xref: /linux/drivers/perf/riscv_pmu_sbi.c (revision 04cf420bbc32a599aa2481725f708435ea19bf3d)
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:0-47");
63 PMU_FORMAT_ATTR(firmware, "config:63");
64 
65 static bool sbi_v2_available;
66 static DEFINE_STATIC_KEY_FALSE(sbi_pmu_snapshot_available);
67 #define sbi_pmu_snapshot_available() \
68 	static_branch_unlikely(&sbi_pmu_snapshot_available)
69 
70 static struct attribute *riscv_arch_formats_attr[] = {
71 	&format_attr_event.attr,
72 	&format_attr_firmware.attr,
73 	NULL,
74 };
75 
76 static struct attribute_group riscv_pmu_format_group = {
77 	.name = "format",
78 	.attrs = riscv_arch_formats_attr,
79 };
80 
81 static const struct attribute_group *riscv_pmu_attr_groups[] = {
82 	&riscv_pmu_format_group,
83 	NULL,
84 };
85 
86 /* Allow user mode access by default */
87 static int sysctl_perf_user_access __read_mostly = SYSCTL_USER_ACCESS;
88 
89 /*
90  * RISC-V doesn't have heterogeneous harts yet. This need to be part of
91  * per_cpu in case of harts with different pmu counters
92  */
93 static union sbi_pmu_ctr_info *pmu_ctr_list;
94 static bool riscv_pmu_use_irq;
95 static unsigned int riscv_pmu_irq_num;
96 static unsigned int riscv_pmu_irq_mask;
97 static unsigned int riscv_pmu_irq;
98 
99 /* Cache the available counters in a bitmask */
100 static unsigned long cmask;
101 
102 struct sbi_pmu_event_data {
103 	union {
104 		union {
105 			struct hw_gen_event {
106 				uint32_t event_code:16;
107 				uint32_t event_type:4;
108 				uint32_t reserved:12;
109 			} hw_gen_event;
110 			struct hw_cache_event {
111 				uint32_t result_id:1;
112 				uint32_t op_id:2;
113 				uint32_t cache_id:13;
114 				uint32_t event_type:4;
115 				uint32_t reserved:12;
116 			} hw_cache_event;
117 		};
118 		uint32_t event_idx;
119 	};
120 };
121 
122 static struct sbi_pmu_event_data pmu_hw_event_map[] = {
123 	[PERF_COUNT_HW_CPU_CYCLES]		= {.hw_gen_event = {
124 							SBI_PMU_HW_CPU_CYCLES,
125 							SBI_PMU_EVENT_TYPE_HW, 0}},
126 	[PERF_COUNT_HW_INSTRUCTIONS]		= {.hw_gen_event = {
127 							SBI_PMU_HW_INSTRUCTIONS,
128 							SBI_PMU_EVENT_TYPE_HW, 0}},
129 	[PERF_COUNT_HW_CACHE_REFERENCES]	= {.hw_gen_event = {
130 							SBI_PMU_HW_CACHE_REFERENCES,
131 							SBI_PMU_EVENT_TYPE_HW, 0}},
132 	[PERF_COUNT_HW_CACHE_MISSES]		= {.hw_gen_event = {
133 							SBI_PMU_HW_CACHE_MISSES,
134 							SBI_PMU_EVENT_TYPE_HW, 0}},
135 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= {.hw_gen_event = {
136 							SBI_PMU_HW_BRANCH_INSTRUCTIONS,
137 							SBI_PMU_EVENT_TYPE_HW, 0}},
138 	[PERF_COUNT_HW_BRANCH_MISSES]		= {.hw_gen_event = {
139 							SBI_PMU_HW_BRANCH_MISSES,
140 							SBI_PMU_EVENT_TYPE_HW, 0}},
141 	[PERF_COUNT_HW_BUS_CYCLES]		= {.hw_gen_event = {
142 							SBI_PMU_HW_BUS_CYCLES,
143 							SBI_PMU_EVENT_TYPE_HW, 0}},
144 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= {.hw_gen_event = {
145 							SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
146 							SBI_PMU_EVENT_TYPE_HW, 0}},
147 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= {.hw_gen_event = {
148 							SBI_PMU_HW_STALLED_CYCLES_BACKEND,
149 							SBI_PMU_EVENT_TYPE_HW, 0}},
150 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= {.hw_gen_event = {
151 							SBI_PMU_HW_REF_CPU_CYCLES,
152 							SBI_PMU_EVENT_TYPE_HW, 0}},
153 };
154 
155 #define C(x) PERF_COUNT_HW_CACHE_##x
156 static struct sbi_pmu_event_data pmu_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
157 [PERF_COUNT_HW_CACHE_OP_MAX]
158 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
159 	[C(L1D)] = {
160 		[C(OP_READ)] = {
161 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
162 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
163 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
164 					C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
165 		},
166 		[C(OP_WRITE)] = {
167 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
168 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
169 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
170 					C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
171 		},
172 		[C(OP_PREFETCH)] = {
173 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
174 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
175 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
176 					C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
177 		},
178 	},
179 	[C(L1I)] = {
180 		[C(OP_READ)] = {
181 			[C(RESULT_ACCESS)] = {.hw_cache_event =	{C(RESULT_ACCESS),
182 					C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
183 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
184 					C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
185 		},
186 		[C(OP_WRITE)] = {
187 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
188 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
189 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
190 					C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
191 		},
192 		[C(OP_PREFETCH)] = {
193 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
194 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
195 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
196 					C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
197 		},
198 	},
199 	[C(LL)] = {
200 		[C(OP_READ)] = {
201 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
202 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
203 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
204 					C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
205 		},
206 		[C(OP_WRITE)] = {
207 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
208 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
209 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
210 					C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
211 		},
212 		[C(OP_PREFETCH)] = {
213 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
214 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
215 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
216 					C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
217 		},
218 	},
219 	[C(DTLB)] = {
220 		[C(OP_READ)] = {
221 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
222 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
223 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
224 					C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
225 		},
226 		[C(OP_WRITE)] = {
227 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
228 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
229 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
230 					C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
231 		},
232 		[C(OP_PREFETCH)] = {
