Linux 4.13.16
[linux/fpc-iii.git] / arch / s390 / kernel / perf_cpum_sf.c
blob7e1e40323b78e1bb7a3910e1191dfbe40027b748
1 /*
2 * Performance event support for the System z CPU-measurement Sampling Facility
4 * Copyright IBM Corp. 2013
5 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
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.
11 #define KMSG_COMPONENT "cpum_sf"
12 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
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/export.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/moduleparam.h>
23 #include <asm/cpu_mf.h>
24 #include <asm/irq.h>
25 #include <asm/debug.h>
26 #include <asm/timex.h>
28 /* Minimum number of sample-data-block-tables:
29 * At least one table is required for the sampling buffer structure.
30 * A single table contains up to 511 pointers to sample-data-blocks.
32 #define CPUM_SF_MIN_SDBT 1
34 /* Number of sample-data-blocks per sample-data-block-table (SDBT):
35 * A table contains SDB pointers (8 bytes) and one table-link entry
36 * that points to the origin of the next SDBT.
38 #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8)
40 /* Maximum page offset for an SDBT table-link entry:
41 * If this page offset is reached, a table-link entry to the next SDBT
42 * must be added.
44 #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8)
45 static inline int require_table_link(const void *sdbt)
47 return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
50 /* Minimum and maximum sampling buffer sizes:
52 * This number represents the maximum size of the sampling buffer taking
53 * the number of sample-data-block-tables into account. Note that these
54 * numbers apply to the basic-sampling function only.
55 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if
56 * the diagnostic-sampling function is active.
58 * Sampling buffer size Buffer characteristics
59 * ---------------------------------------------------
60 * 64KB == 16 pages (4KB per page)
61 * 1 page for SDB-tables
62 * 15 pages for SDBs
64 * 32MB == 8192 pages (4KB per page)
65 * 16 pages for SDB-tables
66 * 8176 pages for SDBs
68 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15;
69 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176;
70 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
72 struct sf_buffer {
73 unsigned long *sdbt; /* Sample-data-block-table origin */
74 /* buffer characteristics (required for buffer increments) */
75 unsigned long num_sdb; /* Number of sample-data-blocks */
76 unsigned long num_sdbt; /* Number of sample-data-block-tables */
77 unsigned long *tail; /* last sample-data-block-table */
80 struct cpu_hw_sf {
81 /* CPU-measurement sampling information block */
82 struct hws_qsi_info_block qsi;
83 /* CPU-measurement sampling control block */
84 struct hws_lsctl_request_block lsctl;
85 struct sf_buffer sfb; /* Sampling buffer */
86 unsigned int flags; /* Status flags */
87 struct perf_event *event; /* Scheduled perf event */
89 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
91 /* Debug feature */
92 static debug_info_t *sfdbg;
95 * sf_disable() - Switch off sampling facility
97 static int sf_disable(void)
99 struct hws_lsctl_request_block sreq;
101 memset(&sreq, 0, sizeof(sreq));
102 return lsctl(&sreq);
106 * sf_buffer_available() - Check for an allocated sampling buffer
108 static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
110 return !!cpuhw->sfb.sdbt;
114 * deallocate sampling facility buffer
116 static void free_sampling_buffer(struct sf_buffer *sfb)
118 unsigned long *sdbt, *curr;
120 if (!sfb->sdbt)
121 return;
123 sdbt = sfb->sdbt;
124 curr = sdbt;
126 /* Free the SDBT after all SDBs are processed... */
127 while (1) {
128 if (!*curr || !sdbt)
129 break;
131 /* Process table-link entries */
132 if (is_link_entry(curr)) {
133 curr = get_next_sdbt(curr);
134 if (sdbt)
135 free_page((unsigned long) sdbt);
137 /* If the origin is reached, sampling buffer is freed */
138 if (curr == sfb->sdbt)
139 break;
140 else
141 sdbt = curr;
142 } else {
143 /* Process SDB pointer */
144 if (*curr) {
145 free_page(*curr);
146 curr++;
151 debug_sprintf_event(sfdbg, 5,
152 "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
153 memset(sfb, 0, sizeof(*sfb));
156 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
158 unsigned long sdb, *trailer;
160 /* Allocate and initialize sample-data-block */
161 sdb = get_zeroed_page(gfp_flags);
162 if (!sdb)
163 return -ENOMEM;
164 trailer = trailer_entry_ptr(sdb);
165 *trailer = SDB_TE_ALERT_REQ_MASK;
167 /* Link SDB into the sample-data-block-table */
168 *sdbt = sdb;
170 return 0;
174 * realloc_sampling_buffer() - extend sampler memory
176 * Allocates new sample-data-blocks and adds them to the specified sampling
177 * buffer memory.
179 * Important: This modifies the sampling buffer and must be called when the
180 * sampling facility is disabled.
182 * Returns zero on success, non-zero otherwise.
184 static int realloc_sampling_buffer(struct sf_buffer *sfb,
185 unsigned long num_sdb, gfp_t gfp_flags)
187 int i, rc;
188 unsigned long *new, *tail;
190 if (!sfb->sdbt || !sfb->tail)
191 return -EINVAL;
193 if (!is_link_entry(sfb->tail))
194 return -EINVAL;
196 /* Append to the existing sampling buffer, overwriting the table-link
197 * register.
198 * The tail variables always points to the "tail" (last and table-link)
199 * entry in an SDB-table.
201 tail = sfb->tail;
203 /* Do a sanity check whether the table-link entry points to
204 * the sampling buffer origin.
206 if (sfb->sdbt != get_next_sdbt(tail)) {
207 debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
208 "sampling buffer is not linked: origin=%p"
209 "tail=%p\n",
210 (void *) sfb->sdbt, (void *) tail);
211 return -EINVAL;
214 /* Allocate remaining SDBs */
215 rc = 0;
216 for (i = 0; i < num_sdb; i++) {
217 /* Allocate a new SDB-table if it is full. */
218 if (require_table_link(tail)) {
219 new = (unsigned long *) get_zeroed_page(gfp_flags);
220 if (!new) {
221 rc = -ENOMEM;
222 break;
224 sfb->num_sdbt++;
225 /* Link current page to tail of chain */
226 *tail = (unsigned long)(void *) new + 1;
227 tail = new;
230 /* Allocate a new sample-data-block.
