xref: /openbmc/linux/arch/arm/mm/cache-l2x0-pmu.c (revision 1edd0337)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * L220/L310 cache controller support
4  *
5  * Copyright (C) 2016 ARM Limited
6  */
7 #include <linux/errno.h>
8 #include <linux/hrtimer.h>
9 #include <linux/io.h>
10 #include <linux/list.h>
11 #include <linux/perf_event.h>
12 #include <linux/printk.h>
13 #include <linux/slab.h>
14 #include <linux/types.h>
15 
16 #include <asm/hardware/cache-l2x0.h>
17 
18 #define PMU_NR_COUNTERS 2
19 
20 static void __iomem *l2x0_base;
21 static struct pmu *l2x0_pmu;
22 static cpumask_t pmu_cpu;
23 
24 static const char *l2x0_name;
25 
26 static ktime_t l2x0_pmu_poll_period;
27 static struct hrtimer l2x0_pmu_hrtimer;
28 
29 /*
30  * The L220/PL310 PMU has two equivalent counters, Counter1 and Counter0.
31  * Registers controlling these are laid out in pairs, in descending order, i.e.
32  * the register for Counter1 comes first, followed by the register for
33  * Counter0.
34  * We ensure that idx 0 -> Counter0, and idx1 -> Counter1.
35  */
36 static struct perf_event *events[PMU_NR_COUNTERS];
37 
38 /* Find an unused counter */
39 static int l2x0_pmu_find_idx(void)
40 {
41 	int i;
42 
43 	for (i = 0; i < PMU_NR_COUNTERS; i++) {
44 		if (!events[i])
45 			return i;
46 	}
47 
48 	return -1;
49 }
50 
51 /* How many counters are allocated? */
52 static int l2x0_pmu_num_active_counters(void)
53 {
54 	int i, cnt = 0;
55 
56 	for (i = 0; i < PMU_NR_COUNTERS; i++) {
57 		if (events[i])
58 			cnt++;
59 	}
60 
61 	return cnt;
62 }
63 
64 static void l2x0_pmu_counter_config_write(int idx, u32 val)
65 {
66 	writel_relaxed(val, l2x0_base + L2X0_EVENT_CNT0_CFG - 4 * idx);
67 }
68 
69 static u32 l2x0_pmu_counter_read(int idx)
70 {
71 	return readl_relaxed(l2x0_base + L2X0_EVENT_CNT0_VAL - 4 * idx);
72 }
73 
74 static void l2x0_pmu_counter_write(int idx, u32 val)
75 {
76 	writel_relaxed(val, l2x0_base + L2X0_EVENT_CNT0_VAL - 4 * idx);
77 }
78 
79 static void __l2x0_pmu_enable(void)
80 {
81 	u32 val = readl_relaxed(l2x0_base + L2X0_EVENT_CNT_CTRL);
82 	val |= L2X0_EVENT_CNT_CTRL_ENABLE;
83 	writel_relaxed(val, l2x0_base + L2X0_EVENT_CNT_CTRL);
84 }
85 
86 static void __l2x0_pmu_disable(void)
87 {
88 	u32 val = readl_relaxed(l2x0_base + L2X0_EVENT_CNT_CTRL);
89 	val &= ~L2X0_EVENT_CNT_CTRL_ENABLE;
90 	writel_relaxed(val, l2x0_base + L2X0_EVENT_CNT_CTRL);
91 }
92 
93 static void l2x0_pmu_enable(struct pmu *pmu)
94 {
95 	if (l2x0_pmu_num_active_counters() == 0)
96 		return;
97 
98 	__l2x0_pmu_enable();
99 }
100 
101 static void l2x0_pmu_disable(struct pmu *pmu)
102 {
103 	if (l2x0_pmu_num_active_counters() == 0)
104 		return;
105 
106 	__l2x0_pmu_disable();
107 }
108 
109 static void warn_if_saturated(u32 count)
110 {
111 	if (count != 0xffffffff)
112 		return;
113 
114 	pr_warn_ratelimited("L2X0 counter saturated. Poll period too long\n");
115 }
116 
117 static void l2x0_pmu_event_read(struct perf_event *event)
118 {
119 	struct hw_perf_event *hw = &event->hw;
120 	u64 prev_count, new_count, mask;
121 
122 	do {
123 		 prev_count = local64_read(&hw->prev_count);
124 		 new_count = l2x0_pmu_counter_read(hw->idx);
125 	} while (local64_xchg(&hw->prev_count, new_count) != prev_count);
126 
127 	mask = GENMASK_ULL(31, 0);
128 	local64_add((new_count - prev_count) & mask, &event->count);
129 
130 	warn_if_saturated(new_count);
131 }
132 
133 static void l2x0_pmu_event_configure(struct perf_event *event)
134 {
135 	struct hw_perf_event *hw = &event->hw;
136 
137 	/*
138 	 * The L2X0 counters saturate at 0xffffffff rather than wrapping, so we
139 	 * will *always* lose some number of events when a counter saturates,
140 	 * and have no way of detecting how many were lost.
