xref: /linux/drivers/perf/arm_spe_pmu.c (revision 7fc7f25419f5a6b09199ba4b5026b94ef184fa79)
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			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_DEV_PROBED			(1UL << 63)
88 	u64					features;
89 
90 	u16					max_record_sz;
91 	u16					align;
92 	struct perf_output_handle __percpu	*handle;
93 };
94 
95 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
96 
97 /* Convert a free-running index from perf into an SPE buffer offset */
98 #define PERF_IDX2OFF(idx, buf)	((idx) % ((buf)->nr_pages << PAGE_SHIFT))
99 
100 /* Keep track of our dynamic hotplug state */
101 static enum cpuhp_state arm_spe_pmu_online;
102 
103 enum arm_spe_pmu_buf_fault_action {
104 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
105 	SPE_PMU_BUF_FAULT_ACT_FATAL,
106 	SPE_PMU_BUF_FAULT_ACT_OK,
107 };
108 
109 /* This sysfs gunk was really good fun to write. */
110 enum arm_spe_pmu_capabilities {
111 	SPE_PMU_CAP_ARCH_INST = 0,
112 	SPE_PMU_CAP_ERND,
113 	SPE_PMU_CAP_FEAT_MAX,
114 	SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
115 	SPE_PMU_CAP_MIN_IVAL,
116 };
117 
118 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
119 	[SPE_PMU_CAP_ARCH_INST]	= SPE_PMU_FEAT_ARCH_INST,
120 	[SPE_PMU_CAP_ERND]	= SPE_PMU_FEAT_ERND,
121 };
122 
123 static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
124 {
125 	if (cap < SPE_PMU_CAP_FEAT_MAX)
126 		return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
127 
128 	switch (cap) {
129 	case SPE_PMU_CAP_CNT_SZ:
130 		return spe_pmu->counter_sz;
131 	case SPE_PMU_CAP_MIN_IVAL:
132 		return spe_pmu->min_period;
133 	default:
134 		WARN(1, "unknown cap %d\n", cap);
135 	}
136 
137 	return 0;
138 }
139 
140 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
141 				    struct device_attribute *attr,
142 				    char *buf)
143 {
144 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
145 	struct dev_ext_attribute *ea =
146 		container_of(attr, struct dev_ext_attribute, attr);
147 	int cap = (long)ea->var;
148 
149 	return sysfs_emit(buf, "%u\n", arm_spe_pmu_cap_get(spe_pmu, cap));
150 }
151 
152 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
153 	&((struct dev_ext_attribute[]) {				\
154 		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
155 	})[0].attr.attr
156 
157 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
158 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
159 
160 static struct attribute *arm_spe_pmu_cap_attr[] = {
161 	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
162 	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
163 	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
164 	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
165 	NULL,
166 };
167 
168 static const struct attribute_group arm_spe_pmu_cap_group = {
169 	.name	= "caps",
170 	.attrs	= arm_spe_pmu_cap_attr,
171 };
172 
173 /* User ABI */
174 #define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
175 #define ATTR_CFG_FLD_ts_enable_LO		0
176 #define ATTR_CFG_FLD_ts_enable_HI		0
177 #define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
178 #define ATTR_CFG_FLD_pa_enable_LO		1
179 #define ATTR_CFG_FLD_pa_enable_HI		1
180 #define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
181 #define ATTR_CFG_FLD_pct_enable_LO		2
182 #define ATTR_CFG_FLD_pct_enable_HI		2
183 #define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
184 #define ATTR_CFG_FLD_jitter_LO			16
185 #define ATTR_CFG_FLD_jitter_HI			16
186 #define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
187 #define ATTR_CFG_FLD_branch_filter_LO		32
188 #define ATTR_CFG_FLD_branch_filter_HI		32
189 #define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
190 #define ATTR_CFG_FLD_load_filter_LO		33
191 #define ATTR_CFG_FLD_load_filter_HI		33
192 #define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
193 #define ATTR_CFG_FLD_store_filter_LO		34
194 #define ATTR_CFG_FLD_store_filter_HI		34
195 
196 #define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
197 #define ATTR_CFG_FLD_event_filter_LO		0
198 #define ATTR_CFG_FLD_event_filter_HI		63
199 
200 #define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
201 #define ATTR_CFG_FLD_min_latency_LO		0
202 #define ATTR_CFG_FLD_min_latency_HI		11
203 
204 /* Why does everything I do descend into this? */
205 #define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
206 	(lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
207 
208 #define _GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
209 	__GEN_PMU_FORMAT_ATTR(cfg, lo, hi)
210 
211 #define GEN_PMU_FORMAT_ATTR(name)					\
212 	PMU_FORMAT_ATTR(name,						\
213 	_GEN_PMU_FORMAT_ATTR(ATTR_CFG_FLD_##name##_CFG,			\
214 			     ATTR_CFG_FLD_##name##_LO,			\
215 			     ATTR_CFG_FLD_##name##_HI))
216 
217 #define _ATTR_CFG_GET_FLD(attr, cfg, lo, hi)				\
218 	((((attr)->cfg) >> lo) & GENMASK(hi - lo, 0))
219 
220 #define ATTR_CFG_GET_FLD(attr, name)					\
221 	_ATTR_CFG_GET_FLD(attr,						\
222 			  ATTR_CFG_FLD_##name##_CFG,			\
