Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * ARM CoreSight Architecture PMU driver.
4 *
5 * This driver adds support for uncore PMU based on ARM CoreSight Performance
6 * Monitoring Unit Architecture. The PMU is accessible via MMIO registers and
7 * like other uncore PMUs, it does not support process specific events and
8 * cannot be used in sampling mode.
9 *
10 * This code is based on other uncore PMUs like ARM DSU PMU. It provides a
11 * generic implementation to operate the PMU according to CoreSight PMU
12 * architecture and ACPI ARM PMU table (APMT) documents below:
13 * - ARM CoreSight PMU architecture document number: ARM IHI 0091 A.a-00bet0.
14 * - APMT document number: ARM DEN0117.
15 *
16 * The user should refer to the vendor technical documentation to get details
17 * about the supported events.
18 *
19 * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
20 *
21 */
22
23#include <linux/acpi.h>
24#include <linux/cacheinfo.h>
25#include <linux/ctype.h>
26#include <linux/interrupt.h>
27#include <linux/io-64-nonatomic-lo-hi.h>
28#include <linux/module.h>
29#include <linux/mutex.h>
30#include <linux/of.h>
31#include <linux/perf_event.h>
32#include <linux/platform_device.h>
33
34#include "arm_cspmu.h"
35
36#define PMUNAME "arm_cspmu"
37#define DRVNAME "arm-cs-arch-pmu"
38
39#define ARM_CSPMU_CPUMASK_ATTR(_name, _config) \
40 ARM_CSPMU_EXT_ATTR(_name, arm_cspmu_cpumask_show, \
41 (unsigned long)_config)
42
43/*
44 * CoreSight PMU Arch register offsets.
45 */
46#define PMEVCNTR_LO 0x0
47#define PMEVCNTR_HI 0x4
48#define PMEVTYPER 0x400
49#define PMCCFILTR 0x47C
50#define PMEVFILTR 0xA00
51#define PMCNTENSET 0xC00
52#define PMCNTENCLR 0xC20
53#define PMINTENSET 0xC40
54#define PMINTENCLR 0xC60
55#define PMOVSCLR 0xC80
56#define PMOVSSET 0xCC0
57#define PMCFGR 0xE00
58#define PMCR 0xE04
59#define PMIIDR 0xE08
60
61/* PMCFGR register field */
62#define PMCFGR_NCG GENMASK(31, 28)
63#define PMCFGR_HDBG BIT(24)
64#define PMCFGR_TRO BIT(23)
65#define PMCFGR_SS BIT(22)
66#define PMCFGR_FZO BIT(21)
67#define PMCFGR_MSI BIT(20)
68#define PMCFGR_UEN BIT(19)
69#define PMCFGR_NA BIT(17)
70#define PMCFGR_EX BIT(16)
71#define PMCFGR_CCD BIT(15)
72#define PMCFGR_CC BIT(14)
73#define PMCFGR_SIZE GENMASK(13, 8)
74#define PMCFGR_N GENMASK(7, 0)
75
76/* PMCR register field */
77#define PMCR_TRO BIT(11)
78#define PMCR_HDBG BIT(10)
79#define PMCR_FZO BIT(9)
80#define PMCR_NA BIT(8)
81#define PMCR_DP BIT(5)
82#define PMCR_X BIT(4)
83#define PMCR_D BIT(3)
84#define PMCR_C BIT(2)
85#define PMCR_P BIT(1)
86#define PMCR_E BIT(0)
87
88/* Each SET/CLR register supports up to 32 counters. */
89#define ARM_CSPMU_SET_CLR_COUNTER_SHIFT 5
90#define ARM_CSPMU_SET_CLR_COUNTER_NUM \
91 (1 << ARM_CSPMU_SET_CLR_COUNTER_SHIFT)
92
93/* Convert counter idx into SET/CLR register number. */
94#define COUNTER_TO_SET_CLR_ID(idx) \
95 (idx >> ARM_CSPMU_SET_CLR_COUNTER_SHIFT)
96
97/* Convert counter idx into SET/CLR register bit. */
98#define COUNTER_TO_SET_CLR_BIT(idx) \
99 (idx & (ARM_CSPMU_SET_CLR_COUNTER_NUM - 1))
100
101#define ARM_CSPMU_ACTIVE_CPU_MASK 0x0
102#define ARM_CSPMU_ASSOCIATED_CPU_MASK 0x1
103
104/*
105 * Maximum poll count for reading counter value using high-low-high sequence.
106 */
107#define HILOHI_MAX_POLL 1000
108
109static unsigned long arm_cspmu_cpuhp_state;
110
111static DEFINE_MUTEX(arm_cspmu_lock);
112
113static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
114 struct hw_perf_event *hwc, u32 filter);
115
116static struct acpi_apmt_node *arm_cspmu_apmt_node(struct device *dev)
117{
118 struct acpi_apmt_node **ptr = dev_get_platdata(dev);
119
120 return ptr ? *ptr : NULL;
121}
122
123/*
124 * In CoreSight PMU architecture, all of the MMIO registers are 32-bit except
125 * counter register. The counter register can be implemented as 32-bit or 64-bit
126 * register depending on the value of PMCFGR.SIZE field. For 64-bit access,
127 * single-copy 64-bit atomic support is implementation defined. APMT node flag
128 * is used to identify if the PMU supports 64-bit single copy atomic. If 64-bit
129 * single copy atomic is not supported, the driver treats the register as a pair
130 * of 32-bit register.
131 */
132
133/*
134 * Read 64-bit register as a pair of 32-bit registers using hi-lo-hi sequence.
135 */
136static u64 read_reg64_hilohi(const void __iomem *addr, u32 max_poll_count)
137{
138 u32 val_lo, val_hi;
139 u64 val;
140
141 /* Use high-low-high sequence to avoid tearing */
142 do {
143 if (max_poll_count-- == 0) {
144 pr_err("ARM CSPMU: timeout hi-low-high sequence\n");
145 return 0;
146 }
147
148 val_hi = readl(addr + 4);
149 val_lo = readl(addr);
150 } while (val_hi != readl(addr + 4));
151
152 val = (((u64)val_hi << 32) | val_lo);
153
154 return val;
155}
156
157/* Check if cycle counter is supported. */
158static inline bool supports_cycle_counter(const struct arm_cspmu *cspmu)
159{
160 return (cspmu->pmcfgr & PMCFGR_CC);
161}
162
163/* Get counter size, which is (PMCFGR_SIZE + 1). */
164static inline u32 counter_size(const struct arm_cspmu *cspmu)
165{
166 return FIELD_GET(PMCFGR_SIZE, cspmu->pmcfgr) + 1;
167}
168
169/* Get counter mask. */
170static inline u64 counter_mask(const struct arm_cspmu *cspmu)
171{
172 return GENMASK_ULL(counter_size(cspmu) - 1, 0);
173}
174
175/* Check if counter is implemented as 64-bit register. */
176static inline bool use_64b_counter_reg(const struct arm_cspmu *cspmu)
177{
178 return (counter_size(cspmu) > 32);
179}
180
181ssize_t arm_cspmu_sysfs_event_show(struct device *dev,
182 struct device_attribute *attr, char *buf)
183{
184 struct perf_pmu_events_attr *pmu_attr;
185
186 pmu_attr = container_of(attr, typeof(*pmu_attr), attr);
187 return sysfs_emit(buf, "event=0x%llx\n", pmu_attr->id);
188}
189EXPORT_SYMBOL_GPL(arm_cspmu_sysfs_event_show);
190
191/* Default event list. */
192static struct attribute *arm_cspmu_event_attrs[] = {
193 ARM_CSPMU_EVENT_ATTR(cycles, ARM_CSPMU_EVT_CYCLES_DEFAULT),
194 NULL,
195};
196
197static struct attribute **
198arm_cspmu_get_event_attrs(const struct arm_cspmu *cspmu)
199{
200 struct attribute **attrs;
201
202 attrs = devm_kmemdup(cspmu->dev, arm_cspmu_event_attrs,
203 sizeof(arm_cspmu_event_attrs), GFP_KERNEL);
204
205 return attrs;
206}
207
208static umode_t
209arm_cspmu_event_attr_is_visible(struct kobject *kobj,
210 struct attribute *attr, int unused)
211{
212 struct device *dev = kobj_to_dev(kobj);
213 struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev));
214 struct perf_pmu_events_attr *eattr;
215
216 eattr = container_of(attr, typeof(*eattr), attr.attr);
217
218 /* Hide cycle event if not supported */
219 if (!supports_cycle_counter(cspmu) &&
220 eattr->id == ARM_CSPMU_EVT_CYCLES_DEFAULT)
221 return 0;
222
223 return attr->mode;
224}
225
226static struct attribute *arm_cspmu_format_attrs[] = {
227 ARM_CSPMU_FORMAT_EVENT_ATTR,
228 ARM_CSPMU_FORMAT_FILTER_ATTR,
229 NULL,
230};
231
232static struct attribute **
233arm_cspmu_get_format_attrs(const struct arm_cspmu *cspmu)
234{
235 struct attribute **attrs;
236
237 attrs = devm_kmemdup(cspmu->dev, arm_cspmu_format_attrs,
238 sizeof(arm_cspmu_format_attrs), GFP_KERNEL);
239
240 return attrs;
241}
242
243static u32 arm_cspmu_event_type(const struct perf_event *event)
244{
245 return event->attr.config & ARM_CSPMU_EVENT_MASK;
246}
247
248static bool arm_cspmu_is_cycle_counter_event(const struct perf_event *event)
249{
250 return (event->attr.config == ARM_CSPMU_EVT_CYCLES_DEFAULT);
251}
252
253static u32 arm_cspmu_event_filter(const struct perf_event *event)
254{
255 return event->attr.config1 & ARM_CSPMU_FILTER_MASK;
256}
257
258static ssize_t arm_cspmu_identifier_show(struct device *dev,
259 struct device_attribute *attr,
260 char *page)
261{
262 struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev));
263
264 return sysfs_emit(page, "%s\n", cspmu->identifier);
265}
266
267static struct device_attribute arm_cspmu_identifier_attr =
268 __ATTR(identifier, 0444, arm_cspmu_identifier_show, NULL);
269
270static struct attribute *arm_cspmu_identifier_attrs[] = {
271 &arm_cspmu_identifier_attr.attr,
272 NULL,
273};
274
275static struct attribute_group arm_cspmu_identifier_attr_group = {
276 .attrs = arm_cspmu_identifier_attrs,
277};
278
279static const char *arm_cspmu_get_identifier(const struct arm_cspmu *cspmu)
280{
281 const char *identifier =
282 devm_kasprintf(cspmu->dev, GFP_KERNEL, "%x",
283 cspmu->impl.pmiidr);
284 return identifier;
285}
286
287static const char *arm_cspmu_type_str[ACPI_APMT_NODE_TYPE_COUNT] = {
288 "mc",
289 "smmu",
290 "pcie",
291 "acpi",
292 "cache",
293};
294
295static const char *arm_cspmu_get_name(const struct arm_cspmu *cspmu)
296{
297 struct device *dev;
298 struct acpi_apmt_node *apmt_node;
299 u8 pmu_type;
300 char *name;
301 char acpi_hid_string[ACPI_ID_LEN] = { 0 };
302 static atomic_t pmu_idx[ACPI_APMT_NODE_TYPE_COUNT] = { 0 };
303
304 dev = cspmu->dev;
305 apmt_node = arm_cspmu_apmt_node(dev);
306 if (!apmt_node)
307 return devm_kasprintf(dev, GFP_KERNEL, PMUNAME "_%u",
308 atomic_fetch_inc(&pmu_idx[0]));
309
310 pmu_type = apmt_node->type;
311
312 if (pmu_type >= ACPI_APMT_NODE_TYPE_COUNT) {
313 dev_err(dev, "unsupported PMU type-%u\n", pmu_type);
314 return NULL;
315 }
316
317 if (pmu_type == ACPI_APMT_NODE_TYPE_ACPI) {
318 memcpy(acpi_hid_string,
319 &apmt_node->inst_primary,
320 sizeof(apmt_node->inst_primary));
321 name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%s_%u", PMUNAME,
322 arm_cspmu_type_str[pmu_type],
323 acpi_hid_string,
324 apmt_node->inst_secondary);
325 } else {
326 name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%d", PMUNAME,
327 arm_cspmu_type_str[pmu_type],
328 atomic_fetch_inc(&pmu_idx[pmu_type]));
329 }
330
331 return name;
332}
333
334static ssize_t arm_cspmu_cpumask_show(struct device *dev,
335 struct device_attribute *attr,
336 char *buf)
337{
338 struct pmu *pmu = dev_get_drvdata(dev);
339 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
340 struct dev_ext_attribute *eattr =
341 container_of(attr, struct dev_ext_attribute, attr);
342 unsigned long mask_id = (unsigned long)eattr->var;
343 const cpumask_t *cpumask;
344
345 switch (mask_id) {
346 case ARM_CSPMU_ACTIVE_CPU_MASK:
347 cpumask = &cspmu->active_cpu;
348 break;
349 case ARM_CSPMU_ASSOCIATED_CPU_MASK:
350 cpumask = &cspmu->associated_cpus;
351 break;
352 default:
353 return 0;
354 }
355 return cpumap_print_to_pagebuf(true, buf, cpumask);
356}
357
358static struct attribute *arm_cspmu_cpumask_attrs[] = {
359 ARM_CSPMU_CPUMASK_ATTR(cpumask, ARM_CSPMU_ACTIVE_CPU_MASK),