141 	 *
142 	 * To minimize the impact of this, we try to maximize the period by
143 	 * always starting counters at zero. To ensure that group ratios are
144 	 * representative, we poll periodically to avoid counters saturating.
145 	 * See l2x0_pmu_poll().
146 	 */
147 	local64_set(&hw->prev_count, 0);
148 	l2x0_pmu_counter_write(hw->idx, 0);
149 }
150 
151 static enum hrtimer_restart l2x0_pmu_poll(struct hrtimer *hrtimer)
152 {
153 	unsigned long flags;
154 	int i;
155 
156 	local_irq_save(flags);
157 	__l2x0_pmu_disable();
158 
159 	for (i = 0; i < PMU_NR_COUNTERS; i++) {
160 		struct perf_event *event = events[i];
161 
162 		if (!event)
163 			continue;
164 
165 		l2x0_pmu_event_read(event);
166 		l2x0_pmu_event_configure(event);
167 	}
168 
169 	__l2x0_pmu_enable();
170 	local_irq_restore(flags);
171 
172 	hrtimer_forward_now(hrtimer, l2x0_pmu_poll_period);
173 	return HRTIMER_RESTART;
174 }
175 
176 
177 static void __l2x0_pmu_event_enable(int idx, u32 event)
178 {
179 	u32 val;
180 
181 	val = event << L2X0_EVENT_CNT_CFG_SRC_SHIFT;
182 	val |= L2X0_EVENT_CNT_CFG_INT_DISABLED;
183 	l2x0_pmu_counter_config_write(idx, val);
184 }
185 
186 static void l2x0_pmu_event_start(struct perf_event *event, int flags)
187 {
188 	struct hw_perf_event *hw = &event->hw;
189 
190 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
191 		return;
192 
193 	if (flags & PERF_EF_RELOAD) {
194 		WARN_ON_ONCE(!(hw->state & PERF_HES_UPTODATE));
195 		l2x0_pmu_event_configure(event);
196 	}
197 
198 	hw->state = 0;
199 
200 	__l2x0_pmu_event_enable(hw->idx, hw->config_base);
201 }
202 
203 static void __l2x0_pmu_event_disable(int idx)
204 {
205 	u32 val;
206 
207 	val = L2X0_EVENT_CNT_CFG_SRC_DISABLED << L2X0_EVENT_CNT_CFG_SRC_SHIFT;
208 	val |= L2X0_EVENT_CNT_CFG_INT_DISABLED;
209 	l2x0_pmu_counter_config_write(idx, val);
210 }
211 
212 static void l2x0_pmu_event_stop(struct perf_event *event, int flags)
213 {
214 	struct hw_perf_event *hw = &event->hw;
215 
216 	if (WARN_ON_ONCE(event->hw.state & PERF_HES_STOPPED))
217 		return;
218 
219 	__l2x0_pmu_event_disable(hw->idx);
220 
221 	hw->state |= PERF_HES_STOPPED;
222 
223 	if (flags & PERF_EF_UPDATE) {
224 		l2x0_pmu_event_read(event);
225 		hw->state |= PERF_HES_UPTODATE;
226 	}
227 }
228 
229 static int l2x0_pmu_event_add(struct perf_event *event, int flags)
230 {
231 	struct hw_perf_event *hw = &event->hw;
232 	int idx = l2x0_pmu_find_idx();
233 
234 	if (idx == -1)
235 		return -EAGAIN;
236 
237 	/*
238 	 * Pin the timer, so that the overflows are handled by the chosen
239 	 * event->cpu (this is the same one as presented in "cpumask"
240 	 * attribute).
