1 // SPDX-License-Identifier: GPL-2.0
3 * RISC-V performance counter support.
5 * Copyright (C) 2021 Western Digital Corporation or its affiliates.
7 * This code is based on ARM perf event code which is in turn based on
8 * sparc64 and x86 code.
11 #define pr_fmt(fmt) "riscv-pmu-sbi: " fmt
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>
20 #include <linux/cpu_pm.h>
21 #include <linux/sched/clock.h>
23 #include <asm/errata_list.h>
25 #include <asm/hwcap.h>
27 #define SYSCTL_NO_USER_ACCESS 0
28 #define SYSCTL_USER_ACCESS 1
29 #define SYSCTL_LEGACY 2
31 #define PERF_EVENT_FLAG_NO_USER_ACCESS BIT(SYSCTL_NO_USER_ACCESS)
32 #define PERF_EVENT_FLAG_USER_ACCESS BIT(SYSCTL_USER_ACCESS)
33 #define PERF_EVENT_FLAG_LEGACY BIT(SYSCTL_LEGACY)
35 PMU_FORMAT_ATTR(event, "config:0-47");
36 PMU_FORMAT_ATTR(firmware, "config:63");
38 static struct attribute *riscv_arch_formats_attr[] = {
39 &format_attr_event.attr,
40 &format_attr_firmware.attr,
44 static struct attribute_group riscv_pmu_format_group = {
46 .attrs = riscv_arch_formats_attr,
49 static const struct attribute_group *riscv_pmu_attr_groups[] = {
50 &riscv_pmu_format_group,
54 /* Allow user mode access by default */
55 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS;
58 * RISC-V doesn't have heterogeneous harts yet. This need to be part of
59 * per_cpu in case of harts with different pmu counters
61 static union sbi_pmu_ctr_info *pmu_ctr_list;
62 static bool riscv_pmu_use_irq;
63 static unsigned int riscv_pmu_irq_num;
64 static unsigned int riscv_pmu_irq;
66 /* Cache the available counters in a bitmask */
67 static unsigned long cmask;
69 struct sbi_pmu_event_data {
73 uint32_t event_code:16;
74 uint32_t event_type:4;
77 struct hw_cache_event {
81 uint32_t event_type:4;
89 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
90 [PERF_COUNT_HW_CPU_CYCLES] = {.hw_gen_event = {
91 SBI_PMU_HW_CPU_CYCLES,
92 SBI_PMU_EVENT_TYPE_HW, 0}},
93 [PERF_COUNT_HW_INSTRUCTIONS] = {.hw_gen_event = {
94 SBI_PMU_HW_INSTRUCTIONS,
95 SBI_PMU_EVENT_TYPE_HW, 0}},
96 [PERF_COUNT_HW_CACHE_REFERENCES] = {.hw_gen_event = {
97 SBI_PMU_HW_CACHE_REFERENCES,
98 SBI_PMU_EVENT_TYPE_HW, 0}},
99 [PERF_COUNT_HW_CACHE_MISSES] = {.hw_gen_event = {
100 SBI_PMU_HW_CACHE_MISSES,
101 SBI_PMU_EVENT_TYPE_HW, 0}},
102 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = {.hw_gen_event = {
103 SBI_PMU_HW_BRANCH_INSTRUCTIONS,
104 SBI_PMU_EVENT_TYPE_HW, 0}},
105 [PERF_COUNT_HW_BRANCH_MISSES] = {.hw_gen_event = {
106 SBI_PMU_HW_BRANCH_MISSES,
107 SBI_PMU_EVENT_TYPE_HW, 0}},
108 [PERF_COUNT_HW_BUS_CYCLES] = {.hw_gen_event = {
109 SBI_PMU_HW_BUS_CYCLES,
110 SBI_PMU_EVENT_TYPE_HW, 0}},
111 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = {.hw_gen_event = {
112 SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
113 SBI_PMU_EVENT_TYPE_HW, 0}},
114 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = {.hw_gen_event = {
115 SBI_PMU_HW_STALLED_CYCLES_BACKEND,
116 SBI_PMU_EVENT_TYPE_HW, 0}},
117 [PERF_COUNT_HW_REF_CPU_CYCLES] = {.hw_gen_event = {
118 SBI_PMU_HW_REF_CPU_CYCLES,
119 SBI_PMU_EVENT_TYPE_HW, 0}},
122 #define C(x) PERF_COUNT_HW_CACHE_##x
123 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
124 [PERF_COUNT_HW_CACHE_OP_MAX]
125 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
128 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
129 C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
130 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
131 C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
134 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
135 C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
136 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
137 C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
140 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
141 C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
142 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
143 C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
148 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
149 C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
150 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
151 C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
154 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
155 C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
156 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
157 C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
160 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
161 C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
162 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
163 C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
168 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
169 C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
170 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
171 C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
174 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
175 C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
176 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
177 C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
180 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
181 C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
