1 // SPDX-License-Identifier: GPL-2.0
3 * CPU PMU driver for the Apple M1 and derivatives
5 * Copyright (C) 2021 Google LLC
7 * Author: Marc Zyngier <maz@kernel.org>
9 * Most of the information used in this driver was provided by the
10 * Asahi Linux project. The rest was experimentally discovered.
14 #include <linux/perf/arm_pmu.h>
15 #include <linux/platform_device.h>
17 #include <asm/apple_m1_pmu.h>
18 #include <asm/irq_regs.h>
19 #include <asm/perf_event.h>
21 #define M1_PMU_NR_COUNTERS 10
23 #define M1_PMU_CFG_EVENT GENMASK(7, 0)
25 #define ANY_BUT_0_1 GENMASK(9, 2)
26 #define ONLY_2_TO_7 GENMASK(7, 2)
27 #define ONLY_2_4_6 (BIT(2) | BIT(4) | BIT(6))
28 #define ONLY_5_6_7 (BIT(5) | BIT(6) | BIT(7))
31 * Description of the events we actually know about, as well as those with
32 * a specific counter affinity. Yes, this is a grand total of two known
33 * counters, and the rest is anybody's guess.
35 * Not all counters can count all events. Counters #0 and #1 are wired to
36 * count cycles and instructions respectively, and some events have
37 * bizarre mappings (every other counter, or even *one* counter). These
38 * restrictions equally apply to both P and E cores.
40 * It is worth noting that the PMUs attached to P and E cores are likely
41 * to be different because the underlying uarches are different. At the
42 * moment, we don't really need to distinguish between the two because we
43 * know next to nothing about the events themselves, and we already have
44 * per cpu-type PMU abstractions.
46 * If we eventually find out that the events are different across
47 * implementations, we'll have to introduce per cpu-type tables.
50 M1_PMU_PERFCTR_UNKNOWN_01 = 0x01,
51 M1_PMU_PERFCTR_CPU_CYCLES = 0x02,
52 M1_PMU_PERFCTR_INSTRUCTIONS = 0x8c,
53 M1_PMU_PERFCTR_UNKNOWN_8d = 0x8d,
54 M1_PMU_PERFCTR_UNKNOWN_8e = 0x8e,
55 M1_PMU_PERFCTR_UNKNOWN_8f = 0x8f,
56 M1_PMU_PERFCTR_UNKNOWN_90 = 0x90,
57 M1_PMU_PERFCTR_UNKNOWN_93 = 0x93,
58 M1_PMU_PERFCTR_UNKNOWN_94 = 0x94,
59 M1_PMU_PERFCTR_UNKNOWN_95 = 0x95,
60 M1_PMU_PERFCTR_UNKNOWN_96 = 0x96,
61 M1_PMU_PERFCTR_UNKNOWN_97 = 0x97,
62 M1_PMU_PERFCTR_UNKNOWN_98 = 0x98,
63 M1_PMU_PERFCTR_UNKNOWN_99 = 0x99,
64 M1_PMU_PERFCTR_UNKNOWN_9a = 0x9a,
65 M1_PMU_PERFCTR_UNKNOWN_9b = 0x9b,
66 M1_PMU_PERFCTR_UNKNOWN_9c = 0x9c,
67 M1_PMU_PERFCTR_UNKNOWN_9f = 0x9f,
68 M1_PMU_PERFCTR_UNKNOWN_bf = 0xbf,
69 M1_PMU_PERFCTR_UNKNOWN_c0 = 0xc0,
70 M1_PMU_PERFCTR_UNKNOWN_c1 = 0xc1,
71 M1_PMU_PERFCTR_UNKNOWN_c4 = 0xc4,
72 M1_PMU_PERFCTR_UNKNOWN_c5 = 0xc5,
73 M1_PMU_PERFCTR_UNKNOWN_c6 = 0xc6,
74 M1_PMU_PERFCTR_UNKNOWN_c8 = 0xc8,
75 M1_PMU_PERFCTR_UNKNOWN_ca = 0xca,
76 M1_PMU_PERFCTR_UNKNOWN_cb = 0xcb,
77 M1_PMU_PERFCTR_UNKNOWN_f5 = 0xf5,
78 M1_PMU_PERFCTR_UNKNOWN_f6 = 0xf6,
79 M1_PMU_PERFCTR_UNKNOWN_f7 = 0xf7,
80 M1_PMU_PERFCTR_UNKNOWN_f8 = 0xf8,
81 M1_PMU_PERFCTR_UNKNOWN_fd = 0xfd,
82 M1_PMU_PERFCTR_LAST = M1_PMU_CFG_EVENT,
85 * From this point onwards, these are not actual HW events,
86 * but attributes that get stored in hw->config_base.
