xref: /openbmc/linux/drivers/perf/riscv_pmu_sbi.c (revision 27e45f2e)
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 PMU_FORMAT_ATTR(event, "config:0-47");
28 PMU_FORMAT_ATTR(firmware, "config:63");
29 
30 static struct attribute *riscv_arch_formats_attr[] = {
31 	&format_attr_event.attr,
32 	&format_attr_firmware.attr,
33 	NULL,
34 };
35 
36 static struct attribute_group riscv_pmu_format_group = {
37 	.name = "format",
38 	.attrs = riscv_arch_formats_attr,
39 };
40 
41 static const struct attribute_group *riscv_pmu_attr_groups[] = {
42 	&riscv_pmu_format_group,
43 	NULL,
44 };
45 
46 /*
47  * RISC-V doesn't have hetergenous harts yet. This need to be part of
48  * per_cpu in case of harts with different pmu counters
49  */
50 static union sbi_pmu_ctr_info *pmu_ctr_list;
51 static bool riscv_pmu_use_irq;
52 static unsigned int riscv_pmu_irq_num;
53 static unsigned int riscv_pmu_irq;
54 
55 struct sbi_pmu_event_data {
56 	union {
57 		union {
58 			struct hw_gen_event {
59 				uint32_t event_code:16;
60 				uint32_t event_type:4;
61 				uint32_t reserved:12;
62 			} hw_gen_event;
63 			struct hw_cache_event {
64 				uint32_t result_id:1;
65 				uint32_t op_id:2;
66 				uint32_t cache_id:13;
67 				uint32_t event_type:4;
68 				uint32_t reserved:12;
69 			} hw_cache_event;
70 		};
71 		uint32_t event_idx;
72 	};
73 };
74 
75 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
76 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
77 							SBI_PMU_HW_CPU_CYCLES,
78 							SBI_PMU_EVENT_TYPE_HW, 0}},
79 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
80 							SBI_PMU_HW_INSTRUCTIONS,
81 							SBI_PMU_EVENT_TYPE_HW, 0}},
82 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
83 							SBI_PMU_HW_CACHE_REFERENCES,
84 							SBI_PMU_EVENT_TYPE_HW, 0}},
85 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
86 							SBI_PMU_HW_CACHE_MISSES,
87 							SBI_PMU_EVENT_TYPE_HW, 0}},
88 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
89 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
90 							SBI_PMU_EVENT_TYPE_HW, 0}},
91 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
92 							SBI_PMU_HW_BRANCH_MISSES,
93 							SBI_PMU_EVENT_TYPE_HW, 0}},
94 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
95 							SBI_PMU_HW_BUS_CYCLES,
96 							SBI_PMU_EVENT_TYPE_HW, 0}},
97 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
98 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
99 							SBI_PMU_EVENT_TYPE_HW, 0}},
100 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
101 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
102 							SBI_PMU_EVENT_TYPE_HW, 0}},
103 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
104 							SBI_PMU_HW_REF_CPU_CYCLES,
105 							SBI_PMU_EVENT_TYPE_HW, 0}},
106 };
107 
108 #define C(x) PERF_COUNT_HW_CACHE_##x
109 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
110 [PERF_COUNT_HW_CACHE_OP_MAX]
111 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
112 	[C(L1D)] = {
113 		[C(OP_READ)] = {
114 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
115 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
116 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
117 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
118 		},
119 		[C(OP_WRITE)] = {
120 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
121 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
122 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
123 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
124 		},
125 		[C(OP_PREFETCH)] = {
126 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
127 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
128 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
129 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
130 		},
131 	},
132 	[C(L1I)] = {
133 		[C(OP_READ)] = {
134 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
135 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
136 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
137 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
138 		},
139 		[C(OP_WRITE)] = {
140 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
141 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
142 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
143 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
144 		},
145 		[C(OP_PREFETCH)] = {
146 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
147 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
148 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
149 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
150 		},
151 	},
152 	[C(LL)] = {
153 		[C(OP_READ)] = {
154 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
155 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
156 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
157 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
158 		},
159 		[C(OP_WRITE)] = {
160 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
161 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
162 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
