MIPS: Octeon: Update L2 Cache code for CN63XX

The CN63XX has a different L2 cache architecture.  Update the helper
functions to reflect this.

Some joining of split lines was also done to improve readability, as
well as reformatting of comments.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
Patchwork: http://patchwork.linux-mips.org/patch/1663/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

arch/mips/include/asm/octeon/cvmx-asm.h

@@ -114,6 +114,17 @@
 #define CVMX_DCACHE_INVALIDATE \
 	{ CVMX_SYNC; asm volatile ("cache 9, 0($0)" : : ); }
 
+#define CVMX_CACHE(op, address, offset)					\
+	asm volatile ("cache " CVMX_TMP_STR(op) ", " CVMX_TMP_STR(offset) "(%[rbase])" \
+		: : [rbase] "d" (address) )
+/* fetch and lock the state. */
+#define CVMX_CACHE_LCKL2(address, offset) CVMX_CACHE(31, address, offset)
+/* unlock the state. */
+#define CVMX_CACHE_WBIL2(address, offset) CVMX_CACHE(23, address, offset)
+/* invalidate the cache block and clear the USED bits for the block */
+#define CVMX_CACHE_WBIL2I(address, offset) CVMX_CACHE(3, address, offset)
+/* load virtual tag and data for the L2 cache block into L2C_TAD0_TAG register */
+#define CVMX_CACHE_LTGL2I(address, offset) CVMX_CACHE(7, address, offset)
 
 #define CVMX_POP(result, input) \
 	asm ("pop %[rd],%[rs]" : [rd] "=d" (result) : [rs] "d" (input))

arch/mips/include/asm/octeon/cvmx-l2c.h

@@ -4,7 +4,7 @@
  * Contact: support@caviumnetworks.com
  * This file is part of the OCTEON SDK
  *
- * Copyright (c) 2003-2008 Cavium Networks
+ * Copyright (c) 2003-2010 Cavium Networks
  *
  * This file is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, Version 2, as
@@ -26,7 +26,6 @@
  ***********************license end**************************************/
 
 /*
- *
  * Interface to the Level 2 Cache (L2C) control, measurement, and debugging
  * facilities.
  */
@@ -34,93 +33,126 @@
 #ifndef __CVMX_L2C_H__
 #define __CVMX_L2C_H__
 
-/* Deprecated macro, use function */
-#define CVMX_L2_ASSOC     cvmx_l2c_get_num_assoc()
-
-/* Deprecated macro, use function */
-#define CVMX_L2_SET_BITS  cvmx_l2c_get_set_bits()
+#define CVMX_L2_ASSOC     cvmx_l2c_get_num_assoc()   /* Deprecated macro, use function */
+#define CVMX_L2_SET_BITS  cvmx_l2c_get_set_bits()    /* Deprecated macro, use function */
+#define CVMX_L2_SETS      cvmx_l2c_get_num_sets()    /* Deprecated macro, use function */
 
-/* Deprecated macro, use function */
-#define CVMX_L2_SETS      cvmx_l2c_get_num_sets()
 
 #define CVMX_L2C_IDX_ADDR_SHIFT 7  /* based on 128 byte cache line size */
 #define CVMX_L2C_IDX_MASK       (cvmx_l2c_get_num_sets() - 1)
 
 /* Defines for index aliasing computations */
-#define CVMX_L2C_TAG_ADDR_ALIAS_SHIFT \
-	(CVMX_L2C_IDX_ADDR_SHIFT + cvmx_l2c_get_set_bits())
+#define CVMX_L2C_TAG_ADDR_ALIAS_SHIFT (CVMX_L2C_IDX_ADDR_SHIFT + cvmx_l2c_get_set_bits())
+#define CVMX_L2C_ALIAS_MASK (CVMX_L2C_IDX_MASK << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT)
+#define CVMX_L2C_MEMBANK_SELECT_SIZE  4096
 
