1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Performance event support for s390x - CPU-measurement Counter Facility
   4 *
   5 *  Copyright IBM Corp. 2012, 2023
   6 *  Author(s): Hendrik Brueckner <brueckner@linux.ibm.com>
   7 *	       Thomas Richter <tmricht@linux.ibm.com>
   8 */
   9#define KMSG_COMPONENT	"cpum_cf"
  10#define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt
  11
  12#include <linux/kernel.h>
  13#include <linux/kernel_stat.h>
  14#include <linux/percpu.h>
  15#include <linux/notifier.h>
  16#include <linux/init.h>
  17#include <linux/export.h>
  18#include <linux/miscdevice.h>
  19#include <linux/perf_event.h>
  20
  21#include <asm/cpu_mf.h>
  22#include <asm/hwctrset.h>
  23#include <asm/debug.h>
  24
 
 
 
 
  25enum cpumf_ctr_set {
  26	CPUMF_CTR_SET_BASIC   = 0,    /* Basic Counter Set */
  27	CPUMF_CTR_SET_USER    = 1,    /* Problem-State Counter Set */
  28	CPUMF_CTR_SET_CRYPTO  = 2,    /* Crypto-Activity Counter Set */
  29	CPUMF_CTR_SET_EXT     = 3,    /* Extended Counter Set */
  30	CPUMF_CTR_SET_MT_DIAG = 4,    /* MT-diagnostic Counter Set */
  31
  32	/* Maximum number of counter sets */
  33	CPUMF_CTR_SET_MAX,
  34};
  35
  36#define CPUMF_LCCTL_ENABLE_SHIFT    16
  37#define CPUMF_LCCTL_ACTCTL_SHIFT     0
  38
  39static inline void ctr_set_enable(u64 *state, u64 ctrsets)
  40{
  41	*state |= ctrsets << CPUMF_LCCTL_ENABLE_SHIFT;
  42}
  43
  44static inline void ctr_set_disable(u64 *state, u64 ctrsets)
  45{
  46	*state &= ~(ctrsets << CPUMF_LCCTL_ENABLE_SHIFT);
  47}
  48
  49static inline void ctr_set_start(u64 *state, u64 ctrsets)
  50{
  51	*state |= ctrsets << CPUMF_LCCTL_ACTCTL_SHIFT;
  52}
  53
  54static inline void ctr_set_stop(u64 *state, u64 ctrsets)
  55{
  56	*state &= ~(ctrsets << CPUMF_LCCTL_ACTCTL_SHIFT);
  57}
  58
  59static inline int ctr_stcctm(enum cpumf_ctr_set set, u64 range, u64 *dest)
  60{
  61	switch (set) {
  62	case CPUMF_CTR_SET_BASIC:
  63		return stcctm(BASIC, range, dest);
  64	case CPUMF_CTR_SET_USER:
  65		return stcctm(PROBLEM_STATE, range, dest);
  66	case CPUMF_CTR_SET_CRYPTO:
  67		return stcctm(CRYPTO_ACTIVITY, range, dest);
  68	case CPUMF_CTR_SET_EXT:
  69		return stcctm(EXTENDED, range, dest);
  70	case CPUMF_CTR_SET_MT_DIAG:
  71		return stcctm(MT_DIAG_CLEARING, range, dest);
  72	case CPUMF_CTR_SET_MAX:
  73		return 3;
  74	}
  75	return 3;
  76}
  77
  78struct cpu_cf_events {
  79	refcount_t refcnt;		/* Reference count */
  80	atomic_t		ctr_set[CPUMF_CTR_SET_MAX];
  81	u64			state;		/* For perf_event_open SVC */
  82	u64			dev_state;	/* For /dev/hwctr */
  83	unsigned int		flags;
  84	size_t used;			/* Bytes used in data */
  85	size_t usedss;			/* Bytes used in start/stop */
  86	unsigned char start[PAGE_SIZE];	/* Counter set at event add */
  87	unsigned char stop[PAGE_SIZE];	/* Counter set at event delete */
  88	unsigned char data[PAGE_SIZE];	/* Counter set at /dev/hwctr */
  89	unsigned int sets;		/* # Counter set saved in memory */
  90};
  91
  92static unsigned int cfdiag_cpu_speed;	/* CPU speed for CF_DIAG trailer */
  93static debug_info_t *cf_dbg;
  94
  95/*
  96 * The CPU Measurement query counter information instruction contains
  97 * information which varies per machine generation, but is constant and
  98 * does not change when running on a particular machine, such as counter
  99 * first and second version number. This is needed to determine the size
 100 * of counter sets. Extract this information at device driver initialization.
