xref: /openbmc/linux/arch/s390/kernel/perf_cpum_sf.c (revision 844f5ed5)
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 <asm-generic/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 	int err;
598 	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
599 
600 	err = 0;
601 	switch (*((int *) flags)) {
602 	case PMC_INIT:
603 		memset(cpusf, 0, sizeof(*cpusf));
604 		err = qsi(&cpusf->qsi);
605 		if (err)
606 			break;
607 		cpusf->flags |= PMU_F_RESERVED;
608 		err = sf_disable();
609 		if (err)
610 			pr_err("Switching off the sampling facility failed "
611 			       "with rc %i\n", err);
612 		break;
613 	case PMC_RELEASE:
614 		cpusf->flags &= ~PMU_F_RESERVED;
615 		err = sf_disable();
616 		if (err) {
617 			pr_err("Switching off the sampling facility failed "
618 			       "with rc %i\n", err);
619 		} else
620 			deallocate_buffers(cpusf);
621 		break;
622 	}
623 	if (err)
624 		*((int *) flags) |= PMC_FAILURE;
625 }
626 
627 static void release_pmc_hardware(void)
628 {
629 	int flags = PMC_RELEASE;
630 
631 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
632 	on_each_cpu(setup_pmc_cpu, &flags, 1);
633 }
634 
635 static int reserve_pmc_hardware(void)
636 {
637 	int flags = PMC_INIT;
638 
639 	on_each_cpu(setup_pmc_cpu, &flags, 1);
640 	if (flags & PMC_FAILURE) {
641 		release_pmc_hardware();
642 		return -ENODEV;
643 	}
644 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
645 
646 	return 0;
647 }
648 
649 static void hw_perf_event_destroy(struct perf_event *event)
650 {
651 	/* Release PMC if this is the last perf event */
652 	if (!atomic_add_unless(&num_events, -1, 1)) {
653 		mutex_lock(&pmc_reserve_mutex);
654 		if (atomic_dec_return(&num_events) == 0)
655 			release_pmc_hardware();
656 		mutex_unlock(&pmc_reserve_mutex);
657 	}
658 }
659 
660 static void hw_init_period(struct hw_perf_event *hwc, u64 period)
661 {
662 	hwc->sample_period = period;
663 	hwc->last_period = hwc->sample_period;
664 	local64_set(&hwc->period_left, hwc->sample_period);
665 }
666 
667 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
668 				   unsigned long rate)
669 {
670 	return clamp_t(unsigned long, rate,
671 		       si->min_sampl_rate, si->max_sampl_rate);
672 }
673 
674 static u32 cpumsf_pid_type(struct perf_event *event,
675 			   u32 pid, enum pid_type type)
676 {
677 	struct task_struct *tsk;
678 
679 	/* Idle process */
680 	if (!pid)
681 		goto out;
682 
683 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
684 	pid = -1;
685 	if (tsk) {
686 		/*
687 		 * Only top level events contain the pid namespace in which
688 		 * they are created.
689 		 */
690 		if (event->parent)
691 			event = event->parent;
692 		pid = __task_pid_nr_ns(tsk, type, event->ns);
693 		/*
694 		 * See also 1d953111b648
695 		 * "perf/core: Don't report zero PIDs for exiting tasks".
696 		 */
697 		if (!pid && !pid_alive(tsk))
698 			pid = -1;
699 	}
700 out:
701 	return pid;
702 }
703 
704 static void cpumsf_output_event_pid(struct perf_event *event,
705 				    struct perf_sample_data *data,
706 				    struct pt_regs *regs)
707 {
708 	u32 pid;
709 	struct perf_event_header header;
710 	struct perf_output_handle handle;
711 
712 	/*
713 	 * Obtain the PID from the basic-sampling data entry and
714 	 * correct the data->tid_entry.pid value.
715 	 */
716 	pid = data->tid_entry.pid;
717 
718 	/* Protect callchain buffers, tasks */
719 	rcu_read_lock();
720 
721 	perf_prepare_sample(data, event, regs);
722 	perf_prepare_header(&header, data, event, regs);
723 	if (perf_output_begin(&handle, data, event, header.size))
724 		goto out;
725 
726 	/* Update the process ID (see also kernel/events/core.c) */
727 	data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID);
728 	data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID);
729 
730 	perf_output_sample(&handle, &header, data, event);
731 	perf_output_end(&handle);
732 out:
733 	rcu_read_unlock();
734 }
735 
736 static unsigned long getrate(bool freq, unsigned long sample,
737 			     struct hws_qsi_info_block *si)
738 {
739 	unsigned long rate;
740 
741 	if (freq) {
742 		rate = freq_to_sample_rate(si, sample);
743 		rate = hw_limit_rate(si, rate);
744 	} else {
745 		/* The min/max sampling rates specifies the valid range
746 		 * of sample periods.  If the specified sample period is
747 		 * out of range, limit the period to the range boundary.
748 		 */
749 		rate = hw_limit_rate(si, sample);
750 
751 		/* The perf core maintains a maximum sample rate that is
752 		 * configurable through the sysctl interface.  Ensure the
753 		 * sampling rate does not exceed this value.  This also helps
754 		 * to avoid throttling when pushing samples with
755 		 * perf_event_overflow().
756 		 */
757 		if (sample_rate_to_freq(si, rate) >
758 		    sysctl_perf_event_sample_rate) {
759 			debug_sprintf_event(sfdbg, 1, "%s: "
760 					    "Sampling rate exceeds maximum "
761 					    "perf sample rate\n", __func__);
762 			rate = 0;
763 		}
764 	}
765 	return rate;
766 }
767 
768 /* The sampling information (si) contains information about the
769  * min/max sampling intervals and the CPU speed.  So calculate the
770  * correct sampling interval and avoid the whole period adjust
771  * feedback loop.
772  *
773  * Since the CPU Measurement sampling facility can not handle frequency
774  * calculate the sampling interval when frequency is specified using
775  * this formula:
776  *	interval := cpu_speed * 1000000 / sample_freq
777  *
778  * Returns errno on bad input and zero on success with parameter interval
779  * set to the correct sampling rate.
780  *
781  * Note: This function turns off freq bit to avoid calling function
782  * perf_adjust_period(). This causes frequency adjustment in the common
783  * code part which causes tremendous variations in the counter values.
784  */
785 static int __hw_perf_event_init_rate(struct perf_event *event,
786 				     struct hws_qsi_info_block *si)
787 {
788 	struct perf_event_attr *attr = &event->attr;
789 	struct hw_perf_event *hwc = &event->hw;
790 	unsigned long rate;
791 
792 	if (attr->freq) {
793 		if (!attr->sample_freq)
794 			return -EINVAL;
795 		rate = getrate(attr->freq, attr->sample_freq, si);
796 		attr->freq = 0;		/* Don't call  perf_adjust_period() */
797 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE;
798 	} else {
799 		rate = getrate(attr->freq, attr->sample_period, si);
800 		if (!rate)
801 			return -EINVAL;
802 	}
803 	attr->sample_period = rate;
804 	SAMPL_RATE(hwc) = rate;
805 	hw_init_period(hwc, SAMPL_RATE(hwc));
806 	debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n",
807 			    __func__, event->cpu, event->attr.sample_period,
808 			    event->attr.freq, SAMPLE_FREQ_MODE(hwc));
809 	return 0;
810 }
811 
812 static int __hw_perf_event_init(struct perf_event *event)
813 {
814 	struct cpu_hw_sf *cpuhw;
815 	struct hws_qsi_info_block si;
816 	struct perf_event_attr *attr = &event->attr;
817 	struct hw_perf_event *hwc = &event->hw;
818 	int cpu, err;
819 
820 	/* Reserve CPU-measurement sampling facility */
821 	err = 0;
822 	if (!atomic_inc_not_zero(&num_events)) {
823 		mutex_lock(&pmc_reserve_mutex);
824 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
825 			err = -EBUSY;
826 		else
827 			atomic_inc(&num_events);
828 		mutex_unlock(&pmc_reserve_mutex);
829 	}
830 	event->destroy = hw_perf_event_destroy;
831 
832 	if (err)
833 		goto out;
834 
835 	/* Access per-CPU sampling information (query sampling info) */
836 	/*
837 	 * The event->cpu value can be -1 to count on every CPU, for example,
838 	 * when attaching to a task.  If this is specified, use the query
839 	 * sampling info from the current CPU, otherwise use event->cpu to
840 	 * retrieve the per-CPU information.
