xref: /linux/drivers/perf/arm_spe_pmu.c (revision 906fd46a65383cd639e5eec72a047efc33045d86)
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, perfmon_capable(),
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) && perfmon_capable())
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_DEV_PROBED			(1UL << 63)
89 	u64					features;
90 
91 	u16					max_record_sz;
92 	u16					align;
93 	struct perf_output_handle __percpu	*handle;
94 };
95 
96 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
97 
98 /* Convert a free-running index from perf into an SPE buffer offset */
99 #define PERF_IDX2OFF(idx, buf)	((idx) % ((buf)->nr_pages << PAGE_SHIFT))
100 
101 /* Keep track of our dynamic hotplug state */
102 static enum cpuhp_state arm_spe_pmu_online;
103 
104 enum arm_spe_pmu_buf_fault_action {
105 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
106 	SPE_PMU_BUF_FAULT_ACT_FATAL,
107 	SPE_PMU_BUF_FAULT_ACT_OK,
108 };
109 
110 /* This sysfs gunk was really good fun to write. */
111 enum arm_spe_pmu_capabilities {
112 	SPE_PMU_CAP_ARCH_INST = 0,
113 	SPE_PMU_CAP_ERND,
114 	SPE_PMU_CAP_FEAT_MAX,
115 	SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
116 	SPE_PMU_CAP_MIN_IVAL,
117 };
118 
119 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
120 	[SPE_PMU_CAP_ARCH_INST]	= SPE_PMU_FEAT_ARCH_INST,
121 	[SPE_PMU_CAP_ERND]	= SPE_PMU_FEAT_ERND,
122 };
123 
124 static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
125 {
126 	if (cap < SPE_PMU_CAP_FEAT_MAX)
127 		return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
128 
129 	switch (cap) {
130 	case SPE_PMU_CAP_CNT_SZ:
131 		return spe_pmu->counter_sz;
132 	case SPE_PMU_CAP_MIN_IVAL:
133 		return spe_pmu->min_period;
134 	default:
135 		WARN(1, "unknown cap %d\n", cap);
136 	}
137 
138 	return 0;
139 }
140 
141 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
142 				    struct device_attribute *attr,
143 				    char *buf)
144 {
145 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
146 	struct dev_ext_attribute *ea =
147 		container_of(attr, struct dev_ext_attribute, attr);
148 	int cap = (long)ea->var;
149 
150 	return sysfs_emit(buf, "%u\n", arm_spe_pmu_cap_get(spe_pmu, cap));
151 }
152 
153 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
154 	&((struct dev_ext_attribute[]) {				\
155 		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
156 	})[0].attr.attr
157 
158 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
159 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
160 
161 static struct attribute *arm_spe_pmu_cap_attr[] = {
162 	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
163 	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
164 	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
165 	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
166 	NULL,
167 };
168 
169 static const struct attribute_group arm_spe_pmu_cap_group = {
170 	.name	= "caps",
171 	.attrs	= arm_spe_pmu_cap_attr,
172 };
173 
174 /* User ABI */
175 #define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
176 #define ATTR_CFG_FLD_ts_enable_LO		0
177 #define ATTR_CFG_FLD_ts_enable_HI		0
178 #define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
179 #define ATTR_CFG_FLD_pa_enable_LO		1
180 #define ATTR_CFG_FLD_pa_enable_HI		1
181 #define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
182 #define ATTR_CFG_FLD_pct_enable_LO		2
183 #define ATTR_CFG_FLD_pct_enable_HI		2
184 #define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
185 #define ATTR_CFG_FLD_jitter_LO			16
186 #define ATTR_CFG_FLD_jitter_HI			16
187 #define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
188 #define ATTR_CFG_FLD_branch_filter_LO		32
189 #define ATTR_CFG_FLD_branch_filter_HI		32
190 #define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
191 #define ATTR_CFG_FLD_load_filter_LO		33
192 #define ATTR_CFG_FLD_load_filter_HI		33
193 #define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
194 #define ATTR_CFG_FLD_store_filter_LO		34
195 #define ATTR_CFG_FLD_store_filter_HI		34
196 
197 #define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
198 #define ATTR_CFG_FLD_event_filter_LO		0
199 #define ATTR_CFG_FLD_event_filter_HI		63
200 
201 #define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
202 #define ATTR_CFG_FLD_min_latency_LO		0
203 #define ATTR_CFG_FLD_min_latency_HI		11
204 
205 #define ATTR_CFG_FLD_inv_event_filter_CFG	config3	/* PMSNEVFR_EL1 */
206 #define ATTR_CFG_FLD_inv_event_filter_LO	0
207 #define ATTR_CFG_FLD_inv_event_filter_HI	63
208 
209 GEN_PMU_FORMAT_ATTR(ts_enable);
210 GEN_PMU_FORMAT_ATTR(pa_enable);
211 GEN_PMU_FORMAT_ATTR(pct_enable);
212 GEN_PMU_FORMAT_ATTR(jitter);
213 GEN_PMU_FORMAT_ATTR(branch_filter);
214 GEN_PMU_FORMAT_ATTR(load_filter);
215 GEN_PMU_FORMAT_ATTR(store_filter);
216 GEN_PMU_FORMAT_ATTR(event_filter);
217 GEN_PMU_FORMAT_ATTR(inv_event_filter);
218 GEN_PMU_FORMAT_ATTR(min_latency);
219 
220 static struct attribute *arm_spe_pmu_formats_attr[] = {
221 	&format_attr_ts_enable.attr,
222 	&format_attr_pa_enable.attr,
223 	&format_attr_pct_enable.attr,
