xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision 8e20ba2e)
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 /* Number of perf events counting hardware events */
518 static atomic_t num_events;
519 /* Used to avoid races in calling reserve/release_cpumf_hardware */
520 static DEFINE_MUTEX(pmc_reserve_mutex);
521 
522 #define PMC_INIT      0
523 #define PMC_RELEASE   1
524 #define PMC_FAILURE   2
525 static void setup_pmc_cpu(void *flags)
526 {
527 	int err;
528 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
529 
530 	err = 0;
531 	switch (*((int *) flags)) {
532 	case PMC_INIT:
533 		memset(cpusf, 0, sizeof(*cpusf));
534 		err = qsi(&cpusf->qsi);
535 		if (err)
536 			break;
537 		cpusf->flags |= PMU_F_RESERVED;
538 		err = sf_disable();
539 		if (err)
540 			pr_err("Switching off the sampling facility failed "
541 			       "with rc=%i\n", err);
542 		debug_sprintf_event(sfdbg, 5,
543 				    "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
544 		break;
545 	case PMC_RELEASE:
546 		cpusf->flags &= ~PMU_F_RESERVED;
547 		err = sf_disable();
548 		if (err) {
549 			pr_err("Switching off the sampling facility failed "
550 			       "with rc=%i\n", err);
551 		} else
552 			deallocate_buffers(cpusf);
553 		debug_sprintf_event(sfdbg, 5,
554 				    "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
555 		break;
556 	}
557 	if (err)
558 		*((int *) flags) |= PMC_FAILURE;
559 }
560 
561 static void release_pmc_hardware(void)
562 {
563 	int flags = PMC_RELEASE;
564 
565 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
566 	on_each_cpu(setup_pmc_cpu, &flags, 1);
567 }
568 
569 static int reserve_pmc_hardware(void)
570 {
571 	int flags = PMC_INIT;
572 
573 	on_each_cpu(setup_pmc_cpu, &flags, 1);
574 	if (flags & PMC_FAILURE) {
575 		release_pmc_hardware();
576 		return -ENODEV;
577 	}
578 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
579 
580 	return 0;
581 }
582 
583 static void hw_perf_event_destroy(struct perf_event *event)
584 {
585 	/* Release PMC if this is the last perf event */
586 	if (!atomic_add_unless(&num_events, -1, 1)) {
587 		mutex_lock(&pmc_reserve_mutex);
588 		if (atomic_dec_return(&num_events) == 0)
589 			release_pmc_hardware();
590 		mutex_unlock(&pmc_reserve_mutex);
591 	}
592 }
593 
594 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
595 {
596 	hwc->sample_period = period;
597 	hwc->last_period = hwc->sample_period;
598 	local64_set(&hwc->period_left, hwc->sample_period);
599 }
600 
601 static void hw_reset_registers(struct hw_perf_event *hwc,
602 			       unsigned long *sdbt_origin)
603 {
604 	/* (Re)set to first sample-data-block-table */
605 	TEAR_REG(hwc) = (unsigned long) sdbt_origin;
606 }
607 
608 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
609 				   unsigned long rate)
610 {
611 	return clamp_t(unsigned long, rate,
612 		       si->min_sampl_rate, si->max_sampl_rate);
613 }
614 
615 static u32 cpumsf_pid_type(struct perf_event *event,
616 			   u32 pid, enum pid_type type)
617 {
618 	struct task_struct *tsk;
619 
620 	/* Idle process */
621 	if (!pid)
622 		goto out;
623 
624 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
625 	pid = -1;
626 	if (tsk) {
627 		/*
628 		 * Only top level events contain the pid namespace in which
629 		 * they are created.
630 		 */
631 		if (event->parent)
632 			event = event->parent;
633 		pid = __task_pid_nr_ns(tsk, type, event->ns);
634 		/*
635 		 * See also 1d953111b648
636 		 * "perf/core: Don't report zero PIDs for exiting tasks".
637 		 */
638 		if (!pid && !pid_alive(tsk))
639 			pid = -1;
640 	}
641 out:
642 	return pid;
643 }
644 
645 static void cpumsf_output_event_pid(struct perf_event *event,
646 				    struct perf_sample_data *data,
647 				    struct pt_regs *regs)
648 {
649 	u32 pid;
650 	struct perf_event_header header;
651 	struct perf_output_handle handle;
652 
653 	/*
654 	 * Obtain the PID from the basic-sampling data entry and
655 	 * correct the data->tid_entry.pid value.
656 	 */
657 	pid = data->tid_entry.pid;
658 
659 	/* Protect callchain buffers, tasks */
660 	rcu_read_lock();
661 
662 	perf_prepare_sample(&header, data, event, regs);
663 	if (perf_output_begin(&handle, event, header.size))
664 		goto out;
665 
666 	/* Update the process ID (see also kernel/events/core.c) */
667 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
668 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
669 
670 	perf_output_sample(&handle, &header, data, event);
671 	perf_output_end(&handle);
672 out:
673 	rcu_read_unlock();
674 }
675 
676 static unsigned long getrate(bool freq, unsigned long sample,
677 			     struct hws_qsi_info_block *si)
678 {
679 	unsigned long rate;
680 
681 	if (freq) {
682 		rate = freq_to_sample_rate(si, sample);
683 		rate = hw_limit_rate(si, rate);
684 	} else {
685 		/* The min/max sampling rates specifies the valid range
686 		 * of sample periods.  If the specified sample period is
687 		 * out of range, limit the period to the range boundary.
688 		 */
689 		rate = hw_limit_rate(si, sample);
690 
691 		/* The perf core maintains a maximum sample rate that is
692 		 * configurable through the sysctl interface.  Ensure the
693 		 * sampling rate does not exceed this value.  This also helps
694 		 * to avoid throttling when pushing samples with
695 		 * perf_event_overflow().
696 		 */
697 		if (sample_rate_to_freq(si, rate) >
698 		    sysctl_perf_event_sample_rate) {
699 			debug_sprintf_event(sfdbg, 1,
700 					    "Sampling rate exceeds maximum "
701 					    "perf sample rate\n");
702 			rate = 0;
703 		}
704 	}
705 	return rate;
706 }
707 
708 /* The sampling information (si) contains information about the
709  * min/max sampling intervals and the CPU speed.  So calculate the
710  * correct sampling interval and avoid the whole period adjust
711  * feedback loop.
712  *
713  * Since the CPU Measurement sampling facility can not handle frequency
714  * calculate the sampling interval when frequency is specified using
715  * this formula:
716  *	interval := cpu_speed * 1000000 / sample_freq
717  *
718  * Returns errno on bad input and zero on success with parameter interval
719  * set to the correct sampling rate.
720  *
721  * Note: This function turns off freq bit to avoid calling function
722  * perf_adjust_period(). This causes frequency adjustment in the common
723  * code part which causes tremendous variations in the counter values.
