xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision cf26e043)
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 	struct hws_trailer_entry *te;
167 	unsigned long sdb;
168 
169 	/* Allocate and initialize sample-data-block */
170 	sdb = get_zeroed_page(gfp_flags);
171 	if (!sdb)
172 		return -ENOMEM;
173 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
174 	te->header.a = 1;
175 
176 	/* Link SDB into the sample-data-block-table */
177 	*sdbt = sdb;
178 
179 	return 0;
180 }
181 
182 /*
183  * realloc_sampling_buffer() - extend sampler memory
184  *
185  * Allocates new sample-data-blocks and adds them to the specified sampling
186  * buffer memory.
187  *
188  * Important: This modifies the sampling buffer and must be called when the
189  *	      sampling facility is disabled.
190  *
191  * Returns zero on success, non-zero otherwise.
192  */
193 static int realloc_sampling_buffer(struct sf_buffer *sfb,
194 				   unsigned long num_sdb, gfp_t gfp_flags)
195 {
196 	int i, rc;
197 	unsigned long *new, *tail, *tail_prev = NULL;
198 
199 	if (!sfb->sdbt || !sfb->tail)
200 		return -EINVAL;
201 
202 	if (!is_link_entry(sfb->tail))
203 		return -EINVAL;
204 
205 	/* Append to the existing sampling buffer, overwriting the table-link
206 	 * register.
207 	 * The tail variables always points to the "tail" (last and table-link)
208 	 * entry in an SDB-table.
209 	 */
210 	tail = sfb->tail;
211 
212 	/* Do a sanity check whether the table-link entry points to
213 	 * the sampling buffer origin.
214 	 */
215 	if (sfb->sdbt != get_next_sdbt(tail)) {
216 		debug_sprintf_event(sfdbg, 3, "%s: "
217 				    "sampling buffer is not linked: origin %#lx"
218 				    " tail %#lx\n", __func__,
219 				    (unsigned long)sfb->sdbt,
220 				    (unsigned long)tail);
221 		return -EINVAL;
222 	}
223 
224 	/* Allocate remaining SDBs */
225 	rc = 0;
226 	for (i = 0; i < num_sdb; i++) {
227 		/* Allocate a new SDB-table if it is full. */
228 		if (require_table_link(tail)) {
229 			new = (unsigned long *) get_zeroed_page(gfp_flags);
230 			if (!new) {
231 				rc = -ENOMEM;
232 				break;
233 			}
234 			sfb->num_sdbt++;
235 			/* Link current page to tail of chain */
236 			*tail = (unsigned long)(void *) new + 1;
237 			tail_prev = tail;
238 			tail = new;
239 		}
240 
241 		/* Allocate a new sample-data-block.
242 		 * If there is not enough memory, stop the realloc process
243 		 * and simply use what was allocated.  If this is a temporary
244 		 * issue, a new realloc call (if required) might succeed.
245 		 */
246 		rc = alloc_sample_data_block(tail, gfp_flags);
247 		if (rc) {
248 			/* Undo last SDBT. An SDBT with no SDB at its first
249 			 * entry but with an SDBT entry instead can not be
250 			 * handled by the interrupt handler code.
251 			 * Avoid this situation.
252 			 */
253 			if (tail_prev) {
254 				sfb->num_sdbt--;
255 				free_page((unsigned long) new);
256 				tail = tail_prev;
257 			}
258 			break;
259 		}
260 		sfb->num_sdb++;
261 		tail++;
262 		tail_prev = new = NULL;	/* Allocated at least one SBD */
263 	}
264 
265 	/* Link sampling buffer to its origin */
266 	*tail = (unsigned long) sfb->sdbt + 1;
267 	sfb->tail = tail;
268 
269 	debug_sprintf_event(sfdbg, 4, "%s: new buffer"
270 			    " settings: sdbt %lu sdb %lu\n", __func__,
271 			    sfb->num_sdbt, sfb->num_sdb);
272 	return rc;
273 }
274 
275 /*
276  * allocate_sampling_buffer() - allocate sampler memory
277  *
278  * Allocates and initializes a sampling buffer structure using the
279  * specified number of sample-data-blocks (SDB).  For each allocation,
280  * a 4K page is used.  The number of sample-data-block-tables (SDBT)
281  * are calculated from SDBs.
282  * Also set the ALERT_REQ mask in each SDBs trailer.
283  *
284  * Returns zero on success, non-zero otherwise.
285  */
286 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
287 {
288 	int rc;
289 
290 	if (sfb->sdbt)
291 		return -EINVAL;
292 
293 	/* Allocate the sample-data-block-table origin */
294 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
295 	if (!sfb->sdbt)
296 		return -ENOMEM;
297 	sfb->num_sdb = 0;
298 	sfb->num_sdbt = 1;
299 
300 	/* Link the table origin to point to itself to prepare for
301 	 * realloc_sampling_buffer() invocation.
302 	 */
303 	sfb->tail = sfb->sdbt;
304 	*sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
305 
306 	/* Allocate requested number of sample-data-blocks */
307 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
308 	if (rc) {
309 		free_sampling_buffer(sfb);
310 		debug_sprintf_event(sfdbg, 4, "%s: "
311 			"realloc_sampling_buffer failed with rc %i\n",
312 			__func__, rc);
313 	} else
314 		debug_sprintf_event(sfdbg, 4,
315 			"%s: tear %#lx dear %#lx\n", __func__,
316 			(unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt);
317 	return rc;
318 }
319 
320 static void sfb_set_limits(unsigned long min, unsigned long max)
321 {
322 	struct hws_qsi_info_block si;
323 
324 	CPUM_SF_MIN_SDB = min;
325 	CPUM_SF_MAX_SDB = max;
326 
327 	memset(&si, 0, sizeof(si));
328 	if (!qsi(&si))
329 		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
330 }
331 
332 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
333 {
334 	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
335 				    : CPUM_SF_MAX_SDB;
336 }
337 
338 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
339 					struct hw_perf_event *hwc)
340 {
341 	if (!sfb->sdbt)
342 		return SFB_ALLOC_REG(hwc);
343 	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
344 		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
345 	return 0;
346 }
347 
348 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
349 				   struct hw_perf_event *hwc)
350 {
351 	return sfb_pending_allocs(sfb, hwc) > 0;
352 }
353 
354 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
355 {
356 	/* Limit the number of SDBs to not exceed the maximum */
357 	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
358 	if (num)
359 		SFB_ALLOC_REG(hwc) += num;
360 }
361 
362 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
363 {
364 	SFB_ALLOC_REG(hwc) = 0;
365 	sfb_account_allocs(num, hwc);
366 }
367 
368 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
369 {
370 	if (cpuhw->sfb.sdbt)
371 		free_sampling_buffer(&cpuhw->sfb);
372 }
373 
374 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
375 {
376 	unsigned long n_sdb, freq;
377 	size_t sample_size;
378 
379 	/* Calculate sampling buffers using 4K pages
380 	 *
381 	 *    1. The sampling size is 32 bytes for basic sampling. This size
382 	 *	 is the same for all machine types. Diagnostic
383 	 *	 sampling uses auxlilary data buffer setup which provides the
384 	 *	 memory for SDBs using linux common code auxiliary trace
385 	 *	 setup.
386 	 *
387 	 *    2. Function alloc_sampling_buffer() sets the Alert Request
388 	 *	 Control indicator to trigger a measurement-alert to harvest
389 	 *	 sample-data-blocks (SDB). This is done per SDB. This
390 	 *	 measurement alert interrupt fires quick enough to handle
391 	 *	 one SDB, on very high frequency and work loads there might
392 	 *	 be 2 to 3 SBDs available for sample processing.
393 	 *	 Currently there is no need for setup alert request on every
394 	 *	 n-th page. This is counterproductive as one IRQ triggers
395 	 *	 a very high number of samples to be processed at one IRQ.
396 	 *
397 	 *    3. Use the sampling frequency as input.
398 	 *	 Compute the number of SDBs and ensure a minimum
399 	 *	 of CPUM_SF_MIN_SDB.  Depending on frequency add some more
400 	 *	 SDBs to handle a higher sampling rate.
401 	 *	 Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples
402 	 *	 (one SDB) for every 10000 HZ frequency increment.
403 	 *
404 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
405 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
406 	 *	 to 511 SDBs).
407 	 */
408 	sample_size = sizeof(struct hws_basic_entry);
409 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
410 	n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000);
411 
412 	/* If there is already a sampling buffer allocated, it is very likely
413 	 * that the sampling facility is enabled too.  If the event to be
414 	 * initialized requires a greater sampling buffer, the allocation must
415 	 * be postponed.  Changing the sampling buffer requires the sampling
416 	 * facility to be in the disabled state.  So, account the number of
417 	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
418 	 * before the event is started.
