xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision 715f23b6)
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, "%s: freed sdbt %#lx\n", __func__,
160 			    (unsigned long)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, *tail_prev = NULL;
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, "%s: "
216 				    "sampling buffer is not linked: origin %#lx"
217 				    " tail %#lx\n", __func__,
218 				    (unsigned long)sfb->sdbt,
219 				    (unsigned long)tail);
220 		return -EINVAL;
221 	}
222 
223 	/* Allocate remaining SDBs */
224 	rc = 0;
225 	for (i = 0; i < num_sdb; i++) {
226 		/* Allocate a new SDB-table if it is full. */
227 		if (require_table_link(tail)) {
228 			new = (unsigned long *) get_zeroed_page(gfp_flags);
229 			if (!new) {
230 				rc = -ENOMEM;
231 				break;
232 			}
233 			sfb->num_sdbt++;
234 			/* Link current page to tail of chain */
235 			*tail = (unsigned long)(void *) new + 1;
236 			tail_prev = tail;
237 			tail = new;
238 		}
239 
240 		/* Allocate a new sample-data-block.
241 		 * If there is not enough memory, stop the realloc process
242 		 * and simply use what was allocated.  If this is a temporary
243 		 * issue, a new realloc call (if required) might succeed.
244 		 */
245 		rc = alloc_sample_data_block(tail, gfp_flags);
246 		if (rc) {
247 			/* Undo last SDBT. An SDBT with no SDB at its first
248 			 * entry but with an SDBT entry instead can not be
249 			 * handled by the interrupt handler code.
250 			 * Avoid this situation.
251 			 */
252 			if (tail_prev) {
253 				sfb->num_sdbt--;
254 				free_page((unsigned long) new);
255 				tail = tail_prev;
256 			}
257 			break;
258 		}
259 		sfb->num_sdb++;
260 		tail++;
261 		tail_prev = new = NULL;	/* Allocated at least one SBD */
262 	}
263 
264 	/* Link sampling buffer to its origin */
265 	*tail = (unsigned long) sfb->sdbt + 1;
266 	sfb->tail = tail;
267 
268 	debug_sprintf_event(sfdbg, 4, "%s: new buffer"
269 			    " settings: sdbt %lu sdb %lu\n", __func__,
270 			    sfb->num_sdbt, sfb->num_sdb);
271 	return rc;
272 }
273 
274 /*
275  * allocate_sampling_buffer() - allocate sampler memory
276  *
277  * Allocates and initializes a sampling buffer structure using the
278  * specified number of sample-data-blocks (SDB).  For each allocation,
279  * a 4K page is used.  The number of sample-data-block-tables (SDBT)
280  * are calculated from SDBs.
281  * Also set the ALERT_REQ mask in each SDBs trailer.
282  *
283  * Returns zero on success, non-zero otherwise.
284  */
285 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
286 {
287 	int rc;
288 
289 	if (sfb->sdbt)
290 		return -EINVAL;
291 
292 	/* Allocate the sample-data-block-table origin */
293 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
294 	if (!sfb->sdbt)
295 		return -ENOMEM;
296 	sfb->num_sdb = 0;
297 	sfb->num_sdbt = 1;
298 
299 	/* Link the table origin to point to itself to prepare for
300 	 * realloc_sampling_buffer() invocation.
301 	 */
302 	sfb->tail = sfb->sdbt;
303 	*sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
304 
305 	/* Allocate requested number of sample-data-blocks */
306 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
307 	if (rc) {
308 		free_sampling_buffer(sfb);
309 		debug_sprintf_event(sfdbg, 4, "%s: "
310 			"realloc_sampling_buffer failed with rc %i\n",
311 			__func__, rc);
312 	} else
313 		debug_sprintf_event(sfdbg, 4,
314 			"%s: tear %#lx dear %#lx\n", __func__,
315 			(unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt);
316 	return rc;
317 }
318 
319 static void sfb_set_limits(unsigned long min, unsigned long max)
320 {
321 	struct hws_qsi_info_block si;
322 
323 	CPUM_SF_MIN_SDB = min;
324 	CPUM_SF_MAX_SDB = max;
325 
326 	memset(&si, 0, sizeof(si));
327 	if (!qsi(&si))
328 		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
329 }
330 
331 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
332 {
333 	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
334 				    : CPUM_SF_MAX_SDB;
335 }
336 
337 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
338 					struct hw_perf_event *hwc)
339 {
340 	if (!sfb->sdbt)
341 		return SFB_ALLOC_REG(hwc);
342 	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
343 		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
344 	return 0;
345 }
346 
347 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
348 				   struct hw_perf_event *hwc)
349 {
350 	return sfb_pending_allocs(sfb, hwc) > 0;
351 }
352 
353 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
354 {
355 	/* Limit the number of SDBs to not exceed the maximum */
356 	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
357 	if (num)
358 		SFB_ALLOC_REG(hwc) += num;
359 }
360 
361 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
362 {
363 	SFB_ALLOC_REG(hwc) = 0;
364 	sfb_account_allocs(num, hwc);
365 }
366 
367 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
368 {
369 	if (cpuhw->sfb.sdbt)
370 		free_sampling_buffer(&cpuhw->sfb);
371 }
372 
373 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
374 {
375 	unsigned long n_sdb, freq, factor;
376 	size_t sample_size;
377 
378 	/* Calculate sampling buffers using 4K pages
379 	 *
380 	 *    1. Determine the sample data size which depends on the used
381 	 *	 sampling functions, for example, basic-sampling or
382 	 *	 basic-sampling with diagnostic-sampling.
383 	 *
384 	 *    2. Use the sampling frequency as input.  The sampling buffer is
385 	 *	 designed for almost one second.  This can be adjusted through
386 	 *	 the "factor" variable.
387 	 *	 In any case, alloc_sampling_buffer() sets the Alert Request
388 	 *	 Control indicator to trigger a measurement-alert to harvest
389 	 *	 sample-data-blocks (sdb).
390 	 *
391 	 *    3. Compute the number of sample-data-blocks and ensure a minimum
392 	 *	 of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
393 	 *	 exceed a "calculated" maximum.  The symbolic maximum is
394 	 *	 designed for basic-sampling only and needs to be increased if
395 	 *	 diagnostic-sampling is active.
396 	 *	 See also the remarks for these symbolic constants.
397 	 *
398 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
399 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
400 	 *	 to 511 SDBs).
401 	 */
402 	sample_size = sizeof(struct hws_basic_entry);
403 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
404 	factor = 1;
405 	n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
406 	if (n_sdb < CPUM_SF_MIN_SDB)
407 		n_sdb = CPUM_SF_MIN_SDB;
408 
409 	/* If there is already a sampling buffer allocated, it is very likely
410 	 * that the sampling facility is enabled too.  If the event to be
411 	 * initialized requires a greater sampling buffer, the allocation must
412 	 * be postponed.  Changing the sampling buffer requires the sampling
413 	 * facility to be in the disabled state.  So, account the number of
414 	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
415 	 * before the event is started.
416 	 */
417 	sfb_init_allocs(n_sdb, hwc);
418 	if (sf_buffer_available(cpuhw))
419 		return 0;
420 
421 	debug_sprintf_event(sfdbg, 3,
422 			    "%s: rate %lu f %lu sdb %lu/%lu"
423 			    " sample_size %lu cpuhw %p\n", __func__,
424 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
425 			    sample_size, cpuhw);
426 
427 	return alloc_sampling_buffer(&cpuhw->sfb,
428 				     sfb_pending_allocs(&cpuhw->sfb, hwc));
429 }
430 
431 static unsigned long min_percent(unsigned int percent, unsigned long base,
432 				 unsigned long min)
433 {
434 	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
435 }
436 
437 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
438 {
439 	/* Use a percentage-based approach to extend the sampling facility
440 	 * buffer.  Accept up to 5% sample data loss.
441 	 * Vary the extents between 1% to 5% of the current number of
442 	 * sample-data-blocks.
443 	 */
444 	if (ratio <= 5)
445 		return 0;
446 	if (ratio <= 25)
447 		return min_percent(1, base, 1);
448 	if (ratio <= 50)
449 		return min_percent(1, base, 1);
450 	if (ratio <= 75)
451 		return min_percent(2, base, 2);
452 	if (ratio <= 100)
453 		return min_percent(3, base, 3);
454 	if (ratio <= 250)
455 		return min_percent(4, base, 4);
456 
457 	return min_percent(5, base, 8);
458 }
459 
460 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
461 				  struct hw_perf_event *hwc)
462 {
463 	unsigned long ratio, num;
464 
465 	if (!OVERFLOW_REG(hwc))
466 		return;
467 
468 	/* The sample_overflow contains the average number of sample data
469 	 * that has been lost because sample-data-blocks were full.
