xref: /openbmc/linux/drivers/perf/arm_spe_pmu.c (revision 53f9cd5c)
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/bitops.h>
16 #include <linux/bug.h>
17 #include <linux/capability.h>
18 #include <linux/cpuhotplug.h>
19 #include <linux/cpumask.h>
20 #include <linux/device.h>
21 #include <linux/errno.h>
22 #include <linux/interrupt.h>
23 #include <linux/irq.h>
24 #include <linux/kernel.h>
25 #include <linux/list.h>
26 #include <linux/module.h>
27 #include <linux/of_address.h>
28 #include <linux/of_device.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			BIT(0)
48 
49 static void set_spe_event_has_cx(struct perf_event *event)
50 {
51 	if (IS_ENABLED(CONFIG_PID_IN_CONTEXTIDR) && perfmon_capable())
52 		event->hw.flags |= SPE_PMU_HW_FLAGS_CX;
53 }
54 
55 static bool get_spe_event_has_cx(struct perf_event *event)
56 {
57 	return !!(event->hw.flags & SPE_PMU_HW_FLAGS_CX);
58 }
59 
60 #define ARM_SPE_BUF_PAD_BYTE			0
61 
62 struct arm_spe_pmu_buf {
63 	int					nr_pages;
64 	bool					snapshot;
65 	void					*base;
66 };
67 
68 struct arm_spe_pmu {
69 	struct pmu				pmu;
70 	struct platform_device			*pdev;
71 	cpumask_t				supported_cpus;
72 	struct hlist_node			hotplug_node;
73 
74 	int					irq; /* PPI */
75 	u16					pmsver;
76 	u16					min_period;
77 	u16					counter_sz;
78 
79 #define SPE_PMU_FEAT_FILT_EVT			(1UL << 0)
80 #define SPE_PMU_FEAT_FILT_TYP			(1UL << 1)
81 #define SPE_PMU_FEAT_FILT_LAT			(1UL << 2)
82 #define SPE_PMU_FEAT_ARCH_INST			(1UL << 3)
83 #define SPE_PMU_FEAT_LDS			(1UL << 4)
84 #define SPE_PMU_FEAT_ERND			(1UL << 5)
85 #define SPE_PMU_FEAT_DEV_PROBED			(1UL << 63)
86 	u64					features;
87 
88 	u16					max_record_sz;
89 	u16					align;
90 	struct perf_output_handle __percpu	*handle;
91 };
92 
93 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
94 
95 /* Convert a free-running index from perf into an SPE buffer offset */
96 #define PERF_IDX2OFF(idx, buf)	((idx) % ((buf)->nr_pages << PAGE_SHIFT))
97 
98 /* Keep track of our dynamic hotplug state */
99 static enum cpuhp_state arm_spe_pmu_online;
100 
101 enum arm_spe_pmu_buf_fault_action {
102 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
103 	SPE_PMU_BUF_FAULT_ACT_FATAL,
104 	SPE_PMU_BUF_FAULT_ACT_OK,
105 };
106 
107 /* This sysfs gunk was really good fun to write. */
108 enum arm_spe_pmu_capabilities {
109 	SPE_PMU_CAP_ARCH_INST = 0,
110 	SPE_PMU_CAP_ERND,
111 	SPE_PMU_CAP_FEAT_MAX,
112 	SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
113 	SPE_PMU_CAP_MIN_IVAL,
114 };
115 
116 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
117 	[SPE_PMU_CAP_ARCH_INST]	= SPE_PMU_FEAT_ARCH_INST,
118 	[SPE_PMU_CAP_ERND]	= SPE_PMU_FEAT_ERND,
119 };
120 
121 static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
122 {
123 	if (cap < SPE_PMU_CAP_FEAT_MAX)
124 		return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
125 
126 	switch (cap) {
127 	case SPE_PMU_CAP_CNT_SZ:
128 		return spe_pmu->counter_sz;
129 	case SPE_PMU_CAP_MIN_IVAL:
130 		return spe_pmu->min_period;
131 	default:
132 		WARN(1, "unknown cap %d\n", cap);
133 	}
134 
135 	return 0;
136 }
137 
138 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
139 				    struct device_attribute *attr,
140 				    char *buf)
141 {
142 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
143 	struct dev_ext_attribute *ea =
144 		container_of(attr, struct dev_ext_attribute, attr);
145 	int cap = (long)ea->var;
146 
147 	return sysfs_emit(buf, "%u\n", arm_spe_pmu_cap_get(spe_pmu, cap));
148 }
149 
150 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
151 	&((struct dev_ext_attribute[]) {				\
152 		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
153 	})[0].attr.attr
154 
155 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
156 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
157 
158 static struct attribute *arm_spe_pmu_cap_attr[] = {
159 	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
160 	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
161 	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
162 	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
163 	NULL,
164 };
165 
166 static const struct attribute_group arm_spe_pmu_cap_group = {
167 	.name	= "caps",
168 	.attrs	= arm_spe_pmu_cap_attr,
169 };
170 
171 /* User ABI */
172 #define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
173 #define ATTR_CFG_FLD_ts_enable_LO		0
174 #define ATTR_CFG_FLD_ts_enable_HI		0
175 #define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
