xref: /linux/drivers/perf/arm_spe_pmu.c (revision aec499c75cf8e0b599be4d559e6922b613085f8f)
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 #define ARM_SPE_BUF_PAD_BYTE			0
43 
44 struct arm_spe_pmu_buf {
45 	int					nr_pages;
46 	bool					snapshot;
47 	void					*base;
48 };
49 
50 struct arm_spe_pmu {
51 	struct pmu				pmu;
52 	struct platform_device			*pdev;
53 	cpumask_t				supported_cpus;
54 	struct hlist_node			hotplug_node;
55 
56 	int					irq; /* PPI */
57 	u16					pmsver;
58 	u16					min_period;
59 	u16					counter_sz;
60 
61 #define SPE_PMU_FEAT_FILT_EVT			(1UL << 0)
62 #define SPE_PMU_FEAT_FILT_TYP			(1UL << 1)
63 #define SPE_PMU_FEAT_FILT_LAT			(1UL << 2)
64 #define SPE_PMU_FEAT_ARCH_INST			(1UL << 3)
65 #define SPE_PMU_FEAT_LDS			(1UL << 4)
66 #define SPE_PMU_FEAT_ERND			(1UL << 5)
67 #define SPE_PMU_FEAT_DEV_PROBED			(1UL << 63)
68 	u64					features;
69 
70 	u16					max_record_sz;
71 	u16					align;
72 	struct perf_output_handle __percpu	*handle;
73 };
74 
75 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
76 
77 /* Convert a free-running index from perf into an SPE buffer offset */
78 #define PERF_IDX2OFF(idx, buf)	((idx) % ((buf)->nr_pages << PAGE_SHIFT))
79 
80 /* Keep track of our dynamic hotplug state */
81 static enum cpuhp_state arm_spe_pmu_online;
82 
83 enum arm_spe_pmu_buf_fault_action {
84 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
85 	SPE_PMU_BUF_FAULT_ACT_FATAL,
86 	SPE_PMU_BUF_FAULT_ACT_OK,
87 };
88 
89 /* This sysfs gunk was really good fun to write. */
90 enum arm_spe_pmu_capabilities {
91 	SPE_PMU_CAP_ARCH_INST = 0,
92 	SPE_PMU_CAP_ERND,
93 	SPE_PMU_CAP_FEAT_MAX,
94 	SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
95 	SPE_PMU_CAP_MIN_IVAL,
96 };
97 
98 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
99 	[SPE_PMU_CAP_ARCH_INST]	= SPE_PMU_FEAT_ARCH_INST,
100 	[SPE_PMU_CAP_ERND]	= SPE_PMU_FEAT_ERND,
101 };
102 
103 static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
104 {
105 	if (cap < SPE_PMU_CAP_FEAT_MAX)
106 		return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
107 
108 	switch (cap) {
109 	case SPE_PMU_CAP_CNT_SZ:
110 		return spe_pmu->counter_sz;
111 	case SPE_PMU_CAP_MIN_IVAL:
112 		return spe_pmu->min_period;
113 	default:
114 		WARN(1, "unknown cap %d\n", cap);
115 	}
116 
117 	return 0;
118 }
119 
120 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
121 				    struct device_attribute *attr,
122 				    char *buf)
123 {
124 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
125 	struct dev_ext_attribute *ea =
126 		container_of(attr, struct dev_ext_attribute, attr);
127 	int cap = (long)ea->var;
128 
129 	return sysfs_emit(buf, "%u\n", arm_spe_pmu_cap_get(spe_pmu, cap));
130 }
131 
132 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
133 	&((struct dev_ext_attribute[]) {				\
134 		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
135 	})[0].attr.attr
136 
137 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
138 	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
139 
140 static struct attribute *arm_spe_pmu_cap_attr[] = {
141 	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
142 	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
143 	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
144 	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
145 	NULL,
146 };
147 
148 static const struct attribute_group arm_spe_pmu_cap_group = {
149 	.name	= "caps",
150 	.attrs	= arm_spe_pmu_cap_attr,
151 };
152 
153 /* User ABI */
154 #define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
155 #define ATTR_CFG_FLD_ts_enable_LO		0
156 #define ATTR_CFG_FLD_ts_enable_HI		0
157 #define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
158 #define ATTR_CFG_FLD_pa_enable_LO		1
159 #define ATTR_CFG_FLD_pa_enable_HI		1
160 #define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
161 #define ATTR_CFG_FLD_pct_enable_LO		2
162 #define ATTR_CFG_FLD_pct_enable_HI		2
163 #define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
164 #define ATTR_CFG_FLD_jitter_LO			16
165 #define ATTR_CFG_FLD_jitter_HI			16
166 #define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
167 #define ATTR_CFG_FLD_branch_filter_LO		32
168 #define ATTR_CFG_FLD_branch_filter_HI		32
169 #define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
170 #define ATTR_CFG_FLD_load_filter_LO		33
171 #define ATTR_CFG_FLD_load_filter_HI		33
172 #define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