470 	 *
471 	 * Calculate the total number of sample data entries that has been
472 	 * discarded.  Then calculate the ratio of lost samples to total samples
473 	 * per second in percent.
474 	 */
475 	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
476 			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
477 
478 	/* Compute number of sample-data-blocks */
479 	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
480 	if (num)
481 		sfb_account_allocs(num, hwc);
482 
483 	debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n",
484 			    __func__, OVERFLOW_REG(hwc), ratio, num);
485 	OVERFLOW_REG(hwc) = 0;
486 }
487 
488 /* extend_sampling_buffer() - Extend sampling buffer
489  * @sfb:	Sampling buffer structure (for local CPU)
490  * @hwc:	Perf event hardware structure
491  *
492  * Use this function to extend the sampling buffer based on the overflow counter
493  * and postponed allocation extents stored in the specified Perf event hardware.
494  *
495  * Important: This function disables the sampling facility in order to safely
496  *	      change the sampling buffer structure.  Do not call this function
497  *	      when the PMU is active.
498  */
499 static void extend_sampling_buffer(struct sf_buffer *sfb,
500 				   struct hw_perf_event *hwc)
501 {
502 	unsigned long num, num_old;
503 	int rc;
504 
505 	num = sfb_pending_allocs(sfb, hwc);
506 	if (!num)
507 		return;
508 	num_old = sfb->num_sdb;
509 
510 	/* Disable the sampling facility to reset any states and also
511 	 * clear pending measurement alerts.
512 	 */
513 	sf_disable();
514 
515 	/* Extend the sampling buffer.
516 	 * This memory allocation typically happens in an atomic context when
517 	 * called by perf.  Because this is a reallocation, it is fine if the
518 	 * new SDB-request cannot be satisfied immediately.
519 	 */
520 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
521 	if (rc)
522 		debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n",
523 				    __func__, rc);
524 
525 	if (sfb_has_pending_allocs(sfb, hwc))
526 		debug_sprintf_event(sfdbg, 5, "%s: "
527 				    "req %lu alloc %lu remaining %lu\n",
528 				    __func__, num, sfb->num_sdb - num_old,
529 				    sfb_pending_allocs(sfb, hwc));
530 }
531 
532 /* Number of perf events counting hardware events */
533 static atomic_t num_events;
534 /* Used to avoid races in calling reserve/release_cpumf_hardware */
535 static DEFINE_MUTEX(pmc_reserve_mutex);
536 
537 #define PMC_INIT      0
538 #define PMC_RELEASE   1
539 #define PMC_FAILURE   2
540 static void setup_pmc_cpu(void *flags)
541 {
542 	int err;
543 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
544 
545 	err = 0;
546 	switch (*((int *) flags)) {
547 	case PMC_INIT:
548 		memset(cpusf, 0, sizeof(*cpusf));
549 		err = qsi(&cpusf->qsi);
550 		if (err)
551 			break;
552 		cpusf->flags |= PMU_F_RESERVED;
553 		err = sf_disable();
554 		if (err)
555 			pr_err("Switching off the sampling facility failed "
556 			       "with rc %i\n", err);
557 		debug_sprintf_event(sfdbg, 5,
558 				    "%s: initialized: cpuhw %p\n", __func__,
559 				    cpusf);
560 		break;
561 	case PMC_RELEASE:
562 		cpusf->flags &= ~PMU_F_RESERVED;
563 		err = sf_disable();
564 		if (err) {
565 			pr_err("Switching off the sampling facility failed "
566 			       "with rc %i\n", err);
567 		} else
568 			deallocate_buffers(cpusf);
569 		debug_sprintf_event(sfdbg, 5,
570 				    "%s: released: cpuhw %p\n", __func__,
571 				    cpusf);
572 		break;
573 	}
574 	if (err)
575 		*((int *) flags) |= PMC_FAILURE;
576 }
577 
578 static void release_pmc_hardware(void)
579 {
580 	int flags = PMC_RELEASE;
581 
582 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
583 	on_each_cpu(setup_pmc_cpu, &flags, 1);
584 }
585 
586 static int reserve_pmc_hardware(void)
587 {
588 	int flags = PMC_INIT;
589 
590 	on_each_cpu(setup_pmc_cpu, &flags, 1);
591 	if (flags & PMC_FAILURE) {
592 		release_pmc_hardware();
593 		return -ENODEV;
594 	}
595 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
596 
597 	return 0;
598 }
599 
600 static void hw_perf_event_destroy(struct perf_event *event)
601 {
602 	/* Release PMC if this is the last perf event */
603 	if (!atomic_add_unless(&num_events, -1, 1)) {
604 		mutex_lock(&pmc_reserve_mutex);
605 		if (atomic_dec_return(&num_events) == 0)
606 			release_pmc_hardware();
607 		mutex_unlock(&pmc_reserve_mutex);
608 	}
609 }
610 
611 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
612 {
613 	hwc->sample_period = period;
614 	hwc->last_period = hwc->sample_period;
615 	local64_set(&hwc->period_left, hwc->sample_period);
616 }
617 
618 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
619 				   unsigned long rate)
620 {
621 	return clamp_t(unsigned long, rate,
622 		       si->min_sampl_rate, si->max_sampl_rate);
623 }
624 
625 static u32 cpumsf_pid_type(struct perf_event *event,
626 			   u32 pid, enum pid_type type)
627 {
628 	struct task_struct *tsk;
629 
630 	/* Idle process */
631 	if (!pid)
632 		goto out;
633 
634 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
635 	pid = -1;
636 	if (tsk) {
637 		/*
638 		 * Only top level events contain the pid namespace in which
639 		 * they are created.
640 		 */
641 		if (event->parent)
642 			event = event->parent;
643 		pid = __task_pid_nr_ns(tsk, type, event->ns);
644 		/*
645 		 * See also 1d953111b648
646 		 * "perf/core: Don't report zero PIDs for exiting tasks".
647 		 */
648 		if (!pid && !pid_alive(tsk))
649 			pid = -1;
650 	}
651 out:
652 	return pid;
653 }
654 
655 static void cpumsf_output_event_pid(struct perf_event *event,
656 				    struct perf_sample_data *data,
657 				    struct pt_regs *regs)
658 {
659 	u32 pid;
660 	struct perf_event_header header;
661 	struct perf_output_handle handle;
662 
663 	/*
664 	 * Obtain the PID from the basic-sampling data entry and
665 	 * correct the data->tid_entry.pid value.
666 	 */
667 	pid = data->tid_entry.pid;
668 
669 	/* Protect callchain buffers, tasks */
670 	rcu_read_lock();
671 
672 	perf_prepare_sample(&header, data, event, regs);
673 	if (perf_output_begin(&handle, event, header.size))
674 		goto out;
675 
676 	/* Update the process ID (see also kernel/events/core.c) */
677 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
678 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
679 
680 	perf_output_sample(&handle, &header, data, event);
681 	perf_output_end(&handle);
682 out:
683 	rcu_read_unlock();
684 }
685 
686 static unsigned long getrate(bool freq, unsigned long sample,
687 			     struct hws_qsi_info_block *si)
688 {
689 	unsigned long rate;
690 
691 	if (freq) {
692 		rate = freq_to_sample_rate(si, sample);
693 		rate = hw_limit_rate(si, rate);
694 	} else {
695 		/* The min/max sampling rates specifies the valid range
696 		 * of sample periods.  If the specified sample period is
697 		 * out of range, limit the period to the range boundary.
698 		 */
699 		rate = hw_limit_rate(si, sample);
700 
701 		/* The perf core maintains a maximum sample rate that is
702 		 * configurable through the sysctl interface.  Ensure the
703 		 * sampling rate does not exceed this value.  This also helps
704 		 * to avoid throttling when pushing samples with
705 		 * perf_event_overflow().
706 		 */
707 		if (sample_rate_to_freq(si, rate) >
708 		    sysctl_perf_event_sample_rate) {
709 			debug_sprintf_event(sfdbg, 1, "%s: "
710 					    "Sampling rate exceeds maximum "
711 					    "perf sample rate\n", __func__);
712 			rate = 0;
713 		}
714 	}
715 	return rate;
716 }
717 
718 /* The sampling information (si) contains information about the
719  * min/max sampling intervals and the CPU speed.  So calculate the
720  * correct sampling interval and avoid the whole period adjust
721  * feedback loop.