360 ARM_CSPMU_CPUMASK_ATTR(associated_cpus, ARM_CSPMU_ASSOCIATED_CPU_MASK),
361 NULL,
362};
363
364static struct attribute_group arm_cspmu_cpumask_attr_group = {
365 .attrs = arm_cspmu_cpumask_attrs,
366};
367
368static struct arm_cspmu_impl_match impl_match[] = {
369 {
370 .module_name = "nvidia_cspmu",
371 .pmiidr_val = ARM_CSPMU_IMPL_ID_NVIDIA,
372 .pmiidr_mask = ARM_CSPMU_PMIIDR_IMPLEMENTER,
373 .module = NULL,
374 .impl_init_ops = NULL,
375 },
376 {
377 .module_name = "ampere_cspmu",
378 .pmiidr_val = ARM_CSPMU_IMPL_ID_AMPERE,
379 .pmiidr_mask = ARM_CSPMU_PMIIDR_IMPLEMENTER,
380 .module = NULL,
381 .impl_init_ops = NULL,
382 },
383
384 {0}
385};
386
387static struct arm_cspmu_impl_match *arm_cspmu_impl_match_get(u32 pmiidr)
388{
389 struct arm_cspmu_impl_match *match = impl_match;
390
391 for (; match->pmiidr_val; match++) {
392 u32 mask = match->pmiidr_mask;
393
394 if ((match->pmiidr_val & mask) == (pmiidr & mask))
395 return match;
396 }
397
398 return NULL;
399}
400
401#define DEFAULT_IMPL_OP(name) .name = arm_cspmu_##name
402
403static int arm_cspmu_init_impl_ops(struct arm_cspmu *cspmu)
404{
405 int ret = 0;
406 struct acpi_apmt_node *apmt_node = arm_cspmu_apmt_node(cspmu->dev);
407 struct arm_cspmu_impl_match *match;
408
409 /* Start with a default PMU implementation */
410 cspmu->impl.module = THIS_MODULE;
411 cspmu->impl.pmiidr = readl(cspmu->base0 + PMIIDR);
412 cspmu->impl.ops = (struct arm_cspmu_impl_ops) {
413 DEFAULT_IMPL_OP(get_event_attrs),
414 DEFAULT_IMPL_OP(get_format_attrs),
415 DEFAULT_IMPL_OP(get_identifier),
416 DEFAULT_IMPL_OP(get_name),
417 DEFAULT_IMPL_OP(is_cycle_counter_event),
418 DEFAULT_IMPL_OP(event_type),
419 DEFAULT_IMPL_OP(event_filter),
420 DEFAULT_IMPL_OP(set_ev_filter),
421 DEFAULT_IMPL_OP(event_attr_is_visible),
422 };
423
424 /* Firmware may override implementer/product ID from PMIIDR */
425 if (apmt_node && apmt_node->impl_id)
426 cspmu->impl.pmiidr = apmt_node->impl_id;
427
428 /* Find implementer specific attribute ops. */
429 match = arm_cspmu_impl_match_get(cspmu->impl.pmiidr);
430
431 /* Load implementer module and initialize the callbacks. */
432 if (match) {
433 mutex_lock(&arm_cspmu_lock);
434
435 if (match->impl_init_ops) {
436 /* Prevent unload until PMU registration is done. */
437 if (try_module_get(match->module)) {
438 cspmu->impl.module = match->module;
439 cspmu->impl.match = match;
440 ret = match->impl_init_ops(cspmu);
441 if (ret)
442 module_put(match->module);
443 } else {
444 WARN(1, "arm_cspmu failed to get module: %s\n",
445 match->module_name);
446 ret = -EINVAL;
447 }
448 } else {
449 request_module_nowait(match->module_name);
450 ret = -EPROBE_DEFER;
451 }
452
453 mutex_unlock(&arm_cspmu_lock);
454 }
455
456 return ret;
457}
458
459static struct attribute_group *
460arm_cspmu_alloc_event_attr_group(struct arm_cspmu *cspmu)
461{
462 struct attribute_group *event_group;
463 struct device *dev = cspmu->dev;
464 const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
465
466 event_group =
467 devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL);
468 if (!event_group)
469 return NULL;
470
471 event_group->name = "events";
472 event_group->is_visible = impl_ops->event_attr_is_visible;
473 event_group->attrs = impl_ops->get_event_attrs(cspmu);
474
475 if (!event_group->attrs)
476 return NULL;
477
478 return event_group;
479}
480
481static struct attribute_group *
482arm_cspmu_alloc_format_attr_group(struct arm_cspmu *cspmu)
483{
484 struct attribute_group *format_group;
485 struct device *dev = cspmu->dev;
486
487 format_group =
488 devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL);
489 if (!format_group)
490 return NULL;
491
492 format_group->name = "format";
493 format_group->attrs = cspmu->impl.ops.get_format_attrs(cspmu);
494
495 if (!format_group->attrs)
496 return NULL;
497
498 return format_group;
499}
500
501static int arm_cspmu_alloc_attr_groups(struct arm_cspmu *cspmu)
502{
503 const struct attribute_group **attr_groups = cspmu->attr_groups;
504 const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
505
506 cspmu->identifier = impl_ops->get_identifier(cspmu);
507 cspmu->name = impl_ops->get_name(cspmu);
508
509 if (!cspmu->identifier || !cspmu->name)
510 return -ENOMEM;
511
512 attr_groups[0] = arm_cspmu_alloc_event_attr_group(cspmu);
513 attr_groups[1] = arm_cspmu_alloc_format_attr_group(cspmu);
514 attr_groups[2] = &arm_cspmu_identifier_attr_group;
515 attr_groups[3] = &arm_cspmu_cpumask_attr_group;
516
517 if (!attr_groups[0] || !attr_groups[1])
518 return -ENOMEM;
519
520 return 0;
521}
522
523static inline void arm_cspmu_reset_counters(struct arm_cspmu *cspmu)
524{
525 writel(PMCR_C | PMCR_P, cspmu->base0 + PMCR);
526}
527
528static inline void arm_cspmu_start_counters(struct arm_cspmu *cspmu)
529{
530 writel(PMCR_E, cspmu->base0 + PMCR);
531}
532
533static inline void arm_cspmu_stop_counters(struct arm_cspmu *cspmu)
534{
535 writel(0, cspmu->base0 + PMCR);
536}
537
538static void arm_cspmu_enable(struct pmu *pmu)
539{
540 bool disabled;
541 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
542
543 disabled = bitmap_empty(cspmu->hw_events.used_ctrs,
544 cspmu->num_logical_ctrs);
545
546 if (disabled)
547 return;
548
549 arm_cspmu_start_counters(cspmu);
550}
551
552static void arm_cspmu_disable(struct pmu *pmu)
553{
554 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
555
556 arm_cspmu_stop_counters(cspmu);
557}
558
559static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
560 struct perf_event *event)
561{
562 int idx, ret;
563 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
564
565 if (supports_cycle_counter(cspmu)) {
566 if (cspmu->impl.ops.is_cycle_counter_event(event)) {
567 /* Search for available cycle counter. */
568 if (test_and_set_bit(cspmu->cycle_counter_logical_idx,
569 hw_events->used_ctrs))
570 return -EAGAIN;
571
572 return cspmu->cycle_counter_logical_idx;
573 }
574
575 /*
576 * Search a regular counter from the used counter bitmap.
577 * The cycle counter divides the bitmap into two parts. Search
578 * the first then second half to exclude the cycle counter bit.
579 */
580 idx = find_first_zero_bit(hw_events->used_ctrs,
581 cspmu->cycle_counter_logical_idx);
582 if (idx >= cspmu->cycle_counter_logical_idx) {
583 idx = find_next_zero_bit(
584 hw_events->used_ctrs,
585 cspmu->num_logical_ctrs,
586 cspmu->cycle_counter_logical_idx + 1);
587 }
588 } else {
589 idx = find_first_zero_bit(hw_events->used_ctrs,
590 cspmu->num_logical_ctrs);
591 }
592
593 if (idx >= cspmu->num_logical_ctrs)
594 return -EAGAIN;
595
596 if (cspmu->impl.ops.validate_event) {
597 ret = cspmu->impl.ops.validate_event(cspmu, event);
598 if (ret)
599 return ret;
600 }
601
602 set_bit(idx, hw_events->used_ctrs);
603
604 return idx;
605}
606
607static bool arm_cspmu_validate_event(struct pmu *pmu,
608 struct arm_cspmu_hw_events *hw_events,
609 struct perf_event *event)
610{
611 if (is_software_event(event))
612 return true;
613
614 /* Reject groups spanning multiple HW PMUs. */
615 if (event->pmu != pmu)
616 return false;
617
618 return (arm_cspmu_get_event_idx(hw_events, event) >= 0);
619}
620
621/*
622 * Make sure the group of events can be scheduled at once
623 * on the PMU.
624 */
625static bool arm_cspmu_validate_group(struct perf_event *event)
626{
627 struct perf_event *sibling, *leader = event->group_leader;
628 struct arm_cspmu_hw_events fake_hw_events;
629
630 if (event->group_leader == event)
631 return true;
632
633 memset(&fake_hw_events, 0, sizeof(fake_hw_events));
634
635 if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events, leader))
636 return false;
637
638 for_each_sibling_event(sibling, leader) {
639 if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events,
640 sibling))
641 return false;
642 }
643
644 return arm_cspmu_validate_event(event->pmu, &fake_hw_events, event);
645}
646
647static int arm_cspmu_event_init(struct perf_event *event)
648{
649 struct arm_cspmu *cspmu;
650 struct hw_perf_event *hwc = &event->hw;
651
652 cspmu = to_arm_cspmu(event->pmu);
653
654 if (event->attr.type != event->pmu->type)
655 return -ENOENT;
656
657 /*
658 * Following other "uncore" PMUs, we do not support sampling mode or
659 * attach to a task (per-process mode).
660 */
661 if (is_sampling_event(event)) {
662 dev_dbg(cspmu->pmu.dev,
663 "Can't support sampling events\n");
664 return -EOPNOTSUPP;
665 }
666
667 if (event->cpu < 0 || event->attach_state & PERF_ATTACH_TASK) {
668 dev_dbg(cspmu->pmu.dev,
669 "Can't support per-task counters\n");
670 return -EINVAL;
671 }
672
673 /*
674 * Make sure the CPU assignment is on one of the CPUs associated with
675 * this PMU.
676 */
677 if (!cpumask_test_cpu(event->cpu, &cspmu->associated_cpus)) {
678 dev_dbg(cspmu->pmu.dev,
679 "Requested cpu is not associated with the PMU\n");
680 return -EINVAL;
681 }
682
683 /* Enforce the current active CPU to handle the events in this PMU. */
684 event->cpu = cpumask_first(&cspmu->active_cpu);
685 if (event->cpu >= nr_cpu_ids)
686 return -EINVAL;
687
688 if (!arm_cspmu_validate_group(event))
689 return -EINVAL;
690
691 /*
692 * The logical counter id is tracked with hw_perf_event.extra_reg.idx.
693 * The physical counter id is tracked with hw_perf_event.idx.
694 * We don't assign an index until we actually place the event onto
695 * hardware. Use -1 to signify that we haven't decided where to put it
696 * yet.
697 */
698 hwc->idx = -1;
699 hwc->extra_reg.idx = -1;
700 hwc->config = cspmu->impl.ops.event_type(event);
701
702 return 0;
703}
704
705static inline u32 counter_offset(u32 reg_sz, u32 ctr_idx)
706{
707 return (PMEVCNTR_LO + (reg_sz * ctr_idx));
708}
709
710static void arm_cspmu_write_counter(struct perf_event *event, u64 val)
711{
712 u32 offset;
713 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
714
715 if (use_64b_counter_reg(cspmu)) {
716 offset = counter_offset(sizeof(u64), event->hw.idx);
717
718 if (cspmu->has_atomic_dword)
719 writeq(val, cspmu->base1 + offset);
720 else
721 lo_hi_writeq(val, cspmu->base1 + offset);
722 } else {
723 offset = counter_offset(sizeof(u32), event->hw.idx);
724
725 writel(lower_32_bits(val), cspmu->base1 + offset);
726 }
727}
728
729static u64 arm_cspmu_read_counter(struct perf_event *event)
730{
731 u32 offset;
732 const void __iomem *counter_addr;
733 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
734
735 if (use_64b_counter_reg(cspmu)) {
736 offset = counter_offset(sizeof(u64), event->hw.idx);
737 counter_addr = cspmu->base1 + offset;
738
739 return cspmu->has_atomic_dword ?
