1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * RISC-V performance counter support. 4 * 5 * Copyright (C) 2021 Western Digital Corporation or its affiliates. 6 * 7 * This code is based on ARM perf event code which is in turn based on 8 * sparc64 and x86 code. 9 */ 10 11 #define pr_fmt(fmt) "riscv-pmu-sbi: " fmt 12 13 #include <linux/mod_devicetable.h> 14 #include <linux/perf/riscv_pmu.h> 15 #include <linux/platform_device.h> 16 #include <linux/irq.h> 17 #include <linux/irqdomain.h> 18 #include <linux/of_irq.h> 19 #include <linux/of.h> 20 #include <linux/cpu_pm.h> 21 #include <linux/sched/clock.h> 22 #include <linux/soc/andes/irq.h> 23 24 #include <asm/errata_list.h> 25 #include <asm/sbi.h> 26 #include <asm/cpufeature.h> 27 28 #define ALT_SBI_PMU_OVERFLOW(__ovl) \ 29 asm volatile(ALTERNATIVE_2( \ 30 "csrr %0, " __stringify(CSR_SCOUNTOVF), \ 31 "csrr %0, " __stringify(THEAD_C9XX_CSR_SCOUNTEROF), \ 32 THEAD_VENDOR_ID, ERRATA_THEAD_PMU, \ 33 CONFIG_ERRATA_THEAD_PMU, \ 34 "csrr %0, " __stringify(ANDES_CSR_SCOUNTEROF), \ 35 0, RISCV_ISA_EXT_XANDESPMU, \ 36 CONFIG_ANDES_CUSTOM_PMU) \ 37 : "=r" (__ovl) : \ 38 : "memory") 39 40 #define ALT_SBI_PMU_OVF_CLEAR_PENDING(__irq_mask) \ 41 asm volatile(ALTERNATIVE( \ 42 "csrc " __stringify(CSR_IP) ", %0\n\t", \ 43 "csrc " __stringify(ANDES_CSR_SLIP) ", %0\n\t", \ 44 0, RISCV_ISA_EXT_XANDESPMU, \ 45 CONFIG_ANDES_CUSTOM_PMU) \ 46 : : "r"(__irq_mask) \ 47 : "memory") 48 49 #define SYSCTL_NO_USER_ACCESS 0 50 #define SYSCTL_USER_ACCESS 1 51 #define SYSCTL_LEGACY 2 52 53 #define PERF_EVENT_FLAG_NO_USER_ACCESS BIT(SYSCTL_NO_USER_ACCESS) 54 #define PERF_EVENT_FLAG_USER_ACCESS BIT(SYSCTL_USER_ACCESS) 55 #define PERF_EVENT_FLAG_LEGACY BIT(SYSCTL_LEGACY) 56 57 PMU_FORMAT_ATTR(event, "config:0-47"); 58 PMU_FORMAT_ATTR(firmware, "config:63"); 59 60 static bool sbi_v2_available; 61 static DEFINE_STATIC_KEY_FALSE(sbi_pmu_snapshot_available); 62 #define sbi_pmu_snapshot_available() \ 63 static_branch_unlikely(&sbi_pmu_snapshot_available) 64 65 static struct attribute *riscv_arch_formats_attr[] = { 66 &format_attr_event.attr, 67 &format_attr_firmware.attr, 68 NULL, 69 }; 70 71 static struct attribute_group riscv_pmu_format_group = { 72 .name = "format", 73 .attrs = riscv_arch_formats_attr, 74 }; 75 76 static const struct attribute_group *riscv_pmu_attr_groups[] = { 77 &riscv_pmu_format_group, 78 NULL, 79 }; 80 81 /* Allow user mode access by default */ 82 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS; 83 84 /* 85 * RISC-V doesn't have heterogeneous harts yet. This need to be part of 86 * per_cpu in case of harts with different pmu counters 87 */ 88 static union sbi_pmu_ctr_info *pmu_ctr_list; 89 static bool riscv_pmu_use_irq; 90 static unsigned int riscv_pmu_irq_num; 91 static unsigned int riscv_pmu_irq_mask; 92 static unsigned int riscv_pmu_irq; 93 94 /* Cache the available counters in a bitmask */ 95 static unsigned long cmask; 96 97 struct sbi_pmu_event_data { 98 union { 99 union { 100 struct hw_gen_event { 101 uint32_t event_code:16; 102 uint32_t event_type:4; 103 uint32_t reserved:12; 104 } hw_gen_event; 105 struct hw_cache_event { 106 uint32_t result_id:1; 107 uint32_t op_id:2; 108 uint32_t cache_id:13; 109 uint32_t event_type:4; 110 uint32_t reserved:12; 111 } hw_cache_event; 112 }; 113 uint32_t event_idx; 114 }; 115 }; 116 117 static const struct sbi_pmu_event_data pmu_hw_event_map[] = { 118 [PERF_COUNT_HW_CPU_CYCLES] = {.hw_gen_event = { 119 SBI_PMU_HW_CPU_CYCLES, 120 SBI_PMU_EVENT_TYPE_HW, 0}}, 121 [PERF_COUNT_HW_INSTRUCTIONS] = {.hw_gen_event = { 122 SBI_PMU_HW_INSTRUCTIONS, 123 SBI_PMU_EVENT_TYPE_HW, 0}}, 124 [PERF_COUNT_HW_CACHE_REFERENCES] = {.hw_gen_event = { 125 SBI_PMU_HW_CACHE_REFERENCES, 126 SBI_PMU_EVENT_TYPE_HW, 0}}, 127 [PERF_COUNT_HW_CACHE_MISSES] = {.hw_gen_event = { 128 SBI_PMU_HW_CACHE_MISSES, 129 SBI_PMU_EVENT_TYPE_HW, 0}}, 130 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = {.hw_gen_event = { 131 SBI_PMU_HW_BRANCH_INSTRUCTIONS, 132 SBI_PMU_EVENT_TYPE_HW, 0}}, 133 [PERF_COUNT_HW_BRANCH_MISSES] = {.hw_gen_event = { 134 SBI_PMU_HW_BRANCH_MISSES, 135 SBI_PMU_EVENT_TYPE_HW, 0}}, 136 [PERF_COUNT_HW_BUS_CYCLES] = {.hw_gen_event = { 137 SBI_PMU_HW_BUS_CYCLES, 138 SBI_PMU_EVENT_TYPE_HW, 0}}, 139 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = {.hw_gen_event = { 140 SBI_PMU_HW_STALLED_CYCLES_FRONTEND, 141 SBI_PMU_EVENT_TYPE_HW, 0}}, 142 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = {.hw_gen_event = { 143 SBI_PMU_HW_STALLED_CYCLES_BACKEND, 144 SBI_PMU_EVENT_TYPE_HW, 0}}, 145 [PERF_COUNT_HW_REF_CPU_CYCLES] = {.hw_gen_event = { 146 SBI_PMU_HW_REF_CPU_CYCLES, 147 SBI_PMU_EVENT_TYPE_HW, 0}}, 148 }; 149 150 #define C(x) PERF_COUNT_HW_CACHE_##x 151 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX] 152 [PERF_COUNT_HW_CACHE_OP_MAX] 153 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 154 [C(L1D)] = { 155 [C(OP_READ)] = { 156 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 157 C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 158 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 159 C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 160 }, 161 [C(OP_WRITE)] = { 162 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 163 C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 164 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 165 C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 166 }, 167 [C(OP_PREFETCH)] = { 168 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 169 C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 170 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 171 C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 172 }, 173 }, 174 [C(L1I)] = { 175 [C(OP_READ)] = { 176 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 177 C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 178 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ), 179 C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 180 }, 181 [C(OP_WRITE)] = { 182 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 183 C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 184 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 185 C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 186 }, 187 [C(OP_PREFETCH)] = { 188 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 189 C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 190 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 191 C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 192 }, 193 }, 194 [C(LL)] = { 195 [C(OP_READ)] = { 196 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 197 C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 198 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 199 C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 200 }, 201 [C(OP_WRITE)] = { 202 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 203 C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 204 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 205 C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 206 }, 207 [C(OP_PREFETCH)] = { 208 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 209 C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 210 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 211 C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 212 }, 213 }, 214 [C(DTLB)] = { 215 [C(OP_READ)] = { 216 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 217 C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 218 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 219 C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 220 }, 221 [C(OP_WRITE)] = { 222 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 223 C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 224 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 225 C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 226 }, 227 [C(OP_PREFETCH)] = { 228 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 229 C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 230 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 231 C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 232 }, 233 }, 234 [C(ITLB)] = { 235 [C(OP_READ)] = { 236 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 237 C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 238 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 239 C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 240 }, 241 [C(OP_WRITE)] = { 242 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 243 C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 244 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 245 C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 246 }, 247 [C(OP_PREFETCH)] = { 248 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 249 C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 250 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 251 C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 252 }, 253 }, 254 [C(BPU)] = { 255 [C(OP_READ)] = { 256 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 257 C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 258 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 259 C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 260 }, 261 [C(OP_WRITE)] = { 262 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 263 C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 264 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 265 C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 266 }, 267 [C(OP_PREFETCH)] = { 268 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 269 C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 270 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 271 C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 272 }, 273 }, 274 [C(NODE)] = { 275 [C(OP_READ)] = { 276 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 277 C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 278 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 279 C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 280 }, 281 [C(OP_WRITE)] = { 282 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 283 C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 284 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 285 C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 286 }, 287 [C(OP_PREFETCH)] = { 288 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS), 289 C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 290 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), 291 C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}}, 292 }, 293 }, 294 }; 295 296 static int pmu_sbi_ctr_get_width(int idx) 297 { 298 return pmu_ctr_list[idx].width; 299 } 300 301 static bool pmu_sbi_ctr_is_fw(int cidx) 302 { 303 union sbi_pmu_ctr_info *info; 304 305 info = &pmu_ctr_list[cidx]; 306 if (!info) 307 return false; 308 309 return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false; 310 } 311 312 /* 313 * Returns the counter width of a programmable counter and number of hardware 314 * counters. As we don't support heterogeneous CPUs yet, it is okay to just 315 * return the counter width of the first programmable counter. 