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
49static 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
55static 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
62struct arm_spe_pmu_buf {
63	int					nr_pages;
64	bool					snapshot;
65	void					*base;
66};
67
68struct 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
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 */
99static enum cpuhp_state arm_spe_pmu_online;
100
101enum 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. */
108enum 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
116static 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
121static 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
138static 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 snprintf(buf, PAGE_SIZE, "%u\n",
148		arm_spe_pmu_cap_get(spe_pmu, cap));
149}
150
151#define SPE_EXT_ATTR_ENTRY(_name, _func, _var)				\
152	&((struct dev_ext_attribute[]) {				\
153		{ __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }	\
154	})[0].attr.attr
155
156#define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)				\
157	SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
158
159static struct attribute *arm_spe_pmu_cap_attr[] = {
160	SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
161	SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
162	SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
163	SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
164	NULL,
165};
166
167static struct attribute_group arm_spe_pmu_cap_group = {
168	.name	= "caps",
169	.attrs	= arm_spe_pmu_cap_attr,
170};
171
172/* User ABI */
173#define ATTR_CFG_FLD_ts_enable_CFG		config	/* PMSCR_EL1.TS */
174#define ATTR_CFG_FLD_ts_enable_LO		0
175#define ATTR_CFG_FLD_ts_enable_HI		0
176#define ATTR_CFG_FLD_pa_enable_CFG		config	/* PMSCR_EL1.PA */
177#define ATTR_CFG_FLD_pa_enable_LO		1
178#define ATTR_CFG_FLD_pa_enable_HI		1
179#define ATTR_CFG_FLD_pct_enable_CFG		config	/* PMSCR_EL1.PCT */
180#define ATTR_CFG_FLD_pct_enable_LO		2
181#define ATTR_CFG_FLD_pct_enable_HI		2
182#define ATTR_CFG_FLD_jitter_CFG			config	/* PMSIRR_EL1.RND */
183#define ATTR_CFG_FLD_jitter_LO			16
184#define ATTR_CFG_FLD_jitter_HI			16
185#define ATTR_CFG_FLD_branch_filter_CFG		config	/* PMSFCR_EL1.B */
186#define ATTR_CFG_FLD_branch_filter_LO		32
187#define ATTR_CFG_FLD_branch_filter_HI		32
188#define ATTR_CFG_FLD_load_filter_CFG		config	/* PMSFCR_EL1.LD */
189#define ATTR_CFG_FLD_load_filter_LO		33
190#define ATTR_CFG_FLD_load_filter_HI		33
191#define ATTR_CFG_FLD_store_filter_CFG		config	/* PMSFCR_EL1.ST */
192#define ATTR_CFG_FLD_store_filter_LO		34
193#define ATTR_CFG_FLD_store_filter_HI		34
194
195#define ATTR_CFG_FLD_event_filter_CFG		config1	/* PMSEVFR_EL1 */
196#define ATTR_CFG_FLD_event_filter_LO		0
197#define ATTR_CFG_FLD_event_filter_HI		63
198
199#define ATTR_CFG_FLD_min_latency_CFG		config2	/* PMSLATFR_EL1.MINLAT */
200#define ATTR_CFG_FLD_min_latency_LO		0
201#define ATTR_CFG_FLD_min_latency_HI		11
202
203/* Why does everything I do descend into this? */
204#define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
205	(lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
206
207#define _GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
208	__GEN_PMU_FORMAT_ATTR(cfg, lo, hi)
209
210#define GEN_PMU_FORMAT_ATTR(name)					\
211	PMU_FORMAT_ATTR(name,						\
212	_GEN_PMU_FORMAT_ATTR(ATTR_CFG_FLD_##name##_CFG,			\
213			     ATTR_CFG_FLD_##name##_LO,			\
214			     ATTR_CFG_FLD_##name##_HI))
215
216#define _ATTR_CFG_GET_FLD(attr, cfg, lo, hi)				\
217	((((attr)->cfg) >> lo) & GENMASK(hi - lo, 0))
218
219#define ATTR_CFG_GET_FLD(attr, name)					\
220	_ATTR_CFG_GET_FLD(attr,						\
221			  ATTR_CFG_FLD_##name##_CFG,			\
