xref: /openbmc/linux/arch/s390/kernel/perf_cpum_cf.c (revision 4da722ca19f30f7db250db808d1ab1703607a932)
1 /*
2  * Performance event support for s390x - CPU-measurement Counter Facility
3  *
4  *  Copyright IBM Corp. 2012, 2017
5  *  Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License (version 2 only)
9  * as published by the Free Software Foundation.
10  */
11 #define KMSG_COMPONENT	"cpum_cf"
12 #define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt
13 
14 #include <linux/kernel.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/perf_event.h>
17 #include <linux/percpu.h>
18 #include <linux/notifier.h>
19 #include <linux/init.h>
20 #include <linux/export.h>
21 #include <asm/ctl_reg.h>
22 #include <asm/irq.h>
23 #include <asm/cpu_mf.h>
24 
25 enum cpumf_ctr_set {
26 	CPUMF_CTR_SET_BASIC   = 0,    /* Basic Counter Set */
27 	CPUMF_CTR_SET_USER    = 1,    /* Problem-State Counter Set */
28 	CPUMF_CTR_SET_CRYPTO  = 2,    /* Crypto-Activity Counter Set */
29 	CPUMF_CTR_SET_EXT     = 3,    /* Extended Counter Set */
30 	CPUMF_CTR_SET_MT_DIAG = 4,    /* MT-diagnostic Counter Set */
31 
32 	/* Maximum number of counter sets */
33 	CPUMF_CTR_SET_MAX,
34 };
35 
36 #define CPUMF_LCCTL_ENABLE_SHIFT    16
37 #define CPUMF_LCCTL_ACTCTL_SHIFT     0
38 static const u64 cpumf_state_ctl[CPUMF_CTR_SET_MAX] = {
39 	[CPUMF_CTR_SET_BASIC]	= 0x02,
40 	[CPUMF_CTR_SET_USER]	= 0x04,
41 	[CPUMF_CTR_SET_CRYPTO]	= 0x08,
42 	[CPUMF_CTR_SET_EXT]	= 0x01,
43 	[CPUMF_CTR_SET_MT_DIAG] = 0x20,
44 };
45 
46 static void ctr_set_enable(u64 *state, int ctr_set)
47 {
48 	*state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT;
49 }
50 static void ctr_set_disable(u64 *state, int ctr_set)
51 {
52 	*state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT);
53 }
54 static void ctr_set_start(u64 *state, int ctr_set)
55 {
56 	*state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT;
57 }
58 static void ctr_set_stop(u64 *state, int ctr_set)
59 {
60 	*state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT);
61 }
62 
63 /* Local CPUMF event structure */
64 struct cpu_hw_events {
65 	struct cpumf_ctr_info	info;
66 	atomic_t		ctr_set[CPUMF_CTR_SET_MAX];
67 	u64			state, tx_state;
68 	unsigned int		flags;
69 	unsigned int		txn_flags;
70 };
71 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
72 	.ctr_set = {
73 		[CPUMF_CTR_SET_BASIC]	= ATOMIC_INIT(0),
74 		[CPUMF_CTR_SET_USER]	= ATOMIC_INIT(0),
75 		[CPUMF_CTR_SET_CRYPTO]	= ATOMIC_INIT(0),
76 		[CPUMF_CTR_SET_EXT]	= ATOMIC_INIT(0),
77 		[CPUMF_CTR_SET_MT_DIAG] = ATOMIC_INIT(0),
78 	},
79 	.state = 0,
80 	.flags = 0,
81 	.txn_flags = 0,
82 };
83 
84 static enum cpumf_ctr_set get_counter_set(u64 event)
85 {
86 	int set = CPUMF_CTR_SET_MAX;
87 
88 	if (event < 32)
89 		set = CPUMF_CTR_SET_BASIC;
90 	else if (event < 64)
91 		set = CPUMF_CTR_SET_USER;
92 	else if (event < 128)
93 		set = CPUMF_CTR_SET_CRYPTO;
94 	else if (event < 256)
95 		set = CPUMF_CTR_SET_EXT;
96 	else if (event >= 448 && event < 496)
97 		set = CPUMF_CTR_SET_MT_DIAG;
98 
99 	return set;
100 }
101 
