xref: /openbmc/linux/drivers/perf/riscv_pmu_sbi.c (revision 747f7a29)
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 
22 #include <asm/sbi.h>
23 #include <asm/hwcap.h>
24 
25 PMU_FORMAT_ATTR(event, "config:0-47");
26 PMU_FORMAT_ATTR(firmware, "config:63");
27 
28 static struct attribute *riscv_arch_formats_attr[] = {
29 	&format_attr_event.attr,
30 	&format_attr_firmware.attr,
31 	NULL,
32 };
33 
34 static struct attribute_group riscv_pmu_format_group = {
35 	.name = "format",
36 	.attrs = riscv_arch_formats_attr,
37 };
38 
39 static const struct attribute_group *riscv_pmu_attr_groups[] = {
40 	&riscv_pmu_format_group,
41 	NULL,
42 };
43 
44 union sbi_pmu_ctr_info {
45 	unsigned long value;
46 	struct {
47 		unsigned long csr:12;
48 		unsigned long width:6;
49 #if __riscv_xlen == 32
50 		unsigned long reserved:13;
51 #else
52 		unsigned long reserved:45;
53 #endif
54 		unsigned long type:1;
55 	};
56 };
57 
58 /*
59  * RISC-V doesn't have hetergenous harts yet. This need to be part of
60  * per_cpu in case of harts with different pmu counters
61  */
62 static union sbi_pmu_ctr_info *pmu_ctr_list;
63 static unsigned int riscv_pmu_irq;
64 
65 struct sbi_pmu_event_data {
66 	union {
67 		union {
68 			struct hw_gen_event {
69 				uint32_t event_code:16;
70 				uint32_t event_type:4;
71 				uint32_t reserved:12;
72 			} hw_gen_event;
73 			struct hw_cache_event {
74 				uint32_t result_id:1;
75 				uint32_t op_id:2;
76 				uint32_t cache_id:13;
77 				uint32_t event_type:4;
78 				uint32_t reserved:12;
79 			} hw_cache_event;
80 		};
81 		uint32_t event_idx;
82 	};
83 };
84 
85 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
86 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
87 							SBI_PMU_HW_CPU_CYCLES,
88 							SBI_PMU_EVENT_TYPE_HW, 0}},
89 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
90 							SBI_PMU_HW_INSTRUCTIONS,
91 							SBI_PMU_EVENT_TYPE_HW, 0}},
92 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
93 							SBI_PMU_HW_CACHE_REFERENCES,
94 							SBI_PMU_EVENT_TYPE_HW, 0}},
95 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
96 							SBI_PMU_HW_CACHE_MISSES,
97 							SBI_PMU_EVENT_TYPE_HW, 0}},
98 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
99 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
100 							SBI_PMU_EVENT_TYPE_HW, 0}},
101 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
102 							SBI_PMU_HW_BRANCH_MISSES,
103 							SBI_PMU_EVENT_TYPE_HW, 0}},
104 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
105 							SBI_PMU_HW_BUS_CYCLES,
106 							SBI_PMU_EVENT_TYPE_HW, 0}},
107 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
108 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
109 							SBI_PMU_EVENT_TYPE_HW, 0}},
110 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
111 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
112 							SBI_PMU_EVENT_TYPE_HW, 0}},
113 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
114 							SBI_PMU_HW_REF_CPU_CYCLES,
115 							SBI_PMU_EVENT_TYPE_HW, 0}},
116 };
117 
118 #define C(x) PERF_COUNT_HW_CACHE_##x
119 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
120 [PERF_COUNT_HW_CACHE_OP_MAX]
121 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
122 	[C(L1D)] = {
123 		[C(OP_READ)] = {
124 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
125 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
126 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
127 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
128 		},
129 		[C(OP_WRITE)] = {
130 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
131 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
132 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
133 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
134 		},
135 		[C(OP_PREFETCH)] = {
136 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
137 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
138 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
139 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
140 		},
141 	},
142 	[C(L1I)] = {
143 		[C(OP_READ)] = {
144 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
145 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
146 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
147 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
148 		},
149 		[C(OP_WRITE)] = {
150 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
151 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
152 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
153 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
154 		},
155 		[C(OP_PREFETCH)] = {
156 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
157 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
158 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
159 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
160 		},
161 	},
162 	[C(LL)] = {
163 		[C(OP_READ)] = {
164 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
165 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
166 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
167 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
168 		},
169 		[C(OP_WRITE)] = {
170 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
171 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