722  *
723  * Since the CPU Measurement sampling facility can not handle frequency
724  * calculate the sampling interval when frequency is specified using
725  * this formula:
726  *	interval := cpu_speed * 1000000 / sample_freq
727  *
728  * Returns errno on bad input and zero on success with parameter interval
729  * set to the correct sampling rate.
730  *
731  * Note: This function turns off freq bit to avoid calling function
732  * perf_adjust_period(). This causes frequency adjustment in the common
733  * code part which causes tremendous variations in the counter values.
734  */
735 static int __hw_perf_event_init_rate(struct perf_event *event,
736 				     struct hws_qsi_info_block *si)
737 {
738 	struct perf_event_attr *attr = &event->attr;
739 	struct hw_perf_event *hwc = &event->hw;
740 	unsigned long rate;
741 
742 	if (attr->freq) {
743 		if (!attr->sample_freq)
744 			return -EINVAL;
745 		rate = getrate(attr->freq, attr->sample_freq, si);
746 		attr->freq = 0;		/* Don't call  perf_adjust_period() */
747 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
748 	} else {
749 		rate = getrate(attr->freq, attr->sample_period, si);
750 		if (!rate)
751 			return -EINVAL;
752 	}
753 	attr->sample_period = rate;
754 	SAMPL_RATE(hwc) = rate;
755 	hw_init_period(hwc, SAMPL_RATE(hwc));
756 	debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n",
757 			    __func__, event->cpu, event->attr.sample_period,
758 			    event->attr.freq, SAMPLE_FREQ_MODE(hwc));
759 	return 0;
760 }
761 
762 static int __hw_perf_event_init(struct perf_event *event)
763 {
764 	struct cpu_hw_sf *cpuhw;
765 	struct hws_qsi_info_block si;
766 	struct perf_event_attr *attr = &event->attr;
767 	struct hw_perf_event *hwc = &event->hw;
768 	int cpu, err;
769 
770 	/* Reserve CPU-measurement sampling facility */
771 	err = 0;
772 	if (!atomic_inc_not_zero(&num_events)) {
773 		mutex_lock(&pmc_reserve_mutex);
774 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
775 			err = -EBUSY;
776 		else
777 			atomic_inc(&num_events);
778 		mutex_unlock(&pmc_reserve_mutex);
779 	}
780 	event->destroy = hw_perf_event_destroy;
781 
782 	if (err)
783 		goto out;
784 
785 	/* Access per-CPU sampling information (query sampling info) */
786 	/*
787 	 * The event->cpu value can be -1 to count on every CPU, for example,
788 	 * when attaching to a task.  If this is specified, use the query
789 	 * sampling info from the current CPU, otherwise use event->cpu to
790 	 * retrieve the per-CPU information.
791 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
792 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
793 	 */
794 	memset(&si, 0, sizeof(si));
795 	cpuhw = NULL;
796 	if (event->cpu == -1)
797 		qsi(&si);
798 	else {
799 		/* Event is pinned to a particular CPU, retrieve the per-CPU
800 		 * sampling structure for accessing the CPU-specific QSI.
801 		 */
802 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
803 		si = cpuhw->qsi;
804 	}
805 
806 	/* Check sampling facility authorization and, if not authorized,
807 	 * fall back to other PMUs.  It is safe to check any CPU because
808 	 * the authorization is identical for all configured CPUs.
809 	 */
810 	if (!si.as) {
811 		err = -ENOENT;
812 		goto out;
813 	}
814 
815 	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
816 		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
817 		err = -EBUSY;
818 		goto out;
819 	}
820 
821 	/* Always enable basic sampling */
822 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
823 
824 	/* Check if diagnostic sampling is requested.  Deny if the required
825 	 * sampling authorization is missing.
826 	 */
827 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
828 		if (!si.ad) {
829 			err = -EPERM;
830 			goto out;
831 		}
832 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
833 	}
834 
835 	/* Check and set other sampling flags */
836 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
837 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
838 
839 	err =  __hw_perf_event_init_rate(event, &si);
840 	if (err)
841 		goto out;
842 
843 	/* Initialize sample data overflow accounting */
844 	hwc->extra_reg.reg = REG_OVERFLOW;
845 	OVERFLOW_REG(hwc) = 0;
846 
847 	/* Use AUX buffer. No need to allocate it by ourself */
848 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
849 		return 0;
850 
851 	/* Allocate the per-CPU sampling buffer using the CPU information
852 	 * from the event.  If the event is not pinned to a particular
853 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
854 	 * buffers for each online CPU.
855 	 */
856 	if (cpuhw)
857 		/* Event is pinned to a particular CPU */
858 		err = allocate_buffers(cpuhw, hwc);
859 	else {
860 		/* Event is not pinned, allocate sampling buffer on
861 		 * each online CPU
862 		 */
863 		for_each_online_cpu(cpu) {
864 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
865 			err = allocate_buffers(cpuhw, hwc);
866 			if (err)
867 				break;
868 		}
869 	}
870 
871 	/* If PID/TID sampling is active, replace the default overflow
872 	 * handler to extract and resolve the PIDs from the basic-sampling
873 	 * data entries.
874 	 */
875 	if (event->attr.sample_type & PERF_SAMPLE_TID)
876 		if (is_default_overflow_handler(event))
877 			event->overflow_handler = cpumsf_output_event_pid;
878 out:
879 	return err;
880 }
881 
882 static int cpumsf_pmu_event_init(struct perf_event *event)
883 {
884 	int err;
885 
886 	/* No support for taken branch sampling */
887 	if (has_branch_stack(event))
888 		return -EOPNOTSUPP;
889 
890 	switch (event->attr.type) {
891 	case PERF_TYPE_RAW:
892 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
893 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
894 			return -ENOENT;
895 		break;
896 	case PERF_TYPE_HARDWARE:
897 		/* Support sampling of CPU cycles in addition to the
898 		 * counter facility.  However, the counter facility
899 		 * is more precise and, hence, restrict this PMU to
900 		 * sampling events only.
901 		 */
902 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
903 			return -ENOENT;
904 		if (!is_sampling_event(event))
905 			return -ENOENT;
906 		break;
907 	default:
908 		return -ENOENT;
909 	}
910 
911 	/* Check online status of the CPU to which the event is pinned */
912 	if (event->cpu >= 0 && !cpu_online(event->cpu))
913 		return -ENODEV;
914 
915 	/* Force reset of idle/hv excludes regardless of what the
916 	 * user requested.
917 	 */
918 	if (event->attr.exclude_hv)
919 		event->attr.exclude_hv = 0;
920 	if (event->attr.exclude_idle)
921 		event->attr.exclude_idle = 0;
922 
923 	err = __hw_perf_event_init(event);
924 	if (unlikely(err))
925 		if (event->destroy)
926 			event->destroy(event);
927 	return err;
928 }
929 
930 static void cpumsf_pmu_enable(struct pmu *pmu)
931 {
932 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
933 	struct hw_perf_event *hwc;
934 	int err;
935 
936 	if (cpuhw->flags & PMU_F_ENABLED)
937 		return;
938 
939 	if (cpuhw->flags & PMU_F_ERR_MASK)
940 		return;
941 
942 	/* Check whether to extent the sampling buffer.
943 	 *
944 	 * Two conditions trigger an increase of the sampling buffer for a
945 	 * perf event:
946 	 *    1. Postponed buffer allocations from the event initialization.
947 	 *    2. Sampling overflows that contribute to pending allocations.
948 	 *
949 	 * Note that the extend_sampling_buffer() function disables the sampling
950 	 * facility, but it can be fully re-enabled using sampling controls that
951 	 * have been saved in cpumsf_pmu_disable().