102 static int validate_ctr_version(const struct hw_perf_event *hwc)
103 {
104 	struct cpu_hw_events *cpuhw;
105 	int err = 0;
106 	u16 mtdiag_ctl;
107 
108 	cpuhw = &get_cpu_var(cpu_hw_events);
109 
110 	/* check required version for counter sets */
111 	switch (hwc->config_base) {
112 	case CPUMF_CTR_SET_BASIC:
113 	case CPUMF_CTR_SET_USER:
114 		if (cpuhw->info.cfvn < 1)
115 			err = -EOPNOTSUPP;
116 		break;
117 	case CPUMF_CTR_SET_CRYPTO:
118 	case CPUMF_CTR_SET_EXT:
119 		if (cpuhw->info.csvn < 1)
120 			err = -EOPNOTSUPP;
121 		if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
122 		    (cpuhw->info.csvn == 2 && hwc->config > 175) ||
123 		    (cpuhw->info.csvn  > 2 && hwc->config > 255))
124 			err = -EOPNOTSUPP;
125 		break;
126 	case CPUMF_CTR_SET_MT_DIAG:
127 		if (cpuhw->info.csvn <= 3)
128 			err = -EOPNOTSUPP;
129 		/*
130 		 * MT-diagnostic counters are read-only.  The counter set
131 		 * is automatically enabled and activated on all CPUs with
132 		 * multithreading (SMT).  Deactivation of multithreading
133 		 * also disables the counter set.  State changes are ignored
134 		 * by lcctl().	Because Linux controls SMT enablement through
135 		 * a kernel parameter only, the counter set is either disabled
136 		 * or enabled and active.
137 		 *
138 		 * Thus, the counters can only be used if SMT is on and the
139 		 * counter set is enabled and active.
140 		 */
141 		mtdiag_ctl = cpumf_state_ctl[CPUMF_CTR_SET_MT_DIAG];
142 		if (!((cpuhw->info.auth_ctl & mtdiag_ctl) &&
143 		      (cpuhw->info.enable_ctl & mtdiag_ctl) &&
144 		      (cpuhw->info.act_ctl & mtdiag_ctl)))
145 			err = -EOPNOTSUPP;
146 		break;
147 	}
148 
149 	put_cpu_var(cpu_hw_events);
150 	return err;
151 }
152 
153 static int validate_ctr_auth(const struct hw_perf_event *hwc)
154 {
155 	struct cpu_hw_events *cpuhw;
156 	u64 ctrs_state;
157 	int err = 0;
158 
159 	cpuhw = &get_cpu_var(cpu_hw_events);
160 
161 	/* Check authorization for cpu counter sets.
162 	 * If the particular CPU counter set is not authorized,
163 	 * return with -ENOENT in order to fall back to other
164 	 * PMUs that might suffice the event request.
165 	 */
166 	ctrs_state = cpumf_state_ctl[hwc->config_base];
167 	if (!(ctrs_state & cpuhw->info.auth_ctl))
168 		err = -ENOENT;
169 
170 	put_cpu_var(cpu_hw_events);
171 	return err;
172 }
173 
174 /*
175  * Change the CPUMF state to active.
176  * Enable and activate the CPU-counter sets according
177  * to the per-cpu control state.
178  */
179 static void cpumf_pmu_enable(struct pmu *pmu)
180 {
181 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
182 	int err;
183 
184 	if (cpuhw->flags & PMU_F_ENABLED)
185 		return;
186 
187 	err = lcctl(cpuhw->state);
188 	if (err) {
189 		pr_err("Enabling the performance measuring unit "
190 		       "failed with rc=%x\n", err);
191 		return;
192 	}
193 
194 	cpuhw->flags |= PMU_F_ENABLED;
195 }
196 
197 /*
198  * Change the CPUMF state to inactive.
199  * Disable and enable (inactive) the CPU-counter sets according
200  * to the per-cpu control state.
