xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision 750afb08)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Performance event support for the System z CPU-measurement Sampling Facility
4  *
5  * Copyright IBM Corp. 2013, 2018
6  * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
7  */
8 #define KMSG_COMPONENT	"cpum_sf"
9 #define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/perf_event.h>
14 #include <linux/percpu.h>
15 #include <linux/pid.h>
16 #include <linux/notifier.h>
17 #include <linux/export.h>
18 #include <linux/slab.h>
19 #include <linux/mm.h>
20 #include <linux/moduleparam.h>
21 #include <asm/cpu_mf.h>
22 #include <asm/irq.h>
23 #include <asm/debug.h>
24 #include <asm/timex.h>
25 
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.
29  */
30 #define CPUM_SF_MIN_SDBT	1
31 
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.
35  */
36 #define CPUM_SF_SDB_PER_TABLE	((PAGE_SIZE - 8) / 8)
37 
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
40  * must be added.
41  */
42 #define CPUM_SF_SDBT_TL_OFFSET	(CPUM_SF_SDB_PER_TABLE * 8)
43 static inline int require_table_link(const void *sdbt)
44 {
45 	return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
46 }
47 
48 /* Minimum and maximum sampling buffer sizes:
49  *
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.
55  *
56  * Sampling buffer size		Buffer characteristics
57  * ---------------------------------------------------
58  *	 64KB		    ==	  16 pages (4KB per page)
59  *				   1 page  for SDB-tables
60  *				  15 pages for SDBs
61  *
62  *  32MB		    ==	8192 pages (4KB per page)
63  *				  16 pages for SDB-tables
64  *				8176 pages for SDBs
65  */
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;
69 
70 struct sf_buffer {
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 */
76 };
77 
78 struct aux_buffer {
79 	struct sf_buffer sfb;
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 */
85 };
86 
87 struct cpu_hw_sf {
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 */
96 };
97 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
98 
99 /* Debug feature */
100 static debug_info_t *sfdbg;
101 
102 /*
103  * sf_disable() - Switch off sampling facility
104  */
105 static int sf_disable(void)
106 {
107 	struct hws_lsctl_request_block sreq;
108 
109 	memset(&sreq, 0, sizeof(sreq));
110 	return lsctl(&sreq);
111 }
112 
113 /*
114  * sf_buffer_available() - Check for an allocated sampling buffer
115  */
116 static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
117 {
118 	return !!cpuhw->sfb.sdbt;
119 }
120 
121 /*
122  * deallocate sampling facility buffer
123  */
124 static void free_sampling_buffer(struct sf_buffer *sfb)
125 {
126 	unsigned long *sdbt, *curr;
127 
128 	if (!sfb->sdbt)
129 		return;
130 
131 	sdbt = sfb->sdbt;
132 	curr = sdbt;
133 
134 	/* Free the SDBT after all SDBs are processed... */
135 	while (1) {
136 		if (!*curr || !sdbt)
137 			break;
138 
139 		/* Process table-link entries */
140 		if (is_link_entry(curr)) {
141 			curr = get_next_sdbt(curr);
142 			if (sdbt)
143 				free_page((unsigned long) sdbt);
144 
145 			/* If the origin is reached, sampling buffer is freed */
146 			if (curr == sfb->sdbt)
147 				break;
148 			else
149 				sdbt = curr;
150 		} else {
151 			/* Process SDB pointer */
152 			if (*curr) {
153 				free_page(*curr);
154 				curr++;
155 			}
156 		}
157 	}
158 
159 	debug_sprintf_event(sfdbg, 5,
160 			    "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
161 	memset(sfb, 0, sizeof(*sfb));
162 }
163 
164 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
165 {
166 	unsigned long sdb, *trailer;
167 
168 	/* Allocate and initialize sample-data-block */
169 	sdb = get_zeroed_page(gfp_flags);
170 	if (!sdb)
171 		return -ENOMEM;
172 	trailer = trailer_entry_ptr(sdb);
173 	*trailer = SDB_TE_ALERT_REQ_MASK;
174 
175 	/* Link SDB into the sample-data-block-table */
176 	*sdbt = sdb;
177 
178 	return 0;
179 }
180 
181 /*
182  * realloc_sampling_buffer() - extend sampler memory
183  *
184  * Allocates new sample-data-blocks and adds them to the specified sampling
185  * buffer memory.
186  *
187  * Important: This modifies the sampling buffer and must be called when the
188  *	      sampling facility is disabled.
189  *
190  * Returns zero on success, non-zero otherwise.
191  */
192 static int realloc_sampling_buffer(struct sf_buffer *sfb,
193 				   unsigned long num_sdb, gfp_t gfp_flags)
194 {
195 	int i, rc;
196 	unsigned long *new, *tail;
197 
198 	if (!sfb->sdbt || !sfb->tail)
199 		return -EINVAL;
200 
201 	if (!is_link_entry(sfb->tail))
202 		return -EINVAL;
203 
204 	/* Append to the existing sampling buffer, overwriting the table-link
205 	 * register.
206 	 * The tail variables always points to the "tail" (last and table-link)
207 	 * entry in an SDB-table.
208 	 */
209 	tail = sfb->tail;
210 
211 	/* Do a sanity check whether the table-link entry points to
212 	 * the sampling buffer origin.
213 	 */
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"
217 				    "tail=%p\n",
218 				    (void *) sfb->sdbt, (void *) tail);
219 		return -EINVAL;
220 	}
221 
222 	/* Allocate remaining SDBs */
223 	rc = 0;
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);
228 			if (!new) {
229 				rc = -ENOMEM;
230 				break;
231 			}
232 			sfb->num_sdbt++;
233 			/* Link current page to tail of chain */
234 			*tail = (unsigned long)(void *) new + 1;
235 			tail = new;
236 		}
237 
238 		/* Allocate a new sample-data-block.
239 		 * If there is not enough memory, stop the realloc process
240 		 * and simply use what was allocated.  If this is a temporary
241 		 * issue, a new realloc call (if required) might succeed.
242 		 */
243 		rc = alloc_sample_data_block(tail, gfp_flags);
244 		if (rc)
245 			break;
246 		sfb->num_sdb++;
247 		tail++;
248 	}
249 
250 	/* Link sampling buffer to its origin */
251 	*tail = (unsigned long) sfb->sdbt + 1;
252 	sfb->tail = tail;
253 
254 	debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
255 			    " settings: sdbt=%lu sdb=%lu\n",
256 			    sfb->num_sdbt, sfb->num_sdb);
257 	return rc;
258 }
259 
260 /*
261  * allocate_sampling_buffer() - allocate sampler memory
262  *
263  * Allocates and initializes a sampling buffer structure using the
264  * specified number of sample-data-blocks (SDB).  For each allocation,
265  * a 4K page is used.  The number of sample-data-block-tables (SDBT)
266  * are calculated from SDBs.
267  * Also set the ALERT_REQ mask in each SDBs trailer.
268  *
269  * Returns zero on success, non-zero otherwise.
270  */
271 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
272 {
273 	int rc;
274 
275 	if (sfb->sdbt)
276 		return -EINVAL;
277 
278 	/* Allocate the sample-data-block-table origin */
279 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
280 	if (!sfb->sdbt)
281 		return -ENOMEM;
282 	sfb->num_sdb = 0;
283 	sfb->num_sdbt = 1;
284 
285 	/* Link the table origin to point to itself to prepare for
286 	 * realloc_sampling_buffer() invocation.
