Loading...
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 (0xFFU << 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
45static 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
51struct 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
56static const struct fsl_ddr_devtype_data imx8_devtype_data;
57
58static const struct fsl_ddr_devtype_data imx8m_devtype_data = {
59 .quirks = DDR_CAP_AXI_ID_FILTER,
60};
61
62static const struct fsl_ddr_devtype_data imx8mq_devtype_data = {
63 .quirks = DDR_CAP_AXI_ID_FILTER,
64 .identifier = "i.MX8MQ",
65};
66
67static const struct fsl_ddr_devtype_data imx8mm_devtype_data = {
68 .quirks = DDR_CAP_AXI_ID_FILTER,
69 .identifier = "i.MX8MM",
70};
71
72static const struct fsl_ddr_devtype_data imx8mn_devtype_data = {
73 .quirks = DDR_CAP_AXI_ID_FILTER,
74 .identifier = "i.MX8MN",
75};
76
77static const struct fsl_ddr_devtype_data imx8mp_devtype_data = {
78 .quirks = DDR_CAP_AXI_ID_FILTER_ENHANCED,
79 .identifier = "i.MX8MP",
80};
81
82static 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};
91MODULE_DEVICE_TABLE(of, imx_ddr_pmu_dt_ids);
92
93struct 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
107static 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
116static 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
128static struct device_attribute ddr_perf_identifier_attr =
129 __ATTR(identifier, 0444, ddr_perf_identifier_show, NULL);
130
131static struct attribute *ddr_perf_identifier_attrs[] = {
132 &ddr_perf_identifier_attr.attr,
133 NULL,
134};
135
136static 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
141enum 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
147static 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
164static 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
184static 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
190static const struct attribute_group ddr_perf_filter_cap_attr_group = {
191 .name = "caps",
192 .attrs = ddr_perf_filter_cap_attr,
193};
194
195static 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
203static struct device_attribute ddr_perf_cpumask_attr =
204 __ATTR(cpumask, 0444, ddr_perf_cpumask_show, NULL);
205
206static struct attribute *ddr_perf_cpumask_attrs[] = {
207 &ddr_perf_cpumask_attr.attr,
208 NULL,
209};
210
211static const struct attribute_group ddr_perf_cpumask_attr_group = {
212 .attrs = ddr_perf_cpumask_attrs,
213};
214
215static ssize_t
216ddr_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 PMU_EVENT_ATTR_ID(_name, ddr_pmu_event_show, _id)
227
228static struct attribute *ddr_perf_events_attrs[] = {
229 IMX8_DDR_PMU_EVENT_ATTR(cycles, EVENT_CYCLES_ID),
230 IMX8_DDR_PMU_EVENT_ATTR(selfresh, 0x01),
231 IMX8_DDR_PMU_EVENT_ATTR(read-accesses, 0x04),
232 IMX8_DDR_PMU_EVENT_ATTR(write-accesses, 0x05),
233 IMX8_DDR_PMU_EVENT_ATTR(read-queue-depth, 0x08),
234 IMX8_DDR_PMU_EVENT_ATTR(write-queue-depth, 0x09),
235 IMX8_DDR_PMU_EVENT_ATTR(lp-read-credit-cnt, 0x10),
236 IMX8_DDR_PMU_EVENT_ATTR(hp-read-credit-cnt, 0x11),
237 IMX8_DDR_PMU_EVENT_ATTR(write-credit-cnt, 0x12),
238 IMX8_DDR_PMU_EVENT_ATTR(read-command, 0x20),
239 IMX8_DDR_PMU_EVENT_ATTR(write-command, 0x21),
240 IMX8_DDR_PMU_EVENT_ATTR(read-modify-write-command, 0x22),
241 IMX8_DDR_PMU_EVENT_ATTR(hp-read, 0x23),
242 IMX8_DDR_PMU_EVENT_ATTR(hp-req-nocredit, 0x24),
243 IMX8_DDR_PMU_EVENT_ATTR(hp-xact-credit, 0x25),
244 IMX8_DDR_PMU_EVENT_ATTR(lp-req-nocredit, 0x26),
245 IMX8_DDR_PMU_EVENT_ATTR(lp-xact-credit, 0x27),
246 IMX8_DDR_PMU_EVENT_ATTR(wr-xact-credit, 0x29),
247 IMX8_DDR_PMU_EVENT_ATTR(read-cycles, 0x2a),
248 IMX8_DDR_PMU_EVENT_ATTR(write-cycles, 0x2b),
249 IMX8_DDR_PMU_EVENT_ATTR(read-write-transition, 0x30),
250 IMX8_DDR_PMU_EVENT_ATTR(precharge, 0x31),
251 IMX8_DDR_PMU_EVENT_ATTR(activate, 0x32),
252 IMX8_DDR_PMU_EVENT_ATTR(load-mode, 0x33),
253 IMX8_DDR_PMU_EVENT_ATTR(perf-mwr, 0x34),
254 IMX8_DDR_PMU_EVENT_ATTR(read, 0x35),
255 IMX8_DDR_PMU_EVENT_ATTR(read-activate, 0x36),
256 IMX8_DDR_PMU_EVENT_ATTR(refresh, 0x37),
257 IMX8_DDR_PMU_EVENT_ATTR(write, 0x38),
258 IMX8_DDR_PMU_EVENT_ATTR(raw-hazard, 0x39),
259 IMX8_DDR_PMU_EVENT_ATTR(axid-read, 0x41),
260 IMX8_DDR_PMU_EVENT_ATTR(axid-write, 0x42),
261 NULL,
262};
263
264static const struct attribute_group ddr_perf_events_attr_group = {
265 .name = "events",
266 .attrs = ddr_perf_events_attrs,
267};
268
269PMU_FORMAT_ATTR(event, "config:0-7");
270PMU_FORMAT_ATTR(axi_id, "config1:0-15");
271PMU_FORMAT_ATTR(axi_mask, "config1:16-31");
272
273static struct attribute *ddr_perf_format_attrs[] = {
274 &format_attr_event.attr,
275 &format_attr_axi_id.attr,
276 &format_attr_axi_mask.attr,
277 NULL,
278};
279
280static const struct attribute_group ddr_perf_format_attr_group = {
281 .name = "format",
282 .attrs = ddr_perf_format_attrs,
283};
284
285static const struct attribute_group *attr_groups[] = {
286 &ddr_perf_events_attr_group,
287 &ddr_perf_format_attr_group,
288 &ddr_perf_cpumask_attr_group,
289 &ddr_perf_filter_cap_attr_group,
290 &ddr_perf_identifier_attr_group,