182 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
183 C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
188 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
189 C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
190 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
191 C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
194 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
195 C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
196 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
197 C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
200 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
201 C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
202 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
203 C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
208 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
209 C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
210 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
211 C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
214 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
215 C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
216 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
217 C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
220 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
221 C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
222 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
223 C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
228 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
229 C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
230 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
231 C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
234 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
235 C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
236 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
237 C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
240 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
241 C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
242 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
243 C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
248 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
249 C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
250 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
251 C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
254 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
255 C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
256 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
257 C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
260 [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
261 C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
262 [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
263 C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
268 static int pmu_sbi_ctr_get_width(int idx)
270 return pmu_ctr_list[idx].width;
273 static bool pmu_sbi_ctr_is_fw(int cidx)
275 union sbi_pmu_ctr_info *info;
277 info = &pmu_ctr_list[cidx];
281 return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
285 * Returns the counter width of a programmable counter and number of hardware
286 * counters. As we don't support heterogeneous CPUs yet, it is okay to just
287 * return the counter width of the first programmable counter.
289 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
292 union sbi_pmu_ctr_info *info;
293 u32 hpm_width = 0, hpm_count = 0;
298 for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
299 info = &pmu_ctr_list[i];
302 if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET)
303 hpm_width = info->width;
304 if (info->type == SBI_PMU_CTR_TYPE_HW)
308 *hw_ctr_width = hpm_width;
309 *num_hw_ctr = hpm_count;
313 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
315 static uint8_t pmu_sbi_csr_index(struct perf_event *event)
317 return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE;
320 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
322 unsigned long cflags = 0;
323 bool guest_events = false;
325 if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
327 if (event->attr.exclude_kernel)
328 cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH;
329 if (event->attr.exclude_user)
330 cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH;
331 if (guest_events && event->attr.exclude_hv)
332 cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
333 if (event->attr.exclude_host)
334 cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH;
335 if (event->attr.exclude_guest)
336 cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH;
341 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
343 struct hw_perf_event *hwc = &event->hw;
344 struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
345 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
348 uint64_t cbase = 0, cmask = rvpmu->cmask;
349 unsigned long cflags = 0;
351 cflags = pmu_sbi_get_filter_flags(event);
354 * In legacy mode, we have to force the fixed counters for those events
355 * but not in the user access mode as we want to use the other counters
356 * that support sampling/filtering.
