xref: /openbmc/linux/drivers/perf/arm_smmuv3_pmu.c (revision 63705da3)
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/perf_event.h>
51 #include <linux/platform_device.h>
52 #include <linux/smp.h>
53 #include <linux/sysfs.h>
54 #include <linux/types.h>
55 
56 #define SMMU_PMCG_EVCNTR0               0x0
57 #define SMMU_PMCG_EVCNTR(n, stride)     (SMMU_PMCG_EVCNTR0 + (n) * (stride))
58 #define SMMU_PMCG_EVTYPER0              0x400
59 #define SMMU_PMCG_EVTYPER(n)            (SMMU_PMCG_EVTYPER0 + (n) * 4)
60 #define SMMU_PMCG_SID_SPAN_SHIFT        29
61 #define SMMU_PMCG_SMR0                  0xA00
62 #define SMMU_PMCG_SMR(n)                (SMMU_PMCG_SMR0 + (n) * 4)
63 #define SMMU_PMCG_CNTENSET0             0xC00
64 #define SMMU_PMCG_CNTENCLR0             0xC20
65 #define SMMU_PMCG_INTENSET0             0xC40
66 #define SMMU_PMCG_INTENCLR0             0xC60
67 #define SMMU_PMCG_OVSCLR0               0xC80
68 #define SMMU_PMCG_OVSSET0               0xCC0
69 #define SMMU_PMCG_CFGR                  0xE00
70 #define SMMU_PMCG_CFGR_SID_FILTER_TYPE  BIT(23)
71 #define SMMU_PMCG_CFGR_MSI              BIT(21)
72 #define SMMU_PMCG_CFGR_RELOC_CTRS       BIT(20)
73 #define SMMU_PMCG_CFGR_SIZE             GENMASK(13, 8)
74 #define SMMU_PMCG_CFGR_NCTR             GENMASK(5, 0)
75 #define SMMU_PMCG_CR                    0xE04
76 #define SMMU_PMCG_CR_ENABLE             BIT(0)
77 #define SMMU_PMCG_IIDR                  0xE08
78 #define SMMU_PMCG_CEID0                 0xE20
79 #define SMMU_PMCG_CEID1                 0xE28
80 #define SMMU_PMCG_IRQ_CTRL              0xE50
81 #define SMMU_PMCG_IRQ_CTRL_IRQEN        BIT(0)
82 #define SMMU_PMCG_IRQ_CFG0              0xE58
83 #define SMMU_PMCG_IRQ_CFG1              0xE60
84 #define SMMU_PMCG_IRQ_CFG2              0xE64
85 
86 /* MSI config fields */
87 #define MSI_CFG0_ADDR_MASK              GENMASK_ULL(51, 2)
88 #define MSI_CFG2_MEMATTR_DEVICE_nGnRE   0x1
89 
90 #define SMMU_PMCG_DEFAULT_FILTER_SPAN   1
91 #define SMMU_PMCG_DEFAULT_FILTER_SID    GENMASK(31, 0)
92 
93 #define SMMU_PMCG_MAX_COUNTERS          64
94 #define SMMU_PMCG_ARCH_MAX_EVENTS       128
95 
96 #define SMMU_PMCG_PA_SHIFT              12
97 
98 #define SMMU_PMCG_EVCNTR_RDONLY         BIT(0)
99 
100 static int cpuhp_state_num;
101 
102 struct smmu_pmu {
103 	struct hlist_node node;
104 	struct perf_event *events[SMMU_PMCG_MAX_COUNTERS];
105 	DECLARE_BITMAP(used_counters, SMMU_PMCG_MAX_COUNTERS);
106 	DECLARE_BITMAP(supported_events, SMMU_PMCG_ARCH_MAX_EVENTS);
107 	unsigned int irq;
108 	unsigned int on_cpu;
109 	struct pmu pmu;
110 	unsigned int num_counters;
111 	struct device *dev;
112 	void __iomem *reg_base;
113 	void __iomem *reloc_base;
114 	u64 counter_mask;
115 	u32 options;
116 	u32 iidr;
117 	bool global_filter;
118 };
119 
120 #define to_smmu_pmu(p) (container_of(p, struct smmu_pmu, pmu))
121 
122 #define SMMU_PMU_EVENT_ATTR_EXTRACTOR(_name, _config, _start, _end)        \
123 	static inline u32 get_##_name(struct perf_event *event)            \
124 	{                                                                  \
125 		return FIELD_GET(GENMASK_ULL(_end, _start),                \
126 				 event->attr._config);                     \
127 	}                                                                  \
128 
129 SMMU_PMU_EVENT_ATTR_EXTRACTOR(event, config, 0, 15);
130 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_stream_id, config1, 0, 31);
131 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_span, config1, 32, 32);
