perf_cpum_sf.c 45.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/*
 * Performance event support for the System z CPU-measurement Sampling Facility
 *
 * Copyright IBM Corp. 2013
 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License (version 2 only)
 * as published by the Free Software Foundation.
 */
#define KMSG_COMPONENT	"cpum_sf"
#define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt

#include <linux/kernel.h>
#include <linux/kernel_stat.h>
#include <linux/perf_event.h>
#include <linux/percpu.h>
#include <linux/notifier.h>
#include <linux/export.h>
20
#include <linux/slab.h>
21 22
#include <linux/mm.h>
#include <linux/moduleparam.h>
23 24 25 26 27 28 29 30 31
#include <asm/cpu_mf.h>
#include <asm/irq.h>
#include <asm/debug.h>
#include <asm/timex.h>

/* Minimum number of sample-data-block-tables:
 * At least one table is required for the sampling buffer structure.
 * A single table contains up to 511 pointers to sample-data-blocks.
 */
32
#define CPUM_SF_MIN_SDBT	1
33

34
/* Number of sample-data-blocks per sample-data-block-table (SDBT):
35 36
 * A table contains SDB pointers (8 bytes) and one table-link entry
 * that points to the origin of the next SDBT.
37
 */
38
#define CPUM_SF_SDB_PER_TABLE	((PAGE_SIZE - 8) / 8)
39

40 41 42 43 44 45 46 47 48 49 50 51
/* Maximum page offset for an SDBT table-link entry:
 * If this page offset is reached, a table-link entry to the next SDBT
 * must be added.
 */
#define CPUM_SF_SDBT_TL_OFFSET	(CPUM_SF_SDB_PER_TABLE * 8)
static inline int require_table_link(const void *sdbt)
{
	return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET;
}

/* Minimum and maximum sampling buffer sizes:
 *
52 53 54 55 56
 * This number represents the maximum size of the sampling buffer taking
 * the number of sample-data-block-tables into account.  Note that these
 * numbers apply to the basic-sampling function only.
 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if
 * the diagnostic-sampling function is active.
57
 *
58 59 60 61 62 63 64 65 66
 * Sampling buffer size		Buffer characteristics
 * ---------------------------------------------------
 *	 64KB		    ==	  16 pages (4KB per page)
 *				   1 page  for SDB-tables
 *				  15 pages for SDBs
 *
 *  32MB		    ==	8192 pages (4KB per page)
 *				  16 pages for SDB-tables
 *				8176 pages for SDBs
67
 */
68 69
static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15;
static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176;
70
static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1;
71 72

struct sf_buffer {
73
	unsigned long	 *sdbt;	    /* Sample-data-block-table origin */
74
	/* buffer characteristics (required for buffer increments) */
75 76 77
	unsigned long  num_sdb;	    /* Number of sample-data-blocks */
	unsigned long num_sdbt;	    /* Number of sample-data-block-tables */
	unsigned long	 *tail;	    /* last sample-data-block-table */
78 79 80 81 82
};

struct cpu_hw_sf {
	/* CPU-measurement sampling information block */
	struct hws_qsi_info_block qsi;
83
	/* CPU-measurement sampling control block */
84 85 86 87 88 89 90 91 92 93
	struct hws_lsctl_request_block lsctl;
	struct sf_buffer sfb;	    /* Sampling buffer */
	unsigned int flags;	    /* Status flags */
	struct perf_event *event;   /* Scheduled perf event */
};
static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf);

/* Debug feature */
static debug_info_t *sfdbg;

94 95 96 97 98 99 100 101 102 103 104
/*
 * sf_disable() - Switch off sampling facility
 */
static int sf_disable(void)
{
	struct hws_lsctl_request_block sreq;

	memset(&sreq, 0, sizeof(sreq));
	return lsctl(&sreq);
}

105 106 107 108 109
/*
 * sf_buffer_available() - Check for an allocated sampling buffer
 */
static int sf_buffer_available(struct cpu_hw_sf *cpuhw)
{
110
	return !!cpuhw->sfb.sdbt;
111 112 113 114 115 116 117
}

/*
 * deallocate sampling facility buffer
 */
static void free_sampling_buffer(struct sf_buffer *sfb)
{
118
	unsigned long *sdbt, *curr;
119 120 121 122 123

	if (!sfb->sdbt)
		return;

	sdbt = sfb->sdbt;
124
	curr = sdbt;
125

126
	/* Free the SDBT after all SDBs are processed... */
127 128 129 130
	while (1) {
		if (!*curr || !sdbt)
			break;

131
		/* Process table-link entries */
132 133 134
		if (is_link_entry(curr)) {
			curr = get_next_sdbt(curr);
			if (sdbt)
135
				free_page((unsigned long) sdbt);
136

137 138
			/* If the origin is reached, sampling buffer is freed */
			if (curr == sfb->sdbt)
139 140
				break;
			else
141
				sdbt = curr;
142
		} else {
143
			/* Process SDB pointer */
144 145 146 147 148 149 150 151
			if (*curr) {
				free_page(*curr);
				curr++;
			}
		}
	}

	debug_sprintf_event(sfdbg, 5,
152
			    "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt);
153 154 155
	memset(sfb, 0, sizeof(*sfb));
}

156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags)
{
	unsigned long sdb, *trailer;

	/* Allocate and initialize sample-data-block */
	sdb = get_zeroed_page(gfp_flags);
	if (!sdb)
		return -ENOMEM;
	trailer = trailer_entry_ptr(sdb);
	*trailer = SDB_TE_ALERT_REQ_MASK;

	/* Link SDB into the sample-data-block-table */
	*sdbt = sdb;

	return 0;
}

/*
 * realloc_sampling_buffer() - extend sampler memory
 *
 * Allocates new sample-data-blocks and adds them to the specified sampling
 * buffer memory.
 *
 * Important: This modifies the sampling buffer and must be called when the
 *	      sampling facility is disabled.
 *
 * Returns zero on success, non-zero otherwise.
 */
static int realloc_sampling_buffer(struct sf_buffer *sfb,
				   unsigned long num_sdb, gfp_t gfp_flags)
{
	int i, rc;
	unsigned long *new, *tail;

	if (!sfb->sdbt || !sfb->tail)
		return -EINVAL;

	if (!is_link_entry(sfb->tail))
		return -EINVAL;

	/* Append to the existing sampling buffer, overwriting the table-link
	 * register.
	 * The tail variables always points to the "tail" (last and table-link)
	 * entry in an SDB-table.
	 */
	tail = sfb->tail;

	/* Do a sanity check whether the table-link entry points to
	 * the sampling buffer origin.
	 */
	if (sfb->sdbt != get_next_sdbt(tail)) {
		debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: "
				    "sampling buffer is not linked: origin=%p"
				    "tail=%p\n",
				    (void *) sfb->sdbt, (void *) tail);
		return -EINVAL;
	}