291 NULL,
292};
293
294static bool ddr_perf_is_filtered(struct perf_event *event)
295{
296 return event->attr.config == 0x41 || event->attr.config == 0x42;
297}
298
299static u32 ddr_perf_filter_val(struct perf_event *event)
300{
301 return event->attr.config1;
302}
303
304static bool ddr_perf_filters_compatible(struct perf_event *a,
305 struct perf_event *b)
306{
307 if (!ddr_perf_is_filtered(a))
308 return true;
309 if (!ddr_perf_is_filtered(b))
310 return true;
311 return ddr_perf_filter_val(a) == ddr_perf_filter_val(b);
312}
313
314static bool ddr_perf_is_enhanced_filtered(struct perf_event *event)
315{
316 unsigned int filt;
317 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
318
319 filt = pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER_ENHANCED;
320 return (filt == DDR_CAP_AXI_ID_FILTER_ENHANCED) &&
321 ddr_perf_is_filtered(event);
322}
323
324static u32 ddr_perf_alloc_counter(struct ddr_pmu *pmu, int event)
325{
326 int i;
327
328 /*
329 * Always map cycle event to counter 0
330 * Cycles counter is dedicated for cycle event
331 * can't used for the other events
332 */
333 if (event == EVENT_CYCLES_ID) {
334 if (pmu->events[EVENT_CYCLES_COUNTER] == NULL)
335 return EVENT_CYCLES_COUNTER;
336 else
337 return -ENOENT;
338 }
339
340 for (i = 1; i < NUM_COUNTERS; i++) {
341 if (pmu->events[i] == NULL)
342 return i;
343 }
344
345 return -ENOENT;
346}
347
348static void ddr_perf_free_counter(struct ddr_pmu *pmu, int counter)
349{
350 pmu->events[counter] = NULL;
351}
352
353static u32 ddr_perf_read_counter(struct ddr_pmu *pmu, int counter)
354{
355 struct perf_event *event = pmu->events[counter];
356 void __iomem *base = pmu->base;
357
358 /*
359 * return bytes instead of bursts from ddr transaction for
360 * axid-read and axid-write event if PMU core supports enhanced
361 * filter.
362 */
363 base += ddr_perf_is_enhanced_filtered(event) ? COUNTER_DPCR1 :
364 COUNTER_READ;
365 return readl_relaxed(base + counter * 4);
366}
367
368static int ddr_perf_event_init(struct perf_event *event)
369{
370 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
371 struct hw_perf_event *hwc = &event->hw;
372 struct perf_event *sibling;
373
374 if (event->attr.type != event->pmu->type)
375 return -ENOENT;
376
377 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
378 return -EOPNOTSUPP;
379
380 if (event->cpu < 0) {
381 dev_warn(pmu->dev, "Can't provide per-task data!\n");
382 return -EOPNOTSUPP;
383 }
384
385 /*
386 * We must NOT create groups containing mixed PMUs, although software
387 * events are acceptable (for example to create a CCN group
388 * periodically read when a hrtimer aka cpu-clock leader triggers).
389 */
390 if (event->group_leader->pmu != event->pmu &&
391 !is_software_event(event->group_leader))
392 return -EINVAL;
393
394 if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
395 if (!ddr_perf_filters_compatible(event, event->group_leader))
396 return -EINVAL;
397 for_each_sibling_event(sibling, event->group_leader) {
398 if (!ddr_perf_filters_compatible(event, sibling))
399 return -EINVAL;
400 }
401 }
402
403 for_each_sibling_event(sibling, event->group_leader) {
404 if (sibling->pmu != event->pmu &&
405 !is_software_event(sibling))
406 return -EINVAL;
407 }
408
409 event->cpu = pmu->cpu;
410 hwc->idx = -1;
411
412 return 0;
413}
414
415static void ddr_perf_counter_enable(struct ddr_pmu *pmu, int config,
416 int counter, bool enable)
417{
418 u8 reg = counter * 4 + COUNTER_CNTL;
419 int val;
420
421 if (enable) {
422 /*
423 * cycle counter is special which should firstly write 0 then
424 * write 1 into CLEAR bit to clear it. Other counters only
425 * need write 0 into CLEAR bit and it turns out to be 1 by
426 * hardware. Below enable flow is harmless for all counters.
427 */
428 writel(0, pmu->base + reg);
429 val = CNTL_EN | CNTL_CLEAR;
430 val |= FIELD_PREP(CNTL_CSV_MASK, config);
431 writel(val, pmu->base + reg);
432 } else {
433 /* Disable counter */
434 val = readl_relaxed(pmu->base + reg) & CNTL_EN_MASK;
435 writel(val, pmu->base + reg);
436 }
437}
438
439static bool ddr_perf_counter_overflow(struct ddr_pmu *pmu, int counter)
440{
441 int val;
442
443 val = readl_relaxed(pmu->base + counter * 4 + COUNTER_CNTL);
444
445 return val & CNTL_OVER;
446}
447
448static void ddr_perf_counter_clear(struct ddr_pmu *pmu, int counter)
449{
450 u8 reg = counter * 4 + COUNTER_CNTL;
451 int val;
452
453 val = readl_relaxed(pmu->base + reg);
454 val &= ~CNTL_CLEAR;
455 writel(val, pmu->base + reg);
456
457 val |= CNTL_CLEAR;
458 writel(val, pmu->base + reg);
459}
460
461static void ddr_perf_event_update(struct perf_event *event)
462{
463 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
464 struct hw_perf_event *hwc = &event->hw;
465 u64 new_raw_count;
466 int counter = hwc->idx;
467 int ret;
468
469 new_raw_count = ddr_perf_read_counter(pmu, counter);
470 local64_add(new_raw_count, &event->count);
471
472 /*
473 * For legacy SoCs: event counter continue counting when overflow,
474 * no need to clear the counter.
