xref: /linux/drivers/perf/riscv_pmu_sbi.c (revision f4fee216df7d28b87d1c9cc60bcebfecb51c1a05)
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 #include <linux/soc/andes/irq.h>
23 
24 #include <asm/errata_list.h>
25 #include <asm/sbi.h>
26 #include <asm/cpufeature.h>
27 
28 #define ALT_SBI_PMU_OVERFLOW(__ovl)					\
29 asm volatile(ALTERNATIVE_2(						\
30 	"csrr %0, " __stringify(CSR_SSCOUNTOVF),			\
31 	"csrr %0, " __stringify(THEAD_C9XX_CSR_SCOUNTEROF),		\
32 		THEAD_VENDOR_ID, ERRATA_THEAD_PMU,			\
33 		CONFIG_ERRATA_THEAD_PMU,				\
34 	"csrr %0, " __stringify(ANDES_CSR_SCOUNTEROF),			\
35 		0, RISCV_ISA_EXT_XANDESPMU,				\
36 		CONFIG_ANDES_CUSTOM_PMU)				\
37 	: "=r" (__ovl) :						\
38 	: "memory")
39 
40 #define ALT_SBI_PMU_OVF_CLEAR_PENDING(__irq_mask)			\
41 asm volatile(ALTERNATIVE(						\
42 	"csrc " __stringify(CSR_IP) ", %0\n\t",				\
43 	"csrc " __stringify(ANDES_CSR_SLIP) ", %0\n\t",			\
44 		0, RISCV_ISA_EXT_XANDESPMU,				\
45 		CONFIG_ANDES_CUSTOM_PMU)				\
46 	: : "r"(__irq_mask)						\
47 	: "memory")
48 
49 #define SYSCTL_NO_USER_ACCESS	0
50 #define SYSCTL_USER_ACCESS	1
51 #define SYSCTL_LEGACY		2
52 
53 #define PERF_EVENT_FLAG_NO_USER_ACCESS	BIT(SYSCTL_NO_USER_ACCESS)
54 #define PERF_EVENT_FLAG_USER_ACCESS	BIT(SYSCTL_USER_ACCESS)
55 #define PERF_EVENT_FLAG_LEGACY		BIT(SYSCTL_LEGACY)
56 
57 PMU_FORMAT_ATTR(event, "config:0-47");
58 PMU_FORMAT_ATTR(firmware, "config:63");
59 
60 static struct attribute *riscv_arch_formats_attr[] = {
61 	&format_attr_event.attr,
62 	&format_attr_firmware.attr,
63 	NULL,
64 };
65 
66 static struct attribute_group riscv_pmu_format_group = {
67 	.name = "format",
68 	.attrs = riscv_arch_formats_attr,
69 };
70 
71 static const struct attribute_group *riscv_pmu_attr_groups[] = {
72 	&riscv_pmu_format_group,
73 	NULL,
74 };
75 
76 /* Allow user mode access by default */
77 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS;
78 
79 /*
80  * RISC-V doesn't have heterogeneous harts yet. This need to be part of
81  * per_cpu in case of harts with different pmu counters
82  */
83 static union sbi_pmu_ctr_info *pmu_ctr_list;
84 static bool riscv_pmu_use_irq;
85 static unsigned int riscv_pmu_irq_num;
86 static unsigned int riscv_pmu_irq_mask;
87 static unsigned int riscv_pmu_irq;
88 
89 /* Cache the available counters in a bitmask */
90 static unsigned long cmask;
91 
92 struct sbi_pmu_event_data {
93 	union {
94 		union {
95 			struct hw_gen_event {
96 				uint32_t event_code:16;
97 				uint32_t event_type:4;
98 				uint32_t reserved:12;
99 			} hw_gen_event;
100 			struct hw_cache_event {
101 				uint32_t result_id:1;
102 				uint32_t op_id:2;
103 				uint32_t cache_id:13;
104 				uint32_t event_type:4;
105 				uint32_t reserved:12;
106 			} hw_cache_event;
107 		};
108 		uint32_t event_idx;
109 	};
110 };
111 
112 static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
113 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
114 							SBI_PMU_HW_CPU_CYCLES,
115 							SBI_PMU_EVENT_TYPE_HW, 0}},
116 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
117 							SBI_PMU_HW_INSTRUCTIONS,
118 							SBI_PMU_EVENT_TYPE_HW, 0}},
119 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
120 							SBI_PMU_HW_CACHE_REFERENCES,
121 							SBI_PMU_EVENT_TYPE_HW, 0}},
122 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
123 							SBI_PMU_HW_CACHE_MISSES,
124 							SBI_PMU_EVENT_TYPE_HW, 0}},
125 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
126 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
127 							SBI_PMU_EVENT_TYPE_HW, 0}},
128 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
129 							SBI_PMU_HW_BRANCH_MISSES,
130 							SBI_PMU_EVENT_TYPE_HW, 0}},
131 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
132 							SBI_PMU_HW_BUS_CYCLES,
133 							SBI_PMU_EVENT_TYPE_HW, 0}},
134 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
135 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
136 							SBI_PMU_EVENT_TYPE_HW, 0}},
137 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
138 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
139 							SBI_PMU_EVENT_TYPE_HW, 0}},
140 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
141 							SBI_PMU_HW_REF_CPU_CYCLES,
142 							SBI_PMU_EVENT_TYPE_HW, 0}},
143 };
144 
145 #define C(x) PERF_COUNT_HW_CACHE_##x
146 static const struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
147 [PERF_COUNT_HW_CACHE_OP_MAX]
148 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
149 	[C(L1D)] = {
150 		[C(OP_READ)] = {
151 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
152 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
153 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
154 					C(OP_READ), C(L1D), 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(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
159 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
160 					C(OP_WRITE), C(L1D), 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(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