	/* Allocate remaining SDBs */
	rc = 0;
	for (i = 0; i < num_sdb; i++) {
		/* Allocate a new SDB-table if it is full. */
		if (require_table_link(tail)) {
			new = (unsigned long *) get_zeroed_page(gfp_flags);
			if (!new) {
				rc = -ENOMEM;
				break;
			}
			sfb->num_sdbt++;
			/* Link current page to tail of chain */
			*tail = (unsigned long)(void *) new + 1;
			tail = new;
		}

		/* Allocate a new sample-data-block.
		 * If there is not enough memory, stop the realloc process
		 * and simply use what was allocated.  If this is a temporary
		 * issue, a new realloc call (if required) might succeed.
		 */
		rc = alloc_sample_data_block(tail, gfp_flags);
		if (rc)
			break;
		sfb->num_sdb++;
		tail++;
	}

	/* Link sampling buffer to its origin */
	*tail = (unsigned long) sfb->sdbt + 1;
	sfb->tail = tail;

	debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer"
			    " settings: sdbt=%lu sdb=%lu\n",
			    sfb->num_sdbt, sfb->num_sdb);
	return rc;
}

252 253 254 255 256 257 258 259 260 261 262 263 264
/*
 * allocate_sampling_buffer() - allocate sampler memory
 *
 * Allocates and initializes a sampling buffer structure using the
 * specified number of sample-data-blocks (SDB).  For each allocation,
 * a 4K page is used.  The number of sample-data-block-tables (SDBT)
 * are calculated from SDBs.
 * Also set the ALERT_REQ mask in each SDBs trailer.
 *
 * Returns zero on success, non-zero otherwise.
 */
static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb)
{
265
	int rc;
266 267 268

	if (sfb->sdbt)
		return -EINVAL;
269 270 271 272 273

	/* Allocate the sample-data-block-table origin */
	sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
	if (!sfb->sdbt)
		return -ENOMEM;
274
	sfb->num_sdb = 0;
275
	sfb->num_sdbt = 1;
276

277 278 279 280 281
	/* Link the table origin to point to itself to prepare for
	 * realloc_sampling_buffer() invocation.
	 */
	sfb->tail = sfb->sdbt;
	*sfb->tail = (unsigned long)(void *) sfb->sdbt + 1;
282

283 284 285 286 287 288 289 290 291 292 293 294
	/* Allocate requested number of sample-data-blocks */
	rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL);
	if (rc) {
		free_sampling_buffer(sfb);
		debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: "
			"realloc_sampling_buffer failed with rc=%i\n", rc);
	} else
		debug_sprintf_event(sfdbg, 4,
			"alloc_sampling_buffer: tear=%p dear=%p\n",
			sfb->sdbt, (void *) *sfb->sdbt);
	return rc;
}
295

296 297
static void sfb_set_limits(unsigned long min, unsigned long max)
{
298 299
	struct hws_qsi_info_block si;

300 301
	CPUM_SF_MIN_SDB = min;
	CPUM_SF_MAX_SDB = max;
302 303 304 305 306 307 308 309 310 311

	memset(&si, 0, sizeof(si));
	if (!qsi(&si))
		CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes);
}

static unsigned long sfb_max_limit(struct hw_perf_event *hwc)
{
	return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR
				    : CPUM_SF_MAX_SDB;
312
}
313

314 315 316 317 318 319 320 321 322
static unsigned long sfb_pending_allocs(struct sf_buffer *sfb,
					struct hw_perf_event *hwc)
{
	if (!sfb->sdbt)
		return SFB_ALLOC_REG(hwc);
	if (SFB_ALLOC_REG(hwc) > sfb->num_sdb)
		return SFB_ALLOC_REG(hwc) - sfb->num_sdb;
	return 0;
}
323

324 325 326 327 328
static int sfb_has_pending_allocs(struct sf_buffer *sfb,
				   struct hw_perf_event *hwc)
{
	return sfb_pending_allocs(sfb, hwc) > 0;
}
329

330 331
static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc)
{
332 333
	/* Limit the number of SDBs to not exceed the maximum */
	num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc));
334 335
	if (num)
		SFB_ALLOC_REG(hwc) += num;
336 337
}

338 339 340 341 342 343
static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc)
{
	SFB_ALLOC_REG(hwc) = 0;
	sfb_account_allocs(num, hwc);
}

344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366
static size_t event_sample_size(struct hw_perf_event *hwc)
{
	struct sf_raw_sample *sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(hwc);
	size_t sample_size;

	/* The sample size depends on the sampling function: The basic-sampling
	 * function must be always enabled, diagnostic-sampling function is
	 * optional.
	 */
	sample_size = sfr->bsdes;
	if (SAMPL_DIAG_MODE(hwc))
		sample_size += sfr->dsdes;

	return sample_size;
}

static void deallocate_buffers(struct cpu_hw_sf *cpuhw)
{
	if (cpuhw->sfb.sdbt)
		free_sampling_buffer(&cpuhw->sfb);
}

static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc)
367
{
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
	unsigned long n_sdb, freq, factor;
	size_t sfr_size, sample_size;
	struct sf_raw_sample *sfr;

	/* Allocate raw sample buffer
	 *
	 *    The raw sample buffer is used to temporarily store sampling data
	 *    entries for perf raw sample processing.  The buffer size mainly
	 *    depends on the size of diagnostic-sampling data entries which is
	 *    machine-specific.  The exact size calculation includes:
	 *	1. The first 4 bytes of diagnostic-sampling data entries are
	 *	   already reflected in the sf_raw_sample structure.  Subtract
	 *	   these bytes.
	 *	2. The perf raw sample data must be 8-byte aligned (u64) and
	 *	   perf's internal data size must be considered too.  So add
	 *	   an additional u32 for correct alignment and subtract before
	 *	   allocating the buffer.
	 *	3. Store the raw sample buffer pointer in the perf event
	 *	   hardware structure.
	 */
	sfr_size = ALIGN((sizeof(*sfr) - sizeof(sfr->diag) + cpuhw->qsi.dsdes) +
			 sizeof(u32), sizeof(u64));
	sfr_size -= sizeof(u32);
	sfr = kzalloc(sfr_size, GFP_KERNEL);
	if (!sfr)
		return -ENOMEM;
	sfr->size = sfr_size;
	sfr->bsdes = cpuhw->qsi.bsdes;
	sfr->dsdes = cpuhw->qsi.dsdes;
	RAWSAMPLE_REG(hwc) = (unsigned long) sfr;
398 399 400