224 	&format_attr_jitter.attr,
225 	&format_attr_branch_filter.attr,
226 	&format_attr_load_filter.attr,
227 	&format_attr_store_filter.attr,
228 	&format_attr_event_filter.attr,
229 	&format_attr_inv_event_filter.attr,
230 	&format_attr_min_latency.attr,
231 	NULL,
232 };
233 
234 static umode_t arm_spe_pmu_format_attr_is_visible(struct kobject *kobj,
235 						  struct attribute *attr,
236 						  int unused)
237 	{
238 	struct device *dev = kobj_to_dev(kobj);
239 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
240 
241 	if (attr == &format_attr_inv_event_filter.attr && !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
242 		return 0;
243 
244 	return attr->mode;
245 }
246 
247 static const struct attribute_group arm_spe_pmu_format_group = {
248 	.name	= "format",
249 	.is_visible = arm_spe_pmu_format_attr_is_visible,
250 	.attrs	= arm_spe_pmu_formats_attr,
251 };
252 
253 static ssize_t cpumask_show(struct device *dev,
254 			    struct device_attribute *attr, char *buf)
255 {
256 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
257 
258 	return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
259 }
260 static DEVICE_ATTR_RO(cpumask);
261 
262 static struct attribute *arm_spe_pmu_attrs[] = {
263 	&dev_attr_cpumask.attr,
264 	NULL,
265 };
266 
267 static const struct attribute_group arm_spe_pmu_group = {
268 	.attrs	= arm_spe_pmu_attrs,
269 };
270 
271 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
272 	&arm_spe_pmu_group,
273 	&arm_spe_pmu_cap_group,
274 	&arm_spe_pmu_format_group,
275 	NULL,
276 };
277 
278 /* Convert between user ABI and register values */
279 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
280 {
281 	struct perf_event_attr *attr = &event->attr;
282 	u64 reg = 0;
283 
284 	reg |= FIELD_PREP(PMSCR_EL1_TS, ATTR_CFG_GET_FLD(attr, ts_enable));
285 	reg |= FIELD_PREP(PMSCR_EL1_PA, ATTR_CFG_GET_FLD(attr, pa_enable));
286 	reg |= FIELD_PREP(PMSCR_EL1_PCT, ATTR_CFG_GET_FLD(attr, pct_enable));
287 
288 	if (!attr->exclude_user)
289 		reg |= PMSCR_EL1_E0SPE;
290 
291 	if (!attr->exclude_kernel)
292 		reg |= PMSCR_EL1_E1SPE;
293 
294 	if (get_spe_event_has_cx(event))
295 		reg |= PMSCR_EL1_CX;
296 
297 	return reg;
298 }
299 
300 static void arm_spe_event_sanitise_period(struct perf_event *event)
301 {
302 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
303 	u64 period = event->hw.sample_period;
304 	u64 max_period = PMSIRR_EL1_INTERVAL_MASK;
305 
306 	if (period < spe_pmu->min_period)
307 		period = spe_pmu->min_period;
308 	else if (period > max_period)
309 		period = max_period;
310 	else
311 		period &= max_period;
312 
313 	event->hw.sample_period = period;
314 }
315 
316 static u64 arm_spe_event_to_pmsirr(struct perf_event *event)
317 {
318 	struct perf_event_attr *attr = &event->attr;
319 	u64 reg = 0;
320 
321 	arm_spe_event_sanitise_period(event);
322 
323 	reg |= FIELD_PREP(PMSIRR_EL1_RND, ATTR_CFG_GET_FLD(attr, jitter));
324 	reg |= event->hw.sample_period;
325 
326 	return reg;
327 }
328 
329 static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
330 {
331 	struct perf_event_attr *attr = &event->attr;
332 	u64 reg = 0;
333 
334 	reg |= FIELD_PREP(PMSFCR_EL1_LD, ATTR_CFG_GET_FLD(attr, load_filter));
335 	reg |= FIELD_PREP(PMSFCR_EL1_ST, ATTR_CFG_GET_FLD(attr, store_filter));
336 	reg |= FIELD_PREP(PMSFCR_EL1_B, ATTR_CFG_GET_FLD(attr, branch_filter));
337 
338 	if (reg)
339 		reg |= PMSFCR_EL1_FT;
340 
341 	if (ATTR_CFG_GET_FLD(attr, event_filter))
342 		reg |= PMSFCR_EL1_FE;
343 
344 	if (ATTR_CFG_GET_FLD(attr, inv_event_filter))
345 		reg |= PMSFCR_EL1_FnE;
346 
347 	if (ATTR_CFG_GET_FLD(attr, min_latency))
348 		reg |= PMSFCR_EL1_FL;
349 
350 	return reg;
351 }
352 
353 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
354 {
355 	struct perf_event_attr *attr = &event->attr;
356 	return ATTR_CFG_GET_FLD(attr, event_filter);
357 }
358 
359 static u64 arm_spe_event_to_pmsnevfr(struct perf_event *event)
360 {
361 	struct perf_event_attr *attr = &event->attr;
362 	return ATTR_CFG_GET_FLD(attr, inv_event_filter);
363 }
364 
365 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
366 {
367 	struct perf_event_attr *attr = &event->attr;
368 	return FIELD_PREP(PMSLATFR_EL1_MINLAT, ATTR_CFG_GET_FLD(attr, min_latency));
369 }
370 
371 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
372 {
373 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
374 	u64 head = PERF_IDX2OFF(handle->head, buf);
375 
376 	memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
377 	if (!buf->snapshot)
378 		perf_aux_output_skip(handle, len);
379 }
380 
381 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
382 {
383 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
384 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
385 	u64 head = PERF_IDX2OFF(handle->head, buf);
386 	u64 limit = buf->nr_pages * PAGE_SIZE;
387 
388 	/*
389 	 * The trace format isn't parseable in reverse, so clamp
390 	 * the limit to half of the buffer size in snapshot mode
391 	 * so that the worst case is half a buffer of records, as
392 	 * opposed to a single record.