241 	 */
242 	if (l2x0_pmu_num_active_counters() == 0)
243 		hrtimer_start(&l2x0_pmu_hrtimer, l2x0_pmu_poll_period,
244 			      HRTIMER_MODE_REL_PINNED);
245 
246 	events[idx] = event;
247 	hw->idx = idx;
248 
249 	l2x0_pmu_event_configure(event);
250 
251 	hw->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
252 
253 	if (flags & PERF_EF_START)
254 		l2x0_pmu_event_start(event, 0);
255 
256 	return 0;
257 }
258 
259 static void l2x0_pmu_event_del(struct perf_event *event, int flags)
260 {
261 	struct hw_perf_event *hw = &event->hw;
262 
263 	l2x0_pmu_event_stop(event, PERF_EF_UPDATE);
264 
265 	events[hw->idx] = NULL;
266 	hw->idx = -1;
267 
268 	if (l2x0_pmu_num_active_counters() == 0)
269 		hrtimer_cancel(&l2x0_pmu_hrtimer);
270 }
271 
272 static bool l2x0_pmu_group_is_valid(struct perf_event *event)
273 {
274 	struct pmu *pmu = event->pmu;
275 	struct perf_event *leader = event->group_leader;
276 	struct perf_event *sibling;
277 	int num_hw = 0;
278 
279 	if (leader->pmu == pmu)
280 		num_hw++;
281 	else if (!is_software_event(leader))
282 		return false;
283 
284 	for_each_sibling_event(sibling, leader) {
285 		if (sibling->pmu == pmu)
286 			num_hw++;
287 		else if (!is_software_event(sibling))
288 			return false;
289 	}
290 
291 	return num_hw <= PMU_NR_COUNTERS;
292 }
293 
294 static int l2x0_pmu_event_init(struct perf_event *event)
295 {
296 	struct hw_perf_event *hw = &event->hw;
297 
298 	if (event->attr.type != l2x0_pmu->type)
299 		return -ENOENT;
300 
301 	if (is_sampling_event(event) ||
302 	    event->attach_state & PERF_ATTACH_TASK)
303 		return -EINVAL;
304 
305 	if (event->cpu < 0)
306 		return -EINVAL;
307 
308 	if (event->attr.config & ~L2X0_EVENT_CNT_CFG_SRC_MASK)
309 		return -EINVAL;
310 
311 	hw->config_base = event->attr.config;
312 
313 	if (!l2x0_pmu_group_is_valid(event))
314 		return -EINVAL;
315 
316 	event->cpu = cpumask_first(&pmu_cpu);
317 
318 	return 0;
319 }
320 
321 struct l2x0_event_attribute {
322 	struct device_attribute attr;
323 	unsigned int config;
324 	bool pl310_only;
325 };
326 
327 #define L2X0_EVENT_ATTR(_name, _config, _pl310_only)				\
328 	(&((struct l2x0_event_attribute[]) {{					\
329 		.attr = __ATTR(_name, S_IRUGO, l2x0_pmu_event_show, NULL),	\
330 		.config = _config,						\
331 		.pl310_only = _pl310_only,					\
332 	}})[0].attr.attr)
333 
334 #define L220_PLUS_EVENT_ATTR(_name, _config)					\
335 	L2X0_EVENT_ATTR(_name, _config, false)
336 
337 #define PL310_EVENT_ATTR(_name, _config)					\
338 	L2X0_EVENT_ATTR(_name, _config, true)
339 
340 static ssize_t l2x0_pmu_event_show(struct device *dev,
341 				   struct device_attribute *attr, char *buf)
342 {
343 	struct l2x0_event_attribute *lattr;
344 
345 	lattr = container_of(attr, typeof(*lattr), attr);
346 	return snprintf(buf, PAGE_SIZE, "config=0x%x\n", lattr->config);
347 }
348 
349 static umode_t l2x0_pmu_event_attr_is_visible(struct kobject *kobj,
350 					      struct attribute *attr,
351 					      int unused)
352 {
353 	struct device *dev = kobj_to_dev(kobj);
354 	struct pmu *pmu = dev_get_drvdata(dev);