222			  ATTR_CFG_FLD_##name##_LO,			\
223			  ATTR_CFG_FLD_##name##_HI)
224
225GEN_PMU_FORMAT_ATTR(ts_enable);
226GEN_PMU_FORMAT_ATTR(pa_enable);
227GEN_PMU_FORMAT_ATTR(pct_enable);
228GEN_PMU_FORMAT_ATTR(jitter);
229GEN_PMU_FORMAT_ATTR(branch_filter);
230GEN_PMU_FORMAT_ATTR(load_filter);
231GEN_PMU_FORMAT_ATTR(store_filter);
232GEN_PMU_FORMAT_ATTR(event_filter);
233GEN_PMU_FORMAT_ATTR(min_latency);
234
235static struct attribute *arm_spe_pmu_formats_attr[] = {
236	&format_attr_ts_enable.attr,
237	&format_attr_pa_enable.attr,
238	&format_attr_pct_enable.attr,
239	&format_attr_jitter.attr,
240	&format_attr_branch_filter.attr,
241	&format_attr_load_filter.attr,
242	&format_attr_store_filter.attr,
243	&format_attr_event_filter.attr,
244	&format_attr_min_latency.attr,
245	NULL,
246};
247
248static struct attribute_group arm_spe_pmu_format_group = {
249	.name	= "format",
250	.attrs	= arm_spe_pmu_formats_attr,
251};
252
253static ssize_t arm_spe_pmu_get_attr_cpumask(struct device *dev,
254					    struct device_attribute *attr,
255					    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}
261static DEVICE_ATTR(cpumask, S_IRUGO, arm_spe_pmu_get_attr_cpumask, NULL);
262
263static struct attribute *arm_spe_pmu_attrs[] = {
264	&dev_attr_cpumask.attr,
265	NULL,
266};
267
268static struct attribute_group arm_spe_pmu_group = {
269	.attrs	= arm_spe_pmu_attrs,
270};
271
272static 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 */
280static 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
301static 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
318static 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
331static 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
352static 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
358static 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
365static 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
375static 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
404static 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);
468no_space:
469	perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
470	perf_aux_output_end(handle, 0);
471	return 0;
472}
473
474static 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
493static 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
520out_write_limit:
521	write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
522}
523
524static 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
538static 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 */
554static enum arm_spe_pmu_buf_fault_action
555arm_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
612out_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
619out_stop:
620	arm_spe_perf_aux_output_end(handle);
621	return ret;
622}
623
624static 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/* Perf callbacks */
677static int arm_spe_pmu_event_init(struct perf_event *event)
678{
679	u64 reg;
680	struct perf_event_attr *attr = &event->attr;
681	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
682
683	/* This is, of course, deeply driver-specific */
684	if (attr->type != event->pmu->type)
685		return -ENOENT;
686
687	if (event->cpu >= 0 &&
688	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
689		return -ENOENT;
690
691	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
692		return -EOPNOTSUPP;
693
694	if (attr->exclude_idle)
695		return -EOPNOTSUPP;
696
697	/*
698	 * Feedback-directed frequency throttling doesn't work when we
699	 * have a buffer of samples. We'd need to manually count the
700	 * samples in the buffer when it fills up and adjust the event
701	 * count to reflect that. Instead, just force the user to specify
702	 * a sample period.