132 SMMU_PMU_EVENT_ATTR_EXTRACTOR(filter_enable, config1, 33, 33);
133 
134 static inline void smmu_pmu_enable(struct pmu *pmu)
135 {
136 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
137 
138 	writel(SMMU_PMCG_IRQ_CTRL_IRQEN,
139 	       smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
140 	writel(SMMU_PMCG_CR_ENABLE, smmu_pmu->reg_base + SMMU_PMCG_CR);
141 }
142 
143 static inline void smmu_pmu_disable(struct pmu *pmu)
144 {
145 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(pmu);
146 
147 	writel(0, smmu_pmu->reg_base + SMMU_PMCG_CR);
148 	writel(0, smmu_pmu->reg_base + SMMU_PMCG_IRQ_CTRL);
149 }
150 
151 static inline void smmu_pmu_counter_set_value(struct smmu_pmu *smmu_pmu,
152 					      u32 idx, u64 value)
153 {
154 	if (smmu_pmu->counter_mask & BIT(32))
155 		writeq(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
156 	else
157 		writel(value, smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
158 }
159 
160 static inline u64 smmu_pmu_counter_get_value(struct smmu_pmu *smmu_pmu, u32 idx)
161 {
162 	u64 value;
163 
164 	if (smmu_pmu->counter_mask & BIT(32))
165 		value = readq(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 8));
166 	else
167 		value = readl(smmu_pmu->reloc_base + SMMU_PMCG_EVCNTR(idx, 4));
168 
169 	return value;
170 }
171 
172 static inline void smmu_pmu_counter_enable(struct smmu_pmu *smmu_pmu, u32 idx)
173 {
174 	writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENSET0);
175 }
176 
177 static inline void smmu_pmu_counter_disable(struct smmu_pmu *smmu_pmu, u32 idx)
178 {
179 	writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
180 }
181 
182 static inline void smmu_pmu_interrupt_enable(struct smmu_pmu *smmu_pmu, u32 idx)
183 {
184 	writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENSET0);
185 }
186 
187 static inline void smmu_pmu_interrupt_disable(struct smmu_pmu *smmu_pmu,
188 					      u32 idx)
189 {
190 	writeq(BIT(idx), smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
191 }
192 
193 static inline void smmu_pmu_set_evtyper(struct smmu_pmu *smmu_pmu, u32 idx,
194 					u32 val)
195 {
196 	writel(val, smmu_pmu->reg_base + SMMU_PMCG_EVTYPER(idx));
197 }
198 
199 static inline void smmu_pmu_set_smr(struct smmu_pmu *smmu_pmu, u32 idx, u32 val)
200 {
201 	writel(val, smmu_pmu->reg_base + SMMU_PMCG_SMR(idx));
202 }
203 
204 static void smmu_pmu_event_update(struct perf_event *event)
205 {
206 	struct hw_perf_event *hwc = &event->hw;
207 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
208 	u64 delta, prev, now;
209 	u32 idx = hwc->idx;
210 
211 	do {
212 		prev = local64_read(&hwc->prev_count);
213 		now = smmu_pmu_counter_get_value(smmu_pmu, idx);
214 	} while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
215 
216 	/* handle overflow. */
217 	delta = now - prev;
218 	delta &= smmu_pmu->counter_mask;
219 
220 	local64_add(delta, &event->count);
221 }
222 
223 static void smmu_pmu_set_period(struct smmu_pmu *smmu_pmu,
224 				struct hw_perf_event *hwc)
225 {
226 	u32 idx = hwc->idx;
227 	u64 new;
228 
229 	if (smmu_pmu->options & SMMU_PMCG_EVCNTR_RDONLY) {
230 		/*
231 		 * On platforms that require this quirk, if the counter starts
232 		 * at < half_counter value and wraps, the current logic of
233 		 * handling the overflow may not work. It is expected that,
234 		 * those platforms will have full 64 counter bits implemented
235 		 * so that such a possibility is remote(eg: HiSilicon HIP08).
