xref: /linux/drivers/perf/riscv_pmu_sbi.c (revision 5e2cb28dd7e182dfa641550dfa225913509ad45d)
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 
23 #include <asm/errata_list.h>
24 #include <asm/sbi.h>
25 #include <asm/hwcap.h>
26 
27 #define SYSCTL_NO_USER_ACCESS	0
28 #define SYSCTL_USER_ACCESS	1
29 #define SYSCTL_LEGACY		2
30 
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)
34 
35 PMU_FORMAT_ATTR(event, "config:0-47");
36 PMU_FORMAT_ATTR(firmware, "config:63");
37 
38 static struct attribute *riscv_arch_formats_attr[] = {
39 	&format_attr_event.attr,
40 	&format_attr_firmware.attr,
41 	NULL,
42 };
43 
44 static struct attribute_group riscv_pmu_format_group = {
45 	.name = "format",
46 	.attrs = riscv_arch_formats_attr,
47 };
48 
49 static const struct attribute_group *riscv_pmu_attr_groups[] = {
50 	&riscv_pmu_format_group,
51 	NULL,
52 };
53 
54 /* Allow user mode access by default */
55 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS;
56 
57 /*
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
60  */
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;
65 
66 /* Cache the available counters in a bitmask */
67 static unsigned long cmask;
68 
69 struct sbi_pmu_event_data {
70 	union {
71 		union {
72 			struct hw_gen_event {
73 				uint32_t event_code:16;
74 				uint32_t event_type:4;
75 				uint32_t reserved:12;
76 			} hw_gen_event;
77 			struct hw_cache_event {
78 				uint32_t result_id:1;
79 				uint32_t op_id:2;
80 				uint32_t cache_id:13;
81 				uint32_t event_type:4;
82 				uint32_t reserved:12;
83 			} hw_cache_event;
84 		};
85 		uint32_t event_idx;
86 	};
87 };
88 
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}},
120 };
121 
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] = {
126 	[C(L1D)] = {
127 		[C(OP_READ)] = {
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}},
132 		},
133 		[C(OP_WRITE)] = {
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}},
138 		},
139 		[C(OP_PREFETCH)] = {
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}},
144 		},
145 	},
146 	[C(L1I)] = {
147 		[C(OP_READ)] = {
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}},
152 		},
153 		[C(OP_WRITE)] = {
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}},
158 		},
159 		[C(OP_PREFETCH)] = {
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}},
164 		},
165 	},
166 	[C(LL)] = {
167 		[C(OP_READ)] = {
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}},
172 		},
173 		[C(OP_WRITE)] = {
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}},
178 		},
179 		[C(OP_PREFETCH)] = {
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}},
184 		},
185 	},
186 	[C(DTLB)] = {
187 		[C(OP_READ)] = {
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}},
192 		},
193 		[C(OP_WRITE)] = {
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}},
198 		},
199 		[C(OP_PREFETCH)] = {
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}},
204 		},
205 	},
206 	[C(ITLB)] = {
207 		[C(OP_READ)] = {
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}},
212 		},
213 		[C(OP_WRITE)] = {
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}},
218 		},
219 		[C(OP_PREFETCH)] = {
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}},
224 		},
225 	},
226 	[C(BPU)] = {
227 		[C(OP_READ)] = {
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}},
232 		},
233 		[C(OP_WRITE)] = {
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}},
238 		},
239 		[C(OP_PREFETCH)] = {
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}},
244 		},
245 	},
246 	[C(NODE)] = {
247 		[C(OP_READ)] = {
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}},
252 		},
253 		[C(OP_WRITE)] = {
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}},
258 		},
259 		[C(OP_PREFETCH)] = {
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}},
264 		},
265 	},
266 };
267 
268 static int pmu_sbi_ctr_get_width(int idx)
269 {
270 	return pmu_ctr_list[idx].width;
271 }
272 
273 static bool pmu_sbi_ctr_is_fw(int cidx)
274 {
275 	union sbi_pmu_ctr_info *info;
276 
277 	info = &pmu_ctr_list[cidx];
278 	if (!info)
279 		return false;
280 
281 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
282 }
283 
284 /*
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.