223 			  ATTR_CFG_FLD_##name##_LO,			\
224 			  ATTR_CFG_FLD_##name##_HI)
225 
226 GEN_PMU_FORMAT_ATTR(ts_enable);
227 GEN_PMU_FORMAT_ATTR(pa_enable);
228 GEN_PMU_FORMAT_ATTR(pct_enable);
229 GEN_PMU_FORMAT_ATTR(jitter);
230 GEN_PMU_FORMAT_ATTR(branch_filter);
231 GEN_PMU_FORMAT_ATTR(load_filter);
232 GEN_PMU_FORMAT_ATTR(store_filter);
233 GEN_PMU_FORMAT_ATTR(event_filter);
234 GEN_PMU_FORMAT_ATTR(min_latency);
235 
236 static struct attribute *arm_spe_pmu_formats_attr[] = {
237 	&format_attr_ts_enable.attr,
238 	&format_attr_pa_enable.attr,
239 	&format_attr_pct_enable.attr,
240 	&format_attr_jitter.attr,
241 	&format_attr_branch_filter.attr,
242 	&format_attr_load_filter.attr,
243 	&format_attr_store_filter.attr,
244 	&format_attr_event_filter.attr,
245 	&format_attr_min_latency.attr,
246 	NULL,
247 };
248 
249 static const struct attribute_group arm_spe_pmu_format_group = {
250 	.name	= "format",
251 	.attrs	= arm_spe_pmu_formats_attr,
252 };
253 
254 static ssize_t cpumask_show(struct device *dev,
255 			    struct device_attribute *attr, char *buf)
256 {
257 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
258 
259 	return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
260 }
261 static DEVICE_ATTR_RO(cpumask);
262 
263 static struct attribute *arm_spe_pmu_attrs[] = {
264 	&dev_attr_cpumask.attr,
265 	NULL,
266 };
267 
268 static const struct attribute_group arm_spe_pmu_group = {
269 	.attrs	= arm_spe_pmu_attrs,
270 };
271 
272 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
273 	&arm_spe_pmu_group,
274 	&arm_spe_pmu_cap_group,
275 	&arm_spe_pmu_format_group,
276 	NULL,
277 };
278 
279 /* Convert between user ABI and register values */
280 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
281 {
282 	struct perf_event_attr *attr = &event->attr;
283 	u64 reg = 0;
284 
285 	reg |= ATTR_CFG_GET_FLD(attr, ts_enable) << SYS_PMSCR_EL1_TS_SHIFT;
286 	reg |= ATTR_CFG_GET_FLD(attr, pa_enable) << SYS_PMSCR_EL1_PA_SHIFT;
287 	reg |= ATTR_CFG_GET_FLD(attr, pct_enable) << SYS_PMSCR_EL1_PCT_SHIFT;
288 
289 	if (!attr->exclude_user)
290 		reg |= BIT(SYS_PMSCR_EL1_E0SPE_SHIFT);
291 
292 	if (!attr->exclude_kernel)
293 		reg |= BIT(SYS_PMSCR_EL1_E1SPE_SHIFT);
294 
295 	if (get_spe_event_has_cx(event))
296 		reg |= BIT(SYS_PMSCR_EL1_CX_SHIFT);
297 
298 	return reg;
299 }
300 
301 static void arm_spe_event_sanitise_period(struct perf_event *event)
302 {
303 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
304 	u64 period = event->hw.sample_period;
305 	u64 max_period = SYS_PMSIRR_EL1_INTERVAL_MASK
306 			 << SYS_PMSIRR_EL1_INTERVAL_SHIFT;
307 
308 	if (period < spe_pmu->min_period)
309 		period = spe_pmu->min_period;
310 	else if (period > max_period)
311 		period = max_period;
312 	else
313 		period &= max_period;
314 
315 	event->hw.sample_period = period;
316 }
317 
318 static u64 arm_spe_event_to_pmsirr(struct perf_event *event)
319 {
320 	struct perf_event_attr *attr = &event->attr;
321 	u64 reg = 0;
322 
323 	arm_spe_event_sanitise_period(event);
324 
325 	reg |= ATTR_CFG_GET_FLD(attr, jitter) << SYS_PMSIRR_EL1_RND_SHIFT;
326 	reg |= event->hw.sample_period;
327 
328 	return reg;
329 }
330 
331 static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
332 {
333 	struct perf_event_attr *attr = &event->attr;
334 	u64 reg = 0;
335 
336 	reg |= ATTR_CFG_GET_FLD(attr, load_filter) << SYS_PMSFCR_EL1_LD_SHIFT;
337 	reg |= ATTR_CFG_GET_FLD(attr, store_filter) << SYS_PMSFCR_EL1_ST_SHIFT;
338 	reg |= ATTR_CFG_GET_FLD(attr, branch_filter) << SYS_PMSFCR_EL1_B_SHIFT;
339 
340 	if (reg)
341 		reg |= BIT(SYS_PMSFCR_EL1_FT_SHIFT);
342 
343 	if (ATTR_CFG_GET_FLD(attr, event_filter))
344 		reg |= BIT(SYS_PMSFCR_EL1_FE_SHIFT);
345 
346 	if (ATTR_CFG_GET_FLD(attr, min_latency))
347 		reg |= BIT(SYS_PMSFCR_EL1_FL_SHIFT);
348 
349 	return reg;
350 }
351 
352 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
353 {
354 	struct perf_event_attr *attr = &event->attr;
355 	return ATTR_CFG_GET_FLD(attr, event_filter);
356 }
357 
358 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
359 {
360 	struct perf_event_attr *attr = &event->attr;
361 	return ATTR_CFG_GET_FLD(attr, min_latency)
362 	       << SYS_PMSLATFR_EL1_MINLAT_SHIFT;
363 }
364 
365 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
366 {
367 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
368 	u64 head = PERF_IDX2OFF(handle->head, buf);
369 
370 	memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
371 	if (!buf->snapshot)
372 		perf_aux_output_skip(handle, len);
373 }
374 
375 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
376 {
377 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
378 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
379 	u64 head = PERF_IDX2OFF(handle->head, buf);
380 	u64 limit = buf->nr_pages * PAGE_SIZE;
381 
382 	/*
383 	 * The trace format isn't parseable in reverse, so clamp
384 	 * the limit to half of the buffer size in snapshot mode
385 	 * so that the worst case is half a buffer of records, as
386 	 * opposed to a single record.