475 * For new SoCs: event counter stop counting when overflow, need
476 * clear counter to let it count again.
477 */
478 if (counter != EVENT_CYCLES_COUNTER) {
479 ret = ddr_perf_counter_overflow(pmu, counter);
480 if (ret)
481 dev_warn_ratelimited(pmu->dev, "events lost due to counter overflow (config 0x%llx)\n",
482 event->attr.config);
483 }
484
485 /* clear counter every time for both cycle counter and event counter */
486 ddr_perf_counter_clear(pmu, counter);
487}
488
489static void ddr_perf_event_start(struct perf_event *event, int flags)
490{
491 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
492 struct hw_perf_event *hwc = &event->hw;
493 int counter = hwc->idx;
494
495 local64_set(&hwc->prev_count, 0);
496
497 ddr_perf_counter_enable(pmu, event->attr.config, counter, true);
498
499 hwc->state = 0;
500}
501
502static int ddr_perf_event_add(struct perf_event *event, int flags)
503{
504 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
505 struct hw_perf_event *hwc = &event->hw;
506 int counter;
507 int cfg = event->attr.config;
508 int cfg1 = event->attr.config1;
509
510 if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
511 int i;
512
513 for (i = 1; i < NUM_COUNTERS; i++) {
514 if (pmu->events[i] &&
515 !ddr_perf_filters_compatible(event, pmu->events[i]))
516 return -EINVAL;
517 }
518
519 if (ddr_perf_is_filtered(event)) {
520 /* revert axi id masking(axi_mask) value */
521 cfg1 ^= AXI_MASKING_REVERT;
522 writel(cfg1, pmu->base + COUNTER_DPCR1);
523 }
524 }
525
526 counter = ddr_perf_alloc_counter(pmu, cfg);
527 if (counter < 0) {
528 dev_dbg(pmu->dev, "There are not enough counters\n");
529 return -EOPNOTSUPP;
530 }
531
532 pmu->events[counter] = event;
533 pmu->active_events++;
534 hwc->idx = counter;
535
536 hwc->state |= PERF_HES_STOPPED;
537
538 if (flags & PERF_EF_START)
539 ddr_perf_event_start(event, flags);
540
541 return 0;
542}
543
544static void ddr_perf_event_stop(struct perf_event *event, int flags)
545{
546 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
547 struct hw_perf_event *hwc = &event->hw;
548 int counter = hwc->idx;
549
550 ddr_perf_counter_enable(pmu, event->attr.config, counter, false);
551 ddr_perf_event_update(event);
552
553 hwc->state |= PERF_HES_STOPPED;
554}
555
556static void ddr_perf_event_del(struct perf_event *event, int flags)
557{
558 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
559 struct hw_perf_event *hwc = &event->hw;
560 int counter = hwc->idx;
561
562 ddr_perf_event_stop(event, PERF_EF_UPDATE);
563
564 ddr_perf_free_counter(pmu, counter);
565 pmu->active_events--;
566 hwc->idx = -1;
567}
568
569static void ddr_perf_pmu_enable(struct pmu *pmu)
570{
571 struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
572
573 /* enable cycle counter if cycle is not active event list */
574 if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
575 ddr_perf_counter_enable(ddr_pmu,
576 EVENT_CYCLES_ID,
577 EVENT_CYCLES_COUNTER,
578 true);
579}
580
581static void ddr_perf_pmu_disable(struct pmu *pmu)
582{
583 struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
584
585 if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
586 ddr_perf_counter_enable(ddr_pmu,
587 EVENT_CYCLES_ID,
588 EVENT_CYCLES_COUNTER,
589 false);
590}
591
592static int ddr_perf_init(struct ddr_pmu *pmu, void __iomem *base,
593 struct device *dev)
594{
595 *pmu = (struct ddr_pmu) {
596 .pmu = (struct pmu) {
597 .module = THIS_MODULE,
598 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
599 .task_ctx_nr = perf_invalid_context,
600 .attr_groups = attr_groups,
601 .event_init = ddr_perf_event_init,
602 .add = ddr_perf_event_add,
603 .del = ddr_perf_event_del,
604 .start = ddr_perf_event_start,
605 .stop = ddr_perf_event_stop,
606 .read = ddr_perf_event_update,
607 .pmu_enable = ddr_perf_pmu_enable,
608 .pmu_disable = ddr_perf_pmu_disable,
609 },
610 .base = base,
611 .dev = dev,
612 };
613
614 pmu->id = ida_alloc(&ddr_ida, GFP_KERNEL);
615 return pmu->id;
616}
617
618static irqreturn_t ddr_perf_irq_handler(int irq, void *p)
619{
620 int i;
621 struct ddr_pmu *pmu = (struct ddr_pmu *) p;
622 struct perf_event *event;
623
624 /* all counter will stop if cycle counter disabled */
625 ddr_perf_counter_enable(pmu,
626 EVENT_CYCLES_ID,
627 EVENT_CYCLES_COUNTER,
628 false);
629 /*
630 * When the cycle counter overflows, all counters are stopped,
631 * and an IRQ is raised. If any other counter overflows, it
632 * continues counting, and no IRQ is raised. But for new SoCs,
633 * such as i.MX8MP, event counter would stop when overflow, so
634 * we need use cycle counter to stop overflow of event counter.
635 *
636 * Cycles occur at least 4 times as often as other events, so we
637 * can update all events on a cycle counter overflow and not
638 * lose events.