173 #define ATTR_CFG_FLD_store_filter_LO		34
174 #define ATTR_CFG_FLD_store_filter_HI		34
175 
176 #define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
177 #define ATTR_CFG_FLD_event_filter_LO		0
178 #define ATTR_CFG_FLD_event_filter_HI		63
179 
180 #define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
181 #define ATTR_CFG_FLD_min_latency_LO		0
182 #define ATTR_CFG_FLD_min_latency_HI		11
183 
184 /* Why does everything I do descend into this? */
185 #define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
186 	(lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
187 
188 #define _GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
189 	__GEN_PMU_FORMAT_ATTR(cfg, lo, hi)
190 
191 #define GEN_PMU_FORMAT_ATTR(name)					\
192 	PMU_FORMAT_ATTR(name,						\
193 	_GEN_PMU_FORMAT_ATTR(ATTR_CFG_FLD_##name##_CFG,			\
194 			     ATTR_CFG_FLD_##name##_LO,			\
195 			     ATTR_CFG_FLD_##name##_HI))
196 
197 #define _ATTR_CFG_GET_FLD(attr, cfg, lo, hi)				\
198 	((((attr)->cfg) >> lo) & GENMASK(hi - lo, 0))
199 
200 #define ATTR_CFG_GET_FLD(attr, name)					\
201 	_ATTR_CFG_GET_FLD(attr,						\
202 			  ATTR_CFG_FLD_##name##_CFG,			\
203 			  ATTR_CFG_FLD_##name##_LO,			\
204 			  ATTR_CFG_FLD_##name##_HI)
205 
206 GEN_PMU_FORMAT_ATTR(ts_enable);
207 GEN_PMU_FORMAT_ATTR(pa_enable);
208 GEN_PMU_FORMAT_ATTR(pct_enable);
209 GEN_PMU_FORMAT_ATTR(jitter);
210 GEN_PMU_FORMAT_ATTR(branch_filter);
211 GEN_PMU_FORMAT_ATTR(load_filter);
212 GEN_PMU_FORMAT_ATTR(store_filter);
213 GEN_PMU_FORMAT_ATTR(event_filter);
214 GEN_PMU_FORMAT_ATTR(min_latency);
215 
216 static struct attribute *arm_spe_pmu_formats_attr[] = {
217 	&format_attr_ts_enable.attr,
218 	&format_attr_pa_enable.attr,
219 	&format_attr_pct_enable.attr,
220 	&format_attr_jitter.attr,
221 	&format_attr_branch_filter.attr,
222 	&format_attr_load_filter.attr,
223 	&format_attr_store_filter.attr,
224 	&format_attr_event_filter.attr,
225 	&format_attr_min_latency.attr,
226 	NULL,
227 };
228 
229 static const struct attribute_group arm_spe_pmu_format_group = {
230 	.name	= "format",
231 	.attrs	= arm_spe_pmu_formats_attr,
232 };
233 
234 static ssize_t cpumask_show(struct device *dev,
235 			    struct device_attribute *attr, char *buf)
236 {
237 	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
238 
239 	return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
240 }
241 static DEVICE_ATTR_RO(cpumask);
242 
243 static struct attribute *arm_spe_pmu_attrs[] = {
244 	&dev_attr_cpumask.attr,
245 	NULL,
246 };
247 
248 static const struct attribute_group arm_spe_pmu_group = {
249 	.attrs	= arm_spe_pmu_attrs,
250 };
251 
252 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
253 	&arm_spe_pmu_group,
254 	&arm_spe_pmu_cap_group,
255 	&arm_spe_pmu_format_group,
256 	NULL,
257 };
258 
259 /* Convert between user ABI and register values */
260 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
261 {
262 	struct perf_event_attr *attr = &event->attr;
263 	u64 reg = 0;
264 
265 	reg |= ATTR_CFG_GET_FLD(attr, ts_enable) << SYS_PMSCR_EL1_TS_SHIFT;
266 	reg |= ATTR_CFG_GET_FLD(attr, pa_enable) << SYS_PMSCR_EL1_PA_SHIFT;
267 	reg |= ATTR_CFG_GET_FLD(attr, pct_enable) << SYS_PMSCR_EL1_PCT_SHIFT;
268 
269 	if (!attr->exclude_user)
270 		reg |= BIT(SYS_PMSCR_EL1_E0SPE_SHIFT);
271 
272 	if (!attr->exclude_kernel)
273 		reg |= BIT(SYS_PMSCR_EL1_E1SPE_SHIFT);
274 
275 	if (IS_ENABLED(CONFIG_PID_IN_CONTEXTIDR) && perfmon_capable())
276 		reg |= BIT(SYS_PMSCR_EL1_CX_SHIFT);
277 
278 	return reg;
279 }
280 
281 static void arm_spe_event_sanitise_period(struct perf_event *event)
282 {
283 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
284 	u64 period = event->hw.sample_period;
285 	u64 max_period = SYS_PMSIRR_EL1_INTERVAL_MASK
286 			 << SYS_PMSIRR_EL1_INTERVAL_SHIFT;
287 
288 	if (period < spe_pmu->min_period)
289 		period = spe_pmu->min_period;
290 	else if (period > max_period)
291 		period = max_period;
292 	else
293 		period &= max_period;
294 
295 	event->hw.sample_period = period;
296 }
297 
298 static u64 arm_spe_event_to_pmsirr(struct perf_event *event)
299 {
300 	struct perf_event_attr *attr = &event->attr;
301 	u64 reg = 0;
302 
303 	arm_spe_event_sanitise_period(event);
304 
305 	reg |= ATTR_CFG_GET_FLD(attr, jitter) << SYS_PMSIRR_EL1_RND_SHIFT;
306 	reg |= event->hw.sample_period;
307 
308 	return reg;
309 }
310 
311 static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
