Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance event support for s390x - CPU-measurement Counter Facility
4 *
5 * Copyright IBM Corp. 2012, 2019
6 * Author(s): Hendrik Brueckner <brueckner@linux.ibm.com>
7 */
8#define KMSG_COMPONENT "cpum_cf"
9#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
10
11#include <linux/kernel.h>
12#include <linux/kernel_stat.h>
13#include <linux/percpu.h>
14#include <linux/notifier.h>
15#include <linux/init.h>
16#include <linux/export.h>
17#include <asm/cpu_mcf.h>
18
19static enum cpumf_ctr_set get_counter_set(u64 event)
20{
21 int set = CPUMF_CTR_SET_MAX;
22
23 if (event < 32)
24 set = CPUMF_CTR_SET_BASIC;
25 else if (event < 64)
26 set = CPUMF_CTR_SET_USER;
27 else if (event < 128)
28 set = CPUMF_CTR_SET_CRYPTO;
29 else if (event < 288)
30 set = CPUMF_CTR_SET_EXT;
31 else if (event >= 448 && event < 496)
32 set = CPUMF_CTR_SET_MT_DIAG;
33
34 return set;
35}
36
37static int validate_ctr_version(const struct hw_perf_event *hwc)
38{
39 struct cpu_cf_events *cpuhw;
40 int err = 0;
41 u16 mtdiag_ctl;
42
43 cpuhw = &get_cpu_var(cpu_cf_events);
44
45 /* check required version for counter sets */
46 switch (hwc->config_base) {
47 case CPUMF_CTR_SET_BASIC:
48 case CPUMF_CTR_SET_USER:
49 if (cpuhw->info.cfvn < 1)
50 err = -EOPNOTSUPP;
51 break;
52 case CPUMF_CTR_SET_CRYPTO:
53 if ((cpuhw->info.csvn >= 1 && cpuhw->info.csvn <= 5 &&
54 hwc->config > 79) ||
55 (cpuhw->info.csvn >= 6 && hwc->config > 83))
56 err = -EOPNOTSUPP;
57 break;
58 case CPUMF_CTR_SET_EXT:
59 if (cpuhw->info.csvn < 1)
60 err = -EOPNOTSUPP;
61 if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
62 (cpuhw->info.csvn == 2 && hwc->config > 175) ||
63 (cpuhw->info.csvn >= 3 && cpuhw->info.csvn <= 5
64 && hwc->config > 255) ||
65 (cpuhw->info.csvn >= 6 && hwc->config > 287))
66 err = -EOPNOTSUPP;
67 break;
68 case CPUMF_CTR_SET_MT_DIAG:
69 if (cpuhw->info.csvn <= 3)
70 err = -EOPNOTSUPP;
71 /*
72 * MT-diagnostic counters are read-only. The counter set
73 * is automatically enabled and activated on all CPUs with
74 * multithreading (SMT). Deactivation of multithreading
75 * also disables the counter set. State changes are ignored
76 * by lcctl(). Because Linux controls SMT enablement through
77 * a kernel parameter only, the counter set is either disabled
78 * or enabled and active.
79 *
80 * Thus, the counters can only be used if SMT is on and the
81 * counter set is enabled and active.
82 */
83 mtdiag_ctl = cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG];
84 if (!((cpuhw->info.auth_ctl & mtdiag_ctl) &&
85 (cpuhw->info.enable_ctl & mtdiag_ctl) &&
86 (cpuhw->info.act_ctl & mtdiag_ctl)))
87 err = -EOPNOTSUPP;
88 break;
89 }
90
91 put_cpu_var(cpu_cf_events);
92 return err;
93}
94
95static int validate_ctr_auth(const struct hw_perf_event *hwc)
96{
97 struct cpu_cf_events *cpuhw;
98 u64 ctrs_state;
99 int err = 0;
100
101 cpuhw = &get_cpu_var(cpu_cf_events);
102
103 /* Check authorization for cpu counter sets.
104 * If the particular CPU counter set is not authorized,
105 * return with -ENOENT in order to fall back to other
106 * PMUs that might suffice the event request.
107 */
108 ctrs_state = cpumf_ctr_ctl[hwc->config_base];
109 if (!(ctrs_state & cpuhw->info.auth_ctl))
110 err = -ENOENT;
111
112 put_cpu_var(cpu_cf_events);
113 return err;
114}
115
116/*
117 * Change the CPUMF state to active.
118 * Enable and activate the CPU-counter sets according
119 * to the per-cpu control state.
120 */
121static void cpumf_pmu_enable(struct pmu *pmu)
122{
123 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
124 int err;
125
126 if (cpuhw->flags & PMU_F_ENABLED)
127 return;
128
129 err = lcctl(cpuhw->state);
130 if (err) {
131 pr_err("Enabling the performance measuring unit "
132 "failed with rc=%x\n", err);
133 return;
134 }
135
136 cpuhw->flags |= PMU_F_ENABLED;
137}
138
139/*
140 * Change the CPUMF state to inactive.