316 */ 317 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr) 318 { 319 int i; 320 union sbi_pmu_ctr_info *info; 321 u32 hpm_width = 0, hpm_count = 0; 322 323 if (!cmask) 324 return -EINVAL; 325 326 for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) { 327 info = &pmu_ctr_list[i]; 328 if (!info) 329 continue; 330 if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET) 331 hpm_width = info->width; 332 if (info->type == SBI_PMU_CTR_TYPE_HW) 333 hpm_count++; 334 } 335 336 *hw_ctr_width = hpm_width; 337 *num_hw_ctr = hpm_count; 338 339 return 0; 340 } 341 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info); 342 343 static uint8_t pmu_sbi_csr_index(struct perf_event *event) 344 { 345 return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE; 346 } 347 348 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event) 349 { 350 unsigned long cflags = 0; 351 bool guest_events = false; 352 353 if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS) 354 guest_events = true; 355 if (event->attr.exclude_kernel) 356 cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH; 357 if (event->attr.exclude_user) 358 cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH; 359 if (guest_events && event->attr.exclude_hv) 360 cflags |= SBI_PMU_CFG_FLAG_SET_SINH; 361 if (event->attr.exclude_host) 362 cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH; 363 if (event->attr.exclude_guest) 364 cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH; 365 366 return cflags; 367 } 368 369 static int pmu_sbi_ctr_get_idx(struct perf_event *event) 370 { 371 struct hw_perf_event *hwc = &event->hw; 372 struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu); 373 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events); 374 struct sbiret ret; 375 int idx; 376 uint64_t cbase = 0, cmask = rvpmu->cmask; 377 unsigned long cflags = 0; 378 379 cflags = pmu_sbi_get_filter_flags(event); 380 381 /* 382 * In legacy mode, we have to force the fixed counters for those events 383 * but not in the user access mode as we want to use the other counters 384 * that support sampling/filtering. 385 */ 386 if (hwc->flags & PERF_EVENT_FLAG_LEGACY) { 387 if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES) { 388 cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH; 389 cmask = 1; 390 } else if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS) { 391 cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH; 392 cmask = BIT(CSR_INSTRET - CSR_CYCLE); 393 } 394 } 395 396 /* retrieve the available counter index */ 397 #if defined(CONFIG_32BIT) 398 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase, 399 cmask, cflags, hwc->event_base, hwc->config, 400 hwc->config >> 32); 401 #else 402 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase, 403 cmask, cflags, hwc->event_base, hwc->config, 0); 404 #endif 405 if (ret.error) { 406 pr_debug("Not able to find a counter for event %lx config %llx\n", 407 hwc->event_base, hwc->config); 408 return sbi_err_map_linux_errno(ret.error); 409 } 410 411 idx = ret.value; 412 if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value) 413 return -ENOENT; 414 415 /* Additional sanity check for the counter id */ 416 if (pmu_sbi_ctr_is_fw(idx)) { 417 if (!test_and_set_bit(idx, cpuc->used_fw_ctrs)) 418 return idx; 419 } else { 420 if (!test_and_set_bit(idx, cpuc->used_hw_ctrs)) 421 return idx; 422 } 423 424 return -ENOENT; 425 } 426 427 static void pmu_sbi_ctr_clear_idx(struct perf_event *event) 428 { 429 430 struct hw_perf_event *hwc = &event->hw; 431 struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu); 432 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events); 433 int idx = hwc->idx; 434 435 if (pmu_sbi_ctr_is_fw(idx)) 436 clear_bit(idx, cpuc->used_fw_ctrs); 437 else 438 clear_bit(idx, cpuc->used_hw_ctrs); 439 } 440 441 static int pmu_event_find_cache(u64 config) 442 { 443 unsigned int cache_type, cache_op, cache_result, ret; 444 445 cache_type = (config >> 0) & 0xff; 446 if (cache_type >= PERF_COUNT_HW_CACHE_MAX) 447 return -EINVAL; 448 449 cache_op = (config >> 8) & 0xff; 450 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) 451 return -EINVAL; 452 453 cache_result = (config >> 16) & 0xff; 454 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 455 return -EINVAL; 456 457 ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx; 458 459 return ret; 460 } 461 462 static bool pmu_sbi_is_fw_event(struct perf_event *event) 463 { 464 u32 type = event->attr.type; 465 u64 config = event->attr.config; 466 467 if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1)) 468 return true; 469 else 470 return false; 471 } 472 473 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig) 474 { 475 u32 type = event->attr.type; 476 u64 config = event->attr.config; 477 int bSoftware; 478 u64 raw_config_val; 479 int ret; 480 481 switch (type) { 482 case PERF_TYPE_HARDWARE: 483 if (config >= PERF_COUNT_HW_MAX) 484 return -EINVAL; 485 ret = pmu_hw_event_map[event->attr.config].event_idx; 486 break; 487 case PERF_TYPE_HW_CACHE: 488 ret = pmu_event_find_cache(config); 489 break; 490 case PERF_TYPE_RAW: 491 /* 492 * As per SBI specification, the upper 16 bits must be unused for 493 * a raw event. Use the MSB (63b) to distinguish between hardware 494 * raw event and firmware events. 