393 	 */
394 	if (head < limit >> 1)
395 		limit >>= 1;
396 
397 	/*
398 	 * If we're within max_record_sz of the limit, we must
399 	 * pad, move the head index and recompute the limit.
400 	 */
401 	if (limit - head < spe_pmu->max_record_sz) {
402 		arm_spe_pmu_pad_buf(handle, limit - head);
403 		handle->head = PERF_IDX2OFF(limit, buf);
404 		limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
405 	}
406 
407 	return limit;
408 }
409 
410 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
411 {
412 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
413 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
414 	const u64 bufsize = buf->nr_pages * PAGE_SIZE;
415 	u64 limit = bufsize;
416 	u64 head, tail, wakeup;
417 
418 	/*
419 	 * The head can be misaligned for two reasons:
420 	 *
421 	 * 1. The hardware left PMBPTR pointing to the first byte after
422 	 *    a record when generating a buffer management event.
423 	 *
424 	 * 2. We used perf_aux_output_skip to consume handle->size bytes
425 	 *    and CIRC_SPACE was used to compute the size, which always
426 	 *    leaves one entry free.
427 	 *
428 	 * Deal with this by padding to the next alignment boundary and
429 	 * moving the head index. If we run out of buffer space, we'll
430 	 * reduce handle->size to zero and end up reporting truncation.
431 	 */
432 	head = PERF_IDX2OFF(handle->head, buf);
433 	if (!IS_ALIGNED(head, spe_pmu->align)) {
434 		unsigned long delta = roundup(head, spe_pmu->align) - head;
435 
436 		delta = min(delta, handle->size);
437 		arm_spe_pmu_pad_buf(handle, delta);
438 		head = PERF_IDX2OFF(handle->head, buf);
439 	}
440 
441 	/* If we've run out of free space, then nothing more to do */
442 	if (!handle->size)
443 		goto no_space;
444 
445 	/* Compute the tail and wakeup indices now that we've aligned head */
446 	tail = PERF_IDX2OFF(handle->head + handle->size, buf);
447 	wakeup = PERF_IDX2OFF(handle->wakeup, buf);
448 
449 	/*
450 	 * Avoid clobbering unconsumed data. We know we have space, so
451 	 * if we see head == tail we know that the buffer is empty. If
452 	 * head > tail, then there's nothing to clobber prior to
453 	 * wrapping.
454 	 */
455 	if (head < tail)
456 		limit = round_down(tail, PAGE_SIZE);
457 
458 	/*
459 	 * Wakeup may be arbitrarily far into the future. If it's not in
460 	 * the current generation, either we'll wrap before hitting it,
461 	 * or it's in the past and has been handled already.
462 	 *
463 	 * If there's a wakeup before we wrap, arrange to be woken up by
464 	 * the page boundary following it. Keep the tail boundary if
465 	 * that's lower.
466 	 */
467 	if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
468 		limit = min(limit, round_up(wakeup, PAGE_SIZE));
469 
470 	if (limit > head)
471 		return limit;
472 
473 	arm_spe_pmu_pad_buf(handle, handle->size);
474 no_space:
475 	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
476 	perf_aux_output_end(handle, 0);
477 	return 0;
478 }
479 
480 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
481 {
482 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
483 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
484 	u64 limit = __arm_spe_pmu_next_off(handle);
485 	u64 head = PERF_IDX2OFF(handle->head, buf);
486 
487 	/*
488 	 * If the head has come too close to the end of the buffer,
489 	 * then pad to the end and recompute the limit.
490 	 */
491 	if (limit && (limit - head < spe_pmu->max_record_sz)) {
492 		arm_spe_pmu_pad_buf(handle, limit - head);
493 		limit = __arm_spe_pmu_next_off(handle);
494 	}
495 
496 	return limit;
497 }
498 
499 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
500 					  struct perf_event *event)
501 {
502 	u64 base, limit;
503 	struct arm_spe_pmu_buf *buf;
504 
505 	/* Start a new aux session */
506 	buf = perf_aux_output_begin(handle, event);
507 	if (!buf) {
508 		event->hw.state |= PERF_HES_STOPPED;
509 		/*
510 		 * We still need to clear the limit pointer, since the
511 		 * profiler might only be disabled by virtue of a fault.