231 * If there is not enough memory, stop the realloc process
232 * and simply use what was allocated. If this is a temporary
233 * issue, a new realloc call (if required) might succeed.
235 rc = alloc_sample_data_block(tail, gfp_flags);
236 if (rc)
237 break;
238 sfb->num_sdb++;
239 tail++;
242 /* Link sampling buffer to its origin */
243 *tail = (unsigned long) sfb->sdbt + 1;
244 sfb->tail = tail;
246 debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
247 " settings: sdbt=%lu sdb=%lu\n",
248 sfb->num_sdbt, sfb->num_sdb);
249 return rc;
253 * allocate_sampling_buffer() - allocate sampler memory
255 * Allocates and initializes a sampling buffer structure using the
256 * specified number of sample-data-blocks (SDB). For each allocation,
257 * a 4K page is used. The number of sample-data-block-tables (SDBT)
258 * are calculated from SDBs.
259 * Also set the ALERT_REQ mask in each SDBs trailer.
261 * Returns zero on success, non-zero otherwise.
263 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
265 int rc;
267 if (sfb->sdbt)
268 return -EINVAL;
270 /* Allocate the sample-data-block-table origin */
271 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
272 if (!sfb->sdbt)
273 return -ENOMEM;
274 sfb->num_sdb = 0;
275 sfb->num_sdbt = 1;
277 /* Link the table origin to point to itself to prepare for
278 * realloc_sampling_buffer() invocation.
280 sfb->tail = sfb->sdbt;
281 *sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
283 /* Allocate requested number of sample-data-blocks */
284 rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
285 if (rc) {
286 free_sampling_buffer(sfb);
287 debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
288 "realloc_sampling_buffer failed with rc=%i\n", rc);
289 } else
290 debug_sprintf_event(sfdbg, 4,
291 "alloc_sampling_buffer: tear=%p dear=%p\n",
292 sfb->sdbt, (void *) *sfb->sdbt);
293 return rc;
296 static void sfb_set_limits(unsigned long min, unsigned long max)
298 struct hws_qsi_info_block si;
300 CPUM_SF_MIN_SDB = min;
301 CPUM_SF_MAX_SDB = max;
303 memset(&si, 0, sizeof(si));
304 if (!qsi(&si))
305 CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
308 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
310 return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
311 : CPUM_SF_MAX_SDB;
314 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
315 struct hw_perf_event *hwc)
317 if (!sfb->sdbt)
318 return SFB_ALLOC_REG(hwc);
319 if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
320 return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
321 return 0;
324 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
325 struct hw_perf_event *hwc)
327 return sfb_pending_allocs(sfb, hwc) > 0;
330 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
332 /* Limit the number of SDBs to not exceed the maximum */
333 num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
334 if (num)
335 SFB_ALLOC_REG(hwc) += num;
338 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
340 SFB_ALLOC_REG(hwc) = 0;
341 sfb_account_allocs(num, hwc);
344 static size_t event_sample_size(struct hw_perf_event *hwc)
346 struct sf_raw_sample *sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(hwc);
347 size_t sample_size;
349 /* The sample size depends on the sampling function: The basic-sampling
350 * function must be always enabled, diagnostic-sampling function is
351 * optional.
353 sample_size = sfr->bsdes;
354 if (SAMPL_DIAG_MODE(hwc))
355 sample_size += sfr->dsdes;
357 return sample_size;
360 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
362 if (cpuhw->sfb.sdbt)
363 free_sampling_buffer(&cpuhw->sfb);
366 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
368 unsigned long n_sdb, freq, factor;
369 size_t sfr_size, sample_size;
370 struct sf_raw_sample *sfr;
372 /* Allocate raw sample buffer
374 * The raw sample buffer is used to temporarily store sampling data
375 * entries for perf raw sample processing. The buffer size mainly
376 * depends on the size of diagnostic-sampling data entries which is
377 * machine-specific. The exact size calculation includes:
378 * 1. The first 4 bytes of diagnostic-sampling data entries are
379 * already reflected in the sf_raw_sample structure. Subtract
380 * these bytes.
381 * 2. The perf raw sample data must be 8-byte aligned (u64) and
382 * perf's internal data size must be considered too. So add
383 * an additional u32 for correct alignment and subtract before
384 * allocating the buffer.
385 * 3. Store the raw sample buffer pointer in the perf event
386 * hardware structure.
388 sfr_size = ALIGN((sizeof(*sfr) - sizeof(sfr->diag) + cpuhw->qsi.dsdes) +
389 sizeof(u32), sizeof(u64));
390 sfr_size -= sizeof(u32);
391 sfr = kzalloc(sfr_size, GFP_KERNEL);
392 if (!sfr)
393 return -ENOMEM;
394 sfr->size = sfr_size;
395 sfr->bsdes = cpuhw->qsi.bsdes;
396 sfr->dsdes = cpuhw->qsi.dsdes;
397 RAWSAMPLE_REG(hwc) = (unsigned long) sfr;
399 /* Calculate sampling buffers using 4K pages
401 * 1. Determine the sample data size which depends on the used
402 * sampling functions, for example, basic-sampling or
403 * basic-sampling with diagnostic-sampling.
405 * 2. Use the sampling frequency as input. The sampling buffer is
406 * designed for almost one second. This can be adjusted through
407 * the "factor" variable.
408 * In any case, alloc_sampling_buffer() sets the Alert Request
409 * Control indicator to trigger a measurement-alert to harvest
410 * sample-data-blocks (sdb).
412 * 3. Compute the number of sample-data-blocks and ensure a minimum
413 * of CPUM_SF_MIN_SDB. Also ensure the upper limit does not
414 * exceed a "calculated" maximum. The symbolic maximum is
415 * designed for basic-sampling only and needs to be increased if
416 * diagnostic-sampling is active.
417 * See also the remarks for these symbolic constants.
419 * 4. Compute the number of sample-data-block-tables (SDBT) and
420 * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
421 * to 511 SDBs).