495 */ 496 bSoftware = config >> 63; 497 raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK; 498 if (bSoftware) { 499 ret = (raw_config_val & 0xFFFF) | 500 (SBI_PMU_EVENT_TYPE_FW << 16); 501 } else { 502 ret = RISCV_PMU_RAW_EVENT_IDX; 503 *econfig = raw_config_val; 504 } 505 break; 506 default: 507 ret = -EINVAL; 508 break; 509 } 510 511 return ret; 512 } 513 514 static void pmu_sbi_snapshot_free(struct riscv_pmu *pmu) 515 { 516 int cpu; 517 518 for_each_possible_cpu(cpu) { 519 struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu); 520 521 if (!cpu_hw_evt->snapshot_addr) 522 continue; 523 524 free_page((unsigned long)cpu_hw_evt->snapshot_addr); 525 cpu_hw_evt->snapshot_addr = NULL; 526 cpu_hw_evt->snapshot_addr_phys = 0; 527 } 528 } 529 530 static int pmu_sbi_snapshot_alloc(struct riscv_pmu *pmu) 531 { 532 int cpu; 533 struct page *snapshot_page; 534 535 for_each_possible_cpu(cpu) { 536 struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu); 537 538 snapshot_page = alloc_page(GFP_ATOMIC | __GFP_ZERO); 539 if (!snapshot_page) { 540 pmu_sbi_snapshot_free(pmu); 541 return -ENOMEM; 542 } 543 cpu_hw_evt->snapshot_addr = page_to_virt(snapshot_page); 544 cpu_hw_evt->snapshot_addr_phys = page_to_phys(snapshot_page); 545 } 546 547 return 0; 548 } 549 550 static int pmu_sbi_snapshot_disable(void) 551 { 552 struct sbiret ret; 553 554 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM, SBI_SHMEM_DISABLE, 555 SBI_SHMEM_DISABLE, 0, 0, 0, 0); 556 if (ret.error) { 557 pr_warn("failed to disable snapshot shared memory\n"); 558 return sbi_err_map_linux_errno(ret.error); 559 } 560 561 return 0; 562 } 563 564 static int pmu_sbi_snapshot_setup(struct riscv_pmu *pmu, int cpu) 565 { 566 struct cpu_hw_events *cpu_hw_evt; 567 struct sbiret ret = {0}; 568 569 cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu); 570 if (!cpu_hw_evt->snapshot_addr_phys) 571 return -EINVAL; 572 573 if (cpu_hw_evt->snapshot_set_done) 574 return 0; 575 576 if (IS_ENABLED(CONFIG_32BIT)) 577 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM, 578 cpu_hw_evt->snapshot_addr_phys, 579 (u64)(cpu_hw_evt->snapshot_addr_phys) >> 32, 0, 0, 0, 0); 580 else 581 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM, 582 cpu_hw_evt->snapshot_addr_phys, 0, 0, 0, 0, 0); 583 584 /* Free up the snapshot area memory and fall back to SBI PMU calls without snapshot */ 585 if (ret.error) { 586 if (ret.error != SBI_ERR_NOT_SUPPORTED) 587 pr_warn("pmu snapshot setup failed with error %ld\n", ret.error); 588 return sbi_err_map_linux_errno(ret.error); 589 } 590 591 memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS); 592 cpu_hw_evt->snapshot_set_done = true; 593 594 return 0; 595 } 596 597 static u64 pmu_sbi_ctr_read(struct perf_event *event) 598 { 599 struct hw_perf_event *hwc = &event->hw; 600 int idx = hwc->idx; 601 struct sbiret ret; 602 u64 val = 0; 603 struct riscv_pmu *pmu = to_riscv_pmu(event->pmu); 604 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events); 605 struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr; 606 union sbi_pmu_ctr_info info = pmu_ctr_list[idx]; 607 608 /* Read the value from the shared memory directly only if counter is stopped */ 609 if (sbi_pmu_snapshot_available() && (hwc->state & PERF_HES_STOPPED)) { 610 val = sdata->ctr_values[idx]; 611 return val; 612 } 613 614 if (pmu_sbi_is_fw_event(event)) { 615 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ, 616 hwc->idx, 0, 0, 0, 0, 0); 617 if (ret.error) 618 return 0; 619 620 val = ret.value; 621 if (IS_ENABLED(CONFIG_32BIT) && sbi_v2_available && info.width >= 32) { 622 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ_HI, 623 hwc->idx, 0, 0, 0, 0, 0); 624 if (!ret.error) 625 val |= ((u64)ret.value << 32); 626 else 627 WARN_ONCE(1, "Unable to read upper 32 bits of firmware counter error: %ld\n", 628 ret.error); 629 } 630 } else { 631 val = riscv_pmu_ctr_read_csr(info.csr); 632 if (IS_ENABLED(CONFIG_32BIT)) 633 val |= ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 32; 634 } 635 636 return val; 637 } 638 639 static void pmu_sbi_set_scounteren(void *arg) 640 { 641 struct perf_event *event = (struct perf_event *)arg; 642 643 if (event->hw.idx != -1) 644 csr_write(CSR_SCOUNTEREN, 645 csr_read(CSR_SCOUNTEREN) | BIT(pmu_sbi_csr_index(event))); 646 } 647 648 static void pmu_sbi_reset_scounteren(void *arg) 649 { 650 struct perf_event *event = (struct perf_event *)arg; 651 652 if (event->hw.idx != -1) 653 csr_write(CSR_SCOUNTEREN, 654 csr_read(CSR_SCOUNTEREN) & ~BIT(pmu_sbi_csr_index(event))); 655 } 656 657 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival) 658 { 659 struct sbiret ret; 660 struct hw_perf_event *hwc = &event->hw; 661 unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE; 662 663 /* There is no benefit setting SNAPSHOT FLAG for a single counter */ 664 #if defined(CONFIG_32BIT) 665 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx, 666 1, flag, ival, ival >> 32, 0); 667 #else 668 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx, 669 1, flag, ival, 0, 0); 670 #endif 671 if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED)) 672 pr_err("Starting counter idx %d failed with error %d\n", 673 hwc->idx, sbi_err_map_linux_errno(ret.error)); 674 675 if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) && 676 (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT)) 677 pmu_sbi_set_scounteren((void *)event); 678 } 679 680 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag) 681 { 682 struct sbiret ret; 683 struct hw_perf_event *hwc = &event->hw; 684 struct riscv_pmu *pmu = to_riscv_pmu(event->pmu); 685 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events); 686 struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr; 687 688 if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) && 689 (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT)) 690 pmu_sbi_reset_scounteren((void *)event); 691 692 if (sbi_pmu_snapshot_available()) 693 flag |= SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT; 694 695 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0); 696 if (!ret.error && sbi_pmu_snapshot_available()) { 697 /* 698 * The counter snapshot is based on the index base specified by hwc->idx. 699 * The actual counter value is updated in shared memory at index 0 when counter 700 * mask is 0x01. To ensure accurate counter values, it's necessary to transfer 701 * the counter value to shared memory. However, if hwc->idx is zero, the counter 702 * value is already correctly updated in shared memory, requiring no further 703 * adjustment. 