1 // SPDX-License-Identifier: GPL-2.0
2
3 /*
4 * This driver adds support for perf events to use the Performance
5 * Monitor Counter Groups (PMCG) associated with an SMMUv3 node
6 * to monitor that node.
7 *
8 * SMMUv3 PMCG devices are named as smmuv3_pmcg_<phys_addr_page> where
9 * <phys_addr_page> is the physical page address of the SMMU PMCG wrapped
10 * to 4K boundary. For example, the PMCG at 0xff88840000 is named
11 * smmuv3_pmcg_ff88840
12 *
13 * Filtering by stream id is done by specifying filtering parameters
14 * with the event. options are:
15 * filter_enable - 0 = no filtering, 1 = filtering enabled
16 * filter_span - 0 = exact match, 1 = pattern match
17 * filter_stream_id - pattern to filter against
18 *
19 * To match a partial StreamID where the X most-significant bits must match
20 * but the Y least-significant bits might differ, STREAMID is programmed
21 * with a value that contains:
22 * STREAMID[Y - 1] == 0.
23 * STREAMID[Y - 2:0] == 1 (where Y > 1).
24 * The remainder of implemented bits of STREAMID (X bits, from bit Y upwards)
25 * contain a value to match from the corresponding bits of event StreamID.
26 *
27 * Example: perf stat -e smmuv3_pmcg_ff88840/transaction,filter_enable=1,
28 * filter_span=1,filter_stream_id=0x42/ -a netperf
29 * Applies filter pattern 0x42 to transaction events, which means events
30 * matching stream ids 0x42 and 0x43 are counted. Further filtering
31 * information is available in the SMMU documentation.
32 *
33 * SMMU events are not attributable to a CPU, so task mode and sampling
34 * are not supported.
35 */
36
37 #include <linux/acpi.h>
38 #include <linux/acpi_iort.h>
39 #include <linux/bitfield.h>
40 #include <linux/bitops.h>
41 #include <linux/cpuhotplug.h>
42 #include <linux/cpumask.h>
43 #include <linux/device.h>
44 #include <linux/errno.h>
45 #include <linux/interrupt.h>
46 #include <linux/irq.h>
47 #include <linux/kernel.h>
48 #include <linux/list.h>
49 #include <linux/msi.h>
50 #include <linux/of.h>
51 #include <linux/perf_event.h>
52 #include <linux/platform_device.h>
53 #include <linux/smp.h>
54 #include <linux/sysfs.h>
55 #include <linux/types.h>
56
57 #define SMMU_PMCG_EVCNTR0 0x0
58 #define SMMU_PMCG_EVCNTR(n, stride) (SMMU_PMCG_EVCNTR0 + (n) * (stride))
59 #define SMMU_PMCG_EVTYPER0 0x400
60 #define SMMU_PMCG_EVTYPER(n) (SMMU_PMCG_EVTYPER0 + (n) * 4)
61 #define SMMU_PMCG_SID_SPAN_SHIFT 29
62 #define SMMU_PMCG_SMR0 0xA00
63 #define SMMU_PMCG_SMR(n) (SMMU_PMCG_SMR0 + (n) * 4)
64 #define SMMU_PMCG_CNTENSET0 0xC00
65 #define SMMU_PMCG_CNTENCLR0 0xC20
66 #define SMMU_PMCG_INTENSET0 0xC40
67 #define SMMU_PMCG_INTENCLR0 0xC60
68 #define SMMU_PMCG_OVSCLR0 0xC80
69 #define SMMU_PMCG_OVSSET0 0xCC0
70 #define SMMU_PMCG_CFGR 0xE00
71 #define SMMU_PMCG_CFGR_SID_FILTER_TYPE BIT(23)
72 #define SMMU_PMCG_CFGR_MSI BIT(21)
73 #define SMMU_PMCG_CFGR_RELOC_CTRS BIT(20)
74 #define SMMU_PMCG_CFGR_SIZE GENMASK(13, 8)
75 #define SMMU_PMCG_CFGR_NCTR GENMASK(5, 0)
76 #define SMMU_PMCG_CR 0xE04
77 #define SMMU_PMCG_CR_ENABLE BIT(0)
78 #define SMMU_PMCG_IIDR 0xE08
79 #define SMMU_PMCG_IIDR_PRODUCTID GENMASK(31, 20)
80 #define SMMU_PMCG_IIDR_VARIANT GENMASK(19, 16)
81 #define SMMU_PMCG_IIDR_REVISION GENMASK(15, 12)
82 #define SMMU_PMCG_IIDR_IMPLEMENTER GENMASK(11, 0)
83 #define SMMU_PMCG_CEID0 0xE20
84 #define SMMU_PMCG_CEID1 0xE28
85 #define SMMU_PMCG_IRQ_CTRL 0xE50
86 #define SMMU_PMCG_IRQ_CTRL_IRQEN BIT(0)
87 #define SMMU_PMCG_IRQ_CFG0 0xE58
88 #define SMMU_PMCG_IRQ_CFG1 0xE60
89 #define SMMU_PMCG_IRQ_CFG2 0xE64
90
91 /* IMP-DEF ID registers */
92 #define SMMU_PMCG_PIDR0 0xFE0
93 #define SMMU_PMCG_PIDR0_PART_0 GENMASK(7, 0)
94 #define SMMU_PMCG_PIDR1 0xFE4
95 #define SMMU_PMCG_PIDR1_DES_0 GENMASK(7, 4)
96 #define SMMU_PMCG_PIDR1_PART_1 GENMASK(3, 0)
97 #define SMMU_PMCG_PIDR2 0xFE8
98 #define SMMU_PMCG_PIDR2_REVISION GENMASK(7, 4)