512 		 */
513 		limit = 0;
514 		goto out_write_limit;
515 	}
516 
517 	limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
518 			      : arm_spe_pmu_next_off(handle);
519 	if (limit)
520 		limit |= PMBLIMITR_EL1_E;
521 
522 	limit += (u64)buf->base;
523 	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
524 	write_sysreg_s(base, SYS_PMBPTR_EL1);
525 
526 out_write_limit:
527 	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
528 }
529 
530 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
531 {
532 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
533 	u64 offset, size;
534 
535 	offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
536 	size = offset - PERF_IDX2OFF(handle->head, buf);
537 
538 	if (buf->snapshot)
539 		handle->head = offset;
540 
541 	perf_aux_output_end(handle, size);
542 }
543 
544 static void arm_spe_pmu_disable_and_drain_local(void)
545 {
546 	/* Disable profiling at EL0 and EL1 */
547 	write_sysreg_s(0, SYS_PMSCR_EL1);
548 	isb();
549 
550 	/* Drain any buffered data */
551 	psb_csync();
552 	dsb(nsh);
553 
554 	/* Disable the profiling buffer */
555 	write_sysreg_s(0, SYS_PMBLIMITR_EL1);
556 	isb();
557 }
558 
559 /* IRQ handling */
560 static enum arm_spe_pmu_buf_fault_action
561 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
562 {
563 	const char *err_str;
564 	u64 pmbsr;
565 	enum arm_spe_pmu_buf_fault_action ret;
566 
567 	/*
568 	 * Ensure new profiling data is visible to the CPU and any external
569 	 * aborts have been resolved.
570 	 */
571 	psb_csync();
572 	dsb(nsh);
573 
574 	/* Ensure hardware updates to PMBPTR_EL1 are visible */
575 	isb();
576 
577 	/* Service required? */
578 	pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
579 	if (!FIELD_GET(PMBSR_EL1_S, pmbsr))
580 		return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
581 
582 	/*
583 	 * If we've lost data, disable profiling and also set the PARTIAL
584 	 * flag to indicate that the last record is corrupted.
585 	 */
586 	if (FIELD_GET(PMBSR_EL1_DL, pmbsr))
587 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
588 					     PERF_AUX_FLAG_PARTIAL);
589 
590 	/* Report collisions to userspace so that it can up the period */
591 	if (FIELD_GET(PMBSR_EL1_COLL, pmbsr))
592 		perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
593 
594 	/* We only expect buffer management events */
595 	switch (FIELD_GET(PMBSR_EL1_EC, pmbsr)) {
596 	case PMBSR_EL1_EC_BUF:
597 		/* Handled below */
598 		break;
599 	case PMBSR_EL1_EC_FAULT_S1:
600 	case PMBSR_EL1_EC_FAULT_S2:
601 		err_str = "Unexpected buffer fault";
602 		goto out_err;
603 	default:
604 		err_str = "Unknown error code";
605 		goto out_err;
606 	}
607 
608 	/* Buffer management event */
609 	switch (FIELD_GET(PMBSR_EL1_BUF_BSC_MASK, pmbsr)) {
610 	case PMBSR_EL1_BUF_BSC_FULL:
611 		ret = SPE_PMU_BUF_FAULT_ACT_OK;
612 		goto out_stop;
613 	default:
614 		err_str = "Unknown buffer status code";
615 	}
616 
617 out_err:
618 	pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
619 			   err_str, smp_processor_id(), pmbsr,
620 			   read_sysreg_s(SYS_PMBPTR_EL1),
621 			   read_sysreg_s(SYS_PMBLIMITR_EL1));
622 	ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
623 
624 out_stop:
625 	arm_spe_perf_aux_output_end(handle);
626 	return ret;
627 }
628 
629 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
630 {
631 	struct perf_output_handle *handle = dev;
632 	struct perf_event *event = handle->event;
633 	enum arm_spe_pmu_buf_fault_action act;
634 
635 	if (!perf_get_aux(handle))
636 		return IRQ_NONE;
637 
638 	act = arm_spe_pmu_buf_get_fault_act(handle);
639 	if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
640 		return IRQ_NONE;
641 
642 	/*
643 	 * Ensure perf callbacks have completed, which may disable the
644 	 * profiling buffer in response to a TRUNCATION flag.
645 	 */
646 	irq_work_run();
647 
648 	switch (act) {
649 	case SPE_PMU_BUF_FAULT_ACT_FATAL:
650 		/*
651 		 * If a fatal exception occurred then leaving the profiling
652 		 * buffer enabled is a recipe waiting to happen. Since
653 		 * fatal faults don't always imply truncation, make sure
654 		 * that the profiling buffer is disabled explicitly before
655 		 * clearing the syndrome register.
656 		 */
657 		arm_spe_pmu_disable_and_drain_local();
658 		break;
659 	case SPE_PMU_BUF_FAULT_ACT_OK:
660 		/*
661 		 * We handled the fault (the buffer was full), so resume
662 		 * profiling as long as we didn't detect truncation.
663 		 * PMBPTR might be misaligned, but we'll burn that bridge
664 		 * when we get to it.
