xref: /linux/arch/s390/kernel/perf_cpum_sf.c (revision 906fd46a65383cd639e5eec72a047efc33045d86)
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 #include <linux/io.h>
26 
27 /* Minimum number of sample-data-block-tables:
28  * At least one table is required for the sampling buffer structure.
29  * A single table contains up to 511 pointers to sample-data-blocks.
30  */
31 #define CPUM_SF_MIN_SDBT	1
32 
33 /* Number of sample-data-blocks per sample-data-block-table (SDBT):
34  * A table contains SDB pointers (8 bytes) and one table-link entry
35  * that points to the origin of the next SDBT.
36  */
37 #define CPUM_SF_SDB_PER_TABLE	((PAGE_SIZE - 8) / 8)
38 
39 /* Maximum page offset for an SDBT table-link entry:
40  * If this page offset is reached, a table-link entry to the next SDBT
41  * must be added.
42  */
43 #define CPUM_SF_SDBT_TL_OFFSET	(CPUM_SF_SDB_PER_TABLE * 8)
44 static inline int require_table_link(const void *sdbt)
45 {
46 	return ((unsigned long)sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
47 }
48 
49 /* Minimum and maximum sampling buffer sizes:
50  *
51  * This number represents the maximum size of the sampling buffer taking
52  * the number of sample-data-block-tables into account.  Note that these
53  * numbers apply to the basic-sampling function only.
54  * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if
55  * the diagnostic-sampling function is active.
56  *
57  * Sampling buffer size		Buffer characteristics
58  * ---------------------------------------------------
59  *	 64KB		    ==	  16 pages (4KB per page)
60  *				   1 page  for SDB-tables
61  *				  15 pages for SDBs
62  *
63  *  32MB		    ==	8192 pages (4KB per page)
64  *				  16 pages for SDB-tables
65  *				8176 pages for SDBs
66  */
67 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15;
68 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176;
69 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
70 
71 struct sf_buffer {
72 	unsigned long	 *sdbt;	    /* Sample-data-block-table origin */
73 	/* buffer characteristics (required for buffer increments) */
74 	unsigned long  num_sdb;	    /* Number of sample-data-blocks */
75 	unsigned long num_sdbt;	    /* Number of sample-data-block-tables */
76 	unsigned long	 *tail;	    /* last sample-data-block-table */
77 };
78 
79 struct aux_buffer {
80 	struct sf_buffer sfb;
81 	unsigned long head;	   /* index of SDB of buffer head */
82 	unsigned long alert_mark;  /* index of SDB of alert request position */
83 	unsigned long empty_mark;  /* mark of SDB not marked full */
84 	unsigned long *sdb_index;  /* SDB address for fast lookup */
85 	unsigned long *sdbt_index; /* SDBT address for fast lookup */
86 };
87 
88 struct cpu_hw_sf {
89 	/* CPU-measurement sampling information block */
90 	struct hws_qsi_info_block qsi;
91 	/* CPU-measurement sampling control block */
92 	struct hws_lsctl_request_block lsctl;
93 	struct sf_buffer sfb;	    /* Sampling buffer */
94 	unsigned int flags;	    /* Status flags */
95 	struct perf_event *event;   /* Scheduled perf event */
96 	struct perf_output_handle handle; /* AUX buffer output handle */
97 };
98 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);
99 
100 /* Debug feature */
101 static debug_info_t *sfdbg;
102 
103 /* Sampling control helper functions */
104 static inline unsigned long freq_to_sample_rate(struct hws_qsi_info_block *qsi,
105 						unsigned long freq)
106 {
107 	return (USEC_PER_SEC / freq) * qsi->cpu_speed;
108 }
109 
110 static inline unsigned long sample_rate_to_freq(struct hws_qsi_info_block *qsi,
111 						unsigned long rate)
112 {
113 	return USEC_PER_SEC * qsi->cpu_speed / rate;
114 }
115 
116 /* Return TOD timestamp contained in an trailer entry */
117 static inline unsigned long long trailer_timestamp(struct hws_trailer_entry *te)
118 {
119 	/* TOD in STCKE format */
120 	if (te->header.t)
121 		return *((unsigned long long *)&te->timestamp[1]);
122 
123 	/* TOD in STCK format */
124 	return *((unsigned long long *)&te->timestamp[0]);
125 }
126 
127 /* Return pointer to trailer entry of an sample data block */
128 static inline struct hws_trailer_entry *trailer_entry_ptr(unsigned long v)
129 {
130 	void *ret;
131 
132 	ret = (void *)v;
133 	ret += PAGE_SIZE;
134 	ret -= sizeof(struct hws_trailer_entry);
135 
136 	return ret;
137 }
138 
139 /*
140  * Return true if the entry in the sample data block table (sdbt)
141  * is a link to the next sdbt
142  */
143 static inline int is_link_entry(unsigned long *s)
144 {
145 	return *s & 0x1UL ? 1 : 0;
146 }
147 
148 /* Return pointer to the linked sdbt */
149 static inline unsigned long *get_next_sdbt(unsigned long *s)
150 {
151 	return phys_to_virt(*s & ~0x1UL);
152 }
153 
154 /*
155  * sf_disable() - Switch off sampling facility
156  */
157 static int sf_disable(void)
158 {
159 	struct hws_lsctl_request_block sreq;
160 
161 	memset(&sreq, 0, sizeof(sreq));
162 	return lsctl(&sreq);
163 }
164 
165 /*
166  * sf_buffer_available() - Check for an allocated sampling buffer
167  */
168 static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
169 {
170 	return !!cpuhw->sfb.sdbt;
171 }
172 
173 /*
174  * deallocate sampling facility buffer
175  */
176 static void free_sampling_buffer(struct sf_buffer *sfb)
177 {
178 	unsigned long *sdbt, *curr;
179 
180 	if (!sfb->sdbt)
181 		return;
182 
183 	sdbt = sfb->sdbt;
184 	curr = sdbt;
185 
186 	/* Free the SDBT after all SDBs are processed... */
187 	while (1) {
188 		if (!*curr || !sdbt)
189 			break;
190 
191 		/* Process table-link entries */
192 		if (is_link_entry(curr)) {
193 			curr = get_next_sdbt(curr);
194 			if (sdbt)
195 				free_page((unsigned long)sdbt);
196 
197 			/* If the origin is reached, sampling buffer is freed */
198 			if (curr == sfb->sdbt)
199 				break;
200 			else
201 				sdbt = curr;
202 		} else {
203 			/* Process SDB pointer */
204 			if (*curr) {
205 				free_page((unsigned long)phys_to_virt(*curr));
206 				curr++;
207 			}
208 		}
209 	}
210 
211 	debug_sprintf_event(sfdbg, 5, "%s: freed sdbt %#lx\n", __func__,
212 			    (unsigned long)sfb->sdbt);
213 	memset(sfb, 0, sizeof(*sfb));
214 }
215 
216 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
217 {
218 	struct hws_trailer_entry *te;
219 	unsigned long sdb;
220 
221 	/* Allocate and initialize sample-data-block */
222 	sdb = get_zeroed_page(gfp_flags);
223 	if (!sdb)
224 		return -ENOMEM;
225 	te = trailer_entry_ptr(sdb);
226 	te->header.a = 1;
227 
228 	/* Link SDB into the sample-data-block-table */
229 	*sdbt = virt_to_phys((void *)sdb);
230 
231 	return 0;
232 }
233 
234 /*
235  * realloc_sampling_buffer() - extend sampler memory
236  *
237  * Allocates new sample-data-blocks and adds them to the specified sampling
238  * buffer memory.
239  *
240  * Important: This modifies the sampling buffer and must be called when the
241  *	      sampling facility is disabled.
242  *
243  * Returns zero on success, non-zero otherwise.
244  */
245 static int realloc_sampling_buffer(struct sf_buffer *sfb,
246 				   unsigned long num_sdb, gfp_t gfp_flags)
247 {
248 	int i, rc;
249 	unsigned long *new, *tail, *tail_prev = NULL;
250 
251 	if (!sfb->sdbt || !sfb->tail)
252 		return -EINVAL;
253 
254 	if (!is_link_entry(sfb->tail))
255 		return -EINVAL;
256 
257 	/* Append to the existing sampling buffer, overwriting the table-link
258 	 * register.
259 	 * The tail variables always points to the "tail" (last and table-link)
260 	 * entry in an SDB-table.
261 	 */
262 	tail = sfb->tail;
263 
264 	/* Do a sanity check whether the table-link entry points to
265 	 * the sampling buffer origin.
266 	 */
267 	if (sfb->sdbt != get_next_sdbt(tail)) {
268 		debug_sprintf_event(sfdbg, 3, "%s: "
269 				    "sampling buffer is not linked: origin %#lx"
270 				    " tail %#lx\n", __func__,
271 				    (unsigned long)sfb->sdbt,
272 				    (unsigned long)tail);
273 		return -EINVAL;
274 	}
275 
276 	/* Allocate remaining SDBs */
277 	rc = 0;
278 	for (i = 0; i < num_sdb; i++) {
279 		/* Allocate a new SDB-table if it is full. */
280 		if (require_table_link(tail)) {
281 			new = (unsigned long *)get_zeroed_page(gfp_flags);
282 			if (!new) {
283 				rc = -ENOMEM;
284 				break;
285 			}
286 			sfb->num_sdbt++;
287 			/* Link current page to tail of chain */
288 			*tail = virt_to_phys((void *)new) + 1;
289 			tail_prev = tail;
290 			tail = new;
291 		}
292 
293 		/* Allocate a new sample-data-block.
294 		 * If there is not enough memory, stop the realloc process
295 		 * and simply use what was allocated.  If this is a temporary
296 		 * issue, a new realloc call (if required) might succeed.