287 	 */
288 	sfb->tail = sfb->sdbt;
289 	*sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
290 
291 	/* Allocate requested number of sample-data-blocks */
292 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
293 	if (rc) {
294 		free_sampling_buffer(sfb);
295 		debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
296 			"realloc_sampling_buffer failed with rc=%i\n", rc);
297 	} else
298 		debug_sprintf_event(sfdbg, 4,
299 			"alloc_sampling_buffer: tear=%p dear=%p\n",
300 			sfb->sdbt, (void *) *sfb->sdbt);
301 	return rc;
302 }
303 
304 static void sfb_set_limits(unsigned long min, unsigned long max)
305 {
306 	struct hws_qsi_info_block si;
307 
308 	CPUM_SF_MIN_SDB = min;
309 	CPUM_SF_MAX_SDB = max;
310 
311 	memset(&si, 0, sizeof(si));
312 	if (!qsi(&si))
313 		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
314 }
315 
316 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
317 {
318 	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
319 				    : CPUM_SF_MAX_SDB;
320 }
321 
322 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
323 					struct hw_perf_event *hwc)
324 {
325 	if (!sfb->sdbt)
326 		return SFB_ALLOC_REG(hwc);
327 	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
328 		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
329 	return 0;
330 }
331 
332 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
333 				   struct hw_perf_event *hwc)
334 {
335 	return sfb_pending_allocs(sfb, hwc) > 0;
336 }
337 
338 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
339 {
340 	/* Limit the number of SDBs to not exceed the maximum */
341 	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
342 	if (num)
343 		SFB_ALLOC_REG(hwc) += num;
344 }
345 
346 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
347 {
348 	SFB_ALLOC_REG(hwc) = 0;
349 	sfb_account_allocs(num, hwc);
350 }
351 
352 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
353 {
354 	if (cpuhw->sfb.sdbt)
355 		free_sampling_buffer(&cpuhw->sfb);
356 }
357 
358 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
359 {
360 	unsigned long n_sdb, freq, factor;
361 	size_t sample_size;
362 
363 	/* Calculate sampling buffers using 4K pages
364 	 *
365 	 *    1. Determine the sample data size which depends on the used
366 	 *	 sampling functions, for example, basic-sampling or
367 	 *	 basic-sampling with diagnostic-sampling.
368 	 *
369 	 *    2. Use the sampling frequency as input.  The sampling buffer is
370 	 *	 designed for almost one second.  This can be adjusted through
371 	 *	 the "factor" variable.
372 	 *	 In any case, alloc_sampling_buffer() sets the Alert Request
373 	 *	 Control indicator to trigger a measurement-alert to harvest
374 	 *	 sample-data-blocks (sdb).
375 	 *
376 	 *    3. Compute the number of sample-data-blocks and ensure a minimum
377 	 *	 of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
378 	 *	 exceed a "calculated" maximum.  The symbolic maximum is
379 	 *	 designed for basic-sampling only and needs to be increased if
380 	 *	 diagnostic-sampling is active.
381 	 *	 See also the remarks for these symbolic constants.
382 	 *
383 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
384 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
385 	 *	 to 511 SDBs).
386 	 */
387 	sample_size = sizeof(struct hws_basic_entry);
388 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
389 	factor = 1;
390 	n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
391 	if (n_sdb < CPUM_SF_MIN_SDB)
392 		n_sdb = CPUM_SF_MIN_SDB;
393 
394 	/* If there is already a sampling buffer allocated, it is very likely
395 	 * that the sampling facility is enabled too.  If the event to be
396 	 * initialized requires a greater sampling buffer, the allocation must
397 	 * be postponed.  Changing the sampling buffer requires the sampling
398 	 * facility to be in the disabled state.  So, account the number of
399 	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
400 	 * before the event is started.
401 	 */
402 	sfb_init_allocs(n_sdb, hwc);
403 	if (sf_buffer_available(cpuhw))
404 		return 0;
405 
406 	debug_sprintf_event(sfdbg, 3,
407 			    "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
408 			    " sample_size=%lu cpuhw=%p\n",
409 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
410 			    sample_size, cpuhw);
411 
412 	return alloc_sampling_buffer(&cpuhw->sfb,
413 				     sfb_pending_allocs(&cpuhw->sfb, hwc));
414 }
415 
416 static unsigned long min_percent(unsigned int percent, unsigned long base,
417 				 unsigned long min)
418 {
419 	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
420 }
421 
422 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
423 {
424 	/* Use a percentage-based approach to extend the sampling facility
425 	 * buffer.  Accept up to 5% sample data loss.
426 	 * Vary the extents between 1% to 5% of the current number of
427 	 * sample-data-blocks.
428 	 */
429 	if (ratio <= 5)
430 		return 0;
431 	if (ratio <= 25)
432 		return min_percent(1, base, 1);
433 	if (ratio <= 50)
434 		return min_percent(1, base, 1);
435 	if (ratio <= 75)
436 		return min_percent(2, base, 2);
437 	if (ratio <= 100)
438 		return min_percent(3, base, 3);
439 	if (ratio <= 250)
440 		return min_percent(4, base, 4);
441 
442 	return min_percent(5, base, 8);
443 }
444 
445 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
446 				  struct hw_perf_event *hwc)
447 {
448 	unsigned long ratio, num;
449 
450 	if (!OVERFLOW_REG(hwc))
451 		return;
452 
453 	/* The sample_overflow contains the average number of sample data
454 	 * that has been lost because sample-data-blocks were full.
455 	 *
456 	 * Calculate the total number of sample data entries that has been
457 	 * discarded.  Then calculate the ratio of lost samples to total samples
458 	 * per second in percent.
459 	 */
460 	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
461 			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
462 
463 	/* Compute number of sample-data-blocks */
464 	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
465 	if (num)
466 		sfb_account_allocs(num, hwc);
467 
468 	debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
469 			    " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
470 	OVERFLOW_REG(hwc) = 0;
471 }
472 
473 /* extend_sampling_buffer() - Extend sampling buffer
474  * @sfb:	Sampling buffer structure (for local CPU)
475  * @hwc:	Perf event hardware structure
476  *
477  * Use this function to extend the sampling buffer based on the overflow counter
478  * and postponed allocation extents stored in the specified Perf event hardware.
479  *
480  * Important: This function disables the sampling facility in order to safely
481  *	      change the sampling buffer structure.  Do not call this function
482  *	      when the PMU is active.
483  */
484 static void extend_sampling_buffer(struct sf_buffer *sfb,
485 				   struct hw_perf_event *hwc)
486 {
487 	unsigned long num, num_old;
488 	int rc;
489 
490 	num = sfb_pending_allocs(sfb, hwc);
491 	if (!num)
492 		return;
493 	num_old = sfb->num_sdb;
494 
495 	/* Disable the sampling facility to reset any states and also
496 	 * clear pending measurement alerts.
497 	 */
498 	sf_disable();
499 
500 	/* Extend the sampling buffer.
501 	 * This memory allocation typically happens in an atomic context when
502 	 * called by perf.  Because this is a reallocation, it is fine if the
503 	 * new SDB-request cannot be satisfied immediately.
504 	 */
505 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
506 	if (rc)
507 		debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
508 				    "failed with rc=%i\n", rc);
509 
510 	if (sfb_has_pending_allocs(sfb, hwc))
511 		debug_sprintf_event(sfdbg, 5, "sfb: extend: "
512 				    "req=%lu alloc=%lu remaining=%lu\n",
513 				    num, sfb->num_sdb - num_old,
514 				    sfb_pending_allocs(sfb, hwc));
515 }
516 
517 
518 /* Number of perf events counting hardware events */
519 static atomic_t num_events;
520 /* Used to avoid races in calling reserve/release_cpumf_hardware */
521 static DEFINE_MUTEX(pmc_reserve_mutex);
522 
523 #define PMC_INIT      0
524 #define PMC_RELEASE   1
525 #define PMC_FAILURE   2
526 static void setup_pmc_cpu(void *flags)
527 {
528 	int err;
529 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
530 
531 	err = 0;
532 	switch (*((int *) flags)) {
533 	case PMC_INIT:
534 		memset(cpusf, 0, sizeof(*cpusf));
535 		err = qsi(&cpusf->qsi);
536 		if (err)
537 			break;
538 		cpusf->flags |= PMU_F_RESERVED;
539 		err = sf_disable();
540 		if (err)
541 			pr_err("Switching off the sampling facility failed "
542 			       "with rc=%i\n", err);
543 		debug_sprintf_event(sfdbg, 5,
544 				    "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
545 		break;
546 	case PMC_RELEASE:
547 		cpusf->flags &= ~PMU_F_RESERVED;
548 		err = sf_disable();
549 		if (err) {
550 			pr_err("Switching off the sampling facility failed "
551 			       "with rc=%i\n", err);
552 		} else
553 			deallocate_buffers(cpusf);
554 		debug_sprintf_event(sfdbg, 5,
555 				    "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
556 		break;
557 	}
558 	if (err)
559 		*((int *) flags) |= PMC_FAILURE;
560 }
561 
562 static void release_pmc_hardware(void)
563 {
564 	int flags = PMC_RELEASE;
565 
566 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
567 	on_each_cpu(setup_pmc_cpu, &flags, 1);
568 }
569 
570 static int reserve_pmc_hardware(void)
571 {
572 	int flags = PMC_INIT;
573 
574 	on_each_cpu(setup_pmc_cpu, &flags, 1);
575 	if (flags & PMC_FAILURE) {
576 		release_pmc_hardware();
577 		return -ENODEV;
578 	}
579 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
580 
581 	return 0;
582 }
583 
584 static void hw_perf_event_destroy(struct perf_event *event)
585 {
586 	/* Release PMC if this is the last perf event */
587 	if (!atomic_add_unless(&num_events, -1, 1)) {
588 		mutex_lock(&pmc_reserve_mutex);
589 		if (atomic_dec_return(&num_events) == 0)
590 			release_pmc_hardware();
591 		mutex_unlock(&pmc_reserve_mutex);
592 	}
593 }
594 
595 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
596 {
597 	hwc->sample_period = period;
598 	hwc->last_period = hwc->sample_period;
599 	local64_set(&hwc->period_left, hwc->sample_period);
600 }
601 
602 static void hw_reset_registers(struct hw_perf_event *hwc,
603 			       unsigned long *sdbt_origin)
604 {
605 	/* (Re)set to first sample-data-block-table */
606 	TEAR_REG(hwc) = (unsigned long) sdbt_origin;
607 }
608 
609 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
610 				   unsigned long rate)
611 {
612 	return clamp_t(unsigned long, rate,
613 		       si->min_sampl_rate, si->max_sampl_rate);
614 }
615 
616 static u32 cpumsf_pid_type(struct perf_event *event,
617 			   u32 pid, enum pid_type type)
618 {
619 	struct task_struct *tsk;
620 
621 	/* Idle process */
622 	if (!pid)
623 		goto out;
624 
625 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
626 	pid = -1;
627 	if (tsk) {
628 		/*
629 		 * Only top level events contain the pid namespace in which
630 		 * they are created.
