1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * CAVIUM THUNDERX2 SoC PMU UNCORE
4 * Copyright (C) 2018 Cavium Inc.
5 * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
6 */
7
8 #include <linux/acpi.h>
9 #include <linux/cpuhotplug.h>
10 #include <linux/perf_event.h>
11 #include <linux/platform_device.h>
12
13 /* Each ThunderX2(TX2) Socket has a L3C and DMC UNCORE PMU device.
14 * Each UNCORE PMU device consists of 4 independent programmable counters.
15 * Counters are 32 bit and do not support overflow interrupt,
16 * they need to be sampled before overflow(i.e, at every 2 seconds).
17 */
18
19 #define TX2_PMU_DMC_L3C_MAX_COUNTERS 4
20 #define TX2_PMU_CCPI2_MAX_COUNTERS 8
21 #define TX2_PMU_MAX_COUNTERS TX2_PMU_CCPI2_MAX_COUNTERS
22
23
24 #define TX2_PMU_DMC_CHANNELS 8
25 #define TX2_PMU_L3_TILES 16
26
27 #define TX2_PMU_HRTIMER_INTERVAL (2 * NSEC_PER_SEC)
28 #define GET_EVENTID(ev, mask) ((ev->hw.config) & mask)
29 #define GET_COUNTERID(ev, mask) ((ev->hw.idx) & mask)
30 /* 1 byte per counter(4 counters).
31 * Event id is encoded in bits [5:1] of a byte,
32 */
33 #define DMC_EVENT_CFG(idx, val) ((val) << (((idx) * 8) + 1))
34
35 /* bits[3:0] to select counters, are indexed from 8 to 15. */
36 #define CCPI2_COUNTER_OFFSET 8
37
38 #define L3C_COUNTER_CTL 0xA8
39 #define L3C_COUNTER_DATA 0xAC
40 #define DMC_COUNTER_CTL 0x234
41 #define DMC_COUNTER_DATA 0x240
42
43 #define CCPI2_PERF_CTL 0x108
44 #define CCPI2_COUNTER_CTL 0x10C
45 #define CCPI2_COUNTER_SEL 0x12c
46 #define CCPI2_COUNTER_DATA_L 0x130
47 #define CCPI2_COUNTER_DATA_H 0x134
48
49 /* L3C event IDs */
50 #define L3_EVENT_READ_REQ 0xD
51 #define L3_EVENT_WRITEBACK_REQ 0xE
52 #define L3_EVENT_INV_N_WRITE_REQ 0xF
53 #define L3_EVENT_INV_REQ 0x10
54 #define L3_EVENT_EVICT_REQ 0x13
55 #define L3_EVENT_INV_N_WRITE_HIT 0x14
56 #define L3_EVENT_INV_HIT 0x15
57 #define L3_EVENT_READ_HIT 0x17
58 #define L3_EVENT_MAX 0x18
59
60 /* DMC event IDs */
61 #define DMC_EVENT_COUNT_CYCLES 0x1
62 #define DMC_EVENT_WRITE_TXNS 0xB
63 #define DMC_EVENT_DATA_TRANSFERS 0xD
64 #define DMC_EVENT_READ_TXNS 0xF
65 #define DMC_EVENT_MAX 0x10
66
67 #define CCPI2_EVENT_REQ_PKT_SENT 0x3D
68 #define CCPI2_EVENT_SNOOP_PKT_SENT 0x65
69 #define CCPI2_EVENT_DATA_PKT_SENT 0x105
70 #define CCPI2_EVENT_GIC_PKT_SENT 0x12D
71 #define CCPI2_EVENT_MAX 0x200
72
73 #define CCPI2_PERF_CTL_ENABLE BIT(0)
74 #define CCPI2_PERF_CTL_START BIT(1)
75 #define CCPI2_PERF_CTL_RESET BIT(4)
76 #define CCPI2_EVENT_LEVEL_RISING_EDGE BIT(10)
77 #define CCPI2_EVENT_TYPE_EDGE_SENSITIVE BIT(11)
78
79 enum tx2_uncore_type {
80 PMU_TYPE_L3C,
81 PMU_TYPE_DMC,
82 PMU_TYPE_CCPI2,
83 PMU_TYPE_INVALID,
84 };
85
86 /*
87 * Each socket has 3 uncore devices associated with a PMU. The DMC and
88 * L3C have 4 32-bit counters and the CCPI2 has 8 64-bit counters.