387 	 */
388 	if (head < limit >> 1)
389 		limit >>= 1;
390 
391 	/*
392 	 * If we're within max_record_sz of the limit, we must
393 	 * pad, move the head index and recompute the limit.
394 	 */
395 	if (limit - head < spe_pmu->max_record_sz) {
396 		arm_spe_pmu_pad_buf(handle, limit - head);
397 		handle->head = PERF_IDX2OFF(limit, buf);
398 		limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
399 	}
400 
401 	return limit;
402 }
403 
404 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
405 {
406 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
407 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
408 	const u64 bufsize = buf->nr_pages * PAGE_SIZE;
409 	u64 limit = bufsize;
410 	u64 head, tail, wakeup;
411 
412 	/*
413 	 * The head can be misaligned for two reasons:
414 	 *
415 	 * 1. The hardware left PMBPTR pointing to the first byte after
416 	 *    a record when generating a buffer management event.
417 	 *
418 	 * 2. We used perf_aux_output_skip to consume handle->size bytes
419 	 *    and CIRC_SPACE was used to compute the size, which always
420 	 *    leaves one entry free.
421 	 *
422 	 * Deal with this by padding to the next alignment boundary and
423 	 * moving the head index. If we run out of buffer space, we'll
424 	 * reduce handle->size to zero and end up reporting truncation.
425 	 */
426 	head = PERF_IDX2OFF(handle->head, buf);
427 	if (!IS_ALIGNED(head, spe_pmu->align)) {
428 		unsigned long delta = roundup(head, spe_pmu->align) - head;
429 
430 		delta = min(delta, handle->size);
431 		arm_spe_pmu_pad_buf(handle, delta);
432 		head = PERF_IDX2OFF(handle->head, buf);
433 	}
434 
435 	/* If we've run out of free space, then nothing more to do */
436 	if (!handle->size)
437 		goto no_space;
438 
439 	/* Compute the tail and wakeup indices now that we've aligned head */
440 	tail = PERF_IDX2OFF(handle->head + handle->size, buf);
441 	wakeup = PERF_IDX2OFF(handle->wakeup, buf);
442 
443 	/*
444 	 * Avoid clobbering unconsumed data. We know we have space, so
445 	 * if we see head == tail we know that the buffer is empty. If
446 	 * head > tail, then there's nothing to clobber prior to
447 	 * wrapping.
448 	 */
449 	if (head < tail)
450 		limit = round_down(tail, PAGE_SIZE);
451 
452 	/*
453 	 * Wakeup may be arbitrarily far into the future. If it's not in
454 	 * the current generation, either we'll wrap before hitting it,
455 	 * or it's in the past and has been handled already.
456 	 *
457 	 * If there's a wakeup before we wrap, arrange to be woken up by
458 	 * the page boundary following it. Keep the tail boundary if
459 	 * that's lower.
460 	 */
461 	if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
462 		limit = min(limit, round_up(wakeup, PAGE_SIZE));
463 
464 	if (limit > head)
465 		return limit;
466 
467 	arm_spe_pmu_pad_buf(handle, handle->size);
468 no_space:
469 	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
470 	perf_aux_output_end(handle, 0);
471 	return 0;
472 }
473 
474 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
475 {
476 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
477 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
478 	u64 limit = __arm_spe_pmu_next_off(handle);
479 	u64 head = PERF_IDX2OFF(handle->head, buf);
480 
481 	/*
482 	 * If the head has come too close to the end of the buffer,
483 	 * then pad to the end and recompute the limit.
484 	 */
485 	if (limit && (limit - head < spe_pmu->max_record_sz)) {
486 		arm_spe_pmu_pad_buf(handle, limit - head);
487 		limit = __arm_spe_pmu_next_off(handle);
488 	}
489 
490 	return limit;
491 }
492 
493 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
494 					  struct perf_event *event)
495 {
496 	u64 base, limit;
497 	struct arm_spe_pmu_buf *buf;
498 
499 	/* Start a new aux session */
500 	buf = perf_aux_output_begin(handle, event);
501 	if (!buf) {
502 		event->hw.state |= PERF_HES_STOPPED;
503 		/*
504 		 * We still need to clear the limit pointer, since the
505 		 * profiler might only be disabled by virtue of a fault.