841 	 * Later, cpuhw indicates whether to allocate sampling buffers for a
842 	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
843 	 */
844 	memset(&si, 0, sizeof(si));
845 	cpuhw = NULL;
846 	if (event->cpu == -1)
847 		qsi(&si);
848 	else {
849 		/* Event is pinned to a particular CPU, retrieve the per-CPU
850 		 * sampling structure for accessing the CPU-specific QSI.
851 		 */
852 		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
853 		si = cpuhw->qsi;
854 	}
855 
856 	/* Check sampling facility authorization and, if not authorized,
857 	 * fall back to other PMUs.  It is safe to check any CPU because
858 	 * the authorization is identical for all configured CPUs.
859 	 */
860 	if (!si.as) {
861 		err = -ENOENT;
862 		goto out;
863 	}
864 
865 	if (si.ribm & CPU_MF_SF_RIBM_NOTAV) {
866 		pr_warn("CPU Measurement Facility sampling is temporarily not available\n");
867 		err = -EBUSY;
868 		goto out;
869 	}
870 
871 	/* Always enable basic sampling */
872 	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;
873 
874 	/* Check if diagnostic sampling is requested.  Deny if the required
875 	 * sampling authorization is missing.
876 	 */
877 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
878 		if (!si.ad) {
879 			err = -EPERM;
880 			goto out;
881 		}
882 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
883 	}
884 
885 	/* Check and set other sampling flags */
886 	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
887 		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;
888 
889 	err =  __hw_perf_event_init_rate(event, &si);
890 	if (err)
891 		goto out;
892 
893 	/* Initialize sample data overflow accounting */
894 	hwc->extra_reg.reg = REG_OVERFLOW;
895 	OVERFLOW_REG(hwc) = 0;
896 
897 	/* Use AUX buffer. No need to allocate it by ourself */
898 	if (attr->config == PERF_EVENT_CPUM_SF_DIAG)
899 		return 0;
900 
901 	/* Allocate the per-CPU sampling buffer using the CPU information
902 	 * from the event.  If the event is not pinned to a particular
903 	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
904 	 * buffers for each online CPU.
905 	 */
906 	if (cpuhw)
907 		/* Event is pinned to a particular CPU */
908 		err = allocate_buffers(cpuhw, hwc);
909 	else {
910 		/* Event is not pinned, allocate sampling buffer on
911 		 * each online CPU
912 		 */
913 		for_each_online_cpu(cpu) {
914 			cpuhw = &per_cpu(cpu_hw_sf, cpu);
915 			err = allocate_buffers(cpuhw, hwc);
916 			if (err)
917 				break;
918 		}
919 	}
920 
921 	/* If PID/TID sampling is active, replace the default overflow
922 	 * handler to extract and resolve the PIDs from the basic-sampling
923 	 * data entries.
924 	 */
925 	if (event->attr.sample_type & PERF_SAMPLE_TID)
926 		if (is_default_overflow_handler(event))
927 			event->overflow_handler = cpumsf_output_event_pid;
928 out:
929 	return err;
930 }
931 
932 static bool is_callchain_event(struct perf_event *event)
933 {
934 	u64 sample_type = event->attr.sample_type;
935 
936 	return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER |
937 			      PERF_SAMPLE_STACK_USER);
938 }
939 
940 static int cpumsf_pmu_event_init(struct perf_event *event)
941 {
942 	int err;
943 
944 	/* No support for taken branch sampling */
945 	/* No support for callchain, stacks and registers */
946 	if (has_branch_stack(event) || is_callchain_event(event))
947 		return -EOPNOTSUPP;
948 
949 	switch (event->attr.type) {
950 	case PERF_TYPE_RAW:
951 		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
952 		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
953 			return -ENOENT;
954 		break;
955 	case PERF_TYPE_HARDWARE:
956 		/* Support sampling of CPU cycles in addition to the
957 		 * counter facility.  However, the counter facility
958 		 * is more precise and, hence, restrict this PMU to
959 		 * sampling events only.
960 		 */
961 		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
962 			return -ENOENT;
963 		if (!is_sampling_event(event))
964 			return -ENOENT;
965 		break;
966 	default:
967 		return -ENOENT;
968 	}
969 
970 	/* Check online status of the CPU to which the event is pinned */
971 	if (event->cpu >= 0 && !cpu_online(event->cpu))
972 		return -ENODEV;
973 
974 	/* Force reset of idle/hv excludes regardless of what the
975 	 * user requested.
976 	 */
977 	if (event->attr.exclude_hv)
978 		event->attr.exclude_hv = 0;
979 	if (event->attr.exclude_idle)
980 		event->attr.exclude_idle = 0;
981 
982 	err = __hw_perf_event_init(event);
983 	if (unlikely(err))
984 		if (event->destroy)
985 			event->destroy(event);
986 	return err;
987 }
988 
989 static void cpumsf_pmu_enable(struct pmu *pmu)
990 {
991 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
992 	struct hw_perf_event *hwc;
993 	int err;
994 
995 	if (cpuhw->flags & PMU_F_ENABLED)
996 		return;
997 
998 	if (cpuhw->flags & PMU_F_ERR_MASK)
999 		return;
1000 
1001 	/* Check whether to extent the sampling buffer.
1002 	 *
1003 	 * Two conditions trigger an increase of the sampling buffer for a
1004 	 * perf event:
1005 	 *    1. Postponed buffer allocations from the event initialization.
1006 	 *    2. Sampling overflows that contribute to pending allocations.
1007 	 *
1008 	 * Note that the extend_sampling_buffer() function disables the sampling
1009 	 * facility, but it can be fully re-enabled using sampling controls that
1010 	 * have been saved in cpumsf_pmu_disable().
1011 	 */
1012 	if (cpuhw->event) {
1013 		hwc = &cpuhw->event->hw;
1014 		if (!(SAMPL_DIAG_MODE(hwc))) {
1015 			/*
1016 			 * Account number of overflow-designated
1017 			 * buffer extents
1018 			 */
1019 			sfb_account_overflows(cpuhw, hwc);
1020 			extend_sampling_buffer(&cpuhw->sfb, hwc);
1021 		}
1022 		/* Rate may be adjusted with ioctl() */
1023 		cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw);
1024 	}
1025 
1026 	/* (Re)enable the PMU and sampling facility */
1027 	cpuhw->flags |= PMU_F_ENABLED;
1028 	barrier();
1029 
1030 	err = lsctl(&cpuhw->lsctl);
1031 	if (err) {
1032 		cpuhw->flags &= ~PMU_F_ENABLED;
1033 		pr_err("Loading sampling controls failed: op %i err %i\n",
1034 			1, err);
1035 		return;
1036 	}
1037 
1038 	/* Load current program parameter */
1039 	lpp(&S390_lowcore.lpp);
1040 
1041 	debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i "
1042 			    "interval %#lx tear %#lx dear %#lx\n", __func__,
1043 			    cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed,
1044 			    cpuhw->lsctl.cd, cpuhw->lsctl.interval,
1045 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
1046 }
1047 
1048 static void cpumsf_pmu_disable(struct pmu *pmu)
1049 {
1050 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1051 	struct hws_lsctl_request_block inactive;
1052 	struct hws_qsi_info_block si;
1053 	int err;
1054 
1055 	if (!(cpuhw->flags & PMU_F_ENABLED))
1056 		return;
1057 
1058 	if (cpuhw->flags & PMU_F_ERR_MASK)
1059 		return;
1060 
1061 	/* Switch off sampling activation control */
1062 	inactive = cpuhw->lsctl;
1063 	inactive.cs = 0;
1064 	inactive.cd = 0;
1065 
1066 	err = lsctl(&inactive);
1067 	if (err) {
1068 		pr_err("Loading sampling controls failed: op %i err %i\n",
1069 			2, err);
1070 		return;
1071 	}
1072 
1073 	/* Save state of TEAR and DEAR register contents */
1074 	err = qsi(&si);
1075 	if (!err) {
1076 		/* TEAR/DEAR values are valid only if the sampling facility is
1077 		 * enabled.  Note that cpumsf_pmu_disable() might be called even
1078 		 * for a disabled sampling facility because cpumsf_pmu_enable()
1079 		 * controls the enable/disable state.