419 	 */
420 	sfb_init_allocs(n_sdb, hwc);
421 	if (sf_buffer_available(cpuhw))
422 		return 0;
423 
424 	debug_sprintf_event(sfdbg, 3,
425 			    "%s: rate %lu f %lu sdb %lu/%lu"
426 			    " sample_size %lu cpuhw %p\n", __func__,
427 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
428 			    sample_size, cpuhw);
429 
430 	return alloc_sampling_buffer(&cpuhw->sfb,
431 				     sfb_pending_allocs(&cpuhw->sfb, hwc));
432 }
433 
434 static unsigned long min_percent(unsigned int percent, unsigned long base,
435 				 unsigned long min)
436 {
437 	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
438 }
439 
440 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
441 {
442 	/* Use a percentage-based approach to extend the sampling facility
443 	 * buffer.  Accept up to 5% sample data loss.
444 	 * Vary the extents between 1% to 5% of the current number of
445 	 * sample-data-blocks.
446 	 */
447 	if (ratio <= 5)
448 		return 0;
449 	if (ratio <= 25)
450 		return min_percent(1, base, 1);
451 	if (ratio <= 50)
452 		return min_percent(1, base, 1);
453 	if (ratio <= 75)
454 		return min_percent(2, base, 2);
455 	if (ratio <= 100)
456 		return min_percent(3, base, 3);
457 	if (ratio <= 250)
458 		return min_percent(4, base, 4);
459 
460 	return min_percent(5, base, 8);
461 }
462 
463 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
464 				  struct hw_perf_event *hwc)
465 {
466 	unsigned long ratio, num;
467 
468 	if (!OVERFLOW_REG(hwc))
469 		return;
470 
471 	/* The sample_overflow contains the average number of sample data
472 	 * that has been lost because sample-data-blocks were full.
473 	 *
474 	 * Calculate the total number of sample data entries that has been
475 	 * discarded.  Then calculate the ratio of lost samples to total samples
476 	 * per second in percent.
477 	 */
478 	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
479 			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
480 
481 	/* Compute number of sample-data-blocks */
482 	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
483 	if (num)
484 		sfb_account_allocs(num, hwc);
485 
486 	debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n",
487 			    __func__, OVERFLOW_REG(hwc), ratio, num);
488 	OVERFLOW_REG(hwc) = 0;
489 }
490 
491 /* extend_sampling_buffer() - Extend sampling buffer
492  * @sfb:	Sampling buffer structure (for local CPU)
493  * @hwc:	Perf event hardware structure
494  *
495  * Use this function to extend the sampling buffer based on the overflow counter
496  * and postponed allocation extents stored in the specified Perf event hardware.
497  *
498  * Important: This function disables the sampling facility in order to safely
499  *	      change the sampling buffer structure.  Do not call this function
500  *	      when the PMU is active.
501  */
502 static void extend_sampling_buffer(struct sf_buffer *sfb,
503 				   struct hw_perf_event *hwc)
504 {
505 	unsigned long num, num_old;
506 	int rc;
507 
508 	num = sfb_pending_allocs(sfb, hwc);
509 	if (!num)
510 		return;
511 	num_old = sfb->num_sdb;
512 
513 	/* Disable the sampling facility to reset any states and also
514 	 * clear pending measurement alerts.
515 	 */
516 	sf_disable();
517 
518 	/* Extend the sampling buffer.
519 	 * This memory allocation typically happens in an atomic context when
520 	 * called by perf.  Because this is a reallocation, it is fine if the
521 	 * new SDB-request cannot be satisfied immediately.
522 	 */
523 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
524 	if (rc)
525 		debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n",
526 				    __func__, rc);
527 
528 	if (sfb_has_pending_allocs(sfb, hwc))
529 		debug_sprintf_event(sfdbg, 5, "%s: "
530 				    "req %lu alloc %lu remaining %lu\n",
531 				    __func__, num, sfb->num_sdb - num_old,
532 				    sfb_pending_allocs(sfb, hwc));
533 }
534 
535 /* Number of perf events counting hardware events */
536 static atomic_t num_events;
537 /* Used to avoid races in calling reserve/release_cpumf_hardware */
538 static DEFINE_MUTEX(pmc_reserve_mutex);
539 
540 #define PMC_INIT      0
541 #define PMC_RELEASE   1
542 #define PMC_FAILURE   2
543 static void setup_pmc_cpu(void *flags)
544 {
545 	int err;
546 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
547 
548 	err = 0;
549 	switch (*((int *) flags)) {
550 	case PMC_INIT:
551 		memset(cpusf, 0, sizeof(*cpusf));
552 		err = qsi(&cpusf->qsi);
553 		if (err)
554 			break;
555 		cpusf->flags |= PMU_F_RESERVED;
556 		err = sf_disable();
557 		if (err)
558 			pr_err("Switching off the sampling facility failed "
559 			       "with rc %i\n", err);
560 		debug_sprintf_event(sfdbg, 5,
561 				    "%s: initialized: cpuhw %p\n", __func__,
562 				    cpusf);
563 		break;
564 	case PMC_RELEASE:
565 		cpusf->flags &= ~PMU_F_RESERVED;
566 		err = sf_disable();
567 		if (err) {
568 			pr_err("Switching off the sampling facility failed "
569 			       "with rc %i\n", err);
570 		} else
571 			deallocate_buffers(cpusf);
572 		debug_sprintf_event(sfdbg, 5,
573 				    "%s: released: cpuhw %p\n", __func__,
574 				    cpusf);
575 		break;
576 	}
577 	if (err)
578 		*((int *) flags) |= PMC_FAILURE;
579 }
580 
581 static void release_pmc_hardware(void)
582 {
583 	int flags = PMC_RELEASE;
584 
585 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
586 	on_each_cpu(setup_pmc_cpu, &flags, 1);
587 }
588 
589 static int reserve_pmc_hardware(void)
590 {
591 	int flags = PMC_INIT;
592 
593 	on_each_cpu(setup_pmc_cpu, &flags, 1);
594 	if (flags & PMC_FAILURE) {
595 		release_pmc_hardware();
596 		return -ENODEV;
597 	}
598 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
599 
600 	return 0;
601 }
602 
603 static void hw_perf_event_destroy(struct perf_event *event)
604 {
605 	/* Release PMC if this is the last perf event */
606 	if (!atomic_add_unless(&num_events, -1, 1)) {
607 		mutex_lock(&pmc_reserve_mutex);
608 		if (atomic_dec_return(&num_events) == 0)
609 			release_pmc_hardware();
610 		mutex_unlock(&pmc_reserve_mutex);
611 	}
612 }
613 
614 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
615 {
616 	hwc->sample_period = period;
617 	hwc->last_period = hwc->sample_period;
618 	local64_set(&hwc->period_left, hwc->sample_period);
619 }
620 
621 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
622 				   unsigned long rate)
623 {
624 	return clamp_t(unsigned long, rate,
625 		       si->min_sampl_rate, si->max_sampl_rate);
626 }
627 
628 static u32 cpumsf_pid_type(struct perf_event *event,
629 			   u32 pid, enum pid_type type)
630 {
631 	struct task_struct *tsk;
632 
633 	/* Idle process */
634 	if (!pid)
635 		goto out;
636 
637 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
638 	pid = -1;
639 	if (tsk) {
640 		/*
641 		 * Only top level events contain the pid namespace in which
642 		 * they are created.
643 		 */
644 		if (event->parent)
645 			event = event->parent;
646 		pid = __task_pid_nr_ns(tsk, type, event->ns);
647 		/*
648 		 * See also 1d953111b648
649 		 * "perf/core: Don't report zero PIDs for exiting tasks".
650 		 */
651 		if (!pid && !pid_alive(tsk))
652 			pid = -1;
653 	}
654 out:
655 	return pid;
656 }
657 
658 static void cpumsf_output_event_pid(struct perf_event *event,
659 				    struct perf_sample_data *data,
660 				    struct pt_regs *regs)
661 {
662 	u32 pid;
663 	struct perf_event_header header;
664 	struct perf_output_handle handle;
665 
666 	/*
667 	 * Obtain the PID from the basic-sampling data entry and
668 	 * correct the data->tid_entry.pid value.
669 	 */
670 	pid = data->tid_entry.pid;
671 
672 	/* Protect callchain buffers, tasks */
673 	rcu_read_lock();
674 
675 	perf_prepare_sample(&header, data, event, regs);
676 	if (perf_output_begin(&handle, data, event, header.size))
677 		goto out;
678 
679 	/* Update the process ID (see also kernel/events/core.c) */
680 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
681 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
682 
683 	perf_output_sample(&handle, &header, data, event);
684 	perf_output_end(&handle);
685 out:
686 	rcu_read_unlock();
687 }
688 
689 static unsigned long getrate(bool freq, unsigned long sample,
690 			     struct hws_qsi_info_block *si)
691 {
692 	unsigned long rate;
693 
694 	if (freq) {
695 		rate = freq_to_sample_rate(si, sample);
696 		rate = hw_limit_rate(si, rate);
697 	} else {
698 		/* The min/max sampling rates specifies the valid range
699 		 * of sample periods.  If the specified sample period is
700 		 * out of range, limit the period to the range boundary.