704 */ 705 if (hwc->idx > 0) { 706 sdata->ctr_values[hwc->idx] = sdata->ctr_values[0]; 707 sdata->ctr_values[0] = 0; 708 } 709 } else if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) && 710 flag != SBI_PMU_STOP_FLAG_RESET) { 711 pr_err("Stopping counter idx %d failed with error %d\n", 712 hwc->idx, sbi_err_map_linux_errno(ret.error)); 713 } 714 } 715 716 static int pmu_sbi_find_num_ctrs(void) 717 { 718 struct sbiret ret; 719 720 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0); 721 if (!ret.error) 722 return ret.value; 723 else 724 return sbi_err_map_linux_errno(ret.error); 725 } 726 727 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask) 728 { 729 struct sbiret ret; 730 int i, num_hw_ctr = 0, num_fw_ctr = 0; 731 union sbi_pmu_ctr_info cinfo; 732 733 pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL); 734 if (!pmu_ctr_list) 735 return -ENOMEM; 736 737 for (i = 0; i < nctr; i++) { 738 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0); 739 if (ret.error) 740 /* The logical counter ids are not expected to be contiguous */ 741 continue; 742 743 *mask |= BIT(i); 744 745 cinfo.value = ret.value; 746 if (cinfo.type == SBI_PMU_CTR_TYPE_FW) 747 num_fw_ctr++; 748 else 749 num_hw_ctr++; 750 pmu_ctr_list[i].value = cinfo.value; 751 } 752 753 pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr); 754 755 return 0; 756 } 757 758 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu) 759 { 760 /* 761 * No need to check the error because we are disabling all the counters 762 * which may include counters that are not enabled yet. 763 */ 764 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 765 0, pmu->cmask, 0, 0, 0, 0); 766 } 767 768 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu) 769 { 770 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events); 771 struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr; 772 unsigned long flag = 0; 773 int i, idx; 774 struct sbiret ret; 775 u64 temp_ctr_overflow_mask = 0; 776 777 if (sbi_pmu_snapshot_available()) 778 flag = SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT; 779 780 /* Reset the shadow copy to avoid save/restore any value from previous overflow */ 781 memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS); 782 783 for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) { 784 /* No need to check the error here as we can't do anything about the error */ 785 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, i * BITS_PER_LONG, 786 cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0); 787 if (!ret.error && sbi_pmu_snapshot_available()) { 788 /* Save the counter values to avoid clobbering */ 789 for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG) 790 cpu_hw_evt->snapshot_cval_shcopy[i * BITS_PER_LONG + idx] = 791 sdata->ctr_values[idx]; 792 /* Save the overflow mask to avoid clobbering */ 793 temp_ctr_overflow_mask |= sdata->ctr_overflow_mask << (i * BITS_PER_LONG); 794 } 795 } 796 797 /* Restore the counter values to the shared memory for used hw counters */ 798 if (sbi_pmu_snapshot_available()) { 799 for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) 800 sdata->ctr_values[idx] = cpu_hw_evt->snapshot_cval_shcopy[idx]; 801 if (temp_ctr_overflow_mask) 802 sdata->ctr_overflow_mask = temp_ctr_overflow_mask; 803 } 804 } 805 806 /* 807 * This function starts all the used counters in two step approach. 808 * Any counter that did not overflow can be start in a single step 809 * while the overflowed counters need to be started with updated initialization 810 * value. 811 */ 812 static inline void pmu_sbi_start_ovf_ctrs_sbi(struct cpu_hw_events *cpu_hw_evt, 813 u64 ctr_ovf_mask) 814 { 815 int idx = 0, i; 816 struct perf_event *event; 817 unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE; 818 unsigned long ctr_start_mask = 0; 819 uint64_t max_period; 820 struct hw_perf_event *hwc; 821 u64 init_val = 0; 822 823 for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) { 824 ctr_start_mask = cpu_hw_evt->used_hw_ctrs[i] & ~ctr_ovf_mask; 825 /* Start all the counters that did not overflow in a single shot */ 826 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, i * BITS_PER_LONG, ctr_start_mask, 827 0, 0, 0, 0); 828 } 829 830 /* Reinitialize and start all the counter that overflowed */ 831 while (ctr_ovf_mask) { 832 if (ctr_ovf_mask & 0x01) { 833 event = cpu_hw_evt->events[idx]; 834 hwc = &event->hw; 835 max_period = riscv_pmu_ctr_get_width_mask(event); 836 init_val = local64_read(&hwc->prev_count) & max_period; 837 #if defined(CONFIG_32BIT) 838 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1, 839 flag, init_val, init_val >> 32, 0); 840 #else 841 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1, 842 flag, init_val, 0, 0); 843 #endif 844 perf_event_update_userpage(event); 845 } 846 ctr_ovf_mask = ctr_ovf_mask >> 1; 847 idx++; 848 } 849 } 850 851 static inline void pmu_sbi_start_ovf_ctrs_snapshot(struct cpu_hw_events *cpu_hw_evt, 852 u64 ctr_ovf_mask) 853 { 854 int i, idx = 0; 855 struct perf_event *event; 856 unsigned long flag = SBI_PMU_START_FLAG_INIT_SNAPSHOT; 857 u64 max_period, init_val = 0; 858 struct hw_perf_event *hwc; 859 struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr; 860 861 for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) { 862 if (ctr_ovf_mask & BIT(idx)) { 863 event = cpu_hw_evt->events[idx]; 864 hwc = &event->hw; 865 max_period = riscv_pmu_ctr_get_width_mask(event); 866 init_val = local64_read(&hwc->prev_count) & max_period; 867 cpu_hw_evt->snapshot_cval_shcopy[idx] = init_val; 868 } 869 /* 870 * We do not need to update the non-overflow counters the previous 871 * value should have been there already. 