141 * Disable and enable (inactive) the CPU-counter sets according
142 * to the per-cpu control state.
143 */
144static void cpumf_pmu_disable(struct pmu *pmu)
145{
146 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
147 int err;
148 u64 inactive;
149
150 if (!(cpuhw->flags & PMU_F_ENABLED))
151 return;
152
153 inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
154 err = lcctl(inactive);
155 if (err) {
156 pr_err("Disabling the performance measuring unit "
157 "failed with rc=%x\n", err);
158 return;
159 }
160
161 cpuhw->flags &= ~PMU_F_ENABLED;
162}
163
164
165/* Number of perf events counting hardware events */
166static atomic_t num_events = ATOMIC_INIT(0);
167/* Used to avoid races in calling reserve/release_cpumf_hardware */
168static DEFINE_MUTEX(pmc_reserve_mutex);
169
170/* Release the PMU if event is the last perf event */
171static void hw_perf_event_destroy(struct perf_event *event)
172{
173 if (!atomic_add_unless(&num_events, -1, 1)) {
174 mutex_lock(&pmc_reserve_mutex);
175 if (atomic_dec_return(&num_events) == 0)
176 __kernel_cpumcf_end();
177 mutex_unlock(&pmc_reserve_mutex);
178 }
179}
180
181/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
182static const int cpumf_generic_events_basic[] = {
183 [PERF_COUNT_HW_CPU_CYCLES] = 0,
184 [PERF_COUNT_HW_INSTRUCTIONS] = 1,
185 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
186 [PERF_COUNT_HW_CACHE_MISSES] = -1,
187 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
188 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
189 [PERF_COUNT_HW_BUS_CYCLES] = -1,
190};
191/* CPUMF <-> perf event mappings for userspace (problem-state set) */
192static const int cpumf_generic_events_user[] = {
193 [PERF_COUNT_HW_CPU_CYCLES] = 32,
194 [PERF_COUNT_HW_INSTRUCTIONS] = 33,
195 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
196 [PERF_COUNT_HW_CACHE_MISSES] = -1,
197 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
198 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
199 [PERF_COUNT_HW_BUS_CYCLES] = -1,
200};
201
202static int __hw_perf_event_init(struct perf_event *event)
203{
204 struct perf_event_attr *attr = &event->attr;
205 struct hw_perf_event *hwc = &event->hw;
206 enum cpumf_ctr_set set;
207 int err = 0;
208 u64 ev;
209
210 switch (attr->type) {
211 case PERF_TYPE_RAW:
212 /* Raw events are used to access counters directly,
213 * hence do not permit excludes */
214 if (attr->exclude_kernel || attr->exclude_user ||
215 attr->exclude_hv)
216 return -EOPNOTSUPP;
217 ev = attr->config;
218 break;
219
220 case PERF_TYPE_HARDWARE:
221 if (is_sampling_event(event)) /* No sampling support */
222 return -ENOENT;
223 ev = attr->config;
224 /* Count user space (problem-state) only */
225 if (!attr->exclude_user && attr->exclude_kernel) {
226 if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
227 return -EOPNOTSUPP;
228 ev = cpumf_generic_events_user[ev];
229
230 /* No support for kernel space counters only */
231 } else if (!attr->exclude_kernel && attr->exclude_user) {
232 return -EOPNOTSUPP;
233
234 /* Count user and kernel space */
235 } else {
236 if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
237 return -EOPNOTSUPP;
238 ev = cpumf_generic_events_basic[ev];
239 }
240 break;
241
242 default:
243 return -ENOENT;
244 }
245
246 if (ev == -1)
247 return -ENOENT;
248
249 if (ev > PERF_CPUM_CF_MAX_CTR)
250 return -ENOENT;
251
252 /* Obtain the counter set to which the specified counter belongs */
253 set = get_counter_set(ev);
254 switch (set) {
255 case CPUMF_CTR_SET_BASIC:
256 case CPUMF_CTR_SET_USER:
257 case CPUMF_CTR_SET_CRYPTO:
258 case CPUMF_CTR_SET_EXT:
259 case CPUMF_CTR_SET_MT_DIAG:
260 /*
261 * Use the hardware perf event structure to store the
262 * counter number in the 'config' member and the counter
263 * set number in the 'config_base'. The counter set number
264 * is then later used to enable/disable the counter(s).
