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 #define SMMU_PMCG_HARDEN_DISABLE BIT(1)
119
120 static int cpuhp_state_num;
121
122 struct smmu_pmu {
123 struct hlist_node node;
124 struct perf_event *events[SMMU_PMCG_MAX_COUNTERS];
125 DECLARE_BITMAP(used_counters, SMMU_PMCG_MAX_COUNTERS);
126 DECLARE_BITMAP(supported_events, SMMU_PMCG_ARCH_MAX_EVENTS);
127 unsigned int irq;
128 unsigned int on_cpu;
129 struct pmu pmu;
130 unsigned int num_counters;
131 struct device *dev;
132 void __iomem *reg_base;
133 void __iomem *reloc_base;
134 u64 counter_mask;
135 u32 options;
136 u32 iidr;
137 bool global_filter;
138 };
139
140 #define to_smmu_pmu(p) (container_of(p, struct smmu_pmu, pmu))
141
142 #define SMMU_PMU_EVENT_ATTR_EXTRACTOR(_name, _config, _start, _end) \
143 static inline u32 get_##_name(struct perf_event *event) \
144 { \
145 return FIELD_GET(GENMASK_ULL(_end, _start), \
146 event->attr._config); \
147 } \
148
149 SMMU_PMU_EVENT_ATTR_EXTRACTOR(event, config, 0, 15);
150 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_stream_id, config1, 0, 31);
151 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_span, config1, 32, 32);
152 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_enable, config1, 33, 33);
153
smmu_pmu_enable(struct pmu * pmu)154 static inline void smmu_pmu_enable(struct pmu *pmu)
155 {
156 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
157
158 writel(SMMU_PMCG_IRQ_CTRL_IRQEN,
159 smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
160 writel(SMMU_PMCG_CR_ENABLE, smmu_pmu->reg_base + SMMU_PMCG_CR);
161 }
162
163 static int smmu_pmu_apply_event_filter(struct smmu_pmu *smmu_pmu,
164 struct perf_event *event, int idx);
165
smmu_pmu_enable_quirk_hip08_09(struct pmu * pmu)166 static inline void smmu_pmu_enable_quirk_hip08_09(struct pmu *pmu)
167 {
168 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
169 unsigned int idx;
170
171 for_each_set_bit(idx, smmu_pmu->used_counters, smmu_pmu->num_counters)
172 smmu_pmu_apply_event_filter(smmu_pmu, smmu_pmu->events[idx], idx);
173
174 smmu_pmu_enable(pmu);
175 }
176
smmu_pmu_disable(struct pmu * pmu)177 static inline void smmu_pmu_disable(struct pmu *pmu)
178 {
179 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
180
181 writel(0, smmu_pmu->reg_base + SMMU_PMCG_CR);
182 writel(0, smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
183 }
184
smmu_pmu_disable_quirk_hip08_09(struct pmu * pmu)185 static inline void smmu_pmu_disable_quirk_hip08_09(struct pmu *pmu)
186 {
187 struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
188 unsigned int idx;
189
190 /*
191 * The global disable of PMU sometimes fail to stop the counting.
192 * Harden this by writing an invalid event type to each used counter
193 * to forcibly stop counting.
194 */
195 for_each_set_bit(idx, smmu_pmu->used_counters, smmu_pmu->num_counters)
196 writel(0xffff, smmu_pmu->reg_base + SMMU_PMCG_EVTYPER(idx));
197
198 smmu_pmu_disable(pmu);
199 }
200
smmu_pmu_counter_set_value(struct smmu_pmu * smmu_pmu,u32 idx,u64 value)201 static inline void smmu_pmu_counter_set_value(struct smmu_pmu *smmu_pmu,
202 u32 idx, u64 value)
203 {
204 if (smmu_pmu->counter_mask & BIT(32))
205 writeq(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
206 else
207 writel(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
208 }
209
smmu_pmu_counter_get_value(struct smmu_pmu * smmu_pmu,u32 idx)210 static inline u64 smmu_pmu_counter_get_value(struct smmu_pmu *smmu_pmu, u32 idx)
211 {
212 u64 value;
213
214 if (smmu_pmu->counter_mask & BIT(32))
215 value = readq(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
216 else
217 value = readl(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
218
219 return value;
220 }
221
smmu_pmu_counter_enable(struct smmu_pmu * smmu_pmu,u32 idx)222 static inline void smmu_pmu_counter_enable(struct smmu_pmu *smmu_pmu, u32 idx)
223 {
224 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENSET0);
225 }
226
smmu_pmu_counter_disable(struct smmu_pmu * smmu_pmu,u32 idx)227 static inline void smmu_pmu_counter_disable(struct smmu_pmu *smmu_pmu, u32 idx)
228 {
229 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
