xref: /openbmc/linux/drivers/perf/riscv_pmu_sbi.c (revision 519b58bb)
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/sbi.h>
24 #include <asm/hwcap.h>
25 
26 PMU_FORMAT_ATTR(event, "config:0-47");
27 PMU_FORMAT_ATTR(firmware, "config:63");
28 
29 static struct attribute *riscv_arch_formats_attr[] = {
30 	&format_attr_event.attr,
31 	&format_attr_firmware.attr,
32 	NULL,
33 };
34 
35 static struct attribute_group riscv_pmu_format_group = {
36 	.name = "format",
37 	.attrs = riscv_arch_formats_attr,
38 };
39 
40 static const struct attribute_group *riscv_pmu_attr_groups[] = {
41 	&riscv_pmu_format_group,
42 	NULL,
43 };
44 
45 /*
46  * RISC-V doesn't have hetergenous harts yet. This need to be part of
47  * per_cpu in case of harts with different pmu counters
48  */
49 static union sbi_pmu_ctr_info *pmu_ctr_list;
50 static unsigned int riscv_pmu_irq;
51 
52 struct sbi_pmu_event_data {
53 	union {
54 		union {
55 			struct hw_gen_event {
56 				uint32_t event_code:16;
57 				uint32_t event_type:4;
58 				uint32_t reserved:12;
59 			} hw_gen_event;
60 			struct hw_cache_event {
61 				uint32_t result_id:1;
62 				uint32_t op_id:2;
63 				uint32_t cache_id:13;
64 				uint32_t event_type:4;
65 				uint32_t reserved:12;
66 			} hw_cache_event;
67 		};
68 		uint32_t event_idx;
69 	};
70 };
71 
72 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
73 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
74 							SBI_PMU_HW_CPU_CYCLES,
75 							SBI_PMU_EVENT_TYPE_HW, 0}},
76 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
77 							SBI_PMU_HW_INSTRUCTIONS,
78 							SBI_PMU_EVENT_TYPE_HW, 0}},
79 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
80 							SBI_PMU_HW_CACHE_REFERENCES,
81 							SBI_PMU_EVENT_TYPE_HW, 0}},
82 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
83 							SBI_PMU_HW_CACHE_MISSES,
84 							SBI_PMU_EVENT_TYPE_HW, 0}},
85 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
86 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
87 							SBI_PMU_EVENT_TYPE_HW, 0}},
88 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
89 							SBI_PMU_HW_BRANCH_MISSES,
90 							SBI_PMU_EVENT_TYPE_HW, 0}},
91 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
92 							SBI_PMU_HW_BUS_CYCLES,
93 							SBI_PMU_EVENT_TYPE_HW, 0}},
94 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
95 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
96 							SBI_PMU_EVENT_TYPE_HW, 0}},
97 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
98 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
99 							SBI_PMU_EVENT_TYPE_HW, 0}},
100 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
101 							SBI_PMU_HW_REF_CPU_CYCLES,
102 							SBI_PMU_EVENT_TYPE_HW, 0}},
103 };
104 
105 #define C(x) PERF_COUNT_HW_CACHE_##x
106 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
107 [PERF_COUNT_HW_CACHE_OP_MAX]
108 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
109 	[C(L1D)] = {
110 		[C(OP_READ)] = {
111 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
112 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
113 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
114 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
115 		},
116 		[C(OP_WRITE)] = {
117 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
118 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
119 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
120 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
121 		},
122 		[C(OP_PREFETCH)] = {
123 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
124 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
125 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
126 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
127 		},
128 	},
129 	[C(L1I)] = {
130 		[C(OP_READ)] = {
131 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
132 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
133 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
134 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
135 		},
136 		[C(OP_WRITE)] = {
137 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
138 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
139 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
140 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
141 		},
142 		[C(OP_PREFETCH)] = {
143 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
144 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
145 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
146 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
147 		},
148 	},
149 	[C(LL)] = {
150 		[C(OP_READ)] = {
151 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
152 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
153 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
154 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
155 		},
156 		[C(OP_WRITE)] = {
157 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
158 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
159 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
160 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
161 		},
162 		[C(OP_PREFETCH)] = {
163 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
164 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
165 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
166 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
167 		},
168 	},
169 	[C(DTLB)] = {
170 		[C(OP_READ)] = {
171 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
172 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
173 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
174 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
175 		},
176 		[C(OP_WRITE)] = {
177 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
178 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
179 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
180 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
181 		},
182 		[C(OP_PREFETCH)] = {
183 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
184 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
185 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
186 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
187 		},
188 	},