724  */
725 static int __hw_perf_event_init_rate(struct perf_event *event,
726 				     struct hws_qsi_info_block *si)
727 {
728 	struct perf_event_attr *attr = &event->attr;
729 	struct hw_perf_event *hwc = &event->hw;
730 	unsigned long rate;
731 
732 	if (attr->freq) {
733 		if (!attr->sample_freq)
734 			return -EINVAL;
735 		rate = getrate(attr->freq, attr->sample_freq, si);
736 		attr->freq = 0;		/* Don't call  perf_adjust_period() */
737 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
738 	} else {
739 		rate = getrate(attr->freq, attr->sample_period, si);
740 		if (!rate)
741 			return -EINVAL;
742 	}
743 	attr->sample_period = rate;
744 	SAMPL_RATE(hwc) = rate;
745 	hw_init_period(hwc, SAMPL_RATE(hwc));
746 	debug_sprintf_event(sfdbg, 4, "__hw_perf_event_init_rate:"
747 			    "cpu:%d period:%llx freq:%d,%#lx\n", event->cpu,
748 			    event->attr.sample_period, event->attr.freq,
749 			    SAMPLE_FREQ_MODE(hwc));
750 	return 0;
751 }
752 
753 static int __hw_perf_event_init(struct perf_event *event)
754 {
755 	struct cpu_hw_sf *cpuhw;
756 	struct hws_qsi_info_block si;
757 	struct perf_event_attr *attr = &event->attr;
758 	struct hw_perf_event *hwc = &event->hw;
759 	int cpu, err;
760 
761 	/* Reserve CPU-measurement sampling facility */
762 	err = 0;
763 	if (!atomic_inc_not_zero(&num_events)) {
764 		mutex_lock(&pmc_reserve_mutex);
765 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
766 			err = -EBUSY;
767 		else
768 			atomic_inc(&num_events);
769 		mutex_unlock(&pmc_reserve_mutex);
770 	}
771 	event->destroy = hw_perf_event_destroy;
772 
773 	if (err)
774 		goto out;
775 
776 	/* Access per-CPU sampling information (query sampling info) */
777 	/*
778 	 * The event->cpu value can be -1 to count on every CPU, for example,
779 	 * when attaching to a task.  If this is specified, use the query
780 	 * sampling info from the current CPU, otherwise use event->cpu to
781 	 * retrieve the per-CPU information.
782 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
783 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
784 	 */
785 	memset(&si, 0, sizeof(si));
786 	cpuhw = NULL;
787 	if (event->cpu == -1)
788 		qsi(&si);
789 	else {
790 		/* Event is pinned to a particular CPU, retrieve the per-CPU
791 		 * sampling structure for accessing the CPU-specific QSI.
792 		 */
793 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
794 		si = cpuhw->qsi;
795 	}
796 
797 	/* Check sampling facility authorization and, if not authorized,
798 	 * fall back to other PMUs.  It is safe to check any CPU because
799 	 * the authorization is identical for all configured CPUs.
800 	 */
801 	if (!si.as) {
802 		err = -ENOENT;
803 		goto out;
804 	}
805 
806 	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
807 		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
808 		err = -EBUSY;
809 		goto out;
810 	}
811 
812 	/* Always enable basic sampling */
813 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
814 
815 	/* Check if diagnostic sampling is requested.  Deny if the required
816 	 * sampling authorization is missing.
817 	 */
818 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
819 		if (!si.ad) {
820 			err = -EPERM;
821 			goto out;
822 		}
823 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
824 	}
825 
826 	/* Check and set other sampling flags */
827 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
828 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
829 
830 	err =  __hw_perf_event_init_rate(event, &si);
831 	if (err)
832 		goto out;
833 
834 	/* Initialize sample data overflow accounting */
835 	hwc->extra_reg.reg = REG_OVERFLOW;
836 	OVERFLOW_REG(hwc) = 0;
837 
838 	/* Use AUX buffer. No need to allocate it by ourself */
839 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
840 		return 0;
841 
842 	/* Allocate the per-CPU sampling buffer using the CPU information
843 	 * from the event.  If the event is not pinned to a particular
844 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
845 	 * buffers for each online CPU.
846 	 */
847 	if (cpuhw)
848 		/* Event is pinned to a particular CPU */
849 		err = allocate_buffers(cpuhw, hwc);
850 	else {
851 		/* Event is not pinned, allocate sampling buffer on
852 		 * each online CPU
853 		 */
854 		for_each_online_cpu(cpu) {
855 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
856 			err = allocate_buffers(cpuhw, hwc);
857 			if (err)
858 				break;
859 		}
860 	}
861 
862 	/* If PID/TID sampling is active, replace the default overflow
863 	 * handler to extract and resolve the PIDs from the basic-sampling
864 	 * data entries.
865 	 */
866 	if (event->attr.sample_type & PERF_SAMPLE_TID)
867 		if (is_default_overflow_handler(event))
868 			event->overflow_handler = cpumsf_output_event_pid;
869 out:
870 	return err;
871 }
872 
873 static int cpumsf_pmu_event_init(struct perf_event *event)
874 {
875 	int err;
876 
877 	/* No support for taken branch sampling */
878 	if (has_branch_stack(event))
879 		return -EOPNOTSUPP;
880 
881 	switch (event->attr.type) {
882 	case PERF_TYPE_RAW:
883 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
884 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
885 			return -ENOENT;
886 		break;
887 	case PERF_TYPE_HARDWARE:
888 		/* Support sampling of CPU cycles in addition to the
889 		 * counter facility.  However, the counter facility
890 		 * is more precise and, hence, restrict this PMU to
891 		 * sampling events only.
892 		 */
893 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
894 			return -ENOENT;
895 		if (!is_sampling_event(event))
896 			return -ENOENT;
897 		break;
898 	default:
899 		return -ENOENT;
900 	}
901 
902 	/* Check online status of the CPU to which the event is pinned */
903 	if (event->cpu >= 0 && !cpu_online(event->cpu))
904 		return -ENODEV;
905 
906 	/* Force reset of idle/hv excludes regardless of what the
907 	 * user requested.
908 	 */
909 	if (event->attr.exclude_hv)
910 		event->attr.exclude_hv = 0;
911 	if (event->attr.exclude_idle)
912 		event->attr.exclude_idle = 0;
913 
914 	err = __hw_perf_event_init(event);
915 	if (unlikely(err))
916 		if (event->destroy)
917 			event->destroy(event);
918 	return err;
919 }
920 
921 static void cpumsf_pmu_enable(struct pmu *pmu)
922 {
923 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
924 	struct hw_perf_event *hwc;
925 	int err;
926 
927 	if (cpuhw->flags & PMU_F_ENABLED)
928 		return;
929 
930 	if (cpuhw->flags & PMU_F_ERR_MASK)
931 		return;
932 
933 	/* Check whether to extent the sampling buffer.
934 	 *
935 	 * Two conditions trigger an increase of the sampling buffer for a
936 	 * perf event:
937 	 *    1. Postponed buffer allocations from the event initialization.
938 	 *    2. Sampling overflows that contribute to pending allocations.
939 	 *
940 	 * Note that the extend_sampling_buffer() function disables the sampling
941 	 * facility, but it can be fully re-enabled using sampling controls that
942 	 * have been saved in cpumsf_pmu_disable().