1080 		 */
1081 		if (si.es) {
1082 			cpuhw->lsctl.tear = si.tear;
1083 			cpuhw->lsctl.dear = si.dear;
1084 		}
1085 	} else
1086 		debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n",
1087 				    __func__, err);
1088 
1089 	cpuhw->flags &= ~PMU_F_ENABLED;
1090 }
1091 
1092 /* perf_exclude_event() - Filter event
1093  * @event:	The perf event
1094  * @regs:	pt_regs structure
1095  * @sde_regs:	Sample-data-entry (sde) regs structure
1096  *
1097  * Filter perf events according to their exclude specification.
1098  *
1099  * Return non-zero if the event shall be excluded.
1100  */
1101 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
1102 			      struct perf_sf_sde_regs *sde_regs)
1103 {
1104 	if (event->attr.exclude_user && user_mode(regs))
1105 		return 1;
1106 	if (event->attr.exclude_kernel && !user_mode(regs))
1107 		return 1;
1108 	if (event->attr.exclude_guest && sde_regs->in_guest)
1109 		return 1;
1110 	if (event->attr.exclude_host && !sde_regs->in_guest)
1111 		return 1;
1112 	return 0;
1113 }
1114 
1115 /* perf_push_sample() - Push samples to perf
1116  * @event:	The perf event
1117  * @sample:	Hardware sample data
1118  *
1119  * Use the hardware sample data to create perf event sample.  The sample
1120  * is the pushed to the event subsystem and the function checks for
1121  * possible event overflows.  If an event overflow occurs, the PMU is
1122  * stopped.
1123  *
1124  * Return non-zero if an event overflow occurred.
1125  */
1126 static int perf_push_sample(struct perf_event *event,
1127 			    struct hws_basic_entry *basic)
1128 {
1129 	int overflow;
1130 	struct pt_regs regs;
1131 	struct perf_sf_sde_regs *sde_regs;
1132 	struct perf_sample_data data;
1133 
1134 	/* Setup perf sample */
1135 	perf_sample_data_init(&data, 0, event->hw.last_period);
1136 
1137 	/* Setup pt_regs to look like an CPU-measurement external interrupt
1138 	 * using the Program Request Alert code.  The regs.int_parm_long
1139 	 * field which is unused contains additional sample-data-entry related
1140 	 * indicators.
1141 	 */
1142 	memset(&regs, 0, sizeof(regs));
1143 	regs.int_code = 0x1407;
1144 	regs.int_parm = CPU_MF_INT_SF_PRA;
1145 	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;
1146 
1147 	psw_bits(regs.psw).ia	= basic->ia;
1148 	psw_bits(regs.psw).dat	= basic->T;
1149 	psw_bits(regs.psw).wait = basic->W;
1150 	psw_bits(regs.psw).pstate = basic->P;
1151 	psw_bits(regs.psw).as	= basic->AS;
1152 
1153 	/*
1154 	 * Use the hardware provided configuration level to decide if the
1155 	 * sample belongs to a guest or host. If that is not available,
1156 	 * fall back to the following heuristics:
1157 	 * A non-zero guest program parameter always indicates a guest
1158 	 * sample. Some early samples or samples from guests without
1159 	 * lpp usage would be misaccounted to the host. We use the asn
1160 	 * value as an addon heuristic to detect most of these guest samples.
1161 	 * If the value differs from 0xffff (the host value), we assume to
1162 	 * be a KVM guest.
1163 	 */
1164 	switch (basic->CL) {
1165 	case 1: /* logical partition */
1166 		sde_regs->in_guest = 0;
1167 		break;
1168 	case 2: /* virtual machine */
1169 		sde_regs->in_guest = 1;
1170 		break;
1171 	default: /* old machine, use heuristics */
1172 		if (basic->gpp || basic->prim_asn != 0xffff)
1173 			sde_regs->in_guest = 1;
1174 		break;
1175 	}
1176 
1177 	/*
1178 	 * Store the PID value from the sample-data-entry to be
1179 	 * processed and resolved by cpumsf_output_event_pid().
1180 	 */
1181 	data.tid_entry.pid = basic->hpp & LPP_PID_MASK;
1182 
1183 	overflow = 0;
1184 	if (perf_exclude_event(event, &regs, sde_regs))
1185 		goto out;
1186 	if (perf_event_overflow(event, &data, &regs)) {
1187 		overflow = 1;
1188 		event->pmu->stop(event, 0);
1189 	}
1190 	perf_event_update_userpage(event);
1191 out:
1192 	return overflow;
1193 }
1194 
1195 static void perf_event_count_update(struct perf_event *event, u64 count)
1196 {
1197 	local64_add(count, &event->count);
1198 }
1199 
1200 /* hw_collect_samples() - Walk through a sample-data-block and collect samples
1201  * @event:	The perf event
1202  * @sdbt:	Sample-data-block table
1203  * @overflow:	Event overflow counter
1204  *
1205  * Walks through a sample-data-block and collects sampling data entries that are
1206  * then pushed to the perf event subsystem.  Depending on the sampling function,
1207  * there can be either basic-sampling or combined-sampling data entries.  A
1208  * combined-sampling data entry consists of a basic- and a diagnostic-sampling
1209  * data entry.	The sampling function is determined by the flags in the perf
1210  * event hardware structure.  The function always works with a combined-sampling
1211  * data entry but ignores the the diagnostic portion if it is not available.
1212  *
1213  * Note that the implementation focuses on basic-sampling data entries and, if
1214  * such an entry is not valid, the entire combined-sampling data entry is
1215  * ignored.
1216  *
1217  * The overflow variables counts the number of samples that has been discarded
1218  * due to a perf event overflow.
1219  */
1220 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
1221 			       unsigned long long *overflow)
1222 {
1223 	struct hws_trailer_entry *te;
1224 	struct hws_basic_entry *sample;
1225 
1226 	te = trailer_entry_ptr((unsigned long)sdbt);
1227 	sample = (struct hws_basic_entry *)sdbt;
1228 	while ((unsigned long *) sample < (unsigned long *) te) {
1229 		/* Check for an empty sample */
1230 		if (!sample->def || sample->LS)
1231 			break;
1232 
1233 		/* Update perf event period */
1234 		perf_event_count_update(event, SAMPL_RATE(&event->hw));
1235 
1236 		/* Check whether sample is valid */
1237 		if (sample->def == 0x0001) {
1238 			/* If an event overflow occurred, the PMU is stopped to
1239 			 * throttle event delivery.  Remaining sample data is
1240 			 * discarded.
1241 			 */
1242 			if (!*overflow) {
1243 				/* Check whether sample is consistent */
1244 				if (sample->I == 0 && sample->W == 0) {
1245 					/* Deliver sample data to perf */
1246 					*overflow = perf_push_sample(event,
1247 								     sample);
1248 				}
1249 			} else
1250 				/* Count discarded samples */
1251 				*overflow += 1;
1252 		} else {
1253 			debug_sprintf_event(sfdbg, 4,
1254 					    "%s: Found unknown"
1255 					    " sampling data entry: te->f %i"
1256 					    " basic.def %#4x (%p)\n", __func__,
1257 					    te->header.f, sample->def, sample);
1258 			/* Sample slot is not yet written or other record.
