xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision e620a1e0)
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 	/* Always enable basic sampling */
807 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
808 
809 	/* Check if diagnostic sampling is requested.  Deny if the required
810 	 * sampling authorization is missing.
811 	 */
812 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
813 		if (!si.ad) {
814 			err = -EPERM;
815 			goto out;
816 		}
817 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
818 	}
819 
820 	/* Check and set other sampling flags */
821 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
822 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
823 
824 	err =  __hw_perf_event_init_rate(event, &si);
825 	if (err)
826 		goto out;
827 
828 	/* Initialize sample data overflow accounting */
829 	hwc->extra_reg.reg = REG_OVERFLOW;
830 	OVERFLOW_REG(hwc) = 0;
831 
832 	/* Use AUX buffer. No need to allocate it by ourself */
833 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
834 		return 0;
835 
836 	/* Allocate the per-CPU sampling buffer using the CPU information
837 	 * from the event.  If the event is not pinned to a particular
838 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
839 	 * buffers for each online CPU.
840 	 */
841 	if (cpuhw)
842 		/* Event is pinned to a particular CPU */
843 		err = allocate_buffers(cpuhw, hwc);
844 	else {
845 		/* Event is not pinned, allocate sampling buffer on
846 		 * each online CPU
847 		 */
848 		for_each_online_cpu(cpu) {
849 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
850 			err = allocate_buffers(cpuhw, hwc);
851 			if (err)
852 				break;
853 		}
854 	}
855 
856 	/* If PID/TID sampling is active, replace the default overflow
857 	 * handler to extract and resolve the PIDs from the basic-sampling
858 	 * data entries.
859 	 */
860 	if (event->attr.sample_type & PERF_SAMPLE_TID)
861 		if (is_default_overflow_handler(event))
862 			event->overflow_handler = cpumsf_output_event_pid;
863 out:
864 	return err;
865 }
866 
867 static int cpumsf_pmu_event_init(struct perf_event *event)
868 {
869 	int err;
870 
871 	/* No support for taken branch sampling */
872 	if (has_branch_stack(event))
873 		return -EOPNOTSUPP;
874 
875 	switch (event->attr.type) {
876 	case PERF_TYPE_RAW:
877 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
878 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
879 			return -ENOENT;
880 		break;
881 	case PERF_TYPE_HARDWARE:
882 		/* Support sampling of CPU cycles in addition to the
883 		 * counter facility.  However, the counter facility
884 		 * is more precise and, hence, restrict this PMU to
885 		 * sampling events only.
886 		 */
887 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
888 			return -ENOENT;
889 		if (!is_sampling_event(event))
890 			return -ENOENT;
891 		break;
892 	default:
893 		return -ENOENT;
894 	}
895 
896 	/* Check online status of the CPU to which the event is pinned */
897 	if (event->cpu >= 0 && !cpu_online(event->cpu))
898 			return -ENODEV;
899 
900 	/* Force reset of idle/hv excludes regardless of what the
901 	 * user requested.
902 	 */
903 	if (event->attr.exclude_hv)
904 		event->attr.exclude_hv = 0;
905 	if (event->attr.exclude_idle)
906 		event->attr.exclude_idle = 0;
907 
908 	err = __hw_perf_event_init(event);
909 	if (unlikely(err))
910 		if (event->destroy)
911 			event->destroy(event);
912 	return err;
913 }
914 
915 static void cpumsf_pmu_enable(struct pmu *pmu)
916 {
917 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
918 	struct hw_perf_event *hwc;
919 	int err;
920 
921 	if (cpuhw->flags & PMU_F_ENABLED)
922 		return;
923 
924 	if (cpuhw->flags & PMU_F_ERR_MASK)
925 		return;
926 
927 	/* Check whether to extent the sampling buffer.
928 	 *
929 	 * Two conditions trigger an increase of the sampling buffer for a
930 	 * perf event:
931 	 *    1. Postponed buffer allocations from the event initialization.
932 	 *    2. Sampling overflows that contribute to pending allocations.
933 	 *
934 	 * Note that the extend_sampling_buffer() function disables the sampling
935 	 * facility, but it can be fully re-enabled using sampling controls that
936 	 * have been saved in cpumsf_pmu_disable().
937 	 */
938 	if (cpuhw->event) {
939 		hwc = &cpuhw->event->hw;
940 		if (!(SAMPL_DIAG_MODE(hwc))) {
941 			/*
942 			 * Account number of overflow-designated
943 			 * buffer extents
944 			 */
945 			sfb_account_overflows(cpuhw, hwc);
946 			if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
947 				extend_sampling_buffer(&cpuhw->sfb, hwc);
948 		}
949 		/* Rate may be adjusted with ioctl() */
950 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
951 	}
952 
953 	/* (Re)enable the PMU and sampling facility */
954 	cpuhw->flags |= PMU_F_ENABLED;
955 	barrier();
956 
957 	err = lsctl(&cpuhw->lsctl);
958 	if (err) {
959 		cpuhw->flags &= ~PMU_F_ENABLED;
960 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
961 			1, err);
962 		return;
963 	}
964 
965 	/* Load current program parameter */
966 	lpp(&S390_lowcore.lpp);
967 
968 	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
969 			    "interval:%lx tear=%p dear=%p\n",
970 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
971 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
972 			    (void *) cpuhw->lsctl.tear,
973 			    (void *) cpuhw->lsctl.dear);
974 }
975 
976 static void cpumsf_pmu_disable(struct pmu *pmu)
977 {
978 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
979 	struct hws_lsctl_request_block inactive;
980 	struct hws_qsi_info_block si;
981 	int err;
982 
983 	if (!(cpuhw->flags & PMU_F_ENABLED))
984 		return;
985 
986 	if (cpuhw->flags & PMU_F_ERR_MASK)
987 		return;
988 
989 	/* Switch off sampling activation control */
990 	inactive = cpuhw->lsctl;
991 	inactive.cs = 0;
992 	inactive.cd = 0;
993 
994 	err = lsctl(&inactive);
995 	if (err) {
996 		pr_err("Loading sampling controls failed: op=%i err=%i\n",
997 			2, err);
998 		return;
999 	}
1000 
1001 	/* Save state of TEAR and DEAR register contents */
1002 	if (!qsi(&si)) {
1003 		/* TEAR/DEAR values are valid only if the sampling facility is
1004 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1005 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1006 		 * controls the enable/disable state.