943 	 */
944 	if (cpuhw->event) {
945 		hwc = &cpuhw->event->hw;
946 		if (!(SAMPL_DIAG_MODE(hwc))) {
947 			/*
948 			 * Account number of overflow-designated
949 			 * buffer extents
950 			 */
951 			sfb_account_overflows(cpuhw, hwc);
952 			if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
953 				extend_sampling_buffer(&cpuhw->sfb, hwc);
954 		}
955 		/* Rate may be adjusted with ioctl() */
956 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
957 	}
958 
959 	/* (Re)enable the PMU and sampling facility */
960 	cpuhw->flags |= PMU_F_ENABLED;
961 	barrier();
962 
963 	err = lsctl(&cpuhw->lsctl);
964 	if (err) {
965 		cpuhw->flags &= ~PMU_F_ENABLED;
966 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
967 			1, err);
968 		return;
969 	}
970 
971 	/* Load current program parameter */
972 	lpp(&S390_lowcore.lpp);
973 
974 	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
975 			    "interval:%lx tear=%p dear=%p\n",
976 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
977 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
978 			    (void *) cpuhw->lsctl.tear,
979 			    (void *) cpuhw->lsctl.dear);
980 }
981 
982 static void cpumsf_pmu_disable(struct pmu *pmu)
983 {
984 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
985 	struct hws_lsctl_request_block inactive;
986 	struct hws_qsi_info_block si;
987 	int err;
988 
989 	if (!(cpuhw->flags & PMU_F_ENABLED))
990 		return;
991 
992 	if (cpuhw->flags & PMU_F_ERR_MASK)
993 		return;
994 
995 	/* Switch off sampling activation control */
996 	inactive = cpuhw->lsctl;
997 	inactive.cs = 0;
998 	inactive.cd = 0;
999 
1000 	err = lsctl(&inactive);
1001 	if (err) {
1002 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
1003 			2, err);
1004 		return;
1005 	}
1006 
1007 	/* Save state of TEAR and DEAR register contents */
1008 	if (!qsi(&si)) {
1009 		/* TEAR/DEAR values are valid only if the sampling facility is
1010 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1011 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1012 		 * controls the enable/disable state.
1013 		 */
1014 		if (si.es) {
1015 			cpuhw->lsctl.tear = si.tear;
1016 			cpuhw->lsctl.dear = si.dear;
1017 		}
1018 	} else
1019 		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
1020 				    "qsi() failed with err=%i\n", err);
1021 
1022 	cpuhw->flags &= ~PMU_F_ENABLED;
1023 }
1024 
1025 /* perf_exclude_event() - Filter event
1026  * @event:	The perf event
1027  * @regs:	pt_regs structure
1028  * @sde_regs:	Sample-data-entry (sde) regs structure
1029  *
1030  * Filter perf events according to their exclude specification.
1031  *
1032  * Return non-zero if the event shall be excluded.
1033  */
1034 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1035 			      struct perf_sf_sde_regs *sde_regs)
1036 {
1037 	if (event->attr.exclude_user && user_mode(regs))
1038 		return 1;
1039 	if (event->attr.exclude_kernel && !user_mode(regs))
1040 		return 1;
1041 	if (event->attr.exclude_guest && sde_regs->in_guest)
1042 		return 1;
1043 	if (event->attr.exclude_host && !sde_regs->in_guest)
1044 		return 1;
1045 	return 0;
1046 }
1047 
1048 /* perf_push_sample() - Push samples to perf
1049  * @event:	The perf event
1050  * @sample:	Hardware sample data
1051  *
1052  * Use the hardware sample data to create perf event sample.  The sample
1053  * is the pushed to the event subsystem and the function checks for
1054  * possible event overflows.  If an event overflow occurs, the PMU is
1055  * stopped.
1056  *
1057  * Return non-zero if an event overflow occurred.
1058  */
1059 static int perf_push_sample(struct perf_event *event,
1060 			    struct hws_basic_entry *basic)
1061 {
1062 	int overflow;
1063 	struct pt_regs regs;
1064 	struct perf_sf_sde_regs *sde_regs;
1065 	struct perf_sample_data data;
1066 
1067 	/* Setup perf sample */
1068 	perf_sample_data_init(&data, 0, event->hw.last_period);
1069 
1070 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1071 	 * using the Program Request Alert code.  The regs.int_parm_long
1072 	 * field which is unused contains additional sample-data-entry related
1073 	 * indicators.
1074 	 */
1075 	memset(&regs, 0, sizeof(regs));
1076 	regs.int_code = 0x1407;
1077 	regs.int_parm = CPU_MF_INT_SF_PRA;
1078 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1079 
1080 	psw_bits(regs.psw).ia	= basic->ia;
1081 	psw_bits(regs.psw).dat	= basic->T;
1082 	psw_bits(regs.psw).wait = basic->W;
1083 	psw_bits(regs.psw).pstate = basic->P;
1084 	psw_bits(regs.psw).as	= basic->AS;
1085 
1086 	/*
1087 	 * Use the hardware provided configuration level to decide if the
1088 	 * sample belongs to a guest or host. If that is not available,
1089 	 * fall back to the following heuristics:
1090 	 * A non-zero guest program parameter always indicates a guest
1091 	 * sample. Some early samples or samples from guests without
1092 	 * lpp usage would be misaccounted to the host. We use the asn
1093 	 * value as an addon heuristic to detect most of these guest samples.
1094 	 * If the value differs from 0xffff (the host value), we assume to
1095 	 * be a KVM guest.
1096 	 */
1097 	switch (basic->CL) {
1098 	case 1: /* logical partition */
1099 		sde_regs->in_guest = 0;
1100 		break;
1101 	case 2: /* virtual machine */
1102 		sde_regs->in_guest = 1;
1103 		break;
1104 	default: /* old machine, use heuristics */
1105 		if (basic->gpp || basic->prim_asn != 0xffff)
1106 			sde_regs->in_guest = 1;
1107 		break;
1108 	}
1109 
1110 	/*
1111 	 * Store the PID value from the sample-data-entry to be
1112 	 * processed and resolved by cpumsf_output_event_pid().
1113 	 */
1114 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1115 
1116 	overflow = 0;
1117 	if (perf_exclude_event(event, &regs, sde_regs))
1118 		goto out;
1119 	if (perf_event_overflow(event, &data, &regs)) {
1120 		overflow = 1;
1121 		event->pmu->stop(event, 0);
1122 	}
1123 	perf_event_update_userpage(event);
1124 out:
1125 	return overflow;
1126 }
1127 
1128 static void perf_event_count_update(struct perf_event *event, u64 count)
1129 {
1130 	local64_add(count, &event->count);
1131 }
1132 
1133 static void debug_sample_entry(struct hws_basic_entry *sample,
1134 			       struct hws_trailer_entry *te)
1135 {
1136 	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1137 			    "sampling data entry: te->f=%i basic.def=%04x "
1138 			    "(%p)\n",
1139 			    te->f, sample->def, sample);
1140 }
1141 
1142 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1143  * @event:	The perf event
1144  * @sdbt:	Sample-data-block table
1145  * @overflow:	Event overflow counter
1146  *
1147  * Walks through a sample-data-block and collects sampling data entries that are
1148  * then pushed to the perf event subsystem.  Depending on the sampling function,
1149  * there can be either basic-sampling or combined-sampling data entries.  A
1150  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1151  * data entry.	The sampling function is determined by the flags in the perf
1152  * event hardware structure.  The function always works with a combined-sampling
1153  * data entry but ignores the the diagnostic portion if it is not available.