163 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
164 		},
165 		[C(OP_PREFETCH)] = {
166 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
167 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
168 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
169 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
170 		},
171 	},
172 	[C(DTLB)] = {
173 		[C(OP_READ)] = {
174 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
175 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
176 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
177 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
178 		},
179 		[C(OP_WRITE)] = {
180 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
181 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
182 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
183 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
184 		},
185 		[C(OP_PREFETCH)] = {
186 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
187 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
188 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
189 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
190 		},
191 	},
192 	[C(ITLB)] = {
193 		[C(OP_READ)] = {
194 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
195 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
196 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
197 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
198 		},
199 		[C(OP_WRITE)] = {
200 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
201 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
202 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
203 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
204 		},
205 		[C(OP_PREFETCH)] = {
206 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
207 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
208 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
209 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
210 		},
211 	},
212 	[C(BPU)] = {
213 		[C(OP_READ)] = {
214 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
215 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
216 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
217 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
218 		},
219 		[C(OP_WRITE)] = {
220 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
221 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
222 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
223 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
224 		},
225 		[C(OP_PREFETCH)] = {
226 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
227 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
228 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
229 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
230 		},
231 	},
232 	[C(NODE)] = {
233 		[C(OP_READ)] = {
234 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
235 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
236 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
237 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
238 		},
239 		[C(OP_WRITE)] = {
240 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
241 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
242 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
243 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
244 		},
245 		[C(OP_PREFETCH)] = {
246 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
247 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
248 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
249 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
250 		},
251 	},
252 };
253 
254 static int pmu_sbi_ctr_get_width(int idx)
255 {
256 	return pmu_ctr_list[idx].width;
257 }
258 
259 static bool pmu_sbi_ctr_is_fw(int cidx)
260 {
261 	union sbi_pmu_ctr_info *info;
262 
263 	info = &pmu_ctr_list[cidx];
264 	if (!info)
265 		return false;
266 
267 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
268 }
269 
270 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
271 {
272 	struct hw_perf_event *hwc = &event->hw;
273 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
274 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
275 	struct sbiret ret;
276 	int idx;
277 	uint64_t cbase = 0;
278 	unsigned long cflags = 0;
279 
280 	if (event->attr.exclude_kernel)
281 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
282 	if (event->attr.exclude_user)
283 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH;
284 
285 	/* retrieve the available counter index */
286 #if defined(CONFIG_32BIT)
287 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
288 			rvpmu->cmask, cflags, hwc->event_base, hwc->config,
289 			hwc->config >> 32);
290 #else
291 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
292 			rvpmu->cmask, cflags, hwc->event_base, hwc->config, 0);
293 #endif
294 	if (ret.error) {
295 		pr_debug("Not able to find a counter for event %lx config %llx\n",
296 			hwc->event_base, hwc->config);
297 		return sbi_err_map_linux_errno(ret.error);
298 	}