230 }
231
smmu_pmu_interrupt_enable(struct smmu_pmu * smmu_pmu,u32 idx)232 static inline void smmu_pmu_interrupt_enable(struct smmu_pmu *smmu_pmu, u32 idx)
233 {
234 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENSET0);
235 }
236
smmu_pmu_interrupt_disable(struct smmu_pmu * smmu_pmu,u32 idx)237 static inline void smmu_pmu_interrupt_disable(struct smmu_pmu *smmu_pmu,
238 u32 idx)
239 {
240 writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
241 }
242
smmu_pmu_set_evtyper(struct smmu_pmu * smmu_pmu,u32 idx,u32 val)243 static inline void smmu_pmu_set_evtyper(struct smmu_pmu *smmu_pmu, u32 idx,
244 u32 val)
245 {
246 writel(val, smmu_pmu->reg_base + SMMU_PMCG_EVTYPER(idx));
247 }
248
smmu_pmu_set_smr(struct smmu_pmu * smmu_pmu,u32 idx,u32 val)249 static inline void smmu_pmu_set_smr(struct smmu_pmu *smmu_pmu, u32 idx, u32 val)
250 {
251 writel(val, smmu_pmu->reg_base + SMMU_PMCG_SMR(idx));
252 }
253
smmu_pmu_event_update(struct perf_event * event)254 static void smmu_pmu_event_update(struct perf_event *event)
255 {
256 struct hw_perf_event *hwc = &event->hw;
257 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
258 u64 delta, prev, now;
259 u32 idx = hwc->idx;
260
261 do {
262 prev = local64_read(&hwc->prev_count);
263 now = smmu_pmu_counter_get_value(smmu_pmu, idx);
264 } while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
265
266 /* handle overflow. */
267 delta = now - prev;
268 delta &= smmu_pmu->counter_mask;
269
270 local64_add(delta, &event->count);
271 }
272
smmu_pmu_set_period(struct smmu_pmu * smmu_pmu,struct hw_perf_event * hwc)273 static void smmu_pmu_set_period(struct smmu_pmu *smmu_pmu,
274 struct hw_perf_event *hwc)
275 {
276 u32 idx = hwc->idx;
277 u64 new;
278
279 if (smmu_pmu->options & SMMU_PMCG_EVCNTR_RDONLY) {
280 /*
281 * On platforms that require this quirk, if the counter starts
282 * at < half_counter value and wraps, the current logic of
283 * handling the overflow may not work. It is expected that,
284 * those platforms will have full 64 counter bits implemented
285 * so that such a possibility is remote(eg: HiSilicon HIP08).
286 */
287 new = smmu_pmu_counter_get_value(smmu_pmu, idx);
288 } else {
289 /*
290 * We limit the max period to half the max counter value
291 * of the counter size, so that even in the case of extreme
292 * interrupt latency the counter will (hopefully) not wrap
293 * past its initial value.
294 */
295 new = smmu_pmu->counter_mask >> 1;
296 smmu_pmu_counter_set_value(smmu_pmu, idx, new);
297 }
298
299 local64_set(&hwc->prev_count, new);
300 }
301
smmu_pmu_set_event_filter(struct perf_event * event,int idx,u32 span,u32 sid)302 static void smmu_pmu_set_event_filter(struct perf_event *event,
303 int idx, u32 span, u32 sid)
304 {
305 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
306 u32 evtyper;
307
308 evtyper = get_event(event) | span << SMMU_PMCG_SID_SPAN_SHIFT;
309 smmu_pmu_set_evtyper(smmu_pmu, idx, evtyper);
310 smmu_pmu_set_smr(smmu_pmu, idx, sid);
311 }
312
smmu_pmu_check_global_filter(struct perf_event * curr,struct perf_event * new)313 static bool smmu_pmu_check_global_filter(struct perf_event *curr,
314 struct perf_event *new)
315 {
316 if (get_filter_enable(new) != get_filter_enable(curr))
317 return false;
318
319 if (!get_filter_enable(new))
320 return true;
321
322 return get_filter_span(new) == get_filter_span(curr) &&
323 get_filter_stream_id(new) == get_filter_stream_id(curr);
324 }
325
smmu_pmu_apply_event_filter(struct smmu_pmu * smmu_pmu,struct perf_event * event,int idx)326 static int smmu_pmu_apply_event_filter(struct smmu_pmu *smmu_pmu,
327 struct perf_event *event, int idx)
328 {
329 u32 span, sid;
330 unsigned int cur_idx, num_ctrs = smmu_pmu->num_counters;
331 bool filter_en = !!get_filter_enable(event);
332
333 span = filter_en ? get_filter_span(event) :
334 SMMU_PMCG_DEFAULT_FILTER_SPAN;
335 sid = filter_en ? get_filter_stream_id(event) :
336 SMMU_PMCG_DEFAULT_FILTER_SID;
337
338 cur_idx = find_first_bit(smmu_pmu->used_counters, num_ctrs);
339 /*
340 * Per-counter filtering, or scheduling the first globally-filtered
341 * event into an empty PMU so idx == 0 and it works out equivalent.
