);
+		else
+			rc = efx_filter_set_ipv4_local(&filter, proto,
+						       ip_entry->ip4dst,
+						       ip_entry->pdst);
+		if (rc)
+			return rc;
 		break;
+	}
+
 	case ETHER_FLOW:
 		/* Must match all of destination, */
 		if (!is_zero_ether_addr(mac_mask->h_dest))
@@ -956,58 +980,24 @@ static int efx_ethtool_set_rx_ntuple(struct net_device *net_dev,
 		if (!is_broadcast_ether_addr(mac_mask->h_source) ||
 		    mac_mask->h_proto != htons(0xffff))
 			return -EINVAL;
+
+		rc = efx_filter_set_eth_local(
+			&filter,
+			(ntuple->fs.vlan_tag_mask == 0xf000) ?
+			ntuple->fs.vlan_tag : EFX_FILTER_VID_UNSPEC,
+			mac_entry->h_dest);
+		if (rc)
+			return rc;
 		break;
+
 	default:
 		return -EINVAL;
 	}
 
-	filter.priority = EFX_FILTER_PRI_MANUAL;
-	filter.flags = 0;
-
-	switch (ntuple->fs.flow_type) {
-	case TCP_V4_FLOW:
-		if (!ip_mask->ip4src)
-			efx_filter_set_rx_tcp_full(&filter,
-						   htonl(ip_entry->ip4src),
-						   htons(ip_entry->psrc),
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		else
-			efx_filter_set_rx_tcp_wild(&filter,
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		break;
-	case UDP_V4_FLOW:
-		if (!ip_mask->ip4src)
-			efx_filter_set_rx_udp_full(&filter,
-						   htonl(ip_entry->ip4src),
-						   htons(ip_entry->psrc),
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		else
-			efx_filter_set_rx_udp_wild(&filter,
-						   htonl(ip_entry->ip4dst),
-						   htons(ip_entry->pdst));
-		break;
-	case ETHER_FLOW:
-		if (ntuple->fs.vlan_tag_mask == 0xf000)
-			efx_filter_set_rx_mac_full(&filter,
-						   ntuple->fs.vlan_tag & 0xfff,
-						   mac_entry->h_dest);
-		else
-			efx_filter_set_rx_mac_wild(&filter, mac_entry->h_dest);
-		break;
-	}
-
-	if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_CLEAR) {
+	if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_CLEAR)
 		return efx_filter_remove_filter(efx, &filter);
-	} else {
-		if (ntuple->fs.action == ETHTOOL_RXNTUPLE_ACTION_DROP)
-			filter.dmaq_id = 0xfff;
-		else
-			filter.dmaq_id = ntuple->fs.action;
+	else
 		return efx_filter_insert_filter(efx, &filter, true);
-	}
 }
 
 static int efx_ethtool_get_rxfh_indir(struct net_device *net_dev,
drivers/net/sfc/filter.c +176 -38

354 lines not on this page · raw message

drivers/net/sfc/filter.h +36 -107

171 lines not on this page · raw message

Signature · 7 lines
1.7.3.2


-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
maintainer
#11 lore raw
In reply to Ben Hutchings

From: Ben Hutchings <[email protected]>
Date: Fri, 10 Dec 2010 17:35:27 +0000

Quoted 6 lines

The following changes since commit defb3519a64141608725e2dac5a5aa9a3c644bae:

net: Abstract away all dst_entry metrics accesses. (2010-12-09 10:46:36 -0800)

are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next-2.6.git for-davem

Pulled, thanks Ben.

Indeed TX push looks intersting.