xref: /linux/drivers/perf/riscv_pmu_sbi.c (revision 68a052239fc4b351e961f698b824f7654a346091)
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 #include <linux/workqueue.h>
24 
25 #include <asm/errata_list.h>
26 #include <asm/sbi.h>
27 #include <asm/cpufeature.h>
28 #include <asm/vendor_extensions.h>
29 #include <asm/vendor_extensions/andes.h>
30 
31 #define ALT_SBI_PMU_OVERFLOW(__ovl)					\
32 asm volatile(ALTERNATIVE_2(						\
33 	"csrr %0, " __stringify(CSR_SCOUNTOVF),				\
34 	"csrr %0, " __stringify(THEAD_C9XX_CSR_SCOUNTEROF),		\
35 		THEAD_VENDOR_ID, ERRATA_THEAD_PMU,			\
36 		CONFIG_ERRATA_THEAD_PMU,				\
37 	"csrr %0, " __stringify(ANDES_CSR_SCOUNTEROF),			\
38 		ANDES_VENDOR_ID,					\
39 		RISCV_ISA_VENDOR_EXT_XANDESPMU + RISCV_VENDOR_EXT_ALTERNATIVES_BASE, \
40 		CONFIG_ANDES_CUSTOM_PMU)				\
41 	: "=r" (__ovl) :						\
42 	: "memory")
43 
44 #define ALT_SBI_PMU_OVF_CLEAR_PENDING(__irq_mask)			\
45 asm volatile(ALTERNATIVE(						\
46 	"csrc " __stringify(CSR_IP) ", %0\n\t",				\
47 	"csrc " __stringify(ANDES_CSR_SLIP) ", %0\n\t",			\
48 		ANDES_VENDOR_ID,					\
49 		RISCV_ISA_VENDOR_EXT_XANDESPMU + RISCV_VENDOR_EXT_ALTERNATIVES_BASE, \
50 		CONFIG_ANDES_CUSTOM_PMU)				\
51 	: : "r"(__irq_mask)						\
52 	: "memory")
53 
54 #define SYSCTL_NO_USER_ACCESS	0
55 #define SYSCTL_USER_ACCESS	1
56 #define SYSCTL_LEGACY		2
57 
58 #define PERF_EVENT_FLAG_NO_USER_ACCESS	BIT(SYSCTL_NO_USER_ACCESS)
59 #define PERF_EVENT_FLAG_USER_ACCESS	BIT(SYSCTL_USER_ACCESS)
60 #define PERF_EVENT_FLAG_LEGACY		BIT(SYSCTL_LEGACY)
61 
62 PMU_FORMAT_ATTR(event, "config:0-55");
63 PMU_FORMAT_ATTR(firmware, "config:62-63");
64 
65 static bool sbi_v2_available;
66 static bool sbi_v3_available;
67 static DEFINE_STATIC_KEY_FALSE(sbi_pmu_snapshot_available);
68 #define sbi_pmu_snapshot_available() \
69 	static_branch_unlikely(&sbi_pmu_snapshot_available)
70 
71 static struct attribute *riscv_arch_formats_attr[] = {
72 	&format_attr_event.attr,
73 	&format_attr_firmware.attr,
74 	NULL,
75 };
76 
77 static struct attribute_group riscv_pmu_format_group = {
78 	.name = "format",
79 	.attrs = riscv_arch_formats_attr,
80 };
81 
82 static const struct attribute_group *riscv_pmu_attr_groups[] = {
83 	&riscv_pmu_format_group,
84 	NULL,
85 };
86 
87 /* Allow user mode access by default */
88 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS;
89 
90 /*
91  * RISC-V doesn't have heterogeneous harts yet. This need to be part of
92  * per_cpu in case of harts with different pmu counters
93  */
94 static union sbi_pmu_ctr_info *pmu_ctr_list;
95 static bool riscv_pmu_use_irq;
96 static unsigned int riscv_pmu_irq_num;
97 static unsigned int riscv_pmu_irq_mask;
98 static unsigned int riscv_pmu_irq;
99 
100 /* Cache the available counters in a bitmask */
101 static unsigned long cmask;
102 
103 static int pmu_event_find_cache(u64 config);
104 struct sbi_pmu_event_data {
105 	union {
106 		union {
107 			struct hw_gen_event {
108 				uint32_t event_code:16;
109 				uint32_t event_type:4;
110 				uint32_t reserved:12;
111 			} hw_gen_event;
112 			struct hw_cache_event {
113 				uint32_t result_id:1;
114 				uint32_t op_id:2;
115 				uint32_t cache_id:13;
116 				uint32_t event_type:4;
117 				uint32_t reserved:12;
118 			} hw_cache_event;
119 		};
120 		uint32_t event_idx;
121 	};
122 };
123 
124 static struct sbi_pmu_event_data pmu_hw_event_map[] = {
125 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
126 							SBI_PMU_HW_CPU_CYCLES,
127 							SBI_PMU_EVENT_TYPE_HW, 0}},
128 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
129 							SBI_PMU_HW_INSTRUCTIONS,
130 							SBI_PMU_EVENT_TYPE_HW, 0}},
131 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
132 							SBI_PMU_HW_CACHE_REFERENCES,
133 							SBI_PMU_EVENT_TYPE_HW, 0}},
134 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
135 							SBI_PMU_HW_CACHE_MISSES,
136 							SBI_PMU_EVENT_TYPE_HW, 0}},
137 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
138 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
139 							SBI_PMU_EVENT_TYPE_HW, 0}},
140 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
141 							SBI_PMU_HW_BRANCH_MISSES,
142 							SBI_PMU_EVENT_TYPE_HW, 0}},
143 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
144 							SBI_PMU_HW_BUS_CYCLES,
145 							SBI_PMU_EVENT_TYPE_HW, 0}},
146 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
147 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
148 							SBI_PMU_EVENT_TYPE_HW, 0}},
149 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
150 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
151 							SBI_PMU_EVENT_TYPE_HW, 0}},
152 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
153 							SBI_PMU_HW_REF_CPU_CYCLES,
154 							SBI_PMU_EVENT_TYPE_HW, 0}},
155 };
156 
157 #define C(x) PERF_COUNT_HW_CACHE_##x
158 static struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
159 [PERF_COUNT_HW_CACHE_OP_MAX]
160 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
161 	[C(L1D)] = {
162 		[C(OP_READ)] = {
163 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
164 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
165 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
166 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
167 		},
168 		[C(OP_WRITE)] = {
169 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
170 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
171 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
172 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
173 		},
174 		[C(OP_PREFETCH)] = {
175 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
176 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
177 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
178 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
179 		},
180 	},
181 	[C(L1I)] = {
182 		[C(OP_READ)] = {
183 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
184 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
185 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
186 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
187 		},
188 		[C(OP_WRITE)] = {
189 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
190 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
191 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
192 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
193 		},
194 		[C(OP_PREFETCH)] = {
195 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
196 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
197 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
198 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
199 		},
200 	},
201 	[C(LL)] = {
202 		[C(OP_READ)] = {
203 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
204 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
205 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
206 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
207 		},
208 		[C(OP_WRITE)] = {
209 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
210 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
211 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
212 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
213 		},
214 		[C(OP_PREFETCH)] = {
215 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
216 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
217 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
218 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
219 		},
220 	},
221 	[C(DTLB)] = {
