1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright 2017 NXP
4  * Copyright 2016 Freescale Semiconductor, Inc.
5  */
6 
7 #include <linux/bitfield.h>
8 #include <linux/init.h>
9 #include <linux/interrupt.h>
10 #include <linux/io.h>
11 #include <linux/module.h>
12 #include <linux/of.h>
13 #include <linux/of_address.h>
14 #include <linux/of_device.h>
15 #include <linux/of_irq.h>
16 #include <linux/perf_event.h>
17 #include <linux/slab.h>
18 
19 #define COUNTER_CNTL		0x0
20 #define COUNTER_READ		0x20
21 
22 #define COUNTER_DPCR1		0x30
23 
24 #define CNTL_OVER		0x1
25 #define CNTL_CLEAR		0x2
26 #define CNTL_EN			0x4
27 #define CNTL_EN_MASK		0xFFFFFFFB
28 #define CNTL_CLEAR_MASK		0xFFFFFFFD
29 #define CNTL_OVER_MASK		0xFFFFFFFE
30 
31 #define CNTL_CSV_SHIFT		24
32 #define CNTL_CSV_MASK		(0xFF << CNTL_CSV_SHIFT)
33 
34 #define EVENT_CYCLES_ID		0
35 #define EVENT_CYCLES_COUNTER	0
36 #define NUM_COUNTERS		4
37 
38 #define AXI_MASKING_REVERT	0xffff0000	/* AXI_MASKING(MSB 16bits) + AXI_ID(LSB 16bits) */
39 
40 #define to_ddr_pmu(p)		container_of(p, struct ddr_pmu, pmu)
41 
42 #define DDR_PERF_DEV_NAME	"imx8_ddr"
43 #define DDR_CPUHP_CB_NAME	DDR_PERF_DEV_NAME "_perf_pmu"
44 
45 static DEFINE_IDA(ddr_ida);
46 
47 /* DDR Perf hardware feature */
48 #define DDR_CAP_AXI_ID_FILTER			0x1     /* support AXI ID filter */
49 #define DDR_CAP_AXI_ID_FILTER_ENHANCED		0x3     /* support enhanced AXI ID filter */
50 
51 struct fsl_ddr_devtype_data {
52 	unsigned int quirks;    /* quirks needed for different DDR Perf core */
53 	const char *identifier;	/* system PMU identifier for userspace */
54 };
55 
56 static const struct fsl_ddr_devtype_data imx8_devtype_data;
57 
58 static const struct fsl_ddr_devtype_data imx8m_devtype_data = {
59 	.quirks = DDR_CAP_AXI_ID_FILTER,
60 };
61 
62 static const struct fsl_ddr_devtype_data imx8mq_devtype_data = {
63 	.quirks = DDR_CAP_AXI_ID_FILTER,
64 	.identifier = "i.MX8MQ",
65 };
66 
67 static const struct fsl_ddr_devtype_data imx8mm_devtype_data = {
68 	.quirks = DDR_CAP_AXI_ID_FILTER,
69 	.identifier = "i.MX8MM",
70 };
71 
72 static const struct fsl_ddr_devtype_data imx8mn_devtype_data = {
73 	.quirks = DDR_CAP_AXI_ID_FILTER,
74 	.identifier = "i.MX8MN",
75 };
76 
77 static const struct fsl_ddr_devtype_data imx8mp_devtype_data = {
78 	.quirks = DDR_CAP_AXI_ID_FILTER_ENHANCED,
79 	.identifier = "i.MX8MP",
80 };
81 
82 static const struct of_device_id imx_ddr_pmu_dt_ids[] = {
83 	{ .compatible = "fsl,imx8-ddr-pmu", .data = &imx8_devtype_data},
84 	{ .compatible = "fsl,imx8m-ddr-pmu", .data = &imx8m_devtype_data},
85 	{ .compatible = "fsl,imx8mq-ddr-pmu", .data = &imx8mq_devtype_data},
86 	{ .compatible = "fsl,imx8mm-ddr-pmu", .data = &imx8mm_devtype_data},
87 	{ .compatible = "fsl,imx8mn-ddr-pmu", .data = &imx8mn_devtype_data},
88 	{ .compatible = "fsl,imx8mp-ddr-pmu", .data = &imx8mp_devtype_data},
89 	{ /* sentinel */ }
90 };
