xref: /linux/drivers/perf/qcom_l2_pmu.c (revision 8f8d5745bb520c76b81abef4a2cb3023d0313bfd)
1 /* Copyright (c) 2015-2017 The Linux Foundation. All rights reserved.
2  *
3  * This program is free software; you can redistribute it and/or modify
4  * it under the terms of the GNU General Public License version 2 and
5  * only version 2 as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful,
8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10  * GNU General Public License for more details.
11  */
12 #include <linux/acpi.h>
13 #include <linux/bitops.h>
14 #include <linux/bug.h>
15 #include <linux/cpuhotplug.h>
16 #include <linux/cpumask.h>
17 #include <linux/device.h>
18 #include <linux/errno.h>
19 #include <linux/interrupt.h>
20 #include <linux/irq.h>
21 #include <linux/kernel.h>
22 #include <linux/list.h>
23 #include <linux/percpu.h>
24 #include <linux/perf_event.h>
25 #include <linux/platform_device.h>
26 #include <linux/smp.h>
27 #include <linux/spinlock.h>
28 #include <linux/sysfs.h>
29 #include <linux/types.h>
30 
31 #include <asm/barrier.h>
32 #include <asm/local64.h>
33 #include <asm/sysreg.h>
34 
35 #define MAX_L2_CTRS             9
36 
37 #define L2PMCR_NUM_EV_SHIFT     11
38 #define L2PMCR_NUM_EV_MASK      0x1F
39 
40 #define L2PMCR                  0x400
41 #define L2PMCNTENCLR            0x403
42 #define L2PMCNTENSET            0x404
43 #define L2PMINTENCLR            0x405
44 #define L2PMINTENSET            0x406
45 #define L2PMOVSCLR              0x407
46 #define L2PMOVSSET              0x408
47 #define L2PMCCNTCR              0x409
48 #define L2PMCCNTR               0x40A
49 #define L2PMCCNTSR              0x40C
50 #define L2PMRESR                0x410
51 #define IA_L2PMXEVCNTCR_BASE    0x420
52 #define IA_L2PMXEVCNTR_BASE     0x421
53 #define IA_L2PMXEVFILTER_BASE   0x423
54 #define IA_L2PMXEVTYPER_BASE    0x424
55 
56 #define IA_L2_REG_OFFSET        0x10
57 
58 #define L2PMXEVFILTER_SUFILTER_ALL      0x000E0000
59 #define L2PMXEVFILTER_ORGFILTER_IDINDEP 0x00000004
60 #define L2PMXEVFILTER_ORGFILTER_ALL     0x00000003
61 
62 #define L2EVTYPER_REG_SHIFT     3
63 
64 #define L2PMRESR_GROUP_BITS     8
65 #define L2PMRESR_GROUP_MASK     GENMASK(7, 0)
66 
67 #define L2CYCLE_CTR_BIT         31
68 #define L2CYCLE_CTR_RAW_CODE    0xFE
69 
70 #define L2PMCR_RESET_ALL        0x6
71 #define L2PMCR_COUNTERS_ENABLE  0x1
72 #define L2PMCR_COUNTERS_DISABLE 0x0
73 
74 #define L2PMRESR_EN             BIT_ULL(63)
75 
76 #define L2_EVT_MASK             0x00000FFF
77 #define L2_EVT_CODE_MASK        0x00000FF0
78 #define L2_EVT_GRP_MASK         0x0000000F
79 #define L2_EVT_CODE_SHIFT       4
80 #define L2_EVT_GRP_SHIFT        0
81 
82 #define L2_EVT_CODE(event)   (((event) & L2_EVT_CODE_MASK) >> L2_EVT_CODE_SHIFT)
83 #define L2_EVT_GROUP(event)  (((event) & L2_EVT_GRP_MASK) >> L2_EVT_GRP_SHIFT)
84 
85 #define L2_EVT_GROUP_MAX        7
86 
87 #define L2_COUNTER_RELOAD       BIT_ULL(31)
88 #define L2_CYCLE_COUNTER_RELOAD BIT_ULL(63)
89 
90 #define L2CPUSRSELR_EL1         sys_reg(3, 3, 15, 0, 6)
91 #define L2CPUSRDR_EL1           sys_reg(3, 3, 15, 0, 7)
92 
93 #define reg_idx(reg, i)         (((i) * IA_L2_REG_OFFSET) + reg##_BASE)
94 
95 /*
96  * Events
97  */
98 #define L2_EVENT_CYCLES                    0xfe
99 #define L2_EVENT_DCACHE_OPS                0x400
100 #define L2_EVENT_ICACHE_OPS                0x401
101 #define L2_EVENT_TLBI                      0x402
102 #define L2_EVENT_BARRIERS                  0x403
103 #define L2_EVENT_TOTAL_READS               0x405
104 #define L2_EVENT_TOTAL_WRITES              0x406
105 #define L2_EVENT_TOTAL_REQUESTS            0x407
106 #define L2_EVENT_LDREX                     0x420
107 #define L2_EVENT_STREX                     0x421
108 #define L2_EVENT_CLREX                     0x422
109 
110 static DEFINE_RAW_SPINLOCK(l2_access_lock);
111 
112 /**
113  * set_l2_indirect_reg: write value to an L2 register
114  * @reg: Address of L2 register.