423 sample_size = event_sample_size(hwc);
424 freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
425 factor = 1;
426 n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
427 if (n_sdb < CPUM_SF_MIN_SDB)
428 n_sdb = CPUM_SF_MIN_SDB;
430 /* If there is already a sampling buffer allocated, it is very likely
431 * that the sampling facility is enabled too. If the event to be
432 * initialized requires a greater sampling buffer, the allocation must
433 * be postponed. Changing the sampling buffer requires the sampling
434 * facility to be in the disabled state. So, account the number of
435 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
436 * before the event is started.
438 sfb_init_allocs(n_sdb, hwc);
439 if (sf_buffer_available(cpuhw))
440 return 0;
442 debug_sprintf_event(sfdbg, 3,
443 "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
444 " sample_size=%lu cpuhw=%p\n",
445 SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
446 sample_size, cpuhw);
448 return alloc_sampling_buffer(&cpuhw->sfb,
449 sfb_pending_allocs(&cpuhw->sfb, hwc));
452 static unsigned long min_percent(unsigned int percent, unsigned long base,
453 unsigned long min)
455 return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
458 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
460 /* Use a percentage-based approach to extend the sampling facility
461 * buffer. Accept up to 5% sample data loss.
462 * Vary the extents between 1% to 5% of the current number of
463 * sample-data-blocks.
465 if (ratio <= 5)
466 return 0;
467 if (ratio <= 25)
468 return min_percent(1, base, 1);
469 if (ratio <= 50)
470 return min_percent(1, base, 1);
471 if (ratio <= 75)
472 return min_percent(2, base, 2);
473 if (ratio <= 100)
474 return min_percent(3, base, 3);
475 if (ratio <= 250)
476 return min_percent(4, base, 4);
478 return min_percent(5, base, 8);
481 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
482 struct hw_perf_event *hwc)
484 unsigned long ratio, num;
486 if (!OVERFLOW_REG(hwc))
487 return;
489 /* The sample_overflow contains the average number of sample data
490 * that has been lost because sample-data-blocks were full.
492 * Calculate the total number of sample data entries that has been
493 * discarded. Then calculate the ratio of lost samples to total samples
494 * per second in percent.
496 ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
497 sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
499 /* Compute number of sample-data-blocks */
500 num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
501 if (num)
502 sfb_account_allocs(num, hwc);
504 debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
505 " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
506 OVERFLOW_REG(hwc) = 0;
509 /* extend_sampling_buffer() - Extend sampling buffer
510 * @sfb: Sampling buffer structure (for local CPU)
511 * @hwc: Perf event hardware structure
513 * Use this function to extend the sampling buffer based on the overflow counter
514 * and postponed allocation extents stored in the specified Perf event hardware.
516 * Important: This function disables the sampling facility in order to safely
517 * change the sampling buffer structure. Do not call this function
518 * when the PMU is active.
520 static void extend_sampling_buffer(struct sf_buffer *sfb,
521 struct hw_perf_event *hwc)
523 unsigned long num, num_old;
524 int rc;
526 num = sfb_pending_allocs(sfb, hwc);
527 if (!num)
528 return;
529 num_old = sfb->num_sdb;
531 /* Disable the sampling facility to reset any states and also
532 * clear pending measurement alerts.
534 sf_disable();
536 /* Extend the sampling buffer.
537 * This memory allocation typically happens in an atomic context when
538 * called by perf. Because this is a reallocation, it is fine if the
539 * new SDB-request cannot be satisfied immediately.
541 rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
542 if (rc)
543 debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
544 "failed with rc=%i\n", rc);
546 if (sfb_has_pending_allocs(sfb, hwc))
547 debug_sprintf_event(sfdbg, 5, "sfb: extend: "
548 "req=%lu alloc=%lu remaining=%lu\n",
549 num, sfb->num_sdb - num_old,
550 sfb_pending_allocs(sfb, hwc));
554 /* Number of perf events counting hardware events */
555 static atomic_t num_events;
556 /* Used to avoid races in calling reserve/release_cpumf_hardware */
557 static DEFINE_MUTEX(pmc_reserve_mutex);
559 #define PMC_INIT 0
560 #define PMC_RELEASE 1
561 #define PMC_FAILURE 2
562 static void setup_pmc_cpu(void *flags)
564 int err;
565 struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
567 err = 0;
568 switch (*((int *) flags)) {
569 case PMC_INIT:
570 memset(cpusf, 0, sizeof(*cpusf));
571 err = qsi(&cpusf->qsi);
572 if (err)
573 break;
574 cpusf->flags |= PMU_F_RESERVED;
575 err = sf_disable();
576 if (err)
577 pr_err("Switching off the sampling facility failed "
578 "with rc=%i\n", err);
579 debug_sprintf_event(sfdbg, 5,
580 "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
581 break;
582 case PMC_RELEASE:
583 cpusf->flags &= ~PMU_F_RESERVED;
584 err = sf_disable();
585 if (err) {
586 pr_err("Switching off the sampling facility failed "
587 "with rc=%i\n", err);
588 } else
589 deallocate_buffers(cpusf);
590 debug_sprintf_event(sfdbg, 5,
591 "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
592 break;
594 if (err)
595 *((int *) flags) |= PMC_FAILURE;
598 static void release_pmc_hardware(void)
600 int flags = PMC_RELEASE;
602 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
603 on_each_cpu(setup_pmc_cpu, &flags, 1);
606 static int reserve_pmc_hardware(void)
608 int flags = PMC_INIT;
610 on_each_cpu(setup_pmc_cpu, &flags, 1);
611 if (flags & PMC_FAILURE) {
612 release_pmc_hardware();
613 return -ENODEV;
615 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
617 return 0;
620 static void hw_perf_event_destroy(struct perf_event *event)
622 /* Free raw sample buffer */
623 if (RAWSAMPLE_REG(&event->hw))
624 kfree((void *) RAWSAMPLE_REG(&event->hw));
626 /* Release PMC if this is the last perf event */
627 if (!atomic_add_unless(&num_events, -1, 1)) {
628 mutex_lock(&pmc_reserve_mutex);
629 if (atomic_dec_return(&num_events) == 0)
630 release_pmc_hardware();
631 mutex_unlock(&pmc_reserve_mutex);
635 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
637 hwc->sample_period = period;
638 hwc->last_period = hwc->sample_period;
639 local64_set(&hwc->period_left, hwc->sample_period);
642 static void hw_reset_registers(struct hw_perf_event *hwc,
643 unsigned long *sdbt_origin)
645 struct sf_raw_sample *sfr;
647 /* (Re)set to first sample-data-block-table */
648 TEAR_REG(hwc) = (unsigned long) sdbt_origin;
650 /* (Re)set raw sampling buffer register */
651 sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(hwc);
652 memset(&sfr->basic, 0, sizeof(sfr->basic));
653 memset(&sfr->diag, 0, sfr->dsdes);
656 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
657 unsigned long rate)
659 return clamp_t(unsigned long, rate,
660 si->min_sampl_rate, si->max_sampl_rate);
663 static int __hw_perf_event_init(struct perf_event *event)
665 struct cpu_hw_sf *cpuhw;
666 struct hws_qsi_info_block si;
667 struct perf_event_attr *attr = &event->attr;
668 struct hw_perf_event *hwc = &event->hw;
669 unsigned long rate;
670 int cpu, err;
672 /* Reserve CPU-measurement sampling facility */
673 err = 0;
674 if (!atomic_inc_not_zero(&num_events)) {
675 mutex_lock(&pmc_reserve_mutex);
676 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
677 err = -EBUSY;
678 else
679 atomic_inc(&num_events);
680 mutex_unlock(&pmc_reserve_mutex);
682 event->destroy = hw_perf_event_destroy;
684 if (err)
685 goto out;
687 /* Access per-CPU sampling information (query sampling info) */
689 * The event->cpu value can be -1 to count on every CPU, for example,
690 * when attaching to a task. If this is specified, use the query
691 * sampling info from the current CPU, otherwise use event->cpu to
692 * retrieve the per-CPU information.