189 	[C(ITLB)] = {
190 		[C(OP_READ)] = {
191 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
192 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
193 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
194 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
195 		},
196 		[C(OP_WRITE)] = {
197 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
198 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
199 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
200 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
201 		},
202 		[C(OP_PREFETCH)] = {
203 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
204 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
205 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
206 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
207 		},
208 	},
209 	[C(BPU)] = {
210 		[C(OP_READ)] = {
211 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
212 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
213 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
214 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
215 		},
216 		[C(OP_WRITE)] = {
217 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
218 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
219 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
220 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
221 		},
222 		[C(OP_PREFETCH)] = {
223 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
224 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
225 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
226 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
227 		},
228 	},
229 	[C(NODE)] = {
230 		[C(OP_READ)] = {
231 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
232 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
233 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
234 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
235 		},
236 		[C(OP_WRITE)] = {
237 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
238 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
239 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
240 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
241 		},
242 		[C(OP_PREFETCH)] = {
243 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
244 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
245 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
246 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
247 		},
248 	},
249 };
250 
251 static int pmu_sbi_ctr_get_width(int idx)
252 {
253 	return pmu_ctr_list[idx].width;
254 }
255 
256 static bool pmu_sbi_ctr_is_fw(int cidx)
257 {
258 	union sbi_pmu_ctr_info *info;
259 
260 	info = &pmu_ctr_list[cidx];
261 	if (!info)
262 		return false;
263 
264 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
265 }
266 
267 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
268 {
269 	struct hw_perf_event *hwc = &event->hw;
270 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
271 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
272 	struct sbiret ret;
273 	int idx;
274 	uint64_t cbase = 0;
275 	unsigned long cflags = 0;
276 
277 	if (event->attr.exclude_kernel)
278 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
279 	if (event->attr.exclude_user)
280 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH;
281 
282 	/* retrieve the available counter index */
283 #if defined(CONFIG_32BIT)
284 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
285 			rvpmu->cmask, cflags, hwc->event_base, hwc->config,
286 			hwc->config >> 32);
287 #else
288 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
289 			rvpmu->cmask, cflags, hwc->event_base, hwc->config, 0);
290 #endif
291 	if (ret.error) {
292 		pr_debug("Not able to find a counter for event %lx config %llx\n",
293 			hwc->event_base, hwc->config);
294 		return sbi_err_map_linux_errno(ret.error);
295 	}
296 
297 	idx = ret.value;
298 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
299 		return -ENOENT;
300 
301 	/* Additional sanity check for the counter id */
302 	if (pmu_sbi_ctr_is_fw(idx)) {
303 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
304 			return idx;
305 	} else {
306 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
307 			return idx;
308 	}
309 
310 	return -ENOENT;
311 }
312 
313 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
314 {
315 
316 	struct hw_perf_event *hwc = &event->hw;
317 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
318 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
319 	int idx = hwc->idx;
320 
321 	if (pmu_sbi_ctr_is_fw(idx))
322 		clear_bit(idx, cpuc->used_fw_ctrs);
323 	else
324 		clear_bit(idx, cpuc->used_hw_ctrs);
325 }
326 
327 static int pmu_event_find_cache(u64 config)
328 {
329 	unsigned int cache_type, cache_op, cache_result, ret;
330 
331 	cache_type = (config >>  0) & 0xff;
332 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
333 		return -EINVAL;
334 
335 	cache_op = (config >>  8) & 0xff;
336 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
337 		return -EINVAL;
338 
339 	cache_result = (config >> 16) & 0xff;
340 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
341 		return -EINVAL;
342 
343 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
344 
345 	return ret;
346 }
347 
348 static bool pmu_sbi_is_fw_event(struct perf_event *event)
349 {
350 	u32 type = event->attr.type;
351 	u64 config = event->attr.config;
352 
353 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
354 		return true;
355 	else
356 		return false;
357 }
358 
359 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
360 {
361 	u32 type = event->attr.type;
362 	u64 config = event->attr.config;
363 	int bSoftware;
364 	u64 raw_config_val;
365 	int ret;
366 
367 	switch (type) {
368 	case PERF_TYPE_HARDWARE:
369 		if (config >= PERF_COUNT_HW_MAX)
370 			return -EINVAL;
371 		ret = pmu_hw_event_map[event->attr.config].event_idx;
372 		break;
373 	case PERF_TYPE_HW_CACHE:
374 		ret = pmu_event_find_cache(config);
375 		break;
376 	case PERF_TYPE_RAW:
377 		/*
378 		 * As per SBI specification, the upper 16 bits must be unused for
379 		 * a raw event. Use the MSB (63b) to distinguish between hardware
380 		 * raw event and firmware events.