89 */
90 struct tx2_uncore_pmu {
91 struct hlist_node hpnode;
92 struct list_head entry;
93 struct pmu pmu;
94 char *name;
95 int node;
96 int cpu;
97 u32 max_counters;
98 u32 counters_mask;
99 u32 prorate_factor;
100 u32 max_events;
101 u32 events_mask;
102 u64 hrtimer_interval;
103 void __iomem *base;
104 DECLARE_BITMAP(active_counters, TX2_PMU_MAX_COUNTERS);
105 struct perf_event *events[TX2_PMU_MAX_COUNTERS];
106 struct device *dev;
107 struct hrtimer hrtimer;
108 const struct attribute_group **attr_groups;
109 enum tx2_uncore_type type;
110 enum hrtimer_restart (*hrtimer_callback)(struct hrtimer *cb);
111 void (*init_cntr_base)(struct perf_event *event,
112 struct tx2_uncore_pmu *tx2_pmu);
113 void (*stop_event)(struct perf_event *event);
114 void (*start_event)(struct perf_event *event, int flags);
115 };
116
117 static LIST_HEAD(tx2_pmus);
118
pmu_to_tx2_pmu(struct pmu * pmu)119 static inline struct tx2_uncore_pmu *pmu_to_tx2_pmu(struct pmu *pmu)
120 {
121 return container_of(pmu, struct tx2_uncore_pmu, pmu);
122 }
123
124 #define TX2_PMU_FORMAT_ATTR(_var, _name, _format) \
125 static ssize_t \
126 __tx2_pmu_##_var##_show(struct device *dev, \
127 struct device_attribute *attr, \
128 char *page) \
129 { \
130 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
131 return sysfs_emit(page, _format "\n"); \
132 } \
133 \
134 static struct device_attribute format_attr_##_var = \
135 __ATTR(_name, 0444, __tx2_pmu_##_var##_show, NULL)
136
137 TX2_PMU_FORMAT_ATTR(event, event, "config:0-4");
138 TX2_PMU_FORMAT_ATTR(event_ccpi2, event, "config:0-9");
139
140 static struct attribute *l3c_pmu_format_attrs[] = {
141 &format_attr_event.attr,
142 NULL,
143 };
144
145 static struct attribute *dmc_pmu_format_attrs[] = {
146 &format_attr_event.attr,
147 NULL,
148 };
149
150 static struct attribute *ccpi2_pmu_format_attrs[] = {
151 &format_attr_event_ccpi2.attr,
152 NULL,
153 };
154
155 static const struct attribute_group l3c_pmu_format_attr_group = {
156 .name = "format",
157 .attrs = l3c_pmu_format_attrs,
158 };
159
160 static const struct attribute_group dmc_pmu_format_attr_group = {
161 .name = "format",
162 .attrs = dmc_pmu_format_attrs,
163 };
164
165 static const struct attribute_group ccpi2_pmu_format_attr_group = {
166 .name = "format",
167 .attrs = ccpi2_pmu_format_attrs,
168 };
169
170 /*
171 * sysfs event attributes
172 */
tx2_pmu_event_show(struct device * dev,struct device_attribute * attr,char * buf)173 static ssize_t tx2_pmu_event_show(struct device *dev,
174 struct device_attribute *attr, char *buf)
175 {
176 struct dev_ext_attribute *eattr;
177
178 eattr = container_of(attr, struct dev_ext_attribute, attr);
179 return sysfs_emit(buf, "event=0x%lx\n", (unsigned long) eattr->var);
180 }
181
182 #define TX2_EVENT_ATTR(name, config) \
183 PMU_EVENT_ATTR(name, tx2_pmu_event_attr_##name, \
184 config, tx2_pmu_event_show)
185
186 TX2_EVENT_ATTR(read_request, L3_EVENT_READ_REQ);
187 TX2_EVENT_ATTR(writeback_request, L3_EVENT_WRITEBACK_REQ);
188 TX2_EVENT_ATTR(inv_nwrite_request, L3_EVENT_INV_N_WRITE_REQ);
189 TX2_EVENT_ATTR(inv_request, L3_EVENT_INV_REQ);
190 TX2_EVENT_ATTR(evict_request, L3_EVENT_EVICT_REQ);
191 TX2_EVENT_ATTR(inv_nwrite_hit, L3_EVENT_INV_N_WRITE_HIT);
192 TX2_EVENT_ATTR(inv_hit, L3_EVENT_INV_HIT);
193 TX2_EVENT_ATTR(read_hit, L3_EVENT_READ_HIT);
194
195 static struct attribute *l3c_pmu_events_attrs[] = {
196 &tx2_pmu_event_attr_read_request.attr.attr,
197 &tx2_pmu_event_attr_writeback_request.attr.attr,
198 &tx2_pmu_event_attr_inv_nwrite_request.attr.attr,
199 &tx2_pmu_event_attr_inv_request.attr.attr,
200 &tx2_pmu_event_attr_evict_request.attr.attr,
201 &tx2_pmu_event_attr_inv_nwrite_hit.attr.attr,
202 &tx2_pmu_event_attr_inv_hit.attr.attr,
203 &tx2_pmu_event_attr_read_hit.attr.attr,
204 NULL,
205 };
206
207 TX2_EVENT_ATTR(cnt_cycles, DMC_EVENT_COUNT_CYCLES);
208 TX2_EVENT_ATTR(write_txns, DMC_EVENT_WRITE_TXNS);
209 TX2_EVENT_ATTR(data_transfers, DMC_EVENT_DATA_TRANSFERS);
210 TX2_EVENT_ATTR(read_txns, DMC_EVENT_READ_TXNS);
211
212 static struct attribute *dmc_pmu_events_attrs[] = {
213 &tx2_pmu_event_attr_cnt_cycles.attr.attr,
214 &tx2_pmu_event_attr_write_txns.attr.attr,
215 &tx2_pmu_event_attr_data_transfers.attr.attr,
216 &tx2_pmu_event_attr_read_txns.attr.attr,
217 NULL,
218 };
219
220 TX2_EVENT_ATTR(req_pktsent, CCPI2_EVENT_REQ_PKT_SENT);
221 TX2_EVENT_ATTR(snoop_pktsent, CCPI2_EVENT_SNOOP_PKT_SENT);
222 TX2_EVENT_ATTR(data_pktsent, CCPI2_EVENT_DATA_PKT_SENT);
223 TX2_EVENT_ATTR(gic_pktsent, CCPI2_EVENT_GIC_PKT_SENT);
224
225 static struct attribute *ccpi2_pmu_events_attrs[] = {
226 &tx2_pmu_event_attr_req_pktsent.attr.attr,
227 &tx2_pmu_event_attr_snoop_pktsent.attr.attr,
228 &tx2_pmu_event_attr_data_pktsent.attr.attr,
229 &tx2_pmu_event_attr_gic_pktsent.attr.attr,
230 NULL,
231 };
232
233 static const struct attribute_group l3c_pmu_events_attr_group = {
234 .name = "events",
235 .attrs = l3c_pmu_events_attrs,
236 };
237
238 static const struct attribute_group dmc_pmu_events_attr_group = {
239 .name = "events",
240 .attrs = dmc_pmu_events_attrs,
241 };
242