297 		 */
298 		rc = alloc_sample_data_block(tail, gfp_flags);
299 		if (rc) {
300 			/* Undo last SDBT. An SDBT with no SDB at its first
301 			 * entry but with an SDBT entry instead can not be
302 			 * handled by the interrupt handler code.
303 			 * Avoid this situation.
304 			 */
305 			if (tail_prev) {
306 				sfb->num_sdbt--;
307 				free_page((unsigned long)new);
308 				tail = tail_prev;
309 			}
310 			break;
311 		}
312 		sfb->num_sdb++;
313 		tail++;
314 		tail_prev = new = NULL;	/* Allocated at least one SBD */
315 	}
316 
317 	/* Link sampling buffer to its origin */
318 	*tail = virt_to_phys(sfb->sdbt) + 1;
319 	sfb->tail = tail;
320 
321 	debug_sprintf_event(sfdbg, 4, "%s: new buffer"
322 			    " settings: sdbt %lu sdb %lu\n", __func__,
323 			    sfb->num_sdbt, sfb->num_sdb);
324 	return rc;
325 }
326 
327 /*
328  * allocate_sampling_buffer() - allocate sampler memory
329  *
330  * Allocates and initializes a sampling buffer structure using the
331  * specified number of sample-data-blocks (SDB).  For each allocation,
332  * a 4K page is used.  The number of sample-data-block-tables (SDBT)
333  * are calculated from SDBs.
334  * Also set the ALERT_REQ mask in each SDBs trailer.
335  *
336  * Returns zero on success, non-zero otherwise.
337  */
338 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
339 {
340 	int rc;
341 
342 	if (sfb->sdbt)
343 		return -EINVAL;
344 
345 	/* Allocate the sample-data-block-table origin */
346 	sfb->sdbt = (unsigned long *)get_zeroed_page(GFP_KERNEL);
347 	if (!sfb->sdbt)
348 		return -ENOMEM;
349 	sfb->num_sdb = 0;
350 	sfb->num_sdbt = 1;
351 
352 	/* Link the table origin to point to itself to prepare for
353 	 * realloc_sampling_buffer() invocation.
354 	 */
355 	sfb->tail = sfb->sdbt;
356 	*sfb->tail = virt_to_phys((void *)sfb->sdbt) + 1;
357 
358 	/* Allocate requested number of sample-data-blocks */
359 	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
360 	if (rc) {
361 		free_sampling_buffer(sfb);
362 		debug_sprintf_event(sfdbg, 4, "%s: "
363 			"realloc_sampling_buffer failed with rc %i\n",
364 			__func__, rc);
365 	} else
366 		debug_sprintf_event(sfdbg, 4,
367 			"%s: tear %#lx dear %#lx\n", __func__,
368 			(unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt);
369 	return rc;
370 }
371 
372 static void sfb_set_limits(unsigned long min, unsigned long max)
373 {
374 	struct hws_qsi_info_block si;
375 
376 	CPUM_SF_MIN_SDB = min;
377 	CPUM_SF_MAX_SDB = max;
378 
379 	memset(&si, 0, sizeof(si));
380 	if (!qsi(&si))
381 		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
382 }
383 
384 static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
385 {
386 	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
387 				    : CPUM_SF_MAX_SDB;
388 }
389 
390 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
391 					struct hw_perf_event *hwc)
392 {
393 	if (!sfb->sdbt)
394 		return SFB_ALLOC_REG(hwc);
395 	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
396 		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
397 	return 0;
398 }
399 
400 static int sfb_has_pending_allocs(struct sf_buffer *sfb,
401 				   struct hw_perf_event *hwc)
402 {
403 	return sfb_pending_allocs(sfb, hwc) > 0;
404 }
405 
406 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
407 {
408 	/* Limit the number of SDBs to not exceed the maximum */
409 	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
410 	if (num)
411 		SFB_ALLOC_REG(hwc) += num;
412 }
413 
414 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
415 {
416 	SFB_ALLOC_REG(hwc) = 0;
417 	sfb_account_allocs(num, hwc);
418 }
419 
420 static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
421 {
422 	if (cpuhw->sfb.sdbt)
423 		free_sampling_buffer(&cpuhw->sfb);
424 }
425 
426 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
427 {
428 	unsigned long n_sdb, freq;
429 	size_t sample_size;
430 
431 	/* Calculate sampling buffers using 4K pages
432 	 *
433 	 *    1. The sampling size is 32 bytes for basic sampling. This size
434 	 *	 is the same for all machine types. Diagnostic
435 	 *	 sampling uses auxlilary data buffer setup which provides the
436 	 *	 memory for SDBs using linux common code auxiliary trace
437 	 *	 setup.
438 	 *
439 	 *    2. Function alloc_sampling_buffer() sets the Alert Request
440 	 *	 Control indicator to trigger a measurement-alert to harvest
441 	 *	 sample-data-blocks (SDB). This is done per SDB. This
442 	 *	 measurement alert interrupt fires quick enough to handle
443 	 *	 one SDB, on very high frequency and work loads there might
444 	 *	 be 2 to 3 SBDs available for sample processing.
445 	 *	 Currently there is no need for setup alert request on every
446 	 *	 n-th page. This is counterproductive as one IRQ triggers
447 	 *	 a very high number of samples to be processed at one IRQ.
448 	 *
449 	 *    3. Use the sampling frequency as input.
450 	 *	 Compute the number of SDBs and ensure a minimum
451 	 *	 of CPUM_SF_MIN_SDB.  Depending on frequency add some more
452 	 *	 SDBs to handle a higher sampling rate.
453 	 *	 Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples
454 	 *	 (one SDB) for every 10000 HZ frequency increment.
455 	 *
456 	 *    4. Compute the number of sample-data-block-tables (SDBT) and
457 	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
458 	 *	 to 511 SDBs).
459 	 */
460 	sample_size = sizeof(struct hws_basic_entry);
461 	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
462 	n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000);
463 
464 	/* If there is already a sampling buffer allocated, it is very likely
465 	 * that the sampling facility is enabled too.  If the event to be
466 	 * initialized requires a greater sampling buffer, the allocation must
467 	 * be postponed.  Changing the sampling buffer requires the sampling
468 	 * facility to be in the disabled state.  So, account the number of
469 	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
470 	 * before the event is started.
471 	 */
472 	sfb_init_allocs(n_sdb, hwc);
473 	if (sf_buffer_available(cpuhw))
474 		return 0;
475 
476 	debug_sprintf_event(sfdbg, 3,
477 			    "%s: rate %lu f %lu sdb %lu/%lu"
478 			    " sample_size %lu cpuhw %p\n", __func__,
479 			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
480 			    sample_size, cpuhw);
481 
482 	return alloc_sampling_buffer(&cpuhw->sfb,
483 				     sfb_pending_allocs(&cpuhw->sfb, hwc));
484 }
485 
486 static unsigned long min_percent(unsigned int percent, unsigned long base,
487 				 unsigned long min)
488 {
489 	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
490 }
491 
492 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
493 {
494 	/* Use a percentage-based approach to extend the sampling facility
495 	 * buffer.  Accept up to 5% sample data loss.
496 	 * Vary the extents between 1% to 5% of the current number of
497 	 * sample-data-blocks.
498 	 */
499 	if (ratio <= 5)
500 		return 0;
501 	if (ratio <= 25)
502 		return min_percent(1, base, 1);
503 	if (ratio <= 50)
504 		return min_percent(1, base, 1);
505 	if (ratio <= 75)
506 		return min_percent(2, base, 2);
507 	if (ratio <= 100)
508 		return min_percent(3, base, 3);
509 	if (ratio <= 250)
510 		return min_percent(4, base, 4);
511 
512 	return min_percent(5, base, 8);
513 }
514 
515 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
516 				  struct hw_perf_event *hwc)
517 {
518 	unsigned long ratio, num;
519 
520 	if (!OVERFLOW_REG(hwc))
521 		return;
522 
523 	/* The sample_overflow contains the average number of sample data
524 	 * that has been lost because sample-data-blocks were full.
525 	 *
526 	 * Calculate the total number of sample data entries that has been
527 	 * discarded.  Then calculate the ratio of lost samples to total samples
528 	 * per second in percent.
529 	 */
530 	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
531 			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));
532 
533 	/* Compute number of sample-data-blocks */
534 	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
535 	if (num)
536 		sfb_account_allocs(num, hwc);
537 
538 	debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n",
539 			    __func__, OVERFLOW_REG(hwc), ratio, num);
540 	OVERFLOW_REG(hwc) = 0;
541 }
542 
543 /* extend_sampling_buffer() - Extend sampling buffer
544  * @sfb:	Sampling buffer structure (for local CPU)
545  * @hwc:	Perf event hardware structure
546  *
547  * Use this function to extend the sampling buffer based on the overflow counter
548  * and postponed allocation extents stored in the specified Perf event hardware.
549  *
550  * Important: This function disables the sampling facility in order to safely
551  *	      change the sampling buffer structure.  Do not call this function
552  *	      when the PMU is active.
553  */
554 static void extend_sampling_buffer(struct sf_buffer *sfb,
555 				   struct hw_perf_event *hwc)
556 {
557 	unsigned long num, num_old;
558 	int rc;
559 
560 	num = sfb_pending_allocs(sfb, hwc);
561 	if (!num)
562 		return;
563 	num_old = sfb->num_sdb;
564 
565 	/* Disable the sampling facility to reset any states and also
566 	 * clear pending measurement alerts.
567 	 */
568 	sf_disable();
569 
570 	/* Extend the sampling buffer.
571 	 * This memory allocation typically happens in an atomic context when
572 	 * called by perf.  Because this is a reallocation, it is fine if the
573 	 * new SDB-request cannot be satisfied immediately.