1154  *
1155  * Note that the implementation focuses on basic-sampling data entries and, if
1156  * such an entry is not valid, the entire combined-sampling data entry is
1157  * ignored.
1158  *
1159  * The overflow variables counts the number of samples that has been discarded
1160  * due to a perf event overflow.
1161  */
1162 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1163 			       unsigned long long *overflow)
1164 {
1165 	struct hws_trailer_entry *te;
1166 	struct hws_basic_entry *sample;
1167 
1168 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1169 	sample = (struct hws_basic_entry *) *sdbt;
1170 	while ((unsigned long *) sample < (unsigned long *) te) {
1171 		/* Check for an empty sample */
1172 		if (!sample->def)
1173 			break;
1174 
1175 		/* Update perf event period */
1176 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1177 
1178 		/* Check whether sample is valid */
1179 		if (sample->def == 0x0001) {
1180 			/* If an event overflow occurred, the PMU is stopped to
1181 			 * throttle event delivery.  Remaining sample data is
1182 			 * discarded.
1183 			 */
1184 			if (!*overflow) {
1185 				/* Check whether sample is consistent */
1186 				if (sample->I == 0 && sample->W == 0) {
1187 					/* Deliver sample data to perf */
1188 					*overflow = perf_push_sample(event,
1189 								     sample);
1190 				}
1191 			} else
1192 				/* Count discarded samples */
1193 				*overflow += 1;
1194 		} else {
1195 			debug_sample_entry(sample, te);
1196 			/* Sample slot is not yet written or other record.
1197 			 *
1198 			 * This condition can occur if the buffer was reused
1199 			 * from a combined basic- and diagnostic-sampling.
1200 			 * If only basic-sampling is then active, entries are
1201 			 * written into the larger diagnostic entries.
1202 			 * This is typically the case for sample-data-blocks
1203 			 * that are not full.  Stop processing if the first
1204 			 * invalid format was detected.
1205 			 */
1206 			if (!te->f)
1207 				break;
1208 		}
1209 
1210 		/* Reset sample slot and advance to next sample */
1211 		sample->def = 0;
1212 		sample++;
1213 	}
1214 }
1215 
1216 /* hw_perf_event_update() - Process sampling buffer
1217  * @event:	The perf event
1218  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1219  *
1220  * Processes the sampling buffer and create perf event samples.
1221  * The sampling buffer position are retrieved and saved in the TEAR_REG
1222  * register of the specified perf event.
1223  *
1224  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1225  * to also walk through partially filled sample-data-blocks.  It is ignored
1226  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1227  * enforces the processing of full sample-data-blocks only (trailer entries
1228  * with the block-full-indicator bit set).
1229  */
1230 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1231 {
1232 	struct hw_perf_event *hwc = &event->hw;
1233 	struct hws_trailer_entry *te;
1234 	unsigned long *sdbt;
1235 	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1236 	int done;
1237 
1238 	/*
1239 	 * AUX buffer is used when in diagnostic sampling mode.
1240 	 * No perf events/samples are created.
1241 	 */
1242 	if (SAMPL_DIAG_MODE(&event->hw))
1243 		return;
1244 
1245 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1246 		flush_all = 0;
1247 
1248 	sdbt = (unsigned long *) TEAR_REG(hwc);
1249 	done = event_overflow = sampl_overflow = num_sdb = 0;
1250 	while (!done) {
1251 		/* Get the trailer entry of the sample-data-block */
1252 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1253 
1254 		/* Leave loop if no more work to do (block full indicator) */
1255 		if (!te->f) {
1256 			done = 1;
1257 			if (!flush_all)
1258 				break;
1259 		}
1260 
1261 		/* Check the sample overflow count */
1262 		if (te->overflow)
1263 			/* Account sample overflows and, if a particular limit
1264 			 * is reached, extend the sampling buffer.
1265 			 * For details, see sfb_account_overflows().
1266 			 */
1267 			sampl_overflow += te->overflow;
1268 
1269 		/* Timestamps are valid for full sample-data-blocks only */
1270 		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1271 				    "overflow=%llu timestamp=%#llx\n",
1272 				    sdbt, te->overflow,
1273 				    (te->f) ? trailer_timestamp(te) : 0ULL);
1274 
1275 		/* Collect all samples from a single sample-data-block and
1276 		 * flag if an (perf) event overflow happened.  If so, the PMU
1277 		 * is stopped and remaining samples will be discarded.
1278 		 */
1279 		hw_collect_samples(event, sdbt, &event_overflow);
1280 		num_sdb++;
1281 
1282 		/* Reset trailer (using compare-double-and-swap) */
1283 		do {
1284 			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1285 			te_flags |= SDB_TE_ALERT_REQ_MASK;
1286 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1287 					 te->flags, te->overflow,
1288 					 te_flags, 0ULL));
1289 
1290 		/* Advance to next sample-data-block */
1291 		sdbt++;
1292 		if (is_link_entry(sdbt))
1293 			sdbt = get_next_sdbt(sdbt);
1294 
1295 		/* Update event hardware registers */
1296 		TEAR_REG(hwc) = (unsigned long) sdbt;
1297 
1298 		/* Stop processing sample-data if all samples of the current
1299 		 * sample-data-block were flushed even if it was not full.
1300 		 */
1301 		if (flush_all && done)
1302 			break;
1303 
1304 		/* If an event overflow happened, discard samples by
1305 		 * processing any remaining sample-data-blocks.
1306 		 */
1307 		if (event_overflow)
1308 			flush_all = 1;
1309 	}
1310 
1311 	/* Account sample overflows in the event hardware structure */
1312 	if (sampl_overflow)
1313 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1314 						 sampl_overflow, 1 + num_sdb);
1315 	if (sampl_overflow || event_overflow)
1316 		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1317 				    "overflow stats: sample=%llu event=%llu\n",
1318 				    sampl_overflow, event_overflow);
1319 }
1320 
1321 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1322 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1323 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1324 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1325 
1326 /*
1327  * Get trailer entry by index of SDB.
1328  */
1329 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1330 						 unsigned long index)
1331 {
1332 	unsigned long sdb;
1333 
1334 	index = AUX_SDB_INDEX(aux, index);
1335 	sdb = aux->sdb_index[index];
1336 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1337 }
1338 
1339 /*
1340  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1341  * disabled. Collect the full SDBs in AUX buffer which have not reached
1342  * the point of alert indicator. And ignore the SDBs which are not
1343  * full.
