xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision 680ef72a)
1 /*
2  * Performance event support for the System z CPU-measurement Sampling Facility
3  *
4  * Copyright IBM Corp. 2013
5  * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License (version 2 only)
9  * as published by the Free Software Foundation.
10  */
11 #define KMSG_COMPONENT	"cpum_sf"
12 #define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt
13 
14 #include <linux/kernel.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/perf_event.h>
17 #include <linux/percpu.h>
18 #include <linux/pid.h>
19 #include <linux/notifier.h>
20 #include <linux/export.h>
21 #include <linux/slab.h>
22 #include <linux/mm.h>
23 #include <linux/moduleparam.h>
24 #include <asm/cpu_mf.h>
25 #include <asm/irq.h>
26 #include <asm/debug.h>
27 #include <asm/timex.h>
28 
29 /* Minimum number of sample-data-block-tables:
30  * At least one table is required for the sampling buffer structure.
31  * A single table contains up to 511 pointers to sample-data-blocks.
32  */
33 #define CPUM_SF_MIN_SDBT	1
34 
35 /* Number of sample-data-blocks per sample-data-block-table (SDBT):
36  * A table contains SDB pointers (8 bytes) and one table-link entry
37  * that points to the origin of the next SDBT.
38  */
39 #define CPUM_SF_SDB_PER_TABLE	((PAGE_SIZE - 8) / 8)
40 
41 /* Maximum page offset for an SDBT table-link entry:
42  * If this page offset is reached, a table-link entry to the next SDBT
43  * must be added.
44  */
45 #define CPUM_SF_SDBT_TL_OFFSET	(CPUM_SF_SDB_PER_TABLE * 8)
46 static inline int require_table_link(const void *sdbt)
47 {
48 	return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
49 }
50 
51 /* Minimum and maximum sampling buffer sizes:
52  *
53  * This number represents the maximum size of the sampling buffer taking
54  * the number of sample-data-block-tables into account.  Note that these
55  * numbers apply to the basic-sampling function only.
56  * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if
57  * the diagnostic-sampling function is active.
58  *
59  * Sampling buffer size		Buffer characteristics
60  * ---------------------------------------------------
61  *	 64KB		    ==	  16 pages (4KB per page)
62  *				   1 page  for SDB-tables
63  *				  15 pages for SDBs
64  *
65  *  32MB		    ==	8192 pages (4KB per page)
66  *				  16 pages for SDB-tables
67  *				8176 pages for SDBs
68  */
69 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15;
70 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176;
71 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
72 
73 struct sf_buffer {
74 	unsigned long	 *sdbt;	    /* Sample-data-block-table origin */
75 	/* buffer characteristics (required for buffer increments) */
76 	unsigned long  num_sdb;	    /* Number of sample-data-blocks */
77 	unsigned long num_sdbt;	    /* Number of sample-data-block-tables */
78 	unsigned long	 *tail;	    /* last sample-data-block-table */
79 };
80 
81 struct aux_buffer {
82 	struct sf_buffer sfb;
83 	unsigned long head;	   /* index of SDB of buffer head */
84 	unsigned long alert_mark;  /* index of SDB of alert request position */
85 	unsigned long empty_mark;  /* mark of SDB not marked full */
86 	unsigned long *sdb_index;  /* SDB address for fast lookup */
87 	unsigned long *sdbt_index; /* SDBT address for fast lookup */
88 };
89 
90 struct cpu_hw_sf {
91 	/* CPU-measurement sampling information block */
92 	struct hws_qsi_info_block qsi;
93 	/* CPU-measurement sampling control block */
94 	struct hws_lsctl_request_block lsctl;
95 	struct sf_buffer sfb;	    /* Sampling buffer */
96 	unsigned int flags;	    /* Status flags */
97 	struct perf_event *event;   /* Scheduled perf event */
98 	struct perf_output_handle handle; /* AUX buffer output handle */
99 };
100 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
101 
102 /* Debug feature */
103 static debug_info_t *sfdbg;
104 
105 /*
106  * sf_disable() - Switch off sampling facility
107  */
108 static int sf_disable(void)
109 {
110 	struct hws_lsctl_request_block sreq;
111 
112 	memset(&sreq, 0, sizeof(sreq));
113 	return lsctl(&sreq);
114 }
115 
116 /*
117  * sf_buffer_available() - Check for an allocated sampling buffer
118  */
119 static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
120 {
121 	return !!cpuhw->sfb.sdbt;
122 }
123 
124 /*
125  * deallocate sampling facility buffer
126  */
127 static void free_sampling_buffer(struct sf_buffer *sfb)
128 {
129 	unsigned long *sdbt, *curr;
130 
131 	if (!sfb->sdbt)
132 		return;
133 
134 	sdbt = sfb->sdbt;
135 	curr = sdbt;
136 
137 	/* Free the SDBT after all SDBs are processed... */
138 	while (1) {
139 		if (!*curr || !sdbt)
140 			break;
141 
142 		/* Process table-link entries */
143 		if (is_link_entry(curr)) {
144 			curr = get_next_sdbt(curr);
145 			if (sdbt)
146 				free_page((unsigned long) sdbt);
147 
148 			/* If the origin is reached, sampling buffer is freed */
149 			if (curr == sfb->sdbt)
150 				break;
151 			else
152 				sdbt = curr;
153 		} else {
154 			/* Process SDB pointer */
155 			if (*curr) {
156 				free_page(*curr);
157 				curr++;
158 			}
159 		}
160 	}
161 
162 	debug_sprintf_event(sfdbg, 5,
163 			    "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
164 	memset(sfb, 0, sizeof(*sfb));
165 }
166 
167 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
168 {
169 	unsigned long sdb, *trailer;
170 
171 	/* Allocate and initialize sample-data-block */
172 	sdb = get_zeroed_page(gfp_flags);
173 	if (!sdb)
174 		return -ENOMEM;
175 	trailer = trailer_entry_ptr(sdb);
176 	*trailer = SDB_TE_ALERT_REQ_MASK;
177 
178 	/* Link SDB into the sample-data-block-table */
179 	*sdbt = sdb;
180 
181 	return 0;
182 }
183 
184 /*
185  * realloc_sampling_buffer() - extend sampler memory
186  *
187  * Allocates new sample-data-blocks and adds them to the specified sampling
188  * buffer memory.
189  *
190  * Important: This modifies the sampling buffer and must be called when the
191  *	      sampling facility is disabled.
192  *
193  * Returns zero on success, non-zero otherwise.
194  */
195 static int realloc_sampling_buffer(struct sf_buffer *sfb,
196 				   unsigned long num_sdb, gfp_t gfp_flags)
197 {
198 	int i, rc;
199 	unsigned long *new, *tail;
200 
201 	if (!sfb->sdbt || !sfb->tail)
202 		return -EINVAL;
203 
204 	if (!is_link_entry(sfb->tail))
205 		return -EINVAL;
206 
207 	/* Append to the existing sampling buffer, overwriting the table-link
208 	 * register.
209 	 * The tail variables always points to the "tail" (last and table-link)
210 	 * entry in an SDB-table.
211 	 */
212 	tail = sfb->tail;
213 
214 	/* Do a sanity check whether the table-link entry points to
215 	 * the sampling buffer origin.
216 	 */
217 	if (sfb->sdbt != get_next_sdbt(tail)) {
218 		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
219 				    "sampling buffer is not linked: origin=%p"
220 				    "tail=%p\n",
221 				    (void *) sfb->sdbt, (void *) tail);
222 		return -EINVAL;
223 	}
224 
225 	/* Allocate remaining SDBs */
226 	rc = 0;
227 	for (i = 0; i < num_sdb; i++) {
228 		/* Allocate a new SDB-table if it is full. */
229 		if (require_table_link(tail)) {
230 			new = (unsigned long *) get_zeroed_page(gfp_flags);
231 			if (!new) {
232 				rc = -ENOMEM;
233 				break;
234 			}
235 			sfb->num_sdbt++;
236 			/* Link current page to tail of chain */
237 			*tail = (unsigned long)(void *) new + 1;
238 			tail = new;
239 		}
240 
241 		/* Allocate a new sample-data-block.
242 		 * If there is not enough memory, stop the realloc process
243 		 * and simply use what was allocated.  If this is a temporary
244 		 * issue, a new realloc call (if required) might succeed.