665 		 */
666 		if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
667 			arm_spe_perf_aux_output_begin(handle, event);
668 			isb();
669 		}
670 		break;
671 	case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
672 		/* We've seen you before, but GCC has the memory of a sieve. */
673 		break;
674 	}
675 
676 	/* The buffer pointers are now sane, so resume profiling. */
677 	write_sysreg_s(0, SYS_PMBSR_EL1);
678 	return IRQ_HANDLED;
679 }
680 
681 static u64 arm_spe_pmsevfr_res0(u16 pmsver)
682 {
683 	switch (pmsver) {
684 	case ID_AA64DFR0_EL1_PMSVer_IMP:
685 		return PMSEVFR_EL1_RES0_IMP;
686 	case ID_AA64DFR0_EL1_PMSVer_V1P1:
687 		return PMSEVFR_EL1_RES0_V1P1;
688 	case ID_AA64DFR0_EL1_PMSVer_V1P2:
689 	/* Return the highest version we support in default */
690 	default:
691 		return PMSEVFR_EL1_RES0_V1P2;
692 	}
693 }
694 
695 /* Perf callbacks */
696 static int arm_spe_pmu_event_init(struct perf_event *event)
697 {
698 	u64 reg;
699 	struct perf_event_attr *attr = &event->attr;
700 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
701 
702 	/* This is, of course, deeply driver-specific */
703 	if (attr->type != event->pmu->type)
704 		return -ENOENT;
705 
706 	if (event->cpu >= 0 &&
707 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
708 		return -ENOENT;
709 
710 	if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
711 		return -EOPNOTSUPP;
712 
713 	if (arm_spe_event_to_pmsnevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
714 		return -EOPNOTSUPP;
715 
716 	if (attr->exclude_idle)
717 		return -EOPNOTSUPP;
718 
719 	/*
720 	 * Feedback-directed frequency throttling doesn't work when we
721 	 * have a buffer of samples. We'd need to manually count the
722 	 * samples in the buffer when it fills up and adjust the event
723 	 * count to reflect that. Instead, just force the user to specify
724 	 * a sample period.
725 	 */
726 	if (attr->freq)
727 		return -EINVAL;
728 
729 	reg = arm_spe_event_to_pmsfcr(event);
730 	if ((FIELD_GET(PMSFCR_EL1_FE, reg)) &&
731 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
732 		return -EOPNOTSUPP;
733 
734 	if ((FIELD_GET(PMSFCR_EL1_FnE, reg)) &&
735 	    !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
736 		return -EOPNOTSUPP;
737 
738 	if ((FIELD_GET(PMSFCR_EL1_FT, reg)) &&
739 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
740 		return -EOPNOTSUPP;
741 
742 	if ((FIELD_GET(PMSFCR_EL1_FL, reg)) &&
743 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
744 		return -EOPNOTSUPP;
745 
746 	set_spe_event_has_cx(event);
747 	reg = arm_spe_event_to_pmscr(event);
748 	if (!perfmon_capable() &&
749 	    (reg & (PMSCR_EL1_PA | PMSCR_EL1_PCT)))
750 		return -EACCES;
751 
752 	return 0;
753 }
754 
755 static void arm_spe_pmu_start(struct perf_event *event, int flags)
756 {
757 	u64 reg;
758 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
759 	struct hw_perf_event *hwc = &event->hw;
760 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
761 
762 	hwc->state = 0;
763 	arm_spe_perf_aux_output_begin(handle, event);
764 	if (hwc->state)
765 		return;
766 
767 	reg = arm_spe_event_to_pmsfcr(event);
768 	write_sysreg_s(reg, SYS_PMSFCR_EL1);
769 
770 	reg = arm_spe_event_to_pmsevfr(event);
771 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
772 
773 	if (spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT) {
774 		reg = arm_spe_event_to_pmsnevfr(event);
775 		write_sysreg_s(reg, SYS_PMSNEVFR_EL1);
776 	}
777 
778 	reg = arm_spe_event_to_pmslatfr(event);
779 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
780 
781 	if (flags & PERF_EF_RELOAD) {
782 		reg = arm_spe_event_to_pmsirr(event);
783 		write_sysreg_s(reg, SYS_PMSIRR_EL1);
784 		isb();
785 		reg = local64_read(&hwc->period_left);
786 		write_sysreg_s(reg, SYS_PMSICR_EL1);
787 	}
788 
789 	reg = arm_spe_event_to_pmscr(event);
790 	isb();
791 	write_sysreg_s(reg, SYS_PMSCR_EL1);
792 }
793 
794 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
795 {
796 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
797 	struct hw_perf_event *hwc = &event->hw;
798 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
799 
800 	/* If we're already stopped, then nothing to do */
801 	if (hwc->state & PERF_HES_STOPPED)
802 		return;
803 
804 	/* Stop all trace generation */
805 	arm_spe_pmu_disable_and_drain_local();
806 
807 	if (flags & PERF_EF_UPDATE) {
808 		/*
809 		 * If there's a fault pending then ensure we contain it
810 		 * to this buffer, since we might be on the context-switch
811 		 * path.
812 		 */
813 		if (perf_get_aux(handle)) {
814 			enum arm_spe_pmu_buf_fault_action act;
815 
816 			act = arm_spe_pmu_buf_get_fault_act(handle);
817 			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
818 				arm_spe_perf_aux_output_end(handle);
819 			else
820 				write_sysreg_s(0, SYS_PMBSR_EL1);
821 		}
822 
823 		/*
824 		 * This may also contain ECOUNT, but nobody else should
825 		 * be looking at period_left, since we forbid frequency
826 		 * based sampling.