1344  *
1345  * 1. Scan SDBs to see how much data is there and consume them.
1346  * 2. Remove alert indicator in the buffer.
1347  */
1348 static void aux_output_end(struct perf_output_handle *handle)
1349 {
1350 	unsigned long i, range_scan, idx;
1351 	struct aux_buffer *aux;
1352 	struct hws_trailer_entry *te;
1353 
1354 	aux = perf_get_aux(handle);
1355 	if (!aux)
1356 		return;
1357 
1358 	range_scan = AUX_SDB_NUM_ALERT(aux);
1359 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1360 		te = aux_sdb_trailer(aux, idx);
1361 		if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1362 			break;
1363 	}
1364 	/* i is num of SDBs which are full */
1365 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1366 
1367 	/* Remove alert indicators in the buffer */
1368 	te = aux_sdb_trailer(aux, aux->alert_mark);
1369 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1370 
1371 	debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1372 }
1373 
1374 /*
1375  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1376  * is first added to the CPU or rescheduled again to the CPU. It is called
1377  * with pmu disabled.
1378  *
1379  * 1. Reset the trailer of SDBs to get ready for new data.
1380  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1381  *    head(tear/dear).
1382  */
1383 static int aux_output_begin(struct perf_output_handle *handle,
1384 			    struct aux_buffer *aux,
1385 			    struct cpu_hw_sf *cpuhw)
1386 {
1387 	unsigned long range;
1388 	unsigned long i, range_scan, idx;
1389 	unsigned long head, base, offset;
1390 	struct hws_trailer_entry *te;
1391 
1392 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1393 		return -EINVAL;
1394 
1395 	aux->head = handle->head >> PAGE_SHIFT;
1396 	range = (handle->size + 1) >> PAGE_SHIFT;
1397 	if (range <= 1)
1398 		return -ENOMEM;
1399 
1400 	/*
1401 	 * SDBs between aux->head and aux->empty_mark are already ready
1402 	 * for new data. range_scan is num of SDBs not within them.
1403 	 */
1404 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1405 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1406 		idx = aux->empty_mark + 1;
1407 		for (i = 0; i < range_scan; i++, idx++) {
1408 			te = aux_sdb_trailer(aux, idx);
1409 			te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1410 			te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK;
1411 			te->overflow = 0;
1412 		}
1413 		/* Save the position of empty SDBs */
1414 		aux->empty_mark = aux->head + range - 1;
1415 	}
1416 
1417 	/* Set alert indicator */
1418 	aux->alert_mark = aux->head + range/2 - 1;
1419 	te = aux_sdb_trailer(aux, aux->alert_mark);
1420 	te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1421 
1422 	/* Reset hardware buffer head */
1423 	head = AUX_SDB_INDEX(aux, aux->head);
1424 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1425 	offset = head % CPUM_SF_SDB_PER_TABLE;
1426 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1427 	cpuhw->lsctl.dear = aux->sdb_index[head];
1428 
1429 	debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1430 			    "head->alert_mark->empty_mark (num_alert, range)"
1431 			    "[%lx -> %lx -> %lx] (%lx, %lx) "
1432 			    "tear index %lx, tear %lx dear %lx\n",
1433 			    aux->head, aux->alert_mark, aux->empty_mark,
1434 			    AUX_SDB_NUM_ALERT(aux), range,
1435 			    head / CPUM_SF_SDB_PER_TABLE,
1436 			    cpuhw->lsctl.tear,
1437 			    cpuhw->lsctl.dear);
1438 
1439 	return 0;
1440 }
1441 
1442 /*
1443  * Set alert indicator on SDB at index @alert_index while sampler is running.
1444  *
1445  * Return true if successfully.
1446  * Return false if full indicator is already set by hardware sampler.
1447  */
1448 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1449 			  unsigned long long *overflow)
1450 {
1451 	unsigned long long orig_overflow, orig_flags, new_flags;
1452 	struct hws_trailer_entry *te;
1453 
1454 	te = aux_sdb_trailer(aux, alert_index);
1455 	do {
1456 		orig_flags = te->flags;
1457 		orig_overflow = te->overflow;
1458 		*overflow = orig_overflow;
1459 		if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1460 			/*
1461 			 * SDB is already set by hardware.
1462 			 * Abort and try to set somewhere
1463 			 * behind.
1464 			 */
1465 			return false;
1466 		}
1467 		new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1468 	} while (!cmpxchg_double(&te->flags, &te->overflow,
1469 				 orig_flags, orig_overflow,
1470 				 new_flags, 0ULL));
1471 	return true;
1472 }
1473 
1474 /*
1475  * aux_reset_buffer() - Scan and setup SDBs for new samples
1476  * @aux:	The AUX buffer to set
1477  * @range:	The range of SDBs to scan started from aux->head
1478  * @overflow:	Set to overflow count
1479  *
1480  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1481  * marked as empty, check if it is already set full by the hardware sampler.
1482  * If yes, that means new data is already there before we can set an alert
1483  * indicator. Caller should try to set alert indicator to some position behind.
1484  *
1485  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1486  * previously and have already been consumed by user space. Reset these SDBs
1487  * (clear full indicator and alert indicator) for new data.
1488  * If aux->alert_mark fall in this area, just set it. Overflow count is
1489  * recorded while scanning.
1490  *
1491  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1492  * and ready for new samples. So scanning on this area could be skipped.
1493  *
1494  * Return true if alert indicator is set successfully and false if not.
1495  */
1496 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1497 			     unsigned long long *overflow)
1498 {
1499 	unsigned long long orig_overflow, orig_flags, new_flags;
1500 	unsigned long i, range_scan, idx;
1501 	struct hws_trailer_entry *te;
1502 
1503 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1504 		/*
1505 		 * No need to scan. All SDBs in range are marked as empty.
1506 		 * Just set alert indicator. Should check race with hardware
1507 		 * sampler.
1508 		 */
1509 		return aux_set_alert(aux, aux->alert_mark, overflow);
1510 
1511 	if (aux->alert_mark <= aux->empty_mark)
1512 		/*
1513 		 * Set alert indicator on empty SDB. Should check race
1514 		 * with hardware sampler.
1515 		 */
1516 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1517 			return false;
1518 
1519 	/*
1520 	 * Scan the SDBs to clear full and alert indicator used previously.
1521 	 * Start scanning from one SDB behind empty_mark. If the new alert
1522 	 * indicator fall into this range, set it.
1523 	 */
1524 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1525 	idx = aux->empty_mark + 1;
1526 	for (i = 0; i < range_scan; i++, idx++) {
1527 		te = aux_sdb_trailer(aux, idx);
1528 		do {
1529 			orig_flags = te->flags;
1530 			orig_overflow = te->overflow;
1531 			new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1532 			if (idx == aux->alert_mark)
1533 				new_flags |= SDB_TE_ALERT_REQ_MASK;
1534 			else
1535 				new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1536 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1537 					 orig_flags, orig_overflow,
1538 					 new_flags, 0ULL));
1539 		*overflow += orig_overflow;
1540 	}
1541 
1542 	/* Update empty_mark to new position */
1543 	aux->empty_mark = aux->head + range - 1;
1544 
1545 	return true;
1546 }
1547 
1548 /*
1549  * Measurement alert handler for diagnostic mode sampling.