701 		 */
702 		rate = hw_limit_rate(si, sample);
703 
704 		/* The perf core maintains a maximum sample rate that is
705 		 * configurable through the sysctl interface.  Ensure the
706 		 * sampling rate does not exceed this value.  This also helps
707 		 * to avoid throttling when pushing samples with
708 		 * perf_event_overflow().
709 		 */
710 		if (sample_rate_to_freq(si, rate) >
711 		    sysctl_perf_event_sample_rate) {
712 			debug_sprintf_event(sfdbg, 1, "%s: "
713 					    "Sampling rate exceeds maximum "
714 					    "perf sample rate\n", __func__);
715 			rate = 0;
716 		}
717 	}
718 	return rate;
719 }
720 
721 /* The sampling information (si) contains information about the
722  * min/max sampling intervals and the CPU speed.  So calculate the
723  * correct sampling interval and avoid the whole period adjust
724  * feedback loop.
725  *
726  * Since the CPU Measurement sampling facility can not handle frequency
727  * calculate the sampling interval when frequency is specified using
728  * this formula:
729  *	interval := cpu_speed * 1000000 / sample_freq
730  *
731  * Returns errno on bad input and zero on success with parameter interval
732  * set to the correct sampling rate.
733  *
734  * Note: This function turns off freq bit to avoid calling function
735  * perf_adjust_period(). This causes frequency adjustment in the common
736  * code part which causes tremendous variations in the counter values.
737  */
738 static int __hw_perf_event_init_rate(struct perf_event *event,
739 				     struct hws_qsi_info_block *si)
740 {
741 	struct perf_event_attr *attr = &event->attr;
742 	struct hw_perf_event *hwc = &event->hw;
743 	unsigned long rate;
744 
745 	if (attr->freq) {
746 		if (!attr->sample_freq)
747 			return -EINVAL;
748 		rate = getrate(attr->freq, attr->sample_freq, si);
749 		attr->freq = 0;		/* Don't call  perf_adjust_period() */
750 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
751 	} else {
752 		rate = getrate(attr->freq, attr->sample_period, si);
753 		if (!rate)
754 			return -EINVAL;
755 	}
756 	attr->sample_period = rate;
757 	SAMPL_RATE(hwc) = rate;
758 	hw_init_period(hwc, SAMPL_RATE(hwc));
759 	debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n",
760 			    __func__, event->cpu, event->attr.sample_period,
761 			    event->attr.freq, SAMPLE_FREQ_MODE(hwc));
762 	return 0;
763 }
764 
765 static int __hw_perf_event_init(struct perf_event *event)
766 {
767 	struct cpu_hw_sf *cpuhw;
768 	struct hws_qsi_info_block si;
769 	struct perf_event_attr *attr = &event->attr;
770 	struct hw_perf_event *hwc = &event->hw;
771 	int cpu, err;
772 
773 	/* Reserve CPU-measurement sampling facility */
774 	err = 0;
775 	if (!atomic_inc_not_zero(&num_events)) {
776 		mutex_lock(&pmc_reserve_mutex);
777 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
778 			err = -EBUSY;
779 		else
780 			atomic_inc(&num_events);
781 		mutex_unlock(&pmc_reserve_mutex);
782 	}
783 	event->destroy = hw_perf_event_destroy;
784 
785 	if (err)
786 		goto out;
787 
788 	/* Access per-CPU sampling information (query sampling info) */
789 	/*
790 	 * The event->cpu value can be -1 to count on every CPU, for example,
791 	 * when attaching to a task.  If this is specified, use the query
792 	 * sampling info from the current CPU, otherwise use event->cpu to
793 	 * retrieve the per-CPU information.
794 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
795 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
796 	 */
797 	memset(&si, 0, sizeof(si));
798 	cpuhw = NULL;
799 	if (event->cpu == -1)
800 		qsi(&si);
801 	else {
802 		/* Event is pinned to a particular CPU, retrieve the per-CPU
803 		 * sampling structure for accessing the CPU-specific QSI.
804 		 */
805 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
806 		si = cpuhw->qsi;
807 	}
808 
809 	/* Check sampling facility authorization and, if not authorized,
810 	 * fall back to other PMUs.  It is safe to check any CPU because
811 	 * the authorization is identical for all configured CPUs.
812 	 */
813 	if (!si.as) {
814 		err = -ENOENT;
815 		goto out;
816 	}
817 
818 	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
819 		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
820 		err = -EBUSY;
821 		goto out;
822 	}
823 
824 	/* Always enable basic sampling */
825 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
826 
827 	/* Check if diagnostic sampling is requested.  Deny if the required
828 	 * sampling authorization is missing.
829 	 */
830 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
831 		if (!si.ad) {
832 			err = -EPERM;
833 			goto out;
834 		}
835 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
836 	}
837 
838 	/* Check and set other sampling flags */
839 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
840 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
841 
842 	err =  __hw_perf_event_init_rate(event, &si);
843 	if (err)
844 		goto out;
845 
846 	/* Initialize sample data overflow accounting */
847 	hwc->extra_reg.reg = REG_OVERFLOW;
848 	OVERFLOW_REG(hwc) = 0;
849 
850 	/* Use AUX buffer. No need to allocate it by ourself */
851 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
852 		return 0;
853 
854 	/* Allocate the per-CPU sampling buffer using the CPU information
855 	 * from the event.  If the event is not pinned to a particular
856 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
857 	 * buffers for each online CPU.
858 	 */
859 	if (cpuhw)
860 		/* Event is pinned to a particular CPU */
861 		err = allocate_buffers(cpuhw, hwc);
862 	else {
863 		/* Event is not pinned, allocate sampling buffer on
864 		 * each online CPU
865 		 */
866 		for_each_online_cpu(cpu) {
867 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
868 			err = allocate_buffers(cpuhw, hwc);
869 			if (err)
870 				break;
871 		}
872 	}
873 
874 	/* If PID/TID sampling is active, replace the default overflow
875 	 * handler to extract and resolve the PIDs from the basic-sampling
876 	 * data entries.
877 	 */
878 	if (event->attr.sample_type & PERF_SAMPLE_TID)
879 		if (is_default_overflow_handler(event))
880 			event->overflow_handler = cpumsf_output_event_pid;
881 out:
882 	return err;
883 }
884 
885 static bool is_callchain_event(struct perf_event *event)
886 {
887 	u64 sample_type = event->attr.sample_type;
888 
889 	return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER |
890 			      PERF_SAMPLE_STACK_USER);
891 }
892 
893 static int cpumsf_pmu_event_init(struct perf_event *event)
894 {
895 	int err;
896 
897 	/* No support for taken branch sampling */
898 	/* No support for callchain, stacks and registers */
899 	if (has_branch_stack(event) || is_callchain_event(event))
900 		return -EOPNOTSUPP;
901 
902 	switch (event->attr.type) {
903 	case PERF_TYPE_RAW:
904 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
905 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
906 			return -ENOENT;
907 		break;
908 	case PERF_TYPE_HARDWARE:
909 		/* Support sampling of CPU cycles in addition to the
910 		 * counter facility.  However, the counter facility
911 		 * is more precise and, hence, restrict this PMU to
912 		 * sampling events only.
913 		 */
914 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
915 			return -ENOENT;
916 		if (!is_sampling_event(event))
917 			return -ENOENT;
918 		break;
919 	default:
920 		return -ENOENT;
921 	}
922 
923 	/* Check online status of the CPU to which the event is pinned */
924 	if (event->cpu >= 0 && !cpu_online(event->cpu))
925 		return -ENODEV;
926 
927 	/* Force reset of idle/hv excludes regardless of what the
928 	 * user requested.
929 	 */
930 	if (event->attr.exclude_hv)
931 		event->attr.exclude_hv = 0;
932 	if (event->attr.exclude_idle)
933 		event->attr.exclude_idle = 0;
934 
935 	err = __hw_perf_event_init(event);
936 	if (unlikely(err))
937 		if (event->destroy)
938 			event->destroy(event);
939 	return err;
940 }
941 
942 static void cpumsf_pmu_enable(struct pmu *pmu)
943 {
944 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
945 	struct hw_perf_event *hwc;
946 	int err;
947 
948 	if (cpuhw->flags & PMU_F_ENABLED)
949 		return;
950 
951 	if (cpuhw->flags & PMU_F_ERR_MASK)
952 		return;
953 
954 	/* Check whether to extent the sampling buffer.
955 	 *
956 	 * Two conditions trigger an increase of the sampling buffer for a
957 	 * perf event:
958 	 *    1. Postponed buffer allocations from the event initialization.
959 	 *    2. Sampling overflows that contribute to pending allocations.
960 	 *
961 	 * Note that the extend_sampling_buffer() function disables the sampling
962 	 * facility, but it can be fully re-enabled using sampling controls that
963 	 * have been saved in cpumsf_pmu_disable().