872 */ 873 } 874 875 for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) { 876 /* Restore the counter values to relative indices for used hw counters */ 877 for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG) 878 sdata->ctr_values[idx] = 879 cpu_hw_evt->snapshot_cval_shcopy[idx + i * BITS_PER_LONG]; 880 /* Start all the counters in a single shot */ 881 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx * BITS_PER_LONG, 882 cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0); 883 } 884 } 885 886 static void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu, 887 u64 ctr_ovf_mask) 888 { 889 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events); 890 891 if (sbi_pmu_snapshot_available()) 892 pmu_sbi_start_ovf_ctrs_snapshot(cpu_hw_evt, ctr_ovf_mask); 893 else 894 pmu_sbi_start_ovf_ctrs_sbi(cpu_hw_evt, ctr_ovf_mask); 895 } 896 897 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev) 898 { 899 struct perf_sample_data data; 900 struct pt_regs *regs; 901 struct hw_perf_event *hw_evt; 902 union sbi_pmu_ctr_info *info; 903 int lidx, hidx, fidx; 904 struct riscv_pmu *pmu; 905 struct perf_event *event; 906 u64 overflow; 907 u64 overflowed_ctrs = 0; 908 struct cpu_hw_events *cpu_hw_evt = dev; 909 u64 start_clock = sched_clock(); 910 struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr; 911 912 if (WARN_ON_ONCE(!cpu_hw_evt)) 913 return IRQ_NONE; 914 915 /* Firmware counter don't support overflow yet */ 916 fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS); 917 if (fidx == RISCV_MAX_COUNTERS) { 918 csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num)); 919 return IRQ_NONE; 920 } 921 922 event = cpu_hw_evt->events[fidx]; 923 if (!event) { 924 ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask); 925 return IRQ_NONE; 926 } 927 928 pmu = to_riscv_pmu(event->pmu); 929 pmu_sbi_stop_hw_ctrs(pmu); 930 931 /* Overflow status register should only be read after counter are stopped */ 932 if (sbi_pmu_snapshot_available()) 933 overflow = sdata->ctr_overflow_mask; 934 else 935 ALT_SBI_PMU_OVERFLOW(overflow); 936 937 /* 938 * Overflow interrupt pending bit should only be cleared after stopping 939 * all the counters to avoid any race condition. 940 */ 941 ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask); 942 943 /* No overflow bit is set */ 944 if (!overflow) 945 return IRQ_NONE; 946 947 regs = get_irq_regs(); 948 949 for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) { 950 struct perf_event *event = cpu_hw_evt->events[lidx]; 951 952 /* Skip if invalid event or user did not request a sampling */ 953 if (!event || !is_sampling_event(event)) 954 continue; 955 956 info = &pmu_ctr_list[lidx]; 957 /* Do a sanity check */ 958 if (!info || info->type != SBI_PMU_CTR_TYPE_HW) 959 continue; 960 961 if (sbi_pmu_snapshot_available()) 962 /* SBI implementation already updated the logical indicies */ 963 hidx = lidx; 964 else 965 /* compute hardware counter index */ 966 hidx = info->csr - CSR_CYCLE; 967 968 /* check if the corresponding bit is set in sscountovf or overflow mask in shmem */ 969 if (!(overflow & BIT(hidx))) 970 continue; 971 972 /* 973 * Keep a track of overflowed counters so that they can be started 974 * with updated initial value. 975 */ 976 overflowed_ctrs |= BIT(lidx); 977 hw_evt = &event->hw; 978 /* Update the event states here so that we know the state while reading */ 979 hw_evt->state |= PERF_HES_STOPPED; 980 riscv_pmu_event_update(event); 981 hw_evt->state |= PERF_HES_UPTODATE; 982 perf_sample_data_init(&data, 0, hw_evt->last_period); 983 if (riscv_pmu_event_set_period(event)) { 984 /* 985 * Unlike other ISAs, RISC-V don't have to disable interrupts 986 * to avoid throttling here. As per the specification, the 987 * interrupt remains disabled until the OF bit is set. 988 * Interrupts are enabled again only during the start. 989 * TODO: We will need to stop the guest counters once 990 * virtualization support is added. 991 */ 992 perf_event_overflow(event, &data, regs); 993 } 994 /* Reset the state as we are going to start the counter after the loop */ 995 hw_evt->state = 0; 996 } 997 998 pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs); 999 perf_sample_event_took(sched_clock() - start_clock); 1000 1001 return IRQ_HANDLED; 1002 } 1003 1004 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node) 1005 { 1006 struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node); 1007 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events); 1008 1009 /* 1010 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the 1011 * legacy option but that will be removed in the future. 