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
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 };
54
55 static const struct fsl_ddr_devtype_data imx8_devtype_data;
56
57 static const struct fsl_ddr_devtype_data imx8m_devtype_data = {
58 .quirks = DDR_CAP_AXI_ID_FILTER,
59 };
60
61 static const struct fsl_ddr_devtype_data imx8mp_devtype_data = {
62 .quirks = DDR_CAP_AXI_ID_FILTER_ENHANCED,
63 };
64
65 static 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 };
71 MODULE_DEVICE_TABLE(of, imx_ddr_pmu_dt_ids);
72
73 struct 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
87 enum 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
ddr_perf_filter_cap_get(struct ddr_pmu * pmu,int cap)93 static 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
ddr_perf_filter_cap_show(struct device * dev,struct device_attribute * attr,char * buf)110 static 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
131 static 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
137 static struct attribute_group ddr_perf_filter_cap_attr_group = {
138 .name = "caps",
139 .attrs = ddr_perf_filter_cap_attr,
140 };
141
ddr_perf_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)142 static 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
150 static struct device_attribute ddr_perf_cpumask_attr =
151 __ATTR(cpumask, 0444, ddr_perf_cpumask_show, NULL);
152
153 static struct attribute *ddr_perf_cpumask_attrs[] = {
154 &ddr_perf_cpumask_attr.attr,
155 NULL,
156 };
157
158 static struct attribute_group ddr_perf_cpumask_attr_group = {
159 .attrs = ddr_perf_cpumask_attrs,
160 };
161
162 static ssize_t
ddr_pmu_event_show(struct device * dev,struct device_attribute * attr,char * page)163 ddr_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
178 static 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
214 static struct attribute_group ddr_perf_events_attr_group = {
215 .name = "events",
216 .attrs = ddr_perf_events_attrs,
217 };
218
219 PMU_FORMAT_ATTR(event, "config:0-7");
220 PMU_FORMAT_ATTR(axi_id, "config1:0-15");
221 PMU_FORMAT_ATTR(axi_mask, "config1:16-31");
222
223 static 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
230 static struct attribute_group ddr_perf_format_attr_group = {
231 .name = "format",
232 .attrs = ddr_perf_format_attrs,
233 };
234
235 static 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
ddr_perf_is_filtered(struct perf_event * event)243 static bool ddr_perf_is_filtered(struct perf_event *event)
244 {
245 return event->attr.config == 0x41 || event->attr.config == 0x42;
246 }
247
ddr_perf_filter_val(struct perf_event * event)248 static u32 ddr_perf_filter_val(struct perf_event *event)
249 {
250 return event->attr.config1;
251 }
252
ddr_perf_filters_compatible(struct perf_event * a,struct perf_event * b)253 static 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
ddr_perf_is_enhanced_filtered(struct perf_event * event)263 static 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
ddr_perf_alloc_counter(struct ddr_pmu * pmu,int event)273 static 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
ddr_perf_free_counter(struct ddr_pmu * pmu,int counter)297 static void ddr_perf_free_counter(struct ddr_pmu *pmu, int counter)
298 {
299 pmu->events[counter] = NULL;
300 }
301
ddr_perf_read_counter(struct ddr_pmu * pmu,int counter)302 static 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
ddr_perf_event_init(struct perf_event * event)317 static 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
ddr_perf_event_update(struct perf_event * event)365 static 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
ddr_perf_counter_enable(struct ddr_pmu * pmu,int config,int counter,bool enable)383 static 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