115  * @value: Value to be written to register.
116  *
117  * Use architecturally required barriers for ordering between system register
118  * accesses
119  */
120 static void set_l2_indirect_reg(u64 reg, u64 val)
121 {
122 	unsigned long flags;
123 
124 	raw_spin_lock_irqsave(&l2_access_lock, flags);
125 	write_sysreg_s(reg, L2CPUSRSELR_EL1);
126 	isb();
127 	write_sysreg_s(val, L2CPUSRDR_EL1);
128 	isb();
129 	raw_spin_unlock_irqrestore(&l2_access_lock, flags);
130 }
131 
132 /**
133  * get_l2_indirect_reg: read an L2 register value
134  * @reg: Address of L2 register.
135  *
136  * Use architecturally required barriers for ordering between system register
137  * accesses
138  */
139 static u64 get_l2_indirect_reg(u64 reg)
140 {
141 	u64 val;
142 	unsigned long flags;
143 
144 	raw_spin_lock_irqsave(&l2_access_lock, flags);
145 	write_sysreg_s(reg, L2CPUSRSELR_EL1);
146 	isb();
147 	val = read_sysreg_s(L2CPUSRDR_EL1);
148 	raw_spin_unlock_irqrestore(&l2_access_lock, flags);
149 
150 	return val;
151 }
152 
153 struct cluster_pmu;
154 
155 /*
156  * Aggregate PMU. Implements the core pmu functions and manages
157  * the hardware PMUs.
158  */
159 struct l2cache_pmu {
160 	struct hlist_node node;
161 	u32 num_pmus;
162 	struct pmu pmu;
163 	int num_counters;
164 	cpumask_t cpumask;
165 	struct platform_device *pdev;
166 	struct cluster_pmu * __percpu *pmu_cluster;
167 	struct list_head clusters;
168 };
169 
170 /*
171  * The cache is made up of one or more clusters, each cluster has its own PMU.
172  * Each cluster is associated with one or more CPUs.
173  * This structure represents one of the hardware PMUs.
174  *
175  * Events can be envisioned as a 2-dimensional array. Each column represents
176  * a group of events. There are 8 groups. Only one entry from each
177  * group can be in use at a time.
178  *
179  * Events are specified as 0xCCG, where CC is 2 hex digits specifying
180  * the code (array row) and G specifies the group (column).
181  *
182  * In addition there is a cycle counter event specified by L2CYCLE_CTR_RAW_CODE
183  * which is outside the above scheme.
184  */
185 struct cluster_pmu {
186 	struct list_head next;
187 	struct perf_event *events[MAX_L2_CTRS];
188 	struct l2cache_pmu *l2cache_pmu;
189 	DECLARE_BITMAP(used_counters, MAX_L2_CTRS);
190 	DECLARE_BITMAP(used_groups, L2_EVT_GROUP_MAX + 1);
191 	int irq;
192 	int cluster_id;
193 	/* The CPU that is used for collecting events on this cluster */
194 	int on_cpu;
195 	/* All the CPUs associated with this cluster */
196 	cpumask_t cluster_cpus;
197 	spinlock_t pmu_lock;
198 };
199 
200 #define to_l2cache_pmu(p) (container_of(p, struct l2cache_pmu, pmu))
201 
202 static u32 l2_cycle_ctr_idx;
203 static u32 l2_counter_present_mask;
204 
205 static inline u32 idx_to_reg_bit(u32 idx)
206 {
207 	if (idx == l2_cycle_ctr_idx)
208 		return BIT(L2CYCLE_CTR_BIT);
209 
210 	return BIT(idx);
211 }
212 
213 static inline struct cluster_pmu *get_cluster_pmu(
214 	struct l2cache_pmu *l2cache_pmu, int cpu)
215 {
216 	return *per_cpu_ptr(l2cache_pmu->pmu_cluster, cpu);
217 }
218 
219 static void cluster_pmu_reset(void)
220 {
221 	/* Reset all counters */
222 	set_l2_indirect_reg(L2PMCR, L2PMCR_RESET_ALL);
223 	set_l2_indirect_reg(L2PMCNTENCLR, l2_counter_present_mask);
224 	set_l2_indirect_reg(L2PMINTENCLR, l2_counter_present_mask);
225 	set_l2_indirect_reg(L2PMOVSCLR, l2_counter_present_mask);
226 }
227 
228 static inline void cluster_pmu_enable(void)
229 {
230 	set_l2_indirect_reg(L2PMCR, L2PMCR_COUNTERS_ENABLE);
231 }
232 
233 static inline void cluster_pmu_disable(void)
234 {
235 	set_l2_indirect_reg(L2PMCR, L2PMCR_COUNTERS_DISABLE);
236 }
237 
238 static inline void cluster_pmu_counter_set_value(u32 idx, u64 value)
239 {
240 	if (idx == l2_cycle_ctr_idx)
241 		set_l2_indirect_reg(L2PMCCNTR, value);
242 	else
243 		set_l2_indirect_reg(reg_idx(IA_L2PMXEVCNTR, idx), value);