88 M1_PMU_CFG_COUNT_USER = BIT(8),
89 M1_PMU_CFG_COUNT_KERNEL = BIT(9),
93 * Per-event affinity table. Most events can be installed on counter
94 * 2-9, but there are a number of exceptions. Note that this table
95 * has been created experimentally, and I wouldn't be surprised if more
96 * counters had strange affinities.
98 static const u16 m1_pmu_event_affinity[M1_PMU_PERFCTR_LAST + 1] = {
99 [0 ... M1_PMU_PERFCTR_LAST] = ANY_BUT_0_1,
100 [M1_PMU_PERFCTR_UNKNOWN_01] = BIT(7),
101 [M1_PMU_PERFCTR_CPU_CYCLES] = ANY_BUT_0_1 | BIT(0),
102 [M1_PMU_PERFCTR_INSTRUCTIONS] = BIT(7) | BIT(1),
103 [M1_PMU_PERFCTR_UNKNOWN_8d] = ONLY_5_6_7,
104 [M1_PMU_PERFCTR_UNKNOWN_8e] = ONLY_5_6_7,
105 [M1_PMU_PERFCTR_UNKNOWN_8f] = ONLY_5_6_7,
106 [M1_PMU_PERFCTR_UNKNOWN_90] = ONLY_5_6_7,
107 [M1_PMU_PERFCTR_UNKNOWN_93] = ONLY_5_6_7,
108 [M1_PMU_PERFCTR_UNKNOWN_94] = ONLY_5_6_7,
109 [M1_PMU_PERFCTR_UNKNOWN_95] = ONLY_5_6_7,
110 [M1_PMU_PERFCTR_UNKNOWN_96] = ONLY_5_6_7,
111 [M1_PMU_PERFCTR_UNKNOWN_97] = BIT(7),
112 [M1_PMU_PERFCTR_UNKNOWN_98] = ONLY_5_6_7,
113 [M1_PMU_PERFCTR_UNKNOWN_99] = ONLY_5_6_7,
114 [M1_PMU_PERFCTR_UNKNOWN_9a] = BIT(7),
115 [M1_PMU_PERFCTR_UNKNOWN_9b] = ONLY_5_6_7,
116 [M1_PMU_PERFCTR_UNKNOWN_9c] = ONLY_5_6_7,
117 [M1_PMU_PERFCTR_UNKNOWN_9f] = BIT(7),
118 [M1_PMU_PERFCTR_UNKNOWN_bf] = ONLY_5_6_7,
119 [M1_PMU_PERFCTR_UNKNOWN_c0] = ONLY_5_6_7,
120 [M1_PMU_PERFCTR_UNKNOWN_c1] = ONLY_5_6_7,
121 [M1_PMU_PERFCTR_UNKNOWN_c4] = ONLY_5_6_7,
122 [M1_PMU_PERFCTR_UNKNOWN_c5] = ONLY_5_6_7,
123 [M1_PMU_PERFCTR_UNKNOWN_c6] = ONLY_5_6_7,
124 [M1_PMU_PERFCTR_UNKNOWN_c8] = ONLY_5_6_7,
125 [M1_PMU_PERFCTR_UNKNOWN_ca] = ONLY_5_6_7,
126 [M1_PMU_PERFCTR_UNKNOWN_cb] = ONLY_5_6_7,
127 [M1_PMU_PERFCTR_UNKNOWN_f5] = ONLY_2_4_6,
128 [M1_PMU_PERFCTR_UNKNOWN_f6] = ONLY_2_4_6,
129 [M1_PMU_PERFCTR_UNKNOWN_f7] = ONLY_2_4_6,
130 [M1_PMU_PERFCTR_UNKNOWN_f8] = ONLY_2_TO_7,
131 [M1_PMU_PERFCTR_UNKNOWN_fd] = ONLY_2_4_6,
134 static const unsigned m1_pmu_perf_map[PERF_COUNT_HW_MAX] = {
135 PERF_MAP_ALL_UNSUPPORTED,
136 [PERF_COUNT_HW_CPU_CYCLES] = M1_PMU_PERFCTR_CPU_CYCLES,
137 [PERF_COUNT_HW_INSTRUCTIONS] = M1_PMU_PERFCTR_INSTRUCTIONS,
138 /* No idea about the rest yet */
141 /* sysfs definitions */
142 static ssize_t m1_pmu_events_sysfs_show(struct device *dev,
143 struct device_attribute *attr,
146 struct perf_pmu_events_attr *pmu_attr;
148 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
150 return sprintf(page, "event=0x%04llx\n", pmu_attr->id);
153 #define M1_PMU_EVENT_ATTR(name, config) \
154 PMU_EVENT_ATTR_ID(name, m1_pmu_events_sysfs_show, config)
156 static struct attribute *m1_pmu_event_attrs[] = {
157 M1_PMU_EVENT_ATTR(cycles, M1_PMU_PERFCTR_CPU_CYCLES),
158 M1_PMU_EVENT_ATTR(instructions, M1_PMU_PERFCTR_INSTRUCTIONS),
162 static const struct attribute_group m1_pmu_events_attr_group = {