243 static const struct attribute_group ccpi2_pmu_events_attr_group = {
244 .name = "events",
245 .attrs = ccpi2_pmu_events_attrs,
246 };
247
248 /*
249 * sysfs cpumask attributes
250 */
cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)251 static ssize_t cpumask_show(struct device *dev, struct device_attribute *attr,
252 char *buf)
253 {
254 struct tx2_uncore_pmu *tx2_pmu;
255
256 tx2_pmu = pmu_to_tx2_pmu(dev_get_drvdata(dev));
257 return cpumap_print_to_pagebuf(true, buf, cpumask_of(tx2_pmu->cpu));
258 }
259 static DEVICE_ATTR_RO(cpumask);
260
261 static struct attribute *tx2_pmu_cpumask_attrs[] = {
262 &dev_attr_cpumask.attr,
263 NULL,
264 };
265
266 static const struct attribute_group pmu_cpumask_attr_group = {
267 .attrs = tx2_pmu_cpumask_attrs,
268 };
269
270 /*
271 * Per PMU device attribute groups
272 */
273 static const struct attribute_group *l3c_pmu_attr_groups[] = {
274 &l3c_pmu_format_attr_group,
275 &pmu_cpumask_attr_group,
276 &l3c_pmu_events_attr_group,
277 NULL
278 };
279
280 static const struct attribute_group *dmc_pmu_attr_groups[] = {
281 &dmc_pmu_format_attr_group,
282 &pmu_cpumask_attr_group,
283 &dmc_pmu_events_attr_group,
284 NULL
285 };
286
287 static const struct attribute_group *ccpi2_pmu_attr_groups[] = {
288 &ccpi2_pmu_format_attr_group,
289 &pmu_cpumask_attr_group,
290 &ccpi2_pmu_events_attr_group,
291 NULL
292 };
293
reg_readl(unsigned long addr)294 static inline u32 reg_readl(unsigned long addr)
295 {
296 return readl((void __iomem *)addr);
297 }
298
reg_writel(u32 val,unsigned long addr)299 static inline void reg_writel(u32 val, unsigned long addr)
300 {
301 writel(val, (void __iomem *)addr);
302 }
303
alloc_counter(struct tx2_uncore_pmu * tx2_pmu)304 static int alloc_counter(struct tx2_uncore_pmu *tx2_pmu)
305 {
306 int counter;
307
308 counter = find_first_zero_bit(tx2_pmu->active_counters,
309 tx2_pmu->max_counters);
310 if (counter == tx2_pmu->max_counters)
311 return -ENOSPC;
312
313 set_bit(counter, tx2_pmu->active_counters);
314 return counter;
315 }
316
free_counter(struct tx2_uncore_pmu * tx2_pmu,int counter)317 static inline void free_counter(struct tx2_uncore_pmu *tx2_pmu, int counter)
318 {
319 clear_bit(counter, tx2_pmu->active_counters);
320 }
321
init_cntr_base_l3c(struct perf_event * event,struct tx2_uncore_pmu * tx2_pmu)322 static void init_cntr_base_l3c(struct perf_event *event,
323 struct tx2_uncore_pmu *tx2_pmu)
324 {
325 struct hw_perf_event *hwc = &event->hw;
326 u32 cmask;
327
328 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
329 cmask = tx2_pmu->counters_mask;
330
331 /* counter ctrl/data reg offset at 8 */
332 hwc->config_base = (unsigned long)tx2_pmu->base
333 + L3C_COUNTER_CTL + (8 * GET_COUNTERID(event, cmask));
334 hwc->event_base = (unsigned long)tx2_pmu->base
335 + L3C_COUNTER_DATA + (8 * GET_COUNTERID(event, cmask));
336 }
337
init_cntr_base_dmc(struct perf_event * event,struct tx2_uncore_pmu * tx2_pmu)338 static void init_cntr_base_dmc(struct perf_event *event,
339 struct tx2_uncore_pmu *tx2_pmu)
340 {
341 struct hw_perf_event *hwc = &event->hw;
342 u32 cmask;
343
344 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
345 cmask = tx2_pmu->counters_mask;
346
347 hwc->config_base = (unsigned long)tx2_pmu->base
348 + DMC_COUNTER_CTL;
349 /* counter data reg offset at 0xc */
350 hwc->event_base = (unsigned long)tx2_pmu->base
351 + DMC_COUNTER_DATA + (0xc * GET_COUNTERID(event, cmask));
352 }
353
init_cntr_base_ccpi2(struct perf_event * event,struct tx2_uncore_pmu * tx2_pmu)354 static void init_cntr_base_ccpi2(struct perf_event *event,
355 struct tx2_uncore_pmu *tx2_pmu)
356 {
357 struct hw_perf_event *hwc = &event->hw;
358 u32 cmask;
359
360 cmask = tx2_pmu->counters_mask;
361
362 hwc->config_base = (unsigned long)tx2_pmu->base
363 + CCPI2_COUNTER_CTL + (4 * GET_COUNTERID(event, cmask));
364 hwc->event_base = (unsigned long)tx2_pmu->base;
365 }
366
uncore_start_event_l3c(struct perf_event * event,int flags)367 static void uncore_start_event_l3c(struct perf_event *event, int flags)
368 {
369 u32 val, emask;
370 struct hw_perf_event *hwc = &event->hw;
371 struct tx2_uncore_pmu *tx2_pmu;
372
373 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
374 emask = tx2_pmu->events_mask;
375
376 /* event id encoded in bits [07:03] */
377 val = GET_EVENTID(event, emask) << 3;
378 reg_writel(val, hwc->config_base);
379 local64_set(&hwc->prev_count, 0);
380 reg_writel(0, hwc->event_base);
381 }
382
uncore_stop_event_l3c(struct perf_event * event)383 static inline void uncore_stop_event_l3c(struct perf_event *event)
384 {
385 reg_writel(0, event->hw.config_base);
386 }
387
uncore_start_event_dmc(struct perf_event * event,int flags)388 static void uncore_start_event_dmc(struct perf_event *event, int flags)
389 {
390 u32 val, cmask, emask;
391 struct hw_perf_event *hwc = &event->hw;
392 struct tx2_uncore_pmu *tx2_pmu;
393 int idx, event_id;
394
395 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
396 cmask = tx2_pmu->counters_mask;
397 emask = tx2_pmu->events_mask;
398
399 idx = GET_COUNTERID(event, cmask);
400 event_id = GET_EVENTID(event, emask);
401
402 /* enable and start counters.