-#define CVMX_L2C_ALIAS_MASK \
-	(CVMX_L2C_IDX_MASK << CVMX_L2C_TAG_ADDR_ALIAS_SHIFT)
+/* Defines for Virtualizations, valid only from Octeon II onwards. */
+#define CVMX_L2C_VRT_MAX_VIRTID_ALLOWED ((OCTEON_IS_MODEL(OCTEON_CN63XX)) ? 64 : 0)
+#define CVMX_L2C_VRT_MAX_MEMSZ_ALLOWED ((OCTEON_IS_MODEL(OCTEON_CN63XX)) ? 32 : 0)
 
 union cvmx_l2c_tag {
 	uint64_t u64;
 	struct {
 		uint64_t reserved:28;
-		uint64_t V:1;	/* Line valid */
-		uint64_t D:1;	/* Line dirty */
-		uint64_t L:1;	/* Line locked */
-		uint64_t U:1;	/* Use, LRU eviction */
+		uint64_t V:1;		/* Line valid */
+		uint64_t D:1;		/* Line dirty */
+		uint64_t L:1;		/* Line locked */
+		uint64_t U:1;		/* Use, LRU eviction */
 		uint64_t addr:32;	/* Phys mem (not all bits valid) */
 	} s;
 };
 
+/* Number of L2C Tag-and-data sections (TADs) that are connected to LMC. */
+#define CVMX_L2C_TADS  1
+
   /* L2C Performance Counter events. */
 enum cvmx_l2c_event {
-	CVMX_L2C_EVENT_CYCLES = 0,
-	CVMX_L2C_EVENT_INSTRUCTION_MISS = 1,
-	CVMX_L2C_EVENT_INSTRUCTION_HIT = 2,
-	CVMX_L2C_EVENT_DATA_MISS = 3,
-	CVMX_L2C_EVENT_DATA_HIT = 4,
-	CVMX_L2C_EVENT_MISS = 5,
-	CVMX_L2C_EVENT_HIT = 6,
-	CVMX_L2C_EVENT_VICTIM_HIT = 7,
-	CVMX_L2C_EVENT_INDEX_CONFLICT = 8,
-	CVMX_L2C_EVENT_TAG_PROBE = 9,
-	CVMX_L2C_EVENT_TAG_UPDATE = 10,
-	CVMX_L2C_EVENT_TAG_COMPLETE = 11,
-	CVMX_L2C_EVENT_TAG_DIRTY = 12,
-	CVMX_L2C_EVENT_DATA_STORE_NOP = 13,
-	CVMX_L2C_EVENT_DATA_STORE_READ = 14,
+	CVMX_L2C_EVENT_CYCLES           =  0,
+	CVMX_L2C_EVENT_INSTRUCTION_MISS =  1,
+	CVMX_L2C_EVENT_INSTRUCTION_HIT  =  2,
+	CVMX_L2C_EVENT_DATA_MISS        =  3,
+	CVMX_L2C_EVENT_DATA_HIT         =  4,
+	CVMX_L2C_EVENT_MISS             =  5,
+	CVMX_L2C_EVENT_HIT              =  6,
+	CVMX_L2C_EVENT_VICTIM_HIT       =  7,
+	CVMX_L2C_EVENT_INDEX_CONFLICT   =  8,
+	CVMX_L2C_EVENT_TAG_PROBE        =  9,
+	CVMX_L2C_EVENT_TAG_UPDATE       = 10,
+	CVMX_L2C_EVENT_TAG_COMPLETE     = 11,
+	CVMX_L2C_EVENT_TAG_DIRTY        = 12,
+	CVMX_L2C_EVENT_DATA_STORE_NOP   = 13,
+	CVMX_L2C_EVENT_DATA_STORE_READ  = 14,
 	CVMX_L2C_EVENT_DATA_STORE_WRITE = 15,
-	CVMX_L2C_EVENT_FILL_DATA_VALID = 16,
-	CVMX_L2C_EVENT_WRITE_REQUEST = 17,
-	CVMX_L2C_EVENT_READ_REQUEST = 18,
+	CVMX_L2C_EVENT_FILL_DATA_VALID  = 16,
+	CVMX_L2C_EVENT_WRITE_REQUEST    = 17,
+	CVMX_L2C_EVENT_READ_REQUEST     = 18,
 	CVMX_L2C_EVENT_WRITE_DATA_VALID = 19,
-	CVMX_L2C_EVENT_XMC_NOP = 20,
-	CVMX_L2C_EVENT_XMC_LDT = 21,
-	CVMX_L2C_EVENT_XMC_LDI = 22,
-	CVMX_L2C_EVENT_XMC_LDD = 23,
-	CVMX_L2C_EVENT_XMC_STF = 24,
-	CVMX_L2C_EVENT_XMC_STT = 25,
-	CVMX_L2C_EVENT_XMC_STP = 26,
-	CVMX_L2C_EVENT_XMC_STC = 27,
-	CVMX_L2C_EVENT_XMC_DWB = 28,
-	CVMX_L2C_EVENT_XMC_PL2 = 29,
-	CVMX_L2C_EVENT_XMC_PSL1 = 30,
-	CVMX_L2C_EVENT_XMC_IOBLD = 31,
-	CVMX_L2C_EVENT_XMC_IOBST = 32,
-	CVMX_L2C_EVENT_XMC_IOBDMA = 33,
-	CVMX_L2C_EVENT_XMC_IOBRSP = 34,
-	CVMX_L2C_EVENT_XMC_BUS_VALID = 35,
-	CVMX_L2C_EVENT_XMC_MEM_DATA = 36,
-	CVMX_L2C_EVENT_XMC_REFL_DATA = 37,
-	CVMX_L2C_EVENT_XMC_IOBRSP_DATA = 38,
-	CVMX_L2C_EVENT_RSC_NOP = 39,
-	CVMX_L2C_EVENT_RSC_STDN = 40,