1259 			 *
1260 			 * This condition can occur if the buffer was reused
1261 			 * from a combined basic- and diagnostic-sampling.
1262 			 * If only basic-sampling is then active, entries are
1263 			 * written into the larger diagnostic entries.
1264 			 * This is typically the case for sample-data-blocks
1265 			 * that are not full.  Stop processing if the first
1266 			 * invalid format was detected.
1267 			 */
1268 			if (!te->header.f)
1269 				break;
1270 		}
1271 
1272 		/* Reset sample slot and advance to next sample */
1273 		sample->def = 0;
1274 		sample++;
1275 	}
1276 }
1277 
1278 static inline __uint128_t __cdsg(__uint128_t *ptr, __uint128_t old, __uint128_t new)
1279 {
1280 	asm volatile(
1281 		"	cdsg	%[old],%[new],%[ptr]\n"
1282 		: [old] "+d" (old), [ptr] "+QS" (*ptr)
1283 		: [new] "d" (new)
1284 		: "memory", "cc");
1285 	return old;
1286 }
1287 
1288 /* hw_perf_event_update() - Process sampling buffer
1289  * @event:	The perf event
1290  * @flush_all:	Flag to also flush partially filled sample-data-blocks
1291  *
1292  * Processes the sampling buffer and create perf event samples.
1293  * The sampling buffer position are retrieved and saved in the TEAR_REG
1294  * register of the specified perf event.
1295  *
1296  * Only full sample-data-blocks are processed.	Specify the flash_all flag
1297  * to also walk through partially filled sample-data-blocks.  It is ignored
1298  * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
1299  * enforces the processing of full sample-data-blocks only (trailer entries
1300  * with the block-full-indicator bit set).
1301  */
1302 static void hw_perf_event_update(struct perf_event *event, int flush_all)
1303 {
1304 	unsigned long long event_overflow, sampl_overflow, num_sdb;
1305 	union hws_trailer_header old, prev, new;
1306 	struct hw_perf_event *hwc = &event->hw;
1307 	struct hws_trailer_entry *te;
1308 	unsigned long *sdbt, sdb;
1309 	int done;
1310 
1311 	/*
1312 	 * AUX buffer is used when in diagnostic sampling mode.
1313 	 * No perf events/samples are created.
1314 	 */
1315 	if (SAMPL_DIAG_MODE(&event->hw))
1316 		return;
1317 
1318 	if (flush_all && SDB_FULL_BLOCKS(hwc))
1319 		flush_all = 0;
1320 
1321 	sdbt = (unsigned long *) TEAR_REG(hwc);
1322 	done = event_overflow = sampl_overflow = num_sdb = 0;
1323 	while (!done) {
1324 		/* Get the trailer entry of the sample-data-block */
1325 		sdb = (unsigned long)phys_to_virt(*sdbt);
1326 		te = trailer_entry_ptr(sdb);
1327 
1328 		/* Leave loop if no more work to do (block full indicator) */
1329 		if (!te->header.f) {
1330 			done = 1;
1331 			if (!flush_all)
1332 				break;
1333 		}
1334 
1335 		/* Check the sample overflow count */
1336 		if (te->header.overflow)
1337 			/* Account sample overflows and, if a particular limit
1338 			 * is reached, extend the sampling buffer.
1339 			 * For details, see sfb_account_overflows().
1340 			 */
1341 			sampl_overflow += te->header.overflow;
1342 
1343 		/* Timestamps are valid for full sample-data-blocks only */
1344 		debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx/%#lx "
1345 				    "overflow %llu timestamp %#llx\n",
1346 				    __func__, sdb, (unsigned long)sdbt,
1347 				    te->header.overflow,
1348 				    (te->header.f) ? trailer_timestamp(te) : 0ULL);
1349 
1350 		/* Collect all samples from a single sample-data-block and
1351 		 * flag if an (perf) event overflow happened.  If so, the PMU
1352 		 * is stopped and remaining samples will be discarded.
1353 		 */
1354 		hw_collect_samples(event, (unsigned long *)sdb, &event_overflow);
1355 		num_sdb++;
1356 
1357 		/* Reset trailer (using compare-double-and-swap) */
1358 		prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1359 		do {
1360 			old.val = prev.val;
1361 			new.val = prev.val;
1362 			new.f = 0;
1363 			new.a = 1;
1364 			new.overflow = 0;
1365 			prev.val = __cdsg(&te->header.val, old.val, new.val);
1366 		} while (prev.val != old.val);
1367 
1368 		/* Advance to next sample-data-block */
1369 		sdbt++;
1370 		if (is_link_entry(sdbt))
1371 			sdbt = get_next_sdbt(sdbt);
1372 
1373 		/* Update event hardware registers */
1374 		TEAR_REG(hwc) = (unsigned long) sdbt;
1375 
1376 		/* Stop processing sample-data if all samples of the current
1377 		 * sample-data-block were flushed even if it was not full.
1378 		 */
1379 		if (flush_all && done)
1380 			break;
1381 	}
1382 
1383 	/* Account sample overflows in the event hardware structure */
1384 	if (sampl_overflow)
1385 		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
1386 						 sampl_overflow, 1 + num_sdb);
1387 
1388 	/* Perf_event_overflow() and perf_event_account_interrupt() limit
1389 	 * the interrupt rate to an upper limit. Roughly 1000 samples per
1390 	 * task tick.
1391 	 * Hitting this limit results in a large number
1392 	 * of throttled REF_REPORT_THROTTLE entries and the samples
1393 	 * are dropped.
1394 	 * Slightly increase the interval to avoid hitting this limit.
1395 	 */
1396 	if (event_overflow) {
1397 		SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10);
1398 		debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n",
1399 				    __func__,
1400 				    DIV_ROUND_UP(SAMPL_RATE(hwc), 10));
1401 	}
1402 
1403 	if (sampl_overflow || event_overflow)
1404 		debug_sprintf_event(sfdbg, 4, "%s: "
1405 				    "overflows: sample %llu event %llu"
1406 				    " total %llu num_sdb %llu\n",
1407 				    __func__, sampl_overflow, event_overflow,
1408 				    OVERFLOW_REG(hwc), num_sdb);
1409 }
1410 
1411 static inline unsigned long aux_sdb_index(struct aux_buffer *aux,
1412 					  unsigned long i)
1413 {
1414 	return i % aux->sfb.num_sdb;
1415 }
1416 
1417 static inline unsigned long aux_sdb_num(unsigned long start, unsigned long end)
1418 {
1419 	return end >= start ? end - start + 1 : 0;
1420 }
1421 
1422 static inline unsigned long aux_sdb_num_alert(struct aux_buffer *aux)
1423 {
1424 	return aux_sdb_num(aux->head, aux->alert_mark);
1425 }
1426 
1427 static inline unsigned long aux_sdb_num_empty(struct aux_buffer *aux)
1428 {
1429 	return aux_sdb_num(aux->head, aux->empty_mark);
1430 }
1431 
1432 /*
1433  * Get trailer entry by index of SDB.
1434  */
1435 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux,
1436 						 unsigned long index)
1437 {
1438 	unsigned long sdb;
1439 
1440 	index = aux_sdb_index(aux, index);
1441 	sdb = aux->sdb_index[index];
1442 	return trailer_entry_ptr(sdb);
1443 }
1444 
1445 /*
1446  * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu
1447  * disabled. Collect the full SDBs in AUX buffer which have not reached
1448  * the point of alert indicator. And ignore the SDBs which are not
1449  * full.
1450  *
1451  * 1. Scan SDBs to see how much data is there and consume them.
1452  * 2. Remove alert indicator in the buffer.