91 MODULE_DEVICE_TABLE(of, imx_ddr_pmu_dt_ids);
92 
93 struct ddr_pmu {
94 	struct pmu pmu;
95 	void __iomem *base;
96 	unsigned int cpu;
97 	struct	hlist_node node;
98 	struct	device *dev;
99 	struct perf_event *events[NUM_COUNTERS];
100 	int active_events;
101 	enum cpuhp_state cpuhp_state;
102 	const struct fsl_ddr_devtype_data *devtype_data;
103 	int irq;
104 	int id;
105 };
106 
107 static ssize_t ddr_perf_identifier_show(struct device *dev,
108 					struct device_attribute *attr,
109 					char *page)
110 {
111 	struct ddr_pmu *pmu = dev_get_drvdata(dev);
112 
113 	return sysfs_emit(page, "%s\n", pmu->devtype_data->identifier);
114 }
115 
116 static umode_t ddr_perf_identifier_attr_visible(struct kobject *kobj,
117 						struct attribute *attr,
118 						int n)
119 {
120 	struct device *dev = kobj_to_dev(kobj);
121 	struct ddr_pmu *pmu = dev_get_drvdata(dev);
122 
123 	if (!pmu->devtype_data->identifier)
124 		return 0;
125 	return attr->mode;
126 };
127 
128 static struct device_attribute ddr_perf_identifier_attr =
129 	__ATTR(identifier, 0444, ddr_perf_identifier_show, NULL);
130 
131 static struct attribute *ddr_perf_identifier_attrs[] = {
132 	&ddr_perf_identifier_attr.attr,
133 	NULL,
134 };
135 
136 static const struct attribute_group ddr_perf_identifier_attr_group = {
137 	.attrs = ddr_perf_identifier_attrs,
138 	.is_visible = ddr_perf_identifier_attr_visible,
139 };
140 
141 enum ddr_perf_filter_capabilities {
142 	PERF_CAP_AXI_ID_FILTER = 0,
143 	PERF_CAP_AXI_ID_FILTER_ENHANCED,
144 	PERF_CAP_AXI_ID_FEAT_MAX,
145 };
146 
147 static u32 ddr_perf_filter_cap_get(struct ddr_pmu *pmu, int cap)
148 {
149 	u32 quirks = pmu->devtype_data->quirks;
150 
151 	switch (cap) {
152 	case PERF_CAP_AXI_ID_FILTER:
153 		return !!(quirks & DDR_CAP_AXI_ID_FILTER);
154 	case PERF_CAP_AXI_ID_FILTER_ENHANCED:
155 		quirks &= DDR_CAP_AXI_ID_FILTER_ENHANCED;
156 		return quirks == DDR_CAP_AXI_ID_FILTER_ENHANCED;
157 	default:
158 		WARN(1, "unknown filter cap %d\n", cap);
159 	}
160 
161 	return 0;
162 }
163 
164 static ssize_t ddr_perf_filter_cap_show(struct device *dev,
165 					struct device_attribute *attr,
166 					char *buf)
167 {
168 	struct ddr_pmu *pmu = dev_get_drvdata(dev);
169 	struct dev_ext_attribute *ea =
170 		container_of(attr, struct dev_ext_attribute, attr);
171 	int cap = (long)ea->var;
172 
173 	return sysfs_emit(buf, "%u\n", ddr_perf_filter_cap_get(pmu, cap));
174 }
175 
176 #define PERF_EXT_ATTR_ENTRY(_name, _func, _var)				\
177 	(&((struct dev_ext_attribute) {					\
178 		__ATTR(_name, 0444, _func, NULL), (void *)_var		\
179 	}).attr.attr)
180 
181 #define PERF_FILTER_EXT_ATTR_ENTRY(_name, _var)				\
182 	PERF_EXT_ATTR_ENTRY(_name, ddr_perf_filter_cap_show, _var)
183 
184 static struct attribute *ddr_perf_filter_cap_attr[] = {
185 	PERF_FILTER_EXT_ATTR_ENTRY(filter, PERF_CAP_AXI_ID_FILTER),
186 	PERF_FILTER_EXT_ATTR_ENTRY(enhanced_filter, PERF_CAP_AXI_ID_FILTER_ENHANCED),
