xref: /linux/drivers/perf/arm_spe_pmu.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Perf support for the Statistical Profiling Extension, introduced as
4  * part of ARMv8.2.
5  *
6  * Copyright (C) 2016 ARM Limited
7  *
8  * Author: Will Deacon <will.deacon@arm.com>
9  */
10 
11 #define PMUNAME					"arm_spe"
12 #define DRVNAME					PMUNAME "_pmu"
13 #define pr_fmt(fmt)				DRVNAME ": " fmt
14 
15 #include <linux/bitfield.h>
16 #include <linux/bitops.h>
17 #include <linux/bug.h>
18 #include <linux/capability.h>
19 #include <linux/cpuhotplug.h>
20 #include <linux/cpumask.h>
21 #include <linux/device.h>
22 #include <linux/errno.h>
23 #include <linux/interrupt.h>
24 #include <linux/irq.h>
25 #include <linux/kernel.h>
26 #include <linux/list.h>
27 #include <linux/module.h>
28 #include <linux/of.h>
29 #include <linux/perf_event.h>
30 #include <linux/perf/arm_pmu.h>
31 #include <linux/platform_device.h>
32 #include <linux/printk.h>
33 #include <linux/slab.h>
34 #include <linux/smp.h>
35 #include <linux/vmalloc.h>
36 
37 #include <asm/barrier.h>
38 #include <asm/cpufeature.h>
39 #include <asm/mmu.h>
40 #include <asm/sysreg.h>
41 
42 /*
43  * Cache if the event is allowed to trace Context information.
44  * This allows us to perform the check, i.e, perf_allow_kernel(),
45  * in the context of the event owner, once, during the event_init().
46  */
47 #define SPE_PMU_HW_FLAGS_CX			0x00001
48 
49 static_assert((PERF_EVENT_FLAG_ARCH & SPE_PMU_HW_FLAGS_CX) == SPE_PMU_HW_FLAGS_CX);
50 
51 static void set_spe_event_has_cx(struct perf_event *event)
52 {
53 	if (IS_ENABLED(CONFIG_PID_IN_CONTEXTIDR) && !perf_allow_kernel())
54 		event->hw.flags |= SPE_PMU_HW_FLAGS_CX;
55 }
56 
57 static bool get_spe_event_has_cx(struct perf_event *event)
58 {
59 	return !!(event->hw.flags & SPE_PMU_HW_FLAGS_CX);
60 }
61 
62 #define ARM_SPE_BUF_PAD_BYTE			0
63 
64 struct arm_spe_pmu_buf {
65 	int					nr_pages;
66 	bool					snapshot;
67 	void					*base;
68 };
69 
70 struct arm_spe_pmu {
71 	struct pmu				pmu;
72 	struct platform_device			*pdev;
73 	cpumask_t				supported_cpus;
74 	struct hlist_node			hotplug_node;
75 
76 	int					irq; /* PPI */
77 	u16					pmsver;
78 	u16					min_period;
79 	u16					counter_sz;
80 
81 #define SPE_PMU_FEAT_FILT_EVT			(1UL << 0)
82 #define SPE_PMU_FEAT_FILT_TYP			(1UL << 1)
83 #define SPE_PMU_FEAT_FILT_LAT			(1UL << 2)
84 #define SPE_PMU_FEAT_ARCH_INST			(1UL << 3)
85 #define SPE_PMU_FEAT_LDS			(1UL << 4)
86 #define SPE_PMU_FEAT_ERND			(1UL << 5)
87 #define SPE_PMU_FEAT_INV_FILT_EVT		(1UL << 6)
88 #define SPE_PMU_FEAT_DISCARD			(1UL << 7)
89 #define SPE_PMU_FEAT_EFT			(1UL << 8)
90 #define SPE_PMU_FEAT_FDS			(1UL << 9)
91 #define SPE_PMU_FEAT_DEV_PROBED			(1UL << 63)
92 	u64					features;
93 
94 	u64					pmsevfr_res0;
95 	u16					max_record_sz;
96 	u16					align;
97 	struct perf_output_handle __percpu	*handle;
98 };
99 
100 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
101 
102 /* Convert a free-running index from perf into an SPE buffer offset */
103 #define PERF_IDX2OFF(idx, buf) \
104 	((idx) % ((unsigned long)(buf)->nr_pages << PAGE_SHIFT))
105 
106 /* Keep track of our dynamic hotplug state */
107 static enum cpuhp_state arm_spe_pmu_online;
108 
109 enum arm_spe_pmu_buf_fault_action {
110 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
111 	SPE_PMU_BUF_FAULT_ACT_FATAL,
112 	SPE_PMU_BUF_FAULT_ACT_OK,
113 };
114 
115 /* This sysfs gunk was really good fun to write. */
116 enum arm_spe_pmu_capabilities {
117 	SPE_PMU_CAP_ARCH_INST = 0,
118 	SPE_PMU_CAP_ERND,
119 	SPE_PMU_CAP_FEAT_MAX,
120 	SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
121 	SPE_PMU_CAP_MIN_IVAL,
122 	SPE_PMU_CAP_EVENT_FILTER,
123 };
124 
125 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
126 	[SPE_PMU_CAP_ARCH_INST]	= SPE_PMU_FEAT_ARCH_INST,
127 	[SPE_PMU_CAP_ERND]	= SPE_PMU_FEAT_ERND,
128 };
129 
130 static u64 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
131 {
132 	if (cap < SPE_PMU_CAP_FEAT_MAX)
133 		return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
134 
135 	switch (cap) {
136 	case SPE_PMU_CAP_CNT_SZ:
137 		return spe_pmu->counter_sz;
138 	case SPE_PMU_CAP_MIN_IVAL:
139 		return spe_pmu->min_period;
140 	case SPE_PMU_CAP_EVENT_FILTER:
141 		return ~spe_pmu->pmsevfr_res0;
142 	default:
143 		WARN(1, "unknown cap %d\n", cap);
144 	}
145 
146 	return 0;
147 }
148 
149 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
150 				    struct device_attribute *attr,
151 				    char *buf)
152 {
153 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
154 	struct dev_ext_attribute *ea =
155 		container_of(attr, struct dev_ext_attribute, attr);
156 	int cap = (long)ea->var;
157 
158 	return sysfs_emit(buf, "%llu\n", arm_spe_pmu_cap_get(spe_pmu, cap));
159 }
160 
161 static ssize_t arm_spe_pmu_cap_show_hex(struct device *dev,
162 					struct device_attribute *attr,
163 					char *buf)
164 {
165 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
166 	struct dev_ext_attribute *ea =
167 		container_of(attr, struct dev_ext_attribute, attr);
168 	int cap = (long)ea->var;
169 
170 	return sysfs_emit(buf, "0x%llx\n", arm_spe_pmu_cap_get(spe_pmu, cap));
171 }
172 
173 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
174 	&((struct dev_ext_attribute[]) {				\
175 		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
176 	})[0].attr.attr
177 
178 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
179 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
180 #define SPE_CAP_EXT_ATTR_ENTRY_HEX(_name, _var)				\
181 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show_hex, _var)