222 		[C(OP_READ)] = {
223 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
224 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
225 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
226 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
227 		},
228 		[C(OP_WRITE)] = {
229 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
230 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
231 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
232 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
233 		},
234 		[C(OP_PREFETCH)] = {
235 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
236 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
237 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
238 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
239 		},
240 	},
241 	[C(ITLB)] = {
242 		[C(OP_READ)] = {
243 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
244 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
245 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
246 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
247 		},
248 		[C(OP_WRITE)] = {
249 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
250 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
251 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
252 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
253 		},
254 		[C(OP_PREFETCH)] = {
255 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
256 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
257 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
258 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
259 		},
260 	},
261 	[C(BPU)] = {
262 		[C(OP_READ)] = {
263 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
264 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
265 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
266 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
267 		},
268 		[C(OP_WRITE)] = {
269 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
270 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
271 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
272 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
273 		},
274 		[C(OP_PREFETCH)] = {
275 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
276 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
277 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
278 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
279 		},
280 	},
281 	[C(NODE)] = {
282 		[C(OP_READ)] = {
283 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
284 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
285 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
286 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
287 		},
288 		[C(OP_WRITE)] = {
289 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
290 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
291 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
292 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
293 		},
294 		[C(OP_PREFETCH)] = {
295 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
296 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
297 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
298 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
299 		},
300 	},
301 };
302 
303 static int pmu_sbi_check_event_info(void)
304 {
305 	int num_events = ARRAY_SIZE(pmu_hw_event_map) + PERF_COUNT_HW_CACHE_MAX *
306 			 PERF_COUNT_HW_CACHE_OP_MAX * PERF_COUNT_HW_CACHE_RESULT_MAX;
307 	struct riscv_pmu_event_info *event_info_shmem;
308 	phys_addr_t base_addr;
309 	int i, j, k, result = 0, count = 0;
310 	struct sbiret ret;
311 
312 	event_info_shmem = kcalloc(num_events, sizeof(*event_info_shmem), GFP_KERNEL);
313 	if (!event_info_shmem)
314 		return -ENOMEM;
315 
316 	for (i = 0; i < ARRAY_SIZE(pmu_hw_event_map); i++)
317 		event_info_shmem[count++].event_idx = pmu_hw_event_map[i].event_idx;
318 
319 	for (i = 0; i < ARRAY_SIZE(pmu_cache_event_map); i++) {
320 		for (j = 0; j < ARRAY_SIZE(pmu_cache_event_map[i]); j++) {
321 			for (k = 0; k < ARRAY_SIZE(pmu_cache_event_map[i][j]); k++)
322 				event_info_shmem[count++].event_idx =
323 							pmu_cache_event_map[i][j][k].event_idx;
324 		}
325 	}
326 
327 	base_addr = __pa(event_info_shmem);
328 	if (IS_ENABLED(CONFIG_32BIT))
329 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_EVENT_GET_INFO, lower_32_bits(base_addr),
330 				upper_32_bits(base_addr), count, 0, 0, 0);
331 	else
332 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_EVENT_GET_INFO, base_addr, 0,
333 				count, 0, 0, 0);
334 	if (ret.error) {
335 		result = -EOPNOTSUPP;
336 		goto free_mem;
337 	}
338 
339 	for (i = 0; i < ARRAY_SIZE(pmu_hw_event_map); i++) {
340 		if (!(event_info_shmem[i].output & RISCV_PMU_EVENT_INFO_OUTPUT_MASK))
341 			pmu_hw_event_map[i].event_idx = -ENOENT;
342 	}
343 
344 	count = ARRAY_SIZE(pmu_hw_event_map);
345 
346 	for (i = 0; i < ARRAY_SIZE(pmu_cache_event_map); i++) {
347 		for (j = 0; j < ARRAY_SIZE(pmu_cache_event_map[i]); j++) {
348 			for (k = 0; k < ARRAY_SIZE(pmu_cache_event_map[i][j]); k++) {
349 				if (!(event_info_shmem[count].output &
350 				      RISCV_PMU_EVENT_INFO_OUTPUT_MASK))
351 					pmu_cache_event_map[i][j][k].event_idx = -ENOENT;
352 				count++;
353 			}
354 		}
355 	}
356 
357 free_mem:
358 	kfree(event_info_shmem);
359 
360 	return result;
361 }
362 
363 static void pmu_sbi_check_event(struct sbi_pmu_event_data *edata)
364 {
365 	struct sbiret ret;
366 
367 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH,
368 			0, cmask, 0, edata->event_idx, 0, 0);
369 	if (!ret.error) {
370 		sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
371 			  ret.value, 0x1, SBI_PMU_STOP_FLAG_RESET, 0, 0, 0);
372 	} else if (ret.error == SBI_ERR_NOT_SUPPORTED) {
373 		/* This event cannot be monitored by any counter */
374 		edata->event_idx = -ENOENT;
375 	}
376 }
377 
378 static void pmu_sbi_check_std_events(struct work_struct *work)
379 {
380 	int ret;
381 
382 	if (sbi_v3_available) {
383 		ret = pmu_sbi_check_event_info();
384 		if (ret)
385 			pr_err("pmu_sbi_check_event_info failed with error %d\n", ret);
386 		return;
387 	}
388 
389 	for (int i = 0; i < ARRAY_SIZE(pmu_hw_event_map); i++)
390 		pmu_sbi_check_event(&pmu_hw_event_map[i]);
391 
392 	for (int i = 0; i < ARRAY_SIZE(pmu_cache_event_map); i++)
393 		for (int j = 0; j < ARRAY_SIZE(pmu_cache_event_map[i]); j++)
394 			for (int k = 0; k < ARRAY_SIZE(pmu_cache_event_map[i][j]); k++)
395 				pmu_sbi_check_event(&pmu_cache_event_map[i][j][k]);
396 }
397 
398 static DECLARE_WORK(check_std_events_work, pmu_sbi_check_std_events);
399 
400 static int pmu_sbi_ctr_get_width(int idx)
401 {
402 	return pmu_ctr_list[idx].width;
403 }
404 
405 static bool pmu_sbi_ctr_is_fw(int cidx)
406 {
407 	union sbi_pmu_ctr_info *info;
408 
409 	info = &pmu_ctr_list[cidx];
410 	if (!info)
411 		return false;
412 
413 	return info->type == SBI_PMU_CTR_TYPE_FW;
414 }
415 
416 int riscv_pmu_get_event_info(u32 type, u64 config, u64 *econfig)
417 {
418 	int ret = -ENOENT;
419 
420 	switch (type) {
421 	case PERF_TYPE_HARDWARE:
422 		if (config >= PERF_COUNT_HW_MAX)
423 			return -EINVAL;
424 		ret = pmu_hw_event_map[config].event_idx;
425 		break;
426 	case PERF_TYPE_HW_CACHE:
427 		ret = pmu_event_find_cache(config);
428 		break;
429 	case PERF_TYPE_RAW:
430 		/*
431 		 * As per SBI v0.3 specification,
432 		 *  -- the upper 16 bits must be unused for a hardware raw event.