187 	NULL,
188 };
189 
190 static const struct attribute_group ddr_perf_filter_cap_attr_group = {
191 	.name = "caps",
192 	.attrs = ddr_perf_filter_cap_attr,
193 };
194 
195 static ssize_t ddr_perf_cpumask_show(struct device *dev,
196 				struct device_attribute *attr, char *buf)
197 {
198 	struct ddr_pmu *pmu = dev_get_drvdata(dev);
199 
200 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(pmu->cpu));
201 }
202 
203 static struct device_attribute ddr_perf_cpumask_attr =
204 	__ATTR(cpumask, 0444, ddr_perf_cpumask_show, NULL);
205 
206 static struct attribute *ddr_perf_cpumask_attrs[] = {
207 	&ddr_perf_cpumask_attr.attr,
208 	NULL,
209 };
210 
211 static const struct attribute_group ddr_perf_cpumask_attr_group = {
212 	.attrs = ddr_perf_cpumask_attrs,
213 };
214 
215 static ssize_t
216 ddr_pmu_event_show(struct device *dev, struct device_attribute *attr,
217 		   char *page)
218 {
219 	struct perf_pmu_events_attr *pmu_attr;
220 
221 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
222 	return sysfs_emit(page, "event=0x%02llx\n", pmu_attr->id);
223 }
224 
225 #define IMX8_DDR_PMU_EVENT_ATTR(_name, _id)				\
226 	(&((struct perf_pmu_events_attr[]) {				\
227 		{ .attr = __ATTR(_name, 0444, ddr_pmu_event_show, NULL),\
228 		  .id = _id, }						\
229 	})[0].attr.attr)
230 
231 static struct attribute *ddr_perf_events_attrs[] = {
232 	IMX8_DDR_PMU_EVENT_ATTR(cycles, EVENT_CYCLES_ID),
233 	IMX8_DDR_PMU_EVENT_ATTR(selfresh, 0x01),
234 	IMX8_DDR_PMU_EVENT_ATTR(read-accesses, 0x04),
235 	IMX8_DDR_PMU_EVENT_ATTR(write-accesses, 0x05),
236 	IMX8_DDR_PMU_EVENT_ATTR(read-queue-depth, 0x08),
237 	IMX8_DDR_PMU_EVENT_ATTR(write-queue-depth, 0x09),
238 	IMX8_DDR_PMU_EVENT_ATTR(lp-read-credit-cnt, 0x10),
239 	IMX8_DDR_PMU_EVENT_ATTR(hp-read-credit-cnt, 0x11),
240 	IMX8_DDR_PMU_EVENT_ATTR(write-credit-cnt, 0x12),
241 	IMX8_DDR_PMU_EVENT_ATTR(read-command, 0x20),
242 	IMX8_DDR_PMU_EVENT_ATTR(write-command, 0x21),
243 	IMX8_DDR_PMU_EVENT_ATTR(read-modify-write-command, 0x22),
244 	IMX8_DDR_PMU_EVENT_ATTR(hp-read, 0x23),
245 	IMX8_DDR_PMU_EVENT_ATTR(hp-req-nocredit, 0x24),
246 	IMX8_DDR_PMU_EVENT_ATTR(hp-xact-credit, 0x25),
247 	IMX8_DDR_PMU_EVENT_ATTR(lp-req-nocredit, 0x26),
248 	IMX8_DDR_PMU_EVENT_ATTR(lp-xact-credit, 0x27),
249 	IMX8_DDR_PMU_EVENT_ATTR(wr-xact-credit, 0x29),
250 	IMX8_DDR_PMU_EVENT_ATTR(read-cycles, 0x2a),
251 	IMX8_DDR_PMU_EVENT_ATTR(write-cycles, 0x2b),
252 	IMX8_DDR_PMU_EVENT_ATTR(read-write-transition, 0x30),
253 	IMX8_DDR_PMU_EVENT_ATTR(precharge, 0x31),
254 	IMX8_DDR_PMU_EVENT_ATTR(activate, 0x32),
255 	IMX8_DDR_PMU_EVENT_ATTR(load-mode, 0x33),
256 	IMX8_DDR_PMU_EVENT_ATTR(perf-mwr, 0x34),
257 	IMX8_DDR_PMU_EVENT_ATTR(read, 0x35),
258 	IMX8_DDR_PMU_EVENT_ATTR(read-activate, 0x36),
259 	IMX8_DDR_PMU_EVENT_ATTR(refresh, 0x37),
260 	IMX8_DDR_PMU_EVENT_ATTR(write, 0x38),
261 	IMX8_DDR_PMU_EVENT_ATTR(raw-hazard, 0x39),