	/* Calculate sampling buffers using 4K pages
	 *
401 402 403 404 405 406 407
	 *    1. Determine the sample data size which depends on the used
	 *	 sampling functions, for example, basic-sampling or
	 *	 basic-sampling with diagnostic-sampling.
	 *
	 *    2. Use the sampling frequency as input.  The sampling buffer is
	 *	 designed for almost one second.  This can be adjusted through
	 *	 the "factor" variable.
408
	 *	 In any case, alloc_sampling_buffer() sets the Alert Request
409
	 *	 Control indicator to trigger a measurement-alert to harvest
410 411
	 *	 sample-data-blocks (sdb).
	 *
412
	 *    3. Compute the number of sample-data-blocks and ensure a minimum
413
	 *	 of CPUM_SF_MIN_SDB.  Also ensure the upper limit does not
414 415 416 417
	 *	 exceed a "calculated" maximum.  The symbolic maximum is
	 *	 designed for basic-sampling only and needs to be increased if
	 *	 diagnostic-sampling is active.
	 *	 See also the remarks for these symbolic constants.
418
	 *
419 420 421
	 *    4. Compute the number of sample-data-block-tables (SDBT) and
	 *	 ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up
	 *	 to 511 SDBs).
422
	 */
423
	sample_size = event_sample_size(hwc);
424 425
	freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc));
	factor = 1;
426
	n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size));
427 428 429
	if (n_sdb < CPUM_SF_MIN_SDB)
		n_sdb = CPUM_SF_MIN_SDB;

430 431 432 433 434 435 436
	/* If there is already a sampling buffer allocated, it is very likely
	 * that the sampling facility is enabled too.  If the event to be
	 * initialized requires a greater sampling buffer, the allocation must
	 * be postponed.  Changing the sampling buffer requires the sampling
	 * facility to be in the disabled state.  So, account the number of
	 * required SDBs and let cpumsf_pmu_enable() resize the buffer just
	 * before the event is started.
437
	 */
438
	sfb_init_allocs(n_sdb, hwc);
439 440 441 442
	if (sf_buffer_available(cpuhw))
		return 0;

	debug_sprintf_event(sfdbg, 3,
443 444 445 446
			    "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu"
			    " sample_size=%lu cpuhw=%p\n",
			    SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc),
			    sample_size, cpuhw);
447 448

	return alloc_sampling_buffer(&cpuhw->sfb,
449
				     sfb_pending_allocs(&cpuhw->sfb, hwc));
450 451
}

452 453 454 455 456
static unsigned long min_percent(unsigned int percent, unsigned long base,
				 unsigned long min)
{
	return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100));
}
457

458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479
static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base)
{
	/* Use a percentage-based approach to extend the sampling facility
	 * buffer.  Accept up to 5% sample data loss.
	 * Vary the extents between 1% to 5% of the current number of
	 * sample-data-blocks.
	 */
	if (ratio <= 5)
		return 0;
	if (ratio <= 25)
		return min_percent(1, base, 1);
	if (ratio <= 50)
		return min_percent(1, base, 1);
	if (ratio <= 75)
		return min_percent(2, base, 2);
	if (ratio <= 100)
		return min_percent(3, base, 3);
	if (ratio <= 250)
		return min_percent(4, base, 4);

	return min_percent(5, base, 8);
}
480

481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
static void sfb_account_overflows(struct cpu_hw_sf *cpuhw,
				  struct hw_perf_event *hwc)
{
	unsigned long ratio, num;

	if (!OVERFLOW_REG(hwc))
		return;

	/* The sample_overflow contains the average number of sample data
	 * that has been lost because sample-data-blocks were full.
	 *
	 * Calculate the total number of sample data entries that has been
	 * discarded.  Then calculate the ratio of lost samples to total samples
	 * per second in percent.
	 */
	ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb,
			     sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)));

	/* Compute number of sample-data-blocks */
	num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb);
	if (num)
		sfb_account_allocs(num, hwc);

	debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu"
			    " num=%lu\n", OVERFLOW_REG(hwc), ratio, num);
	OVERFLOW_REG(hwc) = 0;
}

/* extend_sampling_buffer() - Extend sampling buffer
 * @sfb:	Sampling buffer structure (for local CPU)
 * @hwc:	Perf event hardware structure
 *
 * Use this function to extend the sampling buffer based on the overflow counter
 * and postponed allocation extents stored in the specified Perf event hardware.
 *
 * Important: This function disables the sampling facility in order to safely
 *	      change the sampling buffer structure.  Do not call this function
 *	      when the PMU is active.
519
 */
520 521
static void extend_sampling_buffer(struct sf_buffer *sfb,
				   struct hw_perf_event *hwc)
522
{
523 524
	unsigned long num, num_old;
	int rc;
525

526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550
	num = sfb_pending_allocs(sfb, hwc);
	if (!num)
		return;
	num_old = sfb->num_sdb;

	/* Disable the sampling facility to reset any states and also
	 * clear pending measurement alerts.
	 */
	sf_disable();

	/* Extend the sampling buffer.
	 * This memory allocation typically happens in an atomic context when
	 * called by perf.  Because this is a reallocation, it is fine if the
	 * new SDB-request cannot be satisfied immediately.
	 */
	rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC);
	if (rc)
		debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc "
				    "failed with rc=%i\n", rc);

	if (sfb_has_pending_allocs(sfb, hwc))
		debug_sprintf_event(sfdbg, 5, "sfb: extend: "
				    "req=%lu alloc=%lu remaining=%lu\n",
				    num, sfb->num_sdb - num_old,
				    sfb_pending_allocs(sfb, hwc));
551 552 553
}


554 555 556 557 558
/* Number of perf events counting hardware events */
static atomic_t num_events;
/* Used to avoid races in calling reserve/release_cpumf_hardware */
static DEFINE_MUTEX(pmc_reserve_mutex);

559 560
#define PMC_INIT      0
#define PMC_RELEASE   1
561
#define PMC_FAILURE   2
562 563 564
static void setup_pmc_cpu(void *flags)
{
	int err;
565
	struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf);
566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587

	err = 0;
	switch (*((int *) flags)) {
	case PMC_INIT:
		memset(cpusf, 0, sizeof(*cpusf));
		err = qsi(&cpusf->qsi);
		if (err)
			break;
		cpusf->flags |= PMU_F_RESERVED;
		err = sf_disable();
		if (err)
			pr_err("Switching off the sampling facility failed "
			       "with rc=%i\n", err);
		debug_sprintf_event(sfdbg, 5,
				    "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf);
		break;
	case PMC_RELEASE:
		cpusf->flags &= ~PMU_F_RESERVED;
		err = sf_disable();
		if (err) {
			pr_err("Switching off the sampling facility failed "
			       "with rc=%i\n", err);
588 589
		} else
			deallocate_buffers(cpusf);
590 591 592 593
		debug_sprintf_event(sfdbg, 5,
				    "setup_pmc_cpu: released: cpuhw=%p\n", cpusf);
		break;
	}
594 595
	if (err)
		*((int *) flags) |= PMC_FAILURE;
596 597 598 599 600 601 602 603
}

static void release_pmc_hardware(void)
{
	int flags = PMC_RELEASE;

	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
	on_each_cpu(setup_pmc_cpu, &flags, 1);
604
	perf_release_sampling();
605 606 607 608 609
}

static int reserve_pmc_hardware(void)
{
	int flags = PMC_INIT;
610
	int err;
611

612 613 614
	err = perf_reserve_sampling();
	if (err)
		return err;
615
	on_each_cpu(setup_pmc_cpu, &flags, 1);
616 617 618 619
	if (flags & PMC_FAILURE) {
		release_pmc_hardware();
		return -ENODEV;
	}
620 621 622 623 624 625 626
	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);

	return 0;
}

static void hw_perf_event_destroy(struct perf_event *event)
{
627 628 629 630
	/* Free raw sample buffer */
	if (RAWSAMPLE_REG(&event->hw))
		kfree((void *) RAWSAMPLE_REG(&event->hw));