 101 */
 102static struct cpumf_ctr_info	cpumf_ctr_info;
 103
 104struct cpu_cf_ptr {
 105	struct cpu_cf_events *cpucf;
 106};
 107
 108static struct cpu_cf_root {		/* Anchor to per CPU data */
 109	refcount_t refcnt;		/* Overall active events */
 110	struct cpu_cf_ptr __percpu *cfptr;
 111} cpu_cf_root;
 112
 113/*
 114 * Serialize event initialization and event removal. Both are called from
 115 * user space in task context with perf_event_open() and close()
 116 * system calls.
 117 *
 118 * This mutex serializes functions cpum_cf_alloc_cpu() called at event
 119 * initialization via cpumf_pmu_event_init() and function cpum_cf_free_cpu()
 120 * called at event removal via call back function hw_perf_event_destroy()
 121 * when the event is deleted. They are serialized to enforce correct
 122 * bookkeeping of pointer and reference counts anchored by
 123 * struct cpu_cf_root and the access to cpu_cf_root::refcnt and the
 124 * per CPU pointers stored in cpu_cf_root::cfptr.
 125 */
 126static DEFINE_MUTEX(pmc_reserve_mutex);
 127
 128/*
 129 * Get pointer to per-cpu structure.
 130 *
 131 * Function get_cpu_cfhw() is called from
 132 * - cfset_copy_all(): This function is protected by cpus_read_lock(), so
 133 *   CPU hot plug remove can not happen. Event removal requires a close()
 134 *   first.
 135 *
 136 * Function this_cpu_cfhw() is called from perf common code functions:
 137 * - pmu_{en|dis}able(), pmu_{add|del}()and pmu_{start|stop}():
 138 *   All functions execute with interrupts disabled on that particular CPU.
 139 * - cfset_ioctl_{on|off}, cfset_cpu_read(): see comment cfset_copy_all().
 140 *
 141 * Therefore it is safe to access the CPU specific pointer to the event.
 142 */
 143static struct cpu_cf_events *get_cpu_cfhw(int cpu)
 144{
 145	struct cpu_cf_ptr __percpu *p = cpu_cf_root.cfptr;
 146
 147	if (p) {
 148		struct cpu_cf_ptr *q = per_cpu_ptr(p, cpu);
 149
 150		return q->cpucf;
 151	}
 152	return NULL;
 153}
 154
 155static struct cpu_cf_events *this_cpu_cfhw(void)
 156{
 157	return get_cpu_cfhw(smp_processor_id());
 158}
 159
 160/* Disable counter sets on dedicated CPU */
 161static void cpum_cf_reset_cpu(void *flags)
 162{
 163	lcctl(0);
 164}
 165
 166/* Free per CPU data when the last event is removed. */
 167static void cpum_cf_free_root(void)
 168{
 169	if (!refcount_dec_and_test(&cpu_cf_root.refcnt))
 170		return;
 171	free_percpu(cpu_cf_root.cfptr);
 172	cpu_cf_root.cfptr = NULL;
 173	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
 174	on_each_cpu(cpum_cf_reset_cpu, NULL, 1);
 175	debug_sprintf_event(cf_dbg, 4, "%s root.refcnt %u cfptr %d\n",
 176			    __func__, refcount_read(&cpu_cf_root.refcnt),
 177			    !cpu_cf_root.cfptr);
 178}
 179
 180/*
 181 * On initialization of first event also allocate per CPU data dynamically.
 182 * Start with an array of pointers, the array size is the maximum number of
 183 * CPUs possible, which might be larger than the number of CPUs currently
 184 * online.