952 	 */
953 	if (cpuhw->event) {
954 		hwc = &cpuhw->event->hw;
955 		if (!(SAMPL_DIAG_MODE(hwc))) {
956 			/*
957 			 * Account number of overflow-designated
958 			 * buffer extents
959 			 */
960 			sfb_account_overflows(cpuhw, hwc);
961 			extend_sampling_buffer(&cpuhw->sfb, hwc);
962 		}
963 		/* Rate may be adjusted with ioctl() */
964 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
965 	}
966 
967 	/* (Re)enable the PMU and sampling facility */
968 	cpuhw->flags |= PMU_F_ENABLED;
969 	barrier();
970 
971 	err = lsctl(&cpuhw->lsctl);
972 	if (err) {
973 		cpuhw->flags &= ~PMU_F_ENABLED;
974 		pr_err("Loading sampling controls failed: op %i err %i\n",
975 			1, err);
976 		return;
977 	}
978 
979 	/* Load current program parameter */
980 	lpp(&S390_lowcore.lpp);
981 
982 	debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i "
983 			    "interval %#lx tear %#lx dear %#lx\n", __func__,
984 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
985 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
986 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
987 }
988 
989 static void cpumsf_pmu_disable(struct pmu *pmu)
990 {
991 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
992 	struct hws_lsctl_request_block inactive;
993 	struct hws_qsi_info_block si;
994 	int err;
995 
996 	if (!(cpuhw->flags & PMU_F_ENABLED))
997 		return;
998 
999 	if (cpuhw->flags & PMU_F_ERR_MASK)
1000 		return;
1001 
1002 	/* Switch off sampling activation control */
1003 	inactive = cpuhw->lsctl;
1004 	inactive.cs = 0;
1005 	inactive.cd = 0;
1006 
1007 	err = lsctl(&inactive);
1008 	if (err) {
1009 		pr_err("Loading sampling controls failed: op %i err %i\n",
1010 			2, err);
1011 		return;
1012 	}
1013 
1014 	/* Save state of TEAR and DEAR register contents */
1015 	err = qsi(&si);
1016 	if (!err) {
1017 		/* TEAR/DEAR values are valid only if the sampling facility is
1018 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1019 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1020 		 * controls the enable/disable state.
1021 		 */
1022 		if (si.es) {
1023 			cpuhw->lsctl.tear = si.tear;
1024 			cpuhw->lsctl.dear = si.dear;
1025 		}
1026 	} else
1027 		debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n",
1028 				    __func__, err);
1029 
1030 	cpuhw->flags &= ~PMU_F_ENABLED;
1031 }
1032 
1033 /* perf_exclude_event() - Filter event
1034  * @event:	The perf event
1035  * @regs:	pt_regs structure
1036  * @sde_regs:	Sample-data-entry (sde) regs structure
1037  *
1038  * Filter perf events according to their exclude specification.
1039  *
1040  * Return non-zero if the event shall be excluded.
1041  */
1042 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1043 			      struct perf_sf_sde_regs *sde_regs)
1044 {
1045 	if (event->attr.exclude_user && user_mode(regs))
1046 		return 1;
1047 	if (event->attr.exclude_kernel && !user_mode(regs))
1048 		return 1;
1049 	if (event->attr.exclude_guest && sde_regs->in_guest)
1050 		return 1;
1051 	if (event->attr.exclude_host && !sde_regs->in_guest)
1052 		return 1;
1053 	return 0;
1054 }
1055 
1056 /* perf_push_sample() - Push samples to perf
1057  * @event:	The perf event
1058  * @sample:	Hardware sample data
1059  *
1060  * Use the hardware sample data to create perf event sample.  The sample
1061  * is the pushed to the event subsystem and the function checks for
1062  * possible event overflows.  If an event overflow occurs, the PMU is
1063  * stopped.
1064  *
1065  * Return non-zero if an event overflow occurred.
1066  */
1067 static int perf_push_sample(struct perf_event *event,
1068 			    struct hws_basic_entry *basic)
1069 {
1070 	int overflow;
1071 	struct pt_regs regs;
1072 	struct perf_sf_sde_regs *sde_regs;
1073 	struct perf_sample_data data;
1074 
1075 	/* Setup perf sample */
1076 	perf_sample_data_init(&data, 0, event->hw.last_period);
1077 
1078 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1079 	 * using the Program Request Alert code.  The regs.int_parm_long
1080 	 * field which is unused contains additional sample-data-entry related
1081 	 * indicators.
1082 	 */
1083 	memset(&regs, 0, sizeof(regs));
1084 	regs.int_code = 0x1407;
1085 	regs.int_parm = CPU_MF_INT_SF_PRA;
1086 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1087 
1088 	psw_bits(regs.psw).ia	= basic->ia;
1089 	psw_bits(regs.psw).dat	= basic->T;
1090 	psw_bits(regs.psw).wait = basic->W;
1091 	psw_bits(regs.psw).pstate = basic->P;
1092 	psw_bits(regs.psw).as	= basic->AS;
1093 
1094 	/*
1095 	 * Use the hardware provided configuration level to decide if the
1096 	 * sample belongs to a guest or host. If that is not available,
1097 	 * fall back to the following heuristics:
1098 	 * A non-zero guest program parameter always indicates a guest
1099 	 * sample. Some early samples or samples from guests without
1100 	 * lpp usage would be misaccounted to the host. We use the asn
1101 	 * value as an addon heuristic to detect most of these guest samples.
1102 	 * If the value differs from 0xffff (the host value), we assume to
1103 	 * be a KVM guest.
1104 	 */
1105 	switch (basic->CL) {
1106 	case 1: /* logical partition */
1107 		sde_regs->in_guest = 0;
1108 		break;
1109 	case 2: /* virtual machine */
1110 		sde_regs->in_guest = 1;
1111 		break;
1112 	default: /* old machine, use heuristics */
1113 		if (basic->gpp || basic->prim_asn != 0xffff)
1114 			sde_regs->in_guest = 1;
1115 		break;
1116 	}
1117 
1118 	/*
1119 	 * Store the PID value from the sample-data-entry to be
1120 	 * processed and resolved by cpumsf_output_event_pid().
1121 	 */
1122 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1123 
1124 	overflow = 0;
1125 	if (perf_exclude_event(event, &regs, sde_regs))
1126 		goto out;
1127 	if (perf_event_overflow(event, &data, &regs)) {
1128 		overflow = 1;
1129 		event->pmu->stop(event, 0);
1130 	}
1131 	perf_event_update_userpage(event);
1132 out:
1133 	return overflow;
1134 }
1135 
1136 static void perf_event_count_update(struct perf_event *event, u64 count)
1137 {
1138 	local64_add(count, &event->count);
1139 }
1140 
1141 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1142  * @event:	The perf event
1143  * @sdbt:	Sample-data-block table
1144  * @overflow:	Event overflow counter
1145  *
1146  * Walks through a sample-data-block and collects sampling data entries that are
1147  * then pushed to the perf event subsystem.  Depending on the sampling function,
1148  * there can be either basic-sampling or combined-sampling data entries.  A
1149  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1150  * data entry.	The sampling function is determined by the flags in the perf
1151  * event hardware structure.  The function always works with a combined-sampling
1152  * data entry but ignores the the diagnostic portion if it is not available.
1153  *
1154  * Note that the implementation focuses on basic-sampling data entries and, if
1155  * such an entry is not valid, the entire combined-sampling data entry is
1156  * ignored.
1157  *
1158  * The overflow variables counts the number of samples that has been discarded
1159  * due to a perf event overflow.
1160  */
1161 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1162 			       unsigned long long *overflow)
1163 {
1164 	struct hws_trailer_entry *te;
1165 	struct hws_basic_entry *sample;
1166 
1167 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1168 	sample = (struct hws_basic_entry *) *sdbt;
1169 	while ((unsigned long *) sample < (unsigned long *) te) {
1170 		/* Check for an empty sample */
1171 		if (!sample->def)
1172 			break;
1173 
1174 		/* Update perf event period */
1175 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1176 
1177 		/* Check whether sample is valid */
1178 		if (sample->def == 0x0001) {
1179 			/* If an event overflow occurred, the PMU is stopped to
1180 			 * throttle event delivery.  Remaining sample data is
1181 			 * discarded.
1182 			 */
1183 			if (!*overflow) {
1184 				/* Check whether sample is consistent */
1185 				if (sample->I == 0 && sample->W == 0) {
1186 					/* Deliver sample data to perf */
1187 					*overflow = perf_push_sample(event,
1188 								     sample);
1189 				}
1190 			} else
1191 				/* Count discarded samples */
1192 				*overflow += 1;
1193 		} else {
1194 			debug_sprintf_event(sfdbg, 4,
1195 					    "%s: Found unknown"
1196 					    " sampling data entry: te->f %i"
1197 					    " basic.def %#4x (%p)\n", __func__,
1198 					    te->f, sample->def, sample);
1199 			/* Sample slot is not yet written or other record.
1200 			 *
1201 			 * This condition can occur if the buffer was reused
1202 			 * from a combined basic- and diagnostic-sampling.
1203 			 * If only basic-sampling is then active, entries are
1204 			 * written into the larger diagnostic entries.
