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, 2021
6 * Author(s): Hendrik Brueckner <brueckner@linux.ibm.com>
7 * Thomas Richter <tmricht@linux.ibm.com>
8 */
9#define KMSG_COMPONENT "cpum_cf"
10#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
11
12#include <linux/kernel.h>
13#include <linux/kernel_stat.h>
14#include <linux/percpu.h>
15#include <linux/notifier.h>
16#include <linux/init.h>
17#include <linux/export.h>
18#include <linux/miscdevice.h>
19
20#include <asm/cpu_mcf.h>
21#include <asm/hwctrset.h>
22#include <asm/debug.h>
23
24static unsigned int cfdiag_cpu_speed; /* CPU speed for CF_DIAG trailer */
25static debug_info_t *cf_dbg;
26
27#define CF_DIAG_CTRSET_DEF 0xfeef /* Counter set header mark */
28 /* interval in seconds */
29
30/* Counter sets are stored as data stream in a page sized memory buffer and
31 * exported to user space via raw data attached to the event sample data.
32 * Each counter set starts with an eight byte header consisting of:
33 * - a two byte eye catcher (0xfeef)
34 * - a one byte counter set number
35 * - a two byte counter set size (indicates the number of counters in this set)
36 * - a three byte reserved value (must be zero) to make the header the same
37 * size as a counter value.
38 * All counter values are eight byte in size.
39 *
40 * All counter sets are followed by a 64 byte trailer.
41 * The trailer consists of a:
42 * - flag field indicating valid fields when corresponding bit set
43 * - the counter facility first and second version number
44 * - the CPU speed if nonzero
45 * - the time stamp the counter sets have been collected
46 * - the time of day (TOD) base value
47 * - the machine type.
48 *
49 * The counter sets are saved when the process is prepared to be executed on a
50 * CPU and saved again when the process is going to be removed from a CPU.
51 * The difference of both counter sets are calculated and stored in the event
52 * sample data area.
53 */
54struct cf_ctrset_entry { /* CPU-M CF counter set entry (8 byte) */
55 unsigned int def:16; /* 0-15 Data Entry Format */
56 unsigned int set:16; /* 16-31 Counter set identifier */
57 unsigned int ctr:16; /* 32-47 Number of stored counters */
58 unsigned int res1:16; /* 48-63 Reserved */
59};
60
61struct cf_trailer_entry { /* CPU-M CF_DIAG trailer (64 byte) */
62 /* 0 - 7 */
63 union {
64 struct {
65 unsigned int clock_base:1; /* TOD clock base set */
66 unsigned int speed:1; /* CPU speed set */
67 /* Measurement alerts */
68 unsigned int mtda:1; /* Loss of MT ctr. data alert */
69 unsigned int caca:1; /* Counter auth. change alert */
70 unsigned int lcda:1; /* Loss of counter data alert */
71 };
72 unsigned long flags; /* 0-63 All indicators */
73 };
74 /* 8 - 15 */
75 unsigned int cfvn:16; /* 64-79 Ctr First Version */
76 unsigned int csvn:16; /* 80-95 Ctr Second Version */
77 unsigned int cpu_speed:32; /* 96-127 CPU speed */
78 /* 16 - 23 */
79 unsigned long timestamp; /* 128-191 Timestamp (TOD) */
80 /* 24 - 55 */
81 union {
82 struct {
83 unsigned long progusage1;
84 unsigned long progusage2;
85 unsigned long progusage3;
86 unsigned long tod_base;
87 };
88 unsigned long progusage[4];
89 };
90 /* 56 - 63 */
91 unsigned int mach_type:16; /* Machine type */
92 unsigned int res1:16; /* Reserved */
93 unsigned int res2:32; /* Reserved */
94};
95
96/* Create the trailer data at the end of a page. */
97static void cfdiag_trailer(struct cf_trailer_entry *te)
98{
99 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
100 struct cpuid cpuid;
101
102 te->cfvn = cpuhw->info.cfvn; /* Counter version numbers */
103 te->csvn = cpuhw->info.csvn;
104
105 get_cpu_id(&cpuid); /* Machine type */
106 te->mach_type = cpuid.machine;
107 te->cpu_speed = cfdiag_cpu_speed;
108 if (te->cpu_speed)
109 te->speed = 1;
110 te->clock_base = 1; /* Save clock base */
111 te->tod_base = tod_clock_base.tod;
112 te->timestamp = get_tod_clock_fast();
113}
114
115/* Read a counter set. The counter set number determines the counter set and
116 * the CPUM-CF first and second version number determine the number of
117 * available counters in each counter set.
118 * Each counter set starts with header containing the counter set number and
119 * the number of eight byte counters.
120 *
121 * The functions returns the number of bytes occupied by this counter set
122 * including the header.
123 * If there is no counter in the counter set, this counter set is useless and
124 * zero is returned on this case.
125 *
126 * Note that the counter sets may not be enabled or active and the stcctm
127 * instruction might return error 3. Depending on error_ok value this is ok,
128 * for example when called from cpumf_pmu_start() call back function.
129 */
130static size_t cfdiag_getctrset(struct cf_ctrset_entry *ctrdata, int ctrset,
131 size_t room, bool error_ok)
132{
133 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
134 size_t ctrset_size, need = 0;
135 int rc = 3; /* Assume write failure */
136
137 ctrdata->def = CF_DIAG_CTRSET_DEF;
138 ctrdata->set = ctrset;
139 ctrdata->res1 = 0;
140 ctrset_size = cpum_cf_ctrset_size(ctrset, &cpuhw->info);
141
142 if (ctrset_size) { /* Save data */
143 need = ctrset_size * sizeof(u64) + sizeof(*ctrdata);
144 if (need <= room) {
145 rc = ctr_stcctm(ctrset, ctrset_size,
146 (u64 *)(ctrdata + 1));
147 }
148 if (rc != 3 || error_ok)
149 ctrdata->ctr = ctrset_size;
150 else
151 need = 0;
152 }
153
154 debug_sprintf_event(cf_dbg, 3,
155 "%s ctrset %d ctrset_size %zu cfvn %d csvn %d"
156 " need %zd rc %d\n", __func__, ctrset, ctrset_size,
157 cpuhw->info.cfvn, cpuhw->info.csvn, need, rc);
158 return need;
159}
160
161static const u64 cpumf_ctr_ctl[CPUMF_CTR_SET_MAX] = {
162 [CPUMF_CTR_SET_BASIC] = 0x02,
163 [CPUMF_CTR_SET_USER] = 0x04,
164 [CPUMF_CTR_SET_CRYPTO] = 0x08,
165 [CPUMF_CTR_SET_EXT] = 0x01,
166 [CPUMF_CTR_SET_MT_DIAG] = 0x20,
167};
168
169/* Read out all counter sets and save them in the provided data buffer.
170 * The last 64 byte host an artificial trailer entry.
171 */
172static size_t cfdiag_getctr(void *data, size_t sz, unsigned long auth,
173 bool error_ok)
174{
175 struct cf_trailer_entry *trailer;
176 size_t offset = 0, done;
177 int i;
178
179 memset(data, 0, sz);
180 sz -= sizeof(*trailer); /* Always room for trailer */
181 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
182 struct cf_ctrset_entry *ctrdata = data + offset;
183
184 if (!(auth & cpumf_ctr_ctl[i]))
185 continue; /* Counter set not authorized */
186
187 done = cfdiag_getctrset(ctrdata, i, sz - offset, error_ok);
188 offset += done;
189 }
190 trailer = data + offset;
191 cfdiag_trailer(trailer);
192 return offset + sizeof(*trailer);
193}
194
195/* Calculate the difference for each counter in a counter set. */
196static void cfdiag_diffctrset(u64 *pstart, u64 *pstop, int counters)
197{
198 for (; --counters >= 0; ++pstart, ++pstop)
199 if (*pstop >= *pstart)
200 *pstop -= *pstart;
201 else
202 *pstop = *pstart - *pstop + 1;
203}
204
205/* Scan the counter sets and calculate the difference of each counter
206 * in each set. The result is the increment of each counter during the
207 * period the counter set has been activated.
208 *
209 * Return true on success.
210 */
211static int cfdiag_diffctr(struct cpu_cf_events *cpuhw, unsigned long auth)
212{
213 struct cf_trailer_entry *trailer_start, *trailer_stop;
214 struct cf_ctrset_entry *ctrstart, *ctrstop;
215 size_t offset = 0;
216
217 auth &= (1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1;
218 do {
219 ctrstart = (struct cf_ctrset_entry *)(cpuhw->start + offset);
220 ctrstop = (struct cf_ctrset_entry *)(cpuhw->stop + offset);
221
222 if (memcmp(ctrstop, ctrstart, sizeof(*ctrstop))) {
223 pr_err_once("cpum_cf_diag counter set compare error "
224 "in set %i\n", ctrstart->set);
225 return 0;
226 }
227 auth &= ~cpumf_ctr_ctl[ctrstart->set];
228 if (ctrstart->def == CF_DIAG_CTRSET_DEF) {
229 cfdiag_diffctrset((u64 *)(ctrstart + 1),
230 (u64 *)(ctrstop + 1), ctrstart->ctr);
231 offset += ctrstart->ctr * sizeof(u64) +
232 sizeof(*ctrstart);
233 }
234 } while (ctrstart->def && auth);
235
236 /* Save time_stamp from start of event in stop's trailer */
237 trailer_start = (struct cf_trailer_entry *)(cpuhw->start + offset);
238 trailer_stop = (struct cf_trailer_entry *)(cpuhw->stop + offset);
239 trailer_stop->progusage[0] = trailer_start->timestamp;
240
241 return 1;
242}
243
244static enum cpumf_ctr_set get_counter_set(u64 event)
245{
246 int set = CPUMF_CTR_SET_MAX;
247
248 if (event < 32)
249 set = CPUMF_CTR_SET_BASIC;
250 else if (event < 64)
251 set = CPUMF_CTR_SET_USER;
252 else if (event < 128)
253 set = CPUMF_CTR_SET_CRYPTO;
254 else if (event < 288)
255 set = CPUMF_CTR_SET_EXT;
256 else if (event >= 448 && event < 496)
257 set = CPUMF_CTR_SET_MT_DIAG;
258
259 return set;
260}
261
262static int validate_ctr_version(const struct hw_perf_event *hwc,
263 enum cpumf_ctr_set set)
264{
265 struct cpu_cf_events *cpuhw;
266 int err = 0;
267 u16 mtdiag_ctl;
268
269 cpuhw = &get_cpu_var(cpu_cf_events);
270
271 /* check required version for counter sets */
272 switch (set) {
273 case CPUMF_CTR_SET_BASIC:
274 case CPUMF_CTR_SET_USER:
275 if (cpuhw->info.cfvn < 1)
276 err = -EOPNOTSUPP;
277 break;
278 case CPUMF_CTR_SET_CRYPTO:
279 if ((cpuhw->info.csvn >= 1 && cpuhw->info.csvn <= 5 &&
280 hwc->config > 79) ||
281 (cpuhw->info.csvn >= 6 && hwc->config > 83))
282 err = -EOPNOTSUPP;
283 break;
284 case CPUMF_CTR_SET_EXT:
285 if (cpuhw->info.csvn < 1)
286 err = -EOPNOTSUPP;
287 if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
288 (cpuhw->info.csvn == 2 && hwc->config > 175) ||
289 (cpuhw->info.csvn >= 3 && cpuhw->info.csvn <= 5
290 && hwc->config > 255) ||
291 (cpuhw->info.csvn >= 6 && hwc->config > 287))
292 err = -EOPNOTSUPP;
293 break;
294 case CPUMF_CTR_SET_MT_DIAG:
295 if (cpuhw->info.csvn <= 3)
296 err = -EOPNOTSUPP;
297 /*
298 * MT-diagnostic counters are read-only. The counter set
299 * is automatically enabled and activated on all CPUs with
300 * multithreading (SMT). Deactivation of multithreading
301 * also disables the counter set. State changes are ignored
302 * by lcctl(). Because Linux controls SMT enablement through
303 * a kernel parameter only, the counter set is either disabled
304 * or enabled and active.