703	 */
704	if (attr->freq)
705		return -EINVAL;
706
707	reg = arm_spe_event_to_pmsfcr(event);
708	if ((reg & BIT(SYS_PMSFCR_EL1_FE_SHIFT)) &&
709	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
710		return -EOPNOTSUPP;
711
712	if ((reg & BIT(SYS_PMSFCR_EL1_FT_SHIFT)) &&
713	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
714		return -EOPNOTSUPP;
715
716	if ((reg & BIT(SYS_PMSFCR_EL1_FL_SHIFT)) &&
717	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
718		return -EOPNOTSUPP;
719
720	set_spe_event_has_cx(event);
721	reg = arm_spe_event_to_pmscr(event);
722	if (!perfmon_capable() &&
723	    (reg & (BIT(SYS_PMSCR_EL1_PA_SHIFT) |
724		    BIT(SYS_PMSCR_EL1_PCT_SHIFT))))
725		return -EACCES;
726
727	return 0;
728}
729
730static void arm_spe_pmu_start(struct perf_event *event, int flags)
731{
732	u64 reg;
733	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
734	struct hw_perf_event *hwc = &event->hw;
735	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
736
737	hwc->state = 0;
738	arm_spe_perf_aux_output_begin(handle, event);
739	if (hwc->state)
740		return;
741
742	reg = arm_spe_event_to_pmsfcr(event);
743	write_sysreg_s(reg, SYS_PMSFCR_EL1);
744
745	reg = arm_spe_event_to_pmsevfr(event);
746	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
747
748	reg = arm_spe_event_to_pmslatfr(event);
749	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
750
751	if (flags & PERF_EF_RELOAD) {
752		reg = arm_spe_event_to_pmsirr(event);
753		write_sysreg_s(reg, SYS_PMSIRR_EL1);
754		isb();
755		reg = local64_read(&hwc->period_left);
756		write_sysreg_s(reg, SYS_PMSICR_EL1);
757	}
758
759	reg = arm_spe_event_to_pmscr(event);
760	isb();
761	write_sysreg_s(reg, SYS_PMSCR_EL1);
762}
763
764static void arm_spe_pmu_stop(struct perf_event *event, int flags)
765{
766	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
767	struct hw_perf_event *hwc = &event->hw;
768	struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
769
770	/* If we're already stopped, then nothing to do */
771	if (hwc->state & PERF_HES_STOPPED)
772		return;
773
774	/* Stop all trace generation */
775	arm_spe_pmu_disable_and_drain_local();
776
777	if (flags & PERF_EF_UPDATE) {
778		/*
779		 * If there's a fault pending then ensure we contain it
780		 * to this buffer, since we might be on the context-switch
781		 * path.
782		 */
783		if (perf_get_aux(handle)) {
784			enum arm_spe_pmu_buf_fault_action act;
785
786			act = arm_spe_pmu_buf_get_fault_act(handle);
787			if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
788				arm_spe_perf_aux_output_end(handle);
789			else
790				write_sysreg_s(0, SYS_PMBSR_EL1);
791		}
792
793		/*
794		 * This may also contain ECOUNT, but nobody else should
795		 * be looking at period_left, since we forbid frequency
796		 * based sampling.
797		 */
798		local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
799		hwc->state |= PERF_HES_UPTODATE;
800	}
801
802	hwc->state |= PERF_HES_STOPPED;
803}
804
805static int arm_spe_pmu_add(struct perf_event *event, int flags)
806{
807	int ret = 0;
808	struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
809	struct hw_perf_event *hwc = &event->hw;
810	int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
811
812	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
813		return -ENOENT;
814
815	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
816
817	if (flags & PERF_EF_START) {
818		arm_spe_pmu_start(event, PERF_EF_RELOAD);
819		if (hwc->state & PERF_HES_STOPPED)
820			ret = -EINVAL;
821	}
822
823	return ret;
824}
825
826static void arm_spe_pmu_del(struct perf_event *event, int flags)
827{
828	arm_spe_pmu_stop(event, PERF_EF_UPDATE);
829}
830
831static void arm_spe_pmu_read(struct perf_event *event)
832{
833}
834
835static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
836				   int nr_pages, bool snapshot)
837{
838	int i, cpu = event->cpu;
839	struct page **pglist;
840	struct arm_spe_pmu_buf *buf;
841
842	/* We need at least two pages for this to work. */
843	if (nr_pages < 2)
844		return NULL;
845
846	/*
847	 * We require an even number of pages for snapshot mode, so that
848	 * we can effectively treat the buffer as consisting of two equal
849	 * parts and give userspace a fighting chance of getting some
850	 * useful data out of it.