358 if (hwc->flags & PERF_EVENT_FLAG_LEGACY) {
359 if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
360 cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
362 } else if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS) {
363 cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
364 cmask = 1UL << (CSR_INSTRET - CSR_CYCLE);
368 /* retrieve the available counter index */
369 #if defined(CONFIG_32BIT)
370 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
371 cmask, cflags, hwc->event_base, hwc->config,
374 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
375 cmask, cflags, hwc->event_base, hwc->config, 0);
378 pr_debug("Not able to find a counter for event %lx config %llx\n",
379 hwc->event_base, hwc->config);
380 return sbi_err_map_linux_errno(ret.error);
384 if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
387 /* Additional sanity check for the counter id */
388 if (pmu_sbi_ctr_is_fw(idx)) {
389 if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
392 if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
399 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
402 struct hw_perf_event *hwc = &event->hw;
403 struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
404 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
407 if (pmu_sbi_ctr_is_fw(idx))
408 clear_bit(idx, cpuc->used_fw_ctrs);
410 clear_bit(idx, cpuc->used_hw_ctrs);
413 static int pmu_event_find_cache(u64 config)
415 unsigned int cache_type, cache_op, cache_result, ret;
417 cache_type = (config >> 0) & 0xff;
418 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
421 cache_op = (config >> 8) & 0xff;
422 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
425 cache_result = (config >> 16) & 0xff;
426 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
429 ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
434 static bool pmu_sbi_is_fw_event(struct perf_event *event)
436 u32 type = event->attr.type;
437 u64 config = event->attr.config;
439 if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
445 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
447 u32 type = event->attr.type;
448 u64 config = event->attr.config;
454 case PERF_TYPE_HARDWARE:
455 if (config >= PERF_COUNT_HW_MAX)
457 ret = pmu_hw_event_map[event->attr.config].event_idx;
459 case PERF_TYPE_HW_CACHE:
460 ret = pmu_event_find_cache(config);
464 * As per SBI specification, the upper 16 bits must be unused for
465 * a raw event. Use the MSB (63b) to distinguish between hardware
466 * raw event and firmware events.
468 bSoftware = config >> 63;
469 raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
471 ret = (raw_config_val & 0xFFFF) |
472 (SBI_PMU_EVENT_TYPE_FW << 16);
474 ret = RISCV_PMU_RAW_EVENT_IDX;
475 *econfig = raw_config_val;
486 static u64 pmu_sbi_ctr_read(struct perf_event *event)
488 struct hw_perf_event *hwc = &event->hw;
491 union sbi_pmu_ctr_info info;
494 if (pmu_sbi_is_fw_event(event)) {
495 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
496 hwc->idx, 0, 0, 0, 0, 0);
500 info = pmu_ctr_list[idx];
501 val = riscv_pmu_ctr_read_csr(info.csr);
502 if (IS_ENABLED(CONFIG_32BIT))
503 val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
509 static void pmu_sbi_set_scounteren(void *arg)
511 struct perf_event *event = (struct perf_event *)arg;
513 if (event->hw.idx != -1)
514 csr_write(CSR_SCOUNTEREN,
515 csr_read(CSR_SCOUNTEREN) | (1 << pmu_sbi_csr_index(event)));
518 static void pmu_sbi_reset_scounteren(void *arg)
520 struct perf_event *event = (struct perf_event *)arg;
522 if (event->hw.idx != -1)
523 csr_write(CSR_SCOUNTEREN,
524 csr_read(CSR_SCOUNTEREN) & ~(1 << pmu_sbi_csr_index(event)));
527 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
530 struct hw_perf_event *hwc = &event->hw;
531 unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
533 #if defined(CONFIG_32BIT)
534 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
535 1, flag, ival, ival >> 32, 0);
537 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
538 1, flag, ival, 0, 0);
540 if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
541 pr_err("Starting counter idx %d failed with error %d\n",
542 hwc->idx, sbi_err_map_linux_errno(ret.error));
544 if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
545 (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
546 pmu_sbi_set_scounteren((void *)event);
549 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
552 struct hw_perf_event *hwc = &event->hw;
554 if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
555 (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
556 pmu_sbi_reset_scounteren((void *)event);
558 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
559 if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
560 flag != SBI_PMU_STOP_FLAG_RESET)
561 pr_err("Stopping counter idx %d failed with error %d\n",
562 hwc->idx, sbi_err_map_linux_errno(ret.error));
565 static int pmu_sbi_find_num_ctrs(void)
569 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
573 return sbi_err_map_linux_errno(ret.error);
576 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
579 int i, num_hw_ctr = 0, num_fw_ctr = 0;
580 union sbi_pmu_ctr_info cinfo;
582 pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
586 for (i = 0; i < nctr; i++) {
587 ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
589 /* The logical counter ids are not expected to be contiguous */
594 cinfo.value = ret.value;
595 if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
599 pmu_ctr_list[i].value = cinfo.value;
602 pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
607 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
610 * No need to check the error because we are disabling all the counters
611 * which may include counters that are not enabled yet.
