1 // SPDX-License-Identifier: GPL-2.0
3 * Performance event support for the System z CPU-measurement Sampling Facility
5 * Copyright IBM Corp. 2013, 2018
6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
8 #define KMSG_COMPONENT "cpum_sf"
9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
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/pid.h>
16 #include <linux/notifier.h>
17 #include <linux/export.h>
18 #include <linux/slab.h>
20 #include <linux/moduleparam.h>
21 #include <asm/cpu_mf.h>
23 #include <asm/debug.h>
24 #include <asm/timex.h>
26 /* Minimum number of sample-data-block-tables:
27 * At least one table is required for the sampling buffer structure.
28 * A single table contains up to 511 pointers to sample-data-blocks.
30 #define CPUM_SF_MIN_SDBT 1
32 /* Number of sample-data-blocks per sample-data-block-table (SDBT):
33 * A table contains SDB pointers (8 bytes) and one table-link entry
34 * that points to the origin of the next SDBT.
36 #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8)
38 /* Maximum page offset for an SDBT table-link entry:
39 * If this page offset is reached, a table-link entry to the next SDBT
42 #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8)
43 static inline int require_table_link(const void *sdbt)
45 return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
48 /* Minimum and maximum sampling buffer sizes:
50 * This number represents the maximum size of the sampling buffer taking
51 * the number of sample-data-block-tables into account. Note that these
52 * numbers apply to the basic-sampling function only.
53 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if
54 * the diagnostic-sampling function is active.
56 * Sampling buffer size Buffer characteristics
57 * ---------------------------------------------------
58 * 64KB == 16 pages (4KB per page)
59 * 1 page for SDB-tables
62 * 32MB == 8192 pages (4KB per page)
63 * 16 pages for SDB-tables
66 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15;
67 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176;
68 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
71 unsigned long *sdbt; /* Sample-data-block-table origin */
72 /* buffer characteristics (required for buffer increments) */
73 unsigned long num_sdb; /* Number of sample-data-blocks */
74 unsigned long num_sdbt; /* Number of sample-data-block-tables */
75 unsigned long *tail; /* last sample-data-block-table */
80 unsigned long head; /* index of SDB of buffer head */
81 unsigned long alert_mark; /* index of SDB of alert request position */
82 unsigned long empty_mark; /* mark of SDB not marked full */
83 unsigned long *sdb_index; /* SDB address for fast lookup */
84 unsigned long *sdbt_index; /* SDBT address for fast lookup */
88 /* CPU-measurement sampling information block */
89 struct hws_qsi_info_block qsi;
90 /* CPU-measurement sampling control block */
91 struct hws_lsctl_request_block lsctl;
92 struct sf_buffer sfb; /* Sampling buffer */
93 unsigned int flags; /* Status flags */
94 struct perf_event *event; /* Scheduled perf event */
95 struct perf_output_handle handle; /* AUX buffer output handle */
97 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
100 static debug_info_t *sfdbg;
103 * sf_disable() - Switch off sampling facility
105 static int sf_disable(void)
107 struct hws_lsctl_request_block sreq;
109 memset(&sreq, 0, sizeof(sreq));
114 * sf_buffer_available() - Check for an allocated sampling buffer
116 static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
118 return !!cpuhw->sfb.sdbt;
122 * deallocate sampling facility buffer
124 static void free_sampling_buffer(struct sf_buffer *sfb)
126 unsigned long *sdbt, *curr;
134 /* Free the SDBT after all SDBs are processed... */
139 /* Process table-link entries */
140 if (is_link_entry(curr)) {
141 curr = get_next_sdbt(curr);
143 free_page((unsigned long) sdbt);
145 /* If the origin is reached, sampling buffer is freed */
146 if (curr == sfb->sdbt)
151 /* Process SDB pointer */
159 debug_sprintf_event(sfdbg, 5,
160 "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
161 memset(sfb, 0, sizeof(*sfb));
164 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
166 unsigned long sdb, *trailer;
168 /* Allocate and initialize sample-data-block */
169 sdb = get_zeroed_page(gfp_flags);
172 trailer = trailer_entry_ptr(sdb);
173 *trailer = SDB_TE_ALERT_REQ_MASK;
175 /* Link SDB into the sample-data-block-table */
182 * realloc_sampling_buffer() - extend sampler memory
184 * Allocates new sample-data-blocks and adds them to the specified sampling
187 * Important: This modifies the sampling buffer and must be called when the
188 * sampling facility is disabled.
190 * Returns zero on success, non-zero otherwise.
192 static int realloc_sampling_buffer(struct sf_buffer *sfb,
193 unsigned long num_sdb, gfp_t gfp_flags)
196 unsigned long *new, *tail, *tail_prev = NULL;
198 if (!sfb->sdbt || !sfb->tail)
201 if (!is_link_entry(sfb->tail))
204 /* Append to the existing sampling buffer, overwriting the table-link
206 * The tail variables always points to the "tail" (last and table-link)
207 * entry in an SDB-table.
211 /* Do a sanity check whether the table-link entry points to
212 * the sampling buffer origin.
214 if (sfb->sdbt != get_next_sdbt(tail)) {
215 debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
216 "sampling buffer is not linked: origin=%p"
218 (void *) sfb->sdbt, (void *) tail);
222 /* Allocate remaining SDBs */
224 for (i = 0; i < num_sdb; i++) {
225 /* Allocate a new SDB-table if it is full. */
226 if (require_table_link(tail)) {
227 new = (unsigned long *) get_zeroed_page(gfp_flags);
233 /* Link current page to tail of chain */
234 *tail = (unsigned long)(void *) new + 1;
239 /* Allocate a new sample-data-block.
240 * If there is not enough memory, stop the realloc process
241 * and simply use what was allocated. If this is a temporary
242 * issue, a new realloc call (if required) might succeed.
244 rc = alloc_sample_data_block(tail, gfp_flags);
246 /* Undo last SDBT. An SDBT with no SDB at its first
247 * entry but with an SDBT entry instead can not be
248 * handled by the interrupt handler code.
249 * Avoid this situation.
253 free_page((unsigned long) new);
260 tail_prev = new = NULL; /* Allocated at least one SBD */
263 /* Link sampling buffer to its origin */
264 *tail = (unsigned long) sfb->sdbt + 1;
267 debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
268 " settings: sdbt=%lu sdb=%lu\n",
269 sfb->num_sdbt, sfb->num_sdb);
274 * allocate_sampling_buffer() - allocate sampler memory
276 * Allocates and initializes a sampling buffer structure using the
277 * specified number of sample-data-blocks (SDB). For each allocation,
278 * a 4K page is used. The number of sample-data-block-tables (SDBT)
279 * are calculated from SDBs.