164 .attrs = m1_pmu_event_attrs,
167 PMU_FORMAT_ATTR(event, "config:0-7");
169 static struct attribute *m1_pmu_format_attrs[] = {
170 &format_attr_event.attr,
174 static const struct attribute_group m1_pmu_format_attr_group = {
176 .attrs = m1_pmu_format_attrs,
179 /* Low level accessors. No synchronisation. */
180 #define PMU_READ_COUNTER(_idx) \
181 case _idx: return read_sysreg_s(SYS_IMP_APL_PMC## _idx ##_EL1)
183 #define PMU_WRITE_COUNTER(_val, _idx) \
185 write_sysreg_s(_val, SYS_IMP_APL_PMC## _idx ##_EL1); \
188 static u64 m1_pmu_read_hw_counter(unsigned int index)
206 static void m1_pmu_write_hw_counter(u64 val, unsigned int index)
209 PMU_WRITE_COUNTER(val, 0);
210 PMU_WRITE_COUNTER(val, 1);
211 PMU_WRITE_COUNTER(val, 2);
212 PMU_WRITE_COUNTER(val, 3);
213 PMU_WRITE_COUNTER(val, 4);
214 PMU_WRITE_COUNTER(val, 5);
215 PMU_WRITE_COUNTER(val, 6);
216 PMU_WRITE_COUNTER(val, 7);
217 PMU_WRITE_COUNTER(val, 8);
218 PMU_WRITE_COUNTER(val, 9);
224 #define get_bit_offset(index, mask) (__ffs(mask) + (index))
226 static void __m1_pmu_enable_counter(unsigned int index, bool en)
232 bit = BIT(get_bit_offset(index, PMCR0_CNT_ENABLE_0_7));
235 bit = BIT(get_bit_offset(index - 8, PMCR0_CNT_ENABLE_8_9));
241 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
248 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
251 static void m1_pmu_enable_counter(unsigned int index)
253 __m1_pmu_enable_counter(index, true);
256 static void m1_pmu_disable_counter(unsigned int index)
258 __m1_pmu_enable_counter(index, false);
261 static void __m1_pmu_enable_counter_interrupt(unsigned int index, bool en)
267 bit = BIT(get_bit_offset(index, PMCR0_PMI_ENABLE_0_7));
270 bit = BIT(get_bit_offset(index - 8, PMCR0_PMI_ENABLE_8_9));
276 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
283 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
286 static void m1_pmu_enable_counter_interrupt(unsigned int index)
288 __m1_pmu_enable_counter_interrupt(index, true);
291 static void m1_pmu_disable_counter_interrupt(unsigned int index)
293 __m1_pmu_enable_counter_interrupt(index, false);
296 static void m1_pmu_configure_counter(unsigned int index, u8 event,
297 bool user, bool kernel)
299 u64 val, user_bit, kernel_bit;
304 user_bit = BIT(get_bit_offset(index, PMCR1_COUNT_A64_EL0_0_7));
305 kernel_bit = BIT(get_bit_offset(index, PMCR1_COUNT_A64_EL1_0_7));
308 user_bit = BIT(get_bit_offset(index - 8, PMCR1_COUNT_A64_EL0_8_9));
309 kernel_bit = BIT(get_bit_offset(index - 8, PMCR1_COUNT_A64_EL1_8_9));
315 val = read_sysreg_s(SYS_IMP_APL_PMCR1_EL1);
327 write_sysreg_s(val, SYS_IMP_APL_PMCR1_EL1);
330 * Counters 0 and 1 have fixed events. For anything else,
331 * place the event at the expected location in the relevant
332 * register (PMESR0 holds the event configuration for counters
333 * 2-5, resp. PMESR1 for counters 6-9).