172 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
173 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
174 		},
175 		[C(OP_PREFETCH)] = {
176 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
177 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
178 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
179 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
180 		},
181 	},
182 	[C(DTLB)] = {
183 		[C(OP_READ)] = {
184 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
185 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
186 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
187 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
188 		},
189 		[C(OP_WRITE)] = {
190 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
191 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
192 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
193 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
194 		},
195 		[C(OP_PREFETCH)] = {
196 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
197 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
198 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
199 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
200 		},
201 	},
202 	[C(ITLB)] = {
203 		[C(OP_READ)] = {
204 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
205 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
206 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
207 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
208 		},
209 		[C(OP_WRITE)] = {
210 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
211 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
212 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
213 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
214 		},
215 		[C(OP_PREFETCH)] = {
216 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
217 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
218 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
219 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
220 		},
221 	},
222 	[C(BPU)] = {
223 		[C(OP_READ)] = {
224 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
225 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
226 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
227 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
228 		},
229 		[C(OP_WRITE)] = {
230 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
231 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
232 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
233 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
234 		},
235 		[C(OP_PREFETCH)] = {
236 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
237 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
238 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
239 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
240 		},
241 	},
242 	[C(NODE)] = {
243 		[C(OP_READ)] = {
244 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
245 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
246 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
247 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
248 		},
249 		[C(OP_WRITE)] = {
250 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
251 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
252 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
253 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
254 		},
255 		[C(OP_PREFETCH)] = {
256 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
257 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
258 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
259 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
260 		},
261 	},
262 };
263 
264 static int pmu_sbi_ctr_get_width(int idx)
265 {
266 	return pmu_ctr_list[idx].width;
267 }
268 
269 static bool pmu_sbi_ctr_is_fw(int cidx)
270 {
271 	union sbi_pmu_ctr_info *info;
272 
273 	info = &pmu_ctr_list[cidx];
274 	if (!info)
275 		return false;
276 
277 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
278 }
279 
280 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
281 {
282 	struct hw_perf_event *hwc = &event->hw;
283 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
284 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
285 	struct sbiret ret;
286 	int idx;
287 	uint64_t cbase = 0;
288 	uint64_t cmask = GENMASK_ULL(rvpmu->num_counters - 1, 0);
289 	unsigned long cflags = 0;
290 
291 	if (event->attr.exclude_kernel)
292 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
293 	if (event->attr.exclude_user)
294 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH;
295 
296 	/* retrieve the available counter index */
297 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase, cmask,
298 			cflags, hwc->event_base, hwc->config, 0);
299 	if (ret.error) {
300 		pr_debug("Not able to find a counter for event %lx config %llx\n",
301 			hwc->event_base, hwc->config);
302 		return sbi_err_map_linux_errno(ret.error);
303 	}
304 
305 	idx = ret.value;
306 	if (idx >= rvpmu->num_counters || !pmu_ctr_list[idx].value)
307 		return -ENOENT;
308 
309 	/* Additional sanity check for the counter id */
310 	if (pmu_sbi_ctr_is_fw(idx)) {
311 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
312 			return idx;
313 	} else {