176 #define ATTR_CFG_FLD_pa_enable_LO		1
177 #define ATTR_CFG_FLD_pa_enable_HI		1
178 #define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
179 #define ATTR_CFG_FLD_pct_enable_LO		2
180 #define ATTR_CFG_FLD_pct_enable_HI		2
181 #define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
182 #define ATTR_CFG_FLD_jitter_LO			16
183 #define ATTR_CFG_FLD_jitter_HI			16
184 #define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
185 #define ATTR_CFG_FLD_branch_filter_LO		32
186 #define ATTR_CFG_FLD_branch_filter_HI		32
187 #define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
188 #define ATTR_CFG_FLD_load_filter_LO		33
189 #define ATTR_CFG_FLD_load_filter_HI		33
190 #define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
191 #define ATTR_CFG_FLD_store_filter_LO		34
192 #define ATTR_CFG_FLD_store_filter_HI		34
193 
194 #define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
195 #define ATTR_CFG_FLD_event_filter_LO		0
196 #define ATTR_CFG_FLD_event_filter_HI		63
197 
198 #define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
199 #define ATTR_CFG_FLD_min_latency_LO		0
200 #define ATTR_CFG_FLD_min_latency_HI		11
201 
202 /* Why does everything I do descend into this? */
203 #define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
204 	(lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
205 
206 #define _GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
207 	__GEN_PMU_FORMAT_ATTR(cfg, lo, hi)
208 
209 #define GEN_PMU_FORMAT_ATTR(name)					\
210 	PMU_FORMAT_ATTR(name,						\
211 	_GEN_PMU_FORMAT_ATTR(ATTR_CFG_FLD_##name##_CFG,			\
212 			     ATTR_CFG_FLD_##name##_LO,			\
213 			     ATTR_CFG_FLD_##name##_HI))
214 
215 #define _ATTR_CFG_GET_FLD(attr, cfg, lo, hi)				\
216 	((((attr)->cfg) >> lo) & GENMASK(hi - lo, 0))
217 
218 #define ATTR_CFG_GET_FLD(attr, name)					\
219 	_ATTR_CFG_GET_FLD(attr,						\
220 			  ATTR_CFG_FLD_##name##_CFG,			\
221 			  ATTR_CFG_FLD_##name##_LO,			\
222 			  ATTR_CFG_FLD_##name##_HI)
223 
224 GEN_PMU_FORMAT_ATTR(ts_enable);
225 GEN_PMU_FORMAT_ATTR(pa_enable);
226 GEN_PMU_FORMAT_ATTR(pct_enable);
227 GEN_PMU_FORMAT_ATTR(jitter);
228 GEN_PMU_FORMAT_ATTR(branch_filter);
229 GEN_PMU_FORMAT_ATTR(load_filter);
230 GEN_PMU_FORMAT_ATTR(store_filter);
231 GEN_PMU_FORMAT_ATTR(event_filter);
232 GEN_PMU_FORMAT_ATTR(min_latency);
233 
234 static struct attribute *arm_spe_pmu_formats_attr[] = {
235 	&format_attr_ts_enable.attr,
236 	&format_attr_pa_enable.attr,
237 	&format_attr_pct_enable.attr,
238 	&format_attr_jitter.attr,
239 	&format_attr_branch_filter.attr,
240 	&format_attr_load_filter.attr,
241 	&format_attr_store_filter.attr,
242 	&format_attr_event_filter.attr,
243 	&format_attr_min_latency.attr,
244 	NULL,
245 };
246 
247 static const struct attribute_group arm_spe_pmu_format_group = {
248 	.name	= "format",
249 	.attrs	= arm_spe_pmu_formats_attr,
250 };
251 
252 static ssize_t cpumask_show(struct device *dev,
253 			    struct device_attribute *attr, char *buf)
254 {
255 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
256 
257 	return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
258 }
259 static DEVICE_ATTR_RO(cpumask);
260 
261 static struct attribute *arm_spe_pmu_attrs[] = {
262 	&dev_attr_cpumask.attr,
263 	NULL,
264 };
265 
266 static const struct attribute_group arm_spe_pmu_group = {
267 	.attrs	= arm_spe_pmu_attrs,
268 };
269 
270 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
271 	&arm_spe_pmu_group,
272 	&arm_spe_pmu_cap_group,
273 	&arm_spe_pmu_format_group,
274 	NULL,
275 };
276 
277 /* Convert between user ABI and register values */
278 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
279 {
280 	struct perf_event_attr *attr = &event->attr;
281 	u64 reg = 0;
282 
283 	reg |= ATTR_CFG_GET_FLD(attr, ts_enable) << SYS_PMSCR_EL1_TS_SHIFT;
284 	reg |= ATTR_CFG_GET_FLD(attr, pa_enable) << SYS_PMSCR_EL1_PA_SHIFT;
285 	reg |= ATTR_CFG_GET_FLD(attr, pct_enable) << SYS_PMSCR_EL1_PCT_SHIFT;
286 
287 	if (!attr->exclude_user)
288 		reg |= BIT(SYS_PMSCR_EL1_E0SPE_SHIFT);
289 
290 	if (!attr->exclude_kernel)
291 		reg |= BIT(SYS_PMSCR_EL1_E1SPE_SHIFT);
292 
293 	if (get_spe_event_has_cx(event))
294 		reg |= BIT(SYS_PMSCR_EL1_CX_SHIFT);
295 
296 	return reg;
297 }
298 
299 static void arm_spe_event_sanitise_period(struct perf_event *event)
300 {
301 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
302 	u64 period = event->hw.sample_period;
303 	u64 max_period = SYS_PMSIRR_EL1_INTERVAL_MASK
304 			 << SYS_PMSIRR_EL1_INTERVAL_SHIFT;