280 * Also set the ALERT_REQ mask in each SDBs trailer.
282 * Returns zero on success, non-zero otherwise.
284 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
291 /* Allocate the sample-data-block-table origin */
292 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
298 /* Link the table origin to point to itself to prepare for
299 * realloc_sampling_buffer() invocation.
301 sfb->tail = sfb->sdbt;
302 *sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
304 /* Allocate requested number of sample-data-blocks */
305 rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
307 free_sampling_buffer(sfb);
308 debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
309 "realloc_sampling_buffer failed with rc=%i\n", rc);
311 debug_sprintf_event(sfdbg, 4,
312 "alloc_sampling_buffer: tear=%p dear=%p\n",
313 sfb->sdbt, (void *) *sfb->sdbt);
317 static void sfb_set_limits(unsigned long min, unsigned long max)
319 struct hws_qsi_info_block si;
321 CPUM_SF_MIN_SDB = min;
322 CPUM_SF_MAX_SDB = max;
324 memset(&si, 0, sizeof(si));
326 CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
329 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
331 return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
335 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
336 struct hw_perf_event *hwc)
339 return SFB_ALLOC_REG(hwc);
340 if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
341 return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
345 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
346 struct hw_perf_event *hwc)
348 return sfb_pending_allocs(sfb, hwc) > 0;
351 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
353 /* Limit the number of SDBs to not exceed the maximum */
354 num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
356 SFB_ALLOC_REG(hwc) += num;
359 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
361 SFB_ALLOC_REG(hwc) = 0;
362 sfb_account_allocs(num, hwc);
365 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
368 free_sampling_buffer(&cpuhw->sfb);
371 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
373 unsigned long n_sdb, freq, factor;
376 /* Calculate sampling buffers using 4K pages
378 * 1. Determine the sample data size which depends on the used
379 * sampling functions, for example, basic-sampling or
380 * basic-sampling with diagnostic-sampling.
382 * 2. Use the sampling frequency as input. The sampling buffer is
383 * designed for almost one second. This can be adjusted through
384 * the "factor" variable.
385 * In any case, alloc_sampling_buffer() sets the Alert Request
386 * Control indicator to trigger a measurement-alert to harvest
387 * sample-data-blocks (sdb).
389 * 3. Compute the number of sample-data-blocks and ensure a minimum
390 * of CPUM_SF_MIN_SDB. Also ensure the upper limit does not
391 * exceed a "calculated" maximum. The symbolic maximum is
392 * designed for basic-sampling only and needs to be increased if
393 * diagnostic-sampling is active.
394 * See also the remarks for these symbolic constants.
396 * 4. Compute the number of sample-data-block-tables (SDBT) and
397 * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
400 sample_size = sizeof(struct hws_basic_entry);
401 freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
403 n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
404 if (n_sdb < CPUM_SF_MIN_SDB)
405 n_sdb = CPUM_SF_MIN_SDB;
407 /* If there is already a sampling buffer allocated, it is very likely
408 * that the sampling facility is enabled too. If the event to be
409 * initialized requires a greater sampling buffer, the allocation must
410 * be postponed. Changing the sampling buffer requires the sampling
411 * facility to be in the disabled state. So, account the number of
412 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
413 * before the event is started.
415 sfb_init_allocs(n_sdb, hwc);
416 if (sf_buffer_available(cpuhw))
419 debug_sprintf_event(sfdbg, 3,
420 "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
421 " sample_size=%lu cpuhw=%p\n",
422 SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
425 return alloc_sampling_buffer(&cpuhw->sfb,
426 sfb_pending_allocs(&cpuhw->sfb, hwc));
429 static unsigned long min_percent(unsigned int percent, unsigned long base,
432 return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
435 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
437 /* Use a percentage-based approach to extend the sampling facility
438 * buffer. Accept up to 5% sample data loss.
439 * Vary the extents between 1% to 5% of the current number of
440 * sample-data-blocks.
445 return min_percent(1, base, 1);
447 return min_percent(1, base, 1);
449 return min_percent(2, base, 2);
451 return min_percent(3, base, 3);
453 return min_percent(4, base, 4);
455 return min_percent(5, base, 8);
458 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
459 struct hw_perf_event *hwc)
461 unsigned long ratio, num;
463 if (!OVERFLOW_REG(hwc))
466 /* The sample_overflow contains the average number of sample data
467 * that has been lost because sample-data-blocks were full.
469 * Calculate the total number of sample data entries that has been
470 * discarded. Then calculate the ratio of lost samples to total samples
471 * per second in percent.
473 ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
474 sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
476 /* Compute number of sample-data-blocks */
477 num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
479 sfb_account_allocs(num, hwc);
481 debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
482 " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
483 OVERFLOW_REG(hwc) = 0;
486 /* extend_sampling_buffer() - Extend sampling buffer
487 * @sfb: Sampling buffer structure (for local CPU)
488 * @hwc: Perf event hardware structure
490 * Use this function to extend the sampling buffer based on the overflow counter
491 * and postponed allocation extents stored in the specified Perf event hardware.
493 * Important: This function disables the sampling facility in order to safely
494 * change the sampling buffer structure. Do not call this function
495 * when the PMU is active.
497 static void extend_sampling_buffer(struct sf_buffer *sfb,
498 struct hw_perf_event *hwc)
500 unsigned long num, num_old;
503 num = sfb_pending_allocs(sfb, hwc);
506 num_old = sfb->num_sdb;
508 /* Disable the sampling facility to reset any states and also
509 * clear pending measurement alerts.
513 /* Extend the sampling buffer.
514 * This memory allocation typically happens in an atomic context when
515 * called by perf. Because this is a reallocation, it is fine if the
516 * new SDB-request cannot be satisfied immediately.