339 shift = (index - 2) * 8;
340 val = read_sysreg_s(SYS_IMP_APL_PMESR0_EL1);
341 val &= ~((u64)0xff << shift);
342 val |= (u64)event << shift;
343 write_sysreg_s(val, SYS_IMP_APL_PMESR0_EL1);
346 shift = (index - 6) * 8;
347 val = read_sysreg_s(SYS_IMP_APL_PMESR1_EL1);
348 val &= ~((u64)0xff << shift);
349 val |= (u64)event << shift;
350 write_sysreg_s(val, SYS_IMP_APL_PMESR1_EL1);
355 /* arm_pmu backend */
356 static void m1_pmu_enable_event(struct perf_event *event)
361 evt = event->hw.config_base & M1_PMU_CFG_EVENT;
362 user = event->hw.config_base & M1_PMU_CFG_COUNT_USER;
363 kernel = event->hw.config_base & M1_PMU_CFG_COUNT_KERNEL;
365 m1_pmu_disable_counter_interrupt(event->hw.idx);
366 m1_pmu_disable_counter(event->hw.idx);
369 m1_pmu_configure_counter(event->hw.idx, evt, user, kernel);
370 m1_pmu_enable_counter(event->hw.idx);
371 m1_pmu_enable_counter_interrupt(event->hw.idx);
375 static void m1_pmu_disable_event(struct perf_event *event)
377 m1_pmu_disable_counter_interrupt(event->hw.idx);
378 m1_pmu_disable_counter(event->hw.idx);
382 static irqreturn_t m1_pmu_handle_irq(struct arm_pmu *cpu_pmu)
384 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
385 struct pt_regs *regs;
389 overflow = read_sysreg_s(SYS_IMP_APL_PMSR_EL1);
391 /* Spurious interrupt? */
392 state = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
393 state &= ~PMCR0_IACT;
394 write_sysreg_s(state, SYS_IMP_APL_PMCR0_EL1);
399 cpu_pmu->stop(cpu_pmu);
401 regs = get_irq_regs();
403 for (idx = 0; idx < cpu_pmu->num_events; idx++) {
404 struct perf_event *event = cpuc->events[idx];
405 struct perf_sample_data data;
410 armpmu_event_update(event);
411 perf_sample_data_init(&data, 0, event->hw.last_period);
412 if (!armpmu_event_set_period(event))
415 if (perf_event_overflow(event, &data, regs))
416 m1_pmu_disable_event(event);
419 cpu_pmu->start(cpu_pmu);
424 static u64 m1_pmu_read_counter(struct perf_event *event)
426 return m1_pmu_read_hw_counter(event->hw.idx);
429 static void m1_pmu_write_counter(struct perf_event *event, u64 value)
431 m1_pmu_write_hw_counter(value, event->hw.idx);
435 static int m1_pmu_get_event_idx(struct pmu_hw_events *cpuc,
436 struct perf_event *event)
438 unsigned long evtype = event->hw.config_base & M1_PMU_CFG_EVENT;
439 unsigned long affinity = m1_pmu_event_affinity[evtype];
443 * Place the event on the first free counter that can count
446 * We could do a better job if we had a view of all the events
447 * counting on the PMU at any given time, and by placing the
448 * most constraining events first.
450 for_each_set_bit(idx, &affinity, M1_PMU_NR_COUNTERS) {
451 if (!test_and_set_bit(idx, cpuc->used_mask))
458 static void m1_pmu_clear_event_idx(struct pmu_hw_events *cpuc,
459 struct perf_event *event)
461 clear_bit(event->hw.idx, cpuc->used_mask);
464 static void __m1_pmu_set_mode(u8 mode)
468 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
469 val &= ~(PMCR0_IMODE | PMCR0_IACT);
470 val |= FIELD_PREP(PMCR0_IMODE, mode);
471 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
475 static void m1_pmu_start(struct arm_pmu *cpu_pmu)
477 __m1_pmu_set_mode(PMCR0_IMODE_FIQ);
480 static void m1_pmu_stop(struct arm_pmu *cpu_pmu)
482 __m1_pmu_set_mode(PMCR0_IMODE_OFF);
485 static int m1_pmu_map_event(struct perf_event *event)
488 * Although the counters are 48bit wide, bit 47 is what
489 * triggers the overflow interrupt. Advertise the counters
490 * being 47bit wide to mimick the behaviour of the ARM PMU.