236 		 */
237 		new = smmu_pmu_counter_get_value(smmu_pmu, idx);
238 	} else {
239 		/*
240 		 * We limit the max period to half the max counter value
241 		 * of the counter size, so that even in the case of extreme
242 		 * interrupt latency the counter will (hopefully) not wrap
243 		 * past its initial value.
244 		 */
245 		new = smmu_pmu->counter_mask >> 1;
246 		smmu_pmu_counter_set_value(smmu_pmu, idx, new);
247 	}
248 
249 	local64_set(&hwc->prev_count, new);
250 }
251 
252 static void smmu_pmu_set_event_filter(struct perf_event *event,
253 				      int idx, u32 span, u32 sid)
254 {
255 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
256 	u32 evtyper;
257 
258 	evtyper = get_event(event) | span << SMMU_PMCG_SID_SPAN_SHIFT;
259 	smmu_pmu_set_evtyper(smmu_pmu, idx, evtyper);
260 	smmu_pmu_set_smr(smmu_pmu, idx, sid);
261 }
262 
263 static bool smmu_pmu_check_global_filter(struct perf_event *curr,
264 					 struct perf_event *new)
265 {
266 	if (get_filter_enable(new) != get_filter_enable(curr))
267 		return false;
268 
269 	if (!get_filter_enable(new))
270 		return true;
271 
272 	return get_filter_span(new) == get_filter_span(curr) &&
273 	       get_filter_stream_id(new) == get_filter_stream_id(curr);
274 }
275 
276 static int smmu_pmu_apply_event_filter(struct smmu_pmu *smmu_pmu,
277 				       struct perf_event *event, int idx)
278 {
279 	u32 span, sid;
280 	unsigned int cur_idx, num_ctrs = smmu_pmu->num_counters;
281 	bool filter_en = !!get_filter_enable(event);
282 
283 	span = filter_en ? get_filter_span(event) :
284 			   SMMU_PMCG_DEFAULT_FILTER_SPAN;
285 	sid = filter_en ? get_filter_stream_id(event) :
286 			   SMMU_PMCG_DEFAULT_FILTER_SID;
287 
288 	cur_idx = find_first_bit(smmu_pmu->used_counters, num_ctrs);
289 	/*
290 	 * Per-counter filtering, or scheduling the first globally-filtered
291 	 * event into an empty PMU so idx == 0 and it works out equivalent.
292 	 */
293 	if (!smmu_pmu->global_filter || cur_idx == num_ctrs) {
294 		smmu_pmu_set_event_filter(event, idx, span, sid);
295 		return 0;
296 	}
297 
298 	/* Otherwise, must match whatever's currently scheduled */
299 	if (smmu_pmu_check_global_filter(smmu_pmu->events[cur_idx], event)) {
300 		smmu_pmu_set_evtyper(smmu_pmu, idx, get_event(event));
301 		return 0;
302 	}
303 
304 	return -EAGAIN;
305 }
306 
307 static int smmu_pmu_get_event_idx(struct smmu_pmu *smmu_pmu,
308 				  struct perf_event *event)
309 {
310 	int idx, err;
311 	unsigned int num_ctrs = smmu_pmu->num_counters;
312 
313 	idx = find_first_zero_bit(smmu_pmu->used_counters, num_ctrs);
314 	if (idx == num_ctrs)
315 		/* The counters are all in use. */
316 		return -EAGAIN;
317 
318 	err = smmu_pmu_apply_event_filter(smmu_pmu, event, idx);
319 	if (err)
320 		return err;
321 
322 	set_bit(idx, smmu_pmu->used_counters);
323 
324 	return idx;
325 }
326 
327 static bool smmu_pmu_events_compatible(struct perf_event *curr,
328 				       struct perf_event *new)
329 {
330 	if (new->pmu != curr->pmu)
331 		return false;
332 
333 	if (to_smmu_pmu(new->pmu)->global_filter &&
334 	    !smmu_pmu_check_global_filter(curr, new))
335 		return false;
336 
337 	return true;
338 }
339 
340 /*
341  * Implementation of abstract pmu functionality required by
342  * the core perf events code.