518 rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
520 debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
521 "failed with rc=%i\n", rc);
523 if (sfb_has_pending_allocs(sfb, hwc))
524 debug_sprintf_event(sfdbg, 5, "sfb: extend: "
525 "req=%lu alloc=%lu remaining=%lu\n",
526 num, sfb->num_sdb - num_old,
527 sfb_pending_allocs(sfb, hwc));
531 /* Number of perf events counting hardware events */
532 static atomic_t num_events;
533 /* Used to avoid races in calling reserve/release_cpumf_hardware */
534 static DEFINE_MUTEX(pmc_reserve_mutex);
537 #define PMC_RELEASE 1
538 #define PMC_FAILURE 2
539 static void setup_pmc_cpu(void *flags)
542 struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
545 switch (*((int *) flags)) {
547 memset(cpusf, 0, sizeof(*cpusf));
548 err = qsi(&cpusf->qsi);
551 cpusf->flags |= PMU_F_RESERVED;
554 pr_err("Switching off the sampling facility failed "
555 "with rc=%i\n", err);
556 debug_sprintf_event(sfdbg, 5,
557 "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
560 cpusf->flags &= ~PMU_F_RESERVED;
563 pr_err("Switching off the sampling facility failed "
564 "with rc=%i\n", err);
566 deallocate_buffers(cpusf);
567 debug_sprintf_event(sfdbg, 5,
568 "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
572 *((int *) flags) |= PMC_FAILURE;
575 static void release_pmc_hardware(void)
577 int flags = PMC_RELEASE;
579 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
580 on_each_cpu(setup_pmc_cpu, &flags, 1);
583 static int reserve_pmc_hardware(void)
585 int flags = PMC_INIT;
587 on_each_cpu(setup_pmc_cpu, &flags, 1);
588 if (flags & PMC_FAILURE) {
589 release_pmc_hardware();
592 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
597 static void hw_perf_event_destroy(struct perf_event *event)
599 /* Release PMC if this is the last perf event */
600 if (!atomic_add_unless(&num_events, -1, 1)) {
601 mutex_lock(&pmc_reserve_mutex);
602 if (atomic_dec_return(&num_events) == 0)
603 release_pmc_hardware();
604 mutex_unlock(&pmc_reserve_mutex);
608 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
610 hwc->sample_period = period;
611 hwc->last_period = hwc->sample_period;
612 local64_set(&hwc->period_left, hwc->sample_period);
615 static void hw_reset_registers(struct hw_perf_event *hwc,
616 unsigned long *sdbt_origin)
618 /* (Re)set to first sample-data-block-table */
619 TEAR_REG(hwc) = (unsigned long) sdbt_origin;
622 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
625 return clamp_t(unsigned long, rate,
626 si->min_sampl_rate, si->max_sampl_rate);
629 static u32 cpumsf_pid_type(struct perf_event *event,
630 u32 pid, enum pid_type type)
632 struct task_struct *tsk;
638 tsk = find_task_by_pid_ns(pid, &init_pid_ns);
642 * Only top level events contain the pid namespace in which
646 event = event->parent;
647 pid = __task_pid_nr_ns(tsk, type, event->ns);
649 * See also 1d953111b648
650 * "perf/core: Don't report zero PIDs for exiting tasks".
652 if (!pid && !pid_alive(tsk))
659 static void cpumsf_output_event_pid(struct perf_event *event,
660 struct perf_sample_data *data,
661 struct pt_regs *regs)
664 struct perf_event_header header;
665 struct perf_output_handle handle;
668 * Obtain the PID from the basic-sampling data entry and
669 * correct the data->tid_entry.pid value.
671 pid = data->tid_entry.pid;
673 /* Protect callchain buffers, tasks */
676 perf_prepare_sample(&header, data, event, regs);
677 if (perf_output_begin(&handle, event, header.size))
680 /* Update the process ID (see also kernel/events/core.c) */
681 data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
682 data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
684 perf_output_sample(&handle, &header, data, event);
685 perf_output_end(&handle);
690 static int __hw_perf_event_init(struct perf_event *event)
692 struct cpu_hw_sf *cpuhw;
693 struct hws_qsi_info_block si;
694 struct perf_event_attr *attr = &event->attr;
695 struct hw_perf_event *hwc = &event->hw;
699 /* Reserve CPU-measurement sampling facility */
701 if (!atomic_inc_not_zero(&num_events)) {
702 mutex_lock(&pmc_reserve_mutex);
703 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
706 atomic_inc(&num_events);
707 mutex_unlock(&pmc_reserve_mutex);
709 event->destroy = hw_perf_event_destroy;
714 /* Access per-CPU sampling information (query sampling info) */
716 * The event->cpu value can be -1 to count on every CPU, for example,
717 * when attaching to a task. If this is specified, use the query
718 * sampling info from the current CPU, otherwise use event->cpu to
719 * retrieve the per-CPU information.
720 * Later, cpuhw indicates whether to allocate sampling buffers for a
721 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
723 memset(&si, 0, sizeof(si));
725 if (event->cpu == -1)
728 /* Event is pinned to a particular CPU, retrieve the per-CPU
729 * sampling structure for accessing the CPU-specific QSI.
731 cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
735 /* Check sampling facility authorization and, if not authorized,
736 * fall back to other PMUs. It is safe to check any CPU because
737 * the authorization is identical for all configured CPUs.
744 /* Always enable basic sampling */
745 SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
747 /* Check if diagnostic sampling is requested. Deny if the required
748 * sampling authorization is missing.
750 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
755 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
758 /* Check and set other sampling flags */
759 if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
760 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
762 /* The sampling information (si) contains information about the
763 * min/max sampling intervals and the CPU speed. So calculate the
764 * correct sampling interval and avoid the whole period adjust
769 if (!attr->sample_freq) {
773 rate = freq_to_sample_rate(&si, attr->sample_freq);
774 rate = hw_limit_rate(&si, rate);
776 attr->sample_period = rate;
778 /* The min/max sampling rates specifies the valid range
779 * of sample periods. If the specified sample period is
780 * out of range, limit the period to the range boundary.
782 rate = hw_limit_rate(&si, hwc->sample_period);
784 /* The perf core maintains a maximum sample rate that is
785 * configurable through the sysctl interface. Ensure the
786 * sampling rate does not exceed this value. This also helps
787 * to avoid throttling when pushing samples with
788 * perf_event_overflow().
790 if (sample_rate_to_freq(&si, rate) >
791 sysctl_perf_event_sample_rate) {
793 debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
797 SAMPL_RATE(hwc) = rate;
798 hw_init_period(hwc, SAMPL_RATE(hwc));
800 /* Initialize sample data overflow accounting */
801 hwc->extra_reg.reg = REG_OVERFLOW;
802 OVERFLOW_REG(hwc) = 0;
804 /* Use AUX buffer. No need to allocate it by ourself */
805 if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
808 /* Allocate the per-CPU sampling buffer using the CPU information
809 * from the event. If the event is not pinned to a particular
810 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
811 * buffers for each online CPU.