244 }
245 
246 static inline u64 cluster_pmu_counter_get_value(u32 idx)
247 {
248 	u64 value;
249 
250 	if (idx == l2_cycle_ctr_idx)
251 		value = get_l2_indirect_reg(L2PMCCNTR);
252 	else
253 		value = get_l2_indirect_reg(reg_idx(IA_L2PMXEVCNTR, idx));
254 
255 	return value;
256 }
257 
258 static inline void cluster_pmu_counter_enable(u32 idx)
259 {
260 	set_l2_indirect_reg(L2PMCNTENSET, idx_to_reg_bit(idx));
261 }
262 
263 static inline void cluster_pmu_counter_disable(u32 idx)
264 {
265 	set_l2_indirect_reg(L2PMCNTENCLR, idx_to_reg_bit(idx));
266 }
267 
268 static inline void cluster_pmu_counter_enable_interrupt(u32 idx)
269 {
270 	set_l2_indirect_reg(L2PMINTENSET, idx_to_reg_bit(idx));
271 }
272 
273 static inline void cluster_pmu_counter_disable_interrupt(u32 idx)
274 {
275 	set_l2_indirect_reg(L2PMINTENCLR, idx_to_reg_bit(idx));
276 }
277 
278 static inline void cluster_pmu_set_evccntcr(u32 val)
279 {
280 	set_l2_indirect_reg(L2PMCCNTCR, val);
281 }
282 
283 static inline void cluster_pmu_set_evcntcr(u32 ctr, u32 val)
284 {
285 	set_l2_indirect_reg(reg_idx(IA_L2PMXEVCNTCR, ctr), val);
286 }
287 
288 static inline void cluster_pmu_set_evtyper(u32 ctr, u32 val)
289 {
290 	set_l2_indirect_reg(reg_idx(IA_L2PMXEVTYPER, ctr), val);
291 }
292 
293 static void cluster_pmu_set_resr(struct cluster_pmu *cluster,
294 			       u32 event_group, u32 event_cc)
295 {
296 	u64 field;
297 	u64 resr_val;
298 	u32 shift;
299 	unsigned long flags;
300 
301 	shift = L2PMRESR_GROUP_BITS * event_group;
302 	field = ((u64)(event_cc & L2PMRESR_GROUP_MASK) << shift);
303 
304 	spin_lock_irqsave(&cluster->pmu_lock, flags);
305 
306 	resr_val = get_l2_indirect_reg(L2PMRESR);
307 	resr_val &= ~(L2PMRESR_GROUP_MASK << shift);
308 	resr_val |= field;
309 	resr_val |= L2PMRESR_EN;
310 	set_l2_indirect_reg(L2PMRESR, resr_val);
311 
312 	spin_unlock_irqrestore(&cluster->pmu_lock, flags);
313 }
314 
315 /*
316  * Hardware allows filtering of events based on the originating
317  * CPU. Turn this off by setting filter bits to allow events from
318  * all CPUS, subunits and ID independent events in this cluster.
319  */
320 static inline void cluster_pmu_set_evfilter_sys_mode(u32 ctr)
321 {
322 	u32 val =  L2PMXEVFILTER_SUFILTER_ALL |
323 		   L2PMXEVFILTER_ORGFILTER_IDINDEP |
324 		   L2PMXEVFILTER_ORGFILTER_ALL;
325 
326 	set_l2_indirect_reg(reg_idx(IA_L2PMXEVFILTER, ctr), val);
327 }
328 
329 static inline u32 cluster_pmu_getreset_ovsr(void)
330 {
331 	u32 result = get_l2_indirect_reg(L2PMOVSSET);
332 
333 	set_l2_indirect_reg(L2PMOVSCLR, result);
334 	return result;
335 }
336 
337 static inline bool cluster_pmu_has_overflowed(u32 ovsr)
338 {
339 	return !!(ovsr & l2_counter_present_mask);
340 }
341 
342 static inline bool cluster_pmu_counter_has_overflowed(u32 ovsr, u32 idx)
343 {
344 	return !!(ovsr & idx_to_reg_bit(idx));
345 }
346 
347 static void l2_cache_event_update(struct perf_event *event)
348 {
349 	struct hw_perf_event *hwc = &event->hw;
350 	u64 delta, prev, now;
351 	u32 idx = hwc->idx;
352 
353 	do {
354 		prev = local64_read(&hwc->prev_count);
355 		now = cluster_pmu_counter_get_value(idx);
356 	} while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
357 
358 	/*
359 	 * The cycle counter is 64-bit, but all other counters are
360 	 * 32-bit, and we must handle 32-bit overflow explicitly.
361 	 */
362 	delta = now - prev;
363 	if (idx != l2_cycle_ctr_idx)
364 		delta &= 0xffffffff;
365 
366 	local64_add(delta, &event->count);
367 }
368 
369 static void l2_cache_cluster_set_period(struct cluster_pmu *cluster,
370 				       struct hw_perf_event *hwc)
371 {
372 	u32 idx = hwc->idx;
373 	u64 new;
374 
375 	/*
376 	 * We limit the max period to half the max counter value so
377 	 * that even in the case of extreme interrupt latency the
378 	 * counter will (hopefully) not wrap past its initial value.