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 |= ATTR_CFG_GET_FLD(attr, jitter) << SYS_PMSIRR_EL1_RND_SHIFT;
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 |= ATTR_CFG_GET_FLD(attr, load_filter) << SYS_PMSFCR_EL1_LD_SHIFT;
335 	reg |= ATTR_CFG_GET_FLD(attr, store_filter) << SYS_PMSFCR_EL1_ST_SHIFT;
336 	reg |= ATTR_CFG_GET_FLD(attr, branch_filter) << SYS_PMSFCR_EL1_B_SHIFT;
337 
338 	if (reg)
339 		reg |= BIT(SYS_PMSFCR_EL1_FT_SHIFT);
340 
341 	if (ATTR_CFG_GET_FLD(attr, event_filter))
342 		reg |= BIT(SYS_PMSFCR_EL1_FE_SHIFT);
343 
344 	if (ATTR_CFG_GET_FLD(attr, min_latency))
345 		reg |= BIT(SYS_PMSFCR_EL1_FL_SHIFT);
346 
347 	return reg;
348 }
349 
350 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
351 {
352 	struct perf_event_attr *attr = &event->attr;
353 	return ATTR_CFG_GET_FLD(attr, event_filter);
354 }
355 
356 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
357 {
358 	struct perf_event_attr *attr = &event->attr;
359 	return ATTR_CFG_GET_FLD(attr, min_latency)
360 	       << SYS_PMSLATFR_EL1_MINLAT_SHIFT;
361 }
362 
363 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
364 {
365 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
366 	u64 head = PERF_IDX2OFF(handle->head, buf);
367 
368 	memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
369 	if (!buf->snapshot)
370 		perf_aux_output_skip(handle, len);
371 }
372 
373 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
374 {
375 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
376 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
377 	u64 head = PERF_IDX2OFF(handle->head, buf);
378 	u64 limit = buf->nr_pages * PAGE_SIZE;
379 
380 	/*
381 	 * The trace format isn't parseable in reverse, so clamp
382 	 * the limit to half of the buffer size in snapshot mode
383 	 * so that the worst case is half a buffer of records, as
384 	 * opposed to a single record.
385 	 */
386 	if (head < limit >> 1)
387 		limit >>= 1;
388 
389 	/*
390 	 * If we're within max_record_sz of the limit, we must
391 	 * pad, move the head index and recompute the limit.
392 	 */
393 	if (limit - head < spe_pmu->max_record_sz) {
394 		arm_spe_pmu_pad_buf(handle, limit - head);
395 		handle->head = PERF_IDX2OFF(limit, buf);
396 		limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
397 	}
398 
399 	return limit;
400 }
401 
402 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
403 {
404 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
405 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
406 	const u64 bufsize = buf->nr_pages * PAGE_SIZE;
407 	u64 limit = bufsize;
408 	u64 head, tail, wakeup;
409 
410 	/*
411 	 * The head can be misaligned for two reasons:
412 	 *
413 	 * 1. The hardware left PMBPTR pointing to the first byte after
414 	 *    a record when generating a buffer management event.
415 	 *
416 	 * 2. We used perf_aux_output_skip to consume handle->size bytes
417 	 *    and CIRC_SPACE was used to compute the size, which always
418 	 *    leaves one entry free.
419 	 *
420 	 * Deal with this by padding to the next alignment boundary and
421 	 * moving the head index. If we run out of buffer space, we'll
422 	 * reduce handle->size to zero and end up reporting truncation.
423 	 */
424 	head = PERF_IDX2OFF(handle->head, buf);
425 	if (!IS_ALIGNED(head, spe_pmu->align)) {
426 		unsigned long delta = roundup(head, spe_pmu->align) - head;
427 
428 		delta = min(delta, handle->size);
429 		arm_spe_pmu_pad_buf(handle, delta);
430 		head = PERF_IDX2OFF(handle->head, buf);
431 	}
432 
433 	/* If we've run out of free space, then nothing more to do */
434 	if (!handle->size)
435 		goto no_space;
436 
437 	/* Compute the tail and wakeup indices now that we've aligned head */
438 	tail = PERF_IDX2OFF(handle->head + handle->size, buf);
439 	wakeup = PERF_IDX2OFF(handle->wakeup, buf);
440 
441 	/*
442 	 * Avoid clobbering unconsumed data. We know we have space, so
443 	 * if we see head == tail we know that the buffer is empty. If
444 	 * head > tail, then there's nothing to clobber prior to
445 	 * wrapping.
446 	 */
447 	if (head < tail)
448 		limit = round_down(tail, PAGE_SIZE);
449 
450 	/*
451 	 * Wakeup may be arbitrarily far into the future. If it's not in
452 	 * the current generation, either we'll wrap before hitting it,
453 	 * or it's in the past and has been handled already.
454 	 *
455 	 * If there's a wakeup before we wrap, arrange to be woken up by
456 	 * the page boundary following it. Keep the tail boundary if
457 	 * that's lower.