506 		 */
507 		limit = 0;
508 		goto out_write_limit;
509 	}
510 
511 	limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
512 			      : arm_spe_pmu_next_off(handle);
513 	if (limit)
514 		limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT);
515 
516 	limit += (u64)buf->base;
517 	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
518 	write_sysreg_s(base, SYS_PMBPTR_EL1);
519 
520 out_write_limit:
521 	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
522 }
523 
524 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
525 {
526 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
527 	u64 offset, size;
528 
529 	offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
530 	size = offset - PERF_IDX2OFF(handle->head, buf);
531 
532 	if (buf->snapshot)
533 		handle->head = offset;
534 
535 	perf_aux_output_end(handle, size);
536 }
537 
538 static void arm_spe_pmu_disable_and_drain_local(void)
539 {
540 	/* Disable profiling at EL0 and EL1 */
541 	write_sysreg_s(0, SYS_PMSCR_EL1);
542 	isb();
543 
544 	/* Drain any buffered data */
545 	psb_csync();
546 	dsb(nsh);
547 
548 	/* Disable the profiling buffer */
549 	write_sysreg_s(0, SYS_PMBLIMITR_EL1);
550 	isb();
551 }
552 
553 /* IRQ handling */
554 static enum arm_spe_pmu_buf_fault_action
555 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
556 {
557 	const char *err_str;
558 	u64 pmbsr;
559 	enum arm_spe_pmu_buf_fault_action ret;
560 
561 	/*
562 	 * Ensure new profiling data is visible to the CPU and any external
563 	 * aborts have been resolved.
564 	 */
565 	psb_csync();
566 	dsb(nsh);
567 
568 	/* Ensure hardware updates to PMBPTR_EL1 are visible */
569 	isb();
570 
571 	/* Service required? */
572 	pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
573 	if (!(pmbsr & BIT(SYS_PMBSR_EL1_S_SHIFT)))
574 		return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
575 
576 	/*
577 	 * If we've lost data, disable profiling and also set the PARTIAL
578 	 * flag to indicate that the last record is corrupted.
579 	 */
580 	if (pmbsr & BIT(SYS_PMBSR_EL1_DL_SHIFT))
581 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
582 					     PERF_AUX_FLAG_PARTIAL);
583 
584 	/* Report collisions to userspace so that it can up the period */
585 	if (pmbsr & BIT(SYS_PMBSR_EL1_COLL_SHIFT))
586 		perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
587 
588 	/* We only expect buffer management events */
589 	switch (pmbsr & (SYS_PMBSR_EL1_EC_MASK << SYS_PMBSR_EL1_EC_SHIFT)) {
590 	case SYS_PMBSR_EL1_EC_BUF:
591 		/* Handled below */
592 		break;
593 	case SYS_PMBSR_EL1_EC_FAULT_S1:
594 	case SYS_PMBSR_EL1_EC_FAULT_S2:
595 		err_str = "Unexpected buffer fault";
596 		goto out_err;
597 	default:
598 		err_str = "Unknown error code";
599 		goto out_err;
600 	}
601 
602 	/* Buffer management event */
603 	switch (pmbsr &
604 		(SYS_PMBSR_EL1_BUF_BSC_MASK << SYS_PMBSR_EL1_BUF_BSC_SHIFT)) {
605 	case SYS_PMBSR_EL1_BUF_BSC_FULL:
606 		ret = SPE_PMU_BUF_FAULT_ACT_OK;
607 		goto out_stop;
608 	default:
609 		err_str = "Unknown buffer status code";
610 	}
611 
612 out_err:
613 	pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
614 			   err_str, smp_processor_id(), pmbsr,
615 			   read_sysreg_s(SYS_PMBPTR_EL1),
616 			   read_sysreg_s(SYS_PMBLIMITR_EL1));
617 	ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
618 
619 out_stop:
620 	arm_spe_perf_aux_output_end(handle);
621 	return ret;
622 }
623 
624 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
625 {
626 	struct perf_output_handle *handle = dev;
627 	struct perf_event *event = handle->event;
628 	enum arm_spe_pmu_buf_fault_action act;
629 
630 	if (!perf_get_aux(handle))
631 		return IRQ_NONE;
632 
633 	act = arm_spe_pmu_buf_get_fault_act(handle);
634 	if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
635 		return IRQ_NONE;
636 
637 	/*
638 	 * Ensure perf callbacks have completed, which may disable the
639 	 * profiling buffer in response to a TRUNCATION flag.
640 	 */
641 	irq_work_run();
642 
643 	switch (act) {
644 	case SPE_PMU_BUF_FAULT_ACT_FATAL:
645 		/*
646 		 * If a fatal exception occurred then leaving the profiling
647 		 * buffer enabled is a recipe waiting to happen. Since
648 		 * fatal faults don't always imply truncation, make sure
649 		 * that the profiling buffer is disabled explicitly before
650 		 * clearing the syndrome register.