1012 */ 1013 if (sysctl_perf_user_access == SYSCTL_LEGACY) 1014 csr_write(CSR_SCOUNTEREN, 0x7); 1015 else 1016 csr_write(CSR_SCOUNTEREN, 0x2); 1017 1018 /* Stop all the counters so that they can be enabled from perf */ 1019 pmu_sbi_stop_all(pmu); 1020 1021 if (riscv_pmu_use_irq) { 1022 cpu_hw_evt->irq = riscv_pmu_irq; 1023 ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask); 1024 enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE); 1025 } 1026 1027 if (sbi_pmu_snapshot_available()) 1028 return pmu_sbi_snapshot_setup(pmu, cpu); 1029 1030 return 0; 1031 } 1032 1033 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node) 1034 { 1035 if (riscv_pmu_use_irq) { 1036 disable_percpu_irq(riscv_pmu_irq); 1037 } 1038 1039 /* Disable all counters access for user mode now */ 1040 csr_write(CSR_SCOUNTEREN, 0x0); 1041 1042 if (sbi_pmu_snapshot_available()) 1043 return pmu_sbi_snapshot_disable(); 1044 1045 return 0; 1046 } 1047 1048 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev) 1049 { 1050 int ret; 1051 struct cpu_hw_events __percpu *hw_events = pmu->hw_events; 1052 struct irq_domain *domain = NULL; 1053 1054 if (riscv_isa_extension_available(NULL, SSCOFPMF)) { 1055 riscv_pmu_irq_num = RV_IRQ_PMU; 1056 riscv_pmu_use_irq = true; 1057 } else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) && 1058 riscv_cached_mvendorid(0) == THEAD_VENDOR_ID && 1059 riscv_cached_marchid(0) == 0 && 1060 riscv_cached_mimpid(0) == 0) { 1061 riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU; 1062 riscv_pmu_use_irq = true; 1063 } else if (riscv_isa_extension_available(NULL, XANDESPMU) && 1064 IS_ENABLED(CONFIG_ANDES_CUSTOM_PMU)) { 1065 riscv_pmu_irq_num = ANDES_SLI_CAUSE_BASE + ANDES_RV_IRQ_PMOVI; 1066 riscv_pmu_use_irq = true; 1067 } 1068 1069 riscv_pmu_irq_mask = BIT(riscv_pmu_irq_num % BITS_PER_LONG); 1070 1071 if (!riscv_pmu_use_irq) 1072 return -EOPNOTSUPP; 1073 1074 domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(), 1075 DOMAIN_BUS_ANY); 1076 if (!domain) { 1077 pr_err("Failed to find INTC IRQ root domain\n"); 1078 return -ENODEV; 1079 } 1080 1081 riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num); 1082 if (!riscv_pmu_irq) { 1083 pr_err("Failed to map PMU interrupt for node\n"); 1084 return -ENODEV; 1085 } 1086 1087 ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events); 1088 if (ret) { 1089 pr_err("registering percpu irq failed [%d]\n", ret); 1090 return ret; 1091 } 1092 1093 return 0; 1094 } 1095 1096 #ifdef CONFIG_CPU_PM 1097 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd, 1098 void *v) 1099 { 1100 struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb); 1101 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events); 1102 int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS); 1103 struct perf_event *event; 1104 int idx; 1105 1106 if (!enabled) 1107 return NOTIFY_OK; 1108 1109 for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) { 1110 event = cpuc->events[idx]; 1111 if (!event) 1112 continue; 1113 1114 switch (cmd) { 1115 case CPU_PM_ENTER: 1116 /* 1117 * Stop and update the counter 1118 */ 1119 riscv_pmu_stop(event, PERF_EF_UPDATE); 1120 break; 1121 case CPU_PM_EXIT: 1122 case CPU_PM_ENTER_FAILED: 1123 /* 1124 * Restore and enable the counter. 1125 */ 1126 riscv_pmu_start(event, PERF_EF_RELOAD); 1127 break; 1128 default: 1129 break; 1130 } 1131 } 1132 1133 return NOTIFY_OK; 1134 } 1135 1136 static int riscv_pm_pmu_register(struct riscv_pmu *pmu) 1137 { 1138 pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify; 1139 return cpu_pm_register_notifier(&pmu->riscv_pm_nb); 1140 } 1141 1142 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) 1143 { 1144 cpu_pm_unregister_notifier(&pmu->riscv_pm_nb); 1145 } 1146 #else 1147 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; } 1148 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { } 1149 #endif 1150 1151 static void riscv_pmu_destroy(struct riscv_pmu *pmu) 1152 { 1153 if (sbi_v2_available) { 1154 if (sbi_pmu_snapshot_available()) { 1155 pmu_sbi_snapshot_disable(); 1156 pmu_sbi_snapshot_free(pmu); 1157 } 1158 } 1159 riscv_pm_pmu_unregister(pmu); 1160 cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node); 1161 } 1162 1163 static void pmu_sbi_event_init(struct perf_event *event) 1164 { 1165 /* 1166 * The permissions are set at event_init so that we do not depend 1167 * on the sysctl value that can change. 1168 */ 1169 if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS) 1170 event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS; 1171 else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS) 1172 event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS; 1173 else 1174 event->hw.flags |= PERF_EVENT_FLAG_LEGACY; 1175 } 1176 1177 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm) 1178 { 1179 if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS) 1180 return; 1181 1182 if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) { 1183 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES && 1184 event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) { 1185 return; 1186 } 1187 } 1188 1189 /* 1190 * The user mmapped the event to directly access it: this is where 1191 * we determine based on sysctl_perf_user_access if we grant userspace 1192 * the direct access to this event. That means that within the same 1193 * task, some events may be directly accessible and some other may not, 1194 * if the user changes the value of sysctl_perf_user_accesss in the 1195 * meantime. 1196 */ 1197 1198 event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT; 1199 1200 /* 1201 * We must enable userspace access *before* advertising in the user page 1202 * that it is possible to do so to avoid any race. 1203 * And we must notify all cpus here because threads that currently run 1204 * on other cpus will try to directly access the counter too without 1205 * calling pmu_sbi_ctr_start. 