182 
183 static struct attribute *arm_spe_pmu_cap_attr[] = {
184 	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
185 	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
186 	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
187 	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
188 	SPE_CAP_EXT_ATTR_ENTRY_HEX(event_filter, SPE_PMU_CAP_EVENT_FILTER),
189 	NULL,
190 };
191 
192 static const struct attribute_group arm_spe_pmu_cap_group = {
193 	.name	= "caps",
194 	.attrs	= arm_spe_pmu_cap_attr,
195 };
196 
197 /* User ABI */
198 #define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
199 #define ATTR_CFG_FLD_ts_enable_LO		0
200 #define ATTR_CFG_FLD_ts_enable_HI		0
201 #define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
202 #define ATTR_CFG_FLD_pa_enable_LO		1
203 #define ATTR_CFG_FLD_pa_enable_HI		1
204 #define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
205 #define ATTR_CFG_FLD_pct_enable_LO		2
206 #define ATTR_CFG_FLD_pct_enable_HI		2
207 #define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
208 #define ATTR_CFG_FLD_jitter_LO			16
209 #define ATTR_CFG_FLD_jitter_HI			16
210 #define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
211 #define ATTR_CFG_FLD_branch_filter_LO		32
212 #define ATTR_CFG_FLD_branch_filter_HI		32
213 #define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
214 #define ATTR_CFG_FLD_load_filter_LO		33
215 #define ATTR_CFG_FLD_load_filter_HI		33
216 #define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
217 #define ATTR_CFG_FLD_store_filter_LO		34
218 #define ATTR_CFG_FLD_store_filter_HI		34
219 #define ATTR_CFG_FLD_discard_CFG		config	/* PMBLIMITR_EL1.FM = DISCARD */
220 #define ATTR_CFG_FLD_discard_LO			35
221 #define ATTR_CFG_FLD_discard_HI			35
222 #define ATTR_CFG_FLD_branch_filter_mask_CFG	config	/* PMSFCR_EL1.Bm */
223 #define ATTR_CFG_FLD_branch_filter_mask_LO	36
224 #define ATTR_CFG_FLD_branch_filter_mask_HI	36
225 #define ATTR_CFG_FLD_load_filter_mask_CFG	config	/* PMSFCR_EL1.LDm */
226 #define ATTR_CFG_FLD_load_filter_mask_LO	37
227 #define ATTR_CFG_FLD_load_filter_mask_HI	37
228 #define ATTR_CFG_FLD_store_filter_mask_CFG	config	/* PMSFCR_EL1.STm */
229 #define ATTR_CFG_FLD_store_filter_mask_LO	38
230 #define ATTR_CFG_FLD_store_filter_mask_HI	38
231 #define ATTR_CFG_FLD_simd_filter_CFG		config	/* PMSFCR_EL1.SIMD */
232 #define ATTR_CFG_FLD_simd_filter_LO		39
233 #define ATTR_CFG_FLD_simd_filter_HI		39
234 #define ATTR_CFG_FLD_simd_filter_mask_CFG	config	/* PMSFCR_EL1.SIMDm */
235 #define ATTR_CFG_FLD_simd_filter_mask_LO	40
236 #define ATTR_CFG_FLD_simd_filter_mask_HI	40
237 #define ATTR_CFG_FLD_float_filter_CFG		config	/* PMSFCR_EL1.FP */
238 #define ATTR_CFG_FLD_float_filter_LO		41
239 #define ATTR_CFG_FLD_float_filter_HI		41
240 #define ATTR_CFG_FLD_float_filter_mask_CFG	config	/* PMSFCR_EL1.FPm */
241 #define ATTR_CFG_FLD_float_filter_mask_LO	42
242 #define ATTR_CFG_FLD_float_filter_mask_HI	42
243 
244 #define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
245 #define ATTR_CFG_FLD_event_filter_LO		0
246 #define ATTR_CFG_FLD_event_filter_HI		63
247 
248 #define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
249 #define ATTR_CFG_FLD_min_latency_LO		0
250 #define ATTR_CFG_FLD_min_latency_HI		11
251 
252 #define ATTR_CFG_FLD_inv_event_filter_CFG	config3	/* PMSNEVFR_EL1 */
253 #define ATTR_CFG_FLD_inv_event_filter_LO	0
254 #define ATTR_CFG_FLD_inv_event_filter_HI	63
255 
256 #define ATTR_CFG_FLD_inv_data_src_filter_CFG	config4	/* inverse of PMSDSFR_EL1 */
257 #define ATTR_CFG_FLD_inv_data_src_filter_LO	0
258 #define ATTR_CFG_FLD_inv_data_src_filter_HI	63
259 
260 GEN_PMU_FORMAT_ATTR(ts_enable);
261 GEN_PMU_FORMAT_ATTR(pa_enable);
262 GEN_PMU_FORMAT_ATTR(pct_enable);
263 GEN_PMU_FORMAT_ATTR(jitter);
264 GEN_PMU_FORMAT_ATTR(branch_filter);
265 GEN_PMU_FORMAT_ATTR(branch_filter_mask);
266 GEN_PMU_FORMAT_ATTR(load_filter);
267 GEN_PMU_FORMAT_ATTR(load_filter_mask);
268 GEN_PMU_FORMAT_ATTR(store_filter);
269 GEN_PMU_FORMAT_ATTR(store_filter_mask);
270 GEN_PMU_FORMAT_ATTR(simd_filter);
271 GEN_PMU_FORMAT_ATTR(simd_filter_mask);
272 GEN_PMU_FORMAT_ATTR(float_filter);
273 GEN_PMU_FORMAT_ATTR(float_filter_mask);
274 GEN_PMU_FORMAT_ATTR(event_filter);
275 GEN_PMU_FORMAT_ATTR(inv_event_filter);
276 GEN_PMU_FORMAT_ATTR(inv_data_src_filter);
277 GEN_PMU_FORMAT_ATTR(min_latency);
278 GEN_PMU_FORMAT_ATTR(discard);
279 
280 static struct attribute *arm_spe_pmu_formats_attr[] = {
281 	&format_attr_ts_enable.attr,
282 	&format_attr_pa_enable.attr,
283 	&format_attr_pct_enable.attr,
284 	&format_attr_jitter.attr,
285 	&format_attr_branch_filter.attr,
286 	&format_attr_branch_filter_mask.attr,
287 	&format_attr_load_filter.attr,
288 	&format_attr_load_filter_mask.attr,
289 	&format_attr_store_filter.attr,
290 	&format_attr_store_filter_mask.attr,
291 	&format_attr_simd_filter.attr,
292 	&format_attr_simd_filter_mask.attr,
293 	&format_attr_float_filter.attr,
294 	&format_attr_float_filter_mask.attr,
295 	&format_attr_event_filter.attr,
296 	&format_attr_inv_event_filter.attr,
297 	&format_attr_inv_data_src_filter.attr,
298 	&format_attr_min_latency.attr,
299 	&format_attr_discard.attr,
300 	NULL,
301 };
302 
303 static umode_t arm_spe_pmu_format_attr_is_visible(struct kobject *kobj,
304 						  struct attribute *attr,
305 						  int unused)
306 	{
307 	struct device *dev = kobj_to_dev(kobj);
308 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
309 
310 	if (attr == &format_attr_discard.attr && !(spe_pmu->features & SPE_PMU_FEAT_DISCARD))
311 		return 0;
