xref: /openbmc/linux/drivers/perf/riscv_pmu_sbi.c (revision 7effbd18)
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 heterogeneous 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 /* Cache the available counters in a bitmask */
56 static unsigned long cmask;
57 
58 struct sbi_pmu_event_data {
59 	union {
60 		union {
61 			struct hw_gen_event {
62 				uint32_t event_code:16;
63 				uint32_t event_type:4;
64 				uint32_t reserved:12;
65 			} hw_gen_event;
66 			struct hw_cache_event {
67 				uint32_t result_id:1;
68 				uint32_t op_id:2;
69 				uint32_t cache_id:13;
70 				uint32_t event_type:4;
71 				uint32_t reserved:12;
72 			} hw_cache_event;
73 		};
74 		uint32_t event_idx;
75 	};
76 };
77 
78 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
79 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
80 							SBI_PMU_HW_CPU_CYCLES,
81 							SBI_PMU_EVENT_TYPE_HW, 0}},
82 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
83 							SBI_PMU_HW_INSTRUCTIONS,
84 							SBI_PMU_EVENT_TYPE_HW, 0}},
85 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
86 							SBI_PMU_HW_CACHE_REFERENCES,
87 							SBI_PMU_EVENT_TYPE_HW, 0}},
88 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
89 							SBI_PMU_HW_CACHE_MISSES,
90 							SBI_PMU_EVENT_TYPE_HW, 0}},
91 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
92 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
93 							SBI_PMU_EVENT_TYPE_HW, 0}},
94 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
95 							SBI_PMU_HW_BRANCH_MISSES,
96 							SBI_PMU_EVENT_TYPE_HW, 0}},
97 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
98 							SBI_PMU_HW_BUS_CYCLES,
99 							SBI_PMU_EVENT_TYPE_HW, 0}},
100 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
101 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
102 							SBI_PMU_EVENT_TYPE_HW, 0}},
103 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
104 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
105 							SBI_PMU_EVENT_TYPE_HW, 0}},
106 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
107 							SBI_PMU_HW_REF_CPU_CYCLES,
108 							SBI_PMU_EVENT_TYPE_HW, 0}},
109 };
110 
111 #define C(x) PERF_COUNT_HW_CACHE_##x
112 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
113 [PERF_COUNT_HW_CACHE_OP_MAX]
114 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
115 	[C(L1D)] = {
116 		[C(OP_READ)] = {
117 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
118 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
119 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
120 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
121 		},
122 		[C(OP_WRITE)] = {
123 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
124 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
125 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
126 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
127 		},
128 		[C(OP_PREFETCH)] = {
129 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
130 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
131 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
132 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
133 		},
134 	},
135 	[C(L1I)] = {
136 		[C(OP_READ)] = {
137 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
138 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
139 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
140 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
141 		},
142 		[C(OP_WRITE)] = {
143 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
144 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
145 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
146 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
147 		},
148 		[C(OP_PREFETCH)] = {
149 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
150 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
151 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
152 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
153 		},
154 	},
155 	[C(LL)] = {
156 		[C(OP_READ)] = {
157 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
158 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
159 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
160 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
161 		},
162 		[C(OP_WRITE)] = {
163 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
164 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
165 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
166 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
167 		},
168 		[C(OP_PREFETCH)] = {
169 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
170 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
171 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
172 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
173 		},
174 	},
175 	[C(DTLB)] = {
176 		[C(OP_READ)] = {
177 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
178 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
179 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
180 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
181 		},
182 		[C(OP_WRITE)] = {
183 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
184 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
185 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
186 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
187 		},
188 		[C(OP_PREFETCH)] = {
189 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
190 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
191 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