314 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
315 			return idx;
316 	}
317 
318 	return -ENOENT;
319 }
320 
321 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
322 {
323 
324 	struct hw_perf_event *hwc = &event->hw;
325 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
326 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
327 	int idx = hwc->idx;
328 
329 	if (pmu_sbi_ctr_is_fw(idx))
330 		clear_bit(idx, cpuc->used_fw_ctrs);
331 	else
332 		clear_bit(idx, cpuc->used_hw_ctrs);
333 }
334 
335 static int pmu_event_find_cache(u64 config)
336 {
337 	unsigned int cache_type, cache_op, cache_result, ret;
338 
339 	cache_type = (config >>  0) & 0xff;
340 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
341 		return -EINVAL;
342 
343 	cache_op = (config >>  8) & 0xff;
344 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
345 		return -EINVAL;
346 
347 	cache_result = (config >> 16) & 0xff;
348 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
349 		return -EINVAL;
350 
351 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
352 
353 	return ret;
354 }
355 
356 static bool pmu_sbi_is_fw_event(struct perf_event *event)
357 {
358 	u32 type = event->attr.type;
359 	u64 config = event->attr.config;
360 
361 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
362 		return true;
363 	else
364 		return false;
365 }
366 
367 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
368 {
369 	u32 type = event->attr.type;
370 	u64 config = event->attr.config;
371 	int bSoftware;
372 	u64 raw_config_val;
373 	int ret;
374 
375 	switch (type) {
376 	case PERF_TYPE_HARDWARE:
377 		if (config >= PERF_COUNT_HW_MAX)
378 			return -EINVAL;
379 		ret = pmu_hw_event_map[event->attr.config].event_idx;
380 		break;
381 	case PERF_TYPE_HW_CACHE:
382 		ret = pmu_event_find_cache(config);
383 		break;
384 	case PERF_TYPE_RAW:
385 		/*
386 		 * As per SBI specification, the upper 16 bits must be unused for
387 		 * a raw event. Use the MSB (63b) to distinguish between hardware
388 		 * raw event and firmware events.
389 		 */
390 		bSoftware = config >> 63;
391 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
392 		if (bSoftware) {
393 			if (raw_config_val < SBI_PMU_FW_MAX)
394 				ret = (raw_config_val & 0xFFFF) |
395 				      (SBI_PMU_EVENT_TYPE_FW << 16);
396 			else
397 				return -EINVAL;
398 		} else {
399 			ret = RISCV_PMU_RAW_EVENT_IDX;
400 			*econfig = raw_config_val;
401 		}
402 		break;
403 	default:
404 		ret = -EINVAL;
405 		break;
406 	}
407 
408 	return ret;
409 }
410 
411 static u64 pmu_sbi_ctr_read(struct perf_event *event)
412 {
413 	struct hw_perf_event *hwc = &event->hw;
414 	int idx = hwc->idx;
415 	struct sbiret ret;
416 	union sbi_pmu_ctr_info info;
417 	u64 val = 0;
418 
419 	if (pmu_sbi_is_fw_event(event)) {
420 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
421 				hwc->idx, 0, 0, 0, 0, 0);
422 		if (!ret.error)
423 			val = ret.value;
424 	} else {
425 		info = pmu_ctr_list[idx];
426 		val = riscv_pmu_ctr_read_csr(info.csr);
427 		if (IS_ENABLED(CONFIG_32BIT))
428 			val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
429 	}
430 
431 	return val;
432 }
433 
434 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
435 {
436 	struct sbiret ret;
437 	struct hw_perf_event *hwc = &event->hw;
438 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
439 
440 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
441 			1, flag, ival, ival >> 32, 0);
442 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
443 		pr_err("Starting counter idx %d failed with error %d\n",
444 			hwc->idx, sbi_err_map_linux_errno(ret.error));
445 }
446 
447 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
448 {
449 	struct sbiret ret;
450 	struct hw_perf_event *hwc = &event->hw;
451 
452 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
453 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
454 		flag != SBI_PMU_STOP_FLAG_RESET)
455 		pr_err("Stopping counter idx %d failed with error %d\n",
456 			hwc->idx, sbi_err_map_linux_errno(ret.error));
457 }
458 
459 static int pmu_sbi_find_num_ctrs(void)
460 {
461 	struct sbiret ret;
462 
463 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
464 	if (!ret.error)
465 		return ret.value;
466 	else
467 		return sbi_err_map_linux_errno(ret.error);
468 }
469 
470 static int pmu_sbi_get_ctrinfo(int nctr)
471 {
472 	struct sbiret ret;
473 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
474 	union sbi_pmu_ctr_info cinfo;
475 
476 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
477 	if (!pmu_ctr_list)
478 		return -ENOMEM;
479 
480 	for (i = 0; i <= nctr; i++) {
481 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
482 		if (ret.error)
483 			/* The logical counter ids are not expected to be contiguous */
484 			continue;
485 		cinfo.value = ret.value;
486 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
487 			num_fw_ctr++;
488 		else
489 			num_hw_ctr++;
490 		pmu_ctr_list[i].value = cinfo.value;
491 	}
492 
493 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
494 
495 	return 0;
496 }
497 
498 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
499 {
500 	/*
501 	 * No need to check the error because we are disabling all the counters
502 	 * which may include counters that are not enabled yet.