312 
313 	if (attr == &format_attr_inv_event_filter.attr && !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
314 		return 0;
315 
316 	if (attr == &format_attr_inv_data_src_filter.attr &&
317 	    !(spe_pmu->features & SPE_PMU_FEAT_FDS))
318 		return 0;
319 
320 	if ((attr == &format_attr_branch_filter_mask.attr ||
321 	     attr == &format_attr_load_filter_mask.attr ||
322 	     attr == &format_attr_store_filter_mask.attr ||
323 	     attr == &format_attr_simd_filter.attr ||
324 	     attr == &format_attr_simd_filter_mask.attr ||
325 	     attr == &format_attr_float_filter.attr ||
326 	     attr == &format_attr_float_filter_mask.attr) &&
327 	     !(spe_pmu->features & SPE_PMU_FEAT_EFT))
328 		return 0;
329 
330 	return attr->mode;
331 }
332 
333 static const struct attribute_group arm_spe_pmu_format_group = {
334 	.name	= "format",
335 	.is_visible = arm_spe_pmu_format_attr_is_visible,
336 	.attrs	= arm_spe_pmu_formats_attr,
337 };
338 
339 static ssize_t cpumask_show(struct device *dev,
340 			    struct device_attribute *attr, char *buf)
341 {
342 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
343 
344 	return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
345 }
346 static DEVICE_ATTR_RO(cpumask);
347 
348 static struct attribute *arm_spe_pmu_attrs[] = {
349 	&dev_attr_cpumask.attr,
350 	NULL,
351 };
352 
353 static const struct attribute_group arm_spe_pmu_group = {
354 	.attrs	= arm_spe_pmu_attrs,
355 };
356 
357 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
358 	&arm_spe_pmu_group,
359 	&arm_spe_pmu_cap_group,
360 	&arm_spe_pmu_format_group,
361 	NULL,
362 };
363 
364 /* Convert between user ABI and register values */
365 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
366 {
367 	struct perf_event_attr *attr = &event->attr;
368 	u64 reg = 0;
369 
370 	reg |= FIELD_PREP(PMSCR_EL1_TS, ATTR_CFG_GET_FLD(attr, ts_enable));
371 	reg |= FIELD_PREP(PMSCR_EL1_PA, ATTR_CFG_GET_FLD(attr, pa_enable));
372 	reg |= FIELD_PREP(PMSCR_EL1_PCT, ATTR_CFG_GET_FLD(attr, pct_enable));
373 
374 	if (!attr->exclude_user)
375 		reg |= PMSCR_EL1_E0SPE;
376 
377 	if (!attr->exclude_kernel)
378 		reg |= PMSCR_EL1_E1SPE;
379 
380 	if (get_spe_event_has_cx(event))
381 		reg |= PMSCR_EL1_CX;
382 
383 	return reg;
384 }
385 
386 static void arm_spe_event_sanitise_period(struct perf_event *event)
387 {
388 	u64 period = event->hw.sample_period;
389 	u64 max_period = PMSIRR_EL1_INTERVAL_MASK;
390 
391 	/*
392 	 * The PMSIDR_EL1.Interval field (stored in spe_pmu->min_period) is a
393 	 * recommendation for the minimum interval, not a hardware limitation.
394 	 *
395 	 * According to the Arm ARM (DDI 0487 L.a), section D24.7.12 PMSIRR_EL1,
396 	 * Sampling Interval Reload Register, the INTERVAL field (bits [31:8])
397 	 * states: "Software must set this to a nonzero value". Use 1 as the
398 	 * minimum value.
399 	 */
400 	u64 min_period = FIELD_PREP(PMSIRR_EL1_INTERVAL_MASK, 1);
401 
402 	period = clamp_t(u64, period, min_period, max_period) & max_period;
403 	event->hw.sample_period = period;
404 }
405 
406 static u64 arm_spe_event_to_pmsirr(struct perf_event *event)
407 {
408 	struct perf_event_attr *attr = &event->attr;
409 	u64 reg = 0;
410 
411 	arm_spe_event_sanitise_period(event);
412 
413 	reg |= FIELD_PREP(PMSIRR_EL1_RND, ATTR_CFG_GET_FLD(attr, jitter));
414 	reg |= event->hw.sample_period;
415 
416 	return reg;
417 }
418 
419 static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
420 {
421 	struct perf_event_attr *attr = &event->attr;
422 	u64 reg = 0;
423 
424 	reg |= FIELD_PREP(PMSFCR_EL1_LD, ATTR_CFG_GET_FLD(attr, load_filter));
425 	reg |= FIELD_PREP(PMSFCR_EL1_LDm, ATTR_CFG_GET_FLD(attr, load_filter_mask));
426 	reg |= FIELD_PREP(PMSFCR_EL1_ST, ATTR_CFG_GET_FLD(attr, store_filter));
427 	reg |= FIELD_PREP(PMSFCR_EL1_STm, ATTR_CFG_GET_FLD(attr, store_filter_mask));
428 	reg |= FIELD_PREP(PMSFCR_EL1_B, ATTR_CFG_GET_FLD(attr, branch_filter));
429 	reg |= FIELD_PREP(PMSFCR_EL1_Bm, ATTR_CFG_GET_FLD(attr, branch_filter_mask));
430 	reg |= FIELD_PREP(PMSFCR_EL1_SIMD, ATTR_CFG_GET_FLD(attr, simd_filter));
431 	reg |= FIELD_PREP(PMSFCR_EL1_SIMDm, ATTR_CFG_GET_FLD(attr, simd_filter_mask));
432 	reg |= FIELD_PREP(PMSFCR_EL1_FP, ATTR_CFG_GET_FLD(attr, float_filter));
433 	reg |= FIELD_PREP(PMSFCR_EL1_FPm, ATTR_CFG_GET_FLD(attr, float_filter_mask));
434 
435 	if (reg)
436 		reg |= PMSFCR_EL1_FT;
437 
438 	if (ATTR_CFG_GET_FLD(attr, event_filter))
439 		reg |= PMSFCR_EL1_FE;
440 
441 	if (ATTR_CFG_GET_FLD(attr, inv_event_filter))
442 		reg |= PMSFCR_EL1_FnE;
443 
444 	if (ATTR_CFG_GET_FLD(attr, inv_data_src_filter))
445 		reg |= PMSFCR_EL1_FDS;
446 
447 	if (ATTR_CFG_GET_FLD(attr, min_latency))
448 		reg |= PMSFCR_EL1_FL;
449 
450 	return reg;
451 }
452 
453 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
454 {
455 	struct perf_event_attr *attr = &event->attr;
456 	return ATTR_CFG_GET_FLD(attr, event_filter);
457 }
458 
459 static u64 arm_spe_event_to_pmsnevfr(struct perf_event *event)
460 {
461 	struct perf_event_attr *attr = &event->attr;
462 	return ATTR_CFG_GET_FLD(attr, inv_event_filter);
463 }
464 
465 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
466 {
467 	struct perf_event_attr *attr = &event->attr;
468 	return FIELD_PREP(PMSLATFR_EL1_MINLAT, ATTR_CFG_GET_FLD(attr, min_latency));
469 }
470 
471 static u64 arm_spe_event_to_pmsdsfr(struct perf_event *event)
472 {
473 	struct perf_event_attr *attr = &event->attr;
474 
475 	/*
476 	 * Data src filter is inverted so that the default value of 0 is
477 	 * equivalent to no filtering.