192 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
193 		},
194 	},
195 	[C(ITLB)] = {
196 		[C(OP_READ)] = {
197 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
198 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
199 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
200 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
201 		},
202 		[C(OP_WRITE)] = {
203 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
204 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
205 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
206 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
207 		},
208 		[C(OP_PREFETCH)] = {
209 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
210 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
211 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
212 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
213 		},
214 	},
215 	[C(BPU)] = {
216 		[C(OP_READ)] = {
217 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
218 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
219 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
220 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
221 		},
222 		[C(OP_WRITE)] = {
223 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
224 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
225 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
226 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
227 		},
228 		[C(OP_PREFETCH)] = {
229 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
230 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
231 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
232 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
233 		},
234 	},
235 	[C(NODE)] = {
236 		[C(OP_READ)] = {
237 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
238 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
239 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
240 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
241 		},
242 		[C(OP_WRITE)] = {
243 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
244 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
245 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
246 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
247 		},
248 		[C(OP_PREFETCH)] = {
249 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
250 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
251 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
252 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
253 		},
254 	},
255 };
256 
257 static int pmu_sbi_ctr_get_width(int idx)
258 {
259 	return pmu_ctr_list[idx].width;
260 }
261 
262 static bool pmu_sbi_ctr_is_fw(int cidx)
263 {
264 	union sbi_pmu_ctr_info *info;
265 
266 	info = &pmu_ctr_list[cidx];
267 	if (!info)
268 		return false;
269 
270 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
271 }
272 
273 /*
274  * Returns the counter width of a programmable counter and number of hardware
275  * counters. As we don't support heterogeneous CPUs yet, it is okay to just
276  * return the counter width of the first programmable counter.
277  */
278 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
279 {
280 	int i;
281 	union sbi_pmu_ctr_info *info;
282 	u32 hpm_width = 0, hpm_count = 0;
283 
284 	if (!cmask)
285 		return -EINVAL;
286 
287 	for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
288 		info = &pmu_ctr_list[i];
289 		if (!info)
290 			continue;
291 		if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET)
292 			hpm_width = info->width;
293 		if (info->type == SBI_PMU_CTR_TYPE_HW)
294 			hpm_count++;
295 	}
296 
297 	*hw_ctr_width = hpm_width;
298 	*num_hw_ctr = hpm_count;
299 
300 	return 0;
301 }
302 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
303 
304 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
305 {
306 	unsigned long cflags = 0;
307 	bool guest_events = false;
308 
309 	if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
310 		guest_events = true;
311 	if (event->attr.exclude_kernel)
312 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH;
313 	if (event->attr.exclude_user)
314 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH;
315 	if (guest_events && event->attr.exclude_hv)
316 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
317 	if (event->attr.exclude_host)
318 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH;
319 	if (event->attr.exclude_guest)
320 		cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH;
321 
322 	return cflags;
323 }
324 
325 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
326 {
327 	struct hw_perf_event *hwc = &event->hw;
328 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
329 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
330 	struct sbiret ret;
331 	int idx;
332 	uint64_t cbase = 0;
333 	unsigned long cflags = 0;
334 
335 	cflags = pmu_sbi_get_filter_flags(event);
336 	/* retrieve the available counter index */
337 #if defined(CONFIG_32BIT)
338 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
339 			rvpmu->cmask, cflags, hwc->event_base, hwc->config,
340 			hwc->config >> 32);
341 #else
342 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
343 			rvpmu->cmask, cflags, hwc->event_base, hwc->config, 0);
344 #endif
345 	if (ret.error) {
346 		pr_debug("Not able to find a counter for event %lx config %llx\n",
347 			hwc->event_base, hwc->config);
348 		return sbi_err_map_linux_errno(ret.error);
349 	}
350 
351 	idx = ret.value;
352 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
353 		return -ENOENT;
354 