99 #define SMMU_PMCG_PIDR2_DES_1 GENMASK(2, 0)
100 #define SMMU_PMCG_PIDR3 0xFEC
101 #define SMMU_PMCG_PIDR3_REVAND GENMASK(7, 4)
102 #define SMMU_PMCG_PIDR4 0xFD0
103 #define SMMU_PMCG_PIDR4_DES_2 GENMASK(3, 0)
104
105 /* MSI config fields */
106 #define MSI_CFG0_ADDR_MASK GENMASK_ULL(51, 2)
107 #define MSI_CFG2_MEMATTR_DEVICE_nGnRE 0x1
108
109 #define SMMU_PMCG_DEFAULT_FILTER_SPAN 1
110 #define SMMU_PMCG_DEFAULT_FILTER_SID GENMASK(31, 0)
111
112 #define SMMU_PMCG_MAX_COUNTERS 64
113 #define SMMU_PMCG_ARCH_MAX_EVENTS 128
114
115 #define SMMU_PMCG_PA_SHIFT 12
116
117 #define SMMU_PMCG_EVCNTR_RDONLY BIT(0)
118
119 static int cpuhp_state_num;
120
121 struct smmu_pmu {
122 struct hlist_node node;
123 struct perf_event *events[SMMU_PMCG_MAX_COUNTERS];
124 DECLARE_BITMAP(used_counters, SMMU_PMCG_MAX_COUNTERS);
125 DECLARE_BITMAP(supported_events, SMMU_PMCG_ARCH_MAX_EVENTS);
126 unsigned int irq;
127 unsigned int on_cpu;
128 struct pmu pmu;
129 unsigned int num_counters;
130 struct device *dev;
131 void __iomem *reg_base;
132 void __iomem *reloc_base;
133 u64 counter_mask;
134 u32 options;
135 u32 iidr;
136 bool global_filter;
137 };
138
139 #define to_smmu_pmu(p) (container_of(p, struct smmu_pmu, pmu))
140
141 #define SMMU_PMU_EVENT_ATTR_EXTRACTOR(_name, _config, _start, _end) \
142 static inline u32 get_##_name(struct perf_event *event) \
143 { \
144 return FIELD_GET(GENMASK_ULL(_end, _start), \
145 event->attr._config); \
146 } \
147
148 SMMU_PMU_EVENT_ATTR_EXTRACTOR(event, config, 0, 15);
149 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_stream_id, config1, 0, 31);
150 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_span, config1, 32, 32);
151 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_enable, config1, 33, 33);
152
smmu_pmu_enable(struct pmu * pmu)153 static inline void smmu_pmu_enable(struct pmu *pmu)
154 {
155 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
156
157 writel(SMMU_PMCG_IRQ_CTRL_IRQEN,
158 smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
159 writel(SMMU_PMCG_CR_ENABLE, smmu_pmu->reg_base + SMMU_PMCG_CR);
160 }
161
smmu_pmu_disable(struct pmu * pmu)162 static inline void smmu_pmu_disable(struct pmu *pmu)
163 {
164 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
165
166 writel(0, smmu_pmu->reg_base + SMMU_PMCG_CR);
167 writel(0, smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
168 }
169
smmu_pmu_counter_set_value(struct smmu_pmu * smmu_pmu,u32 idx,u64 value)170 static inline void smmu_pmu_counter_set_value(struct smmu_pmu *smmu_pmu,
171 u32 idx, u64 value)
172 {
173 if (smmu_pmu->counter_mask & BIT(32))
174 writeq(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
175 else
176 writel(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
177 }
178
smmu_pmu_counter_get_value(struct smmu_pmu * smmu_pmu,u32 idx)179 static inline u64 smmu_pmu_counter_get_value(struct smmu_pmu *smmu_pmu, u32 idx)
180 {
181 u64 value;
182
183 if (smmu_pmu->counter_mask & BIT(32))
184 value = readq(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
185 else
186 value = readl(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
187
188 return value;
189 }
190
smmu_pmu_counter_enable(struct smmu_pmu * smmu_pmu,u32 idx)191 static inline void smmu_pmu_counter_enable(struct smmu_pmu *smmu_pmu, u32 idx)
192 {
193 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENSET0);
194 }
195
smmu_pmu_counter_disable(struct smmu_pmu * smmu_pmu,u32 idx)196 static inline void smmu_pmu_counter_disable(struct smmu_pmu *smmu_pmu, u32 idx)
197 {
198 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
199 }
200
smmu_pmu_interrupt_enable(struct smmu_pmu * smmu_pmu,u32 idx)201 static inline void smmu_pmu_interrupt_enable(struct smmu_pmu *smmu_pmu, u32 idx)
202 {
203 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENSET0);
204 }
205