342 */
343 if (!smmu_pmu->global_filter || cur_idx == num_ctrs) {
344 smmu_pmu_set_event_filter(event, idx, span, sid);
345 return 0;
346 }
347
348 /* Otherwise, must match whatever's currently scheduled */
349 if (smmu_pmu_check_global_filter(smmu_pmu->events[cur_idx], event)) {
350 smmu_pmu_set_evtyper(smmu_pmu, idx, get_event(event));
351 return 0;
352 }
353
354 return -EAGAIN;
355 }
356
smmu_pmu_get_event_idx(struct smmu_pmu * smmu_pmu,struct perf_event * event)357 static int smmu_pmu_get_event_idx(struct smmu_pmu *smmu_pmu,
358 struct perf_event *event)
359 {
360 int idx, err;
361 unsigned int num_ctrs = smmu_pmu->num_counters;
362
363 idx = find_first_zero_bit(smmu_pmu->used_counters, num_ctrs);
364 if (idx == num_ctrs)
365 /* The counters are all in use. */
366 return -EAGAIN;
367
368 err = smmu_pmu_apply_event_filter(smmu_pmu, event, idx);
369 if (err)
370 return err;
371
372 set_bit(idx, smmu_pmu->used_counters);
373
374 return idx;
375 }
376
smmu_pmu_events_compatible(struct perf_event * curr,struct perf_event * new)377 static bool smmu_pmu_events_compatible(struct perf_event *curr,
378 struct perf_event *new)
379 {
380 if (new->pmu != curr->pmu)
381 return false;
382
383 if (to_smmu_pmu(new->pmu)->global_filter &&
384 !smmu_pmu_check_global_filter(curr, new))
385 return false;
386
387 return true;
388 }
389
390 /*
391 * Implementation of abstract pmu functionality required by
392 * the core perf events code.
393 */
394
smmu_pmu_event_init(struct perf_event * event)395 static int smmu_pmu_event_init(struct perf_event *event)
396 {
397 struct hw_perf_event *hwc = &event->hw;
398 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
399 struct device *dev = smmu_pmu->dev;
400 struct perf_event *sibling;
401 int group_num_events = 1;
402 u16 event_id;
403
404 if (event->attr.type != event->pmu->type)
405 return -ENOENT;
406
407 if (hwc->sample_period) {
408 dev_dbg(dev, "Sampling not supported\n");
409 return -EOPNOTSUPP;
410 }
411
412 if (event->cpu < 0) {
413 dev_dbg(dev, "Per-task mode not supported\n");
414 return -EOPNOTSUPP;
415 }
416
417 /* Verify specified event is supported on this PMU */
418 event_id = get_event(event);
419 if (event_id < SMMU_PMCG_ARCH_MAX_EVENTS &&
420 (!test_bit(event_id, smmu_pmu->supported_events))) {
421 dev_dbg(dev, "Invalid event %d for this PMU\n", event_id);
422 return -EINVAL;
423 }
424
425 /* Don't allow groups with mixed PMUs, except for s/w events */
426 if (!is_software_event(event->group_leader)) {
427 if (!smmu_pmu_events_compatible(event->group_leader, event))
428 return -EINVAL;
429
430 if (++group_num_events > smmu_pmu->num_counters)
431 return -EINVAL;
432 }
433
434 /*
435 * Ensure all events are on the same cpu so all events are in the
436 * same cpu context, to avoid races on pmu_enable etc.