355 	/* Additional sanity check for the counter id */
356 	if (pmu_sbi_ctr_is_fw(idx)) {
357 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
358 			return idx;
359 	} else {
360 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
361 			return idx;
362 	}
363 
364 	return -ENOENT;
365 }
366 
367 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
368 {
369 
370 	struct hw_perf_event *hwc = &event->hw;
371 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
372 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
373 	int idx = hwc->idx;
374 
375 	if (pmu_sbi_ctr_is_fw(idx))
376 		clear_bit(idx, cpuc->used_fw_ctrs);
377 	else
378 		clear_bit(idx, cpuc->used_hw_ctrs);
379 }
380 
381 static int pmu_event_find_cache(u64 config)
382 {
383 	unsigned int cache_type, cache_op, cache_result, ret;
384 
385 	cache_type = (config >>  0) & 0xff;
386 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
387 		return -EINVAL;
388 
389 	cache_op = (config >>  8) & 0xff;
390 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
391 		return -EINVAL;
392 
393 	cache_result = (config >> 16) & 0xff;
394 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
395 		return -EINVAL;
396 
397 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
398 
399 	return ret;
400 }
401 
402 static bool pmu_sbi_is_fw_event(struct perf_event *event)
403 {
404 	u32 type = event->attr.type;
405 	u64 config = event->attr.config;
406 
407 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
408 		return true;
409 	else
410 		return false;
411 }
412 
413 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
414 {
415 	u32 type = event->attr.type;
416 	u64 config = event->attr.config;
417 	int bSoftware;
418 	u64 raw_config_val;
419 	int ret;
420 
421 	switch (type) {
422 	case PERF_TYPE_HARDWARE:
423 		if (config >= PERF_COUNT_HW_MAX)
424 			return -EINVAL;
425 		ret = pmu_hw_event_map[event->attr.config].event_idx;
426 		break;
427 	case PERF_TYPE_HW_CACHE:
428 		ret = pmu_event_find_cache(config);
429 		break;
430 	case PERF_TYPE_RAW:
431 		/*
432 		 * As per SBI specification, the upper 16 bits must be unused for
433 		 * a raw event. Use the MSB (63b) to distinguish between hardware
434 		 * raw event and firmware events.
435 		 */
436 		bSoftware = config >> 63;
437 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
438 		if (bSoftware) {
439 			ret = (raw_config_val & 0xFFFF) |
440 				(SBI_PMU_EVENT_TYPE_FW << 16);
441 		} else {
442 			ret = RISCV_PMU_RAW_EVENT_IDX;
443 			*econfig = raw_config_val;
444 		}
445 		break;
446 	default:
447 		ret = -EINVAL;
448 		break;
449 	}
450 
451 	return ret;
452 }
453 
454 static u64 pmu_sbi_ctr_read(struct perf_event *event)
455 {
456 	struct hw_perf_event *hwc = &event->hw;
457 	int idx = hwc->idx;
458 	struct sbiret ret;
459 	union sbi_pmu_ctr_info info;
460 	u64 val = 0;
461 
462 	if (pmu_sbi_is_fw_event(event)) {
463 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
464 				hwc->idx, 0, 0, 0, 0, 0);
465 		if (!ret.error)
466 			val = ret.value;
467 	} else {
468 		info = pmu_ctr_list[idx];
469 		val = riscv_pmu_ctr_read_csr(info.csr);
470 		if (IS_ENABLED(CONFIG_32BIT))
471 			val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
472 	}
473 
474 	return val;
475 }
476 
477 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
478 {
479 	struct sbiret ret;
480 	struct hw_perf_event *hwc = &event->hw;
481 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
482 
483 #if defined(CONFIG_32BIT)
484 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
485 			1, flag, ival, ival >> 32, 0);
486 #else
487 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
488 			1, flag, ival, 0, 0);
489 #endif
490 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
491 		pr_err("Starting counter idx %d failed with error %d\n",
492 			hwc->idx, sbi_err_map_linux_errno(ret.error));
493 }
494 
495 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
496 {
497 	struct sbiret ret;
498 	struct hw_perf_event *hwc = &event->hw;
499 
500 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
501 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
502 		flag != SBI_PMU_STOP_FLAG_RESET)
503 		pr_err("Stopping counter idx %d failed with error %d\n",
504 			hwc->idx, sbi_err_map_linux_errno(ret.error));
505 }
506 
507 static int pmu_sbi_find_num_ctrs(void)
508 {
509 	struct sbiret ret;
510 
511 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
512 	if (!ret.error)
513 		return ret.value;
514 	else
515 		return sbi_err_map_linux_errno(ret.error);
516 }
517 
518 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
519 {
520 	struct sbiret ret;
521 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
522 	union sbi_pmu_ctr_info cinfo;
523 
524 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
525 	if (!pmu_ctr_list)
526 		return -ENOMEM;
527 
528 	for (i = 0; i < nctr; i++) {
529 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
530 		if (ret.error)
531 			/* The logical counter ids are not expected to be contiguous */
532 			continue;
533 
534 		*mask |= BIT(i);
535 
536 		cinfo.value = ret.value;
537 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
538 			num_fw_ctr++;
539 		else
540 			num_hw_ctr++;
541 		pmu_ctr_list[i].value = cinfo.value;
542 	}
543 
544 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
545 
546 	return 0;
547 }
548 
549 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
550 {
551 	/*
552 	 * No need to check the error because we are disabling all the counters
553 	 * which may include counters that are not enabled yet.