305 *
306 * Thus, the counters can only be used if SMT is on and the
307 * counter set is enabled and active.
308 */
309 mtdiag_ctl = cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG];
310 if (!((cpuhw->info.auth_ctl & mtdiag_ctl) &&
311 (cpuhw->info.enable_ctl & mtdiag_ctl) &&
312 (cpuhw->info.act_ctl & mtdiag_ctl)))
313 err = -EOPNOTSUPP;
314 break;
315 case CPUMF_CTR_SET_MAX:
316 err = -EOPNOTSUPP;
317 }
318
319 put_cpu_var(cpu_cf_events);
320 return err;
321}
322
323static int validate_ctr_auth(const struct hw_perf_event *hwc)
324{
325 struct cpu_cf_events *cpuhw;
326 int err = 0;
327
328 cpuhw = &get_cpu_var(cpu_cf_events);
329
330 /* Check authorization for cpu counter sets.
331 * If the particular CPU counter set is not authorized,
332 * return with -ENOENT in order to fall back to other
333 * PMUs that might suffice the event request.
334 */
335 if (!(hwc->config_base & cpuhw->info.auth_ctl))
336 err = -ENOENT;
337
338 put_cpu_var(cpu_cf_events);
339 return err;
340}
341
342/*
343 * Change the CPUMF state to active.
344 * Enable and activate the CPU-counter sets according
345 * to the per-cpu control state.
346 */
347static void cpumf_pmu_enable(struct pmu *pmu)
348{
349 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
350 int err;
351
352 if (cpuhw->flags & PMU_F_ENABLED)
353 return;
354
355 err = lcctl(cpuhw->state | cpuhw->dev_state);
356 if (err) {
357 pr_err("Enabling the performance measuring unit "
358 "failed with rc=%x\n", err);
359 return;
360 }
361
362 cpuhw->flags |= PMU_F_ENABLED;
363}
364
365/*
366 * Change the CPUMF state to inactive.
367 * Disable and enable (inactive) the CPU-counter sets according
368 * to the per-cpu control state.
369 */
370static void cpumf_pmu_disable(struct pmu *pmu)
371{
372 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
373 int err;
374 u64 inactive;
375
376 if (!(cpuhw->flags & PMU_F_ENABLED))
377 return;
378
379 inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
380 inactive |= cpuhw->dev_state;
381 err = lcctl(inactive);
382 if (err) {
383 pr_err("Disabling the performance measuring unit "
384 "failed with rc=%x\n", err);
385 return;
386 }
387
388 cpuhw->flags &= ~PMU_F_ENABLED;
389}
390
391
392/* Number of perf events counting hardware events */
393static atomic_t num_events = ATOMIC_INIT(0);
394/* Used to avoid races in calling reserve/release_cpumf_hardware */
395static DEFINE_MUTEX(pmc_reserve_mutex);
396
397/* Release the PMU if event is the last perf event */
398static void hw_perf_event_destroy(struct perf_event *event)
399{
400 if (!atomic_add_unless(&num_events, -1, 1)) {
401 mutex_lock(&pmc_reserve_mutex);
402 if (atomic_dec_return(&num_events) == 0)
403 __kernel_cpumcf_end();
404 mutex_unlock(&pmc_reserve_mutex);
405 }
406}
407
408/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
409static const int cpumf_generic_events_basic[] = {
410 [PERF_COUNT_HW_CPU_CYCLES] = 0,
411 [PERF_COUNT_HW_INSTRUCTIONS] = 1,
412 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
413 [PERF_COUNT_HW_CACHE_MISSES] = -1,
414 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
415 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
416 [PERF_COUNT_HW_BUS_CYCLES] = -1,
417};
418/* CPUMF <-> perf event mappings for userspace (problem-state set) */
419static const int cpumf_generic_events_user[] = {
420 [PERF_COUNT_HW_CPU_CYCLES] = 32,
421 [PERF_COUNT_HW_INSTRUCTIONS] = 33,
422 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
423 [PERF_COUNT_HW_CACHE_MISSES] = -1,
424 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
425 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
426 [PERF_COUNT_HW_BUS_CYCLES] = -1,
427};
428
429static void cpumf_hw_inuse(void)
430{
431 mutex_lock(&pmc_reserve_mutex);
432 if (atomic_inc_return(&num_events) == 1)
433 __kernel_cpumcf_begin();
434 mutex_unlock(&pmc_reserve_mutex);
435}
436
437static int __hw_perf_event_init(struct perf_event *event, unsigned int type)
438{
439 struct perf_event_attr *attr = &event->attr;
440 struct hw_perf_event *hwc = &event->hw;
441 enum cpumf_ctr_set set;
442 int err = 0;
443 u64 ev;
444
445 switch (type) {
446 case PERF_TYPE_RAW:
447 /* Raw events are used to access counters directly,
448 * hence do not permit excludes */
449 if (attr->exclude_kernel || attr->exclude_user ||
450 attr->exclude_hv)
451 return -EOPNOTSUPP;
452 ev = attr->config;
453 break;
454
455 case PERF_TYPE_HARDWARE:
456 if (is_sampling_event(event)) /* No sampling support */
457 return -ENOENT;
458 ev = attr->config;
459 /* Count user space (problem-state) only */
460 if (!attr->exclude_user && attr->exclude_kernel) {
461 if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
462 return -EOPNOTSUPP;
463 ev = cpumf_generic_events_user[ev];
464
465 /* No support for kernel space counters only */
466 } else if (!attr->exclude_kernel && attr->exclude_user) {
467 return -EOPNOTSUPP;
468 } else { /* Count user and kernel space */
469 if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
470 return -EOPNOTSUPP;
471 ev = cpumf_generic_events_basic[ev];
472 }
473 break;
474
475 default:
476 return -ENOENT;
477 }
478
479 if (ev == -1)
480 return -ENOENT;
481
482 if (ev > PERF_CPUM_CF_MAX_CTR)
483 return -ENOENT;
484
485 /* Obtain the counter set to which the specified counter belongs */
486 set = get_counter_set(ev);
487 switch (set) {
488 case CPUMF_CTR_SET_BASIC:
489 case CPUMF_CTR_SET_USER:
490 case CPUMF_CTR_SET_CRYPTO:
491 case CPUMF_CTR_SET_EXT:
492 case CPUMF_CTR_SET_MT_DIAG:
493 /*
494 * Use the hardware perf event structure to store the
495 * counter number in the 'config' member and the counter
496 * set number in the 'config_base' as bit mask.
497 * It is later used to enable/disable the counter(s).
498 */
499 hwc->config = ev;
500 hwc->config_base = cpumf_ctr_ctl[set];
501 break;
502 case CPUMF_CTR_SET_MAX:
503 /* The counter could not be associated to a counter set */
504 return -EINVAL;
505 }
506
507 /* Initialize for using the CPU-measurement counter facility */
508 cpumf_hw_inuse();
509 event->destroy = hw_perf_event_destroy;
510
511 /* Finally, validate version and authorization of the counter set */
512 err = validate_ctr_auth(hwc);
513 if (!err)
514 err = validate_ctr_version(hwc, set);
515
516 return err;
517}
518
519/* Events CPU_CYLCES and INSTRUCTIONS can be submitted with two different
520 * attribute::type values:
521 * - PERF_TYPE_HARDWARE:
522 * - pmu->type:
523 * Handle both type of invocations identical. They address the same hardware.
524 * The result is different when event modifiers exclude_kernel and/or
525 * exclude_user are also set.