312 {
313 	struct perf_event_attr *attr = &event->attr;
314 	u64 reg = 0;
315 
316 	reg |= ATTR_CFG_GET_FLD(attr, load_filter) << SYS_PMSFCR_EL1_LD_SHIFT;
317 	reg |= ATTR_CFG_GET_FLD(attr, store_filter) << SYS_PMSFCR_EL1_ST_SHIFT;
318 	reg |= ATTR_CFG_GET_FLD(attr, branch_filter) << SYS_PMSFCR_EL1_B_SHIFT;
319 
320 	if (reg)
321 		reg |= BIT(SYS_PMSFCR_EL1_FT_SHIFT);
322 
323 	if (ATTR_CFG_GET_FLD(attr, event_filter))
324 		reg |= BIT(SYS_PMSFCR_EL1_FE_SHIFT);
325 
326 	if (ATTR_CFG_GET_FLD(attr, min_latency))
327 		reg |= BIT(SYS_PMSFCR_EL1_FL_SHIFT);
328 
329 	return reg;
330 }
331 
332 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
333 {
334 	struct perf_event_attr *attr = &event->attr;
335 	return ATTR_CFG_GET_FLD(attr, event_filter);
336 }
337 
338 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
339 {
340 	struct perf_event_attr *attr = &event->attr;
341 	return ATTR_CFG_GET_FLD(attr, min_latency)
342 	       << SYS_PMSLATFR_EL1_MINLAT_SHIFT;
343 }
344 
345 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
346 {
347 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
348 	u64 head = PERF_IDX2OFF(handle->head, buf);
349 
350 	memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
351 	if (!buf->snapshot)
352 		perf_aux_output_skip(handle, len);
353 }
354 
355 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
356 {
357 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
358 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
359 	u64 head = PERF_IDX2OFF(handle->head, buf);
360 	u64 limit = buf->nr_pages * PAGE_SIZE;
361 
362 	/*
363 	 * The trace format isn't parseable in reverse, so clamp
364 	 * the limit to half of the buffer size in snapshot mode
365 	 * so that the worst case is half a buffer of records, as
366 	 * opposed to a single record.
367 	 */
368 	if (head < limit >> 1)
369 		limit >>= 1;
370 
371 	/*
372 	 * If we're within max_record_sz of the limit, we must
373 	 * pad, move the head index and recompute the limit.
374 	 */
375 	if (limit - head < spe_pmu->max_record_sz) {
376 		arm_spe_pmu_pad_buf(handle, limit - head);
377 		handle->head = PERF_IDX2OFF(limit, buf);
378 		limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
379 	}
380 
381 	return limit;
382 }
383 
384 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
385 {
386 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
387 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
388 	const u64 bufsize = buf->nr_pages * PAGE_SIZE;
389 	u64 limit = bufsize;
390 	u64 head, tail, wakeup;
391 
392 	/*
393 	 * The head can be misaligned for two reasons:
394 	 *
395 	 * 1. The hardware left PMBPTR pointing to the first byte after
396 	 *    a record when generating a buffer management event.
397 	 *
398 	 * 2. We used perf_aux_output_skip to consume handle->size bytes
399 	 *    and CIRC_SPACE was used to compute the size, which always
400 	 *    leaves one entry free.
401 	 *
402 	 * Deal with this by padding to the next alignment boundary and
403 	 * moving the head index. If we run out of buffer space, we'll
404 	 * reduce handle->size to zero and end up reporting truncation.
405 	 */
406 	head = PERF_IDX2OFF(handle->head, buf);
407 	if (!IS_ALIGNED(head, spe_pmu->align)) {
408 		unsigned long delta = roundup(head, spe_pmu->align) - head;
409 
410 		delta = min(delta, handle->size);
411 		arm_spe_pmu_pad_buf(handle, delta);
412 		head = PERF_IDX2OFF(handle->head, buf);
413 	}
414 
415 	/* If we've run out of free space, then nothing more to do */
416 	if (!handle->size)
417 		goto no_space;
418 
419 	/* Compute the tail and wakeup indices now that we've aligned head */
420 	tail = PERF_IDX2OFF(handle->head + handle->size, buf);
421 	wakeup = PERF_IDX2OFF(handle->wakeup, buf);
422 
423 	/*
424 	 * Avoid clobbering unconsumed data. We know we have space, so
425 	 * if we see head == tail we know that the buffer is empty. If
426 	 * head > tail, then there's nothing to clobber prior to
427 	 * wrapping.
428 	 */
429 	if (head < tail)
430 		limit = round_down(tail, PAGE_SIZE);
431 
432 	/*
433 	 * Wakeup may be arbitrarily far into the future. If it's not in
434 	 * the current generation, either we'll wrap before hitting it,
435 	 * or it's in the past and has been handled already.
436 	 *
437 	 * If there's a wakeup before we wrap, arrange to be woken up by
438 	 * the page boundary following it. Keep the tail boundary if
439 	 * that's lower.