814 /* Event is pinned to a particular CPU */
815 err = allocate_buffers(cpuhw, hwc);
817 /* Event is not pinned, allocate sampling buffer on
820 for_each_online_cpu(cpu) {
821 cpuhw = &per_cpu(cpu_hw_sf, cpu);
822 err = allocate_buffers(cpuhw, hwc);
828 /* If PID/TID sampling is active, replace the default overflow
829 * handler to extract and resolve the PIDs from the basic-sampling
832 if (event->attr.sample_type & PERF_SAMPLE_TID)
833 if (is_default_overflow_handler(event))
834 event->overflow_handler = cpumsf_output_event_pid;
839 static int cpumsf_pmu_event_init(struct perf_event *event)
843 /* No support for taken branch sampling */
844 if (has_branch_stack(event))
847 switch (event->attr.type) {
849 if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
850 (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
853 case PERF_TYPE_HARDWARE:
854 /* Support sampling of CPU cycles in addition to the
855 * counter facility. However, the counter facility
856 * is more precise and, hence, restrict this PMU to
857 * sampling events only.
859 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
861 if (!is_sampling_event(event))
868 /* Check online status of the CPU to which the event is pinned */
869 if (event->cpu >= 0 && !cpu_online(event->cpu))
872 /* Force reset of idle/hv excludes regardless of what the
875 if (event->attr.exclude_hv)
876 event->attr.exclude_hv = 0;
877 if (event->attr.exclude_idle)
878 event->attr.exclude_idle = 0;
880 err = __hw_perf_event_init(event);
883 event->destroy(event);
887 static void cpumsf_pmu_enable(struct pmu *pmu)
889 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
890 struct hw_perf_event *hwc;
893 if (cpuhw->flags & PMU_F_ENABLED)
896 if (cpuhw->flags & PMU_F_ERR_MASK)
899 /* Check whether to extent the sampling buffer.
901 * Two conditions trigger an increase of the sampling buffer for a
903 * 1. Postponed buffer allocations from the event initialization.
904 * 2. Sampling overflows that contribute to pending allocations.
906 * Note that the extend_sampling_buffer() function disables the sampling
907 * facility, but it can be fully re-enabled using sampling controls that
908 * have been saved in cpumsf_pmu_disable().
911 hwc = &cpuhw->event->hw;
912 if (!(SAMPL_DIAG_MODE(hwc))) {
914 * Account number of overflow-designated
917 sfb_account_overflows(cpuhw, hwc);
918 if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
919 extend_sampling_buffer(&cpuhw->sfb, hwc);
923 /* (Re)enable the PMU and sampling facility */
924 cpuhw->flags |= PMU_F_ENABLED;
927 err = lsctl(&cpuhw->lsctl);
929 cpuhw->flags &= ~PMU_F_ENABLED;
930 pr_err("Loading sampling controls failed: op=%i err=%i\n",
935 /* Load current program parameter */
936 lpp(&S390_lowcore.lpp);
938 debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
939 "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
940 cpuhw->lsctl.ed, cpuhw->lsctl.cd,
941 (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
944 static void cpumsf_pmu_disable(struct pmu *pmu)
946 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
947 struct hws_lsctl_request_block inactive;
948 struct hws_qsi_info_block si;
951 if (!(cpuhw->flags & PMU_F_ENABLED))
954 if (cpuhw->flags & PMU_F_ERR_MASK)
957 /* Switch off sampling activation control */
958 inactive = cpuhw->lsctl;
962 err = lsctl(&inactive);
964 pr_err("Loading sampling controls failed: op=%i err=%i\n",
969 /* Save state of TEAR and DEAR register contents */
971 /* TEAR/DEAR values are valid only if the sampling facility is
972 * enabled. Note that cpumsf_pmu_disable() might be called even
973 * for a disabled sampling facility because cpumsf_pmu_enable()
974 * controls the enable/disable state.
977 cpuhw->lsctl.tear = si.tear;
978 cpuhw->lsctl.dear = si.dear;
981 debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
982 "qsi() failed with err=%i\n", err);
984 cpuhw->flags &= ~PMU_F_ENABLED;
987 /* perf_exclude_event() - Filter event
988 * @event: The perf event
989 * @regs: pt_regs structure
990 * @sde_regs: Sample-data-entry (sde) regs structure
992 * Filter perf events according to their exclude specification.
994 * Return non-zero if the event shall be excluded.
996 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
997 struct perf_sf_sde_regs *sde_regs)
999 if (event->attr.exclude_user && user_mode(regs))
1001 if (event->attr.exclude_kernel && !user_mode(regs))
1003 if (event->attr.exclude_guest && sde_regs->in_guest)
1005 if (event->attr.exclude_host && !sde_regs->in_guest)
1010 /* perf_push_sample() - Push samples to perf
1011 * @event: The perf event
1012 * @sample: Hardware sample data
1014 * Use the hardware sample data to create perf event sample. The sample
1015 * is the pushed to the event subsystem and the function checks for
1016 * possible event overflows. If an event overflow occurs, the PMU is
1019 * Return non-zero if an event overflow occurred.
1021 static int perf_push_sample(struct perf_event *event,
1022 struct hws_basic_entry *basic)
1025 struct pt_regs regs;
1026 struct perf_sf_sde_regs *sde_regs;
1027 struct perf_sample_data data;
1029 /* Setup perf sample */
1030 perf_sample_data_init(&data, 0, event->hw.last_period);
1032 /* Setup pt_regs to look like an CPU-measurement external interrupt
1033 * using the Program Request Alert code. The regs.int_parm_long
1034 * field which is unused contains additional sample-data-entry related
1037 memset(®s, 0, sizeof(regs));
1038 regs.int_code = 0x1407;
1039 regs.int_parm = CPU_MF_INT_SF_PRA;
1040 sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long;
1042 psw_bits(regs.psw).ia = basic->ia;
1043 psw_bits(regs.psw).dat = basic->T;
1044 psw_bits(regs.psw).wait = basic->W;
1045 psw_bits(regs.psw).pstate = basic->P;
1046 psw_bits(regs.psw).as = basic->AS;
1049 * Use the hardware provided configuration level to decide if the
1050 * sample belongs to a guest or host. If that is not available,
1051 * fall back to the following heuristics:
1052 * A non-zero guest program parameter always indicates a guest
1053 * sample. Some early samples or samples from guests without
1054 * lpp usage would be misaccounted to the host. We use the asn
1055 * value as an addon heuristic to detect most of these guest samples.
1056 * If the value differs from 0xffff (the host value), we assume to
1059 switch (basic->CL) {
1060 case 1: /* logical partition */
1061 sde_regs->in_guest = 0;
1063 case 2: /* virtual machine */
1064 sde_regs->in_guest = 1;
1066 default: /* old machine, use heuristics */
1067 if (basic->gpp || basic->prim_asn != 0xffff)
1068 sde_regs->in_guest = 1;
1073 * Store the PID value from the sample-data-entry to be
1074 * processed and resolved by cpumsf_output_event_pid().