631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
	/* Release PMC if this is the last perf event */
	if (!atomic_add_unless(&num_events, -1, 1)) {
		mutex_lock(&pmc_reserve_mutex);
		if (atomic_dec_return(&num_events) == 0)
			release_pmc_hardware();
		mutex_unlock(&pmc_reserve_mutex);
	}
}

static void hw_init_period(struct hw_perf_event *hwc, u64 period)
{
	hwc->sample_period = period;
	hwc->last_period = hwc->sample_period;
	local64_set(&hwc->period_left, hwc->sample_period);
}

static void hw_reset_registers(struct hw_perf_event *hwc,
648
			       unsigned long *sdbt_origin)
649
{
650 651
	struct sf_raw_sample *sfr;

652 653
	/* (Re)set to first sample-data-block-table */
	TEAR_REG(hwc) = (unsigned long) sdbt_origin;
654 655 656 657 658

	/* (Re)set raw sampling buffer register */
	sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(hwc);
	memset(&sfr->basic, 0, sizeof(sfr->basic));
	memset(&sfr->diag, 0, sfr->dsdes);
659 660 661 662 663
}

static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si,
				   unsigned long rate)
{
664 665
	return clamp_t(unsigned long, rate,
		       si->min_sampl_rate, si->max_sampl_rate);
666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721
}

static int __hw_perf_event_init(struct perf_event *event)
{
	struct cpu_hw_sf *cpuhw;
	struct hws_qsi_info_block si;
	struct perf_event_attr *attr = &event->attr;
	struct hw_perf_event *hwc = &event->hw;
	unsigned long rate;
	int cpu, err;

	/* Reserve CPU-measurement sampling facility */
	err = 0;
	if (!atomic_inc_not_zero(&num_events)) {
		mutex_lock(&pmc_reserve_mutex);
		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
			err = -EBUSY;
		else
			atomic_inc(&num_events);
		mutex_unlock(&pmc_reserve_mutex);
	}
	event->destroy = hw_perf_event_destroy;

	if (err)
		goto out;

	/* Access per-CPU sampling information (query sampling info) */
	/*
	 * The event->cpu value can be -1 to count on every CPU, for example,
	 * when attaching to a task.  If this is specified, use the query
	 * sampling info from the current CPU, otherwise use event->cpu to
	 * retrieve the per-CPU information.
	 * Later, cpuhw indicates whether to allocate sampling buffers for a
	 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL).
	 */
	memset(&si, 0, sizeof(si));
	cpuhw = NULL;
	if (event->cpu == -1)
		qsi(&si);
	else {
		/* Event is pinned to a particular CPU, retrieve the per-CPU
		 * sampling structure for accessing the CPU-specific QSI.
		 */
		cpuhw = &per_cpu(cpu_hw_sf, event->cpu);
		si = cpuhw->qsi;
	}

	/* Check sampling facility authorization and, if not authorized,
	 * fall back to other PMUs.  It is safe to check any CPU because
	 * the authorization is identical for all configured CPUs.
	 */
	if (!si.as) {
		err = -ENOENT;
		goto out;
	}

722 723 724 725 726 727 728 729 730 731 732 733 734 735
	/* Always enable basic sampling */
	SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE;

	/* Check if diagnostic sampling is requested.  Deny if the required
	 * sampling authorization is missing.
	 */
	if (attr->config == PERF_EVENT_CPUM_SF_DIAG) {
		if (!si.ad) {
			err = -EPERM;
			goto out;
		}
		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE;
	}

736 737 738 739
	/* Check and set other sampling flags */
	if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS)
		SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS;

740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
	/* The sampling information (si) contains information about the
	 * min/max sampling intervals and the CPU speed.  So calculate the
	 * correct sampling interval and avoid the whole period adjust
	 * feedback loop.
	 */
	rate = 0;
	if (attr->freq) {
		rate = freq_to_sample_rate(&si, attr->sample_freq);
		rate = hw_limit_rate(&si, rate);
		attr->freq = 0;
		attr->sample_period = rate;
	} else {
		/* The min/max sampling rates specifies the valid range
		 * of sample periods.  If the specified sample period is
		 * out of range, limit the period to the range boundary.
		 */
		rate = hw_limit_rate(&si, hwc->sample_period);

		/* The perf core maintains a maximum sample rate that is
		 * configurable through the sysctl interface.  Ensure the
		 * sampling rate does not exceed this value.  This also helps
		 * to avoid throttling when pushing samples with
		 * perf_event_overflow().
		 */
		if (sample_rate_to_freq(&si, rate) >
		      sysctl_perf_event_sample_rate) {
			err = -EINVAL;
			debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n");
			goto out;
		}
	}
	SAMPL_RATE(hwc) = rate;
	hw_init_period(hwc, SAMPL_RATE(hwc));

774 775 776 777
	/* Initialize sample data overflow accounting */
	hwc->extra_reg.reg = REG_OVERFLOW;
	OVERFLOW_REG(hwc) = 0;

778 779 780 781 782 783 784
	/* Allocate the per-CPU sampling buffer using the CPU information
	 * from the event.  If the event is not pinned to a particular
	 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling
	 * buffers for each online CPU.
	 */
	if (cpuhw)
		/* Event is pinned to a particular CPU */
785
		err = allocate_buffers(cpuhw, hwc);
786 787 788 789 790 791
	else {
		/* Event is not pinned, allocate sampling buffer on
		 * each online CPU
		 */
		for_each_online_cpu(cpu) {
			cpuhw = &per_cpu(cpu_hw_sf, cpu);
792
			err = allocate_buffers(cpuhw, hwc);
793 794 795 796 797 798 799 800 801 802 803 804
			if (err)
				break;
		}
	}
out:
	return err;
}

static int cpumsf_pmu_event_init(struct perf_event *event)
{
	int err;

805 806 807 808 809 810
	/* No support for taken branch sampling */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

	switch (event->attr.type) {
	case PERF_TYPE_RAW:
811 812
		if ((event->attr.config != PERF_EVENT_CPUM_SF) &&
		    (event->attr.config != PERF_EVENT_CPUM_SF_DIAG))
813 814 815 816 817 818 819 820 821 822 823 824 825 826
			return -ENOENT;
		break;
	case PERF_TYPE_HARDWARE:
		/* Support sampling of CPU cycles in addition to the
		 * counter facility.  However, the counter facility
		 * is more precise and, hence, restrict this PMU to
		 * sampling events only.
		 */
		if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES)
			return -ENOENT;
		if (!is_sampling_event(event))
			return -ENOENT;
		break;
	default:
827
		return -ENOENT;
828
	}
829

830
	/* Check online status of the CPU to which the event is pinned */
831 832 833 834
	if (event->cpu >= nr_cpumask_bits ||
	    (event->cpu >= 0 && !cpu_online(event->cpu)))
		return -ENODEV;