233 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
234 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
235 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
236 					C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
237 		},
238 	},
239 	[C(ITLB)] = {
240 		[C(OP_READ)] = {
241 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
242 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
243 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
244 					C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
245 		},
246 		[C(OP_WRITE)] = {
247 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
248 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
249 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
250 					C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
251 		},
252 		[C(OP_PREFETCH)] = {
253 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
254 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
255 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
256 					C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
257 		},
258 	},
259 	[C(BPU)] = {
260 		[C(OP_READ)] = {
261 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
262 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
263 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
264 					C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
265 		},
266 		[C(OP_WRITE)] = {
267 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
268 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
269 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
270 					C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
271 		},
272 		[C(OP_PREFETCH)] = {
273 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
274 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
275 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
276 					C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
277 		},
278 	},
279 	[C(NODE)] = {
280 		[C(OP_READ)] = {
281 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
282 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
283 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
284 					C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
285 		},
286 		[C(OP_WRITE)] = {
287 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
288 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
289 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
290 					C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
291 		},
292 		[C(OP_PREFETCH)] = {
293 			[C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
294 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
295 			[C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
296 					C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
297 		},
298 	},
299 };
300 
301 static void pmu_sbi_check_event(struct sbi_pmu_event_data *edata)
302 {
303 	struct sbiret ret;
304 
305 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH,
306 			0, cmask, 0, edata->event_idx, 0, 0);
307 	if (!ret.error) {
308 		sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
309 			  ret.value, 0x1, SBI_PMU_STOP_FLAG_RESET, 0, 0, 0);
310 	} else if (ret.error == SBI_ERR_NOT_SUPPORTED) {
311 		/* This event cannot be monitored by any counter */
312 		edata->event_idx = -EINVAL;
313 	}
314 }
315 
316 static void pmu_sbi_check_std_events(struct work_struct *work)
317 {
318 	for (int i = 0; i < ARRAY_SIZE(pmu_hw_event_map); i++)
319 		pmu_sbi_check_event(&pmu_hw_event_map[i]);
320 
321 	for (int i = 0; i < ARRAY_SIZE(pmu_cache_event_map); i++)
322 		for (int j = 0; j < ARRAY_SIZE(pmu_cache_event_map[i]); j++)
323 			for (int k = 0; k < ARRAY_SIZE(pmu_cache_event_map[i][j]); k++)
324 				pmu_sbi_check_event(&pmu_cache_event_map[i][j][k]);
325 }
326 
327 static DECLARE_WORK(check_std_events_work, pmu_sbi_check_std_events);
328 
329 static int pmu_sbi_ctr_get_width(int idx)
330 {
331 	return pmu_ctr_list[idx].width;
332 }
333 
334 static bool pmu_sbi_ctr_is_fw(int cidx)
335 {
336 	union sbi_pmu_ctr_info *info;
337 
338 	info = &pmu_ctr_list[cidx];
339 	if (!info)
340 		return false;
341 
342 	return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
343 }
344 
345 /*
346  * Returns the counter width of a programmable counter and number of hardware
347  * counters. As we don't support heterogeneous CPUs yet, it is okay to just
348  * return the counter width of the first programmable counter.
349  */
350 int riscv_pmu_get_hpm_info(u32 *hw_ctr_width, u32 *num_hw_ctr)
351 {
352 	int i;
353 	union sbi_pmu_ctr_info *info;
354 	u32 hpm_width = 0, hpm_count = 0;
355 
356 	if (!cmask)
357 		return -EINVAL;
358 
359 	for_each_set_bit(i, &cmask, RISCV_MAX_COUNTERS) {
360 		info = &pmu_ctr_list[i];
361 		if (!info)
362 			continue;
363 		if (!hpm_width && info->csr != CSR_CYCLE && info->csr != CSR_INSTRET)
364 			hpm_width = info->width;
365 		if (info->type == SBI_PMU_CTR_TYPE_HW)
366 			hpm_count++;
367 	}
368 
369 	*hw_ctr_width = hpm_width;
370 	*num_hw_ctr = hpm_count;
371 
372 	return 0;
373 }
374 EXPORT_SYMBOL_GPL(riscv_pmu_get_hpm_info);
375 
376 static uint8_t pmu_sbi_csr_index(struct perf_event *event)
377 {
378 	return pmu_ctr_list[event->hw.idx].csr - CSR_CYCLE;
379 }
380 
381 static unsigned long pmu_sbi_get_filter_flags(struct perf_event *event)
382 {
383 	unsigned long cflags = 0;
384 	bool guest_events = false;
385 
386 	if (event->attr.config1 & RISCV_PMU_CONFIG1_GUEST_EVENTS)
387 		guest_events = true;
388 	if (event->attr.exclude_kernel)
389 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VSINH : SBI_PMU_CFG_FLAG_SET_SINH;
390 	if (event->attr.exclude_user)
391 		cflags |= guest_events ? SBI_PMU_CFG_FLAG_SET_VUINH : SBI_PMU_CFG_FLAG_SET_UINH;
392 	if (guest_events && event->attr.exclude_hv)
393 		cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
394 	if (event->attr.exclude_host)
395 		cflags |= SBI_PMU_CFG_FLAG_SET_UINH | SBI_PMU_CFG_FLAG_SET_SINH;
396 	if (event->attr.exclude_guest)
397 		cflags |= SBI_PMU_CFG_FLAG_SET_VSINH | SBI_PMU_CFG_FLAG_SET_VUINH;
398 
399 	return cflags;
400 }
401 
402 static int pmu_sbi_ctr_get_idx(struct perf_event *event)
403 {
404 	struct hw_perf_event *hwc = &event->hw;
405 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
406 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
407 	struct sbiret ret;
408 	int idx;
409 	uint64_t cbase = 0, cmask = rvpmu->cmask;
410 	unsigned long cflags = 0;
411 
412 	cflags = pmu_sbi_get_filter_flags(event);
413 
414 	/*
415 	 * In legacy mode, we have to force the fixed counters for those events
416 	 * but not in the user access mode as we want to use the other counters
417 	 * that support sampling/filtering.