631 		 */
632 		if (event->parent)
633 			event = event->parent;
634 		pid = __task_pid_nr_ns(tsk, type, event->ns);
635 		/*
636 		 * See also 1d953111b648
637 		 * "perf/core: Don't report zero PIDs for exiting tasks".
638 		 */
639 		if (!pid && !pid_alive(tsk))
640 			pid = -1;
641 	}
642 out:
643 	return pid;
644 }
645 
646 static void cpumsf_output_event_pid(struct perf_event *event,
647 				    struct perf_sample_data *data,
648 				    struct pt_regs *regs)
649 {
650 	u32 pid;
651 	struct perf_event_header header;
652 	struct perf_output_handle handle;
653 
654 	/*
655 	 * Obtain the PID from the basic-sampling data entry and
656 	 * correct the data->tid_entry.pid value.
657 	 */
658 	pid = data->tid_entry.pid;
659 
660 	/* Protect callchain buffers, tasks */
661 	rcu_read_lock();
662 
663 	perf_prepare_sample(&header, data, event, regs);
664 	if (perf_output_begin(&handle, event, header.size))
665 		goto out;
666 
667 	/* Update the process ID (see also kernel/events/core.c) */
668 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
669 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
670 
671 	perf_output_sample(&handle, &header, data, event);
672 	perf_output_end(&handle);
673 out:
674 	rcu_read_unlock();
675 }
676 
677 static int __hw_perf_event_init(struct perf_event *event)
678 {
679 	struct cpu_hw_sf *cpuhw;
680 	struct hws_qsi_info_block si;
681 	struct perf_event_attr *attr = &event->attr;
682 	struct hw_perf_event *hwc = &event->hw;
683 	unsigned long rate;
684 	int cpu, err;
685 
686 	/* Reserve CPU-measurement sampling facility */
687 	err = 0;
688 	if (!atomic_inc_not_zero(&num_events)) {
689 		mutex_lock(&pmc_reserve_mutex);
690 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
691 			err = -EBUSY;
692 		else
693 			atomic_inc(&num_events);
694 		mutex_unlock(&pmc_reserve_mutex);
695 	}
696 	event->destroy = hw_perf_event_destroy;
697 
698 	if (err)
699 		goto out;
700 
701 	/* Access per-CPU sampling information (query sampling info) */
702 	/*
703 	 * The event->cpu value can be -1 to count on every CPU, for example,
704 	 * when attaching to a task.  If this is specified, use the query
705 	 * sampling info from the current CPU, otherwise use event->cpu to
706 	 * retrieve the per-CPU information.
707 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
708 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
709 	 */
710 	memset(&si, 0, sizeof(si));
711 	cpuhw = NULL;
712 	if (event->cpu == -1)
713 		qsi(&si);
714 	else {
715 		/* Event is pinned to a particular CPU, retrieve the per-CPU
716 		 * sampling structure for accessing the CPU-specific QSI.
717 		 */
718 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
719 		si = cpuhw->qsi;
720 	}
721 
722 	/* Check sampling facility authorization and, if not authorized,
723 	 * fall back to other PMUs.  It is safe to check any CPU because
724 	 * the authorization is identical for all configured CPUs.
725 	 */
726 	if (!si.as) {
727 		err = -ENOENT;
728 		goto out;
729 	}
730 
731 	/* Always enable basic sampling */
732 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
733 
734 	/* Check if diagnostic sampling is requested.  Deny if the required
735 	 * sampling authorization is missing.
736 	 */
737 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
738 		if (!si.ad) {
739 			err = -EPERM;
740 			goto out;
741 		}
742 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
743 	}
744 
745 	/* Check and set other sampling flags */
746 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
747 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
748 
749 	/* The sampling information (si) contains information about the
750 	 * min/max sampling intervals and the CPU speed.  So calculate the
751 	 * correct sampling interval and avoid the whole period adjust
752 	 * feedback loop.
753 	 */
754 	rate = 0;
755 	if (attr->freq) {
756 		if (!attr->sample_freq) {
757 			err = -EINVAL;
758 			goto out;
759 		}
760 		rate = freq_to_sample_rate(&si, attr->sample_freq);
761 		rate = hw_limit_rate(&si, rate);
762 		attr->freq = 0;
763 		attr->sample_period = rate;
764 	} else {
765 		/* The min/max sampling rates specifies the valid range
766 		 * of sample periods.  If the specified sample period is
767 		 * out of range, limit the period to the range boundary.
768 		 */
769 		rate = hw_limit_rate(&si, hwc->sample_period);
770 
771 		/* The perf core maintains a maximum sample rate that is
772 		 * configurable through the sysctl interface.  Ensure the
773 		 * sampling rate does not exceed this value.  This also helps
774 		 * to avoid throttling when pushing samples with
775 		 * perf_event_overflow().
776 		 */
777 		if (sample_rate_to_freq(&si, rate) >
778 		      sysctl_perf_event_sample_rate) {
779 			err = -EINVAL;
780 			debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
781 			goto out;
782 		}
783 	}
784 	SAMPL_RATE(hwc) = rate;
785 	hw_init_period(hwc, SAMPL_RATE(hwc));
786 
787 	/* Initialize sample data overflow accounting */
788 	hwc->extra_reg.reg = REG_OVERFLOW;
789 	OVERFLOW_REG(hwc) = 0;
790 
791 	/* Use AUX buffer. No need to allocate it by ourself */
792 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
793 		return 0;
794 
795 	/* Allocate the per-CPU sampling buffer using the CPU information
796 	 * from the event.  If the event is not pinned to a particular
797 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
798 	 * buffers for each online CPU.
799 	 */
800 	if (cpuhw)
801 		/* Event is pinned to a particular CPU */
802 		err = allocate_buffers(cpuhw, hwc);
803 	else {
804 		/* Event is not pinned, allocate sampling buffer on
805 		 * each online CPU
806 		 */
807 		for_each_online_cpu(cpu) {
808 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
809 			err = allocate_buffers(cpuhw, hwc);
810 			if (err)
811 				break;
812 		}
813 	}
814 
815 	/* If PID/TID sampling is active, replace the default overflow
816 	 * handler to extract and resolve the PIDs from the basic-sampling
817 	 * data entries.
818 	 */
819 	if (event->attr.sample_type & PERF_SAMPLE_TID)
820 		if (is_default_overflow_handler(event))
821 			event->overflow_handler = cpumsf_output_event_pid;
822 out:
823 	return err;
824 }
825 
826 static int cpumsf_pmu_event_init(struct perf_event *event)
827 {
828 	int err;
829 
830 	/* No support for taken branch sampling */
831 	if (has_branch_stack(event))
832 		return -EOPNOTSUPP;
833 
834 	switch (event->attr.type) {
835 	case PERF_TYPE_RAW:
836 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
837 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
838 			return -ENOENT;
839 		break;
840 	case PERF_TYPE_HARDWARE:
841 		/* Support sampling of CPU cycles in addition to the
842 		 * counter facility.  However, the counter facility
843 		 * is more precise and, hence, restrict this PMU to
844 		 * sampling events only.