381 		 */
382 		bSoftware = config >> 63;
383 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
384 		if (bSoftware) {
385 			if (raw_config_val < SBI_PMU_FW_MAX)
386 				ret = (raw_config_val & 0xFFFF) |
387 				      (SBI_PMU_EVENT_TYPE_FW << 16);
388 			else
389 				return -EINVAL;
390 		} else {
391 			ret = RISCV_PMU_RAW_EVENT_IDX;
392 			*econfig = raw_config_val;
393 		}
394 		break;
395 	default:
396 		ret = -EINVAL;
397 		break;
398 	}
399 
400 	return ret;
401 }
402 
403 static u64 pmu_sbi_ctr_read(struct perf_event *event)
404 {
405 	struct hw_perf_event *hwc = &event->hw;
406 	int idx = hwc->idx;
407 	struct sbiret ret;
408 	union sbi_pmu_ctr_info info;
409 	u64 val = 0;
410 
411 	if (pmu_sbi_is_fw_event(event)) {
412 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
413 				hwc->idx, 0, 0, 0, 0, 0);
414 		if (!ret.error)
415 			val = ret.value;
416 	} else {
417 		info = pmu_ctr_list[idx];
418 		val = riscv_pmu_ctr_read_csr(info.csr);
419 		if (IS_ENABLED(CONFIG_32BIT))
420 			val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
421 	}
422 
423 	return val;
424 }
425 
426 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
427 {
428 	struct sbiret ret;
429 	struct hw_perf_event *hwc = &event->hw;
430 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
431 
432 #if defined(CONFIG_32BIT)
433 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
434 			1, flag, ival, ival >> 32, 0);
435 #else
436 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
437 			1, flag, ival, 0, 0);
438 #endif
439 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
440 		pr_err("Starting counter idx %d failed with error %d\n",
441 			hwc->idx, sbi_err_map_linux_errno(ret.error));
442 }
443 
444 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
445 {
446 	struct sbiret ret;
447 	struct hw_perf_event *hwc = &event->hw;
448 
449 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
450 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
451 		flag != SBI_PMU_STOP_FLAG_RESET)
452 		pr_err("Stopping counter idx %d failed with error %d\n",
453 			hwc->idx, sbi_err_map_linux_errno(ret.error));
454 }
455 
456 static int pmu_sbi_find_num_ctrs(void)
457 {
458 	struct sbiret ret;
459 
460 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
461 	if (!ret.error)
462 		return ret.value;
463 	else
464 		return sbi_err_map_linux_errno(ret.error);
465 }
466 
467 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
468 {
469 	struct sbiret ret;
470 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
471 	union sbi_pmu_ctr_info cinfo;
472 
473 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
474 	if (!pmu_ctr_list)
475 		return -ENOMEM;
476 
477 	for (i = 0; i < nctr; i++) {
478 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
479 		if (ret.error)
480 			/* The logical counter ids are not expected to be contiguous */
481 			continue;
482 
483 		*mask |= BIT(i);
484 
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, pmu->cmask, 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 #if defined(CONFIG_32BIT)
549 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
550 				  flag, init_val, init_val >> 32, 0);
551 #else
552 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
553 				  flag, init_val, 0, 0);
554 #endif
555 			perf_event_update_userpage(event);
556 		}
557 		ctr_ovf_mask = ctr_ovf_mask >> 1;
558 		idx++;
559 	}
560 }
561 
562 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
563 {
564 	struct perf_sample_data data;
565 	struct pt_regs *regs;
566 	struct hw_perf_event *hw_evt;
567 	union sbi_pmu_ctr_info *info;
568 	int lidx, hidx, fidx;
569 	struct riscv_pmu *pmu;
570 	struct perf_event *event;
571 	unsigned long overflow;
572 	unsigned long overflowed_ctrs = 0;
573 	struct cpu_hw_events *cpu_hw_evt = dev;
574 	u64 start_clock = sched_clock();
575 
576 	if (WARN_ON_ONCE(!cpu_hw_evt))
577 		return IRQ_NONE;
578 
579 	/* Firmware counter don't support overflow yet */
580 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
581 	event = cpu_hw_evt->events[fidx];
582 	if (!event) {
583 		csr_clear(CSR_SIP, SIP_LCOFIP);
584 		return IRQ_NONE;
585 	}
586 
587 	pmu = to_riscv_pmu(event->pmu);
588 	pmu_sbi_stop_hw_ctrs(pmu);
589 
590 	/* Overflow status register should only be read after counter are stopped */
591 	overflow = csr_read(CSR_SSCOUNTOVF);
592 
593 	/*
594 	 * Overflow interrupt pending bit should only be cleared after stopping
595 	 * all the counters to avoid any race condition.