1007 		 */
1008 		if (si.es) {
1009 			cpuhw->lsctl.tear = si.tear;
1010 			cpuhw->lsctl.dear = si.dear;
1011 		}
1012 	} else
1013 		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
1014 				    "qsi() failed with err=%i\n", err);
1015 
1016 	cpuhw->flags &= ~PMU_F_ENABLED;
1017 }
1018 
1019 /* perf_exclude_event() - Filter event
1020  * @event:	The perf event
1021  * @regs:	pt_regs structure
1022  * @sde_regs:	Sample-data-entry (sde) regs structure
1023  *
1024  * Filter perf events according to their exclude specification.
1025  *
1026  * Return non-zero if the event shall be excluded.
1027  */
1028 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1029 			      struct perf_sf_sde_regs *sde_regs)
1030 {
1031 	if (event->attr.exclude_user && user_mode(regs))
1032 		return 1;
1033 	if (event->attr.exclude_kernel && !user_mode(regs))
1034 		return 1;
1035 	if (event->attr.exclude_guest && sde_regs->in_guest)
1036 		return 1;
1037 	if (event->attr.exclude_host && !sde_regs->in_guest)
1038 		return 1;
1039 	return 0;
1040 }
1041 
1042 /* perf_push_sample() - Push samples to perf
1043  * @event:	The perf event
1044  * @sample:	Hardware sample data
1045  *
1046  * Use the hardware sample data to create perf event sample.  The sample
1047  * is the pushed to the event subsystem and the function checks for
1048  * possible event overflows.  If an event overflow occurs, the PMU is
1049  * stopped.
1050  *
1051  * Return non-zero if an event overflow occurred.
1052  */
1053 static int perf_push_sample(struct perf_event *event,
1054 			    struct hws_basic_entry *basic)
1055 {
1056 	int overflow;
1057 	struct pt_regs regs;
1058 	struct perf_sf_sde_regs *sde_regs;
1059 	struct perf_sample_data data;
1060 
1061 	/* Setup perf sample */
1062 	perf_sample_data_init(&data, 0, event->hw.last_period);
1063 
1064 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1065 	 * using the Program Request Alert code.  The regs.int_parm_long
1066 	 * field which is unused contains additional sample-data-entry related
1067 	 * indicators.
1068 	 */
1069 	memset(&regs, 0, sizeof(regs));
1070 	regs.int_code = 0x1407;
1071 	regs.int_parm = CPU_MF_INT_SF_PRA;
1072 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1073 
1074 	psw_bits(regs.psw).ia	= basic->ia;
1075 	psw_bits(regs.psw).dat	= basic->T;
1076 	psw_bits(regs.psw).wait = basic->W;
1077 	psw_bits(regs.psw).pstate = basic->P;
1078 	psw_bits(regs.psw).as	= basic->AS;
1079 
1080 	/*
1081 	 * Use the hardware provided configuration level to decide if the
1082 	 * sample belongs to a guest or host. If that is not available,
1083 	 * fall back to the following heuristics:
1084 	 * A non-zero guest program parameter always indicates a guest
1085 	 * sample. Some early samples or samples from guests without
1086 	 * lpp usage would be misaccounted to the host. We use the asn
1087 	 * value as an addon heuristic to detect most of these guest samples.
1088 	 * If the value differs from 0xffff (the host value), we assume to
1089 	 * be a KVM guest.
1090 	 */
1091 	switch (basic->CL) {
1092 	case 1: /* logical partition */
1093 		sde_regs->in_guest = 0;
1094 		break;
1095 	case 2: /* virtual machine */
1096 		sde_regs->in_guest = 1;
1097 		break;
1098 	default: /* old machine, use heuristics */
1099 		if (basic->gpp || basic->prim_asn != 0xffff)
1100 			sde_regs->in_guest = 1;
1101 		break;
1102 	}
1103 
1104 	/*
1105 	 * Store the PID value from the sample-data-entry to be
1106 	 * processed and resolved by cpumsf_output_event_pid().
1107 	 */
1108 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1109 
1110 	overflow = 0;
1111 	if (perf_exclude_event(event, &regs, sde_regs))
1112 		goto out;
1113 	if (perf_event_overflow(event, &data, &regs)) {
1114 		overflow = 1;
1115 		event->pmu->stop(event, 0);
1116 	}
1117 	perf_event_update_userpage(event);
1118 out:
1119 	return overflow;
1120 }
1121 
1122 static void perf_event_count_update(struct perf_event *event, u64 count)
1123 {
1124 	local64_add(count, &event->count);
1125 }
1126 
1127 static void debug_sample_entry(struct hws_basic_entry *sample,
1128 			       struct hws_trailer_entry *te)
1129 {
1130 	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
1131 			    "sampling data entry: te->f=%i basic.def=%04x "
1132 			    "(%p)\n",
1133 			    te->f, sample->def, sample);
1134 }
1135 
1136 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1137  * @event:	The perf event
1138  * @sdbt:	Sample-data-block table
1139  * @overflow:	Event overflow counter
1140  *
1141  * Walks through a sample-data-block and collects sampling data entries that are
1142  * then pushed to the perf event subsystem.  Depending on the sampling function,
1143  * there can be either basic-sampling or combined-sampling data entries.  A
1144  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1145  * data entry.	The sampling function is determined by the flags in the perf
1146  * event hardware structure.  The function always works with a combined-sampling
1147  * data entry but ignores the the diagnostic portion if it is not available.