574 	 */
575 	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
576 	if (rc)
577 		debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n",
578 				    __func__, rc);
579 
580 	if (sfb_has_pending_allocs(sfb, hwc))
581 		debug_sprintf_event(sfdbg, 5, "%s: "
582 				    "req %lu alloc %lu remaining %lu\n",
583 				    __func__, num, sfb->num_sdb - num_old,
584 				    sfb_pending_allocs(sfb, hwc));
585 }
586 
587 /* Number of perf events counting hardware events */
588 static atomic_t num_events;
589 /* Used to avoid races in calling reserve/release_cpumf_hardware */
590 static DEFINE_MUTEX(pmc_reserve_mutex);
591 
592 #define PMC_INIT      0
593 #define PMC_RELEASE   1
594 #define PMC_FAILURE   2
595 static void setup_pmc_cpu(void *flags)
596 {
597 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
598 	int err = 0;
599 
600 	switch (*((int *)flags)) {
601 	case PMC_INIT:
602 		memset(cpusf, 0, sizeof(*cpusf));
603 		err = qsi(&cpusf->qsi);
604 		if (err)
605 			break;
606 		cpusf->flags |= PMU_F_RESERVED;
607 		err = sf_disable();
608 		break;
609 	case PMC_RELEASE:
610 		cpusf->flags &= ~PMU_F_RESERVED;
611 		err = sf_disable();
612 		if (!err)
613 			deallocate_buffers(cpusf);
614 		break;
615 	}
616 	if (err) {
617 		*((int *)flags) |= PMC_FAILURE;
618 		pr_err("Switching off the sampling facility failed with rc %i\n", err);
619 	}
620 }
621 
622 static void release_pmc_hardware(void)
623 {
624 	int flags = PMC_RELEASE;
625 
626 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
627 	on_each_cpu(setup_pmc_cpu, &flags, 1);
628 }
629 
630 static int reserve_pmc_hardware(void)
631 {
632 	int flags = PMC_INIT;
633 
634 	on_each_cpu(setup_pmc_cpu, &flags, 1);
635 	if (flags & PMC_FAILURE) {
636 		release_pmc_hardware();
637 		return -ENODEV;
638 	}
639 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
640 
641 	return 0;
642 }
643 
644 static void hw_perf_event_destroy(struct perf_event *event)
645 {
646 	/* Release PMC if this is the last perf event */
647 	if (!atomic_add_unless(&num_events, -1, 1)) {
648 		mutex_lock(&pmc_reserve_mutex);
649 		if (atomic_dec_return(&num_events) == 0)
650 			release_pmc_hardware();
651 		mutex_unlock(&pmc_reserve_mutex);
652 	}
653 }
654 
655 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
656 {
657 	hwc->sample_period = period;
658 	hwc->last_period = hwc->sample_period;
659 	local64_set(&hwc->period_left, hwc->sample_period);
660 }
661 
662 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
663 				   unsigned long rate)
664 {
665 	return clamp_t(unsigned long, rate,
666 		       si->min_sampl_rate, si->max_sampl_rate);
667 }
668 
669 static u32 cpumsf_pid_type(struct perf_event *event,
670 			   u32 pid, enum pid_type type)
671 {
672 	struct task_struct *tsk;
673 
674 	/* Idle process */
675 	if (!pid)
676 		goto out;
677 
678 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
679 	pid = -1;
680 	if (tsk) {
681 		/*
682 		 * Only top level events contain the pid namespace in which
683 		 * they are created.
684 		 */
685 		if (event->parent)
686 			event = event->parent;
687 		pid = __task_pid_nr_ns(tsk, type, event->ns);
688 		/*
689 		 * See also 1d953111b648
690 		 * "perf/core: Don't report zero PIDs for exiting tasks".
691 		 */
692 		if (!pid && !pid_alive(tsk))
693 			pid = -1;
694 	}
695 out:
696 	return pid;
697 }
698 
699 static void cpumsf_output_event_pid(struct perf_event *event,
700 				    struct perf_sample_data *data,
701 				    struct pt_regs *regs)
702 {
703 	u32 pid;
704 	struct perf_event_header header;
705 	struct perf_output_handle handle;
706 
707 	/*
708 	 * Obtain the PID from the basic-sampling data entry and
709 	 * correct the data->tid_entry.pid value.
710 	 */
711 	pid = data->tid_entry.pid;
712 
713 	/* Protect callchain buffers, tasks */
714 	rcu_read_lock();
715 
716 	perf_prepare_sample(data, event, regs);
717 	perf_prepare_header(&header, data, event, regs);
718 	if (perf_output_begin(&handle, data, event, header.size))
719 		goto out;
720 
721 	/* Update the process ID (see also kernel/events/core.c) */
722 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
723 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
724 
725 	perf_output_sample(&handle, &header, data, event);
726 	perf_output_end(&handle);
727 out:
728 	rcu_read_unlock();
729 }
730 
731 static unsigned long getrate(bool freq, unsigned long sample,
732 			     struct hws_qsi_info_block *si)
733 {
734 	unsigned long rate;
735 
736 	if (freq) {
737 		rate = freq_to_sample_rate(si, sample);
738 		rate = hw_limit_rate(si, rate);
739 	} else {
740 		/* The min/max sampling rates specifies the valid range
741 		 * of sample periods.  If the specified sample period is
742 		 * out of range, limit the period to the range boundary.
743 		 */
744 		rate = hw_limit_rate(si, sample);
745 
746 		/* The perf core maintains a maximum sample rate that is
747 		 * configurable through the sysctl interface.  Ensure the
748 		 * sampling rate does not exceed this value.  This also helps
749 		 * to avoid throttling when pushing samples with
750 		 * perf_event_overflow().
751 		 */
752 		if (sample_rate_to_freq(si, rate) >
753 		    sysctl_perf_event_sample_rate) {
754 			debug_sprintf_event(sfdbg, 1, "%s: "
755 					    "Sampling rate exceeds maximum "
756 					    "perf sample rate\n", __func__);
757 			rate = 0;
758 		}
759 	}
760 	return rate;
761 }
762 
763 /* The sampling information (si) contains information about the
764  * min/max sampling intervals and the CPU speed.  So calculate the
765  * correct sampling interval and avoid the whole period adjust
766  * feedback loop.
767  *
768  * Since the CPU Measurement sampling facility can not handle frequency
769  * calculate the sampling interval when frequency is specified using
770  * this formula:
771  *	interval := cpu_speed * 1000000 / sample_freq
772  *
773  * Returns errno on bad input and zero on success with parameter interval
774  * set to the correct sampling rate.
775  *
776  * Note: This function turns off freq bit to avoid calling function
777  * perf_adjust_period(). This causes frequency adjustment in the common
778  * code part which causes tremendous variations in the counter values.
779  */
780 static int __hw_perf_event_init_rate(struct perf_event *event,
781 				     struct hws_qsi_info_block *si)
782 {
783 	struct perf_event_attr *attr = &event->attr;
784 	struct hw_perf_event *hwc = &event->hw;
785 	unsigned long rate;
786 
787 	if (attr->freq) {
788 		if (!attr->sample_freq)
789 			return -EINVAL;
790 		rate = getrate(attr->freq, attr->sample_freq, si);
791 		attr->freq = 0;		/* Don't call  perf_adjust_period() */
792 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
793 	} else {
794 		rate = getrate(attr->freq, attr->sample_period, si);
795 		if (!rate)
796 			return -EINVAL;
797 	}
798 	attr->sample_period = rate;
799 	SAMPL_RATE(hwc) = rate;
800 	hw_init_period(hwc, SAMPL_RATE(hwc));
801 	debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n",
802 			    __func__, event->cpu, event->attr.sample_period,
803 			    event->attr.freq, SAMPLE_FREQ_MODE(hwc));
804 	return 0;
805 }
806 
807 static int __hw_perf_event_init(struct perf_event *event)
808 {
809 	struct cpu_hw_sf *cpuhw;
810 	struct hws_qsi_info_block si;
811 	struct perf_event_attr *attr = &event->attr;
812 	struct hw_perf_event *hwc = &event->hw;
813 	int cpu, err;
814 
815 	/* Reserve CPU-measurement sampling facility */
816 	err = 0;
817 	if (!atomic_inc_not_zero(&num_events)) {
818 		mutex_lock(&pmc_reserve_mutex);
819 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
820 			err = -EBUSY;
821 		else
822 			atomic_inc(&num_events);
823 		mutex_unlock(&pmc_reserve_mutex);
824 	}
825 	event->destroy = hw_perf_event_destroy;
826 
827 	if (err)
828 		goto out;
829 
830 	/* Access per-CPU sampling information (query sampling info) */
831 	/*
832 	 * The event->cpu value can be -1 to count on every CPU, for example,
833 	 * when attaching to a task.  If this is specified, use the query
834 	 * sampling info from the current CPU, otherwise use event->cpu to
835 	 * retrieve the per-CPU information.
836 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
837 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
838 	 */
839 	memset(&si, 0, sizeof(si));
840 	cpuhw = NULL;
841 	if (event->cpu == -1)
842 		qsi(&si);
843 	else {
844 		/* Event is pinned to a particular CPU, retrieve the per-CPU
845 		 * sampling structure for accessing the CPU-specific QSI.
846 		 */
847 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
848 		si = cpuhw->qsi;
849 	}
850 
851 	/* Check sampling facility authorization and, if not authorized,
852 	 * fall back to other PMUs.  It is safe to check any CPU because
853 	 * the authorization is identical for all configured CPUs.
854 	 */
855 	if (!si.as) {
856 		err = -ENOENT;
857 		goto out;
858 	}
859 
860 	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
861 		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
862 		err = -EBUSY;
863 		goto out;
864 	}
865 
866 	/* Always enable basic sampling */
867 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
868 
869 	/* Check if diagnostic sampling is requested.  Deny if the required
870 	 * sampling authorization is missing.