503 	 */
504 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
505 		  0, GENMASK_ULL(pmu->num_counters - 1, 0), 0, 0, 0, 0);
506 }
507 
508 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
509 {
510 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
511 
512 	/* No need to check the error here as we can't do anything about the error */
513 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
514 		  cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
515 }
516 
517 /*
518  * This function starts all the used counters in two step approach.
519  * Any counter that did not overflow can be start in a single step
520  * while the overflowed counters need to be started with updated initialization
521  * value.
522  */
523 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
524 					       unsigned long ctr_ovf_mask)
525 {
526 	int idx = 0;
527 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
528 	struct perf_event *event;
529 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
530 	unsigned long ctr_start_mask = 0;
531 	uint64_t max_period;
532 	struct hw_perf_event *hwc;
533 	u64 init_val = 0;
534 
535 	ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
536 
537 	/* Start all the counters that did not overflow in a single shot */
538 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
539 		  0, 0, 0, 0);
540 
541 	/* Reinitialize and start all the counter that overflowed */
542 	while (ctr_ovf_mask) {
543 		if (ctr_ovf_mask & 0x01) {
544 			event = cpu_hw_evt->events[idx];
545 			hwc = &event->hw;
546 			max_period = riscv_pmu_ctr_get_width_mask(event);
547 			init_val = local64_read(&hwc->prev_count) & max_period;
548 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
549 				  flag, init_val, 0, 0);
550 		}
551 		ctr_ovf_mask = ctr_ovf_mask >> 1;
552 		idx++;
553 	}
554 }
555 
556 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
557 {
558 	struct perf_sample_data data;
559 	struct pt_regs *regs;
560 	struct hw_perf_event *hw_evt;
561 	union sbi_pmu_ctr_info *info;
562 	int lidx, hidx, fidx;
563 	struct riscv_pmu *pmu;
564 	struct perf_event *event;
565 	unsigned long overflow;
566 	unsigned long overflowed_ctrs = 0;
567 	struct cpu_hw_events *cpu_hw_evt = dev;
568 
569 	if (WARN_ON_ONCE(!cpu_hw_evt))
570 		return IRQ_NONE;
571 
572 	/* Firmware counter don't support overflow yet */
573 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
574 	event = cpu_hw_evt->events[fidx];
575 	if (!event) {
576 		csr_clear(CSR_SIP, SIP_LCOFIP);
577 		return IRQ_NONE;
578 	}
579 
580 	pmu = to_riscv_pmu(event->pmu);
581 	pmu_sbi_stop_hw_ctrs(pmu);
582 
583 	/* Overflow status register should only be read after counter are stopped */
584 	overflow = csr_read(CSR_SSCOUNTOVF);
585 
586 	/*
587 	 * Overflow interrupt pending bit should only be cleared after stopping
588 	 * all the counters to avoid any race condition.
589 	 */
590 	csr_clear(CSR_SIP, SIP_LCOFIP);
591 
592 	/* No overflow bit is set */
593 	if (!overflow)
594 		return IRQ_NONE;
595 
596 	regs = get_irq_regs();
597 
598 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
599 		struct perf_event *event = cpu_hw_evt->events[lidx];
600 
601 		/* Skip if invalid event or user did not request a sampling */
602 		if (!event || !is_sampling_event(event))
603 			continue;
604 
605 		info = &pmu_ctr_list[lidx];
606 		/* Do a sanity check */
607 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
608 			continue;
609 
610 		/* compute hardware counter index */
611 		hidx = info->csr - CSR_CYCLE;
612 		/* check if the corresponding bit is set in sscountovf */
613 		if (!(overflow & (1 << hidx)))
614 			continue;
615 
616 		/*
617 		 * Keep a track of overflowed counters so that they can be started
618 		 * with updated initial value.