1453  */
1454 static void aux_output_end(struct perf_output_handle *handle)
1455 {
1456 	unsigned long i, range_scan, idx;
1457 	struct aux_buffer *aux;
1458 	struct hws_trailer_entry *te;
1459 
1460 	aux = perf_get_aux(handle);
1461 	if (!aux)
1462 		return;
1463 
1464 	range_scan = aux_sdb_num_alert(aux);
1465 	for (i = 0, idx = aux->head; i < range_scan; i++, idx++) {
1466 		te = aux_sdb_trailer(aux, idx);
1467 		if (!te->header.f)
1468 			break;
1469 	}
1470 	/* i is num of SDBs which are full */
1471 	perf_aux_output_end(handle, i << PAGE_SHIFT);
1472 
1473 	/* Remove alert indicators in the buffer */
1474 	te = aux_sdb_trailer(aux, aux->alert_mark);
1475 	te->header.a = 0;
1476 
1477 	debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n",
1478 			    __func__, i, range_scan, aux->head);
1479 }
1480 
1481 /*
1482  * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event
1483  * is first added to the CPU or rescheduled again to the CPU. It is called
1484  * with pmu disabled.
1485  *
1486  * 1. Reset the trailer of SDBs to get ready for new data.
1487  * 2. Tell the hardware where to put the data by reset the SDBs buffer
1488  *    head(tear/dear).
1489  */
1490 static int aux_output_begin(struct perf_output_handle *handle,
1491 			    struct aux_buffer *aux,
1492 			    struct cpu_hw_sf *cpuhw)
1493 {
1494 	unsigned long range, i, range_scan, idx, head, base, offset;
1495 	struct hws_trailer_entry *te;
1496 
1497 	if (WARN_ON_ONCE(handle->head & ~PAGE_MASK))
1498 		return -EINVAL;
1499 
1500 	aux->head = handle->head >> PAGE_SHIFT;
1501 	range = (handle->size + 1) >> PAGE_SHIFT;
1502 	if (range <= 1)
1503 		return -ENOMEM;
1504 
1505 	/*
1506 	 * SDBs between aux->head and aux->empty_mark are already ready
1507 	 * for new data. range_scan is num of SDBs not within them.
1508 	 */
1509 	debug_sprintf_event(sfdbg, 6,
1510 			    "%s: range %ld head %ld alert %ld empty %ld\n",
1511 			    __func__, range, aux->head, aux->alert_mark,
1512 			    aux->empty_mark);
1513 	if (range > aux_sdb_num_empty(aux)) {
1514 		range_scan = range - aux_sdb_num_empty(aux);
1515 		idx = aux->empty_mark + 1;
1516 		for (i = 0; i < range_scan; i++, idx++) {
1517 			te = aux_sdb_trailer(aux, idx);
1518 			te->header.f = 0;
1519 			te->header.a = 0;
1520 			te->header.overflow = 0;
1521 		}
1522 		/* Save the position of empty SDBs */
1523 		aux->empty_mark = aux->head + range - 1;
1524 	}
1525 
1526 	/* Set alert indicator */
1527 	aux->alert_mark = aux->head + range/2 - 1;
1528 	te = aux_sdb_trailer(aux, aux->alert_mark);
1529 	te->header.a = 1;
1530 
1531 	/* Reset hardware buffer head */
1532 	head = aux_sdb_index(aux, aux->head);
1533 	base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE];
1534 	offset = head % CPUM_SF_SDB_PER_TABLE;
1535 	cpuhw->lsctl.tear = virt_to_phys((void *)base) + offset * sizeof(unsigned long);
1536 	cpuhw->lsctl.dear = virt_to_phys((void *)aux->sdb_index[head]);
1537 
1538 	debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld "
1539 			    "index %ld tear %#lx dear %#lx\n", __func__,
1540 			    aux->head, aux->alert_mark, aux->empty_mark,
1541 			    head / CPUM_SF_SDB_PER_TABLE,
1542 			    cpuhw->lsctl.tear, cpuhw->lsctl.dear);
1543 
1544 	return 0;
1545 }
1546 
1547 /*
1548  * Set alert indicator on SDB at index @alert_index while sampler is running.
1549  *
1550  * Return true if successfully.
1551  * Return false if full indicator is already set by hardware sampler.
1552  */
1553 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index,
1554 			  unsigned long long *overflow)
1555 {
1556 	union hws_trailer_header old, prev, new;
1557 	struct hws_trailer_entry *te;
1558 
1559 	te = aux_sdb_trailer(aux, alert_index);
1560 	prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1561 	do {
1562 		old.val = prev.val;
1563 		new.val = prev.val;
1564 		*overflow = old.overflow;
1565 		if (old.f) {
1566 			/*
1567 			 * SDB is already set by hardware.
1568 			 * Abort and try to set somewhere
1569 			 * behind.
1570 			 */
1571 			return false;
1572 		}
1573 		new.a = 1;
1574 		new.overflow = 0;
1575 		prev.val = __cdsg(&te->header.val, old.val, new.val);
1576 	} while (prev.val != old.val);
1577 	return true;
1578 }
1579 
1580 /*
1581  * aux_reset_buffer() - Scan and setup SDBs for new samples
1582  * @aux:	The AUX buffer to set
1583  * @range:	The range of SDBs to scan started from aux->head
1584  * @overflow:	Set to overflow count
1585  *
1586  * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is
1587  * marked as empty, check if it is already set full by the hardware sampler.
1588  * If yes, that means new data is already there before we can set an alert
1589  * indicator. Caller should try to set alert indicator to some position behind.
1590  *
1591  * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used
1592  * previously and have already been consumed by user space. Reset these SDBs
1593  * (clear full indicator and alert indicator) for new data.
1594  * If aux->alert_mark fall in this area, just set it. Overflow count is
1595  * recorded while scanning.
1596  *
1597  * SDBs between aux->head and aux->empty_mark are already reset at last time.
1598  * and ready for new samples. So scanning on this area could be skipped.
1599  *
1600  * Return true if alert indicator is set successfully and false if not.
1601  */
1602 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range,
1603 			     unsigned long long *overflow)
1604 {
1605 	unsigned long i, range_scan, idx, idx_old;
1606 	union hws_trailer_header old, prev, new;
1607 	unsigned long long orig_overflow;
1608 	struct hws_trailer_entry *te;
1609 
1610 	debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld "
1611 			    "empty %ld\n", __func__, range, aux->head,
1612 			    aux->alert_mark, aux->empty_mark);
1613 	if (range <= aux_sdb_num_empty(aux))
1614 		/*
1615 		 * No need to scan. All SDBs in range are marked as empty.
1616 		 * Just set alert indicator. Should check race with hardware
1617 		 * sampler.
1618 		 */
1619 		return aux_set_alert(aux, aux->alert_mark, overflow);
1620 
1621 	if (aux->alert_mark <= aux->empty_mark)
1622 		/*
1623 		 * Set alert indicator on empty SDB. Should check race
1624 		 * with hardware sampler.
1625 		 */
1626 		if (!aux_set_alert(aux, aux->alert_mark, overflow))
1627 			return false;
1628 
1629 	/*
1630 	 * Scan the SDBs to clear full and alert indicator used previously.
1631 	 * Start scanning from one SDB behind empty_mark. If the new alert
1632 	 * indicator fall into this range, set it.
1633 	 */
1634 	range_scan = range - aux_sdb_num_empty(aux);
1635 	idx_old = idx = aux->empty_mark + 1;
1636 	for (i = 0; i < range_scan; i++, idx++) {
1637 		te = aux_sdb_trailer(aux, idx);
1638 		prev.val = READ_ONCE_ALIGNED_128(te->header.val);
1639 		do {
1640 			old.val = prev.val;
1641 			new.val = prev.val;
1642 			orig_overflow = old.overflow;
1643 			new.f = 0;
1644 			new.overflow = 0;
1645 			if (idx == aux->alert_mark)
1646 				new.a = 1;
1647 			else
1648 				new.a = 0;
1649 			prev.val = __cdsg(&te->header.val, old.val, new.val);
1650 		} while (prev.val != old.val);
1651 		*overflow += orig_overflow;
1652 	}
1653 
1654 	/* Update empty_mark to new position */
1655 	aux->empty_mark = aux->head + range - 1;
1656 
1657 	debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld "
1658 			    "empty %ld\n", __func__, range_scan, idx_old,
1659 			    idx - 1, aux->empty_mark);
1660 	return true;
1661 }
1662 
1663 /*
1664  * Measurement alert handler for diagnostic mode sampling.