ddr_perf_event_start(struct perf_event * event,int flags)407 static 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
ddr_perf_event_add(struct perf_event * event,int flags)420 static 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
ddr_perf_event_stop(struct perf_event * event,int flags)462 static 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
ddr_perf_event_del(struct perf_event * event,int flags)474 static 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
ddr_perf_pmu_enable(struct pmu * pmu)487 static 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
ddr_perf_pmu_disable(struct pmu * pmu)499 static 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
ddr_perf_init(struct ddr_pmu * pmu,void __iomem * base,struct device * dev)510 static 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
ddr_perf_irq_handler(int irq,void * p)536 static 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
ddr_perf_offline_cpu(unsigned int cpu,struct hlist_node * node)580 static 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
ddr_perf_probe(struct platform_device * pdev)600 static 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 ret = -ENOMEM;
628 goto cpuhp_state_err;
629 }
630
631 pmu->devtype_data = of_device_get_match_data(&pdev->dev);
632
633 pmu->cpu = raw_smp_processor_id();
634 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
635 DDR_CPUHP_CB_NAME,
636 NULL,
637 ddr_perf_offline_cpu);
638
639 if (ret < 0) {
640 dev_err(&pdev->dev, "cpuhp_setup_state_multi failed\n");
641 goto cpuhp_state_err;
642 }
643
644 pmu->cpuhp_state = ret;
645
646 /* Register the pmu instance for cpu hotplug */
647 ret = cpuhp_state_add_instance_nocalls(pmu->cpuhp_state, &pmu->node);
648 if (ret) {
649 dev_err(&pdev->dev, "Error %d registering hotplug\n", ret);
650 goto cpuhp_instance_err;
651 }
652
653 /* Request irq */
654 irq = of_irq_get(np, 0);
655 if (irq < 0) {
656 dev_err(&pdev->dev, "Failed to get irq: %d", irq);
657 ret = irq;
658 goto ddr_perf_err;
659 }
660
661 ret = devm_request_irq(&pdev->dev, irq,
662 ddr_perf_irq_handler,
663 IRQF_NOBALANCING | IRQF_NO_THREAD,
664 DDR_CPUHP_CB_NAME,
665 pmu);
666 if (ret < 0) {
667 dev_err(&pdev->dev, "Request irq failed: %d", ret);
668 goto ddr_perf_err;
669 }
670
671 pmu->irq = irq;
672 ret = irq_set_affinity_hint(pmu->irq, cpumask_of(pmu->cpu));
673 if (ret) {
674 dev_err(pmu->dev, "Failed to set interrupt affinity!\n");
675 goto ddr_perf_err;
676 }
677
678 ret = perf_pmu_register(&pmu->pmu, name, -1);
679 if (ret)
680 goto ddr_perf_err;
681
682 return 0;
683
684 ddr_perf_err:
685 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
686 cpuhp_instance_err:
687 cpuhp_remove_multi_state(pmu->cpuhp_state);
688 cpuhp_state_err:
689 ida_simple_remove(&ddr_ida, pmu->id);
690 dev_warn(&pdev->dev, "i.MX8 DDR Perf PMU failed (%d), disabled\n", ret);
691 return ret;
692 }
693
ddr_perf_remove(struct platform_device * pdev)694 static int ddr_perf_remove(struct platform_device *pdev)
695 {
696 struct ddr_pmu *pmu = platform_get_drvdata(pdev);
697
698 cpuhp_state_remove_instance_nocalls(pmu->cpuhp_state, &pmu->node);
699 cpuhp_remove_multi_state(pmu->cpuhp_state);
700 irq_set_affinity_hint(pmu->irq, NULL);
701
702 perf_pmu_unregister(&pmu->pmu);
703
704 ida_simple_remove(&ddr_ida, pmu->id);
705 return 0;
706 }
707
708 static struct platform_driver imx_ddr_pmu_driver = {
709 .driver = {
710 .name = "imx-ddr-pmu",
711 .of_match_table = imx_ddr_pmu_dt_ids,
712 .suppress_bind_attrs = true,
713 },
714 .probe = ddr_perf_probe,
715 .remove = ddr_perf_remove,
716 };
717
718 module_platform_driver(imx_ddr_pmu_driver);
719 MODULE_LICENSE("GPL v2");
720