458 	 */
459 	if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
460 		limit = min(limit, round_up(wakeup, PAGE_SIZE));
461 
462 	if (limit > head)
463 		return limit;
464 
465 	arm_spe_pmu_pad_buf(handle, handle->size);
466 no_space:
467 	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
468 	perf_aux_output_end(handle, 0);
469 	return 0;
470 }
471 
472 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
473 {
474 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
475 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
476 	u64 limit = __arm_spe_pmu_next_off(handle);
477 	u64 head = PERF_IDX2OFF(handle->head, buf);
478 
479 	/*
480 	 * If the head has come too close to the end of the buffer,
481 	 * then pad to the end and recompute the limit.
482 	 */
483 	if (limit && (limit - head < spe_pmu->max_record_sz)) {
484 		arm_spe_pmu_pad_buf(handle, limit - head);
485 		limit = __arm_spe_pmu_next_off(handle);
486 	}
487 
488 	return limit;
489 }
490 
491 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
492 					  struct perf_event *event)
493 {
494 	u64 base, limit;
495 	struct arm_spe_pmu_buf *buf;
496 
497 	/* Start a new aux session */
498 	buf = perf_aux_output_begin(handle, event);
499 	if (!buf) {
500 		event->hw.state |= PERF_HES_STOPPED;
501 		/*
502 		 * We still need to clear the limit pointer, since the
503 		 * profiler might only be disabled by virtue of a fault.
504 		 */
505 		limit = 0;
506 		goto out_write_limit;
507 	}
508 
509 	limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
510 			      : arm_spe_pmu_next_off(handle);
511 	if (limit)
512 		limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT);
513 
514 	limit += (u64)buf->base;
515 	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
516 	write_sysreg_s(base, SYS_PMBPTR_EL1);
517 
518 out_write_limit:
519 	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
520 }
521 
522 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
523 {
524 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
525 	u64 offset, size;
526 
527 	offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
528 	size = offset - PERF_IDX2OFF(handle->head, buf);
529 
530 	if (buf->snapshot)
531 		handle->head = offset;
532 
533 	perf_aux_output_end(handle, size);
534 }
535 
536 static void arm_spe_pmu_disable_and_drain_local(void)
537 {
538 	/* Disable profiling at EL0 and EL1 */
539 	write_sysreg_s(0, SYS_PMSCR_EL1);
540 	isb();
541 
542 	/* Drain any buffered data */
543 	psb_csync();
544 	dsb(nsh);
545 
546 	/* Disable the profiling buffer */
547 	write_sysreg_s(0, SYS_PMBLIMITR_EL1);
548 	isb();
549 }
550 
551 /* IRQ handling */
552 static enum arm_spe_pmu_buf_fault_action
553 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
554 {
555 	const char *err_str;
556 	u64 pmbsr;
557 	enum arm_spe_pmu_buf_fault_action ret;
558 
559 	/*
560 	 * Ensure new profiling data is visible to the CPU and any external
561 	 * aborts have been resolved.
562 	 */
563 	psb_csync();
564 	dsb(nsh);
565 
566 	/* Ensure hardware updates to PMBPTR_EL1 are visible */
567 	isb();
568 
569 	/* Service required? */
570 	pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
571 	if (!(pmbsr & BIT(SYS_PMBSR_EL1_S_SHIFT)))
572 		return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
573 
574 	/*
575 	 * If we've lost data, disable profiling and also set the PARTIAL
576 	 * flag to indicate that the last record is corrupted.
577 	 */
578 	if (pmbsr & BIT(SYS_PMBSR_EL1_DL_SHIFT))
579 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
580 					     PERF_AUX_FLAG_PARTIAL);
581 
582 	/* Report collisions to userspace so that it can up the period */
583 	if (pmbsr & BIT(SYS_PMBSR_EL1_COLL_SHIFT))
584 		perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
585 
586 	/* We only expect buffer management events */
587 	switch (pmbsr & (SYS_PMBSR_EL1_EC_MASK << SYS_PMBSR_EL1_EC_SHIFT)) {
588 	case SYS_PMBSR_EL1_EC_BUF:
589 		/* Handled below */
590 		break;
591 	case SYS_PMBSR_EL1_EC_FAULT_S1:
592 	case SYS_PMBSR_EL1_EC_FAULT_S2:
593 		err_str = "Unexpected buffer fault";
594 		goto out_err;
595 	default:
596 		err_str = "Unknown error code";
597 		goto out_err;
598 	}
599 
600 	/* Buffer management event */
601 	switch (pmbsr &
602 		(SYS_PMBSR_EL1_BUF_BSC_MASK << SYS_PMBSR_EL1_BUF_BSC_SHIFT)) {
603 	case SYS_PMBSR_EL1_BUF_BSC_FULL:
604 		ret = SPE_PMU_BUF_FAULT_ACT_OK;
605 		goto out_stop;
606 	default:
607 		err_str = "Unknown buffer status code";
608 	}
609 
610 out_err:
611 	pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
612 			   err_str, smp_processor_id(), pmbsr,
613 			   read_sysreg_s(SYS_PMBPTR_EL1),
614 			   read_sysreg_s(SYS_PMBLIMITR_EL1));
615 	ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
616 
617 out_stop:
618 	arm_spe_perf_aux_output_end(handle);
619 	return ret;
620 }
621 
622 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
623 {
624 	struct perf_output_handle *handle = dev;
625 	struct perf_event *event = handle->event;
626 	enum arm_spe_pmu_buf_fault_action act;
627 
628 	if (!perf_get_aux(handle))
629 		return IRQ_NONE;
630 
631 	act = arm_spe_pmu_buf_get_fault_act(handle);
632 	if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
633 		return IRQ_NONE;
634 
635 	/*
636 	 * Ensure perf callbacks have completed, which may disable the
637 	 * profiling buffer in response to a TRUNCATION flag.