835 836 837 838 839 840 841 842
	/* Force reset of idle/hv excludes regardless of what the
	 * user requested.
	 */
	if (event->attr.exclude_hv)
		event->attr.exclude_hv = 0;
	if (event->attr.exclude_idle)
		event->attr.exclude_idle = 0;

843 844 845 846 847 848 849 850 851
	err = __hw_perf_event_init(event);
	if (unlikely(err))
		if (event->destroy)
			event->destroy(event);
	return err;
}

static void cpumsf_pmu_enable(struct pmu *pmu)
{
852
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
853
	struct hw_perf_event *hwc;
854 855 856 857 858 859 860 861
	int err;

	if (cpuhw->flags & PMU_F_ENABLED)
		return;

	if (cpuhw->flags & PMU_F_ERR_MASK)
		return;

862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881
	/* Check whether to extent the sampling buffer.
	 *
	 * Two conditions trigger an increase of the sampling buffer for a
	 * perf event:
	 *    1. Postponed buffer allocations from the event initialization.
	 *    2. Sampling overflows that contribute to pending allocations.
	 *
	 * Note that the extend_sampling_buffer() function disables the sampling
	 * facility, but it can be fully re-enabled using sampling controls that
	 * have been saved in cpumsf_pmu_disable().
	 */
	if (cpuhw->event) {
		hwc = &cpuhw->event->hw;
		/* Account number of overflow-designated buffer extents */
		sfb_account_overflows(cpuhw, hwc);
		if (sfb_has_pending_allocs(&cpuhw->sfb, hwc))
			extend_sampling_buffer(&cpuhw->sfb, hwc);
	}

	/* (Re)enable the PMU and sampling facility */
882 883 884 885 886 887 888 889 890 891 892
	cpuhw->flags |= PMU_F_ENABLED;
	barrier();

	err = lsctl(&cpuhw->lsctl);
	if (err) {
		cpuhw->flags &= ~PMU_F_ENABLED;
		pr_err("Loading sampling controls failed: op=%i err=%i\n",
			1, err);
		return;
	}

893 894 895
	debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i "
			    "tear=%p dear=%p\n", cpuhw->lsctl.es, cpuhw->lsctl.cs,
			    cpuhw->lsctl.ed, cpuhw->lsctl.cd,
896 897 898 899 900
			    (void *) cpuhw->lsctl.tear, (void *) cpuhw->lsctl.dear);
}

static void cpumsf_pmu_disable(struct pmu *pmu)
{
901
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
902 903 904 905 906 907 908 909 910 911 912 913 914
	struct hws_lsctl_request_block inactive;
	struct hws_qsi_info_block si;
	int err;

	if (!(cpuhw->flags & PMU_F_ENABLED))
		return;

	if (cpuhw->flags & PMU_F_ERR_MASK)
		return;

	/* Switch off sampling activation control */
	inactive = cpuhw->lsctl;
	inactive.cs = 0;
915
	inactive.cd = 0;
916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941

	err = lsctl(&inactive);
	if (err) {
		pr_err("Loading sampling controls failed: op=%i err=%i\n",
			2, err);
		return;
	}

	/* Save state of TEAR and DEAR register contents */
	if (!qsi(&si)) {
		/* TEAR/DEAR values are valid only if the sampling facility is
		 * enabled.  Note that cpumsf_pmu_disable() might be called even
		 * for a disabled sampling facility because cpumsf_pmu_enable()
		 * controls the enable/disable state.
		 */
		if (si.es) {
			cpuhw->lsctl.tear = si.tear;
			cpuhw->lsctl.dear = si.dear;
		}
	} else
		debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: "
				    "qsi() failed with err=%i\n", err);

	cpuhw->flags &= ~PMU_F_ENABLED;
}

942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964
/* perf_exclude_event() - Filter event
 * @event:	The perf event
 * @regs:	pt_regs structure
 * @sde_regs:	Sample-data-entry (sde) regs structure
 *
 * Filter perf events according to their exclude specification.
 *
 * Return non-zero if the event shall be excluded.
 */
static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs,
			      struct perf_sf_sde_regs *sde_regs)
{
	if (event->attr.exclude_user && user_mode(regs))
		return 1;
	if (event->attr.exclude_kernel && !user_mode(regs))
		return 1;
	if (event->attr.exclude_guest && sde_regs->in_guest)
		return 1;
	if (event->attr.exclude_host && !sde_regs->in_guest)
		return 1;
	return 0;
}

965 966 967 968 969 970 971 972 973 974 975
/* perf_push_sample() - Push samples to perf
 * @event:	The perf event
 * @sample:	Hardware sample data
 *
 * Use the hardware sample data to create perf event sample.  The sample
 * is the pushed to the event subsystem and the function checks for
 * possible event overflows.  If an event overflow occurs, the PMU is
 * stopped.
 *
 * Return non-zero if an event overflow occurred.
 */
976
static int perf_push_sample(struct perf_event *event, struct sf_raw_sample *sfr)
977 978 979
{
	int overflow;
	struct pt_regs regs;
980
	struct perf_sf_sde_regs *sde_regs;
981
	struct perf_sample_data data;
982
	struct perf_raw_record raw;
983

984
	/* Setup perf sample */
985
	perf_sample_data_init(&data, 0, event->hw.last_period);
986 987 988
	raw.size = sfr->size;
	raw.data = sfr;
	data.raw = &raw;
989

990 991 992 993 994
	/* Setup pt_regs to look like an CPU-measurement external interrupt
	 * using the Program Request Alert code.  The regs.int_parm_long
	 * field which is unused contains additional sample-data-entry related
	 * indicators.
	 */
995
	memset(&regs, 0, sizeof(regs));
996 997 998 999
	regs.int_code = 0x1407;
	regs.int_parm = CPU_MF_INT_SF_PRA;
	sde_regs = (struct perf_sf_sde_regs *) &regs.int_parm_long;

1000 1001
	regs.psw.addr = sfr->basic.ia;
	if (sfr->basic.T)
1002
		regs.psw.mask |= PSW_MASK_DAT;
1003
	if (sfr->basic.W)
1004
		regs.psw.mask |= PSW_MASK_WAIT;
1005
	if (sfr->basic.P)
1006
		regs.psw.mask |= PSW_MASK_PSTATE;
1007
	switch (sfr->basic.AS) {
1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
	case 0x0:
		regs.psw.mask |= PSW_ASC_PRIMARY;
		break;
	case 0x1:
		regs.psw.mask |= PSW_ASC_ACCREG;
		break;
	case 0x2:
		regs.psw.mask |= PSW_ASC_SECONDARY;
		break;
	case 0x3:
		regs.psw.mask |= PSW_ASC_HOME;
		break;
	}