418 	 */
419 	if ((hwc->flags & PERF_EVENT_FLAG_LEGACY) && (event->attr.type == PERF_TYPE_HARDWARE)) {
420 		if (event->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
421 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
422 			cmask = 1;
423 		} else if (event->attr.config == PERF_COUNT_HW_INSTRUCTIONS) {
424 			cflags |= SBI_PMU_CFG_FLAG_SKIP_MATCH;
425 			cmask = BIT(CSR_INSTRET - CSR_CYCLE);
426 		}
427 	}
428 
429 	/* retrieve the available counter index */
430 #if defined(CONFIG_32BIT)
431 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
432 			cmask, cflags, hwc->event_base, hwc->config,
433 			hwc->config >> 32);
434 #else
435 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase,
436 			cmask, cflags, hwc->event_base, hwc->config, 0);
437 #endif
438 	if (ret.error) {
439 		pr_debug("Not able to find a counter for event %lx config %llx\n",
440 			hwc->event_base, hwc->config);
441 		return sbi_err_map_linux_errno(ret.error);
442 	}
443 
444 	idx = ret.value;
445 	if (!test_bit(idx, &rvpmu->cmask) || !pmu_ctr_list[idx].value)
446 		return -ENOENT;
447 
448 	/* Additional sanity check for the counter id */
449 	if (pmu_sbi_ctr_is_fw(idx)) {
450 		if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
451 			return idx;
452 	} else {
453 		if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
454 			return idx;
455 	}
456 
457 	return -ENOENT;
458 }
459 
460 static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
461 {
462 
463 	struct hw_perf_event *hwc = &event->hw;
464 	struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
465 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
466 	int idx = hwc->idx;
467 
468 	if (pmu_sbi_ctr_is_fw(idx))
469 		clear_bit(idx, cpuc->used_fw_ctrs);
470 	else
471 		clear_bit(idx, cpuc->used_hw_ctrs);
472 }
473 
474 static int pmu_event_find_cache(u64 config)
475 {
476 	unsigned int cache_type, cache_op, cache_result, ret;
477 
478 	cache_type = (config >>  0) & 0xff;
479 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
480 		return -EINVAL;
481 
482 	cache_op = (config >>  8) & 0xff;
483 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
484 		return -EINVAL;
485 
486 	cache_result = (config >> 16) & 0xff;
487 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
488 		return -EINVAL;
489 
490 	ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
491 
492 	return ret;
493 }
494 
495 static bool pmu_sbi_is_fw_event(struct perf_event *event)
496 {
497 	u32 type = event->attr.type;
498 	u64 config = event->attr.config;
499 
500 	if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
501 		return true;
502 	else
503 		return false;
504 }
505 
506 static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
507 {
508 	u32 type = event->attr.type;
509 	u64 config = event->attr.config;
510 	int bSoftware;
511 	u64 raw_config_val;
512 	int ret;
513 
514 	/*
515 	 * Ensure we are finished checking standard hardware events for
516 	 * validity before allowing userspace to configure any events.
517 	 */
518 	flush_work(&check_std_events_work);
519 
520 	switch (type) {
521 	case PERF_TYPE_HARDWARE:
522 		if (config >= PERF_COUNT_HW_MAX)
523 			return -EINVAL;
524 		ret = pmu_hw_event_map[event->attr.config].event_idx;
525 		break;
526 	case PERF_TYPE_HW_CACHE:
527 		ret = pmu_event_find_cache(config);
528 		break;
529 	case PERF_TYPE_RAW:
530 		/*
531 		 * As per SBI specification, the upper 16 bits must be unused for
532 		 * a raw event. Use the MSB (63b) to distinguish between hardware
533 		 * raw event and firmware events.
534 		 */
535 		bSoftware = config >> 63;
536 		raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
537 		if (bSoftware) {
538 			ret = (raw_config_val & 0xFFFF) |
539 				(SBI_PMU_EVENT_TYPE_FW << 16);
540 		} else {
541 			ret = RISCV_PMU_RAW_EVENT_IDX;
542 			*econfig = raw_config_val;
543 		}
544 		break;
545 	default:
546 		ret = -EINVAL;
547 		break;
548 	}
549 
550 	return ret;
551 }
552 
553 static void pmu_sbi_snapshot_free(struct riscv_pmu *pmu)
554 {
555 	int cpu;
556 
557 	for_each_possible_cpu(cpu) {
558 		struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
559 
560 		if (!cpu_hw_evt->snapshot_addr)
561 			continue;
562 
563 		free_page((unsigned long)cpu_hw_evt->snapshot_addr);
564 		cpu_hw_evt->snapshot_addr = NULL;
565 		cpu_hw_evt->snapshot_addr_phys = 0;
566 	}
567 }
568 
569 static int pmu_sbi_snapshot_alloc(struct riscv_pmu *pmu)
570 {
571 	int cpu;
572 	struct page *snapshot_page;
573 
574 	for_each_possible_cpu(cpu) {
575 		struct cpu_hw_events *cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
576 
577 		snapshot_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
578 		if (!snapshot_page) {
579 			pmu_sbi_snapshot_free(pmu);
580 			return -ENOMEM;
581 		}
582 		cpu_hw_evt->snapshot_addr = page_to_virt(snapshot_page);
583 		cpu_hw_evt->snapshot_addr_phys = page_to_phys(snapshot_page);
584 	}
585 
586 	return 0;
587 }
588 
589 static int pmu_sbi_snapshot_disable(void)
590 {
591 	struct sbiret ret;
592 
593 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM, SBI_SHMEM_DISABLE,
594 			SBI_SHMEM_DISABLE, 0, 0, 0, 0);
595 	if (ret.error) {
596 		pr_warn("failed to disable snapshot shared memory\n");
597 		return sbi_err_map_linux_errno(ret.error);
598 	}
599 
600 	return 0;
601 }
602 
603 static int pmu_sbi_snapshot_setup(struct riscv_pmu *pmu, int cpu)
604 {
605 	struct cpu_hw_events *cpu_hw_evt;
606 	struct sbiret ret = {0};
607 
608 	cpu_hw_evt = per_cpu_ptr(pmu->hw_events, cpu);
609 	if (!cpu_hw_evt->snapshot_addr_phys)
610 		return -EINVAL;
611 
612 	if (cpu_hw_evt->snapshot_set_done)
613 		return 0;
614 
615 	if (IS_ENABLED(CONFIG_32BIT))
616 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM,
617 				cpu_hw_evt->snapshot_addr_phys,
618 				(u64)(cpu_hw_evt->snapshot_addr_phys) >> 32, 0, 0, 0, 0);
619 	else
620 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_SNAPSHOT_SET_SHMEM,
621 				cpu_hw_evt->snapshot_addr_phys, 0, 0, 0, 0, 0);
622 