845 		 */
846 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
847 			return -ENOENT;
848 		if (!is_sampling_event(event))
849 			return -ENOENT;
850 		break;
851 	default:
852 		return -ENOENT;
853 	}
854 
855 	/* Check online status of the CPU to which the event is pinned */
856 	if (event->cpu >= 0 && !cpu_online(event->cpu))
857 			return -ENODEV;
858 
859 	/* Force reset of idle/hv excludes regardless of what the
860 	 * user requested.
861 	 */
862 	if (event->attr.exclude_hv)
863 		event->attr.exclude_hv = 0;
864 	if (event->attr.exclude_idle)
865 		event->attr.exclude_idle = 0;
866 
867 	err = __hw_perf_event_init(event);
868 	if (unlikely(err))
869 		if (event->destroy)
870 			event->destroy(event);
871 	return err;
872 }
873 
874 static void cpumsf_pmu_enable(struct pmu *pmu)
875 {
876 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
877 	struct hw_perf_event *hwc;
878 	int err;
879 
880 	if (cpuhw->flags & PMU_F_ENABLED)
881 		return;
882 
883 	if (cpuhw->flags & PMU_F_ERR_MASK)
884 		return;
885 
886 	/* Check whether to extent the sampling buffer.
887 	 *
888 	 * Two conditions trigger an increase of the sampling buffer for a
889 	 * perf event:
890 	 *    1. Postponed buffer allocations from the event initialization.
891 	 *    2. Sampling overflows that contribute to pending allocations.
892 	 *
893 	 * Note that the extend_sampling_buffer() function disables the sampling
894 	 * facility, but it can be fully re-enabled using sampling controls that
895 	 * have been saved in cpumsf_pmu_disable().
896 	 */
897 	if (cpuhw->event) {
898 		hwc = &cpuhw->event->hw;
899 		if (!(SAMPL_DIAG_MODE(hwc))) {
900 			/*
901 			 * Account number of overflow-designated
902 			 * buffer extents
903 			 */
904 			sfb_account_overflows(cpuhw, hwc);
905 			if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
906 				extend_sampling_buffer(&cpuhw->sfb, hwc);
907 		}
908 	}
909 
910 	/* (Re)enable the PMU and sampling facility */
911 	cpuhw->flags |= PMU_F_ENABLED;
912 	barrier();
913 
914 	err = lsctl(&cpuhw->lsctl);
915 	if (err) {
916 		cpuhw->flags &= ~PMU_F_ENABLED;
917 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
918 			1, err);
919 		return;
920 	}
921 
922 	/* Load current program parameter */
923 	lpp(&S390_lowcore.lpp);
924 
925 	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
926 			    "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
927 			    cpuhw->lsctl.ed, cpuhw->lsctl.cd,
928 			    (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
929 }
930 
931 static void cpumsf_pmu_disable(struct pmu *pmu)
932 {
933 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
934 	struct hws_lsctl_request_block inactive;
935 	struct hws_qsi_info_block si;
936 	int err;
937 
938 	if (!(cpuhw->flags & PMU_F_ENABLED))
939 		return;
940 
941 	if (cpuhw->flags & PMU_F_ERR_MASK)
942 		return;
943 
944 	/* Switch off sampling activation control */
945 	inactive = cpuhw->lsctl;
946 	inactive.cs = 0;
947 	inactive.cd = 0;
948 
949 	err = lsctl(&inactive);
950 	if (err) {
951 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
952 			2, err);
953 		return;
954 	}
955 
956 	/* Save state of TEAR and DEAR register contents */
957 	if (!qsi(&si)) {
958 		/* TEAR/DEAR values are valid only if the sampling facility is
959 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
960 		 * for a disabled sampling facility because cpumsf_pmu_enable()
961 		 * controls the enable/disable state.
962 		 */
963 		if (si.es) {
964 			cpuhw->lsctl.tear = si.tear;
965 			cpuhw->lsctl.dear = si.dear;
966 		}
967 	} else
968 		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
969 				    "qsi() failed with err=%i\n", err);
970 
971 	cpuhw->flags &= ~PMU_F_ENABLED;
972 }
973 
974 /* perf_exclude_event() - Filter event
975  * @event:	The perf event
976  * @regs:	pt_regs structure
977  * @sde_regs:	Sample-data-entry (sde) regs structure
978  *
979  * Filter perf events according to their exclude specification.
980  *
981  * Return non-zero if the event shall be excluded.
982  */
983 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
984 			      struct perf_sf_sde_regs *sde_regs)
985 {
986 	if (event->attr.exclude_user && user_mode(regs))
987 		return 1;
988 	if (event->attr.exclude_kernel && !user_mode(regs))
989 		return 1;
990 	if (event->attr.exclude_guest && sde_regs->in_guest)
991 		return 1;
992 	if (event->attr.exclude_host && !sde_regs->in_guest)
993 		return 1;
994 	return 0;
995 }
996 
997 /* perf_push_sample() - Push samples to perf
998  * @event:	The perf event
999  * @sample:	Hardware sample data
1000  *
1001  * Use the hardware sample data to create perf event sample.  The sample
1002  * is the pushed to the event subsystem and the function checks for
1003  * possible event overflows.  If an event overflow occurs, the PMU is
1004  * stopped.
1005  *
1006  * Return non-zero if an event overflow occurred.
1007  */
1008 static int perf_push_sample(struct perf_event *event,
1009 			    struct hws_basic_entry *basic)
1010 {
1011 	int overflow;
1012 	struct pt_regs regs;
1013 	struct perf_sf_sde_regs *sde_regs;
1014 	struct perf_sample_data data;
1015 
1016 	/* Setup perf sample */
1017 	perf_sample_data_init(&data, 0, event->hw.last_period);
1018 
1019 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1020 	 * using the Program Request Alert code.  The regs.int_parm_long
1021 	 * field which is unused contains additional sample-data-entry related
1022 	 * indicators.
1023 	 */
1024 	memset(&regs, 0, sizeof(regs));
1025 	regs.int_code = 0x1407;
1026 	regs.int_parm = CPU_MF_INT_SF_PRA;
1027 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1028 
1029 	psw_bits(regs.psw).ia	= basic->ia;
1030 	psw_bits(regs.psw).dat	= basic->T;
1031 	psw_bits(regs.psw).wait = basic->W;
1032 	psw_bits(regs.psw).pstate = basic->P;
1033 	psw_bits(regs.psw).as	= basic->AS;
1034 
1035 	/*
1036 	 * Use the hardware provided configuration level to decide if the
1037 	 * sample belongs to a guest or host. If that is not available,
1038 	 * fall back to the following heuristics:
1039 	 * A non-zero guest program parameter always indicates a guest
1040 	 * sample. Some early samples or samples from guests without
1041 	 * lpp usage would be misaccounted to the host. We use the asn
1042 	 * value as an addon heuristic to detect most of these guest samples.
1043 	 * If the value differs from 0xffff (the host value), we assume to
1044 	 * be a KVM guest.
1045 	 */
1046 	switch (basic->CL) {
1047 	case 1: /* logical partition */
1048 		sde_regs->in_guest = 0;
1049 		break;
1050 	case 2: /* virtual machine */
1051 		sde_regs->in_guest = 1;
1052 		break;
1053 	default: /* old machine, use heuristics */
1054 		if (basic->gpp || basic->prim_asn != 0xffff)
1055 			sde_regs->in_guest = 1;
1056 		break;
1057 	}
1058 
1059 	/*
1060 	 * Store the PID value from the sample-data-entry to be
1061 	 * processed and resolved by cpumsf_output_event_pid().
1062 	 */
1063 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1064 
1065 	overflow = 0;
1066 	if (perf_exclude_event(event, &regs, sde_regs))
1067 		goto out;
1068 	if (perf_event_overflow(event, &data, &regs)) {
1069 		overflow = 1;
1070 		event->pmu->stop(event, 0);
1071 	}
1072 	perf_event_update_userpage(event);
1073 out:
1074 	return overflow;
1075 }
1076 
1077 static void perf_event_count_update(struct perf_event *event, u64 count)
1078 {
1079 	local64_add(count, &event->count);
1080 }
1081 
1082 static void debug_sample_entry(struct hws_basic_entry *sample,
1083 			       struct hws_trailer_entry *te)
1084 {
1085 	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1086 			    "sampling data entry: te->f=%i basic.def=%04x (%p)\n",
1087 			    te->f, sample->def, sample);
1088 }
1089 
1090 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1091  * @event:	The perf event
1092  * @sdbt:	Sample-data-block table
1093  * @overflow:	Event overflow counter
1094  *
1095  * Walks through a sample-data-block and collects sampling data entries that are
1096  * then pushed to the perf event subsystem.  Depending on the sampling function,
1097  * there can be either basic-sampling or combined-sampling data entries.  A
1098  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1099  * data entry.	The sampling function is determined by the flags in the perf
1100  * event hardware structure.  The function always works with a combined-sampling
1101  * data entry but ignores the the diagnostic portion if it is not available.