651 		 */
652 		arm_spe_pmu_disable_and_drain_local();
653 		break;
654 	case SPE_PMU_BUF_FAULT_ACT_OK:
655 		/*
656 		 * We handled the fault (the buffer was full), so resume
657 		 * profiling as long as we didn't detect truncation.
658 		 * PMBPTR might be misaligned, but we'll burn that bridge
659 		 * when we get to it.
660 		 */
661 		if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
662 			arm_spe_perf_aux_output_begin(handle, event);
663 			isb();
664 		}
665 		break;
666 	case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
667 		/* We've seen you before, but GCC has the memory of a sieve. */
668 		break;
669 	}
670 
671 	/* The buffer pointers are now sane, so resume profiling. */
672 	write_sysreg_s(0, SYS_PMBSR_EL1);
673 	return IRQ_HANDLED;
674 }
675 
676 static u64 arm_spe_pmsevfr_res0(u16 pmsver)
677 {
678 	switch (pmsver) {
679 	case ID_AA64DFR0_EL1_PMSVer_IMP:
680 		return SYS_PMSEVFR_EL1_RES0_8_2;
681 	case ID_AA64DFR0_EL1_PMSVer_V1P1:
682 	/* Return the highest version we support in default */
683 	default:
684 		return SYS_PMSEVFR_EL1_RES0_8_3;
685 	}
686 }
687 
688 /* Perf callbacks */
689 static int arm_spe_pmu_event_init(struct perf_event *event)
690 {
691 	u64 reg;
692 	struct perf_event_attr *attr = &event->attr;
693 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
694 
695 	/* This is, of course, deeply driver-specific */
696 	if (attr->type != event->pmu->type)
697 		return -ENOENT;
698 
699 	if (event->cpu >= 0 &&
700 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
701 		return -ENOENT;
702 
703 	if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
704 		return -EOPNOTSUPP;
705 
706 	if (attr->exclude_idle)
707 		return -EOPNOTSUPP;
708 
709 	/*
710 	 * Feedback-directed frequency throttling doesn't work when we
711 	 * have a buffer of samples. We'd need to manually count the
712 	 * samples in the buffer when it fills up and adjust the event
713 	 * count to reflect that. Instead, just force the user to specify
714 	 * a sample period.
715 	 */
716 	if (attr->freq)
717 		return -EINVAL;
718 
719 	reg = arm_spe_event_to_pmsfcr(event);
720 	if ((reg & BIT(SYS_PMSFCR_EL1_FE_SHIFT)) &&
721 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
722 		return -EOPNOTSUPP;
723 
724 	if ((reg & BIT(SYS_PMSFCR_EL1_FT_SHIFT)) &&
725 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
726 		return -EOPNOTSUPP;
727 
728 	if ((reg & BIT(SYS_PMSFCR_EL1_FL_SHIFT)) &&
729 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
730 		return -EOPNOTSUPP;
731 
732 	set_spe_event_has_cx(event);
733 	reg = arm_spe_event_to_pmscr(event);
734 	if (!perfmon_capable() &&
735 	    (reg & (BIT(SYS_PMSCR_EL1_PA_SHIFT) |
736 		    BIT(SYS_PMSCR_EL1_PCT_SHIFT))))
737 		return -EACCES;
738 
739 	return 0;
740 }
741 
742 static void arm_spe_pmu_start(struct perf_event *event, int flags)
743 {
744 	u64 reg;
745 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
746 	struct hw_perf_event *hwc = &event->hw;
747 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
748 
749 	hwc->state = 0;
750 	arm_spe_perf_aux_output_begin(handle, event);
751 	if (hwc->state)
752 		return;
753 
754 	reg = arm_spe_event_to_pmsfcr(event);
755 	write_sysreg_s(reg, SYS_PMSFCR_EL1);
756 
757 	reg = arm_spe_event_to_pmsevfr(event);
758 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
759 
760 	reg = arm_spe_event_to_pmslatfr(event);
761 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
762 
763 	if (flags & PERF_EF_RELOAD) {
764 		reg = arm_spe_event_to_pmsirr(event);
765 		write_sysreg_s(reg, SYS_PMSIRR_EL1);
766 		isb();
767 		reg = local64_read(&hwc->period_left);
768 		write_sysreg_s(reg, SYS_PMSICR_EL1);
769 	}
770 
771 	reg = arm_spe_event_to_pmscr(event);
772 	isb();
773 	write_sysreg_s(reg, SYS_PMSCR_EL1);
774 }
775 
776 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
777 {
778 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
779 	struct hw_perf_event *hwc = &event->hw;
780 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
781 
782 	/* If we're already stopped, then nothing to do */
783 	if (hwc->state & PERF_HES_STOPPED)
784 		return;
785 
786 	/* Stop all trace generation */
787 	arm_spe_pmu_disable_and_drain_local();
788 
789 	if (flags & PERF_EF_UPDATE) {
790 		/*
791 		 * If there's a fault pending then ensure we contain it
792 		 * to this buffer, since we might be on the context-switch
793 		 * path.