1205 			 * This is typically the case for sample-data-blocks
1206 			 * that are not full.  Stop processing if the first
1207 			 * invalid format was detected.
1208 			 */
1209 			if (!te->f)
1210 				break;
1211 		}
1212 
1213 		/* Reset sample slot and advance to next sample */
1214 		sample->def = 0;
1215 		sample++;
1216 	}
1217 }
1218 
1219 /* hw_perf_event_update() - Process sampling buffer
1220  * @event:	The perf event
1221  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1222  *
1223  * Processes the sampling buffer and create perf event samples.
1224  * The sampling buffer position are retrieved and saved in the TEAR_REG
1225  * register of the specified perf event.
1226  *
1227  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1228  * to also walk through partially filled sample-data-blocks.  It is ignored
1229  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1230  * enforces the processing of full sample-data-blocks only (trailer entries
1231  * with the block-full-indicator bit set).
1232  */
1233 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1234 {
1235 	struct hw_perf_event *hwc = &event->hw;
1236 	struct hws_trailer_entry *te;
1237 	unsigned long *sdbt;
1238 	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1239 	int done;
1240 
1241 	/*
1242 	 * AUX buffer is used when in diagnostic sampling mode.
1243 	 * No perf events/samples are created.
1244 	 */
1245 	if (SAMPL_DIAG_MODE(&event->hw))
1246 		return;
1247 
1248 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1249 		flush_all = 0;
1250 
1251 	sdbt = (unsigned long *) TEAR_REG(hwc);
1252 	done = event_overflow = sampl_overflow = num_sdb = 0;
1253 	while (!done) {
1254 		/* Get the trailer entry of the sample-data-block */
1255 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1256 
1257 		/* Leave loop if no more work to do (block full indicator) */
1258 		if (!te->f) {
1259 			done = 1;
1260 			if (!flush_all)
1261 				break;
1262 		}
1263 
1264 		/* Check the sample overflow count */
1265 		if (te->overflow)
1266 			/* Account sample overflows and, if a particular limit
1267 			 * is reached, extend the sampling buffer.
1268 			 * For details, see sfb_account_overflows().
1269 			 */
1270 			sampl_overflow += te->overflow;
1271 
1272 		/* Timestamps are valid for full sample-data-blocks only */
1273 		debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx "
1274 				    "overflow %llu timestamp %#llx\n",
1275 				    __func__, (unsigned long)sdbt, te->overflow,
1276 				    (te->f) ? trailer_timestamp(te) : 0ULL);
1277 
1278 		/* Collect all samples from a single sample-data-block and
1279 		 * flag if an (perf) event overflow happened.  If so, the PMU
1280 		 * is stopped and remaining samples will be discarded.
1281 		 */
1282 		hw_collect_samples(event, sdbt, &event_overflow);
1283 		num_sdb++;
1284 
1285 		/* Reset trailer (using compare-double-and-swap) */
1286 		do {
1287 			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1288 			te_flags |= SDB_TE_ALERT_REQ_MASK;
1289 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1290 					 te->flags, te->overflow,
1291 					 te_flags, 0ULL));
1292 
1293 		/* Advance to next sample-data-block */
1294 		sdbt++;
1295 		if (is_link_entry(sdbt))
1296 			sdbt = get_next_sdbt(sdbt);
1297 
1298 		/* Update event hardware registers */
1299 		TEAR_REG(hwc) = (unsigned long) sdbt;
1300 
1301 		/* Stop processing sample-data if all samples of the current
1302 		 * sample-data-block were flushed even if it was not full.
1303 		 */
1304 		if (flush_all && done)
1305 			break;
1306 	}
1307 
1308 	/* Account sample overflows in the event hardware structure */
1309 	if (sampl_overflow)
1310 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1311 						 sampl_overflow, 1 + num_sdb);
1312 
1313 	/* Perf_event_overflow() and perf_event_account_interrupt() limit
1314 	 * the interrupt rate to an upper limit. Roughly 1000 samples per
1315 	 * task tick.
1316 	 * Hitting this limit results in a large number
1317 	 * of throttled REF_REPORT_THROTTLE entries and the samples
1318 	 * are dropped.
1319 	 * Slightly increase the interval to avoid hitting this limit.
1320 	 */
1321 	if (event_overflow) {
1322 		SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
1323 		debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n",
1324 				    __func__,
1325 				    DIV_ROUND_UP(SAMPL_RATE(hwc), 10));
1326 	}
1327 
1328 	if (sampl_overflow || event_overflow)
1329 		debug_sprintf_event(sfdbg, 4, "%s: "
1330 				    "overflows: sample %llu event %llu"
1331 				    " total %llu num_sdb %llu\n",
1332 				    __func__, sampl_overflow, event_overflow,
1333 				    OVERFLOW_REG(hwc), num_sdb);
1334 }
1335 
1336 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1337 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1338 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1339 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1340 
1341 /*
1342  * Get trailer entry by index of SDB.
1343  */
1344 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1345 						 unsigned long index)
1346 {
1347 	unsigned long sdb;
1348 
1349 	index = AUX_SDB_INDEX(aux, index);
1350 	sdb = aux->sdb_index[index];
1351 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1352 }
1353 
1354 /*
1355  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1356  * disabled. Collect the full SDBs in AUX buffer which have not reached
1357  * the point of alert indicator. And ignore the SDBs which are not
1358  * full.
1359  *
1360  * 1. Scan SDBs to see how much data is there and consume them.
1361  * 2. Remove alert indicator in the buffer.
1362  */
1363 static void aux_output_end(struct perf_output_handle *handle)
1364 {
1365 	unsigned long i, range_scan, idx;
1366 	struct aux_buffer *aux;
1367 	struct hws_trailer_entry *te;
1368 
1369 	aux = perf_get_aux(handle);
1370 	if (!aux)
1371 		return;
1372 
1373 	range_scan = AUX_SDB_NUM_ALERT(aux);
1374 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1375 		te = aux_sdb_trailer(aux, idx);
1376 		if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1377 			break;
1378 	}
1379 	/* i is num of SDBs which are full */
1380 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1381 
1382 	/* Remove alert indicators in the buffer */
1383 	te = aux_sdb_trailer(aux, aux->alert_mark);
1384 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1385 
1386 	debug_sprintf_event(sfdbg, 6, "%s: collect %#lx SDBs\n", __func__, i);
1387 }
1388 
1389 /*
1390  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1391  * is first added to the CPU or rescheduled again to the CPU. It is called
1392  * with pmu disabled.
1393  *
1394  * 1. Reset the trailer of SDBs to get ready for new data.
1395  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1396  *    head(tear/dear).
1397  */
1398 static int aux_output_begin(struct perf_output_handle *handle,
1399 			    struct aux_buffer *aux,
1400 			    struct cpu_hw_sf *cpuhw)
1401 {
1402 	unsigned long range;
1403 	unsigned long i, range_scan, idx;
1404 	unsigned long head, base, offset;
1405 	struct hws_trailer_entry *te;
1406 
1407 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1408 		return -EINVAL;
1409 
1410 	aux->head = handle->head >> PAGE_SHIFT;
1411 	range = (handle->size + 1) >> PAGE_SHIFT;
1412 	if (range <= 1)
1413 		return -ENOMEM;
1414 
1415 	/*
1416 	 * SDBs between aux->head and aux->empty_mark are already ready
1417 	 * for new data. range_scan is num of SDBs not within them.
1418 	 */
1419 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1420 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1421 		idx = aux->empty_mark + 1;
1422 		for (i = 0; i < range_scan; i++, idx++) {
1423 			te = aux_sdb_trailer(aux, idx);
1424 			te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1425 			te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK;
1426 			te->overflow = 0;
1427 		}
1428 		/* Save the position of empty SDBs */
1429 		aux->empty_mark = aux->head + range - 1;
1430 	}
1431 
1432 	/* Set alert indicator */
1433 	aux->alert_mark = aux->head + range/2 - 1;
1434 	te = aux_sdb_trailer(aux, aux->alert_mark);
1435 	te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1436 
1437 	/* Reset hardware buffer head */
1438 	head = AUX_SDB_INDEX(aux, aux->head);
1439 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1440 	offset = head % CPUM_SF_SDB_PER_TABLE;
1441 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1442 	cpuhw->lsctl.dear = aux->sdb_index[head];
1443 
1444 	debug_sprintf_event(sfdbg, 6, "%s: "
1445 			    "head->alert_mark->empty_mark (num_alert, range)"
1446 			    "[%#lx -> %#lx -> %#lx] (%#lx, %#lx) "
1447 			    "tear index %#lx, tear %#lx dear %#lx\n", __func__,
1448 			    aux->head, aux->alert_mark, aux->empty_mark,
1449 			    AUX_SDB_NUM_ALERT(aux), range,
1450 			    head / CPUM_SF_SDB_PER_TABLE,
1451 			    cpuhw->lsctl.tear,
1452 			    cpuhw->lsctl.dear);
1453 
1454 	return 0;
1455 }
1456 
1457 /*
1458  * Set alert indicator on SDB at index @alert_index while sampler is running.