165 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
166 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
167 		},
168 	},
169 	[C(L1I)] = {
170 		[C(OP_READ)] = {
171 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
172 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
173 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
174 					C(L1I), 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(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
179 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
180 					C(OP_WRITE), C(L1I), 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(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
185 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
186 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
187 		},
188 	},
189 	[C(LL)] = {
190 		[C(OP_READ)] = {
191 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
192 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
193 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
194 					C(OP_READ), C(LL), 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(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
199 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
200 					C(OP_WRITE), C(LL), 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(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
205 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
206 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
207 		},
208 	},
209 	[C(DTLB)] = {
210 		[C(OP_READ)] = {
211 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
212 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
213 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
214 					C(OP_READ), C(DTLB), 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(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
219 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
220 					C(OP_WRITE), C(DTLB), 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(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
225 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
226 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
227 		},
228 	},
229 	[C(ITLB)] = {
230 		[C(OP_READ)] = {
231 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
232 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
233 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
234 					C(OP_READ), C(ITLB), 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(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
239 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
240 					C(OP_WRITE), C(ITLB), 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(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
245 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
246 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
247 		},
248 	},
249 	[C(BPU)] = {
250 		[C(OP_READ)] = {
251 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
252 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
253 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
254 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
255 		},
256 		[C(OP_WRITE)] = {
257 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
258 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
259 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
260 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
261 		},
262 		[C(OP_PREFETCH)] = {
263 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
264 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
265 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
266 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
267 		},
268 	},
269 	[C(NODE)] = {
270 		[C(OP_READ)] = {
271 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
272 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
273 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
274 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
275 		},
276 		[C(OP_WRITE)] = {
277 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
278 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
279 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
280 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
281 		},
282 		[C(OP_PREFETCH)] = {
283 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
284 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
285 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
286 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
287 		},
288 	},
289 };
290 
291 static int pmu_sbi_ctr_get_width(int idx)
292 {
293 	return pmu_ctr_list[idx].width;
294 }
295 
296 static bool pmu_sbi_ctr_is_fw(int cidx)
297 {
298 	union sbi_pmu_ctr_info *info;
299 
300 	info = &pmu_ctr_list[cidx];
301 	if (!info)
302 		return false;
303 
304 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
305 }
306 
307 /*
308  * Returns the counter width of a programmable counter and number of hardware
309  * counters. As we don't support heterogeneous CPUs yet, it is okay to just
310  * return the counter width of the first programmable counter.