440 	 */
441 	if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
442 		limit = min(limit, round_up(wakeup, PAGE_SIZE));
443 
444 	if (limit > head)
445 		return limit;
446 
447 	arm_spe_pmu_pad_buf(handle, handle->size);
448 no_space:
449 	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
450 	perf_aux_output_end(handle, 0);
451 	return 0;
452 }
453 
454 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
455 {
456 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
457 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
458 	u64 limit = __arm_spe_pmu_next_off(handle);
459 	u64 head = PERF_IDX2OFF(handle->head, buf);
460 
461 	/*
462 	 * If the head has come too close to the end of the buffer,
463 	 * then pad to the end and recompute the limit.
464 	 */
465 	if (limit && (limit - head < spe_pmu->max_record_sz)) {
466 		arm_spe_pmu_pad_buf(handle, limit - head);
467 		limit = __arm_spe_pmu_next_off(handle);
468 	}
469 
470 	return limit;
471 }
472 
473 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
474 					  struct perf_event *event)
475 {
476 	u64 base, limit;
477 	struct arm_spe_pmu_buf *buf;
478 
479 	/* Start a new aux session */
480 	buf = perf_aux_output_begin(handle, event);
481 	if (!buf) {
482 		event->hw.state |= PERF_HES_STOPPED;
483 		/*
484 		 * We still need to clear the limit pointer, since the
485 		 * profiler might only be disabled by virtue of a fault.
486 		 */
487 		limit = 0;
488 		goto out_write_limit;
489 	}
490 
491 	limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
492 			      : arm_spe_pmu_next_off(handle);
493 	if (limit)
494 		limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT);
495 
496 	limit += (u64)buf->base;
497 	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
498 	write_sysreg_s(base, SYS_PMBPTR_EL1);
499 
500 out_write_limit:
501 	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
502 }
503 
504 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
505 {
506 	struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
507 	u64 offset, size;
508 
509 	offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
510 	size = offset - PERF_IDX2OFF(handle->head, buf);
511 
512 	if (buf->snapshot)
513 		handle->head = offset;
514 
515 	perf_aux_output_end(handle, size);
516 }
517 
518 static void arm_spe_pmu_disable_and_drain_local(void)
519 {
520 	/* Disable profiling at EL0 and EL1 */
521 	write_sysreg_s(0, SYS_PMSCR_EL1);
522 	isb();
523 
524 	/* Drain any buffered data */
525 	psb_csync();
526 	dsb(nsh);
527 
528 	/* Disable the profiling buffer */
529 	write_sysreg_s(0, SYS_PMBLIMITR_EL1);
530 	isb();
531 }
532 
533 /* IRQ handling */
534 static enum arm_spe_pmu_buf_fault_action
535 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
536 {
537 	const char *err_str;
538 	u64 pmbsr;
539 	enum arm_spe_pmu_buf_fault_action ret;
540 
541 	/*
542 	 * Ensure new profiling data is visible to the CPU and any external
543 	 * aborts have been resolved.
544 	 */
545 	psb_csync();
546 	dsb(nsh);
547 
548 	/* Ensure hardware updates to PMBPTR_EL1 are visible */
549 	isb();
550 
551 	/* Service required? */
552 	pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
553 	if (!(pmbsr & BIT(SYS_PMBSR_EL1_S_SHIFT)))
554 		return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
555 
556 	/*
557 	 * If we've lost data, disable profiling and also set the PARTIAL
558 	 * flag to indicate that the last record is corrupted.
559 	 */
560 	if (pmbsr & BIT(SYS_PMBSR_EL1_DL_SHIFT))
561 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
562 					     PERF_AUX_FLAG_PARTIAL);
563 
564 	/* Report collisions to userspace so that it can up the period */
565 	if (pmbsr & BIT(SYS_PMBSR_EL1_COLL_SHIFT))
566 		perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
567 
568 	/* We only expect buffer management events */
569 	switch (pmbsr & (SYS_PMBSR_EL1_EC_MASK << SYS_PMBSR_EL1_EC_SHIFT)) {
570 	case SYS_PMBSR_EL1_EC_BUF:
571 		/* Handled below */
572 		break;
573 	case SYS_PMBSR_EL1_EC_FAULT_S1:
574 	case SYS_PMBSR_EL1_EC_FAULT_S2:
575 		err_str = "Unexpected buffer fault";
576 		goto out_err;
577 	default:
578 		err_str = "Unknown error code";
579 		goto out_err;
580 	}
581 
582 	/* Buffer management event */
583 	switch (pmbsr &
584 		(SYS_PMBSR_EL1_BUF_BSC_MASK << SYS_PMBSR_EL1_BUF_BSC_SHIFT)) {
585 	case SYS_PMBSR_EL1_BUF_BSC_FULL:
586 		ret = SPE_PMU_BUF_FAULT_ACT_OK;
587 		goto out_stop;
588 	default:
589 		err_str = "Unknown buffer status code";
590 	}
591 
592 out_err:
593 	pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
594 			   err_str, smp_processor_id(), pmbsr,
595 			   read_sysreg_s(SYS_PMBPTR_EL1),
596 			   read_sysreg_s(SYS_PMBLIMITR_EL1));
597 	ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
598 
599 out_stop:
600 	arm_spe_perf_aux_output_end(handle);
601 	return ret;
602 }
603 
604 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
605 {
606 	struct perf_output_handle *handle = dev;
607 	struct perf_event *event = handle->event;
608 	enum arm_spe_pmu_buf_fault_action act;
609 
610 	if (!perf_get_aux(handle))
611 		return IRQ_NONE;
612 
613 	act = arm_spe_pmu_buf_get_fault_act(handle);
614 	if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
615 		return IRQ_NONE;
616 
617 	/*
618 	 * Ensure perf callbacks have completed, which may disable the
619 	 * profiling buffer in response to a TRUNCATION flag.