245 		 */
246 		rc = alloc_sample_data_block(tail, gfp_flags);
247 		if (rc)
248 			break;
249 		sfb->num_sdb++;
250 		tail++;
251 	}
252 
253 	/* Link sampling buffer to its origin */
254 	*tail = (unsigned long) sfb->sdbt + 1;
255 	sfb->tail = tail;
256 
257 	debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
258 			    " settings: sdbt=%lu sdb=%lu\n",
259 			    sfb->num_sdbt, sfb->num_sdb);
260 	return rc;
261 }
262 
263 /*
264  * allocate_sampling_buffer() - allocate sampler memory
265  *
266  * Allocates and initializes a sampling buffer structure using the
267  * specified number of sample-data-blocks (SDB).  For each allocation,
268  * a 4K page is used.  The number of sample-data-block-tables (SDBT)
269  * are calculated from SDBs.
270  * Also set the ALERT_REQ mask in each SDBs trailer.
271  *
272  * Returns zero on success, non-zero otherwise.
273  */
274 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
275 {
276 	int rc;
277 
278 	if (sfb->sdbt)
279 		return -EINVAL;
280 
281 	/* Allocate the sample-data-block-table origin */
282 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
283 	if (!sfb->sdbt)
284 		return -ENOMEM;
285 	sfb->num_sdb = 0;
286 	sfb->num_sdbt = 1;
287 
288 	/* Link the table origin to point to itself to prepare for
289 	 * realloc_sampling_buffer() invocation.
290 	 */
291 	sfb->tail = sfb->sdbt;
292 	*sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
293 
294 	/* Allocate requested number of sample-data-blocks */
295 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
296 	if (rc) {
297 		free_sampling_buffer(sfb);
298 		debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
299 			"realloc_sampling_buffer failed with rc=%i\n", rc);
300 	} else
301 		debug_sprintf_event(sfdbg, 4,
302 			"alloc_sampling_buffer: tear=%p dear=%p\n",
303 			sfb->sdbt, (void *) *sfb->sdbt);
304 	return rc;
305 }
306 
307 static void sfb_set_limits(unsigned long min, unsigned long max)
308 {
309 	struct hws_qsi_info_block si;
310 
311 	CPUM_SF_MIN_SDB = min;
312 	CPUM_SF_MAX_SDB = max;
313 
314 	memset(&si, 0, sizeof(si));
315 	if (!qsi(&si))
316 		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
317 }
318 
319 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
320 {
321 	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
322 				    : CPUM_SF_MAX_SDB;
323 }
324 
325 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
326 					struct hw_perf_event *hwc)
327 {
328 	if (!sfb->sdbt)
329 		return SFB_ALLOC_REG(hwc);
330 	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
331 		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
332 	return 0;
333 }
334 
335 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
336 				   struct hw_perf_event *hwc)
337 {
338 	return sfb_pending_allocs(sfb, hwc) > 0;
339 }
340 
341 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
342 {
343 	/* Limit the number of SDBs to not exceed the maximum */
344 	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
345 	if (num)
346 		SFB_ALLOC_REG(hwc) += num;
347 }
348 
349 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
350 {
351 	SFB_ALLOC_REG(hwc) = 0;
352 	sfb_account_allocs(num, hwc);
353 }
354 
355 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
356 {
357 	if (cpuhw->sfb.sdbt)
358 		free_sampling_buffer(&cpuhw->sfb);
359 }
360 
361 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
362 {
363 	unsigned long n_sdb, freq, factor;
364 	size_t sample_size;
365 
366 	/* Calculate sampling buffers using 4K pages
367 	 *
368 	 *    1. Determine the sample data size which depends on the used
369 	 *	 sampling functions, for example, basic-sampling or
370 	 *	 basic-sampling with diagnostic-sampling.
371 	 *
372 	 *    2. Use the sampling frequency as input.  The sampling buffer is
373 	 *	 designed for almost one second.  This can be adjusted through
374 	 *	 the "factor" variable.
375 	 *	 In any case, alloc_sampling_buffer() sets the Alert Request
376 	 *	 Control indicator to trigger a measurement-alert to harvest
377 	 *	 sample-data-blocks (sdb).
378 	 *
379 	 *    3. Compute the number of sample-data-blocks and ensure a minimum
380 	 *	 of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
381 	 *	 exceed a "calculated" maximum.  The symbolic maximum is
382 	 *	 designed for basic-sampling only and needs to be increased if
383 	 *	 diagnostic-sampling is active.
384 	 *	 See also the remarks for these symbolic constants.
385 	 *
386 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
387 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
388 	 *	 to 511 SDBs).
389 	 */
390 	sample_size = sizeof(struct hws_basic_entry);
391 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
392 	factor = 1;
393 	n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
394 	if (n_sdb < CPUM_SF_MIN_SDB)
395 		n_sdb = CPUM_SF_MIN_SDB;
396 
397 	/* If there is already a sampling buffer allocated, it is very likely
398 	 * that the sampling facility is enabled too.  If the event to be
399 	 * initialized requires a greater sampling buffer, the allocation must
400 	 * be postponed.  Changing the sampling buffer requires the sampling
401 	 * facility to be in the disabled state.  So, account the number of
402 	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
403 	 * before the event is started.
404 	 */
405 	sfb_init_allocs(n_sdb, hwc);
406 	if (sf_buffer_available(cpuhw))
407 		return 0;
408 
409 	debug_sprintf_event(sfdbg, 3,
410 			    "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
411 			    " sample_size=%lu cpuhw=%p\n",
412 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
413 			    sample_size, cpuhw);
414 
415 	return alloc_sampling_buffer(&cpuhw->sfb,
416 				     sfb_pending_allocs(&cpuhw->sfb, hwc));
417 }
418 
419 static unsigned long min_percent(unsigned int percent, unsigned long base,
420 				 unsigned long min)
421 {
422 	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
423 }
424 
425 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
426 {
427 	/* Use a percentage-based approach to extend the sampling facility
428 	 * buffer.  Accept up to 5% sample data loss.
429 	 * Vary the extents between 1% to 5% of the current number of
430 	 * sample-data-blocks.
431 	 */
432 	if (ratio <= 5)
433 		return 0;
434 	if (ratio <= 25)
435 		return min_percent(1, base, 1);
436 	if (ratio <= 50)
437 		return min_percent(1, base, 1);
438 	if (ratio <= 75)
439 		return min_percent(2, base, 2);
440 	if (ratio <= 100)
441 		return min_percent(3, base, 3);
442 	if (ratio <= 250)
443 		return min_percent(4, base, 4);
444 
445 	return min_percent(5, base, 8);
446 }
447 
448 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
449 				  struct hw_perf_event *hwc)
450 {
451 	unsigned long ratio, num;
452 
453 	if (!OVERFLOW_REG(hwc))
454 		return;
455 
456 	/* The sample_overflow contains the average number of sample data
457 	 * that has been lost because sample-data-blocks were full.
458 	 *
459 	 * Calculate the total number of sample data entries that has been
460 	 * discarded.  Then calculate the ratio of lost samples to total samples
461 	 * per second in percent.
462 	 */
463 	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
464 			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
465 
466 	/* Compute number of sample-data-blocks */
467 	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
468 	if (num)
469 		sfb_account_allocs(num, hwc);
470 
471 	debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
472 			    " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
473 	OVERFLOW_REG(hwc) = 0;
474 }
475 
476 /* extend_sampling_buffer() - Extend sampling buffer
477  * @sfb:	Sampling buffer structure (for local CPU)
478  * @hwc:	Perf event hardware structure
479  *
480  * Use this function to extend the sampling buffer based on the overflow counter
481  * and postponed allocation extents stored in the specified Perf event hardware.
482  *
483  * Important: This function disables the sampling facility in order to safely
484  *	      change the sampling buffer structure.  Do not call this function
485  *	      when the PMU is active.
486  */
487 static void extend_sampling_buffer(struct sf_buffer *sfb,
488 				   struct hw_perf_event *hwc)
489 {
490 	unsigned long num, num_old;
491 	int rc;
492 
493 	num = sfb_pending_allocs(sfb, hwc);
494 	if (!num)
495 		return;
496 	num_old = sfb->num_sdb;
497 
498 	/* Disable the sampling facility to reset any states and also
499 	 * clear pending measurement alerts.
500 	 */
501 	sf_disable();
502 
503 	/* Extend the sampling buffer.
504 	 * This memory allocation typically happens in an atomic context when
505 	 * called by perf.  Because this is a reallocation, it is fine if the
506 	 * new SDB-request cannot be satisfied immediately.