693 * Later, cpuhw indicates whether to allocate sampling buffers for a
694 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
696 memset(&si, 0, sizeof(si));
697 cpuhw = NULL;
698 if (event->cpu == -1)
699 qsi(&si);
700 else {
701 /* Event is pinned to a particular CPU, retrieve the per-CPU
702 * sampling structure for accessing the CPU-specific QSI.
704 cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
705 si = cpuhw->qsi;
708 /* Check sampling facility authorization and, if not authorized,
709 * fall back to other PMUs. It is safe to check any CPU because
710 * the authorization is identical for all configured CPUs.
712 if (!si.as) {
713 err = -ENOENT;
714 goto out;
717 /* Always enable basic sampling */
718 SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
720 /* Check if diagnostic sampling is requested. Deny if the required
721 * sampling authorization is missing.
723 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
724 if (!si.ad) {
725 err = -EPERM;
726 goto out;
728 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
731 /* Check and set other sampling flags */
732 if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
733 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
735 /* The sampling information (si) contains information about the
736 * min/max sampling intervals and the CPU speed. So calculate the
737 * correct sampling interval and avoid the whole period adjust
738 * feedback loop.
740 rate = 0;
741 if (attr->freq) {
742 rate = freq_to_sample_rate(&si, attr->sample_freq);
743 rate = hw_limit_rate(&si, rate);
744 attr->freq = 0;
745 attr->sample_period = rate;
746 } else {
747 /* The min/max sampling rates specifies the valid range
748 * of sample periods. If the specified sample period is
749 * out of range, limit the period to the range boundary.
751 rate = hw_limit_rate(&si, hwc->sample_period);
753 /* The perf core maintains a maximum sample rate that is
754 * configurable through the sysctl interface. Ensure the
755 * sampling rate does not exceed this value. This also helps
756 * to avoid throttling when pushing samples with
757 * perf_event_overflow().
759 if (sample_rate_to_freq(&si, rate) >
760 sysctl_perf_event_sample_rate) {
761 err = -EINVAL;
762 debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
763 goto out;
766 SAMPL_RATE(hwc) = rate;
767 hw_init_period(hwc, SAMPL_RATE(hwc));
769 /* Initialize sample data overflow accounting */
770 hwc->extra_reg.reg = REG_OVERFLOW;
771 OVERFLOW_REG(hwc) = 0;
773 /* Allocate the per-CPU sampling buffer using the CPU information
774 * from the event. If the event is not pinned to a particular
775 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
776 * buffers for each online CPU.
778 if (cpuhw)
779 /* Event is pinned to a particular CPU */
780 err = allocate_buffers(cpuhw, hwc);
781 else {
782 /* Event is not pinned, allocate sampling buffer on
783 * each online CPU
785 for_each_online_cpu(cpu) {
786 cpuhw = &per_cpu(cpu_hw_sf, cpu);
787 err = allocate_buffers(cpuhw, hwc);
788 if (err)
789 break;
792 out:
793 return err;
796 static int cpumsf_pmu_event_init(struct perf_event *event)
798 int err;
800 /* No support for taken branch sampling */
801 if (has_branch_stack(event))
802 return -EOPNOTSUPP;
804 switch (event->attr.type) {
805 case PERF_TYPE_RAW:
806 if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
807 (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
808 return -ENOENT;
809 break;
810 case PERF_TYPE_HARDWARE:
811 /* Support sampling of CPU cycles in addition to the
812 * counter facility. However, the counter facility
813 * is more precise and, hence, restrict this PMU to
814 * sampling events only.
816 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
817 return -ENOENT;
818 if (!is_sampling_event(event))
819 return -ENOENT;
820 break;
821 default:
822 return -ENOENT;
825 /* Check online status of the CPU to which the event is pinned */
826 if (event->cpu >= 0) {
827 if ((unsigned int)event->cpu >= nr_cpumask_bits)
828 return -ENODEV;
829 if (!cpu_online(event->cpu))
830 return -ENODEV;
833 /* Force reset of idle/hv excludes regardless of what the
834 * user requested.
836 if (event->attr.exclude_hv)
837 event->attr.exclude_hv = 0;
838 if (event->attr.exclude_idle)
839 event->attr.exclude_idle = 0;
841 err = __hw_perf_event_init(event);
842 if (unlikely(err))
843 if (event->destroy)
844 event->destroy(event);
845 return err;
848 static void cpumsf_pmu_enable(struct pmu *pmu)
850 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
851 struct hw_perf_event *hwc;
852 int err;
854 if (cpuhw->flags & PMU_F_ENABLED)
855 return;
857 if (cpuhw->flags & PMU_F_ERR_MASK)
858 return;
860 /* Check whether to extent the sampling buffer.