554 	 */
555 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
556 		  0, pmu->cmask, 0, 0, 0, 0);
557 }
558 
559 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
560 {
561 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
562 
563 	/* No need to check the error here as we can't do anything about the error */
564 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
565 		  cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
566 }
567 
568 /*
569  * This function starts all the used counters in two step approach.
570  * Any counter that did not overflow can be start in a single step
571  * while the overflowed counters need to be started with updated initialization
572  * value.
573  */
574 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
575 					       unsigned long ctr_ovf_mask)
576 {
577 	int idx = 0;
578 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
579 	struct perf_event *event;
580 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
581 	unsigned long ctr_start_mask = 0;
582 	uint64_t max_period;
583 	struct hw_perf_event *hwc;
584 	u64 init_val = 0;
585 
586 	ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
587 
588 	/* Start all the counters that did not overflow in a single shot */
589 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
590 		  0, 0, 0, 0);
591 
592 	/* Reinitialize and start all the counter that overflowed */
593 	while (ctr_ovf_mask) {
594 		if (ctr_ovf_mask & 0x01) {
595 			event = cpu_hw_evt->events[idx];
596 			hwc = &event->hw;
597 			max_period = riscv_pmu_ctr_get_width_mask(event);
598 			init_val = local64_read(&hwc->prev_count) & max_period;
599 #if defined(CONFIG_32BIT)
600 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
601 				  flag, init_val, init_val >> 32, 0);
602 #else
603 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
604 				  flag, init_val, 0, 0);
605 #endif
606 			perf_event_update_userpage(event);
607 		}
608 		ctr_ovf_mask = ctr_ovf_mask >> 1;
609 		idx++;
610 	}
611 }
612 
613 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
614 {
615 	struct perf_sample_data data;
616 	struct pt_regs *regs;
617 	struct hw_perf_event *hw_evt;
618 	union sbi_pmu_ctr_info *info;
619 	int lidx, hidx, fidx;
620 	struct riscv_pmu *pmu;
621 	struct perf_event *event;
622 	unsigned long overflow;
623 	unsigned long overflowed_ctrs = 0;
624 	struct cpu_hw_events *cpu_hw_evt = dev;
625 	u64 start_clock = sched_clock();
626 
627 	if (WARN_ON_ONCE(!cpu_hw_evt))
628 		return IRQ_NONE;
629 
630 	/* Firmware counter don't support overflow yet */
631 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
632 	event = cpu_hw_evt->events[fidx];
633 	if (!event) {
634 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
635 		return IRQ_NONE;
636 	}
637 
638 	pmu = to_riscv_pmu(event->pmu);
639 	pmu_sbi_stop_hw_ctrs(pmu);
640 
641 	/* Overflow status register should only be read after counter are stopped */
642 	ALT_SBI_PMU_OVERFLOW(overflow);
643 
644 	/*
645 	 * Overflow interrupt pending bit should only be cleared after stopping
646 	 * all the counters to avoid any race condition.
647 	 */
648 	csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
649 
650 	/* No overflow bit is set */
651 	if (!overflow)
652 		return IRQ_NONE;
653 
654 	regs = get_irq_regs();
655 
656 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
657 		struct perf_event *event = cpu_hw_evt->events[lidx];
658 
659 		/* Skip if invalid event or user did not request a sampling */
660 		if (!event || !is_sampling_event(event))
661 			continue;
662 
663 		info = &pmu_ctr_list[lidx];
664 		/* Do a sanity check */
665 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
666 			continue;
667 
668 		/* compute hardware counter index */
669 		hidx = info->csr - CSR_CYCLE;
670 		/* check if the corresponding bit is set in sscountovf */
671 		if (!(overflow & (1 << hidx)))
672 			continue;
673 
674 		/*
675 		 * Keep a track of overflowed counters so that they can be started
676 		 * with updated initial value.