1102  *
1103  * Note that the implementation focuses on basic-sampling data entries and, if
1104  * such an entry is not valid, the entire combined-sampling data entry is
1105  * ignored.
1106  *
1107  * The overflow variables counts the number of samples that has been discarded
1108  * due to a perf event overflow.
1109  */
1110 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1111 			       unsigned long long *overflow)
1112 {
1113 	struct hws_trailer_entry *te;
1114 	struct hws_basic_entry *sample;
1115 
1116 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1117 	sample = (struct hws_basic_entry *) *sdbt;
1118 	while ((unsigned long *) sample < (unsigned long *) te) {
1119 		/* Check for an empty sample */
1120 		if (!sample->def)
1121 			break;
1122 
1123 		/* Update perf event period */
1124 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1125 
1126 		/* Check whether sample is valid */
1127 		if (sample->def == 0x0001) {
1128 			/* If an event overflow occurred, the PMU is stopped to
1129 			 * throttle event delivery.  Remaining sample data is
1130 			 * discarded.
1131 			 */
1132 			if (!*overflow) {
1133 				/* Check whether sample is consistent */
1134 				if (sample->I == 0 && sample->W == 0) {
1135 					/* Deliver sample data to perf */
1136 					*overflow = perf_push_sample(event,
1137 								     sample);
1138 				}
1139 			} else
1140 				/* Count discarded samples */
1141 				*overflow += 1;
1142 		} else {
1143 			debug_sample_entry(sample, te);
1144 			/* Sample slot is not yet written or other record.
1145 			 *
1146 			 * This condition can occur if the buffer was reused
1147 			 * from a combined basic- and diagnostic-sampling.
1148 			 * If only basic-sampling is then active, entries are
1149 			 * written into the larger diagnostic entries.
1150 			 * This is typically the case for sample-data-blocks
1151 			 * that are not full.  Stop processing if the first
1152 			 * invalid format was detected.
1153 			 */
1154 			if (!te->f)
1155 				break;
1156 		}
1157 
1158 		/* Reset sample slot and advance to next sample */
1159 		sample->def = 0;
1160 		sample++;
1161 	}
1162 }
1163 
1164 /* hw_perf_event_update() - Process sampling buffer
1165  * @event:	The perf event
1166  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1167  *
1168  * Processes the sampling buffer and create perf event samples.
1169  * The sampling buffer position are retrieved and saved in the TEAR_REG
1170  * register of the specified perf event.
1171  *
1172  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1173  * to also walk through partially filled sample-data-blocks.  It is ignored
1174  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1175  * enforces the processing of full sample-data-blocks only (trailer entries
1176  * with the block-full-indicator bit set).
1177  */
1178 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1179 {
1180 	struct hw_perf_event *hwc = &event->hw;
1181 	struct hws_trailer_entry *te;
1182 	unsigned long *sdbt;
1183 	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1184 	int done;
1185 
1186 	/*
1187 	 * AUX buffer is used when in diagnostic sampling mode.
1188 	 * No perf events/samples are created.
1189 	 */
1190 	if (SAMPL_DIAG_MODE(&event->hw))
1191 		return;
1192 
1193 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1194 		flush_all = 0;
1195 
1196 	sdbt = (unsigned long *) TEAR_REG(hwc);
1197 	done = event_overflow = sampl_overflow = num_sdb = 0;
1198 	while (!done) {
1199 		/* Get the trailer entry of the sample-data-block */
1200 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1201 
1202 		/* Leave loop if no more work to do (block full indicator) */
1203 		if (!te->f) {
1204 			done = 1;
1205 			if (!flush_all)
1206 				break;
1207 		}
1208 
1209 		/* Check the sample overflow count */
1210 		if (te->overflow)
1211 			/* Account sample overflows and, if a particular limit
1212 			 * is reached, extend the sampling buffer.
1213 			 * For details, see sfb_account_overflows().
1214 			 */
1215 			sampl_overflow += te->overflow;
1216 
1217 		/* Timestamps are valid for full sample-data-blocks only */
1218 		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1219 				    "overflow=%llu timestamp=0x%llx\n",
1220 				    sdbt, te->overflow,
1221 				    (te->f) ? trailer_timestamp(te) : 0ULL);
1222 
1223 		/* Collect all samples from a single sample-data-block and
1224 		 * flag if an (perf) event overflow happened.  If so, the PMU
1225 		 * is stopped and remaining samples will be discarded.
1226 		 */
1227 		hw_collect_samples(event, sdbt, &event_overflow);
1228 		num_sdb++;
1229 
1230 		/* Reset trailer (using compare-double-and-swap) */
1231 		do {
1232 			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1233 			te_flags |= SDB_TE_ALERT_REQ_MASK;
1234 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1235 					 te->flags, te->overflow,
1236 					 te_flags, 0ULL));
1237 
1238 		/* Advance to next sample-data-block */
1239 		sdbt++;
1240 		if (is_link_entry(sdbt))
1241 			sdbt = get_next_sdbt(sdbt);
1242 
1243 		/* Update event hardware registers */
1244 		TEAR_REG(hwc) = (unsigned long) sdbt;
1245 
1246 		/* Stop processing sample-data if all samples of the current
1247 		 * sample-data-block were flushed even if it was not full.
1248 		 */
1249 		if (flush_all && done)
1250 			break;
1251 
1252 		/* If an event overflow happened, discard samples by
1253 		 * processing any remaining sample-data-blocks.
1254 		 */
1255 		if (event_overflow)
1256 			flush_all = 1;
1257 	}
1258 
1259 	/* Account sample overflows in the event hardware structure */
1260 	if (sampl_overflow)
1261 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1262 						 sampl_overflow, 1 + num_sdb);
1263 	if (sampl_overflow || event_overflow)
1264 		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1265 				    "overflow stats: sample=%llu event=%llu\n",
1266 				    sampl_overflow, event_overflow);
1267 }
1268 
1269 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1270 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1271 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1272 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1273 
1274 /*
1275  * Get trailer entry by index of SDB.
1276  */
1277 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1278 						 unsigned long index)
1279 {
1280 	unsigned long sdb;
1281 
1282 	index = AUX_SDB_INDEX(aux, index);
1283 	sdb = aux->sdb_index[index];
1284 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1285 }
1286 
1287 /*
1288  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1289  * disabled. Collect the full SDBs in AUX buffer which have not reached
1290  * the point of alert indicator. And ignore the SDBs which are not
1291  * full.
1292  *
1293  * 1. Scan SDBs to see how much data is there and consume them.
1294  * 2. Remove alert indicator in the buffer.
1295  */
1296 static void aux_output_end(struct perf_output_handle *handle)
1297 {
1298 	unsigned long i, range_scan, idx;
1299 	struct aux_buffer *aux;
1300 	struct hws_trailer_entry *te;
1301 
1302 	aux = perf_get_aux(handle);
1303 	if (!aux)
1304 		return;
1305 
1306 	range_scan = AUX_SDB_NUM_ALERT(aux);
1307 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1308 		te = aux_sdb_trailer(aux, idx);
1309 		if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1310 			break;
1311 	}
1312 	/* i is num of SDBs which are full */
1313 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1314 
1315 	/* Remove alert indicators in the buffer */
1316 	te = aux_sdb_trailer(aux, aux->alert_mark);
1317 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1318 
1319 	debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1320 }
1321 
1322 /*
1323  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1324  * is first added to the CPU or rescheduled again to the CPU. It is called
1325  * with pmu disabled.
1326  *
1327  * 1. Reset the trailer of SDBs to get ready for new data.
1328  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1329  *    head(tear/dear).