871 	 */
872 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
873 		if (!si.ad) {
874 			err = -EPERM;
875 			goto out;
876 		}
877 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
878 	}
879 
880 	err =  __hw_perf_event_init_rate(event, &si);
881 	if (err)
882 		goto out;
883 
884 	/* Initialize sample data overflow accounting */
885 	hwc->extra_reg.reg = REG_OVERFLOW;
886 	OVERFLOW_REG(hwc) = 0;
887 
888 	/* Use AUX buffer. No need to allocate it by ourself */
889 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
890 		return 0;
891 
892 	/* Allocate the per-CPU sampling buffer using the CPU information
893 	 * from the event.  If the event is not pinned to a particular
894 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
895 	 * buffers for each online CPU.
896 	 */
897 	if (cpuhw)
898 		/* Event is pinned to a particular CPU */
899 		err = allocate_buffers(cpuhw, hwc);
900 	else {
901 		/* Event is not pinned, allocate sampling buffer on
902 		 * each online CPU
903 		 */
904 		for_each_online_cpu(cpu) {
905 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
906 			err = allocate_buffers(cpuhw, hwc);
907 			if (err)
908 				break;
909 		}
910 	}
911 
912 	/* If PID/TID sampling is active, replace the default overflow
913 	 * handler to extract and resolve the PIDs from the basic-sampling
914 	 * data entries.
915 	 */
916 	if (event->attr.sample_type & PERF_SAMPLE_TID)
917 		if (is_default_overflow_handler(event))
918 			event->overflow_handler = cpumsf_output_event_pid;
919 out:
920 	return err;
921 }
922 
923 static bool is_callchain_event(struct perf_event *event)
924 {
925 	u64 sample_type = event->attr.sample_type;
926 
927 	return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER |
928 			      PERF_SAMPLE_STACK_USER);
929 }
930 
931 static int cpumsf_pmu_event_init(struct perf_event *event)
932 {
933 	int err;
934 
935 	/* No support for taken branch sampling */
936 	/* No support for callchain, stacks and registers */
937 	if (has_branch_stack(event) || is_callchain_event(event))
938 		return -EOPNOTSUPP;
939 
940 	switch (event->attr.type) {
941 	case PERF_TYPE_RAW:
942 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
943 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
944 			return -ENOENT;
945 		break;
946 	case PERF_TYPE_HARDWARE:
947 		/* Support sampling of CPU cycles in addition to the
948 		 * counter facility.  However, the counter facility
949 		 * is more precise and, hence, restrict this PMU to
950 		 * sampling events only.
951 		 */
952 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
953 			return -ENOENT;
954 		if (!is_sampling_event(event))
955 			return -ENOENT;
956 		break;
957 	default:
958 		return -ENOENT;
959 	}
960 
961 	/* Force reset of idle/hv excludes regardless of what the
962 	 * user requested.
963 	 */
964 	if (event->attr.exclude_hv)
965 		event->attr.exclude_hv = 0;
966 	if (event->attr.exclude_idle)
967 		event->attr.exclude_idle = 0;
968 
969 	err = __hw_perf_event_init(event);
970 	if (unlikely(err))
971 		if (event->destroy)
972 			event->destroy(event);
973 	return err;
974 }
975 
976 static void cpumsf_pmu_enable(struct pmu *pmu)
977 {
978 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
979 	struct hw_perf_event *hwc;
980 	int err;
981 
982 	if (cpuhw->flags & PMU_F_ENABLED)
983 		return;
984 
985 	if (cpuhw->flags & PMU_F_ERR_MASK)
986 		return;
987 
988 	/* Check whether to extent the sampling buffer.
989 	 *
990 	 * Two conditions trigger an increase of the sampling buffer for a
991 	 * perf event:
992 	 *    1. Postponed buffer allocations from the event initialization.
993 	 *    2. Sampling overflows that contribute to pending allocations.
994 	 *
995 	 * Note that the extend_sampling_buffer() function disables the sampling
996 	 * facility, but it can be fully re-enabled using sampling controls that
997 	 * have been saved in cpumsf_pmu_disable().
998 	 */
999 	if (cpuhw->event) {
1000 		hwc = &cpuhw->event->hw;
1001 		if (!(SAMPL_DIAG_MODE(hwc))) {
1002 			/*
1003 			 * Account number of overflow-designated
1004 			 * buffer extents
1005 			 */
1006 			sfb_account_overflows(cpuhw, hwc);
1007 			extend_sampling_buffer(&cpuhw->sfb, hwc);
1008 		}
1009 		/* Rate may be adjusted with ioctl() */
1010 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
1011 	}
1012 
1013 	/* (Re)enable the PMU and sampling facility */
1014 	cpuhw->flags |= PMU_F_ENABLED;
1015 	barrier();
1016 
1017 	err = lsctl(&cpuhw->lsctl);
1018 	if (err) {
1019 		cpuhw->flags &= ~PMU_F_ENABLED;
1020 		pr_err("Loading sampling controls failed: op 1 err %i\n", err);
1021 		return;
1022 	}
1023 
1024 	/* Load current program parameter */
1025 	lpp(&S390_lowcore.lpp);
1026 
1027 	debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i "
1028 			    "interval %#lx tear %#lx dear %#lx\n", __func__,
1029 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
1030 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
1031 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
1032 }
1033 
1034 static void cpumsf_pmu_disable(struct pmu *pmu)
1035 {
1036 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1037 	struct hws_lsctl_request_block inactive;
1038 	struct hws_qsi_info_block si;
1039 	int err;
1040 
1041 	if (!(cpuhw->flags & PMU_F_ENABLED))
1042 		return;
1043 
1044 	if (cpuhw->flags & PMU_F_ERR_MASK)
1045 		return;
1046 
1047 	/* Switch off sampling activation control */
1048 	inactive = cpuhw->lsctl;
1049 	inactive.cs = 0;
1050 	inactive.cd = 0;
1051 
1052 	err = lsctl(&inactive);
1053 	if (err) {
1054 		pr_err("Loading sampling controls failed: op 2 err %i\n", err);
1055 		return;
1056 	}
1057 
1058 	/* Save state of TEAR and DEAR register contents */
1059 	err = qsi(&si);
1060 	if (!err) {
1061 		/* TEAR/DEAR values are valid only if the sampling facility is
1062 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1063 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1064 		 * controls the enable/disable state.
1065 		 */
1066 		if (si.es) {
1067 			cpuhw->lsctl.tear = si.tear;
1068 			cpuhw->lsctl.dear = si.dear;
1069 		}
1070 	} else
1071 		debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n",
1072 				    __func__, err);
1073 
1074 	cpuhw->flags &= ~PMU_F_ENABLED;
1075 }
1076 
1077 /* perf_exclude_event() - Filter event
1078  * @event:	The perf event
1079  * @regs:	pt_regs structure
1080  * @sde_regs:	Sample-data-entry (sde) regs structure
1081  *
1082  * Filter perf events according to their exclude specification.
1083  *
1084  * Return non-zero if the event shall be excluded.
1085  */
1086 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1087 			      struct perf_sf_sde_regs *sde_regs)
1088 {
1089 	if (event->attr.exclude_user && user_mode(regs))
1090 		return 1;
1091 	if (event->attr.exclude_kernel && !user_mode(regs))
1092 		return 1;
1093 	if (event->attr.exclude_guest && sde_regs->in_guest)
1094 		return 1;
1095 	if (event->attr.exclude_host && !sde_regs->in_guest)
1096 		return 1;
1097 	return 0;
1098 }
1099 
1100 /* perf_push_sample() - Push samples to perf
1101  * @event:	The perf event
1102  * @sample:	Hardware sample data
1103  *
1104  * Use the hardware sample data to create perf event sample.  The sample
1105  * is the pushed to the event subsystem and the function checks for
1106  * possible event overflows.  If an event overflow occurs, the PMU is
1107  * stopped.
1108  *
1109  * Return non-zero if an event overflow occurred.
1110  */
1111 static int perf_push_sample(struct perf_event *event,
1112 			    struct hws_basic_entry *basic)
1113 {
1114 	int overflow;
1115 	struct pt_regs regs;
1116 	struct perf_sf_sde_regs *sde_regs;
1117 	struct perf_sample_data data;
1118 
1119 	/* Setup perf sample */
1120 	perf_sample_data_init(&data, 0, event->hw.last_period);
1121 
1122 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1123 	 * using the Program Request Alert code.  The regs.int_parm_long
1124 	 * field which is unused contains additional sample-data-entry related
1125 	 * indicators.
1126 	 */
1127 	memset(&regs, 0, sizeof(regs));
1128 	regs.int_code = 0x1407;
1129 	regs.int_parm = CPU_MF_INT_SF_PRA;
1130 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1131 
1132 	psw_bits(regs.psw).ia	= basic->ia;
1133 	psw_bits(regs.psw).dat	= basic->T;
1134 	psw_bits(regs.psw).wait = basic->W;
1135 	psw_bits(regs.psw).pstate = basic->P;
1136 	psw_bits(regs.psw).as	= basic->AS;
1137 
1138 	/*
1139 	 * Use the hardware provided configuration level to decide if the
1140 	 * sample belongs to a guest or host. If that is not available,
1141 	 * fall back to the following heuristics:
1142 	 * A non-zero guest program parameter always indicates a guest
1143 	 * sample. Some early samples or samples from guests without
1144 	 * lpp usage would be misaccounted to the host. We use the asn
1145 	 * value as an addon heuristic to detect most of these guest samples.
1146 	 * If the value differs from 0xffff (the host value), we assume to
1147 	 * be a KVM guest.