619 		 */
620 		overflowed_ctrs |= 1 << lidx;
621 		hw_evt = &event->hw;
622 		riscv_pmu_event_update(event);
623 		perf_sample_data_init(&data, 0, hw_evt->last_period);
624 		if (riscv_pmu_event_set_period(event)) {
625 			/*
626 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
627 			 * to avoid throttling here. As per the specification, the
628 			 * interrupt remains disabled until the OF bit is set.
629 			 * Interrupts are enabled again only during the start.
630 			 * TODO: We will need to stop the guest counters once
631 			 * virtualization support is added.
632 			 */
633 			perf_event_overflow(event, &data, regs);
634 		}
635 	}
636 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
637 
638 	return IRQ_HANDLED;
639 }
640 
641 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
642 {
643 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
644 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
645 
646 	/* Enable the access for TIME csr only from the user mode now */
647 	csr_write(CSR_SCOUNTEREN, 0x2);
648 
649 	/* Stop all the counters so that they can be enabled from perf */
650 	pmu_sbi_stop_all(pmu);
651 
652 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
653 		cpu_hw_evt->irq = riscv_pmu_irq;
654 		csr_clear(CSR_IP, BIT(RV_IRQ_PMU));
655 		csr_set(CSR_IE, BIT(RV_IRQ_PMU));
656 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
657 	}
658 
659 	return 0;
660 }
661 
662 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
663 {
664 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
665 		disable_percpu_irq(riscv_pmu_irq);
666 		csr_clear(CSR_IE, BIT(RV_IRQ_PMU));
667 	}
668 
669 	/* Disable all counters access for user mode now */
670 	csr_write(CSR_SCOUNTEREN, 0x0);
671 
672 	return 0;
673 }
674 
675 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
676 {
677 	int ret;
678 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
679 	struct device_node *cpu, *child;
680 	struct irq_domain *domain = NULL;
681 
682 	if (!riscv_isa_extension_available(NULL, SSCOFPMF))
683 		return -EOPNOTSUPP;
684 
685 	for_each_of_cpu_node(cpu) {
686 		child = of_get_compatible_child(cpu, "riscv,cpu-intc");
687 		if (!child) {
688 			pr_err("Failed to find INTC node\n");
689 			of_node_put(cpu);
690 			return -ENODEV;
691 		}
692 		domain = irq_find_host(child);
693 		of_node_put(child);
694 		if (domain) {
695 			of_node_put(cpu);
696 			break;
697 		}
698 	}
699 	if (!domain) {
700 		pr_err("Failed to find INTC IRQ root domain\n");
701 		return -ENODEV;
702 	}
703 
704 	riscv_pmu_irq = irq_create_mapping(domain, RV_IRQ_PMU);
705 	if (!riscv_pmu_irq) {
706 		pr_err("Failed to map PMU interrupt for node\n");
707 		return -ENODEV;
708 	}
709 
710 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
711 	if (ret) {
712 		pr_err("registering percpu irq failed [%d]\n", ret);
713 		return ret;
714 	}
715 
716 	return 0;
717 }
718 
719 #ifdef CONFIG_CPU_PM
720 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
721 				void *v)
722 {
723 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
724 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
725 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
726 	struct perf_event *event;
727 	int idx;
728 
729 	if (!enabled)
730 		return NOTIFY_OK;
731 
732 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
733 		event = cpuc->events[idx];
734 		if (!event)
735 			continue;
736 
737 		switch (cmd) {
738 		case CPU_PM_ENTER:
739 			/*
740 			 * Stop and update the counter
741 			 */
742 			riscv_pmu_stop(event, PERF_EF_UPDATE);
743 			break;
744 		case CPU_PM_EXIT:
745 		case CPU_PM_ENTER_FAILED:
746 			/*
747 			 * Restore and enable the counter.