478 	 */
479 	return ~ATTR_CFG_GET_FLD(attr, inv_data_src_filter);
480 }
481 
482 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
483 {
484 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
485 	u64 head = PERF_IDX2OFF(handle->head, buf);
486 
487 	memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
488 	if (!buf->snapshot)
489 		perf_aux_output_skip(handle, len);
490 }
491 
492 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
493 {
494 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
495 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
496 	u64 head = PERF_IDX2OFF(handle->head, buf);
497 	u64 limit = buf->nr_pages * PAGE_SIZE;
498 
499 	/*
500 	 * The trace format isn't parseable in reverse, so clamp
501 	 * the limit to half of the buffer size in snapshot mode
502 	 * so that the worst case is half a buffer of records, as
503 	 * opposed to a single record.
504 	 */
505 	if (head < limit >> 1)
506 		limit >>= 1;
507 
508 	/*
509 	 * If we're within max_record_sz of the limit, we must
510 	 * pad, move the head index and recompute the limit.
511 	 */
512 	if (limit - head < spe_pmu->max_record_sz) {
513 		arm_spe_pmu_pad_buf(handle, limit - head);
514 		handle->head = PERF_IDX2OFF(limit, buf);
515 		limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
516 	}
517 
518 	return limit;
519 }
520 
521 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
522 {
523 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
524 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
525 	const u64 bufsize = buf->nr_pages * PAGE_SIZE;
526 	u64 limit = bufsize;
527 	u64 head, tail, wakeup;
528 
529 	/*
530 	 * The head can be misaligned for two reasons:
531 	 *
532 	 * 1. The hardware left PMBPTR pointing to the first byte after
533 	 *    a record when generating a buffer management event.
534 	 *
535 	 * 2. We used perf_aux_output_skip to consume handle->size bytes
536 	 *    and CIRC_SPACE was used to compute the size, which always
537 	 *    leaves one entry free.
538 	 *
539 	 * Deal with this by padding to the next alignment boundary and
540 	 * moving the head index. If we run out of buffer space, we'll
541 	 * reduce handle->size to zero and end up reporting truncation.
542 	 */
543 	head = PERF_IDX2OFF(handle->head, buf);
544 	if (!IS_ALIGNED(head, spe_pmu->align)) {
545 		unsigned long delta = roundup(head, spe_pmu->align) - head;
546 
547 		delta = min(delta, handle->size);
548 		arm_spe_pmu_pad_buf(handle, delta);
549 		head = PERF_IDX2OFF(handle->head, buf);
550 	}
551 
552 	/* If we've run out of free space, then nothing more to do */
553 	if (!handle->size)
554 		goto no_space;
555 
556 	/* Compute the tail and wakeup indices now that we've aligned head */
557 	tail = PERF_IDX2OFF(handle->head + handle->size, buf);
558 	wakeup = PERF_IDX2OFF(handle->wakeup, buf);
559 
560 	/*
561 	 * Avoid clobbering unconsumed data. We know we have space, so
562 	 * if we see head == tail we know that the buffer is empty. If
563 	 * head > tail, then there's nothing to clobber prior to
564 	 * wrapping.
565 	 */
566 	if (head < tail)
567 		limit = round_down(tail, PAGE_SIZE);
568 
569 	/*
570 	 * Wakeup may be arbitrarily far into the future. If it's not in
571 	 * the current generation, either we'll wrap before hitting it,
572 	 * or it's in the past and has been handled already.
573 	 *
574 	 * If there's a wakeup before we wrap, arrange to be woken up by
575 	 * the page boundary following it. Keep the tail boundary if
576 	 * that's lower.
577 	 */
578 	if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
579 		limit = min(limit, round_up(wakeup, PAGE_SIZE));
580 
581 	if (limit > head)
582 		return limit;
583 
584 	arm_spe_pmu_pad_buf(handle, handle->size);
585 no_space:
586 	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
587 	perf_aux_output_end(handle, 0);
588 	return 0;
589 }
590 
591 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
592 {
593 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
594 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
595 	u64 limit = __arm_spe_pmu_next_off(handle);
596 	u64 head = PERF_IDX2OFF(handle->head, buf);
597 
598 	/*
599 	 * If the head has come too close to the end of the buffer,
600 	 * then pad to the end and recompute the limit.
601 	 */
602 	if (limit && (limit - head < spe_pmu->max_record_sz)) {
603 		arm_spe_pmu_pad_buf(handle, limit - head);
604 		limit = __arm_spe_pmu_next_off(handle);
605 	}
606 
607 	return limit;
608 }
609 
610 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
611 					  struct perf_event *event)
612 {
613 	u64 base, limit;
614 	struct arm_spe_pmu_buf *buf;
615 
616 	if (ATTR_CFG_GET_FLD(&event->attr, discard)) {
617 		limit = FIELD_PREP(PMBLIMITR_EL1_FM, PMBLIMITR_EL1_FM_DISCARD);
618 		limit |= PMBLIMITR_EL1_E;
619 		goto out_write_limit;
620 	}
621 
622 	/* Start a new aux session */
623 	buf = perf_aux_output_begin(handle, event);
624 	if (!buf) {
625 		event->hw.state |= PERF_HES_STOPPED;
626 		/*
627 		 * We still need to clear the limit pointer, since the
628 		 * profiler might only be disabled by virtue of a fault.