343  */
344 
345 static int smmu_pmu_event_init(struct perf_event *event)
346 {
347 	struct hw_perf_event *hwc = &event->hw;
348 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
349 	struct device *dev = smmu_pmu->dev;
350 	struct perf_event *sibling;
351 	int group_num_events = 1;
352 	u16 event_id;
353 
354 	if (event->attr.type != event->pmu->type)
355 		return -ENOENT;
356 
357 	if (hwc->sample_period) {
358 		dev_dbg(dev, "Sampling not supported\n");
359 		return -EOPNOTSUPP;
360 	}
361 
362 	if (event->cpu < 0) {
363 		dev_dbg(dev, "Per-task mode not supported\n");
364 		return -EOPNOTSUPP;
365 	}
366 
367 	/* Verify specified event is supported on this PMU */
368 	event_id = get_event(event);
369 	if (event_id < SMMU_PMCG_ARCH_MAX_EVENTS &&
370 	    (!test_bit(event_id, smmu_pmu->supported_events))) {
371 		dev_dbg(dev, "Invalid event %d for this PMU\n", event_id);
372 		return -EINVAL;
373 	}
374 
375 	/* Don't allow groups with mixed PMUs, except for s/w events */
376 	if (!is_software_event(event->group_leader)) {
377 		if (!smmu_pmu_events_compatible(event->group_leader, event))
378 			return -EINVAL;
379 
380 		if (++group_num_events > smmu_pmu->num_counters)
381 			return -EINVAL;
382 	}
383 
384 	for_each_sibling_event(sibling, event->group_leader) {
385 		if (is_software_event(sibling))
386 			continue;
387 
388 		if (!smmu_pmu_events_compatible(sibling, event))
389 			return -EINVAL;
390 
391 		if (++group_num_events > smmu_pmu->num_counters)
392 			return -EINVAL;
393 	}
394 
395 	hwc->idx = -1;
396 
397 	/*
398 	 * Ensure all events are on the same cpu so all events are in the
399 	 * same cpu context, to avoid races on pmu_enable etc.
400 	 */
401 	event->cpu = smmu_pmu->on_cpu;
402 
403 	return 0;
404 }
405 
406 static void smmu_pmu_event_start(struct perf_event *event, int flags)
407 {
408 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
409 	struct hw_perf_event *hwc = &event->hw;
410 	int idx = hwc->idx;
411 
412 	hwc->state = 0;
413 
414 	smmu_pmu_set_period(smmu_pmu, hwc);
415 
416 	smmu_pmu_counter_enable(smmu_pmu, idx);
417 }
418 
419 static void smmu_pmu_event_stop(struct perf_event *event, int flags)
420 {
421 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
422 	struct hw_perf_event *hwc = &event->hw;
423 	int idx = hwc->idx;
424 
425 	if (hwc->state & PERF_HES_STOPPED)
426 		return;
427 
428 	smmu_pmu_counter_disable(smmu_pmu, idx);
429 	/* As the counter gets updated on _start, ignore PERF_EF_UPDATE */
430 	smmu_pmu_event_update(event);
431 	hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
432 }
433 
434 static int smmu_pmu_event_add(struct perf_event *event, int flags)
435 {
436 	struct hw_perf_event *hwc = &event->hw;
437 	int idx;
438 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
439 
440 	idx = smmu_pmu_get_event_idx(smmu_pmu, event);
441 	if (idx < 0)
442 		return idx;
443 
444 	hwc->idx = idx;
445 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
446 	smmu_pmu->events[idx] = event;
447 	local64_set(&hwc->prev_count, 0);
448 
449 	smmu_pmu_interrupt_enable(smmu_pmu, idx);
450 
451 	if (flags & PERF_EF_START)
452 		smmu_pmu_event_start(event, flags);
453 
454 	/* Propagate changes to the userspace mapping. */
455 	perf_event_update_userpage(event);
456 
457 	return 0;
458 }
459 
460 static void smmu_pmu_event_del(struct perf_event *event, int flags)
461 {
462 	struct hw_perf_event *hwc = &event->hw;