smmu_pmu_interrupt_disable(struct smmu_pmu * smmu_pmu,u32 idx)206 static inline void smmu_pmu_interrupt_disable(struct smmu_pmu *smmu_pmu,
207 u32 idx)
208 {
209 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
210 }
211
smmu_pmu_set_evtyper(struct smmu_pmu * smmu_pmu,u32 idx,u32 val)212 static inline void smmu_pmu_set_evtyper(struct smmu_pmu *smmu_pmu, u32 idx,
213 u32 val)
214 {
215 writel(val, smmu_pmu->reg_base + SMMU_PMCG_EVTYPER(idx));
216 }
217
smmu_pmu_set_smr(struct smmu_pmu * smmu_pmu,u32 idx,u32 val)218 static inline void smmu_pmu_set_smr(struct smmu_pmu *smmu_pmu, u32 idx, u32 val)
219 {
220 writel(val, smmu_pmu->reg_base + SMMU_PMCG_SMR(idx));
221 }
222
smmu_pmu_event_update(struct perf_event * event)223 static void smmu_pmu_event_update(struct perf_event *event)
224 {
225 struct hw_perf_event *hwc = &event->hw;
226 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
227 u64 delta, prev, now;
228 u32 idx = hwc->idx;
229
230 do {
231 prev = local64_read(&hwc->prev_count);
232 now = smmu_pmu_counter_get_value(smmu_pmu, idx);
233 } while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
234
235 /* handle overflow. */
236 delta = now - prev;
237 delta &= smmu_pmu->counter_mask;
238
239 local64_add(delta, &event->count);
240 }
241
smmu_pmu_set_period(struct smmu_pmu * smmu_pmu,struct hw_perf_event * hwc)242 static void smmu_pmu_set_period(struct smmu_pmu *smmu_pmu,
243 struct hw_perf_event *hwc)
244 {
245 u32 idx = hwc->idx;
246 u64 new;
247
248 if (smmu_pmu->options & SMMU_PMCG_EVCNTR_RDONLY) {
249 /*
250 * On platforms that require this quirk, if the counter starts
251 * at < half_counter value and wraps, the current logic of
252 * handling the overflow may not work. It is expected that,
253 * those platforms will have full 64 counter bits implemented
254 * so that such a possibility is remote(eg: HiSilicon HIP08).
255 */
256 new = smmu_pmu_counter_get_value(smmu_pmu, idx);
257 } else {
258 /*
259 * We limit the max period to half the max counter value
260 * of the counter size, so that even in the case of extreme
261 * interrupt latency the counter will (hopefully) not wrap
262 * past its initial value.
263 */
264 new = smmu_pmu->counter_mask >> 1;
265 smmu_pmu_counter_set_value(smmu_pmu, idx, new);
266 }
267
268 local64_set(&hwc->prev_count, new);
269 }
270
smmu_pmu_set_event_filter(struct perf_event * event,int idx,u32 span,u32 sid)271 static void smmu_pmu_set_event_filter(struct perf_event *event,
272 int idx, u32 span, u32 sid)
273 {
274 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
275 u32 evtyper;
276
277 evtyper = get_event(event) | span << SMMU_PMCG_SID_SPAN_SHIFT;
278 smmu_pmu_set_evtyper(smmu_pmu, idx, evtyper);
279 smmu_pmu_set_smr(smmu_pmu, idx, sid);
280 }
281
smmu_pmu_check_global_filter(struct perf_event * curr,struct perf_event * new)282 static bool smmu_pmu_check_global_filter(struct perf_event *curr,
283 struct perf_event *new)
284 {
285 if (get_filter_enable(new) != get_filter_enable(curr))
286 return false;
287
288 if (!get_filter_enable(new))
289 return true;
290
291 return get_filter_span(new) == get_filter_span(curr) &&
292 get_filter_stream_id(new) == get_filter_stream_id(curr);
293 }
294
smmu_pmu_apply_event_filter(struct smmu_pmu * smmu_pmu,struct perf_event * event,int idx)295 static int smmu_pmu_apply_event_filter(struct smmu_pmu *smmu_pmu,
296 struct perf_event *event, int idx)
297 {
298 u32 span, sid;
299 unsigned int cur_idx, num_ctrs = smmu_pmu->num_counters;
300 bool filter_en = !!get_filter_enable(event);
301
302 span = filter_en ? get_filter_span(event) :
303 SMMU_PMCG_DEFAULT_FILTER_SPAN;
304 sid = filter_en ? get_filter_stream_id(event) :
305 SMMU_PMCG_DEFAULT_FILTER_SID;
306
307 cur_idx = find_first_bit(smmu_pmu->used_counters, num_ctrs);
308 /*
309 * Per-counter filtering, or scheduling the first globally-filtered
310 * event into an empty PMU so idx == 0 and it works out equivalent.