1148  *
1149  * Note that the implementation focuses on basic-sampling data entries and, if
1150  * such an entry is not valid, the entire combined-sampling data entry is
1151  * ignored.
1152  *
1153  * The overflow variables counts the number of samples that has been discarded
1154  * due to a perf event overflow.
1155  */
1156 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1157 			       unsigned long long *overflow)
1158 {
1159 	struct hws_trailer_entry *te;
1160 	struct hws_basic_entry *sample;
1161 
1162 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1163 	sample = (struct hws_basic_entry *) *sdbt;
1164 	while ((unsigned long *) sample < (unsigned long *) te) {
1165 		/* Check for an empty sample */
1166 		if (!sample->def)
1167 			break;
1168 
1169 		/* Update perf event period */
1170 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1171 
1172 		/* Check whether sample is valid */
1173 		if (sample->def == 0x0001) {
1174 			/* If an event overflow occurred, the PMU is stopped to
1175 			 * throttle event delivery.  Remaining sample data is
1176 			 * discarded.
1177 			 */
1178 			if (!*overflow) {
1179 				/* Check whether sample is consistent */
1180 				if (sample->I == 0 && sample->W == 0) {
1181 					/* Deliver sample data to perf */
1182 					*overflow = perf_push_sample(event,
1183 								     sample);
1184 				}
1185 			} else
1186 				/* Count discarded samples */
1187 				*overflow += 1;
1188 		} else {
1189 			debug_sample_entry(sample, te);
1190 			/* Sample slot is not yet written or other record.
1191 			 *
1192 			 * This condition can occur if the buffer was reused
1193 			 * from a combined basic- and diagnostic-sampling.
1194 			 * If only basic-sampling is then active, entries are
1195 			 * written into the larger diagnostic entries.
1196 			 * This is typically the case for sample-data-blocks
1197 			 * that are not full.  Stop processing if the first
1198 			 * invalid format was detected.
1199 			 */
1200 			if (!te->f)
1201 				break;
1202 		}
1203 
1204 		/* Reset sample slot and advance to next sample */
1205 		sample->def = 0;
1206 		sample++;
1207 	}
1208 }
1209 
1210 /* hw_perf_event_update() - Process sampling buffer
1211  * @event:	The perf event
1212  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1213  *
1214  * Processes the sampling buffer and create perf event samples.
1215  * The sampling buffer position are retrieved and saved in the TEAR_REG
1216  * register of the specified perf event.
1217  *
1218  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1219  * to also walk through partially filled sample-data-blocks.  It is ignored
1220  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1221  * enforces the processing of full sample-data-blocks only (trailer entries
1222  * with the block-full-indicator bit set).
1223  */
1224 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1225 {
1226 	struct hw_perf_event *hwc = &event->hw;
1227 	struct hws_trailer_entry *te;
1228 	unsigned long *sdbt;
1229 	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1230 	int done;
1231 
1232 	/*
1233 	 * AUX buffer is used when in diagnostic sampling mode.
1234 	 * No perf events/samples are created.
1235 	 */
1236 	if (SAMPL_DIAG_MODE(&event->hw))
1237 		return;
1238 
1239 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1240 		flush_all = 0;
1241 
1242 	sdbt = (unsigned long *) TEAR_REG(hwc);
1243 	done = event_overflow = sampl_overflow = num_sdb = 0;
1244 	while (!done) {
1245 		/* Get the trailer entry of the sample-data-block */
1246 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1247 
1248 		/* Leave loop if no more work to do (block full indicator) */
1249 		if (!te->f) {
1250 			done = 1;
1251 			if (!flush_all)
1252 				break;
1253 		}
1254 
1255 		/* Check the sample overflow count */
1256 		if (te->overflow)
1257 			/* Account sample overflows and, if a particular limit
1258 			 * is reached, extend the sampling buffer.
1259 			 * For details, see sfb_account_overflows().
1260 			 */
1261 			sampl_overflow += te->overflow;
1262 
1263 		/* Timestamps are valid for full sample-data-blocks only */
1264 		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
1265 				    "overflow=%llu timestamp=%#llx\n",
1266 				    sdbt, te->overflow,
1267 				    (te->f) ? trailer_timestamp(te) : 0ULL);
1268 
1269 		/* Collect all samples from a single sample-data-block and
1270 		 * flag if an (perf) event overflow happened.  If so, the PMU
1271 		 * is stopped and remaining samples will be discarded.
1272 		 */
1273 		hw_collect_samples(event, sdbt, &event_overflow);
1274 		num_sdb++;
1275 
1276 		/* Reset trailer (using compare-double-and-swap) */
1277 		do {
1278 			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1279 			te_flags |= SDB_TE_ALERT_REQ_MASK;
1280 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1281 					 te->flags, te->overflow,
1282 					 te_flags, 0ULL));
1283 
1284 		/* Advance to next sample-data-block */
1285 		sdbt++;
1286 		if (is_link_entry(sdbt))
1287 			sdbt = get_next_sdbt(sdbt);
1288 
1289 		/* Update event hardware registers */
1290 		TEAR_REG(hwc) = (unsigned long) sdbt;
1291 
1292 		/* Stop processing sample-data if all samples of the current
1293 		 * sample-data-block were flushed even if it was not full.
1294 		 */
1295 		if (flush_all && done)
1296 			break;
1297 
1298 		/* If an event overflow happened, discard samples by
1299 		 * processing any remaining sample-data-blocks.
1300 		 */
1301 		if (event_overflow)
1302 			flush_all = 1;
1303 	}
1304 
1305 	/* Account sample overflows in the event hardware structure */
1306 	if (sampl_overflow)
1307 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1308 						 sampl_overflow, 1 + num_sdb);
1309 	if (sampl_overflow || event_overflow)
1310 		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
1311 				    "overflow stats: sample=%llu event=%llu\n",
1312 				    sampl_overflow, event_overflow);
1313 }
1314 
1315 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1316 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1317 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1318 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1319 
1320 /*
1321  * Get trailer entry by index of SDB.