463 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
464 	int idx = hwc->idx;
465 
466 	smmu_pmu_event_stop(event, flags | PERF_EF_UPDATE);
467 	smmu_pmu_interrupt_disable(smmu_pmu, idx);
468 	smmu_pmu->events[idx] = NULL;
469 	clear_bit(idx, smmu_pmu->used_counters);
470 
471 	perf_event_update_userpage(event);
472 }
473 
474 static void smmu_pmu_event_read(struct perf_event *event)
475 {
476 	smmu_pmu_event_update(event);
477 }
478 
479 /* cpumask */
480 
481 static ssize_t smmu_pmu_cpumask_show(struct device *dev,
482 				     struct device_attribute *attr,
483 				     char *buf)
484 {
485 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
486 
487 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(smmu_pmu->on_cpu));
488 }
489 
490 static struct device_attribute smmu_pmu_cpumask_attr =
491 		__ATTR(cpumask, 0444, smmu_pmu_cpumask_show, NULL);
492 
493 static struct attribute *smmu_pmu_cpumask_attrs[] = {
494 	&smmu_pmu_cpumask_attr.attr,
495 	NULL
496 };
497 
498 static const struct attribute_group smmu_pmu_cpumask_group = {
499 	.attrs = smmu_pmu_cpumask_attrs,
500 };
501 
502 /* Events */
503 
504 static ssize_t smmu_pmu_event_show(struct device *dev,
505 				   struct device_attribute *attr, char *page)
506 {
507 	struct perf_pmu_events_attr *pmu_attr;
508 
509 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
510 
511 	return sysfs_emit(page, "event=0x%02llx\n", pmu_attr->id);
512 }
513 
514 #define SMMU_EVENT_ATTR(name, config)			\
515 	PMU_EVENT_ATTR_ID(name, smmu_pmu_event_show, config)
516 
517 static struct attribute *smmu_pmu_events[] = {
518 	SMMU_EVENT_ATTR(cycles, 0),
519 	SMMU_EVENT_ATTR(transaction, 1),
520 	SMMU_EVENT_ATTR(tlb_miss, 2),
521 	SMMU_EVENT_ATTR(config_cache_miss, 3),
522 	SMMU_EVENT_ATTR(trans_table_walk_access, 4),
523 	SMMU_EVENT_ATTR(config_struct_access, 5),
524 	SMMU_EVENT_ATTR(pcie_ats_trans_rq, 6),
525 	SMMU_EVENT_ATTR(pcie_ats_trans_passed, 7),
526 	NULL
527 };
528 
529 static umode_t smmu_pmu_event_is_visible(struct kobject *kobj,
530 					 struct attribute *attr, int unused)
531 {
532 	struct device *dev = kobj_to_dev(kobj);
533 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
534 	struct perf_pmu_events_attr *pmu_attr;
535 
536 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
537 
538 	if (test_bit(pmu_attr->id, smmu_pmu->supported_events))
539 		return attr->mode;
540 
541 	return 0;
542 }
543 
544 static const struct attribute_group smmu_pmu_events_group = {
545 	.name = "events",
546 	.attrs = smmu_pmu_events,
547 	.is_visible = smmu_pmu_event_is_visible,
548 };
549 
550 static ssize_t smmu_pmu_identifier_attr_show(struct device *dev,
551 					struct device_attribute *attr,
552 					char *page)
553 {
554 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
555 
556 	return sysfs_emit(page, "0x%08x\n", smmu_pmu->iidr);
557 }
558 
559 static umode_t smmu_pmu_identifier_attr_visible(struct kobject *kobj,
560 						struct attribute *attr,
561 						int n)
562 {
563 	struct device *dev = kobj_to_dev(kobj);
564 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
565 
566 	if (!smmu_pmu->iidr)
567 		return 0;
568 	return attr->mode;
569 }
570 
571 static struct device_attribute smmu_pmu_identifier_attr =
572 	__ATTR(identifier, 0444, smmu_pmu_identifier_attr_show, NULL);
573 
574 static struct attribute *smmu_pmu_identifier_attrs[] = {
575 	&smmu_pmu_identifier_attr.attr,
576 	NULL
577 };
578 