-	CVMX_L2C_EVENT_RSC_FILL = 41,
-	CVMX_L2C_EVENT_RSC_REFL = 42,
-	CVMX_L2C_EVENT_RSC_STIN = 43,
-	CVMX_L2C_EVENT_RSC_SCIN = 44,
-	CVMX_L2C_EVENT_RSC_SCFL = 45,
-	CVMX_L2C_EVENT_RSC_SCDN = 46,
-	CVMX_L2C_EVENT_RSC_DATA_VALID = 47,
-	CVMX_L2C_EVENT_RSC_VALID_FILL = 48,
-	CVMX_L2C_EVENT_RSC_VALID_STRSP = 49,
-	CVMX_L2C_EVENT_RSC_VALID_REFL = 50,
-	CVMX_L2C_EVENT_LRF_REQ = 51,
-	CVMX_L2C_EVENT_DT_RD_ALLOC = 52,
-	CVMX_L2C_EVENT_DT_WR_INVAL = 53
+	CVMX_L2C_EVENT_XMC_NOP          = 20,
+	CVMX_L2C_EVENT_XMC_LDT          = 21,
+	CVMX_L2C_EVENT_XMC_LDI          = 22,
+	CVMX_L2C_EVENT_XMC_LDD          = 23,
+	CVMX_L2C_EVENT_XMC_STF          = 24,
+	CVMX_L2C_EVENT_XMC_STT          = 25,
+	CVMX_L2C_EVENT_XMC_STP          = 26,
+	CVMX_L2C_EVENT_XMC_STC          = 27,
+	CVMX_L2C_EVENT_XMC_DWB          = 28,
+	CVMX_L2C_EVENT_XMC_PL2          = 29,
+	CVMX_L2C_EVENT_XMC_PSL1         = 30,
+	CVMX_L2C_EVENT_XMC_IOBLD        = 31,
+	CVMX_L2C_EVENT_XMC_IOBST        = 32,
+	CVMX_L2C_EVENT_XMC_IOBDMA       = 33,
+	CVMX_L2C_EVENT_XMC_IOBRSP       = 34,
+	CVMX_L2C_EVENT_XMC_BUS_VALID    = 35,
+	CVMX_L2C_EVENT_XMC_MEM_DATA     = 36,
+	CVMX_L2C_EVENT_XMC_REFL_DATA    = 37,
+	CVMX_L2C_EVENT_XMC_IOBRSP_DATA  = 38,
+	CVMX_L2C_EVENT_RSC_NOP          = 39,
+	CVMX_L2C_EVENT_RSC_STDN         = 40,
+	CVMX_L2C_EVENT_RSC_FILL         = 41,
+	CVMX_L2C_EVENT_RSC_REFL         = 42,
+	CVMX_L2C_EVENT_RSC_STIN         = 43,
+	CVMX_L2C_EVENT_RSC_SCIN         = 44,
+	CVMX_L2C_EVENT_RSC_SCFL         = 45,
+	CVMX_L2C_EVENT_RSC_SCDN         = 46,
+	CVMX_L2C_EVENT_RSC_DATA_VALID   = 47,
+	CVMX_L2C_EVENT_RSC_VALID_FILL   = 48,
+	CVMX_L2C_EVENT_RSC_VALID_STRSP  = 49,
+	CVMX_L2C_EVENT_RSC_VALID_REFL   = 50,
+	CVMX_L2C_EVENT_LRF_REQ          = 51,
+	CVMX_L2C_EVENT_DT_RD_ALLOC      = 52,
+	CVMX_L2C_EVENT_DT_WR_INVAL      = 53,
+	CVMX_L2C_EVENT_MAX
+};
+
+/* L2C Performance Counter events for Octeon2. */
+enum cvmx_l2c_tad_event {
+	CVMX_L2C_TAD_EVENT_NONE          = 0,
+	CVMX_L2C_TAD_EVENT_TAG_HIT       = 1,
+	CVMX_L2C_TAD_EVENT_TAG_MISS      = 2,
+	CVMX_L2C_TAD_EVENT_TAG_NOALLOC   = 3,
+	CVMX_L2C_TAD_EVENT_TAG_VICTIM    = 4,
+	CVMX_L2C_TAD_EVENT_SC_FAIL       = 5,
+	CVMX_L2C_TAD_EVENT_SC_PASS       = 6,
+	CVMX_L2C_TAD_EVENT_LFB_VALID     = 7,
+	CVMX_L2C_TAD_EVENT_LFB_WAIT_LFB  = 8,
+	CVMX_L2C_TAD_EVENT_LFB_WAIT_VAB  = 9,
+	CVMX_L2C_TAD_EVENT_QUAD0_INDEX   = 128,
+	CVMX_L2C_TAD_EVENT_QUAD0_READ    = 129,
+	CVMX_L2C_TAD_EVENT_QUAD0_BANK    = 130,
+	CVMX_L2C_TAD_EVENT_QUAD0_WDAT    = 131,
+	CVMX_L2C_TAD_EVENT_QUAD1_INDEX   = 144,
+	CVMX_L2C_TAD_EVENT_QUAD1_READ    = 145,
+	CVMX_L2C_TAD_EVENT_QUAD1_BANK    = 146,
+	CVMX_L2C_TAD_EVENT_QUAD1_WDAT    = 147,
+	CVMX_L2C_TAD_EVENT_QUAD2_INDEX   = 160,
+	CVMX_L2C_TAD_EVENT_QUAD2_READ    = 161,
+	CVMX_L2C_TAD_EVENT_QUAD2_BANK    = 162,
+	CVMX_L2C_TAD_EVENT_QUAD2_WDAT    = 163,
+	CVMX_L2C_TAD_EVENT_QUAD3_INDEX   = 176,
+	CVMX_L2C_TAD_EVENT_QUAD3_READ    = 177,
+	CVMX_L2C_TAD_EVENT_QUAD3_BANK    = 178,
+	CVMX_L2C_TAD_EVENT_QUAD3_WDAT    = 179,
+	CVMX_L2C_TAD_EVENT_MAX
 };
 