596 	 */
597 	csr_clear(CSR_SIP, SIP_LCOFIP);
598 
599 	/* No overflow bit is set */
600 	if (!overflow)
601 		return IRQ_NONE;
602 
603 	regs = get_irq_regs();
604 
605 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
606 		struct perf_event *event = cpu_hw_evt->events[lidx];
607 
608 		/* Skip if invalid event or user did not request a sampling */
609 		if (!event || !is_sampling_event(event))
610 			continue;
611 
612 		info = &pmu_ctr_list[lidx];
613 		/* Do a sanity check */
614 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
615 			continue;
616 
617 		/* compute hardware counter index */
618 		hidx = info->csr - CSR_CYCLE;
619 		/* check if the corresponding bit is set in sscountovf */
620 		if (!(overflow & (1 << hidx)))
621 			continue;
622 
623 		/*
624 		 * Keep a track of overflowed counters so that they can be started
625 		 * with updated initial value.
626 		 */
627 		overflowed_ctrs |= 1 << lidx;
628 		hw_evt = &event->hw;
629 		riscv_pmu_event_update(event);
630 		perf_sample_data_init(&data, 0, hw_evt->last_period);
631 		if (riscv_pmu_event_set_period(event)) {
632 			/*
633 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
634 			 * to avoid throttling here. As per the specification, the
635 			 * interrupt remains disabled until the OF bit is set.
636 			 * Interrupts are enabled again only during the start.
637 			 * TODO: We will need to stop the guest counters once
638 			 * virtualization support is added.
639 			 */
640 			perf_event_overflow(event, &data, regs);
641 		}
642 	}
643 
644 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
645 	perf_sample_event_took(sched_clock() - start_clock);
646 
647 	return IRQ_HANDLED;
648 }
649 
650 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
651 {
652 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
653 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
654 
655 	/*
656 	 * Enable the access for CYCLE, TIME, and INSTRET CSRs from userspace,
657 	 * as is necessary to maintain uABI compatibility.
658 	 */
659 	csr_write(CSR_SCOUNTEREN, 0x7);
660 
661 	/* Stop all the counters so that they can be enabled from perf */
662 	pmu_sbi_stop_all(pmu);
663 
664 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
665 		cpu_hw_evt->irq = riscv_pmu_irq;
666 		csr_clear(CSR_IP, BIT(RV_IRQ_PMU));
667 		csr_set(CSR_IE, BIT(RV_IRQ_PMU));
668 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
669 	}
670 
671 	return 0;
672 }
673 
674 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
675 {
676 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
677 		disable_percpu_irq(riscv_pmu_irq);
678 		csr_clear(CSR_IE, BIT(RV_IRQ_PMU));
679 	}
680 
681 	/* Disable all counters access for user mode now */
682 	csr_write(CSR_SCOUNTEREN, 0x0);
683 
684 	return 0;
685 }
686 
687 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
688 {
689 	int ret;
690 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
691 	struct device_node *cpu, *child;
692 	struct irq_domain *domain = NULL;
693 
694 	if (!riscv_isa_extension_available(NULL, SSCOFPMF))
695 		return -EOPNOTSUPP;
696 
697 	for_each_of_cpu_node(cpu) {
698 		child = of_get_compatible_child(cpu, "riscv,cpu-intc");
699 		if (!child) {
700 			pr_err("Failed to find INTC node\n");
701 			of_node_put(cpu);
702 			return -ENODEV;
703 		}
704 		domain = irq_find_host(child);
705 		of_node_put(child);
706 		if (domain) {
707 			of_node_put(cpu);
708 			break;
709 		}
710 	}
711 	if (!domain) {
712 		pr_err("Failed to find INTC IRQ root domain\n");
713 		return -ENODEV;
714 	}
715 
716 	riscv_pmu_irq = irq_create_mapping(domain, RV_IRQ_PMU);
717 	if (!riscv_pmu_irq) {
718 		pr_err("Failed to map PMU interrupt for node\n");
719 		return -ENODEV;
720 	}
721 
722 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
723 	if (ret) {
724 		pr_err("registering percpu irq failed [%d]\n", ret);
725 		return ret;
726 	}
727 
728 	return 0;
729 }
730 
731 #ifdef CONFIG_CPU_PM
732 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
733 				void *v)
734 {
735 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
736 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
737 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
738 	struct perf_event *event;
739 	int idx;
740 
741 	if (!enabled)
742 		return NOTIFY_OK;
743 
744 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
745 		event = cpuc->events[idx];
746 		if (!event)
747 			continue;
748 
749 		switch (cmd) {
750 		case CPU_PM_ENTER:
751 			/*
752 			 * Stop and update the counter
753 			 */
754 			riscv_pmu_stop(event, PERF_EF_UPDATE);
755 			break;
756 		case CPU_PM_EXIT:
757 		case CPU_PM_ENTER_FAILED:
758 			/*
759 			 * Restore and enable the counter.