677 		 */
678 		overflowed_ctrs |= 1 << lidx;
679 		hw_evt = &event->hw;
680 		riscv_pmu_event_update(event);
681 		perf_sample_data_init(&data, 0, hw_evt->last_period);
682 		if (riscv_pmu_event_set_period(event)) {
683 			/*
684 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
685 			 * to avoid throttling here. As per the specification, the
686 			 * interrupt remains disabled until the OF bit is set.
687 			 * Interrupts are enabled again only during the start.
688 			 * TODO: We will need to stop the guest counters once
689 			 * virtualization support is added.
690 			 */
691 			perf_event_overflow(event, &data, regs);
692 		}
693 	}
694 
695 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
696 	perf_sample_event_took(sched_clock() - start_clock);
697 
698 	return IRQ_HANDLED;
699 }
700 
701 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
702 {
703 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
704 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
705 
706 	/*
707 	 * Enable the access for CYCLE, TIME, and INSTRET CSRs from userspace,
708 	 * as is necessary to maintain uABI compatibility.
709 	 */
710 	csr_write(CSR_SCOUNTEREN, 0x7);
711 
712 	/* Stop all the counters so that they can be enabled from perf */
713 	pmu_sbi_stop_all(pmu);
714 
715 	if (riscv_pmu_use_irq) {
716 		cpu_hw_evt->irq = riscv_pmu_irq;
717 		csr_clear(CSR_IP, BIT(riscv_pmu_irq_num));
718 		csr_set(CSR_IE, BIT(riscv_pmu_irq_num));
719 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
720 	}
721 
722 	return 0;
723 }
724 
725 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
726 {
727 	if (riscv_pmu_use_irq) {
728 		disable_percpu_irq(riscv_pmu_irq);
729 		csr_clear(CSR_IE, BIT(riscv_pmu_irq_num));
730 	}
731 
732 	/* Disable all counters access for user mode now */
733 	csr_write(CSR_SCOUNTEREN, 0x0);
734 
735 	return 0;
736 }
737 
738 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
739 {
740 	int ret;
741 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
742 	struct device_node *cpu, *child;
743 	struct irq_domain *domain = NULL;
744 
745 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
746 		riscv_pmu_irq_num = RV_IRQ_PMU;
747 		riscv_pmu_use_irq = true;
748 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
749 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
750 		   riscv_cached_marchid(0) == 0 &&
751 		   riscv_cached_mimpid(0) == 0) {
752 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
753 		riscv_pmu_use_irq = true;
754 	}
755 
756 	if (!riscv_pmu_use_irq)
757 		return -EOPNOTSUPP;
758 
759 	for_each_of_cpu_node(cpu) {
760 		child = of_get_compatible_child(cpu, "riscv,cpu-intc");
761 		if (!child) {
762 			pr_err("Failed to find INTC node\n");
763 			of_node_put(cpu);
764 			return -ENODEV;
765 		}
766 		domain = irq_find_host(child);
767 		of_node_put(child);
768 		if (domain) {
769 			of_node_put(cpu);
770 			break;
771 		}
772 	}
773 	if (!domain) {
774 		pr_err("Failed to find INTC IRQ root domain\n");
775 		return -ENODEV;
776 	}
777 
778 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
779 	if (!riscv_pmu_irq) {
780 		pr_err("Failed to map PMU interrupt for node\n");
781 		return -ENODEV;
782 	}
783 
784 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
785 	if (ret) {
786 		pr_err("registering percpu irq failed [%d]\n", ret);
787 		return ret;
788 	}
789 
790 	return 0;
791 }
792 
793 #ifdef CONFIG_CPU_PM
794 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
795 				void *v)
796 {
797 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
798 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
799 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
800 	struct perf_event *event;
801 	int idx;
802 
803 	if (!enabled)
804 		return NOTIFY_OK;
805 
806 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
807 		event = cpuc->events[idx];
808 		if (!event)
809 			continue;
810 
811 		switch (cmd) {
812 		case CPU_PM_ENTER:
813 			/*
814 			 * Stop and update the counter
815 			 */
816 			riscv_pmu_stop(event, PERF_EF_UPDATE);
817 			break;
818 		case CPU_PM_EXIT:
819 		case CPU_PM_ENTER_FAILED:
820 			/*
821 			 * Restore and enable the counter.