433 		 * As per SBI v2.0 specification,
434 		 *  -- the upper 8 bits must be unused for a hardware raw event.
435 		 * Bits 63:62 are used to distinguish between raw events
436 		 * 00 - Hardware raw event
437 		 * 10 - SBI firmware events
438 		 * 11 - Risc-V platform specific firmware event
439 		 */
440 		switch (config >> 62) {
441 		case 0:
442 			if (sbi_v3_available) {
443 			/* Return error any bits [56-63] is set  as it is not allowed by the spec */
444 				if (!(config & ~RISCV_PMU_RAW_EVENT_V2_MASK)) {
445 					if (econfig)
446 						*econfig = config & RISCV_PMU_RAW_EVENT_V2_MASK;
447 					ret = RISCV_PMU_RAW_EVENT_V2_IDX;
448 				}
449 			/* Return error any bits [48-63] is set  as it is not allowed by the spec */
450 			} else if (!(config & ~RISCV_PMU_RAW_EVENT_MASK)) {
451 				if (econfig)
452 					*econfig = config & RISCV_PMU_RAW_EVENT_MASK;
453 				ret = RISCV_PMU_RAW_EVENT_IDX;
454 			}
455 			break;
456 		case 2:
457 			ret = (config & 0xFFFF) | (SBI_PMU_EVENT_TYPE_FW << 16);
458 			break;
459 		case 3:
460 			/*
461 			 * For Risc-V platform specific firmware events
462 			 * Event code - 0xFFFF
463 			 * Event data - raw event encoding
464 			 */
465 			ret = SBI_PMU_EVENT_TYPE_FW << 16 | RISCV_PLAT_FW_EVENT;
466 			if (econfig)
467 				*econfig = config & RISCV_PMU_PLAT_FW_EVENT_MASK;
468 			break;
469 		default:
470 			break;
471 		}
472 		break;
473 	default:
474 		break;
475 	}
476 
477 	return ret;
478 }
479 EXPORT_SYMBOL_GPL(riscv_pmu_get_event_info);
480 
481 /*
482  * Returns the counter width of a programmable counter and number of hardware
483  * counters. As we don't support heterogeneous CPUs yet, it is okay to just
484  * return the counter width of the first programmable counter.
485  */
486 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
487 {
488 	int i;
489 	union sbi_pmu_ctr_info *info;
490 	u32 hpm_width = 0, hpm_count = 0;
491 
492 	if (!cmask)
493 		return -EINVAL;
494 
495 	for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
496 		info = &pmu_ctr_list[i];
497 		if (!info)
498 			continue;
499 		if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET)
500 			hpm_width = info->width;
501 		if (info->type == SBI_PMU_CTR_TYPE_HW)
502 			hpm_count++;
503 	}
504 
505 	*hw_ctr_width = hpm_width;
506 	*num_hw_ctr = hpm_count;
507 
508 	return 0;
509 }
510 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
511 
512 static uint8_t pmu_sbi_csr_index(struct perf_event *event)
513 {
514 	return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE;
515 }
516 
517 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
518 {
519 	unsigned long cflags = 0;
520 	bool guest_events = false;
521 
522 	if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
523 		guest_events = true;
524 	if (event->attr.exclude_kernel)
525 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH;
526 	if (event->attr.exclude_user)
527 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH;
528 	if (guest_events && event->attr.exclude_hv)
529 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
530 	if (event->attr.exclude_host)
531 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH;
532 	if (event->attr.exclude_guest)
533 		cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH;
534 
535 	return cflags;
536 }
537 
538 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
539 {
540 	struct hw_perf_event *hwc = &event->hw;
541 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
542 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
543 	struct sbiret ret;
544 	int idx;
545 	uint64_t cbase = 0, cmask = rvpmu->cmask;
546 	unsigned long cflags = 0;
547 
548 	cflags = pmu_sbi_get_filter_flags(event);
549 
550 	/*
551 	 * In legacy mode, we have to force the fixed counters for those events
552 	 * but not in the user access mode as we want to use the other counters
553 	 * that support sampling/filtering.