265 */
266 hwc->config = ev;
267 hwc->config_base = set;
268 break;
269 case CPUMF_CTR_SET_MAX:
270 /* The counter could not be associated to a counter set */
271 return -EINVAL;
272 };
273
274 /* Initialize for using the CPU-measurement counter facility */
275 if (!atomic_inc_not_zero(&num_events)) {
276 mutex_lock(&pmc_reserve_mutex);
277 if (atomic_read(&num_events) == 0 && __kernel_cpumcf_begin())
278 err = -EBUSY;
279 else
280 atomic_inc(&num_events);
281 mutex_unlock(&pmc_reserve_mutex);
282 }
283 if (err)
284 return err;
285 event->destroy = hw_perf_event_destroy;
286
287 /* Finally, validate version and authorization of the counter set */
288 err = validate_ctr_auth(hwc);
289 if (!err)
290 err = validate_ctr_version(hwc);
291
292 return err;
293}
294
295static int cpumf_pmu_event_init(struct perf_event *event)
296{
297 int err;
298
299 switch (event->attr.type) {
300 case PERF_TYPE_HARDWARE:
301 case PERF_TYPE_HW_CACHE:
302 case PERF_TYPE_RAW:
303 err = __hw_perf_event_init(event);
304 break;
305 default:
306 return -ENOENT;
307 }
308
309 if (unlikely(err) && event->destroy)
310 event->destroy(event);
311
312 return err;
313}
314
315static int hw_perf_event_reset(struct perf_event *event)
316{
317 u64 prev, new;
318 int err;
319
320 do {
321 prev = local64_read(&event->hw.prev_count);
322 err = ecctr(event->hw.config, &new);
323 if (err) {
324 if (err != 3)
325 break;
326 /* The counter is not (yet) available. This
327 * might happen if the counter set to which
328 * this counter belongs is in the disabled
329 * state.
330 */
331 new = 0;
332 }
333 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
334
335 return err;
336}
337
338static void hw_perf_event_update(struct perf_event *event)
339{
340 u64 prev, new, delta;
341 int err;
342
343 do {
344 prev = local64_read(&event->hw.prev_count);
345 err = ecctr(event->hw.config, &new);
346 if (err)
347 return;
348 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
349
350 delta = (prev <= new) ? new - prev
351 : (-1ULL - prev) + new + 1; /* overflow */
352 local64_add(delta, &event->count);
353}
354
355static void cpumf_pmu_read(struct perf_event *event)
356{
357 if (event->hw.state & PERF_HES_STOPPED)
358 return;
359
360 hw_perf_event_update(event);
361}
362
363static void cpumf_pmu_start(struct perf_event *event, int flags)
364{
365 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
366 struct hw_perf_event *hwc = &event->hw;
367
368 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
369 return;
370
371 if (WARN_ON_ONCE(hwc->config == -1))
372 return;
373
374 if (flags & PERF_EF_RELOAD)
375 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
376
377 hwc->state = 0;
378
379 /* (Re-)enable and activate the counter set */
380 ctr_set_enable(&cpuhw->state, hwc->config_base);
381 ctr_set_start(&cpuhw->state, hwc->config_base);
382
383 /* The counter set to which this counter belongs can be already active.
384 * Because all counters in a set are active, the event->hw.prev_count
385 * needs to be synchronized. At this point, the counter set can be in
386 * the inactive or disabled state.
387 */
388 hw_perf_event_reset(event);
389
390 /* increment refcount for this counter set */
391 atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
392}
393
394static void cpumf_pmu_stop(struct perf_event *event, int flags)
395{
396 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
397 struct hw_perf_event *hwc = &event->hw;
398
399 if (!(hwc->state & PERF_HES_STOPPED)) {
400 /* Decrement reference count for this counter set and if this
401 * is the last used counter in the set, clear activation
402 * control and set the counter set state to inactive.
403 */
404 if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
405 ctr_set_stop(&cpuhw->state, hwc->config_base);
406 event->hw.state |= PERF_HES_STOPPED;
407 }
408
409 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
410 hw_perf_event_update(event);
411 event->hw.state |= PERF_HES_UPTODATE;
412 }
413}
414
415static int cpumf_pmu_add(struct perf_event *event, int flags)
416{
417 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
418
419 /* Check authorization for the counter set to which this
420 * counter belongs.
421 * For group events transaction, the authorization check is
422 * done in cpumf_pmu_commit_txn().
423 */
424 if (!(cpuhw->txn_flags & PERF_PMU_TXN_ADD))
425 if (validate_ctr_auth(&event->hw))
426 return -ENOENT;
427
428 ctr_set_enable(&cpuhw->state, event->hw.config_base);
429 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
430
431 if (flags & PERF_EF_START)
432 cpumf_pmu_start(event, PERF_EF_RELOAD);
433
434 perf_event_update_userpage(event);
435
436 return 0;
437}
438
439static void cpumf_pmu_del(struct perf_event *event, int flags)
440{
441 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
442
443 cpumf_pmu_stop(event, PERF_EF_UPDATE);
444
445 /* Check if any counter in the counter set is still used. If not used,
446 * change the counter set to the disabled state. This also clears the
447 * content of all counters in the set.
448 *
449 * When a new perf event has been added but not yet started, this can
450 * clear enable control and resets all counters in a set. Therefore,
451 * cpumf_pmu_start() always has to reenable a counter set.
452 */
453 if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
454 ctr_set_disable(&cpuhw->state, event->hw.config_base);
455
456 perf_event_update_userpage(event);
457}
458
459/*
460 * Start group events scheduling transaction.
461 * Set flags to perform a single test at commit time.
462 *
463 * We only support PERF_PMU_TXN_ADD transactions. Save the
464 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
465 * transactions.