740 readq(counter_addr) :
741 read_reg64_hilohi(counter_addr, HILOHI_MAX_POLL);
742 }
743
744 offset = counter_offset(sizeof(u32), event->hw.idx);
745 return readl(cspmu->base1 + offset);
746}
747
748/*
749 * arm_cspmu_set_event_period: Set the period for the counter.
750 *
751 * To handle cases of extreme interrupt latency, we program
752 * the counter with half of the max count for the counters.
753 */
754static void arm_cspmu_set_event_period(struct perf_event *event)
755{
756 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
757 u64 val = counter_mask(cspmu) >> 1ULL;
758
759 local64_set(&event->hw.prev_count, val);
760 arm_cspmu_write_counter(event, val);
761}
762
763static void arm_cspmu_enable_counter(struct arm_cspmu *cspmu, int idx)
764{
765 u32 reg_id, reg_bit, inten_off, cnten_off;
766
767 reg_id = COUNTER_TO_SET_CLR_ID(idx);
768 reg_bit = COUNTER_TO_SET_CLR_BIT(idx);
769
770 inten_off = PMINTENSET + (4 * reg_id);
771 cnten_off = PMCNTENSET + (4 * reg_id);
772
773 writel(BIT(reg_bit), cspmu->base0 + inten_off);
774 writel(BIT(reg_bit), cspmu->base0 + cnten_off);
775}
776
777static void arm_cspmu_disable_counter(struct arm_cspmu *cspmu, int idx)
778{
779 u32 reg_id, reg_bit, inten_off, cnten_off;
780
781 reg_id = COUNTER_TO_SET_CLR_ID(idx);
782 reg_bit = COUNTER_TO_SET_CLR_BIT(idx);
783
784 inten_off = PMINTENCLR + (4 * reg_id);
785 cnten_off = PMCNTENCLR + (4 * reg_id);
786
787 writel(BIT(reg_bit), cspmu->base0 + cnten_off);
788 writel(BIT(reg_bit), cspmu->base0 + inten_off);
789}
790
791static void arm_cspmu_event_update(struct perf_event *event)
792{
793 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
794 struct hw_perf_event *hwc = &event->hw;
795 u64 delta, prev, now;
796
797 do {
798 prev = local64_read(&hwc->prev_count);
799 now = arm_cspmu_read_counter(event);
800 } while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
801
802 delta = (now - prev) & counter_mask(cspmu);
803 local64_add(delta, &event->count);
804}
805
806static inline void arm_cspmu_set_event(struct arm_cspmu *cspmu,
807 struct hw_perf_event *hwc)
808{
809 u32 offset = PMEVTYPER + (4 * hwc->idx);
810
811 writel(hwc->config, cspmu->base0 + offset);
812}
813
814static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
815 struct hw_perf_event *hwc,
816 u32 filter)
817{
818 u32 offset = PMEVFILTR + (4 * hwc->idx);
819
820 writel(filter, cspmu->base0 + offset);
821}
822
823static inline void arm_cspmu_set_cc_filter(struct arm_cspmu *cspmu, u32 filter)
824{
825 u32 offset = PMCCFILTR;
826
827 writel(filter, cspmu->base0 + offset);
828}
829
830static void arm_cspmu_start(struct perf_event *event, int pmu_flags)
831{
832 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
833 struct hw_perf_event *hwc = &event->hw;
834 u32 filter;
835
836 /* We always reprogram the counter */
837 if (pmu_flags & PERF_EF_RELOAD)
838 WARN_ON(!(hwc->state & PERF_HES_UPTODATE));
839
840 arm_cspmu_set_event_period(event);
841
842 filter = cspmu->impl.ops.event_filter(event);
843
844 if (event->hw.extra_reg.idx == cspmu->cycle_counter_logical_idx) {
845 arm_cspmu_set_cc_filter(cspmu, filter);
846 } else {
847 arm_cspmu_set_event(cspmu, hwc);
848 cspmu->impl.ops.set_ev_filter(cspmu, hwc, filter);
849 }
850
851 hwc->state = 0;
852
853 arm_cspmu_enable_counter(cspmu, hwc->idx);
854}
855
856static void arm_cspmu_stop(struct perf_event *event, int pmu_flags)
857{
858 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
859 struct hw_perf_event *hwc = &event->hw;
860
861 if (hwc->state & PERF_HES_STOPPED)
862 return;
863
864 arm_cspmu_disable_counter(cspmu, hwc->idx);
865 arm_cspmu_event_update(event);
866
867 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
868}
869
870static inline u32 to_phys_idx(struct arm_cspmu *cspmu, u32 idx)
871{
872 return (idx == cspmu->cycle_counter_logical_idx) ?
873 ARM_CSPMU_CYCLE_CNTR_IDX : idx;
874}
875
876static int arm_cspmu_add(struct perf_event *event, int flags)
877{
878 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
879 struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events;
880 struct hw_perf_event *hwc = &event->hw;
881 int idx;
882
883 if (WARN_ON_ONCE(!cpumask_test_cpu(smp_processor_id(),
884 &cspmu->associated_cpus)))
885 return -ENOENT;
886
887 idx = arm_cspmu_get_event_idx(hw_events, event);
888 if (idx < 0)
889 return idx;
890
891 hw_events->events[idx] = event;
892 hwc->idx = to_phys_idx(cspmu, idx);
893 hwc->extra_reg.idx = idx;
894 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
895
896 if (flags & PERF_EF_START)
897 arm_cspmu_start(event, PERF_EF_RELOAD);
898
899 /* Propagate changes to the userspace mapping. */
900 perf_event_update_userpage(event);
901
902 return 0;
903}
904
905static void arm_cspmu_del(struct perf_event *event, int flags)
906{
907 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
908 struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events;
909 struct hw_perf_event *hwc = &event->hw;
910 int idx = hwc->extra_reg.idx;
911
912 arm_cspmu_stop(event, PERF_EF_UPDATE);
913
914 hw_events->events[idx] = NULL;
915
916 clear_bit(idx, hw_events->used_ctrs);
917
918 perf_event_update_userpage(event);
919}
920
921static void arm_cspmu_read(struct perf_event *event)
922{
923 arm_cspmu_event_update(event);
924}
925
926static struct arm_cspmu *arm_cspmu_alloc(struct platform_device *pdev)
927{
928 struct acpi_apmt_node *apmt_node;
929 struct arm_cspmu *cspmu;
930 struct device *dev = &pdev->dev;
931
932 cspmu = devm_kzalloc(dev, sizeof(*cspmu), GFP_KERNEL);
933 if (!cspmu)
934 return NULL;
935
936 cspmu->dev = dev;
937 platform_set_drvdata(pdev, cspmu);
938
939 apmt_node = arm_cspmu_apmt_node(dev);
940 if (apmt_node) {
941 cspmu->has_atomic_dword = apmt_node->flags & ACPI_APMT_FLAGS_ATOMIC;
942 } else {
943 u32 width = 0;
944
945 device_property_read_u32(dev, "reg-io-width", &width);
946 cspmu->has_atomic_dword = (width == 8);
947 }
948
949 return cspmu;
950}
951
952static int arm_cspmu_init_mmio(struct arm_cspmu *cspmu)
953{
954 struct device *dev;
955 struct platform_device *pdev;
956
957 dev = cspmu->dev;
958 pdev = to_platform_device(dev);
959
960 /* Base address for page 0. */
961 cspmu->base0 = devm_platform_ioremap_resource(pdev, 0);
962 if (IS_ERR(cspmu->base0)) {
963 dev_err(dev, "ioremap failed for page-0 resource\n");
964 return PTR_ERR(cspmu->base0);
965 }
966
967 /* Base address for page 1 if supported. Otherwise point to page 0. */
968 cspmu->base1 = cspmu->base0;
969 if (platform_get_resource(pdev, IORESOURCE_MEM, 1)) {
970 cspmu->base1 = devm_platform_ioremap_resource(pdev, 1);
971 if (IS_ERR(cspmu->base1)) {
972 dev_err(dev, "ioremap failed for page-1 resource\n");
973 return PTR_ERR(cspmu->base1);
974 }
975 }
976
977 cspmu->pmcfgr = readl(cspmu->base0 + PMCFGR);
978
979 cspmu->num_logical_ctrs = FIELD_GET(PMCFGR_N, cspmu->pmcfgr) + 1;
980
981 cspmu->cycle_counter_logical_idx = ARM_CSPMU_MAX_HW_CNTRS;
982
983 if (supports_cycle_counter(cspmu)) {
984 /*
985 * The last logical counter is mapped to cycle counter if
986 * there is a gap between regular and cycle counter. Otherwise,
987 * logical and physical have 1-to-1 mapping.
988 */
989 cspmu->cycle_counter_logical_idx =
990 (cspmu->num_logical_ctrs <= ARM_CSPMU_CYCLE_CNTR_IDX) ?