1076 data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1079 if (perf_exclude_event(event, ®s, sde_regs))
1081 if (perf_event_overflow(event, &data, ®s)) {
1083 event->pmu->stop(event, 0);
1085 perf_event_update_userpage(event);
1090 static void perf_event_count_update(struct perf_event *event, u64 count)
1092 local64_add(count, &event->count);
1095 static void debug_sample_entry(struct hws_basic_entry *sample,
1096 struct hws_trailer_entry *te)
1098 debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1099 "sampling data entry: te->f=%i basic.def=%04x (%p)\n",
1100 te->f, sample->def, sample);
1103 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1104 * @event: The perf event
1105 * @sdbt: Sample-data-block table
1106 * @overflow: Event overflow counter
1108 * Walks through a sample-data-block and collects sampling data entries that are
1109 * then pushed to the perf event subsystem. Depending on the sampling function,
1110 * there can be either basic-sampling or combined-sampling data entries. A
1111 * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1112 * data entry. The sampling function is determined by the flags in the perf
1113 * event hardware structure. The function always works with a combined-sampling
1114 * data entry but ignores the the diagnostic portion if it is not available.
1116 * Note that the implementation focuses on basic-sampling data entries and, if
1117 * such an entry is not valid, the entire combined-sampling data entry is
1120 * The overflow variables counts the number of samples that has been discarded
1121 * due to a perf event overflow.
1123 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1124 unsigned long long *overflow)
1126 struct hws_trailer_entry *te;
1127 struct hws_basic_entry *sample;
1129 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1130 sample = (struct hws_basic_entry *) *sdbt;
1131 while ((unsigned long *) sample < (unsigned long *) te) {
1132 /* Check for an empty sample */
1136 /* Update perf event period */
1137 perf_event_count_update(event, SAMPL_RATE(&event->hw));
1139 /* Check whether sample is valid */
1140 if (sample->def == 0x0001) {
1141 /* If an event overflow occurred, the PMU is stopped to
1142 * throttle event delivery. Remaining sample data is
1146 /* Check whether sample is consistent */
1147 if (sample->I == 0 && sample->W == 0) {
1148 /* Deliver sample data to perf */
1149 *overflow = perf_push_sample(event,
1153 /* Count discarded samples */
1156 debug_sample_entry(sample, te);
1157 /* Sample slot is not yet written or other record.
1159 * This condition can occur if the buffer was reused
1160 * from a combined basic- and diagnostic-sampling.
1161 * If only basic-sampling is then active, entries are
1162 * written into the larger diagnostic entries.
1163 * This is typically the case for sample-data-blocks
1164 * that are not full. Stop processing if the first
1165 * invalid format was detected.
1171 /* Reset sample slot and advance to next sample */
1177 /* hw_perf_event_update() - Process sampling buffer
1178 * @event: The perf event
1179 * @flush_all: Flag to also flush partially filled sample-data-blocks
1181 * Processes the sampling buffer and create perf event samples.
1182 * The sampling buffer position are retrieved and saved in the TEAR_REG
1183 * register of the specified perf event.
1185 * Only full sample-data-blocks are processed. Specify the flash_all flag
1186 * to also walk through partially filled sample-data-blocks. It is ignored
1187 * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag
1188 * enforces the processing of full sample-data-blocks only (trailer entries
1189 * with the block-full-indicator bit set).
1191 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1193 struct hw_perf_event *hwc = &event->hw;
1194 struct hws_trailer_entry *te;
1195 unsigned long *sdbt;
1196 unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1200 * AUX buffer is used when in diagnostic sampling mode.
1201 * No perf events/samples are created.
1203 if (SAMPL_DIAG_MODE(&event->hw))
1206 if (flush_all && SDB_FULL_BLOCKS(hwc))
1209 sdbt = (unsigned long *) TEAR_REG(hwc);
1210 done = event_overflow = sampl_overflow = num_sdb = 0;
1212 /* Get the trailer entry of the sample-data-block */
1213 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1215 /* Leave loop if no more work to do (block full indicator) */
1222 /* Check the sample overflow count */
1224 /* Account sample overflows and, if a particular limit
1225 * is reached, extend the sampling buffer.
1226 * For details, see sfb_account_overflows().
1228 sampl_overflow += te->overflow;
1230 /* Timestamps are valid for full sample-data-blocks only */
1231 debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1232 "overflow=%llu timestamp=0x%llx\n",
1234 (te->f) ? trailer_timestamp(te) : 0ULL);
1236 /* Collect all samples from a single sample-data-block and
1237 * flag if an (perf) event overflow happened. If so, the PMU
1238 * is stopped and remaining samples will be discarded.
1240 hw_collect_samples(event, sdbt, &event_overflow);
1243 /* Reset trailer (using compare-double-and-swap) */
1245 te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1246 te_flags |= SDB_TE_ALERT_REQ_MASK;
1247 } while (!cmpxchg_double(&te->flags, &te->overflow,
1248 te->flags, te->overflow,
1251 /* Advance to next sample-data-block */
1253 if (is_link_entry(sdbt))
1254 sdbt = get_next_sdbt(sdbt);
1256 /* Update event hardware registers */
1257 TEAR_REG(hwc) = (unsigned long) sdbt;
1259 /* Stop processing sample-data if all samples of the current
1260 * sample-data-block were flushed even if it was not full.
1262 if (flush_all && done)
1266 /* Account sample overflows in the event hardware structure */
1268 OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1269 sampl_overflow, 1 + num_sdb);
1271 /* Perf_event_overflow() and perf_event_account_interrupt() limit
1272 * the interrupt rate to an upper limit. Roughly 1000 samples per
1274 * Hitting this limit results in a large number
1275 * of throttled REF_REPORT_THROTTLE entries and the samples
1277 * Slightly increase the interval to avoid hitting this limit.
1279 if (event_overflow) {
1280 SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
1281 debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n",
1283 DIV_ROUND_UP(SAMPL_RATE(hwc), 10));
1286 if (sampl_overflow || event_overflow)
1287 debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1288 "overflow stats: sample=%llu event=%llu\n",
1289 sampl_overflow, event_overflow);
1292 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1293 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1294 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1295 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1298 * Get trailer entry by index of SDB.
1300 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1301 unsigned long index)
1305 index = AUX_SDB_INDEX(aux, index);
1306 sdb = aux->sdb_index[index];
1307 return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1311 * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1312 * disabled. Collect the full SDBs in AUX buffer which have not reached
1313 * the point of alert indicator. And ignore the SDBs which are not
1316 * 1. Scan SDBs to see how much data is there and consume them.
1317 * 2. Remove alert indicator in the buffer.