579 static const struct attribute_group smmu_pmu_identifier_group = {
580 	.attrs = smmu_pmu_identifier_attrs,
581 	.is_visible = smmu_pmu_identifier_attr_visible,
582 };
583 
584 /* Formats */
585 PMU_FORMAT_ATTR(event,		   "config:0-15");
586 PMU_FORMAT_ATTR(filter_stream_id,  "config1:0-31");
587 PMU_FORMAT_ATTR(filter_span,	   "config1:32");
588 PMU_FORMAT_ATTR(filter_enable,	   "config1:33");
589 
590 static struct attribute *smmu_pmu_formats[] = {
591 	&format_attr_event.attr,
592 	&format_attr_filter_stream_id.attr,
593 	&format_attr_filter_span.attr,
594 	&format_attr_filter_enable.attr,
595 	NULL
596 };
597 
598 static const struct attribute_group smmu_pmu_format_group = {
599 	.name = "format",
600 	.attrs = smmu_pmu_formats,
601 };
602 
603 static const struct attribute_group *smmu_pmu_attr_grps[] = {
604 	&smmu_pmu_cpumask_group,
605 	&smmu_pmu_events_group,
606 	&smmu_pmu_format_group,
607 	&smmu_pmu_identifier_group,
608 	NULL
609 };
610 
611 /*
612  * Generic device handlers
613  */
614 
615 static int smmu_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
616 {
617 	struct smmu_pmu *smmu_pmu;
618 	unsigned int target;
619 
620 	smmu_pmu = hlist_entry_safe(node, struct smmu_pmu, node);
621 	if (cpu != smmu_pmu->on_cpu)
622 		return 0;
623 
624 	target = cpumask_any_but(cpu_online_mask, cpu);
625 	if (target >= nr_cpu_ids)
626 		return 0;
627 
628 	perf_pmu_migrate_context(&smmu_pmu->pmu, cpu, target);
629 	smmu_pmu->on_cpu = target;
630 	WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(target)));
631 
632 	return 0;
633 }
634 
635 static irqreturn_t smmu_pmu_handle_irq(int irq_num, void *data)
636 {
637 	struct smmu_pmu *smmu_pmu = data;
638 	u64 ovsr;
639 	unsigned int idx;
640 
641 	ovsr = readq(smmu_pmu->reloc_base + SMMU_PMCG_OVSSET0);
642 	if (!ovsr)
643 		return IRQ_NONE;
644 
645 	writeq(ovsr, smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
646 
647 	for_each_set_bit(idx, (unsigned long *)&ovsr, smmu_pmu->num_counters) {
648 		struct perf_event *event = smmu_pmu->events[idx];
649 		struct hw_perf_event *hwc;
650 
651 		if (WARN_ON_ONCE(!event))
652 			continue;
653 
654 		smmu_pmu_event_update(event);
655 		hwc = &event->hw;
656 
657 		smmu_pmu_set_period(smmu_pmu, hwc);
658 	}
659 
660 	return IRQ_HANDLED;
661 }
662 
663 static void smmu_pmu_free_msis(void *data)
664 {
665 	struct device *dev = data;
666 
667 	platform_msi_domain_free_irqs(dev);
668 }
669 
670 static void smmu_pmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
671 {
672 	phys_addr_t doorbell;
673 	struct device *dev = msi_desc_to_dev(desc);
674 	struct smmu_pmu *pmu = dev_get_drvdata(dev);
675 
676 	doorbell = (((u64)msg->address_hi) << 32) | msg->address_lo;
677 	doorbell &= MSI_CFG0_ADDR_MASK;
678 
679 	writeq_relaxed(doorbell, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
680 	writel_relaxed(msg->data, pmu->reg_base + SMMU_PMCG_IRQ_CFG1);
681 	writel_relaxed(MSI_CFG2_MEMATTR_DEVICE_nGnRE,
682 		       pmu->reg_base + SMMU_PMCG_IRQ_CFG2);
683 }
684 
685 static void smmu_pmu_setup_msi(struct smmu_pmu *pmu)
686 {
687 	struct msi_desc *desc;
688 	struct device *dev = pmu->dev;
689 	int ret;
690 
691 	/* Clear MSI address reg */
692 	writeq_relaxed(0, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
693 
694 	/* MSI supported or not */
695 	if (!(readl(pmu->reg_base + SMMU_PMCG_CFGR) & SMMU_PMCG_CFGR_MSI))
696 		return;
697 
698 	ret = platform_msi_domain_alloc_irqs(dev, 1, smmu_pmu_write_msi_msg);