991 cspmu->num_logical_ctrs - 1 :
992 ARM_CSPMU_CYCLE_CNTR_IDX;
993 }
994
995 cspmu->num_set_clr_reg =
996 DIV_ROUND_UP(cspmu->num_logical_ctrs,
997 ARM_CSPMU_SET_CLR_COUNTER_NUM);
998
999 cspmu->hw_events.events =
1000 devm_kcalloc(dev, cspmu->num_logical_ctrs,
1001 sizeof(*cspmu->hw_events.events), GFP_KERNEL);
1002
1003 if (!cspmu->hw_events.events)
1004 return -ENOMEM;
1005
1006 return 0;
1007}
1008
1009static inline int arm_cspmu_get_reset_overflow(struct arm_cspmu *cspmu,
1010 u32 *pmovs)
1011{
1012 int i;
1013 u32 pmovclr_offset = PMOVSCLR;
1014 u32 has_overflowed = 0;
1015
1016 for (i = 0; i < cspmu->num_set_clr_reg; ++i) {
1017 pmovs[i] = readl(cspmu->base1 + pmovclr_offset);
1018 has_overflowed |= pmovs[i];
1019 writel(pmovs[i], cspmu->base1 + pmovclr_offset);
1020 pmovclr_offset += sizeof(u32);
1021 }
1022
1023 return has_overflowed != 0;
1024}
1025
1026static irqreturn_t arm_cspmu_handle_irq(int irq_num, void *dev)
1027{
1028 int idx, has_overflowed;
1029 struct perf_event *event;
1030 struct arm_cspmu *cspmu = dev;
1031 DECLARE_BITMAP(pmovs, ARM_CSPMU_MAX_HW_CNTRS);
1032 bool handled = false;
1033
1034 arm_cspmu_stop_counters(cspmu);
1035
1036 has_overflowed = arm_cspmu_get_reset_overflow(cspmu, (u32 *)pmovs);
1037 if (!has_overflowed)
1038 goto done;
1039
1040 for_each_set_bit(idx, cspmu->hw_events.used_ctrs,
1041 cspmu->num_logical_ctrs) {
1042 event = cspmu->hw_events.events[idx];
1043
1044 if (!event)
1045 continue;
1046
1047 if (!test_bit(event->hw.idx, pmovs))
1048 continue;
1049
1050 arm_cspmu_event_update(event);
1051 arm_cspmu_set_event_period(event);
1052
1053 handled = true;
1054 }
1055
1056done:
1057 arm_cspmu_start_counters(cspmu);
1058 return IRQ_RETVAL(handled);
1059}
1060
1061static int arm_cspmu_request_irq(struct arm_cspmu *cspmu)
1062{
1063 int irq, ret;
1064 struct device *dev;
1065 struct platform_device *pdev;
1066
1067 dev = cspmu->dev;
1068 pdev = to_platform_device(dev);
1069
1070 /* Skip IRQ request if the PMU does not support overflow interrupt. */
1071 irq = platform_get_irq_optional(pdev, 0);
1072 if (irq < 0)
1073 return irq == -ENXIO ? 0 : irq;
1074
1075 ret = devm_request_irq(dev, irq, arm_cspmu_handle_irq,
1076 IRQF_NOBALANCING | IRQF_NO_THREAD, dev_name(dev),
1077 cspmu);
1078 if (ret) {
1079 dev_err(dev, "Could not request IRQ %d\n", irq);
1080 return ret;
1081 }
1082
1083 cspmu->irq = irq;
1084
1085 return 0;
1086}
1087
1088#if defined(CONFIG_ACPI) && defined(CONFIG_ARM64)
1089#include <acpi/processor.h>
1090
1091static inline int arm_cspmu_find_cpu_container(int cpu, u32 container_uid)
1092{
1093 struct device *cpu_dev;
1094 struct acpi_device *acpi_dev;
1095
1096 cpu_dev = get_cpu_device(cpu);
1097 if (!cpu_dev)
1098 return -ENODEV;
1099
1100 acpi_dev = ACPI_COMPANION(cpu_dev);
1101 while (acpi_dev) {
1102 if (acpi_dev_hid_uid_match(acpi_dev, ACPI_PROCESSOR_CONTAINER_HID, container_uid))
1103 return 0;
1104
1105 acpi_dev = acpi_dev_parent(acpi_dev);
1106 }
1107
1108 return -ENODEV;
1109}
1110
1111static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu)
1112{
1113 struct acpi_apmt_node *apmt_node;
1114 int affinity_flag;
1115 int cpu;
1116
1117 apmt_node = arm_cspmu_apmt_node(cspmu->dev);
1118 affinity_flag = apmt_node->flags & ACPI_APMT_FLAGS_AFFINITY;
1119
1120 if (affinity_flag == ACPI_APMT_FLAGS_AFFINITY_PROC) {
1121 for_each_possible_cpu(cpu) {
1122 if (apmt_node->proc_affinity ==
1123 get_acpi_id_for_cpu(cpu)) {
1124 cpumask_set_cpu(cpu, &cspmu->associated_cpus);
1125 break;
1126 }
1127 }
1128 } else {
1129 for_each_possible_cpu(cpu) {
1130 if (arm_cspmu_find_cpu_container(
1131 cpu, apmt_node->proc_affinity))
1132 continue;
1133
1134 cpumask_set_cpu(cpu, &cspmu->associated_cpus);
1135 }
1136 }
1137
1138 return 0;
1139}
1140#else
1141static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu)
1142{
1143 return -ENODEV;
1144}
1145#endif
1146
1147static int arm_cspmu_of_get_cpus(struct arm_cspmu *cspmu)
1148{
1149 struct of_phandle_iterator it;
1150 int ret, cpu;
1151
1152 of_for_each_phandle(&it, ret, dev_of_node(cspmu->dev), "cpus", NULL, 0) {
1153 cpu = of_cpu_node_to_id(it.node);
1154 if (cpu < 0)
1155 continue;
1156 cpumask_set_cpu(cpu, &cspmu->associated_cpus);
1157 }
1158 return ret == -ENOENT ? 0 : ret;
1159}
1160
1161static int arm_cspmu_get_cpus(struct arm_cspmu *cspmu)
1162{
1163 int ret = 0;
1164
1165 if (arm_cspmu_apmt_node(cspmu->dev))
1166 ret = arm_cspmu_acpi_get_cpus(cspmu);
1167 else if (device_property_present(cspmu->dev, "cpus"))
1168 ret = arm_cspmu_of_get_cpus(cspmu);
1169 else
1170 cpumask_copy(&cspmu->associated_cpus, cpu_possible_mask);
1171
1172 if (!ret && cpumask_empty(&cspmu->associated_cpus)) {
1173 dev_dbg(cspmu->dev, "No cpu associated with the PMU\n");
1174 ret = -ENODEV;
1175 }
1176 return ret;
1177}
1178
1179static int arm_cspmu_register_pmu(struct arm_cspmu *cspmu)
1180{
1181 int ret, capabilities;
1182
1183 ret = arm_cspmu_alloc_attr_groups(cspmu);
1184 if (ret)
1185 return ret;
1186
1187 ret = cpuhp_state_add_instance(arm_cspmu_cpuhp_state,
1188 &cspmu->cpuhp_node);
1189 if (ret)
1190 return ret;
1191
1192 capabilities = PERF_PMU_CAP_NO_EXCLUDE;
1193 if (cspmu->irq == 0)
1194 capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1195
1196 cspmu->pmu = (struct pmu){
1197 .task_ctx_nr = perf_invalid_context,
1198 .module = cspmu->impl.module,
1199 .parent = cspmu->dev,
1200 .pmu_enable = arm_cspmu_enable,
1201 .pmu_disable = arm_cspmu_disable,
1202 .event_init = arm_cspmu_event_init,
1203 .add = arm_cspmu_add,
1204 .del = arm_cspmu_del,
1205 .start = arm_cspmu_start,
1206 .stop = arm_cspmu_stop,
1207 .read = arm_cspmu_read,
1208 .attr_groups = cspmu->attr_groups,
1209 .capabilities = capabilities,
1210 };
1211
1212 /* Hardware counter init */
1213 arm_cspmu_reset_counters(cspmu);
1214
1215 ret = perf_pmu_register(&cspmu->pmu, cspmu->name, -1);
1216 if (ret) {
1217 cpuhp_state_remove_instance(arm_cspmu_cpuhp_state,
1218 &cspmu->cpuhp_node);
1219 }
1220
1221 return ret;
1222}
1223
1224static int arm_cspmu_device_probe(struct platform_device *pdev)
1225{
1226 int ret;
1227 struct arm_cspmu *cspmu;
1228
1229 cspmu = arm_cspmu_alloc(pdev);
1230 if (!cspmu)
1231 return -ENOMEM;
1232
1233 ret = arm_cspmu_init_mmio(cspmu);
1234 if (ret)
1235 return ret;
1236
1237 ret = arm_cspmu_request_irq(cspmu);
1238 if (ret)
1239 return ret;
1240
1241 ret = arm_cspmu_get_cpus(cspmu);
1242 if (ret)
1243 return ret;
1244
1245 ret = arm_cspmu_init_impl_ops(cspmu);
1246 if (ret)
1247 return ret;
1248
1249 ret = arm_cspmu_register_pmu(cspmu);
1250
1251 /* Matches arm_cspmu_init_impl_ops() above. */
1252 if (cspmu->impl.module != THIS_MODULE)
1253 module_put(cspmu->impl.module);
1254
1255 return ret;
1256}
1257
1258static void arm_cspmu_device_remove(struct platform_device *pdev)
1259{
1260 struct arm_cspmu *cspmu = platform_get_drvdata(pdev);
1261
1262 perf_pmu_unregister(&cspmu->pmu);
1263 cpuhp_state_remove_instance(arm_cspmu_cpuhp_state, &cspmu->cpuhp_node);
1264}
1265
1266static const struct platform_device_id arm_cspmu_id[] = {
1267 {DRVNAME, 0},
1268 { },
1269};
1270MODULE_DEVICE_TABLE(platform, arm_cspmu_id);
1271
1272static const struct of_device_id arm_cspmu_of_match[] = {
1273 { .compatible = "arm,coresight-pmu" },
1274 {}
1275};
1276MODULE_DEVICE_TABLE(of, arm_cspmu_of_match);
1277
1278static struct platform_driver arm_cspmu_driver = {
1279 .driver = {
1280 .name = DRVNAME,
1281 .of_match_table = arm_cspmu_of_match,
1282 .suppress_bind_attrs = true,
1283 },
1284 .probe = arm_cspmu_device_probe,
1285 .remove = arm_cspmu_device_remove,
1286 .id_table = arm_cspmu_id,
1287};
1288
1289static void arm_cspmu_set_active_cpu(int cpu, struct arm_cspmu *cspmu)
1290{
1291 cpumask_set_cpu(cpu, &cspmu->active_cpu);
1292 if (cspmu->irq)
1293 WARN_ON(irq_set_affinity(cspmu->irq, &cspmu->active_cpu));
1294}
1295
1296static int arm_cspmu_cpu_online(unsigned int cpu, struct hlist_node *node)
1297{
1298 struct arm_cspmu *cspmu =
1299 hlist_entry_safe(node, struct arm_cspmu, cpuhp_node);
1300
1301 if (!cpumask_test_cpu(cpu, &cspmu->associated_cpus))
1302 return 0;
1303
1304 /* If the PMU is already managed, there is nothing to do */
1305 if (!cpumask_empty(&cspmu->active_cpu))
1306 return 0;
1307
1308 /* Use this CPU for event counting */
1309 arm_cspmu_set_active_cpu(cpu, cspmu);
1310
1311 return 0;
1312}
1313
1314static int arm_cspmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1315{
1316 unsigned int dst;
1317
1318 struct arm_cspmu *cspmu =
1319 hlist_entry_safe(node, struct arm_cspmu, cpuhp_node);
1320
1321 /* Nothing to do if this CPU doesn't own the PMU */
1322 if (!cpumask_test_and_clear_cpu(cpu, &cspmu->active_cpu))
1323 return 0;
1324
1325 /* Choose a new CPU to migrate ownership of the PMU to */
1326 dst = cpumask_any_and_but(&cspmu->associated_cpus,
1327 cpu_online_mask, cpu);
1328 if (dst >= nr_cpu_ids)
1329 return 0;
1330
1331 /* Use this CPU for event counting */
1332 perf_pmu_migrate_context(&cspmu->pmu, cpu, dst);
1333 arm_cspmu_set_active_cpu(dst, cspmu);
1334
1335 return 0;
1336}
1337
1338static int __init arm_cspmu_init(void)
1339{
1340 int ret;
1341
1342 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1343 "perf/arm/cspmu:online",
1344 arm_cspmu_cpu_online,
1345 arm_cspmu_cpu_teardown);
1346 if (ret < 0)
1347 return ret;
1348 arm_cspmu_cpuhp_state = ret;
1349 return platform_driver_register(&arm_cspmu_driver);
1350}
1351
1352static void __exit arm_cspmu_exit(void)
1353{
1354 platform_driver_unregister(&arm_cspmu_driver);
1355 cpuhp_remove_multi_state(arm_cspmu_cpuhp_state);
1356}
1357
1358int arm_cspmu_impl_register(const struct arm_cspmu_impl_match *impl_match)
1359{
1360 struct arm_cspmu_impl_match *match;
1361 int ret = 0;
1362
1363 match = arm_cspmu_impl_match_get(impl_match->pmiidr_val);
1364
1365 if (match) {
1366 mutex_lock(&arm_cspmu_lock);
1367
1368 if (!match->impl_init_ops) {
1369 match->module = impl_match->module;
1370 match->impl_init_ops = impl_match->impl_init_ops;
1371 } else {
1372 /* Broken match table may contain non-unique entries */
1373 WARN(1, "arm_cspmu backend already registered for module: %s, pmiidr: 0x%x, mask: 0x%x\n",
1374 match->module_name,
1375 match->pmiidr_val,
1376 match->pmiidr_mask);
1377
1378 ret = -EINVAL;
1379 }
1380
1381 mutex_unlock(&arm_cspmu_lock);
1382
1383 if (!ret)
1384 ret = driver_attach(&arm_cspmu_driver.driver);
1385 } else {
1386 pr_err("arm_cspmu reg failed, unable to find a match for pmiidr: 0x%x\n",
1387 impl_match->pmiidr_val);
1388
1389 ret = -EINVAL;
1390 }
1391
1392 return ret;
1393}
1394EXPORT_SYMBOL_GPL(arm_cspmu_impl_register);
1395
1396static int arm_cspmu_match_device(struct device *dev, const void *match)
1397{
1398 struct arm_cspmu *cspmu = platform_get_drvdata(to_platform_device(dev));
1399
1400 return (cspmu && cspmu->impl.match == match) ? 1 : 0;
1401}
1402
1403void arm_cspmu_impl_unregister(const struct arm_cspmu_impl_match *impl_match)
1404{
1405 struct device *dev;
1406 struct arm_cspmu_impl_match *match;
1407
1408 match = arm_cspmu_impl_match_get(impl_match->pmiidr_val);
1409
1410 if (WARN_ON(!match))
1411 return;
1412
1413 /* Unbind the driver from all matching backend devices. */
1414 while ((dev = driver_find_device(&arm_cspmu_driver.driver, NULL,
1415 match, arm_cspmu_match_device)))
1416 device_release_driver(dev);
1417
1418 mutex_lock(&arm_cspmu_lock);
1419
1420 match->module = NULL;
1421 match->impl_init_ops = NULL;
1422
1423 mutex_unlock(&arm_cspmu_lock);
1424}
1425EXPORT_SYMBOL_GPL(arm_cspmu_impl_unregister);
1426
1427module_init(arm_cspmu_init);
1428module_exit(arm_cspmu_exit);
1429
1430MODULE_DESCRIPTION("ARM CoreSight Architecture Performance Monitor Driver");
1431MODULE_LICENSE("GPL v2");
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * ARM CoreSight Architecture PMU driver.