1665  */
1666 static void hw_collect_aux(struct cpu_hw_sf *cpuhw)
1667 {
1668 	struct aux_buffer *aux;
1669 	int done = 0;
1670 	unsigned long range = 0, size;
1671 	unsigned long long overflow = 0;
1672 	struct perf_output_handle *handle = &cpuhw->handle;
1673 	unsigned long num_sdb;
1674 
1675 	aux = perf_get_aux(handle);
1676 	if (WARN_ON_ONCE(!aux))
1677 		return;
1678 
1679 	/* Inform user space new data arrived */
1680 	size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
1681 	debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__,
1682 			    size >> PAGE_SHIFT);
1683 	perf_aux_output_end(handle, size);
1684 
1685 	num_sdb = aux->sfb.num_sdb;
1686 	while (!done) {
1687 		/* Get an output handle */
1688 		aux = perf_aux_output_begin(handle, cpuhw->event);
1689 		if (handle->size == 0) {
1690 			pr_err("The AUX buffer with %lu pages for the "
1691 			       "diagnostic-sampling mode is full\n",
1692 				num_sdb);
1693 			debug_sprintf_event(sfdbg, 1,
1694 					    "%s: AUX buffer used up\n",
1695 					    __func__);
1696 			break;
1697 		}
1698 		if (WARN_ON_ONCE(!aux))
1699 			return;
1700 
1701 		/* Update head and alert_mark to new position */
1702 		aux->head = handle->head >> PAGE_SHIFT;
1703 		range = (handle->size + 1) >> PAGE_SHIFT;
1704 		if (range == 1)
1705 			aux->alert_mark = aux->head;
1706 		else
1707 			aux->alert_mark = aux->head + range/2 - 1;
1708 
1709 		if (aux_reset_buffer(aux, range, &overflow)) {
1710 			if (!overflow) {
1711 				done = 1;
1712 				break;
1713 			}
1714 			size = range << PAGE_SHIFT;
1715 			perf_aux_output_end(&cpuhw->handle, size);
1716 			pr_err("Sample data caused the AUX buffer with %lu "
1717 			       "pages to overflow\n", aux->sfb.num_sdb);
1718 			debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld "
1719 					    "overflow %lld\n", __func__,
1720 					    aux->head, range, overflow);
1721 		} else {
1722 			size = aux_sdb_num_alert(aux) << PAGE_SHIFT;
1723 			perf_aux_output_end(&cpuhw->handle, size);
1724 			debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1725 					    "already full, try another\n",
1726 					    __func__,
1727 					    aux->head, aux->alert_mark);
1728 		}
1729 	}
1730 
1731 	if (done)
1732 		debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld "
1733 				    "empty %ld\n", __func__, aux->head,
1734 				    aux->alert_mark, aux->empty_mark);
1735 }
1736 
1737 /*
1738  * Callback when freeing AUX buffers.
1739  */
1740 static void aux_buffer_free(void *data)
1741 {
1742 	struct aux_buffer *aux = data;
1743 	unsigned long i, num_sdbt;
1744 
1745 	if (!aux)
1746 		return;
1747 
1748 	/* Free SDBT. SDB is freed by the caller */
1749 	num_sdbt = aux->sfb.num_sdbt;
1750 	for (i = 0; i < num_sdbt; i++)
1751 		free_page(aux->sdbt_index[i]);
1752 
1753 	kfree(aux->sdbt_index);
1754 	kfree(aux->sdb_index);
1755 	kfree(aux);
1756 
1757 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt);
1758 }
1759 
1760 static void aux_sdb_init(unsigned long sdb)
1761 {
1762 	struct hws_trailer_entry *te;
1763 
1764 	te = trailer_entry_ptr(sdb);
1765 
1766 	/* Save clock base */
1767 	te->clock_base = 1;
1768 	te->progusage2 = tod_clock_base.tod;
1769 }
1770 
1771 /*
1772  * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
1773  * @event:	Event the buffer is setup for, event->cpu == -1 means current
1774  * @pages:	Array of pointers to buffer pages passed from perf core
1775  * @nr_pages:	Total pages
1776  * @snapshot:	Flag for snapshot mode
1777  *
1778  * This is the callback when setup an event using AUX buffer. Perf tool can
1779  * trigger this by an additional mmap() call on the event. Unlike the buffer
1780  * for basic samples, AUX buffer belongs to the event. It is scheduled with
1781  * the task among online cpus when it is a per-thread event.
1782  *
1783  * Return the private AUX buffer structure if success or NULL if fails.
1784  */
1785 static void *aux_buffer_setup(struct perf_event *event, void **pages,
1786 			      int nr_pages, bool snapshot)
1787 {
1788 	struct sf_buffer *sfb;
1789 	struct aux_buffer *aux;
1790 	unsigned long *new, *tail;
1791 	int i, n_sdbt;
1792 
1793 	if (!nr_pages || !pages)
1794 		return NULL;
1795 
1796 	if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1797 		pr_err("AUX buffer size (%i pages) is larger than the "
1798 		       "maximum sampling buffer limit\n",
1799 		       nr_pages);
1800 		return NULL;
1801 	} else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
1802 		pr_err("AUX buffer size (%i pages) is less than the "
1803 		       "minimum sampling buffer limit\n",
1804 		       nr_pages);
1805 		return NULL;
1806 	}
1807 
1808 	/* Allocate aux_buffer struct for the event */
1809 	aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL);
1810 	if (!aux)
1811 		goto no_aux;
1812 	sfb = &aux->sfb;
1813 
1814 	/* Allocate sdbt_index for fast reference */
1815 	n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE);
1816 	aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
1817 	if (!aux->sdbt_index)
1818 		goto no_sdbt_index;
1819 
1820 	/* Allocate sdb_index for fast reference */
1821 	aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
1822 	if (!aux->sdb_index)
1823 		goto no_sdb_index;
1824 
1825 	/* Allocate the first SDBT */
1826 	sfb->num_sdbt = 0;
1827 	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1828 	if (!sfb->sdbt)
1829 		goto no_sdbt;
1830 	aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
1831 	tail = sfb->tail = sfb->sdbt;
1832 
1833 	/*
1834 	 * Link the provided pages of AUX buffer to SDBT.
1835 	 * Allocate SDBT if needed.
1836 	 */
1837 	for (i = 0; i < nr_pages; i++, tail++) {
1838 		if (require_table_link(tail)) {
1839 			new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
1840 			if (!new)
1841 				goto no_sdbt;
1842 			aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
1843 			/* Link current page to tail of chain */
1844 			*tail = virt_to_phys(new) + 1;
1845 			tail = new;
1846 		}
1847 		/* Tail is the entry in a SDBT */
1848 		*tail = virt_to_phys(pages[i]);
1849 		aux->sdb_index[i] = (unsigned long)pages[i];
1850 		aux_sdb_init((unsigned long)pages[i]);
1851 	}
1852 	sfb->num_sdb = nr_pages;
1853 
1854 	/* Link the last entry in the SDBT to the first SDBT */
1855 	*tail = virt_to_phys(sfb->sdbt) + 1;
1856 	sfb->tail = tail;
1857 
1858 	/*
1859 	 * Initial all SDBs are zeroed. Mark it as empty.
1860 	 * So there is no need to clear the full indicator
1861 	 * when this event is first added.