629 		 */
630 		limit = 0;
631 		goto out_write_limit;
632 	}
633 
634 	limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
635 			      : arm_spe_pmu_next_off(handle);
636 	if (limit)
637 		limit |= PMBLIMITR_EL1_E;
638 
639 	limit += (u64)buf->base;
640 	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
641 	write_sysreg_s(base, SYS_PMBPTR_EL1);
642 
643 out_write_limit:
644 	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
645 }
646 
647 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
648 {
649 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
650 	u64 offset, size;
651 
652 	offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
653 	size = offset - PERF_IDX2OFF(handle->head, buf);
654 
655 	if (buf->snapshot)
656 		handle->head = offset;
657 
658 	perf_aux_output_end(handle, size);
659 }
660 
661 static void arm_spe_pmu_disable_and_drain_local(void)
662 {
663 	/* Disable profiling at EL0 and EL1 */
664 	write_sysreg_s(0, SYS_PMSCR_EL1);
665 	isb();
666 
667 	/* Drain any buffered data */
668 	psb_csync();
669 	dsb(nsh);
670 
671 	/* Disable the profiling buffer */
672 	write_sysreg_s(0, SYS_PMBLIMITR_EL1);
673 	isb();
674 }
675 
676 /* IRQ handling */
677 static enum arm_spe_pmu_buf_fault_action
678 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
679 {
680 	const char *err_str;
681 	u64 pmbsr;
682 	enum arm_spe_pmu_buf_fault_action ret;
683 
684 	/*
685 	 * Ensure new profiling data is visible to the CPU and any external
686 	 * aborts have been resolved.
687 	 */
688 	psb_csync();
689 	dsb(nsh);
690 
691 	/* Ensure hardware updates to PMBPTR_EL1 are visible */
692 	isb();
693 
694 	/* Service required? */
695 	pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
696 	if (!FIELD_GET(PMBSR_EL1_S, pmbsr))
697 		return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
698 
699 	/*
700 	 * If we've lost data, disable profiling and also set the PARTIAL
701 	 * flag to indicate that the last record is corrupted.
702 	 */
703 	if (FIELD_GET(PMBSR_EL1_DL, pmbsr))
704 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
705 					     PERF_AUX_FLAG_PARTIAL);
706 
707 	/* Report collisions to userspace so that it can up the period */
708 	if (FIELD_GET(PMBSR_EL1_COLL, pmbsr))
709 		perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
710 
711 	/* We only expect buffer management events */
712 	switch (FIELD_GET(PMBSR_EL1_EC, pmbsr)) {
713 	case PMBSR_EL1_EC_BUF:
714 		/* Handled below */
715 		break;
716 	case PMBSR_EL1_EC_FAULT_S1:
717 	case PMBSR_EL1_EC_FAULT_S2:
718 		err_str = "Unexpected buffer fault";
719 		goto out_err;
720 	default:
721 		err_str = "Unknown error code";
722 		goto out_err;
723 	}
724 
725 	/* Buffer management event */
726 	switch (FIELD_GET(PMBSR_EL1_BUF_BSC_MASK, pmbsr)) {
727 	case PMBSR_EL1_BUF_BSC_FULL:
728 		ret = SPE_PMU_BUF_FAULT_ACT_OK;
729 		goto out_stop;
730 	default:
731 		err_str = "Unknown buffer status code";
732 	}
733 
734 out_err:
735 	pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
736 			   err_str, smp_processor_id(), pmbsr,
737 			   read_sysreg_s(SYS_PMBPTR_EL1),
738 			   read_sysreg_s(SYS_PMBLIMITR_EL1));
739 	ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
740 
741 out_stop:
742 	arm_spe_perf_aux_output_end(handle);
743 	return ret;
744 }
745 
746 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
747 {
748 	struct perf_output_handle *handle = dev;
749 	struct perf_event *event = handle->event;
750 	enum arm_spe_pmu_buf_fault_action act;
751 
752 	if (!perf_get_aux(handle))
753 		return IRQ_NONE;
754 
755 	act = arm_spe_pmu_buf_get_fault_act(handle);
756 	if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
757 		return IRQ_NONE;
758 
759 	/*
760 	 * Ensure perf callbacks have completed, which may disable the
761 	 * profiling buffer in response to a TRUNCATION flag.
762 	 */
763 	irq_work_run();
764 
765 	switch (act) {
766 	case SPE_PMU_BUF_FAULT_ACT_FATAL:
767 		/*
768 		 * If a fatal exception occurred then leaving the profiling
769 		 * buffer enabled is a recipe waiting to happen. Since
770 		 * fatal faults don't always imply truncation, make sure
771 		 * that the profiling buffer is disabled explicitly before
772 		 * clearing the syndrome register.
773 		 */
774 		arm_spe_pmu_disable_and_drain_local();
775 		break;
776 	case SPE_PMU_BUF_FAULT_ACT_OK:
777 		/*
778 		 * We handled the fault (the buffer was full), so resume
779 		 * profiling as long as we didn't detect truncation.
780 		 * PMBPTR might be misaligned, but we'll burn that bridge
781 		 * when we get to it.
782 		 */
783 		if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
784 			arm_spe_perf_aux_output_begin(handle, event);
785 			isb();
786 		}
787 		break;
788 	case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
789 		/* We've seen you before, but GCC has the memory of a sieve. */
790 		break;
791 	}
792 
793 	/* The buffer pointers are now sane, so resume profiling. */
794 	write_sysreg_s(0, SYS_PMBSR_EL1);
795 	return IRQ_HANDLED;
796 }
797 
798 /* Perf callbacks */
799 static int arm_spe_pmu_event_init(struct perf_event *event)
800 {
801 	u64 reg;
802 	struct perf_event_attr *attr = &event->attr;
803 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
804 
805 	/* This is, of course, deeply driver-specific */
806 	if (attr->type != event->pmu->type)
807 		return -ENOENT;
808 
809 	if (event->cpu >= 0 &&
810 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
811 		return -ENOENT;
812 
813 	if (arm_spe_event_to_pmsevfr(event) & spe_pmu->pmsevfr_res0)
814 		return -EOPNOTSUPP;
815 
816 	if (arm_spe_event_to_pmsnevfr(event) & spe_pmu->pmsevfr_res0)
817 		return -EOPNOTSUPP;
818 
819 	if (arm_spe_event_to_pmsdsfr(event) != U64_MAX &&
820 	    !(spe_pmu->features & SPE_PMU_FEAT_FDS))
821 		return -EOPNOTSUPP;
822 
823 	if (attr->exclude_idle)
824 		return -EOPNOTSUPP;
825 
826 	/*
827 	 * Feedback-directed frequency throttling doesn't work when we
828 	 * have a buffer of samples. We'd need to manually count the
829 	 * samples in the buffer when it fills up and adjust the event
830 	 * count to reflect that. Instead, just force the user to specify
831 	 * a sample period.