507 	 */
508 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
509 	if (rc)
510 		debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
511 				    "failed with rc=%i\n", rc);
512 
513 	if (sfb_has_pending_allocs(sfb, hwc))
514 		debug_sprintf_event(sfdbg, 5, "sfb: extend: "
515 				    "req=%lu alloc=%lu remaining=%lu\n",
516 				    num, sfb->num_sdb - num_old,
517 				    sfb_pending_allocs(sfb, hwc));
518 }
519 
520 
521 /* Number of perf events counting hardware events */
522 static atomic_t num_events;
523 /* Used to avoid races in calling reserve/release_cpumf_hardware */
524 static DEFINE_MUTEX(pmc_reserve_mutex);
525 
526 #define PMC_INIT      0
527 #define PMC_RELEASE   1
528 #define PMC_FAILURE   2
529 static void setup_pmc_cpu(void *flags)
530 {
531 	int err;
532 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
533 
534 	err = 0;
535 	switch (*((int *) flags)) {
536 	case PMC_INIT:
537 		memset(cpusf, 0, sizeof(*cpusf));
538 		err = qsi(&cpusf->qsi);
539 		if (err)
540 			break;
541 		cpusf->flags |= PMU_F_RESERVED;
542 		err = sf_disable();
543 		if (err)
544 			pr_err("Switching off the sampling facility failed "
545 			       "with rc=%i\n", err);
546 		debug_sprintf_event(sfdbg, 5,
547 				    "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
548 		break;
549 	case PMC_RELEASE:
550 		cpusf->flags &= ~PMU_F_RESERVED;
551 		err = sf_disable();
552 		if (err) {
553 			pr_err("Switching off the sampling facility failed "
554 			       "with rc=%i\n", err);
555 		} else
556 			deallocate_buffers(cpusf);
557 		debug_sprintf_event(sfdbg, 5,
558 				    "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
559 		break;
560 	}
561 	if (err)
562 		*((int *) flags) |= PMC_FAILURE;
563 }
564 
565 static void release_pmc_hardware(void)
566 {
567 	int flags = PMC_RELEASE;
568 
569 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
570 	on_each_cpu(setup_pmc_cpu, &flags, 1);
571 }
572 
573 static int reserve_pmc_hardware(void)
574 {
575 	int flags = PMC_INIT;
576 
577 	on_each_cpu(setup_pmc_cpu, &flags, 1);
578 	if (flags & PMC_FAILURE) {
579 		release_pmc_hardware();
580 		return -ENODEV;
581 	}
582 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
583 
584 	return 0;
585 }
586 
587 static void hw_perf_event_destroy(struct perf_event *event)
588 {
589 	/* Release PMC if this is the last perf event */
590 	if (!atomic_add_unless(&num_events, -1, 1)) {
591 		mutex_lock(&pmc_reserve_mutex);
592 		if (atomic_dec_return(&num_events) == 0)
593 			release_pmc_hardware();
594 		mutex_unlock(&pmc_reserve_mutex);
595 	}
596 }
597 
598 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
599 {
600 	hwc->sample_period = period;
601 	hwc->last_period = hwc->sample_period;
602 	local64_set(&hwc->period_left, hwc->sample_period);
603 }
604 
605 static void hw_reset_registers(struct hw_perf_event *hwc,
606 			       unsigned long *sdbt_origin)
607 {
608 	/* (Re)set to first sample-data-block-table */
609 	TEAR_REG(hwc) = (unsigned long) sdbt_origin;
610 }
611 
612 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
613 				   unsigned long rate)
614 {
615 	return clamp_t(unsigned long, rate,
616 		       si->min_sampl_rate, si->max_sampl_rate);
617 }
618 
619 static u32 cpumsf_pid_type(struct perf_event *event,
620 			   u32 pid, enum pid_type type)
621 {
622 	struct task_struct *tsk;
623 
624 	/* Idle process */
625 	if (!pid)
626 		goto out;
627 
628 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
629 	pid = -1;
630 	if (tsk) {
631 		/*
632 		 * Only top level events contain the pid namespace in which
633 		 * they are created.
634 		 */
635 		if (event->parent)
636 			event = event->parent;
637 		pid = __task_pid_nr_ns(tsk, type, event->ns);
638 		/*
639 		 * See also 1d953111b648
640 		 * "perf/core: Don't report zero PIDs for exiting tasks".
641 		 */
642 		if (!pid && !pid_alive(tsk))
643 			pid = -1;
644 	}
645 out:
646 	return pid;
647 }
648 
649 static void cpumsf_output_event_pid(struct perf_event *event,
650 				    struct perf_sample_data *data,
651 				    struct pt_regs *regs)
652 {
653 	u32 pid;
654 	struct perf_event_header header;
655 	struct perf_output_handle handle;
656 
657 	/*
658 	 * Obtain the PID from the basic-sampling data entry and
659 	 * correct the data->tid_entry.pid value.
660 	 */
661 	pid = data->tid_entry.pid;
662 
663 	/* Protect callchain buffers, tasks */
664 	rcu_read_lock();
665 
666 	perf_prepare_sample(&header, data, event, regs);
667 	if (perf_output_begin(&handle, event, header.size))
668 		goto out;
669 
670 	/* Update the process ID (see also kernel/events/core.c) */
671 	data->tid_entry.pid = cpumsf_pid_type(event, pid, __PIDTYPE_TGID);
672 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
673 
674 	perf_output_sample(&handle, &header, data, event);
675 	perf_output_end(&handle);
676 out:
677 	rcu_read_unlock();
678 }
679 
680 static int __hw_perf_event_init(struct perf_event *event)
681 {
682 	struct cpu_hw_sf *cpuhw;
683 	struct hws_qsi_info_block si;
684 	struct perf_event_attr *attr = &event->attr;
685 	struct hw_perf_event *hwc = &event->hw;
686 	unsigned long rate;
687 	int cpu, err;
688 
689 	/* Reserve CPU-measurement sampling facility */
690 	err = 0;
691 	if (!atomic_inc_not_zero(&num_events)) {
692 		mutex_lock(&pmc_reserve_mutex);
693 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
694 			err = -EBUSY;
695 		else
696 			atomic_inc(&num_events);
697 		mutex_unlock(&pmc_reserve_mutex);
698 	}
699 	event->destroy = hw_perf_event_destroy;
700 
701 	if (err)
702 		goto out;
703 
704 	/* Access per-CPU sampling information (query sampling info) */
705 	/*
706 	 * The event->cpu value can be -1 to count on every CPU, for example,
707 	 * when attaching to a task.  If this is specified, use the query
708 	 * sampling info from the current CPU, otherwise use event->cpu to
709 	 * retrieve the per-CPU information.
710 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
711 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
712 	 */
713 	memset(&si, 0, sizeof(si));
714 	cpuhw = NULL;
715 	if (event->cpu == -1)
716 		qsi(&si);
717 	else {
718 		/* Event is pinned to a particular CPU, retrieve the per-CPU
719 		 * sampling structure for accessing the CPU-specific QSI.
720 		 */
721 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
722 		si = cpuhw->qsi;
723 	}
724 
725 	/* Check sampling facility authorization and, if not authorized,
726 	 * fall back to other PMUs.  It is safe to check any CPU because
727 	 * the authorization is identical for all configured CPUs.
728 	 */
729 	if (!si.as) {
730 		err = -ENOENT;
731 		goto out;
732 	}
733 
734 	/* Always enable basic sampling */
735 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
736 
737 	/* Check if diagnostic sampling is requested.  Deny if the required
738 	 * sampling authorization is missing.
739 	 */
740 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
741 		if (!si.ad) {
742 			err = -EPERM;
743 			goto out;
744 		}
745 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
746 	}
747 
748 	/* Check and set other sampling flags */
749 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
750 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
751 
752 	/* The sampling information (si) contains information about the
753 	 * min/max sampling intervals and the CPU speed.  So calculate the
754 	 * correct sampling interval and avoid the whole period adjust
755 	 * feedback loop.
756 	 */
757 	rate = 0;
758 	if (attr->freq) {
759 		rate = freq_to_sample_rate(&si, attr->sample_freq);
760 		rate = hw_limit_rate(&si, rate);
761 		attr->freq = 0;
762 		attr->sample_period = rate;
763 	} else {
764 		/* The min/max sampling rates specifies the valid range
765 		 * of sample periods.  If the specified sample period is
766 		 * out of range, limit the period to the range boundary.
767 		 */
768 		rate = hw_limit_rate(&si, hwc->sample_period);
769 
770 		/* The perf core maintains a maximum sample rate that is
771 		 * configurable through the sysctl interface.  Ensure the
772 		 * sampling rate does not exceed this value.  This also helps
773 		 * to avoid throttling when pushing samples with
774 		 * perf_event_overflow().