466 */
467static void cpumf_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
468{
469 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
470
471 WARN_ON_ONCE(cpuhw->txn_flags); /* txn already in flight */
472
473 cpuhw->txn_flags = txn_flags;
474 if (txn_flags & ~PERF_PMU_TXN_ADD)
475 return;
476
477 perf_pmu_disable(pmu);
478 cpuhw->tx_state = cpuhw->state;
479}
480
481/*
482 * Stop and cancel a group events scheduling tranctions.
483 * Assumes cpumf_pmu_del() is called for each successful added
484 * cpumf_pmu_add() during the transaction.
485 */
486static void cpumf_pmu_cancel_txn(struct pmu *pmu)
487{
488 unsigned int txn_flags;
489 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
490
491 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
492
493 txn_flags = cpuhw->txn_flags;
494 cpuhw->txn_flags = 0;
495 if (txn_flags & ~PERF_PMU_TXN_ADD)
496 return;
497
498 WARN_ON(cpuhw->tx_state != cpuhw->state);
499
500 perf_pmu_enable(pmu);
501}
502
503/*
504 * Commit the group events scheduling transaction. On success, the
505 * transaction is closed. On error, the transaction is kept open
506 * until cpumf_pmu_cancel_txn() is called.
507 */
508static int cpumf_pmu_commit_txn(struct pmu *pmu)
509{
510 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
511 u64 state;
512
513 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
514
515 if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
516 cpuhw->txn_flags = 0;
517 return 0;
518 }
519
520 /* check if the updated state can be scheduled */
521 state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
522 state >>= CPUMF_LCCTL_ENABLE_SHIFT;
523 if ((state & cpuhw->info.auth_ctl) != state)
524 return -ENOENT;
525
526 cpuhw->txn_flags = 0;
527 perf_pmu_enable(pmu);
528 return 0;
529}
530
531/* Performance monitoring unit for s390x */
532static struct pmu cpumf_pmu = {
533 .task_ctx_nr = perf_sw_context,
534 .capabilities = PERF_PMU_CAP_NO_INTERRUPT,
535 .pmu_enable = cpumf_pmu_enable,
536 .pmu_disable = cpumf_pmu_disable,
537 .event_init = cpumf_pmu_event_init,
538 .add = cpumf_pmu_add,
539 .del = cpumf_pmu_del,
540 .start = cpumf_pmu_start,
541 .stop = cpumf_pmu_stop,
542 .read = cpumf_pmu_read,
543 .start_txn = cpumf_pmu_start_txn,
544 .commit_txn = cpumf_pmu_commit_txn,
545 .cancel_txn = cpumf_pmu_cancel_txn,
546};
547
548static int __init cpumf_pmu_init(void)
549{
550 int rc;
551
552 if (!kernel_cpumcf_avail())
553 return -ENODEV;
554
555 cpumf_pmu.attr_groups = cpumf_cf_event_group();
556 rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
557 if (rc)
558 pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
559 return rc;
560}
561subsys_initcall(cpumf_pmu_init);
1/*
2 * Performance event support for s390x - CPU-measurement Counter Facility
3 *
4 * Copyright IBM Corp. 2012
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/* CPU-measurement counter facility supports these CPU counter sets:
26 * For CPU counter sets:
27 * Basic counter set: 0-31
28 * Problem-state counter set: 32-63
29 * Crypto-activity counter set: 64-127
30 * Extented counter set: 128-159
31 */
32enum cpumf_ctr_set {
33 /* CPU counter sets */
34 CPUMF_CTR_SET_BASIC = 0,
35 CPUMF_CTR_SET_USER = 1,
36 CPUMF_CTR_SET_CRYPTO = 2,
37 CPUMF_CTR_SET_EXT = 3,
38
39 /* Maximum number of counter sets */
40 CPUMF_CTR_SET_MAX,
41};
42
43#define CPUMF_LCCTL_ENABLE_SHIFT 16
44#define CPUMF_LCCTL_ACTCTL_SHIFT 0
45static const u64 cpumf_state_ctl[CPUMF_CTR_SET_MAX] = {
46 [CPUMF_CTR_SET_BASIC] = 0x02,
47 [CPUMF_CTR_SET_USER] = 0x04,
48 [CPUMF_CTR_SET_CRYPTO] = 0x08,
49 [CPUMF_CTR_SET_EXT] = 0x01,
50};
51
52static void ctr_set_enable(u64 *state, int ctr_set)
53{
54 *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT;
55}