832 	 */
833 	if (attr->freq)
834 		return -EINVAL;
835 
836 	reg = arm_spe_event_to_pmsfcr(event);
837 	if ((FIELD_GET(PMSFCR_EL1_FE, reg)) &&
838 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
839 		return -EOPNOTSUPP;
840 
841 	if ((FIELD_GET(PMSFCR_EL1_FnE, reg)) &&
842 	    !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
843 		return -EOPNOTSUPP;
844 
845 	if ((FIELD_GET(PMSFCR_EL1_FT, reg)) &&
846 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
847 		return -EOPNOTSUPP;
848 
849 	if ((FIELD_GET(PMSFCR_EL1_FL, reg)) &&
850 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
851 		return -EOPNOTSUPP;
852 
853 	if ((FIELD_GET(PMSFCR_EL1_LDm, reg) ||
854 	     FIELD_GET(PMSFCR_EL1_STm, reg) ||
855 	     FIELD_GET(PMSFCR_EL1_Bm, reg) ||
856 	     FIELD_GET(PMSFCR_EL1_SIMD, reg) ||
857 	     FIELD_GET(PMSFCR_EL1_SIMDm, reg) ||
858 	     FIELD_GET(PMSFCR_EL1_FP, reg) ||
859 	     FIELD_GET(PMSFCR_EL1_FPm, reg)) &&
860 	    !(spe_pmu->features & SPE_PMU_FEAT_EFT))
861 		return -EOPNOTSUPP;
862 
863 	if (ATTR_CFG_GET_FLD(&event->attr, discard) &&
864 	    !(spe_pmu->features & SPE_PMU_FEAT_DISCARD))
865 		return -EOPNOTSUPP;
866 
867 	set_spe_event_has_cx(event);
868 	reg = arm_spe_event_to_pmscr(event);
869 	if (reg & (PMSCR_EL1_PA | PMSCR_EL1_PCT))
870 		return perf_allow_kernel();
871 
872 	return 0;
873 }
874 
875 static void arm_spe_pmu_start(struct perf_event *event, int flags)
876 {
877 	u64 reg;
878 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
879 	struct hw_perf_event *hwc = &event->hw;
880 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
881 
882 	hwc->state = 0;
883 	arm_spe_perf_aux_output_begin(handle, event);
884 	if (hwc->state)
885 		return;
886 
887 	reg = arm_spe_event_to_pmsfcr(event);
888 	write_sysreg_s(reg, SYS_PMSFCR_EL1);
889 
890 	reg = arm_spe_event_to_pmsevfr(event);
891 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
892 
893 	if (spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT) {
894 		reg = arm_spe_event_to_pmsnevfr(event);
895 		write_sysreg_s(reg, SYS_PMSNEVFR_EL1);
896 	}
897 
898 	if (spe_pmu->features & SPE_PMU_FEAT_FDS) {
899 		reg = arm_spe_event_to_pmsdsfr(event);
900 		write_sysreg_s(reg, SYS_PMSDSFR_EL1);
901 	}
902 
903 	reg = arm_spe_event_to_pmslatfr(event);
904 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
905 
906 	if (flags & PERF_EF_RELOAD) {
907 		reg = arm_spe_event_to_pmsirr(event);
908 		write_sysreg_s(reg, SYS_PMSIRR_EL1);
909 		isb();
910 		reg = local64_read(&hwc->period_left);
911 		write_sysreg_s(reg, SYS_PMSICR_EL1);
912 	}
913 
914 	reg = arm_spe_event_to_pmscr(event);
915 	isb();
916 	write_sysreg_s(reg, SYS_PMSCR_EL1);
917 }
918 
919 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
920 {
921 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
922 	struct hw_perf_event *hwc = &event->hw;
923 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
924 
925 	/* If we're already stopped, then nothing to do */
926 	if (hwc->state & PERF_HES_STOPPED)
927 		return;
928 
929 	/* Stop all trace generation */
930 	arm_spe_pmu_disable_and_drain_local();
931 
932 	if (flags & PERF_EF_UPDATE) {
933 		/*
934 		 * If there's a fault pending then ensure we contain it
935 		 * to this buffer, since we might be on the context-switch
936 		 * path.
937 		 */
938 		if (perf_get_aux(handle)) {
939 			enum arm_spe_pmu_buf_fault_action act;
940 
941 			act = arm_spe_pmu_buf_get_fault_act(handle);
942 			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
943 				arm_spe_perf_aux_output_end(handle);
944 			else
945 				write_sysreg_s(0, SYS_PMBSR_EL1);
946 		}
947 
948 		/*
949 		 * This may also contain ECOUNT, but nobody else should
950 		 * be looking at period_left, since we forbid frequency
951 		 * based sampling.
952 		 */
953 		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
954 		hwc->state |= PERF_HES_UPTODATE;
955 	}
956 
957 	hwc->state |= PERF_HES_STOPPED;
958 }
959 
960 static int arm_spe_pmu_add(struct perf_event *event, int flags)
961 {
962 	int ret = 0;
963 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
964 	struct hw_perf_event *hwc = &event->hw;
965 	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
966 
967 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
968 		return -ENOENT;
969 
970 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
971 
972 	if (flags & PERF_EF_START) {
973 		arm_spe_pmu_start(event, PERF_EF_RELOAD);
974 		if (hwc->state & PERF_HES_STOPPED)
975 			ret = -EINVAL;
976 	}
977 
978 	return ret;
979 }
980 
981 static void arm_spe_pmu_del(struct perf_event *event, int flags)
982 {
983 	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
984 }
985 
986 static void arm_spe_pmu_read(struct perf_event *event)
987 {
988 }
989 
990 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
991 				   int nr_pages, bool snapshot)
992 {
993 	int i, cpu = event->cpu;
994 	struct page **pglist;
995 	struct arm_spe_pmu_buf *buf;
996 
997 	/* We need at least two pages for this to work. */
998 	if (nr_pages < 2)
999 		return NULL;
1000 
1001 	/*
1002 	 * We require an even number of pages for snapshot mode, so that
1003 	 * we can effectively treat the buffer as consisting of two equal
1004 	 * parts and give userspace a fighting chance of getting some
1005 	 * useful data out of it.
1006 	 */
1007 	if (snapshot && (nr_pages & 1))
1008 		return NULL;
1009 
1010 	if (cpu == -1)
1011 		cpu = raw_smp_processor_id();
1012 
1013 	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
1014 	if (!buf)
1015 		return NULL;
1016 
1017 	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
1018 	if (!pglist)
1019 		goto out_free_buf;
1020 
1021 	for (i = 0; i < nr_pages; ++i)
1022 		pglist[i] = virt_to_page(pages[i]);
1023 
1024 	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
1025 	if (!buf->base)
1026 		goto out_free_pglist;
1027 
1028 	buf->nr_pages	= nr_pages;
1029 	buf->snapshot	= snapshot;
1030 
1031 	kfree(pglist);
1032 	return buf;
1033 
1034 out_free_pglist:
1035 	kfree(pglist);
1036 out_free_buf:
1037 	kfree(buf);
1038 	return NULL;
1039 }
1040 
1041 static void arm_spe_pmu_free_aux(void *aux)
1042 {
1043 	struct arm_spe_pmu_buf *buf = aux;
1044 
1045 	vunmap(buf->base);
1046 	kfree(buf);
1047 }
1048 
1049 /* Initialisation and teardown functions */
1050 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
1051 {
1052 	static atomic_t pmu_idx = ATOMIC_INIT(-1);
1053 
1054 	int idx;
1055 	char *name;
1056 	struct device *dev = &spe_pmu->pdev->dev;
1057 
1058 	spe_pmu->pmu = (struct pmu) {
1059 		.module = THIS_MODULE,
1060 		.parent		= &spe_pmu->pdev->dev,
1061 		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
1062 		.attr_groups	= arm_spe_pmu_attr_groups,
1063 		/*
1064 		 * We hitch a ride on the software context here, so that
1065 		 * we can support per-task profiling (which is not possible
1066 		 * with the invalid context as it doesn't get sched callbacks).