638 	 */
639 	irq_work_run();
640 
641 	switch (act) {
642 	case SPE_PMU_BUF_FAULT_ACT_FATAL:
643 		/*
644 		 * If a fatal exception occurred then leaving the profiling
645 		 * buffer enabled is a recipe waiting to happen. Since
646 		 * fatal faults don't always imply truncation, make sure
647 		 * that the profiling buffer is disabled explicitly before
648 		 * clearing the syndrome register.
649 		 */
650 		arm_spe_pmu_disable_and_drain_local();
651 		break;
652 	case SPE_PMU_BUF_FAULT_ACT_OK:
653 		/*
654 		 * We handled the fault (the buffer was full), so resume
655 		 * profiling as long as we didn't detect truncation.
656 		 * PMBPTR might be misaligned, but we'll burn that bridge
657 		 * when we get to it.
658 		 */
659 		if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
660 			arm_spe_perf_aux_output_begin(handle, event);
661 			isb();
662 		}
663 		break;
664 	case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
665 		/* We've seen you before, but GCC has the memory of a sieve. */
666 		break;
667 	}
668 
669 	/* The buffer pointers are now sane, so resume profiling. */
670 	write_sysreg_s(0, SYS_PMBSR_EL1);
671 	return IRQ_HANDLED;
672 }
673 
674 static u64 arm_spe_pmsevfr_res0(u16 pmsver)
675 {
676 	switch (pmsver) {
677 	case ID_AA64DFR0_PMSVER_8_2:
678 		return SYS_PMSEVFR_EL1_RES0_8_2;
679 	case ID_AA64DFR0_PMSVER_8_3:
680 	/* Return the highest version we support in default */
681 	default:
682 		return SYS_PMSEVFR_EL1_RES0_8_3;
683 	}
684 }
685 
686 /* Perf callbacks */
687 static int arm_spe_pmu_event_init(struct perf_event *event)
688 {
689 	u64 reg;
690 	struct perf_event_attr *attr = &event->attr;
691 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
692 
693 	/* This is, of course, deeply driver-specific */
694 	if (attr->type != event->pmu->type)
695 		return -ENOENT;
696 
697 	if (event->cpu >= 0 &&
698 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
699 		return -ENOENT;
700 
701 	if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
702 		return -EOPNOTSUPP;
703 
704 	if (attr->exclude_idle)
705 		return -EOPNOTSUPP;
706 
707 	/*
708 	 * Feedback-directed frequency throttling doesn't work when we
709 	 * have a buffer of samples. We'd need to manually count the
710 	 * samples in the buffer when it fills up and adjust the event
711 	 * count to reflect that. Instead, just force the user to specify
712 	 * a sample period.
713 	 */
714 	if (attr->freq)
715 		return -EINVAL;
716 
717 	reg = arm_spe_event_to_pmsfcr(event);
718 	if ((reg & BIT(SYS_PMSFCR_EL1_FE_SHIFT)) &&
719 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
720 		return -EOPNOTSUPP;
721 
722 	if ((reg & BIT(SYS_PMSFCR_EL1_FT_SHIFT)) &&
723 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
724 		return -EOPNOTSUPP;
725 
726 	if ((reg & BIT(SYS_PMSFCR_EL1_FL_SHIFT)) &&
727 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
728 		return -EOPNOTSUPP;
729 
730 	set_spe_event_has_cx(event);
731 	reg = arm_spe_event_to_pmscr(event);
732 	if (!perfmon_capable() &&
733 	    (reg & (BIT(SYS_PMSCR_EL1_PA_SHIFT) |
734 		    BIT(SYS_PMSCR_EL1_PCT_SHIFT))))
735 		return -EACCES;
736 
737 	return 0;
738 }
739 
740 static void arm_spe_pmu_start(struct perf_event *event, int flags)
741 {
742 	u64 reg;
743 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
744 	struct hw_perf_event *hwc = &event->hw;
745 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
746 
747 	hwc->state = 0;
748 	arm_spe_perf_aux_output_begin(handle, event);
749 	if (hwc->state)
750 		return;
751 
752 	reg = arm_spe_event_to_pmsfcr(event);
753 	write_sysreg_s(reg, SYS_PMSFCR_EL1);
754 
755 	reg = arm_spe_event_to_pmsevfr(event);
756 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
757 
758 	reg = arm_spe_event_to_pmslatfr(event);
759 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
760 
761 	if (flags & PERF_EF_RELOAD) {
762 		reg = arm_spe_event_to_pmsirr(event);
763 		write_sysreg_s(reg, SYS_PMSIRR_EL1);
764 		isb();
765 		reg = local64_read(&hwc->period_left);
766 		write_sysreg_s(reg, SYS_PMSICR_EL1);
767 	}
768 
769 	reg = arm_spe_event_to_pmscr(event);
770 	isb();
771 	write_sysreg_s(reg, SYS_PMSCR_EL1);
772 }
773 
774 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
775 {
776 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
777 	struct hw_perf_event *hwc = &event->hw;
778 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
779 
780 	/* If we're already stopped, then nothing to do */
781 	if (hwc->state & PERF_HES_STOPPED)
782 		return;
783 
784 	/* Stop all trace generation */
785 	arm_spe_pmu_disable_and_drain_local();
786 
787 	if (flags & PERF_EF_UPDATE) {
788 		/*
789 		 * If there's a fault pending then ensure we contain it
790 		 * to this buffer, since we might be on the context-switch
791 		 * path.