964 	 */
965 	if (cpuhw->event) {
966 		hwc = &cpuhw->event->hw;
967 		if (!(SAMPL_DIAG_MODE(hwc))) {
968 			/*
969 			 * Account number of overflow-designated
970 			 * buffer extents
971 			 */
972 			sfb_account_overflows(cpuhw, hwc);
973 			extend_sampling_buffer(&cpuhw->sfb, hwc);
974 		}
975 		/* Rate may be adjusted with ioctl() */
976 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
977 	}
978 
979 	/* (Re)enable the PMU and sampling facility */
980 	cpuhw->flags |= PMU_F_ENABLED;
981 	barrier();
982 
983 	err = lsctl(&cpuhw->lsctl);
984 	if (err) {
985 		cpuhw->flags &= ~PMU_F_ENABLED;
986 		pr_err("Loading sampling controls failed: op %i err %i\n",
987 			1, err);
988 		return;
989 	}
990 
991 	/* Load current program parameter */
992 	lpp(&S390_lowcore.lpp);
993 
994 	debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i "
995 			    "interval %#lx tear %#lx dear %#lx\n", __func__,
996 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
997 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
998 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
999 }
1000 
1001 static void cpumsf_pmu_disable(struct pmu *pmu)
1002 {
1003 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1004 	struct hws_lsctl_request_block inactive;
1005 	struct hws_qsi_info_block si;
1006 	int err;
1007 
1008 	if (!(cpuhw->flags & PMU_F_ENABLED))
1009 		return;
1010 
1011 	if (cpuhw->flags & PMU_F_ERR_MASK)
1012 		return;
1013 
1014 	/* Switch off sampling activation control */
1015 	inactive = cpuhw->lsctl;
1016 	inactive.cs = 0;
1017 	inactive.cd = 0;
1018 
1019 	err = lsctl(&inactive);
1020 	if (err) {
1021 		pr_err("Loading sampling controls failed: op %i err %i\n",
1022 			2, err);
1023 		return;
1024 	}
1025 
1026 	/* Save state of TEAR and DEAR register contents */
1027 	err = qsi(&si);
1028 	if (!err) {
1029 		/* TEAR/DEAR values are valid only if the sampling facility is
1030 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1031 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1032 		 * controls the enable/disable state.
1033 		 */
1034 		if (si.es) {
1035 			cpuhw->lsctl.tear = si.tear;
1036 			cpuhw->lsctl.dear = si.dear;
1037 		}
1038 	} else
1039 		debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n",
1040 				    __func__, err);
1041 
1042 	cpuhw->flags &= ~PMU_F_ENABLED;
1043 }
1044 
1045 /* perf_exclude_event() - Filter event
1046  * @event:	The perf event
1047  * @regs:	pt_regs structure
1048  * @sde_regs:	Sample-data-entry (sde) regs structure
1049  *
1050  * Filter perf events according to their exclude specification.
1051  *
1052  * Return non-zero if the event shall be excluded.
1053  */
1054 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1055 			      struct perf_sf_sde_regs *sde_regs)
1056 {
1057 	if (event->attr.exclude_user && user_mode(regs))
1058 		return 1;
1059 	if (event->attr.exclude_kernel && !user_mode(regs))
1060 		return 1;
1061 	if (event->attr.exclude_guest && sde_regs->in_guest)
1062 		return 1;
1063 	if (event->attr.exclude_host && !sde_regs->in_guest)
1064 		return 1;
1065 	return 0;
1066 }
1067 
1068 /* perf_push_sample() - Push samples to perf
1069  * @event:	The perf event
1070  * @sample:	Hardware sample data
1071  *
1072  * Use the hardware sample data to create perf event sample.  The sample
1073  * is the pushed to the event subsystem and the function checks for
1074  * possible event overflows.  If an event overflow occurs, the PMU is
1075  * stopped.
1076  *
1077  * Return non-zero if an event overflow occurred.
1078  */
1079 static int perf_push_sample(struct perf_event *event,
1080 			    struct hws_basic_entry *basic)
1081 {
1082 	int overflow;
1083 	struct pt_regs regs;
1084 	struct perf_sf_sde_regs *sde_regs;
1085 	struct perf_sample_data data;
1086 
1087 	/* Setup perf sample */
1088 	perf_sample_data_init(&data, 0, event->hw.last_period);
1089 
1090 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1091 	 * using the Program Request Alert code.  The regs.int_parm_long
1092 	 * field which is unused contains additional sample-data-entry related
1093 	 * indicators.
1094 	 */
1095 	memset(&regs, 0, sizeof(regs));
1096 	regs.int_code = 0x1407;
1097 	regs.int_parm = CPU_MF_INT_SF_PRA;
1098 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1099 
1100 	psw_bits(regs.psw).ia	= basic->ia;
1101 	psw_bits(regs.psw).dat	= basic->T;
1102 	psw_bits(regs.psw).wait = basic->W;
1103 	psw_bits(regs.psw).pstate = basic->P;
1104 	psw_bits(regs.psw).as	= basic->AS;
1105 
1106 	/*
1107 	 * Use the hardware provided configuration level to decide if the
1108 	 * sample belongs to a guest or host. If that is not available,
1109 	 * fall back to the following heuristics:
1110 	 * A non-zero guest program parameter always indicates a guest
1111 	 * sample. Some early samples or samples from guests without
1112 	 * lpp usage would be misaccounted to the host. We use the asn
1113 	 * value as an addon heuristic to detect most of these guest samples.
1114 	 * If the value differs from 0xffff (the host value), we assume to
1115 	 * be a KVM guest.
1116 	 */
1117 	switch (basic->CL) {
1118 	case 1: /* logical partition */
1119 		sde_regs->in_guest = 0;
1120 		break;
1121 	case 2: /* virtual machine */
1122 		sde_regs->in_guest = 1;
1123 		break;
1124 	default: /* old machine, use heuristics */
1125 		if (basic->gpp || basic->prim_asn != 0xffff)
1126 			sde_regs->in_guest = 1;
1127 		break;
1128 	}
1129 
1130 	/*
1131 	 * Store the PID value from the sample-data-entry to be
1132 	 * processed and resolved by cpumsf_output_event_pid().
1133 	 */
1134 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1135 
1136 	overflow = 0;
1137 	if (perf_exclude_event(event, &regs, sde_regs))
1138 		goto out;
1139 	if (perf_event_overflow(event, &data, &regs)) {
1140 		overflow = 1;
1141 		event->pmu->stop(event, 0);
1142 	}
1143 	perf_event_update_userpage(event);
1144 out:
1145 	return overflow;
1146 }
1147 
1148 static void perf_event_count_update(struct perf_event *event, u64 count)
1149 {
1150 	local64_add(count, &event->count);
1151 }
1152 
1153 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1154  * @event:	The perf event
1155  * @sdbt:	Sample-data-block table
1156  * @overflow:	Event overflow counter
1157  *
1158  * Walks through a sample-data-block and collects sampling data entries that are
1159  * then pushed to the perf event subsystem.  Depending on the sampling function,
1160  * there can be either basic-sampling or combined-sampling data entries.  A
1161  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1162  * data entry.	The sampling function is determined by the flags in the perf
1163  * event hardware structure.  The function always works with a combined-sampling
1164  * data entry but ignores the the diagnostic portion if it is not available.
1165  *
1166  * Note that the implementation focuses on basic-sampling data entries and, if
1167  * such an entry is not valid, the entire combined-sampling data entry is
1168  * ignored.
1169  *
1170  * The overflow variables counts the number of samples that has been discarded
1171  * due to a perf event overflow.
1172  */
1173 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1174 			       unsigned long long *overflow)
1175 {
1176 	struct hws_trailer_entry *te;
1177 	struct hws_basic_entry *sample;
1178 
1179 	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1180 	sample = (struct hws_basic_entry *) *sdbt;
1181 	while ((unsigned long *) sample < (unsigned long *) te) {
1182 		/* Check for an empty sample */
1183 		if (!sample->def || sample->LS)
1184 			break;
1185 
1186 		/* Update perf event period */
1187 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1188 
1189 		/* Check whether sample is valid */
1190 		if (sample->def == 0x0001) {
1191 			/* If an event overflow occurred, the PMU is stopped to
1192 			 * throttle event delivery.  Remaining sample data is
1193 			 * discarded.
1194 			 */
1195 			if (!*overflow) {
1196 				/* Check whether sample is consistent */
1197 				if (sample->I == 0 && sample->W == 0) {
1198 					/* Deliver sample data to perf */
1199 					*overflow = perf_push_sample(event,
1200 								     sample);
1201 				}
1202 			} else
1203 				/* Count discarded samples */
1204 				*overflow += 1;
1205 		} else {
1206 			debug_sprintf_event(sfdbg, 4,
1207 					    "%s: Found unknown"
1208 					    " sampling data entry: te->f %i"
1209 					    " basic.def %#4x (%p)\n", __func__,
1210 					    te->header.f, sample->def, sample);
1211 			/* Sample slot is not yet written or other record.