 185 */
 186static int cpum_cf_alloc_root(void)
 187{
 188	int rc = 0;
 189
 190	if (refcount_inc_not_zero(&cpu_cf_root.refcnt))
 191		return rc;
 192
 193	/* The memory is already zeroed. */
 194	cpu_cf_root.cfptr = alloc_percpu(struct cpu_cf_ptr);
 195	if (cpu_cf_root.cfptr) {
 196		refcount_set(&cpu_cf_root.refcnt, 1);
 197		on_each_cpu(cpum_cf_reset_cpu, NULL, 1);
 198		irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
 199	} else {
 200		rc = -ENOMEM;
 201	}
 202
 203	return rc;
 204}
 205
 206/* Free CPU counter data structure for a PMU */
 207static void cpum_cf_free_cpu(int cpu)
 208{
 209	struct cpu_cf_events *cpuhw;
 210	struct cpu_cf_ptr *p;
 211
 212	mutex_lock(&pmc_reserve_mutex);
 213	/*
 214	 * When invoked via CPU hotplug handler, there might be no events
 215	 * installed or that particular CPU might not have an
 216	 * event installed. This anchor pointer can be NULL!
 217	 */
 218	if (!cpu_cf_root.cfptr)
 219		goto out;
 220	p = per_cpu_ptr(cpu_cf_root.cfptr, cpu);
 221	cpuhw = p->cpucf;
 222	/*
 223	 * Might be zero when called from CPU hotplug handler and no event
 224	 * installed on that CPU, but on different CPUs.
 225	 */
 226	if (!cpuhw)
 227		goto out;
 228
 229	if (refcount_dec_and_test(&cpuhw->refcnt)) {
 230		kfree(cpuhw);
 231		p->cpucf = NULL;
 232	}
 233	cpum_cf_free_root();
 234out:
 235	mutex_unlock(&pmc_reserve_mutex);
 236}
 237
 238/* Allocate CPU counter data structure for a PMU. Called under mutex lock. */
 239static int cpum_cf_alloc_cpu(int cpu)
 240{
 241	struct cpu_cf_events *cpuhw;
 242	struct cpu_cf_ptr *p;
 243	int rc;
 244
 245	mutex_lock(&pmc_reserve_mutex);
 246	rc = cpum_cf_alloc_root();
 247	if (rc)
 248		goto unlock;
 249	p = per_cpu_ptr(cpu_cf_root.cfptr, cpu);
 250	cpuhw = p->cpucf;
 251
 252	if (!cpuhw) {
 253		cpuhw = kzalloc(sizeof(*cpuhw), GFP_KERNEL);
 254		if (cpuhw) {
 255			p->cpucf = cpuhw;
 256			refcount_set(&cpuhw->refcnt, 1);
 257		} else {
 258			rc = -ENOMEM;
 259		}
 260	} else {
 261		refcount_inc(&cpuhw->refcnt);
 262	}
 263	if (rc) {
 264		/*
 265		 * Error in allocation of event, decrement anchor. Since
 266		 * cpu_cf_event in not created, its destroy() function is not
 267		 * invoked. Adjust the reference counter for the anchor.
 268		 */
 269		cpum_cf_free_root();
 270	}
 271unlock:
 272	mutex_unlock(&pmc_reserve_mutex);
 273	return rc;
 274}
 275
 276/*
 277 * Create/delete per CPU data structures for /dev/hwctr interface and events
 278 * created by perf_event_open().
 279 * If cpu is -1, track task on all available CPUs. This requires
 280 * allocation of hardware data structures for all CPUs. This setup handles
 281 * perf_event_open() with task context and /dev/hwctr interface.
 282 * If cpu is non-zero install event on this CPU only. This setup handles
 283 * perf_event_open() with CPU context.
 284 */
 285static int cpum_cf_alloc(int cpu)
 286{
 287	cpumask_var_t mask;
 288	int rc;
 289
 290	if (cpu == -1) {
 291		if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
 292			return -ENOMEM;
 293		for_each_online_cpu(cpu) {
 294			rc = cpum_cf_alloc_cpu(cpu);
 295			if (rc) {
 296				for_each_cpu(cpu, mask)
 297					cpum_cf_free_cpu(cpu);
 298				break;
 299			}
 300			cpumask_set_cpu(cpu, mask);
 301		}
 302		free_cpumask_var(mask);
 303	} else {
 304		rc = cpum_cf_alloc_cpu(cpu);
 305	}
 306	return rc;
 307}
 308
 309static void cpum_cf_free(int cpu)
 310{
 311	if (cpu == -1) {
 312		for_each_online_cpu(cpu)
 313			cpum_cf_free_cpu(cpu);
 314	} else {
 315		cpum_cf_free_cpu(cpu);
 316	}
 317}
 318
 319#define	CF_DIAG_CTRSET_DEF		0xfeef	/* Counter set header mark */
 320						/* interval in seconds */
 321
 322/* Counter sets are stored as data stream in a page sized memory buffer and
 323 * exported to user space via raw data attached to the event sample data.