1322  */
1323 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1324 						 unsigned long index)
1325 {
1326 	unsigned long sdb;
1327 
1328 	index = AUX_SDB_INDEX(aux, index);
1329 	sdb = aux->sdb_index[index];
1330 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1331 }
1332 
1333 /*
1334  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1335  * disabled. Collect the full SDBs in AUX buffer which have not reached
1336  * the point of alert indicator. And ignore the SDBs which are not
1337  * full.
1338  *
1339  * 1. Scan SDBs to see how much data is there and consume them.
1340  * 2. Remove alert indicator in the buffer.
1341  */
1342 static void aux_output_end(struct perf_output_handle *handle)
1343 {
1344 	unsigned long i, range_scan, idx;
1345 	struct aux_buffer *aux;
1346 	struct hws_trailer_entry *te;
1347 
1348 	aux = perf_get_aux(handle);
1349 	if (!aux)
1350 		return;
1351 
1352 	range_scan = AUX_SDB_NUM_ALERT(aux);
1353 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1354 		te = aux_sdb_trailer(aux, idx);
1355 		if (!(te->flags & SDB_TE_BUFFER_FULL_MASK))
1356 			break;
1357 	}
1358 	/* i is num of SDBs which are full */
1359 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1360 
1361 	/* Remove alert indicators in the buffer */
1362 	te = aux_sdb_trailer(aux, aux->alert_mark);
1363 	te->flags &= ~SDB_TE_ALERT_REQ_MASK;
1364 
1365 	debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i);
1366 }
1367 
1368 /*
1369  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1370  * is first added to the CPU or rescheduled again to the CPU. It is called
1371  * with pmu disabled.
1372  *
1373  * 1. Reset the trailer of SDBs to get ready for new data.
1374  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1375  *    head(tear/dear).
1376  */
1377 static int aux_output_begin(struct perf_output_handle *handle,
1378 			    struct aux_buffer *aux,
1379 			    struct cpu_hw_sf *cpuhw)
1380 {
1381 	unsigned long range;
1382 	unsigned long i, range_scan, idx;
1383 	unsigned long head, base, offset;
1384 	struct hws_trailer_entry *te;
1385 
1386 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1387 		return -EINVAL;
1388 
1389 	aux->head = handle->head >> PAGE_SHIFT;
1390 	range = (handle->size + 1) >> PAGE_SHIFT;
1391 	if (range <= 1)
1392 		return -ENOMEM;
1393 
1394 	/*
1395 	 * SDBs between aux->head and aux->empty_mark are already ready
1396 	 * for new data. range_scan is num of SDBs not within them.
1397 	 */
1398 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1399 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1400 		idx = aux->empty_mark + 1;
1401 		for (i = 0; i < range_scan; i++, idx++) {
1402 			te = aux_sdb_trailer(aux, idx);
1403 			te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
1404 			te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK;
1405 			te->overflow = 0;
1406 		}
1407 		/* Save the position of empty SDBs */
1408 		aux->empty_mark = aux->head + range - 1;
1409 	}
1410 
1411 	/* Set alert indicator */
1412 	aux->alert_mark = aux->head + range/2 - 1;
1413 	te = aux_sdb_trailer(aux, aux->alert_mark);
1414 	te->flags = te->flags | SDB_TE_ALERT_REQ_MASK;
1415 
1416 	/* Reset hardware buffer head */
1417 	head = AUX_SDB_INDEX(aux, aux->head);
1418 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1419 	offset = head % CPUM_SF_SDB_PER_TABLE;
1420 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1421 	cpuhw->lsctl.dear = aux->sdb_index[head];
1422 
1423 	debug_sprintf_event(sfdbg, 6, "aux_output_begin: "
1424 			    "head->alert_mark->empty_mark (num_alert, range)"
1425 			    "[%lx -> %lx -> %lx] (%lx, %lx) "
1426 			    "tear index %lx, tear %lx dear %lx\n",
1427 			    aux->head, aux->alert_mark, aux->empty_mark,
1428 			    AUX_SDB_NUM_ALERT(aux), range,
1429 			    head / CPUM_SF_SDB_PER_TABLE,
1430 			    cpuhw->lsctl.tear,
1431 			    cpuhw->lsctl.dear);
1432 
1433 	return 0;
1434 }
1435 
1436 /*
1437  * Set alert indicator on SDB at index @alert_index while sampler is running.
1438  *
1439  * Return true if successfully.
1440  * Return false if full indicator is already set by hardware sampler.
1441  */
1442 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1443 			  unsigned long long *overflow)
1444 {
1445 	unsigned long long orig_overflow, orig_flags, new_flags;
1446 	struct hws_trailer_entry *te;
1447 
1448 	te = aux_sdb_trailer(aux, alert_index);
1449 	do {
1450 		orig_flags = te->flags;
1451 		orig_overflow = te->overflow;
1452 		*overflow = orig_overflow;
1453 		if (orig_flags & SDB_TE_BUFFER_FULL_MASK) {
1454 			/*
1455 			 * SDB is already set by hardware.
1456 			 * Abort and try to set somewhere
1457 			 * behind.
1458 			 */
1459 			return false;
1460 		}
1461 		new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK;
1462 	} while (!cmpxchg_double(&te->flags, &te->overflow,
1463 				 orig_flags, orig_overflow,
1464 				 new_flags, 0ULL));
1465 	return true;
1466 }
1467 
1468 /*
1469  * aux_reset_buffer() - Scan and setup SDBs for new samples
1470  * @aux:	The AUX buffer to set
1471  * @range:	The range of SDBs to scan started from aux->head
1472  * @overflow:	Set to overflow count
1473  *
1474  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1475  * marked as empty, check if it is already set full by the hardware sampler.
1476  * If yes, that means new data is already there before we can set an alert
1477  * indicator. Caller should try to set alert indicator to some position behind.