311  */
312 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
313 {
314 	int i;
315 	union sbi_pmu_ctr_info *info;
316 	u32 hpm_width = 0, hpm_count = 0;
317 
318 	if (!cmask)
319 		return -EINVAL;
320 
321 	for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
322 		info = &pmu_ctr_list[i];
323 		if (!info)
324 			continue;
325 		if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET)
326 			hpm_width = info->width;
327 		if (info->type == SBI_PMU_CTR_TYPE_HW)
328 			hpm_count++;
329 	}
330 
331 	*hw_ctr_width = hpm_width;
332 	*num_hw_ctr = hpm_count;
333 
334 	return 0;
335 }
336 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
337 
338 static uint8_t pmu_sbi_csr_index(struct perf_event *event)
339 {
340 	return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE;
341 }
342 
343 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
344 {
345 	unsigned long cflags = 0;
346 	bool guest_events = false;
347 
348 	if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
349 		guest_events = true;
350 	if (event->attr.exclude_kernel)
351 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH;
352 	if (event->attr.exclude_user)
353 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH;
354 	if (guest_events && event->attr.exclude_hv)
355 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
356 	if (event->attr.exclude_host)
357 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH;
358 	if (event->attr.exclude_guest)
359 		cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH;
360 
361 	return cflags;
362 }
363 
364 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
365 {
366 	struct hw_perf_event *hwc = &event->hw;
367 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
368 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
369 	struct sbiret ret;
370 	int idx;
371 	uint64_t cbase = 0, cmask = rvpmu->cmask;
372 	unsigned long cflags = 0;
373 
374 	cflags = pmu_sbi_get_filter_flags(event);
375 
376 	/*
377 	 * In legacy mode, we have to force the fixed counters for those events
378 	 * but not in the user access mode as we want to use the other counters
379 	 * that support sampling/filtering.
380 	 */
381 	if (hwc->flags & PERF_EVENT_FLAG_LEGACY) {
382 		if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
383 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
384 			cmask = 1;
385 		} else if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS) {
386 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
387 			cmask = 1UL << (CSR_INSTRET - CSR_CYCLE);
388 		}
389 	}
390 
391 	/* retrieve the available counter index */
392 #if defined(CONFIG_32BIT)
393 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
394 			cmask, cflags, hwc->event_base, hwc->config,
395 			hwc->config >> 32);
396 #else
397 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
398 			cmask, cflags, hwc->event_base, hwc->config, 0);
399 #endif
400 	if (ret.error) {
401 		pr_debug("Not able to find a counter for event %lx config %llx\n",
402 			hwc->event_base, hwc->config);
403 		return sbi_err_map_linux_errno(ret.error);
404 	}
405 
406 	idx = ret.value;
407 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
408 		return -ENOENT;
409 
410 	/* Additional sanity check for the counter id */
411 	if (pmu_sbi_ctr_is_fw(idx)) {
412 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
413 			return idx;
414 	} else {
415 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
416 			return idx;
417 	}
418 
419 	return -ENOENT;
420 }
421 
422 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
423 {
424 
425 	struct hw_perf_event *hwc = &event->hw;
426 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
427 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
428 	int idx = hwc->idx;
429 
430 	if (pmu_sbi_ctr_is_fw(idx))
431 		clear_bit(idx, cpuc->used_fw_ctrs);
432 	else
433 		clear_bit(idx, cpuc->used_hw_ctrs);
434 }
435 
436 static int pmu_event_find_cache(u64 config)
437 {
438 	unsigned int cache_type, cache_op, cache_result, ret;
439 
440 	cache_type = (config >>  0) & 0xff;
441 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
442 		return -EINVAL;
443 
444 	cache_op = (config >>  8) & 0xff;
445 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
446 		return -EINVAL;
447 
448 	cache_result = (config >> 16) & 0xff;
449 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
450 		return -EINVAL;
451 
452 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
453 
454 	return ret;
455 }
456 
457 static bool pmu_sbi_is_fw_event(struct perf_event *event)
458 {
459 	u32 type = event->attr.type;
460 	u64 config = event->attr.config;
461 
462 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
463 		return true;
464 	else
465 		return false;
466 }
467 
468 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
469 {
470 	u32 type = event->attr.type;
471 	u64 config = event->attr.config;
472 	int bSoftware;
473 	u64 raw_config_val;
474 	int ret;
475 
476 	switch (type) {
477 	case PERF_TYPE_HARDWARE:
478 		if (config >= PERF_COUNT_HW_MAX)
479 			return -EINVAL;
480 		ret = pmu_hw_event_map[event->attr.config].event_idx;
481 		break;
482 	case PERF_TYPE_HW_CACHE:
483 		ret = pmu_event_find_cache(config);
484 		break;
485 	case PERF_TYPE_RAW:
486 		/*
487 		 * As per SBI specification, the upper 16 bits must be unused for
488 		 * a raw event. Use the MSB (63b) to distinguish between hardware
489 		 * raw event and firmware events.