526 */
527static int cpumf_pmu_event_type(struct perf_event *event)
528{
529 u64 ev = event->attr.config;
530
531 if (cpumf_generic_events_basic[PERF_COUNT_HW_CPU_CYCLES] == ev ||
532 cpumf_generic_events_basic[PERF_COUNT_HW_INSTRUCTIONS] == ev ||
533 cpumf_generic_events_user[PERF_COUNT_HW_CPU_CYCLES] == ev ||
534 cpumf_generic_events_user[PERF_COUNT_HW_INSTRUCTIONS] == ev)
535 return PERF_TYPE_HARDWARE;
536 return PERF_TYPE_RAW;
537}
538
539static int cpumf_pmu_event_init(struct perf_event *event)
540{
541 unsigned int type = event->attr.type;
542 int err;
543
544 if (type == PERF_TYPE_HARDWARE || type == PERF_TYPE_RAW)
545 err = __hw_perf_event_init(event, type);
546 else if (event->pmu->type == type)
547 /* Registered as unknown PMU */
548 err = __hw_perf_event_init(event, cpumf_pmu_event_type(event));
549 else
550 return -ENOENT;
551
552 if (unlikely(err) && event->destroy)
553 event->destroy(event);
554
555 return err;
556}
557
558static int hw_perf_event_reset(struct perf_event *event)
559{
560 u64 prev, new;
561 int err;
562
563 do {
564 prev = local64_read(&event->hw.prev_count);
565 err = ecctr(event->hw.config, &new);
566 if (err) {
567 if (err != 3)
568 break;
569 /* The counter is not (yet) available. This
570 * might happen if the counter set to which
571 * this counter belongs is in the disabled
572 * state.
573 */
574 new = 0;
575 }
576 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
577
578 return err;
579}
580
581static void hw_perf_event_update(struct perf_event *event)
582{
583 u64 prev, new, delta;
584 int err;
585
586 do {
587 prev = local64_read(&event->hw.prev_count);
588 err = ecctr(event->hw.config, &new);
589 if (err)
590 return;
591 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
592
593 delta = (prev <= new) ? new - prev
594 : (-1ULL - prev) + new + 1; /* overflow */
595 local64_add(delta, &event->count);
596}
597
598static void cpumf_pmu_read(struct perf_event *event)
599{
600 if (event->hw.state & PERF_HES_STOPPED)
601 return;
602
603 hw_perf_event_update(event);
604}
605
606static void cpumf_pmu_start(struct perf_event *event, int flags)
607{
608 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
609 struct hw_perf_event *hwc = &event->hw;
610 int i;
611
612 if (!(hwc->state & PERF_HES_STOPPED))
613 return;
614
615 hwc->state = 0;
616
617 /* (Re-)enable and activate the counter set */
618 ctr_set_enable(&cpuhw->state, hwc->config_base);
619 ctr_set_start(&cpuhw->state, hwc->config_base);
620
621 /* The counter set to which this counter belongs can be already active.
622 * Because all counters in a set are active, the event->hw.prev_count
623 * needs to be synchronized. At this point, the counter set can be in
624 * the inactive or disabled state.
625 */
626 if (hwc->config == PERF_EVENT_CPUM_CF_DIAG) {
627 cpuhw->usedss = cfdiag_getctr(cpuhw->start,
628 sizeof(cpuhw->start),
629 hwc->config_base, true);
630 } else {
631 hw_perf_event_reset(event);
632 }
633
634 /* Increment refcount for counter sets */
635 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i)
636 if ((hwc->config_base & cpumf_ctr_ctl[i]))
637 atomic_inc(&cpuhw->ctr_set[i]);
638}
639
640/* Create perf event sample with the counter sets as raw data. The sample
641 * is then pushed to the event subsystem and the function checks for
642 * possible event overflows. If an event overflow occurs, the PMU is
643 * stopped.
644 *
645 * Return non-zero if an event overflow occurred.
646 */
647static int cfdiag_push_sample(struct perf_event *event,
648 struct cpu_cf_events *cpuhw)
649{
650 struct perf_sample_data data;
651 struct perf_raw_record raw;
652 struct pt_regs regs;
653 int overflow;
654
655 /* Setup perf sample */
656 perf_sample_data_init(&data, 0, event->hw.last_period);
657 memset(®s, 0, sizeof(regs));
658 memset(&raw, 0, sizeof(raw));
659
660 if (event->attr.sample_type & PERF_SAMPLE_CPU)
661 data.cpu_entry.cpu = event->cpu;
662 if (event->attr.sample_type & PERF_SAMPLE_RAW) {
663 raw.frag.size = cpuhw->usedss;
664 raw.frag.data = cpuhw->stop;
665 raw.size = raw.frag.size;
666 data.raw = &raw;
667 data.sample_flags |= PERF_SAMPLE_RAW;
668 }
669
670 overflow = perf_event_overflow(event, &data, ®s);
671 debug_sprintf_event(cf_dbg, 3,
672 "%s event %#llx sample_type %#llx raw %d ov %d\n",
673 __func__, event->hw.config,
674 event->attr.sample_type, raw.size, overflow);
675 if (overflow)
676 event->pmu->stop(event, 0);
677
678 perf_event_update_userpage(event);
679 return overflow;
680}
681
682static void cpumf_pmu_stop(struct perf_event *event, int flags)
683{
684 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
685 struct hw_perf_event *hwc = &event->hw;
686 int i;
687
688 if (!(hwc->state & PERF_HES_STOPPED)) {
689 /* Decrement reference count for this counter set and if this
690 * is the last used counter in the set, clear activation
691 * control and set the counter set state to inactive.
692 */
693 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
694 if (!(hwc->config_base & cpumf_ctr_ctl[i]))
695 continue;
696 if (!atomic_dec_return(&cpuhw->ctr_set[i]))
697 ctr_set_stop(&cpuhw->state, cpumf_ctr_ctl[i]);
698 }
699 hwc->state |= PERF_HES_STOPPED;
700 }
701
702 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
703 if (hwc->config == PERF_EVENT_CPUM_CF_DIAG) {
704 local64_inc(&event->count);
705 cpuhw->usedss = cfdiag_getctr(cpuhw->stop,
706 sizeof(cpuhw->stop),
707 event->hw.config_base,
708 false);
709 if (cfdiag_diffctr(cpuhw, event->hw.config_base))
710 cfdiag_push_sample(event, cpuhw);
711 } else if (cpuhw->flags & PMU_F_RESERVED) {
712 /* Only update when PMU not hotplugged off */
713 hw_perf_event_update(event);
714 }
715 hwc->state |= PERF_HES_UPTODATE;
716 }
717}
718
719static int cpumf_pmu_add(struct perf_event *event, int flags)
720{
721 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
722
723 ctr_set_enable(&cpuhw->state, event->hw.config_base);
724 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
725
726 if (flags & PERF_EF_START)
727 cpumf_pmu_start(event, PERF_EF_RELOAD);
728
729 return 0;
730}
731
732static void cpumf_pmu_del(struct perf_event *event, int flags)
733{
734 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
735 int i;
736
737 cpumf_pmu_stop(event, PERF_EF_UPDATE);
738
739 /* Check if any counter in the counter set is still used. If not used,
740 * change the counter set to the disabled state. This also clears the
741 * content of all counters in the set.
742 *
743 * When a new perf event has been added but not yet started, this can
744 * clear enable control and resets all counters in a set. Therefore,
745 * cpumf_pmu_start() always has to reenable a counter set.
746 */
747 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i)
748 if (!atomic_read(&cpuhw->ctr_set[i]))
749 ctr_set_disable(&cpuhw->state, cpumf_ctr_ctl[i]);
750}
751
752/* Performance monitoring unit for s390x */
753static struct pmu cpumf_pmu = {
754 .task_ctx_nr = perf_sw_context,
755 .capabilities = PERF_PMU_CAP_NO_INTERRUPT,
756 .pmu_enable = cpumf_pmu_enable,
757 .pmu_disable = cpumf_pmu_disable,
758 .event_init = cpumf_pmu_event_init,
759 .add = cpumf_pmu_add,
760 .del = cpumf_pmu_del,
761 .start = cpumf_pmu_start,
762 .stop = cpumf_pmu_stop,
763 .read = cpumf_pmu_read,
764};
765
766static int cfset_init(void);
767static int __init cpumf_pmu_init(void)
768{
769 int rc;
770
771 if (!kernel_cpumcf_avail())
772 return -ENODEV;
773
774 /* Setup s390dbf facility */
775 cf_dbg = debug_register(KMSG_COMPONENT, 2, 1, 128);
776 if (!cf_dbg) {
777 pr_err("Registration of s390dbf(cpum_cf) failed\n");
778 return -ENOMEM;
779 }
780 debug_register_view(cf_dbg, &debug_sprintf_view);
781
782 cpumf_pmu.attr_groups = cpumf_cf_event_group();
783 rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", -1);
784 if (rc) {
785 debug_unregister_view(cf_dbg, &debug_sprintf_view);
786 debug_unregister(cf_dbg);
787 pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
788 } else if (stccm_avail()) { /* Setup counter set device */
789 cfset_init();
790 }
791 return rc;
792}
793
794/* Support for the CPU Measurement Facility counter set extraction using
795 * device /dev/hwctr. This allows user space programs to extract complete
796 * counter set via normal file operations.
797 */
798
799static atomic_t cfset_opencnt = ATOMIC_INIT(0); /* Access count */
800static DEFINE_MUTEX(cfset_ctrset_mutex);/* Synchronize access to hardware */
801struct cfset_call_on_cpu_parm { /* Parm struct for smp_call_on_cpu */
802 unsigned int sets; /* Counter set bit mask */
803 atomic_t cpus_ack; /* # CPUs successfully executed func */
804};
805
806static struct cfset_session { /* CPUs and counter set bit mask */
807 struct list_head head; /* Head of list of active processes */
808} cfset_session = {
809 .head = LIST_HEAD_INIT(cfset_session.head)
810};
811
812struct cfset_request { /* CPUs and counter set bit mask */
813 unsigned long ctrset; /* Bit mask of counter set to read */
814 cpumask_t mask; /* CPU mask to read from */
815 struct list_head node; /* Chain to cfset_session.head */
816};
817
818static void cfset_session_init(void)
819{
820 INIT_LIST_HEAD(&cfset_session.head);
821}
822
823/* Remove current request from global bookkeeping. Maintain a counter set bit
824 * mask on a per CPU basis.
825 * Done in process context under mutex protection.
826 */
827static void cfset_session_del(struct cfset_request *p)
828{
829 list_del(&p->node);
830}
831
832/* Add current request to global bookkeeping. Maintain a counter set bit mask
833 * on a per CPU basis.
834 * Done in process context under mutex protection.
835 */
836static void cfset_session_add(struct cfset_request *p)
837{
838 list_add(&p->node, &cfset_session.head);
839}
840
841/* The /dev/hwctr device access uses PMU_F_IN_USE to mark the device access
842 * path is currently used.