299 
300 	idx = ret.value;
301 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
302 		return -ENOENT;
303 
304 	/* Additional sanity check for the counter id */
305 	if (pmu_sbi_ctr_is_fw(idx)) {
306 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
307 			return idx;
308 	} else {
309 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
310 			return idx;
311 	}
312 
313 	return -ENOENT;
314 }
315 
316 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
317 {
318 
319 	struct hw_perf_event *hwc = &event->hw;
320 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
321 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
322 	int idx = hwc->idx;
323 
324 	if (pmu_sbi_ctr_is_fw(idx))
325 		clear_bit(idx, cpuc->used_fw_ctrs);
326 	else
327 		clear_bit(idx, cpuc->used_hw_ctrs);
328 }
329 
330 static int pmu_event_find_cache(u64 config)
331 {
332 	unsigned int cache_type, cache_op, cache_result, ret;
333 
334 	cache_type = (config >>  0) & 0xff;
335 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
336 		return -EINVAL;
337 
338 	cache_op = (config >>  8) & 0xff;
339 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
340 		return -EINVAL;
341 
342 	cache_result = (config >> 16) & 0xff;
343 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
344 		return -EINVAL;
345 
346 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
347 
348 	return ret;
349 }
350 
351 static bool pmu_sbi_is_fw_event(struct perf_event *event)
352 {
353 	u32 type = event->attr.type;
354 	u64 config = event->attr.config;
355 
356 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
357 		return true;
358 	else
359 		return false;
360 }
361 
362 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
363 {
364 	u32 type = event->attr.type;
365 	u64 config = event->attr.config;
366 	int bSoftware;
367 	u64 raw_config_val;
368 	int ret;
369 
370 	switch (type) {
371 	case PERF_TYPE_HARDWARE:
372 		if (config >= PERF_COUNT_HW_MAX)
373 			return -EINVAL;
374 		ret = pmu_hw_event_map[event->attr.config].event_idx;
375 		break;
376 	case PERF_TYPE_HW_CACHE:
377 		ret = pmu_event_find_cache(config);
378 		break;
379 	case PERF_TYPE_RAW:
380 		/*
381 		 * As per SBI specification, the upper 16 bits must be unused for
382 		 * a raw event. Use the MSB (63b) to distinguish between hardware
383 		 * raw event and firmware events.
384 		 */
385 		bSoftware = config >> 63;
386 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
387 		if (bSoftware) {
388 			if (raw_config_val < SBI_PMU_FW_MAX)
389 				ret = (raw_config_val & 0xFFFF) |
390 				      (SBI_PMU_EVENT_TYPE_FW << 16);
391 			else
392 				return -EINVAL;
393 		} else {
394 			ret = RISCV_PMU_RAW_EVENT_IDX;
395 			*econfig = raw_config_val;
396 		}
397 		break;
398 	default:
399 		ret = -EINVAL;
400 		break;
401 	}
402 
403 	return ret;
404 }
405 
406 static u64 pmu_sbi_ctr_read(struct perf_event *event)
407 {
408 	struct hw_perf_event *hwc = &event->hw;
409 	int idx = hwc->idx;
410 	struct sbiret ret;
411 	union sbi_pmu_ctr_info info;
412 	u64 val = 0;
413 
414 	if (pmu_sbi_is_fw_event(event)) {
415 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
416 				hwc->idx, 0, 0, 0, 0, 0);
417 		if (!ret.error)
418 			val = ret.value;
419 	} else {
420 		info = pmu_ctr_list[idx];
421 		val = riscv_pmu_ctr_read_csr(info.csr);
422 		if (IS_ENABLED(CONFIG_32BIT))
423 			val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
424 	}
425 
426 	return val;
427 }
428 
429 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
430 {
431 	struct sbiret ret;
432 	struct hw_perf_event *hwc = &event->hw;
433 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
434 
435 #if defined(CONFIG_32BIT)
436 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
437 			1, flag, ival, ival >> 32, 0);
438 #else
439 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
440 			1, flag, ival, 0, 0);
441 #endif
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, unsigned long *mask)
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 
486 		*mask |= BIT(i);
487 
488 		cinfo.value = ret.value;
489 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
490 			num_fw_ctr++;
491 		else
492 			num_hw_ctr++;
493 		pmu_ctr_list[i].value = cinfo.value;
494 	}
495 
496 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
497 
498 	return 0;
499 }
500 
501 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
502 {
503 	/*
504 	 * No need to check the error because we are disabling all the counters
505 	 * which may include counters that are not enabled yet.
506 	 */
507 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
508 		  0, pmu->cmask, 0, 0, 0, 0);
509 }
510 
511 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
512 {
513 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
514 
515 	/* No need to check the error here as we can't do anything about the error */
516 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
517 		  cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
518 }
519 
520 /*
521  * This function starts all the used counters in two step approach.