490 		 */
491 		bSoftware = config >> 63;
492 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
493 		if (bSoftware) {
494 			ret = (raw_config_val & 0xFFFF) |
495 				(SBI_PMU_EVENT_TYPE_FW << 16);
496 		} else {
497 			ret = RISCV_PMU_RAW_EVENT_IDX;
498 			*econfig = raw_config_val;
499 		}
500 		break;
501 	default:
502 		ret = -EINVAL;
503 		break;
504 	}
505 
506 	return ret;
507 }
508 
509 static u64 pmu_sbi_ctr_read(struct perf_event *event)
510 {
511 	struct hw_perf_event *hwc = &event->hw;
512 	int idx = hwc->idx;
513 	struct sbiret ret;
514 	union sbi_pmu_ctr_info info;
515 	u64 val = 0;
516 
517 	if (pmu_sbi_is_fw_event(event)) {
518 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
519 				hwc->idx, 0, 0, 0, 0, 0);
520 		if (!ret.error)
521 			val = ret.value;
522 	} else {
523 		info = pmu_ctr_list[idx];
524 		val = riscv_pmu_ctr_read_csr(info.csr);
525 		if (IS_ENABLED(CONFIG_32BIT))
526 			val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
527 	}
528 
529 	return val;
530 }
531 
532 static void pmu_sbi_set_scounteren(void *arg)
533 {
534 	struct perf_event *event = (struct perf_event *)arg;
535 
536 	if (event->hw.idx != -1)
537 		csr_write(CSR_SCOUNTEREN,
538 			  csr_read(CSR_SCOUNTEREN) | BIT(pmu_sbi_csr_index(event)));
539 }
540 
541 static void pmu_sbi_reset_scounteren(void *arg)
542 {
543 	struct perf_event *event = (struct perf_event *)arg;
544 
545 	if (event->hw.idx != -1)
546 		csr_write(CSR_SCOUNTEREN,
547 			  csr_read(CSR_SCOUNTEREN) & ~BIT(pmu_sbi_csr_index(event)));
548 }
549 
550 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
551 {
552 	struct sbiret ret;
553 	struct hw_perf_event *hwc = &event->hw;
554 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
555 
556 #if defined(CONFIG_32BIT)
557 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
558 			1, flag, ival, ival >> 32, 0);
559 #else
560 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
561 			1, flag, ival, 0, 0);
562 #endif
563 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
564 		pr_err("Starting counter idx %d failed with error %d\n",
565 			hwc->idx, sbi_err_map_linux_errno(ret.error));
566 
567 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
568 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
569 		pmu_sbi_set_scounteren((void *)event);
570 }
571 
572 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
573 {
574 	struct sbiret ret;
575 	struct hw_perf_event *hwc = &event->hw;
576 
577 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
578 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
579 		pmu_sbi_reset_scounteren((void *)event);
580 
581 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
582 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
583 		flag != SBI_PMU_STOP_FLAG_RESET)
584 		pr_err("Stopping counter idx %d failed with error %d\n",
585 			hwc->idx, sbi_err_map_linux_errno(ret.error));
586 }
587 
588 static int pmu_sbi_find_num_ctrs(void)
589 {
590 	struct sbiret ret;
591 
592 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
593 	if (!ret.error)
594 		return ret.value;
595 	else
596 		return sbi_err_map_linux_errno(ret.error);
597 }
598 
599 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
600 {
601 	struct sbiret ret;
602 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
603 	union sbi_pmu_ctr_info cinfo;
604 
605 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
606 	if (!pmu_ctr_list)
607 		return -ENOMEM;
608 
609 	for (i = 0; i < nctr; i++) {
610 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
611 		if (ret.error)
612 			/* The logical counter ids are not expected to be contiguous */
613 			continue;
614 
615 		*mask |= BIT(i);
616 
617 		cinfo.value = ret.value;
618 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
619 			num_fw_ctr++;
620 		else
621 			num_hw_ctr++;
622 		pmu_ctr_list[i].value = cinfo.value;
623 	}
624 
625 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
626 
627 	return 0;
628 }
629 
630 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
631 {
632 	/*
633 	 * No need to check the error because we are disabling all the counters
634 	 * which may include counters that are not enabled yet.