437 */
438 event->cpu = smmu_pmu->on_cpu;
439
440 hwc->idx = -1;
441
442 if (event->group_leader == event)
443 return 0;
444
445 for_each_sibling_event(sibling, event->group_leader) {
446 if (is_software_event(sibling))
447 continue;
448
449 if (!smmu_pmu_events_compatible(sibling, event))
450 return -EINVAL;
451
452 if (++group_num_events > smmu_pmu->num_counters)
453 return -EINVAL;
454 }
455
456 return 0;
457 }
458
smmu_pmu_event_start(struct perf_event * event,int flags)459 static void smmu_pmu_event_start(struct perf_event *event, int flags)
460 {
461 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
462 struct hw_perf_event *hwc = &event->hw;
463 int idx = hwc->idx;
464
465 hwc->state = 0;
466
467 smmu_pmu_set_period(smmu_pmu, hwc);
468
469 smmu_pmu_counter_enable(smmu_pmu, idx);
470 }
471
smmu_pmu_event_stop(struct perf_event * event,int flags)472 static void smmu_pmu_event_stop(struct perf_event *event, int flags)
473 {
474 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
475 struct hw_perf_event *hwc = &event->hw;
476 int idx = hwc->idx;
477
478 if (hwc->state & PERF_HES_STOPPED)
479 return;
480
481 smmu_pmu_counter_disable(smmu_pmu, idx);
482 /* As the counter gets updated on _start, ignore PERF_EF_UPDATE */
483 smmu_pmu_event_update(event);
484 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
485 }
486
smmu_pmu_event_add(struct perf_event * event,int flags)487 static int smmu_pmu_event_add(struct perf_event *event, int flags)
488 {
489 struct hw_perf_event *hwc = &event->hw;
490 int idx;
491 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
492
493 idx = smmu_pmu_get_event_idx(smmu_pmu, event);
494 if (idx < 0)
495 return idx;
496
497 hwc->idx = idx;
498 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
499 smmu_pmu->events[idx] = event;
500 local64_set(&hwc->prev_count, 0);
501
502 smmu_pmu_interrupt_enable(smmu_pmu, idx);
503
504 if (flags & PERF_EF_START)
505 smmu_pmu_event_start(event, flags);
506
507 /* Propagate changes to the userspace mapping. */
508 perf_event_update_userpage(event);
509
510 return 0;
511 }
512
smmu_pmu_event_del(struct perf_event * event,int flags)513 static void smmu_pmu_event_del(struct perf_event *event, int flags)
514 {
515 struct hw_perf_event *hwc = &event->hw;
516 struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
517 int idx = hwc->idx;
518
519 smmu_pmu_event_stop(event, flags | PERF_EF_UPDATE);
520 smmu_pmu_interrupt_disable(smmu_pmu, idx);
521 smmu_pmu->events[idx] = NULL;
522 clear_bit(idx, smmu_pmu->used_counters);
523
524 perf_event_update_userpage(event);
525 }
526
smmu_pmu_event_read(struct perf_event * event)527 static void smmu_pmu_event_read(struct perf_event *event)
528 {
529 smmu_pmu_event_update(event);
530 }
531
532 /* cpumask */
533
smmu_pmu_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)534 static ssize_t smmu_pmu_cpumask_show(struct device *dev,
535 struct device_attribute *attr,
536 char *buf)
537 {
538 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
539
540 return cpumap_print_to_pagebuf(true, buf, cpumask_of(smmu_pmu->on_cpu));
541 }
542
543 static struct device_attribute smmu_pmu_cpumask_attr =
544 __ATTR(cpumask, 0444, smmu_pmu_cpumask_show, NULL);
545
546 static struct attribute *smmu_pmu_cpumask_attrs[] = {
547 &smmu_pmu_cpumask_attr.attr,
548 NULL
549 };
550
551 static const struct attribute_group smmu_pmu_cpumask_group = {
552 .attrs = smmu_pmu_cpumask_attrs,
553 };
554
555 /* Events */
556
smmu_pmu_event_show(struct device * dev,struct device_attribute * attr,char * page)557 static ssize_t smmu_pmu_event_show(struct device *dev,
558 struct device_attribute *attr, char *page)
559 {
560 struct perf_pmu_events_attr *pmu_attr;
561
562 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
563
564 return sysfs_emit(page, "event=0x%02llx\n", pmu_attr->id);
565 }
566
567 #define SMMU_EVENT_ATTR(name, config) \
568 PMU_EVENT_ATTR_ID(name, smmu_pmu_event_show, config)
