xref: /linux/drivers/perf/arm_smmuv3_pmu.c (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 	for_each_sibling_event(sibling, event->group_leader) {
435 		if (is_software_event(sibling))
436 			continue;
437 
438 		if (!smmu_pmu_events_compatible(sibling, event))
439 			return -EINVAL;
440 
441 		if (++group_num_events > smmu_pmu->num_counters)
442 			return -EINVAL;
443 	}
444 
445 	hwc->idx = -1;
446 
447 	/*
448 	 * Ensure all events are on the same cpu so all events are in the
449 	 * same cpu context, to avoid races on pmu_enable etc.
450 	 */
451 	event->cpu = smmu_pmu->on_cpu;
452 
453 	return 0;
454 }
455 
456 static void smmu_pmu_event_start(struct perf_event *event, int flags)
457 {
458 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
459 	struct hw_perf_event *hwc = &event->hw;
460 	int idx = hwc->idx;
461 
462 	hwc->state = 0;
463 
464 	smmu_pmu_set_period(smmu_pmu, hwc);
465 
466 	smmu_pmu_counter_enable(smmu_pmu, idx);
467 }
468 
469 static void smmu_pmu_event_stop(struct perf_event *event, int flags)
470 {
471 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
472 	struct hw_perf_event *hwc = &event->hw;
473 	int idx = hwc->idx;
474 
475 	if (hwc->state & PERF_HES_STOPPED)
476 		return;
477 
478 	smmu_pmu_counter_disable(smmu_pmu, idx);
479 	/* As the counter gets updated on _start, ignore PERF_EF_UPDATE */
480 	smmu_pmu_event_update(event);
481 	hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
482 }
483 
484 static int smmu_pmu_event_add(struct perf_event *event, int flags)
485 {
486 	struct hw_perf_event *hwc = &event->hw;
487 	int idx;
488 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
489 
490 	idx = smmu_pmu_get_event_idx(smmu_pmu, event);
491 	if (idx < 0)
492 		return idx;
493 
494 	hwc->idx = idx;
495 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
496 	smmu_pmu->events[idx] = event;
497 	local64_set(&hwc->prev_count, 0);
498 
499 	smmu_pmu_interrupt_enable(smmu_pmu, idx);
500 
501 	if (flags & PERF_EF_START)
502 		smmu_pmu_event_start(event, flags);
503 
504 	/* Propagate changes to the userspace mapping. */
505 	perf_event_update_userpage(event);
506 
507 	return 0;
508 }
509 
510 static void smmu_pmu_event_del(struct perf_event *event, int flags)
511 {
512 	struct hw_perf_event *hwc = &event->hw;
513 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(event->pmu);
514 	int idx = hwc->idx;
515 
516 	smmu_pmu_event_stop(event, flags | PERF_EF_UPDATE);
517 	smmu_pmu_interrupt_disable(smmu_pmu, idx);
518 	smmu_pmu->events[idx] = NULL;
519 	clear_bit(idx, smmu_pmu->used_counters);
520 
521 	perf_event_update_userpage(event);
522 }
523 
524 static void smmu_pmu_event_read(struct perf_event *event)
525 {
526 	smmu_pmu_event_update(event);
527 }
528 
529 /* cpumask */
530 
531 static ssize_t smmu_pmu_cpumask_show(struct device *dev,
532 				     struct device_attribute *attr,
533 				     char *buf)
534 {
535 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
536 
537 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(smmu_pmu->on_cpu));
538 }
539 