262 	IMX8_DDR_PMU_EVENT_ATTR(axid-read, 0x41),
263 	IMX8_DDR_PMU_EVENT_ATTR(axid-write, 0x42),
264 	NULL,
265 };
266 
267 static const struct attribute_group ddr_perf_events_attr_group = {
268 	.name = "events",
269 	.attrs = ddr_perf_events_attrs,
270 };
271 
272 PMU_FORMAT_ATTR(event, "config:0-7");
273 PMU_FORMAT_ATTR(axi_id, "config1:0-15");
274 PMU_FORMAT_ATTR(axi_mask, "config1:16-31");
275 
276 static struct attribute *ddr_perf_format_attrs[] = {
277 	&format_attr_event.attr,
278 	&format_attr_axi_id.attr,
279 	&format_attr_axi_mask.attr,
280 	NULL,
281 };
282 
283 static const struct attribute_group ddr_perf_format_attr_group = {
284 	.name = "format",
285 	.attrs = ddr_perf_format_attrs,
286 };
287 
288 static const struct attribute_group *attr_groups[] = {
289 	&ddr_perf_events_attr_group,
290 	&ddr_perf_format_attr_group,
291 	&ddr_perf_cpumask_attr_group,
292 	&ddr_perf_filter_cap_attr_group,
293 	&ddr_perf_identifier_attr_group,
294 	NULL,
295 };
296 
297 static bool ddr_perf_is_filtered(struct perf_event *event)
298 {
299 	return event->attr.config == 0x41 || event->attr.config == 0x42;
300 }
301 
302 static u32 ddr_perf_filter_val(struct perf_event *event)
303 {
304 	return event->attr.config1;
305 }
306 
307 static bool ddr_perf_filters_compatible(struct perf_event *a,
308 					struct perf_event *b)
309 {
310 	if (!ddr_perf_is_filtered(a))
311 		return true;
312 	if (!ddr_perf_is_filtered(b))
313 		return true;
314 	return ddr_perf_filter_val(a) == ddr_perf_filter_val(b);
315 }
316 
317 static bool ddr_perf_is_enhanced_filtered(struct perf_event *event)
318 {
319 	unsigned int filt;
320 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
321 
322 	filt = pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER_ENHANCED;
323 	return (filt == DDR_CAP_AXI_ID_FILTER_ENHANCED) &&
324 		ddr_perf_is_filtered(event);
325 }
326 
327 static u32 ddr_perf_alloc_counter(struct ddr_pmu *pmu, int event)
328 {
329 	int i;
330 
331 	/*
332 	 * Always map cycle event to counter 0
333 	 * Cycles counter is dedicated for cycle event
334 	 * can't used for the other events
335 	 */
336 	if (event == EVENT_CYCLES_ID) {
337 		if (pmu->events[EVENT_CYCLES_COUNTER] == NULL)
338 			return EVENT_CYCLES_COUNTER;
339 		else
340 			return -ENOENT;
341 	}
342 
343 	for (i = 1; i < NUM_COUNTERS; i++) {
344 		if (pmu->events[i] == NULL)
345 			return i;
346 	}
347 
348 	return -ENOENT;
349 }
350 
351 static void ddr_perf_free_counter(struct ddr_pmu *pmu, int counter)
352 {
353 	pmu->events[counter] = NULL;
354 }
355 
356 static u32 ddr_perf_read_counter(struct ddr_pmu *pmu, int counter)
357 {
358 	struct perf_event *event = pmu->events[counter];
359 	void __iomem *base = pmu->base;
360 
361 	/*
362 	 * return bytes instead of bursts from ddr transaction for
363 	 * axid-read and axid-write event if PMU core supports enhanced
364 	 * filter.