623 	/* Free up the snapshot area memory and fall back to SBI PMU calls without snapshot */
624 	if (ret.error) {
625 		if (ret.error != SBI_ERR_NOT_SUPPORTED)
626 			pr_warn("pmu snapshot setup failed with error %ld\n", ret.error);
627 		return sbi_err_map_linux_errno(ret.error);
628 	}
629 
630 	memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS);
631 	cpu_hw_evt->snapshot_set_done = true;
632 
633 	return 0;
634 }
635 
636 static u64 pmu_sbi_ctr_read(struct perf_event *event)
637 {
638 	struct hw_perf_event *hwc = &event->hw;
639 	int idx = hwc->idx;
640 	struct sbiret ret;
641 	u64 val = 0;
642 	struct riscv_pmu *pmu = to_riscv_pmu(event->pmu);
643 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
644 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
645 	union sbi_pmu_ctr_info info = pmu_ctr_list[idx];
646 
647 	/* Read the value from the shared memory directly only if counter is stopped */
648 	if (sbi_pmu_snapshot_available() && (hwc->state & PERF_HES_STOPPED)) {
649 		val = sdata->ctr_values[idx];
650 		return val;
651 	}
652 
653 	if (pmu_sbi_is_fw_event(event)) {
654 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
655 				hwc->idx, 0, 0, 0, 0, 0);
656 		if (ret.error)
657 			return 0;
658 
659 		val = ret.value;
660 		if (IS_ENABLED(CONFIG_32BIT) && sbi_v2_available && info.width >= 32) {
661 			ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ_HI,
662 					hwc->idx, 0, 0, 0, 0, 0);
663 			if (!ret.error)
664 				val |= ((u64)ret.value << 32);
665 			else
666 				WARN_ONCE(1, "Unable to read upper 32 bits of firmware counter error: %ld\n",
667 					  ret.error);
668 		}
669 	} else {
670 		val = riscv_pmu_ctr_read_csr(info.csr);
671 		if (IS_ENABLED(CONFIG_32BIT))
672 			val |= ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 32;
673 	}
674 
675 	return val;
676 }
677 
678 static void pmu_sbi_set_scounteren(void *arg)
679 {
680 	struct perf_event *event = (struct perf_event *)arg;
681 
682 	if (event->hw.idx != -1)
683 		csr_write(CSR_SCOUNTEREN,
684 			  csr_read(CSR_SCOUNTEREN) | BIT(pmu_sbi_csr_index(event)));
685 }
686 
687 static void pmu_sbi_reset_scounteren(void *arg)
688 {
689 	struct perf_event *event = (struct perf_event *)arg;
690 
691 	if (event->hw.idx != -1)
692 		csr_write(CSR_SCOUNTEREN,
693 			  csr_read(CSR_SCOUNTEREN) & ~BIT(pmu_sbi_csr_index(event)));
694 }
695 
696 static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
697 {
698 	struct sbiret ret;
699 	struct hw_perf_event *hwc = &event->hw;
700 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
701 
702 	/* There is no benefit setting SNAPSHOT FLAG for a single counter */
703 #if defined(CONFIG_32BIT)
704 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
705 			1, flag, ival, ival >> 32, 0);
706 #else
707 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
708 			1, flag, ival, 0, 0);
709 #endif
710 	if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
711 		pr_err("Starting counter idx %d failed with error %d\n",
712 			hwc->idx, sbi_err_map_linux_errno(ret.error));
713 
714 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
715 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
716 		pmu_sbi_set_scounteren((void *)event);
717 }
718 
719 static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
720 {
721 	struct sbiret ret;
722 	struct hw_perf_event *hwc = &event->hw;
723 	struct riscv_pmu *pmu = to_riscv_pmu(event->pmu);
724 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
725 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
726 
727 	if ((hwc->flags & PERF_EVENT_FLAG_USER_ACCESS) &&
728 	    (hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
729 		pmu_sbi_reset_scounteren((void *)event);
730 
731 	if (sbi_pmu_snapshot_available())
732 		flag |= SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT;
733 
734 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
735 	if (!ret.error && sbi_pmu_snapshot_available()) {
736 		/*
737 		 * The counter snapshot is based on the index base specified by hwc->idx.
738 		 * The actual counter value is updated in shared memory at index 0 when counter
739 		 * mask is 0x01. To ensure accurate counter values, it's necessary to transfer
740 		 * the counter value to shared memory. However, if hwc->idx is zero, the counter
741 		 * value is already correctly updated in shared memory, requiring no further
742 		 * adjustment.
743 		 */
744 		if (hwc->idx > 0) {
745 			sdata->ctr_values[hwc->idx] = sdata->ctr_values[0];
746 			sdata->ctr_values[0] = 0;
747 		}
748 	} else if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
749 		flag != SBI_PMU_STOP_FLAG_RESET) {
750 		pr_err("Stopping counter idx %d failed with error %d\n",
751 			hwc->idx, sbi_err_map_linux_errno(ret.error));
752 	}
753 }
754 
755 static int pmu_sbi_find_num_ctrs(void)
756 {
757 	struct sbiret ret;
758 
759 	ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
760 	if (!ret.error)
761 		return ret.value;
762 	else
763 		return sbi_err_map_linux_errno(ret.error);
764 }
765 
766 static int pmu_sbi_get_ctrinfo(int nctr, unsigned long *mask)
767 {
768 	struct sbiret ret;
769 	int i, num_hw_ctr = 0, num_fw_ctr = 0;
770 	union sbi_pmu_ctr_info cinfo;
771 
772 	pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
773 	if (!pmu_ctr_list)
774 		return -ENOMEM;
775 
776 	for (i = 0; i < nctr; i++) {
777 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
778 		if (ret.error)
779 			/* The logical counter ids are not expected to be contiguous */
780 			continue;
781 
782 		*mask |= BIT(i);
783 
784 		cinfo.value = ret.value;
785 		if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
786 			num_fw_ctr++;
787 		else
788 			num_hw_ctr++;
789 		pmu_ctr_list[i].value = cinfo.value;
790 	}
791 
792 	pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
793 
794 	return 0;
795 }
796 
797 static inline void pmu_sbi_stop_all(struct riscv_pmu *pmu)
798 {
799 	/*
800 	 * No need to check the error because we are disabling all the counters
801 	 * which may include counters that are not enabled yet.