379 	 */
380 	if (idx == l2_cycle_ctr_idx)
381 		new = L2_CYCLE_COUNTER_RELOAD;
382 	else
383 		new = L2_COUNTER_RELOAD;
384 
385 	local64_set(&hwc->prev_count, new);
386 	cluster_pmu_counter_set_value(idx, new);
387 }
388 
389 static int l2_cache_get_event_idx(struct cluster_pmu *cluster,
390 				   struct perf_event *event)
391 {
392 	struct hw_perf_event *hwc = &event->hw;
393 	int idx;
394 	int num_ctrs = cluster->l2cache_pmu->num_counters - 1;
395 	unsigned int group;
396 
397 	if (hwc->config_base == L2CYCLE_CTR_RAW_CODE) {
398 		if (test_and_set_bit(l2_cycle_ctr_idx, cluster->used_counters))
399 			return -EAGAIN;
400 
401 		return l2_cycle_ctr_idx;
402 	}
403 
404 	idx = find_first_zero_bit(cluster->used_counters, num_ctrs);
405 	if (idx == num_ctrs)
406 		/* The counters are all in use. */
407 		return -EAGAIN;
408 
409 	/*
410 	 * Check for column exclusion: event column already in use by another
411 	 * event. This is for events which are not in the same group.
412 	 * Conflicting events in the same group are detected in event_init.
413 	 */
414 	group = L2_EVT_GROUP(hwc->config_base);
415 	if (test_bit(group, cluster->used_groups))
416 		return -EAGAIN;
417 
418 	set_bit(idx, cluster->used_counters);
419 	set_bit(group, cluster->used_groups);
420 
421 	return idx;
422 }
423 
424 static void l2_cache_clear_event_idx(struct cluster_pmu *cluster,
425 				      struct perf_event *event)
426 {
427 	struct hw_perf_event *hwc = &event->hw;
428 	int idx = hwc->idx;
429 
430 	clear_bit(idx, cluster->used_counters);
431 	if (hwc->config_base != L2CYCLE_CTR_RAW_CODE)
432 		clear_bit(L2_EVT_GROUP(hwc->config_base), cluster->used_groups);
433 }
434 
435 static irqreturn_t l2_cache_handle_irq(int irq_num, void *data)
436 {
437 	struct cluster_pmu *cluster = data;
438 	int num_counters = cluster->l2cache_pmu->num_counters;
439 	u32 ovsr;
440 	int idx;
441 
442 	ovsr = cluster_pmu_getreset_ovsr();
443 	if (!cluster_pmu_has_overflowed(ovsr))
444 		return IRQ_NONE;
445 
446 	for_each_set_bit(idx, cluster->used_counters, num_counters) {
447 		struct perf_event *event = cluster->events[idx];
448 		struct hw_perf_event *hwc;
449 
450 		if (WARN_ON_ONCE(!event))
451 			continue;
452 
453 		if (!cluster_pmu_counter_has_overflowed(ovsr, idx))
454 			continue;
455 
456 		l2_cache_event_update(event);
457 		hwc = &event->hw;
458 
459 		l2_cache_cluster_set_period(cluster, hwc);
460 	}
461 
462 	return IRQ_HANDLED;
463 }
464 
465 /*
466  * Implementation of abstract pmu functionality required by
467  * the core perf events code.
468  */
469 
470 static void l2_cache_pmu_enable(struct pmu *pmu)
471 {
472 	/*
473 	 * Although there is only one PMU (per socket) controlling multiple
474 	 * physical PMUs (per cluster), because we do not support per-task mode
475 	 * each event is associated with a CPU. Each event has pmu_enable
476 	 * called on its CPU, so here it is only necessary to enable the
477 	 * counters for the current CPU.
478 	 */
479 
480 	cluster_pmu_enable();
481 }
482 
483 static void l2_cache_pmu_disable(struct pmu *pmu)
484 {
485 	cluster_pmu_disable();
486 }
487 
488 static int l2_cache_event_init(struct perf_event *event)
489 {
490 	struct hw_perf_event *hwc = &event->hw;
491 	struct cluster_pmu *cluster;
492 	struct perf_event *sibling;
493 	struct l2cache_pmu *l2cache_pmu;
494 
495 	if (event->attr.type != event->pmu->type)
496 		return -ENOENT;
497 
498 	l2cache_pmu = to_l2cache_pmu(event->pmu);
499 
500 	if (hwc->sample_period) {
501 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
502 				    "Sampling not supported\n");
503 		return -EOPNOTSUPP;
504 	}
505 
506 	if (event->cpu < 0) {
507 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
508 				    "Per-task mode not supported\n");
509 		return -EOPNOTSUPP;
510 	}
511 
512 	if (((L2_EVT_GROUP(event->attr.config) > L2_EVT_GROUP_MAX) ||
513 	     ((event->attr.config & ~L2_EVT_MASK) != 0)) &&
514 	    (event->attr.config != L2CYCLE_CTR_RAW_CODE)) {
515 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
516 				    "Invalid config %llx\n",
517 				    event->attr.config);
518 		return -EINVAL;
519 	}
520 
521 	/* Don't allow groups with mixed PMUs, except for s/w events */
522 	if (event->group_leader->pmu != event->pmu &&