365 	 */
366 	base += ddr_perf_is_enhanced_filtered(event) ? COUNTER_DPCR1 :
367 						       COUNTER_READ;
368 	return readl_relaxed(base + counter * 4);
369 }
370 
371 static int ddr_perf_event_init(struct perf_event *event)
372 {
373 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
374 	struct hw_perf_event *hwc = &event->hw;
375 	struct perf_event *sibling;
376 
377 	if (event->attr.type != event->pmu->type)
378 		return -ENOENT;
379 
380 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
381 		return -EOPNOTSUPP;
382 
383 	if (event->cpu < 0) {
384 		dev_warn(pmu->dev, "Can't provide per-task data!\n");
385 		return -EOPNOTSUPP;
386 	}
387 
388 	/*
389 	 * We must NOT create groups containing mixed PMUs, although software
390 	 * events are acceptable (for example to create a CCN group
391 	 * periodically read when a hrtimer aka cpu-clock leader triggers).
392 	 */
393 	if (event->group_leader->pmu != event->pmu &&
394 			!is_software_event(event->group_leader))
395 		return -EINVAL;
396 
397 	if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
398 		if (!ddr_perf_filters_compatible(event, event->group_leader))
399 			return -EINVAL;
400 		for_each_sibling_event(sibling, event->group_leader) {
401 			if (!ddr_perf_filters_compatible(event, sibling))
402 				return -EINVAL;
403 		}
404 	}
405 
406 	for_each_sibling_event(sibling, event->group_leader) {
407 		if (sibling->pmu != event->pmu &&
408 				!is_software_event(sibling))
409 			return -EINVAL;
410 	}
411 
412 	event->cpu = pmu->cpu;
413 	hwc->idx = -1;
414 
415 	return 0;
416 }
417 
418 static void ddr_perf_counter_enable(struct ddr_pmu *pmu, int config,
419 				  int counter, bool enable)
420 {
421 	u8 reg = counter * 4 + COUNTER_CNTL;
422 	int val;
423 
424 	if (enable) {
425 		/*
426 		 * cycle counter is special which should firstly write 0 then
427 		 * write 1 into CLEAR bit to clear it. Other counters only
428 		 * need write 0 into CLEAR bit and it turns out to be 1 by
429 		 * hardware. Below enable flow is harmless for all counters.
430 		 */
431 		writel(0, pmu->base + reg);
432 		val = CNTL_EN | CNTL_CLEAR;
433 		val |= FIELD_PREP(CNTL_CSV_MASK, config);
434 		writel(val, pmu->base + reg);
435 	} else {
436 		/* Disable counter */
437 		val = readl_relaxed(pmu->base + reg) & CNTL_EN_MASK;
438 		writel(val, pmu->base + reg);
439 	}
440 }
441 
442 static bool ddr_perf_counter_overflow(struct ddr_pmu *pmu, int counter)
443 {
444 	int val;
445 
446 	val = readl_relaxed(pmu->base + counter * 4 + COUNTER_CNTL);
447 
448 	return val & CNTL_OVER;
449 }
450 
451 static void ddr_perf_counter_clear(struct ddr_pmu *pmu, int counter)
452 {
453 	u8 reg = counter * 4 + COUNTER_CNTL;
454 	int val;
455 
456 	val = readl_relaxed(pmu->base + reg);
457 	val &= ~CNTL_CLEAR;
458 	writel(val, pmu->base + reg);
459 
460 	val |= CNTL_CLEAR;
461 	writel(val, pmu->base + reg);
462 }
463 
464 static void ddr_perf_event_update(struct perf_event *event)
465 {
466 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
467 	struct hw_perf_event *hwc = &event->hw;
468 	u64 new_raw_count;
469 	int counter = hwc->idx;
470 	int ret;
471 
472 	new_raw_count = ddr_perf_read_counter(pmu, counter);
473 	local64_add(new_raw_count, &event->count);
474 
475 	/*
476 	 * For legacy SoCs: event counter continue counting when overflow,
477 	 *                  no need to clear the counter.