794 		 */
795 		if (perf_get_aux(handle)) {
796 			enum arm_spe_pmu_buf_fault_action act;
797 
798 			act = arm_spe_pmu_buf_get_fault_act(handle);
799 			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
800 				arm_spe_perf_aux_output_end(handle);
801 			else
802 				write_sysreg_s(0, SYS_PMBSR_EL1);
803 		}
804 
805 		/*
806 		 * This may also contain ECOUNT, but nobody else should
807 		 * be looking at period_left, since we forbid frequency
808 		 * based sampling.
809 		 */
810 		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
811 		hwc->state |= PERF_HES_UPTODATE;
812 	}
813 
814 	hwc->state |= PERF_HES_STOPPED;
815 }
816 
817 static int arm_spe_pmu_add(struct perf_event *event, int flags)
818 {
819 	int ret = 0;
820 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
821 	struct hw_perf_event *hwc = &event->hw;
822 	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
823 
824 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
825 		return -ENOENT;
826 
827 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
828 
829 	if (flags & PERF_EF_START) {
830 		arm_spe_pmu_start(event, PERF_EF_RELOAD);
831 		if (hwc->state & PERF_HES_STOPPED)
832 			ret = -EINVAL;
833 	}
834 
835 	return ret;
836 }
837 
838 static void arm_spe_pmu_del(struct perf_event *event, int flags)
839 {
840 	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
841 }
842 
843 static void arm_spe_pmu_read(struct perf_event *event)
844 {
845 }
846 
847 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
848 				   int nr_pages, bool snapshot)
849 {
850 	int i, cpu = event->cpu;
851 	struct page **pglist;
852 	struct arm_spe_pmu_buf *buf;
853 
854 	/* We need at least two pages for this to work. */
855 	if (nr_pages < 2)
856 		return NULL;
857 
858 	/*
859 	 * We require an even number of pages for snapshot mode, so that
860 	 * we can effectively treat the buffer as consisting of two equal
861 	 * parts and give userspace a fighting chance of getting some
862 	 * useful data out of it.
863 	 */
864 	if (snapshot && (nr_pages & 1))
865 		return NULL;
866 
867 	if (cpu == -1)
868 		cpu = raw_smp_processor_id();
869 
870 	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
871 	if (!buf)
872 		return NULL;
873 
874 	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
875 	if (!pglist)
876 		goto out_free_buf;
877 
878 	for (i = 0; i < nr_pages; ++i)
879 		pglist[i] = virt_to_page(pages[i]);
880 
881 	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
882 	if (!buf->base)
883 		goto out_free_pglist;
884 
885 	buf->nr_pages	= nr_pages;
886 	buf->snapshot	= snapshot;
887 
888 	kfree(pglist);
889 	return buf;
890 
891 out_free_pglist:
892 	kfree(pglist);
893 out_free_buf:
894 	kfree(buf);
895 	return NULL;
896 }
897 
898 static void arm_spe_pmu_free_aux(void *aux)
899 {
900 	struct arm_spe_pmu_buf *buf = aux;
901 
902 	vunmap(buf->base);
903 	kfree(buf);
904 }
905 
906 /* Initialisation and teardown functions */
907 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
908 {
909 	static atomic_t pmu_idx = ATOMIC_INIT(-1);
910 
911 	int idx;
912 	char *name;
913 	struct device *dev = &spe_pmu->pdev->dev;
914 
915 	spe_pmu->pmu = (struct pmu) {
916 		.module = THIS_MODULE,
917 		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
918 		.attr_groups	= arm_spe_pmu_attr_groups,
919 		/*
920 		 * We hitch a ride on the software context here, so that
921 		 * we can support per-task profiling (which is not possible
922 		 * with the invalid context as it doesn't get sched callbacks).
923 		 * This requires that userspace either uses a dummy event for
924 		 * perf_event_open, since the aux buffer is not setup until
925 		 * a subsequent mmap, or creates the profiling event in a
926 		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
927 		 * once the buffer has been created.