1459  *
1460  * Return true if successfully.
1461  * Return false if full indicator is already set by hardware sampler.
1462  */
1463 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1464 			  unsigned long long *overflow)
1465 {
1466 	unsigned long long orig_overflow, orig_flags, new_flags;
1467 	struct hws_trailer_entry *te;
1468 
1469 	te = aux_sdb_trailer(aux, alert_index);
1470 	do {
1471 		orig_flags = te->flags;
1472 		orig_overflow = te->overflow;
1473 		*overflow = orig_overflow;
1474 		if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1475 			/*
1476 			 * SDB is already set by hardware.
1477 			 * Abort and try to set somewhere
1478 			 * behind.
1479 			 */
1480 			return false;
1481 		}
1482 		new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1483 	} while (!cmpxchg_double(&te->flags, &te->overflow,
1484 				 orig_flags, orig_overflow,
1485 				 new_flags, 0ULL));
1486 	return true;
1487 }
1488 
1489 /*
1490  * aux_reset_buffer() - Scan and setup SDBs for new samples
1491  * @aux:	The AUX buffer to set
1492  * @range:	The range of SDBs to scan started from aux->head
1493  * @overflow:	Set to overflow count
1494  *
1495  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1496  * marked as empty, check if it is already set full by the hardware sampler.
1497  * If yes, that means new data is already there before we can set an alert
1498  * indicator. Caller should try to set alert indicator to some position behind.
1499  *
1500  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1501  * previously and have already been consumed by user space. Reset these SDBs
1502  * (clear full indicator and alert indicator) for new data.
1503  * If aux->alert_mark fall in this area, just set it. Overflow count is
1504  * recorded while scanning.
1505  *
1506  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1507  * and ready for new samples. So scanning on this area could be skipped.
1508  *
1509  * Return true if alert indicator is set successfully and false if not.
1510  */
1511 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1512 			     unsigned long long *overflow)
1513 {
1514 	unsigned long long orig_overflow, orig_flags, new_flags;
1515 	unsigned long i, range_scan, idx;
1516 	struct hws_trailer_entry *te;
1517 
1518 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1519 		/*
1520 		 * No need to scan. All SDBs in range are marked as empty.
1521 		 * Just set alert indicator. Should check race with hardware
1522 		 * sampler.
1523 		 */
1524 		return aux_set_alert(aux, aux->alert_mark, overflow);
1525 
1526 	if (aux->alert_mark <= aux->empty_mark)
1527 		/*
1528 		 * Set alert indicator on empty SDB. Should check race
1529 		 * with hardware sampler.
1530 		 */
1531 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1532 			return false;
1533 
1534 	/*
1535 	 * Scan the SDBs to clear full and alert indicator used previously.
1536 	 * Start scanning from one SDB behind empty_mark. If the new alert
1537 	 * indicator fall into this range, set it.
1538 	 */
1539 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1540 	idx = aux->empty_mark + 1;
1541 	for (i = 0; i < range_scan; i++, idx++) {
1542 		te = aux_sdb_trailer(aux, idx);
1543 		do {
1544 			orig_flags = te->flags;
1545 			orig_overflow = te->overflow;
1546 			new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1547 			if (idx == aux->alert_mark)
1548 				new_flags |= SDB_TE_ALERT_REQ_MASK;
1549 			else
1550 				new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1551 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1552 					 orig_flags, orig_overflow,
1553 					 new_flags, 0ULL));
1554 		*overflow += orig_overflow;
1555 	}
1556 
1557 	/* Update empty_mark to new position */
1558 	aux->empty_mark = aux->head + range - 1;
1559 
1560 	return true;
1561 }
1562 
1563 /*
1564  * Measurement alert handler for diagnostic mode sampling.
1565  */
1566 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1567 {
1568 	struct aux_buffer *aux;
1569 	int done = 0;
1570 	unsigned long range = 0, size;
1571 	unsigned long long overflow = 0;
1572 	struct perf_output_handle *handle = &cpuhw->handle;
1573 	unsigned long num_sdb;
1574 
1575 	aux = perf_get_aux(handle);
1576 	if (WARN_ON_ONCE(!aux))
1577 		return;
1578 
1579 	/* Inform user space new data arrived */
1580 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1581 	perf_aux_output_end(handle, size);
1582 	num_sdb = aux->sfb.num_sdb;
1583 
1584 	while (!done) {
1585 		/* Get an output handle */
1586 		aux = perf_aux_output_begin(handle, cpuhw->event);
1587 		if (handle->size == 0) {
1588 			pr_err("The AUX buffer with %lu pages for the "
1589 			       "diagnostic-sampling mode is full\n",
1590 				num_sdb);
1591 			debug_sprintf_event(sfdbg, 1,
1592 					    "%s: AUX buffer used up\n",
1593 					    __func__);
1594 			break;
1595 		}
1596 		if (WARN_ON_ONCE(!aux))
1597 			return;
1598 
1599 		/* Update head and alert_mark to new position */
1600 		aux->head = handle->head >> PAGE_SHIFT;
1601 		range = (handle->size + 1) >> PAGE_SHIFT;
1602 		if (range == 1)
1603 			aux->alert_mark = aux->head;
1604 		else
1605 			aux->alert_mark = aux->head + range/2 - 1;
1606 
1607 		if (aux_reset_buffer(aux, range, &overflow)) {
1608 			if (!overflow) {
1609 				done = 1;
1610 				break;
1611 			}
1612 			size = range << PAGE_SHIFT;
1613 			perf_aux_output_end(&cpuhw->handle, size);
1614 			pr_err("Sample data caused the AUX buffer with %lu "
1615 			       "pages to overflow\n", num_sdb);
1616 			debug_sprintf_event(sfdbg, 1, "%s: head %#lx range %#lx "
1617 					    "overflow %#llx\n", __func__,
1618 					    aux->head, range, overflow);
1619 		} else {
1620 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1621 			perf_aux_output_end(&cpuhw->handle, size);
1622 			debug_sprintf_event(sfdbg, 6, "%s: head %#lx alert %#lx "
1623 					    "already full, try another\n",
1624 					    __func__,
1625 					    aux->head, aux->alert_mark);
1626 		}
1627 	}
1628 
1629 	if (done)
1630 		debug_sprintf_event(sfdbg, 6, "%s: aux_reset_buffer "
1631 				    "[%#lx -> %#lx -> %#lx] (%#lx, %#lx)\n",
1632 				    __func__, aux->head, aux->alert_mark,
1633 				    aux->empty_mark, AUX_SDB_NUM_ALERT(aux),
1634 				    range);
1635 }
1636 
1637 /*
1638  * Callback when freeing AUX buffers.
1639  */
1640 static void aux_buffer_free(void *data)
1641 {
1642 	struct aux_buffer *aux = data;
1643 	unsigned long i, num_sdbt;
1644 
1645 	if (!aux)
1646 		return;
1647 
1648 	/* Free SDBT. SDB is freed by the caller */
1649 	num_sdbt = aux->sfb.num_sdbt;
1650 	for (i = 0; i < num_sdbt; i++)
1651 		free_page(aux->sdbt_index[i]);
1652 
1653 	kfree(aux->sdbt_index);
1654 	kfree(aux->sdb_index);
1655 	kfree(aux);
1656 
1657 	debug_sprintf_event(sfdbg, 4, "%s: free "
1658 			    "%lu SDBTs\n", __func__, num_sdbt);
1659 }
1660 
1661 static void aux_sdb_init(unsigned long sdb)
1662 {
1663 	struct hws_trailer_entry *te;
1664 
1665 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1666 
1667 	/* Save clock base */
1668 	te->clock_base = 1;
1669 	memcpy(&te->progusage2, &tod_clock_base[1], 8);
1670 }
1671 
1672 /*
1673  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1674  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1675  * @pages:	Array of pointers to buffer pages passed from perf core
1676  * @nr_pages:	Total pages
1677  * @snapshot:	Flag for snapshot mode
1678  *
1679  * This is the callback when setup an event using AUX buffer. Perf tool can
1680  * trigger this by an additional mmap() call on the event. Unlike the buffer
1681  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1682  * the task among online cpus when it is a per-thread event.