775 		 */
776 		if (sample_rate_to_freq(&si, rate) >
777 		      sysctl_perf_event_sample_rate) {
778 			err = -EINVAL;
779 			debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
780 			goto out;
781 		}
782 	}
783 	SAMPL_RATE(hwc) = rate;
784 	hw_init_period(hwc, SAMPL_RATE(hwc));
785 
786 	/* Initialize sample data overflow accounting */
787 	hwc->extra_reg.reg = REG_OVERFLOW;
788 	OVERFLOW_REG(hwc) = 0;
789 
790 	/* Use AUX buffer. No need to allocate it by ourself */
791 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
792 		return 0;
793 
794 	/* Allocate the per-CPU sampling buffer using the CPU information
795 	 * from the event.  If the event is not pinned to a particular
796 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
797 	 * buffers for each online CPU.
798 	 */
799 	if (cpuhw)
800 		/* Event is pinned to a particular CPU */
801 		err = allocate_buffers(cpuhw, hwc);
802 	else {
803 		/* Event is not pinned, allocate sampling buffer on
804 		 * each online CPU
805 		 */
806 		for_each_online_cpu(cpu) {
807 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
808 			err = allocate_buffers(cpuhw, hwc);
809 			if (err)
810 				break;
811 		}
812 	}
813 
814 	/* If PID/TID sampling is active, replace the default overflow
815 	 * handler to extract and resolve the PIDs from the basic-sampling
816 	 * data entries.
817 	 */
818 	if (event->attr.sample_type & PERF_SAMPLE_TID)
819 		if (is_default_overflow_handler(event))
820 			event->overflow_handler = cpumsf_output_event_pid;
821 out:
822 	return err;
823 }
824 
825 static int cpumsf_pmu_event_init(struct perf_event *event)
826 {
827 	int err;
828 
829 	/* No support for taken branch sampling */
830 	if (has_branch_stack(event))
831 		return -EOPNOTSUPP;
832 
833 	switch (event->attr.type) {
834 	case PERF_TYPE_RAW:
835 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
836 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
837 			return -ENOENT;
838 		break;
839 	case PERF_TYPE_HARDWARE:
840 		/* Support sampling of CPU cycles in addition to the
841 		 * counter facility.  However, the counter facility
842 		 * is more precise and, hence, restrict this PMU to
843 		 * sampling events only.
844 		 */
845 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
846 			return -ENOENT;
847 		if (!is_sampling_event(event))
848 			return -ENOENT;
849 		break;
850 	default:
851 		return -ENOENT;
852 	}
853 
854 	/* Check online status of the CPU to which the event is pinned */
855 	if (event->cpu >= 0 && !cpu_online(event->cpu))
856 			return -ENODEV;
857 
858 	/* Force reset of idle/hv excludes regardless of what the
859 	 * user requested.
860 	 */
861 	if (event->attr.exclude_hv)
862 		event->attr.exclude_hv = 0;
863 	if (event->attr.exclude_idle)
864 		event->attr.exclude_idle = 0;
865 
866 	err = __hw_perf_event_init(event);
867 	if (unlikely(err))
868 		if (event->destroy)
869 			event->destroy(event);
870 	return err;
871 }
872 
873 static void cpumsf_pmu_enable(struct pmu *pmu)
874 {
875 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
876 	struct hw_perf_event *hwc;
877 	int err;
878 
879 	if (cpuhw->flags & PMU_F_ENABLED)
880 		return;
881 
882 	if (cpuhw->flags & PMU_F_ERR_MASK)
883 		return;
884 
885 	/* Check whether to extent the sampling buffer.
886 	 *
887 	 * Two conditions trigger an increase of the sampling buffer for a
888 	 * perf event:
889 	 *    1. Postponed buffer allocations from the event initialization.
890 	 *    2. Sampling overflows that contribute to pending allocations.
891 	 *
892 	 * Note that the extend_sampling_buffer() function disables the sampling
893 	 * facility, but it can be fully re-enabled using sampling controls that
894 	 * have been saved in cpumsf_pmu_disable().
895 	 */
896 	if (cpuhw->event) {
897 		hwc = &cpuhw->event->hw;
898 		if (!(SAMPL_DIAG_MODE(hwc))) {
899 			/*
900 			 * Account number of overflow-designated
901 			 * buffer extents
902 			 */
903 			sfb_account_overflows(cpuhw, hwc);
904 			if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
905 				extend_sampling_buffer(&cpuhw->sfb, hwc);
906 		}
907 	}
908 
909 	/* (Re)enable the PMU and sampling facility */
910 	cpuhw->flags |= PMU_F_ENABLED;
911 	barrier();
912 
913 	err = lsctl(&cpuhw->lsctl);
914 	if (err) {
915 		cpuhw->flags &= ~PMU_F_ENABLED;
916 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
917 			1, err);
918 		return;
919 	}
920 
921 	/* Load current program parameter */
922 	lpp(&S390_lowcore.lpp);
923 
924 	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
925 			    "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
926 			    cpuhw->lsctl.ed, cpuhw->lsctl.cd,
927 			    (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
928 }
929 
930 static void cpumsf_pmu_disable(struct pmu *pmu)
931 {
932 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
933 	struct hws_lsctl_request_block inactive;
934 	struct hws_qsi_info_block si;
935 	int err;
936 
937 	if (!(cpuhw->flags & PMU_F_ENABLED))
938 		return;
939 
940 	if (cpuhw->flags & PMU_F_ERR_MASK)
941 		return;
942 
943 	/* Switch off sampling activation control */
944 	inactive = cpuhw->lsctl;
945 	inactive.cs = 0;
946 	inactive.cd = 0;
947 
948 	err = lsctl(&inactive);
949 	if (err) {
950 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
951 			2, err);
952 		return;
953 	}
954 
955 	/* Save state of TEAR and DEAR register contents */
956 	if (!qsi(&si)) {
957 		/* TEAR/DEAR values are valid only if the sampling facility is
958 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
959 		 * for a disabled sampling facility because cpumsf_pmu_enable()
960 		 * controls the enable/disable state.
961 		 */
962 		if (si.es) {
963 			cpuhw->lsctl.tear = si.tear;
964 			cpuhw->lsctl.dear = si.dear;
965 		}
966 	} else
967 		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
968 				    "qsi() failed with err=%i\n", err);
969 
970 	cpuhw->flags &= ~PMU_F_ENABLED;
971 }
972 
973 /* perf_exclude_event() - Filter event
974  * @event:	The perf event
975  * @regs:	pt_regs structure
976  * @sde_regs:	Sample-data-entry (sde) regs structure
977  *
978  * Filter perf events according to their exclude specification.
979  *
980  * Return non-zero if the event shall be excluded.
981  */
982 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
983 			      struct perf_sf_sde_regs *sde_regs)
984 {
985 	if (event->attr.exclude_user && user_mode(regs))
986 		return 1;
987 	if (event->attr.exclude_kernel && !user_mode(regs))
988 		return 1;
989 	if (event->attr.exclude_guest && sde_regs->in_guest)
990 		return 1;
991 	if (event->attr.exclude_host && !sde_regs->in_guest)
992 		return 1;
993 	return 0;
994 }
995 
996 /* perf_push_sample() - Push samples to perf
997  * @event:	The perf event
998  * @sample:	Hardware sample data
999  *
1000  * Use the hardware sample data to create perf event sample.  The sample
1001  * is the pushed to the event subsystem and the function checks for
1002  * possible event overflows.  If an event overflow occurs, the PMU is
1003  * stopped.
1004  *
1005  * Return non-zero if an event overflow occurred.
1006  */
1007 static int perf_push_sample(struct perf_event *event,
1008 			    struct hws_basic_entry *basic)
1009 {
1010 	int overflow;
1011 	struct pt_regs regs;
1012 	struct perf_sf_sde_regs *sde_regs;
1013 	struct perf_sample_data data;
1014 
1015 	/* Setup perf sample */
1016 	perf_sample_data_init(&data, 0, event->hw.last_period);
1017 
1018 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1019 	 * using the Program Request Alert code.  The regs.int_parm_long
1020 	 * field which is unused contains additional sample-data-entry related
1021 	 * indicators.
1022 	 */
1023 	memset(&regs, 0, sizeof(regs));
1024 	regs.int_code = 0x1407;
1025 	regs.int_parm = CPU_MF_INT_SF_PRA;
1026 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1027 
1028 	psw_bits(regs.psw).ia	= basic->ia;
1029 	psw_bits(regs.psw).dat	= basic->T;
1030 	psw_bits(regs.psw).wait = basic->W;
1031 	psw_bits(regs.psw).pstate = basic->P;
1032 	psw_bits(regs.psw).as	= basic->AS;
1033 
1034 	/*
1035 	 * Use the hardware provided configuration level to decide if the
1036 	 * sample belongs to a guest or host. If that is not available,
1037 	 * fall back to the following heuristics:
1038 	 * A non-zero guest program parameter always indicates a guest
1039 	 * sample. Some early samples or samples from guests without
1040 	 * lpp usage would be misaccounted to the host. We use the asn
1041 	 * value as an addon heuristic to detect most of these guest samples.