843 * The cpu_cf_events::dev_state is used to denote counter sets in use by this
844 * interface. It is always or'ed in. If this interface is not active, its
845 * value is zero and no additional counter sets will be included.
846 *
847 * The cpu_cf_events::state is used by the perf_event_open SVC and remains
848 * unchanged.
849 *
850 * perf_pmu_enable() and perf_pmu_enable() and its call backs
851 * cpumf_pmu_enable() and cpumf_pmu_disable() are called by the
852 * performance measurement subsystem to enable per process
853 * CPU Measurement counter facility.
854 * The XXX_enable() and XXX_disable functions are used to turn off
855 * x86 performance monitoring interrupt (PMI) during scheduling.
856 * s390 uses these calls to temporarily stop and resume the active CPU
857 * counters sets during scheduling.
858 *
859 * We do allow concurrent access of perf_event_open() SVC and /dev/hwctr
860 * device access. The perf_event_open() SVC interface makes a lot of effort
861 * to only run the counters while the calling process is actively scheduled
862 * to run.
863 * When /dev/hwctr interface is also used at the same time, the counter sets
864 * will keep running, even when the process is scheduled off a CPU.
865 * However this is not a problem and does not lead to wrong counter values
866 * for the perf_event_open() SVC. The current counter value will be recorded
867 * during schedule-in. At schedule-out time the current counter value is
868 * extracted again and the delta is calculated and added to the event.
869 */
870/* Stop all counter sets via ioctl interface */
871static void cfset_ioctl_off(void *parm)
872{
873 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
874 struct cfset_call_on_cpu_parm *p = parm;
875 int rc;
876
877 /* Check if any counter set used by /dev/hwc */
878 for (rc = CPUMF_CTR_SET_BASIC; rc < CPUMF_CTR_SET_MAX; ++rc)
879 if ((p->sets & cpumf_ctr_ctl[rc])) {
880 if (!atomic_dec_return(&cpuhw->ctr_set[rc])) {
881 ctr_set_disable(&cpuhw->dev_state,
882 cpumf_ctr_ctl[rc]);
883 ctr_set_stop(&cpuhw->dev_state,
884 cpumf_ctr_ctl[rc]);
885 }
886 }
887 /* Keep perf_event_open counter sets */
888 rc = lcctl(cpuhw->dev_state | cpuhw->state);
889 if (rc)
890 pr_err("Counter set stop %#llx of /dev/%s failed rc=%i\n",
891 cpuhw->state, S390_HWCTR_DEVICE, rc);
892 if (!cpuhw->dev_state)
893 cpuhw->flags &= ~PMU_F_IN_USE;
894 debug_sprintf_event(cf_dbg, 4, "%s rc %d state %#llx dev_state %#llx\n",
895 __func__, rc, cpuhw->state, cpuhw->dev_state);
896}
897
898/* Start counter sets on particular CPU */
899static void cfset_ioctl_on(void *parm)
900{
901 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
902 struct cfset_call_on_cpu_parm *p = parm;
903 int rc;
904
905 cpuhw->flags |= PMU_F_IN_USE;
906 ctr_set_enable(&cpuhw->dev_state, p->sets);
907 ctr_set_start(&cpuhw->dev_state, p->sets);
908 for (rc = CPUMF_CTR_SET_BASIC; rc < CPUMF_CTR_SET_MAX; ++rc)
909 if ((p->sets & cpumf_ctr_ctl[rc]))
910 atomic_inc(&cpuhw->ctr_set[rc]);
911 rc = lcctl(cpuhw->dev_state | cpuhw->state); /* Start counter sets */
912 if (!rc)
913 atomic_inc(&p->cpus_ack);
914 else
915 pr_err("Counter set start %#llx of /dev/%s failed rc=%i\n",
916 cpuhw->dev_state | cpuhw->state, S390_HWCTR_DEVICE, rc);
917 debug_sprintf_event(cf_dbg, 4, "%s rc %d state %#llx dev_state %#llx\n",
918 __func__, rc, cpuhw->state, cpuhw->dev_state);
919}
920
921static void cfset_release_cpu(void *p)
922{
923 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
924 int rc;
925
926 debug_sprintf_event(cf_dbg, 4, "%s state %#llx dev_state %#llx\n",
927 __func__, cpuhw->state, cpuhw->dev_state);
928 cpuhw->dev_state = 0;
929 rc = lcctl(cpuhw->state); /* Keep perf_event_open counter sets */
930 if (rc)
931 pr_err("Counter set release %#llx of /dev/%s failed rc=%i\n",
932 cpuhw->state, S390_HWCTR_DEVICE, rc);
933}
934
935/* This modifies the process CPU mask to adopt it to the currently online
936 * CPUs. Offline CPUs can not be addresses. This call terminates the access
937 * and is usually followed by close() or a new iotcl(..., START, ...) which
938 * creates a new request structure.
939 */
940static void cfset_all_stop(struct cfset_request *req)
941{
942 struct cfset_call_on_cpu_parm p = {
943 .sets = req->ctrset,
944 };
945
946 cpumask_and(&req->mask, &req->mask, cpu_online_mask);
947 on_each_cpu_mask(&req->mask, cfset_ioctl_off, &p, 1);
948}
949
950/* Release function is also called when application gets terminated without
951 * doing a proper ioctl(..., S390_HWCTR_STOP, ...) command.
952 */
953static int cfset_release(struct inode *inode, struct file *file)
954{
955 mutex_lock(&cfset_ctrset_mutex);
956 /* Open followed by close/exit has no private_data */
957 if (file->private_data) {
958 cfset_all_stop(file->private_data);
959 cfset_session_del(file->private_data);
960 kfree(file->private_data);
961 file->private_data = NULL;
962 }
963 if (!atomic_dec_return(&cfset_opencnt))
964 on_each_cpu(cfset_release_cpu, NULL, 1);
965 mutex_unlock(&cfset_ctrset_mutex);
966
967 hw_perf_event_destroy(NULL);
968 return 0;
969}
970
971static int cfset_open(struct inode *inode, struct file *file)
972{
973 if (!capable(CAP_SYS_ADMIN))
974 return -EPERM;
975 mutex_lock(&cfset_ctrset_mutex);
976 if (atomic_inc_return(&cfset_opencnt) == 1)
977 cfset_session_init();
978 mutex_unlock(&cfset_ctrset_mutex);
979
980 cpumf_hw_inuse();
981 file->private_data = NULL;
982 /* nonseekable_open() never fails */
983 return nonseekable_open(inode, file);
984}
985
986static int cfset_all_start(struct cfset_request *req)
987{
988 struct cfset_call_on_cpu_parm p = {
989 .sets = req->ctrset,
990 .cpus_ack = ATOMIC_INIT(0),
991 };
992 cpumask_var_t mask;
993 int rc = 0;
994
995 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
996 return -ENOMEM;
997 cpumask_and(mask, &req->mask, cpu_online_mask);
998 on_each_cpu_mask(mask, cfset_ioctl_on, &p, 1);
999 if (atomic_read(&p.cpus_ack) != cpumask_weight(mask)) {
1000 on_each_cpu_mask(mask, cfset_ioctl_off, &p, 1);
1001 rc = -EIO;
1002 debug_sprintf_event(cf_dbg, 4, "%s CPUs missing", __func__);
1003 }
1004 free_cpumask_var(mask);
1005 return rc;
1006}
1007
1008
1009/* Return the maximum required space for all possible CPUs in case one
1010 * CPU will be onlined during the START, READ, STOP cycles.
1011 * To find out the size of the counter sets, any one CPU will do. They
1012 * all have the same counter sets.
