@@ -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,
Document exactly which registers and functions have special behaviour, and why races on writes to descriptor pointers are safe. Signed-off-by: Ben Hutchings <bhutchings@solarflare.com> --- drivers/net/sfc/io.h | 98 +++++++++++++++++++++++++------------------------- 1 files changed, 49 insertions(+), 49 deletions(-)