 324 * Each counter set starts with an eight byte header consisting of:
 325 * - a two byte eye catcher (0xfeef)
 326 * - a one byte counter set number
 327 * - a two byte counter set size (indicates the number of counters in this set)
 328 * - a three byte reserved value (must be zero) to make the header the same
 329 *   size as a counter value.
 330 * All counter values are eight byte in size.
 331 *
 332 * All counter sets are followed by a 64 byte trailer.
 333 * The trailer consists of a:
 334 * - flag field indicating valid fields when corresponding bit set
 335 * - the counter facility first and second version number
 336 * - the CPU speed if nonzero
 337 * - the time stamp the counter sets have been collected
 338 * - the time of day (TOD) base value
 339 * - the machine type.
 340 *
 341 * The counter sets are saved when the process is prepared to be executed on a
 342 * CPU and saved again when the process is going to be removed from a CPU.
 343 * The difference of both counter sets are calculated and stored in the event
 344 * sample data area.
 345 */
 346struct cf_ctrset_entry {	/* CPU-M CF counter set entry (8 byte) */
 347	unsigned int def:16;	/* 0-15  Data Entry Format */
 348	unsigned int set:16;	/* 16-31 Counter set identifier */
 349	unsigned int ctr:16;	/* 32-47 Number of stored counters */
 350	unsigned int res1:16;	/* 48-63 Reserved */
 351};
 352
 353struct cf_trailer_entry {	/* CPU-M CF_DIAG trailer (64 byte) */
 354	/* 0 - 7 */
 355	union {
 356		struct {
 357			unsigned int clock_base:1;	/* TOD clock base set */
 358			unsigned int speed:1;		/* CPU speed set */
 359			/* Measurement alerts */
 360			unsigned int mtda:1;	/* Loss of MT ctr. data alert */
 361			unsigned int caca:1;	/* Counter auth. change alert */
 362			unsigned int lcda:1;	/* Loss of counter data alert */
 363		};
 364		unsigned long flags;	/* 0-63    All indicators */
 365	};
 366	/* 8 - 15 */
 367	unsigned int cfvn:16;			/* 64-79   Ctr First Version */
 368	unsigned int csvn:16;			/* 80-95   Ctr Second Version */
 369	unsigned int cpu_speed:32;		/* 96-127  CPU speed */
 370	/* 16 - 23 */
 371	unsigned long timestamp;		/* 128-191 Timestamp (TOD) */
 372	/* 24 - 55 */
 373	union {
 374		struct {
 375			unsigned long progusage1;
 376			unsigned long progusage2;
 377			unsigned long progusage3;
 378			unsigned long tod_base;
 379		};
 380		unsigned long progusage[4];
 381	};
 382	/* 56 - 63 */
 383	unsigned int mach_type:16;		/* Machine type */
 384	unsigned int res1:16;			/* Reserved */
 385	unsigned int res2:32;			/* Reserved */
 386};
 387
 388/* Create the trailer data at the end of a page. */
 389static void cfdiag_trailer(struct cf_trailer_entry *te)
 390{
 391	struct cpuid cpuid;
 392
 393	te->cfvn = cpumf_ctr_info.cfvn;		/* Counter version numbers */
 394	te->csvn = cpumf_ctr_info.csvn;
 395
 396	get_cpu_id(&cpuid);			/* Machine type */
 397	te->mach_type = cpuid.machine;
 398	te->cpu_speed = cfdiag_cpu_speed;
 399	if (te->cpu_speed)
 400		te->speed = 1;
 401	te->clock_base = 1;			/* Save clock base */
 402	te->tod_base = tod_clock_base.tod;
 403	te->timestamp = get_tod_clock_fast();
 404}
 405
 406/*
 407 * The number of counters per counter set varies between machine generations,
 408 * but is constant when running on a particular machine generation.
 409 * Determine each counter set size at device driver initialization and
 410 * retrieve it later.