201  */
202 static void cpumf_pmu_disable(struct pmu *pmu)
203 {
204 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
205 	int err;
206 	u64 inactive;
207 
208 	if (!(cpuhw->flags & PMU_F_ENABLED))
209 		return;
210 
211 	inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
212 	err = lcctl(inactive);
213 	if (err) {
214 		pr_err("Disabling the performance measuring unit "
215 		       "failed with rc=%x\n", err);
216 		return;
217 	}
218 
219 	cpuhw->flags &= ~PMU_F_ENABLED;
220 }
221 
222 
223 /* Number of perf events counting hardware events */
224 static atomic_t num_events = ATOMIC_INIT(0);
225 /* Used to avoid races in calling reserve/release_cpumf_hardware */
226 static DEFINE_MUTEX(pmc_reserve_mutex);
227 
228 /* CPU-measurement alerts for the counter facility */
229 static void cpumf_measurement_alert(struct ext_code ext_code,
230 				    unsigned int alert, unsigned long unused)
231 {
232 	struct cpu_hw_events *cpuhw;
233 
234 	if (!(alert & CPU_MF_INT_CF_MASK))
235 		return;
236 
237 	inc_irq_stat(IRQEXT_CMC);
238 	cpuhw = this_cpu_ptr(&cpu_hw_events);
239 
240 	/* Measurement alerts are shared and might happen when the PMU
241 	 * is not reserved.  Ignore these alerts in this case. */
242 	if (!(cpuhw->flags & PMU_F_RESERVED))
243 		return;
244 
245 	/* counter authorization change alert */
246 	if (alert & CPU_MF_INT_CF_CACA)
247 		qctri(&cpuhw->info);
248 
249 	/* loss of counter data alert */
250 	if (alert & CPU_MF_INT_CF_LCDA)
251 		pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
252 
253 	/* loss of MT counter data alert */
254 	if (alert & CPU_MF_INT_CF_MTDA)
255 		pr_warn("CPU[%i] MT counter data was lost\n",
256 			smp_processor_id());
257 }
258 
259 #define PMC_INIT      0
260 #define PMC_RELEASE   1
261 static void setup_pmc_cpu(void *flags)
262 {
263 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
264 
265 	switch (*((int *) flags)) {
266 	case PMC_INIT:
267 		memset(&cpuhw->info, 0, sizeof(cpuhw->info));
268 		qctri(&cpuhw->info);
269 		cpuhw->flags |= PMU_F_RESERVED;
270 		break;
271 
272 	case PMC_RELEASE:
273 		cpuhw->flags &= ~PMU_F_RESERVED;
274 		break;
275 	}
276 
277 	/* Disable CPU counter sets */
278 	lcctl(0);
279 }
280 
281 /* Initialize the CPU-measurement facility */
282 static int reserve_pmc_hardware(void)
283 {
284 	int flags = PMC_INIT;
285 
286 	on_each_cpu(setup_pmc_cpu, &flags, 1);
287 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
288 
289 	return 0;
290 }
291 
292 /* Release the CPU-measurement facility */
293 static void release_pmc_hardware(void)
294 {
295 	int flags = PMC_RELEASE;
296 
297 	on_each_cpu(setup_pmc_cpu, &flags, 1);
298 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
299 }
300 
301 /* Release the PMU if event is the last perf event */
302 static void hw_perf_event_destroy(struct perf_event *event)
303 {
304 	if (!atomic_add_unless(&num_events, -1, 1)) {
305 		mutex_lock(&pmc_reserve_mutex);
306 		if (atomic_dec_return(&num_events) == 0)
307 			release_pmc_hardware();
308 		mutex_unlock(&pmc_reserve_mutex);
309 	}
310 }
311 
312 /* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
313 static const int cpumf_generic_events_basic[] = {
314 	[PERF_COUNT_HW_CPU_CYCLES]	    = 0,
315 	[PERF_COUNT_HW_INSTRUCTIONS]	    = 1,
316 	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
317 	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
318 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
319 	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
320 	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
321 };
322 /* CPUMF <-> perf event mappings for userspace (problem-state set) */
323 static const int cpumf_generic_events_user[] = {
324 	[PERF_COUNT_HW_CPU_CYCLES]	    = 32,
325 	[PERF_COUNT_HW_INSTRUCTIONS]	    = 33,