639 *
640 */
641 for (i = 0; i < NUM_COUNTERS; i++) {
642
643 if (!pmu->events[i])
644 continue;
645
646 event = pmu->events[i];
647
648 ddr_perf_event_update(event);
649 }
650
651 ddr_perf_counter_enable(pmu,
652 EVENT_CYCLES_ID,
653 EVENT_CYCLES_COUNTER,
654 true);
655
656 return IRQ_HANDLED;
657}
658
659static int ddr_perf_offline_cpu(unsigned int cpu, struct hlist_node *node)
660{
661 struct ddr_pmu *pmu = hlist_entry_safe(node, struct ddr_pmu, node);
662 int target;
663
664 if (cpu != pmu->cpu)
665 return 0;
666
667 target = cpumask_any_but(cpu_online_mask, cpu);
668 if (target >= nr_cpu_ids)
669 return 0;
670
671 perf_pmu_migrate_context(&pmu->pmu, cpu, target);
672 pmu->cpu = target;
673
674 WARN_ON(irq_set_affinity(pmu->irq, cpumask_of(pmu->cpu)));
675
676 return 0;
677}
678
679static int ddr_perf_probe(struct platform_device *pdev)
680{
681 struct ddr_pmu *pmu;
682 struct device_node *np;
683 void __iomem *base;
684 char *name;
685 int num;
686 int ret;
687 int irq;
688
689 base = devm_platform_ioremap_resource(pdev, 0);
690 if (IS_ERR(base))
691 return PTR_ERR(base);
692
693 np = pdev->dev.of_node;
694
695 pmu = devm_kzalloc(&pdev->dev, sizeof(*pmu), GFP_KERNEL);
696 if (!pmu)
697 return -ENOMEM;
698
699 num = ddr_perf_init(pmu, base, &pdev->dev);
700
701 platform_set_drvdata(pdev, pmu);
702
703 name = devm_kasprintf(&pdev->dev, GFP_KERNEL, DDR_PERF_DEV_NAME "%d",
704 num);
705 if (!name) {
706 ret = -ENOMEM;
707 goto cpuhp_state_err;
708 }
709
710 pmu->devtype_data = of_device_get_match_data(&pdev->dev);
711
712 pmu->cpu = raw_smp_processor_id();
713 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
714 DDR_CPUHP_CB_NAME,
715 NULL,
716 ddr_perf_offline_cpu);
717
718 if (ret < 0) {
719 dev_err(&pdev->dev, "cpuhp_setup_state_multi failed\n");
720 goto cpuhp_state_err;
721 }
722
723 pmu->cpuhp_state = ret;
724
725 /* Register the pmu instance for cpu hotplug */
726 ret = cpuhp_state_add_instance_nocalls(pmu->cpuhp_state, &pmu->node);
727 if (ret) {
728 dev_err(&pdev->dev, "Error %d registering hotplug\n", ret);
729 goto cpuhp_instance_err;
730 }
731
732 /* Request irq */
733 irq = of_irq_get(np, 0);
734 if (irq < 0) {
735 dev_err(&pdev->dev, "Failed to get irq: %d", irq);
736 ret = irq;
737 goto ddr_perf_err;
738 }
739
740 ret = devm_request_irq(&pdev->dev, irq,
741 ddr_perf_irq_handler,
742 IRQF_NOBALANCING | IRQF_NO_THREAD,
743 DDR_CPUHP_CB_NAME,
744 pmu);
745 if (ret < 0) {
746 dev_err(&pdev->dev, "Request irq failed: %d", ret);
747 goto ddr_perf_err;
748 }
749
750 pmu->irq = irq;
751 ret = irq_set_affinity(pmu->irq, cpumask_of(pmu->cpu));
752 if (ret) {
753 dev_err(pmu->dev, "Failed to set interrupt affinity!\n");
754 goto ddr_perf_err;
755 }
756
757 ret = perf_pmu_register(&pmu->pmu, name, -1);
758 if (ret)
759 goto ddr_perf_err;
760
761 return 0;
762
763ddr_perf_err:
764 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
765cpuhp_instance_err:
766 cpuhp_remove_multi_state(pmu->cpuhp_state);
767cpuhp_state_err:
768 ida_free(&ddr_ida, pmu->id);
769 dev_warn(&pdev->dev, "i.MX8 DDR Perf PMU failed (%d), disabled\n", ret);
770 return ret;
771}
772
773static int ddr_perf_remove(struct platform_device *pdev)
774{
775 struct ddr_pmu *pmu = platform_get_drvdata(pdev);
776
777 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
778 cpuhp_remove_multi_state(pmu->cpuhp_state);
779
780 perf_pmu_unregister(&pmu->pmu);
781
782 ida_free(&ddr_ida, pmu->id);
783 return 0;
784}
785
786static struct platform_driver imx_ddr_pmu_driver = {
787 .driver = {
788 .name = "imx-ddr-pmu",
789 .of_match_table = imx_ddr_pmu_dt_ids,
790 .suppress_bind_attrs = true,
791 },
792 .probe = ddr_perf_probe,
793 .remove = ddr_perf_remove,
794};
795
796module_platform_driver(imx_ddr_pmu_driver);
797MODULE_LICENSE("GPL v2");
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
45static 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
51struct fsl_ddr_devtype_data {
52 unsigned int quirks; /* quirks needed for different DDR Perf core */
53};
54
55static const struct fsl_ddr_devtype_data imx8_devtype_data;
56
57static const struct fsl_ddr_devtype_data imx8m_devtype_data = {
58 .quirks = DDR_CAP_AXI_ID_FILTER,
59};
60
61static const struct fsl_ddr_devtype_data imx8mp_devtype_data = {