1478  *
1479  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1480  * previously and have already been consumed by user space. Reset these SDBs
1481  * (clear full indicator and alert indicator) for new data.
1482  * If aux->alert_mark fall in this area, just set it. Overflow count is
1483  * recorded while scanning.
1484  *
1485  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1486  * and ready for new samples. So scanning on this area could be skipped.
1487  *
1488  * Return true if alert indicator is set successfully and false if not.
1489  */
1490 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1491 			     unsigned long long *overflow)
1492 {
1493 	unsigned long long orig_overflow, orig_flags, new_flags;
1494 	unsigned long i, range_scan, idx;
1495 	struct hws_trailer_entry *te;
1496 
1497 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1498 		/*
1499 		 * No need to scan. All SDBs in range are marked as empty.
1500 		 * Just set alert indicator. Should check race with hardware
1501 		 * sampler.
1502 		 */
1503 		return aux_set_alert(aux, aux->alert_mark, overflow);
1504 
1505 	if (aux->alert_mark <= aux->empty_mark)
1506 		/*
1507 		 * Set alert indicator on empty SDB. Should check race
1508 		 * with hardware sampler.
1509 		 */
1510 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1511 			return false;
1512 
1513 	/*
1514 	 * Scan the SDBs to clear full and alert indicator used previously.
1515 	 * Start scanning from one SDB behind empty_mark. If the new alert
1516 	 * indicator fall into this range, set it.
1517 	 */
1518 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1519 	idx = aux->empty_mark + 1;
1520 	for (i = 0; i < range_scan; i++, idx++) {
1521 		te = aux_sdb_trailer(aux, idx);
1522 		do {
1523 			orig_flags = te->flags;
1524 			orig_overflow = te->overflow;
1525 			new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK;
1526 			if (idx == aux->alert_mark)
1527 				new_flags |= SDB_TE_ALERT_REQ_MASK;
1528 			else
1529 				new_flags &= ~SDB_TE_ALERT_REQ_MASK;
1530 		} while (!cmpxchg_double(&te->flags, &te->overflow,
1531 					 orig_flags, orig_overflow,
1532 					 new_flags, 0ULL));
1533 		*overflow += orig_overflow;
1534 	}
1535 
1536 	/* Update empty_mark to new position */
1537 	aux->empty_mark = aux->head + range - 1;
1538 
1539 	return true;
1540 }
1541 
1542 /*
1543  * Measurement alert handler for diagnostic mode sampling.
1544  */
1545 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1546 {
1547 	struct aux_buffer *aux;
1548 	int done = 0;
1549 	unsigned long range = 0, size;
1550 	unsigned long long overflow = 0;
1551 	struct perf_output_handle *handle = &cpuhw->handle;
1552 	unsigned long num_sdb;
1553 
1554 	aux = perf_get_aux(handle);
1555 	if (WARN_ON_ONCE(!aux))
1556 		return;
1557 
1558 	/* Inform user space new data arrived */
1559 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1560 	perf_aux_output_end(handle, size);
1561 	num_sdb = aux->sfb.num_sdb;
1562 
1563 	while (!done) {
1564 		/* Get an output handle */
1565 		aux = perf_aux_output_begin(handle, cpuhw->event);
1566 		if (handle->size == 0) {
1567 			pr_err("The AUX buffer with %lu pages for the "
1568 			       "diagnostic-sampling mode is full\n",
1569 				num_sdb);
1570 			debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n");
1571 			break;
1572 		}
1573 		if (WARN_ON_ONCE(!aux))
1574 			return;
1575 
1576 		/* Update head and alert_mark to new position */
1577 		aux->head = handle->head >> PAGE_SHIFT;
1578 		range = (handle->size + 1) >> PAGE_SHIFT;
1579 		if (range == 1)
1580 			aux->alert_mark = aux->head;
1581 		else
1582 			aux->alert_mark = aux->head + range/2 - 1;
1583 
1584 		if (aux_reset_buffer(aux, range, &overflow)) {
1585 			if (!overflow) {
1586 				done = 1;
1587 				break;
1588 			}
1589 			size = range << PAGE_SHIFT;
1590 			perf_aux_output_end(&cpuhw->handle, size);
1591 			pr_err("Sample data caused the AUX buffer with %lu "
1592 			       "pages to overflow\n", num_sdb);
1593 			debug_sprintf_event(sfdbg, 1, "head %lx range %lx "
1594 					    "overflow %llx\n",
1595 					    aux->head, range, overflow);
1596 		} else {
1597 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1598 			perf_aux_output_end(&cpuhw->handle, size);
1599 			debug_sprintf_event(sfdbg, 6, "head %lx alert %lx "
1600 					    "already full, try another\n",
1601 					    aux->head, aux->alert_mark);
1602 		}
1603 	}
1604 
1605 	if (done)
1606 		debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: "
1607 				    "[%lx -> %lx -> %lx] (%lx, %lx)\n",
1608 				    aux->head, aux->alert_mark, aux->empty_mark,
1609 				    AUX_SDB_NUM_ALERT(aux), range);
1610 }
1611 
1612 /*
1613  * Callback when freeing AUX buffers.