492 event->hw.flags |= ARMPMU_EVT_47BIT;
493 return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
496 static int m2_pmu_map_event(struct perf_event *event)
499 * Same deal as the above, except that M2 has 64bit counters.
500 * Which, as far as we're concerned, actually means 63 bits.
501 * Yes, this is getting awkward.
503 event->hw.flags |= ARMPMU_EVT_63BIT;
504 return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
507 static void m1_pmu_reset(void *info)
511 __m1_pmu_set_mode(PMCR0_IMODE_OFF);
513 for (i = 0; i < M1_PMU_NR_COUNTERS; i++) {
514 m1_pmu_disable_counter(i);
515 m1_pmu_disable_counter_interrupt(i);
516 m1_pmu_write_hw_counter(0, i);
522 static int m1_pmu_set_event_filter(struct hw_perf_event *event,
523 struct perf_event_attr *attr)
525 unsigned long config_base = 0;
527 if (!attr->exclude_guest)
529 if (!attr->exclude_kernel)
530 config_base |= M1_PMU_CFG_COUNT_KERNEL;
531 if (!attr->exclude_user)
532 config_base |= M1_PMU_CFG_COUNT_USER;
534 event->config_base = config_base;
539 static int m1_pmu_init(struct arm_pmu *cpu_pmu, u32 flags)
541 cpu_pmu->handle_irq = m1_pmu_handle_irq;
542 cpu_pmu->enable = m1_pmu_enable_event;
543 cpu_pmu->disable = m1_pmu_disable_event;
544 cpu_pmu->read_counter = m1_pmu_read_counter;
545 cpu_pmu->write_counter = m1_pmu_write_counter;
546 cpu_pmu->get_event_idx = m1_pmu_get_event_idx;
547 cpu_pmu->clear_event_idx = m1_pmu_clear_event_idx;
548 cpu_pmu->start = m1_pmu_start;
549 cpu_pmu->stop = m1_pmu_stop;
551 if (flags & ARMPMU_EVT_47BIT)
552 cpu_pmu->map_event = m1_pmu_map_event;
553 else if (flags & ARMPMU_EVT_63BIT)
554 cpu_pmu->map_event = m2_pmu_map_event;
556 return WARN_ON(-EINVAL);
558 cpu_pmu->reset = m1_pmu_reset;
559 cpu_pmu->set_event_filter = m1_pmu_set_event_filter;
561 cpu_pmu->num_events = M1_PMU_NR_COUNTERS;
562 cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = &m1_pmu_events_attr_group;
563 cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_FORMATS] = &m1_pmu_format_attr_group;
567 /* Device driver gunk */
568 static int m1_pmu_ice_init(struct arm_pmu *cpu_pmu)
570 cpu_pmu->name = "apple_icestorm_pmu";
571 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_47BIT);
574 static int m1_pmu_fire_init(struct arm_pmu *cpu_pmu)
576 cpu_pmu->name = "apple_firestorm_pmu";
577 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_47BIT);
580 static int m2_pmu_avalanche_init(struct arm_pmu *cpu_pmu)
582 cpu_pmu->name = "apple_avalanche_pmu";
583 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_63BIT);
586 static int m2_pmu_blizzard_init(struct arm_pmu *cpu_pmu)
588 cpu_pmu->name = "apple_blizzard_pmu";
589 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_63BIT);
592 static const struct of_device_id m1_pmu_of_device_ids[] = {
593 { .compatible = "apple,avalanche-pmu", .data = m2_pmu_avalanche_init, },
594 { .compatible = "apple,blizzard-pmu", .data = m2_pmu_blizzard_init, },
595 { .compatible = "apple,icestorm-pmu", .data = m1_pmu_ice_init, },
596 { .compatible = "apple,firestorm-pmu", .data = m1_pmu_fire_init, },
599 MODULE_DEVICE_TABLE(of, m1_pmu_of_device_ids);
601 static int m1_pmu_device_probe(struct platform_device *pdev)
603 return arm_pmu_device_probe(pdev, m1_pmu_of_device_ids, NULL);
606 static struct platform_driver m1_pmu_driver = {
608 .name = "apple-m1-cpu-pmu",
609 .of_match_table = m1_pmu_of_device_ids,
610 .suppress_bind_attrs = true,
612 .probe = m1_pmu_device_probe,
615 module_platform_driver(m1_pmu_driver);