523 	    !is_software_event(event->group_leader)) {
524 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
525 			 "Can't create mixed PMU group\n");
526 		return -EINVAL;
527 	}
528 
529 	for_each_sibling_event(sibling, event->group_leader) {
530 		if (sibling->pmu != event->pmu &&
531 		    !is_software_event(sibling)) {
532 			dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
533 				 "Can't create mixed PMU group\n");
534 			return -EINVAL;
535 		}
536 	}
537 
538 	cluster = get_cluster_pmu(l2cache_pmu, event->cpu);
539 	if (!cluster) {
540 		/* CPU has not been initialised */
541 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
542 			"CPU%d not associated with L2 cluster\n", event->cpu);
543 		return -EINVAL;
544 	}
545 
546 	/* Ensure all events in a group are on the same cpu */
547 	if ((event->group_leader != event) &&
548 	    (cluster->on_cpu != event->group_leader->cpu)) {
549 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
550 			 "Can't create group on CPUs %d and %d",
551 			 event->cpu, event->group_leader->cpu);
552 		return -EINVAL;
553 	}
554 
555 	if ((event != event->group_leader) &&
556 	    !is_software_event(event->group_leader) &&
557 	    (L2_EVT_GROUP(event->group_leader->attr.config) ==
558 	     L2_EVT_GROUP(event->attr.config))) {
559 		dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
560 			 "Column exclusion: conflicting events %llx %llx\n",
561 		       event->group_leader->attr.config,
562 		       event->attr.config);
563 		return -EINVAL;
564 	}
565 
566 	for_each_sibling_event(sibling, event->group_leader) {
567 		if ((sibling != event) &&
568 		    !is_software_event(sibling) &&
569 		    (L2_EVT_GROUP(sibling->attr.config) ==
570 		     L2_EVT_GROUP(event->attr.config))) {
571 			dev_dbg_ratelimited(&l2cache_pmu->pdev->dev,
572 			     "Column exclusion: conflicting events %llx %llx\n",
573 					    sibling->attr.config,
574 					    event->attr.config);
575 			return -EINVAL;
576 		}
577 	}
578 
579 	hwc->idx = -1;
580 	hwc->config_base = event->attr.config;
581 
582 	/*
583 	 * Ensure all events are on the same cpu so all events are in the
584 	 * same cpu context, to avoid races on pmu_enable etc.
585 	 */
586 	event->cpu = cluster->on_cpu;
587 
588 	return 0;
589 }
590 
591 static void l2_cache_event_start(struct perf_event *event, int flags)
592 {
593 	struct cluster_pmu *cluster;
594 	struct hw_perf_event *hwc = &event->hw;
595 	int idx = hwc->idx;
596 	u32 config;
597 	u32 event_cc, event_group;
598 
599 	hwc->state = 0;
600 
601 	cluster = get_cluster_pmu(to_l2cache_pmu(event->pmu), event->cpu);
602 
603 	l2_cache_cluster_set_period(cluster, hwc);
604 
605 	if (hwc->config_base == L2CYCLE_CTR_RAW_CODE) {
606 		cluster_pmu_set_evccntcr(0);
607 	} else {
608 		config = hwc->config_base;
609 		event_cc    = L2_EVT_CODE(config);
610 		event_group = L2_EVT_GROUP(config);
611 
612 		cluster_pmu_set_evcntcr(idx, 0);
613 		cluster_pmu_set_evtyper(idx, event_group);
614 		cluster_pmu_set_resr(cluster, event_group, event_cc);
615 		cluster_pmu_set_evfilter_sys_mode(idx);
616 	}
617 
618 	cluster_pmu_counter_enable_interrupt(idx);
619 	cluster_pmu_counter_enable(idx);
620 }
621 
622 static void l2_cache_event_stop(struct perf_event *event, int flags)
623 {
624 	struct hw_perf_event *hwc = &event->hw;
625 	int idx = hwc->idx;
626 
627 	if (hwc->state & PERF_HES_STOPPED)
628 		return;
629 
630 	cluster_pmu_counter_disable_interrupt(idx);
631 	cluster_pmu_counter_disable(idx);
632 
633 	if (flags & PERF_EF_UPDATE)
634 		l2_cache_event_update(event);
635 	hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
636 }
637 
638 static int l2_cache_event_add(struct perf_event *event, int flags)
639 {
640 	struct hw_perf_event *hwc = &event->hw;
641 	int idx;
642 	int err = 0;
643 	struct cluster_pmu *cluster;
644 
645 	cluster = get_cluster_pmu(to_l2cache_pmu(event->pmu), event->cpu);
646 
647 	idx = l2_cache_get_event_idx(cluster, event);
648 	if (idx < 0)
649 		return idx;
650 
651 	hwc->idx = idx;
652 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
653 	cluster->events[idx] = event;
654 	local64_set(&hwc->prev_count, 0);
655 
656 	if (flags & PERF_EF_START)
657 		l2_cache_event_start(event, flags);
658 