928 		 */
929 		.task_ctx_nr	= perf_sw_context,
930 		.event_init	= arm_spe_pmu_event_init,
931 		.add		= arm_spe_pmu_add,
932 		.del		= arm_spe_pmu_del,
933 		.start		= arm_spe_pmu_start,
934 		.stop		= arm_spe_pmu_stop,
935 		.read		= arm_spe_pmu_read,
936 		.setup_aux	= arm_spe_pmu_setup_aux,
937 		.free_aux	= arm_spe_pmu_free_aux,
938 	};
939 
940 	idx = atomic_inc_return(&pmu_idx);
941 	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
942 	if (!name) {
943 		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
944 		return -ENOMEM;
945 	}
946 
947 	return perf_pmu_register(&spe_pmu->pmu, name, -1);
948 }
949 
950 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
951 {
952 	perf_pmu_unregister(&spe_pmu->pmu);
953 }
954 
955 static void __arm_spe_pmu_dev_probe(void *info)
956 {
957 	int fld;
958 	u64 reg;
959 	struct arm_spe_pmu *spe_pmu = info;
960 	struct device *dev = &spe_pmu->pdev->dev;
961 
962 	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
963 						   ID_AA64DFR0_EL1_PMSVer_SHIFT);
964 	if (!fld) {
965 		dev_err(dev,
966 			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
967 			fld, smp_processor_id());
968 		return;
969 	}
970 	spe_pmu->pmsver = (u16)fld;
971 
972 	/* Read PMBIDR first to determine whether or not we have access */
973 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
974 	if (reg & BIT(SYS_PMBIDR_EL1_P_SHIFT)) {
975 		dev_err(dev,
976 			"profiling buffer owned by higher exception level\n");
977 		return;
978 	}
979 
980 	/* Minimum alignment. If it's out-of-range, then fail the probe */
981 	fld = reg >> SYS_PMBIDR_EL1_ALIGN_SHIFT & SYS_PMBIDR_EL1_ALIGN_MASK;
982 	spe_pmu->align = 1 << fld;
983 	if (spe_pmu->align > SZ_2K) {
984 		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
985 			fld, smp_processor_id());
986 		return;
987 	}
988 
989 	/* It's now safe to read PMSIDR and figure out what we've got */
990 	reg = read_sysreg_s(SYS_PMSIDR_EL1);
991 	if (reg & BIT(SYS_PMSIDR_EL1_FE_SHIFT))
992 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
993 
994 	if (reg & BIT(SYS_PMSIDR_EL1_FT_SHIFT))
995 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
996 
997 	if (reg & BIT(SYS_PMSIDR_EL1_FL_SHIFT))
998 		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
999 
1000 	if (reg & BIT(SYS_PMSIDR_EL1_ARCHINST_SHIFT))
1001 		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
1002 
1003 	if (reg & BIT(SYS_PMSIDR_EL1_LDS_SHIFT))
1004 		spe_pmu->features |= SPE_PMU_FEAT_LDS;
1005 
1006 	if (reg & BIT(SYS_PMSIDR_EL1_ERND_SHIFT))
1007 		spe_pmu->features |= SPE_PMU_FEAT_ERND;
1008 
1009 	/* This field has a spaced out encoding, so just use a look-up */
1010 	fld = reg >> SYS_PMSIDR_EL1_INTERVAL_SHIFT & SYS_PMSIDR_EL1_INTERVAL_MASK;
1011 	switch (fld) {
1012 	case 0:
1013 		spe_pmu->min_period = 256;
1014 		break;
1015 	case 2:
1016 		spe_pmu->min_period = 512;
1017 		break;
1018 	case 3:
1019 		spe_pmu->min_period = 768;
1020 		break;
1021 	case 4:
1022 		spe_pmu->min_period = 1024;
1023 		break;
1024 	case 5:
1025 		spe_pmu->min_period = 1536;
1026 		break;
1027 	case 6:
1028 		spe_pmu->min_period = 2048;
1029 		break;
1030 	case 7:
1031 		spe_pmu->min_period = 3072;
1032 		break;
1033 	default:
1034 		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1035 			 fld);
1036 		fallthrough;
1037 	case 8:
1038 		spe_pmu->min_period = 4096;
1039 	}
1040 
1041 	/* Maximum record size. If it's out-of-range, then fail the probe */
1042 	fld = reg >> SYS_PMSIDR_EL1_MAXSIZE_SHIFT & SYS_PMSIDR_EL1_MAXSIZE_MASK;
1043 	spe_pmu->max_record_sz = 1 << fld;
1044 	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1045 		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1046 			fld, smp_processor_id());
1047 		return;
1048 	}
1049 
1050 	fld = reg >> SYS_PMSIDR_EL1_COUNTSIZE_SHIFT & SYS_PMSIDR_EL1_COUNTSIZE_MASK;
1051 	switch (fld) {
1052 	default:
1053 		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1054 			 fld);
1055 		fallthrough;
1056 	case 2:
1057 		spe_pmu->counter_sz = 12;
1058 		break;
1059 	case 3:
1060 		spe_pmu->counter_sz = 16;
1061 	}
1062 
1063 	dev_info(dev,
1064 		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1065 		 cpumask_pr_args(&spe_pmu->supported_cpus),
1066 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1067 
1068 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1069 }
1070 
1071 static void __arm_spe_pmu_reset_local(void)
1072 {
1073 	/*
1074 	 * This is probably overkill, as we have no idea where we're
1075 	 * draining any buffered data to...
1076 	 */
1077 	arm_spe_pmu_disable_and_drain_local();
1078 
1079 	/* Reset the buffer base pointer */
1080 	write_sysreg_s(0, SYS_PMBPTR_EL1);
1081 	isb();
1082 
1083 	/* Clear any pending management interrupts */
1084 	write_sysreg_s(0, SYS_PMBSR_EL1);
1085 	isb();
1086 }
1087 
1088 static void __arm_spe_pmu_setup_one(void *info)
1089 {
1090 	struct arm_spe_pmu *spe_pmu = info;
1091 
1092 	__arm_spe_pmu_reset_local();
1093 	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1094 }
1095 
1096 static void __arm_spe_pmu_stop_one(void *info)
1097 {
1098 	struct arm_spe_pmu *spe_pmu = info;
1099 
1100 	disable_percpu_irq(spe_pmu->irq);
1101 	__arm_spe_pmu_reset_local();
1102 }
1103 
1104 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1105 {
1106 	struct arm_spe_pmu *spe_pmu;
1107 
1108 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1109 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1110 		return 0;
1111 
1112 	__arm_spe_pmu_setup_one(spe_pmu);
1113 	return 0;
1114 }
1115 
1116 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1117 {
1118 	struct arm_spe_pmu *spe_pmu;
1119 
1120 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1121 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1122 		return 0;
1123 
1124 	__arm_spe_pmu_stop_one(spe_pmu);
1125 	return 0;
1126 }
1127 
1128 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1129 {
1130 	int ret;
1131 	cpumask_t *mask = &spe_pmu->supported_cpus;
1132 
1133 	/* Make sure we probe the hardware on a relevant CPU */
1134 	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1135 	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1136 		return -ENXIO;
1137 
1138 	/* Request our PPIs (note that the IRQ is still disabled) */
1139 	ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1140 				 spe_pmu->handle);
1141 	if (ret)
1142 		return ret;
1143 
1144 	/*
1145 	 * Register our hotplug notifier now so we don't miss any events.