478 	 * For new SoCs: event counter stop counting when overflow, need
479 	 *               clear counter to let it count again.
480 	 */
481 	if (counter != EVENT_CYCLES_COUNTER) {
482 		ret = ddr_perf_counter_overflow(pmu, counter);
483 		if (ret)
484 			dev_warn_ratelimited(pmu->dev,  "events lost due to counter overflow (config 0x%llx)\n",
485 					     event->attr.config);
486 	}
487 
488 	/* clear counter every time for both cycle counter and event counter */
489 	ddr_perf_counter_clear(pmu, counter);
490 }
491 
492 static void ddr_perf_event_start(struct perf_event *event, int flags)
493 {
494 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
495 	struct hw_perf_event *hwc = &event->hw;
496 	int counter = hwc->idx;
497 
498 	local64_set(&hwc->prev_count, 0);
499 
500 	ddr_perf_counter_enable(pmu, event->attr.config, counter, true);
501 
502 	hwc->state = 0;
503 }
504 
505 static int ddr_perf_event_add(struct perf_event *event, int flags)
506 {
507 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
508 	struct hw_perf_event *hwc = &event->hw;
509 	int counter;
510 	int cfg = event->attr.config;
511 	int cfg1 = event->attr.config1;
512 
513 	if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
514 		int i;
515 
516 		for (i = 1; i < NUM_COUNTERS; i++) {
517 			if (pmu->events[i] &&
518 			    !ddr_perf_filters_compatible(event, pmu->events[i]))
519 				return -EINVAL;
520 		}
521 
522 		if (ddr_perf_is_filtered(event)) {
523 			/* revert axi id masking(axi_mask) value */
524 			cfg1 ^= AXI_MASKING_REVERT;
525 			writel(cfg1, pmu->base + COUNTER_DPCR1);
526 		}
527 	}
528 
529 	counter = ddr_perf_alloc_counter(pmu, cfg);
530 	if (counter < 0) {
531 		dev_dbg(pmu->dev, "There are not enough counters\n");
532 		return -EOPNOTSUPP;
533 	}
534 
535 	pmu->events[counter] = event;
536 	pmu->active_events++;
537 	hwc->idx = counter;
538 
539 	hwc->state |= PERF_HES_STOPPED;
540 
541 	if (flags & PERF_EF_START)
542 		ddr_perf_event_start(event, flags);
543 
544 	return 0;
545 }
546 
547 static void ddr_perf_event_stop(struct perf_event *event, int flags)
548 {
549 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
550 	struct hw_perf_event *hwc = &event->hw;
551 	int counter = hwc->idx;
552 
553 	ddr_perf_counter_enable(pmu, event->attr.config, counter, false);
554 	ddr_perf_event_update(event);
555 
556 	hwc->state |= PERF_HES_STOPPED;
557 }
558 
559 static void ddr_perf_event_del(struct perf_event *event, int flags)
560 {
561 	struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
562 	struct hw_perf_event *hwc = &event->hw;
563 	int counter = hwc->idx;
564 
565 	ddr_perf_event_stop(event, PERF_EF_UPDATE);
566 
567 	ddr_perf_free_counter(pmu, counter);
568 	pmu->active_events--;
569 	hwc->idx = -1;
570 }
571 
572 static void ddr_perf_pmu_enable(struct pmu *pmu)
573 {
574 	struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
575 
576 	/* enable cycle counter if cycle is not active event list */
577 	if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
578 		ddr_perf_counter_enable(ddr_pmu,
579 				      EVENT_CYCLES_ID,