 411 */
 412static size_t cpumf_ctr_setsizes[CPUMF_CTR_SET_MAX];
 413static void cpum_cf_make_setsize(enum cpumf_ctr_set ctrset)
 414{
 415	size_t ctrset_size = 0;
 416
 417	switch (ctrset) {
 418	case CPUMF_CTR_SET_BASIC:
 419		if (cpumf_ctr_info.cfvn >= 1)
 420			ctrset_size = 6;
 421		break;
 422	case CPUMF_CTR_SET_USER:
 423		if (cpumf_ctr_info.cfvn == 1)
 424			ctrset_size = 6;
 425		else if (cpumf_ctr_info.cfvn >= 3)
 426			ctrset_size = 2;
 427		break;
 428	case CPUMF_CTR_SET_CRYPTO:
 429		if (cpumf_ctr_info.csvn >= 1 && cpumf_ctr_info.csvn <= 5)
 430			ctrset_size = 16;
 431		else if (cpumf_ctr_info.csvn == 6 || cpumf_ctr_info.csvn == 7)
 432			ctrset_size = 20;
 433		break;
 434	case CPUMF_CTR_SET_EXT:
 435		if (cpumf_ctr_info.csvn == 1)
 436			ctrset_size = 32;
 437		else if (cpumf_ctr_info.csvn == 2)
 438			ctrset_size = 48;
 439		else if (cpumf_ctr_info.csvn >= 3 && cpumf_ctr_info.csvn <= 5)
 440			ctrset_size = 128;
 441		else if (cpumf_ctr_info.csvn == 6 || cpumf_ctr_info.csvn == 7)
 442			ctrset_size = 160;
 443		break;
 444	case CPUMF_CTR_SET_MT_DIAG:
 445		if (cpumf_ctr_info.csvn > 3)
 446			ctrset_size = 48;
 447		break;
 448	case CPUMF_CTR_SET_MAX:
 449		break;
 450	}
 451	cpumf_ctr_setsizes[ctrset] = ctrset_size;
 452}
 453
 454/*
 455 * Return the maximum possible counter set size (in number of 8 byte counters)
 456 * depending on type and model number.
 457 */
 458static size_t cpum_cf_read_setsize(enum cpumf_ctr_set ctrset)
 459{
 460	return cpumf_ctr_setsizes[ctrset];
 461}
 462
 463/* Read a counter set. The counter set number determines the counter set and
 464 * the CPUM-CF first and second version number determine the number of
 465 * available counters in each counter set.
 466 * Each counter set starts with header containing the counter set number and
 467 * the number of eight byte counters.
 468 *
 469 * The functions returns the number of bytes occupied by this counter set
 470 * including the header.
 471 * If there is no counter in the counter set, this counter set is useless and
 472 * zero is returned on this case.
 473 *
 474 * Note that the counter sets may not be enabled or active and the stcctm
 475 * instruction might return error 3. Depending on error_ok value this is ok,
 476 * for example when called from cpumf_pmu_start() call back function.
 477 */
 478static size_t cfdiag_getctrset(struct cf_ctrset_entry *ctrdata, int ctrset,
 479			       size_t room, bool error_ok)
 480{
 481	size_t ctrset_size, need = 0;
 482	int rc = 3;				/* Assume write failure */
 483
 484	ctrdata->def = CF_DIAG_CTRSET_DEF;
 485	ctrdata->set = ctrset;
 486	ctrdata->res1 = 0;
 487	ctrset_size = cpum_cf_read_setsize(ctrset);
 488
 489	if (ctrset_size) {			/* Save data */
 490		need = ctrset_size * sizeof(u64) + sizeof(*ctrdata);
 491		if (need <= room) {
 492			rc = ctr_stcctm(ctrset, ctrset_size,
 493					(u64 *)(ctrdata + 1));
 494		}
 495		if (rc != 3 || error_ok)
 496			ctrdata->ctr = ctrset_size;
 497		else
 498			need = 0;
 499	}
 500
 501	return need;
 502}
 503
 504static const u64 cpumf_ctr_ctl[CPUMF_CTR_SET_MAX] = {
 505	[CPUMF_CTR_SET_BASIC]	= 0x02,
 506	[CPUMF_CTR_SET_USER]	= 0x04,
 507	[CPUMF_CTR_SET_CRYPTO]	= 0x08,
 508	[CPUMF_CTR_SET_EXT]	= 0x01,
 509	[CPUMF_CTR_SET_MT_DIAG] = 0x20,
 510};
 511
 512/* Read out all counter sets and save them in the provided data buffer.