311 */
312 if (!smmu_pmu->global_filter || cur_idx == num_ctrs) {
313 smmu_pmu_set_event_filter(event, idx, span, sid);
314 return 0;
315 }
316
317 /* Otherwise, must match whatever's currently scheduled */
318 if (smmu_pmu_check_global_filter(smmu_pmu->events[cur_idx], event)) {
319 smmu_pmu_set_evtyper(smmu_pmu, idx, get_event(event));
320 return 0;
321 }
322
323 return -EAGAIN;
324 }
325
smmu_pmu_get_event_idx(struct smmu_pmu * smmu_pmu,struct perf_event * event)326 static int smmu_pmu_get_event_idx(struct smmu_pmu *smmu_pmu,
327 struct perf_event *event)
328 {
329 int idx, err;
330 unsigned int num_ctrs = smmu_pmu->num_counters;
331
332 idx = find_first_zero_bit(smmu_pmu->used_counters, num_ctrs);
333 if (idx == num_ctrs)
334 /* The counters are all in use. */
335 return -EAGAIN;
336
337 err = smmu_pmu_apply_event_filter(smmu_pmu, event, idx);
338 if (err)
339 return err;
340
341 set_bit(idx, smmu_pmu->used_counters);
342
343 return idx;
344 }
345
smmu_pmu_events_compatible(struct perf_event * curr,struct perf_event * new)346 static bool smmu_pmu_events_compatible(struct perf_event *curr,
347 struct perf_event *new)
348 {
349 if (new->pmu != curr->pmu)
350 return false;
351
352 if (to_smmu_pmu(new->pmu)->global_filter &&
353 !smmu_pmu_check_global_filter(curr, new))
354 return false;
355
356 return true;
357 }
358
359 /*
360 * Implementation of abstract pmu functionality required by
361 * the core perf events code.
362 */
363
smmu_pmu_event_init(struct perf_event * event)364 static int smmu_pmu_event_init(struct perf_event *event)
365 {
366 struct hw_perf_event *hwc = &event->hw;
367 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
368 struct device *dev = smmu_pmu->dev;
369 struct perf_event *sibling;
370 int group_num_events = 1;
371 u16 event_id;
372
373 if (event->attr.type != event->pmu->type)
374 return -ENOENT;
375
376 if (hwc->sample_period) {
377 dev_dbg(dev, "Sampling not supported\n");
378 return -EOPNOTSUPP;
379 }
380
381 if (event->cpu < 0) {
382 dev_dbg(dev, "Per-task mode not supported\n");
383 return -EOPNOTSUPP;
384 }
385
386 /* Verify specified event is supported on this PMU */
387 event_id = get_event(event);
388 if (event_id < SMMU_PMCG_ARCH_MAX_EVENTS &&
389 (!test_bit(event_id, smmu_pmu->supported_events))) {
390 dev_dbg(dev, "Invalid event %d for this PMU\n", event_id);
391 return -EINVAL;
392 }
393
394 /* Don't allow groups with mixed PMUs, except for s/w events */
395 if (!is_software_event(event->group_leader)) {
396 if (!smmu_pmu_events_compatible(event->group_leader, event))
397 return -EINVAL;
398
399 if (++group_num_events > smmu_pmu->num_counters)
400 return -EINVAL;
401 }
402
403 for_each_sibling_event(sibling, event->group_leader) {
404 if (is_software_event(sibling))
405 continue;
406
407 if (!smmu_pmu_events_compatible(sibling, event))
408 return -EINVAL;
409
410 if (++group_num_events > smmu_pmu->num_counters)
411 return -EINVAL;
412 }
413
414 hwc->idx = -1;
415
416 /*
417 * Ensure all events are on the same cpu so all events are in the
418 * same cpu context, to avoid races on pmu_enable etc.
419 */
420 event->cpu = smmu_pmu->on_cpu;
421
422 return 0;
423 }
424
smmu_pmu_event_start(struct perf_event * event,int flags)425 static void smmu_pmu_event_start(struct perf_event *event, int flags)
426 {
427 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
428 struct hw_perf_event *hwc = &event->hw;
429 int idx = hwc->idx;
430
431 hwc->state = 0;
432
433 smmu_pmu_set_period(smmu_pmu, hwc);
434
435 smmu_pmu_counter_enable(smmu_pmu, idx);
436 }
437
smmu_pmu_event_stop(struct perf_event * event,int flags)438 static void smmu_pmu_event_stop(struct perf_event *event, int flags)
439 {
440 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
441 struct hw_perf_event *hwc = &event->hw;
442 int idx = hwc->idx;
443
444 if (hwc->state & PERF_HES_STOPPED)
445 return;
446
447 smmu_pmu_counter_disable(smmu_pmu, idx);
448 /* As the counter gets updated on _start, ignore PERF_EF_UPDATE */
449 smmu_pmu_event_update(event);
450 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
451 }
452
smmu_pmu_event_add(struct perf_event * event,int flags)453 static int smmu_pmu_event_add(struct perf_event *event, int flags)