4 *
5 * This driver adds support for uncore PMU based on ARM CoreSight Performance
6 * Monitoring Unit Architecture. The PMU is accessible via MMIO registers and
7 * like other uncore PMUs, it does not support process specific events and
8 * cannot be used in sampling mode.
9 *
10 * This code is based on other uncore PMUs like ARM DSU PMU. It provides a
11 * generic implementation to operate the PMU according to CoreSight PMU
12 * architecture and ACPI ARM PMU table (APMT) documents below:
13 * - ARM CoreSight PMU architecture document number: ARM IHI 0091 A.a-00bet0.
14 * - APMT document number: ARM DEN0117.
15 *
16 * The user should refer to the vendor technical documentation to get details
17 * about the supported events.
18 *
19 * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
20 *
21 */
22
23#include <linux/acpi.h>
24#include <linux/cacheinfo.h>
25#include <linux/ctype.h>
26#include <linux/interrupt.h>
27#include <linux/io-64-nonatomic-lo-hi.h>
28#include <linux/module.h>
29#include <linux/mutex.h>
30#include <linux/perf_event.h>
31#include <linux/platform_device.h>
32
33#include "arm_cspmu.h"
34
35#define PMUNAME "arm_cspmu"
36#define DRVNAME "arm-cs-arch-pmu"
37
38#define ARM_CSPMU_CPUMASK_ATTR(_name, _config) \
39 ARM_CSPMU_EXT_ATTR(_name, arm_cspmu_cpumask_show, \
40 (unsigned long)_config)
41
42/*
43 * CoreSight PMU Arch register offsets.
44 */
45#define PMEVCNTR_LO 0x0
46#define PMEVCNTR_HI 0x4
47#define PMEVTYPER 0x400
48#define PMCCFILTR 0x47C
49#define PMEVFILTR 0xA00
50#define PMCNTENSET 0xC00
51#define PMCNTENCLR 0xC20
52#define PMINTENSET 0xC40
53#define PMINTENCLR 0xC60
54#define PMOVSCLR 0xC80
55#define PMOVSSET 0xCC0
56#define PMCFGR 0xE00
57#define PMCR 0xE04
58#define PMIIDR 0xE08
59
60/* PMCFGR register field */
61#define PMCFGR_NCG GENMASK(31, 28)
62#define PMCFGR_HDBG BIT(24)
63#define PMCFGR_TRO BIT(23)
64#define PMCFGR_SS BIT(22)
65#define PMCFGR_FZO BIT(21)
66#define PMCFGR_MSI BIT(20)
67#define PMCFGR_UEN BIT(19)
68#define PMCFGR_NA BIT(17)
69#define PMCFGR_EX BIT(16)
70#define PMCFGR_CCD BIT(15)
71#define PMCFGR_CC BIT(14)
72#define PMCFGR_SIZE GENMASK(13, 8)
73#define PMCFGR_N GENMASK(7, 0)
74
75/* PMCR register field */
76#define PMCR_TRO BIT(11)
77#define PMCR_HDBG BIT(10)
78#define PMCR_FZO BIT(9)
79#define PMCR_NA BIT(8)
80#define PMCR_DP BIT(5)
81#define PMCR_X BIT(4)
82#define PMCR_D BIT(3)
83#define PMCR_C BIT(2)
84#define PMCR_P BIT(1)
85#define PMCR_E BIT(0)
86
87/* Each SET/CLR register supports up to 32 counters. */
88#define ARM_CSPMU_SET_CLR_COUNTER_SHIFT 5
89#define ARM_CSPMU_SET_CLR_COUNTER_NUM \
90 (1 << ARM_CSPMU_SET_CLR_COUNTER_SHIFT)
91
92/* Convert counter idx into SET/CLR register number. */
93#define COUNTER_TO_SET_CLR_ID(idx) \
94 (idx >> ARM_CSPMU_SET_CLR_COUNTER_SHIFT)
95
96/* Convert counter idx into SET/CLR register bit. */
97#define COUNTER_TO_SET_CLR_BIT(idx) \
98 (idx & (ARM_CSPMU_SET_CLR_COUNTER_NUM - 1))
99
100#define ARM_CSPMU_ACTIVE_CPU_MASK 0x0
101#define ARM_CSPMU_ASSOCIATED_CPU_MASK 0x1
102
103/* Check and use default if implementer doesn't provide attribute callback */
104#define CHECK_DEFAULT_IMPL_OPS(ops, callback) \
105 do { \
106 if (!ops->callback) \
107 ops->callback = arm_cspmu_ ## callback; \
108 } while (0)
109
110/*
111 * Maximum poll count for reading counter value using high-low-high sequence.
112 */
113#define HILOHI_MAX_POLL 1000
114
115static unsigned long arm_cspmu_cpuhp_state;
116
117static DEFINE_MUTEX(arm_cspmu_lock);
118
119static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
120 struct hw_perf_event *hwc, u32 filter);
121
122static struct acpi_apmt_node *arm_cspmu_apmt_node(struct device *dev)
123{
124 return *(struct acpi_apmt_node **)dev_get_platdata(dev);
125}
126
127/*
128 * In CoreSight PMU architecture, all of the MMIO registers are 32-bit except
129 * counter register. The counter register can be implemented as 32-bit or 64-bit
130 * register depending on the value of PMCFGR.SIZE field. For 64-bit access,
131 * single-copy 64-bit atomic support is implementation defined. APMT node flag
132 * is used to identify if the PMU supports 64-bit single copy atomic. If 64-bit
133 * single copy atomic is not supported, the driver treats the register as a pair
134 * of 32-bit register.
135 */
136
137/*
138 * Read 64-bit register as a pair of 32-bit registers using hi-lo-hi sequence.
139 */
140static u64 read_reg64_hilohi(const void __iomem *addr, u32 max_poll_count)
141{
142 u32 val_lo, val_hi;
143 u64 val;
144
145 /* Use high-low-high sequence to avoid tearing */
146 do {
147 if (max_poll_count-- == 0) {
148 pr_err("ARM CSPMU: timeout hi-low-high sequence\n");
149 return 0;
150 }
151
152 val_hi = readl(addr + 4);
153 val_lo = readl(addr);
154 } while (val_hi != readl(addr + 4));
155
156 val = (((u64)val_hi << 32) | val_lo);
157
158 return val;
159}
160
161/* Check if cycle counter is supported. */
162static inline bool supports_cycle_counter(const struct arm_cspmu *cspmu)
163{
164 return (cspmu->pmcfgr & PMCFGR_CC);
165}
166
167/* Get counter size, which is (PMCFGR_SIZE + 1). */
168static inline u32 counter_size(const struct arm_cspmu *cspmu)
169{
170 return FIELD_GET(PMCFGR_SIZE, cspmu->pmcfgr) + 1;
171}
172
173/* Get counter mask. */
174static inline u64 counter_mask(const struct arm_cspmu *cspmu)
175{
176 return GENMASK_ULL(counter_size(cspmu) - 1, 0);
177}
178
179/* Check if counter is implemented as 64-bit register. */
180static inline bool use_64b_counter_reg(const struct arm_cspmu *cspmu)
181{
182 return (counter_size(cspmu) > 32);
183}
184
185ssize_t arm_cspmu_sysfs_event_show(struct device *dev,
186 struct device_attribute *attr, char *buf)
187{
188 struct perf_pmu_events_attr *pmu_attr;
189
190 pmu_attr = container_of(attr, typeof(*pmu_attr), attr);
191 return sysfs_emit(buf, "event=0x%llx\n", pmu_attr->id);
192}
193EXPORT_SYMBOL_GPL(arm_cspmu_sysfs_event_show);
194
195/* Default event list. */
196static struct attribute *arm_cspmu_event_attrs[] = {
197 ARM_CSPMU_EVENT_ATTR(cycles, ARM_CSPMU_EVT_CYCLES_DEFAULT),
198 NULL,
199};
200
201static struct attribute **
202arm_cspmu_get_event_attrs(const struct arm_cspmu *cspmu)
203{
204 struct attribute **attrs;
205
206 attrs = devm_kmemdup(cspmu->dev, arm_cspmu_event_attrs,
207 sizeof(arm_cspmu_event_attrs), GFP_KERNEL);
208
209 return attrs;
210}
211
212static umode_t
213arm_cspmu_event_attr_is_visible(struct kobject *kobj,
214 struct attribute *attr, int unused)
215{
216 struct device *dev = kobj_to_dev(kobj);
217 struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev));
218 struct perf_pmu_events_attr *eattr;
219
220 eattr = container_of(attr, typeof(*eattr), attr.attr);
221
222 /* Hide cycle event if not supported */
223 if (!supports_cycle_counter(cspmu) &&
224 eattr->id == ARM_CSPMU_EVT_CYCLES_DEFAULT)
225 return 0;
226
227 return attr->mode;
228}
229
230ssize_t arm_cspmu_sysfs_format_show(struct device *dev,
231 struct device_attribute *attr,
232 char *buf)
233{
234 struct dev_ext_attribute *eattr =
235 container_of(attr, struct dev_ext_attribute, attr);
236 return sysfs_emit(buf, "%s\n", (char *)eattr->var);
237}
238EXPORT_SYMBOL_GPL(arm_cspmu_sysfs_format_show);
239
240static struct attribute *arm_cspmu_format_attrs[] = {
241 ARM_CSPMU_FORMAT_EVENT_ATTR,
242 ARM_CSPMU_FORMAT_FILTER_ATTR,
243 NULL,
244};
245
246static struct attribute **
247arm_cspmu_get_format_attrs(const struct arm_cspmu *cspmu)
248{
249 struct attribute **attrs;
250
251 attrs = devm_kmemdup(cspmu->dev, arm_cspmu_format_attrs,
252 sizeof(arm_cspmu_format_attrs), GFP_KERNEL);
253
254 return attrs;
255}
256
257static u32 arm_cspmu_event_type(const struct perf_event *event)
258{
259 return event->attr.config & ARM_CSPMU_EVENT_MASK;
260}
261
262static bool arm_cspmu_is_cycle_counter_event(const struct perf_event *event)
263{
264 return (event->attr.config == ARM_CSPMU_EVT_CYCLES_DEFAULT);
265}
266
267static u32 arm_cspmu_event_filter(const struct perf_event *event)
268{
269 return event->attr.config1 & ARM_CSPMU_FILTER_MASK;
270}
271
272static ssize_t arm_cspmu_identifier_show(struct device *dev,
273 struct device_attribute *attr,
274 char *page)
275{
276 struct arm_cspmu *cspmu = to_arm_cspmu(dev_get_drvdata(dev));
277
278 return sysfs_emit(page, "%s\n", cspmu->identifier);
279}
280
281static struct device_attribute arm_cspmu_identifier_attr =
282 __ATTR(identifier, 0444, arm_cspmu_identifier_show, NULL);
283
284static struct attribute *arm_cspmu_identifier_attrs[] = {
285 &arm_cspmu_identifier_attr.attr,
286 NULL,
287};
288
289static struct attribute_group arm_cspmu_identifier_attr_group = {
290 .attrs = arm_cspmu_identifier_attrs,
291};
292
293static const char *arm_cspmu_get_identifier(const struct arm_cspmu *cspmu)
294{
295 const char *identifier =
296 devm_kasprintf(cspmu->dev, GFP_KERNEL, "%x",
297 cspmu->impl.pmiidr);
298 return identifier;
299}
300
301static const char *arm_cspmu_type_str[ACPI_APMT_NODE_TYPE_COUNT] = {
302 "mc",
303 "smmu",
304 "pcie",
305 "acpi",
306 "cache",
307};
308
309static const char *arm_cspmu_get_name(const struct arm_cspmu *cspmu)
310{
311 struct device *dev;
312 struct acpi_apmt_node *apmt_node;
313 u8 pmu_type;
314 char *name;
315 char acpi_hid_string[ACPI_ID_LEN] = { 0 };
316 static atomic_t pmu_idx[ACPI_APMT_NODE_TYPE_COUNT] = { 0 };
317
318 dev = cspmu->dev;
319 apmt_node = arm_cspmu_apmt_node(dev);
320 pmu_type = apmt_node->type;
321
322 if (pmu_type >= ACPI_APMT_NODE_TYPE_COUNT) {
323 dev_err(dev, "unsupported PMU type-%u\n", pmu_type);
324 return NULL;
325 }
326
327 if (pmu_type == ACPI_APMT_NODE_TYPE_ACPI) {
328 memcpy(acpi_hid_string,
329 &apmt_node->inst_primary,
330 sizeof(apmt_node->inst_primary));
331 name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%s_%u", PMUNAME,
332 arm_cspmu_type_str[pmu_type],
333 acpi_hid_string,
334 apmt_node->inst_secondary);
335 } else {
336 name = devm_kasprintf(dev, GFP_KERNEL, "%s_%s_%d", PMUNAME,
337 arm_cspmu_type_str[pmu_type],
338 atomic_fetch_inc(&pmu_idx[pmu_type]));
339 }
340
341 return name;
342}
343
344static ssize_t arm_cspmu_cpumask_show(struct device *dev,
345 struct device_attribute *attr,
346 char *buf)
347{
348 struct pmu *pmu = dev_get_drvdata(dev);
349 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
350 struct dev_ext_attribute *eattr =
351 container_of(attr, struct dev_ext_attribute, attr);
352 unsigned long mask_id = (unsigned long)eattr->var;
353 const cpumask_t *cpumask;
354
355 switch (mask_id) {
356 case ARM_CSPMU_ACTIVE_CPU_MASK:
357 cpumask = &cspmu->active_cpu;
358 break;
359 case ARM_CSPMU_ASSOCIATED_CPU_MASK:
360 cpumask = &cspmu->associated_cpus;
361 break;
362 default:
363 return 0;
364 }
365 return cpumap_print_to_pagebuf(true, buf, cpumask);
366}
367
368static struct attribute *arm_cspmu_cpumask_attrs[] = {
369 ARM_CSPMU_CPUMASK_ATTR(cpumask, ARM_CSPMU_ACTIVE_CPU_MASK),
370 ARM_CSPMU_CPUMASK_ATTR(associated_cpus, ARM_CSPMU_ASSOCIATED_CPU_MASK),
371 NULL,
372};
373
374static struct attribute_group arm_cspmu_cpumask_attr_group = {
375 .attrs = arm_cspmu_cpumask_attrs,
376};
377
378static struct arm_cspmu_impl_match impl_match[] = {
379 {
380 .module_name = "nvidia_cspmu",
381 .pmiidr_val = ARM_CSPMU_IMPL_ID_NVIDIA,
382 .pmiidr_mask = ARM_CSPMU_PMIIDR_IMPLEMENTER,
383 .module = NULL,
384 .impl_init_ops = NULL,
385 },
386 {
387 .module_name = "ampere_cspmu",
388 .pmiidr_val = ARM_CSPMU_IMPL_ID_AMPERE,
389 .pmiidr_mask = ARM_CSPMU_PMIIDR_IMPLEMENTER,
390 .module = NULL,
391 .impl_init_ops = NULL,
392 },
393
394 {0}
395};
396
397static struct arm_cspmu_impl_match *arm_cspmu_impl_match_get(u32 pmiidr)
398{
399 struct arm_cspmu_impl_match *match = impl_match;
400
401 for (; match->pmiidr_val; match++) {
402 u32 mask = match->pmiidr_mask;
403
404 if ((match->pmiidr_val & mask) == (pmiidr & mask))
405 return match;
406 }
407
408 return NULL;
409}
410
411static int arm_cspmu_init_impl_ops(struct arm_cspmu *cspmu)
412{
413 int ret = 0;
414 struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
415 struct acpi_apmt_node *apmt_node = arm_cspmu_apmt_node(cspmu->dev);
416 struct arm_cspmu_impl_match *match;
417
418 /*
419 * Get PMU implementer and product id from APMT node.