620 	 */
621 	irq_work_run();
622 
623 	switch (act) {
624 	case SPE_PMU_BUF_FAULT_ACT_FATAL:
625 		/*
626 		 * If a fatal exception occurred then leaving the profiling
627 		 * buffer enabled is a recipe waiting to happen. Since
628 		 * fatal faults don't always imply truncation, make sure
629 		 * that the profiling buffer is disabled explicitly before
630 		 * clearing the syndrome register.
631 		 */
632 		arm_spe_pmu_disable_and_drain_local();
633 		break;
634 	case SPE_PMU_BUF_FAULT_ACT_OK:
635 		/*
636 		 * We handled the fault (the buffer was full), so resume
637 		 * profiling as long as we didn't detect truncation.
638 		 * PMBPTR might be misaligned, but we'll burn that bridge
639 		 * when we get to it.
640 		 */
641 		if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
642 			arm_spe_perf_aux_output_begin(handle, event);
643 			isb();
644 		}
645 		break;
646 	case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
647 		/* We've seen you before, but GCC has the memory of a sieve. */
648 		break;
649 	}
650 
651 	/* The buffer pointers are now sane, so resume profiling. */
652 	write_sysreg_s(0, SYS_PMBSR_EL1);
653 	return IRQ_HANDLED;
654 }
655 
656 static u64 arm_spe_pmsevfr_res0(u16 pmsver)
657 {
658 	switch (pmsver) {
659 	case ID_AA64DFR0_PMSVER_8_2:
660 		return SYS_PMSEVFR_EL1_RES0_8_2;
661 	case ID_AA64DFR0_PMSVER_8_3:
662 	/* Return the highest version we support in default */
663 	default:
664 		return SYS_PMSEVFR_EL1_RES0_8_3;
665 	}
666 }
667 
668 /* Perf callbacks */
669 static int arm_spe_pmu_event_init(struct perf_event *event)
670 {
671 	u64 reg;
672 	struct perf_event_attr *attr = &event->attr;
673 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
674 
675 	/* This is, of course, deeply driver-specific */
676 	if (attr->type != event->pmu->type)
677 		return -ENOENT;
678 
679 	if (event->cpu >= 0 &&
680 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
681 		return -ENOENT;
682 
683 	if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
684 		return -EOPNOTSUPP;
685 
686 	if (attr->exclude_idle)
687 		return -EOPNOTSUPP;
688 
689 	/*
690 	 * Feedback-directed frequency throttling doesn't work when we
691 	 * have a buffer of samples. We'd need to manually count the
692 	 * samples in the buffer when it fills up and adjust the event
693 	 * count to reflect that. Instead, just force the user to specify
694 	 * a sample period.
695 	 */
696 	if (attr->freq)
697 		return -EINVAL;
698 
699 	reg = arm_spe_event_to_pmsfcr(event);
700 	if ((reg & BIT(SYS_PMSFCR_EL1_FE_SHIFT)) &&
701 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
702 		return -EOPNOTSUPP;
703 
704 	if ((reg & BIT(SYS_PMSFCR_EL1_FT_SHIFT)) &&
705 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
706 		return -EOPNOTSUPP;
707 
708 	if ((reg & BIT(SYS_PMSFCR_EL1_FL_SHIFT)) &&
709 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
710 		return -EOPNOTSUPP;
711 
712 	reg = arm_spe_event_to_pmscr(event);
713 	if (!perfmon_capable() &&
714 	    (reg & (BIT(SYS_PMSCR_EL1_PA_SHIFT) |
715 		    BIT(SYS_PMSCR_EL1_CX_SHIFT) |
716 		    BIT(SYS_PMSCR_EL1_PCT_SHIFT))))
717 		return -EACCES;
718 
719 	return 0;
720 }
721 
722 static void arm_spe_pmu_start(struct perf_event *event, int flags)
723 {
724 	u64 reg;
725 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
726 	struct hw_perf_event *hwc = &event->hw;
727 	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
728 
729 	hwc->state = 0;
730 	arm_spe_perf_aux_output_begin(handle, event);
731 	if (hwc->state)
732 		return;
733 
734 	reg = arm_spe_event_to_pmsfcr(event);
735 	write_sysreg_s(reg, SYS_PMSFCR_EL1);
736 
737 	reg = arm_spe_event_to_pmsevfr(event);
738 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
739 
740 	reg = arm_spe_event_to_pmslatfr(event);
741 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
742 
743 	if (flags & PERF_EF_RELOAD) {
744 		reg = arm_spe_event_to_pmsirr(event);
745 		write_sysreg_s(reg, SYS_PMSIRR_EL1);
746 		isb();
747 		reg = local64_read(&hwc->period_left);
748 		write_sysreg_s(reg, SYS_PMSICR_EL1);
749 	}
750 
751 	reg = arm_spe_event_to_pmscr(event);
752 	isb();
753 	write_sysreg_s(reg, SYS_PMSCR_EL1);
754 }
755 
756 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
757 {
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 	/* If we're already stopped, then nothing to do */
763 	if (hwc->state & PERF_HES_STOPPED)
764 		return;
765 
766 	/* Stop all trace generation */
767 	arm_spe_pmu_disable_and_drain_local();
768 
769 	if (flags & PERF_EF_UPDATE) {
770 		/*
771 		 * If there's a fault pending then ensure we contain it
772 		 * to this buffer, since we might be on the context-switch
773 		 * path.