748 			 *
749 			 * Requires RCU read locking to be functional,
750 			 * wrap the call within RCU_NONIDLE to make the
751 			 * RCU subsystem aware this cpu is not idle from
752 			 * an RCU perspective for the riscv_pmu_start() call
753 			 * duration.
754 			 */
755 			RCU_NONIDLE(riscv_pmu_start(event, PERF_EF_RELOAD));
756 			break;
757 		default:
758 			break;
759 		}
760 	}
761 
762 	return NOTIFY_OK;
763 }
764 
765 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
766 {
767 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
768 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
769 }
770 
771 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
772 {
773 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
774 }
775 #else
776 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
777 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
778 #endif
779 
780 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
781 {
782 	riscv_pm_pmu_unregister(pmu);
783 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
784 }
785 
786 static int pmu_sbi_device_probe(struct platform_device *pdev)
787 {
788 	struct riscv_pmu *pmu = NULL;
789 	int num_counters;
790 	int ret = -ENODEV;
791 
792 	pr_info("SBI PMU extension is available\n");
793 	pmu = riscv_pmu_alloc();
794 	if (!pmu)
795 		return -ENOMEM;
796 
797 	num_counters = pmu_sbi_find_num_ctrs();
798 	if (num_counters < 0) {
799 		pr_err("SBI PMU extension doesn't provide any counters\n");
800 		goto out_free;
801 	}
802 
803 	/* cache all the information about counters now */
804 	if (pmu_sbi_get_ctrinfo(num_counters))
805 		goto out_free;
806 
807 	ret = pmu_sbi_setup_irqs(pmu, pdev);
808 	if (ret < 0) {
809 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
810 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
811 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
812 	}
813 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
814 	pmu->num_counters = num_counters;
815 	pmu->ctr_start = pmu_sbi_ctr_start;
816 	pmu->ctr_stop = pmu_sbi_ctr_stop;
817 	pmu->event_map = pmu_sbi_event_map;
818 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
819 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
820 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
821 	pmu->ctr_read = pmu_sbi_ctr_read;
822 
823 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
824 	if (ret)
825 		return ret;
826 
827 	ret = riscv_pm_pmu_register(pmu);
828 	if (ret)
829 		goto out_unregister;
830 
831 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
832 	if (ret)
833 		goto out_unregister;
834 
835 	return 0;
836 
837 out_unregister:
838 	riscv_pmu_destroy(pmu);
839 
840 out_free:
841 	kfree(pmu);
842 	return ret;
843 }
844 
845 static struct platform_driver pmu_sbi_driver = {
846 	.probe		= pmu_sbi_device_probe,
847 	.driver		= {
848 		.name	= RISCV_PMU_PDEV_NAME,
849 	},
850 };
851 
852 static int __init pmu_sbi_devinit(void)
853 {
854 	int ret;
855 	struct platform_device *pdev;
856 
857 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
858 	    sbi_probe_extension(SBI_EXT_PMU) <= 0) {
859 		return 0;
860 	}
861 
862 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
863 				      "perf/riscv/pmu:starting",
864 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
865 	if (ret) {
866 		pr_err("CPU hotplug notifier could not be registered: %d\n",
867 		       ret);
868 		return ret;
869 	}
870 
871 	ret = platform_driver_register(&pmu_sbi_driver);
872 	if (ret)
873 		return ret;
874 
875 	pdev = platform_device_register_simple(RISCV_PMU_PDEV_NAME, -1, NULL, 0);
876 	if (IS_ERR(pdev)) {
877 		platform_driver_unregister(&pmu_sbi_driver);
878 		return PTR_ERR(pdev);
879 	}
880 
881 	/* Notify legacy implementation that SBI pmu is available*/
882 	riscv_pmu_legacy_skip_init();
883 
884 	return ret;
885 }
886 device_initcall(pmu_sbi_devinit)
887