851	 */
852	if (snapshot && (nr_pages & 1))
853		return NULL;
854
855	if (cpu == -1)
856		cpu = raw_smp_processor_id();
857
858	buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
859	if (!buf)
860		return NULL;
861
862	pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
863	if (!pglist)
864		goto out_free_buf;
865
866	for (i = 0; i < nr_pages; ++i)
867		pglist[i] = virt_to_page(pages[i]);
868
869	buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
870	if (!buf->base)
871		goto out_free_pglist;
872
873	buf->nr_pages	= nr_pages;
874	buf->snapshot	= snapshot;
875
876	kfree(pglist);
877	return buf;
878
879out_free_pglist:
880	kfree(pglist);
881out_free_buf:
882	kfree(buf);
883	return NULL;
884}
885
886static void arm_spe_pmu_free_aux(void *aux)
887{
888	struct arm_spe_pmu_buf *buf = aux;
889
890	vunmap(buf->base);
891	kfree(buf);
892}
893
894/* Initialisation and teardown functions */
895static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
896{
897	static atomic_t pmu_idx = ATOMIC_INIT(-1);
898
899	int idx;
900	char *name;
901	struct device *dev = &spe_pmu->pdev->dev;
902
903	spe_pmu->pmu = (struct pmu) {
904		.module = THIS_MODULE,
905		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
906		.attr_groups	= arm_spe_pmu_attr_groups,
907		/*
908		 * We hitch a ride on the software context here, so that
909		 * we can support per-task profiling (which is not possible
910		 * with the invalid context as it doesn't get sched callbacks).
911		 * This requires that userspace either uses a dummy event for
912		 * perf_event_open, since the aux buffer is not setup until
913		 * a subsequent mmap, or creates the profiling event in a
914		 * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
915		 * once the buffer has been created.
916		 */
917		.task_ctx_nr	= perf_sw_context,
918		.event_init	= arm_spe_pmu_event_init,
919		.add		= arm_spe_pmu_add,
920		.del		= arm_spe_pmu_del,
921		.start		= arm_spe_pmu_start,
922		.stop		= arm_spe_pmu_stop,
923		.read		= arm_spe_pmu_read,
924		.setup_aux	= arm_spe_pmu_setup_aux,
925		.free_aux	= arm_spe_pmu_free_aux,
926	};
927
928	idx = atomic_inc_return(&pmu_idx);
929	name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
930	if (!name) {
931		dev_err(dev, "failed to allocate name for pmu %d\n", idx);
932		return -ENOMEM;
933	}
934
935	return perf_pmu_register(&spe_pmu->pmu, name, -1);
936}
937
938static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
939{
940	perf_pmu_unregister(&spe_pmu->pmu);
941}
942
943static void __arm_spe_pmu_dev_probe(void *info)
944{
945	int fld;
946	u64 reg;
947	struct arm_spe_pmu *spe_pmu = info;
948	struct device *dev = &spe_pmu->pdev->dev;
949
950	fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
951						   ID_AA64DFR0_PMSVER_SHIFT);
952	if (!fld) {
953		dev_err(dev,
954			"unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
955			fld, smp_processor_id());
956		return;
957	}
958
959	/* Read PMBIDR first to determine whether or not we have access */
960	reg = read_sysreg_s(SYS_PMBIDR_EL1);
961	if (reg & BIT(SYS_PMBIDR_EL1_P_SHIFT)) {
962		dev_err(dev,
963			"profiling buffer owned by higher exception level\n");
964		return;
965	}
966
967	/* Minimum alignment. If it's out-of-range, then fail the probe */
968	fld = reg >> SYS_PMBIDR_EL1_ALIGN_SHIFT & SYS_PMBIDR_EL1_ALIGN_MASK;
969	spe_pmu->align = 1 << fld;
970	if (spe_pmu->align > SZ_2K) {
971		dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
972			fld, smp_processor_id());
973		return;
974	}
975
976	/* It's now safe to read PMSIDR and figure out what we've got */
977	reg = read_sysreg_s(SYS_PMSIDR_EL1);
978	if (reg & BIT(SYS_PMSIDR_EL1_FE_SHIFT))
979		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
980
981	if (reg & BIT(SYS_PMSIDR_EL1_FT_SHIFT))
982		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
983
984	if (reg & BIT(SYS_PMSIDR_EL1_FL_SHIFT))
985		spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
986
987	if (reg & BIT(SYS_PMSIDR_EL1_ARCHINST_SHIFT))
988		spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