1013 */
1014static size_t cfset_needspace(unsigned int sets)
1015{
1016 struct cpu_cf_events *cpuhw = get_cpu_ptr(&cpu_cf_events);
1017 size_t bytes = 0;
1018 int i;
1019
1020 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1021 if (!(sets & cpumf_ctr_ctl[i]))
1022 continue;
1023 bytes += cpum_cf_ctrset_size(i, &cpuhw->info) * sizeof(u64) +
1024 sizeof(((struct s390_ctrset_setdata *)0)->set) +
1025 sizeof(((struct s390_ctrset_setdata *)0)->no_cnts);
1026 }
1027 bytes = sizeof(((struct s390_ctrset_read *)0)->no_cpus) + nr_cpu_ids *
1028 (bytes + sizeof(((struct s390_ctrset_cpudata *)0)->cpu_nr) +
1029 sizeof(((struct s390_ctrset_cpudata *)0)->no_sets));
1030 put_cpu_ptr(&cpu_cf_events);
1031 return bytes;
1032}
1033
1034static int cfset_all_copy(unsigned long arg, cpumask_t *mask)
1035{
1036 struct s390_ctrset_read __user *ctrset_read;
1037 unsigned int cpu, cpus, rc;
1038 void __user *uptr;
1039
1040 ctrset_read = (struct s390_ctrset_read __user *)arg;
1041 uptr = ctrset_read->data;
1042 for_each_cpu(cpu, mask) {
1043 struct cpu_cf_events *cpuhw = per_cpu_ptr(&cpu_cf_events, cpu);
1044 struct s390_ctrset_cpudata __user *ctrset_cpudata;
1045
1046 ctrset_cpudata = uptr;
1047 rc = put_user(cpu, &ctrset_cpudata->cpu_nr);
1048 rc |= put_user(cpuhw->sets, &ctrset_cpudata->no_sets);
1049 rc |= copy_to_user(ctrset_cpudata->data, cpuhw->data,
1050 cpuhw->used);
1051 if (rc)
1052 return -EFAULT;
1053 uptr += sizeof(struct s390_ctrset_cpudata) + cpuhw->used;
1054 cond_resched();
1055 }
1056 cpus = cpumask_weight(mask);
1057 if (put_user(cpus, &ctrset_read->no_cpus))
1058 return -EFAULT;
1059 debug_sprintf_event(cf_dbg, 4, "%s copied %ld\n", __func__,
1060 uptr - (void __user *)ctrset_read->data);
1061 return 0;
1062}
1063
1064static size_t cfset_cpuset_read(struct s390_ctrset_setdata *p, int ctrset,
1065 int ctrset_size, size_t room)
1066{
1067 size_t need = 0;
1068 int rc = -1;
1069
1070 need = sizeof(*p) + sizeof(u64) * ctrset_size;
1071 if (need <= room) {
1072 p->set = cpumf_ctr_ctl[ctrset];
1073 p->no_cnts = ctrset_size;
1074 rc = ctr_stcctm(ctrset, ctrset_size, (u64 *)p->cv);
1075 if (rc == 3) /* Nothing stored */
1076 need = 0;
1077 }
1078 return need;
1079}
1080
1081/* Read all counter sets. */
1082static void cfset_cpu_read(void *parm)
1083{
1084 struct cpu_cf_events *cpuhw = this_cpu_ptr(&cpu_cf_events);
1085 struct cfset_call_on_cpu_parm *p = parm;
1086 int set, set_size;
1087 size_t space;
1088
1089 /* No data saved yet */
1090 cpuhw->used = 0;
1091 cpuhw->sets = 0;
1092 memset(cpuhw->data, 0, sizeof(cpuhw->data));
1093
1094 /* Scan the counter sets */
1095 for (set = CPUMF_CTR_SET_BASIC; set < CPUMF_CTR_SET_MAX; ++set) {
1096 struct s390_ctrset_setdata *sp = (void *)cpuhw->data +
1097 cpuhw->used;
1098
1099 if (!(p->sets & cpumf_ctr_ctl[set]))
1100 continue; /* Counter set not in list */
1101 set_size = cpum_cf_ctrset_size(set, &cpuhw->info);
1102 space = sizeof(cpuhw->data) - cpuhw->used;
1103 space = cfset_cpuset_read(sp, set, set_size, space);
1104 if (space) {
1105 cpuhw->used += space;
1106 cpuhw->sets += 1;
1107 }
1108 }
1109 debug_sprintf_event(cf_dbg, 4, "%s sets %d used %zd\n", __func__,
1110 cpuhw->sets, cpuhw->used);
1111}
1112
1113static int cfset_all_read(unsigned long arg, struct cfset_request *req)
1114{
1115 struct cfset_call_on_cpu_parm p;
1116 cpumask_var_t mask;
1117 int rc;
1118
1119 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
1120 return -ENOMEM;
1121
1122 p.sets = req->ctrset;
1123 cpumask_and(mask, &req->mask, cpu_online_mask);
1124 on_each_cpu_mask(mask, cfset_cpu_read, &p, 1);
1125 rc = cfset_all_copy(arg, mask);
1126 free_cpumask_var(mask);
1127 return rc;
1128}
1129
1130static long cfset_ioctl_read(unsigned long arg, struct cfset_request *req)
1131{
1132 struct s390_ctrset_read read;
1133 int ret = -ENODATA;
1134
1135 if (req && req->ctrset) {
1136 if (copy_from_user(&read, (char __user *)arg, sizeof(read)))
1137 return -EFAULT;
1138 ret = cfset_all_read(arg, req);
1139 }
1140 return ret;
1141}
1142
1143static long cfset_ioctl_stop(struct file *file)
1144{
1145 struct cfset_request *req = file->private_data;
1146 int ret = -ENXIO;
1147
1148 if (req) {
1149 cfset_all_stop(req);
1150 cfset_session_del(req);
1151 kfree(req);
1152 file->private_data = NULL;
1153 ret = 0;
1154 }
1155 return ret;
1156}
1157
1158static long cfset_ioctl_start(unsigned long arg, struct file *file)
1159{
1160 struct s390_ctrset_start __user *ustart;
1161 struct s390_ctrset_start start;
1162 struct cfset_request *preq;
1163 void __user *umask;
1164 unsigned int len;
1165 int ret = 0;
1166 size_t need;
1167
1168 if (file->private_data)
1169 return -EBUSY;
1170 ustart = (struct s390_ctrset_start __user *)arg;
1171 if (copy_from_user(&start, ustart, sizeof(start)))
1172 return -EFAULT;
1173 if (start.version != S390_HWCTR_START_VERSION)
1174 return -EINVAL;
1175 if (start.counter_sets & ~(cpumf_ctr_ctl[CPUMF_CTR_SET_BASIC] |
1176 cpumf_ctr_ctl[CPUMF_CTR_SET_USER] |
1177 cpumf_ctr_ctl[CPUMF_CTR_SET_CRYPTO] |
1178 cpumf_ctr_ctl[CPUMF_CTR_SET_EXT] |
1179 cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG]))
1180 return -EINVAL; /* Invalid counter set */
1181 if (!start.counter_sets)
1182 return -EINVAL; /* No counter set at all? */
1183
1184 preq = kzalloc(sizeof(*preq), GFP_KERNEL);
1185 if (!preq)
1186 return -ENOMEM;
1187 cpumask_clear(&preq->mask);
1188 len = min_t(u64, start.cpumask_len, cpumask_size());
1189 umask = (void __user *)start.cpumask;
1190 if (copy_from_user(&preq->mask, umask, len)) {
1191 kfree(preq);
1192 return -EFAULT;
1193 }
1194 if (cpumask_empty(&preq->mask)) {
1195 kfree(preq);
1196 return -EINVAL;
1197 }
1198 need = cfset_needspace(start.counter_sets);
1199 if (put_user(need, &ustart->data_bytes)) {
1200 kfree(preq);
1201 return -EFAULT;
1202 }
1203 preq->ctrset = start.counter_sets;
1204 ret = cfset_all_start(preq);
1205 if (!ret) {
1206 cfset_session_add(preq);
1207 file->private_data = preq;
1208 debug_sprintf_event(cf_dbg, 4, "%s set %#lx need %ld ret %d\n",
1209 __func__, preq->ctrset, need, ret);
1210 } else {
1211 kfree(preq);
1212 }
1213 return ret;
1214}
1215
1216/* Entry point to the /dev/hwctr device interface.
1217 * The ioctl system call supports three subcommands:
1218 * S390_HWCTR_START: Start the specified counter sets on a CPU list. The
1219 * counter set keeps running until explicitly stopped. Returns the number
1220 * of bytes needed to store the counter values. If another S390_HWCTR_START
1221 * ioctl subcommand is called without a previous S390_HWCTR_STOP stop
1222 * command on the same file descriptor, -EBUSY is returned.
1223 * S390_HWCTR_READ: Read the counter set values from specified CPU list given
1224 * with the S390_HWCTR_START command.
1225 * S390_HWCTR_STOP: Stops the counter sets on the CPU list given with the
1226 * previous S390_HWCTR_START subcommand.
1227 */
1228static long cfset_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1229{
1230 int ret;
1231
1232 cpus_read_lock();
1233 mutex_lock(&cfset_ctrset_mutex);
1234 switch (cmd) {
1235 case S390_HWCTR_START:
1236 ret = cfset_ioctl_start(arg, file);
1237 break;
1238 case S390_HWCTR_STOP:
1239 ret = cfset_ioctl_stop(file);
1240 break;
1241 case S390_HWCTR_READ:
1242 ret = cfset_ioctl_read(arg, file->private_data);
1243 break;
1244 default:
1245 ret = -ENOTTY;
1246 break;
1247 }
1248 mutex_unlock(&cfset_ctrset_mutex);
1249 cpus_read_unlock();
1250 return ret;
1251}
1252
1253static const struct file_operations cfset_fops = {
1254 .owner = THIS_MODULE,
1255 .open = cfset_open,
1256 .release = cfset_release,
1257 .unlocked_ioctl = cfset_ioctl,
1258 .compat_ioctl = cfset_ioctl,
1259 .llseek = no_llseek
1260};
1261
1262static struct miscdevice cfset_dev = {
1263 .name = S390_HWCTR_DEVICE,
1264 .minor = MISC_DYNAMIC_MINOR,
1265 .fops = &cfset_fops,
1266};
1267
1268/* Hotplug add of a CPU. Scan through all active processes and add
1269 * that CPU to the list of CPUs supplied with ioctl(..., START, ...).
1270 */
1271int cfset_online_cpu(unsigned int cpu)
1272{
1273 struct cfset_call_on_cpu_parm p;
1274 struct cfset_request *rp;
1275
1276 mutex_lock(&cfset_ctrset_mutex);
1277 if (!list_empty(&cfset_session.head)) {
1278 list_for_each_entry(rp, &cfset_session.head, node) {
1279 p.sets = rp->ctrset;
1280 cfset_ioctl_on(&p);
1281 cpumask_set_cpu(cpu, &rp->mask);
1282 }
1283 }
1284 mutex_unlock(&cfset_ctrset_mutex);
1285 return 0;
1286}
1287
1288/* Hotplug remove of a CPU. Scan through all active processes and clear
1289 * that CPU from the list of CPUs supplied with ioctl(..., START, ...).
1290 */
1291int cfset_offline_cpu(unsigned int cpu)
1292{
1293 struct cfset_call_on_cpu_parm p;
1294 struct cfset_request *rp;
1295
1296 mutex_lock(&cfset_ctrset_mutex);
1297 if (!list_empty(&cfset_session.head)) {
1298 list_for_each_entry(rp, &cfset_session.head, node) {
1299 p.sets = rp->ctrset;
1300 cfset_ioctl_off(&p);
1301 cpumask_clear_cpu(cpu, &rp->mask);
1302 }
1303 }
1304 mutex_unlock(&cfset_ctrset_mutex);
1305 return 0;
1306}
1307
1308static void cfdiag_read(struct perf_event *event)
1309{
1310 debug_sprintf_event(cf_dbg, 3, "%s event %#llx count %ld\n", __func__,
1311 event->attr.config, local64_read(&event->count));
1312}
1313
1314static int get_authctrsets(void)
1315{
1316 struct cpu_cf_events *cpuhw;
1317 unsigned long auth = 0;
1318 enum cpumf_ctr_set i;
1319
1320 cpuhw = &get_cpu_var(cpu_cf_events);
1321 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1322 if (cpuhw->info.auth_ctl & cpumf_ctr_ctl[i])
1323 auth |= cpumf_ctr_ctl[i];
1324 }
1325 put_cpu_var(cpu_cf_events);
1326 return auth;
1327}
1328
1329/* Setup the event. Test for authorized counter sets and only include counter
1330 * sets which are authorized at the time of the setup. Including unauthorized
1331 * counter sets result in specification exception (and panic).