827 		 */
828 		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
829 		hwc->state |= PERF_HES_UPTODATE;
830 	}
831 
832 	hwc->state |= PERF_HES_STOPPED;
833 }
834 
835 static int arm_spe_pmu_add(struct perf_event *event, int flags)
836 {
837 	int ret = 0;
838 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
839 	struct hw_perf_event *hwc = &event->hw;
840 	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
841 
842 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
843 		return -ENOENT;
844 
845 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
846 
847 	if (flags & PERF_EF_START) {
848 		arm_spe_pmu_start(event, PERF_EF_RELOAD);
849 		if (hwc->state & PERF_HES_STOPPED)
850 			ret = -EINVAL;
851 	}
852 
853 	return ret;
854 }
855 
856 static void arm_spe_pmu_del(struct perf_event *event, int flags)
857 {
858 	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
859 }
860 
861 static void arm_spe_pmu_read(struct perf_event *event)
862 {
863 }
864 
865 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
866 				   int nr_pages, bool snapshot)
867 {
868 	int i, cpu = event->cpu;
869 	struct page **pglist;
870 	struct arm_spe_pmu_buf *buf;
871 
872 	/* We need at least two pages for this to work. */
873 	if (nr_pages < 2)
874 		return NULL;
875 
876 	/*
877 	 * We require an even number of pages for snapshot mode, so that
878 	 * we can effectively treat the buffer as consisting of two equal
879 	 * parts and give userspace a fighting chance of getting some
880 	 * useful data out of it.
881 	 */
882 	if (snapshot && (nr_pages & 1))
883 		return NULL;
884 
885 	if (cpu == -1)
886 		cpu = raw_smp_processor_id();
887 
888 	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
889 	if (!buf)
890 		return NULL;
891 
892 	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
893 	if (!pglist)
894 		goto out_free_buf;
895 
896 	for (i = 0; i < nr_pages; ++i)
897 		pglist[i] = virt_to_page(pages[i]);
898 
899 	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
900 	if (!buf->base)
901 		goto out_free_pglist;
902 
903 	buf->nr_pages	= nr_pages;
904 	buf->snapshot	= snapshot;
905 
906 	kfree(pglist);
907 	return buf;
908 
909 out_free_pglist:
910 	kfree(pglist);
911 out_free_buf:
912 	kfree(buf);
913 	return NULL;
914 }
915 
916 static void arm_spe_pmu_free_aux(void *aux)
917 {
918 	struct arm_spe_pmu_buf *buf = aux;
919 
920 	vunmap(buf->base);
921 	kfree(buf);
922 }
923 
924 /* Initialisation and teardown functions */
925 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
926 {
927 	static atomic_t pmu_idx = ATOMIC_INIT(-1);
928 
929 	int idx;
930 	char *name;
931 	struct device *dev = &spe_pmu->pdev->dev;
932 
933 	spe_pmu->pmu = (struct pmu) {
934 		.module = THIS_MODULE,
935 		.parent		= &spe_pmu->pdev->dev,
936 		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
937 		.attr_groups	= arm_spe_pmu_attr_groups,
938 		/*
939 		 * We hitch a ride on the software context here, so that
940 		 * we can support per-task profiling (which is not possible
941 		 * with the invalid context as it doesn't get sched callbacks).
942 		 * This requires that userspace either uses a dummy event for
943 		 * perf_event_open, since the aux buffer is not setup until
944 		 * a subsequent mmap, or creates the profiling event in a
945 		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
946 		 * once the buffer has been created.
947 		 */
948 		.task_ctx_nr	= perf_sw_context,
949 		.event_init	= arm_spe_pmu_event_init,
950 		.add		= arm_spe_pmu_add,
951 		.del		= arm_spe_pmu_del,
952 		.start		= arm_spe_pmu_start,
953 		.stop		= arm_spe_pmu_stop,
954 		.read		= arm_spe_pmu_read,
955 		.setup_aux	= arm_spe_pmu_setup_aux,
956 		.free_aux	= arm_spe_pmu_free_aux,
957 	};
958 
959 	idx = atomic_inc_return(&pmu_idx);
960 	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
961 	if (!name) {
962 		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
963 		return -ENOMEM;
964 	}
965 
966 	return perf_pmu_register(&spe_pmu->pmu, name, -1);
967 }
968 
969 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
970 {
971 	perf_pmu_unregister(&spe_pmu->pmu);
972 }
973 
974 static void __arm_spe_pmu_dev_probe(void *info)
975 {
976 	int fld;
977 	u64 reg;
978 	struct arm_spe_pmu *spe_pmu = info;
979 	struct device *dev = &spe_pmu->pdev->dev;
980 
981 	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
982 						   ID_AA64DFR0_EL1_PMSVer_SHIFT);
983 	if (!fld) {
984 		dev_err(dev,
985 			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
986 			fld, smp_processor_id());
987 		return;
988 	}
989 	spe_pmu->pmsver = (u16)fld;
990 
991 	/* Read PMBIDR first to determine whether or not we have access */
992 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
993 	if (FIELD_GET(PMBIDR_EL1_P, reg)) {
994 		dev_err(dev,
995 			"profiling buffer owned by higher exception level\n");
996 		return;
997 	}
998 
999 	/* Minimum alignment. If it's out-of-range, then fail the probe */
1000 	fld = FIELD_GET(PMBIDR_EL1_ALIGN, reg);
1001 	spe_pmu->align = 1 << fld;