554 	 */
555 	if ((hwc->flags & PERF_EVENT_FLAG_LEGACY) && (event->attr.type == PERF_TYPE_HARDWARE)) {
556 		if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
557 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
558 			cmask = 1;
559 		} else if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS) {
560 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
561 			cmask = BIT(CSR_INSTRET - CSR_CYCLE);
562 		}
563 	}
564 
565 	/* retrieve the available counter index */
566 #if defined(CONFIG_32BIT)
567 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
568 			cmask, cflags, hwc->event_base, hwc->config,
569 			hwc->config >> 32);
570 #else
571 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
572 			cmask, cflags, hwc->event_base, hwc->config, 0);
573 #endif
574 	if (ret.error) {
575 		pr_debug("Not able to find a counter for event %lx config %llx\n",
576 			hwc->event_base, hwc->config);
577 		return sbi_err_map_linux_errno(ret.error);
578 	}
579 
580 	idx = ret.value;
581 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
582 		return -ENOENT;
583 
584 	/* Additional sanity check for the counter id */
585 	if (pmu_sbi_ctr_is_fw(idx)) {
586 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
587 			return idx;
588 	} else {
589 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
590 			return idx;
591 	}
592 
593 	return -ENOENT;
594 }
595 
596 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
597 {
598 
599 	struct hw_perf_event *hwc = &event->hw;
600 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
601 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
602 	int idx = hwc->idx;
603 
604 	if (pmu_sbi_ctr_is_fw(idx))
605 		clear_bit(idx, cpuc->used_fw_ctrs);
606 	else
607 		clear_bit(idx, cpuc->used_hw_ctrs);
608 }
609 
610 static int pmu_event_find_cache(u64 config)
611 {
612 	unsigned int cache_type, cache_op, cache_result, ret;
613 
614 	cache_type = (config >>  0) & 0xff;
615 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
616 		return -EINVAL;
617 
618 	cache_op = (config >>  8) & 0xff;
619 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
620 		return -EINVAL;
621 
622 	cache_result = (config >> 16) & 0xff;
623 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
624 		return -EINVAL;
625 
626 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
627 
628 	return ret;
629 }
630 
631 static bool pmu_sbi_is_fw_event(struct perf_event *event)
632 {
633 	u32 type = event->attr.type;
634 	u64 config = event->attr.config;
635 
636 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
637 		return true;
638 	else
639 		return false;
640 }
641 
642 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
643 {
644 	u32 type = event->attr.type;
645 	u64 config = event->attr.config;
646 
647 	/*
648 	 * Ensure we are finished checking standard hardware events for
649 	 * validity before allowing userspace to configure any events.
650 	 */
651 	flush_work(&check_std_events_work);
652 
653 	return riscv_pmu_get_event_info(type, config, econfig);
654 }
655 
656 static void pmu_sbi_snapshot_free(struct riscv_pmu *pmu)
657 {
658 	int cpu;
659 
660 	for_each_possible_cpu(cpu) {
661 		struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
662 
663 		if (!cpu_hw_evt->snapshot_addr)
664 			continue;
665 
666 		free_page((unsigned long)cpu_hw_evt->snapshot_addr);
667 		cpu_hw_evt->snapshot_addr = NULL;
668 		cpu_hw_evt->snapshot_addr_phys = 0;
669 	}
670 }
671 
672 static int pmu_sbi_snapshot_alloc(struct riscv_pmu *pmu)
673 {
674 	int cpu;
675 	struct page *snapshot_page;
676 
677 	for_each_possible_cpu(cpu) {
678 		struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
679 
680 		snapshot_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
681 		if (!snapshot_page) {
682 			pmu_sbi_snapshot_free(pmu);
683 			return -ENOMEM;
684 		}
685 		cpu_hw_evt->snapshot_addr = page_to_virt(snapshot_page);
686 		cpu_hw_evt->snapshot_addr_phys = page_to_phys(snapshot_page);
687 	}
688 
689 	return 0;
690 }
691 
692 static int pmu_sbi_snapshot_disable(void)
693 {
694 	struct sbiret ret;
695 
696 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM, SBI_SHMEM_DISABLE,
697 			SBI_SHMEM_DISABLE, 0, 0, 0, 0);
698 	if (ret.error) {
699 		pr_warn("failed to disable snapshot shared memory\n");
700 		return sbi_err_map_linux_errno(ret.error);
701 	}
702 
703 	return 0;
704 }
705 
706 static int pmu_sbi_snapshot_setup(struct riscv_pmu *pmu, int cpu)
707 {
708 	struct cpu_hw_events *cpu_hw_evt;
709 	struct sbiret ret = {0};
710 
711 	cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
712 	if (!cpu_hw_evt->snapshot_addr_phys)
713 		return -EINVAL;
714 
715 	if (cpu_hw_evt->snapshot_set_done)
716 		return 0;
717 
718 	if (IS_ENABLED(CONFIG_32BIT))
719 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM,
720 				cpu_hw_evt->snapshot_addr_phys,
721 				(u64)(cpu_hw_evt->snapshot_addr_phys) >> 32, 0, 0, 0, 0);
722 	else
723 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM,
724 				cpu_hw_evt->snapshot_addr_phys, 0, 0, 0, 0, 0);
725 
726 	/* Free up the snapshot area memory and fall back to SBI PMU calls without snapshot */
727 	if (ret.error) {
728 		if (ret.error != SBI_ERR_NOT_SUPPORTED)
729 			pr_warn("pmu snapshot setup failed with error %ld\n", ret.error);
730 		return sbi_err_map_linux_errno(ret.error);
731 	}
732 
733 	memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS);
734 	cpu_hw_evt->snapshot_set_done = true;
735 
736 	return 0;
737 }
738 
739 static u64 pmu_sbi_ctr_read(struct perf_event *event)
740 {
741 	struct hw_perf_event *hwc = &event->hw;
742 	int idx = hwc->idx;
743 	struct sbiret ret;
744 	u64 val = 0;
745 	struct riscv_pmu *pmu = to_riscv_pmu(event->pmu);
746 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
747 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
748 	union sbi_pmu_ctr_info info = pmu_ctr_list[idx];
749 
750 	/* Read the value from the shared memory directly only if counter is stopped */
751 	if (sbi_pmu_snapshot_available() && (hwc->state & PERF_HES_STOPPED)) {
752 		val = sdata->ctr_values[idx];
753 		return val;
754 	}
755 
756 	if (pmu_sbi_is_fw_event(event)) {
757 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
758 				hwc->idx, 0, 0, 0, 0, 0);
759 		if (ret.error)
760 			return 0;
761 
762 		val = ret.value;
763 		if (IS_ENABLED(CONFIG_32BIT) && sbi_v2_available && info.width >= 32) {
764 			ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ_HI,
765 					hwc->idx, 0, 0, 0, 0, 0);
766 			if (!ret.error)
767 				val |= ((u64)ret.value << 32);
768 			else
769 				WARN_ONCE(1, "Unable to read upper 32 bits of firmware counter error: %ld\n",
770 					  ret.error);
771 		}
772 	} else {
773 		val = riscv_pmu_ctr_read_csr(info.csr);
774 		if (IS_ENABLED(CONFIG_32BIT))
775 			val |= ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 32;
776 	}
777 
778 	return val;
779 }
780 
781 static void pmu_sbi_set_scounteren(void *arg)
782 {
783 	struct perf_event *event = (struct perf_event *)arg;
784 
785 	if (event->hw.idx != -1)
786 		csr_write(CSR_SCOUNTEREN,
787 			  csr_read(CSR_SCOUNTEREN) | BIT(pmu_sbi_csr_index(event)));
788 }
789 
790 static void pmu_sbi_reset_scounteren(void *arg)
791 {
792 	struct perf_event *event = (struct perf_event *)arg;
793 
794 	if (event->hw.idx != -1)
795 		csr_write(CSR_SCOUNTEREN,
796 			  csr_read(CSR_SCOUNTEREN) & ~BIT(pmu_sbi_csr_index(event)));
797 }
798 
799 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
800 {
801 	struct sbiret ret;
802 	struct hw_perf_event *hwc = &event->hw;
803 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
804 
805 	/* There is no benefit setting SNAPSHOT FLAG for a single counter */
806 #if defined(CONFIG_32BIT)
807 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
808 			1, flag, ival, ival >> 32, 0);
809 #else
810 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
811 			1, flag, ival, 0, 0);
812 #endif
813 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
814 		pr_err("Starting counter idx %d failed with error %d\n",
815 			hwc->idx, sbi_err_map_linux_errno(ret.error));
816 
817 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
818 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
819 		pmu_sbi_set_scounteren((void *)event);
820 }
821 
822 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
823 {
824 	struct sbiret ret;
825 	struct hw_perf_event *hwc = &event->hw;
826 	struct riscv_pmu *pmu = to_riscv_pmu(event->pmu);
827 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
828 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
829 
830 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
831 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
832 		pmu_sbi_reset_scounteren((void *)event);
833 
834 	if (sbi_pmu_snapshot_available())
835 		flag |= SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT;
836 
837 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
838 	if (!ret.error && sbi_pmu_snapshot_available()) {
839 		/*
840 		 * The counter snapshot is based on the index base specified by hwc->idx.