56static void ctr_set_disable(u64 *state, int ctr_set)
57{
58 *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT);
59}
60static void ctr_set_start(u64 *state, int ctr_set)
61{
62 *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT;
63}
64static void ctr_set_stop(u64 *state, int ctr_set)
65{
66 *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT);
67}
68
69/* Local CPUMF event structure */
70struct cpu_hw_events {
71 struct cpumf_ctr_info info;
72 atomic_t ctr_set[CPUMF_CTR_SET_MAX];
73 u64 state, tx_state;
74 unsigned int flags;
75 unsigned int txn_flags;
76};
77static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
78 .ctr_set = {
79 [CPUMF_CTR_SET_BASIC] = ATOMIC_INIT(0),
80 [CPUMF_CTR_SET_USER] = ATOMIC_INIT(0),
81 [CPUMF_CTR_SET_CRYPTO] = ATOMIC_INIT(0),
82 [CPUMF_CTR_SET_EXT] = ATOMIC_INIT(0),
83 },
84 .state = 0,
85 .flags = 0,
86 .txn_flags = 0,
87};
88
89static int get_counter_set(u64 event)
90{
91 int set = -1;
92
93 if (event < 32)
94 set = CPUMF_CTR_SET_BASIC;
95 else if (event < 64)
96 set = CPUMF_CTR_SET_USER;
97 else if (event < 128)
98 set = CPUMF_CTR_SET_CRYPTO;
99 else if (event < 256)
100 set = CPUMF_CTR_SET_EXT;
101
102 return set;
103}
104
105static int validate_event(const struct hw_perf_event *hwc)
106{
107 switch (hwc->config_base) {
108 case CPUMF_CTR_SET_BASIC:
109 case CPUMF_CTR_SET_USER:
110 case CPUMF_CTR_SET_CRYPTO:
111 case CPUMF_CTR_SET_EXT:
112 /* check for reserved counters */
113 if ((hwc->config >= 6 && hwc->config <= 31) ||
114 (hwc->config >= 38 && hwc->config <= 63) ||
115 (hwc->config >= 80 && hwc->config <= 127))
116 return -EOPNOTSUPP;
117 break;
118 default:
119 return -EINVAL;
120 }
121
122 return 0;
123}
124
125static int validate_ctr_version(const struct hw_perf_event *hwc)
126{
127 struct cpu_hw_events *cpuhw;
128 int err = 0;
129
130 cpuhw = &get_cpu_var(cpu_hw_events);
131
132 /* check required version for counter sets */
133 switch (hwc->config_base) {
134 case CPUMF_CTR_SET_BASIC:
135 case CPUMF_CTR_SET_USER:
136 if (cpuhw->info.cfvn < 1)
137 err = -EOPNOTSUPP;
138 break;
139 case CPUMF_CTR_SET_CRYPTO:
140 case CPUMF_CTR_SET_EXT:
141 if (cpuhw->info.csvn < 1)
142 err = -EOPNOTSUPP;
143 if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
144 (cpuhw->info.csvn == 2 && hwc->config > 175) ||
145 (cpuhw->info.csvn > 2 && hwc->config > 255))
146 err = -EOPNOTSUPP;
147 break;
148 }
149
150 put_cpu_var(cpu_hw_events);
151 return err;
152}
153
154static int validate_ctr_auth(const struct hw_perf_event *hwc)
155{
156 struct cpu_hw_events *cpuhw;
157 u64 ctrs_state;
158 int err = 0;
159
160 cpuhw = &get_cpu_var(cpu_hw_events);
161
162 /* Check authorization for cpu counter sets.
163 * If the particular CPU counter set is not authorized,
164 * return with -ENOENT in order to fall back to other
165 * PMUs that might suffice the event request.
166 */
167 ctrs_state = cpumf_state_ctl[hwc->config_base];
168 if (!(ctrs_state & cpuhw->info.auth_ctl))
169 err = -ENOENT;
170
171 put_cpu_var(cpu_hw_events);
172 return err;
173}
174
175/*
176 * Change the CPUMF state to active.
177 * Enable and activate the CPU-counter sets according
178 * to the per-cpu control state.
179 */
180static void cpumf_pmu_enable(struct pmu *pmu)
181{
182 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
183 int err;
184
185 if (cpuhw->flags & PMU_F_ENABLED)
186 return;
187
188 err = lcctl(cpuhw->state);
189 if (err) {
190 pr_err("Enabling the performance measuring unit "
191 "failed with rc=%x\n", err);
192 return;
193 }
194
195 cpuhw->flags |= PMU_F_ENABLED;
196}
197
198/*
199 * Change the CPUMF state to inactive.
200 * Disable and enable (inactive) the CPU-counter sets according
201 * to the per-cpu control state.