1319 static void aux_output_end(struct perf_output_handle *handle)
1321 unsigned long i, range_scan, idx;
1322 struct aux_buffer *aux;
1323 struct hws_trailer_entry *te;
1325 aux = perf_get_aux(handle);
1329 range_scan = AUX_SDB_NUM_ALERT(aux);
1330 for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1331 te = aux_sdb_trailer(aux, idx);
1332 if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1335 /* i is num of SDBs which are full */
1336 perf_aux_output_end(handle, i << PAGE_SHIFT);
1338 /* Remove alert indicators in the buffer */
1339 te = aux_sdb_trailer(aux, aux->alert_mark);
1340 te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1342 debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1346 * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1347 * is first added to the CPU or rescheduled again to the CPU. It is called
1348 * with pmu disabled.
1350 * 1. Reset the trailer of SDBs to get ready for new data.
1351 * 2. Tell the hardware where to put the data by reset the SDBs buffer
1354 static int aux_output_begin(struct perf_output_handle *handle,
1355 struct aux_buffer *aux,
1356 struct cpu_hw_sf *cpuhw)
1358 unsigned long range;
1359 unsigned long i, range_scan, idx;
1360 unsigned long head, base, offset;
1361 struct hws_trailer_entry *te;
1363 if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1366 aux->head = handle->head >> PAGE_SHIFT;
1367 range = (handle->size + 1) >> PAGE_SHIFT;
1372 * SDBs between aux->head and aux->empty_mark are already ready
1373 * for new data. range_scan is num of SDBs not within them.
1375 if (range > AUX_SDB_NUM_EMPTY(aux)) {
1376 range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1377 idx = aux->empty_mark + 1;
1378 for (i = 0; i < range_scan; i++, idx++) {
1379 te = aux_sdb_trailer(aux, idx);
1380 te->flags &= ~(SDB_TE_BUFFER_FULL_MASK |
1381 SDB_TE_ALERT_REQ_MASK);
1384 /* Save the position of empty SDBs */
1385 aux->empty_mark = aux->head + range - 1;
1388 /* Set alert indicator */
1389 aux->alert_mark = aux->head + range/2 - 1;
1390 te = aux_sdb_trailer(aux, aux->alert_mark);
1391 te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1393 /* Reset hardware buffer head */
1394 head = AUX_SDB_INDEX(aux, aux->head);
1395 base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1396 offset = head % CPUM_SF_SDB_PER_TABLE;
1397 cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1398 cpuhw->lsctl.dear = aux->sdb_index[head];
1400 debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1401 "head->alert_mark->empty_mark (num_alert, range)"
1402 "[%lx -> %lx -> %lx] (%lx, %lx) "
1403 "tear index %lx, tear %lx dear %lx\n",
1404 aux->head, aux->alert_mark, aux->empty_mark,
1405 AUX_SDB_NUM_ALERT(aux), range,
1406 head / CPUM_SF_SDB_PER_TABLE,
1414 * Set alert indicator on SDB at index @alert_index while sampler is running.
1416 * Return true if successfully.
1417 * Return false if full indicator is already set by hardware sampler.
1419 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1420 unsigned long long *overflow)
1422 unsigned long long orig_overflow, orig_flags, new_flags;
1423 struct hws_trailer_entry *te;
1425 te = aux_sdb_trailer(aux, alert_index);
1427 orig_flags = te->flags;
1428 *overflow = orig_overflow = te->overflow;
1429 if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1431 * SDB is already set by hardware.
1432 * Abort and try to set somewhere
1437 new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1438 } while (!cmpxchg_double(&te->flags, &te->overflow,
1439 orig_flags, orig_overflow,
1445 * aux_reset_buffer() - Scan and setup SDBs for new samples
1446 * @aux: The AUX buffer to set
1447 * @range: The range of SDBs to scan started from aux->head
1448 * @overflow: Set to overflow count
1450 * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1451 * marked as empty, check if it is already set full by the hardware sampler.
1452 * If yes, that means new data is already there before we can set an alert
1453 * indicator. Caller should try to set alert indicator to some position behind.
1455 * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1456 * previously and have already been consumed by user space. Reset these SDBs
1457 * (clear full indicator and alert indicator) for new data.
1458 * If aux->alert_mark fall in this area, just set it. Overflow count is
1459 * recorded while scanning.
1461 * SDBs between aux->head and aux->empty_mark are already reset at last time.
1462 * and ready for new samples. So scanning on this area could be skipped.
1464 * Return true if alert indicator is set successfully and false if not.
1466 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1467 unsigned long long *overflow)
1469 unsigned long long orig_overflow, orig_flags, new_flags;
1470 unsigned long i, range_scan, idx;
1471 struct hws_trailer_entry *te;
1473 if (range <= AUX_SDB_NUM_EMPTY(aux))
1475 * No need to scan. All SDBs in range are marked as empty.
1476 * Just set alert indicator. Should check race with hardware
1479 return aux_set_alert(aux, aux->alert_mark, overflow);
1481 if (aux->alert_mark <= aux->empty_mark)
1483 * Set alert indicator on empty SDB. Should check race
1484 * with hardware sampler.
1486 if (!aux_set_alert(aux, aux->alert_mark, overflow))
1490 * Scan the SDBs to clear full and alert indicator used previously.
1491 * Start scanning from one SDB behind empty_mark. If the new alert
1492 * indicator fall into this range, set it.
1494 range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1495 idx = aux->empty_mark + 1;
1496 for (i = 0; i < range_scan; i++, idx++) {
1497 te = aux_sdb_trailer(aux, idx);
1499 orig_flags = te->flags;
1500 orig_overflow = te->overflow;
1501 new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1502 if (idx == aux->alert_mark)
1503 new_flags |= SDB_TE_ALERT_REQ_MASK;
1505 new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1506 } while (!cmpxchg_double(&te->flags, &te->overflow,
1507 orig_flags, orig_overflow,
1509 *overflow += orig_overflow;
1512 /* Update empty_mark to new position */
1513 aux->empty_mark = aux->head + range - 1;
1519 * Measurement alert handler for diagnostic mode sampling.