 /**
@@ -132,10 +164,10 @@ enum cvmx_l2c_event {
  * @clear_on_read:  When asserted, any read of the performance counter
  *                       clears the counter.
  *
- * The routine does not clear the counter.
+ * @note The routine does not clear the counter.
  */
-void cvmx_l2c_config_perf(uint32_t counter,
-			  enum cvmx_l2c_event event, uint32_t clear_on_read);
+void cvmx_l2c_config_perf(uint32_t counter, enum cvmx_l2c_event event, uint32_t clear_on_read);
+
 /**
  * Read the given L2 Cache performance counter. The counter must be configured
  * before reading, but this routine does not enforce this requirement.
@@ -160,18 +192,18 @@ int cvmx_l2c_get_core_way_partition(uint32_t core);
 /**
  * Partitions the L2 cache for a core
  *
- * @core:  The core that the partitioning applies to.
+ * @core: The core that the partitioning applies to.
+ * @mask: The partitioning of the ways expressed as a binary
+ *             mask. A 0 bit allows the core to evict cache lines from
+ *             a way, while a 1 bit blocks the core from evicting any
+ *             lines from that way. There must be at least one allowed
+ *             way (0 bit) in the mask.
  *
- * @mask: The partitioning of the ways expressed as a binary mask. A 0
- *        bit allows the core to evict cache lines from a way, while a
- *        1 bit blocks the core from evicting any lines from that
- *        way. There must be at least one allowed way (0 bit) in the
- *        mask.
- *
- * If any ways are blocked for all cores and the HW blocks, then those
- * ways will never have any cache lines evicted from them.  All cores
- * and the hardware blocks are free to read from all ways regardless
- * of the partitioning.
+
+ * @note If any ways are blocked for all cores and the HW blocks, then
+ *       those ways will never have any cache lines evicted from them.
+ *       All cores and the hardware blocks are free to read from all
+ *       ways regardless of the partitioning.
  */
 int cvmx_l2c_set_core_way_partition(uint32_t core, uint32_t mask);
 