326 	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
327 	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
328 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
329 	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
330 	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
331 };
332 
333 static int __hw_perf_event_init(struct perf_event *event)
334 {
335 	struct perf_event_attr *attr = &event->attr;
336 	struct hw_perf_event *hwc = &event->hw;
337 	enum cpumf_ctr_set set;
338 	int err;
339 	u64 ev;
340 
341 	switch (attr->type) {
342 	case PERF_TYPE_RAW:
343 		/* Raw events are used to access counters directly,
344 		 * hence do not permit excludes */
345 		if (attr->exclude_kernel || attr->exclude_user ||
346 		    attr->exclude_hv)
347 			return -EOPNOTSUPP;
348 		ev = attr->config;
349 		break;
350 
351 	case PERF_TYPE_HARDWARE:
352 		ev = attr->config;
353 		/* Count user space (problem-state) only */
354 		if (!attr->exclude_user && attr->exclude_kernel) {
355 			if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
356 				return -EOPNOTSUPP;
357 			ev = cpumf_generic_events_user[ev];
358 
359 		/* No support for kernel space counters only */
360 		} else if (!attr->exclude_kernel && attr->exclude_user) {
361 			return -EOPNOTSUPP;
362 
363 		/* Count user and kernel space */
364 		} else {
365 			if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
366 				return -EOPNOTSUPP;
367 			ev = cpumf_generic_events_basic[ev];
368 		}
369 		break;
370 
371 	default:
372 		return -ENOENT;
373 	}
374 
375 	if (ev == -1)
376 		return -ENOENT;
377 
378 	if (ev > PERF_CPUM_CF_MAX_CTR)
379 		return -EINVAL;
380 
381 	/* Obtain the counter set to which the specified counter belongs */
382 	set = get_counter_set(ev);
383 	switch (set) {
384 	case CPUMF_CTR_SET_BASIC:
385 	case CPUMF_CTR_SET_USER:
386 	case CPUMF_CTR_SET_CRYPTO:
387 	case CPUMF_CTR_SET_EXT:
388 	case CPUMF_CTR_SET_MT_DIAG:
389 		/*
390 		 * Use the hardware perf event structure to store the
391 		 * counter number in the 'config' member and the counter
392 		 * set number in the 'config_base'.  The counter set number
393 		 * is then later used to enable/disable the counter(s).
394 		 */
395 		hwc->config = ev;
396 		hwc->config_base = set;
397 		break;
398 	case CPUMF_CTR_SET_MAX:
399 		/* The counter could not be associated to a counter set */
400 		return -EINVAL;
401 	};
402 
403 	/* Initialize for using the CPU-measurement counter facility */
404 	if (!atomic_inc_not_zero(&num_events)) {
405 		mutex_lock(&pmc_reserve_mutex);
406 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
407 			err = -EBUSY;
408 		else
409 			atomic_inc(&num_events);
410 		mutex_unlock(&pmc_reserve_mutex);
411 	}
412 	event->destroy = hw_perf_event_destroy;
413 
414 	/* Finally, validate version and authorization of the counter set */
415 	err = validate_ctr_auth(hwc);
416 	if (!err)
417 		err = validate_ctr_version(hwc);
418 
419 	return err;
420 }
421 
422 static int cpumf_pmu_event_init(struct perf_event *event)
423 {
424 	int err;
425 
426 	switch (event->attr.type) {
427 	case PERF_TYPE_HARDWARE:
428 	case PERF_TYPE_HW_CACHE:
429 	case PERF_TYPE_RAW:
430 		err = __hw_perf_event_init(event);
431 		break;
432 	default:
433 		return -ENOENT;
434 	}
435 
436 	if (unlikely(err) && event->destroy)
437 		event->destroy(event);
438 
439 	return err;
440 }
441 
442 static int hw_perf_event_reset(struct perf_event *event)
443 {
444 	u64 prev, new;
445 	int err;
446 
447 	do {
448 		prev = local64_read(&event->hw.prev_count);
449 		err = ecctr(event->hw.config, &new);
450 		if (err) {
451 			if (err != 3)
452 				break;
453 			/* The counter is not (yet) available. This
454 			 * might happen if the counter set to which
455 			 * this counter belongs is in the disabled
456 			 * state.