1148 	 */
1149 	switch (basic->CL) {
1150 	case 1: /* logical partition */
1151 		sde_regs->in_guest = 0;
1152 		break;
1153 	case 2: /* virtual machine */
1154 		sde_regs->in_guest = 1;
1155 		break;
1156 	default: /* old machine, use heuristics */
1157 		if (basic->gpp || basic->prim_asn != 0xffff)
1158 			sde_regs->in_guest = 1;
1159 		break;
1160 	}
1161 
1162 	/*
1163 	 * Store the PID value from the sample-data-entry to be
1164 	 * processed and resolved by cpumsf_output_event_pid().
1165 	 */
1166 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1167 
1168 	overflow = 0;
1169 	if (perf_exclude_event(event, &regs, sde_regs))
1170 		goto out;
1171 	if (perf_event_overflow(event, &data, &regs)) {
1172 		overflow = 1;
1173 		event->pmu->stop(event, 0);
1174 	}
1175 	perf_event_update_userpage(event);
1176 out:
1177 	return overflow;
1178 }
1179 
1180 static void perf_event_count_update(struct perf_event *event, u64 count)
1181 {
1182 	local64_add(count, &event->count);
1183 }
1184 
1185 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1186  * @event:	The perf event
1187  * @sdbt:	Sample-data-block table
1188  * @overflow:	Event overflow counter
1189  *
1190  * Walks through a sample-data-block and collects sampling data entries that are
1191  * then pushed to the perf event subsystem.  Depending on the sampling function,
1192  * there can be either basic-sampling or combined-sampling data entries.  A
1193  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1194  * data entry.	The sampling function is determined by the flags in the perf
1195  * event hardware structure.  The function always works with a combined-sampling
1196  * data entry but ignores the the diagnostic portion if it is not available.
1197  *
1198  * Note that the implementation focuses on basic-sampling data entries and, if
1199  * such an entry is not valid, the entire combined-sampling data entry is
1200  * ignored.
1201  *
1202  * The overflow variables counts the number of samples that has been discarded
1203  * due to a perf event overflow.
1204  */
1205 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1206 			       unsigned long long *overflow)
1207 {
1208 	struct hws_trailer_entry *te;
1209 	struct hws_basic_entry *sample;
1210 
1211 	te = trailer_entry_ptr((unsigned long)sdbt);
1212 	sample = (struct hws_basic_entry *)sdbt;
1213 	while ((unsigned long *)sample < (unsigned long *)te) {
1214 		/* Check for an empty sample */
1215 		if (!sample->def || sample->LS)
1216 			break;
1217 
1218 		/* Update perf event period */
1219 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1220 
1221 		/* Check whether sample is valid */
1222 		if (sample->def == 0x0001) {
1223 			/* If an event overflow occurred, the PMU is stopped to
1224 			 * throttle event delivery.  Remaining sample data is
1225 			 * discarded.
1226 			 */
1227 			if (!*overflow) {
1228 				/* Check whether sample is consistent */
1229 				if (sample->I == 0 && sample->W == 0) {
1230 					/* Deliver sample data to perf */
1231 					*overflow = perf_push_sample(event,
1232 								     sample);
1233 				}
1234 			} else
1235 				/* Count discarded samples */
1236 				*overflow += 1;
1237 		} else {
1238 			debug_sprintf_event(sfdbg, 4,
1239 					    "%s: Found unknown"
1240 					    " sampling data entry: te->f %i"
1241 					    " basic.def %#4x (%p)\n", __func__,
1242 					    te->header.f, sample->def, sample);
1243 			/* Sample slot is not yet written or other record.
1244 			 *
1245 			 * This condition can occur if the buffer was reused
1246 			 * from a combined basic- and diagnostic-sampling.
1247 			 * If only basic-sampling is then active, entries are
1248 			 * written into the larger diagnostic entries.
1249 			 * This is typically the case for sample-data-blocks
1250 			 * that are not full.  Stop processing if the first
1251 			 * invalid format was detected.
1252 			 */
1253 			if (!te->header.f)
1254 				break;
1255 		}
1256 
1257 		/* Reset sample slot and advance to next sample */
1258 		sample->def = 0;
1259 		sample++;
1260 	}
1261 }
1262 
1263 /* hw_perf_event_update() - Process sampling buffer
1264  * @event:	The perf event
1265  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1266  *
1267  * Processes the sampling buffer and create perf event samples.
1268  * The sampling buffer position are retrieved and saved in the TEAR_REG
1269  * register of the specified perf event.
1270  *
1271  * Only full sample-data-blocks are processed.	Specify the flush_all flag
1272  * to also walk through partially filled sample-data-blocks.
1273  */
1274 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1275 {
1276 	unsigned long long event_overflow, sampl_overflow, num_sdb;
1277 	union hws_trailer_header old, prev, new;
1278 	struct hw_perf_event *hwc = &event->hw;
1279 	struct hws_trailer_entry *te;
1280 	unsigned long *sdbt, sdb;
1281 	int done;
1282 
1283 	/*
1284 	 * AUX buffer is used when in diagnostic sampling mode.
1285 	 * No perf events/samples are created.
1286 	 */
1287 	if (SAMPL_DIAG_MODE(&event->hw))
1288 		return;
1289 
1290 	sdbt = (unsigned long *)TEAR_REG(hwc);
1291 	done = event_overflow = sampl_overflow = num_sdb = 0;
1292 	while (!done) {
1293 		/* Get the trailer entry of the sample-data-block */
1294 		sdb = (unsigned long)phys_to_virt(*sdbt);
1295 		te = trailer_entry_ptr(sdb);
1296 
1297 		/* Leave loop if no more work to do (block full indicator) */
1298 		if (!te->header.f) {
1299 			done = 1;
1300 			if (!flush_all)
1301 				break;
1302 		}
1303 
1304 		/* Check the sample overflow count */
1305 		if (te->header.overflow)
1306 			/* Account sample overflows and, if a particular limit
1307 			 * is reached, extend the sampling buffer.
1308 			 * For details, see sfb_account_overflows().
1309 			 */
1310 			sampl_overflow += te->header.overflow;
1311 
1312 		/* Timestamps are valid for full sample-data-blocks only */
1313 		debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx/%#lx "
1314 				    "overflow %llu timestamp %#llx\n",
1315 				    __func__, sdb, (unsigned long)sdbt,
1316 				    te->header.overflow,
1317 				    (te->header.f) ? trailer_timestamp(te) : 0ULL);
1318 
1319 		/* Collect all samples from a single sample-data-block and
1320 		 * flag if an (perf) event overflow happened.  If so, the PMU
1321 		 * is stopped and remaining samples will be discarded.
1322 		 */
1323 		hw_collect_samples(event, (unsigned long *)sdb, &event_overflow);
1324 		num_sdb++;
1325 
1326 		/* Reset trailer (using compare-double-and-swap) */
1327 		prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1328 		do {
1329 			old.val = prev.val;
1330 			new.val = prev.val;
1331 			new.f = 0;
1332 			new.a = 1;
1333 			new.overflow = 0;
1334 			prev.val = cmpxchg128(&te->header.val, old.val, new.val);
1335 		} while (prev.val != old.val);
1336 
1337 		/* Advance to next sample-data-block */
1338 		sdbt++;
1339 		if (is_link_entry(sdbt))
1340 			sdbt = get_next_sdbt(sdbt);
1341 
1342 		/* Update event hardware registers */
1343 		TEAR_REG(hwc) = (unsigned long) sdbt;
1344 
1345 		/* Stop processing sample-data if all samples of the current
1346 		 * sample-data-block were flushed even if it was not full.
1347 		 */
1348 		if (flush_all && done)
1349 			break;
1350 	}
1351 
1352 	/* Account sample overflows in the event hardware structure */
1353 	if (sampl_overflow)
1354 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1355 						 sampl_overflow, 1 + num_sdb);
1356 
1357 	/* Perf_event_overflow() and perf_event_account_interrupt() limit
1358 	 * the interrupt rate to an upper limit. Roughly 1000 samples per
1359 	 * task tick.
1360 	 * Hitting this limit results in a large number
1361 	 * of throttled REF_REPORT_THROTTLE entries and the samples
1362 	 * are dropped.
1363 	 * Slightly increase the interval to avoid hitting this limit.
1364 	 */
1365 	if (event_overflow) {
1366 		SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
1367 		debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n",
1368 				    __func__,
1369 				    DIV_ROUND_UP(SAMPL_RATE(hwc), 10));
1370 	}
1371 
1372 	if (sampl_overflow || event_overflow)
1373 		debug_sprintf_event(sfdbg, 4, "%s: "
1374 				    "overflows: sample %llu event %llu"
1375 				    " total %llu num_sdb %llu\n",
1376 				    __func__, sampl_overflow, event_overflow,
1377 				    OVERFLOW_REG(hwc), num_sdb);
1378 }
1379 
1380 static inline unsigned long aux_sdb_index(struct aux_buffer *aux,
1381 					  unsigned long i)
1382 {
1383 	return i % aux->sfb.num_sdb;
1384 }
1385 
1386 static inline unsigned long aux_sdb_num(unsigned long start, unsigned long end)
1387 {
1388 	return end >= start ? end - start + 1 : 0;
1389 }
1390 
1391 static inline unsigned long aux_sdb_num_alert(struct aux_buffer *aux)
1392 {
1393 	return aux_sdb_num(aux->head, aux->alert_mark);
1394 }
1395 
1396 static inline unsigned long aux_sdb_num_empty(struct aux_buffer *aux)
1397 {
1398 	return aux_sdb_num(aux->head, aux->empty_mark);
1399 }
1400 
1401 /*
1402  * Get trailer entry by index of SDB.
1403  */
1404 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1405 						 unsigned long index)
1406 {
1407 	unsigned long sdb;
1408 
1409 	index = aux_sdb_index(aux, index);
1410 	sdb = aux->sdb_index[index];
1411 	return trailer_entry_ptr(sdb);
1412 }
1413 
1414 /*
1415  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1416  * disabled. Collect the full SDBs in AUX buffer which have not reached
1417  * the point of alert indicator. And ignore the SDBs which are not
1418  * full.