1332 */
1333static int cfdiag_event_init2(struct perf_event *event)
1334{
1335 struct perf_event_attr *attr = &event->attr;
1336 int err = 0;
1337
1338 /* Set sample_period to indicate sampling */
1339 event->hw.config = attr->config;
1340 event->hw.sample_period = attr->sample_period;
1341 local64_set(&event->hw.period_left, event->hw.sample_period);
1342 local64_set(&event->count, 0);
1343 event->hw.last_period = event->hw.sample_period;
1344
1345 /* Add all authorized counter sets to config_base. The
1346 * the hardware init function is either called per-cpu or just once
1347 * for all CPUS (event->cpu == -1). This depends on the whether
1348 * counting is started for all CPUs or on a per workload base where
1349 * the perf event moves from one CPU to another CPU.
1350 * Checking the authorization on any CPU is fine as the hardware
1351 * applies the same authorization settings to all CPUs.
1352 */
1353 event->hw.config_base = get_authctrsets();
1354
1355 /* No authorized counter sets, nothing to count/sample */
1356 if (!event->hw.config_base)
1357 err = -EINVAL;
1358
1359 debug_sprintf_event(cf_dbg, 5, "%s err %d config_base %#lx\n",
1360 __func__, err, event->hw.config_base);
1361 return err;
1362}
1363
1364static int cfdiag_event_init(struct perf_event *event)
1365{
1366 struct perf_event_attr *attr = &event->attr;
1367 int err = -ENOENT;
1368
1369 if (event->attr.config != PERF_EVENT_CPUM_CF_DIAG ||
1370 event->attr.type != event->pmu->type)
1371 goto out;
1372
1373 /* Raw events are used to access counters directly,
1374 * hence do not permit excludes.
1375 * This event is useless without PERF_SAMPLE_RAW to return counter set
1376 * values as raw data.
1377 */
1378 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv ||
1379 !(attr->sample_type & (PERF_SAMPLE_CPU | PERF_SAMPLE_RAW))) {
1380 err = -EOPNOTSUPP;
1381 goto out;
1382 }
1383
1384 /* Initialize for using the CPU-measurement counter facility */
1385 cpumf_hw_inuse();
1386 event->destroy = hw_perf_event_destroy;
1387
1388 err = cfdiag_event_init2(event);
1389 if (unlikely(err))
1390 event->destroy(event);
1391out:
1392 return err;
1393}
1394
1395/* Create cf_diag/events/CF_DIAG event sysfs file. This counter is used
1396 * to collect the complete counter sets for a scheduled process. Target
1397 * are complete counter sets attached as raw data to the artificial event.
1398 * This results in complete counter sets available when a process is
1399 * scheduled. Contains the delta of every counter while the process was
1400 * running.
1401 */
1402CPUMF_EVENT_ATTR(CF_DIAG, CF_DIAG, PERF_EVENT_CPUM_CF_DIAG);
1403
1404static struct attribute *cfdiag_events_attr[] = {
1405 CPUMF_EVENT_PTR(CF_DIAG, CF_DIAG),
1406 NULL,
1407};
1408
1409PMU_FORMAT_ATTR(event, "config:0-63");
1410
1411static struct attribute *cfdiag_format_attr[] = {
1412 &format_attr_event.attr,
1413 NULL,
1414};
1415
1416static struct attribute_group cfdiag_events_group = {
1417 .name = "events",
1418 .attrs = cfdiag_events_attr,
1419};
1420static struct attribute_group cfdiag_format_group = {
1421 .name = "format",
1422 .attrs = cfdiag_format_attr,
1423};
1424static const struct attribute_group *cfdiag_attr_groups[] = {
1425 &cfdiag_events_group,
1426 &cfdiag_format_group,
1427 NULL,
1428};
1429
1430/* Performance monitoring unit for event CF_DIAG. Since this event
1431 * is also started and stopped via the perf_event_open() system call, use
1432 * the same event enable/disable call back functions. They do not
1433 * have a pointer to the perf_event strcture as first parameter.
1434 *
1435 * The functions XXX_add, XXX_del, XXX_start and XXX_stop are also common.
1436 * Reuse them and distinguish the event (always first parameter) via
1437 * 'config' member.
1438 */
1439static struct pmu cf_diag = {
1440 .task_ctx_nr = perf_sw_context,
1441 .event_init = cfdiag_event_init,
1442 .pmu_enable = cpumf_pmu_enable,
1443 .pmu_disable = cpumf_pmu_disable,
1444 .add = cpumf_pmu_add,
1445 .del = cpumf_pmu_del,
1446 .start = cpumf_pmu_start,
1447 .stop = cpumf_pmu_stop,
1448 .read = cfdiag_read,
1449
1450 .attr_groups = cfdiag_attr_groups
1451};
1452
1453/* Calculate memory needed to store all counter sets together with header and
1454 * trailer data. This is independent of the counter set authorization which
1455 * can vary depending on the configuration.
1456 */
1457static size_t cfdiag_maxsize(struct cpumf_ctr_info *info)
1458{
1459 size_t max_size = sizeof(struct cf_trailer_entry);
1460 enum cpumf_ctr_set i;
1461
1462 for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1463 size_t size = cpum_cf_ctrset_size(i, info);
1464
1465 if (size)
1466 max_size += size * sizeof(u64) +
1467 sizeof(struct cf_ctrset_entry);
1468 }
1469 return max_size;
1470}
1471
1472/* Get the CPU speed, try sampling facility first and CPU attributes second. */
1473static void cfdiag_get_cpu_speed(void)
1474{
1475 unsigned long mhz;
1476
1477 if (cpum_sf_avail()) { /* Sampling facility first */
1478 struct hws_qsi_info_block si;
1479
1480 memset(&si, 0, sizeof(si));
1481 if (!qsi(&si)) {
1482 cfdiag_cpu_speed = si.cpu_speed;
1483 return;
1484 }
1485 }
1486
1487 /* Fallback: CPU speed extract static part. Used in case
1488 * CPU Measurement Sampling Facility is turned off.
1489 */
1490 mhz = __ecag(ECAG_CPU_ATTRIBUTE, 0);
1491 if (mhz != -1UL)
1492 cfdiag_cpu_speed = mhz & 0xffffffff;
1493}
1494
1495static int cfset_init(void)
1496{
1497 struct cpumf_ctr_info info;
1498 size_t need;
1499 int rc;
1500
1501 if (qctri(&info))
1502 return -ENODEV;
1503
1504 cfdiag_get_cpu_speed();
1505 /* Make sure the counter set data fits into predefined buffer. */
1506 need = cfdiag_maxsize(&info);
1507 if (need > sizeof(((struct cpu_cf_events *)0)->start)) {
1508 pr_err("Insufficient memory for PMU(cpum_cf_diag) need=%zu\n",
1509 need);
1510 return -ENOMEM;
1511 }
1512
1513 rc = misc_register(&cfset_dev);
1514 if (rc) {
1515 pr_err("Registration of /dev/%s failed rc=%i\n",
1516 cfset_dev.name, rc);
1517 goto out;
1518 }
1519
1520 rc = perf_pmu_register(&cf_diag, "cpum_cf_diag", -1);
1521 if (rc) {
1522 misc_deregister(&cfset_dev);
1523 pr_err("Registration of PMU(cpum_cf_diag) failed with rc=%i\n",
1524 rc);
1525 }
1526out:
1527 return rc;
1528}
1529
1530device_initcall(cpumf_pmu_init);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance event support for s390x - CPU-measurement Counter Facility
4 *
5 * Copyright IBM Corp. 2012, 2017
6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.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/perf_event.h>
14#include <linux/percpu.h>
15#include <linux/notifier.h>
16#include <linux/init.h>
17#include <linux/export.h>
18#include <asm/ctl_reg.h>
19#include <asm/irq.h>
20#include <asm/cpu_mf.h>
21
22enum cpumf_ctr_set {
23 CPUMF_CTR_SET_BASIC = 0, /* Basic Counter Set */
24 CPUMF_CTR_SET_USER = 1, /* Problem-State Counter Set */
25 CPUMF_CTR_SET_CRYPTO = 2, /* Crypto-Activity Counter Set */
26 CPUMF_CTR_SET_EXT = 3, /* Extended Counter Set */
27 CPUMF_CTR_SET_MT_DIAG = 4, /* MT-diagnostic Counter Set */
28
29 /* Maximum number of counter sets */
30 CPUMF_CTR_SET_MAX,
31};
32
33#define CPUMF_LCCTL_ENABLE_SHIFT 16
34#define CPUMF_LCCTL_ACTCTL_SHIFT 0
35static const u64 cpumf_state_ctl[CPUMF_CTR_SET_MAX] = {
36 [CPUMF_CTR_SET_BASIC] = 0x02,
37 [CPUMF_CTR_SET_USER] = 0x04,
38 [CPUMF_CTR_SET_CRYPTO] = 0x08,
39 [CPUMF_CTR_SET_EXT] = 0x01,
40 [CPUMF_CTR_SET_MT_DIAG] = 0x20,
41};
42
43static void ctr_set_enable(u64 *state, int ctr_set)
44{
45 *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT;
46}
47static void ctr_set_disable(u64 *state, int ctr_set)
48{
49 *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT);
50}
51static void ctr_set_start(u64 *state, int ctr_set)
52{
53 *state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT;
54}
55static void ctr_set_stop(u64 *state, int ctr_set)
56{
57 *state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT);
58}
59
60/* Local CPUMF event structure */
61struct cpu_hw_events {
62 struct cpumf_ctr_info info;
63 atomic_t ctr_set[CPUMF_CTR_SET_MAX];
64 u64 state, tx_state;
65 unsigned int flags;
66 unsigned int txn_flags;
67};
68static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
69 .ctr_set = {
70 [CPUMF_CTR_SET_BASIC] = ATOMIC_INIT(0),
71 [CPUMF_CTR_SET_USER] = ATOMIC_INIT(0),
72 [CPUMF_CTR_SET_CRYPTO] = ATOMIC_INIT(0),
73 [CPUMF_CTR_SET_EXT] = ATOMIC_INIT(0),
74 [CPUMF_CTR_SET_MT_DIAG] = ATOMIC_INIT(0),
75 },
76 .state = 0,
77 .flags = 0,
78 .txn_flags = 0,
79};
80
81static enum cpumf_ctr_set get_counter_set(u64 event)
82{
83 int set = CPUMF_CTR_SET_MAX;
84
85 if (event < 32)
86 set = CPUMF_CTR_SET_BASIC;
87 else if (event < 64)
88 set = CPUMF_CTR_SET_USER;
89 else if (event < 128)
90 set = CPUMF_CTR_SET_CRYPTO;
91 else if (event < 256)
92 set = CPUMF_CTR_SET_EXT;
93 else if (event >= 448 && event < 496)
94 set = CPUMF_CTR_SET_MT_DIAG;
95
96 return set;
97}
98
99static int validate_ctr_version(const struct hw_perf_event *hwc)
100{
101 struct cpu_hw_events *cpuhw;
102 int err = 0;
103 u16 mtdiag_ctl;
104
105 cpuhw = &get_cpu_var(cpu_hw_events);
106
107 /* check required version for counter sets */
108 switch (hwc->config_base) {
109 case CPUMF_CTR_SET_BASIC:
110 case CPUMF_CTR_SET_USER:
111 if (cpuhw->info.cfvn < 1)
112 err = -EOPNOTSUPP;
113 break;
114 case CPUMF_CTR_SET_CRYPTO:
115 case CPUMF_CTR_SET_EXT:
116 if (cpuhw->info.csvn < 1)
117 err = -EOPNOTSUPP;
118 if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
119 (cpuhw->info.csvn == 2 && hwc->config > 175) ||
120 (cpuhw->info.csvn > 2 && hwc->config > 255))
121 err = -EOPNOTSUPP;
122 break;
123 case CPUMF_CTR_SET_MT_DIAG:
124 if (cpuhw->info.csvn <= 3)
125 err = -EOPNOTSUPP;
126 /*
127 * MT-diagnostic counters are read-only. The counter set
128 * is automatically enabled and activated on all CPUs with
129 * multithreading (SMT). Deactivation of multithreading
130 * also disables the counter set. State changes are ignored
131 * by lcctl(). Because Linux controls SMT enablement through
132 * a kernel parameter only, the counter set is either disabled
133 * or enabled and active.