522  * Any counter that did not overflow can be start in a single step
523  * while the overflowed counters need to be started with updated initialization
524  * value.
525  */
526 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
527 					       unsigned long ctr_ovf_mask)
528 {
529 	int idx = 0;
530 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
531 	struct perf_event *event;
532 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
533 	unsigned long ctr_start_mask = 0;
534 	uint64_t max_period;
535 	struct hw_perf_event *hwc;
536 	u64 init_val = 0;
537 
538 	ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
539 
540 	/* Start all the counters that did not overflow in a single shot */
541 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
542 		  0, 0, 0, 0);
543 
544 	/* Reinitialize and start all the counter that overflowed */
545 	while (ctr_ovf_mask) {
546 		if (ctr_ovf_mask & 0x01) {
547 			event = cpu_hw_evt->events[idx];
548 			hwc = &event->hw;
549 			max_period = riscv_pmu_ctr_get_width_mask(event);
550 			init_val = local64_read(&hwc->prev_count) & max_period;
551 #if defined(CONFIG_32BIT)
552 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
553 				  flag, init_val, init_val >> 32, 0);
554 #else
555 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
556 				  flag, init_val, 0, 0);
557 #endif
558 			perf_event_update_userpage(event);
559 		}
560 		ctr_ovf_mask = ctr_ovf_mask >> 1;
561 		idx++;
562 	}
563 }
564 
565 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
566 {
567 	struct perf_sample_data data;
568 	struct pt_regs *regs;
569 	struct hw_perf_event *hw_evt;
570 	union sbi_pmu_ctr_info *info;
571 	int lidx, hidx, fidx;
572 	struct riscv_pmu *pmu;
573 	struct perf_event *event;
574 	unsigned long overflow;
575 	unsigned long overflowed_ctrs = 0;
576 	struct cpu_hw_events *cpu_hw_evt = dev;
577 	u64 start_clock = sched_clock();
578 
579 	if (WARN_ON_ONCE(!cpu_hw_evt))
580 		return IRQ_NONE;
581 
582 	/* Firmware counter don't support overflow yet */
583 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
584 	event = cpu_hw_evt->events[fidx];
585 	if (!event) {
586 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
587 		return IRQ_NONE;
588 	}
589 
590 	pmu = to_riscv_pmu(event->pmu);
591 	pmu_sbi_stop_hw_ctrs(pmu);
592 
593 	/* Overflow status register should only be read after counter are stopped */
594 	ALT_SBI_PMU_OVERFLOW(overflow);
595 
596 	/*
597 	 * Overflow interrupt pending bit should only be cleared after stopping
598 	 * all the counters to avoid any race condition.
599 	 */
600 	csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
601 
602 	/* No overflow bit is set */
603 	if (!overflow)
604 		return IRQ_NONE;
605 
606 	regs = get_irq_regs();
607 
608 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
609 		struct perf_event *event = cpu_hw_evt->events[lidx];
610 
611 		/* Skip if invalid event or user did not request a sampling */
612 		if (!event || !is_sampling_event(event))
613 			continue;
614 
615 		info = &pmu_ctr_list[lidx];
616 		/* Do a sanity check */
617 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
618 			continue;
619 
620 		/* compute hardware counter index */
621 		hidx = info->csr - CSR_CYCLE;
622 		/* check if the corresponding bit is set in sscountovf */
623 		if (!(overflow & (1 << hidx)))
624 			continue;
625 
626 		/*
627 		 * Keep a track of overflowed counters so that they can be started
628 		 * with updated initial value.
629 		 */
630 		overflowed_ctrs |= 1 << lidx;
631 		hw_evt = &event->hw;
632 		riscv_pmu_event_update(event);
633 		perf_sample_data_init(&data, 0, hw_evt->last_period);
634 		if (riscv_pmu_event_set_period(event)) {
635 			/*
636 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
637 			 * to avoid throttling here. As per the specification, the
638 			 * interrupt remains disabled until the OF bit is set.