288  */
289 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
290 {
291 	int i;
292 	union sbi_pmu_ctr_info *info;
293 	u32 hpm_width = 0, hpm_count = 0;
294 
295 	if (!cmask)
296 		return -EINVAL;
297 
298 	for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
299 		info = &pmu_ctr_list[i];
300 		if (!info)
301 			continue;
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)
305 			hpm_count++;
306 	}
307 
308 	*hw_ctr_width = hpm_width;
309 	*num_hw_ctr = hpm_count;
310 
311 	return 0;
312 }
313 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
314 
315 static uint8_t pmu_sbi_csr_index(struct perf_event *event)
316 {
317 	return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE;
318 }
319 
320 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
321 {
322 	unsigned long cflags = 0;
323 	bool guest_events = false;
324 
325 	if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
326 		guest_events = true;
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;
337 
338 	return cflags;
339 }
340 
341 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
342 {
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);
346 	struct sbiret ret;
347 	int idx;
348 	uint64_t cbase = 0, cmask = rvpmu->cmask;
349 	unsigned long cflags = 0;
350 
351 	cflags = pmu_sbi_get_filter_flags(event);
352 
353 	/*
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.
357 	 */
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;
361 			cmask = 1;
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);
365 		}
366 	}
367 
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,
372 			hwc->config >> 32);
373 #else
374 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
375 			cmask, cflags, hwc->event_base, hwc->config, 0);
376 #endif
377 	if (ret.error) {
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);
381 	}
382 
383 	idx = ret.value;
384 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
385 		return -ENOENT;
386 
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))
390 			return idx;
391 	} else {
392 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
393 			return idx;
394 	}
395 
396 	return -ENOENT;
397 }
398 
399 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
400 {
401 
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);
405 	int idx = hwc->idx;
406 
407 	if (pmu_sbi_ctr_is_fw(idx))
408 		clear_bit(idx, cpuc->used_fw_ctrs);
409 	else
410 		clear_bit(idx, cpuc->used_hw_ctrs);
411 }
412 
413 static int pmu_event_find_cache(u64 config)
414 {
415 	unsigned int cache_type, cache_op, cache_result, ret;
416 
417 	cache_type = (config >>  0) & 0xff;
418 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
419 		return -EINVAL;
420 
421 	cache_op = (config >>  8) & 0xff;
422 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
423 		return -EINVAL;
424 
425 	cache_result = (config >> 16) & 0xff;
426 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
427 		return -EINVAL;
428 
429 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
430 
431 	return ret;
432 }
433 
434 static bool pmu_sbi_is_fw_event(struct perf_event *event)
435 {
436 	u32 type = event->attr.type;
437 	u64 config = event->attr.config;
438 
439 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
440 		return true;
441 	else
442 		return false;
443 }
444 
445 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
446 {
447 	u32 type = event->attr.type;
448 	u64 config = event->attr.config;
449 	int bSoftware;
450 	u64 raw_config_val;
451 	int ret;
452 
453 	switch (type) {
454 	case PERF_TYPE_HARDWARE:
455 		if (config >= PERF_COUNT_HW_MAX)
456 			return -EINVAL;
457 		ret = pmu_hw_event_map[event->attr.config].event_idx;
458 		break;
459 	case PERF_TYPE_HW_CACHE:
460 		ret = pmu_event_find_cache(config);
461 		break;
462 	case PERF_TYPE_RAW:
463 		/*
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.
467 		 */
468 		bSoftware = config >> 63;
469 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
470 		if (bSoftware) {
471 			ret = (raw_config_val & 0xFFFF) |
472 				(SBI_PMU_EVENT_TYPE_FW << 16);
473 		} else {
474 			ret = RISCV_PMU_RAW_EVENT_IDX;
475 			*econfig = raw_config_val;
476 		}
477 		break;
478 	default:
479 		ret = -EINVAL;
480 		break;
481 	}
482 
483 	return ret;
484 }
485 
486 static u64 pmu_sbi_ctr_read(struct perf_event *event)
487 {
488 	struct hw_perf_event *hwc = &event->hw;
489 	int idx = hwc->idx;
490 	struct sbiret ret;
491 	union sbi_pmu_ctr_info info;
492 	u64 val = 0;
493 
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);
497 		if (!ret.error)
498 			val = ret.value;
499 	} else {
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;
504 	}
505 
506 	return val;
507 }
508 
509 static void pmu_sbi_set_scounteren(void *arg)
510 {
511 	struct perf_event *event = (struct perf_event *)arg;
512 
513 	if (event->hw.idx != -1)
514 		csr_write(CSR_SCOUNTEREN,
515 			  csr_read(CSR_SCOUNTEREN) | (1 << pmu_sbi_csr_index(event)));
516 }
517 
518 static void pmu_sbi_reset_scounteren(void *arg)
519 {
520 	struct perf_event *event = (struct perf_event *)arg;
521 
522 	if (event->hw.idx != -1)
523 		csr_write(CSR_SCOUNTEREN,
524 			  csr_read(CSR_SCOUNTEREN) & ~(1 << pmu_sbi_csr_index(event)));
525 }
526 
527 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
528 {
529 	struct sbiret ret;
530 	struct hw_perf_event *hwc = &event->hw;
531 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
532 
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);