1206 */ 1207 if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS) 1208 on_each_cpu_mask(mm_cpumask(mm), 1209 pmu_sbi_set_scounteren, (void *)event, 1); 1210 } 1211 1212 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm) 1213 { 1214 if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS) 1215 return; 1216 1217 if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) { 1218 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES && 1219 event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) { 1220 return; 1221 } 1222 } 1223 1224 /* 1225 * Here we can directly remove user access since the user does not have 1226 * access to the user page anymore so we avoid the racy window where the 1227 * user could have read cap_user_rdpmc to true right before we disable 1228 * it. 1229 */ 1230 event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT; 1231 1232 if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS) 1233 on_each_cpu_mask(mm_cpumask(mm), 1234 pmu_sbi_reset_scounteren, (void *)event, 1); 1235 } 1236 1237 static void riscv_pmu_update_counter_access(void *info) 1238 { 1239 if (sysctl_perf_user_access == SYSCTL_LEGACY) 1240 csr_write(CSR_SCOUNTEREN, 0x7); 1241 else 1242 csr_write(CSR_SCOUNTEREN, 0x2); 1243 } 1244 1245 static int riscv_pmu_proc_user_access_handler(struct ctl_table *table, 1246 int write, void *buffer, 1247 size_t *lenp, loff_t *ppos) 1248 { 1249 int prev = sysctl_perf_user_access; 1250 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1251 1252 /* 1253 * Test against the previous value since we clear SCOUNTEREN when 1254 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should 1255 * not do that if that was already the case. 1256 */ 1257 if (ret || !write || prev == sysctl_perf_user_access) 1258 return ret; 1259 1260 on_each_cpu(riscv_pmu_update_counter_access, NULL, 1); 1261 1262 return 0; 1263 } 1264 1265 static struct ctl_table sbi_pmu_sysctl_table[] = { 1266 { 1267 .procname = "perf_user_access", 1268 .data = &sysctl_perf_user_access, 1269 .maxlen = sizeof(unsigned int), 1270 .mode = 0644, 1271 .proc_handler = riscv_pmu_proc_user_access_handler, 1272 .extra1 = SYSCTL_ZERO, 1273 .extra2 = SYSCTL_TWO, 1274 }, 1275 }; 1276 1277 static int pmu_sbi_device_probe(struct platform_device *pdev) 1278 { 1279 struct riscv_pmu *pmu = NULL; 1280 int ret = -ENODEV; 1281 int num_counters; 1282 1283 pr_info("SBI PMU extension is available\n"); 1284 pmu = riscv_pmu_alloc(); 1285 if (!pmu) 1286 return -ENOMEM; 1287 1288 num_counters = pmu_sbi_find_num_ctrs(); 1289 if (num_counters < 0) { 1290 pr_err("SBI PMU extension doesn't provide any counters\n"); 1291 goto out_free; 1292 } 1293 1294 /* It is possible to get from SBI more than max number of counters */ 1295 if (num_counters > RISCV_MAX_COUNTERS) { 1296 num_counters = RISCV_MAX_COUNTERS; 1297 pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters); 1298 } 1299 1300 /* cache all the information about counters now */ 1301 if (pmu_sbi_get_ctrinfo(num_counters, &cmask)) 1302 goto out_free; 1303 1304 ret = pmu_sbi_setup_irqs(pmu, pdev); 1305 if (ret < 0) { 1306 pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n"); 1307 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT; 1308 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE; 1309 } 1310 1311 pmu->pmu.attr_groups = riscv_pmu_attr_groups; 1312 pmu->pmu.parent = &pdev->dev; 1313 pmu->cmask = cmask; 1314 pmu->ctr_start = pmu_sbi_ctr_start; 1315 pmu->ctr_stop = pmu_sbi_ctr_stop; 1316 pmu->event_map = pmu_sbi_event_map; 1317 pmu->ctr_get_idx = pmu_sbi_ctr_get_idx; 1318 pmu->ctr_get_width = pmu_sbi_ctr_get_width; 1319 pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx; 1320 pmu->ctr_read = pmu_sbi_ctr_read; 1321 pmu->event_init = pmu_sbi_event_init; 1322 pmu->event_mapped = pmu_sbi_event_mapped; 1323 pmu->event_unmapped = pmu_sbi_event_unmapped; 1324 pmu->csr_index = pmu_sbi_csr_index; 1325 1326 ret = riscv_pm_pmu_register(pmu); 1327 if (ret) 1328 goto out_unregister; 1329 1330 ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW); 1331 if (ret) 1332 goto out_unregister; 1333 1334 /* SBI PMU Snapsphot is only available in SBI v2.0 */ 1335 if (sbi_v2_available) { 1336 ret = pmu_sbi_snapshot_alloc(pmu); 1337 if (ret) 1338 goto out_unregister; 1339 1340 ret = pmu_sbi_snapshot_setup(pmu, smp_processor_id()); 1341 if (ret) { 1342 /* Snapshot is an optional feature. Continue if not available */ 1343 pmu_sbi_snapshot_free(pmu); 1344 } else { 1345 pr_info("SBI PMU snapshot detected\n"); 1346 /* 1347 * We enable it once here for the boot cpu. If snapshot shmem setup 1348 * fails during cpu hotplug process, it will fail to start the cpu 1349 * as we can not handle hetergenous PMUs with different snapshot 1350 * capability. 1351 */ 1352 static_branch_enable(&sbi_pmu_snapshot_available); 1353 } 1354 } 1355 1356 register_sysctl("kernel", sbi_pmu_sysctl_table); 1357 1358 ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node); 1359 if (ret) 1360 goto out_unregister; 1361 1362 return 0; 1363 1364 out_unregister: 1365 riscv_pmu_destroy(pmu); 1366 1367 out_free: 1368 kfree(pmu); 1369 return ret; 1370 } 1371 1372 static struct platform_driver pmu_sbi_driver = { 1373 .probe = pmu_sbi_device_probe, 1374 .driver = { 1375 .name = RISCV_PMU_SBI_PDEV_NAME, 1376 }, 1377 }; 1378 1379 static int __init pmu_sbi_devinit(void) 1380 { 1381 int ret; 1382 struct platform_device *pdev; 1383 1384 if (sbi_spec_version < sbi_mk_version(0, 3) || 1385 !sbi_probe_extension(SBI_EXT_PMU)) { 1386 return 0; 1387 } 1388 1389 if (sbi_spec_version >= sbi_mk_version(2, 0)) 1390 sbi_v2_available = true; 1391 1392 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING, 1393 "perf/riscv/pmu:starting", 1394 pmu_sbi_starting_cpu, pmu_sbi_dying_cpu); 1395 if (ret) { 1396 pr_err("CPU hotplug notifier could not be registered: %d\n", 1397 ret); 1398 return ret; 1399 } 1400 1401 ret = platform_driver_register(&pmu_sbi_driver); 1402 if (ret) 1403 return ret; 1404 1405 pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0); 1406 if (IS_ERR(pdev)) { 1407 platform_driver_unregister(&pmu_sbi_driver); 1408 return PTR_ERR(pdev); 1409 } 1410 1411 /* Notify legacy implementation that SBI pmu is available*/ 1412 riscv_pmu_legacy_skip_init(); 1413 1414 return ret; 1415 } 1416 device_initcall(pmu_sbi_devinit) 1417