1042 	 * If the value differs from 0xffff (the host value), we assume to
1043 	 * be a KVM guest.
1044 	 */
1045 	switch (basic->CL) {
1046 	case 1: /* logical partition */
1047 		sde_regs->in_guest = 0;
1048 		break;
1049 	case 2: /* virtual machine */
1050 		sde_regs->in_guest = 1;
1051 		break;
1052 	default: /* old machine, use heuristics */
1053 		if (basic->gpp || basic->prim_asn != 0xffff)
1054 			sde_regs->in_guest = 1;
1055 		break;
1056 	}
1057 
1058 	/*
1059 	 * Store the PID value from the sample-data-entry to be
1060 	 * processed and resolved by cpumsf_output_event_pid().
1061 	 */
1062 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1063 
1064 	overflow = 0;
1065 	if (perf_exclude_event(event, &regs, sde_regs))
1066 		goto out;
1067 	if (perf_event_overflow(event, &data, &regs)) {
1068 		overflow = 1;
1069 		event->pmu->stop(event, 0);
1070 	}
1071 	perf_event_update_userpage(event);
1072 out:
1073 	return overflow;
1074 }
1075 
1076 static void perf_event_count_update(struct perf_event *event, u64 count)
1077 {
1078 	local64_add(count, &event->count);
1079 }
1080 
1081 static void debug_sample_entry(struct hws_basic_entry *sample,
1082 			       struct hws_trailer_entry *te)
1083 {
1084 	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1085 			    "sampling data entry: te->f=%i basic.def=%04x (%p)\n",
1086 			    te->f, sample->def, sample);
1087 }
1088 
1089 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1090  * @event:	The perf event
1091  * @sdbt:	Sample-data-block table
1092  * @overflow:	Event overflow counter
1093  *
1094  * Walks through a sample-data-block and collects sampling data entries that are
1095  * then pushed to the perf event subsystem.  Depending on the sampling function,
1096  * there can be either basic-sampling or combined-sampling data entries.  A
1097  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1098  * data entry.	The sampling function is determined by the flags in the perf
1099  * event hardware structure.  The function always works with a combined-sampling
1100  * data entry but ignores the the diagnostic portion if it is not available.
1101  *
1102  * Note that the implementation focuses on basic-sampling data entries and, if
1103  * such an entry is not valid, the entire combined-sampling data entry is
1104  * ignored.
1105  *
1106  * The overflow variables counts the number of samples that has been discarded
1107  * due to a perf event overflow.
1108  */
1109 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1110 			       unsigned long long *overflow)
1111 {
1112 	struct hws_trailer_entry *te;
1113 	struct hws_basic_entry *sample;
1114 
1115 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1116 	sample = (struct hws_basic_entry *) *sdbt;
1117 	while ((unsigned long *) sample < (unsigned long *) te) {
1118 		/* Check for an empty sample */
1119 		if (!sample->def)
1120 			break;
1121 
1122 		/* Update perf event period */
1123 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1124 
1125 		/* Check whether sample is valid */
1126 		if (sample->def == 0x0001) {
1127 			/* If an event overflow occurred, the PMU is stopped to
1128 			 * throttle event delivery.  Remaining sample data is
1129 			 * discarded.
1130 			 */
1131 			if (!*overflow) {
1132 				/* Check whether sample is consistent */
1133 				if (sample->I == 0 && sample->W == 0) {
1134 					/* Deliver sample data to perf */
1135 					*overflow = perf_push_sample(event,
1136 								     sample);
1137 				}
1138 			} else
1139 				/* Count discarded samples */
1140 				*overflow += 1;
1141 		} else {
1142 			debug_sample_entry(sample, te);
1143 			/* Sample slot is not yet written or other record.
1144 			 *
1145 			 * This condition can occur if the buffer was reused
1146 			 * from a combined basic- and diagnostic-sampling.
1147 			 * If only basic-sampling is then active, entries are
1148 			 * written into the larger diagnostic entries.
1149 			 * This is typically the case for sample-data-blocks
1150 			 * that are not full.  Stop processing if the first
1151 			 * invalid format was detected.
1152 			 */
1153 			if (!te->f)
1154 				break;
1155 		}
1156 
1157 		/* Reset sample slot and advance to next sample */
1158 		sample->def = 0;
1159 		sample++;
1160 	}
1161 }
1162 
1163 /* hw_perf_event_update() - Process sampling buffer
1164  * @event:	The perf event
1165  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1166  *
1167  * Processes the sampling buffer and create perf event samples.
1168  * The sampling buffer position are retrieved and saved in the TEAR_REG
1169  * register of the specified perf event.
1170  *
1171  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1172  * to also walk through partially filled sample-data-blocks.  It is ignored
1173  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1174  * enforces the processing of full sample-data-blocks only (trailer entries
1175  * with the block-full-indicator bit set).
1176  */
1177 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1178 {
1179 	struct hw_perf_event *hwc = &event->hw;
1180 	struct hws_trailer_entry *te;
1181 	unsigned long *sdbt;
1182 	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1183 	int done;
1184 
1185 	/*
1186 	 * AUX buffer is used when in diagnostic sampling mode.
1187 	 * No perf events/samples are created.
1188 	 */
1189 	if (SAMPL_DIAG_MODE(&event->hw))
1190 		return;
1191 
1192 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1193 		flush_all = 0;
1194 
1195 	sdbt = (unsigned long *) TEAR_REG(hwc);
1196 	done = event_overflow = sampl_overflow = num_sdb = 0;
1197 	while (!done) {
1198 		/* Get the trailer entry of the sample-data-block */
1199 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1200 
1201 		/* Leave loop if no more work to do (block full indicator) */
1202 		if (!te->f) {
1203 			done = 1;
1204 			if (!flush_all)
1205 				break;
1206 		}
1207 
1208 		/* Check the sample overflow count */
1209 		if (te->overflow)
1210 			/* Account sample overflows and, if a particular limit
1211 			 * is reached, extend the sampling buffer.
1212 			 * For details, see sfb_account_overflows().
1213 			 */
1214 			sampl_overflow += te->overflow;
1215 
1216 		/* Timestamps are valid for full sample-data-blocks only */
1217 		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1218 				    "overflow=%llu timestamp=0x%llx\n",
1219 				    sdbt, te->overflow,
1220 				    (te->f) ? trailer_timestamp(te) : 0ULL);
1221 
1222 		/* Collect all samples from a single sample-data-block and
1223 		 * flag if an (perf) event overflow happened.  If so, the PMU
1224 		 * is stopped and remaining samples will be discarded.
1225 		 */
1226 		hw_collect_samples(event, sdbt, &event_overflow);
1227 		num_sdb++;
1228 
1229 		/* Reset trailer (using compare-double-and-swap) */
1230 		do {
1231 			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1232 			te_flags |= SDB_TE_ALERT_REQ_MASK;
1233 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1234 					 te->flags, te->overflow,
1235 					 te_flags, 0ULL));
1236 
1237 		/* Advance to next sample-data-block */
1238 		sdbt++;
1239 		if (is_link_entry(sdbt))
1240 			sdbt = get_next_sdbt(sdbt);
1241 
1242 		/* Update event hardware registers */
1243 		TEAR_REG(hwc) = (unsigned long) sdbt;
1244 
1245 		/* Stop processing sample-data if all samples of the current
1246 		 * sample-data-block were flushed even if it was not full.
1247 		 */
1248 		if (flush_all && done)
1249 			break;
1250 
1251 		/* If an event overflow happened, discard samples by
1252 		 * processing any remaining sample-data-blocks.
1253 		 */
1254 		if (event_overflow)
1255 			flush_all = 1;
1256 	}
1257 
1258 	/* Account sample overflows in the event hardware structure */
1259 	if (sampl_overflow)
1260 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1261 						 sampl_overflow, 1 + num_sdb);
1262 	if (sampl_overflow || event_overflow)
1263 		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1264 				    "overflow stats: sample=%llu event=%llu\n",
1265 				    sampl_overflow, event_overflow);
1266 }
1267 
1268 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1269 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1270 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1271 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1272 
1273 /*
1274  * Get trailer entry by index of SDB.