1330  */
1331 static int aux_output_begin(struct perf_output_handle *handle,
1332 			    struct aux_buffer *aux,
1333 			    struct cpu_hw_sf *cpuhw)
1334 {
1335 	unsigned long range;
1336 	unsigned long i, range_scan, idx;
1337 	unsigned long head, base, offset;
1338 	struct hws_trailer_entry *te;
1339 
1340 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1341 		return -EINVAL;
1342 
1343 	aux->head = handle->head >> PAGE_SHIFT;
1344 	range = (handle->size + 1) >> PAGE_SHIFT;
1345 	if (range <= 1)
1346 		return -ENOMEM;
1347 
1348 	/*
1349 	 * SDBs between aux->head and aux->empty_mark are already ready
1350 	 * for new data. range_scan is num of SDBs not within them.
1351 	 */
1352 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1353 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1354 		idx = aux->empty_mark + 1;
1355 		for (i = 0; i < range_scan; i++, idx++) {
1356 			te = aux_sdb_trailer(aux, idx);
1357 			te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1358 			te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK;
1359 			te->overflow = 0;
1360 		}
1361 		/* Save the position of empty SDBs */
1362 		aux->empty_mark = aux->head + range - 1;
1363 	}
1364 
1365 	/* Set alert indicator */
1366 	aux->alert_mark = aux->head + range/2 - 1;
1367 	te = aux_sdb_trailer(aux, aux->alert_mark);
1368 	te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1369 
1370 	/* Reset hardware buffer head */
1371 	head = AUX_SDB_INDEX(aux, aux->head);
1372 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1373 	offset = head % CPUM_SF_SDB_PER_TABLE;
1374 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1375 	cpuhw->lsctl.dear = aux->sdb_index[head];
1376 
1377 	debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1378 			    "head->alert_mark->empty_mark (num_alert, range)"
1379 			    "[%lx -> %lx -> %lx] (%lx, %lx) "
1380 			    "tear index %lx, tear %lx dear %lx\n",
1381 			    aux->head, aux->alert_mark, aux->empty_mark,
1382 			    AUX_SDB_NUM_ALERT(aux), range,
1383 			    head / CPUM_SF_SDB_PER_TABLE,
1384 			    cpuhw->lsctl.tear,
1385 			    cpuhw->lsctl.dear);
1386 
1387 	return 0;
1388 }
1389 
1390 /*
1391  * Set alert indicator on SDB at index @alert_index while sampler is running.
1392  *
1393  * Return true if successfully.
1394  * Return false if full indicator is already set by hardware sampler.
1395  */
1396 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1397 			  unsigned long long *overflow)
1398 {
1399 	unsigned long long orig_overflow, orig_flags, new_flags;
1400 	struct hws_trailer_entry *te;
1401 
1402 	te = aux_sdb_trailer(aux, alert_index);
1403 	do {
1404 		orig_flags = te->flags;
1405 		orig_overflow = te->overflow;
1406 		*overflow = orig_overflow;
1407 		if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1408 			/*
1409 			 * SDB is already set by hardware.
1410 			 * Abort and try to set somewhere
1411 			 * behind.
1412 			 */
1413 			return false;
1414 		}
1415 		new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1416 	} while (!cmpxchg_double(&te->flags, &te->overflow,
1417 				 orig_flags, orig_overflow,
1418 				 new_flags, 0ULL));
1419 	return true;
1420 }
1421 
1422 /*
1423  * aux_reset_buffer() - Scan and setup SDBs for new samples
1424  * @aux:	The AUX buffer to set
1425  * @range:	The range of SDBs to scan started from aux->head
1426  * @overflow:	Set to overflow count
1427  *
1428  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1429  * marked as empty, check if it is already set full by the hardware sampler.
1430  * If yes, that means new data is already there before we can set an alert
1431  * indicator. Caller should try to set alert indicator to some position behind.
1432  *
1433  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1434  * previously and have already been consumed by user space. Reset these SDBs
1435  * (clear full indicator and alert indicator) for new data.
1436  * If aux->alert_mark fall in this area, just set it. Overflow count is
1437  * recorded while scanning.
1438  *
1439  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1440  * and ready for new samples. So scanning on this area could be skipped.
1441  *
1442  * Return true if alert indicator is set successfully and false if not.
1443  */
1444 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1445 			     unsigned long long *overflow)
1446 {
1447 	unsigned long long orig_overflow, orig_flags, new_flags;
1448 	unsigned long i, range_scan, idx;
1449 	struct hws_trailer_entry *te;
1450 
1451 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1452 		/*
1453 		 * No need to scan. All SDBs in range are marked as empty.
1454 		 * Just set alert indicator. Should check race with hardware
1455 		 * sampler.
1456 		 */
1457 		return aux_set_alert(aux, aux->alert_mark, overflow);
1458 
1459 	if (aux->alert_mark <= aux->empty_mark)
1460 		/*
1461 		 * Set alert indicator on empty SDB. Should check race
1462 		 * with hardware sampler.
1463 		 */
1464 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1465 			return false;
1466 
1467 	/*
1468 	 * Scan the SDBs to clear full and alert indicator used previously.
1469 	 * Start scanning from one SDB behind empty_mark. If the new alert
1470 	 * indicator fall into this range, set it.
1471 	 */
1472 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1473 	idx = aux->empty_mark + 1;
1474 	for (i = 0; i < range_scan; i++, idx++) {
1475 		te = aux_sdb_trailer(aux, idx);
1476 		do {
1477 			orig_flags = te->flags;
1478 			orig_overflow = te->overflow;
1479 			new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1480 			if (idx == aux->alert_mark)
1481 				new_flags |= SDB_TE_ALERT_REQ_MASK;
1482 			else
1483 				new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1484 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1485 					 orig_flags, orig_overflow,
1486 					 new_flags, 0ULL));
1487 		*overflow += orig_overflow;
1488 	}
1489 
1490 	/* Update empty_mark to new position */
1491 	aux->empty_mark = aux->head + range - 1;
1492 
1493 	return true;
1494 }
1495 
1496 /*
1497  * Measurement alert handler for diagnostic mode sampling.
1498  */
1499 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1500 {
1501 	struct aux_buffer *aux;
1502 	int done = 0;
1503 	unsigned long range = 0, size;
1504 	unsigned long long overflow = 0;
1505 	struct perf_output_handle *handle = &cpuhw->handle;
1506 	unsigned long num_sdb;
1507 
1508 	aux = perf_get_aux(handle);
1509 	if (WARN_ON_ONCE(!aux))
1510 		return;
1511 
1512 	/* Inform user space new data arrived */
1513 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1514 	perf_aux_output_end(handle, size);
1515 	num_sdb = aux->sfb.num_sdb;
1516 
1517 	while (!done) {
1518 		/* Get an output handle */
1519 		aux = perf_aux_output_begin(handle, cpuhw->event);
1520 		if (handle->size == 0) {
1521 			pr_err("The AUX buffer with %lu pages for the "
1522 			       "diagnostic-sampling mode is full\n",
1523 				num_sdb);
1524 			debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1525 			break;
1526 		}
1527 		if (WARN_ON_ONCE(!aux))
1528 			return;
1529 
1530 		/* Update head and alert_mark to new position */
1531 		aux->head = handle->head >> PAGE_SHIFT;
1532 		range = (handle->size + 1) >> PAGE_SHIFT;
1533 		if (range == 1)
1534 			aux->alert_mark = aux->head;
1535 		else
1536 			aux->alert_mark = aux->head + range/2 - 1;
1537 
1538 		if (aux_reset_buffer(aux, range, &overflow)) {
1539 			if (!overflow) {
1540 				done = 1;
1541 				break;
1542 			}
1543 			size = range << PAGE_SHIFT;
1544 			perf_aux_output_end(&cpuhw->handle, size);
1545 			pr_err("Sample data caused the AUX buffer with %lu "
1546 			       "pages to overflow\n", num_sdb);
1547 			debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1548 					    "overflow %llx\n",
1549 					    aux->head, range, overflow);
1550 		} else {
1551 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1552 			perf_aux_output_end(&cpuhw->handle, size);
1553 			debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1554 					    "already full, try another\n",
1555 					    aux->head, aux->alert_mark);
1556 		}
1557 	}
1558 
1559 	if (done)
1560 		debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1561 				    "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1562 				    aux->head, aux->alert_mark, aux->empty_mark,
1563 				    AUX_SDB_NUM_ALERT(aux), range);
1564 }
1565 
1566 /*
1567  * Callback when freeing AUX buffers.