569
570 static struct attribute *smmu_pmu_events[] = {
571 SMMU_EVENT_ATTR(cycles, 0),
572 SMMU_EVENT_ATTR(transaction, 1),
573 SMMU_EVENT_ATTR(tlb_miss, 2),
574 SMMU_EVENT_ATTR(config_cache_miss, 3),
575 SMMU_EVENT_ATTR(trans_table_walk_access, 4),
576 SMMU_EVENT_ATTR(config_struct_access, 5),
577 SMMU_EVENT_ATTR(pcie_ats_trans_rq, 6),
578 SMMU_EVENT_ATTR(pcie_ats_trans_passed, 7),
579 NULL
580 };
581
smmu_pmu_event_is_visible(struct kobject * kobj,struct attribute * attr,int unused)582 static umode_t smmu_pmu_event_is_visible(struct kobject *kobj,
583 struct attribute *attr, int unused)
584 {
585 struct device *dev = kobj_to_dev(kobj);
586 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
587 struct perf_pmu_events_attr *pmu_attr;
588
589 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
590
591 if (test_bit(pmu_attr->id, smmu_pmu->supported_events))
592 return attr->mode;
593
594 return 0;
595 }
596
597 static const struct attribute_group smmu_pmu_events_group = {
598 .name = "events",
599 .attrs = smmu_pmu_events,
600 .is_visible = smmu_pmu_event_is_visible,
601 };
602
smmu_pmu_identifier_attr_show(struct device * dev,struct device_attribute * attr,char * page)603 static ssize_t smmu_pmu_identifier_attr_show(struct device *dev,
604 struct device_attribute *attr,
605 char *page)
606 {
607 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
608
609 return sysfs_emit(page, "0x%08x\n", smmu_pmu->iidr);
610 }
611
smmu_pmu_identifier_attr_visible(struct kobject * kobj,struct attribute * attr,int n)612 static umode_t smmu_pmu_identifier_attr_visible(struct kobject *kobj,
613 struct attribute *attr,
614 int n)
615 {
616 struct device *dev = kobj_to_dev(kobj);
617 struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
618
619 if (!smmu_pmu->iidr)
620 return 0;
621 return attr->mode;
622 }
623
624 static struct device_attribute smmu_pmu_identifier_attr =
625 __ATTR(identifier, 0444, smmu_pmu_identifier_attr_show, NULL);
626
627 static struct attribute *smmu_pmu_identifier_attrs[] = {
628 &smmu_pmu_identifier_attr.attr,
629 NULL
630 };
631
632 static const struct attribute_group smmu_pmu_identifier_group = {
633 .attrs = smmu_pmu_identifier_attrs,
634 .is_visible = smmu_pmu_identifier_attr_visible,
635 };
636
637 /* Formats */
638 PMU_FORMAT_ATTR(event, "config:0-15");
639 PMU_FORMAT_ATTR(filter_stream_id, "config1:0-31");
640 PMU_FORMAT_ATTR(filter_span, "config1:32");
641 PMU_FORMAT_ATTR(filter_enable, "config1:33");
642
643 static struct attribute *smmu_pmu_formats[] = {
644 &format_attr_event.attr,
645 &format_attr_filter_stream_id.attr,
646 &format_attr_filter_span.attr,
647 &format_attr_filter_enable.attr,
648 NULL
649 };
650
651 static const struct attribute_group smmu_pmu_format_group = {
652 .name = "format",
653 .attrs = smmu_pmu_formats,
654 };
655
656 static const struct attribute_group *smmu_pmu_attr_grps[] = {
657 &smmu_pmu_cpumask_group,
658 &smmu_pmu_events_group,
659 &smmu_pmu_format_group,
660 &smmu_pmu_identifier_group,
661 NULL
662 };
663
664 /*
665 * Generic device handlers
666 */
667
smmu_pmu_offline_cpu(unsigned int cpu,struct hlist_node * node)668 static int smmu_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
669 {
670 struct smmu_pmu *smmu_pmu;
671 unsigned int target;
672
673 smmu_pmu = hlist_entry_safe(node, struct smmu_pmu, node);
674 if (cpu != smmu_pmu->on_cpu)
675 return 0;
676
677 target = cpumask_any_but(cpu_online_mask, cpu);
678 if (target >= nr_cpu_ids)
679 return 0;
680
681 perf_pmu_migrate_context(&smmu_pmu->pmu, cpu, target);
682 smmu_pmu->on_cpu = target;
683 WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(target)));
684
685 return 0;
686 }
687
smmu_pmu_handle_irq(int irq_num,void * data)688 static irqreturn_t smmu_pmu_handle_irq(int irq_num, void *data)
689 {
690 struct smmu_pmu *smmu_pmu = data;
691 DECLARE_BITMAP(ovs, BITS_PER_TYPE(u64));
692 u64 ovsr;
693 unsigned int idx;
694
695 ovsr = readq(smmu_pmu->reloc_base + SMMU_PMCG_OVSSET0);