792 		 */
793 		if (perf_get_aux(handle)) {
794 			enum arm_spe_pmu_buf_fault_action act;
795 
796 			act = arm_spe_pmu_buf_get_fault_act(handle);
797 			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
798 				arm_spe_perf_aux_output_end(handle);
799 			else
800 				write_sysreg_s(0, SYS_PMBSR_EL1);
801 		}
802 
803 		/*
804 		 * This may also contain ECOUNT, but nobody else should
805 		 * be looking at period_left, since we forbid frequency
806 		 * based sampling.
807 		 */
808 		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
809 		hwc->state |= PERF_HES_UPTODATE;
810 	}
811 
812 	hwc->state |= PERF_HES_STOPPED;
813 }
814 
815 static int arm_spe_pmu_add(struct perf_event *event, int flags)
816 {
817 	int ret = 0;
818 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
819 	struct hw_perf_event *hwc = &event->hw;
820 	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
821 
822 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
823 		return -ENOENT;
824 
825 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
826 
827 	if (flags & PERF_EF_START) {
828 		arm_spe_pmu_start(event, PERF_EF_RELOAD);
829 		if (hwc->state & PERF_HES_STOPPED)
830 			ret = -EINVAL;
831 	}
832 
833 	return ret;
834 }
835 
836 static void arm_spe_pmu_del(struct perf_event *event, int flags)
837 {
838 	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
839 }
840 
841 static void arm_spe_pmu_read(struct perf_event *event)
842 {
843 }
844 
845 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
846 				   int nr_pages, bool snapshot)
847 {
848 	int i, cpu = event->cpu;
849 	struct page **pglist;
850 	struct arm_spe_pmu_buf *buf;
851 
852 	/* We need at least two pages for this to work. */
853 	if (nr_pages < 2)
854 		return NULL;
855 
856 	/*
857 	 * We require an even number of pages for snapshot mode, so that
858 	 * we can effectively treat the buffer as consisting of two equal
859 	 * parts and give userspace a fighting chance of getting some
860 	 * useful data out of it.
861 	 */
862 	if (snapshot && (nr_pages & 1))
863 		return NULL;
864 
865 	if (cpu == -1)
866 		cpu = raw_smp_processor_id();
867 
868 	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
869 	if (!buf)
870 		return NULL;
871 
872 	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
873 	if (!pglist)
874 		goto out_free_buf;
875 
876 	for (i = 0; i < nr_pages; ++i)
877 		pglist[i] = virt_to_page(pages[i]);
878 
879 	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
880 	if (!buf->base)
881 		goto out_free_pglist;
882 
883 	buf->nr_pages	= nr_pages;
884 	buf->snapshot	= snapshot;
885 
886 	kfree(pglist);
887 	return buf;
888 
889 out_free_pglist:
890 	kfree(pglist);
891 out_free_buf:
892 	kfree(buf);
893 	return NULL;
894 }
895 
896 static void arm_spe_pmu_free_aux(void *aux)
897 {
898 	struct arm_spe_pmu_buf *buf = aux;
899 
900 	vunmap(buf->base);
901 	kfree(buf);
902 }
903 
904 /* Initialisation and teardown functions */
905 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
906 {
907 	static atomic_t pmu_idx = ATOMIC_INIT(-1);
908 
909 	int idx;
910 	char *name;
911 	struct device *dev = &spe_pmu->pdev->dev;
912 
913 	spe_pmu->pmu = (struct pmu) {
914 		.module = THIS_MODULE,
915 		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
916 		.attr_groups	= arm_spe_pmu_attr_groups,
917 		/*
918 		 * We hitch a ride on the software context here, so that
919 		 * we can support per-task profiling (which is not possible
920 		 * with the invalid context as it doesn't get sched callbacks).
921 		 * This requires that userspace either uses a dummy event for
922 		 * perf_event_open, since the aux buffer is not setup until
923 		 * a subsequent mmap, or creates the profiling event in a
924 		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
925 		 * once the buffer has been created.