1212 			 *
1213 			 * This condition can occur if the buffer was reused
1214 			 * from a combined basic- and diagnostic-sampling.
1215 			 * If only basic-sampling is then active, entries are
1216 			 * written into the larger diagnostic entries.
1217 			 * This is typically the case for sample-data-blocks
1218 			 * that are not full.  Stop processing if the first
1219 			 * invalid format was detected.
1220 			 */
1221 			if (!te->header.f)
1222 				break;
1223 		}
1224 
1225 		/* Reset sample slot and advance to next sample */
1226 		sample->def = 0;
1227 		sample++;
1228 	}
1229 }
1230 
1231 static inline __uint128_t __cdsg(__uint128_t *ptr, __uint128_t old, __uint128_t new)
1232 {
1233 	asm volatile(
1234 		"	cdsg	%[old],%[new],%[ptr]\n"
1235 		: [old] "+d" (old), [ptr] "+QS" (*ptr)
1236 		: [new] "d" (new)
1237 		: "memory", "cc");
1238 	return old;
1239 }
1240 
1241 /* hw_perf_event_update() - Process sampling buffer
1242  * @event:	The perf event
1243  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1244  *
1245  * Processes the sampling buffer and create perf event samples.
1246  * The sampling buffer position are retrieved and saved in the TEAR_REG
1247  * register of the specified perf event.
1248  *
1249  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1250  * to also walk through partially filled sample-data-blocks.  It is ignored
1251  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1252  * enforces the processing of full sample-data-blocks only (trailer entries
1253  * with the block-full-indicator bit set).
1254  */
1255 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1256 {
1257 	unsigned long long event_overflow, sampl_overflow, num_sdb;
1258 	union hws_trailer_header old, prev, new;
1259 	struct hw_perf_event *hwc = &event->hw;
1260 	struct hws_trailer_entry *te;
1261 	unsigned long *sdbt;
1262 	int done;
1263 
1264 	/*
1265 	 * AUX buffer is used when in diagnostic sampling mode.
1266 	 * No perf events/samples are created.
1267 	 */
1268 	if (SAMPL_DIAG_MODE(&event->hw))
1269 		return;
1270 
1271 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1272 		flush_all = 0;
1273 
1274 	sdbt = (unsigned long *) TEAR_REG(hwc);
1275 	done = event_overflow = sampl_overflow = num_sdb = 0;
1276 	while (!done) {
1277 		/* Get the trailer entry of the sample-data-block */
1278 		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
1279 
1280 		/* Leave loop if no more work to do (block full indicator) */
1281 		if (!te->header.f) {
1282 			done = 1;
1283 			if (!flush_all)
1284 				break;
1285 		}
1286 
1287 		/* Check the sample overflow count */
1288 		if (te->header.overflow)
1289 			/* Account sample overflows and, if a particular limit
1290 			 * is reached, extend the sampling buffer.
1291 			 * For details, see sfb_account_overflows().
1292 			 */
1293 			sampl_overflow += te->header.overflow;
1294 
1295 		/* Timestamps are valid for full sample-data-blocks only */
1296 		debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx "
1297 				    "overflow %llu timestamp %#llx\n",
1298 				    __func__, (unsigned long)sdbt, te->header.overflow,
1299 				    (te->header.f) ? trailer_timestamp(te) : 0ULL);
1300 
1301 		/* Collect all samples from a single sample-data-block and
1302 		 * flag if an (perf) event overflow happened.  If so, the PMU
1303 		 * is stopped and remaining samples will be discarded.
1304 		 */
1305 		hw_collect_samples(event, sdbt, &event_overflow);
1306 		num_sdb++;
1307 
1308 		/* Reset trailer (using compare-double-and-swap) */
1309 		/* READ_ONCE() 16 byte header */
1310 		prev.val = __cdsg(&te->header.val, 0, 0);
1311 		do {
1312 			old.val = prev.val;
1313 			new.val = prev.val;
1314 			new.f = 0;
1315 			new.a = 1;
1316 			new.overflow = 0;
1317 			prev.val = __cdsg(&te->header.val, old.val, new.val);
1318 		} while (prev.val != old.val);
1319 
1320 		/* Advance to next sample-data-block */
1321 		sdbt++;
1322 		if (is_link_entry(sdbt))
1323 			sdbt = get_next_sdbt(sdbt);
1324 
1325 		/* Update event hardware registers */
1326 		TEAR_REG(hwc) = (unsigned long) sdbt;
1327 
1328 		/* Stop processing sample-data if all samples of the current
1329 		 * sample-data-block were flushed even if it was not full.
1330 		 */
1331 		if (flush_all && done)
1332 			break;
1333 	}
1334 
1335 	/* Account sample overflows in the event hardware structure */
1336 	if (sampl_overflow)
1337 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1338 						 sampl_overflow, 1 + num_sdb);
1339 
1340 	/* Perf_event_overflow() and perf_event_account_interrupt() limit
1341 	 * the interrupt rate to an upper limit. Roughly 1000 samples per
1342 	 * task tick.
1343 	 * Hitting this limit results in a large number
1344 	 * of throttled REF_REPORT_THROTTLE entries and the samples
1345 	 * are dropped.
1346 	 * Slightly increase the interval to avoid hitting this limit.
1347 	 */
1348 	if (event_overflow) {
1349 		SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
1350 		debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n",
1351 				    __func__,
1352 				    DIV_ROUND_UP(SAMPL_RATE(hwc), 10));
1353 	}
1354 
1355 	if (sampl_overflow || event_overflow)
1356 		debug_sprintf_event(sfdbg, 4, "%s: "
1357 				    "overflows: sample %llu event %llu"
1358 				    " total %llu num_sdb %llu\n",
1359 				    __func__, sampl_overflow, event_overflow,
1360 				    OVERFLOW_REG(hwc), num_sdb);
1361 }
1362 
1363 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb)
1364 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0)
1365 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark)
1366 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark)
1367 
1368 /*
1369  * Get trailer entry by index of SDB.
1370  */
1371 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1372 						 unsigned long index)
1373 {
1374 	unsigned long sdb;
1375 
1376 	index = AUX_SDB_INDEX(aux, index);
1377 	sdb = aux->sdb_index[index];
1378 	return (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1379 }
1380 
1381 /*
1382  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1383  * disabled. Collect the full SDBs in AUX buffer which have not reached
1384  * the point of alert indicator. And ignore the SDBs which are not
1385  * full.
1386  *
1387  * 1. Scan SDBs to see how much data is there and consume them.
1388  * 2. Remove alert indicator in the buffer.
1389  */
1390 static void aux_output_end(struct perf_output_handle *handle)
1391 {
1392 	unsigned long i, range_scan, idx;
1393 	struct aux_buffer *aux;
1394 	struct hws_trailer_entry *te;
1395 
1396 	aux = perf_get_aux(handle);
1397 	if (!aux)
1398 		return;
1399 
1400 	range_scan = AUX_SDB_NUM_ALERT(aux);
1401 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1402 		te = aux_sdb_trailer(aux, idx);
1403 		if (!te->header.f)
1404 			break;
1405 	}
1406 	/* i is num of SDBs which are full */
1407 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1408 
1409 	/* Remove alert indicators in the buffer */
1410 	te = aux_sdb_trailer(aux, aux->alert_mark);
1411 	te->header.a = 0;
1412 
1413 	debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n",
1414 			    __func__, i, range_scan, aux->head);
1415 }
1416 
1417 /*
1418  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1419  * is first added to the CPU or rescheduled again to the CPU. It is called
1420  * with pmu disabled.
1421  *
1422  * 1. Reset the trailer of SDBs to get ready for new data.
1423  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1424  *    head(tear/dear).
1425  */
1426 static int aux_output_begin(struct perf_output_handle *handle,
1427 			    struct aux_buffer *aux,
1428 			    struct cpu_hw_sf *cpuhw)
1429 {
1430 	unsigned long range;
1431 	unsigned long i, range_scan, idx;
1432 	unsigned long head, base, offset;
1433 	struct hws_trailer_entry *te;
1434 
1435 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1436 		return -EINVAL;
1437 
1438 	aux->head = handle->head >> PAGE_SHIFT;
1439 	range = (handle->size + 1) >> PAGE_SHIFT;
1440 	if (range <= 1)
1441 		return -ENOMEM;
1442 
1443 	/*
1444 	 * SDBs between aux->head and aux->empty_mark are already ready
1445 	 * for new data. range_scan is num of SDBs not within them.