613 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
614 0, pmu->cmask, 0, 0, 0, 0);
617 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
619 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
621 /* No need to check the error here as we can't do anything about the error */
622 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
623 cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
627 * This function starts all the used counters in two step approach.
628 * Any counter that did not overflow can be start in a single step
629 * while the overflowed counters need to be started with updated initialization
632 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
633 unsigned long ctr_ovf_mask)
636 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
637 struct perf_event *event;
638 unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
639 unsigned long ctr_start_mask = 0;
641 struct hw_perf_event *hwc;
644 ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
646 /* Start all the counters that did not overflow in a single shot */
647 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
650 /* Reinitialize and start all the counter that overflowed */
651 while (ctr_ovf_mask) {
652 if (ctr_ovf_mask & 0x01) {
653 event = cpu_hw_evt->events[idx];
655 max_period = riscv_pmu_ctr_get_width_mask(event);
656 init_val = local64_read(&hwc->prev_count) & max_period;
657 #if defined(CONFIG_32BIT)
658 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
659 flag, init_val, init_val >> 32, 0);
661 sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
662 flag, init_val, 0, 0);
664 perf_event_update_userpage(event);
666 ctr_ovf_mask = ctr_ovf_mask >> 1;
671 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
673 struct perf_sample_data data;
674 struct pt_regs *regs;
675 struct hw_perf_event *hw_evt;
676 union sbi_pmu_ctr_info *info;
677 int lidx, hidx, fidx;
678 struct riscv_pmu *pmu;
679 struct perf_event *event;
680 unsigned long overflow;
681 unsigned long overflowed_ctrs = 0;
682 struct cpu_hw_events *cpu_hw_evt = dev;
683 u64 start_clock = sched_clock();
685 if (WARN_ON_ONCE(!cpu_hw_evt))
688 /* Firmware counter don't support overflow yet */
689 fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
690 if (fidx == RISCV_MAX_COUNTERS) {
691 csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
695 event = cpu_hw_evt->events[fidx];
697 csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
701 pmu = to_riscv_pmu(event->pmu);
702 pmu_sbi_stop_hw_ctrs(pmu);
704 /* Overflow status register should only be read after counter are stopped */
705 ALT_SBI_PMU_OVERFLOW(overflow);
708 * Overflow interrupt pending bit should only be cleared after stopping
709 * all the counters to avoid any race condition.
711 csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
713 /* No overflow bit is set */
717 regs = get_irq_regs();
719 for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
720 struct perf_event *event = cpu_hw_evt->events[lidx];
722 /* Skip if invalid event or user did not request a sampling */
723 if (!event || !is_sampling_event(event))
726 info = &pmu_ctr_list[lidx];
727 /* Do a sanity check */
728 if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
731 /* compute hardware counter index */
732 hidx = info->csr - CSR_CYCLE;
733 /* check if the corresponding bit is set in sscountovf */
734 if (!(overflow & (1 << hidx)))
738 * Keep a track of overflowed counters so that they can be started
739 * with updated initial value.
741 overflowed_ctrs |= 1 << lidx;
743 riscv_pmu_event_update(event);
744 perf_sample_data_init(&data, 0, hw_evt->last_period);
745 if (riscv_pmu_event_set_period(event)) {
747 * Unlike other ISAs, RISC-V don't have to disable interrupts
748 * to avoid throttling here. As per the specification, the
749 * interrupt remains disabled until the OF bit is set.
750 * Interrupts are enabled again only during the start.