699 	if (ret) {
700 		dev_warn(dev, "failed to allocate MSIs\n");
701 		return;
702 	}
703 
704 	desc = first_msi_entry(dev);
705 	if (desc)
706 		pmu->irq = desc->irq;
707 
708 	/* Add callback to free MSIs on teardown */
709 	devm_add_action(dev, smmu_pmu_free_msis, dev);
710 }
711 
712 static int smmu_pmu_setup_irq(struct smmu_pmu *pmu)
713 {
714 	unsigned long flags = IRQF_NOBALANCING | IRQF_SHARED | IRQF_NO_THREAD;
715 	int irq, ret = -ENXIO;
716 
717 	smmu_pmu_setup_msi(pmu);
718 
719 	irq = pmu->irq;
720 	if (irq)
721 		ret = devm_request_irq(pmu->dev, irq, smmu_pmu_handle_irq,
722 				       flags, "smmuv3-pmu", pmu);
723 	return ret;
724 }
725 
726 static void smmu_pmu_reset(struct smmu_pmu *smmu_pmu)
727 {
728 	u64 counter_present_mask = GENMASK_ULL(smmu_pmu->num_counters - 1, 0);
729 
730 	smmu_pmu_disable(&smmu_pmu->pmu);
731 
732 	/* Disable counter and interrupt */
733 	writeq_relaxed(counter_present_mask,
734 		       smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
735 	writeq_relaxed(counter_present_mask,
736 		       smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
737 	writeq_relaxed(counter_present_mask,
738 		       smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
739 }
740 
741 static void smmu_pmu_get_acpi_options(struct smmu_pmu *smmu_pmu)
742 {
743 	u32 model;
744 
745 	model = *(u32 *)dev_get_platdata(smmu_pmu->dev);
746 
747 	switch (model) {
748 	case IORT_SMMU_V3_PMCG_HISI_HIP08:
749 		/* HiSilicon Erratum 162001800 */
750 		smmu_pmu->options |= SMMU_PMCG_EVCNTR_RDONLY;
751 		break;
752 	}
753 
754 	dev_notice(smmu_pmu->dev, "option mask 0x%x\n", smmu_pmu->options);
755 }
756 
757 static int smmu_pmu_probe(struct platform_device *pdev)
758 {
759 	struct smmu_pmu *smmu_pmu;
760 	struct resource *res_0;
761 	u32 cfgr, reg_size;
762 	u64 ceid_64[2];
763 	int irq, err;
764 	char *name;
765 	struct device *dev = &pdev->dev;
766 
767 	smmu_pmu = devm_kzalloc(dev, sizeof(*smmu_pmu), GFP_KERNEL);
768 	if (!smmu_pmu)
769 		return -ENOMEM;
770 
771 	smmu_pmu->dev = dev;
772 	platform_set_drvdata(pdev, smmu_pmu);
773 
774 	smmu_pmu->pmu = (struct pmu) {
775 		.module		= THIS_MODULE,
776 		.task_ctx_nr    = perf_invalid_context,
777 		.pmu_enable	= smmu_pmu_enable,
778 		.pmu_disable	= smmu_pmu_disable,
779 		.event_init	= smmu_pmu_event_init,
780 		.add		= smmu_pmu_event_add,
781 		.del		= smmu_pmu_event_del,
782 		.start		= smmu_pmu_event_start,
783 		.stop		= smmu_pmu_event_stop,
784 		.read		= smmu_pmu_event_read,
785 		.attr_groups	= smmu_pmu_attr_grps,
786 		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
787 	};
788 
789 	smmu_pmu->reg_base = devm_platform_get_and_ioremap_resource(pdev, 0, &res_0);
790 	if (IS_ERR(smmu_pmu->reg_base))
791 		return PTR_ERR(smmu_pmu->reg_base);
792 
793 	cfgr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CFGR);
794 
795 	/* Determine if page 1 is present */
796 	if (cfgr & SMMU_PMCG_CFGR_RELOC_CTRS) {
797 		smmu_pmu->reloc_base = devm_platform_ioremap_resource(pdev, 1);
798 		if (IS_ERR(smmu_pmu->reloc_base))
799 			return PTR_ERR(smmu_pmu->reloc_base);
800 	} else {
801 		smmu_pmu->reloc_base = smmu_pmu->reg_base;
802 	}
803 
804 	irq = platform_get_irq_optional(pdev, 0);
805 	if (irq > 0)
806 		smmu_pmu->irq = irq;
807 
808 	ceid_64[0] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID0);
809 	ceid_64[1] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID1);