1419  *
1420  * 1. Scan SDBs to see how much data is there and consume them.
1421  * 2. Remove alert indicator in the buffer.
1422  */
1423 static void aux_output_end(struct perf_output_handle *handle)
1424 {
1425 	unsigned long i, range_scan, idx;
1426 	struct aux_buffer *aux;
1427 	struct hws_trailer_entry *te;
1428 
1429 	aux = perf_get_aux(handle);
1430 	if (!aux)
1431 		return;
1432 
1433 	range_scan = aux_sdb_num_alert(aux);
1434 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1435 		te = aux_sdb_trailer(aux, idx);
1436 		if (!te->header.f)
1437 			break;
1438 	}
1439 	/* i is num of SDBs which are full */
1440 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1441 
1442 	/* Remove alert indicators in the buffer */
1443 	te = aux_sdb_trailer(aux, aux->alert_mark);
1444 	te->header.a = 0;
1445 
1446 	debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n",
1447 			    __func__, i, range_scan, aux->head);
1448 }
1449 
1450 /*
1451  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1452  * is first added to the CPU or rescheduled again to the CPU. It is called
1453  * with pmu disabled.
1454  *
1455  * 1. Reset the trailer of SDBs to get ready for new data.
1456  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1457  *    head(tear/dear).
1458  */
1459 static int aux_output_begin(struct perf_output_handle *handle,
1460 			    struct aux_buffer *aux,
1461 			    struct cpu_hw_sf *cpuhw)
1462 {
1463 	unsigned long range, i, range_scan, idx, head, base, offset;
1464 	struct hws_trailer_entry *te;
1465 
1466 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1467 		return -EINVAL;
1468 
1469 	aux->head = handle->head >> PAGE_SHIFT;
1470 	range = (handle->size + 1) >> PAGE_SHIFT;
1471 	if (range <= 1)
1472 		return -ENOMEM;
1473 
1474 	/*
1475 	 * SDBs between aux->head and aux->empty_mark are already ready
1476 	 * for new data. range_scan is num of SDBs not within them.
1477 	 */
1478 	debug_sprintf_event(sfdbg, 6,
1479 			    "%s: range %ld head %ld alert %ld empty %ld\n",
1480 			    __func__, range, aux->head, aux->alert_mark,
1481 			    aux->empty_mark);
1482 	if (range > aux_sdb_num_empty(aux)) {
1483 		range_scan = range - aux_sdb_num_empty(aux);
1484 		idx = aux->empty_mark + 1;
1485 		for (i = 0; i < range_scan; i++, idx++) {
1486 			te = aux_sdb_trailer(aux, idx);
1487 			te->header.f = 0;
1488 			te->header.a = 0;
1489 			te->header.overflow = 0;
1490 		}
1491 		/* Save the position of empty SDBs */
1492 		aux->empty_mark = aux->head + range - 1;
1493 	}
1494 
1495 	/* Set alert indicator */
1496 	aux->alert_mark = aux->head + range/2 - 1;
1497 	te = aux_sdb_trailer(aux, aux->alert_mark);
1498 	te->header.a = 1;
1499 
1500 	/* Reset hardware buffer head */
1501 	head = aux_sdb_index(aux, aux->head);
1502 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1503 	offset = head % CPUM_SF_SDB_PER_TABLE;
1504 	cpuhw->lsctl.tear = virt_to_phys((void *)base) + offset * sizeof(unsigned long);
1505 	cpuhw->lsctl.dear = virt_to_phys((void *)aux->sdb_index[head]);
1506 
1507 	debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld "
1508 			    "index %ld tear %#lx dear %#lx\n", __func__,
1509 			    aux->head, aux->alert_mark, aux->empty_mark,
1510 			    head / CPUM_SF_SDB_PER_TABLE,
1511 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
1512 
1513 	return 0;
1514 }
1515 
1516 /*
1517  * Set alert indicator on SDB at index @alert_index while sampler is running.
1518  *
1519  * Return true if successfully.
1520  * Return false if full indicator is already set by hardware sampler.
1521  */
1522 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1523 			  unsigned long long *overflow)
1524 {
1525 	union hws_trailer_header old, prev, new;
1526 	struct hws_trailer_entry *te;
1527 
1528 	te = aux_sdb_trailer(aux, alert_index);
1529 	prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1530 	do {
1531 		old.val = prev.val;
1532 		new.val = prev.val;
1533 		*overflow = old.overflow;
1534 		if (old.f) {
1535 			/*
1536 			 * SDB is already set by hardware.
1537 			 * Abort and try to set somewhere
1538 			 * behind.
1539 			 */
1540 			return false;
1541 		}
1542 		new.a = 1;
1543 		new.overflow = 0;
1544 		prev.val = cmpxchg128(&te->header.val, old.val, new.val);
1545 	} while (prev.val != old.val);
1546 	return true;
1547 }
1548 
1549 /*
1550  * aux_reset_buffer() - Scan and setup SDBs for new samples
1551  * @aux:	The AUX buffer to set
1552  * @range:	The range of SDBs to scan started from aux->head
1553  * @overflow:	Set to overflow count
1554  *
1555  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1556  * marked as empty, check if it is already set full by the hardware sampler.
1557  * If yes, that means new data is already there before we can set an alert
1558  * indicator. Caller should try to set alert indicator to some position behind.
1559  *
1560  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1561  * previously and have already been consumed by user space. Reset these SDBs
1562  * (clear full indicator and alert indicator) for new data.
1563  * If aux->alert_mark fall in this area, just set it. Overflow count is
1564  * recorded while scanning.
1565  *
1566  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1567  * and ready for new samples. So scanning on this area could be skipped.
1568  *
1569  * Return true if alert indicator is set successfully and false if not.
1570  */
1571 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1572 			     unsigned long long *overflow)
1573 {
1574 	unsigned long i, range_scan, idx, idx_old;
1575 	union hws_trailer_header old, prev, new;
1576 	unsigned long long orig_overflow;
1577 	struct hws_trailer_entry *te;
1578 
1579 	debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld "
1580 			    "empty %ld\n", __func__, range, aux->head,
1581 			    aux->alert_mark, aux->empty_mark);
1582 	if (range <= aux_sdb_num_empty(aux))
1583 		/*
1584 		 * No need to scan. All SDBs in range are marked as empty.
1585 		 * Just set alert indicator. Should check race with hardware
1586 		 * sampler.
1587 		 */
1588 		return aux_set_alert(aux, aux->alert_mark, overflow);
1589 
1590 	if (aux->alert_mark <= aux->empty_mark)
1591 		/*
1592 		 * Set alert indicator on empty SDB. Should check race
1593 		 * with hardware sampler.
1594 		 */
1595 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1596 			return false;
1597 
1598 	/*
1599 	 * Scan the SDBs to clear full and alert indicator used previously.
1600 	 * Start scanning from one SDB behind empty_mark. If the new alert
1601 	 * indicator fall into this range, set it.
1602 	 */
1603 	range_scan = range - aux_sdb_num_empty(aux);
1604 	idx_old = idx = aux->empty_mark + 1;
1605 	for (i = 0; i < range_scan; i++, idx++) {
1606 		te = aux_sdb_trailer(aux, idx);
1607 		prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1608 		do {
1609 			old.val = prev.val;
1610 			new.val = prev.val;
1611 			orig_overflow = old.overflow;
1612 			new.f = 0;
1613 			new.overflow = 0;
1614 			if (idx == aux->alert_mark)
1615 				new.a = 1;
1616 			else
1617 				new.a = 0;
1618 			prev.val = cmpxchg128(&te->header.val, old.val, new.val);
1619 		} while (prev.val != old.val);
1620 		*overflow += orig_overflow;
1621 	}
1622 
1623 	/* Update empty_mark to new position */
1624 	aux->empty_mark = aux->head + range - 1;
1625 
1626 	debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld "
1627 			    "empty %ld\n", __func__, range_scan, idx_old,
1628 			    idx - 1, aux->empty_mark);
1629 	return true;
1630 }
1631 
1632 /*
1633  * Measurement alert handler for diagnostic mode sampling.
1634  */
1635 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1636 {
1637 	struct aux_buffer *aux;
1638 	int done = 0;
1639 	unsigned long range = 0, size;
1640 	unsigned long long overflow = 0;
1641 	struct perf_output_handle *handle = &cpuhw->handle;
1642 	unsigned long num_sdb;
1643 
1644 	aux = perf_get_aux(handle);
1645 	if (WARN_ON_ONCE(!aux))
1646 		return;
1647 
1648 	/* Inform user space new data arrived */
1649 	size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
1650 	debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__,
1651 			    size >> PAGE_SHIFT);
1652 	perf_aux_output_end(handle, size);
1653 
1654 	num_sdb = aux->sfb.num_sdb;
1655 	while (!done) {
1656 		/* Get an output handle */
1657 		aux = perf_aux_output_begin(handle, cpuhw->event);
1658 		if (handle->size == 0) {
1659 			pr_err("The AUX buffer with %lu pages for the "
1660 			       "diagnostic-sampling mode is full\n",
1661 				num_sdb);
1662 			break;
1663 		}
1664 		if (WARN_ON_ONCE(!aux))
1665 			return;
1666 
1667 		/* Update head and alert_mark to new position */
1668 		aux->head = handle->head >> PAGE_SHIFT;
1669 		range = (handle->size + 1) >> PAGE_SHIFT;
1670 		if (range == 1)
1671 			aux->alert_mark = aux->head;
1672 		else
1673 			aux->alert_mark = aux->head + range/2 - 1;
1674 
1675 		if (aux_reset_buffer(aux, range, &overflow)) {
1676 			if (!overflow) {
1677 				done = 1;
1678 				break;
1679 			}
1680 			size = range << PAGE_SHIFT;
1681 			perf_aux_output_end(&cpuhw->handle, size);
1682 			pr_err("Sample data caused the AUX buffer with %lu "
1683 			       "pages to overflow\n", aux->sfb.num_sdb);
1684 			debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld "
1685 					    "overflow %lld\n", __func__,
1686 					    aux->head, range, overflow);
1687 		} else {
1688 			size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
1689 			perf_aux_output_end(&cpuhw->handle, size);
1690 			debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1691 					    "already full, try another\n",
1692 					    __func__,
1693 					    aux->head, aux->alert_mark);
1694 		}
1695 	}
1696 
1697 	if (done)
1698 		debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1699 				    "empty %ld\n", __func__, aux->head,
1700 				    aux->alert_mark, aux->empty_mark);
1701 }
1702 
1703 /*
1704  * Callback when freeing AUX buffers.