540 static struct device_attribute smmu_pmu_cpumask_attr =
541 		__ATTR(cpumask, 0444, smmu_pmu_cpumask_show, NULL);
542 
543 static struct attribute *smmu_pmu_cpumask_attrs[] = {
544 	&smmu_pmu_cpumask_attr.attr,
545 	NULL
546 };
547 
548 static const struct attribute_group smmu_pmu_cpumask_group = {
549 	.attrs = smmu_pmu_cpumask_attrs,
550 };
551 
552 /* Events */
553 
554 static ssize_t smmu_pmu_event_show(struct device *dev,
555 				   struct device_attribute *attr, char *page)
556 {
557 	struct perf_pmu_events_attr *pmu_attr;
558 
559 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
560 
561 	return sysfs_emit(page, "event=0x%02llx\n", pmu_attr->id);
562 }
563 
564 #define SMMU_EVENT_ATTR(name, config)			\
565 	PMU_EVENT_ATTR_ID(name, smmu_pmu_event_show, config)
566 
567 static struct attribute *smmu_pmu_events[] = {
568 	SMMU_EVENT_ATTR(cycles, 0),
569 	SMMU_EVENT_ATTR(transaction, 1),
570 	SMMU_EVENT_ATTR(tlb_miss, 2),
571 	SMMU_EVENT_ATTR(config_cache_miss, 3),
572 	SMMU_EVENT_ATTR(trans_table_walk_access, 4),
573 	SMMU_EVENT_ATTR(config_struct_access, 5),
574 	SMMU_EVENT_ATTR(pcie_ats_trans_rq, 6),
575 	SMMU_EVENT_ATTR(pcie_ats_trans_passed, 7),
576 	NULL
577 };
578 
579 static umode_t smmu_pmu_event_is_visible(struct kobject *kobj,
580 					 struct attribute *attr, int unused)
581 {
582 	struct device *dev = kobj_to_dev(kobj);
583 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
584 	struct perf_pmu_events_attr *pmu_attr;
585 
586 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
587 
588 	if (test_bit(pmu_attr->id, smmu_pmu->supported_events))
589 		return attr->mode;
590 
591 	return 0;
592 }
593 
594 static const struct attribute_group smmu_pmu_events_group = {
595 	.name = "events",
596 	.attrs = smmu_pmu_events,
597 	.is_visible = smmu_pmu_event_is_visible,
598 };
599 
600 static ssize_t smmu_pmu_identifier_attr_show(struct device *dev,
601 					struct device_attribute *attr,
602 					char *page)
603 {
604 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
605 
606 	return sysfs_emit(page, "0x%08x\n", smmu_pmu->iidr);
607 }
608 
609 static umode_t smmu_pmu_identifier_attr_visible(struct kobject *kobj,
610 						struct attribute *attr,
611 						int n)
612 {
613 	struct device *dev = kobj_to_dev(kobj);
614 	struct smmu_pmu *smmu_pmu = to_smmu_pmu(dev_get_drvdata(dev));
615 
616 	if (!smmu_pmu->iidr)
617 		return 0;
618 	return attr->mode;
619 }
620 
621 static struct device_attribute smmu_pmu_identifier_attr =
622 	__ATTR(identifier, 0444, smmu_pmu_identifier_attr_show, NULL);
623 
624 static struct attribute *smmu_pmu_identifier_attrs[] = {
625 	&smmu_pmu_identifier_attr.attr,
626 	NULL
627 };
628 
629 static const struct attribute_group smmu_pmu_identifier_group = {
630 	.attrs = smmu_pmu_identifier_attrs,
631 	.is_visible = smmu_pmu_identifier_attr_visible,
632 };
633 
634 /* Formats */
635 PMU_FORMAT_ATTR(event,		   "config:0-15");
636 PMU_FORMAT_ATTR(filter_stream_id,  "config1:0-31");
637 PMU_FORMAT_ATTR(filter_span,	   "config1:32");
638 PMU_FORMAT_ATTR(filter_enable,	   "config1:33");