403 * 8 bits for each counter, bits[05:01] of a counter to set event type.
404 */
405 val = reg_readl(hwc->config_base);
406 val &= ~DMC_EVENT_CFG(idx, 0x1f);
407 val |= DMC_EVENT_CFG(idx, event_id);
408 reg_writel(val, hwc->config_base);
409 local64_set(&hwc->prev_count, 0);
410 reg_writel(0, hwc->event_base);
411 }
412
uncore_stop_event_dmc(struct perf_event * event)413 static void uncore_stop_event_dmc(struct perf_event *event)
414 {
415 u32 val, cmask;
416 struct hw_perf_event *hwc = &event->hw;
417 struct tx2_uncore_pmu *tx2_pmu;
418 int idx;
419
420 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
421 cmask = tx2_pmu->counters_mask;
422 idx = GET_COUNTERID(event, cmask);
423
424 /* clear event type(bits[05:01]) to stop counter */
425 val = reg_readl(hwc->config_base);
426 val &= ~DMC_EVENT_CFG(idx, 0x1f);
427 reg_writel(val, hwc->config_base);
428 }
429
uncore_start_event_ccpi2(struct perf_event * event,int flags)430 static void uncore_start_event_ccpi2(struct perf_event *event, int flags)
431 {
432 u32 emask;
433 struct hw_perf_event *hwc = &event->hw;
434 struct tx2_uncore_pmu *tx2_pmu;
435
436 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
437 emask = tx2_pmu->events_mask;
438
439 /* Bit [09:00] to set event id.
440 * Bits [10], set level to rising edge.
441 * Bits [11], set type to edge sensitive.
442 */
443 reg_writel((CCPI2_EVENT_TYPE_EDGE_SENSITIVE |
444 CCPI2_EVENT_LEVEL_RISING_EDGE |
445 GET_EVENTID(event, emask)), hwc->config_base);
446
447 /* reset[4], enable[0] and start[1] counters */
448 reg_writel(CCPI2_PERF_CTL_RESET |
449 CCPI2_PERF_CTL_START |
450 CCPI2_PERF_CTL_ENABLE,
451 hwc->event_base + CCPI2_PERF_CTL);
452 local64_set(&event->hw.prev_count, 0ULL);
453 }
454
uncore_stop_event_ccpi2(struct perf_event * event)455 static void uncore_stop_event_ccpi2(struct perf_event *event)
456 {
457 struct hw_perf_event *hwc = &event->hw;
458
459 /* disable and stop counter */
460 reg_writel(0, hwc->event_base + CCPI2_PERF_CTL);
461 }
462
tx2_uncore_event_update(struct perf_event * event)463 static void tx2_uncore_event_update(struct perf_event *event)
464 {
465 u64 prev, delta, new = 0;
466 struct hw_perf_event *hwc = &event->hw;
467 struct tx2_uncore_pmu *tx2_pmu;
468 enum tx2_uncore_type type;
469 u32 prorate_factor;
470 u32 cmask, emask;
471
472 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
473 type = tx2_pmu->type;
474 cmask = tx2_pmu->counters_mask;
475 emask = tx2_pmu->events_mask;
476 prorate_factor = tx2_pmu->prorate_factor;
477 if (type == PMU_TYPE_CCPI2) {
478 reg_writel(CCPI2_COUNTER_OFFSET +
479 GET_COUNTERID(event, cmask),
480 hwc->event_base + CCPI2_COUNTER_SEL);
481 new = reg_readl(hwc->event_base + CCPI2_COUNTER_DATA_H);
482 new = (new << 32) +
483 reg_readl(hwc->event_base + CCPI2_COUNTER_DATA_L);
484 prev = local64_xchg(&hwc->prev_count, new);
485 delta = new - prev;
486 } else {
487 new = reg_readl(hwc->event_base);
488 prev = local64_xchg(&hwc->prev_count, new);
489 /* handles rollover of 32 bit counter */
490 delta = (u32)(((1ULL << 32) - prev) + new);
491 }
492
493 /* DMC event data_transfers granularity is 16 Bytes, convert it to 64 */
494 if (type == PMU_TYPE_DMC &&
495 GET_EVENTID(event, emask) == DMC_EVENT_DATA_TRANSFERS)
496 delta = delta/4;
497
498 /* L3C and DMC has 16 and 8 interleave channels respectively.