862 * Two conditions trigger an increase of the sampling buffer for a
863 * perf event:
864 * 1. Postponed buffer allocations from the event initialization.
865 * 2. Sampling overflows that contribute to pending allocations.
867 * Note that the extend_sampling_buffer() function disables the sampling
868 * facility, but it can be fully re-enabled using sampling controls that
869 * have been saved in cpumsf_pmu_disable().
871 if (cpuhw->event) {
872 hwc = &cpuhw->event->hw;
873 /* Account number of overflow-designated buffer extents */
874 sfb_account_overflows(cpuhw, hwc);
875 if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
876 extend_sampling_buffer(&cpuhw->sfb, hwc);
879 /* (Re)enable the PMU and sampling facility */
880 cpuhw->flags |= PMU_F_ENABLED;
881 barrier();
883 err = lsctl(&cpuhw->lsctl);
884 if (err) {
885 cpuhw->flags &= ~PMU_F_ENABLED;
886 pr_err("Loading sampling controls failed: op=%i err=%i\n",
887 1, err);
888 return;
891 debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
892 "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
893 cpuhw->lsctl.ed, cpuhw->lsctl.cd,
894 (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
897 static void cpumsf_pmu_disable(struct pmu *pmu)
899 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
900 struct hws_lsctl_request_block inactive;
901 struct hws_qsi_info_block si;
902 int err;
904 if (!(cpuhw->flags & PMU_F_ENABLED))
905 return;
907 if (cpuhw->flags & PMU_F_ERR_MASK)
908 return;
910 /* Switch off sampling activation control */
911 inactive = cpuhw->lsctl;
912 inactive.cs = 0;
913 inactive.cd = 0;
915 err = lsctl(&inactive);
916 if (err) {
917 pr_err("Loading sampling controls failed: op=%i err=%i\n",
918 2, err);
919 return;
922 /* Save state of TEAR and DEAR register contents */
923 if (!qsi(&si)) {
924 /* TEAR/DEAR values are valid only if the sampling facility is
925 * enabled. Note that cpumsf_pmu_disable() might be called even
926 * for a disabled sampling facility because cpumsf_pmu_enable()
927 * controls the enable/disable state.
929 if (si.es) {
930 cpuhw->lsctl.tear = si.tear;
931 cpuhw->lsctl.dear = si.dear;
933 } else
934 debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
935 "qsi() failed with err=%i\n", err);
937 cpuhw->flags &= ~PMU_F_ENABLED;
940 /* perf_exclude_event() - Filter event
941 * @event: The perf event
942 * @regs: pt_regs structure
943 * @sde_regs: Sample-data-entry (sde) regs structure
945 * Filter perf events according to their exclude specification.
947 * Return non-zero if the event shall be excluded.
949 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
950 struct perf_sf_sde_regs *sde_regs)
952 if (event->attr.exclude_user && user_mode(regs))
953 return 1;
954 if (event->attr.exclude_kernel && !user_mode(regs))
955 return 1;
956 if (event->attr.exclude_guest && sde_regs->in_guest)
957 return 1;
958 if (event->attr.exclude_host && !sde_regs->in_guest)
959 return 1;
960 return 0;
963 /* perf_push_sample() - Push samples to perf
964 * @event: The perf event
965 * @sample: Hardware sample data
967 * Use the hardware sample data to create perf event sample. The sample
968 * is the pushed to the event subsystem and the function checks for
969 * possible event overflows. If an event overflow occurs, the PMU is
970 * stopped.
972 * Return non-zero if an event overflow occurred.
974 static int perf_push_sample(struct perf_event *event, struct sf_raw_sample *sfr)
976 int overflow;
977 struct pt_regs regs;
978 struct perf_sf_sde_regs *sde_regs;
979 struct perf_sample_data data;
980 struct perf_raw_record raw = {
981 .frag = {
982 .size = sfr->size,
983 .data = sfr,
987 /* Setup perf sample */
988 perf_sample_data_init(&data, 0, event->hw.last_period);
989 data.raw = &raw;
991 /* Setup pt_regs to look like an CPU-measurement external interrupt
992 * using the Program Request Alert code. The regs.int_parm_long
993 * field which is unused contains additional sample-data-entry related
994 * indicators.
996 memset(&regs, 0, sizeof(regs));
997 regs.int_code = 0x1407;
998 regs.int_parm = CPU_MF_INT_SF_PRA;
999 sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1001 psw_bits(regs.psw).ia = sfr->basic.ia;
1002 psw_bits(regs.psw).dat = sfr->basic.T;
1003 psw_bits(regs.psw).wait = sfr->basic.W;
1004 psw_bits(regs.psw).pstate = sfr->basic.P;
1005 psw_bits(regs.psw).as = sfr->basic.AS;
1008 * Use the hardware provided configuration level to decide if the
1009 * sample belongs to a guest or host. If that is not available,
1010 * fall back to the following heuristics:
1011 * A non-zero guest program parameter always indicates a guest
1012 * sample. Some early samples or samples from guests without
1013 * lpp usage would be misaccounted to the host. We use the asn
1014 * value as an addon heuristic to detect most of these guest samples.
1015 * If the value differs from 0xffff (the host value), we assume to
1016 * be a KVM guest.
1018 switch (sfr->basic.CL) {
1019 case 1: /* logical partition */
1020 sde_regs->in_guest = 0;
1021 break;
1022 case 2: /* virtual machine */
1023 sde_regs->in_guest = 1;
1024 break;
1025 default: /* old machine, use heuristics */
1026 if (sfr->basic.gpp || sfr->basic.prim_asn != 0xffff)
1027 sde_regs->in_guest = 1;
1028 break;
1031 overflow = 0;
1032 if (perf_exclude_event(event, &regs, sde_regs))
1033 goto out;
1034 if (perf_event_overflow(event, &data, &regs)) {
1035 overflow = 1;
1036 event->pmu->stop(event, 0);
1038 perf_event_update_userpage(event);
1039 out:
1040 return overflow;
1043 static void perf_event_count_update(struct perf_event *event, u64 count)
1045 local64_add(count, &event->count);
1048 static int sample_format_is_valid(struct hws_combined_entry *sample,
1049 unsigned int flags)
1051 if (likely(flags & PERF_CPUM_SF_BASIC_MODE))
1052 /* Only basic-sampling data entries with data-entry-format
1053 * version of 0x0001 can be processed.