1022 1023 1024 1025 1026 1027 1028
	/* The host-program-parameter (hpp) contains the sie control
	 * block that is set by sie64a() in entry64.S.	Check if hpp
	 * refers to a valid control block and set sde_regs flags
	 * accordingly.  This would allow to use hpp values for other
	 * purposes too.
	 * For now, simply use a non-zero value as guest indicator.
	 */
1029
	if (sfr->basic.hpp)
1030 1031
		sde_regs->in_guest = 1;

1032
	overflow = 0;
1033 1034
	if (perf_exclude_event(event, &regs, sde_regs))
		goto out;
1035 1036 1037 1038 1039
	if (perf_event_overflow(event, &data, &regs)) {
		overflow = 1;
		event->pmu->stop(event, 0);
	}
	perf_event_update_userpage(event);
1040
out:
1041 1042 1043 1044 1045 1046 1047 1048
	return overflow;
}

static void perf_event_count_update(struct perf_event *event, u64 count)
{
	local64_add(count, &event->count);
}

1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119
static int sample_format_is_valid(struct hws_combined_entry *sample,
				   unsigned int flags)
{
	if (likely(flags & PERF_CPUM_SF_BASIC_MODE))
		/* Only basic-sampling data entries with data-entry-format
		 * version of 0x0001 can be processed.
		 */
		if (sample->basic.def != 0x0001)
			return 0;
	if (flags & PERF_CPUM_SF_DIAG_MODE)
		/* The data-entry-format number of diagnostic-sampling data
		 * entries can vary.  Because diagnostic data is just passed
		 * through, do only a sanity check on the DEF.
		 */
		if (sample->diag.def < 0x8001)
			return 0;
	return 1;
}

static int sample_is_consistent(struct hws_combined_entry *sample,
				unsigned long flags)
{
	/* This check applies only to basic-sampling data entries of potentially
	 * combined-sampling data entries.  Invalid entries cannot be processed
	 * by the PMU and, thus, do not deliver an associated
	 * diagnostic-sampling data entry.
	 */
	if (unlikely(!(flags & PERF_CPUM_SF_BASIC_MODE)))
		return 0;
	/*
	 * Samples are skipped, if they are invalid or for which the
	 * instruction address is not predictable, i.e., the wait-state bit is
	 * set.
	 */
	if (sample->basic.I || sample->basic.W)
		return 0;
	return 1;
}

static void reset_sample_slot(struct hws_combined_entry *sample,
			      unsigned long flags)
{
	if (likely(flags & PERF_CPUM_SF_BASIC_MODE))
		sample->basic.def = 0;
	if (flags & PERF_CPUM_SF_DIAG_MODE)
		sample->diag.def = 0;
}

static void sfr_store_sample(struct sf_raw_sample *sfr,
			     struct hws_combined_entry *sample)
{
	if (likely(sfr->format & PERF_CPUM_SF_BASIC_MODE))
		sfr->basic = sample->basic;
	if (sfr->format & PERF_CPUM_SF_DIAG_MODE)
		memcpy(&sfr->diag, &sample->diag, sfr->dsdes);
}

static void debug_sample_entry(struct hws_combined_entry *sample,
			       struct hws_trailer_entry *te,
			       unsigned long flags)
{
	debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown "
			    "sampling data entry: te->f=%i basic.def=%04x (%p)"
			    " diag.def=%04x (%p)\n", te->f,
			    sample->basic.def, &sample->basic,
			    (flags & PERF_CPUM_SF_DIAG_MODE)
					? sample->diag.def : 0xFFFF,
			    (flags & PERF_CPUM_SF_DIAG_MODE)
					?  &sample->diag : NULL);
}

1120 1121 1122 1123 1124
/* hw_collect_samples() - Walk through a sample-data-block and collect samples
 * @event:	The perf event
 * @sdbt:	Sample-data-block table
 * @overflow:	Event overflow counter
 *
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
 * Walks through a sample-data-block and collects sampling data entries that are
 * then pushed to the perf event subsystem.  Depending on the sampling function,
 * there can be either basic-sampling or combined-sampling data entries.  A
 * combined-sampling data entry consists of a basic- and a diagnostic-sampling
 * data entry.	The sampling function is determined by the flags in the perf
 * event hardware structure.  The function always works with a combined-sampling
 * data entry but ignores the the diagnostic portion if it is not available.
 *
 * Note that the implementation focuses on basic-sampling data entries and, if
 * such an entry is not valid, the entire combined-sampling data entry is
 * ignored.
 *
 * The overflow variables counts the number of samples that has been discarded
 * due to a perf event overflow.
1139 1140 1141 1142
 */
static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt,
			       unsigned long long *overflow)
{
1143 1144 1145 1146 1147 1148 1149 1150 1151
	unsigned long flags = SAMPL_FLAGS(&event->hw);
	struct hws_combined_entry *sample;
	struct hws_trailer_entry *te;
	struct sf_raw_sample *sfr;
	size_t sample_size;

	/* Prepare and initialize raw sample data */
	sfr = (struct sf_raw_sample *) RAWSAMPLE_REG(&event->hw);
	sfr->format = flags & PERF_CPUM_SF_MODE_MASK;
1152

1153 1154 1155 1156
	sample_size = event_sample_size(&event->hw);
	te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);
	sample = (struct hws_combined_entry *) *sdbt;
	while ((unsigned long *) sample < (unsigned long *) te) {
1157
		/* Check for an empty sample */
1158
		if (!sample->basic.def)
1159 1160 1161 1162 1163
			break;

		/* Update perf event period */
		perf_event_count_update(event, SAMPL_RATE(&event->hw));

1164 1165
		/* Check sampling data entry */
		if (sample_format_is_valid(sample, flags)) {
1166 1167 1168 1169
			/* If an event overflow occurred, the PMU is stopped to
			 * throttle event delivery.  Remaining sample data is
			 * discarded.
			 */
1170 1171 1172 1173 1174 1175 1176
			if (!*overflow) {
				if (sample_is_consistent(sample, flags)) {
					/* Deliver sample data to perf */
					sfr_store_sample(sfr, sample);
					*overflow = perf_push_sample(event, sfr);
				}
			} else
1177 1178
				/* Count discarded samples */
				*overflow += 1;
1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193
		} else {
			debug_sample_entry(sample, te, flags);
			/* Sample slot is not yet written or other record.
			 *
			 * This condition can occur if the buffer was reused
			 * from a combined basic- and diagnostic-sampling.
			 * If only basic-sampling is then active, entries are
			 * written into the larger diagnostic entries.
			 * This is typically the case for sample-data-blocks
			 * that are not full.  Stop processing if the first
			 * invalid format was detected.
			 */
			if (!te->f)
				break;
		}
1194 1195

		/* Reset sample slot and advance to next sample */
1196 1197
		reset_sample_slot(sample, flags);
		sample += sample_size;
1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209
	}
}