499 * The sampled value is for channel 0 and multiplied with
500 * prorate_factor to get the count for a device.
501 */
502 local64_add(delta * prorate_factor, &event->count);
503 }
504
get_tx2_pmu_type(struct acpi_device * adev)505 static enum tx2_uncore_type get_tx2_pmu_type(struct acpi_device *adev)
506 {
507 struct acpi_device_id devices[] = {
508 {"CAV901D", PMU_TYPE_L3C},
509 {"CAV901F", PMU_TYPE_DMC},
510 {"CAV901E", PMU_TYPE_CCPI2},
511 {}
512 };
513 const struct acpi_device_id *id;
514
515 id = acpi_match_acpi_device(devices, adev);
516 if (!id)
517 return PMU_TYPE_INVALID;
518
519 return (enum tx2_uncore_type)id->driver_data;
520 }
521
tx2_uncore_validate_event(struct pmu * pmu,struct perf_event * event,int * counters)522 static bool tx2_uncore_validate_event(struct pmu *pmu,
523 struct perf_event *event, int *counters)
524 {
525 if (is_software_event(event))
526 return true;
527 /* Reject groups spanning multiple HW PMUs. */
528 if (event->pmu != pmu)
529 return false;
530
531 *counters = *counters + 1;
532 return true;
533 }
534
535 /*
536 * Make sure the group of events can be scheduled at once
537 * on the PMU.
538 */
tx2_uncore_validate_event_group(struct perf_event * event,int max_counters)539 static bool tx2_uncore_validate_event_group(struct perf_event *event,
540 int max_counters)
541 {
542 struct perf_event *sibling, *leader = event->group_leader;
543 int counters = 0;
544
545 if (event->group_leader == event)
546 return true;
547
548 if (!tx2_uncore_validate_event(event->pmu, leader, &counters))
549 return false;
550
551 for_each_sibling_event(sibling, leader) {
552 if (!tx2_uncore_validate_event(event->pmu, sibling, &counters))
553 return false;
554 }
555
556 if (!tx2_uncore_validate_event(event->pmu, event, &counters))
557 return false;
558
559 /*
560 * If the group requires more counters than the HW has,
561 * it cannot ever be scheduled.
562 */
563 return counters <= max_counters;
564 }
565
566
tx2_uncore_event_init(struct perf_event * event)567 static int tx2_uncore_event_init(struct perf_event *event)
568 {
569 struct hw_perf_event *hwc = &event->hw;
570 struct tx2_uncore_pmu *tx2_pmu;
571
572 /* Test the event attr type check for PMU enumeration */
573 if (event->attr.type != event->pmu->type)
574 return -ENOENT;
575
576 /*
577 * SOC PMU counters are shared across all cores.
578 * Therefore, it does not support per-process mode.
579 * Also, it does not support event sampling mode.
580 */
581 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
582 return -EINVAL;
583
584 if (event->cpu < 0)
585 return -EINVAL;
586
587 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
588 if (tx2_pmu->cpu >= nr_cpu_ids)
589 return -EINVAL;
590 event->cpu = tx2_pmu->cpu;
591
592 if (event->attr.config >= tx2_pmu->max_events)
593 return -EINVAL;
594
595 /* store event id */
596 hwc->config = event->attr.config;
597
598 /* Validate the group */
599 if (!tx2_uncore_validate_event_group(event, tx2_pmu->max_counters))
600 return -EINVAL;
601
602 return 0;
603 }
604
tx2_uncore_event_start(struct perf_event * event,int flags)605 static void tx2_uncore_event_start(struct perf_event *event, int flags)
606 {
607 struct hw_perf_event *hwc = &event->hw;
608 struct tx2_uncore_pmu *tx2_pmu;
609
610 hwc->state = 0;
611 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
612
613 tx2_pmu->start_event(event, flags);
614 perf_event_update_userpage(event);
615
616 /* No hrtimer needed for CCPI2, 64-bit counters */
617 if (!tx2_pmu->hrtimer_callback)
618 return;
619
620 /* Start timer for first event */
621 if (bitmap_weight(tx2_pmu->active_counters,
622 tx2_pmu->max_counters) == 1) {
623 hrtimer_start(&tx2_pmu->hrtimer,
624 ns_to_ktime(tx2_pmu->hrtimer_interval),
625 HRTIMER_MODE_REL_PINNED);
626 }
627 }
628
tx2_uncore_event_stop(struct perf_event * event,int flags)629 static void tx2_uncore_event_stop(struct perf_event *event, int flags)
630 {
631 struct hw_perf_event *hwc = &event->hw;
632 struct tx2_uncore_pmu *tx2_pmu;
633
634 if (hwc->state & PERF_HES_UPTODATE)
635 return;
636
637 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
638 tx2_pmu->stop_event(event);
639 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
640 hwc->state |= PERF_HES_STOPPED;
641 if (flags & PERF_EF_UPDATE) {
642 tx2_uncore_event_update(event);
643 hwc->state |= PERF_HES_UPTODATE;