841 		 * The actual counter value is updated in shared memory at index 0 when counter
842 		 * mask is 0x01. To ensure accurate counter values, it's necessary to transfer
843 		 * the counter value to shared memory. However, if hwc->idx is zero, the counter
844 		 * value is already correctly updated in shared memory, requiring no further
845 		 * adjustment.
846 		 */
847 		if (hwc->idx > 0) {
848 			sdata->ctr_values[hwc->idx] = sdata->ctr_values[0];
849 			sdata->ctr_values[0] = 0;
850 		}
851 	} else if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
852 		flag != SBI_PMU_STOP_FLAG_RESET) {
853 		pr_err("Stopping counter idx %d failed with error %d\n",
854 			hwc->idx, sbi_err_map_linux_errno(ret.error));
855 	}
856 }
857 
858 static int pmu_sbi_find_num_ctrs(void)
859 {
860 	struct sbiret ret;
861 
862 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
863 	if (!ret.error)
864 		return ret.value;
865 	else
866 		return sbi_err_map_linux_errno(ret.error);
867 }
868 
869 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
870 {
871 	struct sbiret ret;
872 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
873 	union sbi_pmu_ctr_info cinfo;
874 
875 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
876 	if (!pmu_ctr_list)
877 		return -ENOMEM;
878 
879 	for (i = 0; i < nctr; i++) {
880 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
881 		if (ret.error)
882 			/* The logical counter ids are not expected to be contiguous */
883 			continue;
884 
885 		*mask |= BIT(i);
886 
887 		cinfo.value = ret.value;
888 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
889 			num_fw_ctr++;
890 		else
891 			num_hw_ctr++;
892 		pmu_ctr_list[i].value = cinfo.value;
893 	}
894 
895 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
896 
897 	return 0;
898 }
899 
900 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
901 {
902 	/*
903 	 * No need to check the error because we are disabling all the counters
904 	 * which may include counters that are not enabled yet.
905 	 */
906 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
907 		  0, pmu->cmask, SBI_PMU_STOP_FLAG_RESET, 0, 0, 0);
908 }
909 
910 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
911 {
912 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
913 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
914 	unsigned long flag = 0;
915 	int i, idx;
916 	struct sbiret ret;
917 	u64 temp_ctr_overflow_mask = 0;
918 
919 	if (sbi_pmu_snapshot_available())
920 		flag = SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT;
921 
922 	/* Reset the shadow copy to avoid save/restore any value from previous overflow */
923 	memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS);
924 
925 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
926 		/* No need to check the error here as we can't do anything about the error */
927 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, i * BITS_PER_LONG,
928 				cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0);
929 		if (!ret.error && sbi_pmu_snapshot_available()) {
930 			/* Save the counter values to avoid clobbering */
931 			for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG)
932 				cpu_hw_evt->snapshot_cval_shcopy[i * BITS_PER_LONG + idx] =
933 							sdata->ctr_values[idx];
934 			/* Save the overflow mask to avoid clobbering */
935 			temp_ctr_overflow_mask |= sdata->ctr_overflow_mask << (i * BITS_PER_LONG);
936 		}
937 	}
938 
939 	/* Restore the counter values to the shared memory for used hw counters */
940 	if (sbi_pmu_snapshot_available()) {
941 		for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS)
942 			sdata->ctr_values[idx] = cpu_hw_evt->snapshot_cval_shcopy[idx];
943 		if (temp_ctr_overflow_mask)
944 			sdata->ctr_overflow_mask = temp_ctr_overflow_mask;
945 	}
946 }
947 
948 /*
949  * This function starts all the used counters in two step approach.
950  * Any counter that did not overflow can be start in a single step
951  * while the overflowed counters need to be started with updated initialization
952  * value.
953  */
954 static inline void pmu_sbi_start_ovf_ctrs_sbi(struct cpu_hw_events *cpu_hw_evt,
955 					      u64 ctr_ovf_mask)
956 {
957 	int idx = 0, i;
958 	struct perf_event *event;
959 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
960 	unsigned long ctr_start_mask = 0;
961 	uint64_t max_period;
962 	struct hw_perf_event *hwc;
963 	u64 init_val = 0;
964 
965 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
966 		ctr_start_mask = cpu_hw_evt->used_hw_ctrs[i] & ~ctr_ovf_mask;
967 		/* Start all the counters that did not overflow in a single shot */
968 		if (ctr_start_mask) {
969 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, i * BITS_PER_LONG,
970 				  ctr_start_mask, 0, 0, 0, 0);
971 		}
972 	}
973 
974 	/* Reinitialize and start all the counter that overflowed */
975 	while (ctr_ovf_mask) {
976 		if (ctr_ovf_mask & 0x01) {
977 			event = cpu_hw_evt->events[idx];
978 			hwc = &event->hw;
979 			max_period = riscv_pmu_ctr_get_width_mask(event);
980 			init_val = local64_read(&hwc->prev_count) & max_period;
981 #if defined(CONFIG_32BIT)
982 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
983 				  flag, init_val, init_val >> 32, 0);
984 #else
985 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
986 				  flag, init_val, 0, 0);
987 #endif
988 			perf_event_update_userpage(event);
989 		}
990 		ctr_ovf_mask = ctr_ovf_mask >> 1;
991 		idx++;
992 	}
993 }
994 
995 static inline void pmu_sbi_start_ovf_ctrs_snapshot(struct cpu_hw_events *cpu_hw_evt,
996 						   u64 ctr_ovf_mask)
997 {
998 	int i, idx = 0;
999 	struct perf_event *event;
1000 	unsigned long flag = SBI_PMU_START_FLAG_INIT_SNAPSHOT;
1001 	u64 max_period, init_val = 0;
1002 	struct hw_perf_event *hwc;
1003 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
1004 
1005 	for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
1006 		if (ctr_ovf_mask & BIT(idx)) {
1007 			event = cpu_hw_evt->events[idx];
1008 			hwc = &event->hw;
1009 			max_period = riscv_pmu_ctr_get_width_mask(event);
1010 			init_val = local64_read(&hwc->prev_count) & max_period;
1011 			cpu_hw_evt->snapshot_cval_shcopy[idx] = init_val;
1012 		}
1013 		/*
1014 		 * We do not need to update the non-overflow counters the previous
1015 		 * value should have been there already.