/* hw_perf_event_update() - Process sampling buffer
 * @event:	The perf event
 * @flush_all:	Flag to also flush partially filled sample-data-blocks
 *
 * Processes the sampling buffer and create perf event samples.
 * The sampling buffer position are retrieved and saved in the TEAR_REG
 * register of the specified perf event.
 *
 * Only full sample-data-blocks are processed.	Specify the flash_all flag
1210 1211 1212 1213
 * to also walk through partially filled sample-data-blocks.  It is ignored
 * if PERF_CPUM_SF_FULL_BLOCKS is set.	The PERF_CPUM_SF_FULL_BLOCKS flag
 * enforces the processing of full sample-data-blocks only (trailer entries
 * with the block-full-indicator bit set).
1214 1215 1216 1217 1218 1219
 */
static void hw_perf_event_update(struct perf_event *event, int flush_all)
{
	struct hw_perf_event *hwc = &event->hw;
	struct hws_trailer_entry *te;
	unsigned long *sdbt;
1220
	unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags;
1221 1222
	int done;

1223 1224 1225
	if (flush_all && SDB_FULL_BLOCKS(hwc))
		flush_all = 0;

1226
	sdbt = (unsigned long *) TEAR_REG(hwc);
1227
	done = event_overflow = sampl_overflow = num_sdb = 0;
1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
	while (!done) {
		/* Get the trailer entry of the sample-data-block */
		te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt);

		/* Leave loop if no more work to do (block full indicator) */
		if (!te->f) {
			done = 1;
			if (!flush_all)
				break;
		}

1239 1240 1241 1242 1243
		/* Check the sample overflow count */
		if (te->overflow)
			/* Account sample overflows and, if a particular limit
			 * is reached, extend the sampling buffer.
			 * For details, see sfb_account_overflows().
1244
			 */
1245
			sampl_overflow += te->overflow;
1246 1247 1248 1249 1250

		/* Timestamps are valid for full sample-data-blocks only */
		debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p "
				    "overflow=%llu timestamp=0x%llx\n",
				    sdbt, te->overflow,
1251
				    (te->f) ? trailer_timestamp(te) : 0ULL);
1252 1253 1254 1255 1256 1257

		/* Collect all samples from a single sample-data-block and
		 * flag if an (perf) event overflow happened.  If so, the PMU
		 * is stopped and remaining samples will be discarded.
		 */
		hw_collect_samples(event, sdbt, &event_overflow);
1258
		num_sdb++;
1259

1260 1261 1262 1263 1264 1265 1266
		/* Reset trailer (using compare-double-and-swap) */
		do {
			te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK;
			te_flags |= SDB_TE_ALERT_REQ_MASK;
		} while (!cmpxchg_double(&te->flags, &te->overflow,
					 te->flags, te->overflow,
					 te_flags, 0ULL));
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288

		/* Advance to next sample-data-block */
		sdbt++;
		if (is_link_entry(sdbt))
			sdbt = get_next_sdbt(sdbt);

		/* Update event hardware registers */
		TEAR_REG(hwc) = (unsigned long) sdbt;

		/* Stop processing sample-data if all samples of the current
		 * sample-data-block were flushed even if it was not full.
		 */
		if (flush_all && done)
			break;

		/* If an event overflow happened, discard samples by
		 * processing any remaining sample-data-blocks.
		 */
		if (event_overflow)
			flush_all = 1;
	}

1289 1290 1291 1292
	/* Account sample overflows in the event hardware structure */
	if (sampl_overflow)
		OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) +
						 sampl_overflow, 1 + num_sdb);
1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308
	if (sampl_overflow || event_overflow)
		debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: "
				    "overflow stats: sample=%llu event=%llu\n",
				    sampl_overflow, event_overflow);
}

static void cpumsf_pmu_read(struct perf_event *event)
{
	/* Nothing to do ... updates are interrupt-driven */
}

/* Activate sampling control.
 * Next call of pmu_enable() starts sampling.
 */
static void cpumsf_pmu_start(struct perf_event *event, int flags)
{
1309
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1310 1311 1312 1313 1314 1315 1316 1317 1318 1319

	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
		return;

	if (flags & PERF_EF_RELOAD)
		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));

	perf_pmu_disable(event->pmu);
	event->hw.state = 0;
	cpuhw->lsctl.cs = 1;
1320 1321
	if (SAMPL_DIAG_MODE(&event->hw))
		cpuhw->lsctl.cd = 1;
1322 1323 1324 1325 1326 1327 1328 1329
	perf_pmu_enable(event->pmu);
}

/* Deactivate sampling control.
 * Next call of pmu_enable() stops sampling.
 */
static void cpumsf_pmu_stop(struct perf_event *event, int flags)
{
1330
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1331 1332 1333 1334 1335 1336

	if (event->hw.state & PERF_HES_STOPPED)
		return;

	perf_pmu_disable(event->pmu);
	cpuhw->lsctl.cs = 0;
1337
	cpuhw->lsctl.cd = 0;
1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348
	event->hw.state |= PERF_HES_STOPPED;

	if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
		hw_perf_event_update(event, 1);
		event->hw.state |= PERF_HES_UPTODATE;
	}
	perf_pmu_enable(event->pmu);
}

static int cpumsf_pmu_add(struct perf_event *event, int flags)
{
1349
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
	int err;

	if (cpuhw->flags & PMU_F_IN_USE)
		return -EAGAIN;

	if (!cpuhw->sfb.sdbt)
		return -EINVAL;

	err = 0;
	perf_pmu_disable(event->pmu);

	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;

	/* Set up sampling controls.  Always program the sampling register
	 * using the SDB-table start.  Reset TEAR_REG event hardware register
	 * that is used by hw_perf_event_update() to store the sampling buffer
	 * position after samples have been flushed.
	 */
	cpuhw->lsctl.s = 0;
	cpuhw->lsctl.h = 1;
1370
	cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt;
1371 1372 1373 1374 1375 1376
	cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt;
	cpuhw->lsctl.interval = SAMPL_RATE(&event->hw);
	hw_reset_registers(&event->hw, cpuhw->sfb.sdbt);

	/* Ensure sampling functions are in the disabled state.  If disabled,
	 * switch on sampling enable control. */
1377
	if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) {
1378 1379 1380 1381
		err = -EAGAIN;
		goto out;
	}
	cpuhw->lsctl.es = 1;
1382 1383
	if (SAMPL_DIAG_MODE(&event->hw))
		cpuhw->lsctl.ed = 1;
1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398

	/* Set in_use flag and store event */
	cpuhw->event = event;
	cpuhw->flags |= PMU_F_IN_USE;

	if (flags & PERF_EF_START)
		cpumsf_pmu_start(event, PERF_EF_RELOAD);
out:
	perf_event_update_userpage(event);
	perf_pmu_enable(event->pmu);
	return err;
}

static void cpumsf_pmu_del(struct perf_event *event, int flags)
{
1399
	struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf);
1400 1401 1402 1403 1404

	perf_pmu_disable(event->pmu);
	cpumsf_pmu_stop(event, PERF_EF_UPDATE);

	cpuhw->lsctl.es = 0;
1405
	cpuhw->lsctl.ed = 0;
1406 1407 1408 1409 1410 1411 1412 1413
	cpuhw->flags &= ~PMU_F_IN_USE;
	cpuhw->event = NULL;

	perf_event_update_userpage(event);
	perf_pmu_enable(event->pmu);
}

CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF);
1414
CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG);
1415 1416 1417

static struct attribute *cpumsf_pmu_events_attr[] = {
	CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC),
1418
	CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG),
1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465
	NULL,
};