639 
640 static struct attribute *smmu_pmu_formats[] = {
641 	&format_attr_event.attr,
642 	&format_attr_filter_stream_id.attr,
643 	&format_attr_filter_span.attr,
644 	&format_attr_filter_enable.attr,
645 	NULL
646 };
647 
648 static const struct attribute_group smmu_pmu_format_group = {
649 	.name = "format",
650 	.attrs = smmu_pmu_formats,
651 };
652 
653 static const struct attribute_group *smmu_pmu_attr_grps[] = {
654 	&smmu_pmu_cpumask_group,
655 	&smmu_pmu_events_group,
656 	&smmu_pmu_format_group,
657 	&smmu_pmu_identifier_group,
658 	NULL
659 };
660 
661 /*
662  * Generic device handlers
663  */
664 
665 static int smmu_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
666 {
667 	struct smmu_pmu *smmu_pmu;
668 	unsigned int target;
669 
670 	smmu_pmu = hlist_entry_safe(node, struct smmu_pmu, node);
671 	if (cpu != smmu_pmu->on_cpu)
672 		return 0;
673 
674 	target = cpumask_any_but(cpu_online_mask, cpu);
675 	if (target >= nr_cpu_ids)
676 		return 0;
677 
678 	perf_pmu_migrate_context(&smmu_pmu->pmu, cpu, target);
679 	smmu_pmu->on_cpu = target;
680 	WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(target)));
681 
682 	return 0;
683 }
684 
685 static irqreturn_t smmu_pmu_handle_irq(int irq_num, void *data)
686 {
687 	struct smmu_pmu *smmu_pmu = data;
688 	DECLARE_BITMAP(ovs, BITS_PER_TYPE(u64));
689 	u64 ovsr;
690 	unsigned int idx;
691 
692 	ovsr = readq(smmu_pmu->reloc_base + SMMU_PMCG_OVSSET0);
693 	if (!ovsr)
694 		return IRQ_NONE;
695 
696 	writeq(ovsr, smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
697 
698 	bitmap_from_u64(ovs, ovsr);
699 	for_each_set_bit(idx, ovs, smmu_pmu->num_counters) {
700 		struct perf_event *event = smmu_pmu->events[idx];
701 		struct hw_perf_event *hwc;
702 
703 		if (WARN_ON_ONCE(!event))
704 			continue;
705 
706 		smmu_pmu_event_update(event);
707 		hwc = &event->hw;
708 
709 		smmu_pmu_set_period(smmu_pmu, hwc);
710 	}
711 
712 	return IRQ_HANDLED;
713 }
714 
715 static void smmu_pmu_free_msis(void *data)
716 {
717 	struct device *dev = data;
718 
719 	platform_msi_domain_free_irqs(dev);
720 }
721 
722 static void smmu_pmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
723 {
724 	phys_addr_t doorbell;
725 	struct device *dev = msi_desc_to_dev(desc);
726 	struct smmu_pmu *pmu = dev_get_drvdata(dev);
727 
728 	doorbell = (((u64)msg->address_hi) << 32) | msg->address_lo;
729 	doorbell &= MSI_CFG0_ADDR_MASK;
730 
731 	writeq_relaxed(doorbell, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
732 	writel_relaxed(msg->data, pmu->reg_base + SMMU_PMCG_IRQ_CFG1);
733 	writel_relaxed(MSI_CFG2_MEMATTR_DEVICE_nGnRE,
734 		       pmu->reg_base + SMMU_PMCG_IRQ_CFG2);
735 }
736 
737 static void smmu_pmu_setup_msi(struct smmu_pmu *pmu)
738 {
739 	struct device *dev = pmu->dev;
740 	int ret;
741 
742 	/* Clear MSI address reg */
743 	writeq_relaxed(0, pmu->reg_base + SMMU_PMCG_IRQ_CFG0);
744 
745 	/* MSI supported or not */
746 	if (!(readl(pmu->reg_base + SMMU_PMCG_CFGR) & SMMU_PMCG_CFGR_MSI))
747 		return;
748 
749 	ret = platform_msi_domain_alloc_irqs(dev, 1, smmu_pmu_write_msi_msg);