659 	/* Propagate changes to the userspace mapping. */
660 	perf_event_update_userpage(event);
661 
662 	return err;
663 }
664 
665 static void l2_cache_event_del(struct perf_event *event, int flags)
666 {
667 	struct hw_perf_event *hwc = &event->hw;
668 	struct cluster_pmu *cluster;
669 	int idx = hwc->idx;
670 
671 	cluster = get_cluster_pmu(to_l2cache_pmu(event->pmu), event->cpu);
672 
673 	l2_cache_event_stop(event, flags | PERF_EF_UPDATE);
674 	cluster->events[idx] = NULL;
675 	l2_cache_clear_event_idx(cluster, event);
676 
677 	perf_event_update_userpage(event);
678 }
679 
680 static void l2_cache_event_read(struct perf_event *event)
681 {
682 	l2_cache_event_update(event);
683 }
684 
685 static ssize_t l2_cache_pmu_cpumask_show(struct device *dev,
686 					 struct device_attribute *attr,
687 					 char *buf)
688 {
689 	struct l2cache_pmu *l2cache_pmu = to_l2cache_pmu(dev_get_drvdata(dev));
690 
691 	return cpumap_print_to_pagebuf(true, buf, &l2cache_pmu->cpumask);
692 }
693 
694 static struct device_attribute l2_cache_pmu_cpumask_attr =
695 		__ATTR(cpumask, S_IRUGO, l2_cache_pmu_cpumask_show, NULL);
696 
697 static struct attribute *l2_cache_pmu_cpumask_attrs[] = {
698 	&l2_cache_pmu_cpumask_attr.attr,
699 	NULL,
700 };
701 
702 static struct attribute_group l2_cache_pmu_cpumask_group = {
703 	.attrs = l2_cache_pmu_cpumask_attrs,
704 };
705 
706 /* CCG format for perf RAW codes. */
707 PMU_FORMAT_ATTR(l2_code,   "config:4-11");
708 PMU_FORMAT_ATTR(l2_group,  "config:0-3");
709 PMU_FORMAT_ATTR(event,     "config:0-11");
710 
711 static struct attribute *l2_cache_pmu_formats[] = {
712 	&format_attr_l2_code.attr,
713 	&format_attr_l2_group.attr,
714 	&format_attr_event.attr,
715 	NULL,
716 };
717 
718 static struct attribute_group l2_cache_pmu_format_group = {
719 	.name = "format",
720 	.attrs = l2_cache_pmu_formats,
721 };
722 
723 static ssize_t l2cache_pmu_event_show(struct device *dev,
724 				      struct device_attribute *attr, char *page)
725 {
726 	struct perf_pmu_events_attr *pmu_attr;
727 
728 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
729 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
730 }
731 
732 #define L2CACHE_EVENT_ATTR(_name, _id)					     \
733 	(&((struct perf_pmu_events_attr[]) {				     \
734 		{ .attr = __ATTR(_name, 0444, l2cache_pmu_event_show, NULL), \
735 		  .id = _id, }						     \
736 	})[0].attr.attr)
737 
738 static struct attribute *l2_cache_pmu_events[] = {
739 	L2CACHE_EVENT_ATTR(cycles, L2_EVENT_CYCLES),
740 	L2CACHE_EVENT_ATTR(dcache-ops, L2_EVENT_DCACHE_OPS),
741 	L2CACHE_EVENT_ATTR(icache-ops, L2_EVENT_ICACHE_OPS),
742 	L2CACHE_EVENT_ATTR(tlbi, L2_EVENT_TLBI),
743 	L2CACHE_EVENT_ATTR(barriers, L2_EVENT_BARRIERS),
744 	L2CACHE_EVENT_ATTR(total-reads, L2_EVENT_TOTAL_READS),
745 	L2CACHE_EVENT_ATTR(total-writes, L2_EVENT_TOTAL_WRITES),
746 	L2CACHE_EVENT_ATTR(total-requests, L2_EVENT_TOTAL_REQUESTS),
747 	L2CACHE_EVENT_ATTR(ldrex, L2_EVENT_LDREX),
748 	L2CACHE_EVENT_ATTR(strex, L2_EVENT_STREX),
749 	L2CACHE_EVENT_ATTR(clrex, L2_EVENT_CLREX),
750 	NULL
751 };
752 
753 static struct attribute_group l2_cache_pmu_events_group = {
754 	.name = "events",
755 	.attrs = l2_cache_pmu_events,
756 };
757 
758 static const struct attribute_group *l2_cache_pmu_attr_grps[] = {
759 	&l2_cache_pmu_format_group,
760 	&l2_cache_pmu_cpumask_group,
761 	&l2_cache_pmu_events_group,
762 	NULL,
763 };
764 
765 /*
766  * Generic device handlers
767  */
768 
769 static const struct acpi_device_id l2_cache_pmu_acpi_match[] = {
770 	{ "QCOM8130", },
771 	{ }
772 };
773 
774 static int get_num_counters(void)
775 {
776 	int val;
777 
778 	val = get_l2_indirect_reg(L2PMCR);
779 
780 	/*
781 	 * Read number of counters from L2PMCR and add 1
782 	 * for the cycle counter.
783 	 */
784 	return ((val >> L2PMCR_NUM_EV_SHIFT) & L2PMCR_NUM_EV_MASK) + 1;
785 }
786 
787 static struct cluster_pmu *l2_cache_associate_cpu_with_cluster(
788 	struct l2cache_pmu *l2cache_pmu, int cpu)
789 {
790 	u64 mpidr;
791 	int cpu_cluster_id;
792 	struct cluster_pmu *cluster = NULL;
793 
794 	/*
795 	 * This assumes that the cluster_id is in MPIDR[aff1] for
796 	 * single-threaded cores, and MPIDR[aff2] for multi-threaded
797 	 * cores. This logic will have to be updated if this changes.