822 			 */
823 			riscv_pmu_start(event, PERF_EF_RELOAD);
824 			break;
825 		default:
826 			break;
827 		}
828 	}
829 
830 	return NOTIFY_OK;
831 }
832 
833 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
834 {
835 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
836 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
837 }
838 
839 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
840 {
841 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
842 }
843 #else
844 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
845 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
846 #endif
847 
848 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
849 {
850 	riscv_pm_pmu_unregister(pmu);
851 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
852 }
853 
854 static int pmu_sbi_device_probe(struct platform_device *pdev)
855 {
856 	struct riscv_pmu *pmu = NULL;
857 	int ret = -ENODEV;
858 	int num_counters;
859 
860 	pr_info("SBI PMU extension is available\n");
861 	pmu = riscv_pmu_alloc();
862 	if (!pmu)
863 		return -ENOMEM;
864 
865 	num_counters = pmu_sbi_find_num_ctrs();
866 	if (num_counters < 0) {
867 		pr_err("SBI PMU extension doesn't provide any counters\n");
868 		goto out_free;
869 	}
870 
871 	/* cache all the information about counters now */
872 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
873 		goto out_free;
874 
875 	ret = pmu_sbi_setup_irqs(pmu, pdev);
876 	if (ret < 0) {
877 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
878 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
879 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
880 	}
881 
882 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
883 	pmu->cmask = cmask;
884 	pmu->ctr_start = pmu_sbi_ctr_start;
885 	pmu->ctr_stop = pmu_sbi_ctr_stop;
886 	pmu->event_map = pmu_sbi_event_map;
887 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
888 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
889 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
890 	pmu->ctr_read = pmu_sbi_ctr_read;
891 
892 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
893 	if (ret)
894 		return ret;
895 
896 	ret = riscv_pm_pmu_register(pmu);
897 	if (ret)
898 		goto out_unregister;
899 
900 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
901 	if (ret)
902 		goto out_unregister;
903 
904 	return 0;
905 
906 out_unregister:
907 	riscv_pmu_destroy(pmu);
908 
909 out_free:
910 	kfree(pmu);
911 	return ret;
912 }
913 
914 static struct platform_driver pmu_sbi_driver = {
915 	.probe		= pmu_sbi_device_probe,
916 	.driver		= {
917 		.name	= RISCV_PMU_PDEV_NAME,
918 	},
919 };
920 
921 static int __init pmu_sbi_devinit(void)
922 {
923 	int ret;
924 	struct platform_device *pdev;
925 
926 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
927 	    sbi_probe_extension(SBI_EXT_PMU) <= 0) {
928 		return 0;
929 	}
930 
931 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
932 				      "perf/riscv/pmu:starting",
933 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
934 	if (ret) {
935 		pr_err("CPU hotplug notifier could not be registered: %d\n",
936 		       ret);
937 		return ret;
938 	}
939 
940 	ret = platform_driver_register(&pmu_sbi_driver);
941 	if (ret)
942 		return ret;
943 
944 	pdev = platform_device_register_simple(RISCV_PMU_PDEV_NAME, -1, NULL, 0);
945 	if (IS_ERR(pdev)) {
946 		platform_driver_unregister(&pmu_sbi_driver);
947 		return PTR_ERR(pdev);
948 	}
949 
950 	/* Notify legacy implementation that SBI pmu is available*/
951 	riscv_pmu_legacy_skip_init();
952 
953 	return ret;
954 }
955 device_initcall(pmu_sbi_devinit)
956