1705  */
1706 static void aux_buffer_free(void *data)
1707 {
1708 	struct aux_buffer *aux = data;
1709 	unsigned long i, num_sdbt;
1710 
1711 	if (!aux)
1712 		return;
1713 
1714 	/* Free SDBT. SDB is freed by the caller */
1715 	num_sdbt = aux->sfb.num_sdbt;
1716 	for (i = 0; i < num_sdbt; i++)
1717 		free_page(aux->sdbt_index[i]);
1718 
1719 	kfree(aux->sdbt_index);
1720 	kfree(aux->sdb_index);
1721 	kfree(aux);
1722 
1723 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt);
1724 }
1725 
1726 static void aux_sdb_init(unsigned long sdb)
1727 {
1728 	struct hws_trailer_entry *te;
1729 
1730 	te = trailer_entry_ptr(sdb);
1731 
1732 	/* Save clock base */
1733 	te->clock_base = 1;
1734 	te->progusage2 = tod_clock_base.tod;
1735 }
1736 
1737 /*
1738  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1739  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1740  * @pages:	Array of pointers to buffer pages passed from perf core
1741  * @nr_pages:	Total pages
1742  * @snapshot:	Flag for snapshot mode
1743  *
1744  * This is the callback when setup an event using AUX buffer. Perf tool can
1745  * trigger this by an additional mmap() call on the event. Unlike the buffer
1746  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1747  * the task among online cpus when it is a per-thread event.
1748  *
1749  * Return the private AUX buffer structure if success or NULL if fails.
1750  */
1751 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1752 			      int nr_pages, bool snapshot)
1753 {
1754 	struct sf_buffer *sfb;
1755 	struct aux_buffer *aux;
1756 	unsigned long *new, *tail;
1757 	int i, n_sdbt;
1758 
1759 	if (!nr_pages || !pages)
1760 		return NULL;
1761 
1762 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1763 		pr_err("AUX buffer size (%i pages) is larger than the "
1764 		       "maximum sampling buffer limit\n",
1765 		       nr_pages);
1766 		return NULL;
1767 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1768 		pr_err("AUX buffer size (%i pages) is less than the "
1769 		       "minimum sampling buffer limit\n",
1770 		       nr_pages);
1771 		return NULL;
1772 	}
1773 
1774 	/* Allocate aux_buffer struct for the event */
1775 	aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1776 	if (!aux)
1777 		goto no_aux;
1778 	sfb = &aux->sfb;
1779 
1780 	/* Allocate sdbt_index for fast reference */
1781 	n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE);
1782 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1783 	if (!aux->sdbt_index)
1784 		goto no_sdbt_index;
1785 
1786 	/* Allocate sdb_index for fast reference */
1787 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1788 	if (!aux->sdb_index)
1789 		goto no_sdb_index;
1790 
1791 	/* Allocate the first SDBT */
1792 	sfb->num_sdbt = 0;
1793 	sfb->sdbt = (unsigned long *)get_zeroed_page(GFP_KERNEL);
1794 	if (!sfb->sdbt)
1795 		goto no_sdbt;
1796 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1797 	tail = sfb->tail = sfb->sdbt;
1798 
1799 	/*
1800 	 * Link the provided pages of AUX buffer to SDBT.
1801 	 * Allocate SDBT if needed.
1802 	 */
1803 	for (i = 0; i < nr_pages; i++, tail++) {
1804 		if (require_table_link(tail)) {
1805 			new = (unsigned long *)get_zeroed_page(GFP_KERNEL);
1806 			if (!new)
1807 				goto no_sdbt;
1808 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1809 			/* Link current page to tail of chain */
1810 			*tail = virt_to_phys(new) + 1;
1811 			tail = new;
1812 		}
1813 		/* Tail is the entry in a SDBT */
1814 		*tail = virt_to_phys(pages[i]);
1815 		aux->sdb_index[i] = (unsigned long)pages[i];
1816 		aux_sdb_init((unsigned long)pages[i]);
1817 	}
1818 	sfb->num_sdb = nr_pages;
1819 
1820 	/* Link the last entry in the SDBT to the first SDBT */
1821 	*tail = virt_to_phys(sfb->sdbt) + 1;
1822 	sfb->tail = tail;
1823 
1824 	/*
1825 	 * Initial all SDBs are zeroed. Mark it as empty.
1826 	 * So there is no need to clear the full indicator
1827 	 * when this event is first added.
1828 	 */
1829 	aux->empty_mark = sfb->num_sdb - 1;
1830 
1831 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__,
1832 			    sfb->num_sdbt, sfb->num_sdb);
1833 
1834 	return aux;
1835 
1836 no_sdbt:
1837 	/* SDBs (AUX buffer pages) are freed by caller */
1838 	for (i = 0; i < sfb->num_sdbt; i++)
1839 		free_page(aux->sdbt_index[i]);
1840 	kfree(aux->sdb_index);
1841 no_sdb_index:
1842 	kfree(aux->sdbt_index);
1843 no_sdbt_index:
1844 	kfree(aux);
1845 no_aux:
1846 	return NULL;
1847 }
1848 
1849 static void cpumsf_pmu_read(struct perf_event *event)
1850 {
1851 	/* Nothing to do ... updates are interrupt-driven */
1852 }
1853 
1854 /* Check if the new sampling period/frequency is appropriate.
1855  *
1856  * Return non-zero on error and zero on passed checks.
1857  */
1858 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1859 {
1860 	struct hws_qsi_info_block si;
1861 	unsigned long rate;
1862 	bool do_freq;
1863 
1864 	memset(&si, 0, sizeof(si));
1865 	if (event->cpu == -1) {
1866 		if (qsi(&si))
1867 			return -ENODEV;
1868 	} else {
1869 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1870 		 * sampling structure for accessing the CPU-specific QSI.
1871 		 */
1872 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1873 
1874 		si = cpuhw->qsi;
1875 	}
1876 
1877 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1878 	rate = getrate(do_freq, value, &si);
1879 	if (!rate)
1880 		return -EINVAL;
1881 
1882 	event->attr.sample_period = rate;
1883 	SAMPL_RATE(&event->hw) = rate;
1884 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1885 	debug_sprintf_event(sfdbg, 4, "%s:"
1886 			    " cpu %d value %#llx period %#llx freq %d\n",
1887 			    __func__, event->cpu, value,
1888 			    event->attr.sample_period, do_freq);
1889 	return 0;
1890 }
1891 
1892 /* Activate sampling control.
1893  * Next call of pmu_enable() starts sampling.
1894  */
1895 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1896 {
1897 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1898 
1899 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1900 		return;
1901 
1902 	if (flags & PERF_EF_RELOAD)
1903 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1904 
1905 	perf_pmu_disable(event->pmu);
1906 	event->hw.state = 0;
1907 	cpuhw->lsctl.cs = 1;
1908 	if (SAMPL_DIAG_MODE(&event->hw))
1909 		cpuhw->lsctl.cd = 1;
1910 	perf_pmu_enable(event->pmu);
1911 }
1912 
1913 /* Deactivate sampling control.
1914  * Next call of pmu_enable() stops sampling.
1915  */
1916 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1917 {
1918 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1919 
1920 	if (event->hw.state & PERF_HES_STOPPED)
1921 		return;
1922 
1923 	perf_pmu_disable(event->pmu);
1924 	cpuhw->lsctl.cs = 0;
1925 	cpuhw->lsctl.cd = 0;
1926 	event->hw.state |= PERF_HES_STOPPED;
1927 
1928 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1929 		hw_perf_event_update(event, 1);
1930 		event->hw.state |= PERF_HES_UPTODATE;
1931 	}
1932 	perf_pmu_enable(event->pmu);
1933 }
1934 
1935 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1936 {
1937 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1938 	struct aux_buffer *aux;
1939 	int err;
1940 
1941 	if (cpuhw->flags & PMU_F_IN_USE)
1942 		return -EAGAIN;
1943 
1944 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1945 		return -EINVAL;
1946 
1947 	err = 0;
1948 	perf_pmu_disable(event->pmu);
1949 
1950 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1951 
1952 	/* Set up sampling controls.  Always program the sampling register
1953 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1954 	 * that is used by hw_perf_event_update() to store the sampling buffer
1955 	 * position after samples have been flushed.