798 	 */
799 	mpidr = read_cpuid_mpidr();
800 	if (mpidr & MPIDR_MT_BITMASK)
801 		cpu_cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 2);
802 	else
803 		cpu_cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
804 
805 	list_for_each_entry(cluster, &l2cache_pmu->clusters, next) {
806 		if (cluster->cluster_id != cpu_cluster_id)
807 			continue;
808 
809 		dev_info(&l2cache_pmu->pdev->dev,
810 			 "CPU%d associated with cluster %d\n", cpu,
811 			 cluster->cluster_id);
812 		cpumask_set_cpu(cpu, &cluster->cluster_cpus);
813 		*per_cpu_ptr(l2cache_pmu->pmu_cluster, cpu) = cluster;
814 		break;
815 	}
816 
817 	return cluster;
818 }
819 
820 static int l2cache_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
821 {
822 	struct cluster_pmu *cluster;
823 	struct l2cache_pmu *l2cache_pmu;
824 
825 	l2cache_pmu = hlist_entry_safe(node, struct l2cache_pmu, node);
826 	cluster = get_cluster_pmu(l2cache_pmu, cpu);
827 	if (!cluster) {
828 		/* First time this CPU has come online */
829 		cluster = l2_cache_associate_cpu_with_cluster(l2cache_pmu, cpu);
830 		if (!cluster) {
831 			/* Only if broken firmware doesn't list every cluster */
832 			WARN_ONCE(1, "No L2 cache cluster for CPU%d\n", cpu);
833 			return 0;
834 		}
835 	}
836 
837 	/* If another CPU is managing this cluster, we're done */
838 	if (cluster->on_cpu != -1)
839 		return 0;
840 
841 	/*
842 	 * All CPUs on this cluster were down, use this one.
843 	 * Reset to put it into sane state.
844 	 */
845 	cluster->on_cpu = cpu;
846 	cpumask_set_cpu(cpu, &l2cache_pmu->cpumask);
847 	cluster_pmu_reset();
848 
849 	WARN_ON(irq_set_affinity(cluster->irq, cpumask_of(cpu)));
850 	enable_irq(cluster->irq);
851 
852 	return 0;
853 }
854 
855 static int l2cache_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
856 {
857 	struct cluster_pmu *cluster;
858 	struct l2cache_pmu *l2cache_pmu;
859 	cpumask_t cluster_online_cpus;
860 	unsigned int target;
861 
862 	l2cache_pmu = hlist_entry_safe(node, struct l2cache_pmu, node);
863 	cluster = get_cluster_pmu(l2cache_pmu, cpu);
864 	if (!cluster)
865 		return 0;
866 
867 	/* If this CPU is not managing the cluster, we're done */
868 	if (cluster->on_cpu != cpu)
869 		return 0;
870 
871 	/* Give up ownership of cluster */
872 	cpumask_clear_cpu(cpu, &l2cache_pmu->cpumask);
873 	cluster->on_cpu = -1;
874 
875 	/* Any other CPU for this cluster which is still online */
876 	cpumask_and(&cluster_online_cpus, &cluster->cluster_cpus,
877 		    cpu_online_mask);
878 	target = cpumask_any_but(&cluster_online_cpus, cpu);
879 	if (target >= nr_cpu_ids) {
880 		disable_irq(cluster->irq);
881 		return 0;
882 	}
883 
884 	perf_pmu_migrate_context(&l2cache_pmu->pmu, cpu, target);
885 	cluster->on_cpu = target;
886 	cpumask_set_cpu(target, &l2cache_pmu->cpumask);
887 	WARN_ON(irq_set_affinity(cluster->irq, cpumask_of(target)));
888 
889 	return 0;
890 }
891 
892 static int l2_cache_pmu_probe_cluster(struct device *dev, void *data)
893 {
894 	struct platform_device *pdev = to_platform_device(dev->parent);
895 	struct platform_device *sdev = to_platform_device(dev);
896 	struct l2cache_pmu *l2cache_pmu = data;
897 	struct cluster_pmu *cluster;
898 	struct acpi_device *device;
899 	unsigned long fw_cluster_id;
900 	int err;
901 	int irq;
902 
903 	if (acpi_bus_get_device(ACPI_HANDLE(dev), &device))
904 		return -ENODEV;
905 
906 	if (kstrtoul(device->pnp.unique_id, 10, &fw_cluster_id) < 0) {
907 		dev_err(&pdev->dev, "unable to read ACPI uid\n");
908 		return -ENODEV;
909 	}
910 
911 	cluster = devm_kzalloc(&pdev->dev, sizeof(*cluster), GFP_KERNEL);
912 	if (!cluster)
913 		return -ENOMEM;
914 
915 	INIT_LIST_HEAD(&cluster->next);
916 	list_add(&cluster->next, &l2cache_pmu->clusters);
917 	cluster->cluster_id = fw_cluster_id;
918 
919 	irq = platform_get_irq(sdev, 0);
920 	if (irq < 0) {
921 		dev_err(&pdev->dev,
922 			"Failed to get valid irq for cluster %ld\n",
923 			fw_cluster_id);
924 		return irq;
925 	}
926 	irq_set_status_flags(irq, IRQ_NOAUTOEN);
927 	cluster->irq = irq;
928 
929 	cluster->l2cache_pmu = l2cache_pmu;