1055 if (sample->basic.def != 0x0001)
1056 return 0;
1057 if (flags & PERF_CPUM_SF_DIAG_MODE)
1058 /* The data-entry-format number of diagnostic-sampling data
1059 * entries can vary. Because diagnostic data is just passed
1060 * through, do only a sanity check on the DEF.
1062 if (sample->diag.def < 0x8001)
1063 return 0;
1064 return 1;
1067 static int sample_is_consistent(struct hws_combined_entry *sample,
1068 unsigned long flags)
1070 /* This check applies only to basic-sampling data entries of potentially
1071 * combined-sampling data entries. Invalid entries cannot be processed
1072 * by the PMU and, thus, do not deliver an associated
1073 * diagnostic-sampling data entry.
1075 if (unlikely(!(flags & PERF_CPUM_SF_BASIC_MODE)))
1076 return 0;
1078 * Samples are skipped, if they are invalid or for which the
1079 * instruction address is not predictable, i.e., the wait-state bit is
1080 * set.
1082 if (sample->basic.I || sample->basic.W)
1083 return 0;
1084 return 1;
1087 static void reset_sample_slot(struct hws_combined_entry *sample,
1088 unsigned long flags)
1090 if (likely(flags & PERF_CPUM_SF_BASIC_MODE))
1091 sample->basic.def = 0;
1092 if (flags & PERF_CPUM_SF_DIAG_MODE)
1093 sample->diag.def = 0;
1096 static void sfr_store_sample(struct sf_raw_sample *sfr,
1097 struct hws_combined_entry *sample)
1099 if (likely(sfr->format & PERF_CPUM_SF_BASIC_MODE))
1100 sfr->basic = sample->basic;
1101 if (sfr->format & PERF_CPUM_SF_DIAG_MODE)
1102 memcpy(&sfr->diag, &sample->diag, sfr->dsdes);
1105 static void debug_sample_entry(struct hws_combined_entry *sample,
1106 struct hws_trailer_entry *te,
1107 unsigned long flags)
1109 debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1110 "sampling data entry: te->f=%i basic.def=%04x (%p)"
1111 " diag.def=%04x (%p)\n", te->f,
1112 sample->basic.def, &sample->basic,
1113 (flags & PERF_CPUM_SF_DIAG_MODE)
1114 ? sample->diag.def : 0xFFFF,
1115 (flags & PERF_CPUM_SF_DIAG_MODE)
1116 ? &sample->diag : NULL);
1119 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1120 * @event: The perf event
1121 * @sdbt: Sample-data-block table
1122 * @overflow: Event overflow counter
1124 * Walks through a sample-data-block and collects sampling data entries that are
1125 * then pushed to the perf event subsystem. Depending on the sampling function,
1126 * there can be either basic-sampling or combined-sampling data entries. A
1127 * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1128 * data entry. The sampling function is determined by the flags in the perf
1129 * event hardware structure. The function always works with a combined-sampling
1130 * data entry but ignores the the diagnostic portion if it is not available.
1132 * Note that the implementation focuses on basic-sampling data entries and, if
1133 * such an entry is not valid, the entire combined-sampling data entry is
1134 * ignored.
1136 * The overflow variables counts the number of samples that has been discarded
1137 * due to a perf event overflow.
1139 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1140 unsigned long long *overflow)
1142 unsigned long flags = SAMPL_FLAGS(&event->hw);
1143 struct hws_combined_entry *sample;
1144 struct hws_trailer_entry *te;
1145 struct sf_raw_sample *sfr;
1146 size_t sample_size;
1148 /* Prepare and initialize raw sample data */
1149 sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(&event->hw);
1150 sfr->format = flags & PERF_CPUM_SF_MODE_MASK;
1152 sample_size = event_sample_size(&event->hw);
1153 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1154 sample = (struct hws_combined_entry *) *sdbt;
1155 while ((unsigned long *) sample < (unsigned long *) te) {
1156 /* Check for an empty sample */
1157 if (!sample->basic.def)
1158 break;
1160 /* Update perf event period */
1161 perf_event_count_update(event, SAMPL_RATE(&event->hw));
1163 /* Check sampling data entry */
1164 if (sample_format_is_valid(sample, flags)) {
1165 /* If an event overflow occurred, the PMU is stopped to
1166 * throttle event delivery. Remaining sample data is
1167 * discarded.
1169 if (!*overflow) {
1170 if (sample_is_consistent(sample, flags)) {
1171 /* Deliver sample data to perf */
1172 sfr_store_sample(sfr, sample);
1173 *overflow = perf_push_sample(event, sfr);
1175 } else
1176 /* Count discarded samples */
1177 *overflow += 1;
1178 } else {
1179 debug_sample_entry(sample, te, flags);
1180 /* Sample slot is not yet written or other record.
1182 * This condition can occur if the buffer was reused
1183 * from a combined basic- and diagnostic-sampling.
1184 * If only basic-sampling is then active, entries are
1185 * written into the larger diagnostic entries.
1186 * This is typically the case for sample-data-blocks
1187 * that are not full. Stop processing if the first
1188 * invalid format was detected.
1190 if (!te->f)
1191 break;
1194 /* Reset sample slot and advance to next sample */
1195 reset_sample_slot(sample, flags);
1196 sample += sample_size;
1200 /* hw_perf_event_update() - Process sampling buffer
1201 * @event: The perf event
1202 * @flush_all: Flag to also flush partially filled sample-data-blocks
1204 * Processes the sampling buffer and create perf event samples.
1205 * The sampling buffer position are retrieved and saved in the TEAR_REG
1206 * register of the specified perf event.
1208 * Only full sample-data-blocks are processed. Specify the flash_all flag
1209 * to also walk through partially filled sample-data-blocks. It is ignored
1210 * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag
1211 * enforces the processing of full sample-data-blocks only (trailer entries
1212 * with the block-full-indicator bit set).