750 	if (ret) {
751 		dev_warn(dev, "failed to allocate MSIs\n");
752 		return;
753 	}
754 
755 	pmu->irq = msi_get_virq(dev, 0);
756 
757 	/* Add callback to free MSIs on teardown */
758 	devm_add_action(dev, smmu_pmu_free_msis, dev);
759 }
760 
761 static int smmu_pmu_setup_irq(struct smmu_pmu *pmu)
762 {
763 	unsigned long flags = IRQF_NOBALANCING | IRQF_SHARED | IRQF_NO_THREAD;
764 	int irq, ret = -ENXIO;
765 
766 	smmu_pmu_setup_msi(pmu);
767 
768 	irq = pmu->irq;
769 	if (irq)
770 		ret = devm_request_irq(pmu->dev, irq, smmu_pmu_handle_irq,
771 				       flags, "smmuv3-pmu", pmu);
772 	return ret;
773 }
774 
775 static void smmu_pmu_reset(struct smmu_pmu *smmu_pmu)
776 {
777 	u64 counter_present_mask = GENMASK_ULL(smmu_pmu->num_counters - 1, 0);
778 
779 	smmu_pmu_disable(&smmu_pmu->pmu);
780 
781 	/* Disable counter and interrupt */
782 	writeq_relaxed(counter_present_mask,
783 		       smmu_pmu->reg_base + SMMU_PMCG_CNTENCLR0);
784 	writeq_relaxed(counter_present_mask,
785 		       smmu_pmu->reg_base + SMMU_PMCG_INTENCLR0);
786 	writeq_relaxed(counter_present_mask,
787 		       smmu_pmu->reloc_base + SMMU_PMCG_OVSCLR0);
788 }
789 
790 static void smmu_pmu_get_acpi_options(struct smmu_pmu *smmu_pmu)
791 {
792 	u32 model;
793 
794 	model = *(u32 *)dev_get_platdata(smmu_pmu->dev);
795 
796 	switch (model) {
797 	case IORT_SMMU_V3_PMCG_HISI_HIP08:
798 		/* HiSilicon Erratum 162001800 */
799 		smmu_pmu->options |= SMMU_PMCG_EVCNTR_RDONLY | SMMU_PMCG_HARDEN_DISABLE;
800 		break;
801 	case IORT_SMMU_V3_PMCG_HISI_HIP09:
802 		smmu_pmu->options |= SMMU_PMCG_HARDEN_DISABLE;
803 		break;
804 	}
805 
806 	dev_notice(smmu_pmu->dev, "option mask 0x%x\n", smmu_pmu->options);
807 }
808 
809 static bool smmu_pmu_coresight_id_regs(struct smmu_pmu *smmu_pmu)
810 {
811 	return of_device_is_compatible(smmu_pmu->dev->of_node,
812 				       "arm,mmu-600-pmcg");
813 }
814 
815 static void smmu_pmu_get_iidr(struct smmu_pmu *smmu_pmu)
816 {
817 	u32 iidr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_IIDR);
818 
819 	if (!iidr && smmu_pmu_coresight_id_regs(smmu_pmu)) {
820 		u32 pidr0 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR0);
821 		u32 pidr1 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR1);
822 		u32 pidr2 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR2);
823 		u32 pidr3 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR3);
824 		u32 pidr4 = readl(smmu_pmu->reg_base + SMMU_PMCG_PIDR4);
825 
826 		u32 productid = FIELD_GET(SMMU_PMCG_PIDR0_PART_0, pidr0) |
827 				(FIELD_GET(SMMU_PMCG_PIDR1_PART_1, pidr1) << 8);
828 		u32 variant = FIELD_GET(SMMU_PMCG_PIDR2_REVISION, pidr2);
829 		u32 revision = FIELD_GET(SMMU_PMCG_PIDR3_REVAND, pidr3);
830 		u32 implementer =
831 			FIELD_GET(SMMU_PMCG_PIDR1_DES_0, pidr1) |
832 			(FIELD_GET(SMMU_PMCG_PIDR2_DES_1, pidr2) << 4) |
833 			(FIELD_GET(SMMU_PMCG_PIDR4_DES_2, pidr4) << 8);
834 
835 		iidr = FIELD_PREP(SMMU_PMCG_IIDR_PRODUCTID, productid) |
836 		       FIELD_PREP(SMMU_PMCG_IIDR_VARIANT, variant) |