62 .quirks = DDR_CAP_AXI_ID_FILTER_ENHANCED,
63};
64
65static const struct of_device_id imx_ddr_pmu_dt_ids[] = {
66 { .compatible = "fsl,imx8-ddr-pmu", .data = &imx8_devtype_data},
67 { .compatible = "fsl,imx8m-ddr-pmu", .data = &imx8m_devtype_data},
68 { .compatible = "fsl,imx8mp-ddr-pmu", .data = &imx8mp_devtype_data},
69 { /* sentinel */ }
70};
71MODULE_DEVICE_TABLE(of, imx_ddr_pmu_dt_ids);
72
73struct ddr_pmu {
74 struct pmu pmu;
75 void __iomem *base;
76 unsigned int cpu;
77 struct hlist_node node;
78 struct device *dev;
79 struct perf_event *events[NUM_COUNTERS];
80 int active_events;
81 enum cpuhp_state cpuhp_state;
82 const struct fsl_ddr_devtype_data *devtype_data;
83 int irq;
84 int id;
85};
86
87enum ddr_perf_filter_capabilities {
88 PERF_CAP_AXI_ID_FILTER = 0,
89 PERF_CAP_AXI_ID_FILTER_ENHANCED,
90 PERF_CAP_AXI_ID_FEAT_MAX,
91};
92
93static u32 ddr_perf_filter_cap_get(struct ddr_pmu *pmu, int cap)
94{
95 u32 quirks = pmu->devtype_data->quirks;
96
97 switch (cap) {
98 case PERF_CAP_AXI_ID_FILTER:
99 return !!(quirks & DDR_CAP_AXI_ID_FILTER);
100 case PERF_CAP_AXI_ID_FILTER_ENHANCED:
101 quirks &= DDR_CAP_AXI_ID_FILTER_ENHANCED;
102 return quirks == DDR_CAP_AXI_ID_FILTER_ENHANCED;
103 default:
104 WARN(1, "unknown filter cap %d\n", cap);
105 }
106
107 return 0;
108}
109
110static ssize_t ddr_perf_filter_cap_show(struct device *dev,
111 struct device_attribute *attr,
112 char *buf)
113{
114 struct ddr_pmu *pmu = dev_get_drvdata(dev);
115 struct dev_ext_attribute *ea =
116 container_of(attr, struct dev_ext_attribute, attr);
117 int cap = (long)ea->var;
118
119 return snprintf(buf, PAGE_SIZE, "%u\n",
120 ddr_perf_filter_cap_get(pmu, cap));
121}
122
123#define PERF_EXT_ATTR_ENTRY(_name, _func, _var) \
124 (&((struct dev_ext_attribute) { \
125 __ATTR(_name, 0444, _func, NULL), (void *)_var \
126 }).attr.attr)
127
128#define PERF_FILTER_EXT_ATTR_ENTRY(_name, _var) \
129 PERF_EXT_ATTR_ENTRY(_name, ddr_perf_filter_cap_show, _var)
130
131static struct attribute *ddr_perf_filter_cap_attr[] = {
132 PERF_FILTER_EXT_ATTR_ENTRY(filter, PERF_CAP_AXI_ID_FILTER),
133 PERF_FILTER_EXT_ATTR_ENTRY(enhanced_filter, PERF_CAP_AXI_ID_FILTER_ENHANCED),
134 NULL,
135};
136
137static struct attribute_group ddr_perf_filter_cap_attr_group = {
138 .name = "caps",
139 .attrs = ddr_perf_filter_cap_attr,
140};
141
142static ssize_t ddr_perf_cpumask_show(struct device *dev,
143 struct device_attribute *attr, char *buf)
144{
145 struct ddr_pmu *pmu = dev_get_drvdata(dev);
146
147 return cpumap_print_to_pagebuf(true, buf, cpumask_of(pmu->cpu));
148}
149
150static struct device_attribute ddr_perf_cpumask_attr =
151 __ATTR(cpumask, 0444, ddr_perf_cpumask_show, NULL);
152
153static struct attribute *ddr_perf_cpumask_attrs[] = {
154 &ddr_perf_cpumask_attr.attr,
155 NULL,
156};
157
158static struct attribute_group ddr_perf_cpumask_attr_group = {
159 .attrs = ddr_perf_cpumask_attrs,
160};
161
162static ssize_t
163ddr_pmu_event_show(struct device *dev, struct device_attribute *attr,
164 char *page)
165{
166 struct perf_pmu_events_attr *pmu_attr;
167
168 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
169 return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
170}
171
172#define IMX8_DDR_PMU_EVENT_ATTR(_name, _id) \
173 (&((struct perf_pmu_events_attr[]) { \
174 { .attr = __ATTR(_name, 0444, ddr_pmu_event_show, NULL),\
175 .id = _id, } \
176 })[0].attr.attr)
177
178static struct attribute *ddr_perf_events_attrs[] = {
179 IMX8_DDR_PMU_EVENT_ATTR(cycles, EVENT_CYCLES_ID),
180 IMX8_DDR_PMU_EVENT_ATTR(selfresh, 0x01),
181 IMX8_DDR_PMU_EVENT_ATTR(read-accesses, 0x04),
182 IMX8_DDR_PMU_EVENT_ATTR(write-accesses, 0x05),
183 IMX8_DDR_PMU_EVENT_ATTR(read-queue-depth, 0x08),
184 IMX8_DDR_PMU_EVENT_ATTR(write-queue-depth, 0x09),
185 IMX8_DDR_PMU_EVENT_ATTR(lp-read-credit-cnt, 0x10),
186 IMX8_DDR_PMU_EVENT_ATTR(hp-read-credit-cnt, 0x11),
187 IMX8_DDR_PMU_EVENT_ATTR(write-credit-cnt, 0x12),
188 IMX8_DDR_PMU_EVENT_ATTR(read-command, 0x20),
189 IMX8_DDR_PMU_EVENT_ATTR(write-command, 0x21),
190 IMX8_DDR_PMU_EVENT_ATTR(read-modify-write-command, 0x22),
191 IMX8_DDR_PMU_EVENT_ATTR(hp-read, 0x23),
192 IMX8_DDR_PMU_EVENT_ATTR(hp-req-nocredit, 0x24),
193 IMX8_DDR_PMU_EVENT_ATTR(hp-xact-credit, 0x25),