536 #else
537 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
538 			1, flag, ival, 0, 0);
539 #endif
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));
543 
544 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
545 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
546 		on_each_cpu_mask(mm_cpumask(event->owner->mm),
547 				 pmu_sbi_set_scounteren, (void *)event, 1);
548 }
549 
550 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
551 {
552 	struct sbiret ret;
553 	struct hw_perf_event *hwc = &event->hw;
554 
555 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
556 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
557 		on_each_cpu_mask(mm_cpumask(event->owner->mm),
558 				 pmu_sbi_reset_scounteren, (void *)event, 1);
559 
560 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
561 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
562 		flag != SBI_PMU_STOP_FLAG_RESET)
563 		pr_err("Stopping counter idx %d failed with error %d\n",
564 			hwc->idx, sbi_err_map_linux_errno(ret.error));
565 }
566 
567 static int pmu_sbi_find_num_ctrs(void)
568 {
569 	struct sbiret ret;
570 
571 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
572 	if (!ret.error)
573 		return ret.value;
574 	else
575 		return sbi_err_map_linux_errno(ret.error);
576 }
577 
578 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
579 {
580 	struct sbiret ret;
581 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
582 	union sbi_pmu_ctr_info cinfo;
583 
584 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
585 	if (!pmu_ctr_list)
586 		return -ENOMEM;
587 
588 	for (i = 0; i < nctr; i++) {
589 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
590 		if (ret.error)
591 			/* The logical counter ids are not expected to be contiguous */
592 			continue;
593 
594 		*mask |= BIT(i);
595 
596 		cinfo.value = ret.value;
597 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
598 			num_fw_ctr++;
599 		else
600 			num_hw_ctr++;
601 		pmu_ctr_list[i].value = cinfo.value;
602 	}
603 
604 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
605 
606 	return 0;
607 }
608 
609 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
610 {
611 	/*
612 	 * No need to check the error because we are disabling all the counters
613 	 * which may include counters that are not enabled yet.
614 	 */
615 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
616 		  0, pmu->cmask, 0, 0, 0, 0);
617 }
618 
619 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
620 {
621 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
622 
623 	/* No need to check the error here as we can't do anything about the error */
624 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
625 		  cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
626 }
627 
628 /*
629  * This function starts all the used counters in two step approach.
630  * Any counter that did not overflow can be start in a single step
631  * while the overflowed counters need to be started with updated initialization
632  * value.
633  */
634 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
635 					       unsigned long ctr_ovf_mask)
636 {
637 	int idx = 0;
638 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
639 	struct perf_event *event;
640 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
641 	unsigned long ctr_start_mask = 0;
642 	uint64_t max_period;
643 	struct hw_perf_event *hwc;
644 	u64 init_val = 0;
645 
646 	ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
647 
648 	/* Start all the counters that did not overflow in a single shot */
649 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
650 		  0, 0, 0, 0);
651 
652 	/* Reinitialize and start all the counter that overflowed */
653 	while (ctr_ovf_mask) {
654 		if (ctr_ovf_mask & 0x01) {
655 			event = cpu_hw_evt->events[idx];
656 			hwc = &event->hw;
657 			max_period = riscv_pmu_ctr_get_width_mask(event);
658 			init_val = local64_read(&hwc->prev_count) & max_period;
659 #if defined(CONFIG_32BIT)
660 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
661 				  flag, init_val, init_val >> 32, 0);
662 #else
663 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
664 				  flag, init_val, 0, 0);
665 #endif
666 			perf_event_update_userpage(event);
667 		}
668 		ctr_ovf_mask = ctr_ovf_mask >> 1;
669 		idx++;
670 	}
671 }
672 
673 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
674 {
675 	struct perf_sample_data data;
676 	struct pt_regs *regs;
677 	struct hw_perf_event *hw_evt;
678 	union sbi_pmu_ctr_info *info;
679 	int lidx, hidx, fidx;
680 	struct riscv_pmu *pmu;
681 	struct perf_event *event;
682 	unsigned long overflow;
683 	unsigned long overflowed_ctrs = 0;
684 	struct cpu_hw_events *cpu_hw_evt = dev;
685 	u64 start_clock = sched_clock();
686 
687 	if (WARN_ON_ONCE(!cpu_hw_evt))
688 		return IRQ_NONE;
689 
690 	/* Firmware counter don't support overflow yet */
691 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
692 	event = cpu_hw_evt->events[fidx];
693 	if (!event) {
694 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
695 		return IRQ_NONE;
696 	}
697 
698 	pmu = to_riscv_pmu(event->pmu);
699 	pmu_sbi_stop_hw_ctrs(pmu);
700 
701 	/* Overflow status register should only be read after counter are stopped */
702 	ALT_SBI_PMU_OVERFLOW(overflow);
703 
704 	/*
705 	 * Overflow interrupt pending bit should only be cleared after stopping
706 	 * all the counters to avoid any race condition.