696 if (!ovsr)
697 return IRQ_NONE;
698
699 writeq(ovsr, smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
700
701 bitmap_from_u64(ovs, ovsr);
702 for_each_set_bit(idx, ovs, smmu_pmu->num_counters) {
703 struct perf_event *event = smmu_pmu->events[idx];
704 struct hw_perf_event *hwc;
705
706 if (WARN_ON_ONCE(!event))
707 continue;
708
709 smmu_pmu_event_update(event);
710 hwc = &event->hw;
711
712 smmu_pmu_set_period(smmu_pmu, hwc);
713 }
714
715 return IRQ_HANDLED;
716 }
717
smmu_pmu_free_msis(void * data)718 static void smmu_pmu_free_msis(void *data)
719 {
720 struct device *dev = data;
721
722 platform_device_msi_free_irqs_all(dev);
723 }
724
smmu_pmu_write_msi_msg(struct msi_desc * desc,struct msi_msg * msg)725 static void smmu_pmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
726 {
727 phys_addr_t doorbell;
728 struct device *dev = msi_desc_to_dev(desc);
729 struct smmu_pmu *pmu = dev_get_drvdata(dev);
730
731 doorbell = (((u64)msg->address_hi) << 32) | msg->address_lo;
732 doorbell &= MSI_CFG0_ADDR_MASK;
733
734 writeq_relaxed(doorbell, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
735 writel_relaxed(msg->data, pmu->reg_base + SMMU_PMCG_IRQ_CFG1);
736 writel_relaxed(MSI_CFG2_MEMATTR_DEVICE_nGnRE,
737 pmu->reg_base + SMMU_PMCG_IRQ_CFG2);
738 }
739
smmu_pmu_setup_msi(struct smmu_pmu * pmu)740 static void smmu_pmu_setup_msi(struct smmu_pmu *pmu)
741 {
742 struct device *dev = pmu->dev;
743 int ret;
744
745 /* Clear MSI address reg */
746 writeq_relaxed(0, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
747
748 /* MSI supported or not */
749 if (!(readl(pmu->reg_base + SMMU_PMCG_CFGR) & SMMU_PMCG_CFGR_MSI))
750 return;
751
752 ret = platform_device_msi_init_and_alloc_irqs(dev, 1, smmu_pmu_write_msi_msg);
753 if (ret) {
754 dev_warn(dev, "failed to allocate MSIs\n");
755 return;
756 }
757
758 pmu->irq = msi_get_virq(dev, 0);
759
760 /* Add callback to free MSIs on teardown */
761 devm_add_action(dev, smmu_pmu_free_msis, dev);
762 }
763
smmu_pmu_setup_irq(struct smmu_pmu * pmu)764 static int smmu_pmu_setup_irq(struct smmu_pmu *pmu)
765 {
766 unsigned long flags = IRQF_NOBALANCING | IRQF_SHARED | IRQF_NO_THREAD;
767 int irq, ret = -ENXIO;
768
769 smmu_pmu_setup_msi(pmu);
770
771 irq = pmu->irq;
772 if (irq)
773 ret = devm_request_irq(pmu->dev, irq, smmu_pmu_handle_irq,
774 flags, "smmuv3-pmu", pmu);
775 return ret;
776 }
777
smmu_pmu_reset(struct smmu_pmu * smmu_pmu)778 static void smmu_pmu_reset(struct smmu_pmu *smmu_pmu)
779 {
780 u64 counter_present_mask = GENMASK_ULL(smmu_pmu->num_counters - 1, 0);
781
782 smmu_pmu_disable(&smmu_pmu->pmu);
783
784 /* Disable counter and interrupt */
785 writeq_relaxed(counter_present_mask,
786 smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
787 writeq_relaxed(counter_present_mask,
788 smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
789 writeq_relaxed(counter_present_mask,
790 smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
791 }
792
smmu_pmu_get_acpi_options(struct smmu_pmu * smmu_pmu)793 static void smmu_pmu_get_acpi_options(struct smmu_pmu *smmu_pmu)
794 {
795 u32 model;
796
797 model = *(u32 *)dev_get_platdata(smmu_pmu->dev);
798
799 switch (model) {
800 case IORT_SMMU_V3_PMCG_HISI_HIP08:
801 /* HiSilicon Erratum 162001800 */
802 smmu_pmu->options |= SMMU_PMCG_EVCNTR_RDONLY | SMMU_PMCG_HARDEN_DISABLE;
803 break;
804 case IORT_SMMU_V3_PMCG_HISI_HIP09:
805 smmu_pmu->options |= SMMU_PMCG_HARDEN_DISABLE;
806 break;
807 }
808
809 dev_notice(smmu_pmu->dev, "option mask 0x%x\n", smmu_pmu->options);
810 }
811
smmu_pmu_coresight_id_regs(struct smmu_pmu * smmu_pmu)812 static bool smmu_pmu_coresight_id_regs(struct smmu_pmu *smmu_pmu)
813 {
814 return of_device_is_compatible(smmu_pmu->dev->of_node,
815 "arm,mmu-600-pmcg");
816 }
817
smmu_pmu_get_iidr(struct smmu_pmu * smmu_pmu)818 static void smmu_pmu_get_iidr(struct smmu_pmu *smmu_pmu)
819 {