1446 	 */
1447 	debug_sprintf_event(sfdbg, 6,
1448 			    "%s: range %ld head %ld alert %ld empty %ld\n",
1449 			    __func__, range, aux->head, aux->alert_mark,
1450 			    aux->empty_mark);
1451 	if (range > AUX_SDB_NUM_EMPTY(aux)) {
1452 		range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1453 		idx = aux->empty_mark + 1;
1454 		for (i = 0; i < range_scan; i++, idx++) {
1455 			te = aux_sdb_trailer(aux, idx);
1456 			te->header.f = 0;
1457 			te->header.a = 0;
1458 			te->header.overflow = 0;
1459 		}
1460 		/* Save the position of empty SDBs */
1461 		aux->empty_mark = aux->head + range - 1;
1462 	}
1463 
1464 	/* Set alert indicator */
1465 	aux->alert_mark = aux->head + range/2 - 1;
1466 	te = aux_sdb_trailer(aux, aux->alert_mark);
1467 	te->header.a = 1;
1468 
1469 	/* Reset hardware buffer head */
1470 	head = AUX_SDB_INDEX(aux, aux->head);
1471 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1472 	offset = head % CPUM_SF_SDB_PER_TABLE;
1473 	cpuhw->lsctl.tear = base + offset * sizeof(unsigned long);
1474 	cpuhw->lsctl.dear = aux->sdb_index[head];
1475 
1476 	debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld "
1477 			    "index %ld tear %#lx dear %#lx\n", __func__,
1478 			    aux->head, aux->alert_mark, aux->empty_mark,
1479 			    head / CPUM_SF_SDB_PER_TABLE,
1480 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
1481 
1482 	return 0;
1483 }
1484 
1485 /*
1486  * Set alert indicator on SDB at index @alert_index while sampler is running.
1487  *
1488  * Return true if successfully.
1489  * Return false if full indicator is already set by hardware sampler.
1490  */
1491 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1492 			  unsigned long long *overflow)
1493 {
1494 	union hws_trailer_header old, prev, new;
1495 	struct hws_trailer_entry *te;
1496 
1497 	te = aux_sdb_trailer(aux, alert_index);
1498 	/* READ_ONCE() 16 byte header */
1499 	prev.val = __cdsg(&te->header.val, 0, 0);
1500 	do {
1501 		old.val = prev.val;
1502 		new.val = prev.val;
1503 		*overflow = old.overflow;
1504 		if (old.f) {
1505 			/*
1506 			 * SDB is already set by hardware.
1507 			 * Abort and try to set somewhere
1508 			 * behind.
1509 			 */
1510 			return false;
1511 		}
1512 		new.a = 1;
1513 		new.overflow = 0;
1514 		prev.val = __cdsg(&te->header.val, old.val, new.val);
1515 	} while (prev.val != old.val);
1516 	return true;
1517 }
1518 
1519 /*
1520  * aux_reset_buffer() - Scan and setup SDBs for new samples
1521  * @aux:	The AUX buffer to set
1522  * @range:	The range of SDBs to scan started from aux->head
1523  * @overflow:	Set to overflow count
1524  *
1525  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1526  * marked as empty, check if it is already set full by the hardware sampler.
1527  * If yes, that means new data is already there before we can set an alert
1528  * indicator. Caller should try to set alert indicator to some position behind.
1529  *
1530  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1531  * previously and have already been consumed by user space. Reset these SDBs
1532  * (clear full indicator and alert indicator) for new data.
1533  * If aux->alert_mark fall in this area, just set it. Overflow count is
1534  * recorded while scanning.
1535  *
1536  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1537  * and ready for new samples. So scanning on this area could be skipped.
1538  *
1539  * Return true if alert indicator is set successfully and false if not.
1540  */
1541 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1542 			     unsigned long long *overflow)
1543 {
1544 	unsigned long i, range_scan, idx, idx_old;
1545 	union hws_trailer_header old, prev, new;
1546 	unsigned long long orig_overflow;
1547 	struct hws_trailer_entry *te;
1548 
1549 	debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld "
1550 			    "empty %ld\n", __func__, range, aux->head,
1551 			    aux->alert_mark, aux->empty_mark);
1552 	if (range <= AUX_SDB_NUM_EMPTY(aux))
1553 		/*
1554 		 * No need to scan. All SDBs in range are marked as empty.
1555 		 * Just set alert indicator. Should check race with hardware
1556 		 * sampler.
1557 		 */
1558 		return aux_set_alert(aux, aux->alert_mark, overflow);
1559 
1560 	if (aux->alert_mark <= aux->empty_mark)
1561 		/*
1562 		 * Set alert indicator on empty SDB. Should check race
1563 		 * with hardware sampler.
1564 		 */
1565 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1566 			return false;
1567 
1568 	/*
1569 	 * Scan the SDBs to clear full and alert indicator used previously.
1570 	 * Start scanning from one SDB behind empty_mark. If the new alert
1571 	 * indicator fall into this range, set it.
1572 	 */
1573 	range_scan = range - AUX_SDB_NUM_EMPTY(aux);
1574 	idx_old = idx = aux->empty_mark + 1;
1575 	for (i = 0; i < range_scan; i++, idx++) {
1576 		te = aux_sdb_trailer(aux, idx);
1577 		/* READ_ONCE() 16 byte header */
1578 		prev.val = __cdsg(&te->header.val, 0, 0);
1579 		do {
1580 			old.val = prev.val;
1581 			new.val = prev.val;
1582 			orig_overflow = old.overflow;
1583 			new.f = 0;
1584 			new.overflow = 0;
1585 			if (idx == aux->alert_mark)
1586 				new.a = 1;
1587 			else
1588 				new.a = 0;
1589 			prev.val = __cdsg(&te->header.val, old.val, new.val);
1590 		} while (prev.val != old.val);
1591 		*overflow += orig_overflow;
1592 	}
1593 
1594 	/* Update empty_mark to new position */
1595 	aux->empty_mark = aux->head + range - 1;
1596 
1597 	debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld "
1598 			    "empty %ld\n", __func__, range_scan, idx_old,
1599 			    idx - 1, aux->empty_mark);
1600 	return true;
1601 }
1602 
1603 /*
1604  * Measurement alert handler for diagnostic mode sampling.
1605  */
1606 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1607 {
1608 	struct aux_buffer *aux;
1609 	int done = 0;
1610 	unsigned long range = 0, size;
1611 	unsigned long long overflow = 0;
1612 	struct perf_output_handle *handle = &cpuhw->handle;
1613 	unsigned long num_sdb;
1614 
1615 	aux = perf_get_aux(handle);
1616 	if (WARN_ON_ONCE(!aux))
1617 		return;
1618 
1619 	/* Inform user space new data arrived */
1620 	size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1621 	debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__,
1622 			    size >> PAGE_SHIFT);
1623 	perf_aux_output_end(handle, size);
1624 
1625 	num_sdb = aux->sfb.num_sdb;
1626 	while (!done) {
1627 		/* Get an output handle */
1628 		aux = perf_aux_output_begin(handle, cpuhw->event);
1629 		if (handle->size == 0) {
1630 			pr_err("The AUX buffer with %lu pages for the "
1631 			       "diagnostic-sampling mode is full\n",
1632 				num_sdb);
1633 			debug_sprintf_event(sfdbg, 1,
1634 					    "%s: AUX buffer used up\n",
1635 					    __func__);
1636 			break;
1637 		}
1638 		if (WARN_ON_ONCE(!aux))
1639 			return;
1640 
1641 		/* Update head and alert_mark to new position */
1642 		aux->head = handle->head >> PAGE_SHIFT;
1643 		range = (handle->size + 1) >> PAGE_SHIFT;
1644 		if (range == 1)
1645 			aux->alert_mark = aux->head;
1646 		else
1647 			aux->alert_mark = aux->head + range/2 - 1;
1648 
1649 		if (aux_reset_buffer(aux, range, &overflow)) {
1650 			if (!overflow) {
1651 				done = 1;
1652 				break;
1653 			}
1654 			size = range << PAGE_SHIFT;
1655 			perf_aux_output_end(&cpuhw->handle, size);
1656 			pr_err("Sample data caused the AUX buffer with %lu "
1657 			       "pages to overflow\n", aux->sfb.num_sdb);
1658 			debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld "
1659 					    "overflow %lld\n", __func__,
1660 					    aux->head, range, overflow);
1661 		} else {
1662 			size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT;
1663 			perf_aux_output_end(&cpuhw->handle, size);
1664 			debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1665 					    "already full, try another\n",
1666 					    __func__,
1667 					    aux->head, aux->alert_mark);
1668 		}
1669 	}
1670 
1671 	if (done)
1672 		debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1673 				    "empty %ld\n", __func__, aux->head,
1674 				    aux->alert_mark, aux->empty_mark);
1675 }
1676 
1677 /*
1678  * Callback when freeing AUX buffers.