774 		 */
775 		if (perf_get_aux(handle)) {
776 			enum arm_spe_pmu_buf_fault_action act;
777 
778 			act = arm_spe_pmu_buf_get_fault_act(handle);
779 			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
780 				arm_spe_perf_aux_output_end(handle);
781 			else
782 				write_sysreg_s(0, SYS_PMBSR_EL1);
783 		}
784 
785 		/*
786 		 * This may also contain ECOUNT, but nobody else should
787 		 * be looking at period_left, since we forbid frequency
788 		 * based sampling.
789 		 */
790 		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
791 		hwc->state |= PERF_HES_UPTODATE;
792 	}
793 
794 	hwc->state |= PERF_HES_STOPPED;
795 }
796 
797 static int arm_spe_pmu_add(struct perf_event *event, int flags)
798 {
799 	int ret = 0;
800 	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
801 	struct hw_perf_event *hwc = &event->hw;
802 	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
803 
804 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
805 		return -ENOENT;
806 
807 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
808 
809 	if (flags & PERF_EF_START) {
810 		arm_spe_pmu_start(event, PERF_EF_RELOAD);
811 		if (hwc->state & PERF_HES_STOPPED)
812 			ret = -EINVAL;
813 	}
814 
815 	return ret;
816 }
817 
818 static void arm_spe_pmu_del(struct perf_event *event, int flags)
819 {
820 	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
821 }
822 
823 static void arm_spe_pmu_read(struct perf_event *event)
824 {
825 }
826 
827 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
828 				   int nr_pages, bool snapshot)
829 {
830 	int i, cpu = event->cpu;
831 	struct page **pglist;
832 	struct arm_spe_pmu_buf *buf;
833 
834 	/* We need at least two pages for this to work. */
835 	if (nr_pages < 2)
836 		return NULL;
837 
838 	/*
839 	 * We require an even number of pages for snapshot mode, so that
840 	 * we can effectively treat the buffer as consisting of two equal
841 	 * parts and give userspace a fighting chance of getting some
842 	 * useful data out of it.
843 	 */
844 	if (snapshot && (nr_pages & 1))
845 		return NULL;
846 
847 	if (cpu == -1)
848 		cpu = raw_smp_processor_id();
849 
850 	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
851 	if (!buf)
852 		return NULL;
853 
854 	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
855 	if (!pglist)
856 		goto out_free_buf;
857 
858 	for (i = 0; i < nr_pages; ++i)
859 		pglist[i] = virt_to_page(pages[i]);
860 
861 	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
862 	if (!buf->base)
863 		goto out_free_pglist;
864 
865 	buf->nr_pages	= nr_pages;
866 	buf->snapshot	= snapshot;
867 
868 	kfree(pglist);
869 	return buf;
870 
871 out_free_pglist:
872 	kfree(pglist);
873 out_free_buf:
874 	kfree(buf);
875 	return NULL;
876 }
877 
878 static void arm_spe_pmu_free_aux(void *aux)
879 {
880 	struct arm_spe_pmu_buf *buf = aux;
881 
882 	vunmap(buf->base);
883 	kfree(buf);
884 }
885 
886 /* Initialisation and teardown functions */
887 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
888 {
889 	static atomic_t pmu_idx = ATOMIC_INIT(-1);
890 
891 	int idx;
892 	char *name;
893 	struct device *dev = &spe_pmu->pdev->dev;
894 
895 	spe_pmu->pmu = (struct pmu) {
896 		.module = THIS_MODULE,
897 		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
898 		.attr_groups	= arm_spe_pmu_attr_groups,
899 		/*
900 		 * We hitch a ride on the software context here, so that
901 		 * we can support per-task profiling (which is not possible
902 		 * with the invalid context as it doesn't get sched callbacks).
903 		 * This requires that userspace either uses a dummy event for
904 		 * perf_event_open, since the aux buffer is not setup until
905 		 * a subsequent mmap, or creates the profiling event in a
906 		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
907 		 * once the buffer has been created.