194 IMX8_DDR_PMU_EVENT_ATTR(lp-req-nocredit, 0x26),
195 IMX8_DDR_PMU_EVENT_ATTR(lp-xact-credit, 0x27),
196 IMX8_DDR_PMU_EVENT_ATTR(wr-xact-credit, 0x29),
197 IMX8_DDR_PMU_EVENT_ATTR(read-cycles, 0x2a),
198 IMX8_DDR_PMU_EVENT_ATTR(write-cycles, 0x2b),
199 IMX8_DDR_PMU_EVENT_ATTR(read-write-transition, 0x30),
200 IMX8_DDR_PMU_EVENT_ATTR(precharge, 0x31),
201 IMX8_DDR_PMU_EVENT_ATTR(activate, 0x32),
202 IMX8_DDR_PMU_EVENT_ATTR(load-mode, 0x33),
203 IMX8_DDR_PMU_EVENT_ATTR(perf-mwr, 0x34),
204 IMX8_DDR_PMU_EVENT_ATTR(read, 0x35),
205 IMX8_DDR_PMU_EVENT_ATTR(read-activate, 0x36),
206 IMX8_DDR_PMU_EVENT_ATTR(refresh, 0x37),
207 IMX8_DDR_PMU_EVENT_ATTR(write, 0x38),
208 IMX8_DDR_PMU_EVENT_ATTR(raw-hazard, 0x39),
209 IMX8_DDR_PMU_EVENT_ATTR(axid-read, 0x41),
210 IMX8_DDR_PMU_EVENT_ATTR(axid-write, 0x42),
211 NULL,
212};
213
214static struct attribute_group ddr_perf_events_attr_group = {
215 .name = "events",
216 .attrs = ddr_perf_events_attrs,
217};
218
219PMU_FORMAT_ATTR(event, "config:0-7");
220PMU_FORMAT_ATTR(axi_id, "config1:0-15");
221PMU_FORMAT_ATTR(axi_mask, "config1:16-31");
222
223static struct attribute *ddr_perf_format_attrs[] = {
224 &format_attr_event.attr,
225 &format_attr_axi_id.attr,
226 &format_attr_axi_mask.attr,
227 NULL,
228};
229
230static struct attribute_group ddr_perf_format_attr_group = {
231 .name = "format",
232 .attrs = ddr_perf_format_attrs,
233};
234
235static const struct attribute_group *attr_groups[] = {
236 &ddr_perf_events_attr_group,
237 &ddr_perf_format_attr_group,
238 &ddr_perf_cpumask_attr_group,
239 &ddr_perf_filter_cap_attr_group,
240 NULL,
241};
242
243static bool ddr_perf_is_filtered(struct perf_event *event)
244{
245 return event->attr.config == 0x41 || event->attr.config == 0x42;
246}
247
248static u32 ddr_perf_filter_val(struct perf_event *event)
249{
250 return event->attr.config1;
251}
252
253static bool ddr_perf_filters_compatible(struct perf_event *a,
254 struct perf_event *b)
255{
256 if (!ddr_perf_is_filtered(a))
257 return true;
258 if (!ddr_perf_is_filtered(b))
259 return true;
260 return ddr_perf_filter_val(a) == ddr_perf_filter_val(b);
261}
262
263static bool ddr_perf_is_enhanced_filtered(struct perf_event *event)
264{
265 unsigned int filt;
266 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
267
268 filt = pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER_ENHANCED;
269 return (filt == DDR_CAP_AXI_ID_FILTER_ENHANCED) &&
270 ddr_perf_is_filtered(event);
271}
272
273static u32 ddr_perf_alloc_counter(struct ddr_pmu *pmu, int event)
274{
275 int i;
276
277 /*
278 * Always map cycle event to counter 0
279 * Cycles counter is dedicated for cycle event
280 * can't used for the other events
281 */
282 if (event == EVENT_CYCLES_ID) {
283 if (pmu->events[EVENT_CYCLES_COUNTER] == NULL)
284 return EVENT_CYCLES_COUNTER;
285 else
286 return -ENOENT;
287 }
288
289 for (i = 1; i < NUM_COUNTERS; i++) {
290 if (pmu->events[i] == NULL)
291 return i;
292 }
293
294 return -ENOENT;
295}
296
297static void ddr_perf_free_counter(struct ddr_pmu *pmu, int counter)
298{
299 pmu->events[counter] = NULL;
300}
301
302static u32 ddr_perf_read_counter(struct ddr_pmu *pmu, int counter)
303{
304 struct perf_event *event = pmu->events[counter];
305 void __iomem *base = pmu->base;
306
307 /*
308 * return bytes instead of bursts from ddr transaction for
309 * axid-read and axid-write event if PMU core supports enhanced
310 * filter.
311 */
312 base += ddr_perf_is_enhanced_filtered(event) ? COUNTER_DPCR1 :
313 COUNTER_READ;
314 return readl_relaxed(base + counter * 4);
315}
316
317static int ddr_perf_event_init(struct perf_event *event)
318{
319 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
320 struct hw_perf_event *hwc = &event->hw;
321 struct perf_event *sibling;
322
323 if (event->attr.type != event->pmu->type)
324 return -ENOENT;
325
326 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
327 return -EOPNOTSUPP;
328
329 if (event->cpu < 0) {
330 dev_warn(pmu->dev, "Can't provide per-task data!\n");
331 return -EOPNOTSUPP;
332 }
333
334 /*
335 * We must NOT create groups containing mixed PMUs, although software
336 * events are acceptable (for example to create a CCN group
337 * periodically read when a hrtimer aka cpu-clock leader triggers).