810 	bitmap_from_arr32(smmu_pmu->supported_events, (u32 *)ceid_64,
811 			  SMMU_PMCG_ARCH_MAX_EVENTS);
812 
813 	smmu_pmu->num_counters = FIELD_GET(SMMU_PMCG_CFGR_NCTR, cfgr) + 1;
814 
815 	smmu_pmu->global_filter = !!(cfgr & SMMU_PMCG_CFGR_SID_FILTER_TYPE);
816 
817 	reg_size = FIELD_GET(SMMU_PMCG_CFGR_SIZE, cfgr);
818 	smmu_pmu->counter_mask = GENMASK_ULL(reg_size, 0);
819 
820 	smmu_pmu_reset(smmu_pmu);
821 
822 	err = smmu_pmu_setup_irq(smmu_pmu);
823 	if (err) {
824 		dev_err(dev, "Setup irq failed, PMU @%pa\n", &res_0->start);
825 		return err;
826 	}
827 
828 	smmu_pmu->iidr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_IIDR);
829 
830 	name = devm_kasprintf(&pdev->dev, GFP_KERNEL, "smmuv3_pmcg_%llx",
831 			      (res_0->start) >> SMMU_PMCG_PA_SHIFT);
832 	if (!name) {
833 		dev_err(dev, "Create name failed, PMU @%pa\n", &res_0->start);
834 		return -EINVAL;
835 	}
836 
837 	smmu_pmu_get_acpi_options(smmu_pmu);
838 
839 	/* Pick one CPU to be the preferred one to use */
840 	smmu_pmu->on_cpu = raw_smp_processor_id();
841 	WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(smmu_pmu->on_cpu)));
842 
843 	err = cpuhp_state_add_instance_nocalls(cpuhp_state_num,
844 					       &smmu_pmu->node);
845 	if (err) {
846 		dev_err(dev, "Error %d registering hotplug, PMU @%pa\n",
847 			err, &res_0->start);
848 		return err;
849 	}
850 
851 	err = perf_pmu_register(&smmu_pmu->pmu, name, -1);
852 	if (err) {
853 		dev_err(dev, "Error %d registering PMU @%pa\n",
854 			err, &res_0->start);
855 		goto out_unregister;
856 	}
857 
858 	dev_info(dev, "Registered PMU @ %pa using %d counters with %s filter settings\n",
859 		 &res_0->start, smmu_pmu->num_counters,
860 		 smmu_pmu->global_filter ? "Global(Counter0)" :
861 		 "Individual");
862 
863 	return 0;
864 
865 out_unregister:
866 	cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
867 	return err;
868 }
869 
870 static int smmu_pmu_remove(struct platform_device *pdev)
871 {
872 	struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
873 
874 	perf_pmu_unregister(&smmu_pmu->pmu);
875 	cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
876 
877 	return 0;
878 }
879 
880 static void smmu_pmu_shutdown(struct platform_device *pdev)
881 {
882 	struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
883 
884 	smmu_pmu_disable(&smmu_pmu->pmu);
885 }
886 
887 static struct platform_driver smmu_pmu_driver = {
888 	.driver = {
889 		.name = "arm-smmu-v3-pmcg",
890 		.suppress_bind_attrs = true,
891 	},
892 	.probe = smmu_pmu_probe,
893 	.remove = smmu_pmu_remove,
894 	.shutdown = smmu_pmu_shutdown,
895 };
896 
897 static int __init arm_smmu_pmu_init(void)
898 {
899 	cpuhp_state_num = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
900 						  "perf/arm/pmcg:online",
901 						  NULL,
902 						  smmu_pmu_offline_cpu);
903 	if (cpuhp_state_num < 0)
904 		return cpuhp_state_num;
905 
906 	return platform_driver_register(&smmu_pmu_driver);
907 }
908 module_init(arm_smmu_pmu_init);
909 
910 static void __exit arm_smmu_pmu_exit(void)
911 {
912 	platform_driver_unregister(&smmu_pmu_driver);
913 	cpuhp_remove_multi_state(cpuhp_state_num);
914 }
915 
916 module_exit(arm_smmu_pmu_exit);
917 
918 MODULE_DESCRIPTION("PMU driver for ARM SMMUv3 Performance Monitors Extension");
919 MODULE_AUTHOR("Neil Leeder <nleeder@codeaurora.org>");
920 MODULE_AUTHOR("Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>");
921 MODULE_LICENSE("GPL v2");
922