1275  */
1276 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1277 						 unsigned long index)
1278 {
1279 	unsigned long sdb;
1280 
1281 	index = AUX_SDB_INDEX(aux, index);
1282 	sdb = aux->sdb_index[index];
1283 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1284 }
1285 
1286 /*
1287  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1288  * disabled. Collect the full SDBs in AUX buffer which have not reached
1289  * the point of alert indicator. And ignore the SDBs which are not
1290  * full.
1291  *
1292  * 1. Scan SDBs to see how much data is there and consume them.
1293  * 2. Remove alert indicator in the buffer.
1294  */
1295 static void aux_output_end(struct perf_output_handle *handle)
1296 {
1297 	unsigned long i, range_scan, idx;
1298 	struct aux_buffer *aux;
1299 	struct hws_trailer_entry *te;
1300 
1301 	aux = perf_get_aux(handle);
1302 	if (!aux)
1303 		return;
1304 
1305 	range_scan = AUX_SDB_NUM_ALERT(aux);
1306 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1307 		te = aux_sdb_trailer(aux, idx);
1308 		if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1309 			break;
1310 	}
1311 	/* i is num of SDBs which are full */
1312 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1313 
1314 	/* Remove alert indicators in the buffer */
1315 	te = aux_sdb_trailer(aux, aux->alert_mark);
1316 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1317 
1318 	debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1319 }
1320 
1321 /*
1322  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1323  * is first added to the CPU or rescheduled again to the CPU. It is called
1324  * with pmu disabled.
1325  *
1326  * 1. Reset the trailer of SDBs to get ready for new data.
1327  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1328  *    head(tear/dear).
1329  */
1330 static int aux_output_begin(struct perf_output_handle *handle,
1331 			    struct aux_buffer *aux,
1332 			    struct cpu_hw_sf *cpuhw)
1333 {
1334 	unsigned long range;
1335 	unsigned long i, range_scan, idx;
1336 	unsigned long head, base, offset;
1337 	struct hws_trailer_entry *te;
1338 
1339 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1340 		return -EINVAL;
1341 
1342 	aux->head = handle->head >> PAGE_SHIFT;
1343 	range = (handle->size + 1) >> PAGE_SHIFT;
1344 	if (range <= 1)
1345 		return -ENOMEM;
1346 
1347 	/*
1348 	 * SDBs between aux->head and aux->empty_mark are already ready
1349 	 * for new data. range_scan is num of SDBs not within them.
1350 	 */
1351 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1352 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1353 		idx = aux->empty_mark + 1;
1354 		for (i = 0; i < range_scan; i++, idx++) {
1355 			te = aux_sdb_trailer(aux, idx);
1356 			te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1357 			te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK;
1358 			te->overflow = 0;
1359 		}
1360 		/* Save the position of empty SDBs */
1361 		aux->empty_mark = aux->head + range - 1;
1362 	}
1363 
1364 	/* Set alert indicator */
1365 	aux->alert_mark = aux->head + range/2 - 1;
1366 	te = aux_sdb_trailer(aux, aux->alert_mark);
1367 	te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1368 
1369 	/* Reset hardware buffer head */
1370 	head = AUX_SDB_INDEX(aux, aux->head);
1371 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1372 	offset = head % CPUM_SF_SDB_PER_TABLE;
1373 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1374 	cpuhw->lsctl.dear = aux->sdb_index[head];
1375 
1376 	debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1377 			    "head->alert_mark->empty_mark (num_alert, range)"
1378 			    "[%lx -> %lx -> %lx] (%lx, %lx) "
1379 			    "tear index %lx, tear %lx dear %lx\n",
1380 			    aux->head, aux->alert_mark, aux->empty_mark,
1381 			    AUX_SDB_NUM_ALERT(aux), range,
1382 			    head / CPUM_SF_SDB_PER_TABLE,
1383 			    cpuhw->lsctl.tear,
1384 			    cpuhw->lsctl.dear);
1385 
1386 	return 0;
1387 }
1388 
1389 /*
1390  * Set alert indicator on SDB at index @alert_index while sampler is running.
1391  *
1392  * Return true if successfully.
1393  * Return false if full indicator is already set by hardware sampler.
1394  */
1395 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1396 			  unsigned long long *overflow)
1397 {
1398 	unsigned long long orig_overflow, orig_flags, new_flags;
1399 	struct hws_trailer_entry *te;
1400 
1401 	te = aux_sdb_trailer(aux, alert_index);
1402 	do {
1403 		orig_flags = te->flags;
1404 		orig_overflow = te->overflow;
1405 		*overflow = orig_overflow;
1406 		if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1407 			/*
1408 			 * SDB is already set by hardware.
1409 			 * Abort and try to set somewhere
1410 			 * behind.
1411 			 */
1412 			return false;
1413 		}
1414 		new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1415 	} while (!cmpxchg_double(&te->flags, &te->overflow,
1416 				 orig_flags, orig_overflow,
1417 				 new_flags, 0ULL));
1418 	return true;
1419 }
1420 
1421 /*
1422  * aux_reset_buffer() - Scan and setup SDBs for new samples
1423  * @aux:	The AUX buffer to set
1424  * @range:	The range of SDBs to scan started from aux->head
1425  * @overflow:	Set to overflow count
1426  *
1427  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1428  * marked as empty, check if it is already set full by the hardware sampler.
1429  * If yes, that means new data is already there before we can set an alert
1430  * indicator. Caller should try to set alert indicator to some position behind.
1431  *
1432  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1433  * previously and have already been consumed by user space. Reset these SDBs
1434  * (clear full indicator and alert indicator) for new data.
1435  * If aux->alert_mark fall in this area, just set it. Overflow count is
1436  * recorded while scanning.
1437  *
1438  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1439  * and ready for new samples. So scanning on this area could be skipped.
1440  *
1441  * Return true if alert indicator is set successfully and false if not.
1442  */
1443 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1444 			     unsigned long long *overflow)
1445 {
1446 	unsigned long long orig_overflow, orig_flags, new_flags;
1447 	unsigned long i, range_scan, idx;
1448 	struct hws_trailer_entry *te;
1449 
1450 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1451 		/*
1452 		 * No need to scan. All SDBs in range are marked as empty.
1453 		 * Just set alert indicator. Should check race with hardware
1454 		 * sampler.
1455 		 */
1456 		return aux_set_alert(aux, aux->alert_mark, overflow);
1457 
1458 	if (aux->alert_mark <= aux->empty_mark)
1459 		/*
1460 		 * Set alert indicator on empty SDB. Should check race
1461 		 * with hardware sampler.
1462 		 */
1463 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1464 			return false;
1465 
1466 	/*
1467 	 * Scan the SDBs to clear full and alert indicator used previously.
1468 	 * Start scanning from one SDB behind empty_mark. If the new alert
1469 	 * indicator fall into this range, set it.
1470 	 */
1471 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1472 	idx = aux->empty_mark + 1;
1473 	for (i = 0; i < range_scan; i++, idx++) {
1474 		te = aux_sdb_trailer(aux, idx);
1475 		do {
1476 			orig_flags = te->flags;
1477 			orig_overflow = te->overflow;
1478 			new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1479 			if (idx == aux->alert_mark)
1480 				new_flags |= SDB_TE_ALERT_REQ_MASK;
1481 			else
1482 				new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1483 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1484 					 orig_flags, orig_overflow,
1485 					 new_flags, 0ULL));
1486 		*overflow += orig_overflow;
1487 	}
1488 
1489 	/* Update empty_mark to new position */
1490 	aux->empty_mark = aux->head + range - 1;
1491 
1492 	return true;
1493 }
1494 
1495 /*
1496  * Measurement alert handler for diagnostic mode sampling.