338 */
339 if (event->group_leader->pmu != event->pmu &&
340 !is_software_event(event->group_leader))
341 return -EINVAL;
342
343 if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
344 if (!ddr_perf_filters_compatible(event, event->group_leader))
345 return -EINVAL;
346 for_each_sibling_event(sibling, event->group_leader) {
347 if (!ddr_perf_filters_compatible(event, sibling))
348 return -EINVAL;
349 }
350 }
351
352 for_each_sibling_event(sibling, event->group_leader) {
353 if (sibling->pmu != event->pmu &&
354 !is_software_event(sibling))
355 return -EINVAL;
356 }
357
358 event->cpu = pmu->cpu;
359 hwc->idx = -1;
360
361 return 0;
362}
363
364
365static void ddr_perf_event_update(struct perf_event *event)
366{
367 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
368 struct hw_perf_event *hwc = &event->hw;
369 u64 delta, prev_raw_count, new_raw_count;
370 int counter = hwc->idx;
371
372 do {
373 prev_raw_count = local64_read(&hwc->prev_count);
374 new_raw_count = ddr_perf_read_counter(pmu, counter);
375 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
376 new_raw_count) != prev_raw_count);
377
378 delta = (new_raw_count - prev_raw_count) & 0xFFFFFFFF;
379
380 local64_add(delta, &event->count);
381}
382
383static void ddr_perf_counter_enable(struct ddr_pmu *pmu, int config,
384 int counter, bool enable)
385{
386 u8 reg = counter * 4 + COUNTER_CNTL;
387 int val;
388
389 if (enable) {
390 /*
391 * cycle counter is special which should firstly write 0 then
392 * write 1 into CLEAR bit to clear it. Other counters only
393 * need write 0 into CLEAR bit and it turns out to be 1 by
394 * hardware. Below enable flow is harmless for all counters.
395 */
396 writel(0, pmu->base + reg);
397 val = CNTL_EN | CNTL_CLEAR;
398 val |= FIELD_PREP(CNTL_CSV_MASK, config);
399 writel(val, pmu->base + reg);
400 } else {
401 /* Disable counter */
402 val = readl_relaxed(pmu->base + reg) & CNTL_EN_MASK;
403 writel(val, pmu->base + reg);
404 }
405}
406
407static void ddr_perf_event_start(struct perf_event *event, int flags)
408{
409 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
410 struct hw_perf_event *hwc = &event->hw;
411 int counter = hwc->idx;
412
413 local64_set(&hwc->prev_count, 0);
414
415 ddr_perf_counter_enable(pmu, event->attr.config, counter, true);
416
417 hwc->state = 0;
418}
419
420static int ddr_perf_event_add(struct perf_event *event, int flags)
421{
422 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
423 struct hw_perf_event *hwc = &event->hw;
424 int counter;
425 int cfg = event->attr.config;
426 int cfg1 = event->attr.config1;
427
428 if (pmu->devtype_data->quirks & DDR_CAP_AXI_ID_FILTER) {
429 int i;
430
431 for (i = 1; i < NUM_COUNTERS; i++) {
432 if (pmu->events[i] &&
433 !ddr_perf_filters_compatible(event, pmu->events[i]))
434 return -EINVAL;
435 }
436
437 if (ddr_perf_is_filtered(event)) {
438 /* revert axi id masking(axi_mask) value */
439 cfg1 ^= AXI_MASKING_REVERT;
440 writel(cfg1, pmu->base + COUNTER_DPCR1);
441 }
442 }
443
444 counter = ddr_perf_alloc_counter(pmu, cfg);
445 if (counter < 0) {
446 dev_dbg(pmu->dev, "There are not enough counters\n");
447 return -EOPNOTSUPP;
448 }
449
450 pmu->events[counter] = event;
451 pmu->active_events++;
452 hwc->idx = counter;
453
454 hwc->state |= PERF_HES_STOPPED;
455
456 if (flags & PERF_EF_START)
457 ddr_perf_event_start(event, flags);
458
459 return 0;
460}
461
462static void ddr_perf_event_stop(struct perf_event *event, int flags)
463{
464 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
465 struct hw_perf_event *hwc = &event->hw;
466 int counter = hwc->idx;
467
468 ddr_perf_counter_enable(pmu, event->attr.config, counter, false);
469 ddr_perf_event_update(event);
470
471 hwc->state |= PERF_HES_STOPPED;
472}
473
474static void ddr_perf_event_del(struct perf_event *event, int flags)
475{
476 struct ddr_pmu *pmu = to_ddr_pmu(event->pmu);
477 struct hw_perf_event *hwc = &event->hw;
478 int counter = hwc->idx;
479
480 ddr_perf_event_stop(event, PERF_EF_UPDATE);
481
482 ddr_perf_free_counter(pmu, counter);
483 pmu->active_events--;
484 hwc->idx = -1;
485}
486
487static void ddr_perf_pmu_enable(struct pmu *pmu)
488{
489 struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
490
491 /* enable cycle counter if cycle is not active event list */
492 if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
493 ddr_perf_counter_enable(ddr_pmu,
494 EVENT_CYCLES_ID,
495 EVENT_CYCLES_COUNTER,
496 true);
497}
498
499static void ddr_perf_pmu_disable(struct pmu *pmu)
500{
501 struct ddr_pmu *ddr_pmu = to_ddr_pmu(pmu);
502
503 if (ddr_pmu->events[EVENT_CYCLES_COUNTER] == NULL)
504 ddr_perf_counter_enable(ddr_pmu,
505 EVENT_CYCLES_ID,
506 EVENT_CYCLES_COUNTER,
507 false);
508}
509
510static int ddr_perf_init(struct ddr_pmu *pmu, void __iomem *base,
511 struct device *dev)
512{
513 *pmu = (struct ddr_pmu) {
514 .pmu = (struct pmu) {
515 .module = THIS_MODULE,
516 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
517 .task_ctx_nr = perf_invalid_context,
518 .attr_groups = attr_groups,
519 .event_init = ddr_perf_event_init,
520 .add = ddr_perf_event_add,
521 .del = ddr_perf_event_del,
522 .start = ddr_perf_event_start,
523 .stop = ddr_perf_event_stop,
524 .read = ddr_perf_event_update,
525 .pmu_enable = ddr_perf_pmu_enable,
526 .pmu_disable = ddr_perf_pmu_disable,
527 },
528 .base = base,
529 .dev = dev,
530 };
531
532 pmu->id = ida_simple_get(&ddr_ida, 0, 0, GFP_KERNEL);
533 return pmu->id;
534}
535
536static irqreturn_t ddr_perf_irq_handler(int irq, void *p)
537{
538 int i;
539 struct ddr_pmu *pmu = (struct ddr_pmu *) p;
540 struct perf_event *event, *cycle_event = NULL;
541
542 /* all counter will stop if cycle counter disabled */
543 ddr_perf_counter_enable(pmu,
544 EVENT_CYCLES_ID,
545 EVENT_CYCLES_COUNTER,
546 false);
547 /*
548 * When the cycle counter overflows, all counters are stopped,
549 * and an IRQ is raised. If any other counter overflows, it
550 * continues counting, and no IRQ is raised.