635 	 */
636 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
637 		  0, pmu->cmask, 0, 0, 0, 0);
638 }
639 
640 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
641 {
642 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
643 
644 	/* No need to check the error here as we can't do anything about the error */
645 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, 0,
646 		  cpu_hw_evt->used_hw_ctrs[0], 0, 0, 0, 0);
647 }
648 
649 /*
650  * This function starts all the used counters in two step approach.
651  * Any counter that did not overflow can be start in a single step
652  * while the overflowed counters need to be started with updated initialization
653  * value.
654  */
655 static inline void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
656 					       unsigned long ctr_ovf_mask)
657 {
658 	int idx = 0;
659 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
660 	struct perf_event *event;
661 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
662 	unsigned long ctr_start_mask = 0;
663 	uint64_t max_period;
664 	struct hw_perf_event *hwc;
665 	u64 init_val = 0;
666 
667 	ctr_start_mask = cpu_hw_evt->used_hw_ctrs[0] & ~ctr_ovf_mask;
668 
669 	/* Start all the counters that did not overflow in a single shot */
670 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, 0, ctr_start_mask,
671 		  0, 0, 0, 0);
672 
673 	/* Reinitialize and start all the counter that overflowed */
674 	while (ctr_ovf_mask) {
675 		if (ctr_ovf_mask & 0x01) {
676 			event = cpu_hw_evt->events[idx];
677 			hwc = &event->hw;
678 			max_period = riscv_pmu_ctr_get_width_mask(event);
679 			init_val = local64_read(&hwc->prev_count) & max_period;
680 #if defined(CONFIG_32BIT)
681 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
682 				  flag, init_val, init_val >> 32, 0);
683 #else
684 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
685 				  flag, init_val, 0, 0);
686 #endif
687 			perf_event_update_userpage(event);
688 		}
689 		ctr_ovf_mask = ctr_ovf_mask >> 1;
690 		idx++;
691 	}
692 }
693 
694 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
695 {
696 	struct perf_sample_data data;
697 	struct pt_regs *regs;
698 	struct hw_perf_event *hw_evt;
699 	union sbi_pmu_ctr_info *info;
700 	int lidx, hidx, fidx;
701 	struct riscv_pmu *pmu;
702 	struct perf_event *event;
703 	unsigned long overflow;
704 	unsigned long overflowed_ctrs = 0;
705 	struct cpu_hw_events *cpu_hw_evt = dev;
706 	u64 start_clock = sched_clock();
707 
708 	if (WARN_ON_ONCE(!cpu_hw_evt))
709 		return IRQ_NONE;
710 
711 	/* Firmware counter don't support overflow yet */
712 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
713 	if (fidx == RISCV_MAX_COUNTERS) {
714 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
715 		return IRQ_NONE;
716 	}
717 
718 	event = cpu_hw_evt->events[fidx];
719 	if (!event) {
720 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
721 		return IRQ_NONE;
722 	}
723 
724 	pmu = to_riscv_pmu(event->pmu);
725 	pmu_sbi_stop_hw_ctrs(pmu);
726 
727 	/* Overflow status register should only be read after counter are stopped */
728 	ALT_SBI_PMU_OVERFLOW(overflow);
729 
730 	/*
731 	 * Overflow interrupt pending bit should only be cleared after stopping
732 	 * all the counters to avoid any race condition.