930 	cluster->on_cpu = -1;
931 
932 	err = devm_request_irq(&pdev->dev, irq, l2_cache_handle_irq,
933 			       IRQF_NOBALANCING | IRQF_NO_THREAD,
934 			       "l2-cache-pmu", cluster);
935 	if (err) {
936 		dev_err(&pdev->dev,
937 			"Unable to request IRQ%d for L2 PMU counters\n", irq);
938 		return err;
939 	}
940 
941 	dev_info(&pdev->dev,
942 		"Registered L2 cache PMU cluster %ld\n", fw_cluster_id);
943 
944 	spin_lock_init(&cluster->pmu_lock);
945 
946 	l2cache_pmu->num_pmus++;
947 
948 	return 0;
949 }
950 
951 static int l2_cache_pmu_probe(struct platform_device *pdev)
952 {
953 	int err;
954 	struct l2cache_pmu *l2cache_pmu;
955 
956 	l2cache_pmu =
957 		devm_kzalloc(&pdev->dev, sizeof(*l2cache_pmu), GFP_KERNEL);
958 	if (!l2cache_pmu)
959 		return -ENOMEM;
960 
961 	INIT_LIST_HEAD(&l2cache_pmu->clusters);
962 
963 	platform_set_drvdata(pdev, l2cache_pmu);
964 	l2cache_pmu->pmu = (struct pmu) {
965 		/* suffix is instance id for future use with multiple sockets */
966 		.name		= "l2cache_0",
967 		.task_ctx_nr    = perf_invalid_context,
968 		.pmu_enable	= l2_cache_pmu_enable,
969 		.pmu_disable	= l2_cache_pmu_disable,
970 		.event_init	= l2_cache_event_init,
971 		.add		= l2_cache_event_add,
972 		.del		= l2_cache_event_del,
973 		.start		= l2_cache_event_start,
974 		.stop		= l2_cache_event_stop,
975 		.read		= l2_cache_event_read,
976 		.attr_groups	= l2_cache_pmu_attr_grps,
977 		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
978 	};
979 
980 	l2cache_pmu->num_counters = get_num_counters();
981 	l2cache_pmu->pdev = pdev;
982 	l2cache_pmu->pmu_cluster = devm_alloc_percpu(&pdev->dev,
983 						     struct cluster_pmu *);
984 	if (!l2cache_pmu->pmu_cluster)
985 		return -ENOMEM;
986 
987 	l2_cycle_ctr_idx = l2cache_pmu->num_counters - 1;
988 	l2_counter_present_mask = GENMASK(l2cache_pmu->num_counters - 2, 0) |
989 		BIT(L2CYCLE_CTR_BIT);
990 
991 	cpumask_clear(&l2cache_pmu->cpumask);
992 
993 	/* Read cluster info and initialize each cluster */
994 	err = device_for_each_child(&pdev->dev, l2cache_pmu,
995 				    l2_cache_pmu_probe_cluster);
996 	if (err)
997 		return err;
998 
999 	if (l2cache_pmu->num_pmus == 0) {
1000 		dev_err(&pdev->dev, "No hardware L2 cache PMUs found\n");
1001 		return -ENODEV;
1002 	}
1003 
1004 	err = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_QCOM_L2_ONLINE,
1005 				       &l2cache_pmu->node);
1006 	if (err) {
1007 		dev_err(&pdev->dev, "Error %d registering hotplug", err);
1008 		return err;
1009 	}
1010 
1011 	err = perf_pmu_register(&l2cache_pmu->pmu, l2cache_pmu->pmu.name, -1);
1012 	if (err) {
1013 		dev_err(&pdev->dev, "Error %d registering L2 cache PMU\n", err);
1014 		goto out_unregister;
1015 	}
1016 
1017 	dev_info(&pdev->dev, "Registered L2 cache PMU using %d HW PMUs\n",
1018 		 l2cache_pmu->num_pmus);
1019 
1020 	return err;
1021 
1022 out_unregister:
1023 	cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_QCOM_L2_ONLINE,
1024 				    &l2cache_pmu->node);
1025 	return err;
1026 }
1027 
1028 static int l2_cache_pmu_remove(struct platform_device *pdev)
1029 {
1030 	struct l2cache_pmu *l2cache_pmu =
1031 		to_l2cache_pmu(platform_get_drvdata(pdev));
1032 
1033 	perf_pmu_unregister(&l2cache_pmu->pmu);
1034 	cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_QCOM_L2_ONLINE,
1035 				    &l2cache_pmu->node);
1036 	return 0;
1037 }
1038 
1039 static struct platform_driver l2_cache_pmu_driver = {
1040 	.driver = {
1041 		.name = "qcom-l2cache-pmu",
1042 		.acpi_match_table = ACPI_PTR(l2_cache_pmu_acpi_match),
1043 	},
1044 	.probe = l2_cache_pmu_probe,
1045 	.remove = l2_cache_pmu_remove,
1046 };
1047 
1048 static int __init register_l2_cache_pmu_driver(void)
1049 {
1050 	int err;
1051 
1052 	err = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_QCOM_L2_ONLINE,
1053 				      "AP_PERF_ARM_QCOM_L2_ONLINE",
1054 				      l2cache_pmu_online_cpu,
1055 				      l2cache_pmu_offline_cpu);
1056 	if (err)
1057 		return err;
1058 
1059 	return platform_driver_register(&l2_cache_pmu_driver);
1060 }
1061 device_initcall(register_l2_cache_pmu_driver);
1062