989
990	if (reg & BIT(SYS_PMSIDR_EL1_LDS_SHIFT))
991		spe_pmu->features |= SPE_PMU_FEAT_LDS;
992
993	if (reg & BIT(SYS_PMSIDR_EL1_ERND_SHIFT))
994		spe_pmu->features |= SPE_PMU_FEAT_ERND;
995
996	/* This field has a spaced out encoding, so just use a look-up */
997	fld = reg >> SYS_PMSIDR_EL1_INTERVAL_SHIFT & SYS_PMSIDR_EL1_INTERVAL_MASK;
998	switch (fld) {
999	case 0:
1000		spe_pmu->min_period = 256;
1001		break;
1002	case 2:
1003		spe_pmu->min_period = 512;
1004		break;
1005	case 3:
1006		spe_pmu->min_period = 768;
1007		break;
1008	case 4:
1009		spe_pmu->min_period = 1024;
1010		break;
1011	case 5:
1012		spe_pmu->min_period = 1536;
1013		break;
1014	case 6:
1015		spe_pmu->min_period = 2048;
1016		break;
1017	case 7:
1018		spe_pmu->min_period = 3072;
1019		break;
1020	default:
1021		dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1022			 fld);
1023		fallthrough;
1024	case 8:
1025		spe_pmu->min_period = 4096;
1026	}
1027
1028	/* Maximum record size. If it's out-of-range, then fail the probe */
1029	fld = reg >> SYS_PMSIDR_EL1_MAXSIZE_SHIFT & SYS_PMSIDR_EL1_MAXSIZE_MASK;
1030	spe_pmu->max_record_sz = 1 << fld;
1031	if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1032		dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1033			fld, smp_processor_id());
1034		return;
1035	}
1036
1037	fld = reg >> SYS_PMSIDR_EL1_COUNTSIZE_SHIFT & SYS_PMSIDR_EL1_COUNTSIZE_MASK;
1038	switch (fld) {
1039	default:
1040		dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1041			 fld);
1042		fallthrough;
1043	case 2:
1044		spe_pmu->counter_sz = 12;
1045	}
1046
1047	dev_info(dev,
1048		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1049		 cpumask_pr_args(&spe_pmu->supported_cpus),
1050		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1051
1052	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1053	return;
1054}
1055
1056static void __arm_spe_pmu_reset_local(void)
1057{
1058	/*
1059	 * This is probably overkill, as we have no idea where we're
1060	 * draining any buffered data to...
1061	 */
1062	arm_spe_pmu_disable_and_drain_local();
1063
1064	/* Reset the buffer base pointer */
1065	write_sysreg_s(0, SYS_PMBPTR_EL1);
1066	isb();
1067
1068	/* Clear any pending management interrupts */
1069	write_sysreg_s(0, SYS_PMBSR_EL1);
1070	isb();
1071}
1072
1073static void __arm_spe_pmu_setup_one(void *info)
1074{
1075	struct arm_spe_pmu *spe_pmu = info;
1076
1077	__arm_spe_pmu_reset_local();
1078	enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1079}
1080
1081static void __arm_spe_pmu_stop_one(void *info)
1082{
1083	struct arm_spe_pmu *spe_pmu = info;
1084
1085	disable_percpu_irq(spe_pmu->irq);
1086	__arm_spe_pmu_reset_local();
1087}
1088
1089static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1090{
1091	struct arm_spe_pmu *spe_pmu;
1092
1093	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1094	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1095		return 0;
1096
1097	__arm_spe_pmu_setup_one(spe_pmu);
1098	return 0;
1099}
1100
1101static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1102{
1103	struct arm_spe_pmu *spe_pmu;
1104
1105	spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1106	if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1107		return 0;
1108
1109	__arm_spe_pmu_stop_one(spe_pmu);
1110	return 0;
1111}
1112
1113static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1114{
1115	int ret;
1116	cpumask_t *mask = &spe_pmu->supported_cpus;
1117
1118	/* Make sure we probe the hardware on a relevant CPU */
1119	ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1120	if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1121		return -ENXIO;
1122
1123	/* Request our PPIs (note that the IRQ is still disabled) */
1124	ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1125				 spe_pmu->handle);
1126	if (ret)
1127		return ret;
1128
1129	/*
1130	 * Register our hotplug notifier now so we don't miss any events.
1131	 * This will enable the IRQ for any supported CPUs that are already
1132	 * up.