644 }
645 }
646
tx2_uncore_event_add(struct perf_event * event,int flags)647 static int tx2_uncore_event_add(struct perf_event *event, int flags)
648 {
649 struct hw_perf_event *hwc = &event->hw;
650 struct tx2_uncore_pmu *tx2_pmu;
651
652 tx2_pmu = pmu_to_tx2_pmu(event->pmu);
653
654 /* Allocate a free counter */
655 hwc->idx = alloc_counter(tx2_pmu);
656 if (hwc->idx < 0)
657 return -EAGAIN;
658
659 tx2_pmu->events[hwc->idx] = event;
660 /* set counter control and data registers base address */
661 tx2_pmu->init_cntr_base(event, tx2_pmu);
662
663 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
664 if (flags & PERF_EF_START)
665 tx2_uncore_event_start(event, flags);
666
667 return 0;
668 }
669
tx2_uncore_event_del(struct perf_event * event,int flags)670 static void tx2_uncore_event_del(struct perf_event *event, int flags)
671 {
672 struct tx2_uncore_pmu *tx2_pmu = pmu_to_tx2_pmu(event->pmu);
673 struct hw_perf_event *hwc = &event->hw;
674 u32 cmask;
675
676 cmask = tx2_pmu->counters_mask;
677 tx2_uncore_event_stop(event, PERF_EF_UPDATE);
678
679 /* clear the assigned counter */
680 free_counter(tx2_pmu, GET_COUNTERID(event, cmask));
681
682 perf_event_update_userpage(event);
683 tx2_pmu->events[hwc->idx] = NULL;
684 hwc->idx = -1;
685
686 if (!tx2_pmu->hrtimer_callback)
687 return;
688
689 if (bitmap_empty(tx2_pmu->active_counters, tx2_pmu->max_counters))
690 hrtimer_cancel(&tx2_pmu->hrtimer);
691 }
692
tx2_uncore_event_read(struct perf_event * event)693 static void tx2_uncore_event_read(struct perf_event *event)
694 {
695 tx2_uncore_event_update(event);
696 }
697
tx2_hrtimer_callback(struct hrtimer * timer)698 static enum hrtimer_restart tx2_hrtimer_callback(struct hrtimer *timer)
699 {
700 struct tx2_uncore_pmu *tx2_pmu;
701 int max_counters, idx;
702
703 tx2_pmu = container_of(timer, struct tx2_uncore_pmu, hrtimer);
704 max_counters = tx2_pmu->max_counters;
705
706 if (bitmap_empty(tx2_pmu->active_counters, max_counters))
707 return HRTIMER_NORESTART;
708
709 for_each_set_bit(idx, tx2_pmu->active_counters, max_counters) {
710 struct perf_event *event = tx2_pmu->events[idx];
711
712 tx2_uncore_event_update(event);
713 }
714 hrtimer_forward_now(timer, ns_to_ktime(tx2_pmu->hrtimer_interval));
715 return HRTIMER_RESTART;
716 }
717
tx2_uncore_pmu_register(struct tx2_uncore_pmu * tx2_pmu)718 static int tx2_uncore_pmu_register(
719 struct tx2_uncore_pmu *tx2_pmu)
720 {
721 struct device *dev = tx2_pmu->dev;
722 char *name = tx2_pmu->name;
723
724 /* Perf event registration */
725 tx2_pmu->pmu = (struct pmu) {
726 .module = THIS_MODULE,
727 .parent = tx2_pmu->dev,
728 .attr_groups = tx2_pmu->attr_groups,
729 .task_ctx_nr = perf_invalid_context,
730 .event_init = tx2_uncore_event_init,
731 .add = tx2_uncore_event_add,
732 .del = tx2_uncore_event_del,
733 .start = tx2_uncore_event_start,
734 .stop = tx2_uncore_event_stop,
735 .read = tx2_uncore_event_read,
736 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
737 };
738
739 tx2_pmu->pmu.name = devm_kasprintf(dev, GFP_KERNEL,
740 "%s", name);
741
742 return perf_pmu_register(&tx2_pmu->pmu, tx2_pmu->pmu.name, -1);
743 }
744
tx2_uncore_pmu_add_dev(struct tx2_uncore_pmu * tx2_pmu)745 static int tx2_uncore_pmu_add_dev(struct tx2_uncore_pmu *tx2_pmu)
746 {
747 int ret, cpu;
748
749 cpu = cpumask_any_and(cpumask_of_node(tx2_pmu->node),
750 cpu_online_mask);
751
752 tx2_pmu->cpu = cpu;
753
754 if (tx2_pmu->hrtimer_callback) {
755 hrtimer_init(&tx2_pmu->hrtimer,
756 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
757 tx2_pmu->hrtimer.function = tx2_pmu->hrtimer_callback;
758 }
759
760 ret = tx2_uncore_pmu_register(tx2_pmu);
761 if (ret) {
762 dev_err(tx2_pmu->dev, "%s PMU: Failed to init driver\n",
763 tx2_pmu->name);
764 return -ENODEV;
765 }
766
767 /* register hotplug callback for the pmu */
768 ret = cpuhp_state_add_instance(
769 CPUHP_AP_PERF_ARM_CAVIUM_TX2_UNCORE_ONLINE,
770 &tx2_pmu->hpnode);
771 if (ret) {
772 dev_err(tx2_pmu->dev, "Error %d registering hotplug", ret);
773 return ret;
774 }
775
776 /* Add to list */
777 list_add(&tx2_pmu->entry, &tx2_pmus);
778
779 dev_dbg(tx2_pmu->dev, "%s PMU UNCORE registered\n",
780 tx2_pmu->pmu.name);
781 return ret;
782 }
783
tx2_uncore_pmu_init_dev(struct device * dev,acpi_handle handle,struct acpi_device * adev,u32 type)784 static struct tx2_uncore_pmu *tx2_uncore_pmu_init_dev(struct device *dev,