454 {
455 struct hw_perf_event *hwc = &event->hw;
456 int idx;
457 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
458
459 idx = smmu_pmu_get_event_idx(smmu_pmu, event);
460 if (idx < 0)
461 return idx;
462
463 hwc->idx = idx;
464 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
465 smmu_pmu->events[idx] = event;
466 local64_set(&hwc->prev_count, 0);
467
468 smmu_pmu_interrupt_enable(smmu_pmu, idx);
469
470 if (flags & PERF_EF_START)
471 smmu_pmu_event_start(event, flags);
472
473 /* Propagate changes to the userspace mapping. */
474 perf_event_update_userpage(event);
475
476 return 0;
477 }
478
smmu_pmu_event_del(struct perf_event * event,int flags)479 static void smmu_pmu_event_del(struct perf_event *event, int flags)
480 {
481 struct hw_perf_event *hwc = &event->hw;
482 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
483 int idx = hwc->idx;
484
485 smmu_pmu_event_stop(event, flags | PERF_EF_UPDATE);
486 smmu_pmu_interrupt_disable(smmu_pmu, idx);
487 smmu_pmu->events[idx] = NULL;
488 clear_bit(idx, smmu_pmu->used_counters);
489
490 perf_event_update_userpage(event);
491 }
492
smmu_pmu_event_read(struct perf_event * event)493 static void smmu_pmu_event_read(struct perf_event *event)
494 {
495 smmu_pmu_event_update(event);
496 }
497
498 /* cpumask */
499
smmu_pmu_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)500 static ssize_t smmu_pmu_cpumask_show(struct device *dev,
501 struct device_attribute *attr,
502 char *buf)
503 {
504 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
505
506 return cpumap_print_to_pagebuf(true, buf, cpumask_of(smmu_pmu->on_cpu));
507 }
508
509 static struct device_attribute smmu_pmu_cpumask_attr =
510 __ATTR(cpumask, 0444, smmu_pmu_cpumask_show, NULL);
511
512 static struct attribute *smmu_pmu_cpumask_attrs[] = {
513 &smmu_pmu_cpumask_attr.attr,
514 NULL
515 };
516
517 static const struct attribute_group smmu_pmu_cpumask_group = {
518 .attrs = smmu_pmu_cpumask_attrs,
519 };
520
521 /* Events */
522
smmu_pmu_event_show(struct device * dev,struct device_attribute * attr,char * page)523 static ssize_t smmu_pmu_event_show(struct device *dev,
524 struct device_attribute *attr, char *page)
525 {
526 struct perf_pmu_events_attr *pmu_attr;
527
528 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
529
530 return sysfs_emit(page, "event=0x%02llx\n", pmu_attr->id);
531 }
532
533 #define SMMU_EVENT_ATTR(name, config) \
534 PMU_EVENT_ATTR_ID(name, smmu_pmu_event_show, config)
535
536 static struct attribute *smmu_pmu_events[] = {
537 SMMU_EVENT_ATTR(cycles, 0),
538 SMMU_EVENT_ATTR(transaction, 1),
539 SMMU_EVENT_ATTR(tlb_miss, 2),
540 SMMU_EVENT_ATTR(config_cache_miss, 3),
541 SMMU_EVENT_ATTR(trans_table_walk_access, 4),
542 SMMU_EVENT_ATTR(config_struct_access, 5),
543 SMMU_EVENT_ATTR(pcie_ats_trans_rq, 6),
544 SMMU_EVENT_ATTR(pcie_ats_trans_passed, 7),
545 NULL
546 };
547
smmu_pmu_event_is_visible(struct kobject * kobj,struct attribute * attr,int unused)548 static umode_t smmu_pmu_event_is_visible(struct kobject *kobj,
549 struct attribute *attr, int unused)
550 {
551 struct device *dev = kobj_to_dev(kobj);
552 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
553 struct perf_pmu_events_attr *pmu_attr;
554
555 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
556
557 if (test_bit(pmu_attr->id, smmu_pmu->supported_events))
558 return attr->mode;
559
560 return 0;
561 }
562
563 static const struct attribute_group smmu_pmu_events_group = {
564 .name = "events",
565 .attrs = smmu_pmu_events,
566 .is_visible = smmu_pmu_event_is_visible,
567 };
568
smmu_pmu_identifier_attr_show(struct device * dev,struct device_attribute * attr,char * page)569 static ssize_t smmu_pmu_identifier_attr_show(struct device *dev,
570 struct device_attribute *attr,
571 char *page)
572 {
573 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
574
575 return sysfs_emit(page, "0x%08x\n", smmu_pmu->iidr);
576 }
577
smmu_pmu_identifier_attr_visible(struct kobject * kobj,struct attribute * attr,int n)578 static umode_t smmu_pmu_identifier_attr_visible(struct kobject *kobj,
579 struct attribute *attr,
580 int n)
581 {
582 struct device *dev = kobj_to_dev(kobj);
583 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