202 */
203static void cpumf_pmu_disable(struct pmu *pmu)
204{
205 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
206 int err;
207 u64 inactive;
208
209 if (!(cpuhw->flags & PMU_F_ENABLED))
210 return;
211
212 inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
213 err = lcctl(inactive);
214 if (err) {
215 pr_err("Disabling the performance measuring unit "
216 "failed with rc=%x\n", err);
217 return;
218 }
219
220 cpuhw->flags &= ~PMU_F_ENABLED;
221}
222
223
224/* Number of perf events counting hardware events */
225static atomic_t num_events = ATOMIC_INIT(0);
226/* Used to avoid races in calling reserve/release_cpumf_hardware */
227static DEFINE_MUTEX(pmc_reserve_mutex);
228
229/* CPU-measurement alerts for the counter facility */
230static void cpumf_measurement_alert(struct ext_code ext_code,
231 unsigned int alert, unsigned long unused)
232{
233 struct cpu_hw_events *cpuhw;
234
235 if (!(alert & CPU_MF_INT_CF_MASK))
236 return;
237
238 inc_irq_stat(IRQEXT_CMC);
239 cpuhw = this_cpu_ptr(&cpu_hw_events);
240
241 /* Measurement alerts are shared and might happen when the PMU
242 * is not reserved. Ignore these alerts in this case. */
243 if (!(cpuhw->flags & PMU_F_RESERVED))
244 return;
245
246 /* counter authorization change alert */
247 if (alert & CPU_MF_INT_CF_CACA)
248 qctri(&cpuhw->info);
249
250 /* loss of counter data alert */
251 if (alert & CPU_MF_INT_CF_LCDA)
252 pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
253}
254
255#define PMC_INIT 0
256#define PMC_RELEASE 1
257static void setup_pmc_cpu(void *flags)
258{
259 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
260
261 switch (*((int *) flags)) {
262 case PMC_INIT:
263 memset(&cpuhw->info, 0, sizeof(cpuhw->info));
264 qctri(&cpuhw->info);
265 cpuhw->flags |= PMU_F_RESERVED;
266 break;
267
268 case PMC_RELEASE:
269 cpuhw->flags &= ~PMU_F_RESERVED;
270 break;
271 }
272
273 /* Disable CPU counter sets */
274 lcctl(0);
275}
276
277/* Initialize the CPU-measurement facility */
278static int reserve_pmc_hardware(void)
279{
280 int flags = PMC_INIT;
281
282 on_each_cpu(setup_pmc_cpu, &flags, 1);
283 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
284
285 return 0;
286}
287
288/* Release the CPU-measurement facility */
289static void release_pmc_hardware(void)
290{
291 int flags = PMC_RELEASE;
292
293 on_each_cpu(setup_pmc_cpu, &flags, 1);
294 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
295}
296
297/* Release the PMU if event is the last perf event */
298static void hw_perf_event_destroy(struct perf_event *event)
299{
300 if (!atomic_add_unless(&num_events, -1, 1)) {
301 mutex_lock(&pmc_reserve_mutex);
302 if (atomic_dec_return(&num_events) == 0)
303 release_pmc_hardware();
304 mutex_unlock(&pmc_reserve_mutex);
305 }
306}
307
308/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
309static const int cpumf_generic_events_basic[] = {
310 [PERF_COUNT_HW_CPU_CYCLES] = 0,
311 [PERF_COUNT_HW_INSTRUCTIONS] = 1,
312 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
313 [PERF_COUNT_HW_CACHE_MISSES] = -1,
314 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
315 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
316 [PERF_COUNT_HW_BUS_CYCLES] = -1,
317};
318/* CPUMF <-> perf event mappings for userspace (problem-state set) */
319static const int cpumf_generic_events_user[] = {
320 [PERF_COUNT_HW_CPU_CYCLES] = 32,
321 [PERF_COUNT_HW_INSTRUCTIONS] = 33,
322 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
323 [PERF_COUNT_HW_CACHE_MISSES] = -1,
324 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
325 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
326 [PERF_COUNT_HW_BUS_CYCLES] = -1,
327};
328
329static int __hw_perf_event_init(struct perf_event *event)
330{
331 struct perf_event_attr *attr = &event->attr;
332 struct hw_perf_event *hwc = &event->hw;
333 int err;
334 u64 ev;
335
336 switch (attr->type) {
337 case PERF_TYPE_RAW:
338 /* Raw events are used to access counters directly,
339 * hence do not permit excludes */
340 if (attr->exclude_kernel || attr->exclude_user ||
341 attr->exclude_hv)
342 return -EOPNOTSUPP;
343 ev = attr->config;
344 break;
345
346 case PERF_TYPE_HARDWARE:
347 ev = attr->config;
348 /* Count user space (problem-state) only */
349 if (!attr->exclude_user && attr->exclude_kernel) {
350 if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
351 return -EOPNOTSUPP;
352 ev = cpumf_generic_events_user[ev];
353
354 /* No support for kernel space counters only */
355 } else if (!attr->exclude_kernel && attr->exclude_user) {
356 return -EOPNOTSUPP;
357
358 /* Count user and kernel space */
359 } else {
360 if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
361 return -EOPNOTSUPP;
362 ev = cpumf_generic_events_basic[ev];
363 }
364 break;
365
366 default:
367 return -ENOENT;
368 }
369
370 if (ev == -1)
371 return -ENOENT;
372
373 if (ev >= PERF_CPUM_CF_MAX_CTR)
374 return -EINVAL;
375
376 /* Use the hardware perf event structure to store the counter number
377 * in 'config' member and the counter set to which the counter belongs
378 * in the 'config_base'. The counter set (config_base) is then used
379 * to enable/disable the counters.
380 */
381 hwc->config = ev;
382 hwc->config_base = get_counter_set(ev);
383
384 /* Validate the counter that is assigned to this event.