1550  */
1551 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1552 {
1553 	struct aux_buffer *aux;
1554 	int done = 0;
1555 	unsigned long range = 0, size;
1556 	unsigned long long overflow = 0;
1557 	struct perf_output_handle *handle = &cpuhw->handle;
1558 	unsigned long num_sdb;
1559 
1560 	aux = perf_get_aux(handle);
1561 	if (WARN_ON_ONCE(!aux))
1562 		return;
1563 
1564 	/* Inform user space new data arrived */
1565 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1566 	perf_aux_output_end(handle, size);
1567 	num_sdb = aux->sfb.num_sdb;
1568 
1569 	while (!done) {
1570 		/* Get an output handle */
1571 		aux = perf_aux_output_begin(handle, cpuhw->event);
1572 		if (handle->size == 0) {
1573 			pr_err("The AUX buffer with %lu pages for the "
1574 			       "diagnostic-sampling mode is full\n",
1575 				num_sdb);
1576 			debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1577 			break;
1578 		}
1579 		if (WARN_ON_ONCE(!aux))
1580 			return;
1581 
1582 		/* Update head and alert_mark to new position */
1583 		aux->head = handle->head >> PAGE_SHIFT;
1584 		range = (handle->size + 1) >> PAGE_SHIFT;
1585 		if (range == 1)
1586 			aux->alert_mark = aux->head;
1587 		else
1588 			aux->alert_mark = aux->head + range/2 - 1;
1589 
1590 		if (aux_reset_buffer(aux, range, &overflow)) {
1591 			if (!overflow) {
1592 				done = 1;
1593 				break;
1594 			}
1595 			size = range << PAGE_SHIFT;
1596 			perf_aux_output_end(&cpuhw->handle, size);
1597 			pr_err("Sample data caused the AUX buffer with %lu "
1598 			       "pages to overflow\n", num_sdb);
1599 			debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1600 					    "overflow %llx\n",
1601 					    aux->head, range, overflow);
1602 		} else {
1603 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1604 			perf_aux_output_end(&cpuhw->handle, size);
1605 			debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1606 					    "already full, try another\n",
1607 					    aux->head, aux->alert_mark);
1608 		}
1609 	}
1610 
1611 	if (done)
1612 		debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1613 				    "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1614 				    aux->head, aux->alert_mark, aux->empty_mark,
1615 				    AUX_SDB_NUM_ALERT(aux), range);
1616 }
1617 
1618 /*
1619  * Callback when freeing AUX buffers.
1620  */
1621 static void aux_buffer_free(void *data)
1622 {
1623 	struct aux_buffer *aux = data;
1624 	unsigned long i, num_sdbt;
1625 
1626 	if (!aux)
1627 		return;
1628 
1629 	/* Free SDBT. SDB is freed by the caller */
1630 	num_sdbt = aux->sfb.num_sdbt;
1631 	for (i = 0; i < num_sdbt; i++)
1632 		free_page(aux->sdbt_index[i]);
1633 
1634 	kfree(aux->sdbt_index);
1635 	kfree(aux->sdb_index);
1636 	kfree(aux);
1637 
1638 	debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1639 			    "%lu SDBTs\n", num_sdbt);
1640 }
1641 
1642 static void aux_sdb_init(unsigned long sdb)
1643 {
1644 	struct hws_trailer_entry *te;
1645 
1646 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1647 
1648 	/* Save clock base */
1649 	te->clock_base = 1;
1650 	memcpy(&te->progusage2, &tod_clock_base[1], 8);
1651 }
1652 
1653 /*
1654  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1655  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1656  * @pages:	Array of pointers to buffer pages passed from perf core
1657  * @nr_pages:	Total pages
1658  * @snapshot:	Flag for snapshot mode
1659  *
1660  * This is the callback when setup an event using AUX buffer. Perf tool can
1661  * trigger this by an additional mmap() call on the event. Unlike the buffer
1662  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1663  * the task among online cpus when it is a per-thread event.
1664  *
1665  * Return the private AUX buffer structure if success or NULL if fails.
1666  */
1667 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1668 			      int nr_pages, bool snapshot)
1669 {
1670 	struct sf_buffer *sfb;
1671 	struct aux_buffer *aux;
1672 	unsigned long *new, *tail;
1673 	int i, n_sdbt;
1674 
1675 	if (!nr_pages || !pages)
1676 		return NULL;
1677 
1678 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1679 		pr_err("AUX buffer size (%i pages) is larger than the "
1680 		       "maximum sampling buffer limit\n",
1681 		       nr_pages);
1682 		return NULL;
1683 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1684 		pr_err("AUX buffer size (%i pages) is less than the "
1685 		       "minimum sampling buffer limit\n",
1686 		       nr_pages);
1687 		return NULL;
1688 	}
1689 
1690 	/* Allocate aux_buffer struct for the event */
1691 	aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1692 	if (!aux)
1693 		goto no_aux;
1694 	sfb = &aux->sfb;
1695 
1696 	/* Allocate sdbt_index for fast reference */
1697 	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1698 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1699 	if (!aux->sdbt_index)
1700 		goto no_sdbt_index;
1701 
1702 	/* Allocate sdb_index for fast reference */
1703 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1704 	if (!aux->sdb_index)
1705 		goto no_sdb_index;
1706 
1707 	/* Allocate the first SDBT */
1708 	sfb->num_sdbt = 0;
1709 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1710 	if (!sfb->sdbt)
1711 		goto no_sdbt;
1712 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1713 	tail = sfb->tail = sfb->sdbt;
1714 
1715 	/*
1716 	 * Link the provided pages of AUX buffer to SDBT.
1717 	 * Allocate SDBT if needed.
1718 	 */
1719 	for (i = 0; i < nr_pages; i++, tail++) {
1720 		if (require_table_link(tail)) {
1721 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1722 			if (!new)
1723 				goto no_sdbt;
1724 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1725 			/* Link current page to tail of chain */
1726 			*tail = (unsigned long)(void *) new + 1;
1727 			tail = new;
1728 		}
1729 		/* Tail is the entry in a SDBT */
1730 		*tail = (unsigned long)pages[i];
1731 		aux->sdb_index[i] = (unsigned long)pages[i];
1732 		aux_sdb_init((unsigned long)pages[i]);
1733 	}
1734 	sfb->num_sdb = nr_pages;
1735 
1736 	/* Link the last entry in the SDBT to the first SDBT */
1737 	*tail = (unsigned long) sfb->sdbt + 1;
1738 	sfb->tail = tail;
1739 
1740 	/*
1741 	 * Initial all SDBs are zeroed. Mark it as empty.
1742 	 * So there is no need to clear the full indicator
1743 	 * when this event is first added.