908 		 */
909 		.task_ctx_nr	= perf_sw_context,
910 		.event_init	= arm_spe_pmu_event_init,
911 		.add		= arm_spe_pmu_add,
912 		.del		= arm_spe_pmu_del,
913 		.start		= arm_spe_pmu_start,
914 		.stop		= arm_spe_pmu_stop,
915 		.read		= arm_spe_pmu_read,
916 		.setup_aux	= arm_spe_pmu_setup_aux,
917 		.free_aux	= arm_spe_pmu_free_aux,
918 	};
919 
920 	idx = atomic_inc_return(&pmu_idx);
921 	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
922 	if (!name) {
923 		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
924 		return -ENOMEM;
925 	}
926 
927 	return perf_pmu_register(&spe_pmu->pmu, name, -1);
928 }
929 
930 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
931 {
932 	perf_pmu_unregister(&spe_pmu->pmu);
933 }
934 
935 static void __arm_spe_pmu_dev_probe(void *info)
936 {
937 	int fld;
938 	u64 reg;
939 	struct arm_spe_pmu *spe_pmu = info;
940 	struct device *dev = &spe_pmu->pdev->dev;
941 
942 	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
943 						   ID_AA64DFR0_PMSVER_SHIFT);
944 	if (!fld) {
945 		dev_err(dev,
946 			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
947 			fld, smp_processor_id());
948 		return;
949 	}
950 	spe_pmu->pmsver = (u16)fld;
951 
952 	/* Read PMBIDR first to determine whether or not we have access */
953 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
954 	if (reg & BIT(SYS_PMBIDR_EL1_P_SHIFT)) {
955 		dev_err(dev,
956 			"profiling buffer owned by higher exception level\n");
957 		return;
958 	}
959 
960 	/* Minimum alignment. If it's out-of-range, then fail the probe */
961 	fld = reg >> SYS_PMBIDR_EL1_ALIGN_SHIFT & SYS_PMBIDR_EL1_ALIGN_MASK;
962 	spe_pmu->align = 1 << fld;
963 	if (spe_pmu->align > SZ_2K) {
964 		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
965 			fld, smp_processor_id());
966 		return;
967 	}
968 
969 	/* It's now safe to read PMSIDR and figure out what we've got */
970 	reg = read_sysreg_s(SYS_PMSIDR_EL1);
971 	if (reg & BIT(SYS_PMSIDR_EL1_FE_SHIFT))
972 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
973 
974 	if (reg & BIT(SYS_PMSIDR_EL1_FT_SHIFT))
975 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
976 
977 	if (reg & BIT(SYS_PMSIDR_EL1_FL_SHIFT))
978 		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
979 
980 	if (reg & BIT(SYS_PMSIDR_EL1_ARCHINST_SHIFT))
981 		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
982 
983 	if (reg & BIT(SYS_PMSIDR_EL1_LDS_SHIFT))
984 		spe_pmu->features |= SPE_PMU_FEAT_LDS;
985 
986 	if (reg & BIT(SYS_PMSIDR_EL1_ERND_SHIFT))
987 		spe_pmu->features |= SPE_PMU_FEAT_ERND;
988 
989 	/* This field has a spaced out encoding, so just use a look-up */
990 	fld = reg >> SYS_PMSIDR_EL1_INTERVAL_SHIFT & SYS_PMSIDR_EL1_INTERVAL_MASK;
991 	switch (fld) {
992 	case 0:
993 		spe_pmu->min_period = 256;
994 		break;
995 	case 2:
996 		spe_pmu->min_period = 512;
997 		break;
998 	case 3:
999 		spe_pmu->min_period = 768;
1000 		break;
1001 	case 4:
1002 		spe_pmu->min_period = 1024;
1003 		break;
1004 	case 5:
1005 		spe_pmu->min_period = 1536;
1006 		break;
1007 	case 6:
1008 		spe_pmu->min_period = 2048;
1009 		break;
1010 	case 7:
1011 		spe_pmu->min_period = 3072;
1012 		break;
1013 	default:
1014 		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1015 			 fld);
1016 		fallthrough;
1017 	case 8:
1018 		spe_pmu->min_period = 4096;
1019 	}
1020 
1021 	/* Maximum record size. If it's out-of-range, then fail the probe */
1022 	fld = reg >> SYS_PMSIDR_EL1_MAXSIZE_SHIFT & SYS_PMSIDR_EL1_MAXSIZE_MASK;
1023 	spe_pmu->max_record_sz = 1 << fld;
1024 	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1025 		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1026 			fld, smp_processor_id());
1027 		return;
1028 	}
1029 
1030 	fld = reg >> SYS_PMSIDR_EL1_COUNTSIZE_SHIFT & SYS_PMSIDR_EL1_COUNTSIZE_MASK;
1031 	switch (fld) {
1032 	default:
1033 		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1034 			 fld);
1035 		fallthrough;
1036 	case 2:
1037 		spe_pmu->counter_sz = 12;
1038 	}
1039 
1040 	dev_info(dev,
1041 		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1042 		 cpumask_pr_args(&spe_pmu->supported_cpus),
1043 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1044 
1045 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1046 }
1047 
1048 static void __arm_spe_pmu_reset_local(void)
1049 {
1050 	/*
1051 	 * This is probably overkill, as we have no idea where we're
1052 	 * draining any buffered data to...
1053 	 */
1054 	arm_spe_pmu_disable_and_drain_local();
1055 
1056 	/* Reset the buffer base pointer */
1057 	write_sysreg_s(0, SYS_PMBPTR_EL1);
1058 	isb();
1059 
1060 	/* Clear any pending management interrupts */
1061 	write_sysreg_s(0, SYS_PMBSR_EL1);
1062 	isb();
1063 }
1064 
1065 static void __arm_spe_pmu_setup_one(void *info)
1066 {
1067 	struct arm_spe_pmu *spe_pmu = info;
1068 
1069 	__arm_spe_pmu_reset_local();
1070 	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1071 }
1072 
1073 static void __arm_spe_pmu_stop_one(void *info)
1074 {
1075 	struct arm_spe_pmu *spe_pmu = info;
1076 
1077 	disable_percpu_irq(spe_pmu->irq);
1078 	__arm_spe_pmu_reset_local();
1079 }
1080 
1081 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1082 {
1083 	struct arm_spe_pmu *spe_pmu;
1084 
1085 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1086 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1087 		return 0;
1088 
1089 	__arm_spe_pmu_setup_one(spe_pmu);
1090 	return 0;
1091 }
1092 
1093 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1094 {
1095 	struct arm_spe_pmu *spe_pmu;
1096 
1097 	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1098 	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1099 		return 0;
1100 
1101 	__arm_spe_pmu_stop_one(spe_pmu);
1102 	return 0;
1103 }
1104 
1105 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1106 {
1107 	int ret;
1108 	cpumask_t *mask = &spe_pmu->supported_cpus;
1109 
1110 	/* Make sure we probe the hardware on a relevant CPU */
1111 	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1112 	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1113 		return -ENXIO;
1114 
1115 	/* Request our PPIs (note that the IRQ is still disabled) */
1116 	ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1117 				 spe_pmu->handle);
1118 	if (ret)
1119 		return ret;
1120 
1121 	/*
1122 	 * Register our hotplug notifier now so we don't miss any events.