802 	 */
803 	sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
804 		  0, pmu->cmask, SBI_PMU_STOP_FLAG_RESET, 0, 0, 0);
805 }
806 
807 static inline void pmu_sbi_stop_hw_ctrs(struct riscv_pmu *pmu)
808 {
809 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
810 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
811 	unsigned long flag = 0;
812 	int i, idx;
813 	struct sbiret ret;
814 	u64 temp_ctr_overflow_mask = 0;
815 
816 	if (sbi_pmu_snapshot_available())
817 		flag = SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT;
818 
819 	/* Reset the shadow copy to avoid save/restore any value from previous overflow */
820 	memset(cpu_hw_evt->snapshot_cval_shcopy, 0, sizeof(u64) * RISCV_MAX_COUNTERS);
821 
822 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
823 		/* No need to check the error here as we can't do anything about the error */
824 		ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, i * BITS_PER_LONG,
825 				cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0);
826 		if (!ret.error && sbi_pmu_snapshot_available()) {
827 			/* Save the counter values to avoid clobbering */
828 			for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG)
829 				cpu_hw_evt->snapshot_cval_shcopy[i * BITS_PER_LONG + idx] =
830 							sdata->ctr_values[idx];
831 			/* Save the overflow mask to avoid clobbering */
832 			temp_ctr_overflow_mask |= sdata->ctr_overflow_mask << (i * BITS_PER_LONG);
833 		}
834 	}
835 
836 	/* Restore the counter values to the shared memory for used hw counters */
837 	if (sbi_pmu_snapshot_available()) {
838 		for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS)
839 			sdata->ctr_values[idx] = cpu_hw_evt->snapshot_cval_shcopy[idx];
840 		if (temp_ctr_overflow_mask)
841 			sdata->ctr_overflow_mask = temp_ctr_overflow_mask;
842 	}
843 }
844 
845 /*
846  * This function starts all the used counters in two step approach.
847  * Any counter that did not overflow can be start in a single step
848  * while the overflowed counters need to be started with updated initialization
849  * value.
850  */
851 static inline void pmu_sbi_start_ovf_ctrs_sbi(struct cpu_hw_events *cpu_hw_evt,
852 					      u64 ctr_ovf_mask)
853 {
854 	int idx = 0, i;
855 	struct perf_event *event;
856 	unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
857 	unsigned long ctr_start_mask = 0;
858 	uint64_t max_period;
859 	struct hw_perf_event *hwc;
860 	u64 init_val = 0;
861 
862 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
863 		ctr_start_mask = cpu_hw_evt->used_hw_ctrs[i] & ~ctr_ovf_mask;
864 		/* Start all the counters that did not overflow in a single shot */
865 		sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, i * BITS_PER_LONG, ctr_start_mask,
866 			0, 0, 0, 0);
867 	}
868 
869 	/* Reinitialize and start all the counter that overflowed */
870 	while (ctr_ovf_mask) {
871 		if (ctr_ovf_mask & 0x01) {
872 			event = cpu_hw_evt->events[idx];
873 			hwc = &event->hw;
874 			max_period = riscv_pmu_ctr_get_width_mask(event);
875 			init_val = local64_read(&hwc->prev_count) & max_period;
876 #if defined(CONFIG_32BIT)
877 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
878 				  flag, init_val, init_val >> 32, 0);
879 #else
880 			sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx, 1,
881 				  flag, init_val, 0, 0);
882 #endif
883 			perf_event_update_userpage(event);
884 		}
885 		ctr_ovf_mask = ctr_ovf_mask >> 1;
886 		idx++;
887 	}
888 }
889 
890 static inline void pmu_sbi_start_ovf_ctrs_snapshot(struct cpu_hw_events *cpu_hw_evt,
891 						   u64 ctr_ovf_mask)
892 {
893 	int i, idx = 0;
894 	struct perf_event *event;
895 	unsigned long flag = SBI_PMU_START_FLAG_INIT_SNAPSHOT;
896 	u64 max_period, init_val = 0;
897 	struct hw_perf_event *hwc;
898 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
899 
900 	for_each_set_bit(idx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
901 		if (ctr_ovf_mask & BIT(idx)) {
902 			event = cpu_hw_evt->events[idx];
903 			hwc = &event->hw;
904 			max_period = riscv_pmu_ctr_get_width_mask(event);
905 			init_val = local64_read(&hwc->prev_count) & max_period;
906 			cpu_hw_evt->snapshot_cval_shcopy[idx] = init_val;
907 		}
908 		/*
909 		 * We do not need to update the non-overflow counters the previous
910 		 * value should have been there already.
911 		 */
912 	}
913 
914 	for (i = 0; i < BITS_TO_LONGS(RISCV_MAX_COUNTERS); i++) {
915 		/* Restore the counter values to relative indices for used hw counters */
916 		for_each_set_bit(idx, &cpu_hw_evt->used_hw_ctrs[i], BITS_PER_LONG)
917 			sdata->ctr_values[idx] =
918 					cpu_hw_evt->snapshot_cval_shcopy[idx + i * BITS_PER_LONG];
919 		/* Start all the counters in a single shot */
920 		sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, idx * BITS_PER_LONG,
921 			  cpu_hw_evt->used_hw_ctrs[i], flag, 0, 0, 0);
922 	}
923 }
924 
925 static void pmu_sbi_start_overflow_mask(struct riscv_pmu *pmu,
926 					u64 ctr_ovf_mask)
927 {
928 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
929 
930 	if (sbi_pmu_snapshot_available())
931 		pmu_sbi_start_ovf_ctrs_snapshot(cpu_hw_evt, ctr_ovf_mask);
932 	else
933 		pmu_sbi_start_ovf_ctrs_sbi(cpu_hw_evt, ctr_ovf_mask);
934 }
935 
936 static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)
937 {
938 	struct perf_sample_data data;
939 	struct pt_regs *regs;
940 	struct hw_perf_event *hw_evt;
941 	union sbi_pmu_ctr_info *info;
942 	int lidx, hidx, fidx;
943 	struct riscv_pmu *pmu;
944 	struct perf_event *event;
945 	u64 overflow;
946 	u64 overflowed_ctrs = 0;
947 	struct cpu_hw_events *cpu_hw_evt = dev;
948 	u64 start_clock = sched_clock();
949 	struct riscv_pmu_snapshot_data *sdata = cpu_hw_evt->snapshot_addr;
950 
951 	if (WARN_ON_ONCE(!cpu_hw_evt))
952 		return IRQ_NONE;
953 
954 	/* Firmware counter don't support overflow yet */
955 	fidx = find_first_bit(cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS);
956 	if (fidx == RISCV_MAX_COUNTERS) {
957 		csr_clear(CSR_SIP, BIT(riscv_pmu_irq_num));
958 		return IRQ_NONE;
959 	}
960 
961 	event = cpu_hw_evt->events[fidx];
962 	if (!event) {
963 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
964 		return IRQ_NONE;
965 	}
966 
967 	pmu = to_riscv_pmu(event->pmu);
968 	pmu_sbi_stop_hw_ctrs(pmu);
969 
970 	/* Overflow status register should only be read after counter are stopped */
971 	if (sbi_pmu_snapshot_available())
972 		overflow = sdata->ctr_overflow_mask;
973 	else
974 		ALT_SBI_PMU_OVERFLOW(overflow);
975 
976 	/*
977 	 * Overflow interrupt pending bit should only be cleared after stopping
978 	 * all the counters to avoid any race condition.
