1 /* 2 * Performance events support for SH-4A performance counters 3 * 4 * Copyright (C) 2009, 2010 Paul Mundt 5 * 6 * This file is subject to the terms and conditions of the GNU General Public 7 * License. See the file "COPYING" in the main directory of this archive 8 * for more details. 9 */ 10 #include <linux/kernel.h> 11 #include <linux/init.h> 12 #include <linux/io.h> 13 #include <linux/irq.h> 14 #include <linux/perf_event.h> 15 #include <asm/processor.h> 16 17 #define PPC_CCBR(idx) (0xff200800 + (sizeof(u32) * idx)) 18 #define PPC_PMCTR(idx) (0xfc100000 + (sizeof(u32) * idx)) 19 20 #define CCBR_CIT_MASK (0x7ff << 6) 21 #define CCBR_DUC (1 << 3) 22 #define CCBR_CMDS (1 << 1) 23 #define CCBR_PPCE (1 << 0) 24 25 #ifdef CONFIG_CPU_SHX3 26 /* 27 * The PMCAT location for SH-X3 CPUs was quietly moved, while the CCBR 28 * and PMCTR locations remains tentatively constant. This change remains 29 * wholly undocumented, and was simply found through trial and error. 30 * 31 * Early cuts of SH-X3 still appear to use the SH-X/SH-X2 locations, and 32 * it's unclear when this ceased to be the case. For now we always use 33 * the new location (if future parts keep up with this trend then 34 * scanning for them at runtime also remains a viable option.) 35 * 36 * The gap in the register space also suggests that there are other 37 * undocumented counters, so this will need to be revisited at a later 38 * point in time. 39 */ 40 #define PPC_PMCAT 0xfc100240 41 #else 42 #define PPC_PMCAT 0xfc100080 43 #endif 44 45 #define PMCAT_OVF3 (1 << 27) 46 #define PMCAT_CNN3 (1 << 26) 47 #define PMCAT_CLR3 (1 << 25) 48 #define PMCAT_OVF2 (1 << 19) 49 #define PMCAT_CLR2 (1 << 17) 50 #define PMCAT_OVF1 (1 << 11) 51 #define PMCAT_CNN1 (1 << 10) 52 #define PMCAT_CLR1 (1 << 9) 53 #define PMCAT_OVF0 (1 << 3) 54 #define PMCAT_CLR0 (1 << 1) 55 56 static struct sh_pmu sh4a_pmu; 57 58 /* 59 * Supported raw event codes: 60 * 61 * Event Code Description 62 * ---------- ----------- 63 * 64 * 0x0000 number of elapsed cycles 65 * 0x0200 number of elapsed cycles in privileged mode 66 * 0x0280 number of elapsed cycles while SR.BL is asserted 67 * 0x0202 instruction execution 68 * 0x0203 instruction execution in parallel 69 * 0x0204 number of unconditional branches 70 * 0x0208 number of exceptions 71 * 0x0209 number of interrupts 72 * 0x0220 UTLB miss caused by instruction fetch 73 * 0x0222 UTLB miss caused by operand access 74 * 0x02a0 number of ITLB misses 75 * 0x0028 number of accesses to instruction memories 76 * 0x0029 number of accesses to instruction cache 77 * 0x002a instruction cache miss 78 * 0x022e number of access to instruction X/Y memory 79 * 0x0030 number of reads to operand memories 80 * 0x0038 number