1133	 */
1134	ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1135				       &spe_pmu->hotplug_node);
1136	if (ret)
1137		free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1138
1139	return ret;
1140}
1141
1142static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1143{
1144	cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1145	free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1146}
1147
1148/* Driver and device probing */
1149static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1150{
1151	struct platform_device *pdev = spe_pmu->pdev;
1152	int irq = platform_get_irq(pdev, 0);
1153
1154	if (irq < 0)
1155		return -ENXIO;
1156
1157	if (!irq_is_percpu(irq)) {
1158		dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1159		return -EINVAL;
1160	}
1161
1162	if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1163		dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1164		return -EINVAL;
1165	}
1166
1167	spe_pmu->irq = irq;
1168	return 0;
1169}
1170
1171static const struct of_device_id arm_spe_pmu_of_match[] = {
1172	{ .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1173	{ /* Sentinel */ },
1174};
1175MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1176
1177static const struct platform_device_id arm_spe_match[] = {
1178	{ ARMV8_SPE_PDEV_NAME, 0},
1179	{ }
1180};
1181MODULE_DEVICE_TABLE(platform, arm_spe_match);
1182
1183static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1184{
1185	int ret;
1186	struct arm_spe_pmu *spe_pmu;
1187	struct device *dev = &pdev->dev;
1188
1189	/*
1190	 * If kernelspace is unmapped when running at EL0, then the SPE
1191	 * buffer will fault and prematurely terminate the AUX session.
1192	 */
1193	if (arm64_kernel_unmapped_at_el0()) {
1194		dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1195		return -EPERM;
1196	}
1197
1198	spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1199	if (!spe_pmu) {
1200		dev_err(dev, "failed to allocate spe_pmu\n");
1201		return -ENOMEM;
1202	}
1203
1204	spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1205	if (!spe_pmu->handle)
1206		return -ENOMEM;
1207
1208	spe_pmu->pdev = pdev;
1209	platform_set_drvdata(pdev, spe_pmu);
1210
1211	ret = arm_spe_pmu_irq_probe(spe_pmu);
1212	if (ret)
1213		goto out_free_handle;
1214
1215	ret = arm_spe_pmu_dev_init(spe_pmu);
1216	if (ret)
1217		goto out_free_handle;
1218
1219	ret = arm_spe_pmu_perf_init(spe_pmu);
1220	if (ret)
1221		goto out_teardown_dev;
1222
1223	return 0;
1224
1225out_teardown_dev:
1226	arm_spe_pmu_dev_teardown(spe_pmu);
1227out_free_handle:
1228	free_percpu(spe_pmu->handle);
1229	return ret;
1230}
1231
1232static int arm_spe_pmu_device_remove(struct platform_device *pdev)
1233{
1234	struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1235
1236	arm_spe_pmu_perf_destroy(spe_pmu);
1237	arm_spe_pmu_dev_teardown(spe_pmu);
1238	free_percpu(spe_pmu->handle);
1239	return 0;
1240}
1241
1242static struct platform_driver arm_spe_pmu_driver = {
1243	.id_table = arm_spe_match,
1244	.driver	= {
1245		.name		= DRVNAME,
1246		.of_match_table	= of_match_ptr(arm_spe_pmu_of_match),
1247		.suppress_bind_attrs = true,
1248	},
1249	.probe	= arm_spe_pmu_device_probe,
1250	.remove	= arm_spe_pmu_device_remove,
1251};
1252
1253static int __init arm_spe_pmu_init(void)
1254{
1255	int ret;
1256
1257	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1258				      arm_spe_pmu_cpu_startup,
1259				      arm_spe_pmu_cpu_teardown);
1260	if (ret < 0)
1261		return ret;
1262	arm_spe_pmu_online = ret;
1263
1264	ret = platform_driver_register(&arm_spe_pmu_driver);
1265	if (ret)
1266		cpuhp_remove_multi_state(arm_spe_pmu_online);
1267
1268	return ret;
1269}
1270
1271static void __exit arm_spe_pmu_exit(void)
1272{
1273	platform_driver_unregister(&arm_spe_pmu_driver);
1274	cpuhp_remove_multi_state(arm_spe_pmu_online);
1275}
1276
1277module_init(arm_spe_pmu_init);
1278module_exit(arm_spe_pmu_exit);
1279
1280MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1281MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
1282MODULE_LICENSE("GPL v2");
1283