1002 	if (spe_pmu->align > SZ_2K) {
1003 		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
1004 			fld, smp_processor_id());
1005 		return;
1006 	}
1007 
1008 	/* It's now safe to read PMSIDR and figure out what we've got */
1009 	reg = read_sysreg_s(SYS_PMSIDR_EL1);
1010 	if (FIELD_GET(PMSIDR_EL1_FE, reg))
1011 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
1012 
1013 	if (FIELD_GET(PMSIDR_EL1_FnE, reg))
1014 		spe_pmu->features |= SPE_PMU_FEAT_INV_FILT_EVT;
1015 
1016 	if (FIELD_GET(PMSIDR_EL1_FT, reg))
1017 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
1018 
1019 	if (FIELD_GET(PMSIDR_EL1_FL, reg))
1020 		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
1021 
1022 	if (FIELD_GET(PMSIDR_EL1_ARCHINST, reg))
1023 		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
1024 
1025 	if (FIELD_GET(PMSIDR_EL1_LDS, reg))
1026 		spe_pmu->features |= SPE_PMU_FEAT_LDS;
1027 
1028 	if (FIELD_GET(PMSIDR_EL1_ERND, reg))
1029 		spe_pmu->features |= SPE_PMU_FEAT_ERND;
1030 
1031 	/* This field has a spaced out encoding, so just use a look-up */
1032 	fld = FIELD_GET(PMSIDR_EL1_INTERVAL, reg);
1033 	switch (fld) {
1034 	case PMSIDR_EL1_INTERVAL_256:
1035 		spe_pmu->min_period = 256;
1036 		break;
1037 	case PMSIDR_EL1_INTERVAL_512:
1038 		spe_pmu->min_period = 512;
1039 		break;
1040 	case PMSIDR_EL1_INTERVAL_768:
1041 		spe_pmu->min_period = 768;
1042 		break;
1043 	case PMSIDR_EL1_INTERVAL_1024:
1044 		spe_pmu->min_period = 1024;
1045 		break;
1046 	case PMSIDR_EL1_INTERVAL_1536:
1047 		spe_pmu->min_period = 1536;
1048 		break;
1049 	case PMSIDR_EL1_INTERVAL_2048:
1050 		spe_pmu->min_period = 2048;
1051 		break;
1052 	case PMSIDR_EL1_INTERVAL_3072:
1053 		spe_pmu->min_period = 3072;
1054 		break;
1055 	default:
1056 		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1057 			 fld);
1058 		fallthrough;
1059 	case PMSIDR_EL1_INTERVAL_4096:
1060 		spe_pmu->min_period = 4096;
1061 	}
1062 
1063 	/* Maximum record size. If it's out-of-range, then fail the probe */
1064 	fld = FIELD_GET(PMSIDR_EL1_MAXSIZE, reg);
1065 	spe_pmu->max_record_sz = 1 << fld;
1066 	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1067 		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1068 			fld, smp_processor_id());
1069 		return;
1070 	}
1071 
1072 	fld = FIELD_GET(PMSIDR_EL1_COUNTSIZE, reg);
1073 	switch (fld) {
1074 	default:
1075 		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1076 			 fld);
1077 		fallthrough;
1078 	case PMSIDR_EL1_COUNTSIZE_12_BIT_SAT:
1079 		spe_pmu->counter_sz = 12;
1080 		break;
1081 	case PMSIDR_EL1_COUNTSIZE_16_BIT_SAT:
1082 		spe_pmu->counter_sz = 16;
1083 	}
1084 
1085 	dev_info(dev,
1086 		 "probed SPEv1.%d for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1087 		 spe_pmu->pmsver - 1, cpumask_pr_args(&spe_pmu->supported_cpus),
1088 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1089 
1090 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1091 }
1092 
1093 static void __arm_spe_pmu_reset_local(void)
1094 {
1095 	/*
1096 	 * This is probably overkill, as we have no idea where we're
1097 	 * draining any buffered data to...
1098 	 */
1099 	arm_spe_pmu_disable_and_drain_local();
1100 
1101 	/* Reset the buffer base pointer */
1102 	write_sysreg_s(0, SYS_PMBPTR_EL1);
1103 	isb();
1104 
1105 	/* Clear any pending management interrupts */
1106 	write_sysreg_s(0, SYS_PMBSR_EL1);
1107 	isb();
1108 }
1109 
1110 static void __arm_spe_pmu_setup_one(void *info)
1111 {
1112 	struct arm_spe_pmu *spe_pmu = info;
1113 
1114 	__arm_spe_pmu_reset_local();
1115 	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1116 }
1117 
1118 static void __arm_spe_pmu_stop_one(void *info)
1119 {
1120 	struct arm_spe_pmu *spe_pmu = info;
1121 
1122 	disable_percpu_irq(spe_pmu->irq);
1123 	__arm_spe_pmu_reset_local();
1124 }
1125 
1126 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1127 {
1128 	struct arm_spe_pmu *spe_pmu;
1129 
1130 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1131 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1132 		return 0;
1133 
1134 	__arm_spe_pmu_setup_one(spe_pmu);
1135 	return 0;
1136 }
1137 
1138 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1139 {
1140 	struct arm_spe_pmu *spe_pmu;
1141 
1142 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1143 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1144 		return 0;
1145 
1146 	__arm_spe_pmu_stop_one(spe_pmu);
1147 	return 0;
1148 }
1149 
1150 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1151 {
1152 	int ret;
1153 	cpumask_t *mask = &spe_pmu->supported_cpus;
1154 
1155 	/* Make sure we probe the hardware on a relevant CPU */
1156 	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1157 	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1158 		return -ENXIO;
1159 
1160 	/* Request our PPIs (note that the IRQ is still disabled) */
1161 	ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1162 				 spe_pmu->handle);
1163 	if (ret)
1164 		return ret;
1165 
1166 	/*
1167 	 * Register our hotplug notifier now so we don't miss any events.