1214 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1216 struct hw_perf_event *hwc = &event->hw;
1217 struct hws_trailer_entry *te;
1218 unsigned long *sdbt;
1219 unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1220 int done;
1222 if (flush_all && SDB_FULL_BLOCKS(hwc))
1223 flush_all = 0;
1225 sdbt = (unsigned long *) TEAR_REG(hwc);
1226 done = event_overflow = sampl_overflow = num_sdb = 0;
1227 while (!done) {
1228 /* Get the trailer entry of the sample-data-block */
1229 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1231 /* Leave loop if no more work to do (block full indicator) */
1232 if (!te->f) {
1233 done = 1;
1234 if (!flush_all)
1235 break;
1238 /* Check the sample overflow count */
1239 if (te->overflow)
1240 /* Account sample overflows and, if a particular limit
1241 * is reached, extend the sampling buffer.
1242 * For details, see sfb_account_overflows().
1244 sampl_overflow += te->overflow;
1246 /* Timestamps are valid for full sample-data-blocks only */
1247 debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1248 "overflow=%llu timestamp=0x%llx\n",
1249 sdbt, te->overflow,
1250 (te->f) ? trailer_timestamp(te) : 0ULL);
1252 /* Collect all samples from a single sample-data-block and
1253 * flag if an (perf) event overflow happened. If so, the PMU
1254 * is stopped and remaining samples will be discarded.
1256 hw_collect_samples(event, sdbt, &event_overflow);
1257 num_sdb++;
1259 /* Reset trailer (using compare-double-and-swap) */
1260 do {
1261 te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1262 te_flags |= SDB_TE_ALERT_REQ_MASK;
1263 } while (!cmpxchg_double(&te->flags, &te->overflow,
1264 te->flags, te->overflow,
1265 te_flags, 0ULL));
1267 /* Advance to next sample-data-block */
1268 sdbt++;
1269 if (is_link_entry(sdbt))
1270 sdbt = get_next_sdbt(sdbt);
1272 /* Update event hardware registers */
1273 TEAR_REG(hwc) = (unsigned long) sdbt;
1275 /* Stop processing sample-data if all samples of the current
1276 * sample-data-block were flushed even if it was not full.
1278 if (flush_all && done)
1279 break;
1281 /* If an event overflow happened, discard samples by
1282 * processing any remaining sample-data-blocks.
1284 if (event_overflow)
1285 flush_all = 1;
1288 /* Account sample overflows in the event hardware structure */
1289 if (sampl_overflow)
1290 OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1291 sampl_overflow, 1 + num_sdb);
1292 if (sampl_overflow || event_overflow)
1293 debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1294 "overflow stats: sample=%llu event=%llu\n",
1295 sampl_overflow, event_overflow);
1298 static void cpumsf_pmu_read(struct perf_event *event)
1300 /* Nothing to do ... updates are interrupt-driven */
1303 /* Activate sampling control.
1304 * Next call of pmu_enable() starts sampling.
1306 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1308 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1310 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1311 return;
1313 if (flags & PERF_EF_RELOAD)
1314 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1316 perf_pmu_disable(event->pmu);
1317 event->hw.state = 0;
1318 cpuhw->lsctl.cs = 1;
1319 if (SAMPL_DIAG_MODE(&event->hw))
1320 cpuhw->lsctl.cd = 1;
1321 perf_pmu_enable(event->pmu);
1324 /* Deactivate sampling control.
1325 * Next call of pmu_enable() stops sampling.
1327 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1329 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1331 if (event->hw.state & PERF_HES_STOPPED)
1332 return;
1334 perf_pmu_disable(event->pmu);
1335 cpuhw->lsctl.cs = 0;
1336 cpuhw->lsctl.cd = 0;
1337 event->hw.state |= PERF_HES_STOPPED;
1339 if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1340 hw_perf_event_update(event, 1);
1341 event->hw.state |= PERF_HES_UPTODATE;
1343 perf_pmu_enable(event->pmu);
1346 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1348 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1349 int err;
1351 if (cpuhw->flags & PMU_F_IN_USE)
1352 return -EAGAIN;
1354 if (!cpuhw->sfb.sdbt)
1355 return -EINVAL;
1357 err = 0;
1358 perf_pmu_disable(event->pmu);
1360 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1362 /* Set up sampling controls. Always program the sampling register
1363 * using the SDB-table start. Reset TEAR_REG event hardware register
1364 * that is used by hw_perf_event_update() to store the sampling buffer
1365 * position after samples have been flushed.