1123 	 * This will enable the IRQ for any supported CPUs that are already
1124 	 * up.
1125 	 */
1126 	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1127 				       &spe_pmu->hotplug_node);
1128 	if (ret)
1129 		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1130 
1131 	return ret;
1132 }
1133 
1134 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1135 {
1136 	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1137 	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1138 }
1139 
1140 /* Driver and device probing */
1141 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1142 {
1143 	struct platform_device *pdev = spe_pmu->pdev;
1144 	int irq = platform_get_irq(pdev, 0);
1145 
1146 	if (irq < 0)
1147 		return -ENXIO;
1148 
1149 	if (!irq_is_percpu(irq)) {
1150 		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1151 		return -EINVAL;
1152 	}
1153 
1154 	if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1155 		dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1156 		return -EINVAL;
1157 	}
1158 
1159 	spe_pmu->irq = irq;
1160 	return 0;
1161 }
1162 
1163 static const struct of_device_id arm_spe_pmu_of_match[] = {
1164 	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1165 	{ /* Sentinel */ },
1166 };
1167 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1168 
1169 static const struct platform_device_id arm_spe_match[] = {
1170 	{ ARMV8_SPE_PDEV_NAME, 0},
1171 	{ }
1172 };
1173 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1174 
1175 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1176 {
1177 	int ret;
1178 	struct arm_spe_pmu *spe_pmu;
1179 	struct device *dev = &pdev->dev;
1180 
1181 	/*
1182 	 * If kernelspace is unmapped when running at EL0, then the SPE
1183 	 * buffer will fault and prematurely terminate the AUX session.
1184 	 */
1185 	if (arm64_kernel_unmapped_at_el0()) {
1186 		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1187 		return -EPERM;
1188 	}
1189 
1190 	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1191 	if (!spe_pmu)
1192 		return -ENOMEM;
1193 
1194 	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1195 	if (!spe_pmu->handle)
1196 		return -ENOMEM;
1197 
1198 	spe_pmu->pdev = pdev;
1199 	platform_set_drvdata(pdev, spe_pmu);
1200 
1201 	ret = arm_spe_pmu_irq_probe(spe_pmu);
1202 	if (ret)
1203 		goto out_free_handle;
1204 
1205 	ret = arm_spe_pmu_dev_init(spe_pmu);
1206 	if (ret)
1207 		goto out_free_handle;
1208 
1209 	ret = arm_spe_pmu_perf_init(spe_pmu);
1210 	if (ret)
1211 		goto out_teardown_dev;
1212 
1213 	return 0;
1214 
1215 out_teardown_dev:
1216 	arm_spe_pmu_dev_teardown(spe_pmu);
1217 out_free_handle:
1218 	free_percpu(spe_pmu->handle);
1219 	return ret;
1220 }
1221 
1222 static int arm_spe_pmu_device_remove(struct platform_device *pdev)
1223 {
1224 	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1225 
1226 	arm_spe_pmu_perf_destroy(spe_pmu);
1227 	arm_spe_pmu_dev_teardown(spe_pmu);
1228 	free_percpu(spe_pmu->handle);
1229 	return 0;
1230 }
1231 
1232 static struct platform_driver arm_spe_pmu_driver = {
1233 	.id_table = arm_spe_match,
1234 	.driver	= {
1235 		.name		= DRVNAME,
1236 		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1237 		.suppress_bind_attrs = true,
1238 	},
1239 	.probe	= arm_spe_pmu_device_probe,
1240 	.remove	= arm_spe_pmu_device_remove,
1241 };
1242 
1243 static int __init arm_spe_pmu_init(void)
1244 {
1245 	int ret;
1246 
1247 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1248 				      arm_spe_pmu_cpu_startup,
1249 				      arm_spe_pmu_cpu_teardown);
1250 	if (ret < 0)
1251 		return ret;
1252 	arm_spe_pmu_online = ret;
1253 
1254 	ret = platform_driver_register(&arm_spe_pmu_driver);
1255 	if (ret)
1256 		cpuhp_remove_multi_state(arm_spe_pmu_online);
1257 
1258 	return ret;
1259 }
1260 
1261 static void __exit arm_spe_pmu_exit(void)
1262 {
1263 	platform_driver_unregister(&arm_spe_pmu_driver);
1264 	cpuhp_remove_multi_state(arm_spe_pmu_online);
1265 }
1266 
1267 module_init(arm_spe_pmu_init);
1268 module_exit(arm_spe_pmu_exit);
1269 
1270 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1271 MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1272 MODULE_LICENSE("GPL v2");
1273