580 				      EVENT_CYCLES_COUNTER,
581 				      true);
582 }
583 
584 static void ddr_perf_pmu_disable(struct pmu *pmu)
585 {
586 	struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
587 
588 	if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
589 		ddr_perf_counter_enable(ddr_pmu,
590 				      EVENT_CYCLES_ID,
591 				      EVENT_CYCLES_COUNTER,
592 				      false);
593 }
594 
595 static int ddr_perf_init(struct ddr_pmu *pmu, void __iomem *base,
596 			 struct device *dev)
597 {
598 	*pmu = (struct ddr_pmu) {
599 		.pmu = (struct pmu) {
600 			.module	      = THIS_MODULE,
601 			.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
602 			.task_ctx_nr = perf_invalid_context,
603 			.attr_groups = attr_groups,
604 			.event_init  = ddr_perf_event_init,
605 			.add	     = ddr_perf_event_add,
606 			.del	     = ddr_perf_event_del,
607 			.start	     = ddr_perf_event_start,
608 			.stop	     = ddr_perf_event_stop,
609 			.read	     = ddr_perf_event_update,
610 			.pmu_enable  = ddr_perf_pmu_enable,
611 			.pmu_disable = ddr_perf_pmu_disable,
612 		},
613 		.base = base,
614 		.dev = dev,
615 	};
616 
617 	pmu->id = ida_simple_get(&ddr_ida, 0, 0, GFP_KERNEL);
618 	return pmu->id;
619 }
620 
621 static irqreturn_t ddr_perf_irq_handler(int irq, void *p)
622 {
623 	int i;
624 	struct ddr_pmu *pmu = (struct ddr_pmu *) p;
625 	struct perf_event *event;
626 
627 	/* all counter will stop if cycle counter disabled */
628 	ddr_perf_counter_enable(pmu,
629 			      EVENT_CYCLES_ID,
630 			      EVENT_CYCLES_COUNTER,
631 			      false);
632 	/*
633 	 * When the cycle counter overflows, all counters are stopped,
634 	 * and an IRQ is raised. If any other counter overflows, it
635 	 * continues counting, and no IRQ is raised. But for new SoCs,
636 	 * such as i.MX8MP, event counter would stop when overflow, so
637 	 * we need use cycle counter to stop overflow of event counter.
638 	 *
639 	 * Cycles occur at least 4 times as often as other events, so we
640 	 * can update all events on a cycle counter overflow and not
641 	 * lose events.
642 	 *
643 	 */
644 	for (i = 0; i < NUM_COUNTERS; i++) {
645 
646 		if (!pmu->events[i])
647 			continue;
648 
649 		event = pmu->events[i];
650 
651 		ddr_perf_event_update(event);
652 	}
653 
654 	ddr_perf_counter_enable(pmu,
655 			      EVENT_CYCLES_ID,
656 			      EVENT_CYCLES_COUNTER,
657 			      true);
658 
659 	return IRQ_HANDLED;
660 }
661 
662 static int ddr_perf_offline_cpu(unsigned int cpu, struct hlist_node *node)
663 {
664 	struct ddr_pmu *pmu = hlist_entry_safe(node, struct ddr_pmu, node);
665 	int target;
666 
667 	if (cpu != pmu->cpu)
668 		return 0;
669 
670 	target = cpumask_any_but(cpu_online_mask, cpu);
671 	if (target >= nr_cpu_ids)
672 		return 0;
673 
674 	perf_pmu_migrate_context(&pmu->pmu, cpu, target);
675 	pmu->cpu = target;
676 
677 	WARN_ON(irq_set_affinity_hint(pmu->irq, cpumask_of(pmu->cpu)));
678 
679 	return 0;
680 }
681 
682 static int ddr_perf_probe(struct platform_device *pdev)
683 {
684 	struct ddr_pmu *pmu;
685 	struct device_node *np;
686 	void __iomem *base;