979 	 */
980 	ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
981 
982 	/* No overflow bit is set */
983 	if (!overflow)
984 		return IRQ_NONE;
985 
986 	regs = get_irq_regs();
987 
988 	for_each_set_bit(lidx, cpu_hw_evt->used_hw_ctrs, RISCV_MAX_COUNTERS) {
989 		struct perf_event *event = cpu_hw_evt->events[lidx];
990 
991 		/* Skip if invalid event or user did not request a sampling */
992 		if (!event || !is_sampling_event(event))
993 			continue;
994 
995 		info = &pmu_ctr_list[lidx];
996 		/* Do a sanity check */
997 		if (!info || info->type != SBI_PMU_CTR_TYPE_HW)
998 			continue;
999 
1000 		if (sbi_pmu_snapshot_available())
1001 			/* SBI implementation already updated the logical indicies */
1002 			hidx = lidx;
1003 		else
1004 			/* compute hardware counter index */
1005 			hidx = info->csr - CSR_CYCLE;
1006 
1007 		/* check if the corresponding bit is set in sscountovf or overflow mask in shmem */
1008 		if (!(overflow & BIT(hidx)))
1009 			continue;
1010 
1011 		/*
1012 		 * Keep a track of overflowed counters so that they can be started
1013 		 * with updated initial value.
1014 		 */
1015 		overflowed_ctrs |= BIT(lidx);
1016 		hw_evt = &event->hw;
1017 		/* Update the event states here so that we know the state while reading */
1018 		hw_evt->state |= PERF_HES_STOPPED;
1019 		riscv_pmu_event_update(event);
1020 		hw_evt->state |= PERF_HES_UPTODATE;
1021 		perf_sample_data_init(&data, 0, hw_evt->last_period);
1022 		if (riscv_pmu_event_set_period(event)) {
1023 			/*
1024 			 * Unlike other ISAs, RISC-V don't have to disable interrupts
1025 			 * to avoid throttling here. As per the specification, the
1026 			 * interrupt remains disabled until the OF bit is set.
1027 			 * Interrupts are enabled again only during the start.
1028 			 * TODO: We will need to stop the guest counters once
1029 			 * virtualization support is added.
1030 			 */
1031 			perf_event_overflow(event, &data, regs);
1032 		}
1033 		/* Reset the state as we are going to start the counter after the loop */
1034 		hw_evt->state = 0;
1035 	}
1036 
1037 	pmu_sbi_start_overflow_mask(pmu, overflowed_ctrs);
1038 	perf_sample_event_took(sched_clock() - start_clock);
1039 
1040 	return IRQ_HANDLED;
1041 }
1042 
1043 static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
1044 {
1045 	struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
1046 	struct cpu_hw_events *cpu_hw_evt = this_cpu_ptr(pmu->hw_events);
1047 
1048 	/*
1049 	 * We keep enabling userspace access to CYCLE, TIME and INSTRET via the
1050 	 * legacy option but that will be removed in the future.
1051 	 */
1052 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
1053 		csr_write(CSR_SCOUNTEREN, 0x7);
1054 	else
1055 		csr_write(CSR_SCOUNTEREN, 0x2);
1056 
1057 	/* Stop all the counters so that they can be enabled from perf */
1058 	pmu_sbi_stop_all(pmu);
1059 
1060 	if (riscv_pmu_use_irq) {
1061 		cpu_hw_evt->irq = riscv_pmu_irq;
1062 		ALT_SBI_PMU_OVF_CLEAR_PENDING(riscv_pmu_irq_mask);
1063 		enable_percpu_irq(riscv_pmu_irq, IRQ_TYPE_NONE);
1064 	}
1065 
1066 	if (sbi_pmu_snapshot_available())
1067 		return pmu_sbi_snapshot_setup(pmu, cpu);
1068 
1069 	return 0;
1070 }
1071 
1072 static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
1073 {
1074 	if (riscv_pmu_use_irq) {
1075 		disable_percpu_irq(riscv_pmu_irq);
1076 	}
1077 
1078 	/* Disable all counters access for user mode now */
1079 	csr_write(CSR_SCOUNTEREN, 0x0);
1080 
1081 	if (sbi_pmu_snapshot_available())
1082 		return pmu_sbi_snapshot_disable();
1083 
1084 	return 0;
1085 }
1086 
1087 static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdev)
1088 {
1089 	int ret;
1090 	struct cpu_hw_events __percpu *hw_events = pmu->hw_events;
1091 	struct irq_domain *domain = NULL;
1092 
1093 	if (riscv_isa_extension_available(NULL, SSCOFPMF)) {
1094 		riscv_pmu_irq_num = RV_IRQ_PMU;
1095 		riscv_pmu_use_irq = true;
1096 	} else if (IS_ENABLED(CONFIG_ERRATA_THEAD_PMU) &&
1097 		   riscv_cached_mvendorid(0) == THEAD_VENDOR_ID &&
1098 		   riscv_cached_marchid(0) == 0 &&
1099 		   riscv_cached_mimpid(0) == 0) {
1100 		riscv_pmu_irq_num = THEAD_C9XX_RV_IRQ_PMU;
1101 		riscv_pmu_use_irq = true;
1102 	} else if (riscv_has_vendor_extension_unlikely(ANDES_VENDOR_ID,
1103 						       RISCV_ISA_VENDOR_EXT_XANDESPMU) &&
1104 		   IS_ENABLED(CONFIG_ANDES_CUSTOM_PMU)) {
1105 		riscv_pmu_irq_num = ANDES_SLI_CAUSE_BASE + ANDES_RV_IRQ_PMOVI;
1106 		riscv_pmu_use_irq = true;
1107 	}
1108 
1109 	riscv_pmu_irq_mask = BIT(riscv_pmu_irq_num % BITS_PER_LONG);
1110 
1111 	if (!riscv_pmu_use_irq)
1112 		return -EOPNOTSUPP;
1113 
1114 	domain = irq_find_matching_fwnode(riscv_get_intc_hwnode(),
1115 					  DOMAIN_BUS_ANY);
1116 	if (!domain) {
1117 		pr_err("Failed to find INTC IRQ root domain\n");
1118 		return -ENODEV;
1119 	}
1120 
1121 	riscv_pmu_irq = irq_create_mapping(domain, riscv_pmu_irq_num);
1122 	if (!riscv_pmu_irq) {
1123 		pr_err("Failed to map PMU interrupt for node\n");
1124 		return -ENODEV;
1125 	}
1126 
1127 	ret = request_percpu_irq(riscv_pmu_irq, pmu_sbi_ovf_handler, "riscv-pmu", hw_events);
1128 	if (ret) {
1129 		pr_err("registering percpu irq failed [%d]\n", ret);
1130 		return ret;
1131 	}
1132 
1133 	return 0;
1134 }
1135 
1136 #ifdef CONFIG_CPU_PM
1137 static int riscv_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
1138 				void *v)
1139 {
1140 	struct riscv_pmu *rvpmu = container_of(b, struct riscv_pmu, riscv_pm_nb);
1141 	struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
1142 	int enabled = bitmap_weight(cpuc->used_hw_ctrs, RISCV_MAX_COUNTERS);
1143 	struct perf_event *event;
1144 	int idx;
1145 
1146 	if (!enabled)
1147 		return NOTIFY_OK;
1148 
1149 	for (idx = 0; idx < RISCV_MAX_COUNTERS; idx++) {
1150 		event = cpuc->events[idx];
1151 		if (!event)
1152 			continue;
1153 
1154 		switch (cmd) {
1155 		case CPU_PM_ENTER:
1156 			/*
1157 			 * Stop and update the counter
1158 			 */
1159 			riscv_pmu_stop(event, PERF_EF_UPDATE);
1160 			break;
1161 		case CPU_PM_EXIT:
1162 		case CPU_PM_ENTER_FAILED:
1163 			/*
1164 			 * Restore and enable the counter.