1679  */
1680 static void aux_buffer_free(void *data)
1681 {
1682 	struct aux_buffer *aux = data;
1683 	unsigned long i, num_sdbt;
1684 
1685 	if (!aux)
1686 		return;
1687 
1688 	/* Free SDBT. SDB is freed by the caller */
1689 	num_sdbt = aux->sfb.num_sdbt;
1690 	for (i = 0; i < num_sdbt; i++)
1691 		free_page(aux->sdbt_index[i]);
1692 
1693 	kfree(aux->sdbt_index);
1694 	kfree(aux->sdb_index);
1695 	kfree(aux);
1696 
1697 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt);
1698 }
1699 
1700 static void aux_sdb_init(unsigned long sdb)
1701 {
1702 	struct hws_trailer_entry *te;
1703 
1704 	te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb);
1705 
1706 	/* Save clock base */
1707 	te->clock_base = 1;
1708 	te->progusage2 = tod_clock_base.tod;
1709 }
1710 
1711 /*
1712  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1713  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1714  * @pages:	Array of pointers to buffer pages passed from perf core
1715  * @nr_pages:	Total pages
1716  * @snapshot:	Flag for snapshot mode
1717  *
1718  * This is the callback when setup an event using AUX buffer. Perf tool can
1719  * trigger this by an additional mmap() call on the event. Unlike the buffer
1720  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1721  * the task among online cpus when it is a per-thread event.
1722  *
1723  * Return the private AUX buffer structure if success or NULL if fails.
1724  */
1725 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1726 			      int nr_pages, bool snapshot)
1727 {
1728 	struct sf_buffer *sfb;
1729 	struct aux_buffer *aux;
1730 	unsigned long *new, *tail;
1731 	int i, n_sdbt;
1732 
1733 	if (!nr_pages || !pages)
1734 		return NULL;
1735 
1736 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1737 		pr_err("AUX buffer size (%i pages) is larger than the "
1738 		       "maximum sampling buffer limit\n",
1739 		       nr_pages);
1740 		return NULL;
1741 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1742 		pr_err("AUX buffer size (%i pages) is less than the "
1743 		       "minimum sampling buffer limit\n",
1744 		       nr_pages);
1745 		return NULL;
1746 	}
1747 
1748 	/* Allocate aux_buffer struct for the event */
1749 	aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1750 	if (!aux)
1751 		goto no_aux;
1752 	sfb = &aux->sfb;
1753 
1754 	/* Allocate sdbt_index for fast reference */
1755 	n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE);
1756 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1757 	if (!aux->sdbt_index)
1758 		goto no_sdbt_index;
1759 
1760 	/* Allocate sdb_index for fast reference */
1761 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1762 	if (!aux->sdb_index)
1763 		goto no_sdb_index;
1764 
1765 	/* Allocate the first SDBT */
1766 	sfb->num_sdbt = 0;
1767 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1768 	if (!sfb->sdbt)
1769 		goto no_sdbt;
1770 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1771 	tail = sfb->tail = sfb->sdbt;
1772 
1773 	/*
1774 	 * Link the provided pages of AUX buffer to SDBT.
1775 	 * Allocate SDBT if needed.
1776 	 */
1777 	for (i = 0; i < nr_pages; i++, tail++) {
1778 		if (require_table_link(tail)) {
1779 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1780 			if (!new)
1781 				goto no_sdbt;
1782 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1783 			/* Link current page to tail of chain */
1784 			*tail = (unsigned long)(void *) new + 1;
1785 			tail = new;
1786 		}
1787 		/* Tail is the entry in a SDBT */
1788 		*tail = (unsigned long)pages[i];
1789 		aux->sdb_index[i] = (unsigned long)pages[i];
1790 		aux_sdb_init((unsigned long)pages[i]);
1791 	}
1792 	sfb->num_sdb = nr_pages;
1793 
1794 	/* Link the last entry in the SDBT to the first SDBT */
1795 	*tail = (unsigned long) sfb->sdbt + 1;
1796 	sfb->tail = tail;
1797 
1798 	/*
1799 	 * Initial all SDBs are zeroed. Mark it as empty.
1800 	 * So there is no need to clear the full indicator
1801 	 * when this event is first added.
1802 	 */
1803 	aux->empty_mark = sfb->num_sdb - 1;
1804 
1805 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__,
1806 			    sfb->num_sdbt, sfb->num_sdb);
1807 
1808 	return aux;
1809 
1810 no_sdbt:
1811 	/* SDBs (AUX buffer pages) are freed by caller */
1812 	for (i = 0; i < sfb->num_sdbt; i++)
1813 		free_page(aux->sdbt_index[i]);
1814 	kfree(aux->sdb_index);
1815 no_sdb_index:
1816 	kfree(aux->sdbt_index);
1817 no_sdbt_index:
1818 	kfree(aux);
1819 no_aux:
1820 	return NULL;
1821 }
1822 
1823 static void cpumsf_pmu_read(struct perf_event *event)
1824 {
1825 	/* Nothing to do ... updates are interrupt-driven */
1826 }
1827 
1828 /* Check if the new sampling period/freqeuncy is appropriate.
1829  *
1830  * Return non-zero on error and zero on passed checks.
1831  */
1832 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1833 {
1834 	struct hws_qsi_info_block si;
1835 	unsigned long rate;
1836 	bool do_freq;
1837 
1838 	memset(&si, 0, sizeof(si));
1839 	if (event->cpu == -1) {
1840 		if (qsi(&si))
1841 			return -ENODEV;
1842 	} else {
1843 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1844 		 * sampling structure for accessing the CPU-specific QSI.
1845 		 */
1846 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1847 
1848 		si = cpuhw->qsi;
1849 	}
1850 
1851 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1852 	rate = getrate(do_freq, value, &si);
1853 	if (!rate)
1854 		return -EINVAL;
1855 
1856 	event->attr.sample_period = rate;
1857 	SAMPL_RATE(&event->hw) = rate;
1858 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1859 	debug_sprintf_event(sfdbg, 4, "%s:"
1860 			    " cpu %d value %#llx period %#llx freq %d\n",
1861 			    __func__, event->cpu, value,
1862 			    event->attr.sample_period, do_freq);
1863 	return 0;
1864 }
1865 
1866 /* Activate sampling control.
1867  * Next call of pmu_enable() starts sampling.
1868  */
1869 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1870 {
1871 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1872 
1873 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1874 		return;
1875 
1876 	if (flags & PERF_EF_RELOAD)
1877 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1878 
1879 	perf_pmu_disable(event->pmu);
1880 	event->hw.state = 0;
1881 	cpuhw->lsctl.cs = 1;
1882 	if (SAMPL_DIAG_MODE(&event->hw))
1883 		cpuhw->lsctl.cd = 1;
1884 	perf_pmu_enable(event->pmu);
1885 }
1886 
1887 /* Deactivate sampling control.
1888  * Next call of pmu_enable() stops sampling.
1889  */
1890 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1891 {
1892 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1893 
1894 	if (event->hw.state & PERF_HES_STOPPED)
1895 		return;
1896 
1897 	perf_pmu_disable(event->pmu);
1898 	cpuhw->lsctl.cs = 0;
1899 	cpuhw->lsctl.cd = 0;
1900 	event->hw.state |= PERF_HES_STOPPED;
1901 
1902 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1903 		hw_perf_event_update(event, 1);
1904 		event->hw.state |= PERF_HES_UPTODATE;
1905 	}
1906 	perf_pmu_enable(event->pmu);
1907 }
1908 
1909 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1910 {
1911 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1912 	struct aux_buffer *aux;
1913 	int err;
1914 
1915 	if (cpuhw->flags & PMU_F_IN_USE)
1916 		return -EAGAIN;
1917 
1918 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1919 		return -EINVAL;
1920 
1921 	err = 0;
1922 	perf_pmu_disable(event->pmu);
1923 
1924 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1925 
1926 	/* Set up sampling controls.  Always program the sampling register
1927 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1928 	 * that is used by hw_perf_event_update() to store the sampling buffer
1929 	 * position after samples have been flushed.
1930 	 */
1931 	cpuhw->lsctl.s = 0;
1932 	cpuhw->lsctl.h = 1;
1933 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1934 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1935 		cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1936 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1937 		TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt;
1938 	}
1939 
1940 	/* Ensure sampling functions are in the disabled state.  If disabled,
1941 	 * switch on sampling enable control. */
1942 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1943 		err = -EAGAIN;
1944 		goto out;
1945 	}
1946 	if (SAMPL_DIAG_MODE(&event->hw)) {
1947 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1948 		if (!aux) {
1949 			err = -EINVAL;
1950 			goto out;
1951 		}
1952 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1953 		if (err)
1954 			goto out;
1955 		cpuhw->lsctl.ed = 1;
1956 	}
1957 	cpuhw->lsctl.es = 1;
1958 
1959 	/* Set in_use flag and store event */
1960 	cpuhw->event = event;
1961 	cpuhw->flags |= PMU_F_IN_USE;
1962 
1963 	if (flags & PERF_EF_START)
1964 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1965 out:
1966 	perf_event_update_userpage(event);
1967 	perf_pmu_enable(event->pmu);
1968 	return err;
1969 }
1970 
1971 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1972 {
1973 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1974 
1975 	perf_pmu_disable(event->pmu);
1976 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
1977 
1978 	cpuhw->lsctl.es = 0;
1979 	cpuhw->lsctl.ed = 0;
1980 	cpuhw->flags &= ~PMU_F_IN_USE;
1981 	cpuhw->event = NULL;
1982 
1983 	if (SAMPL_DIAG_MODE(&event->hw))
1984 		aux_output_end(&cpuhw->handle);
1985 	perf_event_update_userpage(event);
1986 	perf_pmu_enable(event->pmu);
1987 }
1988 
1989 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1990 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1991 
1992 /* Attribute list for CPU_SF.