584
585 if (!smmu_pmu->iidr)
586 return 0;
587 return attr->mode;
588 }
589
590 static struct device_attribute smmu_pmu_identifier_attr =
591 __ATTR(identifier, 0444, smmu_pmu_identifier_attr_show, NULL);
592
593 static struct attribute *smmu_pmu_identifier_attrs[] = {
594 &smmu_pmu_identifier_attr.attr,
595 NULL
596 };
597
598 static const struct attribute_group smmu_pmu_identifier_group = {
599 .attrs = smmu_pmu_identifier_attrs,
600 .is_visible = smmu_pmu_identifier_attr_visible,
601 };
602
603 /* Formats */
604 PMU_FORMAT_ATTR(event, "config:0-15");
605 PMU_FORMAT_ATTR(filter_stream_id, "config1:0-31");
606 PMU_FORMAT_ATTR(filter_span, "config1:32");
607 PMU_FORMAT_ATTR(filter_enable, "config1:33");
608
609 static struct attribute *smmu_pmu_formats[] = {
610 &format_attr_event.attr,
611 &format_attr_filter_stream_id.attr,
612 &format_attr_filter_span.attr,
613 &format_attr_filter_enable.attr,
614 NULL
615 };
616
617 static const struct attribute_group smmu_pmu_format_group = {
618 .name = "format",
619 .attrs = smmu_pmu_formats,
620 };
621
622 static const struct attribute_group *smmu_pmu_attr_grps[] = {
623 &smmu_pmu_cpumask_group,
624 &smmu_pmu_events_group,
625 &smmu_pmu_format_group,
626 &smmu_pmu_identifier_group,
627 NULL
628 };
629
630 /*
631 * Generic device handlers
632 */
633
smmu_pmu_offline_cpu(unsigned int cpu,struct hlist_node * node)634 static int smmu_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
635 {
636 struct smmu_pmu *smmu_pmu;
637 unsigned int target;
638
639 smmu_pmu = hlist_entry_safe(node, struct smmu_pmu, node);
640 if (cpu != smmu_pmu->on_cpu)
641 return 0;
642
643 target = cpumask_any_but(cpu_online_mask, cpu);
644 if (target >= nr_cpu_ids)
645 return 0;
646
647 perf_pmu_migrate_context(&smmu_pmu->pmu, cpu, target);
648 smmu_pmu->on_cpu = target;
649 WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(target)));
650
651 return 0;
652 }
653
smmu_pmu_handle_irq(int irq_num,void * data)654 static irqreturn_t smmu_pmu_handle_irq(int irq_num, void *data)
655 {
656 struct smmu_pmu *smmu_pmu = data;
657 DECLARE_BITMAP(ovs, BITS_PER_TYPE(u64));
658 u64 ovsr;
659 unsigned int idx;
660
661 ovsr = readq(smmu_pmu->reloc_base + SMMU_PMCG_OVSSET0);
662 if (!ovsr)
663 return IRQ_NONE;
664
665 writeq(ovsr, smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
666
667 bitmap_from_u64(ovs, ovsr);
668 for_each_set_bit(idx, ovs, smmu_pmu->num_counters) {
669 struct perf_event *event = smmu_pmu->events[idx];
670 struct hw_perf_event *hwc;
671
672 if (WARN_ON_ONCE(!event))
673 continue;
674
675 smmu_pmu_event_update(event);
676 hwc = &event->hw;
677
678 smmu_pmu_set_period(smmu_pmu, hwc);
679 }
680
681 return IRQ_HANDLED;
682 }
683
smmu_pmu_free_msis(void * data)684 static void smmu_pmu_free_msis(void *data)
685 {
686 struct device *dev = data;
687
688 platform_msi_domain_free_irqs(dev);
689 }
690
smmu_pmu_write_msi_msg(struct msi_desc * desc,struct msi_msg * msg)691 static void smmu_pmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
692 {
693 phys_addr_t doorbell;
694 struct device *dev = msi_desc_to_dev(desc);
695 struct smmu_pmu *pmu = dev_get_drvdata(dev);
696
697 doorbell = (((u64)msg->address_hi) << 32) | msg->address_lo;
698 doorbell &= MSI_CFG0_ADDR_MASK;
699
700 writeq_relaxed(doorbell, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
701 writel_relaxed(msg->data, pmu->reg_base + SMMU_PMCG_IRQ_CFG1);
702 writel_relaxed(MSI_CFG2_MEMATTR_DEVICE_nGnRE,
703 pmu->reg_base + SMMU_PMCG_IRQ_CFG2);
704 }
705
smmu_pmu_setup_msi(struct smmu_pmu * pmu)706 static void smmu_pmu_setup_msi(struct smmu_pmu *pmu)
707 {
708 struct device *dev = pmu->dev;
709 int ret;
710
711 /* Clear MSI address reg */
712 writeq_relaxed(0, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
713
714 /* MSI supported or not */
715 if (!(readl(pmu->reg_base + SMMU_PMCG_CFGR) & SMMU_PMCG_CFGR_MSI))
716 return;
717
718 ret = platform_msi_domain_alloc_irqs(dev, 1, smmu_pmu_write_msi_msg);
719 if (ret) {
720 dev_warn(dev, "failed to allocate MSIs\n");
721 return;
722 }
723
724 pmu->irq = msi_get_virq(dev, 0);
725
726 /* Add callback to free MSIs on teardown */
727 devm_add_action(dev, smmu_pmu_free_msis, dev);
728 }
729