457 			 */
458 			new = 0;
459 		}
460 	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
461 
462 	return err;
463 }
464 
465 static void hw_perf_event_update(struct perf_event *event)
466 {
467 	u64 prev, new, delta;
468 	int err;
469 
470 	do {
471 		prev = local64_read(&event->hw.prev_count);
472 		err = ecctr(event->hw.config, &new);
473 		if (err)
474 			return;
475 	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
476 
477 	delta = (prev <= new) ? new - prev
478 			      : (-1ULL - prev) + new + 1;	 /* overflow */
479 	local64_add(delta, &event->count);
480 }
481 
482 static void cpumf_pmu_read(struct perf_event *event)
483 {
484 	if (event->hw.state & PERF_HES_STOPPED)
485 		return;
486 
487 	hw_perf_event_update(event);
488 }
489 
490 static void cpumf_pmu_start(struct perf_event *event, int flags)
491 {
492 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
493 	struct hw_perf_event *hwc = &event->hw;
494 
495 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
496 		return;
497 
498 	if (WARN_ON_ONCE(hwc->config == -1))
499 		return;
500 
501 	if (flags & PERF_EF_RELOAD)
502 		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
503 
504 	hwc->state = 0;
505 
506 	/* (Re-)enable and activate the counter set */
507 	ctr_set_enable(&cpuhw->state, hwc->config_base);
508 	ctr_set_start(&cpuhw->state, hwc->config_base);
509 
510 	/* The counter set to which this counter belongs can be already active.
511 	 * Because all counters in a set are active, the event->hw.prev_count
512 	 * needs to be synchronized.  At this point, the counter set can be in
513 	 * the inactive or disabled state.
514 	 */
515 	hw_perf_event_reset(event);
516 
517 	/* increment refcount for this counter set */
518 	atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
519 }
520 
521 static void cpumf_pmu_stop(struct perf_event *event, int flags)
522 {
523 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
524 	struct hw_perf_event *hwc = &event->hw;
525 
526 	if (!(hwc->state & PERF_HES_STOPPED)) {
527 		/* Decrement reference count for this counter set and if this
528 		 * is the last used counter in the set, clear activation
529 		 * control and set the counter set state to inactive.
530 		 */
531 		if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
532 			ctr_set_stop(&cpuhw->state, hwc->config_base);
533 		event->hw.state |= PERF_HES_STOPPED;
534 	}
535 
536 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
537 		hw_perf_event_update(event);
538 		event->hw.state |= PERF_HES_UPTODATE;
539 	}
540 }
541 
542 static int cpumf_pmu_add(struct perf_event *event, int flags)
543 {
544 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
545 
546 	/* Check authorization for the counter set to which this
547 	 * counter belongs.
548 	 * For group events transaction, the authorization check is
549 	 * done in cpumf_pmu_commit_txn().
550 	 */
551 	if (!(cpuhw->txn_flags & PERF_PMU_TXN_ADD))
552 		if (validate_ctr_auth(&event->hw))
553 			return -ENOENT;
554 
555 	ctr_set_enable(&cpuhw->state, event->hw.config_base);
556 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
557 
558 	if (flags & PERF_EF_START)
559 		cpumf_pmu_start(event, PERF_EF_RELOAD);
560 
561 	perf_event_update_userpage(event);
562 
563 	return 0;
564 }
565 
566 static void cpumf_pmu_del(struct perf_event *event, int flags)
567 {
568 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
569 
570 	cpumf_pmu_stop(event, PERF_EF_UPDATE);
571 
572 	/* Check if any counter in the counter set is still used.  If not used,
573 	 * change the counter set to the disabled state.  This also clears the
574 	 * content of all counters in the set.
575 	 *
576 	 * When a new perf event has been added but not yet started, this can
577 	 * clear enable control and resets all counters in a set.  Therefore,
578 	 * cpumf_pmu_start() always has to reenable a counter set.
579 	 */
580 	if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
581 		ctr_set_disable(&cpuhw->state, event->hw.config_base);
582 
583 	perf_event_update_userpage(event);
584 }
585 
586 /*
587  * Start group events scheduling transaction.
588  * Set flags to perform a single test at commit time.
589  *
590  * We only support PERF_PMU_TXN_ADD transactions. Save the
591  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
592  * transactions.
593  */
594 static void cpumf_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
595 {
596 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
597 
598 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
599 
600 	cpuhw->txn_flags = txn_flags;
601 	if (txn_flags & ~PERF_PMU_TXN_ADD)
602 		return;
603 
604 	perf_pmu_disable(pmu);
605 	cpuhw->tx_state = cpuhw->state;
606 }
607 
608 /*
609  * Stop and cancel a group events scheduling tranctions.