1521 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1523 struct aux_buffer *aux;
1525 unsigned long range = 0, size;
1526 unsigned long long overflow = 0;
1527 struct perf_output_handle *handle = &cpuhw->handle;
1528 unsigned long num_sdb;
1530 aux = perf_get_aux(handle);
1531 if (WARN_ON_ONCE(!aux))
1534 /* Inform user space new data arrived */
1535 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1536 perf_aux_output_end(handle, size);
1537 num_sdb = aux->sfb.num_sdb;
1539 num_sdb = aux->sfb.num_sdb;
1541 /* Get an output handle */
1542 aux = perf_aux_output_begin(handle, cpuhw->event);
1543 if (handle->size == 0) {
1544 pr_err("The AUX buffer with %lu pages for the "
1545 "diagnostic-sampling mode is full\n",
1547 debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1550 if (WARN_ON_ONCE(!aux))
1553 /* Update head and alert_mark to new position */
1554 aux->head = handle->head >> PAGE_SHIFT;
1555 range = (handle->size + 1) >> PAGE_SHIFT;
1557 aux->alert_mark = aux->head;
1559 aux->alert_mark = aux->head + range/2 - 1;
1561 if (aux_reset_buffer(aux, range, &overflow)) {
1566 size = range << PAGE_SHIFT;
1567 perf_aux_output_end(&cpuhw->handle, size);
1568 pr_err("Sample data caused the AUX buffer with %lu "
1569 "pages to overflow\n", num_sdb);
1570 debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1572 aux->head, range, overflow);
1574 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1575 perf_aux_output_end(&cpuhw->handle, size);
1576 debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1577 "already full, try another\n",
1578 aux->head, aux->alert_mark);
1583 debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1584 "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1585 aux->head, aux->alert_mark, aux->empty_mark,
1586 AUX_SDB_NUM_ALERT(aux), range);
1590 * Callback when freeing AUX buffers.
1592 static void aux_buffer_free(void *data)
1594 struct aux_buffer *aux = data;
1595 unsigned long i, num_sdbt;
1600 /* Free SDBT. SDB is freed by the caller */
1601 num_sdbt = aux->sfb.num_sdbt;
1602 for (i = 0; i < num_sdbt; i++)
1603 free_page(aux->sdbt_index[i]);
1605 kfree(aux->sdbt_index);
1606 kfree(aux->sdb_index);
1609 debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1610 "%lu SDBTs\n", num_sdbt);
1613 static void aux_sdb_init(unsigned long sdb)
1615 struct hws_trailer_entry *te;
1617 te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1619 /* Save clock base */
1621 memcpy(&te->progusage2, &tod_clock_base[1], 8);
1625 * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1626 * @event: Event the buffer is setup for, event->cpu == -1 means current
1627 * @pages: Array of pointers to buffer pages passed from perf core
1628 * @nr_pages: Total pages
1629 * @snapshot: Flag for snapshot mode
1631 * This is the callback when setup an event using AUX buffer. Perf tool can
1632 * trigger this by an additional mmap() call on the event. Unlike the buffer
1633 * for basic samples, AUX buffer belongs to the event. It is scheduled with
1634 * the task among online cpus when it is a per-thread event.
1636 * Return the private AUX buffer structure if success or NULL if fails.
1638 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1639 int nr_pages, bool snapshot)
1641 struct sf_buffer *sfb;
1642 struct aux_buffer *aux;
1643 unsigned long *new, *tail;
1646 if (!nr_pages || !pages)
1649 if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1650 pr_err("AUX buffer size (%i pages) is larger than the "
1651 "maximum sampling buffer limit\n",
1654 } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1655 pr_err("AUX buffer size (%i pages) is less than the "
1656 "minimum sampling buffer limit\n",
1661 /* Allocate aux_buffer struct for the event */
1662 aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1667 /* Allocate sdbt_index for fast reference */
1668 n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1669 aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1670 if (!aux->sdbt_index)
1673 /* Allocate sdb_index for fast reference */
1674 aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1675 if (!aux->sdb_index)
1678 /* Allocate the first SDBT */
1680 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1683 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1684 tail = sfb->tail = sfb->sdbt;
1687 * Link the provided pages of AUX buffer to SDBT.
1688 * Allocate SDBT if needed.
1690 for (i = 0; i < nr_pages; i++, tail++) {
1691 if (require_table_link(tail)) {
1692 new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1695 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1696 /* Link current page to tail of chain */
1697 *tail = (unsigned long)(void *) new + 1;
1700 /* Tail is the entry in a SDBT */
1701 *tail = (unsigned long)pages[i];
1702 aux->sdb_index[i] = (unsigned long)pages[i];
1703 aux_sdb_init((unsigned long)pages[i]);
1705 sfb->num_sdb = nr_pages;
1707 /* Link the last entry in the SDBT to the first SDBT */
1708 *tail = (unsigned long) sfb->sdbt + 1;
1712 * Initial all SDBs are zeroed. Mark it as empty.
1713 * So there is no need to clear the full indicator
1714 * when this event is first added.
1716 aux->empty_mark = sfb->num_sdb - 1;
1718 debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1720 sfb->num_sdbt, sfb->num_sdb);
1725 /* SDBs (AUX buffer pages) are freed by caller */
1726 for (i = 0; i < sfb->num_sdbt; i++)
1727 free_page(aux->sdbt_index[i]);
1728 kfree(aux->sdb_index);
1730 kfree(aux->sdbt_index);
1737 static void cpumsf_pmu_read(struct perf_event *event)
1739 /* Nothing to do ... updates are interrupt-driven */
1742 /* Activate sampling control.
1743 * Next call of pmu_enable() starts sampling.
1745 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1747 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1749 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1752 if (flags & PERF_EF_RELOAD)
1753 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1755 perf_pmu_disable(event->pmu);
1756 event->hw.state = 0;
1757 cpuhw->lsctl.cs = 1;
1758 if (SAMPL_DIAG_MODE(&event->hw))
1759 cpuhw->lsctl.cd = 1;
1760 perf_pmu_enable(event->pmu);
1763 /* Deactivate sampling control.
1764 * Next call of pmu_enable() stops sampling.
1766 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1768 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1770 if (event->hw.state & PERF_HES_STOPPED)
1773 perf_pmu_disable(event->pmu);
1774 cpuhw->lsctl.cs = 0;
1775 cpuhw->lsctl.cd = 0;
1776 event->hw.state |= PERF_HES_STOPPED;
1778 if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1779 hw_perf_event_update(event, 1);
1780 event->hw.state |= PERF_HES_UPTODATE;
1782 perf_pmu_enable(event->pmu);
1785 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1787 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1788 struct aux_buffer *aux;
1791 if (cpuhw->flags & PMU_F_IN_USE)
1794 if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1798 perf_pmu_disable(event->pmu);
1800 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1802 /* Set up sampling controls. Always program the sampling register
1803 * using the SDB-table start. Reset TEAR_REG event hardware register
1804 * that is used by hw_perf_event_update() to store the sampling buffer
1805 * position after samples have been flushed.