smmu_pmu_setup_irq(struct smmu_pmu * pmu)730 static int smmu_pmu_setup_irq(struct smmu_pmu *pmu)
731 {
732 unsigned long flags = IRQF_NOBALANCING | IRQF_SHARED | IRQF_NO_THREAD;
733 int irq, ret = -ENXIO;
734
735 smmu_pmu_setup_msi(pmu);
736
737 irq = pmu->irq;
738 if (irq)
739 ret = devm_request_irq(pmu->dev, irq, smmu_pmu_handle_irq,
740 flags, "smmuv3-pmu", pmu);
741 return ret;
742 }
743
smmu_pmu_reset(struct smmu_pmu * smmu_pmu)744 static void smmu_pmu_reset(struct smmu_pmu *smmu_pmu)
745 {
746 u64 counter_present_mask = GENMASK_ULL(smmu_pmu->num_counters - 1, 0);
747
748 smmu_pmu_disable(&smmu_pmu->pmu);
749
750 /* Disable counter and interrupt */
751 writeq_relaxed(counter_present_mask,
752 smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
753 writeq_relaxed(counter_present_mask,
754 smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
755 writeq_relaxed(counter_present_mask,
756 smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
757 }
758
smmu_pmu_get_acpi_options(struct smmu_pmu * smmu_pmu)759 static void smmu_pmu_get_acpi_options(struct smmu_pmu *smmu_pmu)
760 {
761 u32 model;
762
763 model = *(u32 *)dev_get_platdata(smmu_pmu->dev);
764
765 switch (model) {
766 case IORT_SMMU_V3_PMCG_HISI_HIP08:
767 /* HiSilicon Erratum 162001800 */
768 smmu_pmu->options |= SMMU_PMCG_EVCNTR_RDONLY;
769 break;
770 }
771
772 dev_notice(smmu_pmu->dev, "option mask 0x%x\n", smmu_pmu->options);
773 }
774
smmu_pmu_coresight_id_regs(struct smmu_pmu * smmu_pmu)775 static bool smmu_pmu_coresight_id_regs(struct smmu_pmu *smmu_pmu)
776 {
777 return of_device_is_compatible(smmu_pmu->dev->of_node,
778 "arm,mmu-600-pmcg");
779 }
780
smmu_pmu_get_iidr(struct smmu_pmu * smmu_pmu)781 static void smmu_pmu_get_iidr(struct smmu_pmu *smmu_pmu)
782 {
783 u32 iidr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_IIDR);
784
785 if (!iidr && smmu_pmu_coresight_id_regs(smmu_pmu)) {
786 u32 pidr0 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR0);
787 u32 pidr1 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR1);
788 u32 pidr2 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR2);
789 u32 pidr3 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR3);
790 u32 pidr4 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR4);
791
792 u32 productid = FIELD_GET(SMMU_PMCG_PIDR0_PART_0, pidr0) |
793 (FIELD_GET(SMMU_PMCG_PIDR1_PART_1, pidr1) << 8);
794 u32 variant = FIELD_GET(SMMU_PMCG_PIDR2_REVISION, pidr2);
795 u32 revision = FIELD_GET(SMMU_PMCG_PIDR3_REVAND, pidr3);
796 u32 implementer =
797 FIELD_GET(SMMU_PMCG_PIDR1_DES_0, pidr1) |
798 (FIELD_GET(SMMU_PMCG_PIDR2_DES_1, pidr2) << 4) |
799 (FIELD_GET(SMMU_PMCG_PIDR4_DES_2, pidr4) << 8);
800
801 iidr = FIELD_PREP(SMMU_PMCG_IIDR_PRODUCTID, productid) |
802 FIELD_PREP(SMMU_PMCG_IIDR_VARIANT, variant) |
803 FIELD_PREP(SMMU_PMCG_IIDR_REVISION, revision) |
804 FIELD_PREP(SMMU_PMCG_IIDR_IMPLEMENTER, implementer);
805 }
806
807 smmu_pmu->iidr = iidr;
808 }
809
smmu_pmu_probe(struct platform_device * pdev)810 static int smmu_pmu_probe(struct platform_device *pdev)
811 {
812 struct smmu_pmu *smmu_pmu;
813 struct resource *res_0;
814 u32 cfgr, reg_size;
815 u64 ceid_64[2];
816 int irq, err;
817 char *name;
818 struct device *dev = &pdev->dev;
819
820 smmu_pmu = devm_kzalloc(dev, sizeof(*smmu_pmu), GFP_KERNEL);
821 if (!smmu_pmu)
822 return -ENOMEM;
823
824 smmu_pmu->dev = dev;
825 platform_set_drvdata(pdev, smmu_pmu);
826
827 smmu_pmu->pmu = (struct pmu) {
828 .module = THIS_MODULE,
829 .task_ctx_nr = perf_invalid_context,
830 .pmu_enable = smmu_pmu_enable,
831 .pmu_disable = smmu_pmu_disable,
832 .event_init = smmu_pmu_event_init,
833 .add = smmu_pmu_event_add,
834 .del = smmu_pmu_event_del,
835 .start = smmu_pmu_event_start,
836 .stop = smmu_pmu_event_stop,
837 .read = smmu_pmu_event_read,
838 .attr_groups = smmu_pmu_attr_grps,
839 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
840 };
841
842 smmu_pmu->reg_base = devm_platform_get_and_ioremap_resource(pdev, 0, &res_0);
843 if (IS_ERR(smmu_pmu->reg_base))
844 return PTR_ERR(smmu_pmu->reg_base);
845
846 cfgr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CFGR);
847
848 /* Determine if page 1 is present */
849 if (cfgr & SMMU_PMCG_CFGR_RELOC_CTRS) {