1862 	 */
1863 	aux->empty_mark = sfb->num_sdb - 1;
1864 
1865 	debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__,
1866 			    sfb->num_sdbt, sfb->num_sdb);
1867 
1868 	return aux;
1869 
1870 no_sdbt:
1871 	/* SDBs (AUX buffer pages) are freed by caller */
1872 	for (i = 0; i < sfb->num_sdbt; i++)
1873 		free_page(aux->sdbt_index[i]);
1874 	kfree(aux->sdb_index);
1875 no_sdb_index:
1876 	kfree(aux->sdbt_index);
1877 no_sdbt_index:
1878 	kfree(aux);
1879 no_aux:
1880 	return NULL;
1881 }
1882 
1883 static void cpumsf_pmu_read(struct perf_event *event)
1884 {
1885 	/* Nothing to do ... updates are interrupt-driven */
1886 }
1887 
1888 /* Check if the new sampling period/freqeuncy is appropriate.
1889  *
1890  * Return non-zero on error and zero on passed checks.
1891  */
1892 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value)
1893 {
1894 	struct hws_qsi_info_block si;
1895 	unsigned long rate;
1896 	bool do_freq;
1897 
1898 	memset(&si, 0, sizeof(si));
1899 	if (event->cpu == -1) {
1900 		if (qsi(&si))
1901 			return -ENODEV;
1902 	} else {
1903 		/* Event is pinned to a particular CPU, retrieve the per-CPU
1904 		 * sampling structure for accessing the CPU-specific QSI.
1905 		 */
1906 		struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
1907 
1908 		si = cpuhw->qsi;
1909 	}
1910 
1911 	do_freq = !!SAMPLE_FREQ_MODE(&event->hw);
1912 	rate = getrate(do_freq, value, &si);
1913 	if (!rate)
1914 		return -EINVAL;
1915 
1916 	event->attr.sample_period = rate;
1917 	SAMPL_RATE(&event->hw) = rate;
1918 	hw_init_period(&event->hw, SAMPL_RATE(&event->hw));
1919 	debug_sprintf_event(sfdbg, 4, "%s:"
1920 			    " cpu %d value %#llx period %#llx freq %d\n",
1921 			    __func__, event->cpu, value,
1922 			    event->attr.sample_period, do_freq);
1923 	return 0;
1924 }
1925 
1926 /* Activate sampling control.
1927  * Next call of pmu_enable() starts sampling.
1928  */
1929 static void cpumsf_pmu_start(struct perf_event *event, int flags)
1930 {
1931 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1932 
1933 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1934 		return;
1935 
1936 	if (flags & PERF_EF_RELOAD)
1937 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1938 
1939 	perf_pmu_disable(event->pmu);
1940 	event->hw.state = 0;
1941 	cpuhw->lsctl.cs = 1;
1942 	if (SAMPL_DIAG_MODE(&event->hw))
1943 		cpuhw->lsctl.cd = 1;
1944 	perf_pmu_enable(event->pmu);
1945 }
1946 
1947 /* Deactivate sampling control.
1948  * Next call of pmu_enable() stops sampling.
1949  */
1950 static void cpumsf_pmu_stop(struct perf_event *event, int flags)
1951 {
1952 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1953 
1954 	if (event->hw.state & PERF_HES_STOPPED)
1955 		return;
1956 
1957 	perf_pmu_disable(event->pmu);
1958 	cpuhw->lsctl.cs = 0;
1959 	cpuhw->lsctl.cd = 0;
1960 	event->hw.state |= PERF_HES_STOPPED;
1961 
1962 	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
1963 		hw_perf_event_update(event, 1);
1964 		event->hw.state |= PERF_HES_UPTODATE;
1965 	}
1966 	perf_pmu_enable(event->pmu);
1967 }
1968 
1969 static int cpumsf_pmu_add(struct perf_event *event, int flags)
1970 {
1971 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1972 	struct aux_buffer *aux;
1973 	int err;
1974 
1975 	if (cpuhw->flags & PMU_F_IN_USE)
1976 		return -EAGAIN;
1977 
1978 	if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt)
1979 		return -EINVAL;
1980 
1981 	err = 0;
1982 	perf_pmu_disable(event->pmu);
1983 
1984 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1985 
1986 	/* Set up sampling controls.  Always program the sampling register
1987 	 * using the SDB-table start.  Reset TEAR_REG event hardware register
1988 	 * that is used by hw_perf_event_update() to store the sampling buffer
1989 	 * position after samples have been flushed.
1990 	 */
1991 	cpuhw->lsctl.s = 0;
1992 	cpuhw->lsctl.h = 1;
1993 	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
1994 	if (!SAMPL_DIAG_MODE(&event->hw)) {
1995 		cpuhw->lsctl.tear = virt_to_phys(cpuhw->sfb.sdbt);
1996 		cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
1997 		TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt;
1998 	}
1999 
2000 	/* Ensure sampling functions are in the disabled state.  If disabled,
2001 	 * switch on sampling enable control. */
2002 	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
2003 		err = -EAGAIN;
2004 		goto out;
2005 	}
2006 	if (SAMPL_DIAG_MODE(&event->hw)) {
2007 		aux = perf_aux_output_begin(&cpuhw->handle, event);
2008 		if (!aux) {
2009 			err = -EINVAL;
2010 			goto out;
2011 		}
2012 		err = aux_output_begin(&cpuhw->handle, aux, cpuhw);
2013 		if (err)
2014 			goto out;
2015 		cpuhw->lsctl.ed = 1;
2016 	}
2017 	cpuhw->lsctl.es = 1;
2018 
2019 	/* Set in_use flag and store event */
2020 	cpuhw->event = event;
2021 	cpuhw->flags |= PMU_F_IN_USE;
2022 
2023 	if (flags & PERF_EF_START)
2024 		cpumsf_pmu_start(event, PERF_EF_RELOAD);
2025 out:
2026 	perf_event_update_userpage(event);
2027 	perf_pmu_enable(event->pmu);
2028 	return err;
2029 }
2030 
2031 static void cpumsf_pmu_del(struct perf_event *event, int flags)
2032 {
2033 	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
2034 
2035 	perf_pmu_disable(event->pmu);
2036 	cpumsf_pmu_stop(event, PERF_EF_UPDATE);
2037 
2038 	cpuhw->lsctl.es = 0;
2039 	cpuhw->lsctl.ed = 0;
2040 	cpuhw->flags &= ~PMU_F_IN_USE;
2041 	cpuhw->event = NULL;
2042 
2043 	if (SAMPL_DIAG_MODE(&event->hw))
2044 		aux_output_end(&cpuhw->handle);
2045 	perf_event_update_userpage(event);
2046 	perf_pmu_enable(event->pmu);
2047 }
2048 
2049 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
2050 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
2051 
2052 /* Attribute list for CPU_SF.
2053  *
2054  * The availablitiy depends on the CPU_MF sampling facility authorization
2055  * for basic + diagnositic samples. This is determined at initialization
2056  * time by the sampling facility device driver.
2057  * If the authorization for basic samples is turned off, it should be
2058  * also turned off for diagnostic sampling.
2059  *
2060  * During initialization of the device driver, check the authorization
2061  * level for diagnostic sampling and installs the attribute
2062  * file for diagnostic sampling if necessary.
2063  *
2064  * For now install a placeholder to reference all possible attributes:
2065  * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG.
2066  * Add another entry for the final NULL pointer.