1016 		 */
1017 	}
1018 
1019 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
1020 		/* Restore the counter values to relative indices for used hw counters */
1021 		for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG)
1022 			sdata->ctr_values[idx] =
1023 					cpu_hw_evt->snapshot_cval_shcopy[idx + i * BITS_PER_LONG];
1024 		/* Start all the counters in a single shot */
1025 		sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx * BITS_PER_LONG,
1026 			  cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0);
1027 	}
1028 }
1029 
1030 static void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
1031 					u64 ctr_ovf_mask)
1032 {
1033 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
1034 
1035 	if (sbi_pmu_snapshot_available())
1036 		pmu_sbi_start_ovf_ctrs_snapshot(cpu_hw_evt, ctr_ovf_mask);
1037 	else
1038 		pmu_sbi_start_ovf_ctrs_sbi(cpu_hw_evt, ctr_ovf_mask);
1039 }
1040 
1041 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
1042 {
1043 	struct perf_sample_data data;
1044 	struct pt_regs *regs;
1045 	struct hw_perf_event *hw_evt;
1046 	union sbi_pmu_ctr_info *info;
1047 	int lidx, hidx, fidx;
1048 	struct riscv_pmu *pmu;
1049 	struct perf_event *event;
1050 	u64 overflow;
1051 	u64 overflowed_ctrs = 0;
1052 	struct cpu_hw_events *cpu_hw_evt = dev;
1053 	u64 start_clock = sched_clock();
1054 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
1055 
1056 	if (WARN_ON_ONCE(!cpu_hw_evt))
1057 		return IRQ_NONE;
1058 
1059 	/* Firmware counter don't support overflow yet */
1060 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
1061 	if (fidx == RISCV_MAX_COUNTERS) {
1062 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
1063 		return IRQ_NONE;
1064 	}
1065 
1066 	event = cpu_hw_evt->events[fidx];
1067 	if (!event) {
1068 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
1069 		return IRQ_NONE;
1070 	}
1071 
1072 	pmu = to_riscv_pmu(event->pmu);
1073 	pmu_sbi_stop_hw_ctrs(pmu);
1074 
1075 	/* Overflow status register should only be read after counter are stopped */
1076 	if (sbi_pmu_snapshot_available())
1077 		overflow = sdata->ctr_overflow_mask;
1078 	else
1079 		ALT_SBI_PMU_OVERFLOW(overflow);
1080 
1081 	/*
1082 	 * Overflow interrupt pending bit should only be cleared after stopping
1083 	 * all the counters to avoid any race condition.
1084 	 */
1085 	ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
1086 
1087 	/* No overflow bit is set */
1088 	if (!overflow)
1089 		return IRQ_NONE;
1090 
1091 	regs = get_irq_regs();
1092 
1093 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
1094 		struct perf_event *event = cpu_hw_evt->events[lidx];
1095 
1096 		/* Skip if invalid event or user did not request a sampling */
1097 		if (!event || !is_sampling_event(event))
1098 			continue;
1099 
1100 		info = &pmu_ctr_list[lidx];
1101 		/* Do a sanity check */
1102 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
1103 			continue;
1104 
1105 		if (sbi_pmu_snapshot_available())
1106 			/* SBI implementation already updated the logical indicies */
1107 			hidx = lidx;
1108 		else
1109 			/* compute hardware counter index */
1110 			hidx = info->csr - CSR_CYCLE;
1111 
1112 		/* check if the corresponding bit is set in sscountovf or overflow mask in shmem */
1113 		if (!(overflow & BIT(hidx)))
1114 			continue;
1115 
1116 		/*
1117 		 * Keep a track of overflowed counters so that they can be started
1118 		 * with updated initial value.
1119 		 */
1120 		overflowed_ctrs |= BIT(lidx);
1121 		hw_evt = &event->hw;
1122 		/* Update the event states here so that we know the state while reading */
1123 		hw_evt->state |= PERF_HES_STOPPED;
1124 		riscv_pmu_event_update(event);
1125 		hw_evt->state |= PERF_HES_UPTODATE;
1126 		perf_sample_data_init(&data, 0, hw_evt->last_period);
1127 		if (riscv_pmu_event_set_period(event)) {
1128 			/*
1129 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
1130 			 * to avoid throttling here. As per the specification, the
1131 			 * interrupt remains disabled until the OF bit is set.
1132 			 * Interrupts are enabled again only during the start.
1133 			 * TODO: We will need to stop the guest counters once
1134 			 * virtualization support is added.
1135 			 */
1136 			perf_event_overflow(event, &data, regs);
1137 		}
1138 		/* Reset the state as we are going to start the counter after the loop */
1139 		hw_evt->state = 0;
1140 	}
1141 
1142 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
1143 	perf_sample_event_took(sched_clock() - start_clock);
1144 
1145 	return IRQ_HANDLED;
1146 }
1147 
1148 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
1149 {
1150 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
1151 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
1152 
1153 	/*
1154 	 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the
1155 	 * legacy option but that will be removed in the future.