837 		       FIELD_PREP(SMMU_PMCG_IIDR_REVISION, revision) |
838 		       FIELD_PREP(SMMU_PMCG_IIDR_IMPLEMENTER, implementer);
839 	}
840 
841 	smmu_pmu->iidr = iidr;
842 }
843 
844 static int smmu_pmu_probe(struct platform_device *pdev)
845 {
846 	struct smmu_pmu *smmu_pmu;
847 	struct resource *res_0;
848 	u32 cfgr, reg_size;
849 	u64 ceid_64[2];
850 	int irq, err;
851 	char *name;
852 	struct device *dev = &pdev->dev;
853 
854 	smmu_pmu = devm_kzalloc(dev, sizeof(*smmu_pmu), GFP_KERNEL);
855 	if (!smmu_pmu)
856 		return -ENOMEM;
857 
858 	smmu_pmu->dev = dev;
859 	platform_set_drvdata(pdev, smmu_pmu);
860 
861 	smmu_pmu->pmu = (struct pmu) {
862 		.module		= THIS_MODULE,
863 		.task_ctx_nr    = perf_invalid_context,
864 		.pmu_enable	= smmu_pmu_enable,
865 		.pmu_disable	= smmu_pmu_disable,
866 		.event_init	= smmu_pmu_event_init,
867 		.add		= smmu_pmu_event_add,
868 		.del		= smmu_pmu_event_del,
869 		.start		= smmu_pmu_event_start,
870 		.stop		= smmu_pmu_event_stop,
871 		.read		= smmu_pmu_event_read,
872 		.attr_groups	= smmu_pmu_attr_grps,
873 		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
874 	};
875 
876 	smmu_pmu->reg_base = devm_platform_get_and_ioremap_resource(pdev, 0, &res_0);
877 	if (IS_ERR(smmu_pmu->reg_base))
878 		return PTR_ERR(smmu_pmu->reg_base);
879 
880 	cfgr = readl_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CFGR);
881 
882 	/* Determine if page 1 is present */
883 	if (cfgr & SMMU_PMCG_CFGR_RELOC_CTRS) {
884 		smmu_pmu->reloc_base = devm_platform_ioremap_resource(pdev, 1);
885 		if (IS_ERR(smmu_pmu->reloc_base))
886 			return PTR_ERR(smmu_pmu->reloc_base);
887 	} else {
888 		smmu_pmu->reloc_base = smmu_pmu->reg_base;
889 	}
890 
891 	irq = platform_get_irq_optional(pdev, 0);
892 	if (irq > 0)
893 		smmu_pmu->irq = irq;
894 
895 	ceid_64[0] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID0);
896 	ceid_64[1] = readq_relaxed(smmu_pmu->reg_base + SMMU_PMCG_CEID1);
897 	bitmap_from_arr32(smmu_pmu->supported_events, (u32 *)ceid_64,
898 			  SMMU_PMCG_ARCH_MAX_EVENTS);
899 
900 	smmu_pmu->num_counters = FIELD_GET(SMMU_PMCG_CFGR_NCTR, cfgr) + 1;
901 
902 	smmu_pmu->global_filter = !!(cfgr & SMMU_PMCG_CFGR_SID_FILTER_TYPE);
903 
904 	reg_size = FIELD_GET(SMMU_PMCG_CFGR_SIZE, cfgr);
905 	smmu_pmu->counter_mask = GENMASK_ULL(reg_size, 0);
906 
907 	smmu_pmu_reset(smmu_pmu);
908 
909 	err = smmu_pmu_setup_irq(smmu_pmu);
910 	if (err) {
911 		dev_err(dev, "Setup irq failed, PMU @%pa\n", &res_0->start);
912 		return err;
913 	}
914 
915 	smmu_pmu_get_iidr(smmu_pmu);
916 
917 	name = devm_kasprintf(&pdev->dev, GFP_KERNEL, "smmuv3_pmcg_%llx",
918 			      (res_0->start) >> SMMU_PMCG_PA_SHIFT);
919 	if (!name) {
920 		dev_err(dev, "Create name failed, PMU @%pa\n", &res_0->start);
921 		return -EINVAL;
922 	}
923 
924 	if (!dev->of_node)
925 		smmu_pmu_get_acpi_options(smmu_pmu);
926 
927 	/*
928 	 * For platforms suffer this quirk, the PMU disable sometimes fails to
929 	 * stop the counters. This will leads to inaccurate or error counting.
930 	 * Forcibly disable the counters with these quirk handler.