639 			 * Interrupts are enabled again only during the start.
640 			 * TODO: We will need to stop the guest counters once
641 			 * virtualization support is added.
642 			 */
643 			perf_event_overflow(event, &data, regs);
644 		}
645 	}
646 
647 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
648 	perf_sample_event_took(sched_clock() - start_clock);
649 
650 	return IRQ_HANDLED;
651 }
652 
653 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
654 {
655 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
656 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
657 
658 	/*
659 	 * Enable the access for CYCLE, TIME, and INSTRET CSRs from userspace,
660 	 * as is necessary to maintain uABI compatibility.
661 	 */
662 	csr_write(CSR_SCOUNTEREN, 0x7);
663 
664 	/* Stop all the counters so that they can be enabled from perf */
665 	pmu_sbi_stop_all(pmu);
666 
667 	if (riscv_pmu_use_irq) {
668 		cpu_hw_evt->irq = riscv_pmu_irq;
669 		csr_clear(CSR_IP, BIT(riscv_pmu_irq_num));
670 		csr_set(CSR_IE, BIT(riscv_pmu_irq_num));
671 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
672 	}
673 
674 	return 0;
675 }
676 
677 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
678 {
679 	if (riscv_pmu_use_irq) {
680 		disable_percpu_irq(riscv_pmu_irq);
681 		csr_clear(CSR_IE, BIT(riscv_pmu_irq_num));
682 	}
683 
684 	/* Disable all counters access for user mode now */
685 	csr_write(CSR_SCOUNTEREN, 0x0);
686 
687 	return 0;
688 }
689 
690 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
691 {
692 	int ret;
693 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
694 	struct device_node *cpu, *child;
695 	struct irq_domain *domain = NULL;
696 
697 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
698 		riscv_pmu_irq_num = RV_IRQ_PMU;
699 		riscv_pmu_use_irq = true;
700 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
701 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
702 		   riscv_cached_marchid(0) == 0 &&
703 		   riscv_cached_mimpid(0) == 0) {
704 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
705 		riscv_pmu_use_irq = true;
706 	}
707 
708 	if (!riscv_pmu_use_irq)
709 		return -EOPNOTSUPP;
710 
711 	for_each_of_cpu_node(cpu) {
712 		child = of_get_compatible_child(cpu, "riscv,cpu-intc");
713 		if (!child) {
714 			pr_err("Failed to find INTC node\n");
715 			of_node_put(cpu);
716 			return -ENODEV;
717 		}
718 		domain = irq_find_host(child);
719 		of_node_put(child);
720 		if (domain) {
721 			of_node_put(cpu);
722 			break;
723 		}
724 	}
725 	if (!domain) {
726 		pr_err("Failed to find INTC IRQ root domain\n");
727 		return -ENODEV;
728 	}
729 
730 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
731 	if (!riscv_pmu_irq) {
732 		pr_err("Failed to map PMU interrupt for node\n");
733 		return -ENODEV;
734 	}
735 
736 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
737 	if (ret) {
738 		pr_err("registering percpu irq failed [%d]\n", ret);
739 		return ret;
740 	}
741 
742 	return 0;
743 }
744 
745 #ifdef CONFIG_CPU_PM
746 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
747 				void *v)
748 {
749 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
750 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
751 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
752 	struct perf_event *event;
753 	int idx;
754 
755 	if (!enabled)
756 		return NOTIFY_OK;
757 
758 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
759 		event = cpuc->events[idx];
760 		if (!event)
761 			continue;
762 
763 		switch (cmd) {
764 		case CPU_PM_ENTER:
765 			/*
766 			 * Stop and update the counter
767 			 */
768 			riscv_pmu_stop(event, PERF_EF_UPDATE);
769 			break;
770 		case CPU_PM_EXIT:
771 		case CPU_PM_ENTER_FAILED:
772 			/*
773 			 * Restore and enable the counter.