733 	 */
734 	ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
735 
736 	/* No overflow bit is set */
737 	if (!overflow)
738 		return IRQ_NONE;
739 
740 	regs = get_irq_regs();
741 
742 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
743 		struct perf_event *event = cpu_hw_evt->events[lidx];
744 
745 		/* Skip if invalid event or user did not request a sampling */
746 		if (!event || !is_sampling_event(event))
747 			continue;
748 
749 		info = &pmu_ctr_list[lidx];
750 		/* Do a sanity check */
751 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
752 			continue;
753 
754 		/* compute hardware counter index */
755 		hidx = info->csr - CSR_CYCLE;
756 		/* check if the corresponding bit is set in sscountovf */
757 		if (!(overflow & BIT(hidx)))
758 			continue;
759 
760 		/*
761 		 * Keep a track of overflowed counters so that they can be started
762 		 * with updated initial value.
763 		 */
764 		overflowed_ctrs |= BIT(lidx);
765 		hw_evt = &event->hw;
766 		riscv_pmu_event_update(event);
767 		perf_sample_data_init(&data, 0, hw_evt->last_period);
768 		if (riscv_pmu_event_set_period(event)) {
769 			/*
770 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
771 			 * to avoid throttling here. As per the specification, the
772 			 * interrupt remains disabled until the OF bit is set.
773 			 * Interrupts are enabled again only during the start.
774 			 * TODO: We will need to stop the guest counters once
775 			 * virtualization support is added.
776 			 */
777 			perf_event_overflow(event, &data, regs);
778 		}
779 	}
780 
781 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
782 	perf_sample_event_took(sched_clock() - start_clock);
783 
784 	return IRQ_HANDLED;
785 }
786 
787 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
788 {
789 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
790 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
791 
792 	/*
793 	 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the
794 	 * legacy option but that will be removed in the future.
795 	 */
796 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
797 		csr_write(CSR_SCOUNTEREN, 0x7);
798 	else
799 		csr_write(CSR_SCOUNTEREN, 0x2);
800 
801 	/* Stop all the counters so that they can be enabled from perf */
802 	pmu_sbi_stop_all(pmu);
803 
804 	if (riscv_pmu_use_irq) {
805 		cpu_hw_evt->irq = riscv_pmu_irq;
806 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
807 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
808 	}
809 
810 	return 0;
811 }
812 
813 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
814 {
815 	if (riscv_pmu_use_irq) {
816 		disable_percpu_irq(riscv_pmu_irq);
817 	}
818 
819 	/* Disable all counters access for user mode now */
820 	csr_write(CSR_SCOUNTEREN, 0x0);
821 
822 	return 0;
823 }
824 
825 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
826 {
827 	int ret;
828 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
829 	struct irq_domain *domain = NULL;
830 
831 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
832 		riscv_pmu_irq_num = RV_IRQ_PMU;
833 		riscv_pmu_use_irq = true;
834 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
835 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
836 		   riscv_cached_marchid(0) == 0 &&
837 		   riscv_cached_mimpid(0) == 0) {
838 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
839 		riscv_pmu_use_irq = true;
840 	} else if (riscv_isa_extension_available(NULL, XANDESPMU) &&
841 		   IS_ENABLED(CONFIG_ANDES_CUSTOM_PMU)) {
842 		riscv_pmu_irq_num = ANDES_SLI_CAUSE_BASE + ANDES_RV_IRQ_PMOVI;
843 		riscv_pmu_use_irq = true;
844 	}
845 
846 	riscv_pmu_irq_mask = BIT(riscv_pmu_irq_num % BITS_PER_LONG);
847 
848 	if (!riscv_pmu_use_irq)
849 		return -EOPNOTSUPP;
850 
851 	domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
852 					  DOMAIN_BUS_ANY);
853 	if (!domain) {
854 		pr_err("Failed to find INTC IRQ root domain\n");
855 		return -ENODEV;
856 	}
857 
858 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
859 	if (!riscv_pmu_irq) {
860 		pr_err("Failed to map PMU interrupt for node\n");
861 		return -ENODEV;
862 	}
863 
864 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
865 	if (ret) {
866 		pr_err("registering percpu irq failed [%d]\n", ret);
867 		return ret;
868 	}
869 
870 	return 0;
871 }
872 
873 #ifdef CONFIG_CPU_PM
874 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
875 				void *v)
876 {
877 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
878 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
879 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
880 	struct perf_event *event;
881 	int idx;
882 
883 	if (!enabled)
884 		return NOTIFY_OK;
885 
886 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
887 		event = cpuc->events[idx];
888 		if (!event)
889 			continue;
890 
891 		switch (cmd) {
892 		case CPU_PM_ENTER:
893 			/*
894 			 * Stop and update the counter
895 			 */
896 			riscv_pmu_stop(event, PERF_EF_UPDATE);
897 			break;
898 		case CPU_PM_EXIT:
899 		case CPU_PM_ENTER_FAILED:
900 			/*
901 			 * Restore and enable the counter.