751 * TODO: We will need to stop the guest counters once
752 * virtualization support is added.
754 perf_event_overflow(event, &data, regs);
758 pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
759 perf_sample_event_took(sched_clock() - start_clock);
764 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
766 struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
767 struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
770 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the
771 * legacy option but that will be removed in the future.
773 if (sysctl_perf_user_access == SYSCTL_LEGACY)
774 csr_write(CSR_SCOUNTEREN, 0x7);
776 csr_write(CSR_SCOUNTEREN, 0x2);
778 /* Stop all the counters so that they can be enabled from perf */
779 pmu_sbi_stop_all(pmu);
781 if (riscv_pmu_use_irq) {
782 cpu_hw_evt->irq = riscv_pmu_irq;
783 csr_clear(CSR_IP, BIT(riscv_pmu_irq_num));
784 csr_set(CSR_IE, BIT(riscv_pmu_irq_num));
785 enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
791 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
793 if (riscv_pmu_use_irq) {
794 disable_percpu_irq(riscv_pmu_irq);
795 csr_clear(CSR_IE, BIT(riscv_pmu_irq_num));
798 /* Disable all counters access for user mode now */
799 csr_write(CSR_SCOUNTEREN, 0x0);
804 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
807 struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
808 struct irq_domain *domain = NULL;
810 if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
811 riscv_pmu_irq_num = RV_IRQ_PMU;
812 riscv_pmu_use_irq = true;
813 } else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
814 riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
815 riscv_cached_marchid(0) == 0 &&
816 riscv_cached_mimpid(0) == 0) {
817 riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
818 riscv_pmu_use_irq = true;
821 if (!riscv_pmu_use_irq)
824 domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
827 pr_err("Failed to find INTC IRQ root domain\n");
831 riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
832 if (!riscv_pmu_irq) {
833 pr_err("Failed to map PMU interrupt for node\n");
837 ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
839 pr_err("registering percpu irq failed [%d]\n", ret);
847 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
850 struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
851 struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
852 int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
853 struct perf_event *event;
859 for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
860 event = cpuc->events[idx];
867 * Stop and update the counter
869 riscv_pmu_stop(event, PERF_EF_UPDATE);
872 case CPU_PM_ENTER_FAILED:
874 * Restore and enable the counter.
876 riscv_pmu_start(event, PERF_EF_RELOAD);
886 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
888 pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
889 return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
892 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
894 cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
897 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
898 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
901 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
903 riscv_pm_pmu_unregister(pmu);
904 cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
907 static void pmu_sbi_event_init(struct perf_event *event)
910 * The permissions are set at event_init so that we do not depend
911 * on the sysctl value that can change.
913 if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS)
914 event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS;
915 else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS)
916 event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS;
918 event->hw.flags |= PERF_EVENT_FLAG_LEGACY;
921 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm)
923 if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
926 if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
927 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
928 event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
934 * The user mmapped the event to directly access it: this is where
935 * we determine based on sysctl_perf_user_access if we grant userspace
936 * the direct access to this event. That means that within the same
937 * task, some events may be directly accessible and some other may not,
938 * if the user changes the value of sysctl_perf_user_accesss in the
942 event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
945 * We must enable userspace access *before* advertising in the user page
946 * that it is possible to do so to avoid any race.
947 * And we must notify all cpus here because threads that currently run
948 * on other cpus will try to directly access the counter too without
949 * calling pmu_sbi_ctr_start.