707 	 */
708 	csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
709 
710 	/* No overflow bit is set */
711 	if (!overflow)
712 		return IRQ_NONE;
713 
714 	regs = get_irq_regs();
715 
716 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
717 		struct perf_event *event = cpu_hw_evt->events[lidx];
718 
719 		/* Skip if invalid event or user did not request a sampling */
720 		if (!event || !is_sampling_event(event))
721 			continue;
722 
723 		info = &pmu_ctr_list[lidx];
724 		/* Do a sanity check */
725 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
726 			continue;
727 
728 		/* compute hardware counter index */
729 		hidx = info->csr - CSR_CYCLE;
730 		/* check if the corresponding bit is set in sscountovf */
731 		if (!(overflow & (1 << hidx)))
732 			continue;
733 
734 		/*
735 		 * Keep a track of overflowed counters so that they can be started
736 		 * with updated initial value.
737 		 */
738 		overflowed_ctrs |= 1 << lidx;
739 		hw_evt = &event->hw;
740 		riscv_pmu_event_update(event);
741 		perf_sample_data_init(&data, 0, hw_evt->last_period);
742 		if (riscv_pmu_event_set_period(event)) {
743 			/*
744 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
745 			 * to avoid throttling here. As per the specification, the
746 			 * interrupt remains disabled until the OF bit is set.
747 			 * Interrupts are enabled again only during the start.
748 			 * TODO: We will need to stop the guest counters once
749 			 * virtualization support is added.
750 			 */
751 			perf_event_overflow(event, &data, regs);
752 		}
753 	}
754 
755 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
756 	perf_sample_event_took(sched_clock() - start_clock);
757 
758 	return IRQ_HANDLED;
759 }
760 
761 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
762 {
763 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
764 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
765 
766 	/*
767 	 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the
768 	 * legacy option but that will be removed in the future.
769 	 */
770 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
771 		csr_write(CSR_SCOUNTEREN, 0x7);
772 	else
773 		csr_write(CSR_SCOUNTEREN, 0x2);
774 
775 	/* Stop all the counters so that they can be enabled from perf */
776 	pmu_sbi_stop_all(pmu);
777 
778 	if (riscv_pmu_use_irq) {
779 		cpu_hw_evt->irq = riscv_pmu_irq;
780 		csr_clear(CSR_IP, BIT(riscv_pmu_irq_num));
781 		csr_set(CSR_IE, BIT(riscv_pmu_irq_num));
782 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
783 	}
784 
785 	return 0;
786 }
787 
788 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
789 {
790 	if (riscv_pmu_use_irq) {
791 		disable_percpu_irq(riscv_pmu_irq);
792 		csr_clear(CSR_IE, BIT(riscv_pmu_irq_num));
793 	}
794 
795 	/* Disable all counters access for user mode now */
796 	csr_write(CSR_SCOUNTEREN, 0x0);
797 
798 	return 0;
799 }
800 
801 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
802 {
803 	int ret;
804 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
805 	struct irq_domain *domain = NULL;
806 
807 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
808 		riscv_pmu_irq_num = RV_IRQ_PMU;
809 		riscv_pmu_use_irq = true;
810 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
811 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
812 		   riscv_cached_marchid(0) == 0 &&
813 		   riscv_cached_mimpid(0) == 0) {
814 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
815 		riscv_pmu_use_irq = true;
816 	}
817 
818 	if (!riscv_pmu_use_irq)
819 		return -EOPNOTSUPP;
820 
821 	domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
822 					  DOMAIN_BUS_ANY);
823 	if (!domain) {
824 		pr_err("Failed to find INTC IRQ root domain\n");
825 		return -ENODEV;
826 	}
827 
828 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
829 	if (!riscv_pmu_irq) {
830 		pr_err("Failed to map PMU interrupt for node\n");
831 		return -ENODEV;
832 	}
833 
834 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
835 	if (ret) {
836 		pr_err("registering percpu irq failed [%d]\n", ret);
837 		return ret;
838 	}
839 
840 	return 0;
841 }
842 
843 #ifdef CONFIG_CPU_PM
844 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
845 				void *v)
846 {
847 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
848 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
849 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
850 	struct perf_event *event;
851 	int idx;
852 
853 	if (!enabled)
854 		return NOTIFY_OK;
855 
856 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
857 		event = cpuc->events[idx];
858 		if (!event)
859 			continue;
860 
861 		switch (cmd) {
862 		case CPU_PM_ENTER:
863 			/*
864 			 * Stop and update the counter
865 			 */
866 			riscv_pmu_stop(event, PERF_EF_UPDATE);
867 			break;
868 		case CPU_PM_EXIT:
869 		case CPU_PM_ENTER_FAILED:
870 			/*
871 			 * Restore and enable the counter.