850 smmu_pmu->reloc_base = devm_platform_ioremap_resource(pdev, 1);
851 if (IS_ERR(smmu_pmu->reloc_base))
852 return PTR_ERR(smmu_pmu->reloc_base);
853 } else {
854 smmu_pmu->reloc_base = smmu_pmu->reg_base;
855 }
856
857 irq = platform_get_irq_optional(pdev, 0);
858 if (irq > 0)
859 smmu_pmu->irq = irq;
860
861 ceid_64[0] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID0);
862 ceid_64[1] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID1);
863 bitmap_from_arr32(smmu_pmu->supported_events, (u32 *)ceid_64,
864 SMMU_PMCG_ARCH_MAX_EVENTS);
865
866 smmu_pmu->num_counters = FIELD_GET(SMMU_PMCG_CFGR_NCTR, cfgr) + 1;
867
868 smmu_pmu->global_filter = !!(cfgr & SMMU_PMCG_CFGR_SID_FILTER_TYPE);
869
870 reg_size = FIELD_GET(SMMU_PMCG_CFGR_SIZE, cfgr);
871 smmu_pmu->counter_mask = GENMASK_ULL(reg_size, 0);
872
873 smmu_pmu_reset(smmu_pmu);
874
875 err = smmu_pmu_setup_irq(smmu_pmu);
876 if (err) {
877 dev_err(dev, "Setup irq failed, PMU @%pa\n", &res_0->start);
878 return err;
879 }
880
881 smmu_pmu_get_iidr(smmu_pmu);
882
883 name = devm_kasprintf(&pdev->dev, GFP_KERNEL, "smmuv3_pmcg_%llx",
884 (res_0->start) >> SMMU_PMCG_PA_SHIFT);
885 if (!name) {
886 dev_err(dev, "Create name failed, PMU @%pa\n", &res_0->start);
887 return -EINVAL;
888 }
889
890 if (!dev->of_node)
891 smmu_pmu_get_acpi_options(smmu_pmu);
892
893 /* Pick one CPU to be the preferred one to use */
894 smmu_pmu->on_cpu = raw_smp_processor_id();
895 WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(smmu_pmu->on_cpu)));
896
897 err = cpuhp_state_add_instance_nocalls(cpuhp_state_num,
898 &smmu_pmu->node);
899 if (err) {
900 dev_err(dev, "Error %d registering hotplug, PMU @%pa\n",
901 err, &res_0->start);
902 return err;
903 }
904
905 err = perf_pmu_register(&smmu_pmu->pmu, name, -1);
906 if (err) {
907 dev_err(dev, "Error %d registering PMU @%pa\n",
908 err, &res_0->start);
909 goto out_unregister;
910 }
911
912 dev_info(dev, "Registered PMU @ %pa using %d counters with %s filter settings\n",
913 &res_0->start, smmu_pmu->num_counters,
914 smmu_pmu->global_filter ? "Global(Counter0)" :
915 "Individual");
916
917 return 0;
918
919 out_unregister:
920 cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
921 return err;
922 }
923
smmu_pmu_remove(struct platform_device * pdev)924 static int smmu_pmu_remove(struct platform_device *pdev)
925 {
926 struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
927
928 perf_pmu_unregister(&smmu_pmu->pmu);
929 cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
930
931 return 0;
932 }
933
smmu_pmu_shutdown(struct platform_device * pdev)934 static void smmu_pmu_shutdown(struct platform_device *pdev)
935 {
936 struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
937
938 smmu_pmu_disable(&smmu_pmu->pmu);
939 }
940
941 #ifdef CONFIG_OF
942 static const struct of_device_id smmu_pmu_of_match[] = {
943 { .compatible = "arm,smmu-v3-pmcg" },
944 {}
945 };
946 MODULE_DEVICE_TABLE(of, smmu_pmu_of_match);
947 #endif
948
949 static struct platform_driver smmu_pmu_driver = {
950 .driver = {
951 .name = "arm-smmu-v3-pmcg",
952 .of_match_table = of_match_ptr(smmu_pmu_of_match),
953 .suppress_bind_attrs = true,
954 },
955 .probe = smmu_pmu_probe,
956 .remove = smmu_pmu_remove,
957 .shutdown = smmu_pmu_shutdown,
958 };
959
arm_smmu_pmu_init(void)960 static int __init arm_smmu_pmu_init(void)
961 {
962 int ret;
963
964 cpuhp_state_num = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
965 "perf/arm/pmcg:online",
966 NULL,
967 smmu_pmu_offline_cpu);
968 if (cpuhp_state_num < 0)
969 return cpuhp_state_num;
970
971 ret = platform_driver_register(&smmu_pmu_driver);
972 if (ret)
973 cpuhp_remove_multi_state(cpuhp_state_num);
974
975 return ret;
976 }
977 module_init(arm_smmu_pmu_init);
978
arm_smmu_pmu_exit(void)979 static void __exit arm_smmu_pmu_exit(void)
980 {
981 platform_driver_unregister(&smmu_pmu_driver);
982 cpuhp_remove_multi_state(cpuhp_state_num);
983 }
984
985 module_exit(arm_smmu_pmu_exit);
986
987 MODULE_DESCRIPTION("PMU driver for ARM SMMUv3 Performance Monitors Extension");
988 MODULE_AUTHOR("Neil Leeder <nleeder@codeaurora.org>");
989 MODULE_AUTHOR("Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>");
990 MODULE_LICENSE("GPL v2");
991