1156 	 */
1157 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
1158 		csr_write(CSR_SCOUNTEREN, 0x7);
1159 	else
1160 		csr_write(CSR_SCOUNTEREN, 0x2);
1161 
1162 	/* Stop all the counters so that they can be enabled from perf */
1163 	pmu_sbi_stop_all(pmu);
1164 
1165 	if (riscv_pmu_use_irq) {
1166 		cpu_hw_evt->irq = riscv_pmu_irq;
1167 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
1168 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
1169 	}
1170 
1171 	if (sbi_pmu_snapshot_available())
1172 		return pmu_sbi_snapshot_setup(pmu, cpu);
1173 
1174 	return 0;
1175 }
1176 
1177 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
1178 {
1179 	if (riscv_pmu_use_irq) {
1180 		disable_percpu_irq(riscv_pmu_irq);
1181 	}
1182 
1183 	/* Disable all counters access for user mode now */
1184 	csr_write(CSR_SCOUNTEREN, 0x0);
1185 
1186 	if (sbi_pmu_snapshot_available())
1187 		return pmu_sbi_snapshot_disable();
1188 
1189 	return 0;
1190 }
1191 
1192 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
1193 {
1194 	int ret;
1195 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
1196 	struct irq_domain *domain = NULL;
1197 
1198 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
1199 		riscv_pmu_irq_num = RV_IRQ_PMU;
1200 		riscv_pmu_use_irq = true;
1201 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
1202 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
1203 		   riscv_cached_marchid(0) == 0 &&
1204 		   riscv_cached_mimpid(0) == 0) {
1205 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
1206 		riscv_pmu_use_irq = true;
1207 	} else if (riscv_has_vendor_extension_unlikely(ANDES_VENDOR_ID,
1208 						       RISCV_ISA_VENDOR_EXT_XANDESPMU) &&
1209 		   IS_ENABLED(CONFIG_ANDES_CUSTOM_PMU)) {
1210 		riscv_pmu_irq_num = ANDES_SLI_CAUSE_BASE + ANDES_RV_IRQ_PMOVI;
1211 		riscv_pmu_use_irq = true;
1212 	}
1213 
1214 	riscv_pmu_irq_mask = BIT(riscv_pmu_irq_num % BITS_PER_LONG);
1215 
1216 	if (!riscv_pmu_use_irq)
1217 		return -EOPNOTSUPP;
1218 
1219 	domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
1220 					  DOMAIN_BUS_ANY);
1221 	if (!domain) {
1222 		pr_err("Failed to find INTC IRQ root domain\n");
1223 		return -ENODEV;
1224 	}
1225 
1226 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
1227 	if (!riscv_pmu_irq) {
1228 		pr_err("Failed to map PMU interrupt for node\n");
1229 		return -ENODEV;
1230 	}
1231 
1232 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
1233 	if (ret) {
1234 		pr_err("registering percpu irq failed [%d]\n", ret);
1235 		return ret;
1236 	}
1237 
1238 	return 0;
1239 }
1240 
1241 #ifdef CONFIG_CPU_PM
1242 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
1243 				void *v)
1244 {
1245 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
1246 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
1247 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
1248 	struct perf_event *event;
1249 	int idx;
1250 
1251 	if (!enabled)
1252 		return NOTIFY_OK;
1253 
1254 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
1255 		event = cpuc->events[idx];
1256 		if (!event)
1257 			continue;
1258 
1259 		switch (cmd) {
1260 		case CPU_PM_ENTER:
1261 			/*
1262 			 * Stop and update the counter
1263 			 */
1264 			riscv_pmu_stop(event, PERF_EF_UPDATE);
1265 			break;
1266 		case CPU_PM_EXIT:
1267 		case CPU_PM_ENTER_FAILED:
1268 			/*
1269 			 * Restore and enable the counter.
1270 			 */
1271 			riscv_pmu_start(event, PERF_EF_RELOAD);
1272 			break;
1273 		default:
1274 			break;
1275 		}
1276 	}
1277 
1278 	return NOTIFY_OK;
1279 }
1280 
1281 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
1282 {
1283 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
1284 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
1285 }
1286 
1287 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
1288 {
1289 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
1290 }
1291 #else
1292 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
1293 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
1294 #endif
1295 
1296 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
1297 {
1298 	if (sbi_v2_available) {
1299 		if (sbi_pmu_snapshot_available()) {
1300 			pmu_sbi_snapshot_disable();
1301 			pmu_sbi_snapshot_free(pmu);
1302 		}
1303 	}
1304 	riscv_pm_pmu_unregister(pmu);
1305 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1306 }
1307 
1308 static void pmu_sbi_event_init(struct perf_event *event)
1309 {
1310 	/*
1311 	 * The permissions are set at event_init so that we do not depend
1312 	 * on the sysctl value that can change.
1313 	 */
1314 	if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS)
1315 		event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS;
1316 	else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS)
1317 		event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS;
1318 	else
1319 		event->hw.flags |= PERF_EVENT_FLAG_LEGACY;
1320 }
1321 
1322 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm)
1323 {
1324 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
1325 		return;
1326 
1327 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
1328 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
1329 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
1330 			return;
1331 		}
1332 	}
1333 
1334 	/*
1335 	 * The user mmapped the event to directly access it: this is where
1336 	 * we determine based on sysctl_perf_user_access if we grant userspace
1337 	 * the direct access to this event. That means that within the same
1338 	 * task, some events may be directly accessible and some other may not,
1339 	 * if the user changes the value of sysctl_perf_user_accesss in the
1340 	 * meantime.
1341 	 */
1342 
1343 	event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
1344 
1345 	/*
1346 	 * We must enable userspace access *before* advertising in the user page
1347 	 * that it is possible to do so to avoid any race.
1348 	 * And we must notify all cpus here because threads that currently run
1349 	 * on other cpus will try to directly access the counter too without
1350 	 * calling pmu_sbi_ctr_start.
1351 	 */
1352 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
1353 		on_each_cpu_mask(mm_cpumask(mm),
1354 				 pmu_sbi_set_scounteren, (void *)event, 1);
1355 }
1356 
1357 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm)
1358 {
1359 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
1360 		return;
1361 
1362 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
1363 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
1364 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
1365 			return;
1366 		}
1367 	}
1368 
1369 	/*
1370 	 * Here we can directly remove user access since the user does not have
1371 	 * access to the user page anymore so we avoid the racy window where the
1372 	 * user could have read cap_user_rdpmc to true right before we disable
1373 	 * it.