 513 * The last 64 byte host an artificial trailer entry.
 514 */
 515static size_t cfdiag_getctr(void *data, size_t sz, unsigned long auth,
 516			    bool error_ok)
 517{
 518	struct cf_trailer_entry *trailer;
 519	size_t offset = 0, done;
 520	int i;
 521
 522	memset(data, 0, sz);
 523	sz -= sizeof(*trailer);		/* Always room for trailer */
 524	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
 525		struct cf_ctrset_entry *ctrdata = data + offset;
 526
 527		if (!(auth & cpumf_ctr_ctl[i]))
 528			continue;	/* Counter set not authorized */
 529
 530		done = cfdiag_getctrset(ctrdata, i, sz - offset, error_ok);
 531		offset += done;
 532	}
 533	trailer = data + offset;
 534	cfdiag_trailer(trailer);
 535	return offset + sizeof(*trailer);
 536}
 537
 538/* Calculate the difference for each counter in a counter set. */
 539static void cfdiag_diffctrset(u64 *pstart, u64 *pstop, int counters)
 540{
 541	for (; --counters >= 0; ++pstart, ++pstop)
 542		if (*pstop >= *pstart)
 543			*pstop -= *pstart;
 544		else
 545			*pstop = *pstart - *pstop + 1;
 546}
 547
 548/* Scan the counter sets and calculate the difference of each counter
 549 * in each set. The result is the increment of each counter during the
 550 * period the counter set has been activated.
 551 *
 552 * Return true on success.
 553 */
 554static int cfdiag_diffctr(struct cpu_cf_events *cpuhw, unsigned long auth)
 555{
 556	struct cf_trailer_entry *trailer_start, *trailer_stop;
 557	struct cf_ctrset_entry *ctrstart, *ctrstop;
 558	size_t offset = 0;
 
 559
 560	auth &= (1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1;
 561	do {
 562		ctrstart = (struct cf_ctrset_entry *)(cpuhw->start + offset);
 563		ctrstop = (struct cf_ctrset_entry *)(cpuhw->stop + offset);
 564
 
 
 
 
 
 
 
 565		if (memcmp(ctrstop, ctrstart, sizeof(*ctrstop))) {
 566			pr_err_once("cpum_cf_diag counter set compare error "
 567				    "in set %i\n", ctrstart->set);
 568			return 0;
 569		}
 570		auth &= ~cpumf_ctr_ctl[ctrstart->set];
 571		if (ctrstart->def == CF_DIAG_CTRSET_DEF) {
 572			cfdiag_diffctrset((u64 *)(ctrstart + 1),
 573					  (u64 *)(ctrstop + 1), ctrstart->ctr);
 574			offset += ctrstart->ctr * sizeof(u64) +
 575							sizeof(*ctrstart);
 576		}
 577	} while (ctrstart->def && auth);
 578
 579	/* Save time_stamp from start of event in stop's trailer */
 580	trailer_start = (struct cf_trailer_entry *)(cpuhw->start + offset);
 581	trailer_stop = (struct cf_trailer_entry *)(cpuhw->stop + offset);
 582	trailer_stop->progusage[0] = trailer_start->timestamp;
 583
 584	return 1;
 585}
 586
 587static enum cpumf_ctr_set get_counter_set(u64 event)
 588{
 589	int set = CPUMF_CTR_SET_MAX;
 590
 591	if (event < 32)
 592		set = CPUMF_CTR_SET_BASIC;
 593	else if (event < 64)
 594		set = CPUMF_CTR_SET_USER;
 595	else if (event < 128)
 596		set = CPUMF_CTR_SET_CRYPTO;