355 	struct l2x0_event_attribute *lattr;
356 
357 	lattr = container_of(attr, typeof(*lattr), attr.attr);
358 
359 	if (!lattr->pl310_only || strcmp("l2c_310", pmu->name) == 0)
360 		return attr->mode;
361 
362 	return 0;
363 }
364 
365 static struct attribute *l2x0_pmu_event_attrs[] = {
366 	L220_PLUS_EVENT_ATTR(co,	0x1),
367 	L220_PLUS_EVENT_ATTR(drhit,	0x2),
368 	L220_PLUS_EVENT_ATTR(drreq,	0x3),
369 	L220_PLUS_EVENT_ATTR(dwhit,	0x4),
370 	L220_PLUS_EVENT_ATTR(dwreq,	0x5),
371 	L220_PLUS_EVENT_ATTR(dwtreq,	0x6),
372 	L220_PLUS_EVENT_ATTR(irhit,	0x7),
373 	L220_PLUS_EVENT_ATTR(irreq,	0x8),
374 	L220_PLUS_EVENT_ATTR(wa,	0x9),
375 	PL310_EVENT_ATTR(ipfalloc,	0xa),
376 	PL310_EVENT_ATTR(epfhit,	0xb),
377 	PL310_EVENT_ATTR(epfalloc,	0xc),
378 	PL310_EVENT_ATTR(srrcvd,	0xd),
379 	PL310_EVENT_ATTR(srconf,	0xe),
380 	PL310_EVENT_ATTR(epfrcvd,	0xf),
381 	NULL
382 };
383 
384 static struct attribute_group l2x0_pmu_event_attrs_group = {
385 	.name = "events",
386 	.attrs = l2x0_pmu_event_attrs,
387 	.is_visible = l2x0_pmu_event_attr_is_visible,
388 };
389 
390 static ssize_t l2x0_pmu_cpumask_show(struct device *dev,
391 				     struct device_attribute *attr, char *buf)
392 {
393 	return cpumap_print_to_pagebuf(true, buf, &pmu_cpu);
394 }
395 
396 static struct device_attribute l2x0_pmu_cpumask_attr =
397 		__ATTR(cpumask, S_IRUGO, l2x0_pmu_cpumask_show, NULL);
398 
399 static struct attribute *l2x0_pmu_cpumask_attrs[] = {
400 	&l2x0_pmu_cpumask_attr.attr,
401 	NULL,
402 };
403 
404 static struct attribute_group l2x0_pmu_cpumask_attr_group = {
405 	.attrs = l2x0_pmu_cpumask_attrs,
406 };
407 
408 static const struct attribute_group *l2x0_pmu_attr_groups[] = {
409 	&l2x0_pmu_event_attrs_group,
410 	&l2x0_pmu_cpumask_attr_group,
411 	NULL,
412 };
413 
414 static void l2x0_pmu_reset(void)
415 {
416 	int i;
417 
418 	__l2x0_pmu_disable();
419 
420 	for (i = 0; i < PMU_NR_COUNTERS; i++)
421 		__l2x0_pmu_event_disable(i);
422 }
423 
424 static int l2x0_pmu_offline_cpu(unsigned int cpu)
425 {
426 	unsigned int target;
427 
428 	if (!cpumask_test_and_clear_cpu(cpu, &pmu_cpu))
429 		return 0;
430 
431 	target = cpumask_any_but(cpu_online_mask, cpu);
432 	if (target >= nr_cpu_ids)
433 		return 0;
434 
435 	perf_pmu_migrate_context(l2x0_pmu, cpu, target);
436 	cpumask_set_cpu(target, &pmu_cpu);
437 
438 	return 0;
439 }
440 
441 void l2x0_pmu_suspend(void)
442 {
443 	int i;
444 
445 	if (!l2x0_pmu)
446 		return;
447 
448 	l2x0_pmu_disable(l2x0_pmu);
449 
450 	for (i = 0; i < PMU_NR_COUNTERS; i++) {
451 		if (events[i])
452 			l2x0_pmu_event_stop(events[i], PERF_EF_UPDATE);
453 	}
454 
455 }
456 
457 void l2x0_pmu_resume(void)
458 {
459 	int i;
460 
461 	if (!l2x0_pmu)
462 		return;
463 
464 	l2x0_pmu_reset();
465 
466 	for (i = 0; i < PMU_NR_COUNTERS; i++) {
467 		if (events[i])
468 			l2x0_pmu_event_start(events[i], PERF_EF_RELOAD);
469 	}
470 
471 	l2x0_pmu_enable(l2x0_pmu);
472 }
473 
474 void __init l2x0_pmu_register(void __iomem *base, u32 part)
475 {
476 	/*
477 	 * Determine whether we support the PMU, and choose the name for sysfs.