872 			 */
873 			riscv_pmu_start(event, PERF_EF_RELOAD);
874 			break;
875 		default:
876 			break;
877 		}
878 	}
879 
880 	return NOTIFY_OK;
881 }
882 
883 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
884 {
885 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
886 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
887 }
888 
889 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
890 {
891 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
892 }
893 #else
894 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
895 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
896 #endif
897 
898 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
899 {
900 	riscv_pm_pmu_unregister(pmu);
901 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
902 }
903 
904 static void pmu_sbi_event_init(struct perf_event *event)
905 {
906 	/*
907 	 * The permissions are set at event_init so that we do not depend
908 	 * on the sysctl value that can change.
909 	 */
910 	if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS)
911 		event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS;
912 	else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS)
913 		event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS;
914 	else
915 		event->hw.flags |= PERF_EVENT_FLAG_LEGACY;
916 }
917 
918 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm)
919 {
920 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
921 		return;
922 
923 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
924 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
925 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
926 			return;
927 		}
928 	}
929 
930 	/*
931 	 * The user mmapped the event to directly access it: this is where
932 	 * we determine based on sysctl_perf_user_access if we grant userspace
933 	 * the direct access to this event. That means that within the same
934 	 * task, some events may be directly accessible and some other may not,
935 	 * if the user changes the value of sysctl_perf_user_accesss in the
936 	 * meantime.
937 	 */
938 
939 	event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
940 
941 	/*
942 	 * We must enable userspace access *before* advertising in the user page
943 	 * that it is possible to do so to avoid any race.
944 	 * And we must notify all cpus here because threads that currently run
945 	 * on other cpus will try to directly access the counter too without
946 	 * calling pmu_sbi_ctr_start.
947 	 */
948 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
949 		on_each_cpu_mask(mm_cpumask(mm),
950 				 pmu_sbi_set_scounteren, (void *)event, 1);
951 }
952 
953 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm)
954 {
955 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
956 		return;
957 
958 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
959 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
960 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
961 			return;
962 		}
963 	}
964 
965 	/*
966 	 * Here we can directly remove user access since the user does not have
967 	 * access to the user page anymore so we avoid the racy window where the
968 	 * user could have read cap_user_rdpmc to true right before we disable
969 	 * it.
970 	 */
971 	event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT;
972 
973 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
974 		on_each_cpu_mask(mm_cpumask(mm),
975 				 pmu_sbi_reset_scounteren, (void *)event, 1);
976 }
977 
978 static void riscv_pmu_update_counter_access(void *info)
979 {
980 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
981 		csr_write(CSR_SCOUNTEREN, 0x7);
982 	else
983 		csr_write(CSR_SCOUNTEREN, 0x2);
984 }
985 
986 static int riscv_pmu_proc_user_access_handler(struct ctl_table *table,
987 					      int write, void *buffer,
988 					      size_t *lenp, loff_t *ppos)
989 {
990 	int prev = sysctl_perf_user_access;
991 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
992 
993 	/*
994 	 * Test against the previous value since we clear SCOUNTEREN when
995 	 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should
996 	 * not do that if that was already the case.