1165 			 */
1166 			riscv_pmu_start(event, PERF_EF_RELOAD);
1167 			break;
1168 		default:
1169 			break;
1170 		}
1171 	}
1172 
1173 	return NOTIFY_OK;
1174 }
1175 
1176 static int riscv_pm_pmu_register(struct riscv_pmu *pmu)
1177 {
1178 	pmu->riscv_pm_nb.notifier_call = riscv_pm_pmu_notify;
1179 	return cpu_pm_register_notifier(&pmu->riscv_pm_nb);
1180 }
1181 
1182 static void riscv_pm_pmu_unregister(struct riscv_pmu *pmu)
1183 {
1184 	cpu_pm_unregister_notifier(&pmu->riscv_pm_nb);
1185 }
1186 #else
1187 static inline int riscv_pm_pmu_register(struct riscv_pmu *pmu) { return 0; }
1188 static inline void riscv_pm_pmu_unregister(struct riscv_pmu *pmu) { }
1189 #endif
1190 
1191 static void riscv_pmu_destroy(struct riscv_pmu *pmu)
1192 {
1193 	if (sbi_v2_available) {
1194 		if (sbi_pmu_snapshot_available()) {
1195 			pmu_sbi_snapshot_disable();
1196 			pmu_sbi_snapshot_free(pmu);
1197 		}
1198 	}
1199 	riscv_pm_pmu_unregister(pmu);
1200 	cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1201 }
1202 
1203 static void pmu_sbi_event_init(struct perf_event *event)
1204 {
1205 	/*
1206 	 * The permissions are set at event_init so that we do not depend
1207 	 * on the sysctl value that can change.
1208 	 */
1209 	if (sysctl_perf_user_access == SYSCTL_NO_USER_ACCESS)
1210 		event->hw.flags |= PERF_EVENT_FLAG_NO_USER_ACCESS;
1211 	else if (sysctl_perf_user_access == SYSCTL_USER_ACCESS)
1212 		event->hw.flags |= PERF_EVENT_FLAG_USER_ACCESS;
1213 	else
1214 		event->hw.flags |= PERF_EVENT_FLAG_LEGACY;
1215 }
1216 
1217 static void pmu_sbi_event_mapped(struct perf_event *event, struct mm_struct *mm)
1218 {
1219 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
1220 		return;
1221 
1222 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
1223 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
1224 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
1225 			return;
1226 		}
1227 	}
1228 
1229 	/*
1230 	 * The user mmapped the event to directly access it: this is where
1231 	 * we determine based on sysctl_perf_user_access if we grant userspace
1232 	 * the direct access to this event. That means that within the same
1233 	 * task, some events may be directly accessible and some other may not,
1234 	 * if the user changes the value of sysctl_perf_user_accesss in the
1235 	 * meantime.
1236 	 */
1237 
1238 	event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
1239 
1240 	/*
1241 	 * We must enable userspace access *before* advertising in the user page
1242 	 * that it is possible to do so to avoid any race.
1243 	 * And we must notify all cpus here because threads that currently run
1244 	 * on other cpus will try to directly access the counter too without
1245 	 * calling pmu_sbi_ctr_start.
1246 	 */
1247 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
1248 		on_each_cpu_mask(mm_cpumask(mm),
1249 				 pmu_sbi_set_scounteren, (void *)event, 1);
1250 }
1251 
1252 static void pmu_sbi_event_unmapped(struct perf_event *event, struct mm_struct *mm)
1253 {
1254 	if (event->hw.flags & PERF_EVENT_FLAG_NO_USER_ACCESS)
1255 		return;
1256 
1257 	if (event->hw.flags & PERF_EVENT_FLAG_LEGACY) {
1258 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES &&
1259 		    event->attr.config != PERF_COUNT_HW_INSTRUCTIONS) {
1260 			return;
1261 		}
1262 	}
1263 
1264 	/*
1265 	 * Here we can directly remove user access since the user does not have
1266 	 * access to the user page anymore so we avoid the racy window where the
1267 	 * user could have read cap_user_rdpmc to true right before we disable
1268 	 * it.
1269 	 */
1270 	event->hw.flags &= ~PERF_EVENT_FLAG_USER_READ_CNT;
1271 
1272 	if (event->hw.flags & PERF_EVENT_FLAG_USER_ACCESS)
1273 		on_each_cpu_mask(mm_cpumask(mm),
1274 				 pmu_sbi_reset_scounteren, (void *)event, 1);
1275 }
1276 
1277 static void riscv_pmu_update_counter_access(void *info)
1278 {
1279 	if (sysctl_perf_user_access == SYSCTL_LEGACY)
1280 		csr_write(CSR_SCOUNTEREN, 0x7);
1281 	else
1282 		csr_write(CSR_SCOUNTEREN, 0x2);
1283 }
1284 
1285 static int riscv_pmu_proc_user_access_handler(const struct ctl_table *table,
1286 					      int write, void *buffer,
1287 					      size_t *lenp, loff_t *ppos)
1288 {
1289 	int prev = sysctl_perf_user_access;
1290 	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1291 
1292 	/*
1293 	 * Test against the previous value since we clear SCOUNTEREN when
1294 	 * sysctl_perf_user_access is set to SYSCTL_USER_ACCESS, but we should
1295 	 * not do that if that was already the case.