785 acpi_handle handle, struct acpi_device *adev, u32 type)
786 {
787 struct tx2_uncore_pmu *tx2_pmu;
788 void __iomem *base;
789 struct resource res;
790 struct resource_entry *rentry;
791 struct list_head list;
792 int ret;
793
794 INIT_LIST_HEAD(&list);
795 ret = acpi_dev_get_resources(adev, &list, NULL, NULL);
796 if (ret <= 0) {
797 dev_err(dev, "failed to parse _CRS method, error %d\n", ret);
798 return NULL;
799 }
800
801 list_for_each_entry(rentry, &list, node) {
802 if (resource_type(rentry->res) == IORESOURCE_MEM) {
803 res = *rentry->res;
804 rentry = NULL;
805 break;
806 }
807 }
808 acpi_dev_free_resource_list(&list);
809
810 if (rentry) {
811 dev_err(dev, "PMU type %d: Fail to find resource\n", type);
812 return NULL;
813 }
814
815 base = devm_ioremap_resource(dev, &res);
816 if (IS_ERR(base))
817 return NULL;
818
819 tx2_pmu = devm_kzalloc(dev, sizeof(*tx2_pmu), GFP_KERNEL);
820 if (!tx2_pmu)
821 return NULL;
822
823 tx2_pmu->dev = dev;
824 tx2_pmu->type = type;
825 tx2_pmu->base = base;
826 tx2_pmu->node = dev_to_node(dev);
827 INIT_LIST_HEAD(&tx2_pmu->entry);
828
829 switch (tx2_pmu->type) {
830 case PMU_TYPE_L3C:
831 tx2_pmu->max_counters = TX2_PMU_DMC_L3C_MAX_COUNTERS;
832 tx2_pmu->counters_mask = 0x3;
833 tx2_pmu->prorate_factor = TX2_PMU_L3_TILES;
834 tx2_pmu->max_events = L3_EVENT_MAX;
835 tx2_pmu->events_mask = 0x1f;
836 tx2_pmu->hrtimer_interval = TX2_PMU_HRTIMER_INTERVAL;
837 tx2_pmu->hrtimer_callback = tx2_hrtimer_callback;
838 tx2_pmu->attr_groups = l3c_pmu_attr_groups;
839 tx2_pmu->name = devm_kasprintf(dev, GFP_KERNEL,
840 "uncore_l3c_%d", tx2_pmu->node);
841 tx2_pmu->init_cntr_base = init_cntr_base_l3c;
842 tx2_pmu->start_event = uncore_start_event_l3c;
843 tx2_pmu->stop_event = uncore_stop_event_l3c;
844 break;
845 case PMU_TYPE_DMC:
846 tx2_pmu->max_counters = TX2_PMU_DMC_L3C_MAX_COUNTERS;
847 tx2_pmu->counters_mask = 0x3;
848 tx2_pmu->prorate_factor = TX2_PMU_DMC_CHANNELS;
849 tx2_pmu->max_events = DMC_EVENT_MAX;
850 tx2_pmu->events_mask = 0x1f;
851 tx2_pmu->hrtimer_interval = TX2_PMU_HRTIMER_INTERVAL;
852 tx2_pmu->hrtimer_callback = tx2_hrtimer_callback;
853 tx2_pmu->attr_groups = dmc_pmu_attr_groups;
854 tx2_pmu->name = devm_kasprintf(dev, GFP_KERNEL,
855 "uncore_dmc_%d", tx2_pmu->node);
856 tx2_pmu->init_cntr_base = init_cntr_base_dmc;
857 tx2_pmu->start_event = uncore_start_event_dmc;
858 tx2_pmu->stop_event = uncore_stop_event_dmc;
859 break;
860 case PMU_TYPE_CCPI2:
861 /* CCPI2 has 8 counters */
862 tx2_pmu->max_counters = TX2_PMU_CCPI2_MAX_COUNTERS;
863 tx2_pmu->counters_mask = 0x7;
864 tx2_pmu->prorate_factor = 1;
865 tx2_pmu->max_events = CCPI2_EVENT_MAX;
866 tx2_pmu->events_mask = 0x1ff;
867 tx2_pmu->attr_groups = ccpi2_pmu_attr_groups;
868 tx2_pmu->name = devm_kasprintf(dev, GFP_KERNEL,
869 "uncore_ccpi2_%d", tx2_pmu->node);
870 tx2_pmu->init_cntr_base = init_cntr_base_ccpi2;
871 tx2_pmu->start_event = uncore_start_event_ccpi2;
872 tx2_pmu->stop_event = uncore_stop_event_ccpi2;
873 tx2_pmu->hrtimer_callback = NULL;
874 break;
875 case PMU_TYPE_INVALID:
876 devm_kfree(dev, tx2_pmu);
877 return NULL;
878 }
879
880 return tx2_pmu;
881 }
882
tx2_uncore_pmu_add(acpi_handle handle,u32 level,void * data,void ** return_value)883 static acpi_status tx2_uncore_pmu_add(acpi_handle handle, u32 level,
884 void *data, void **return_value)
885 {
886 struct acpi_device *adev = acpi_fetch_acpi_dev(handle);
887 struct tx2_uncore_pmu *tx2_pmu;
888 enum tx2_uncore_type type;
889
890 if (!adev || acpi_bus_get_status(adev) || !adev->status.present)
891 return AE_OK;
892
893 type = get_tx2_pmu_type(adev);
894 if (type == PMU_TYPE_INVALID)
895 return AE_OK;
896
897 tx2_pmu = tx2_uncore_pmu_init_dev((struct device *)data,
898 handle, adev, type);
899
900 if (!tx2_pmu)
901 return AE_ERROR;
902
903 if (tx2_uncore_pmu_add_dev(tx2_pmu)) {
904 /* Can't add the PMU device, abort */
905 return AE_ERROR;
906 }
907 return AE_OK;
908 }
909
tx2_uncore_pmu_online_cpu(unsigned int cpu,struct hlist_node * hpnode)910 static int tx2_uncore_pmu_online_cpu(unsigned int cpu,
911 struct hlist_node *hpnode)
912 {
913 struct tx2_uncore_pmu *tx2_pmu;
914
915 tx2_pmu = hlist_entry_safe(hpnode,
916 struct tx2_uncore_pmu, hpnode);
917
918 /* Pick this CPU, If there is no CPU/PMU association and both are
919 * from same node.