997 	 */
998 	if (ret || !write || prev == sysctl_perf_user_access)
999 		return ret;
1000 
1001 	on_each_cpu(riscv_pmu_update_counter_access, NULL, 1);
1002 
1003 	return 0;
1004 }
1005 
1006 static struct ctl_table sbi_pmu_sysctl_table[] = {
1007 	{
1008 		.procname       = "perf_user_access",
1009 		.data		= &sysctl_perf_user_access,
1010 		.maxlen		= sizeof(unsigned int),
1011 		.mode           = 0644,
1012 		.proc_handler	= riscv_pmu_proc_user_access_handler,
1013 		.extra1		= SYSCTL_ZERO,
1014 		.extra2		= SYSCTL_TWO,
1015 	},
1016 	{ }
1017 };
1018 
1019 static int pmu_sbi_device_probe(struct platform_device *pdev)
1020 {
1021 	struct riscv_pmu *pmu = NULL;
1022 	int ret = -ENODEV;
1023 	int num_counters;
1024 
1025 	pr_info("SBI PMU extension is available\n");
1026 	pmu = riscv_pmu_alloc();
1027 	if (!pmu)
1028 		return -ENOMEM;
1029 
1030 	num_counters = pmu_sbi_find_num_ctrs();
1031 	if (num_counters < 0) {
1032 		pr_err("SBI PMU extension doesn't provide any counters\n");
1033 		goto out_free;
1034 	}
1035 
1036 	/* It is possible to get from SBI more than max number of counters */
1037 	if (num_counters > RISCV_MAX_COUNTERS) {
1038 		num_counters = RISCV_MAX_COUNTERS;
1039 		pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters);
1040 	}
1041 
1042 	/* cache all the information about counters now */
1043 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
1044 		goto out_free;
1045 
1046 	ret = pmu_sbi_setup_irqs(pmu, pdev);
1047 	if (ret < 0) {
1048 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
1049 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1050 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
1051 	}
1052 
1053 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
1054 	pmu->cmask = cmask;
1055 	pmu->ctr_start = pmu_sbi_ctr_start;
1056 	pmu->ctr_stop = pmu_sbi_ctr_stop;
1057 	pmu->event_map = pmu_sbi_event_map;
1058 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
1059 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
1060 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
1061 	pmu->ctr_read = pmu_sbi_ctr_read;
1062 	pmu->event_init = pmu_sbi_event_init;
1063 	pmu->event_mapped = pmu_sbi_event_mapped;
1064 	pmu->event_unmapped = pmu_sbi_event_unmapped;
1065 	pmu->csr_index = pmu_sbi_csr_index;
1066 
1067 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1068 	if (ret)
1069 		return ret;
1070 
1071 	ret = riscv_pm_pmu_register(pmu);
1072 	if (ret)
1073 		goto out_unregister;
1074 
1075 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
1076 	if (ret)
1077 		goto out_unregister;
1078 
1079 	register_sysctl("kernel", sbi_pmu_sysctl_table);
1080 
1081 	return 0;
1082 
1083 out_unregister:
1084 	riscv_pmu_destroy(pmu);
1085 
1086 out_free:
1087 	kfree(pmu);
1088 	return ret;
1089 }
1090 
1091 static struct platform_driver pmu_sbi_driver = {
1092 	.probe		= pmu_sbi_device_probe,
1093 	.driver		= {
1094 		.name	= RISCV_PMU_SBI_PDEV_NAME,
1095 	},
1096 };
1097 
1098 static int __init pmu_sbi_devinit(void)
1099 {
1100 	int ret;
1101 	struct platform_device *pdev;
1102 
1103 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
1104 	    !sbi_probe_extension(SBI_EXT_PMU)) {
1105 		return 0;
1106 	}
1107 
1108 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
1109 				      "perf/riscv/pmu:starting",
1110 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
1111 	if (ret) {
1112 		pr_err("CPU hotplug notifier could not be registered: %d\n",
1113 		       ret);
1114 		return ret;
1115 	}
1116 
1117 	ret = platform_driver_register(&pmu_sbi_driver);
1118 	if (ret)
1119 		return ret;
1120 
1121 	pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0);
1122 	if (IS_ERR(pdev)) {
1123 		platform_driver_unregister(&pmu_sbi_driver);
1124 		return PTR_ERR(pdev);
1125 	}
1126 
1127 	/* Notify legacy implementation that SBI pmu is available*/
1128 	riscv_pmu_legacy_skip_init();
1129 
1130 	return ret;
1131 }
1132 device_initcall(pmu_sbi_devinit)
1133