902 			 */
903 			riscv_pmu_start(event, PERF_EF_RELOAD);
904 			break;
905 		default:
906 			break;
907 		}
908 	}
909 
910 	return NOTIFY_OK;
911 }
912 
913 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
914 {
915 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
916 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
917 }
918 
919 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
920 {
921 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
922 }
923 #else
924 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
925 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
926 #endif
927 
928 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
929 {
930 	riscv_pm_pmu_unregister(pmu);
931 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
932 }
933 
934 static void pmu_sbi_event_init(struct perf_event *event)
935 {
936 	/*
937 	 * The permissions are set at event_init so that we do not depend
938 	 * on the sysctl value that can change.
939 	 */
940 	if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS)
941 		event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS;
942 	else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS)
943 		event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS;
944 	else
945 		event->hw.flags |= PERF_EVENT_FLAG_LEGACY;
946 }
947 
948 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm)
949 {
950 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
951 		return;
952 
953 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
954 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
955 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
956 			return;
957 		}
958 	}
959 
960 	/*
961 	 * The user mmapped the event to directly access it: this is where
962 	 * we determine based on sysctl_perf_user_access if we grant userspace
963 	 * the direct access to this event. That means that within the same
964 	 * task, some events may be directly accessible and some other may not,
965 	 * if the user changes the value of sysctl_perf_user_accesss in the
966 	 * meantime.
967 	 */
968 
969 	event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
970 
971 	/*
972 	 * We must enable userspace access *before* advertising in the user page
973 	 * that it is possible to do so to avoid any race.
974 	 * And we must notify all cpus here because threads that currently run
975 	 * on other cpus will try to directly access the counter too without
976 	 * calling pmu_sbi_ctr_start.
977 	 */
978 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
979 		on_each_cpu_mask(mm_cpumask(mm),
980 				 pmu_sbi_set_scounteren, (void *)event, 1);
981 }
982 
983 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm)
984 {
985 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
986 		return;
987 
988 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
989 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
990 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
991 			return;
992 		}
993 	}
994 
995 	/*
996 	 * Here we can directly remove user access since the user does not have
997 	 * access to the user page anymore so we avoid the racy window where the
998 	 * user could have read cap_user_rdpmc to true right before we disable
999 	 * it.
1000 	 */
1001 	event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT;
1002 
1003 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
1004 		on_each_cpu_mask(mm_cpumask(mm),
1005 				 pmu_sbi_reset_scounteren, (void *)event, 1);
1006 }
1007 
1008 static void riscv_pmu_update_counter_access(void *info)
1009 {
1010 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
1011 		csr_write(CSR_SCOUNTEREN, 0x7);
1012 	else
1013 		csr_write(CSR_SCOUNTEREN, 0x2);
1014 }
1015 
1016 static int riscv_pmu_proc_user_access_handler(struct ctl_table *table,
1017 					      int write, void *buffer,
1018 					      size_t *lenp, loff_t *ppos)
1019 {
1020 	int prev = sysctl_perf_user_access;
1021 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1022 
1023 	/*
1024 	 * Test against the previous value since we clear SCOUNTEREN when
1025 	 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should
1026 	 * not do that if that was already the case.