420 * If APMT node doesn't have implementer/product id, try get it
421 * from PMIIDR.
422 */
423 cspmu->impl.pmiidr =
424 (apmt_node->impl_id) ? apmt_node->impl_id :
425 readl(cspmu->base0 + PMIIDR);
426
427 /* Find implementer specific attribute ops. */
428 match = arm_cspmu_impl_match_get(cspmu->impl.pmiidr);
429
430 /* Load implementer module and initialize the callbacks. */
431 if (match) {
432 mutex_lock(&arm_cspmu_lock);
433
434 if (match->impl_init_ops) {
435 /* Prevent unload until PMU registration is done. */
436 if (try_module_get(match->module)) {
437 cspmu->impl.module = match->module;
438 cspmu->impl.match = match;
439 ret = match->impl_init_ops(cspmu);
440 if (ret)
441 module_put(match->module);
442 } else {
443 WARN(1, "arm_cspmu failed to get module: %s\n",
444 match->module_name);
445 ret = -EINVAL;
446 }
447 } else {
448 request_module_nowait(match->module_name);
449 ret = -EPROBE_DEFER;
450 }
451
452 mutex_unlock(&arm_cspmu_lock);
453
454 if (ret)
455 return ret;
456 } else
457 cspmu->impl.module = THIS_MODULE;
458
459 /* Use default callbacks if implementer doesn't provide one. */
460 CHECK_DEFAULT_IMPL_OPS(impl_ops, get_event_attrs);
461 CHECK_DEFAULT_IMPL_OPS(impl_ops, get_format_attrs);
462 CHECK_DEFAULT_IMPL_OPS(impl_ops, get_identifier);
463 CHECK_DEFAULT_IMPL_OPS(impl_ops, get_name);
464 CHECK_DEFAULT_IMPL_OPS(impl_ops, is_cycle_counter_event);
465 CHECK_DEFAULT_IMPL_OPS(impl_ops, event_type);
466 CHECK_DEFAULT_IMPL_OPS(impl_ops, event_filter);
467 CHECK_DEFAULT_IMPL_OPS(impl_ops, event_attr_is_visible);
468 CHECK_DEFAULT_IMPL_OPS(impl_ops, set_ev_filter);
469
470 return 0;
471}
472
473static struct attribute_group *
474arm_cspmu_alloc_event_attr_group(struct arm_cspmu *cspmu)
475{
476 struct attribute_group *event_group;
477 struct device *dev = cspmu->dev;
478 const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
479
480 event_group =
481 devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL);
482 if (!event_group)
483 return NULL;
484
485 event_group->name = "events";
486 event_group->is_visible = impl_ops->event_attr_is_visible;
487 event_group->attrs = impl_ops->get_event_attrs(cspmu);
488
489 if (!event_group->attrs)
490 return NULL;
491
492 return event_group;
493}
494
495static struct attribute_group *
496arm_cspmu_alloc_format_attr_group(struct arm_cspmu *cspmu)
497{
498 struct attribute_group *format_group;
499 struct device *dev = cspmu->dev;
500
501 format_group =
502 devm_kzalloc(dev, sizeof(struct attribute_group), GFP_KERNEL);
503 if (!format_group)
504 return NULL;
505
506 format_group->name = "format";
507 format_group->attrs = cspmu->impl.ops.get_format_attrs(cspmu);
508
509 if (!format_group->attrs)
510 return NULL;
511
512 return format_group;
513}
514
515static struct attribute_group **
516arm_cspmu_alloc_attr_group(struct arm_cspmu *cspmu)
517{
518 struct attribute_group **attr_groups = NULL;
519 struct device *dev = cspmu->dev;
520 const struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
521
522 cspmu->identifier = impl_ops->get_identifier(cspmu);
523 cspmu->name = impl_ops->get_name(cspmu);
524
525 if (!cspmu->identifier || !cspmu->name)
526 return NULL;
527
528 attr_groups = devm_kcalloc(dev, 5, sizeof(struct attribute_group *),
529 GFP_KERNEL);
530 if (!attr_groups)
531 return NULL;
532
533 attr_groups[0] = arm_cspmu_alloc_event_attr_group(cspmu);
534 attr_groups[1] = arm_cspmu_alloc_format_attr_group(cspmu);
535 attr_groups[2] = &arm_cspmu_identifier_attr_group;
536 attr_groups[3] = &arm_cspmu_cpumask_attr_group;
537
538 if (!attr_groups[0] || !attr_groups[1])
539 return NULL;
540
541 return attr_groups;
542}
543
544static inline void arm_cspmu_reset_counters(struct arm_cspmu *cspmu)
545{
546 u32 pmcr = 0;
547
548 pmcr |= PMCR_P;
549 pmcr |= PMCR_C;
550 writel(pmcr, cspmu->base0 + PMCR);
551}
552
553static inline void arm_cspmu_start_counters(struct arm_cspmu *cspmu)
554{
555 writel(PMCR_E, cspmu->base0 + PMCR);
556}
557
558static inline void arm_cspmu_stop_counters(struct arm_cspmu *cspmu)
559{
560 writel(0, cspmu->base0 + PMCR);
561}
562
563static void arm_cspmu_enable(struct pmu *pmu)
564{
565 bool disabled;
566 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
567
568 disabled = bitmap_empty(cspmu->hw_events.used_ctrs,
569 cspmu->num_logical_ctrs);
570
571 if (disabled)
572 return;
573
574 arm_cspmu_start_counters(cspmu);
575}
576
577static void arm_cspmu_disable(struct pmu *pmu)
578{
579 struct arm_cspmu *cspmu = to_arm_cspmu(pmu);
580
581 arm_cspmu_stop_counters(cspmu);
582}
583
584static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
585 struct perf_event *event)
586{
587 int idx, ret;
588 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
589
590 if (supports_cycle_counter(cspmu)) {
591 if (cspmu->impl.ops.is_cycle_counter_event(event)) {
592 /* Search for available cycle counter. */
593 if (test_and_set_bit(cspmu->cycle_counter_logical_idx,
594 hw_events->used_ctrs))
595 return -EAGAIN;
596
597 return cspmu->cycle_counter_logical_idx;
598 }
599
600 /*
601 * Search a regular counter from the used counter bitmap.
602 * The cycle counter divides the bitmap into two parts. Search
603 * the first then second half to exclude the cycle counter bit.
604 */
605 idx = find_first_zero_bit(hw_events->used_ctrs,
606 cspmu->cycle_counter_logical_idx);
607 if (idx >= cspmu->cycle_counter_logical_idx) {
608 idx = find_next_zero_bit(
609 hw_events->used_ctrs,
610 cspmu->num_logical_ctrs,
611 cspmu->cycle_counter_logical_idx + 1);
612 }
613 } else {
614 idx = find_first_zero_bit(hw_events->used_ctrs,
615 cspmu->num_logical_ctrs);
616 }
617
618 if (idx >= cspmu->num_logical_ctrs)
619 return -EAGAIN;
620
621 if (cspmu->impl.ops.validate_event) {
622 ret = cspmu->impl.ops.validate_event(cspmu, event);
623 if (ret)
624 return ret;
625 }
626
627 set_bit(idx, hw_events->used_ctrs);
628
629 return idx;
630}
631
632static bool arm_cspmu_validate_event(struct pmu *pmu,
633 struct arm_cspmu_hw_events *hw_events,
634 struct perf_event *event)
635{
636 if (is_software_event(event))
637 return true;
638
639 /* Reject groups spanning multiple HW PMUs. */
640 if (event->pmu != pmu)
641 return false;
642
643 return (arm_cspmu_get_event_idx(hw_events, event) >= 0);
644}
645
646/*
647 * Make sure the group of events can be scheduled at once
648 * on the PMU.
649 */
650static bool arm_cspmu_validate_group(struct perf_event *event)
651{
652 struct perf_event *sibling, *leader = event->group_leader;
653 struct arm_cspmu_hw_events fake_hw_events;
654
655 if (event->group_leader == event)
656 return true;
657
658 memset(&fake_hw_events, 0, sizeof(fake_hw_events));
659
660 if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events, leader))
661 return false;
662
663 for_each_sibling_event(sibling, leader) {
664 if (!arm_cspmu_validate_event(event->pmu, &fake_hw_events,
665 sibling))
666 return false;
667 }
668
669 return arm_cspmu_validate_event(event->pmu, &fake_hw_events, event);
670}
671
672static int arm_cspmu_event_init(struct perf_event *event)
673{
674 struct arm_cspmu *cspmu;
675 struct hw_perf_event *hwc = &event->hw;
676
677 cspmu = to_arm_cspmu(event->pmu);
678
679 if (event->attr.type != event->pmu->type)
680 return -ENOENT;
681
682 /*
683 * Following other "uncore" PMUs, we do not support sampling mode or
684 * attach to a task (per-process mode).
685 */
686 if (is_sampling_event(event)) {
687 dev_dbg(cspmu->pmu.dev,
688 "Can't support sampling events\n");
689 return -EOPNOTSUPP;
690 }
691
692 if (event->cpu < 0 || event->attach_state & PERF_ATTACH_TASK) {
693 dev_dbg(cspmu->pmu.dev,
694 "Can't support per-task counters\n");
695 return -EINVAL;
696 }
697
698 /*
699 * Make sure the CPU assignment is on one of the CPUs associated with
700 * this PMU.
701 */
702 if (!cpumask_test_cpu(event->cpu, &cspmu->associated_cpus)) {
703 dev_dbg(cspmu->pmu.dev,
704 "Requested cpu is not associated with the PMU\n");
705 return -EINVAL;
706 }
707
708 /* Enforce the current active CPU to handle the events in this PMU. */
709 event->cpu = cpumask_first(&cspmu->active_cpu);
710 if (event->cpu >= nr_cpu_ids)
711 return -EINVAL;
712
713 if (!arm_cspmu_validate_group(event))
714 return -EINVAL;
715
716 /*
717 * The logical counter id is tracked with hw_perf_event.extra_reg.idx.