926 		 */
927 		.task_ctx_nr	= perf_sw_context,
928 		.event_init	= arm_spe_pmu_event_init,
929 		.add		= arm_spe_pmu_add,
930 		.del		= arm_spe_pmu_del,
931 		.start		= arm_spe_pmu_start,
932 		.stop		= arm_spe_pmu_stop,
933 		.read		= arm_spe_pmu_read,
934 		.setup_aux	= arm_spe_pmu_setup_aux,
935 		.free_aux	= arm_spe_pmu_free_aux,
936 	};
937 
938 	idx = atomic_inc_return(&pmu_idx);
939 	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
940 	if (!name) {
941 		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
942 		return -ENOMEM;
943 	}
944 
945 	return perf_pmu_register(&spe_pmu->pmu, name, -1);
946 }
947 
948 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
949 {
950 	perf_pmu_unregister(&spe_pmu->pmu);
951 }
952 
953 static void __arm_spe_pmu_dev_probe(void *info)
954 {
955 	int fld;
956 	u64 reg;
957 	struct arm_spe_pmu *spe_pmu = info;
958 	struct device *dev = &spe_pmu->pdev->dev;
959 
960 	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
961 						   ID_AA64DFR0_PMSVER_SHIFT);
962 	if (!fld) {
963 		dev_err(dev,
964 			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
965 			fld, smp_processor_id());
966 		return;
967 	}
968 	spe_pmu->pmsver = (u16)fld;
969 
970 	/* Read PMBIDR first to determine whether or not we have access */
971 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
972 	if (reg & BIT(SYS_PMBIDR_EL1_P_SHIFT)) {
973 		dev_err(dev,
974 			"profiling buffer owned by higher exception level\n");
975 		return;
976 	}
977 
978 	/* Minimum alignment. If it's out-of-range, then fail the probe */
979 	fld = reg >> SYS_PMBIDR_EL1_ALIGN_SHIFT & SYS_PMBIDR_EL1_ALIGN_MASK;
980 	spe_pmu->align = 1 << fld;
981 	if (spe_pmu->align > SZ_2K) {
982 		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
983 			fld, smp_processor_id());
984 		return;
985 	}
986 
987 	/* It's now safe to read PMSIDR and figure out what we've got */
988 	reg = read_sysreg_s(SYS_PMSIDR_EL1);
989 	if (reg & BIT(SYS_PMSIDR_EL1_FE_SHIFT))
990 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
991 
992 	if (reg & BIT(SYS_PMSIDR_EL1_FT_SHIFT))
993 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
994 
995 	if (reg & BIT(SYS_PMSIDR_EL1_FL_SHIFT))
996 		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
997 
998 	if (reg & BIT(SYS_PMSIDR_EL1_ARCHINST_SHIFT))
999 		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
1000 
1001 	if (reg & BIT(SYS_PMSIDR_EL1_LDS_SHIFT))
1002 		spe_pmu->features |= SPE_PMU_FEAT_LDS;
1003 
1004 	if (reg & BIT(SYS_PMSIDR_EL1_ERND_SHIFT))
1005 		spe_pmu->features |= SPE_PMU_FEAT_ERND;
1006 
1007 	/* This field has a spaced out encoding, so just use a look-up */
1008 	fld = reg >> SYS_PMSIDR_EL1_INTERVAL_SHIFT & SYS_PMSIDR_EL1_INTERVAL_MASK;
1009 	switch (fld) {
1010 	case 0:
1011 		spe_pmu->min_period = 256;
1012 		break;
1013 	case 2:
1014 		spe_pmu->min_period = 512;
1015 		break;
1016 	case 3:
1017 		spe_pmu->min_period = 768;
1018 		break;
1019 	case 4:
1020 		spe_pmu->min_period = 1024;
1021 		break;
1022 	case 5:
1023 		spe_pmu->min_period = 1536;
1024 		break;
1025 	case 6:
1026 		spe_pmu->min_period = 2048;
1027 		break;
1028 	case 7:
1029 		spe_pmu->min_period = 3072;
1030 		break;
1031 	default:
1032 		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1033 			 fld);
1034 		fallthrough;
1035 	case 8:
1036 		spe_pmu->min_period = 4096;
1037 	}
1038 
1039 	/* Maximum record size. If it's out-of-range, then fail the probe */
1040 	fld = reg >> SYS_PMSIDR_EL1_MAXSIZE_SHIFT & SYS_PMSIDR_EL1_MAXSIZE_MASK;
1041 	spe_pmu->max_record_sz = 1 << fld;
1042 	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1043 		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1044 			fld, smp_processor_id());
1045 		return;
1046 	}
1047 
1048 	fld = reg >> SYS_PMSIDR_EL1_COUNTSIZE_SHIFT & SYS_PMSIDR_EL1_COUNTSIZE_MASK;
1049 	switch (fld) {
1050 	default:
1051 		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1052 			 fld);
1053 		fallthrough;
1054 	case 2:
1055 		spe_pmu->counter_sz = 12;
1056 		break;
1057 	case 3:
1058 		spe_pmu->counter_sz = 16;
1059 	}
1060 
1061 	dev_info(dev,
1062 		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1063 		 cpumask_pr_args(&spe_pmu->supported_cpus),
1064 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1065 
1066 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1067 }
1068 
1069 static void __arm_spe_pmu_reset_local(void)
1070 {
1071 	/*
1072 	 * This is probably overkill, as we have no idea where we're
1073 	 * draining any buffered data to...
1074 	 */
1075 	arm_spe_pmu_disable_and_drain_local();
1076 
1077 	/* Reset the buffer base pointer */
1078 	write_sysreg_s(0, SYS_PMBPTR_EL1);
1079 	isb();
1080 
1081 	/* Clear any pending management interrupts */
1082 	write_sysreg_s(0, SYS_PMBSR_EL1);
1083 	isb();
1084 }
1085 
1086 static void __arm_spe_pmu_setup_one(void *info)
1087 {
1088 	struct arm_spe_pmu *spe_pmu = info;
1089 
1090 	__arm_spe_pmu_reset_local();
1091 	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1092 }
1093 
1094 static void __arm_spe_pmu_stop_one(void *info)
1095 {
1096 	struct arm_spe_pmu *spe_pmu = info;
1097 
1098 	disable_percpu_irq(spe_pmu->irq);
1099 	__arm_spe_pmu_reset_local();
1100 }
1101 
1102 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1103 {
1104 	struct arm_spe_pmu *spe_pmu;
1105 
1106 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1107 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1108 		return 0;
1109 
1110 	__arm_spe_pmu_setup_one(spe_pmu);
1111 	return 0;
1112 }
1113 
1114 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1115 {
1116 	struct arm_spe_pmu *spe_pmu;
1117 
1118 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1119 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1120 		return 0;
1121 
1122 	__arm_spe_pmu_stop_one(spe_pmu);
1123 	return 0;
1124 }
1125 
1126 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1127 {
1128 	int ret;
1129 	cpumask_t *mask = &spe_pmu->supported_cpus;
1130 
1131 	/* Make sure we probe the hardware on a relevant CPU */
1132 	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1133 	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1134 		return -ENXIO;
1135 
1136 	/* Request our PPIs (note that the IRQ is still disabled) */
1137 	ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1138 				 spe_pmu->handle);
1139 	if (ret)
1140 		return ret;
1141 
1142 	/*
1143 	 * Register our hotplug notifier now so we don't miss any events.