1744 	 */
1745 	aux->empty_mark = sfb->num_sdb - 1;
1746 
1747 	debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1748 			    " and %lu SDBs\n",
1749 			    sfb->num_sdbt, sfb->num_sdb);
1750 
1751 	return aux;
1752 
1753 no_sdbt:
1754 	/* SDBs (AUX buffer pages) are freed by caller */
1755 	for (i = 0; i < sfb->num_sdbt; i++)
1756 		free_page(aux->sdbt_index[i]);
1757 	kfree(aux->sdb_index);
1758 no_sdb_index:
1759 	kfree(aux->sdbt_index);
1760 no_sdbt_index:
1761 	kfree(aux);
1762 no_aux:
1763 	return NULL;
1764 }
1765 
1766 static void cpumsf_pmu_read(struct perf_event *event)
1767 {
1768 	/* Nothing to do ... updates are interrupt-driven */
1769 }
1770 
1771 /* Check if the new sampling period/freqeuncy is appropriate.
1772  *
1773  * Return non-zero on error and zero on passed checks.
1774  */
1775 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1776 {
1777 	struct hws_qsi_info_block si;
1778 	unsigned long rate;
1779 	bool do_freq;
1780 
1781 	memset(&si, 0, sizeof(si));
1782 	if (event->cpu == -1) {
1783 		if (qsi(&si))
1784 			return -ENODEV;
1785 	} else {
1786 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1787 		 * sampling structure for accessing the CPU-specific QSI.
1788 		 */
1789 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1790 
1791 		si = cpuhw->qsi;
1792 	}
1793 
1794 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1795 	rate = getrate(do_freq, value, &si);
1796 	if (!rate)
1797 		return -EINVAL;
1798 
1799 	event->attr.sample_period = rate;
1800 	SAMPL_RATE(&event->hw) = rate;
1801 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1802 	debug_sprintf_event(sfdbg, 4, "cpumsf_pmu_check_period:"
1803 			    "cpu:%d value:%llx period:%llx freq:%d\n",
1804 			    event->cpu, value,
1805 			    event->attr.sample_period, do_freq);
1806 	return 0;
1807 }
1808 
1809 /* Activate sampling control.
1810  * Next call of pmu_enable() starts sampling.
1811  */
1812 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1813 {
1814 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1815 
1816 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1817 		return;
1818 
1819 	if (flags & PERF_EF_RELOAD)
1820 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1821 
1822 	perf_pmu_disable(event->pmu);
1823 	event->hw.state = 0;
1824 	cpuhw->lsctl.cs = 1;
1825 	if (SAMPL_DIAG_MODE(&event->hw))
1826 		cpuhw->lsctl.cd = 1;
1827 	perf_pmu_enable(event->pmu);
1828 }
1829 
1830 /* Deactivate sampling control.
1831  * Next call of pmu_enable() stops sampling.
1832  */
1833 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1834 {
1835 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1836 
1837 	if (event->hw.state & PERF_HES_STOPPED)
1838 		return;
1839 
1840 	perf_pmu_disable(event->pmu);
1841 	cpuhw->lsctl.cs = 0;
1842 	cpuhw->lsctl.cd = 0;
1843 	event->hw.state |= PERF_HES_STOPPED;
1844 
1845 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1846 		hw_perf_event_update(event, 1);
1847 		event->hw.state |= PERF_HES_UPTODATE;
1848 	}
1849 	perf_pmu_enable(event->pmu);
1850 }
1851 
1852 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1853 {
1854 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1855 	struct aux_buffer *aux;
1856 	int err;
1857 
1858 	if (cpuhw->flags & PMU_F_IN_USE)
1859 		return -EAGAIN;
1860 
1861 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1862 		return -EINVAL;
1863 
1864 	err = 0;
1865 	perf_pmu_disable(event->pmu);
1866 
1867 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1868 
1869 	/* Set up sampling controls.  Always program the sampling register
1870 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1871 	 * that is used by hw_perf_event_update() to store the sampling buffer
1872 	 * position after samples have been flushed.
1873 	 */
1874 	cpuhw->lsctl.s = 0;
1875 	cpuhw->lsctl.h = 1;
1876 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1877 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1878 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1879 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1880 		hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1881 	}
1882 
1883 	/* Ensure sampling functions are in the disabled state.  If disabled,
1884 	 * switch on sampling enable control. */
1885 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1886 		err = -EAGAIN;
1887 		goto out;
1888 	}
1889 	if (SAMPL_DIAG_MODE(&event->hw)) {
1890 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1891 		if (!aux) {
1892 			err = -EINVAL;
1893 			goto out;
1894 		}
1895 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1896 		if (err)
1897 			goto out;
1898 		cpuhw->lsctl.ed = 1;
1899 	}
1900 	cpuhw->lsctl.es = 1;
1901 
1902 	/* Set in_use flag and store event */
1903 	cpuhw->event = event;
1904 	cpuhw->flags |= PMU_F_IN_USE;
1905 
1906 	if (flags & PERF_EF_START)
1907 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1908 out:
1909 	perf_event_update_userpage(event);
1910 	perf_pmu_enable(event->pmu);
1911 	return err;
1912 }
1913 
1914 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1915 {
1916 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1917 
1918 	perf_pmu_disable(event->pmu);
1919 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1920 
1921 	cpuhw->lsctl.es = 0;
1922 	cpuhw->lsctl.ed = 0;
1923 	cpuhw->flags &= ~PMU_F_IN_USE;
1924 	cpuhw->event = NULL;
1925 
1926 	if (SAMPL_DIAG_MODE(&event->hw))
1927 		aux_output_end(&cpuhw->handle);
1928 	perf_event_update_userpage(event);
1929 	perf_pmu_enable(event->pmu);
1930 }
1931 
1932 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1933 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1934 
1935 /* Attribute list for CPU_SF.
1936  *
1937  * The availablitiy depends on the CPU_MF sampling facility authorization
1938  * for basic + diagnositic samples. This is determined at initialization
1939  * time by the sampling facility device driver.
1940  * If the authorization for basic samples is turned off, it should be
1941  * also turned off for diagnostic sampling.
1942  *
1943  * During initialization of the device driver, check the authorization
1944  * level for diagnostic sampling and installs the attribute
1945  * file for diagnostic sampling if necessary.
1946  *
1947  * For now install a placeholder to reference all possible attributes:
1948  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
1949  * Add another entry for the final NULL pointer.