1067 		 * This requires that userspace either uses a dummy event for
1068 		 * perf_event_open, since the aux buffer is not setup until
1069 		 * a subsequent mmap, or creates the profiling event in a
1070 		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
1071 		 * once the buffer has been created.
1072 		 */
1073 		.task_ctx_nr	= perf_sw_context,
1074 		.event_init	= arm_spe_pmu_event_init,
1075 		.add		= arm_spe_pmu_add,
1076 		.del		= arm_spe_pmu_del,
1077 		.start		= arm_spe_pmu_start,
1078 		.stop		= arm_spe_pmu_stop,
1079 		.read		= arm_spe_pmu_read,
1080 		.setup_aux	= arm_spe_pmu_setup_aux,
1081 		.free_aux	= arm_spe_pmu_free_aux,
1082 	};
1083 
1084 	idx = atomic_inc_return(&pmu_idx);
1085 	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
1086 	if (!name) {
1087 		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
1088 		return -ENOMEM;
1089 	}
1090 
1091 	return perf_pmu_register(&spe_pmu->pmu, name, -1);
1092 }
1093 
1094 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
1095 {
1096 	perf_pmu_unregister(&spe_pmu->pmu);
1097 }
1098 
1099 static void __arm_spe_pmu_dev_probe(void *info)
1100 {
1101 	int fld;
1102 	u64 reg;
1103 	struct arm_spe_pmu *spe_pmu = info;
1104 	struct device *dev = &spe_pmu->pdev->dev;
1105 
1106 	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
1107 						   ID_AA64DFR0_EL1_PMSVer_SHIFT);
1108 	if (!fld) {
1109 		dev_err(dev,
1110 			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
1111 			fld, smp_processor_id());
1112 		return;
1113 	}
1114 	spe_pmu->pmsver = (u16)fld;
1115 
1116 	/* Read PMBIDR first to determine whether or not we have access */
1117 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
1118 	if (FIELD_GET(PMBIDR_EL1_P, reg)) {
1119 		dev_err(dev,
1120 			"profiling buffer owned by higher exception level\n");
1121 		return;
1122 	}
1123 
1124 	/* Minimum alignment. If it's out-of-range, then fail the probe */
1125 	fld = FIELD_GET(PMBIDR_EL1_ALIGN, reg);
1126 	spe_pmu->align = 1 << fld;
1127 	if (spe_pmu->align > SZ_2K) {
1128 		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
1129 			fld, smp_processor_id());
1130 		return;
1131 	}
1132 
1133 	/* It's now safe to read PMSIDR and figure out what we've got */
1134 	reg = read_sysreg_s(SYS_PMSIDR_EL1);
1135 	if (FIELD_GET(PMSIDR_EL1_FE, reg))
1136 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
1137 
1138 	if (FIELD_GET(PMSIDR_EL1_FnE, reg))
1139 		spe_pmu->features |= SPE_PMU_FEAT_INV_FILT_EVT;
1140 
1141 	if (FIELD_GET(PMSIDR_EL1_FT, reg))
1142 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
1143 
1144 	if (FIELD_GET(PMSIDR_EL1_FL, reg))
1145 		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
1146 
1147 	if (FIELD_GET(PMSIDR_EL1_ARCHINST, reg))
1148 		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
1149 
1150 	if (FIELD_GET(PMSIDR_EL1_LDS, reg))
1151 		spe_pmu->features |= SPE_PMU_FEAT_LDS;
1152 
1153 	if (FIELD_GET(PMSIDR_EL1_ERND, reg))
1154 		spe_pmu->features |= SPE_PMU_FEAT_ERND;
1155 
1156 	if (spe_pmu->pmsver >= ID_AA64DFR0_EL1_PMSVer_V1P2)
1157 		spe_pmu->features |= SPE_PMU_FEAT_DISCARD;
1158 
1159 	if (FIELD_GET(PMSIDR_EL1_EFT, reg))
1160 		spe_pmu->features |= SPE_PMU_FEAT_EFT;
1161 
1162 	if (FIELD_GET(PMSIDR_EL1_FDS, reg))
1163 		spe_pmu->features |= SPE_PMU_FEAT_FDS;
1164 
1165 	/* This field has a spaced out encoding, so just use a look-up */
1166 	fld = FIELD_GET(PMSIDR_EL1_INTERVAL, reg);
1167 	switch (fld) {
1168 	case PMSIDR_EL1_INTERVAL_256:
1169 		spe_pmu->min_period = 256;
1170 		break;
1171 	case PMSIDR_EL1_INTERVAL_512:
1172 		spe_pmu->min_period = 512;
1173 		break;
1174 	case PMSIDR_EL1_INTERVAL_768:
1175 		spe_pmu->min_period = 768;
1176 		break;
1177 	case PMSIDR_EL1_INTERVAL_1024:
1178 		spe_pmu->min_period = 1024;
1179 		break;
1180 	case PMSIDR_EL1_INTERVAL_1536:
1181 		spe_pmu->min_period = 1536;
1182 		break;
1183 	case PMSIDR_EL1_INTERVAL_2048:
1184 		spe_pmu->min_period = 2048;
1185 		break;
1186 	case PMSIDR_EL1_INTERVAL_3072:
1187 		spe_pmu->min_period = 3072;
1188 		break;
1189 	default:
1190 		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1191 			 fld);
1192 		fallthrough;
1193 	case PMSIDR_EL1_INTERVAL_4096:
1194 		spe_pmu->min_period = 4096;
1195 	}
1196 
1197 	/* Maximum record size. If it's out-of-range, then fail the probe */
1198 	fld = FIELD_GET(PMSIDR_EL1_MAXSIZE, reg);
1199 	spe_pmu->max_record_sz = 1 << fld;
1200 	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1201 		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1202 			fld, smp_processor_id());
1203 		return;
1204 	}
1205 
1206 	fld = FIELD_GET(PMSIDR_EL1_COUNTSIZE, reg);
1207 	switch (fld) {
1208 	default:
1209 		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1210 			 fld);
1211 		fallthrough;
1212 	case PMSIDR_EL1_COUNTSIZE_12_BIT_SAT:
1213 		spe_pmu->counter_sz = 12;
1214 		break;
1215 	case PMSIDR_EL1_COUNTSIZE_16_BIT_SAT:
1216 		spe_pmu->counter_sz = 16;
1217 	}
1218 
1219 	/* Write all 1s and then read back. Unsupported filter bits are RAZ/WI. */
1220 	write_sysreg_s(U64_MAX, SYS_PMSEVFR_EL1);
1221 	spe_pmu->pmsevfr_res0 = ~read_sysreg_s(SYS_PMSEVFR_EL1);
1222 
1223 	dev_info(dev,
1224 		 "probed SPEv1.%d for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1225 		 spe_pmu->pmsver - 1, cpumask_pr_args(&spe_pmu->supported_cpus),
1226 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1227 
1228 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1229 }
1230 
1231 static void __arm_spe_pmu_reset_local(void)
1232 {
1233 	/*
1234 	 * This is probably overkill, as we have no idea where we're
1235 	 * draining any buffered data to...