1146 	 * This will enable the IRQ for any supported CPUs that are already
1147 	 * up.
1148 	 */
1149 	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1150 				       &spe_pmu->hotplug_node);
1151 	if (ret)
1152 		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1153 
1154 	return ret;
1155 }
1156 
1157 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1158 {
1159 	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1160 	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1161 }
1162 
1163 /* Driver and device probing */
1164 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1165 {
1166 	struct platform_device *pdev = spe_pmu->pdev;
1167 	int irq = platform_get_irq(pdev, 0);
1168 
1169 	if (irq < 0)
1170 		return -ENXIO;
1171 
1172 	if (!irq_is_percpu(irq)) {
1173 		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1174 		return -EINVAL;
1175 	}
1176 
1177 	if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1178 		dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1179 		return -EINVAL;
1180 	}
1181 
1182 	spe_pmu->irq = irq;
1183 	return 0;
1184 }
1185 
1186 static const struct of_device_id arm_spe_pmu_of_match[] = {
1187 	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1188 	{ /* Sentinel */ },
1189 };
1190 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1191 
1192 static const struct platform_device_id arm_spe_match[] = {
1193 	{ ARMV8_SPE_PDEV_NAME, 0},
1194 	{ }
1195 };
1196 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1197 
1198 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1199 {
1200 	int ret;
1201 	struct arm_spe_pmu *spe_pmu;
1202 	struct device *dev = &pdev->dev;
1203 
1204 	/*
1205 	 * If kernelspace is unmapped when running at EL0, then the SPE
1206 	 * buffer will fault and prematurely terminate the AUX session.
1207 	 */
1208 	if (arm64_kernel_unmapped_at_el0()) {
1209 		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1210 		return -EPERM;
1211 	}
1212 
1213 	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1214 	if (!spe_pmu)
1215 		return -ENOMEM;
1216 
1217 	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1218 	if (!spe_pmu->handle)
1219 		return -ENOMEM;
1220 
1221 	spe_pmu->pdev = pdev;
1222 	platform_set_drvdata(pdev, spe_pmu);
1223 
1224 	ret = arm_spe_pmu_irq_probe(spe_pmu);
1225 	if (ret)
1226 		goto out_free_handle;
1227 
1228 	ret = arm_spe_pmu_dev_init(spe_pmu);
1229 	if (ret)
1230 		goto out_free_handle;
1231 
1232 	ret = arm_spe_pmu_perf_init(spe_pmu);
1233 	if (ret)
1234 		goto out_teardown_dev;
1235 
1236 	return 0;
1237 
1238 out_teardown_dev:
1239 	arm_spe_pmu_dev_teardown(spe_pmu);
1240 out_free_handle:
1241 	free_percpu(spe_pmu->handle);
1242 	return ret;
1243 }
1244 
1245 static int arm_spe_pmu_device_remove(struct platform_device *pdev)
1246 {
1247 	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1248 
1249 	arm_spe_pmu_perf_destroy(spe_pmu);
1250 	arm_spe_pmu_dev_teardown(spe_pmu);
1251 	free_percpu(spe_pmu->handle);
1252 	return 0;
1253 }
1254 
1255 static struct platform_driver arm_spe_pmu_driver = {
1256 	.id_table = arm_spe_match,
1257 	.driver	= {
1258 		.name		= DRVNAME,
1259 		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1260 		.suppress_bind_attrs = true,
1261 	},
1262 	.probe	= arm_spe_pmu_device_probe,
1263 	.remove	= arm_spe_pmu_device_remove,
1264 };
1265 
1266 static int __init arm_spe_pmu_init(void)
1267 {
1268 	int ret;
1269 
1270 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1271 				      arm_spe_pmu_cpu_startup,
1272 				      arm_spe_pmu_cpu_teardown);
1273 	if (ret < 0)
1274 		return ret;
1275 	arm_spe_pmu_online = ret;
1276 
1277 	ret = platform_driver_register(&arm_spe_pmu_driver);
1278 	if (ret)
1279 		cpuhp_remove_multi_state(arm_spe_pmu_online);
1280 
1281 	return ret;
1282 }
1283 
1284 static void __exit arm_spe_pmu_exit(void)
1285 {
1286 	platform_driver_unregister(&arm_spe_pmu_driver);
1287 	cpuhp_remove_multi_state(arm_spe_pmu_online);
1288 }
1289 
1290 module_init(arm_spe_pmu_init);
1291 module_exit(arm_spe_pmu_exit);
1292 
1293 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1294 MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1295 MODULE_LICENSE("GPL v2");
1296