of writes to operand memories 81 * 0x0031 number of operand cache read accesses 82 * 0x0039 number of operand cache write accesses 83 * 0x0032 operand cache read miss 84 * 0x003a operand cache write miss 85 * 0x0236 number of reads to operand X/Y memory 86 * 0x023e number of writes to operand X/Y memory 87 * 0x0237 number of reads to operand U memory 88 * 0x023f number of writes to operand U memory 89 * 0x0337 number of U memory read buffer misses 90 * 0x02b4 number of wait cycles due to operand read access 91 * 0x02bc number of wait cycles due to operand write access 92 * 0x0033 number of wait cycles due to operand cache read miss 93 * 0x003b number of wait cycles due to operand cache write miss 94 */ 95 96 /* 97 * Special reserved bits used by hardware emulators, read values will 98 * vary, but writes must always be 0. 99 */ 100 #define PMCAT_EMU_CLR_MASK ((1 << 24) | (1 << 16) | (1 << 8) | (1 << 0)) 101 102 static const int sh4a_general_events[] = { 103 [PERF_COUNT_HW_CPU_CYCLES] = 0x0000, 104 [PERF_COUNT_HW_INSTRUCTIONS] = 0x0202, 105 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0029, /* I-cache */ 106 [PERF_COUNT_HW_CACHE_MISSES] = 0x002a, /* I-cache */ 107 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x0204, 108 [PERF_COUNT_HW_BRANCH_MISSES] = -1, 109 [PERF_COUNT_HW_BUS_CYCLES] = -1, 110 }; 111 112 #define C(x) PERF_COUNT_HW_CACHE_##x 113 114 static const int sh4a_cache_events 115 [PERF_COUNT_HW_CACHE_MAX] 116 [PERF_COUNT_HW_CACHE_OP_MAX] 117 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 118 { 119 [ C(L1D) ] = { 120 [ C(OP_READ) ] = { 121 [ C(RESULT_ACCESS) ] = 0x0031, 122 [ C(RESULT_MISS) ] = 0x0032, 123 }, 124 [ C(OP_WRITE) ] = { 125 [ C(RESULT_ACCESS) ] = 0x0039, 126 [ C(RESULT_MISS) ] = 0x003a, 127 }, 128 [ C(OP_PREFETCH) ] = { 129 [ C(RESULT_ACCESS) ] = 0, 130 [ C(RESULT_MISS) ] = 0, 131 }, 132 }, 133 134 [ C(L1I) ] = { 135 [ C(OP_READ) ] = { 136 [ C(RESULT_ACCESS) ] = 0x0029, 137 [ C(RESULT_MISS) ] = 0x002a, 138 }, 139 [ C(OP_WRITE) ] = { 140 [ C(RESULT_ACCESS) ] = -1, 141 [ C(RESULT_MISS) ] = -1, 142 }, 143 [ C(OP_PREFETCH) ] = { 144 [ C(RESULT_ACCESS) ] = 0, 145 [ C(RESULT_MISS) ] = 0, 146 }, 147 }, 148 149 [ C(LL) ] = { 150 [ C(OP_READ) ] = { 151 [ C(RESULT_ACCESS) ] = 0x0030, 152 [ C(RESULT_MISS) ] = 0, 153 }, 154 [ C(OP_WRITE) ] = { 155 [ C(RESULT_ACCESS) ] = 0x0038, 156 [ C(RESULT_MISS) ] = 0, 157 }, 158 [ C(OP_PREFETCH) ] = { 159 [ C(RESULT_ACCESS) ] = 0, 160 [ C(RESULT_MISS) ] = 0, 161 }, 162 }, 163 164 [ C(DTLB) ] = { 165 [ C(OP_READ) ] = { 166 [ C(RESULT_ACCESS) ] = 0x0222, 167 [ C(RESULT_MISS) ] = 0x0220, 168 }, 169 [ C(OP_WRITE) ] = { 170 [ C(RESULT_ACCESS) ] = 0, 171 [ C(RESULT_MISS) ] = 0, 172 }, 173 [ C(OP_PREFETCH) ] = { 