774 			 *
775 			 * Requires RCU read locking to be functional,
776 			 * wrap the call within RCU_NONIDLE to make the
777 			 * RCU subsystem aware this cpu is not idle from
778 			 * an RCU perspective for the riscv_pmu_start() call
779 			 * duration.
780 			 */
781 			RCU_NONIDLE(riscv_pmu_start(event, PERF_EF_RELOAD));
782 			break;
783 		default:
784 			break;
785 		}
786 	}
787 
788 	return NOTIFY_OK;
789 }
790 
791 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
792 {
793 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
794 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
795 }
796 
797 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
798 {
799 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
800 }
801 #else
802 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
803 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
804 #endif
805 
806 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
807 {
808 	riscv_pm_pmu_unregister(pmu);
809 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
810 }
811 
812 static int pmu_sbi_device_probe(struct platform_device *pdev)
813 {
814 	struct riscv_pmu *pmu = NULL;
815 	unsigned long cmask = 0;
816 	int ret = -ENODEV;
817 	int num_counters;
818 
819 	pr_info("SBI PMU extension is available\n");
820 	pmu = riscv_pmu_alloc();
821 	if (!pmu)
822 		return -ENOMEM;
823 
824 	num_counters = pmu_sbi_find_num_ctrs();
825 	if (num_counters < 0) {
826 		pr_err("SBI PMU extension doesn't provide any counters\n");
827 		goto out_free;
828 	}
829 
830 	/* cache all the information about counters now */
831 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
832 		goto out_free;
833 
834 	ret = pmu_sbi_setup_irqs(pmu, pdev);
835 	if (ret < 0) {
836 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
837 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
838 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
839 	}
840 
841 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
842 	pmu->cmask = cmask;
843 	pmu->ctr_start = pmu_sbi_ctr_start;
844 	pmu->ctr_stop = pmu_sbi_ctr_stop;
845 	pmu->event_map = pmu_sbi_event_map;
846 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
847 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
848 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
849 	pmu->ctr_read = pmu_sbi_ctr_read;
850 
851 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
852 	if (ret)
853 		return ret;
854 
855 	ret = riscv_pm_pmu_register(pmu);
856 	if (ret)
857 		goto out_unregister;
858 
859 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
860 	if (ret)
861 		goto out_unregister;
862 
863 	return 0;
864 
865 out_unregister:
866 	riscv_pmu_destroy(pmu);
867 
868 out_free:
869 	kfree(pmu);
870 	return ret;
871 }
872 
873 static struct platform_driver pmu_sbi_driver = {
874 	.probe		= pmu_sbi_device_probe,
875 	.driver		= {
876 		.name	= RISCV_PMU_PDEV_NAME,
877 	},
878 };
879 
880 static int __init pmu_sbi_devinit(void)
881 {
882 	int ret;
883 	struct platform_device *pdev;
884 
885 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
886 	    sbi_probe_extension(SBI_EXT_PMU) <= 0) {
887 		return 0;
888 	}
889 
890 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
891 				      "perf/riscv/pmu:starting",
892 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
893 	if (ret) {
894 		pr_err("CPU hotplug notifier could not be registered: %d\n",
895 		       ret);
896 		return ret;
897 	}
898 
899 	ret = platform_driver_register(&pmu_sbi_driver);
900 	if (ret)
901 		return ret;
902 
903 	pdev = platform_device_register_simple(RISCV_PMU_PDEV_NAME, -1, NULL, 0);
904 	if (IS_ERR(pdev)) {
905 		platform_driver_unregister(&pmu_sbi_driver);
906 		return PTR_ERR(pdev);
907 	}
908 
909 	/* Notify legacy implementation that SBI pmu is available*/
910 	riscv_pmu_legacy_skip_init();
911 
912 	return ret;
913 }
914 device_initcall(pmu_sbi_devinit)
915