1236 	 */
1237 	arm_spe_pmu_disable_and_drain_local();
1238 
1239 	/* Reset the buffer base pointer */
1240 	write_sysreg_s(0, SYS_PMBPTR_EL1);
1241 	isb();
1242 
1243 	/* Clear any pending management interrupts */
1244 	write_sysreg_s(0, SYS_PMBSR_EL1);
1245 	isb();
1246 }
1247 
1248 static void __arm_spe_pmu_setup_one(void *info)
1249 {
1250 	struct arm_spe_pmu *spe_pmu = info;
1251 
1252 	__arm_spe_pmu_reset_local();
1253 	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1254 }
1255 
1256 static void __arm_spe_pmu_stop_one(void *info)
1257 {
1258 	struct arm_spe_pmu *spe_pmu = info;
1259 
1260 	disable_percpu_irq(spe_pmu->irq);
1261 	__arm_spe_pmu_reset_local();
1262 }
1263 
1264 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1265 {
1266 	struct arm_spe_pmu *spe_pmu;
1267 
1268 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1269 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1270 		return 0;
1271 
1272 	__arm_spe_pmu_setup_one(spe_pmu);
1273 	return 0;
1274 }
1275 
1276 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1277 {
1278 	struct arm_spe_pmu *spe_pmu;
1279 
1280 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1281 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1282 		return 0;
1283 
1284 	__arm_spe_pmu_stop_one(spe_pmu);
1285 	return 0;
1286 }
1287 
1288 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1289 {
1290 	int ret;
1291 	cpumask_t *mask = &spe_pmu->supported_cpus;
1292 
1293 	/* Make sure we probe the hardware on a relevant CPU */
1294 	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1295 	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1296 		return -ENXIO;
1297 
1298 	/* Request our PPIs (note that the IRQ is still disabled) */
1299 	ret = request_percpu_irq_affinity(spe_pmu->irq, arm_spe_pmu_irq_handler,
1300 					  DRVNAME, mask, spe_pmu->handle);
1301 	if (ret)
1302 		return ret;
1303 
1304 	/*
1305 	 * Register our hotplug notifier now so we don't miss any events.
1306 	 * This will enable the IRQ for any supported CPUs that are already
1307 	 * up.
1308 	 */
1309 	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1310 				       &spe_pmu->hotplug_node);
1311 	if (ret)
1312 		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1313 
1314 	return ret;
1315 }
1316 
1317 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1318 {
1319 	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1320 	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1321 }
1322 
1323 /* Driver and device probing */
1324 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1325 {
1326 	struct platform_device *pdev = spe_pmu->pdev;
1327 	const struct cpumask *affinity;
1328 	int irq;
1329 
1330 	irq = platform_get_irq_affinity(pdev, 0, &affinity);
1331 	if (irq < 0)
1332 		return -ENXIO;
1333 
1334 	if (!irq_is_percpu(irq)) {
1335 		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1336 		return -EINVAL;
1337 	}
1338 
1339 	cpumask_copy(&spe_pmu->supported_cpus, affinity);
1340 
1341 	spe_pmu->irq = irq;
1342 	return 0;
1343 }
1344 
1345 static const struct of_device_id arm_spe_pmu_of_match[] = {
1346 	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1347 	{ /* Sentinel */ },
1348 };
1349 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1350 
1351 static const struct platform_device_id arm_spe_match[] = {
1352 	{ ARMV8_SPE_PDEV_NAME, 0},
1353 	{ }
1354 };
1355 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1356 
1357 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1358 {
1359 	int ret;
1360 	struct arm_spe_pmu *spe_pmu;
1361 	struct device *dev = &pdev->dev;
1362 
1363 	/*
1364 	 * If kernelspace is unmapped when running at EL0, then the SPE
1365 	 * buffer will fault and prematurely terminate the AUX session.
1366 	 */
1367 	if (arm64_kernel_unmapped_at_el0()) {
1368 		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1369 		return -EPERM;
1370 	}
1371 
1372 	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1373 	if (!spe_pmu)
1374 		return -ENOMEM;
1375 
1376 	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1377 	if (!spe_pmu->handle)
1378 		return -ENOMEM;
1379 
1380 	spe_pmu->pdev = pdev;
1381 	platform_set_drvdata(pdev, spe_pmu);
1382 
1383 	ret = arm_spe_pmu_irq_probe(spe_pmu);
1384 	if (ret)
1385 		goto out_free_handle;
1386 
1387 	ret = arm_spe_pmu_dev_init(spe_pmu);
1388 	if (ret)
1389 		goto out_free_handle;
1390 
1391 	ret = arm_spe_pmu_perf_init(spe_pmu);
1392 	if (ret)
1393 		goto out_teardown_dev;
1394 
1395 	return 0;
1396 
1397 out_teardown_dev:
1398 	arm_spe_pmu_dev_teardown(spe_pmu);
1399 out_free_handle:
1400 	free_percpu(spe_pmu->handle);
1401 	return ret;
1402 }
1403 
1404 static void arm_spe_pmu_device_remove(struct platform_device *pdev)
1405 {
1406 	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1407 
1408 	arm_spe_pmu_perf_destroy(spe_pmu);
1409 	arm_spe_pmu_dev_teardown(spe_pmu);
1410 	free_percpu(spe_pmu->handle);
1411 }
1412 
1413 static struct platform_driver arm_spe_pmu_driver = {
1414 	.id_table = arm_spe_match,
1415 	.driver	= {
1416 		.name		= DRVNAME,
1417 		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1418 		.suppress_bind_attrs = true,
1419 	},
1420 	.probe	= arm_spe_pmu_device_probe,
1421 	.remove = arm_spe_pmu_device_remove,
1422 };
1423 
1424 static int __init arm_spe_pmu_init(void)
1425 {
1426 	int ret;
1427 
1428 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1429 				      arm_spe_pmu_cpu_startup,
1430 				      arm_spe_pmu_cpu_teardown);
1431 	if (ret < 0)
1432 		return ret;
1433 	arm_spe_pmu_online = ret;
1434 
1435 	ret = platform_driver_register(&arm_spe_pmu_driver);
1436 	if (ret)
1437 		cpuhp_remove_multi_state(arm_spe_pmu_online);
1438 
1439 	return ret;
1440 }
1441 
1442 static void __exit arm_spe_pmu_exit(void)
1443 {
1444 	platform_driver_unregister(&arm_spe_pmu_driver);
1445 	cpuhp_remove_multi_state(arm_spe_pmu_online);
1446 }
1447 
1448 module_init(arm_spe_pmu_init);
1449 module_exit(arm_spe_pmu_exit);
1450 
1451 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1452 MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1453 MODULE_LICENSE("GPL v2");
1454