718 * The physical counter id is tracked with hw_perf_event.idx.
719 * We don't assign an index until we actually place the event onto
720 * hardware. Use -1 to signify that we haven't decided where to put it
721 * yet.
722 */
723 hwc->idx = -1;
724 hwc->extra_reg.idx = -1;
725 hwc->config = cspmu->impl.ops.event_type(event);
726
727 return 0;
728}
729
730static inline u32 counter_offset(u32 reg_sz, u32 ctr_idx)
731{
732 return (PMEVCNTR_LO + (reg_sz * ctr_idx));
733}
734
735static void arm_cspmu_write_counter(struct perf_event *event, u64 val)
736{
737 u32 offset;
738 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
739
740 if (use_64b_counter_reg(cspmu)) {
741 offset = counter_offset(sizeof(u64), event->hw.idx);
742
743 if (cspmu->has_atomic_dword)
744 writeq(val, cspmu->base1 + offset);
745 else
746 lo_hi_writeq(val, cspmu->base1 + offset);
747 } else {
748 offset = counter_offset(sizeof(u32), event->hw.idx);
749
750 writel(lower_32_bits(val), cspmu->base1 + offset);
751 }
752}
753
754static u64 arm_cspmu_read_counter(struct perf_event *event)
755{
756 u32 offset;
757 const void __iomem *counter_addr;
758 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
759
760 if (use_64b_counter_reg(cspmu)) {
761 offset = counter_offset(sizeof(u64), event->hw.idx);
762 counter_addr = cspmu->base1 + offset;
763
764 return cspmu->has_atomic_dword ?
765 readq(counter_addr) :
766 read_reg64_hilohi(counter_addr, HILOHI_MAX_POLL);
767 }
768
769 offset = counter_offset(sizeof(u32), event->hw.idx);
770 return readl(cspmu->base1 + offset);
771}
772
773/*
774 * arm_cspmu_set_event_period: Set the period for the counter.
775 *
776 * To handle cases of extreme interrupt latency, we program
777 * the counter with half of the max count for the counters.
778 */
779static void arm_cspmu_set_event_period(struct perf_event *event)
780{
781 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
782 u64 val = counter_mask(cspmu) >> 1ULL;
783
784 local64_set(&event->hw.prev_count, val);
785 arm_cspmu_write_counter(event, val);
786}
787
788static void arm_cspmu_enable_counter(struct arm_cspmu *cspmu, int idx)
789{
790 u32 reg_id, reg_bit, inten_off, cnten_off;
791
792 reg_id = COUNTER_TO_SET_CLR_ID(idx);
793 reg_bit = COUNTER_TO_SET_CLR_BIT(idx);
794
795 inten_off = PMINTENSET + (4 * reg_id);
796 cnten_off = PMCNTENSET + (4 * reg_id);
797
798 writel(BIT(reg_bit), cspmu->base0 + inten_off);
799 writel(BIT(reg_bit), cspmu->base0 + cnten_off);
800}
801
802static void arm_cspmu_disable_counter(struct arm_cspmu *cspmu, int idx)
803{
804 u32 reg_id, reg_bit, inten_off, cnten_off;
805
806 reg_id = COUNTER_TO_SET_CLR_ID(idx);
807 reg_bit = COUNTER_TO_SET_CLR_BIT(idx);
808
809 inten_off = PMINTENCLR + (4 * reg_id);
810 cnten_off = PMCNTENCLR + (4 * reg_id);
811
812 writel(BIT(reg_bit), cspmu->base0 + cnten_off);
813 writel(BIT(reg_bit), cspmu->base0 + inten_off);
814}
815
816static void arm_cspmu_event_update(struct perf_event *event)
817{
818 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
819 struct hw_perf_event *hwc = &event->hw;
820 u64 delta, prev, now;
821
822 do {
823 prev = local64_read(&hwc->prev_count);
824 now = arm_cspmu_read_counter(event);
825 } while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
826
827 delta = (now - prev) & counter_mask(cspmu);
828 local64_add(delta, &event->count);
829}
830
831static inline void arm_cspmu_set_event(struct arm_cspmu *cspmu,
832 struct hw_perf_event *hwc)
833{
834 u32 offset = PMEVTYPER + (4 * hwc->idx);
835
836 writel(hwc->config, cspmu->base0 + offset);
837}
838
839static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
840 struct hw_perf_event *hwc,
841 u32 filter)
842{
843 u32 offset = PMEVFILTR + (4 * hwc->idx);
844
845 writel(filter, cspmu->base0 + offset);
846}
847
848static inline void arm_cspmu_set_cc_filter(struct arm_cspmu *cspmu, u32 filter)
849{
850 u32 offset = PMCCFILTR;
851
852 writel(filter, cspmu->base0 + offset);
853}
854
855static void arm_cspmu_start(struct perf_event *event, int pmu_flags)
856{
857 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
858 struct hw_perf_event *hwc = &event->hw;
859 u32 filter;
860
861 /* We always reprogram the counter */
862 if (pmu_flags & PERF_EF_RELOAD)
863 WARN_ON(!(hwc->state & PERF_HES_UPTODATE));
864
865 arm_cspmu_set_event_period(event);
866
867 filter = cspmu->impl.ops.event_filter(event);
868
869 if (event->hw.extra_reg.idx == cspmu->cycle_counter_logical_idx) {
870 arm_cspmu_set_cc_filter(cspmu, filter);
871 } else {
872 arm_cspmu_set_event(cspmu, hwc);
873 cspmu->impl.ops.set_ev_filter(cspmu, hwc, filter);
874 }
875
876 hwc->state = 0;
877
878 arm_cspmu_enable_counter(cspmu, hwc->idx);
879}
880
881static void arm_cspmu_stop(struct perf_event *event, int pmu_flags)
882{
883 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
884 struct hw_perf_event *hwc = &event->hw;
885
886 if (hwc->state & PERF_HES_STOPPED)
887 return;
888
889 arm_cspmu_disable_counter(cspmu, hwc->idx);
890 arm_cspmu_event_update(event);
891
892 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
893}
894
895static inline u32 to_phys_idx(struct arm_cspmu *cspmu, u32 idx)
896{
897 return (idx == cspmu->cycle_counter_logical_idx) ?
898 ARM_CSPMU_CYCLE_CNTR_IDX : idx;
899}
900
901static int arm_cspmu_add(struct perf_event *event, int flags)
902{
903 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
904 struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events;
905 struct hw_perf_event *hwc = &event->hw;
906 int idx;
907
908 if (WARN_ON_ONCE(!cpumask_test_cpu(smp_processor_id(),
909 &cspmu->associated_cpus)))
910 return -ENOENT;
911
912 idx = arm_cspmu_get_event_idx(hw_events, event);
913 if (idx < 0)
914 return idx;
915
916 hw_events->events[idx] = event;
917 hwc->idx = to_phys_idx(cspmu, idx);
918 hwc->extra_reg.idx = idx;
919 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
920
921 if (flags & PERF_EF_START)
922 arm_cspmu_start(event, PERF_EF_RELOAD);
923
924 /* Propagate changes to the userspace mapping. */
925 perf_event_update_userpage(event);
926
927 return 0;
928}
929
930static void arm_cspmu_del(struct perf_event *event, int flags)
931{
932 struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
933 struct arm_cspmu_hw_events *hw_events = &cspmu->hw_events;
934 struct hw_perf_event *hwc = &event->hw;
935 int idx = hwc->extra_reg.idx;
936
937 arm_cspmu_stop(event, PERF_EF_UPDATE);
938
939 hw_events->events[idx] = NULL;
940
941 clear_bit(idx, hw_events->used_ctrs);
942
943 perf_event_update_userpage(event);
944}
945
946static void arm_cspmu_read(struct perf_event *event)
947{
948 arm_cspmu_event_update(event);
949}
950
951static struct arm_cspmu *arm_cspmu_alloc(struct platform_device *pdev)
952{
953 struct acpi_apmt_node *apmt_node;
954 struct arm_cspmu *cspmu;
955 struct device *dev = &pdev->dev;
956
957 cspmu = devm_kzalloc(dev, sizeof(*cspmu), GFP_KERNEL);
958 if (!cspmu)
959 return NULL;
960
961 cspmu->dev = dev;
962 platform_set_drvdata(pdev, cspmu);
963
964 apmt_node = arm_cspmu_apmt_node(dev);
965 cspmu->has_atomic_dword = apmt_node->flags & ACPI_APMT_FLAGS_ATOMIC;
966
967 return cspmu;
968}
969
970static int arm_cspmu_init_mmio(struct arm_cspmu *cspmu)
971{
972 struct device *dev;
973 struct platform_device *pdev;
974
975 dev = cspmu->dev;
976 pdev = to_platform_device(dev);
977
978 /* Base address for page 0. */
979 cspmu->base0 = devm_platform_ioremap_resource(pdev, 0);
980 if (IS_ERR(cspmu->base0)) {
981 dev_err(dev, "ioremap failed for page-0 resource\n");
982 return PTR_ERR(cspmu->base0);
983 }
984
985 /* Base address for page 1 if supported. Otherwise point to page 0. */
986 cspmu->base1 = cspmu->base0;
987 if (platform_get_resource(pdev, IORESOURCE_MEM, 1)) {
988 cspmu->base1 = devm_platform_ioremap_resource(pdev, 1);
989 if (IS_ERR(cspmu->base1)) {
990 dev_err(dev, "ioremap failed for page-1 resource\n");
991 return PTR_ERR(cspmu->base1);
992 }
993 }
994
995 cspmu->pmcfgr = readl(cspmu->base0 + PMCFGR);
996
997 cspmu->num_logical_ctrs = FIELD_GET(PMCFGR_N, cspmu->pmcfgr) + 1;
998
999 cspmu->cycle_counter_logical_idx = ARM_CSPMU_MAX_HW_CNTRS;
1000
1001 if (supports_cycle_counter(cspmu)) {
1002 /*
1003 * The last logical counter is mapped to cycle counter if
1004 * there is a gap between regular and cycle counter. Otherwise,
1005 * logical and physical have 1-to-1 mapping.
1006 */
1007 cspmu->cycle_counter_logical_idx =
1008 (cspmu->num_logical_ctrs <= ARM_CSPMU_CYCLE_CNTR_IDX) ?