385 * Because the counter facility can use numerous counters at the
386 * same time without constraints, it is not necessary to explicitly
387 * validate event groups (event->group_leader != event).
388 */
389 err = validate_event(hwc);
390 if (err)
391 return err;
392
393 /* Initialize for using the CPU-measurement counter facility */
394 if (!atomic_inc_not_zero(&num_events)) {
395 mutex_lock(&pmc_reserve_mutex);
396 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
397 err = -EBUSY;
398 else
399 atomic_inc(&num_events);
400 mutex_unlock(&pmc_reserve_mutex);
401 }
402 event->destroy = hw_perf_event_destroy;
403
404 /* Finally, validate version and authorization of the counter set */
405 err = validate_ctr_auth(hwc);
406 if (!err)
407 err = validate_ctr_version(hwc);
408
409 return err;
410}
411
412static int cpumf_pmu_event_init(struct perf_event *event)
413{
414 int err;
415
416 switch (event->attr.type) {
417 case PERF_TYPE_HARDWARE:
418 case PERF_TYPE_HW_CACHE:
419 case PERF_TYPE_RAW:
420 err = __hw_perf_event_init(event);
421 break;
422 default:
423 return -ENOENT;
424 }
425
426 if (unlikely(err) && event->destroy)
427 event->destroy(event);
428
429 return err;
430}
431
432static int hw_perf_event_reset(struct perf_event *event)
433{
434 u64 prev, new;
435 int err;
436
437 do {
438 prev = local64_read(&event->hw.prev_count);
439 err = ecctr(event->hw.config, &new);
440 if (err) {
441 if (err != 3)
442 break;
443 /* The counter is not (yet) available. This
444 * might happen if the counter set to which
445 * this counter belongs is in the disabled
446 * state.
447 */
448 new = 0;
449 }
450 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
451
452 return err;
453}
454
455static int hw_perf_event_update(struct perf_event *event)
456{
457 u64 prev, new, delta;
458 int err;
459
460 do {
461 prev = local64_read(&event->hw.prev_count);
462 err = ecctr(event->hw.config, &new);
463 if (err)
464 goto out;
465 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
466
467 delta = (prev <= new) ? new - prev
468 : (-1ULL - prev) + new + 1; /* overflow */
469 local64_add(delta, &event->count);
470out:
471 return err;
472}
473
474static void cpumf_pmu_read(struct perf_event *event)
475{
476 if (event->hw.state & PERF_HES_STOPPED)
477 return;
478
479 hw_perf_event_update(event);
480}
481
482static void cpumf_pmu_start(struct perf_event *event, int flags)
483{
484 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
485 struct hw_perf_event *hwc = &event->hw;
486
487 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
488 return;
489
490 if (WARN_ON_ONCE(hwc->config == -1))
491 return;
492
493 if (flags & PERF_EF_RELOAD)
494 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
495
496 hwc->state = 0;
497
498 /* (Re-)enable and activate the counter set */
499 ctr_set_enable(&cpuhw->state, hwc->config_base);
500 ctr_set_start(&cpuhw->state, hwc->config_base);
501
502 /* The counter set to which this counter belongs can be already active.
503 * Because all counters in a set are active, the event->hw.prev_count
504 * needs to be synchronized. At this point, the counter set can be in
505 * the inactive or disabled state.
506 */
507 hw_perf_event_reset(event);
508
509 /* increment refcount for this counter set */
510 atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
511}
512
513static void cpumf_pmu_stop(struct perf_event *event, int flags)
514{
515 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
516 struct hw_perf_event *hwc = &event->hw;
517
518 if (!(hwc->state & PERF_HES_STOPPED)) {
519 /* Decrement reference count for this counter set and if this
520 * is the last used counter in the set, clear activation
521 * control and set the counter set state to inactive.
522 */
523 if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
524 ctr_set_stop(&cpuhw->state, hwc->config_base);
525 event->hw.state |= PERF_HES_STOPPED;
526 }
527
528 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
529 hw_perf_event_update(event);
530 event->hw.state |= PERF_HES_UPTODATE;
531 }
532}
533
534static int cpumf_pmu_add(struct perf_event *event, int flags)
535{
536 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
537
538 /* Check authorization for the counter set to which this
539 * counter belongs.
540 * For group events transaction, the authorization check is
541 * done in cpumf_pmu_commit_txn().
542 */
543 if (!(cpuhw->txn_flags & PERF_PMU_TXN_ADD))
544 if (validate_ctr_auth(&event->hw))
545 return -ENOENT;
546
547 ctr_set_enable(&cpuhw->state, event->hw.config_base);
548 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
549
550 if (flags & PERF_EF_START)
551 cpumf_pmu_start(event, PERF_EF_RELOAD);
552
553 perf_event_update_userpage(event);
554
555 return 0;
556}
557
558static void cpumf_pmu_del(struct perf_event *event, int flags)
559{
560 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
561
562 cpumf_pmu_stop(event, PERF_EF_UPDATE);
563
564 /* Check if any counter in the counter set is still used. If not used,
565 * change the counter set to the disabled state. This also clears the
566 * content of all counters in the set.