2067  */
2068 enum {
2069 	SF_CYCLES_BASIC_ATTR_IDX = 0,
2070 	SF_CYCLES_BASIC_DIAG_ATTR_IDX,
2071 	SF_CYCLES_ATTR_MAX
2072 };
2073 
2074 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = {
2075 	[SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC)
2076 };
2077 
2078 PMU_FORMAT_ATTR(event, "config:0-63");
2079 
2080 static struct attribute *cpumsf_pmu_format_attr[] = {
2081 	&format_attr_event.attr,
2082 	NULL,
2083 };
2084 
2085 static struct attribute_group cpumsf_pmu_events_group = {
2086 	.name = "events",
2087 	.attrs = cpumsf_pmu_events_attr,
2088 };
2089 
2090 static struct attribute_group cpumsf_pmu_format_group = {
2091 	.name = "format",
2092 	.attrs = cpumsf_pmu_format_attr,
2093 };
2094 
2095 static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
2096 	&cpumsf_pmu_events_group,
2097 	&cpumsf_pmu_format_group,
2098 	NULL,
2099 };
2100 
2101 static struct pmu cpumf_sampling = {
2102 	.pmu_enable   = cpumsf_pmu_enable,
2103 	.pmu_disable  = cpumsf_pmu_disable,
2104 
2105 	.event_init   = cpumsf_pmu_event_init,
2106 	.add	      = cpumsf_pmu_add,
2107 	.del	      = cpumsf_pmu_del,
2108 
2109 	.start	      = cpumsf_pmu_start,
2110 	.stop	      = cpumsf_pmu_stop,
2111 	.read	      = cpumsf_pmu_read,
2112 
2113 	.attr_groups  = cpumsf_pmu_attr_groups,
2114 
2115 	.setup_aux    = aux_buffer_setup,
2116 	.free_aux     = aux_buffer_free,
2117 
2118 	.check_period = cpumsf_pmu_check_period,
2119 };
2120 
2121 static void cpumf_measurement_alert(struct ext_code ext_code,
2122 				    unsigned int alert, unsigned long unused)
2123 {
2124 	struct cpu_hw_sf *cpuhw;
2125 
2126 	if (!(alert & CPU_MF_INT_SF_MASK))
2127 		return;
2128 	inc_irq_stat(IRQEXT_CMS);
2129 	cpuhw = this_cpu_ptr(&cpu_hw_sf);
2130 
2131 	/* Measurement alerts are shared and might happen when the PMU
2132 	 * is not reserved.  Ignore these alerts in this case. */
2133 	if (!(cpuhw->flags & PMU_F_RESERVED))
2134 		return;
2135 
2136 	/* The processing below must take care of multiple alert events that
2137 	 * might be indicated concurrently. */
2138 
2139 	/* Program alert request */
2140 	if (alert & CPU_MF_INT_SF_PRA) {
2141 		if (cpuhw->flags & PMU_F_IN_USE)
2142 			if (SAMPL_DIAG_MODE(&cpuhw->event->hw))
2143 				hw_collect_aux(cpuhw);
2144 			else
2145 				hw_perf_event_update(cpuhw->event, 0);
2146 		else
2147 			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
2148 	}
2149 
2150 	/* Report measurement alerts only for non-PRA codes */
2151 	if (alert != CPU_MF_INT_SF_PRA)
2152 		debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__,
2153 				    alert);
2154 
2155 	/* Sampling authorization change request */
2156 	if (alert & CPU_MF_INT_SF_SACA)
2157 		qsi(&cpuhw->qsi);
2158 
2159 	/* Loss of sample data due to high-priority machine activities */
2160 	if (alert & CPU_MF_INT_SF_LSDA) {
2161 		pr_err("Sample data was lost\n");
2162 		cpuhw->flags |= PMU_F_ERR_LSDA;
2163 		sf_disable();
2164 	}
2165 
2166 	/* Invalid sampling buffer entry */
2167 	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
2168 		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
2169 		       alert);
2170 		cpuhw->flags |= PMU_F_ERR_IBE;
2171 		sf_disable();
2172 	}
2173 }
2174 
2175 static int cpusf_pmu_setup(unsigned int cpu, int flags)
2176 {
2177 	/* Ignore the notification if no events are scheduled on the PMU.
2178 	 * This might be racy...
2179 	 */
2180 	if (!atomic_read(&num_events))
2181 		return 0;
2182 
2183 	local_irq_disable();
2184 	setup_pmc_cpu(&flags);
2185 	local_irq_enable();
2186 	return 0;
2187 }
2188 
2189 static int s390_pmu_sf_online_cpu(unsigned int cpu)
2190 {
2191 	return cpusf_pmu_setup(cpu, PMC_INIT);
2192 }
2193 
2194 static int s390_pmu_sf_offline_cpu(unsigned int cpu)
2195 {
2196 	return cpusf_pmu_setup(cpu, PMC_RELEASE);
2197 }
2198 
2199 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
2200 {
2201 	if (!cpum_sf_avail())
2202 		return -ENODEV;
2203 	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2204 }
2205 
2206 static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
2207 {
2208 	int rc;
2209 	unsigned long min, max;
2210 
2211 	if (!cpum_sf_avail())
2212 		return -ENODEV;
2213 	if (!val || !strlen(val))
2214 		return -EINVAL;
2215 
2216 	/* Valid parameter values: "min,max" or "max" */
2217 	min = CPUM_SF_MIN_SDB;
2218 	max = CPUM_SF_MAX_SDB;
2219 	if (strchr(val, ','))
2220 		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
2221 	else
2222 		rc = kstrtoul(val, 10, &max);
2223 
2224 	if (min < 2 || min >= max || max > get_num_physpages())
2225 		rc = -EINVAL;
2226 	if (rc)
2227 		return rc;
2228 
2229 	sfb_set_limits(min, max);
2230 	pr_info("The sampling buffer limits have changed to: "
2231 		"min %lu max %lu (diag %lu)\n",
2232 		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
2233 	return 0;
2234 }
2235 
2236 #define param_check_sfb_size(name, p) __param_check(name, p, void)
2237 static const struct kernel_param_ops param_ops_sfb_size = {
2238 	.set = param_set_sfb_size,
2239 	.get = param_get_sfb_size,
2240 };
2241 
2242 #define RS_INIT_FAILURE_QSI	  0x0001
2243 #define RS_INIT_FAILURE_BSDES	  0x0002
2244 #define RS_INIT_FAILURE_ALRT	  0x0003
2245 #define RS_INIT_FAILURE_PERF	  0x0004
2246 static void __init pr_cpumsf_err(unsigned int reason)
2247 {
2248 	pr_err("Sampling facility support for perf is not available: "
2249 	       "reason %#x\n", reason);
2250 }
2251 
2252 static int __init init_cpum_sampling_pmu(void)
2253 {
2254 	struct hws_qsi_info_block si;
2255 	int err;
2256 
2257 	if (!cpum_sf_avail())
2258 		return -ENODEV;
2259 
2260 	memset(&si, 0, sizeof(si));
2261 	if (qsi(&si)) {
2262 		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
2263 		return -ENODEV;
2264 	}
2265 
2266 	if (!si.as && !si.ad)
2267 		return -ENODEV;
2268 
2269 	if (si.bsdes != sizeof(struct hws_basic_entry)) {
2270 		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
2271 		return -EINVAL;
2272 	}
2273 
2274 	if (si.ad) {
2275 		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
2276 		/* Sampling of diagnostic data authorized,
2277 		 * install event into attribute list of PMU device.
2278 		 */
2279 		cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] =
2280 			CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG);
2281 	}
2282 
2283 	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
2284 	if (!sfdbg) {
2285 		pr_err("Registering for s390dbf failed\n");
2286 		return -ENOMEM;
2287 	}
2288 	debug_register_view(sfdbg, &debug_sprintf_view);
2289 
2290 	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
2291 				    cpumf_measurement_alert);
2292 	if (err) {
2293 		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
2294 		debug_unregister(sfdbg);
2295 		goto out;
2296 	}
2297 
2298 	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
2299 	if (err) {
2300 		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
2301 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
2302 					cpumf_measurement_alert);
2303 		debug_unregister(sfdbg);
2304 		goto out;
2305 	}
2306 
2307 	cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online",
2308 			  s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu);
2309 out:
2310 	return err;
2311 }
2312 
2313 arch_initcall(init_cpum_sampling_pmu);
2314 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644);
2315