1956 	 */
1957 	cpuhw->lsctl.s = 0;
1958 	cpuhw->lsctl.h = 1;
1959 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1960 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1961 		cpuhw->lsctl.tear = virt_to_phys(cpuhw->sfb.sdbt);
1962 		cpuhw->lsctl.dear = *(unsigned long *)cpuhw->sfb.sdbt;
1963 		TEAR_REG(&event->hw) = (unsigned long)cpuhw->sfb.sdbt;
1964 	}
1965 
1966 	/* Ensure sampling functions are in the disabled state.  If disabled,
1967 	 * switch on sampling enable control. */
1968 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1969 		err = -EAGAIN;
1970 		goto out;
1971 	}
1972 	if (SAMPL_DIAG_MODE(&event->hw)) {
1973 		aux = perf_aux_output_begin(&cpuhw->handle, event);
1974 		if (!aux) {
1975 			err = -EINVAL;
1976 			goto out;
1977 		}
1978 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
1979 		if (err)
1980 			goto out;
1981 		cpuhw->lsctl.ed = 1;
1982 	}
1983 	cpuhw->lsctl.es = 1;
1984 
1985 	/* Set in_use flag and store event */
1986 	cpuhw->event = event;
1987 	cpuhw->flags |= PMU_F_IN_USE;
1988 
1989 	if (flags & PERF_EF_START)
1990 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
1991 out:
1992 	perf_event_update_userpage(event);
1993 	perf_pmu_enable(event->pmu);
1994 	return err;
1995 }
1996 
1997 static void cpumsf_pmu_del(struct perf_event *event, int flags)
1998 {
1999 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
2000 
2001 	perf_pmu_disable(event->pmu);
2002 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
2003 
2004 	cpuhw->lsctl.es = 0;
2005 	cpuhw->lsctl.ed = 0;
2006 	cpuhw->flags &= ~PMU_F_IN_USE;
2007 	cpuhw->event = NULL;
2008 
2009 	if (SAMPL_DIAG_MODE(&event->hw))
2010 		aux_output_end(&cpuhw->handle);
2011 	perf_event_update_userpage(event);
2012 	perf_pmu_enable(event->pmu);
2013 }
2014 
2015 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
2016 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
2017 
2018 /* Attribute list for CPU_SF.
2019  *
2020  * The availablitiy depends on the CPU_MF sampling facility authorization
2021  * for basic + diagnositic samples. This is determined at initialization
2022  * time by the sampling facility device driver.
2023  * If the authorization for basic samples is turned off, it should be
2024  * also turned off for diagnostic sampling.
2025  *
2026  * During initialization of the device driver, check the authorization
2027  * level for diagnostic sampling and installs the attribute
2028  * file for diagnostic sampling if necessary.
2029  *
2030  * For now install a placeholder to reference all possible attributes:
2031  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
2032  * Add another entry for the final NULL pointer.
2033  */
2034 enum {
2035 	SF_CYCLES_BASIC_ATTR_IDX = 0,
2036 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
2037 	SF_CYCLES_ATTR_MAX
2038 };
2039 
2040 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
2041 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
2042 };
2043 
2044 PMU_FORMAT_ATTR(event, "config:0-63");
2045 
2046 static struct attribute *cpumsf_pmu_format_attr[] = {
2047 	&format_attr_event.attr,
2048 	NULL,
2049 };
2050 
2051 static struct attribute_group cpumsf_pmu_events_group = {
2052 	.name = "events",
2053 	.attrs = cpumsf_pmu_events_attr,
2054 };
2055 
2056 static struct attribute_group cpumsf_pmu_format_group = {
2057 	.name = "format",
2058 	.attrs = cpumsf_pmu_format_attr,
2059 };
2060 
2061 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
2062 	&cpumsf_pmu_events_group,
2063 	&cpumsf_pmu_format_group,
2064 	NULL,
2065 };
2066 
2067 static struct pmu cpumf_sampling = {
2068 	.pmu_enable   = cpumsf_pmu_enable,
2069 	.pmu_disable  = cpumsf_pmu_disable,
2070 
2071 	.event_init   = cpumsf_pmu_event_init,
2072 	.add	      = cpumsf_pmu_add,
2073 	.del	      = cpumsf_pmu_del,
2074 
2075 	.start	      = cpumsf_pmu_start,
2076 	.stop	      = cpumsf_pmu_stop,
2077 	.read	      = cpumsf_pmu_read,
2078 
2079 	.attr_groups  = cpumsf_pmu_attr_groups,
2080 
2081 	.setup_aux    = aux_buffer_setup,
2082 	.free_aux     = aux_buffer_free,
2083 
2084 	.check_period = cpumsf_pmu_check_period,
2085 };
2086 
2087 static void cpumf_measurement_alert(struct ext_code ext_code,
2088 				    unsigned int alert, unsigned long unused)
2089 {
2090 	struct cpu_hw_sf *cpuhw;
2091 
2092 	if (!(alert & CPU_MF_INT_SF_MASK))
2093 		return;
2094 	inc_irq_stat(IRQEXT_CMS);
2095 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2096 
2097 	/* Measurement alerts are shared and might happen when the PMU
2098 	 * is not reserved.  Ignore these alerts in this case. */
2099 	if (!(cpuhw->flags & PMU_F_RESERVED))
2100 		return;
2101 
2102 	/* The processing below must take care of multiple alert events that
2103 	 * might be indicated concurrently. */
2104 
2105 	/* Program alert request */
2106 	if (alert & CPU_MF_INT_SF_PRA) {
2107 		if (cpuhw->flags & PMU_F_IN_USE)
2108 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2109 				hw_collect_aux(cpuhw);
2110 			else
2111 				hw_perf_event_update(cpuhw->event, 0);
2112 		else
2113 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2114 	}
2115 
2116 	/* Report measurement alerts only for non-PRA codes */
2117 	if (alert != CPU_MF_INT_SF_PRA)
2118 		debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__,
2119 				    alert);
2120 
2121 	/* Sampling authorization change request */
2122 	if (alert & CPU_MF_INT_SF_SACA)
2123 		qsi(&cpuhw->qsi);
2124 
2125 	/* Loss of sample data due to high-priority machine activities */
2126 	if (alert & CPU_MF_INT_SF_LSDA) {
2127 		pr_err("Sample data was lost\n");
2128 		cpuhw->flags |= PMU_F_ERR_LSDA;
2129 		sf_disable();
2130 	}
2131 
2132 	/* Invalid sampling buffer entry */
2133 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2134 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2135 		       alert);
2136 		cpuhw->flags |= PMU_F_ERR_IBE;
2137 		sf_disable();
2138 	}
2139 }
2140 
2141 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2142 {
2143 	/* Ignore the notification if no events are scheduled on the PMU.
2144 	 * This might be racy...
2145 	 */
2146 	if (!atomic_read(&num_events))
2147 		return 0;
2148 
2149 	local_irq_disable();
2150 	setup_pmc_cpu(&flags);
2151 	local_irq_enable();
2152 	return 0;
2153 }
2154 
2155 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2156 {
2157 	return cpusf_pmu_setup(cpu, PMC_INIT);
2158 }
2159 
2160 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2161 {
2162 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2163 }
2164 
2165 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2166 {
2167 	if (!cpum_sf_avail())
2168 		return -ENODEV;
2169 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2170 }
2171 
2172 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2173 {
2174 	int rc;
2175 	unsigned long min, max;
2176 
2177 	if (!cpum_sf_avail())
2178 		return -ENODEV;
2179 	if (!val || !strlen(val))
2180 		return -EINVAL;
2181 
2182 	/* Valid parameter values: "min,max" or "max" */
2183 	min = CPUM_SF_MIN_SDB;
2184 	max = CPUM_SF_MAX_SDB;
2185 	if (strchr(val, ','))
2186 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2187 	else
2188 		rc = kstrtoul(val, 10, &max);
2189 
2190 	if (min < 2 || min >= max || max > get_num_physpages())
2191 		rc = -EINVAL;
2192 	if (rc)
2193 		return rc;
2194 
2195 	sfb_set_limits(min, max);
2196 	pr_info("The sampling buffer limits have changed to: "
2197 		"min %lu max %lu (diag %lu)\n",
2198 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2199 	return 0;
2200 }
2201 
2202 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2203 static const struct kernel_param_ops param_ops_sfb_size = {
2204 	.set = param_set_sfb_size,
2205 	.get = param_get_sfb_size,
2206 };
2207 
2208 #define RS_INIT_FAILURE_QSI	  0x0001
2209 #define RS_INIT_FAILURE_BSDES	  0x0002
2210 #define RS_INIT_FAILURE_ALRT	  0x0003
2211 #define RS_INIT_FAILURE_PERF	  0x0004
2212 static void __init pr_cpumsf_err(unsigned int reason)
2213 {
2214 	pr_err("Sampling facility support for perf is not available: "
2215 	       "reason %#x\n", reason);
2216 }
2217 
2218 static int __init init_cpum_sampling_pmu(void)
2219 {
2220 	struct hws_qsi_info_block si;
2221 	int err;
2222 
2223 	if (!cpum_sf_avail())
2224 		return -ENODEV;
2225 
2226 	memset(&si, 0, sizeof(si));
2227 	if (qsi(&si)) {
2228 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2229 		return -ENODEV;
2230 	}
2231 
2232 	if (!si.as && !si.ad)
2233 		return -ENODEV;
2234 
2235 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2236 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2237 		return -EINVAL;
2238 	}
2239 
2240 	if (si.ad) {
2241 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2242 		/* Sampling of diagnostic data authorized,
2243 		 * install event into attribute list of PMU device.
2244 		 */
2245 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2246 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2247 	}
2248 
2249 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2250 	if (!sfdbg) {
2251 		pr_err("Registering for s390dbf failed\n");
2252 		return -ENOMEM;
2253 	}
2254 	debug_register_view(sfdbg, &debug_sprintf_view);
2255 
2256 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2257 				    cpumf_measurement_alert);
2258 	if (err) {
2259 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2260 		debug_unregister(sfdbg);
2261 		goto out;
2262 	}
2263 
2264 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2265 	if (err) {
2266 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2267 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2268 					cpumf_measurement_alert);
2269 		debug_unregister(sfdbg);
2270 		goto out;
2271 	}
2272 
2273 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2274 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2275 out:
2276 	return err;
2277 }
2278 
2279 arch_initcall(init_cpum_sampling_pmu);
2280 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644);
2281