1009 cspmu->num_logical_ctrs - 1 :
1010 ARM_CSPMU_CYCLE_CNTR_IDX;
1011 }
1012
1013 cspmu->num_set_clr_reg =
1014 DIV_ROUND_UP(cspmu->num_logical_ctrs,
1015 ARM_CSPMU_SET_CLR_COUNTER_NUM);
1016
1017 cspmu->hw_events.events =
1018 devm_kcalloc(dev, cspmu->num_logical_ctrs,
1019 sizeof(*cspmu->hw_events.events), GFP_KERNEL);
1020
1021 if (!cspmu->hw_events.events)
1022 return -ENOMEM;
1023
1024 return 0;
1025}
1026
1027static inline int arm_cspmu_get_reset_overflow(struct arm_cspmu *cspmu,
1028 u32 *pmovs)
1029{
1030 int i;
1031 u32 pmovclr_offset = PMOVSCLR;
1032 u32 has_overflowed = 0;
1033
1034 for (i = 0; i < cspmu->num_set_clr_reg; ++i) {
1035 pmovs[i] = readl(cspmu->base1 + pmovclr_offset);
1036 has_overflowed |= pmovs[i];
1037 writel(pmovs[i], cspmu->base1 + pmovclr_offset);
1038 pmovclr_offset += sizeof(u32);
1039 }
1040
1041 return has_overflowed != 0;
1042}
1043
1044static irqreturn_t arm_cspmu_handle_irq(int irq_num, void *dev)
1045{
1046 int idx, has_overflowed;
1047 struct perf_event *event;
1048 struct arm_cspmu *cspmu = dev;
1049 DECLARE_BITMAP(pmovs, ARM_CSPMU_MAX_HW_CNTRS);
1050 bool handled = false;
1051
1052 arm_cspmu_stop_counters(cspmu);
1053
1054 has_overflowed = arm_cspmu_get_reset_overflow(cspmu, (u32 *)pmovs);
1055 if (!has_overflowed)
1056 goto done;
1057
1058 for_each_set_bit(idx, cspmu->hw_events.used_ctrs,
1059 cspmu->num_logical_ctrs) {
1060 event = cspmu->hw_events.events[idx];
1061
1062 if (!event)
1063 continue;
1064
1065 if (!test_bit(event->hw.idx, pmovs))
1066 continue;
1067
1068 arm_cspmu_event_update(event);
1069 arm_cspmu_set_event_period(event);
1070
1071 handled = true;
1072 }
1073
1074done:
1075 arm_cspmu_start_counters(cspmu);
1076 return IRQ_RETVAL(handled);
1077}
1078
1079static int arm_cspmu_request_irq(struct arm_cspmu *cspmu)
1080{
1081 int irq, ret;
1082 struct device *dev;
1083 struct platform_device *pdev;
1084
1085 dev = cspmu->dev;
1086 pdev = to_platform_device(dev);
1087
1088 /* Skip IRQ request if the PMU does not support overflow interrupt. */
1089 irq = platform_get_irq_optional(pdev, 0);
1090 if (irq < 0)
1091 return irq == -ENXIO ? 0 : irq;
1092
1093 ret = devm_request_irq(dev, irq, arm_cspmu_handle_irq,
1094 IRQF_NOBALANCING | IRQF_NO_THREAD, dev_name(dev),
1095 cspmu);
1096 if (ret) {
1097 dev_err(dev, "Could not request IRQ %d\n", irq);
1098 return ret;
1099 }
1100
1101 cspmu->irq = irq;
1102
1103 return 0;
1104}
1105
1106#if defined(CONFIG_ACPI) && defined(CONFIG_ARM64)
1107#include <acpi/processor.h>
1108
1109static inline int arm_cspmu_find_cpu_container(int cpu, u32 container_uid)
1110{
1111 struct device *cpu_dev;
1112 struct acpi_device *acpi_dev;
1113
1114 cpu_dev = get_cpu_device(cpu);
1115 if (!cpu_dev)
1116 return -ENODEV;
1117
1118 acpi_dev = ACPI_COMPANION(cpu_dev);
1119 while (acpi_dev) {
1120 if (acpi_dev_hid_uid_match(acpi_dev, ACPI_PROCESSOR_CONTAINER_HID, container_uid))
1121 return 0;
1122
1123 acpi_dev = acpi_dev_parent(acpi_dev);
1124 }
1125
1126 return -ENODEV;
1127}
1128
1129static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu)
1130{
1131 struct acpi_apmt_node *apmt_node;
1132 int affinity_flag;
1133 int cpu;
1134
1135 apmt_node = arm_cspmu_apmt_node(cspmu->dev);
1136 affinity_flag = apmt_node->flags & ACPI_APMT_FLAGS_AFFINITY;
1137
1138 if (affinity_flag == ACPI_APMT_FLAGS_AFFINITY_PROC) {
1139 for_each_possible_cpu(cpu) {
1140 if (apmt_node->proc_affinity ==
1141 get_acpi_id_for_cpu(cpu)) {
1142 cpumask_set_cpu(cpu, &cspmu->associated_cpus);
1143 break;
1144 }
1145 }
1146 } else {
1147 for_each_possible_cpu(cpu) {
1148 if (arm_cspmu_find_cpu_container(
1149 cpu, apmt_node->proc_affinity))
1150 continue;
1151
1152 cpumask_set_cpu(cpu, &cspmu->associated_cpus);
1153 }
1154 }
1155
1156 if (cpumask_empty(&cspmu->associated_cpus)) {
1157 dev_dbg(cspmu->dev, "No cpu associated with the PMU\n");
1158 return -ENODEV;
1159 }
1160
1161 return 0;
1162}
1163#else
1164static int arm_cspmu_acpi_get_cpus(struct arm_cspmu *cspmu)
1165{
1166 return -ENODEV;
1167}
1168#endif
1169
1170static int arm_cspmu_get_cpus(struct arm_cspmu *cspmu)
1171{
1172 return arm_cspmu_acpi_get_cpus(cspmu);
1173}
1174
1175static int arm_cspmu_register_pmu(struct arm_cspmu *cspmu)
1176{
1177 int ret, capabilities;
1178 struct attribute_group **attr_groups;
1179
1180 attr_groups = arm_cspmu_alloc_attr_group(cspmu);
1181 if (!attr_groups)
1182 return -ENOMEM;
1183
1184 ret = cpuhp_state_add_instance(arm_cspmu_cpuhp_state,
1185 &cspmu->cpuhp_node);
1186 if (ret)
1187 return ret;
1188
1189 capabilities = PERF_PMU_CAP_NO_EXCLUDE;
1190 if (cspmu->irq == 0)
1191 capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1192
1193 cspmu->pmu = (struct pmu){
1194 .task_ctx_nr = perf_invalid_context,
1195 .module = cspmu->impl.module,
1196 .pmu_enable = arm_cspmu_enable,
1197 .pmu_disable = arm_cspmu_disable,
1198 .event_init = arm_cspmu_event_init,
1199 .add = arm_cspmu_add,
1200 .del = arm_cspmu_del,
1201 .start = arm_cspmu_start,
1202 .stop = arm_cspmu_stop,
1203 .read = arm_cspmu_read,
1204 .attr_groups = (const struct attribute_group **)attr_groups,
1205 .capabilities = capabilities,
1206 };
1207
1208 /* Hardware counter init */
1209 arm_cspmu_stop_counters(cspmu);
1210 arm_cspmu_reset_counters(cspmu);
1211
1212 ret = perf_pmu_register(&cspmu->pmu, cspmu->name, -1);
1213 if (ret) {
1214 cpuhp_state_remove_instance(arm_cspmu_cpuhp_state,
1215 &cspmu->cpuhp_node);
1216 }
1217
1218 return ret;
1219}
1220
1221static int arm_cspmu_device_probe(struct platform_device *pdev)
1222{
1223 int ret;
1224 struct arm_cspmu *cspmu;
1225
1226 cspmu = arm_cspmu_alloc(pdev);
1227 if (!cspmu)
1228 return -ENOMEM;
1229
1230 ret = arm_cspmu_init_mmio(cspmu);
1231 if (ret)
1232 return ret;
1233
1234 ret = arm_cspmu_request_irq(cspmu);
1235 if (ret)
1236 return ret;
1237
1238 ret = arm_cspmu_get_cpus(cspmu);
1239 if (ret)
1240 return ret;
1241
1242 ret = arm_cspmu_init_impl_ops(cspmu);
1243 if (ret)
1244 return ret;
1245
1246 ret = arm_cspmu_register_pmu(cspmu);
1247
1248 /* Matches arm_cspmu_init_impl_ops() above. */
1249 if (cspmu->impl.module != THIS_MODULE)
1250 module_put(cspmu->impl.module);
1251
1252 return ret;
1253}
1254
1255static int arm_cspmu_device_remove(struct platform_device *pdev)
1256{
1257 struct arm_cspmu *cspmu = platform_get_drvdata(pdev);
1258
1259 perf_pmu_unregister(&cspmu->pmu);
1260 cpuhp_state_remove_instance(arm_cspmu_cpuhp_state, &cspmu->cpuhp_node);
1261
1262 return 0;
1263}
1264
1265static const struct platform_device_id arm_cspmu_id[] = {
1266 {DRVNAME, 0},
1267 { },
1268};
1269MODULE_DEVICE_TABLE(platform, arm_cspmu_id);
1270
1271static struct platform_driver arm_cspmu_driver = {
1272 .driver = {
1273 .name = DRVNAME,
1274 .suppress_bind_attrs = true,
1275 },
1276 .probe = arm_cspmu_device_probe,
1277 .remove = arm_cspmu_device_remove,
1278 .id_table = arm_cspmu_id,
1279};
1280
1281static void arm_cspmu_set_active_cpu(int cpu, struct arm_cspmu *cspmu)
1282{
1283 cpumask_set_cpu(cpu, &cspmu->active_cpu);
1284 if (cspmu->irq)
1285 WARN_ON(irq_set_affinity(cspmu->irq, &cspmu->active_cpu));
1286}
1287
1288static int arm_cspmu_cpu_online(unsigned int cpu, struct hlist_node *node)
1289{
1290 struct arm_cspmu *cspmu =
1291 hlist_entry_safe(node, struct arm_cspmu, cpuhp_node);
1292
1293 if (!cpumask_test_cpu(cpu, &cspmu->associated_cpus))
1294 return 0;
1295
1296 /* If the PMU is already managed, there is nothing to do */
1297 if (!cpumask_empty(&cspmu->active_cpu))
1298 return 0;
1299
1300 /* Use this CPU for event counting */
1301 arm_cspmu_set_active_cpu(cpu, cspmu);
1302
1303 return 0;
1304}
1305
1306static int arm_cspmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1307{
1308 int dst;
1309 struct cpumask online_supported;
1310
1311 struct arm_cspmu *cspmu =
1312 hlist_entry_safe(node, struct arm_cspmu, cpuhp_node);
1313
1314 /* Nothing to do if this CPU doesn't own the PMU */
1315 if (!cpumask_test_and_clear_cpu(cpu, &cspmu->active_cpu))
1316 return 0;
1317
1318 /* Choose a new CPU to migrate ownership of the PMU to */
1319 cpumask_and(&online_supported, &cspmu->associated_cpus,
1320 cpu_online_mask);
1321 dst = cpumask_any_but(&online_supported, cpu);
1322 if (dst >= nr_cpu_ids)
1323 return 0;
1324
1325 /* Use this CPU for event counting */
1326 perf_pmu_migrate_context(&cspmu->pmu, cpu, dst);
1327 arm_cspmu_set_active_cpu(dst, cspmu);
1328
1329 return 0;
1330}
1331
1332static int __init arm_cspmu_init(void)
1333{
1334 int ret;
1335
1336 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1337 "perf/arm/cspmu:online",
1338 arm_cspmu_cpu_online,
1339 arm_cspmu_cpu_teardown);
1340 if (ret < 0)
1341 return ret;
1342 arm_cspmu_cpuhp_state = ret;
1343 return platform_driver_register(&arm_cspmu_driver);
1344}
1345
1346static void __exit arm_cspmu_exit(void)
1347{
1348 platform_driver_unregister(&arm_cspmu_driver);
1349 cpuhp_remove_multi_state(arm_cspmu_cpuhp_state);
1350}
1351
1352int arm_cspmu_impl_register(const struct arm_cspmu_impl_match *impl_match)
1353{
1354 struct arm_cspmu_impl_match *match;
1355 int ret = 0;
1356
1357 match = arm_cspmu_impl_match_get(impl_match->pmiidr_val);
1358
1359 if (match) {
1360 mutex_lock(&arm_cspmu_lock);
1361
1362 if (!match->impl_init_ops) {
1363 match->module = impl_match->module;
1364 match->impl_init_ops = impl_match->impl_init_ops;
1365 } else {
1366 /* Broken match table may contain non-unique entries */
1367 WARN(1, "arm_cspmu backend already registered for module: %s, pmiidr: 0x%x, mask: 0x%x\n",
1368 match->module_name,
1369 match->pmiidr_val,
1370 match->pmiidr_mask);
1371
1372 ret = -EINVAL;
1373 }
1374
1375 mutex_unlock(&arm_cspmu_lock);
1376
1377 if (!ret)
1378 ret = driver_attach(&arm_cspmu_driver.driver);
1379 } else {
1380 pr_err("arm_cspmu reg failed, unable to find a match for pmiidr: 0x%x\n",
1381 impl_match->pmiidr_val);
1382
1383 ret = -EINVAL;
1384 }
1385
1386 return ret;
1387}
1388EXPORT_SYMBOL_GPL(arm_cspmu_impl_register);
1389
1390static int arm_cspmu_match_device(struct device *dev, const void *match)
1391{
1392 struct arm_cspmu *cspmu = platform_get_drvdata(to_platform_device(dev));
1393
1394 return (cspmu && cspmu->impl.match == match) ? 1 : 0;
1395}
1396
1397void arm_cspmu_impl_unregister(const struct arm_cspmu_impl_match *impl_match)
1398{
1399 struct device *dev;
1400 struct arm_cspmu_impl_match *match;
1401
1402 match = arm_cspmu_impl_match_get(impl_match->pmiidr_val);
1403
1404 if (WARN_ON(!match))
1405 return;
1406
1407 /* Unbind the driver from all matching backend devices. */
1408 while ((dev = driver_find_device(&arm_cspmu_driver.driver, NULL,
1409 match, arm_cspmu_match_device)))
1410 device_release_driver(dev);
1411
1412 mutex_lock(&arm_cspmu_lock);
1413
1414 match->module = NULL;
1415 match->impl_init_ops = NULL;
1416
1417 mutex_unlock(&arm_cspmu_lock);
1418}
1419EXPORT_SYMBOL_GPL(arm_cspmu_impl_unregister);
1420
1421module_init(arm_cspmu_init);
1422module_exit(arm_cspmu_exit);
1423
1424MODULE_LICENSE("GPL v2");