1809 cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1810 if (!SAMPL_DIAG_MODE(&event->hw)) {
1811 cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1812 cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1813 hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1816 /* Ensure sampling functions are in the disabled state. If disabled,
1817 * switch on sampling enable control. */
1818 if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1822 if (SAMPL_DIAG_MODE(&event->hw)) {
1823 aux = perf_aux_output_begin(&cpuhw->handle, event);
1828 err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1831 cpuhw->lsctl.ed = 1;
1833 cpuhw->lsctl.es = 1;
1835 /* Set in_use flag and store event */
1836 cpuhw->event = event;
1837 cpuhw->flags |= PMU_F_IN_USE;
1839 if (flags & PERF_EF_START)
1840 cpumsf_pmu_start(event, PERF_EF_RELOAD);
1842 perf_event_update_userpage(event);
1843 perf_pmu_enable(event->pmu);
1847 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1849 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1851 perf_pmu_disable(event->pmu);
1852 cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1854 cpuhw->lsctl.es = 0;
1855 cpuhw->lsctl.ed = 0;
1856 cpuhw->flags &= ~PMU_F_IN_USE;
1857 cpuhw->event = NULL;
1859 if (SAMPL_DIAG_MODE(&event->hw))
1860 aux_output_end(&cpuhw->handle);
1861 perf_event_update_userpage(event);
1862 perf_pmu_enable(event->pmu);
1865 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1866 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1868 static struct attribute *cpumsf_pmu_events_attr[] = {
1869 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
1874 PMU_FORMAT_ATTR(event, "config:0-63");
1876 static struct attribute *cpumsf_pmu_format_attr[] = {
1877 &format_attr_event.attr,
1881 static struct attribute_group cpumsf_pmu_events_group = {
1883 .attrs = cpumsf_pmu_events_attr,
1885 static struct attribute_group cpumsf_pmu_format_group = {
1887 .attrs = cpumsf_pmu_format_attr,
1889 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1890 &cpumsf_pmu_events_group,
1891 &cpumsf_pmu_format_group,
1895 static struct pmu cpumf_sampling = {
1896 .pmu_enable = cpumsf_pmu_enable,
1897 .pmu_disable = cpumsf_pmu_disable,
1899 .event_init = cpumsf_pmu_event_init,
1900 .add = cpumsf_pmu_add,
1901 .del = cpumsf_pmu_del,
1903 .start = cpumsf_pmu_start,
1904 .stop = cpumsf_pmu_stop,
1905 .read = cpumsf_pmu_read,
1907 .attr_groups = cpumsf_pmu_attr_groups,
1909 .setup_aux = aux_buffer_setup,
1910 .free_aux = aux_buffer_free,
1913 static void cpumf_measurement_alert(struct ext_code ext_code,
1914 unsigned int alert, unsigned long unused)
1916 struct cpu_hw_sf *cpuhw;
1918 if (!(alert & CPU_MF_INT_SF_MASK))
1920 inc_irq_stat(IRQEXT_CMS);
1921 cpuhw = this_cpu_ptr(&cpu_hw_sf);
1923 /* Measurement alerts are shared and might happen when the PMU
1924 * is not reserved. Ignore these alerts in this case. */
1925 if (!(cpuhw->flags & PMU_F_RESERVED))
1928 /* The processing below must take care of multiple alert events that
1929 * might be indicated concurrently. */
1931 /* Program alert request */
1932 if (alert & CPU_MF_INT_SF_PRA) {
1933 if (cpuhw->flags & PMU_F_IN_USE)
1934 if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
1935 hw_collect_aux(cpuhw);
1937 hw_perf_event_update(cpuhw->event, 0);
1939 WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
1942 /* Report measurement alerts only for non-PRA codes */
1943 if (alert != CPU_MF_INT_SF_PRA)
1944 debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
1946 /* Sampling authorization change request */
1947 if (alert & CPU_MF_INT_SF_SACA)
1950 /* Loss of sample data due to high-priority machine activities */
1951 if (alert & CPU_MF_INT_SF_LSDA) {
1952 pr_err("Sample data was lost\n");
1953 cpuhw->flags |= PMU_F_ERR_LSDA;
1957 /* Invalid sampling buffer entry */
1958 if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
1959 pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
1961 cpuhw->flags |= PMU_F_ERR_IBE;
1965 static int cpusf_pmu_setup(unsigned int cpu, int flags)
1967 /* Ignore the notification if no events are scheduled on the PMU.
1968 * This might be racy...
1970 if (!atomic_read(&num_events))
1973 local_irq_disable();
1974 setup_pmc_cpu(&flags);
1979 static int s390_pmu_sf_online_cpu(unsigned int cpu)
1981 return cpusf_pmu_setup(cpu, PMC_INIT);
1984 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
1986 return cpusf_pmu_setup(cpu, PMC_RELEASE);
1989 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
1991 if (!cpum_sf_avail())
1993 return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
1996 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
1999 unsigned long min, max;
2001 if (!cpum_sf_avail())
2003 if (!val || !strlen(val))
2006 /* Valid parameter values: "min,max" or "max" */
2007 min = CPUM_SF_MIN_SDB;
2008 max = CPUM_SF_MAX_SDB;
2009 if (strchr(val, ','))
2010 rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2012 rc = kstrtoul(val, 10, &max);
2014 if (min < 2 || min >= max || max > get_num_physpages())
2019 sfb_set_limits(min, max);
2020 pr_info("The sampling buffer limits have changed to: "
2021 "min=%lu max=%lu (diag=x%lu)\n",
2022 CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2026 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2027 static const struct kernel_param_ops param_ops_sfb_size = {
2028 .set = param_set_sfb_size,
2029 .get = param_get_sfb_size,
2032 #define RS_INIT_FAILURE_QSI 0x0001
2033 #define RS_INIT_FAILURE_BSDES 0x0002
2034 #define RS_INIT_FAILURE_ALRT 0x0003
2035 #define RS_INIT_FAILURE_PERF 0x0004
2036 static void __init pr_cpumsf_err(unsigned int reason)
2038 pr_err("Sampling facility support for perf is not available: "
2039 "reason=%04x\n", reason);
2042 static int __init init_cpum_sampling_pmu(void)
2044 struct hws_qsi_info_block si;
2047 if (!cpum_sf_avail())
2050 memset(&si, 0, sizeof(si));
2052 pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2056 if (!si.as && !si.ad)
2059 if (si.bsdes != sizeof(struct hws_basic_entry)) {
2060 pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2065 sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2066 cpumsf_pmu_events_attr[1] =
2067 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2070 sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2072 pr_err("Registering for s390dbf failed\n");
2075 debug_register_view(sfdbg, &debug_sprintf_view);
2077 err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2078 cpumf_measurement_alert);
2080 pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2081 debug_unregister(sfdbg);
2085 err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2087 pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2088 unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2089 cpumf_measurement_alert);
2090 debug_unregister(sfdbg);
2094 cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2095 s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2099 arch_initcall(init_cpum_sampling_pmu);
2100 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644);