760 			 *
761 			 * Requires RCU read locking to be functional,
762 			 * wrap the call within RCU_NONIDLE to make the
763 			 * RCU subsystem aware this cpu is not idle from
764 			 * an RCU perspective for the riscv_pmu_start() call
765 			 * duration.
766 			 */
767 			RCU_NONIDLE(riscv_pmu_start(event, PERF_EF_RELOAD));
768 			break;
769 		default:
770 			break;
771 		}
772 	}
773 
774 	return NOTIFY_OK;
775 }
776 
777 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
778 {
779 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
780 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
781 }
782 
783 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
784 {
785 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
786 }
787 #else
788 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
789 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
790 #endif
791 
792 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
793 {
794 	riscv_pm_pmu_unregister(pmu);
795 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
796 }
797 
798 static int pmu_sbi_device_probe(struct platform_device *pdev)
799 {
800 	struct riscv_pmu *pmu = NULL;
801 	unsigned long cmask = 0;
802 	int ret = -ENODEV;
803 	int num_counters;
804 
805 	pr_info("SBI PMU extension is available\n");
806 	pmu = riscv_pmu_alloc();
807 	if (!pmu)
808 		return -ENOMEM;
809 
810 	num_counters = pmu_sbi_find_num_ctrs();
811 	if (num_counters < 0) {
812 		pr_err("SBI PMU extension doesn't provide any counters\n");
813 		goto out_free;
814 	}
815 
816 	/* cache all the information about counters now */
817 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
818 		goto out_free;
819 
820 	ret = pmu_sbi_setup_irqs(pmu, pdev);
821 	if (ret < 0) {
822 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
823 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
824 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
825 	}
826 
827 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
828 	pmu->cmask = cmask;
829 	pmu->ctr_start = pmu_sbi_ctr_start;
830 	pmu->ctr_stop = pmu_sbi_ctr_stop;
831 	pmu->event_map = pmu_sbi_event_map;
832 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
833 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
834 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
835 	pmu->ctr_read = pmu_sbi_ctr_read;
836 
837 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
838 	if (ret)
839 		return ret;
840 
841 	ret = riscv_pm_pmu_register(pmu);
842 	if (ret)
843 		goto out_unregister;
844 
845 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
846 	if (ret)
847 		goto out_unregister;
848 
849 	return 0;
850 
851 out_unregister:
852 	riscv_pmu_destroy(pmu);
853 
854 out_free:
855 	kfree(pmu);
856 	return ret;
857 }
858 
859 static struct platform_driver pmu_sbi_driver = {
860 	.probe		= pmu_sbi_device_probe,
861 	.driver		= {
862 		.name	= RISCV_PMU_PDEV_NAME,
863 	},
864 };
865 
866 static int __init pmu_sbi_devinit(void)
867 {
868 	int ret;
869 	struct platform_device *pdev;
870 
871 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
872 	    sbi_probe_extension(SBI_EXT_PMU) <= 0) {
873 		return 0;
874 	}
875 
876 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
877 				      "perf/riscv/pmu:starting",
878 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
879 	if (ret) {
880 		pr_err("CPU hotplug notifier could not be registered: %d\n",
881 		       ret);
882 		return ret;
883 	}
884 
885 	ret = platform_driver_register(&pmu_sbi_driver);
886 	if (ret)
887 		return ret;
888 
889 	pdev = platform_device_register_simple(RISCV_PMU_PDEV_NAME, -1, NULL, 0);
890 	if (IS_ERR(pdev)) {
891 		platform_driver_unregister(&pmu_sbi_driver);
892 		return PTR_ERR(pdev);
893 	}
894 
895 	/* Notify legacy implementation that SBI pmu is available*/
896 	riscv_pmu_legacy_skip_init();
897 
898 	return ret;
899 }
900 device_initcall(pmu_sbi_devinit)
901