1027 	 */
1028 	if (ret || !write || prev == sysctl_perf_user_access)
1029 		return ret;
1030 
1031 	on_each_cpu(riscv_pmu_update_counter_access, NULL, 1);
1032 
1033 	return 0;
1034 }
1035 
1036 static struct ctl_table sbi_pmu_sysctl_table[] = {
1037 	{
1038 		.procname       = "perf_user_access",
1039 		.data		= &sysctl_perf_user_access,
1040 		.maxlen		= sizeof(unsigned int),
1041 		.mode           = 0644,
1042 		.proc_handler	= riscv_pmu_proc_user_access_handler,
1043 		.extra1		= SYSCTL_ZERO,
1044 		.extra2		= SYSCTL_TWO,
1045 	},
1046 	{ }
1047 };
1048 
1049 static int pmu_sbi_device_probe(struct platform_device *pdev)
1050 {
1051 	struct riscv_pmu *pmu = NULL;
1052 	int ret = -ENODEV;
1053 	int num_counters;
1054 
1055 	pr_info("SBI PMU extension is available\n");
1056 	pmu = riscv_pmu_alloc();
1057 	if (!pmu)
1058 		return -ENOMEM;
1059 
1060 	num_counters = pmu_sbi_find_num_ctrs();
1061 	if (num_counters < 0) {
1062 		pr_err("SBI PMU extension doesn't provide any counters\n");
1063 		goto out_free;
1064 	}
1065 
1066 	/* It is possible to get from SBI more than max number of counters */
1067 	if (num_counters > RISCV_MAX_COUNTERS) {
1068 		num_counters = RISCV_MAX_COUNTERS;
1069 		pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters);
1070 	}
1071 
1072 	/* cache all the information about counters now */
1073 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
1074 		goto out_free;
1075 
1076 	ret = pmu_sbi_setup_irqs(pmu, pdev);
1077 	if (ret < 0) {
1078 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
1079 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1080 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
1081 	}
1082 
1083 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
1084 	pmu->cmask = cmask;
1085 	pmu->ctr_start = pmu_sbi_ctr_start;
1086 	pmu->ctr_stop = pmu_sbi_ctr_stop;
1087 	pmu->event_map = pmu_sbi_event_map;
1088 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
1089 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
1090 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
1091 	pmu->ctr_read = pmu_sbi_ctr_read;
1092 	pmu->event_init = pmu_sbi_event_init;
1093 	pmu->event_mapped = pmu_sbi_event_mapped;
1094 	pmu->event_unmapped = pmu_sbi_event_unmapped;
1095 	pmu->csr_index = pmu_sbi_csr_index;
1096 
1097 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1098 	if (ret)
1099 		return ret;
1100 
1101 	ret = riscv_pm_pmu_register(pmu);
1102 	if (ret)
1103 		goto out_unregister;
1104 
1105 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
1106 	if (ret)
1107 		goto out_unregister;
1108 
1109 	register_sysctl("kernel", sbi_pmu_sysctl_table);
1110 
1111 	return 0;
1112 
1113 out_unregister:
1114 	riscv_pmu_destroy(pmu);
1115 
1116 out_free:
1117 	kfree(pmu);
1118 	return ret;
1119 }
1120 
1121 static struct platform_driver pmu_sbi_driver = {
1122 	.probe		= pmu_sbi_device_probe,
1123 	.driver		= {
1124 		.name	= RISCV_PMU_SBI_PDEV_NAME,
1125 	},
1126 };
1127 
1128 static int __init pmu_sbi_devinit(void)
1129 {
1130 	int ret;
1131 	struct platform_device *pdev;
1132 
1133 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
1134 	    !sbi_probe_extension(SBI_EXT_PMU)) {
1135 		return 0;
1136 	}
1137 
1138 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
1139 				      "perf/riscv/pmu:starting",
1140 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
1141 	if (ret) {
1142 		pr_err("CPU hotplug notifier could not be registered: %d\n",
1143 		       ret);
1144 		return ret;
1145 	}
1146 
1147 	ret = platform_driver_register(&pmu_sbi_driver);
1148 	if (ret)
1149 		return ret;
1150 
1151 	pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0);
1152 	if (IS_ERR(pdev)) {
1153 		platform_driver_unregister(&pmu_sbi_driver);
1154 		return PTR_ERR(pdev);
1155 	}
1156 
1157 	/* Notify legacy implementation that SBI pmu is available*/
1158 	riscv_pmu_legacy_skip_init();
1159 
1160 	return ret;
1161 }
1162 device_initcall(pmu_sbi_devinit)
1163