478 	 * This is also used by l2x0_pmu_event_attr_is_visible to determine
479 	 * which events to display, as the PL310 PMU supports a superset of
480 	 * L220 events.
481 	 *
482 	 * The L210 PMU has a different programmer's interface, and is not
483 	 * supported by this driver.
484 	 *
485 	 * We must defer registering the PMU until the perf subsystem is up and
486 	 * running, so just stash the name and base, and leave that to another
487 	 * initcall.
488 	 */
489 	switch (part & L2X0_CACHE_ID_PART_MASK) {
490 	case L2X0_CACHE_ID_PART_L220:
491 		l2x0_name = "l2c_220";
492 		break;
493 	case L2X0_CACHE_ID_PART_L310:
494 		l2x0_name = "l2c_310";
495 		break;
496 	default:
497 		return;
498 	}
499 
500 	l2x0_base = base;
501 }
502 
503 static __init int l2x0_pmu_init(void)
504 {
505 	int ret;
506 
507 	if (!l2x0_base)
508 		return 0;
509 
510 	l2x0_pmu = kzalloc(sizeof(*l2x0_pmu), GFP_KERNEL);
511 	if (!l2x0_pmu) {
512 		pr_warn("Unable to allocate L2x0 PMU\n");
513 		return -ENOMEM;
514 	}
515 
516 	*l2x0_pmu = (struct pmu) {
517 		.task_ctx_nr = perf_invalid_context,
518 		.pmu_enable = l2x0_pmu_enable,
519 		.pmu_disable = l2x0_pmu_disable,
520 		.read = l2x0_pmu_event_read,
521 		.start = l2x0_pmu_event_start,
522 		.stop = l2x0_pmu_event_stop,
523 		.add = l2x0_pmu_event_add,
524 		.del = l2x0_pmu_event_del,
525 		.event_init = l2x0_pmu_event_init,
526 		.attr_groups = l2x0_pmu_attr_groups,
527 		.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
528 	};
529 
530 	l2x0_pmu_reset();
531 
532 	/*
533 	 * We always use a hrtimer rather than an interrupt.
534 	 * See comments in l2x0_pmu_event_configure and l2x0_pmu_poll.
535 	 *
536 	 * Polling once a second allows the counters to fill up to 1/128th on a
537 	 * quad-core test chip with cores clocked at 400MHz. Hopefully this
538 	 * leaves sufficient headroom to avoid overflow on production silicon
539 	 * at higher frequencies.
540 	 */
541 	l2x0_pmu_poll_period = ms_to_ktime(1000);
542 	hrtimer_init(&l2x0_pmu_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
543 	l2x0_pmu_hrtimer.function = l2x0_pmu_poll;
544 
545 	cpumask_set_cpu(0, &pmu_cpu);
546 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_L2X0_ONLINE,
547 					"perf/arm/l2x0:online", NULL,
548 					l2x0_pmu_offline_cpu);
549 	if (ret)
550 		goto out_pmu;
551 
552 	ret = perf_pmu_register(l2x0_pmu, l2x0_name, -1);
553 	if (ret)
554 		goto out_cpuhp;
555 
556 	return 0;
557 
558 out_cpuhp:
559 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_ARM_L2X0_ONLINE);
560 out_pmu:
561 	kfree(l2x0_pmu);
562 	l2x0_pmu = NULL;
563 	return ret;
564 }
565 device_initcall(l2x0_pmu_init);
566