567 *
568 * When a new perf event has been added but not yet started, this can
569 * clear enable control and resets all counters in a set. Therefore,
570 * cpumf_pmu_start() always has to reenable a counter set.
571 */
572 if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
573 ctr_set_disable(&cpuhw->state, event->hw.config_base);
574
575 perf_event_update_userpage(event);
576}
577
578/*
579 * Start group events scheduling transaction.
580 * Set flags to perform a single test at commit time.
581 *
582 * We only support PERF_PMU_TXN_ADD transactions. Save the
583 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
584 * transactions.
585 */
586static void cpumf_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
587{
588 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
589
590 WARN_ON_ONCE(cpuhw->txn_flags); /* txn already in flight */
591
592 cpuhw->txn_flags = txn_flags;
593 if (txn_flags & ~PERF_PMU_TXN_ADD)
594 return;
595
596 perf_pmu_disable(pmu);
597 cpuhw->tx_state = cpuhw->state;
598}
599
600/*
601 * Stop and cancel a group events scheduling tranctions.
602 * Assumes cpumf_pmu_del() is called for each successful added
603 * cpumf_pmu_add() during the transaction.
604 */
605static void cpumf_pmu_cancel_txn(struct pmu *pmu)
606{
607 unsigned int txn_flags;
608 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
609
610 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
611
612 txn_flags = cpuhw->txn_flags;
613 cpuhw->txn_flags = 0;
614 if (txn_flags & ~PERF_PMU_TXN_ADD)
615 return;
616
617 WARN_ON(cpuhw->tx_state != cpuhw->state);
618
619 perf_pmu_enable(pmu);
620}
621
622/*
623 * Commit the group events scheduling transaction. On success, the
624 * transaction is closed. On error, the transaction is kept open
625 * until cpumf_pmu_cancel_txn() is called.
626 */
627static int cpumf_pmu_commit_txn(struct pmu *pmu)
628{
629 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
630 u64 state;
631
632 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
633
634 if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
635 cpuhw->txn_flags = 0;
636 return 0;
637 }
638
639 /* check if the updated state can be scheduled */
640 state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
641 state >>= CPUMF_LCCTL_ENABLE_SHIFT;
642 if ((state & cpuhw->info.auth_ctl) != state)
643 return -ENOENT;
644
645 cpuhw->txn_flags = 0;
646 perf_pmu_enable(pmu);
647 return 0;
648}
649
650/* Performance monitoring unit for s390x */
651static struct pmu cpumf_pmu = {
652 .task_ctx_nr = perf_sw_context,
653 .capabilities = PERF_PMU_CAP_NO_INTERRUPT,
654 .pmu_enable = cpumf_pmu_enable,
655 .pmu_disable = cpumf_pmu_disable,
656 .event_init = cpumf_pmu_event_init,
657 .add = cpumf_pmu_add,
658 .del = cpumf_pmu_del,
659 .start = cpumf_pmu_start,
660 .stop = cpumf_pmu_stop,
661 .read = cpumf_pmu_read,
662 .start_txn = cpumf_pmu_start_txn,
663 .commit_txn = cpumf_pmu_commit_txn,
664 .cancel_txn = cpumf_pmu_cancel_txn,
665};
666
667static int cpumf_pmf_setup(unsigned int cpu, int flags)
668{
669 local_irq_disable();
670 setup_pmc_cpu(&flags);
671 local_irq_enable();
672 return 0;
673}
674
675static int s390_pmu_online_cpu(unsigned int cpu)
676{
677 return cpumf_pmf_setup(cpu, PMC_INIT);
678}
679
680static int s390_pmu_offline_cpu(unsigned int cpu)
681{
682 return cpumf_pmf_setup(cpu, PMC_RELEASE);
683}
684
685static int __init cpumf_pmu_init(void)
686{
687 int rc;
688
689 if (!cpum_cf_avail())
690 return -ENODEV;
691
692 /* clear bit 15 of cr0 to unauthorize problem-state to
693 * extract measurement counters */
694 ctl_clear_bit(0, 48);
695
696 /* register handler for measurement-alert interruptions */
697 rc = register_external_irq(EXT_IRQ_MEASURE_ALERT,
698 cpumf_measurement_alert);
699 if (rc) {
700 pr_err("Registering for CPU-measurement alerts "
701 "failed with rc=%i\n", rc);
702 return rc;
703 }
704
705 cpumf_pmu.attr_groups = cpumf_cf_event_group();
706 rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
707 if (rc) {
708 pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
709 unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
710 cpumf_measurement_alert);
711 return rc;
712 }
713 return cpuhp_setup_state(CPUHP_AP_PERF_S390_CF_ONLINE,
714 "perf/s390/cf:online",
715 s390_pmu_online_cpu, s390_pmu_offline_cpu);
716}
717early_initcall(cpumf_pmu_init);