1168 	 * This will enable the IRQ for any supported CPUs that are already
1169 	 * up.
1170 	 */
1171 	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1172 				       &spe_pmu->hotplug_node);
1173 	if (ret)
1174 		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1175 
1176 	return ret;
1177 }
1178 
1179 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1180 {
1181 	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1182 	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1183 }
1184 
1185 /* Driver and device probing */
1186 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1187 {
1188 	struct platform_device *pdev = spe_pmu->pdev;
1189 	int irq = platform_get_irq(pdev, 0);
1190 
1191 	if (irq < 0)
1192 		return -ENXIO;
1193 
1194 	if (!irq_is_percpu(irq)) {
1195 		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1196 		return -EINVAL;
1197 	}
1198 
1199 	if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1200 		dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1201 		return -EINVAL;
1202 	}
1203 
1204 	spe_pmu->irq = irq;
1205 	return 0;
1206 }
1207 
1208 static const struct of_device_id arm_spe_pmu_of_match[] = {
1209 	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1210 	{ /* Sentinel */ },
1211 };
1212 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1213 
1214 static const struct platform_device_id arm_spe_match[] = {
1215 	{ ARMV8_SPE_PDEV_NAME, 0},
1216 	{ }
1217 };
1218 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1219 
1220 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1221 {
1222 	int ret;
1223 	struct arm_spe_pmu *spe_pmu;
1224 	struct device *dev = &pdev->dev;
1225 
1226 	/*
1227 	 * If kernelspace is unmapped when running at EL0, then the SPE
1228 	 * buffer will fault and prematurely terminate the AUX session.
1229 	 */
1230 	if (arm64_kernel_unmapped_at_el0()) {
1231 		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1232 		return -EPERM;
1233 	}
1234 
1235 	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1236 	if (!spe_pmu)
1237 		return -ENOMEM;
1238 
1239 	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1240 	if (!spe_pmu->handle)
1241 		return -ENOMEM;
1242 
1243 	spe_pmu->pdev = pdev;
1244 	platform_set_drvdata(pdev, spe_pmu);
1245 
1246 	ret = arm_spe_pmu_irq_probe(spe_pmu);
1247 	if (ret)
1248 		goto out_free_handle;
1249 
1250 	ret = arm_spe_pmu_dev_init(spe_pmu);
1251 	if (ret)
1252 		goto out_free_handle;
1253 
1254 	ret = arm_spe_pmu_perf_init(spe_pmu);
1255 	if (ret)
1256 		goto out_teardown_dev;
1257 
1258 	return 0;
1259 
1260 out_teardown_dev:
1261 	arm_spe_pmu_dev_teardown(spe_pmu);
1262 out_free_handle:
1263 	free_percpu(spe_pmu->handle);
1264 	return ret;
1265 }
1266 
1267 static void arm_spe_pmu_device_remove(struct platform_device *pdev)
1268 {
1269 	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1270 
1271 	arm_spe_pmu_perf_destroy(spe_pmu);
1272 	arm_spe_pmu_dev_teardown(spe_pmu);
1273 	free_percpu(spe_pmu->handle);
1274 }
1275 
1276 static struct platform_driver arm_spe_pmu_driver = {
1277 	.id_table = arm_spe_match,
1278 	.driver	= {
1279 		.name		= DRVNAME,
1280 		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1281 		.suppress_bind_attrs = true,
1282 	},
1283 	.probe	= arm_spe_pmu_device_probe,
1284 	.remove_new = arm_spe_pmu_device_remove,
1285 };
1286 
1287 static int __init arm_spe_pmu_init(void)
1288 {
1289 	int ret;
1290 
1291 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1292 				      arm_spe_pmu_cpu_startup,
1293 				      arm_spe_pmu_cpu_teardown);
1294 	if (ret < 0)
1295 		return ret;
1296 	arm_spe_pmu_online = ret;
1297 
1298 	ret = platform_driver_register(&arm_spe_pmu_driver);
1299 	if (ret)
1300 		cpuhp_remove_multi_state(arm_spe_pmu_online);
1301 
1302 	return ret;
1303 }
1304 
1305 static void __exit arm_spe_pmu_exit(void)
1306 {
1307 	platform_driver_unregister(&arm_spe_pmu_driver);
1308 	cpuhp_remove_multi_state(arm_spe_pmu_online);
1309 }
1310 
1311 module_init(arm_spe_pmu_init);
1312 module_exit(arm_spe_pmu_exit);
1313 
1314 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1315 MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1316 MODULE_LICENSE("GPL v2");
1317