134 *
135 * Thus, the counters can only be used if SMT is on and the
136 * counter set is enabled and active.
137 */
138 mtdiag_ctl = cpumf_state_ctl[CPUMF_CTR_SET_MT_DIAG];
139 if (!((cpuhw->info.auth_ctl & mtdiag_ctl) &&
140 (cpuhw->info.enable_ctl & mtdiag_ctl) &&
141 (cpuhw->info.act_ctl & mtdiag_ctl)))
142 err = -EOPNOTSUPP;
143 break;
144 }
145
146 put_cpu_var(cpu_hw_events);
147 return err;
148}
149
150static int validate_ctr_auth(const struct hw_perf_event *hwc)
151{
152 struct cpu_hw_events *cpuhw;
153 u64 ctrs_state;
154 int err = 0;
155
156 cpuhw = &get_cpu_var(cpu_hw_events);
157
158 /* Check authorization for cpu counter sets.
159 * If the particular CPU counter set is not authorized,
160 * return with -ENOENT in order to fall back to other
161 * PMUs that might suffice the event request.
162 */
163 ctrs_state = cpumf_state_ctl[hwc->config_base];
164 if (!(ctrs_state & cpuhw->info.auth_ctl))
165 err = -ENOENT;
166
167 put_cpu_var(cpu_hw_events);
168 return err;
169}
170
171/*
172 * Change the CPUMF state to active.
173 * Enable and activate the CPU-counter sets according
174 * to the per-cpu control state.
175 */
176static void cpumf_pmu_enable(struct pmu *pmu)
177{
178 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
179 int err;
180
181 if (cpuhw->flags & PMU_F_ENABLED)
182 return;
183
184 err = lcctl(cpuhw->state);
185 if (err) {
186 pr_err("Enabling the performance measuring unit "
187 "failed with rc=%x\n", err);
188 return;
189 }
190
191 cpuhw->flags |= PMU_F_ENABLED;
192}
193
194/*
195 * Change the CPUMF state to inactive.
196 * Disable and enable (inactive) the CPU-counter sets according
197 * to the per-cpu control state.
198 */
199static void cpumf_pmu_disable(struct pmu *pmu)
200{
201 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
202 int err;
203 u64 inactive;
204
205 if (!(cpuhw->flags & PMU_F_ENABLED))
206 return;
207
208 inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
209 err = lcctl(inactive);
210 if (err) {
211 pr_err("Disabling the performance measuring unit "
212 "failed with rc=%x\n", err);
213 return;
214 }
215
216 cpuhw->flags &= ~PMU_F_ENABLED;
217}
218
219
220/* Number of perf events counting hardware events */
221static atomic_t num_events = ATOMIC_INIT(0);
222/* Used to avoid races in calling reserve/release_cpumf_hardware */
223static DEFINE_MUTEX(pmc_reserve_mutex);
224
225/* CPU-measurement alerts for the counter facility */
226static void cpumf_measurement_alert(struct ext_code ext_code,
227 unsigned int alert, unsigned long unused)
228{
229 struct cpu_hw_events *cpuhw;
230
231 if (!(alert & CPU_MF_INT_CF_MASK))
232 return;
233
234 inc_irq_stat(IRQEXT_CMC);
235 cpuhw = this_cpu_ptr(&cpu_hw_events);
236
237 /* Measurement alerts are shared and might happen when the PMU
238 * is not reserved. Ignore these alerts in this case. */
239 if (!(cpuhw->flags & PMU_F_RESERVED))
240 return;
241
242 /* counter authorization change alert */
243 if (alert & CPU_MF_INT_CF_CACA)
244 qctri(&cpuhw->info);
245
246 /* loss of counter data alert */
247 if (alert & CPU_MF_INT_CF_LCDA)
248 pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
249
250 /* loss of MT counter data alert */
251 if (alert & CPU_MF_INT_CF_MTDA)
252 pr_warn("CPU[%i] MT counter data was lost\n",
253 smp_processor_id());
254}
255
256#define PMC_INIT 0
257#define PMC_RELEASE 1
258static void setup_pmc_cpu(void *flags)
259{
260 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
261
262 switch (*((int *) flags)) {
263 case PMC_INIT:
264 memset(&cpuhw->info, 0, sizeof(cpuhw->info));
265 qctri(&cpuhw->info);
266 cpuhw->flags |= PMU_F_RESERVED;
267 break;
268
269 case PMC_RELEASE:
270 cpuhw->flags &= ~PMU_F_RESERVED;
271 break;
272 }
273
274 /* Disable CPU counter sets */
275 lcctl(0);
276}
277
278/* Initialize the CPU-measurement facility */
279static int reserve_pmc_hardware(void)
280{
281 int flags = PMC_INIT;
282
283 on_each_cpu(setup_pmc_cpu, &flags, 1);
284 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
285
286 return 0;
287}
288
289/* Release the CPU-measurement facility */
290static void release_pmc_hardware(void)
291{
292 int flags = PMC_RELEASE;
293
294 on_each_cpu(setup_pmc_cpu, &flags, 1);
295 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
296}
297
298/* Release the PMU if event is the last perf event */
299static void hw_perf_event_destroy(struct perf_event *event)
300{
301 if (!atomic_add_unless(&num_events, -1, 1)) {
302 mutex_lock(&pmc_reserve_mutex);
303 if (atomic_dec_return(&num_events) == 0)
304 release_pmc_hardware();
305 mutex_unlock(&pmc_reserve_mutex);
306 }
307}
308
309/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
310static const int cpumf_generic_events_basic[] = {
311 [PERF_COUNT_HW_CPU_CYCLES] = 0,
312 [PERF_COUNT_HW_INSTRUCTIONS] = 1,
313 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
314 [PERF_COUNT_HW_CACHE_MISSES] = -1,
315 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
316 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
317 [PERF_COUNT_HW_BUS_CYCLES] = -1,
318};
319/* CPUMF <-> perf event mappings for userspace (problem-state set) */
320static const int cpumf_generic_events_user[] = {
321 [PERF_COUNT_HW_CPU_CYCLES] = 32,
322 [PERF_COUNT_HW_INSTRUCTIONS] = 33,
323 [PERF_COUNT_HW_CACHE_REFERENCES] = -1,
324 [PERF_COUNT_HW_CACHE_MISSES] = -1,
325 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
326 [PERF_COUNT_HW_BRANCH_MISSES] = -1,
327 [PERF_COUNT_HW_BUS_CYCLES] = -1,
328};
329
330static int __hw_perf_event_init(struct perf_event *event)
331{
332 struct perf_event_attr *attr = &event->attr;
333 struct hw_perf_event *hwc = &event->hw;
334 enum cpumf_ctr_set set;
335 int err;
336 u64 ev;
337
338 switch (attr->type) {
339 case PERF_TYPE_RAW:
340 /* Raw events are used to access counters directly,
341 * hence do not permit excludes */
342 if (attr->exclude_kernel || attr->exclude_user ||
343 attr->exclude_hv)
344 return -EOPNOTSUPP;
345 ev = attr->config;
346 break;
347
348 case PERF_TYPE_HARDWARE:
349 ev = attr->config;
350 /* Count user space (problem-state) only */
351 if (!attr->exclude_user && attr->exclude_kernel) {
352 if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
353 return -EOPNOTSUPP;
354 ev = cpumf_generic_events_user[ev];
355
356 /* No support for kernel space counters only */
357 } else if (!attr->exclude_kernel && attr->exclude_user) {
358 return -EOPNOTSUPP;
359
360 /* Count user and kernel space */
361 } else {
362 if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
363 return -EOPNOTSUPP;
364 ev = cpumf_generic_events_basic[ev];
365 }
366 break;
367
368 default:
369 return -ENOENT;
370 }
371
372 if (ev == -1)
373 return -ENOENT;
374
375 if (ev > PERF_CPUM_CF_MAX_CTR)
376 return -EINVAL;
377
378 /* Obtain the counter set to which the specified counter belongs */
379 set = get_counter_set(ev);
380 switch (set) {
381 case CPUMF_CTR_SET_BASIC:
382 case CPUMF_CTR_SET_USER:
383 case CPUMF_CTR_SET_CRYPTO:
384 case CPUMF_CTR_SET_EXT:
385 case CPUMF_CTR_SET_MT_DIAG:
386 /*
387 * Use the hardware perf event structure to store the
388 * counter number in the 'config' member and the counter
389 * set number in the 'config_base'. The counter set number
390 * is then later used to enable/disable the counter(s).