1568  */
1569 static void aux_buffer_free(void *data)
1570 {
1571 	struct aux_buffer *aux = data;
1572 	unsigned long i, num_sdbt;
1573 
1574 	if (!aux)
1575 		return;
1576 
1577 	/* Free SDBT. SDB is freed by the caller */
1578 	num_sdbt = aux->sfb.num_sdbt;
1579 	for (i = 0; i < num_sdbt; i++)
1580 		free_page(aux->sdbt_index[i]);
1581 
1582 	kfree(aux->sdbt_index);
1583 	kfree(aux->sdb_index);
1584 	kfree(aux);
1585 
1586 	debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1587 			    "%lu SDBTs\n", num_sdbt);
1588 }
1589 
1590 static void aux_sdb_init(unsigned long sdb)
1591 {
1592 	struct hws_trailer_entry *te;
1593 
1594 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1595 
1596 	/* Save clock base */
1597 	te->clock_base = 1;
1598 	memcpy(&te->progusage2, &tod_clock_base[1], 8);
1599 }
1600 
1601 /*
1602  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1603  * @cpu:	On which to allocate, -1 means current
1604  * @pages:	Array of pointers to buffer pages passed from perf core
1605  * @nr_pages:	Total pages
1606  * @snapshot:	Flag for snapshot mode
1607  *
1608  * This is the callback when setup an event using AUX buffer. Perf tool can
1609  * trigger this by an additional mmap() call on the event. Unlike the buffer
1610  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1611  * the task among online cpus when it is a per-thread event.
1612  *
1613  * Return the private AUX buffer structure if success or NULL if fails.
1614  */
1615 static void *aux_buffer_setup(int cpu, void **pages, int nr_pages,
1616 			      bool snapshot)
1617 {
1618 	struct sf_buffer *sfb;
1619 	struct aux_buffer *aux;
1620 	unsigned long *new, *tail;
1621 	int i, n_sdbt;
1622 
1623 	if (!nr_pages || !pages)
1624 		return NULL;
1625 
1626 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1627 		pr_err("AUX buffer size (%i pages) is larger than the "
1628 		       "maximum sampling buffer limit\n",
1629 		       nr_pages);
1630 		return NULL;
1631 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1632 		pr_err("AUX buffer size (%i pages) is less than the "
1633 		       "minimum sampling buffer limit\n",
1634 		       nr_pages);
1635 		return NULL;
1636 	}
1637 
1638 	/* Allocate aux_buffer struct for the event */
1639 	aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1640 	if (!aux)
1641 		goto no_aux;
1642 	sfb = &aux->sfb;
1643 
1644 	/* Allocate sdbt_index for fast reference */
1645 	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1646 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1647 	if (!aux->sdbt_index)
1648 		goto no_sdbt_index;
1649 
1650 	/* Allocate sdb_index for fast reference */
1651 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1652 	if (!aux->sdb_index)
1653 		goto no_sdb_index;
1654 
1655 	/* Allocate the first SDBT */
1656 	sfb->num_sdbt = 0;
1657 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1658 	if (!sfb->sdbt)
1659 		goto no_sdbt;
1660 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1661 	tail = sfb->tail = sfb->sdbt;
1662 
1663 	/*
1664 	 * Link the provided pages of AUX buffer to SDBT.
1665 	 * Allocate SDBT if needed.
1666 	 */
1667 	for (i = 0; i < nr_pages; i++, tail++) {
1668 		if (require_table_link(tail)) {
1669 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1670 			if (!new)
1671 				goto no_sdbt;
1672 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1673 			/* Link current page to tail of chain */
1674 			*tail = (unsigned long)(void *) new + 1;
1675 			tail = new;
1676 		}
1677 		/* Tail is the entry in a SDBT */
1678 		*tail = (unsigned long)pages[i];
1679 		aux->sdb_index[i] = (unsigned long)pages[i];
1680 		aux_sdb_init((unsigned long)pages[i]);
1681 	}
1682 	sfb->num_sdb = nr_pages;
1683 
1684 	/* Link the last entry in the SDBT to the first SDBT */
1685 	*tail = (unsigned long) sfb->sdbt + 1;
1686 	sfb->tail = tail;
1687 
1688 	/*
1689 	 * Initial all SDBs are zeroed. Mark it as empty.
1690 	 * So there is no need to clear the full indicator
1691 	 * when this event is first added.
1692 	 */
1693 	aux->empty_mark = sfb->num_sdb - 1;
1694 
1695 	debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1696 			    " and %lu SDBs\n",
1697 			    sfb->num_sdbt, sfb->num_sdb);
1698 
1699 	return aux;
1700 
1701 no_sdbt:
1702 	/* SDBs (AUX buffer pages) are freed by caller */
1703 	for (i = 0; i < sfb->num_sdbt; i++)
1704 		free_page(aux->sdbt_index[i]);
1705 	kfree(aux->sdb_index);
1706 no_sdb_index:
1707 	kfree(aux->sdbt_index);
1708 no_sdbt_index:
1709 	kfree(aux);
1710 no_aux:
1711 	return NULL;
1712 }
1713 
1714 static void cpumsf_pmu_read(struct perf_event *event)
1715 {
1716 	/* Nothing to do ... updates are interrupt-driven */
1717 }
1718 
1719 /* Activate sampling control.
1720  * Next call of pmu_enable() starts sampling.
1721  */
1722 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1723 {
1724 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1725 
1726 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1727 		return;
1728 
1729 	if (flags & PERF_EF_RELOAD)
1730 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1731 
1732 	perf_pmu_disable(event->pmu);
1733 	event->hw.state = 0;
1734 	cpuhw->lsctl.cs = 1;
1735 	if (SAMPL_DIAG_MODE(&event->hw))
1736 		cpuhw->lsctl.cd = 1;
1737 	perf_pmu_enable(event->pmu);
1738 }
1739 
1740 /* Deactivate sampling control.
1741  * Next call of pmu_enable() stops sampling.
1742  */
1743 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1744 {
1745 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1746 
1747 	if (event->hw.state & PERF_HES_STOPPED)
1748 		return;
1749 
1750 	perf_pmu_disable(event->pmu);
1751 	cpuhw->lsctl.cs = 0;
1752 	cpuhw->lsctl.cd = 0;
1753 	event->hw.state |= PERF_HES_STOPPED;
1754 
1755 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1756 		hw_perf_event_update(event, 1);
1757 		event->hw.state |= PERF_HES_UPTODATE;
1758 	}
1759 	perf_pmu_enable(event->pmu);
1760 }
1761 
1762 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1763 {
1764 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1765 	struct aux_buffer *aux;
1766 	int err;
1767 
1768 	if (cpuhw->flags & PMU_F_IN_USE)
1769 		return -EAGAIN;
1770 
1771 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1772 		return -EINVAL;
1773 
1774 	err = 0;
1775 	perf_pmu_disable(event->pmu);
1776 
1777 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1778 
1779 	/* Set up sampling controls.  Always program the sampling register
1780 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1781 	 * that is used by hw_perf_event_update() to store the sampling buffer
1782 	 * position after samples have been flushed.
1783 	 */
1784 	cpuhw->lsctl.s = 0;
1785 	cpuhw->lsctl.h = 1;
1786 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1787 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1788 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1789 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1790 		hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1791 	}
1792 
1793 	/* Ensure sampling functions are in the disabled state.  If disabled,
1794 	 * switch on sampling enable control. */
1795 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1796 		err = -EAGAIN;
1797 		goto out;
1798 	}
1799 	if (SAMPL_DIAG_MODE(&event->hw)) {
1800 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1801 		if (!aux) {
1802 			err = -EINVAL;
1803 			goto out;
1804 		}
1805 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1806 		if (err)
1807 			goto out;
1808 		cpuhw->lsctl.ed = 1;
1809 	}
1810 	cpuhw->lsctl.es = 1;
1811 
1812 	/* Set in_use flag and store event */
1813 	cpuhw->event = event;
1814 	cpuhw->flags |= PMU_F_IN_USE;
1815 
1816 	if (flags & PERF_EF_START)
1817 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1818 out:
1819 	perf_event_update_userpage(event);
1820 	perf_pmu_enable(event->pmu);
1821 	return err;
1822 }
1823 
1824 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1825 {
1826 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1827 
1828 	perf_pmu_disable(event->pmu);
1829 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1830 
1831 	cpuhw->lsctl.es = 0;
1832 	cpuhw->lsctl.ed = 0;
1833 	cpuhw->flags &= ~PMU_F_IN_USE;
1834 	cpuhw->event = NULL;
1835 
1836 	if (SAMPL_DIAG_MODE(&event->hw))
1837 		aux_output_end(&cpuhw->handle);
1838 	perf_event_update_userpage(event);
1839 	perf_pmu_enable(event->pmu);
1840 }
1841 
1842 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1843 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1844 
1845 /* Attribute list for CPU_SF.