 597	else if (event < 288)
 598		set = CPUMF_CTR_SET_EXT;
 599	else if (event >= 448 && event < 496)
 600		set = CPUMF_CTR_SET_MT_DIAG;
 601
 602	return set;
 603}
 604
 605static int validate_ctr_version(const u64 config, enum cpumf_ctr_set set)
 606{
 607	u16 mtdiag_ctl;
 608	int err = 0;
 609
 610	/* check required version for counter sets */
 611	switch (set) {
 612	case CPUMF_CTR_SET_BASIC:
 613	case CPUMF_CTR_SET_USER:
 614		if (cpumf_ctr_info.cfvn < 1)
 615			err = -EOPNOTSUPP;
 616		break;
 617	case CPUMF_CTR_SET_CRYPTO:
 618		if ((cpumf_ctr_info.csvn >= 1 && cpumf_ctr_info.csvn <= 5 &&
 619		     config > 79) || (cpumf_ctr_info.csvn >= 6 && config > 83))
 620			err = -EOPNOTSUPP;
 621		break;
 622	case CPUMF_CTR_SET_EXT:
 623		if (cpumf_ctr_info.csvn < 1)
 624			err = -EOPNOTSUPP;
 625		if ((cpumf_ctr_info.csvn == 1 && config > 159) ||
 626		    (cpumf_ctr_info.csvn == 2 && config > 175) ||
 627		    (cpumf_ctr_info.csvn >= 3 && cpumf_ctr_info.csvn <= 5 &&
 628		     config > 255) ||
 629		    (cpumf_ctr_info.csvn >= 6 && config > 287))
 630			err = -EOPNOTSUPP;
 631		break;
 632	case CPUMF_CTR_SET_MT_DIAG:
 633		if (cpumf_ctr_info.csvn <= 3)
 634			err = -EOPNOTSUPP;
 635		/*
 636		 * MT-diagnostic counters are read-only.  The counter set
 637		 * is automatically enabled and activated on all CPUs with
 638		 * multithreading (SMT).  Deactivation of multithreading
 639		 * also disables the counter set.  State changes are ignored
 640		 * by lcctl().	Because Linux controls SMT enablement through
 641		 * a kernel parameter only, the counter set is either disabled
 642		 * or enabled and active.
 643		 *
 644		 * Thus, the counters can only be used if SMT is on and the
 645		 * counter set is enabled and active.
 646		 */
 647		mtdiag_ctl = cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG];
 648		if (!((cpumf_ctr_info.auth_ctl & mtdiag_ctl) &&
 649		      (cpumf_ctr_info.enable_ctl & mtdiag_ctl) &&
 650		      (cpumf_ctr_info.act_ctl & mtdiag_ctl)))
 651			err = -EOPNOTSUPP;
 652		break;
 653	case CPUMF_CTR_SET_MAX:
 654		err = -EOPNOTSUPP;
 655	}
 656
 657	return err;
 658}
 659
 660/*
 661 * Change the CPUMF state to active.
 662 * Enable and activate the CPU-counter sets according
 663 * to the per-cpu control state.
 664 */
 665static void cpumf_pmu_enable(struct pmu *pmu)
 666{
 667	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
 668	int err;
 669
 670	if (!cpuhw || (cpuhw->flags & PMU_F_ENABLED))
 671		return;
 672
 673	err = lcctl(cpuhw->state | cpuhw->dev_state);
 674	if (err)
 675		pr_err("Enabling the performance measuring unit failed with rc=%x\n", err);
 676	else
 677		cpuhw->flags |= PMU_F_ENABLED;
 678}
 679
 680/*
 681 * Change the CPUMF state to inactive.
 682 * Disable and enable (inactive) the CPU-counter sets according
 683 * to the per-cpu control state.