551 *
552 * Cycles occur at least 4 times as often as other events, so we
553 * can update all events on a cycle counter overflow and not
554 * lose events.
555 *
556 */
557 for (i = 0; i < NUM_COUNTERS; i++) {
558
559 if (!pmu->events[i])
560 continue;
561
562 event = pmu->events[i];
563
564 ddr_perf_event_update(event);
565
566 if (event->hw.idx == EVENT_CYCLES_COUNTER)
567 cycle_event = event;
568 }
569
570 ddr_perf_counter_enable(pmu,
571 EVENT_CYCLES_ID,
572 EVENT_CYCLES_COUNTER,
573 true);
574 if (cycle_event)
575 ddr_perf_event_update(cycle_event);
576
577 return IRQ_HANDLED;
578}
579
580static int ddr_perf_offline_cpu(unsigned int cpu, struct hlist_node *node)
581{
582 struct ddr_pmu *pmu = hlist_entry_safe(node, struct ddr_pmu, node);
583 int target;
584
585 if (cpu != pmu->cpu)
586 return 0;
587
588 target = cpumask_any_but(cpu_online_mask, cpu);
589 if (target >= nr_cpu_ids)
590 return 0;
591
592 perf_pmu_migrate_context(&pmu->pmu, cpu, target);
593 pmu->cpu = target;
594
595 WARN_ON(irq_set_affinity_hint(pmu->irq, cpumask_of(pmu->cpu)));
596
597 return 0;
598}
599
600static int ddr_perf_probe(struct platform_device *pdev)
601{
602 struct ddr_pmu *pmu;
603 struct device_node *np;
604 void __iomem *base;
605 char *name;
606 int num;
607 int ret;
608 int irq;
609
610 base = devm_platform_ioremap_resource(pdev, 0);
611 if (IS_ERR(base))
612 return PTR_ERR(base);
613
614 np = pdev->dev.of_node;
615
616 pmu = devm_kzalloc(&pdev->dev, sizeof(*pmu), GFP_KERNEL);
617 if (!pmu)
618 return -ENOMEM;
619
620 num = ddr_perf_init(pmu, base, &pdev->dev);
621
622 platform_set_drvdata(pdev, pmu);
623
624 name = devm_kasprintf(&pdev->dev, GFP_KERNEL, DDR_PERF_DEV_NAME "%d",
625 num);
626 if (!name)
627 return -ENOMEM;
628
629 pmu->devtype_data = of_device_get_match_data(&pdev->dev);
630
631 pmu->cpu = raw_smp_processor_id();
632 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
633 DDR_CPUHP_CB_NAME,
634 NULL,
635 ddr_perf_offline_cpu);
636
637 if (ret < 0) {
638 dev_err(&pdev->dev, "cpuhp_setup_state_multi failed\n");
639 goto cpuhp_state_err;
640 }
641
642 pmu->cpuhp_state = ret;
643
644 /* Register the pmu instance for cpu hotplug */
645 ret = cpuhp_state_add_instance_nocalls(pmu->cpuhp_state, &pmu->node);
646 if (ret) {
647 dev_err(&pdev->dev, "Error %d registering hotplug\n", ret);
648 goto cpuhp_instance_err;
649 }
650
651 /* Request irq */
652 irq = of_irq_get(np, 0);
653 if (irq < 0) {
654 dev_err(&pdev->dev, "Failed to get irq: %d", irq);
655 ret = irq;
656 goto ddr_perf_err;
657 }
658
659 ret = devm_request_irq(&pdev->dev, irq,
660 ddr_perf_irq_handler,
661 IRQF_NOBALANCING | IRQF_NO_THREAD,
662 DDR_CPUHP_CB_NAME,
663 pmu);
664 if (ret < 0) {
665 dev_err(&pdev->dev, "Request irq failed: %d", ret);
666 goto ddr_perf_err;
667 }
668
669 pmu->irq = irq;
670 ret = irq_set_affinity_hint(pmu->irq, cpumask_of(pmu->cpu));
671 if (ret) {
672 dev_err(pmu->dev, "Failed to set interrupt affinity!\n");
673 goto ddr_perf_err;
674 }
675
676 ret = perf_pmu_register(&pmu->pmu, name, -1);
677 if (ret)
678 goto ddr_perf_err;
679
680 return 0;
681
682ddr_perf_err:
683 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
684cpuhp_instance_err:
685 cpuhp_remove_multi_state(pmu->cpuhp_state);
686cpuhp_state_err:
687 ida_simple_remove(&ddr_ida, pmu->id);
688 dev_warn(&pdev->dev, "i.MX8 DDR Perf PMU failed (%d), disabled\n", ret);
689 return ret;
690}
691
692static int ddr_perf_remove(struct platform_device *pdev)
693{
694 struct ddr_pmu *pmu = platform_get_drvdata(pdev);
695
696 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
697 cpuhp_remove_multi_state(pmu->cpuhp_state);
698 irq_set_affinity_hint(pmu->irq, NULL);
699
700 perf_pmu_unregister(&pmu->pmu);
701
702 ida_simple_remove(&ddr_ida, pmu->id);
703 return 0;
704}
705
706static struct platform_driver imx_ddr_pmu_driver = {
707 .driver = {
708 .name = "imx-ddr-pmu",
709 .of_match_table = imx_ddr_pmu_dt_ids,
710 .suppress_bind_attrs = true,
711 },
712 .probe = ddr_perf_probe,
713 .remove = ddr_perf_remove,
714};
715
716module_platform_driver(imx_ddr_pmu_driver);
717MODULE_LICENSE("GPL v2");