@@ -187,19 +219,21 @@ int cvmx_l2c_get_hw_way_partition(void);
 /**
  * Partitions the L2 cache for the hardware blocks.
  *
- * @mask: The partitioning of the ways expressed as a binary mask. A 0
- *        bit allows the core to evict cache lines from a way, while a
- *        1 bit blocks the core from evicting any lines from that
- *        way. There must be at least one allowed way (0 bit) in the
- *        mask.
+ * @mask: The partitioning of the ways expressed as a binary
+ *             mask. A 0 bit allows the core to evict cache lines from
+ *             a way, while a 1 bit blocks the core from evicting any
+ *             lines from that way. There must be at least one allowed
+ *             way (0 bit) in the mask.
  *
- * If any ways are blocked for all cores and the HW blocks, then those
- * ways will never have any cache lines evicted from them.  All cores
- * and the hardware blocks are free to read from all ways regardless
- * of the partitioning.
+
+ * @note If any ways are blocked for all cores and the HW blocks, then
+ *       those ways will never have any cache lines evicted from them.
+ *       All cores and the hardware blocks are free to read from all
+ *       ways regardless of the partitioning.
  */
 int cvmx_l2c_set_hw_way_partition(uint32_t mask);
 
+
 /**
  * Locks a line in the L2 cache at the specified physical address
  *
@@ -263,13 +297,14 @@ int cvmx_l2c_unlock_mem_region(uint64_t start, uint64_t len);
  */
 union cvmx_l2c_tag cvmx_l2c_get_tag(uint32_t association, uint32_t index);
 
-/* Wrapper around deprecated old function name */
-static inline union cvmx_l2c_tag cvmx_get_l2c_tag(uint32_t association,
-					      uint32_t index)
+/* Wrapper providing a deprecated old function name */
+static inline union cvmx_l2c_tag cvmx_get_l2c_tag(uint32_t association, uint32_t index) __attribute__((deprecated));
+static inline union cvmx_l2c_tag cvmx_get_l2c_tag(uint32_t association, uint32_t index)
 {
 	return cvmx_l2c_get_tag(association, index);
 }
 
+
 /**
  * Returns the cache index for a given physical address
  *