174 [ C(RESULT_ACCESS) ] = 0, 175 [ C(RESULT_MISS) ] = 0, 176 }, 177 }, 178 179 [ C(ITLB) ] = { 180 [ C(OP_READ) ] = { 181 [ C(RESULT_ACCESS) ] = 0, 182 [ C(RESULT_MISS) ] = 0x02a0, 183 }, 184 [ C(OP_WRITE) ] = { 185 [ C(RESULT_ACCESS) ] = -1, 186 [ C(RESULT_MISS) ] = -1, 187 }, 188 [ C(OP_PREFETCH) ] = { 189 [ C(RESULT_ACCESS) ] = -1, 190 [ C(RESULT_MISS) ] = -1, 191 }, 192 }, 193 194 [ C(BPU) ] = { 195 [ C(OP_READ) ] = { 196 [ C(RESULT_ACCESS) ] = -1, 197 [ C(RESULT_MISS) ] = -1, 198 }, 199 [ C(OP_WRITE) ] = { 200 [ C(RESULT_ACCESS) ] = -1, 201 [ C(RESULT_MISS) ] = -1, 202 }, 203 [ C(OP_PREFETCH) ] = { 204 [ C(RESULT_ACCESS) ] = -1, 205 [ C(RESULT_MISS) ] = -1, 206 }, 207 }, 208 }; 209 210 static int sh4a_event_map(int event) 211 { 212 return sh4a_general_events[event]; 213 } 214 215 static u64 sh4a_pmu_read(int idx) 216 { 217 return __raw_readl(PPC_PMCTR(idx)); 218 } 219 220 static void sh4a_pmu_disable(struct hw_perf_event *hwc, int idx) 221 { 222 unsigned int tmp; 223 224 tmp = __raw_readl(PPC_CCBR(idx)); 225 tmp &= ~(CCBR_CIT_MASK | CCBR_DUC); 226 __raw_writel(tmp, PPC_CCBR(idx)); 227 } 228 229 static void sh4a_pmu_enable(struct hw_perf_event *hwc, int idx) 230 { 231 unsigned int tmp; 232 233 tmp = __raw_readl(PPC_PMCAT); 234 tmp &= ~PMCAT_EMU_CLR_MASK; 235 tmp |= idx ? PMCAT_CLR1 : PMCAT_CLR0; 236 __raw_writel(tmp, PPC_PMCAT); 237 238 tmp = __raw_readl(PPC_CCBR(idx)); 239 tmp |= (hwc->config << 6) | CCBR_CMDS | CCBR_PPCE; 240 __raw_writel(tmp, PPC_CCBR(idx)); 241 242 __raw_writel(__raw_readl(PPC_CCBR(idx)) | CCBR_DUC, PPC_CCBR(idx)); 243 } 244 245 static void sh4a_pmu_disable_all(void) 246 { 247 int i; 248 249 for (i = 0; i < sh4a_pmu.num_events; i++) 250 __raw_writel(__raw_readl(PPC_CCBR(i)) & ~CCBR_DUC, PPC_CCBR(i)); 251 } 252 253 static void sh4a_pmu_enable_all(void) 254 { 255 int i; 256 257 for (i = 0; i < sh4a_pmu.num_events; i++) 258 __raw_writel(__raw_readl(PPC_CCBR(i)) | CCBR_DUC, PPC_CCBR(i)); 259 } 260 261 static struct sh_pmu sh4a_pmu = { 262 .name = "sh4a", 263 .num_events = 2, 264 .event_map = sh4a_event_map, 265 .max_events = ARRAY_SIZE(sh4a_general_events), 266 .raw_event_mask = 0x3ff, 267 .cache_events = &sh4a_cache_events, 268 .read = sh4a_pmu_read, 269 .disable = sh4a_pmu_disable, 270 .enable = sh4a_pmu_enable, 271 .disable_all = sh4a_pmu_disable_all, 272 .enable_all = sh4a_pmu_enable_all, 273 }; 274 275 static int __init sh4a_pmu_init(void) 276 { 277 /* 278 * Make sure this CPU actually has perf counters. 279 */ 280 if (!(boot_cpu_data.flags & CPU_HAS_PERF_COUNTER)) { 281 pr_notice("HW perf events unsupported, software events only.\n"); 282 return -ENODEV; 283 } 284 285 return register_sh_pmu(&sh4a_pmu); 286 } 287 early_initcall(sh4a_pmu_init); 288