1683  *
1684  * Return the private AUX buffer structure if success or NULL if fails.
1685  */
1686 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1687 			      int nr_pages, bool snapshot)
1688 {
1689 	struct sf_buffer *sfb;
1690 	struct aux_buffer *aux;
1691 	unsigned long *new, *tail;
1692 	int i, n_sdbt;
1693 
1694 	if (!nr_pages || !pages)
1695 		return NULL;
1696 
1697 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1698 		pr_err("AUX buffer size (%i pages) is larger than the "
1699 		       "maximum sampling buffer limit\n",
1700 		       nr_pages);
1701 		return NULL;
1702 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1703 		pr_err("AUX buffer size (%i pages) is less than the "
1704 		       "minimum sampling buffer limit\n",
1705 		       nr_pages);
1706 		return NULL;
1707 	}
1708 
1709 	/* Allocate aux_buffer struct for the event */
1710 	aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1711 	if (!aux)
1712 		goto no_aux;
1713 	sfb = &aux->sfb;
1714 
1715 	/* Allocate sdbt_index for fast reference */
1716 	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1717 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1718 	if (!aux->sdbt_index)
1719 		goto no_sdbt_index;
1720 
1721 	/* Allocate sdb_index for fast reference */
1722 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1723 	if (!aux->sdb_index)
1724 		goto no_sdb_index;
1725 
1726 	/* Allocate the first SDBT */
1727 	sfb->num_sdbt = 0;
1728 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1729 	if (!sfb->sdbt)
1730 		goto no_sdbt;
1731 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1732 	tail = sfb->tail = sfb->sdbt;
1733 
1734 	/*
1735 	 * Link the provided pages of AUX buffer to SDBT.
1736 	 * Allocate SDBT if needed.
1737 	 */
1738 	for (i = 0; i < nr_pages; i++, tail++) {
1739 		if (require_table_link(tail)) {
1740 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1741 			if (!new)
1742 				goto no_sdbt;
1743 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1744 			/* Link current page to tail of chain */
1745 			*tail = (unsigned long)(void *) new + 1;
1746 			tail = new;
1747 		}
1748 		/* Tail is the entry in a SDBT */
1749 		*tail = (unsigned long)pages[i];
1750 		aux->sdb_index[i] = (unsigned long)pages[i];
1751 		aux_sdb_init((unsigned long)pages[i]);
1752 	}
1753 	sfb->num_sdb = nr_pages;
1754 
1755 	/* Link the last entry in the SDBT to the first SDBT */
1756 	*tail = (unsigned long) sfb->sdbt + 1;
1757 	sfb->tail = tail;
1758 
1759 	/*
1760 	 * Initial all SDBs are zeroed. Mark it as empty.
1761 	 * So there is no need to clear the full indicator
1762 	 * when this event is first added.
1763 	 */
1764 	aux->empty_mark = sfb->num_sdb - 1;
1765 
1766 	debug_sprintf_event(sfdbg, 4, "%s: setup %lu SDBTs and %lu SDBs\n",
1767 			    __func__, sfb->num_sdbt, sfb->num_sdb);
1768 
1769 	return aux;
1770 
1771 no_sdbt:
1772 	/* SDBs (AUX buffer pages) are freed by caller */
1773 	for (i = 0; i < sfb->num_sdbt; i++)
1774 		free_page(aux->sdbt_index[i]);
1775 	kfree(aux->sdb_index);
1776 no_sdb_index:
1777 	kfree(aux->sdbt_index);
1778 no_sdbt_index:
1779 	kfree(aux);
1780 no_aux:
1781 	return NULL;
1782 }
1783 
1784 static void cpumsf_pmu_read(struct perf_event *event)
1785 {
1786 	/* Nothing to do ... updates are interrupt-driven */
1787 }
1788 
1789 /* Check if the new sampling period/freqeuncy is appropriate.
1790  *
1791  * Return non-zero on error and zero on passed checks.
1792  */
1793 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1794 {
1795 	struct hws_qsi_info_block si;
1796 	unsigned long rate;
1797 	bool do_freq;
1798 
1799 	memset(&si, 0, sizeof(si));
1800 	if (event->cpu == -1) {
1801 		if (qsi(&si))
1802 			return -ENODEV;
1803 	} else {
1804 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1805 		 * sampling structure for accessing the CPU-specific QSI.
1806 		 */
1807 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1808 
1809 		si = cpuhw->qsi;
1810 	}
1811 
1812 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1813 	rate = getrate(do_freq, value, &si);
1814 	if (!rate)
1815 		return -EINVAL;
1816 
1817 	event->attr.sample_period = rate;
1818 	SAMPL_RATE(&event->hw) = rate;
1819 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1820 	debug_sprintf_event(sfdbg, 4, "%s:"
1821 			    " cpu %d value %#llx period %#llx freq %d\n",
1822 			    __func__, event->cpu, value,
1823 			    event->attr.sample_period, do_freq);
1824 	return 0;
1825 }
1826 
1827 /* Activate sampling control.
1828  * Next call of pmu_enable() starts sampling.
1829  */
1830 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1831 {
1832 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1833 
1834 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1835 		return;
1836 
1837 	if (flags & PERF_EF_RELOAD)
1838 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1839 
1840 	perf_pmu_disable(event->pmu);
1841 	event->hw.state = 0;
1842 	cpuhw->lsctl.cs = 1;
1843 	if (SAMPL_DIAG_MODE(&event->hw))
1844 		cpuhw->lsctl.cd = 1;
1845 	perf_pmu_enable(event->pmu);
1846 }
1847 
1848 /* Deactivate sampling control.
1849  * Next call of pmu_enable() stops sampling.
1850  */
1851 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1852 {
1853 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1854 
1855 	if (event->hw.state & PERF_HES_STOPPED)
1856 		return;
1857 
1858 	perf_pmu_disable(event->pmu);
1859 	cpuhw->lsctl.cs = 0;
1860 	cpuhw->lsctl.cd = 0;
1861 	event->hw.state |= PERF_HES_STOPPED;
1862 
1863 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1864 		hw_perf_event_update(event, 1);
1865 		event->hw.state |= PERF_HES_UPTODATE;
1866 	}
1867 	perf_pmu_enable(event->pmu);
1868 }
1869 
1870 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1871 {
1872 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1873 	struct aux_buffer *aux;
1874 	int err;
1875 
1876 	if (cpuhw->flags & PMU_F_IN_USE)
1877 		return -EAGAIN;
1878 
1879 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1880 		return -EINVAL;
1881 
1882 	err = 0;
1883 	perf_pmu_disable(event->pmu);
1884 
1885 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1886 
1887 	/* Set up sampling controls.  Always program the sampling register
1888 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1889 	 * that is used by hw_perf_event_update() to store the sampling buffer
1890 	 * position after samples have been flushed.
1891 	 */
1892 	cpuhw->lsctl.s = 0;
1893 	cpuhw->lsctl.h = 1;
1894 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1895 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1896 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1897 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1898 		TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt;
1899 	}
1900 
1901 	/* Ensure sampling functions are in the disabled state.  If disabled,
1902 	 * switch on sampling enable control. */
1903 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1904 		err = -EAGAIN;
1905 		goto out;
1906 	}
1907 	if (SAMPL_DIAG_MODE(&event->hw)) {
1908 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1909 		if (!aux) {
1910 			err = -EINVAL;
1911 			goto out;
1912 		}
1913 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1914 		if (err)
1915 			goto out;
1916 		cpuhw->lsctl.ed = 1;
1917 	}
1918 	cpuhw->lsctl.es = 1;
1919 
1920 	/* Set in_use flag and store event */
1921 	cpuhw->event = event;
1922 	cpuhw->flags |= PMU_F_IN_USE;
1923 
1924 	if (flags & PERF_EF_START)
1925 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1926 out:
1927 	perf_event_update_userpage(event);
1928 	perf_pmu_enable(event->pmu);
1929 	return err;
1930 }
1931 
1932 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1933 {
1934 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1935 
1936 	perf_pmu_disable(event->pmu);
1937 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1938 
1939 	cpuhw->lsctl.es = 0;
1940 	cpuhw->lsctl.ed = 0;
1941 	cpuhw->flags &= ~PMU_F_IN_USE;
1942 	cpuhw->event = NULL;
1943 
1944 	if (SAMPL_DIAG_MODE(&event->hw))
1945 		aux_output_end(&cpuhw->handle);
1946 	perf_event_update_userpage(event);
1947 	perf_pmu_enable(event->pmu);
1948 }
1949 
1950 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1951 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1952 
1953 /* Attribute list for CPU_SF.