610  * Assumes cpumf_pmu_del() is called for each successful added
611  * cpumf_pmu_add() during the transaction.
612  */
613 static void cpumf_pmu_cancel_txn(struct pmu *pmu)
614 {
615 	unsigned int txn_flags;
616 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
617 
618 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
619 
620 	txn_flags = cpuhw->txn_flags;
621 	cpuhw->txn_flags = 0;
622 	if (txn_flags & ~PERF_PMU_TXN_ADD)
623 		return;
624 
625 	WARN_ON(cpuhw->tx_state != cpuhw->state);
626 
627 	perf_pmu_enable(pmu);
628 }
629 
630 /*
631  * Commit the group events scheduling transaction.  On success, the
632  * transaction is closed.   On error, the transaction is kept open
633  * until cpumf_pmu_cancel_txn() is called.
634  */
635 static int cpumf_pmu_commit_txn(struct pmu *pmu)
636 {
637 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
638 	u64 state;
639 
640 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
641 
642 	if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
643 		cpuhw->txn_flags = 0;
644 		return 0;
645 	}
646 
647 	/* check if the updated state can be scheduled */
648 	state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
649 	state >>= CPUMF_LCCTL_ENABLE_SHIFT;
650 	if ((state & cpuhw->info.auth_ctl) != state)
651 		return -ENOENT;
652 
653 	cpuhw->txn_flags = 0;
654 	perf_pmu_enable(pmu);
655 	return 0;
656 }
657 
658 /* Performance monitoring unit for s390x */
659 static struct pmu cpumf_pmu = {
660 	.task_ctx_nr  = perf_sw_context,
661 	.capabilities = PERF_PMU_CAP_NO_INTERRUPT,
662 	.pmu_enable   = cpumf_pmu_enable,
663 	.pmu_disable  = cpumf_pmu_disable,
664 	.event_init   = cpumf_pmu_event_init,
665 	.add	      = cpumf_pmu_add,
666 	.del	      = cpumf_pmu_del,
667 	.start	      = cpumf_pmu_start,
668 	.stop	      = cpumf_pmu_stop,
669 	.read	      = cpumf_pmu_read,
670 	.start_txn    = cpumf_pmu_start_txn,
671 	.commit_txn   = cpumf_pmu_commit_txn,
672 	.cancel_txn   = cpumf_pmu_cancel_txn,
673 };
674 
675 static int cpumf_pmf_setup(unsigned int cpu, int flags)
676 {
677 	local_irq_disable();
678 	setup_pmc_cpu(&flags);
679 	local_irq_enable();
680 	return 0;
681 }
682 
683 static int s390_pmu_online_cpu(unsigned int cpu)
684 {
685 	return cpumf_pmf_setup(cpu, PMC_INIT);
686 }
687 
688 static int s390_pmu_offline_cpu(unsigned int cpu)
689 {
690 	return cpumf_pmf_setup(cpu, PMC_RELEASE);
691 }
692 
693 static int __init cpumf_pmu_init(void)
694 {
695 	int rc;
696 
697 	if (!cpum_cf_avail())
698 		return -ENODEV;
699 
700 	/* clear bit 15 of cr0 to unauthorize problem-state to
701 	 * extract measurement counters */
702 	ctl_clear_bit(0, 48);
703 
704 	/* register handler for measurement-alert interruptions */
705 	rc = register_external_irq(EXT_IRQ_MEASURE_ALERT,
706 				   cpumf_measurement_alert);
707 	if (rc) {
708 		pr_err("Registering for CPU-measurement alerts "
709 		       "failed with rc=%i\n", rc);
710 		return rc;
711 	}
712 
713 	cpumf_pmu.attr_groups = cpumf_cf_event_group();
714 	rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
715 	if (rc) {
716 		pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
717 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
718 					cpumf_measurement_alert);
719 		return rc;
720 	}
721 	return cpuhp_setup_state(CPUHP_AP_PERF_S390_CF_ONLINE,
722 				 "perf/s390/cf:online",
723 				 s390_pmu_online_cpu, s390_pmu_offline_cpu);
724 }
725 early_initcall(cpumf_pmu_init);
726