951 if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
952 on_each_cpu_mask(mm_cpumask(mm),
953 pmu_sbi_set_scounteren, (void *)event, 1);
956 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm)
958 if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
961 if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
962 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
963 event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
969 * Here we can directly remove user access since the user does not have
970 * access to the user page anymore so we avoid the racy window where the
971 * user could have read cap_user_rdpmc to true right before we disable
974 event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT;
976 if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
977 on_each_cpu_mask(mm_cpumask(mm),
978 pmu_sbi_reset_scounteren, (void *)event, 1);
981 static void riscv_pmu_update_counter_access(void *info)
983 if (sysctl_perf_user_access == SYSCTL_LEGACY)
984 csr_write(CSR_SCOUNTEREN, 0x7);
986 csr_write(CSR_SCOUNTEREN, 0x2);
989 static int riscv_pmu_proc_user_access_handler(struct ctl_table *table,
990 int write, void *buffer,
991 size_t *lenp, loff_t *ppos)
993 int prev = sysctl_perf_user_access;
994 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
997 * Test against the previous value since we clear SCOUNTEREN when
998 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should
999 * not do that if that was already the case.
1001 if (ret || !write || prev == sysctl_perf_user_access)
1004 on_each_cpu(riscv_pmu_update_counter_access, NULL, 1);
1009 static struct ctl_table sbi_pmu_sysctl_table[] = {
1011 .procname = "perf_user_access",
1012 .data = &sysctl_perf_user_access,
1013 .maxlen = sizeof(unsigned int),
1015 .proc_handler = riscv_pmu_proc_user_access_handler,
1016 .extra1 = SYSCTL_ZERO,
1017 .extra2 = SYSCTL_TWO,
1022 static int pmu_sbi_device_probe(struct platform_device *pdev)
1024 struct riscv_pmu *pmu = NULL;
1028 pr_info("SBI PMU extension is available\n");
1029 pmu = riscv_pmu_alloc();
1033 num_counters = pmu_sbi_find_num_ctrs();
1034 if (num_counters < 0) {
1035 pr_err("SBI PMU extension doesn't provide any counters\n");
1039 /* It is possible to get from SBI more than max number of counters */
1040 if (num_counters > RISCV_MAX_COUNTERS) {
1041 num_counters = RISCV_MAX_COUNTERS;
1042 pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters);
1045 /* cache all the information about counters now */
1046 if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
1049 ret = pmu_sbi_setup_irqs(pmu, pdev);
1051 pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
1052 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1053 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
1056 pmu->pmu.attr_groups = riscv_pmu_attr_groups;
1058 pmu->ctr_start = pmu_sbi_ctr_start;
1059 pmu->ctr_stop = pmu_sbi_ctr_stop;
1060 pmu->event_map = pmu_sbi_event_map;
1061 pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
1062 pmu->ctr_get_width = pmu_sbi_ctr_get_width;
1063 pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
1064 pmu->ctr_read = pmu_sbi_ctr_read;
1065 pmu->event_init = pmu_sbi_event_init;
1066 pmu->event_mapped = pmu_sbi_event_mapped;
1067 pmu->event_unmapped = pmu_sbi_event_unmapped;
1068 pmu->csr_index = pmu_sbi_csr_index;
1070 ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1074 ret = riscv_pm_pmu_register(pmu);
1076 goto out_unregister;
1078 ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
1080 goto out_unregister;
1082 register_sysctl("kernel", sbi_pmu_sysctl_table);
1087 riscv_pmu_destroy(pmu);
1094 static struct platform_driver pmu_sbi_driver = {
1095 .probe = pmu_sbi_device_probe,
1097 .name = RISCV_PMU_SBI_PDEV_NAME,
1101 static int __init pmu_sbi_devinit(void)
1104 struct platform_device *pdev;
1106 if (sbi_spec_version < sbi_mk_version(0, 3) ||
1107 !sbi_probe_extension(SBI_EXT_PMU)) {
1111 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
1112 "perf/riscv/pmu:starting",
1113 pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
1115 pr_err("CPU hotplug notifier could not be registered: %d\n",
1120 ret = platform_driver_register(&pmu_sbi_driver);
1124 pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0);
1126 platform_driver_unregister(&pmu_sbi_driver);
1127 return PTR_ERR(pdev);
1130 /* Notify legacy implementation that SBI pmu is available*/
1131 riscv_pmu_legacy_skip_init();
1135 device_initcall(pmu_sbi_devinit)