1954  *
1955  * The availablitiy depends on the CPU_MF sampling facility authorization
1956  * for basic + diagnositic samples. This is determined at initialization
1957  * time by the sampling facility device driver.
1958  * If the authorization for basic samples is turned off, it should be
1959  * also turned off for diagnostic sampling.
1960  *
1961  * During initialization of the device driver, check the authorization
1962  * level for diagnostic sampling and installs the attribute
1963  * file for diagnostic sampling if necessary.
1964  *
1965  * For now install a placeholder to reference all possible attributes:
1966  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
1967  * Add another entry for the final NULL pointer.
1968  */
1969 enum {
1970 	SF_CYCLES_BASIC_ATTR_IDX = 0,
1971 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
1972 	SF_CYCLES_ATTR_MAX
1973 };
1974 
1975 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
1976 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
1977 };
1978 
1979 PMU_FORMAT_ATTR(event, "config:0-63");
1980 
1981 static struct attribute *cpumsf_pmu_format_attr[] = {
1982 	&format_attr_event.attr,
1983 	NULL,
1984 };
1985 
1986 static struct attribute_group cpumsf_pmu_events_group = {
1987 	.name = "events",
1988 	.attrs = cpumsf_pmu_events_attr,
1989 };
1990 
1991 static struct attribute_group cpumsf_pmu_format_group = {
1992 	.name = "format",
1993 	.attrs = cpumsf_pmu_format_attr,
1994 };
1995 
1996 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1997 	&cpumsf_pmu_events_group,
1998 	&cpumsf_pmu_format_group,
1999 	NULL,
2000 };
2001 
2002 static struct pmu cpumf_sampling = {
2003 	.pmu_enable   = cpumsf_pmu_enable,
2004 	.pmu_disable  = cpumsf_pmu_disable,
2005 
2006 	.event_init   = cpumsf_pmu_event_init,
2007 	.add	      = cpumsf_pmu_add,
2008 	.del	      = cpumsf_pmu_del,
2009 
2010 	.start	      = cpumsf_pmu_start,
2011 	.stop	      = cpumsf_pmu_stop,
2012 	.read	      = cpumsf_pmu_read,
2013 
2014 	.attr_groups  = cpumsf_pmu_attr_groups,
2015 
2016 	.setup_aux    = aux_buffer_setup,
2017 	.free_aux     = aux_buffer_free,
2018 
2019 	.check_period = cpumsf_pmu_check_period,
2020 };
2021 
2022 static void cpumf_measurement_alert(struct ext_code ext_code,
2023 				    unsigned int alert, unsigned long unused)
2024 {
2025 	struct cpu_hw_sf *cpuhw;
2026 
2027 	if (!(alert & CPU_MF_INT_SF_MASK))
2028 		return;
2029 	inc_irq_stat(IRQEXT_CMS);
2030 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2031 
2032 	/* Measurement alerts are shared and might happen when the PMU
2033 	 * is not reserved.  Ignore these alerts in this case. */
2034 	if (!(cpuhw->flags & PMU_F_RESERVED))
2035 		return;
2036 
2037 	/* The processing below must take care of multiple alert events that
2038 	 * might be indicated concurrently. */
2039 
2040 	/* Program alert request */
2041 	if (alert & CPU_MF_INT_SF_PRA) {
2042 		if (cpuhw->flags & PMU_F_IN_USE)
2043 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2044 				hw_collect_aux(cpuhw);
2045 			else
2046 				hw_perf_event_update(cpuhw->event, 0);
2047 		else
2048 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2049 	}
2050 
2051 	/* Report measurement alerts only for non-PRA codes */
2052 	if (alert != CPU_MF_INT_SF_PRA)
2053 		debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__,
2054 				    alert);
2055 
2056 	/* Sampling authorization change request */
2057 	if (alert & CPU_MF_INT_SF_SACA)
2058 		qsi(&cpuhw->qsi);
2059 
2060 	/* Loss of sample data due to high-priority machine activities */
2061 	if (alert & CPU_MF_INT_SF_LSDA) {
2062 		pr_err("Sample data was lost\n");
2063 		cpuhw->flags |= PMU_F_ERR_LSDA;
2064 		sf_disable();
2065 	}
2066 
2067 	/* Invalid sampling buffer entry */
2068 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2069 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2070 		       alert);
2071 		cpuhw->flags |= PMU_F_ERR_IBE;
2072 		sf_disable();
2073 	}
2074 }
2075 
2076 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2077 {
2078 	/* Ignore the notification if no events are scheduled on the PMU.
2079 	 * This might be racy...
2080 	 */
2081 	if (!atomic_read(&num_events))
2082 		return 0;
2083 
2084 	local_irq_disable();
2085 	setup_pmc_cpu(&flags);
2086 	local_irq_enable();
2087 	return 0;
2088 }
2089 
2090 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2091 {
2092 	return cpusf_pmu_setup(cpu, PMC_INIT);
2093 }
2094 
2095 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2096 {
2097 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2098 }
2099 
2100 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2101 {
2102 	if (!cpum_sf_avail())
2103 		return -ENODEV;
2104 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2105 }
2106 
2107 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2108 {
2109 	int rc;
2110 	unsigned long min, max;
2111 
2112 	if (!cpum_sf_avail())
2113 		return -ENODEV;
2114 	if (!val || !strlen(val))
2115 		return -EINVAL;
2116 
2117 	/* Valid parameter values: "min,max" or "max" */
2118 	min = CPUM_SF_MIN_SDB;
2119 	max = CPUM_SF_MAX_SDB;
2120 	if (strchr(val, ','))
2121 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2122 	else
2123 		rc = kstrtoul(val, 10, &max);
2124 
2125 	if (min < 2 || min >= max || max > get_num_physpages())
2126 		rc = -EINVAL;
2127 	if (rc)
2128 		return rc;
2129 
2130 	sfb_set_limits(min, max);
2131 	pr_info("The sampling buffer limits have changed to: "
2132 		"min %lu max %lu (diag %lu)\n",
2133 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2134 	return 0;
2135 }
2136 
2137 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2138 static const struct kernel_param_ops param_ops_sfb_size = {
2139 	.set = param_set_sfb_size,
2140 	.get = param_get_sfb_size,
2141 };
2142 
2143 #define RS_INIT_FAILURE_QSI	  0x0001
2144 #define RS_INIT_FAILURE_BSDES	  0x0002
2145 #define RS_INIT_FAILURE_ALRT	  0x0003
2146 #define RS_INIT_FAILURE_PERF	  0x0004
2147 static void __init pr_cpumsf_err(unsigned int reason)
2148 {
2149 	pr_err("Sampling facility support for perf is not available: "
2150 	       "reason %#x\n", reason);
2151 }
2152 
2153 static int __init init_cpum_sampling_pmu(void)
2154 {
2155 	struct hws_qsi_info_block si;
2156 	int err;
2157 
2158 	if (!cpum_sf_avail())
2159 		return -ENODEV;
2160 
2161 	memset(&si, 0, sizeof(si));
2162 	if (qsi(&si)) {
2163 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2164 		return -ENODEV;
2165 	}
2166 
2167 	if (!si.as && !si.ad)
2168 		return -ENODEV;
2169 
2170 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2171 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2172 		return -EINVAL;
2173 	}
2174 
2175 	if (si.ad) {
2176 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2177 		/* Sampling of diagnostic data authorized,
2178 		 * install event into attribute list of PMU device.
2179 		 */
2180 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2181 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2182 	}
2183 
2184 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2185 	if (!sfdbg) {
2186 		pr_err("Registering for s390dbf failed\n");
2187 		return -ENOMEM;
2188 	}
2189 	debug_register_view(sfdbg, &debug_sprintf_view);
2190 
2191 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2192 				    cpumf_measurement_alert);
2193 	if (err) {
2194 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2195 		debug_unregister(sfdbg);
2196 		goto out;
2197 	}
2198 
2199 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2200 	if (err) {
2201 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2202 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2203 					cpumf_measurement_alert);
2204 		debug_unregister(sfdbg);
2205 		goto out;
2206 	}
2207 
2208 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2209 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2210 out:
2211 	return err;
2212 }
2213 
2214 arch_initcall(init_cpum_sampling_pmu);
2215 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);
2216