931 	 */
932 	if (smmu_pmu->options & SMMU_PMCG_HARDEN_DISABLE) {
933 		smmu_pmu->pmu.pmu_enable = smmu_pmu_enable_quirk_hip08_09;
934 		smmu_pmu->pmu.pmu_disable = smmu_pmu_disable_quirk_hip08_09;
935 	}
936 
937 	/* Pick one CPU to be the preferred one to use */
938 	smmu_pmu->on_cpu = raw_smp_processor_id();
939 	WARN_ON(irq_set_affinity(smmu_pmu->irq, cpumask_of(smmu_pmu->on_cpu)));
940 
941 	err = cpuhp_state_add_instance_nocalls(cpuhp_state_num,
942 					       &smmu_pmu->node);
943 	if (err) {
944 		dev_err(dev, "Error %d registering hotplug, PMU @%pa\n",
945 			err, &res_0->start);
946 		return err;
947 	}
948 
949 	err = perf_pmu_register(&smmu_pmu->pmu, name, -1);
950 	if (err) {
951 		dev_err(dev, "Error %d registering PMU @%pa\n",
952 			err, &res_0->start);
953 		goto out_unregister;
954 	}
955 
956 	dev_info(dev, "Registered PMU @ %pa using %d counters with %s filter settings\n",
957 		 &res_0->start, smmu_pmu->num_counters,
958 		 smmu_pmu->global_filter ? "Global(Counter0)" :
959 		 "Individual");
960 
961 	return 0;
962 
963 out_unregister:
964 	cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
965 	return err;
966 }
967 
968 static int smmu_pmu_remove(struct platform_device *pdev)
969 {
970 	struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
971 
972 	perf_pmu_unregister(&smmu_pmu->pmu);
973 	cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &smmu_pmu->node);
974 
975 	return 0;
976 }
977 
978 static void smmu_pmu_shutdown(struct platform_device *pdev)
979 {
980 	struct smmu_pmu *smmu_pmu = platform_get_drvdata(pdev);
981 
982 	smmu_pmu_disable(&smmu_pmu->pmu);
983 }
984 
985 #ifdef CONFIG_OF
986 static const struct of_device_id smmu_pmu_of_match[] = {
987 	{ .compatible = "arm,smmu-v3-pmcg" },
988 	{}
989 };
990 MODULE_DEVICE_TABLE(of, smmu_pmu_of_match);
991 #endif
992 
993 static struct platform_driver smmu_pmu_driver = {
994 	.driver = {
995 		.name = "arm-smmu-v3-pmcg",
996 		.of_match_table = of_match_ptr(smmu_pmu_of_match),
997 		.suppress_bind_attrs = true,
998 	},
999 	.probe = smmu_pmu_probe,
1000 	.remove = smmu_pmu_remove,
1001 	.shutdown = smmu_pmu_shutdown,
1002 };
1003 
1004 static int __init arm_smmu_pmu_init(void)
1005 {
1006 	int ret;
1007 
1008 	cpuhp_state_num = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1009 						  "perf/arm/pmcg:online",
1010 						  NULL,
1011 						  smmu_pmu_offline_cpu);
1012 	if (cpuhp_state_num < 0)
1013 		return cpuhp_state_num;
1014 
1015 	ret = platform_driver_register(&smmu_pmu_driver);
1016 	if (ret)
1017 		cpuhp_remove_multi_state(cpuhp_state_num);
1018 
1019 	return ret;
1020 }
1021 module_init(arm_smmu_pmu_init);
1022 
1023 static void __exit arm_smmu_pmu_exit(void)
1024 {
1025 	platform_driver_unregister(&smmu_pmu_driver);
1026 	cpuhp_remove_multi_state(cpuhp_state_num);
1027 }
1028 
1029 module_exit(arm_smmu_pmu_exit);
1030 
1031 MODULE_ALIAS("platform:arm-smmu-v3-pmcg");
1032 MODULE_DESCRIPTION("PMU driver for ARM SMMUv3 Performance Monitors Extension");
1033 MODULE_AUTHOR("Neil Leeder <nleeder@codeaurora.org>");
1034 MODULE_AUTHOR("Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>");
1035 MODULE_LICENSE("GPL v2");
1036