1950  */
1951 enum {
1952 	SF_CYCLES_BASIC_ATTR_IDX = 0,
1953 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
1954 	SF_CYCLES_ATTR_MAX
1955 };
1956 
1957 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
1958 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
1959 };
1960 
1961 PMU_FORMAT_ATTR(event, "config:0-63");
1962 
1963 static struct attribute *cpumsf_pmu_format_attr[] = {
1964 	&format_attr_event.attr,
1965 	NULL,
1966 };
1967 
1968 static struct attribute_group cpumsf_pmu_events_group = {
1969 	.name = "events",
1970 	.attrs = cpumsf_pmu_events_attr,
1971 };
1972 
1973 static struct attribute_group cpumsf_pmu_format_group = {
1974 	.name = "format",
1975 	.attrs = cpumsf_pmu_format_attr,
1976 };
1977 
1978 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1979 	&cpumsf_pmu_events_group,
1980 	&cpumsf_pmu_format_group,
1981 	NULL,
1982 };
1983 
1984 static struct pmu cpumf_sampling = {
1985 	.pmu_enable   = cpumsf_pmu_enable,
1986 	.pmu_disable  = cpumsf_pmu_disable,
1987 
1988 	.event_init   = cpumsf_pmu_event_init,
1989 	.add	      = cpumsf_pmu_add,
1990 	.del	      = cpumsf_pmu_del,
1991 
1992 	.start	      = cpumsf_pmu_start,
1993 	.stop	      = cpumsf_pmu_stop,
1994 	.read	      = cpumsf_pmu_read,
1995 
1996 	.attr_groups  = cpumsf_pmu_attr_groups,
1997 
1998 	.setup_aux    = aux_buffer_setup,
1999 	.free_aux     = aux_buffer_free,
2000 
2001 	.check_period = cpumsf_pmu_check_period,
2002 };
2003 
2004 static void cpumf_measurement_alert(struct ext_code ext_code,
2005 				    unsigned int alert, unsigned long unused)
2006 {
2007 	struct cpu_hw_sf *cpuhw;
2008 
2009 	if (!(alert & CPU_MF_INT_SF_MASK))
2010 		return;
2011 	inc_irq_stat(IRQEXT_CMS);
2012 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2013 
2014 	/* Measurement alerts are shared and might happen when the PMU
2015 	 * is not reserved.  Ignore these alerts in this case. */
2016 	if (!(cpuhw->flags & PMU_F_RESERVED))
2017 		return;
2018 
2019 	/* The processing below must take care of multiple alert events that
2020 	 * might be indicated concurrently. */
2021 
2022 	/* Program alert request */
2023 	if (alert & CPU_MF_INT_SF_PRA) {
2024 		if (cpuhw->flags & PMU_F_IN_USE)
2025 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2026 				hw_collect_aux(cpuhw);
2027 			else
2028 				hw_perf_event_update(cpuhw->event, 0);
2029 		else
2030 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2031 	}
2032 
2033 	/* Report measurement alerts only for non-PRA codes */
2034 	if (alert != CPU_MF_INT_SF_PRA)
2035 		debug_sprintf_event(sfdbg, 6, "measurement alert: %#x\n",
2036 				    alert);
2037 
2038 	/* Sampling authorization change request */
2039 	if (alert & CPU_MF_INT_SF_SACA)
2040 		qsi(&cpuhw->qsi);
2041 
2042 	/* Loss of sample data due to high-priority machine activities */
2043 	if (alert & CPU_MF_INT_SF_LSDA) {
2044 		pr_err("Sample data was lost\n");
2045 		cpuhw->flags |= PMU_F_ERR_LSDA;
2046 		sf_disable();
2047 	}
2048 
2049 	/* Invalid sampling buffer entry */
2050 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2051 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2052 		       alert);
2053 		cpuhw->flags |= PMU_F_ERR_IBE;
2054 		sf_disable();
2055 	}
2056 }
2057 
2058 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2059 {
2060 	/* Ignore the notification if no events are scheduled on the PMU.
2061 	 * This might be racy...
2062 	 */
2063 	if (!atomic_read(&num_events))
2064 		return 0;
2065 
2066 	local_irq_disable();
2067 	setup_pmc_cpu(&flags);
2068 	local_irq_enable();
2069 	return 0;
2070 }
2071 
2072 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2073 {
2074 	return cpusf_pmu_setup(cpu, PMC_INIT);
2075 }
2076 
2077 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2078 {
2079 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2080 }
2081 
2082 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2083 {
2084 	if (!cpum_sf_avail())
2085 		return -ENODEV;
2086 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2087 }
2088 
2089 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2090 {
2091 	int rc;
2092 	unsigned long min, max;
2093 
2094 	if (!cpum_sf_avail())
2095 		return -ENODEV;
2096 	if (!val || !strlen(val))
2097 		return -EINVAL;
2098 
2099 	/* Valid parameter values: "min,max" or "max" */
2100 	min = CPUM_SF_MIN_SDB;
2101 	max = CPUM_SF_MAX_SDB;
2102 	if (strchr(val, ','))
2103 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2104 	else
2105 		rc = kstrtoul(val, 10, &max);
2106 
2107 	if (min < 2 || min >= max || max > get_num_physpages())
2108 		rc = -EINVAL;
2109 	if (rc)
2110 		return rc;
2111 
2112 	sfb_set_limits(min, max);
2113 	pr_info("The sampling buffer limits have changed to: "
2114 		"min=%lu max=%lu (diag=x%lu)\n",
2115 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2116 	return 0;
2117 }
2118 
2119 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2120 static const struct kernel_param_ops param_ops_sfb_size = {
2121 	.set = param_set_sfb_size,
2122 	.get = param_get_sfb_size,
2123 };
2124 
2125 #define RS_INIT_FAILURE_QSI	  0x0001
2126 #define RS_INIT_FAILURE_BSDES	  0x0002
2127 #define RS_INIT_FAILURE_ALRT	  0x0003
2128 #define RS_INIT_FAILURE_PERF	  0x0004
2129 static void __init pr_cpumsf_err(unsigned int reason)
2130 {
2131 	pr_err("Sampling facility support for perf is not available: "
2132 	       "reason=%04x\n", reason);
2133 }
2134 
2135 static int __init init_cpum_sampling_pmu(void)
2136 {
2137 	struct hws_qsi_info_block si;
2138 	int err;
2139 
2140 	if (!cpum_sf_avail())
2141 		return -ENODEV;
2142 
2143 	memset(&si, 0, sizeof(si));
2144 	if (qsi(&si)) {
2145 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2146 		return -ENODEV;
2147 	}
2148 
2149 	if (!si.as && !si.ad)
2150 		return -ENODEV;
2151 
2152 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2153 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2154 		return -EINVAL;
2155 	}
2156 
2157 	if (si.ad) {
2158 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2159 		/* Sampling of diagnostic data authorized,
2160 		 * install event into attribute list of PMU device.
2161 		 */
2162 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2163 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2164 	}
2165 
2166 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2167 	if (!sfdbg) {
2168 		pr_err("Registering for s390dbf failed\n");
2169 		return -ENOMEM;
2170 	}
2171 	debug_register_view(sfdbg, &debug_sprintf_view);
2172 
2173 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2174 				    cpumf_measurement_alert);
2175 	if (err) {
2176 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2177 		debug_unregister(sfdbg);
2178 		goto out;
2179 	}
2180 
2181 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2182 	if (err) {
2183 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2184 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2185 					cpumf_measurement_alert);
2186 		debug_unregister(sfdbg);
2187 		goto out;
2188 	}
2189 
2190 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2191 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2192 out:
2193 	return err;
2194 }
2195 
2196 arch_initcall(init_cpum_sampling_pmu);
2197 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);
2198