1614  */
1615 static void aux_buffer_free(void *data)
1616 {
1617 	struct aux_buffer *aux = data;
1618 	unsigned long i, num_sdbt;
1619 
1620 	if (!aux)
1621 		return;
1622 
1623 	/* Free SDBT. SDB is freed by the caller */
1624 	num_sdbt = aux->sfb.num_sdbt;
1625 	for (i = 0; i < num_sdbt; i++)
1626 		free_page(aux->sdbt_index[i]);
1627 
1628 	kfree(aux->sdbt_index);
1629 	kfree(aux->sdb_index);
1630 	kfree(aux);
1631 
1632 	debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free "
1633 			    "%lu SDBTs\n", num_sdbt);
1634 }
1635 
1636 static void aux_sdb_init(unsigned long sdb)
1637 {
1638 	struct hws_trailer_entry *te;
1639 
1640 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1641 
1642 	/* Save clock base */
1643 	te->clock_base = 1;
1644 	memcpy(&te->progusage2, &tod_clock_base[1], 8);
1645 }
1646 
1647 /*
1648  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1649  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1650  * @pages:	Array of pointers to buffer pages passed from perf core
1651  * @nr_pages:	Total pages
1652  * @snapshot:	Flag for snapshot mode
1653  *
1654  * This is the callback when setup an event using AUX buffer. Perf tool can
1655  * trigger this by an additional mmap() call on the event. Unlike the buffer
1656  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1657  * the task among online cpus when it is a per-thread event.
1658  *
1659  * Return the private AUX buffer structure if success or NULL if fails.
1660  */
1661 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1662 			      int nr_pages, bool snapshot)
1663 {
1664 	struct sf_buffer *sfb;
1665 	struct aux_buffer *aux;
1666 	unsigned long *new, *tail;
1667 	int i, n_sdbt;
1668 
1669 	if (!nr_pages || !pages)
1670 		return NULL;
1671 
1672 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1673 		pr_err("AUX buffer size (%i pages) is larger than the "
1674 		       "maximum sampling buffer limit\n",
1675 		       nr_pages);
1676 		return NULL;
1677 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1678 		pr_err("AUX buffer size (%i pages) is less than the "
1679 		       "minimum sampling buffer limit\n",
1680 		       nr_pages);
1681 		return NULL;
1682 	}
1683 
1684 	/* Allocate aux_buffer struct for the event */
1685 	aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1686 	if (!aux)
1687 		goto no_aux;
1688 	sfb = &aux->sfb;
1689 
1690 	/* Allocate sdbt_index for fast reference */
1691 	n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
1692 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1693 	if (!aux->sdbt_index)
1694 		goto no_sdbt_index;
1695 
1696 	/* Allocate sdb_index for fast reference */
1697 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1698 	if (!aux->sdb_index)
1699 		goto no_sdb_index;
1700 
1701 	/* Allocate the first SDBT */
1702 	sfb->num_sdbt = 0;
1703 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1704 	if (!sfb->sdbt)
1705 		goto no_sdbt;
1706 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1707 	tail = sfb->tail = sfb->sdbt;
1708 
1709 	/*
1710 	 * Link the provided pages of AUX buffer to SDBT.
1711 	 * Allocate SDBT if needed.
1712 	 */
1713 	for (i = 0; i < nr_pages; i++, tail++) {
1714 		if (require_table_link(tail)) {
1715 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1716 			if (!new)
1717 				goto no_sdbt;
1718 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1719 			/* Link current page to tail of chain */
1720 			*tail = (unsigned long)(void *) new + 1;
1721 			tail = new;
1722 		}
1723 		/* Tail is the entry in a SDBT */
1724 		*tail = (unsigned long)pages[i];
1725 		aux->sdb_index[i] = (unsigned long)pages[i];
1726 		aux_sdb_init((unsigned long)pages[i]);
1727 	}
1728 	sfb->num_sdb = nr_pages;
1729 
1730 	/* Link the last entry in the SDBT to the first SDBT */
1731 	*tail = (unsigned long) sfb->sdbt + 1;
1732 	sfb->tail = tail;
1733 
1734 	/*
1735 	 * Initial all SDBs are zeroed. Mark it as empty.
1736 	 * So there is no need to clear the full indicator
1737 	 * when this event is first added.
1738 	 */
1739 	aux->empty_mark = sfb->num_sdb - 1;
1740 
1741 	debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
1742 			    " and %lu SDBs\n",
1743 			    sfb->num_sdbt, sfb->num_sdb);
1744 
1745 	return aux;
1746 
1747 no_sdbt:
1748 	/* SDBs (AUX buffer pages) are freed by caller */
1749 	for (i = 0; i < sfb->num_sdbt; i++)
1750 		free_page(aux->sdbt_index[i]);
1751 	kfree(aux->sdb_index);
1752 no_sdb_index:
1753 	kfree(aux->sdbt_index);
1754 no_sdbt_index:
1755 	kfree(aux);
1756 no_aux:
1757 	return NULL;
1758 }
1759 
1760 static void cpumsf_pmu_read(struct perf_event *event)
1761 {
1762 	/* Nothing to do ... updates are interrupt-driven */
1763 }
1764 
1765 /* Check if the new sampling period/freqeuncy is appropriate.
1766  *
1767  * Return non-zero on error and zero on passed checks.
1768  */
1769 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1770 {
1771 	struct hws_qsi_info_block si;
1772 	unsigned long rate;
1773 	bool do_freq;
1774 
1775 	memset(&si, 0, sizeof(si));
1776 	if (event->cpu == -1) {
1777 		if (qsi(&si))
1778 			return -ENODEV;
1779 	} else {
1780 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1781 		 * sampling structure for accessing the CPU-specific QSI.
1782 		 */
1783 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1784 
1785 		si = cpuhw->qsi;
1786 	}
1787 
1788 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1789 	rate = getrate(do_freq, value, &si);
1790 	if (!rate)
1791 		return -EINVAL;
1792 
1793 	event->attr.sample_period = rate;
1794 	SAMPL_RATE(&event->hw) = rate;
1795 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1796 	debug_sprintf_event(sfdbg, 4, "cpumsf_pmu_check_period:"
1797 			    "cpu:%d value:%llx period:%llx freq:%d\n",
1798 			    event->cpu, value,
1799 			    event->attr.sample_period, do_freq);
1800 	return 0;
1801 }
1802 
1803 /* Activate sampling control.
1804  * Next call of pmu_enable() starts sampling.