820 u32 iidr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_IIDR);
821
822 if (!iidr && smmu_pmu_coresight_id_regs(smmu_pmu)) {
823 u32 pidr0 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR0);
824 u32 pidr1 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR1);
825 u32 pidr2 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR2);
826 u32 pidr3 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR3);
827 u32 pidr4 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR4);
828
829 u32 productid = FIELD_GET(SMMU_PMCG_PIDR0_PART_0, pidr0) |
830 (FIELD_GET(SMMU_PMCG_PIDR1_PART_1, pidr1) << 8);
831 u32 variant = FIELD_GET(SMMU_PMCG_PIDR2_REVISION, pidr2);
832 u32 revision = FIELD_GET(SMMU_PMCG_PIDR3_REVAND, pidr3);
833 u32 implementer =
834 FIELD_GET(SMMU_PMCG_PIDR1_DES_0, pidr1) |
835 (FIELD_GET(SMMU_PMCG_PIDR2_DES_1, pidr2) << 4) |
836 (FIELD_GET(SMMU_PMCG_PIDR4_DES_2, pidr4) << 8);
837
838 iidr = FIELD_PREP(SMMU_PMCG_IIDR_PRODUCTID, productid) |
839 FIELD_PREP(SMMU_PMCG_IIDR_VARIANT, variant) |
840 FIELD_PREP(SMMU_PMCG_IIDR_REVISION, revision) |
841 FIELD_PREP(SMMU_PMCG_IIDR_IMPLEMENTER, implementer);
842 }
843
844 smmu_pmu->iidr = iidr;
845 }
846
smmu_pmu_probe(struct platform_device * pdev)847 static int smmu_pmu_probe(struct platform_device *pdev)
848 {
849 struct smmu_pmu *smmu_pmu;
850 struct resource *res_0;
851 u32 cfgr, reg_size;
852 u64 ceid_64[2];
853 int irq, err;
854 char *name;
855 struct device *dev = &pdev->dev;
856
857 smmu_pmu = devm_kzalloc(dev, sizeof(*smmu_pmu), GFP_KERNEL);
858 if (!smmu_pmu)
859 return -ENOMEM;
860
861 smmu_pmu->dev = dev;
862 platform_set_drvdata(pdev, smmu_pmu);
863
864 smmu_pmu->pmu = (struct pmu) {
865 .module = THIS_MODULE,
866 .parent = &pdev->dev,
867 .task_ctx_nr = perf_invalid_context,
868 .pmu_enable = smmu_pmu_enable,
869 .pmu_disable = smmu_pmu_disable,
870 .event_init = smmu_pmu_event_init,
871 .add = smmu_pmu_event_add,
872 .del = smmu_pmu_event_del,
873 .start = smmu_pmu_event_start,
874 .stop = smmu_pmu_event_stop,
875 .read = smmu_pmu_event_read,
876 .attr_groups = smmu_pmu_attr_grps,
877 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
878 };
879
880 smmu_pmu->reg_base = devm_platform_get_and_ioremap_resource(pdev, 0, &res_0);
881 if (IS_ERR(smmu_pmu->reg_base))
882 return PTR_ERR(smmu_pmu->reg_base);
883
884 cfgr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CFGR);
885
886 /* Determine if page 1 is present */
887 if (cfgr & SMMU_PMCG_CFGR_RELOC_CTRS) {
888 smmu_pmu->reloc_base = devm_platform_ioremap_resource(pdev, 1);
889 if (IS_ERR(smmu_pmu->reloc_base))
890 return PTR_ERR(smmu_pmu->reloc_base);
891 } else {
892 smmu_pmu->reloc_base = smmu_pmu->reg_base;
893 }
894
895 irq = platform_get_irq_optional(pdev, 0);
896 if (irq > 0)
897 smmu_pmu->irq = irq;
898
899 ceid_64[0] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID0);
900 ceid_64[1] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID1);
901 bitmap_from_arr32(smmu_pmu->supported_events, (u32 *)ceid_64,
902 SMMU_PMCG_ARCH_MAX_EVENTS);
903
904 smmu_pmu->num_counters = FIELD_GET(SMMU_PMCG_CFGR_NCTR, cfgr) + 1;
905
906 smmu_pmu->global_filter = !!(cfgr & SMMU_PMCG_CFGR_SID_FILTER_TYPE);
907
908 reg_size = FIELD_GET(SMMU_PMCG_CFGR_SIZE, cfgr);
909 smmu_pmu->counter_mask = GENMASK_ULL(reg_size, 0);
910
911 smmu_pmu_reset(smmu_pmu);
912
913 err = smmu_pmu_setup_irq(smmu_pmu);
914 if (err) {
915 dev_err(dev, "Setup irq failed, PMU @%pa\n", &res_0->start);
916 return err;
917 }
918
919 smmu_pmu_get_iidr(smmu_pmu);
920
921 name = devm_kasprintf(&pdev->dev, GFP_KERNEL, "smmuv3_pmcg_%llx",
922 (res_0->start) >> SMMU_PMCG_PA_SHIFT);
923 if (!name) {
924 dev_err(dev, "Create name failed, PMU @%pa\n", &res_0->start);
925 return -EINVAL;
926 }
927
928 if (!dev->of_node)
929 smmu_pmu_get_acpi_options(smmu_pmu);
930
931 /*
932 * For platforms suffer this quirk, the PMU disable sometimes fails to
933 * stop the counters. This will leads to inaccurate or error counting.