1367 cpuhw->lsctl.s = 0;
1368 cpuhw->lsctl.h = 1;
1369 cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1370 cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1371 cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1372 hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1374 /* Ensure sampling functions are in the disabled state. If disabled,
1375 * switch on sampling enable control. */
1376 if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1377 err = -EAGAIN;
1378 goto out;
1380 cpuhw->lsctl.es = 1;
1381 if (SAMPL_DIAG_MODE(&event->hw))
1382 cpuhw->lsctl.ed = 1;
1384 /* Set in_use flag and store event */
1385 cpuhw->event = event;
1386 cpuhw->flags |= PMU_F_IN_USE;
1388 if (flags & PERF_EF_START)
1389 cpumsf_pmu_start(event, PERF_EF_RELOAD);
1390 out:
1391 perf_event_update_userpage(event);
1392 perf_pmu_enable(event->pmu);
1393 return err;
1396 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1398 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1400 perf_pmu_disable(event->pmu);
1401 cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1403 cpuhw->lsctl.es = 0;
1404 cpuhw->lsctl.ed = 0;
1405 cpuhw->flags &= ~PMU_F_IN_USE;
1406 cpuhw->event = NULL;
1408 perf_event_update_userpage(event);
1409 perf_pmu_enable(event->pmu);
1412 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1413 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1415 static struct attribute *cpumsf_pmu_events_attr[] = {
1416 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
1417 NULL,
1418 NULL,
1421 PMU_FORMAT_ATTR(event, "config:0-63");
1423 static struct attribute *cpumsf_pmu_format_attr[] = {
1424 &format_attr_event.attr,
1425 NULL,
1428 static struct attribute_group cpumsf_pmu_events_group = {
1429 .name = "events",
1430 .attrs = cpumsf_pmu_events_attr,
1432 static struct attribute_group cpumsf_pmu_format_group = {
1433 .name = "format",
1434 .attrs = cpumsf_pmu_format_attr,
1436 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1437 &cpumsf_pmu_events_group,
1438 &cpumsf_pmu_format_group,
1439 NULL,
1442 static struct pmu cpumf_sampling = {
1443 .pmu_enable = cpumsf_pmu_enable,
1444 .pmu_disable = cpumsf_pmu_disable,
1446 .event_init = cpumsf_pmu_event_init,
1447 .add = cpumsf_pmu_add,
1448 .del = cpumsf_pmu_del,
1450 .start = cpumsf_pmu_start,
1451 .stop = cpumsf_pmu_stop,
1452 .read = cpumsf_pmu_read,
1454 .attr_groups = cpumsf_pmu_attr_groups,
1457 static void cpumf_measurement_alert(struct ext_code ext_code,
1458 unsigned int alert, unsigned long unused)
1460 struct cpu_hw_sf *cpuhw;
1462 if (!(alert & CPU_MF_INT_SF_MASK))
1463 return;
1464 inc_irq_stat(IRQEXT_CMS);
1465 cpuhw = this_cpu_ptr(&cpu_hw_sf);
1467 /* Measurement alerts are shared and might happen when the PMU
1468 * is not reserved. Ignore these alerts in this case. */
1469 if (!(cpuhw->flags & PMU_F_RESERVED))
1470 return;
1472 /* The processing below must take care of multiple alert events that
1473 * might be indicated concurrently. */
1475 /* Program alert request */
1476 if (alert & CPU_MF_INT_SF_PRA) {
1477 if (cpuhw->flags & PMU_F_IN_USE)
1478 hw_perf_event_update(cpuhw->event, 0);
1479 else
1480 WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
1483 /* Report measurement alerts only for non-PRA codes */
1484 if (alert != CPU_MF_INT_SF_PRA)
1485 debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
1487 /* Sampling authorization change request */
1488 if (alert & CPU_MF_INT_SF_SACA)
1489 qsi(&cpuhw->qsi);
1491 /* Loss of sample data due to high-priority machine activities */
1492 if (alert & CPU_MF_INT_SF_LSDA) {
1493 pr_err("Sample data was lost\n");
1494 cpuhw->flags |= PMU_F_ERR_LSDA;
1495 sf_disable();
1498 /* Invalid sampling buffer entry */
1499 if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
1500 pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
1501 alert);
1502 cpuhw->flags |= PMU_F_ERR_IBE;
1503 sf_disable();
1506 static int cpusf_pmu_setup(unsigned int cpu, int flags)
1508 /* Ignore the notification if no events are scheduled on the PMU.
1509 * This might be racy...
1511 if (!atomic_read(&num_events))
1512 return 0;
1514 local_irq_disable();
1515 setup_pmc_cpu(&flags);
1516 local_irq_enable();
1517 return 0;
1520 static int s390_pmu_sf_online_cpu(unsigned int cpu)
1522 return cpusf_pmu_setup(cpu, PMC_INIT);
1525 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
1527 return cpusf_pmu_setup(cpu, PMC_RELEASE);
1530 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
1532 if (!cpum_sf_avail())
1533 return -ENODEV;
1534 return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
1537 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
1539 int rc;
1540 unsigned long min, max;
1542 if (!cpum_sf_avail())
1543 return -ENODEV;
1544 if (!val || !strlen(val))
1545 return -EINVAL;
1547 /* Valid parameter values: "min,max" or "max" */
1548 min = CPUM_SF_MIN_SDB;
1549 max = CPUM_SF_MAX_SDB;
1550 if (strchr(val, ','))
1551 rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
1552 else
1553 rc = kstrtoul(val, 10, &max);
1555 if (min < 2 || min >= max || max > get_num_physpages())
1556 rc = -EINVAL;
1557 if (rc)
1558 return rc;
1560 sfb_set_limits(min, max);
1561 pr_info("The sampling buffer limits have changed to: "
1562 "min=%lu max=%lu (diag=x%lu)\n",
1563 CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
1564 return 0;
1567 #define param_check_sfb_size(name, p) __param_check(name, p, void)
1568 static const struct kernel_param_ops param_ops_sfb_size = {
1569 .set = param_set_sfb_size,
1570 .get = param_get_sfb_size,
1573 #define RS_INIT_FAILURE_QSI 0x0001
1574 #define RS_INIT_FAILURE_BSDES 0x0002
1575 #define RS_INIT_FAILURE_ALRT 0x0003
1576 #define RS_INIT_FAILURE_PERF 0x0004
1577 static void __init pr_cpumsf_err(unsigned int reason)
1579 pr_err("Sampling facility support for perf is not available: "
1580 "reason=%04x\n", reason);
1583 static int __init init_cpum_sampling_pmu(void)
1585 struct hws_qsi_info_block si;
1586 int err;
1588 if (!cpum_sf_avail())
1589 return -ENODEV;
1591 memset(&si, 0, sizeof(si));
1592 if (qsi(&si)) {
1593 pr_cpumsf_err(RS_INIT_FAILURE_QSI);
1594 return -ENODEV;
1597 if (si.bsdes != sizeof(struct hws_basic_entry)) {
1598 pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
1599 return -EINVAL;
1602 if (si.ad) {
1603 sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
1604 cpumsf_pmu_events_attr[1] =
1605 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
1608 sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
1609 if (!sfdbg)
1610 pr_err("Registering for s390dbf failed\n");
1611 debug_register_view(sfdbg, &debug_sprintf_view);
1613 err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
1614 cpumf_measurement_alert);
1615 if (err) {
1616 pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
1617 goto out;
1620 err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
1621 if (err) {
1622 pr_cpumsf_err(RS_INIT_FAILURE_PERF);
1623 unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
1624 cpumf_measurement_alert);
1625 goto out;
1628 cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
1629 s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
1630 out:
1631 return err;
1633 arch_initcall(init_cpum_sampling_pmu);
1634 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);