391 */
392 hwc->config = ev;
393 hwc->config_base = set;
394 break;
395 case CPUMF_CTR_SET_MAX:
396 /* The counter could not be associated to a counter set */
397 return -EINVAL;
398 };
399
400 /* Initialize for using the CPU-measurement counter facility */
401 if (!atomic_inc_not_zero(&num_events)) {
402 mutex_lock(&pmc_reserve_mutex);
403 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
404 err = -EBUSY;
405 else
406 atomic_inc(&num_events);
407 mutex_unlock(&pmc_reserve_mutex);
408 }
409 event->destroy = hw_perf_event_destroy;
410
411 /* Finally, validate version and authorization of the counter set */
412 err = validate_ctr_auth(hwc);
413 if (!err)
414 err = validate_ctr_version(hwc);
415
416 return err;
417}
418
419static int cpumf_pmu_event_init(struct perf_event *event)
420{
421 int err;
422
423 switch (event->attr.type) {
424 case PERF_TYPE_HARDWARE:
425 case PERF_TYPE_HW_CACHE:
426 case PERF_TYPE_RAW:
427 err = __hw_perf_event_init(event);
428 break;
429 default:
430 return -ENOENT;
431 }
432
433 if (unlikely(err) && event->destroy)
434 event->destroy(event);
435
436 return err;
437}
438
439static int hw_perf_event_reset(struct perf_event *event)
440{
441 u64 prev, new;
442 int err;
443
444 do {
445 prev = local64_read(&event->hw.prev_count);
446 err = ecctr(event->hw.config, &new);
447 if (err) {
448 if (err != 3)
449 break;
450 /* The counter is not (yet) available. This
451 * might happen if the counter set to which
452 * this counter belongs is in the disabled
453 * state.
454 */
455 new = 0;
456 }
457 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
458
459 return err;
460}
461
462static void hw_perf_event_update(struct perf_event *event)
463{
464 u64 prev, new, delta;
465 int err;
466
467 do {
468 prev = local64_read(&event->hw.prev_count);
469 err = ecctr(event->hw.config, &new);
470 if (err)
471 return;
472 } while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
473
474 delta = (prev <= new) ? new - prev
475 : (-1ULL - prev) + new + 1; /* overflow */
476 local64_add(delta, &event->count);
477}
478
479static void cpumf_pmu_read(struct perf_event *event)
480{
481 if (event->hw.state & PERF_HES_STOPPED)
482 return;
483
484 hw_perf_event_update(event);
485}
486
487static void cpumf_pmu_start(struct perf_event *event, int flags)
488{
489 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
490 struct hw_perf_event *hwc = &event->hw;
491
492 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
493 return;
494
495 if (WARN_ON_ONCE(hwc->config == -1))
496 return;
497
498 if (flags & PERF_EF_RELOAD)
499 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
500
501 hwc->state = 0;
502
503 /* (Re-)enable and activate the counter set */
504 ctr_set_enable(&cpuhw->state, hwc->config_base);
505 ctr_set_start(&cpuhw->state, hwc->config_base);
506
507 /* The counter set to which this counter belongs can be already active.
508 * Because all counters in a set are active, the event->hw.prev_count
509 * needs to be synchronized. At this point, the counter set can be in
510 * the inactive or disabled state.
511 */
512 hw_perf_event_reset(event);
513
514 /* increment refcount for this counter set */
515 atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
516}
517
518static void cpumf_pmu_stop(struct perf_event *event, int flags)
519{
520 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
521 struct hw_perf_event *hwc = &event->hw;
522
523 if (!(hwc->state & PERF_HES_STOPPED)) {
524 /* Decrement reference count for this counter set and if this
525 * is the last used counter in the set, clear activation
526 * control and set the counter set state to inactive.
527 */
528 if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
529 ctr_set_stop(&cpuhw->state, hwc->config_base);
530 event->hw.state |= PERF_HES_STOPPED;
531 }
532
533 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
534 hw_perf_event_update(event);
535 event->hw.state |= PERF_HES_UPTODATE;
536 }
537}
538
539static int cpumf_pmu_add(struct perf_event *event, int flags)
540{
541 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
542
543 /* Check authorization for the counter set to which this
544 * counter belongs.
545 * For group events transaction, the authorization check is
546 * done in cpumf_pmu_commit_txn().
547 */
548 if (!(cpuhw->txn_flags & PERF_PMU_TXN_ADD))
549 if (validate_ctr_auth(&event->hw))
550 return -ENOENT;
551
552 ctr_set_enable(&cpuhw->state, event->hw.config_base);
553 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
554
555 if (flags & PERF_EF_START)
556 cpumf_pmu_start(event, PERF_EF_RELOAD);
557
558 perf_event_update_userpage(event);
559
560 return 0;
561}
562
563static void cpumf_pmu_del(struct perf_event *event, int flags)
564{
565 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
566
567 cpumf_pmu_stop(event, PERF_EF_UPDATE);
568
569 /* Check if any counter in the counter set is still used. If not used,
570 * change the counter set to the disabled state. This also clears the
571 * content of all counters in the set.
572 *
573 * When a new perf event has been added but not yet started, this can
574 * clear enable control and resets all counters in a set. Therefore,
575 * cpumf_pmu_start() always has to reenable a counter set.
576 */
577 if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
578 ctr_set_disable(&cpuhw->state, event->hw.config_base);
579
580 perf_event_update_userpage(event);
581}
582
583/*
584 * Start group events scheduling transaction.
585 * Set flags to perform a single test at commit time.
586 *
587 * We only support PERF_PMU_TXN_ADD transactions. Save the
588 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
589 * transactions.
590 */
591static void cpumf_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
592{
593 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
594
595 WARN_ON_ONCE(cpuhw->txn_flags); /* txn already in flight */
596
597 cpuhw->txn_flags = txn_flags;
598 if (txn_flags & ~PERF_PMU_TXN_ADD)
599 return;
600
601 perf_pmu_disable(pmu);
602 cpuhw->tx_state = cpuhw->state;
603}
604
605/*
606 * Stop and cancel a group events scheduling tranctions.
607 * Assumes cpumf_pmu_del() is called for each successful added
608 * cpumf_pmu_add() during the transaction.
609 */
610static void cpumf_pmu_cancel_txn(struct pmu *pmu)
611{
612 unsigned int txn_flags;
613 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
614
615 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
616
617 txn_flags = cpuhw->txn_flags;
618 cpuhw->txn_flags = 0;
619 if (txn_flags & ~PERF_PMU_TXN_ADD)
620 return;
621
622 WARN_ON(cpuhw->tx_state != cpuhw->state);
623
624 perf_pmu_enable(pmu);
625}
626
627/*
628 * Commit the group events scheduling transaction. On success, the
629 * transaction is closed. On error, the transaction is kept open
630 * until cpumf_pmu_cancel_txn() is called.
631 */
632static int cpumf_pmu_commit_txn(struct pmu *pmu)
633{
634 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
635 u64 state;
636
637 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
638
639 if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
640 cpuhw->txn_flags = 0;
641 return 0;
642 }
643
644 /* check if the updated state can be scheduled */
645 state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
646 state >>= CPUMF_LCCTL_ENABLE_SHIFT;
647 if ((state & cpuhw->info.auth_ctl) != state)
648 return -ENOENT;
649
650 cpuhw->txn_flags = 0;
651 perf_pmu_enable(pmu);
652 return 0;
653}
654
655/* Performance monitoring unit for s390x */
656static struct pmu cpumf_pmu = {
657 .task_ctx_nr = perf_sw_context,
658 .capabilities = PERF_PMU_CAP_NO_INTERRUPT,
659 .pmu_enable = cpumf_pmu_enable,
660 .pmu_disable = cpumf_pmu_disable,
661 .event_init = cpumf_pmu_event_init,
662 .add = cpumf_pmu_add,
663 .del = cpumf_pmu_del,
664 .start = cpumf_pmu_start,
665 .stop = cpumf_pmu_stop,
666 .read = cpumf_pmu_read,
667 .start_txn = cpumf_pmu_start_txn,
668 .commit_txn = cpumf_pmu_commit_txn,
669 .cancel_txn = cpumf_pmu_cancel_txn,
670};
671
672static int cpumf_pmf_setup(unsigned int cpu, int flags)
673{
674 local_irq_disable();
675 setup_pmc_cpu(&flags);
676 local_irq_enable();
677 return 0;
678}
679
680static int s390_pmu_online_cpu(unsigned int cpu)
681{
682 return cpumf_pmf_setup(cpu, PMC_INIT);
683}
684
685static int s390_pmu_offline_cpu(unsigned int cpu)
686{
687 return cpumf_pmf_setup(cpu, PMC_RELEASE);
688}
689
690static int __init cpumf_pmu_init(void)
691{
692 int rc;
693
694 if (!cpum_cf_avail())
695 return -ENODEV;
696
697 /* clear bit 15 of cr0 to unauthorize problem-state to
698 * extract measurement counters */
699 ctl_clear_bit(0, 48);
700
701 /* register handler for measurement-alert interruptions */
702 rc = register_external_irq(EXT_IRQ_MEASURE_ALERT,
703 cpumf_measurement_alert);
704 if (rc) {
705 pr_err("Registering for CPU-measurement alerts "
706 "failed with rc=%i\n", rc);
707 return rc;
708 }
709
710 cpumf_pmu.attr_groups = cpumf_cf_event_group();
711 rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
712 if (rc) {
713 pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
714 unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
715 cpumf_measurement_alert);
716 return rc;
717 }
718 return cpuhp_setup_state(CPUHP_AP_PERF_S390_CF_ONLINE,
719 "perf/s390/cf:online",
720 s390_pmu_online_cpu, s390_pmu_offline_cpu);
721}
722early_initcall(cpumf_pmu_init);