1805  */
1806 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1807 {
1808 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1809 
1810 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1811 		return;
1812 
1813 	if (flags & PERF_EF_RELOAD)
1814 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1815 
1816 	perf_pmu_disable(event->pmu);
1817 	event->hw.state = 0;
1818 	cpuhw->lsctl.cs = 1;
1819 	if (SAMPL_DIAG_MODE(&event->hw))
1820 		cpuhw->lsctl.cd = 1;
1821 	perf_pmu_enable(event->pmu);
1822 }
1823 
1824 /* Deactivate sampling control.
1825  * Next call of pmu_enable() stops sampling.
1826  */
1827 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1828 {
1829 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1830 
1831 	if (event->hw.state & PERF_HES_STOPPED)
1832 		return;
1833 
1834 	perf_pmu_disable(event->pmu);
1835 	cpuhw->lsctl.cs = 0;
1836 	cpuhw->lsctl.cd = 0;
1837 	event->hw.state |= PERF_HES_STOPPED;
1838 
1839 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1840 		hw_perf_event_update(event, 1);
1841 		event->hw.state |= PERF_HES_UPTODATE;
1842 	}
1843 	perf_pmu_enable(event->pmu);
1844 }
1845 
1846 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1847 {
1848 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1849 	struct aux_buffer *aux;
1850 	int err;
1851 
1852 	if (cpuhw->flags & PMU_F_IN_USE)
1853 		return -EAGAIN;
1854 
1855 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1856 		return -EINVAL;
1857 
1858 	err = 0;
1859 	perf_pmu_disable(event->pmu);
1860 
1861 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1862 
1863 	/* Set up sampling controls.  Always program the sampling register
1864 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1865 	 * that is used by hw_perf_event_update() to store the sampling buffer
1866 	 * position after samples have been flushed.
1867 	 */
1868 	cpuhw->lsctl.s = 0;
1869 	cpuhw->lsctl.h = 1;
1870 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1871 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1872 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1873 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1874 		hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);
1875 	}
1876 
1877 	/* Ensure sampling functions are in the disabled state.  If disabled,
1878 	 * switch on sampling enable control. */
1879 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1880 		err = -EAGAIN;
1881 		goto out;
1882 	}
1883 	if (SAMPL_DIAG_MODE(&event->hw)) {
1884 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1885 		if (!aux) {
1886 			err = -EINVAL;
1887 			goto out;
1888 		}
1889 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1890 		if (err)
1891 			goto out;
1892 		cpuhw->lsctl.ed = 1;
1893 	}
1894 	cpuhw->lsctl.es = 1;
1895 
1896 	/* Set in_use flag and store event */
1897 	cpuhw->event = event;
1898 	cpuhw->flags |= PMU_F_IN_USE;
1899 
1900 	if (flags & PERF_EF_START)
1901 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1902 out:
1903 	perf_event_update_userpage(event);
1904 	perf_pmu_enable(event->pmu);
1905 	return err;
1906 }
1907 
1908 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1909 {
1910 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1911 
1912 	perf_pmu_disable(event->pmu);
1913 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1914 
1915 	cpuhw->lsctl.es = 0;
1916 	cpuhw->lsctl.ed = 0;
1917 	cpuhw->flags &= ~PMU_F_IN_USE;
1918 	cpuhw->event = NULL;
1919 
1920 	if (SAMPL_DIAG_MODE(&event->hw))
1921 		aux_output_end(&cpuhw->handle);
1922 	perf_event_update_userpage(event);
1923 	perf_pmu_enable(event->pmu);
1924 }
1925 
1926 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1927 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1928 
1929 /* Attribute list for CPU_SF.
1930  *
1931  * The availablitiy depends on the CPU_MF sampling facility authorization
1932  * for basic + diagnositic samples. This is determined at initialization
1933  * time by the sampling facility device driver.
1934  * If the authorization for basic samples is turned off, it should be
1935  * also turned off for diagnostic sampling.
1936  *
1937  * During initialization of the device driver, check the authorization
1938  * level for diagnostic sampling and installs the attribute
1939  * file for diagnostic sampling if necessary.
1940  *
1941  * For now install a placeholder to reference all possible attributes:
1942  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
1943  * Add another entry for the final NULL pointer.
1944  */
1945 enum {
1946 	SF_CYCLES_BASIC_ATTR_IDX = 0,
1947 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
1948 	SF_CYCLES_ATTR_MAX
1949 };
1950 
1951 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
1952 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
1953 };
1954 
1955 PMU_FORMAT_ATTR(event, "config:0-63");
1956 
1957 static struct attribute *cpumsf_pmu_format_attr[] = {
1958 	&format_attr_event.attr,
1959 	NULL,
1960 };
1961 
1962 static struct attribute_group cpumsf_pmu_events_group = {
1963 	.name = "events",
1964 	.attrs = cpumsf_pmu_events_attr,
1965 };
1966 
1967 static struct attribute_group cpumsf_pmu_format_group = {
1968 	.name = "format",
1969 	.attrs = cpumsf_pmu_format_attr,
1970 };
1971 
1972 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
1973 	&cpumsf_pmu_events_group,
1974 	&cpumsf_pmu_format_group,
1975 	NULL,
1976 };
1977 
1978 static struct pmu cpumf_sampling = {
1979 	.pmu_enable   = cpumsf_pmu_enable,
1980 	.pmu_disable  = cpumsf_pmu_disable,
1981 
1982 	.event_init   = cpumsf_pmu_event_init,
1983 	.add	      = cpumsf_pmu_add,
1984 	.del	      = cpumsf_pmu_del,
1985 
1986 	.start	      = cpumsf_pmu_start,
1987 	.stop	      = cpumsf_pmu_stop,
1988 	.read	      = cpumsf_pmu_read,
1989 
1990 	.attr_groups  = cpumsf_pmu_attr_groups,
1991 
1992 	.setup_aux    = aux_buffer_setup,
1993 	.free_aux     = aux_buffer_free,