PMU_FORMAT_ATTR(event, "config:0-63");

static struct attribute *cpumsf_pmu_format_attr[] = {
	&format_attr_event.attr,
	NULL,
};

static struct attribute_group cpumsf_pmu_events_group = {
	.name = "events",
	.attrs = cpumsf_pmu_events_attr,
};
static struct attribute_group cpumsf_pmu_format_group = {
	.name = "format",
	.attrs = cpumsf_pmu_format_attr,
};
static const struct attribute_group *cpumsf_pmu_attr_groups[] = {
	&cpumsf_pmu_events_group,
	&cpumsf_pmu_format_group,
	NULL,
};

static struct pmu cpumf_sampling = {
	.pmu_enable   = cpumsf_pmu_enable,
	.pmu_disable  = cpumsf_pmu_disable,

	.event_init   = cpumsf_pmu_event_init,
	.add	      = cpumsf_pmu_add,
	.del	      = cpumsf_pmu_del,

	.start	      = cpumsf_pmu_start,
	.stop	      = cpumsf_pmu_stop,
	.read	      = cpumsf_pmu_read,

	.attr_groups  = cpumsf_pmu_attr_groups,
};

static void cpumf_measurement_alert(struct ext_code ext_code,
				    unsigned int alert, unsigned long unused)
{
	struct cpu_hw_sf *cpuhw;

	if (!(alert & CPU_MF_INT_SF_MASK))
		return;
	inc_irq_stat(IRQEXT_CMS);
1466
	cpuhw = this_cpu_ptr(&cpu_hw_sf);
1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507

	/* Measurement alerts are shared and might happen when the PMU
	 * is not reserved.  Ignore these alerts in this case. */
	if (!(cpuhw->flags & PMU_F_RESERVED))
		return;

	/* The processing below must take care of multiple alert events that
	 * might be indicated concurrently. */

	/* Program alert request */
	if (alert & CPU_MF_INT_SF_PRA) {
		if (cpuhw->flags & PMU_F_IN_USE)
			hw_perf_event_update(cpuhw->event, 0);
		else
			WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE));
	}

	/* Report measurement alerts only for non-PRA codes */
	if (alert != CPU_MF_INT_SF_PRA)
		debug_sprintf_event(sfdbg, 6, "measurement alert: 0x%x\n", alert);

	/* Sampling authorization change request */
	if (alert & CPU_MF_INT_SF_SACA)
		qsi(&cpuhw->qsi);

	/* Loss of sample data due to high-priority machine activities */
	if (alert & CPU_MF_INT_SF_LSDA) {
		pr_err("Sample data was lost\n");
		cpuhw->flags |= PMU_F_ERR_LSDA;
		sf_disable();
	}

	/* Invalid sampling buffer entry */
	if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) {
		pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n",
		       alert);
		cpuhw->flags |= PMU_F_ERR_IBE;
		sf_disable();
	}
}

1508 1509
static int cpumf_pmu_notifier(struct notifier_block *self,
			      unsigned long action, void *hcpu)
1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
{
	unsigned int cpu = (long) hcpu;
	int flags;

	/* Ignore the notification if no events are scheduled on the PMU.
	 * This might be racy...
	 */
	if (!atomic_read(&num_events))
		return NOTIFY_OK;

	switch (action & ~CPU_TASKS_FROZEN) {
	case CPU_ONLINE:
	case CPU_ONLINE_FROZEN:
		flags = PMC_INIT;
		smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
		break;
	case CPU_DOWN_PREPARE:
		flags = PMC_RELEASE;
		smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567
static int param_get_sfb_size(char *buffer, const struct kernel_param *kp)
{
	if (!cpum_sf_avail())
		return -ENODEV;
	return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);
}

static int param_set_sfb_size(const char *val, const struct kernel_param *kp)
{
	int rc;
	unsigned long min, max;

	if (!cpum_sf_avail())
		return -ENODEV;
	if (!val || !strlen(val))
		return -EINVAL;

	/* Valid parameter values: "min,max" or "max" */
	min = CPUM_SF_MIN_SDB;
	max = CPUM_SF_MAX_SDB;
	if (strchr(val, ','))
		rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL;
	else
		rc = kstrtoul(val, 10, &max);

	if (min < 2 || min >= max || max > get_num_physpages())
		rc = -EINVAL;
	if (rc)
		return rc;

	sfb_set_limits(min, max);
1568 1569 1570
	pr_info("The sampling buffer limits have changed to: "
		"min=%lu max=%lu (diag=x%lu)\n",
		CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR);
1571 1572 1573 1574 1575 1576 1577 1578 1579
	return 0;
}

#define param_check_sfb_size(name, p) __param_check(name, p, void)
static struct kernel_param_ops param_ops_sfb_size = {
	.set = param_set_sfb_size,
	.get = param_get_sfb_size,
};

1580 1581 1582 1583 1584 1585 1586 1587 1588 1589
#define RS_INIT_FAILURE_QSI	  0x0001
#define RS_INIT_FAILURE_BSDES	  0x0002
#define RS_INIT_FAILURE_ALRT	  0x0003
#define RS_INIT_FAILURE_PERF	  0x0004
static void __init pr_cpumsf_err(unsigned int reason)
{
	pr_err("Sampling facility support for perf is not available: "
	       "reason=%04x\n", reason);
}

1590 1591
static int __init init_cpum_sampling_pmu(void)
{
1592
	struct hws_qsi_info_block si;
1593 1594 1595 1596 1597
	int err;

	if (!cpum_sf_avail())
		return -ENODEV;

1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611
	memset(&si, 0, sizeof(si));
	if (qsi(&si)) {
		pr_cpumsf_err(RS_INIT_FAILURE_QSI);
		return -ENODEV;
	}

	if (si.bsdes != sizeof(struct hws_basic_entry)) {
		pr_cpumsf_err(RS_INIT_FAILURE_BSDES);
		return -EINVAL;
	}

	if (si.ad)
		sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB);

1612 1613 1614 1615 1616
	sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80);
	if (!sfdbg)
		pr_err("Registering for s390dbf failed\n");
	debug_register_view(sfdbg, &debug_sprintf_view);

1617 1618
	err = register_external_irq(EXT_IRQ_MEASURE_ALERT,
				    cpumf_measurement_alert);
1619
	if (err) {
1620
		pr_cpumsf_err(RS_INIT_FAILURE_ALRT);
1621 1622 1623 1624 1625
		goto out;
	}

	err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW);
	if (err) {
1626
		pr_cpumsf_err(RS_INIT_FAILURE_PERF);
1627 1628
		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
					cpumf_measurement_alert);
1629 1630 1631 1632 1633 1634 1635
		goto out;
	}
	perf_cpu_notifier(cpumf_pmu_notifier);
out:
	return err;
}
arch_initcall(init_cpum_sampling_pmu);
1636
core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640);