1846  *
1847  * The availablitiy depends on the CPU_MF sampling facility authorization
1848  * for basic + diagnositic samples. This is determined at initialization
1849  * time by the sampling facility device driver.
1850  * If the authorization for basic samples is turned off, it should be
1851  * also turned off for diagnostic sampling.
1852  *
1853  * During initialization of the device driver, check the authorization
1854  * level for diagnostic sampling and installs the attribute
1855  * file for diagnostic sampling if necessary.
1856  *
1857  * For now install a placeholder to reference all possible attributes:
1858  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
1859  * Add another entry for the final NULL pointer.
1860  */
1861 enum {
1862 	SF_CYCLES_BASIC_ATTR_IDX = 0,
1863 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
1864 	SF_CYCLES_ATTR_MAX
1865 };
1866 
1867 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
1868 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
1869 };
1870 
1871 PMU_FORMAT_ATTR(event, "config:0-63");
1872 
1873 static struct attribute *cpumsf_pmu_format_attr[] = {
1874 	&format_attr_event.attr,
1875 	NULL,
1876 };
1877 
1878 static struct attribute_group cpumsf_pmu_events_group = {
1879 	.name = "events",
1880 	.attrs = cpumsf_pmu_events_attr,
1881 };
1882 static struct attribute_group cpumsf_pmu_format_group = {
1883 	.name = "format",
1884 	.attrs = cpumsf_pmu_format_attr,
1885 };
1886 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1887 	&cpumsf_pmu_events_group,
1888 	&cpumsf_pmu_format_group,
1889 	NULL,
1890 };
1891 
1892 static struct pmu cpumf_sampling = {
1893 	.pmu_enable   = cpumsf_pmu_enable,
1894 	.pmu_disable  = cpumsf_pmu_disable,
1895 
1896 	.event_init   = cpumsf_pmu_event_init,
1897 	.add	      = cpumsf_pmu_add,
1898 	.del	      = cpumsf_pmu_del,
1899 
1900 	.start	      = cpumsf_pmu_start,
1901 	.stop	      = cpumsf_pmu_stop,
1902 	.read	      = cpumsf_pmu_read,
1903 
1904 	.attr_groups  = cpumsf_pmu_attr_groups,
1905 
1906 	.setup_aux    = aux_buffer_setup,
1907 	.free_aux     = aux_buffer_free,
1908 };
1909 
1910 static void cpumf_measurement_alert(struct ext_code ext_code,
1911 				    unsigned int alert, unsigned long unused)
1912 {
1913 	struct cpu_hw_sf *cpuhw;
1914 
1915 	if (!(alert & CPU_MF_INT_SF_MASK))
1916 		return;
1917 	inc_irq_stat(IRQEXT_CMS);
1918 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
1919 
1920 	/* Measurement alerts are shared and might happen when the PMU
1921 	 * is not reserved.  Ignore these alerts in this case. */
1922 	if (!(cpuhw->flags & PMU_F_RESERVED))
1923 		return;
1924 
1925 	/* The processing below must take care of multiple alert events that
1926 	 * might be indicated concurrently. */
1927 
1928 	/* Program alert request */
1929 	if (alert & CPU_MF_INT_SF_PRA) {
1930 		if (cpuhw->flags & PMU_F_IN_USE)
1931 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
1932 				hw_collect_aux(cpuhw);
1933 			else
1934 				hw_perf_event_update(cpuhw->event, 0);
1935 		else
1936 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
1937 	}
1938 
1939 	/* Report measurement alerts only for non-PRA codes */
1940 	if (alert != CPU_MF_INT_SF_PRA)
1941 		debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);
1942 
1943 	/* Sampling authorization change request */
1944 	if (alert & CPU_MF_INT_SF_SACA)
1945 		qsi(&cpuhw->qsi);
1946 
1947 	/* Loss of sample data due to high-priority machine activities */
1948 	if (alert & CPU_MF_INT_SF_LSDA) {
1949 		pr_err("Sample data was lost\n");
1950 		cpuhw->flags |= PMU_F_ERR_LSDA;
1951 		sf_disable();
1952 	}
1953 
1954 	/* Invalid sampling buffer entry */
1955 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
1956 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
1957 		       alert);
1958 		cpuhw->flags |= PMU_F_ERR_IBE;
1959 		sf_disable();
1960 	}
1961 }
1962 static int cpusf_pmu_setup(unsigned int cpu, int flags)
1963 {
1964 	/* Ignore the notification if no events are scheduled on the PMU.
1965 	 * This might be racy...
1966 	 */
1967 	if (!atomic_read(&num_events))
1968 		return 0;
1969 
1970 	local_irq_disable();
1971 	setup_pmc_cpu(&flags);
1972 	local_irq_enable();
1973 	return 0;
1974 }
1975 
1976 static int s390_pmu_sf_online_cpu(unsigned int cpu)
1977 {
1978 	return cpusf_pmu_setup(cpu, PMC_INIT);
1979 }
1980 
1981 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
1982 {
1983 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
1984 }
1985 
1986 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
1987 {
1988 	if (!cpum_sf_avail())
1989 		return -ENODEV;
1990 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
1991 }
1992 
1993 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
1994 {
1995 	int rc;
1996 	unsigned long min, max;
1997 
1998 	if (!cpum_sf_avail())
1999 		return -ENODEV;
2000 	if (!val || !strlen(val))
2001 		return -EINVAL;
2002 
2003 	/* Valid parameter values: "min,max" or "max" */
2004 	min = CPUM_SF_MIN_SDB;
2005 	max = CPUM_SF_MAX_SDB;
2006 	if (strchr(val, ','))
2007 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2008 	else
2009 		rc = kstrtoul(val, 10, &max);
2010 
2011 	if (min < 2 || min >= max || max > get_num_physpages())
2012 		rc = -EINVAL;
2013 	if (rc)
2014 		return rc;
2015 
2016 	sfb_set_limits(min, max);
2017 	pr_info("The sampling buffer limits have changed to: "
2018 		"min=%lu max=%lu (diag=x%lu)\n",
2019 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2020 	return 0;
2021 }
2022 
2023 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2024 static const struct kernel_param_ops param_ops_sfb_size = {
2025 	.set = param_set_sfb_size,
2026 	.get = param_get_sfb_size,
2027 };
2028 
2029 #define RS_INIT_FAILURE_QSI	  0x0001
2030 #define RS_INIT_FAILURE_BSDES	  0x0002
2031 #define RS_INIT_FAILURE_ALRT	  0x0003
2032 #define RS_INIT_FAILURE_PERF	  0x0004
2033 static void __init pr_cpumsf_err(unsigned int reason)
2034 {
2035 	pr_err("Sampling facility support for perf is not available: "
2036 	       "reason=%04x\n", reason);
2037 }
2038 
2039 static int __init init_cpum_sampling_pmu(void)
2040 {
2041 	struct hws_qsi_info_block si;
2042 	int err;
2043 
2044 	if (!cpum_sf_avail())
2045 		return -ENODEV;
2046 
2047 	memset(&si, 0, sizeof(si));
2048 	if (qsi(&si)) {
2049 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2050 		return -ENODEV;
2051 	}
2052 
2053 	if (!si.as && !si.ad)
2054 		return -ENODEV;
2055 
2056 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2057 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2058 		return -EINVAL;
2059 	}
2060 
2061 	if (si.ad) {
2062 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2063 		/* Sampling of diagnostic data authorized,
2064 		 * install event into attribute list of PMU device.
2065 		 */
2066 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2067 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2068 	}
2069 
2070 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2071 	if (!sfdbg) {
2072 		pr_err("Registering for s390dbf failed\n");
2073 		return -ENOMEM;
2074 	}
2075 	debug_register_view(sfdbg, &debug_sprintf_view);
2076 
2077 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2078 				    cpumf_measurement_alert);
2079 	if (err) {
2080 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2081 		debug_unregister(sfdbg);
2082 		goto out;
2083 	}
2084 
2085 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2086 	if (err) {
2087 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2088 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2089 					cpumf_measurement_alert);
2090 		debug_unregister(sfdbg);
2091 		goto out;
2092 	}
2093 
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);
2096 out:
2097 	return err;
2098 }
2099 arch_initcall(init_cpum_sampling_pmu);
2100 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);
2101