1994 
1995 	.check_period = cpumsf_pmu_check_period,
1996 };
1997 
1998 static void cpumf_measurement_alert(struct ext_code ext_code,
1999 				    unsigned int alert, unsigned long unused)
2000 {
2001 	struct cpu_hw_sf *cpuhw;
2002 
2003 	if (!(alert & CPU_MF_INT_SF_MASK))
2004 		return;
2005 	inc_irq_stat(IRQEXT_CMS);
2006 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2007 
2008 	/* Measurement alerts are shared and might happen when the PMU
2009 	 * is not reserved.  Ignore these alerts in this case. */
2010 	if (!(cpuhw->flags & PMU_F_RESERVED))
2011 		return;
2012 
2013 	/* The processing below must take care of multiple alert events that
2014 	 * might be indicated concurrently. */
2015 
2016 	/* Program alert request */
2017 	if (alert & CPU_MF_INT_SF_PRA) {
2018 		if (cpuhw->flags & PMU_F_IN_USE)
2019 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2020 				hw_collect_aux(cpuhw);
2021 			else
2022 				hw_perf_event_update(cpuhw->event, 0);
2023 		else
2024 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2025 	}
2026 
2027 	/* Report measurement alerts only for non-PRA codes */
2028 	if (alert != CPU_MF_INT_SF_PRA)
2029 		debug_sprintf_event(sfdbg, 6, "measurement alert: %#x\n",
2030 				    alert);
2031 
2032 	/* Sampling authorization change request */
2033 	if (alert & CPU_MF_INT_SF_SACA)
2034 		qsi(&cpuhw->qsi);
2035 
2036 	/* Loss of sample data due to high-priority machine activities */
2037 	if (alert & CPU_MF_INT_SF_LSDA) {
2038 		pr_err("Sample data was lost\n");
2039 		cpuhw->flags |= PMU_F_ERR_LSDA;
2040 		sf_disable();
2041 	}
2042 
2043 	/* Invalid sampling buffer entry */
2044 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2045 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2046 		       alert);
2047 		cpuhw->flags |= PMU_F_ERR_IBE;
2048 		sf_disable();
2049 	}
2050 }
2051 
2052 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2053 {
2054 	/* Ignore the notification if no events are scheduled on the PMU.
2055 	 * This might be racy...
2056 	 */
2057 	if (!atomic_read(&num_events))
2058 		return 0;
2059 
2060 	local_irq_disable();
2061 	setup_pmc_cpu(&flags);
2062 	local_irq_enable();
2063 	return 0;
2064 }
2065 
2066 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2067 {
2068 	return cpusf_pmu_setup(cpu, PMC_INIT);
2069 }
2070 
2071 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2072 {
2073 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2074 }
2075 
2076 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2077 {
2078 	if (!cpum_sf_avail())
2079 		return -ENODEV;
2080 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2081 }
2082 
2083 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2084 {
2085 	int rc;
2086 	unsigned long min, max;
2087 
2088 	if (!cpum_sf_avail())
2089 		return -ENODEV;
2090 	if (!val || !strlen(val))
2091 		return -EINVAL;
2092 
2093 	/* Valid parameter values: "min,max" or "max" */
2094 	min = CPUM_SF_MIN_SDB;
2095 	max = CPUM_SF_MAX_SDB;
2096 	if (strchr(val, ','))
2097 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2098 	else
2099 		rc = kstrtoul(val, 10, &max);
2100 
2101 	if (min < 2 || min >= max || max > get_num_physpages())
2102 		rc = -EINVAL;
2103 	if (rc)
2104 		return rc;
2105 
2106 	sfb_set_limits(min, max);
2107 	pr_info("The sampling buffer limits have changed to: "
2108 		"min=%lu max=%lu (diag=x%lu)\n",
2109 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2110 	return 0;
2111 }
2112 
2113 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2114 static const struct kernel_param_ops param_ops_sfb_size = {
2115 	.set = param_set_sfb_size,
2116 	.get = param_get_sfb_size,
2117 };
2118 
2119 #define RS_INIT_FAILURE_QSI	  0x0001
2120 #define RS_INIT_FAILURE_BSDES	  0x0002
2121 #define RS_INIT_FAILURE_ALRT	  0x0003
2122 #define RS_INIT_FAILURE_PERF	  0x0004
2123 static void __init pr_cpumsf_err(unsigned int reason)
2124 {
2125 	pr_err("Sampling facility support for perf is not available: "
2126 	       "reason=%04x\n", reason);
2127 }
2128 
2129 static int __init init_cpum_sampling_pmu(void)
2130 {
2131 	struct hws_qsi_info_block si;
2132 	int err;
2133 
2134 	if (!cpum_sf_avail())
2135 		return -ENODEV;
2136 
2137 	memset(&si, 0, sizeof(si));
2138 	if (qsi(&si)) {
2139 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2140 		return -ENODEV;
2141 	}
2142 
2143 	if (!si.as && !si.ad)
2144 		return -ENODEV;
2145 
2146 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2147 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2148 		return -EINVAL;
2149 	}
2150 
2151 	if (si.ad) {
2152 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2153 		/* Sampling of diagnostic data authorized,
2154 		 * install event into attribute list of PMU device.
2155 		 */
2156 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2157 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2158 	}
2159 
2160 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2161 	if (!sfdbg) {
2162 		pr_err("Registering for s390dbf failed\n");
2163 		return -ENOMEM;
2164 	}
2165 	debug_register_view(sfdbg, &debug_sprintf_view);
2166 
2167 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2168 				    cpumf_measurement_alert);
2169 	if (err) {
2170 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2171 		debug_unregister(sfdbg);
2172 		goto out;
2173 	}
2174 
2175 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2176 	if (err) {
2177 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2178 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2179 					cpumf_measurement_alert);
2180 		debug_unregister(sfdbg);
2181 		goto out;
2182 	}
2183 
2184 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2185 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2186 out:
2187 	return err;
2188 }
2189 
2190 arch_initcall(init_cpum_sampling_pmu);
2191 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);
2192