920 */
921 if ((tx2_pmu->cpu >= nr_cpu_ids) &&
922 (tx2_pmu->node == cpu_to_node(cpu)))
923 tx2_pmu->cpu = cpu;
924
925 return 0;
926 }
927
tx2_uncore_pmu_offline_cpu(unsigned int cpu,struct hlist_node * hpnode)928 static int tx2_uncore_pmu_offline_cpu(unsigned int cpu,
929 struct hlist_node *hpnode)
930 {
931 struct tx2_uncore_pmu *tx2_pmu;
932 unsigned int new_cpu;
933
934 tx2_pmu = hlist_entry_safe(hpnode,
935 struct tx2_uncore_pmu, hpnode);
936
937 if (cpu != tx2_pmu->cpu)
938 return 0;
939
940 if (tx2_pmu->hrtimer_callback)
941 hrtimer_cancel(&tx2_pmu->hrtimer);
942
943 new_cpu = cpumask_any_and_but(cpumask_of_node(tx2_pmu->node),
944 cpu_online_mask, cpu);
945
946 tx2_pmu->cpu = new_cpu;
947 if (new_cpu >= nr_cpu_ids)
948 return 0;
949 perf_pmu_migrate_context(&tx2_pmu->pmu, cpu, new_cpu);
950
951 return 0;
952 }
953
954 static const struct acpi_device_id tx2_uncore_acpi_match[] = {
955 {"CAV901C", 0},
956 {},
957 };
958 MODULE_DEVICE_TABLE(acpi, tx2_uncore_acpi_match);
959
tx2_uncore_probe(struct platform_device * pdev)960 static int tx2_uncore_probe(struct platform_device *pdev)
961 {
962 struct device *dev = &pdev->dev;
963 acpi_handle handle;
964 acpi_status status;
965
966 set_dev_node(dev, acpi_get_node(ACPI_HANDLE(dev)));
967
968 if (!has_acpi_companion(dev))
969 return -ENODEV;
970
971 handle = ACPI_HANDLE(dev);
972 if (!handle)
973 return -EINVAL;
974
975 /* Walk through the tree for all PMU UNCORE devices */
976 status = acpi_walk_namespace(ACPI_TYPE_DEVICE, handle, 1,
977 tx2_uncore_pmu_add,
978 NULL, dev, NULL);
979 if (ACPI_FAILURE(status)) {
980 dev_err(dev, "failed to probe PMU devices\n");
981 return_ACPI_STATUS(status);
982 }
983
984 dev_info(dev, "node%d: pmu uncore registered\n", dev_to_node(dev));
985 return 0;
986 }
987
tx2_uncore_remove(struct platform_device * pdev)988 static void tx2_uncore_remove(struct platform_device *pdev)
989 {
990 struct tx2_uncore_pmu *tx2_pmu, *temp;
991 struct device *dev = &pdev->dev;
992
993 if (!list_empty(&tx2_pmus)) {
994 list_for_each_entry_safe(tx2_pmu, temp, &tx2_pmus, entry) {
995 if (tx2_pmu->node == dev_to_node(dev)) {
996 cpuhp_state_remove_instance_nocalls(
997 CPUHP_AP_PERF_ARM_CAVIUM_TX2_UNCORE_ONLINE,
998 &tx2_pmu->hpnode);
999 perf_pmu_unregister(&tx2_pmu->pmu);
1000 list_del(&tx2_pmu->entry);
1001 }
1002 }
1003 }
1004 }
1005
1006 static struct platform_driver tx2_uncore_driver = {
1007 .driver = {
1008 .name = "tx2-uncore-pmu",
1009 .acpi_match_table = ACPI_PTR(tx2_uncore_acpi_match),
1010 .suppress_bind_attrs = true,
1011 },
1012 .probe = tx2_uncore_probe,
1013 .remove_new = tx2_uncore_remove,
1014 };
1015
tx2_uncore_driver_init(void)1016 static int __init tx2_uncore_driver_init(void)
1017 {
1018 int ret;
1019
1020 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CAVIUM_TX2_UNCORE_ONLINE,
1021 "perf/tx2/uncore:online",
1022 tx2_uncore_pmu_online_cpu,
1023 tx2_uncore_pmu_offline_cpu);
1024 if (ret) {
1025 pr_err("TX2 PMU: setup hotplug failed(%d)\n", ret);
1026 return ret;
1027 }
1028 ret = platform_driver_register(&tx2_uncore_driver);
1029 if (ret)
1030 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CAVIUM_TX2_UNCORE_ONLINE);
1031
1032 return ret;
1033 }
1034 module_init(tx2_uncore_driver_init);
1035
tx2_uncore_driver_exit(void)1036 static void __exit tx2_uncore_driver_exit(void)
1037 {
1038 platform_driver_unregister(&tx2_uncore_driver);
1039 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CAVIUM_TX2_UNCORE_ONLINE);
1040 }
1041 module_exit(tx2_uncore_driver_exit);
1042
1043 MODULE_DESCRIPTION("ThunderX2 UNCORE PMU driver");
1044 MODULE_LICENSE("GPL v2");
1045 MODULE_AUTHOR("Ganapatrao Kulkarni <gkulkarni@cavium.com>");
1046