1374 	 */
1375 	event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT;
1376 
1377 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
1378 		on_each_cpu_mask(mm_cpumask(mm),
1379 				 pmu_sbi_reset_scounteren, (void *)event, 1);
1380 }
1381 
1382 static void riscv_pmu_update_counter_access(void *info)
1383 {
1384 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
1385 		csr_write(CSR_SCOUNTEREN, 0x7);
1386 	else
1387 		csr_write(CSR_SCOUNTEREN, 0x2);
1388 }
1389 
1390 static int riscv_pmu_proc_user_access_handler(const struct ctl_table *table,
1391 					      int write, void *buffer,
1392 					      size_t *lenp, loff_t *ppos)
1393 {
1394 	int prev = sysctl_perf_user_access;
1395 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1396 
1397 	/*
1398 	 * Test against the previous value since we clear SCOUNTEREN when
1399 	 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should
1400 	 * not do that if that was already the case.
1401 	 */
1402 	if (ret || !write || prev == sysctl_perf_user_access)
1403 		return ret;
1404 
1405 	on_each_cpu(riscv_pmu_update_counter_access, NULL, 1);
1406 
1407 	return 0;
1408 }
1409 
1410 static const struct ctl_table sbi_pmu_sysctl_table[] = {
1411 	{
1412 		.procname       = "perf_user_access",
1413 		.data		= &sysctl_perf_user_access,
1414 		.maxlen		= sizeof(unsigned int),
1415 		.mode           = 0644,
1416 		.proc_handler	= riscv_pmu_proc_user_access_handler,
1417 		.extra1		= SYSCTL_ZERO,
1418 		.extra2		= SYSCTL_TWO,
1419 	},
1420 };
1421 
1422 static int pmu_sbi_device_probe(struct platform_device *pdev)
1423 {
1424 	struct riscv_pmu *pmu = NULL;
1425 	int ret = -ENODEV;
1426 	int num_counters;
1427 
1428 	pr_info("SBI PMU extension is available\n");
1429 	pmu = riscv_pmu_alloc();
1430 	if (!pmu)
1431 		return -ENOMEM;
1432 
1433 	num_counters = pmu_sbi_find_num_ctrs();
1434 	if (num_counters < 0) {
1435 		pr_err("SBI PMU extension doesn't provide any counters\n");
1436 		goto out_free;
1437 	}
1438 
1439 	/* It is possible to get from SBI more than max number of counters */
1440 	if (num_counters > RISCV_MAX_COUNTERS) {
1441 		num_counters = RISCV_MAX_COUNTERS;
1442 		pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters);
1443 	}
1444 
1445 	/* cache all the information about counters now */
1446 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
1447 		goto out_free;
1448 
1449 	ret = pmu_sbi_setup_irqs(pmu, pdev);
1450 	if (ret < 0) {
1451 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
1452 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1453 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
1454 	}
1455 
1456 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
1457 	pmu->pmu.parent = &pdev->dev;
1458 	pmu->cmask = cmask;
1459 	pmu->ctr_start = pmu_sbi_ctr_start;
1460 	pmu->ctr_stop = pmu_sbi_ctr_stop;
1461 	pmu->event_map = pmu_sbi_event_map;
1462 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
1463 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
1464 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
1465 	pmu->ctr_read = pmu_sbi_ctr_read;
1466 	pmu->event_init = pmu_sbi_event_init;
1467 	pmu->event_mapped = pmu_sbi_event_mapped;
1468 	pmu->event_unmapped = pmu_sbi_event_unmapped;
1469 	pmu->csr_index = pmu_sbi_csr_index;
1470 
1471 	ret = riscv_pm_pmu_register(pmu);
1472 	if (ret)
1473 		goto out_unregister;
1474 
1475 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
1476 	if (ret)
1477 		goto out_unregister;
1478 
1479 	/* SBI PMU Snapsphot is only available in SBI v2.0 */
1480 	if (sbi_v2_available) {
1481 		int cpu;
1482 
1483 		ret = pmu_sbi_snapshot_alloc(pmu);
1484 		if (ret)
1485 			goto out_unregister;
1486 
1487 		cpu = get_cpu();
1488 		ret = pmu_sbi_snapshot_setup(pmu, cpu);
1489 		put_cpu();
1490 
1491 		if (ret) {
1492 			/* Snapshot is an optional feature. Continue if not available */
1493 			pmu_sbi_snapshot_free(pmu);
1494 		} else {
1495 			pr_info("SBI PMU snapshot detected\n");
1496 			/*
1497 			 * We enable it once here for the boot cpu. If snapshot shmem setup
1498 			 * fails during cpu hotplug process, it will fail to start the cpu
1499 			 * as we can not handle hetergenous PMUs with different snapshot
1500 			 * capability.
1501 			 */
1502 			static_branch_enable(&sbi_pmu_snapshot_available);
1503 		}
1504 	}
1505 
1506 	register_sysctl("kernel", sbi_pmu_sysctl_table);
1507 
1508 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1509 	if (ret)
1510 		goto out_unregister;
1511 
1512 	/* Asynchronously check which standard events are available */
1513 	schedule_work(&check_std_events_work);
1514 
1515 	return 0;
1516 
1517 out_unregister:
1518 	riscv_pmu_destroy(pmu);
1519 
1520 out_free:
1521 	kfree(pmu);
1522 	return ret;
1523 }
1524 
1525 static struct platform_driver pmu_sbi_driver = {
1526 	.probe		= pmu_sbi_device_probe,
1527 	.driver		= {
1528 		.name	= RISCV_PMU_SBI_PDEV_NAME,
1529 	},
1530 };
1531 
1532 static int __init pmu_sbi_devinit(void)
1533 {
1534 	int ret;
1535 	struct platform_device *pdev;
1536 
1537 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
1538 	    !sbi_probe_extension(SBI_EXT_PMU)) {
1539 		return 0;
1540 	}
1541 
1542 	if (sbi_spec_version >= sbi_mk_version(2, 0))
1543 		sbi_v2_available = true;
1544 
1545 	if (sbi_spec_version >= sbi_mk_version(3, 0))
1546 		sbi_v3_available = true;
1547 
1548 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
1549 				      "perf/riscv/pmu:starting",
1550 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
1551 	if (ret) {
1552 		pr_err("CPU hotplug notifier could not be registered: %d\n",
1553 		       ret);
1554 		return ret;
1555 	}
1556 
1557 	ret = platform_driver_register(&pmu_sbi_driver);
1558 	if (ret)
1559 		return ret;
1560 
1561 	pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0);
1562 	if (IS_ERR(pdev)) {
1563 		platform_driver_unregister(&pmu_sbi_driver);
1564 		return PTR_ERR(pdev);
1565 	}
1566 
1567 	/* Notify legacy implementation that SBI pmu is available*/
1568 	riscv_pmu_legacy_skip_init();
1569 
1570 	return ret;
1571 }
1572 device_initcall(pmu_sbi_devinit)
1573