1296 	 */
1297 	if (ret || !write || prev == sysctl_perf_user_access)
1298 		return ret;
1299 
1300 	on_each_cpu(riscv_pmu_update_counter_access, NULL, 1);
1301 
1302 	return 0;
1303 }
1304 
1305 static struct ctl_table sbi_pmu_sysctl_table[] = {
1306 	{
1307 		.procname       = "perf_user_access",
1308 		.data		= &sysctl_perf_user_access,
1309 		.maxlen		= sizeof(unsigned int),
1310 		.mode           = 0644,
1311 		.proc_handler	= riscv_pmu_proc_user_access_handler,
1312 		.extra1		= SYSCTL_ZERO,
1313 		.extra2		= SYSCTL_TWO,
1314 	},
1315 };
1316 
1317 static int pmu_sbi_device_probe(struct platform_device *pdev)
1318 {
1319 	struct riscv_pmu *pmu = NULL;
1320 	int ret = -ENODEV;
1321 	int num_counters;
1322 
1323 	pr_info("SBI PMU extension is available\n");
1324 	pmu = riscv_pmu_alloc();
1325 	if (!pmu)
1326 		return -ENOMEM;
1327 
1328 	num_counters = pmu_sbi_find_num_ctrs();
1329 	if (num_counters < 0) {
1330 		pr_err("SBI PMU extension doesn't provide any counters\n");
1331 		goto out_free;
1332 	}
1333 
1334 	/* It is possible to get from SBI more than max number of counters */
1335 	if (num_counters > RISCV_MAX_COUNTERS) {
1336 		num_counters = RISCV_MAX_COUNTERS;
1337 		pr_info("SBI returned more than maximum number of counters. Limiting the number of counters to %d\n", num_counters);
1338 	}
1339 
1340 	/* cache all the information about counters now */
1341 	if (pmu_sbi_get_ctrinfo(num_counters, &cmask))
1342 		goto out_free;
1343 
1344 	ret = pmu_sbi_setup_irqs(pmu, pdev);
1345 	if (ret < 0) {
1346 		pr_info("Perf sampling/filtering is not supported as sscof extension is not available\n");
1347 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1348 		pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
1349 	}
1350 
1351 	pmu->pmu.attr_groups = riscv_pmu_attr_groups;
1352 	pmu->pmu.parent = &pdev->dev;
1353 	pmu->cmask = cmask;
1354 	pmu->ctr_start = pmu_sbi_ctr_start;
1355 	pmu->ctr_stop = pmu_sbi_ctr_stop;
1356 	pmu->event_map = pmu_sbi_event_map;
1357 	pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
1358 	pmu->ctr_get_width = pmu_sbi_ctr_get_width;
1359 	pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
1360 	pmu->ctr_read = pmu_sbi_ctr_read;
1361 	pmu->event_init = pmu_sbi_event_init;
1362 	pmu->event_mapped = pmu_sbi_event_mapped;
1363 	pmu->event_unmapped = pmu_sbi_event_unmapped;
1364 	pmu->csr_index = pmu_sbi_csr_index;
1365 
1366 	ret = riscv_pm_pmu_register(pmu);
1367 	if (ret)
1368 		goto out_unregister;
1369 
1370 	ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
1371 	if (ret)
1372 		goto out_unregister;
1373 
1374 	/* SBI PMU Snapsphot is only available in SBI v2.0 */
1375 	if (sbi_v2_available) {
1376 		ret = pmu_sbi_snapshot_alloc(pmu);
1377 		if (ret)
1378 			goto out_unregister;
1379 
1380 		ret = pmu_sbi_snapshot_setup(pmu, smp_processor_id());
1381 		if (ret) {
1382 			/* Snapshot is an optional feature. Continue if not available */
1383 			pmu_sbi_snapshot_free(pmu);
1384 		} else {
1385 			pr_info("SBI PMU snapshot detected\n");
1386 			/*
1387 			 * We enable it once here for the boot cpu. If snapshot shmem setup
1388 			 * fails during cpu hotplug process, it will fail to start the cpu
1389 			 * as we can not handle hetergenous PMUs with different snapshot
1390 			 * capability.
1391 			 */
1392 			static_branch_enable(&sbi_pmu_snapshot_available);
1393 		}
1394 	}
1395 
1396 	register_sysctl("kernel", sbi_pmu_sysctl_table);
1397 
1398 	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
1399 	if (ret)
1400 		goto out_unregister;
1401 
1402 	/* Asynchronously check which standard events are available */
1403 	schedule_work(&check_std_events_work);
1404 
1405 	return 0;
1406 
1407 out_unregister:
1408 	riscv_pmu_destroy(pmu);
1409 
1410 out_free:
1411 	kfree(pmu);
1412 	return ret;
1413 }
1414 
1415 static struct platform_driver pmu_sbi_driver = {
1416 	.probe		= pmu_sbi_device_probe,
1417 	.driver		= {
1418 		.name	= RISCV_PMU_SBI_PDEV_NAME,
1419 	},
1420 };
1421 
1422 static int __init pmu_sbi_devinit(void)
1423 {
1424 	int ret;
1425 	struct platform_device *pdev;
1426 
1427 	if (sbi_spec_version < sbi_mk_version(0, 3) ||
1428 	    !sbi_probe_extension(SBI_EXT_PMU)) {
1429 		return 0;
1430 	}
1431 
1432 	if (sbi_spec_version >= sbi_mk_version(2, 0))
1433 		sbi_v2_available = true;
1434 
1435 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
1436 				      "perf/riscv/pmu:starting",
1437 				      pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
1438 	if (ret) {
1439 		pr_err("CPU hotplug notifier could not be registered: %d\n",
1440 		       ret);
1441 		return ret;
1442 	}
1443 
1444 	ret = platform_driver_register(&pmu_sbi_driver);
1445 	if (ret)
1446 		return ret;
1447 
1448 	pdev = platform_device_register_simple(RISCV_PMU_SBI_PDEV_NAME, -1, NULL, 0);
1449 	if (IS_ERR(pdev)) {
1450 		platform_driver_unregister(&pmu_sbi_driver);
1451 		return PTR_ERR(pdev);
1452 	}
1453 
1454 	/* Notify legacy implementation that SBI pmu is available*/
1455 	riscv_pmu_legacy_skip_init();
1456 
1457 	return ret;
1458 }
1459 device_initcall(pmu_sbi_devinit)
1460