1144 	 * This will enable the IRQ for any supported CPUs that are already
1145 	 * up.
1146 	 */
1147 	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1148 				       &spe_pmu->hotplug_node);
1149 	if (ret)
1150 		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1151 
1152 	return ret;
1153 }
1154 
1155 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1156 {
1157 	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1158 	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1159 }
1160 
1161 /* Driver and device probing */
1162 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1163 {
1164 	struct platform_device *pdev = spe_pmu->pdev;
1165 	int irq = platform_get_irq(pdev, 0);
1166 
1167 	if (irq < 0)
1168 		return -ENXIO;
1169 
1170 	if (!irq_is_percpu(irq)) {
1171 		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1172 		return -EINVAL;
1173 	}
1174 
1175 	if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1176 		dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1177 		return -EINVAL;
1178 	}
1179 
1180 	spe_pmu->irq = irq;
1181 	return 0;
1182 }
1183 
1184 static const struct of_device_id arm_spe_pmu_of_match[] = {
1185 	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1186 	{ /* Sentinel */ },
1187 };
1188 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1189 
1190 static const struct platform_device_id arm_spe_match[] = {
1191 	{ ARMV8_SPE_PDEV_NAME, 0},
1192 	{ }
1193 };
1194 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1195 
1196 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1197 {
1198 	int ret;
1199 	struct arm_spe_pmu *spe_pmu;
1200 	struct device *dev = &pdev->dev;
1201 
1202 	/*
1203 	 * If kernelspace is unmapped when running at EL0, then the SPE
1204 	 * buffer will fault and prematurely terminate the AUX session.
1205 	 */
1206 	if (arm64_kernel_unmapped_at_el0()) {
1207 		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1208 		return -EPERM;
1209 	}
1210 
1211 	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1212 	if (!spe_pmu)
1213 		return -ENOMEM;
1214 
1215 	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1216 	if (!spe_pmu->handle)
1217 		return -ENOMEM;
1218 
1219 	spe_pmu->pdev = pdev;
1220 	platform_set_drvdata(pdev, spe_pmu);
1221 
1222 	ret = arm_spe_pmu_irq_probe(spe_pmu);
1223 	if (ret)
1224 		goto out_free_handle;
1225 
1226 	ret = arm_spe_pmu_dev_init(spe_pmu);
1227 	if (ret)
1228 		goto out_free_handle;
1229 
1230 	ret = arm_spe_pmu_perf_init(spe_pmu);
1231 	if (ret)
1232 		goto out_teardown_dev;
1233 
1234 	return 0;
1235 
1236 out_teardown_dev:
1237 	arm_spe_pmu_dev_teardown(spe_pmu);
1238 out_free_handle:
1239 	free_percpu(spe_pmu->handle);
1240 	return ret;
1241 }
1242 
1243 static int arm_spe_pmu_device_remove(struct platform_device *pdev)
1244 {
1245 	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1246 
1247 	arm_spe_pmu_perf_destroy(spe_pmu);
1248 	arm_spe_pmu_dev_teardown(spe_pmu);
1249 	free_percpu(spe_pmu->handle);
1250 	return 0;
1251 }
1252 
1253 static struct platform_driver arm_spe_pmu_driver = {
1254 	.id_table = arm_spe_match,
1255 	.driver	= {
1256 		.name		= DRVNAME,
1257 		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1258 		.suppress_bind_attrs = true,
1259 	},
1260 	.probe	= arm_spe_pmu_device_probe,
1261 	.remove	= arm_spe_pmu_device_remove,
1262 };
1263 
1264 static int __init arm_spe_pmu_init(void)
1265 {
1266 	int ret;
1267 
1268 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1269 				      arm_spe_pmu_cpu_startup,
1270 				      arm_spe_pmu_cpu_teardown);
1271 	if (ret < 0)
1272 		return ret;
1273 	arm_spe_pmu_online = ret;
1274 
1275 	ret = platform_driver_register(&arm_spe_pmu_driver);
1276 	if (ret)
1277 		cpuhp_remove_multi_state(arm_spe_pmu_online);
1278 
1279 	return ret;
1280 }
1281 
1282 static void __exit arm_spe_pmu_exit(void)
1283 {
1284 	platform_driver_unregister(&arm_spe_pmu_driver);
1285 	cpuhp_remove_multi_state(arm_spe_pmu_online);
1286 }
1287 
1288 module_init(arm_spe_pmu_init);
1289 module_exit(arm_spe_pmu_exit);
1290 
1291 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1292 MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1293 MODULE_LICENSE("GPL v2");
1294