1993  *
1994  * The availablitiy depends on the CPU_MF sampling facility authorization
1995  * for basic + diagnositic samples. This is determined at initialization
1996  * time by the sampling facility device driver.
1997  * If the authorization for basic samples is turned off, it should be
1998  * also turned off for diagnostic sampling.
1999  *
2000  * During initialization of the device driver, check the authorization
2001  * level for diagnostic sampling and installs the attribute
2002  * file for diagnostic sampling if necessary.
2003  *
2004  * For now install a placeholder to reference all possible attributes:
2005  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
2006  * Add another entry for the final NULL pointer.
2007  */
2008 enum {
2009 	SF_CYCLES_BASIC_ATTR_IDX = 0,
2010 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
2011 	SF_CYCLES_ATTR_MAX
2012 };
2013 
2014 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
2015 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
2016 };
2017 
2018 PMU_FORMAT_ATTR(event, "config:0-63");
2019 
2020 static struct attribute *cpumsf_pmu_format_attr[] = {
2021 	&format_attr_event.attr,
2022 	NULL,
2023 };
2024 
2025 static struct attribute_group cpumsf_pmu_events_group = {
2026 	.name = "events",
2027 	.attrs = cpumsf_pmu_events_attr,
2028 };
2029 
2030 static struct attribute_group cpumsf_pmu_format_group = {
2031 	.name = "format",
2032 	.attrs = cpumsf_pmu_format_attr,
2033 };
2034 
2035 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
2036 	&cpumsf_pmu_events_group,
2037 	&cpumsf_pmu_format_group,
2038 	NULL,
2039 };
2040 
2041 static struct pmu cpumf_sampling = {
2042 	.pmu_enable   = cpumsf_pmu_enable,
2043 	.pmu_disable  = cpumsf_pmu_disable,
2044 
2045 	.event_init   = cpumsf_pmu_event_init,
2046 	.add	      = cpumsf_pmu_add,
2047 	.del	      = cpumsf_pmu_del,
2048 
2049 	.start	      = cpumsf_pmu_start,
2050 	.stop	      = cpumsf_pmu_stop,
2051 	.read	      = cpumsf_pmu_read,
2052 
2053 	.attr_groups  = cpumsf_pmu_attr_groups,
2054 
2055 	.setup_aux    = aux_buffer_setup,
2056 	.free_aux     = aux_buffer_free,
2057 
2058 	.check_period = cpumsf_pmu_check_period,
2059 };
2060 
2061 static void cpumf_measurement_alert(struct ext_code ext_code,
2062 				    unsigned int alert, unsigned long unused)
2063 {
2064 	struct cpu_hw_sf *cpuhw;
2065 
2066 	if (!(alert & CPU_MF_INT_SF_MASK))
2067 		return;
2068 	inc_irq_stat(IRQEXT_CMS);
2069 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2070 
2071 	/* Measurement alerts are shared and might happen when the PMU
2072 	 * is not reserved.  Ignore these alerts in this case. */
2073 	if (!(cpuhw->flags & PMU_F_RESERVED))
2074 		return;
2075 
2076 	/* The processing below must take care of multiple alert events that
2077 	 * might be indicated concurrently. */
2078 
2079 	/* Program alert request */
2080 	if (alert & CPU_MF_INT_SF_PRA) {
2081 		if (cpuhw->flags & PMU_F_IN_USE)
2082 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2083 				hw_collect_aux(cpuhw);
2084 			else
2085 				hw_perf_event_update(cpuhw->event, 0);
2086 		else
2087 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2088 	}
2089 
2090 	/* Report measurement alerts only for non-PRA codes */
2091 	if (alert != CPU_MF_INT_SF_PRA)
2092 		debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__,
2093 				    alert);
2094 
2095 	/* Sampling authorization change request */
2096 	if (alert & CPU_MF_INT_SF_SACA)
2097 		qsi(&cpuhw->qsi);
2098 
2099 	/* Loss of sample data due to high-priority machine activities */
2100 	if (alert & CPU_MF_INT_SF_LSDA) {
2101 		pr_err("Sample data was lost\n");
2102 		cpuhw->flags |= PMU_F_ERR_LSDA;
2103 		sf_disable();
2104 	}
2105 
2106 	/* Invalid sampling buffer entry */
2107 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2108 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2109 		       alert);
2110 		cpuhw->flags |= PMU_F_ERR_IBE;
2111 		sf_disable();
2112 	}
2113 }
2114 
2115 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2116 {
2117 	/* Ignore the notification if no events are scheduled on the PMU.
2118 	 * This might be racy...
2119 	 */
2120 	if (!atomic_read(&num_events))
2121 		return 0;
2122 
2123 	local_irq_disable();
2124 	setup_pmc_cpu(&flags);
2125 	local_irq_enable();
2126 	return 0;
2127 }
2128 
2129 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2130 {
2131 	return cpusf_pmu_setup(cpu, PMC_INIT);
2132 }
2133 
2134 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2135 {
2136 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2137 }
2138 
2139 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2140 {
2141 	if (!cpum_sf_avail())
2142 		return -ENODEV;
2143 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2144 }
2145 
2146 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2147 {
2148 	int rc;
2149 	unsigned long min, max;
2150 
2151 	if (!cpum_sf_avail())
2152 		return -ENODEV;
2153 	if (!val || !strlen(val))
2154 		return -EINVAL;
2155 
2156 	/* Valid parameter values: "min,max" or "max" */
2157 	min = CPUM_SF_MIN_SDB;
2158 	max = CPUM_SF_MAX_SDB;
2159 	if (strchr(val, ','))
2160 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2161 	else
2162 		rc = kstrtoul(val, 10, &max);
2163 
2164 	if (min < 2 || min >= max || max > get_num_physpages())
2165 		rc = -EINVAL;
2166 	if (rc)
2167 		return rc;
2168 
2169 	sfb_set_limits(min, max);
2170 	pr_info("The sampling buffer limits have changed to: "
2171 		"min %lu max %lu (diag %lu)\n",
2172 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2173 	return 0;
2174 }
2175 
2176 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2177 static const struct kernel_param_ops param_ops_sfb_size = {
2178 	.set = param_set_sfb_size,
2179 	.get = param_get_sfb_size,
2180 };
2181 
2182 #define RS_INIT_FAILURE_QSI	  0x0001
2183 #define RS_INIT_FAILURE_BSDES	  0x0002
2184 #define RS_INIT_FAILURE_ALRT	  0x0003
2185 #define RS_INIT_FAILURE_PERF	  0x0004
2186 static void __init pr_cpumsf_err(unsigned int reason)
2187 {
2188 	pr_err("Sampling facility support for perf is not available: "
2189 	       "reason %#x\n", reason);
2190 }
2191 
2192 static int __init init_cpum_sampling_pmu(void)
2193 {
2194 	struct hws_qsi_info_block si;
2195 	int err;
2196 
2197 	if (!cpum_sf_avail())
2198 		return -ENODEV;
2199 
2200 	memset(&si, 0, sizeof(si));
2201 	if (qsi(&si)) {
2202 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2203 		return -ENODEV;
2204 	}
2205 
2206 	if (!si.as && !si.ad)
2207 		return -ENODEV;
2208 
2209 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2210 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2211 		return -EINVAL;
2212 	}
2213 
2214 	if (si.ad) {
2215 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2216 		/* Sampling of diagnostic data authorized,
2217 		 * install event into attribute list of PMU device.
2218 		 */
2219 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2220 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2221 	}
2222 
2223 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2224 	if (!sfdbg) {
2225 		pr_err("Registering for s390dbf failed\n");
2226 		return -ENOMEM;
2227 	}
2228 	debug_register_view(sfdbg, &debug_sprintf_view);
2229 
2230 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2231 				    cpumf_measurement_alert);
2232 	if (err) {
2233 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2234 		debug_unregister(sfdbg);
2235 		goto out;
2236 	}
2237 
2238 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2239 	if (err) {
2240 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2241 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2242 					cpumf_measurement_alert);
2243 		debug_unregister(sfdbg);
2244 		goto out;
2245 	}
2246 
2247 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2248 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2249 out:
2250 	return err;
2251 }
2252 
2253 arch_initcall(init_cpum_sampling_pmu);
2254 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644);
2255