934 * Forcibly disable the counters with these quirk handler.
935 */
936 if (smmu_pmu->options & SMMU_PMCG_HARDEN_DISABLE) {
937 smmu_pmu->pmu.pmu_enable = smmu_pmu_enable_quirk_hip08_09;
938 smmu_pmu->pmu.pmu_disable = smmu_pmu_disable_quirk_hip08_09;
939 }
940
941 /* Pick one CPU to be the preferred one to use */
942 smmu_pmu->on_cpu = raw_smp_processor_id();
943 WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(smmu_pmu->on_cpu)));
944
945 err = cpuhp_state_add_instance_nocalls(cpuhp_state_num,
946 &smmu_pmu->node);
947 if (err) {
948 dev_err(dev, "Error %d registering hotplug, PMU @%pa\n",
949 err, &res_0->start);
950 return err;
951 }
952
953 err = perf_pmu_register(&smmu_pmu->pmu, name, -1);
954 if (err) {
955 dev_err(dev, "Error %d registering PMU @%pa\n",
956 err, &res_0->start);
957 goto out_unregister;
958 }
959
960 dev_info(dev, "Registered PMU @ %pa using %d counters with %s filter settings\n",
961 &res_0->start, smmu_pmu->num_counters,
962 smmu_pmu->global_filter ? "Global(Counter0)" :
963 "Individual");
964
965 return 0;
966
967 out_unregister:
968 cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
969 return err;
970 }
971
smmu_pmu_remove(struct platform_device * pdev)972 static void smmu_pmu_remove(struct platform_device *pdev)
973 {
974 struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
975
976 perf_pmu_unregister(&smmu_pmu->pmu);
977 cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
978 }
979
smmu_pmu_shutdown(struct platform_device * pdev)980 static void smmu_pmu_shutdown(struct platform_device *pdev)
981 {
982 struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
983
984 smmu_pmu_disable(&smmu_pmu->pmu);
985 }
986
987 #ifdef CONFIG_OF
988 static const struct of_device_id smmu_pmu_of_match[] = {
989 { .compatible = "arm,smmu-v3-pmcg" },
990 {}
991 };
992 MODULE_DEVICE_TABLE(of, smmu_pmu_of_match);
993 #endif
994
995 static struct platform_driver smmu_pmu_driver = {
996 .driver = {
997 .name = "arm-smmu-v3-pmcg",
998 .of_match_table = of_match_ptr(smmu_pmu_of_match),
999 .suppress_bind_attrs = true,
1000 },
1001 .probe = smmu_pmu_probe,
1002 .remove = smmu_pmu_remove,
1003 .shutdown = smmu_pmu_shutdown,
1004 };
1005
arm_smmu_pmu_init(void)1006 static int __init arm_smmu_pmu_init(void)
1007 {
1008 int ret;
1009
1010 cpuhp_state_num = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1011 "perf/arm/pmcg:online",
1012 NULL,
1013 smmu_pmu_offline_cpu);
1014 if (cpuhp_state_num < 0)
1015 return cpuhp_state_num;
1016
1017 ret = platform_driver_register(&smmu_pmu_driver);
1018 if (ret)
1019 cpuhp_remove_multi_state(cpuhp_state_num);
1020
1021 return ret;
1022 }
1023 module_init(arm_smmu_pmu_init);
1024
arm_smmu_pmu_exit(void)1025 static void __exit arm_smmu_pmu_exit(void)
1026 {
1027 platform_driver_unregister(&smmu_pmu_driver);
1028 cpuhp_remove_multi_state(cpuhp_state_num);
1029 }
1030
1031 module_exit(arm_smmu_pmu_exit);
1032
1033 MODULE_ALIAS("platform:arm-smmu-v3-pmcg");
1034 MODULE_DESCRIPTION("PMU driver for ARM SMMUv3 Performance Monitors Extension");
1035 MODULE_AUTHOR("Neil Leeder <nleeder@codeaurora.org>");
1036 MODULE_AUTHOR("Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>");
1037 MODULE_LICENSE("GPL v2");
1038