1497  */
1498 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1499 {
1500 	struct aux_buffer *aux;
1501 	int done = 0;
1502 	unsigned long range = 0, size;
1503 	unsigned long long overflow = 0;
1504 	struct perf_output_handle *handle = &cpuhw->handle;
1505 	unsigned long num_sdb;
1506 
1507 	aux = perf_get_aux(handle);
1508 	if (WARN_ON_ONCE(!aux))
1509 		return;
1510 
1511 	/* Inform user space new data arrived */
1512 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1513 	perf_aux_output_end(handle, size);
1514 	num_sdb = aux->sfb.num_sdb;
1515 
1516 	while (!done) {
1517 		/* Get an output handle */
1518 		aux = perf_aux_output_begin(handle, cpuhw->event);
1519 		if (handle->size == 0) {
1520 			pr_err("The AUX buffer with %lu pages for the "
1521 			       "diagnostic-sampling mode is full\n",
1522 				num_sdb);
1523 			debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1524 			break;
1525 		}
1526 		if (WARN_ON_ONCE(!aux))
1527 			return;
1528 
1529 		/* Update head and alert_mark to new position */
1530 		aux->head = handle->head >> PAGE_SHIFT;
1531 		range = (handle->size + 1) >> PAGE_SHIFT;
1532 		if (range == 1)
1533 			aux->alert_mark = aux->head;
1534 		else
1535 			aux->alert_mark = aux->head + range/2 - 1;
1536 
1537 		if (aux_reset_buffer(aux, range, &overflow)) {
1538 			if (!overflow) {
1539 				done = 1;
1540 				break;
1541 			}
1542 			size = range << PAGE_SHIFT;
1543 			perf_aux_output_end(&cpuhw->handle, size);
1544 			pr_err("Sample data caused the AUX buffer with %lu "
1545 			       "pages to overflow\n", num_sdb);
1546 			debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1547 					    "overflow %llx\n",
1548 					    aux->head, range, overflow);
1549 		} else {
1550 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1551 			perf_aux_output_end(&cpuhw->handle, size);
1552 			debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1553 					    "already full, try another\n",
1554 					    aux->head, aux->alert_mark);
1555 		}
1556 	}
1557 
1558 	if (done)
1559 		debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1560 				    "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1561 				    aux->head, aux->alert_mark, aux->empty_mark,
1562 				    AUX_SDB_NUM_ALERT(aux), range);
1563 }
1564 
1565 /*
1566  * Callback when freeing AUX buffers.
1567  */
1568 static void aux_buffer_free(void *data)
1569 {
1570 	struct aux_buffer *aux = data;
1571 	unsigned long i, num_sdbt;
1572 
1573 	if (!aux)
1574 		return;
1575 
1576 	/* Free SDBT. SDB is freed by the caller */
1577 	num_sdbt = aux->sfb.num_sdbt;
1578 	for (i = 0; i < num_sdbt; i++)
1579 		free_page(aux->sdbt_index[i]);
1580 
1581 	kfree(aux->sdbt_index);
1582 	kfree(aux->sdb_index);
1583 	kfree(aux);
1584 
1585 	debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1586 			    "%lu SDBTs\n", num_sdbt);
1587 }
1588 
1589 /*
1590  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1591  * @cpu:	On which to allocate, -1 means current
1592  * @pages:	Array of pointers to buffer pages passed from perf core
1593  * @nr_pages:	Total pages
1594  * @snapshot:	Flag for snapshot mode
1595  *
1596  * This is the callback when setup an event using AUX buffer. Perf tool can
1597  * trigger this by an additional mmap() call on the event. Unlike the buffer
1598  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1599  * the task among online cpus when it is a per-thread event.
1600  *
1601  * Return the private AUX buffer structure if success or NULL if fails.
1602  */
1603 static void *aux_buffer_setup(int cpu, void **pages, int nr_pages,
1604 			      bool snapshot)
1605 {
1606 	struct sf_buffer *sfb;
1607 	struct aux_buffer *aux;
1608 	unsigned long *new, *tail;
1609 	int i, n_sdbt;
1610 
1611 	if (!nr_pages || !pages)
1612 		return NULL;
1613 
1614 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1615 		pr_err("AUX buffer size (%i pages) is larger than the "
1616 		       "maximum sampling buffer limit\n",
1617 		       nr_pages);
1618 		return NULL;
1619 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1620 		pr_err("AUX buffer size (%i pages) is less than the "
1621 		       "minimum sampling buffer limit\n",
1622 		       nr_pages);
1623 		return NULL;
1624 	}
1625 
1626 	/* Allocate aux_buffer struct for the event */
1627 	aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1628 	if (!aux)
1629 		goto no_aux;
1630 	sfb = &aux->sfb;
1631 
1632 	/* Allocate sdbt_index for fast reference */
1633 	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1634 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1635 	if (!aux->sdbt_index)
1636 		goto no_sdbt_index;
1637 
1638 	/* Allocate sdb_index for fast reference */
1639 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1640 	if (!aux->sdb_index)
1641 		goto no_sdb_index;
1642 
1643 	/* Allocate the first SDBT */
1644 	sfb->num_sdbt = 0;
1645 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1646 	if (!sfb->sdbt)
1647 		goto no_sdbt;
1648 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1649 	tail = sfb->tail = sfb->sdbt;
1650 
1651 	/*
1652 	 * Link the provided pages of AUX buffer to SDBT.
1653 	 * Allocate SDBT if needed.
1654 	 */
1655 	for (i = 0; i < nr_pages; i++, tail++) {
1656 		if (require_table_link(tail)) {
1657 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1658 			if (!new)
1659 				goto no_sdbt;
1660 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1661 			/* Link current page to tail of chain */
1662 			*tail = (unsigned long)(void *) new + 1;
1663 			tail = new;
1664 		}
1665 		/* Tail is the entry in a SDBT */
1666 		*tail = (unsigned long)pages[i];
1667 		aux->sdb_index[i] = (unsigned long)pages[i];
1668 	}
1669 	sfb->num_sdb = nr_pages;
1670 
1671 	/* Link the last entry in the SDBT to the first SDBT */
1672 	*tail = (unsigned long) sfb->sdbt + 1;
1673 	sfb->tail = tail;
1674 
1675 	/*
1676 	 * Initial all SDBs are zeroed. Mark it as empty.
1677 	 * So there is no need to clear the full indicator
1678 	 * when this event is first added.
1679 	 */
1680 	aux->empty_mark = sfb->num_sdb - 1;
1681 
1682 	debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1683 			    " and %lu SDBs\n",
1684 			    sfb->num_sdbt, sfb->num_sdb);
1685 
1686 	return aux;
1687 
1688 no_sdbt:
1689 	/* SDBs (AUX buffer pages) are freed by caller */
1690 	for (i = 0; i < sfb->num_sdbt; i++)
1691 		free_page(aux->sdbt_index[i]);
1692 	kfree(aux->sdb_index);
1693 no_sdb_index:
1694 	kfree(aux->sdbt_index);
1695 no_sdbt_index:
1696 	kfree(aux);
1697 no_aux:
1698 	return NULL;
1699 }
1700 
1701 static void cpumsf_pmu_read(struct perf_event *event)
1702 {
1703 	/* Nothing to do ... updates are interrupt-driven */
1704 }
1705 
1706 /* Activate sampling control.
1707  * Next call of pmu_enable() starts sampling.
1708  */
1709 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1710 {
1711 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1712 
1713 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1714 		return;
1715 
1716 	if (flags & PERF_EF_RELOAD)
1717 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1718 
1719 	perf_pmu_disable(event->pmu);
1720 	event->hw.state = 0;
1721 	cpuhw->lsctl.cs = 1;
1722 	if (SAMPL_DIAG_MODE(&event->hw))
1723 		cpuhw->lsctl.cd = 1;
1724 	perf_pmu_enable(event->pmu);
1725 }
1726 
1727 /* Deactivate sampling control.
1728  * Next call of pmu_enable() stops sampling.
1729  */
1730 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1731 {
1732 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1733 
1734 	if (event->hw.state & PERF_HES_STOPPED)
1735 		return;
1736 
1737 	perf_pmu_disable(event->pmu);
1738 	cpuhw->lsctl.cs = 0;
1739 	cpuhw->lsctl.cd = 0;
1740 	event->hw.state |= PERF_HES_STOPPED;
1741 
1742 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1743 		hw_perf_event_update(event, 1);
1744 		event->hw.state |= PERF_HES_UPTODATE;
1745 	}
1746 	perf_pmu_enable(event->pmu);
1747 }
1748 
1749 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1750 {
1751 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1752 	struct aux_buffer *aux;
1753 	int err;
1754 
1755 	if (cpuhw->flags & PMU_F_IN_USE)
1756 		return -EAGAIN;
1757 
1758 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1759 		return -EINVAL;
1760 
1761 	err = 0;
1762 	perf_pmu_disable(event->pmu);
1763 
1764 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1765 
1766 	/* Set up sampling controls.  Always program the sampling register
1767 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1768 	 * that is used by hw_perf_event_update() to store the sampling buffer
1769 	 * position after samples have been flushed.