687 	char *name;
688 	int num;
689 	int ret;
690 	int irq;
691 
692 	base = devm_platform_ioremap_resource(pdev, 0);
693 	if (IS_ERR(base))
694 		return PTR_ERR(base);
695 
696 	np = pdev->dev.of_node;
697 
698 	pmu = devm_kzalloc(&pdev->dev, sizeof(*pmu), GFP_KERNEL);
699 	if (!pmu)
700 		return -ENOMEM;
701 
702 	num = ddr_perf_init(pmu, base, &pdev->dev);
703 
704 	platform_set_drvdata(pdev, pmu);
705 
706 	name = devm_kasprintf(&pdev->dev, GFP_KERNEL, DDR_PERF_DEV_NAME "%d",
707 			      num);
708 	if (!name)
709 		return -ENOMEM;
710 
711 	pmu->devtype_data = of_device_get_match_data(&pdev->dev);
712 
713 	pmu->cpu = raw_smp_processor_id();
714 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
715 				      DDR_CPUHP_CB_NAME,
716 				      NULL,
717 				      ddr_perf_offline_cpu);
718 
719 	if (ret < 0) {
720 		dev_err(&pdev->dev, "cpuhp_setup_state_multi failed\n");
721 		goto cpuhp_state_err;
722 	}
723 
724 	pmu->cpuhp_state = ret;
725 
726 	/* Register the pmu instance for cpu hotplug */
727 	ret = cpuhp_state_add_instance_nocalls(pmu->cpuhp_state, &pmu->node);
728 	if (ret) {
729 		dev_err(&pdev->dev, "Error %d registering hotplug\n", ret);
730 		goto cpuhp_instance_err;
731 	}
732 
733 	/* Request irq */
734 	irq = of_irq_get(np, 0);
735 	if (irq < 0) {
736 		dev_err(&pdev->dev, "Failed to get irq: %d", irq);
737 		ret = irq;
738 		goto ddr_perf_err;
739 	}
740 
741 	ret = devm_request_irq(&pdev->dev, irq,
742 					ddr_perf_irq_handler,
743 					IRQF_NOBALANCING | IRQF_NO_THREAD,
744 					DDR_CPUHP_CB_NAME,
745 					pmu);
746 	if (ret < 0) {
747 		dev_err(&pdev->dev, "Request irq failed: %d", ret);
748 		goto ddr_perf_err;
749 	}
750 
751 	pmu->irq = irq;
752 	ret = irq_set_affinity_hint(pmu->irq, cpumask_of(pmu->cpu));
753 	if (ret) {
754 		dev_err(pmu->dev, "Failed to set interrupt affinity!\n");
755 		goto ddr_perf_err;
756 	}
757 
758 	ret = perf_pmu_register(&pmu->pmu, name, -1);
759 	if (ret)
760 		goto ddr_perf_err;
761 
762 	return 0;
763 
764 ddr_perf_err:
765 	cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
766 cpuhp_instance_err:
767 	cpuhp_remove_multi_state(pmu->cpuhp_state);
768 cpuhp_state_err:
769 	ida_simple_remove(&ddr_ida, pmu->id);
770 	dev_warn(&pdev->dev, "i.MX8 DDR Perf PMU failed (%d), disabled\n", ret);
771 	return ret;
772 }
773 
774 static int ddr_perf_remove(struct platform_device *pdev)
775 {
776 	struct ddr_pmu *pmu = platform_get_drvdata(pdev);
777 
778 	cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
779 	cpuhp_remove_multi_state(pmu->cpuhp_state);
780 	irq_set_affinity_hint(pmu->irq, NULL);
781 
782 	perf_pmu_unregister(&pmu->pmu);
783 
784 	ida_simple_remove(&ddr_ida, pmu->id);
785 	return 0;
786 }
787 
788 static struct platform_driver imx_ddr_pmu_driver = {
789 	.driver         = {
790 		.name   = "imx-ddr-pmu",
791 		.of_match_table = imx_ddr_pmu_dt_ids,
792 		.suppress_bind_attrs = true,
793 	},
794 	.probe          = ddr_perf_probe,
795 	.remove         = ddr_perf_remove,
796 };
797 
798 module_platform_driver(imx_ddr_pmu_driver);
799 MODULE_LICENSE("GPL v2");
800