1770 	 */
1771 	cpuhw->lsctl.s = 0;
1772 	cpuhw->lsctl.h = 1;
1773 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1774 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1775 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1776 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1777 		hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1778 	}
1779 
1780 	/* Ensure sampling functions are in the disabled state.  If disabled,
1781 	 * switch on sampling enable control. */
1782 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1783 		err = -EAGAIN;
1784 		goto out;
1785 	}
1786 	if (SAMPL_DIAG_MODE(&event->hw)) {
1787 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1788 		if (!aux) {
1789 			err = -EINVAL;
1790 			goto out;
1791 		}
1792 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1793 		if (err)
1794 			goto out;
1795 		cpuhw->lsctl.ed = 1;
1796 	}
1797 	cpuhw->lsctl.es = 1;
1798 
1799 	/* Set in_use flag and store event */
1800 	cpuhw->event = event;
1801 	cpuhw->flags |= PMU_F_IN_USE;
1802 
1803 	if (flags & PERF_EF_START)
1804 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1805 out:
1806 	perf_event_update_userpage(event);
1807 	perf_pmu_enable(event->pmu);
1808 	return err;
1809 }
1810 
1811 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1812 {
1813 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1814 
1815 	perf_pmu_disable(event->pmu);
1816 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1817 
1818 	cpuhw->lsctl.es = 0;
1819 	cpuhw->lsctl.ed = 0;
1820 	cpuhw->flags &= ~PMU_F_IN_USE;
1821 	cpuhw->event = NULL;
1822 
1823 	if (SAMPL_DIAG_MODE(&event->hw))
1824 		aux_output_end(&cpuhw->handle);
1825 	perf_event_update_userpage(event);
1826 	perf_pmu_enable(event->pmu);
1827 }
1828 
1829 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1830 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1831 
1832 static struct attribute *cpumsf_pmu_events_attr[] = {
1833 	CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
1834 	NULL,
1835 	NULL,
1836 };
1837 
1838 PMU_FORMAT_ATTR(event, "config:0-63");
1839 
1840 static struct attribute *cpumsf_pmu_format_attr[] = {
1841 	&format_attr_event.attr,
1842 	NULL,
1843 };
1844 
1845 static struct attribute_group cpumsf_pmu_events_group = {
1846 	.name = "events",
1847 	.attrs = cpumsf_pmu_events_attr,
1848 };
1849 static struct attribute_group cpumsf_pmu_format_group = {
1850 	.name = "format",
1851 	.attrs = cpumsf_pmu_format_attr,
1852 };
1853 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1854 	&cpumsf_pmu_events_group,
1855 	&cpumsf_pmu_format_group,
1856 	NULL,
1857 };
1858 
1859 static struct pmu cpumf_sampling = {
1860 	.pmu_enable   = cpumsf_pmu_enable,
1861 	.pmu_disable  = cpumsf_pmu_disable,
1862 
1863 	.event_init   = cpumsf_pmu_event_init,
1864 	.add	      = cpumsf_pmu_add,
1865 	.del	      = cpumsf_pmu_del,
1866 
1867 	.start	      = cpumsf_pmu_start,
1868 	.stop	      = cpumsf_pmu_stop,
1869 	.read	      = cpumsf_pmu_read,
1870 
1871 	.attr_groups  = cpumsf_pmu_attr_groups,
1872 
1873 	.setup_aux    = aux_buffer_setup,
1874 	.free_aux     = aux_buffer_free,
1875 };
1876 
1877 static void cpumf_measurement_alert(struct ext_code ext_code,
1878 				    unsigned int alert, unsigned long unused)
1879 {
1880 	struct cpu_hw_sf *cpuhw;
1881 
1882 	if (!(alert & CPU_MF_INT_SF_MASK))
1883 		return;
1884 	inc_irq_stat(IRQEXT_CMS);
1885 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
1886 
1887 	/* Measurement alerts are shared and might happen when the PMU
1888 	 * is not reserved.  Ignore these alerts in this case. */
1889 	if (!(cpuhw->flags & PMU_F_RESERVED))
1890 		return;
1891 
1892 	/* The processing below must take care of multiple alert events that
1893 	 * might be indicated concurrently. */
1894 
1895 	/* Program alert request */
1896 	if (alert & CPU_MF_INT_SF_PRA) {
1897 		if (cpuhw->flags & PMU_F_IN_USE)
1898 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
1899 				hw_collect_aux(cpuhw);
1900 			else
1901 				hw_perf_event_update(cpuhw->event, 0);
1902 		else
1903 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
1904 	}
1905 
1906 	/* Report measurement alerts only for non-PRA codes */
1907 	if (alert != CPU_MF_INT_SF_PRA)
1908 		debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
1909 
1910 	/* Sampling authorization change request */
1911 	if (alert & CPU_MF_INT_SF_SACA)
1912 		qsi(&cpuhw->qsi);
1913 
1914 	/* Loss of sample data due to high-priority machine activities */
1915 	if (alert & CPU_MF_INT_SF_LSDA) {
1916 		pr_err("Sample data was lost\n");
1917 		cpuhw->flags |= PMU_F_ERR_LSDA;
1918 		sf_disable();
1919 	}
1920 
1921 	/* Invalid sampling buffer entry */
1922 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
1923 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
1924 		       alert);
1925 		cpuhw->flags |= PMU_F_ERR_IBE;
1926 		sf_disable();
1927 	}
1928 }
1929 static int cpusf_pmu_setup(unsigned int cpu, int flags)
1930 {
1931 	/* Ignore the notification if no events are scheduled on the PMU.
1932 	 * This might be racy...
1933 	 */
1934 	if (!atomic_read(&num_events))
1935 		return 0;
1936 
1937 	local_irq_disable();
1938 	setup_pmc_cpu(&flags);
1939 	local_irq_enable();
1940 	return 0;
1941 }
1942 
1943 static int s390_pmu_sf_online_cpu(unsigned int cpu)
1944 {
1945 	return cpusf_pmu_setup(cpu, PMC_INIT);
1946 }
1947 
1948 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
1949 {
1950 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
1951 }
1952 
1953 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
1954 {
1955 	if (!cpum_sf_avail())
1956 		return -ENODEV;
1957 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
1958 }
1959 
1960 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
1961 {
1962 	int rc;
1963 	unsigned long min, max;
1964 
1965 	if (!cpum_sf_avail())
1966 		return -ENODEV;
1967 	if (!val || !strlen(val))
1968 		return -EINVAL;
1969 
1970 	/* Valid parameter values: "min,max" or "max" */
1971 	min = CPUM_SF_MIN_SDB;
1972 	max = CPUM_SF_MAX_SDB;
1973 	if (strchr(val, ','))
1974 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
1975 	else
1976 		rc = kstrtoul(val, 10, &max);
1977 
1978 	if (min < 2 || min >= max || max > get_num_physpages())
1979 		rc = -EINVAL;
1980 	if (rc)
1981 		return rc;
1982 
1983 	sfb_set_limits(min, max);
1984 	pr_info("The sampling buffer limits have changed to: "
1985 		"min=%lu max=%lu (diag=x%lu)\n",
1986 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
1987 	return 0;
1988 }
1989 
1990 #define param_check_sfb_size(name, p) __param_check(name, p, void)
1991 static const struct kernel_param_ops param_ops_sfb_size = {
1992 	.set = param_set_sfb_size,
1993 	.get = param_get_sfb_size,
1994 };
1995 
1996 #define RS_INIT_FAILURE_QSI	  0x0001
1997 #define RS_INIT_FAILURE_BSDES	  0x0002
1998 #define RS_INIT_FAILURE_ALRT	  0x0003
1999 #define RS_INIT_FAILURE_PERF	  0x0004
2000 static void __init pr_cpumsf_err(unsigned int reason)
2001 {
2002 	pr_err("Sampling facility support for perf is not available: "
2003 	       "reason=%04x\n", reason);
2004 }
2005 
2006 static int __init init_cpum_sampling_pmu(void)
2007 {
2008 	struct hws_qsi_info_block si;
2009 	int err;
2010 
2011 	if (!cpum_sf_avail())
2012 		return -ENODEV;
2013 
2014 	memset(&si, 0, sizeof(si));
2015 	if (qsi(&si)) {
2016 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2017 		return -ENODEV;
2018 	}
2019 
2020 	if (!si.as && !si.ad)
2021 		return -ENODEV;
2022 
2023 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2024 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2025 		return -EINVAL;
2026 	}
2027 
2028 	if (si.ad) {
2029 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2030 		cpumsf_pmu_events_attr[1] =
2031 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2032 	}
2033 
2034 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2035 	if (!sfdbg)
2036 		pr_err("Registering for s390dbf failed\n");
2037 	debug_register_view(sfdbg, &debug_sprintf_view);
2038 
2039 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2040 				    cpumf_measurement_alert);
2041 	if (err) {
2042 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2043 		goto out;
2044 	}
2045 
2046 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2047 	if (err) {
2048 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2049 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2050 					cpumf_measurement_alert);
2051 		goto out;
2052 	}
2053 
2054 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2055 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2056 out:
2057 	return err;
2058 }
2059 arch_initcall(init_cpum_sampling_pmu);
2060 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);
2061