 684 */
 685static void cpumf_pmu_disable(struct pmu *pmu)
 686{
 687	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
 688	u64 inactive;
 689	int err;
 690
 691	if (!cpuhw || !(cpuhw->flags & PMU_F_ENABLED))
 692		return;
 693
 694	inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
 695	inactive |= cpuhw->dev_state;
 696	err = lcctl(inactive);
 697	if (err)
 698		pr_err("Disabling the performance measuring unit failed with rc=%x\n", err);
 699	else
 700		cpuhw->flags &= ~PMU_F_ENABLED;
 701}
 702
 703/* Release the PMU if event is the last perf event */
 704static void hw_perf_event_destroy(struct perf_event *event)
 705{
 706	cpum_cf_free(event->cpu);
 707}
 708
 709/* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
 710static const int cpumf_generic_events_basic[] = {
 711	[PERF_COUNT_HW_CPU_CYCLES]	    = 0,
 712	[PERF_COUNT_HW_INSTRUCTIONS]	    = 1,
 713	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
 714	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
 715	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
 716	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
 717	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
 718};
 719/* CPUMF <-> perf event mappings for userspace (problem-state set) */
 720static const int cpumf_generic_events_user[] = {
 721	[PERF_COUNT_HW_CPU_CYCLES]	    = 32,
 722	[PERF_COUNT_HW_INSTRUCTIONS]	    = 33,
 723	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
 724	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
 725	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
 726	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
 727	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
 728};
 729
 730static int is_userspace_event(u64 ev)
 731{
 732	return cpumf_generic_events_user[PERF_COUNT_HW_CPU_CYCLES] == ev ||
 733	       cpumf_generic_events_user[PERF_COUNT_HW_INSTRUCTIONS] == ev;
 734}
 735
 736static int __hw_perf_event_init(struct perf_event *event, unsigned int type)
 737{
 738	struct perf_event_attr *attr = &event->attr;
 739	struct hw_perf_event *hwc = &event->hw;
 740	enum cpumf_ctr_set set;
 741	u64 ev;
 742
 743	switch (type) {
 744	case PERF_TYPE_RAW:
 745		/* Raw events are used to access counters directly,
 746		 * hence do not permit excludes */
 747		if (attr->exclude_kernel || attr->exclude_user ||
 748		    attr->exclude_hv)
 749			return -EOPNOTSUPP;
 750		ev = attr->config;
 751		break;
 752
 753	case PERF_TYPE_HARDWARE:
 754		if (is_sampling_event(event))	/* No sampling support */
 755			return -ENOENT;
 756		ev = attr->config;
 757		if (!attr->exclude_user && attr->exclude_kernel) {
 758			/*
 759			 * Count user space (problem-state) only
 760			 * Handle events 32 and 33 as 0:u and 1:u
 761			 */
 762			if (!is_userspace_event(ev)) {
 763				if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
 764					return -EOPNOTSUPP;
 765				ev = cpumf_generic_events_user[ev];
 766			}
 767		} else if (!attr->exclude_kernel && attr->exclude_user) {
 768			/* No support for kernel space counters only */
 769			return -EOPNOTSUPP;
 770		} else {
 771			/* Count user and kernel space, incl. events 32 + 33 */
 772			if (!is_userspace_event(ev)) {
 773				if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
 774					return -EOPNOTSUPP;
 775				ev = cpumf_generic_events_basic[ev];
 776			}
 777		}
 778		break;
 779
 780	default:
 781		return -ENOENT;
 782	}
 783
 784	if (ev == -1)
 785		return -ENOENT;
 786
 787	if (ev > PERF_CPUM_CF_MAX_CTR)
 788		return -ENOENT;
 789
 790	/* Obtain the counter set to which the specified counter belongs */
 791	set = get_counter_set(ev);
 792	switch (set) {
 793	case CPUMF_CTR_SET_BASIC:
 794	case CPUMF_CTR_SET_USER:
 795	case CPUMF_CTR_SET_CRYPTO:
 796	case CPUMF_CTR_SET_EXT:
 797	case CPUMF_CTR_SET_MT_DIAG:
 798		/*
 799		 * Use the hardware perf event structure to store the
 800		 * counter number in the 'config' member and the counter
 801		 * set number in the 'config_base' as bit mask.
 802		 * It is later used to enable/disable the counter(s).
 803		 */
 804		hwc->config = ev;
 805		hwc->config_base = cpumf_ctr_ctl[set];
 806		break;
 807	case CPUMF_CTR_SET_MAX:
 808		/* The counter could not be associated to a counter set */
 809		return -EINVAL;
 810	}
 811
 812	/* Initialize for using the CPU-measurement counter facility */
 813	if (cpum_cf_alloc(event->cpu))
 814		return -ENOMEM;
 815	event->destroy = hw_perf_event_destroy;
 816
 817	/*
 818	 * Finally, validate version and authorization of the counter set.
 819	 * If the particular CPU counter set is not authorized,
 820	 * return with -ENOENT in order to fall back to other
 821	 * PMUs that might suffice the event request.
 822	 */
 823	if (!(hwc->config_base & cpumf_ctr_info.auth_ctl))
 824		return -ENOENT;
 825	return validate_ctr_version(hwc->config, set);
 826}
 827
 828/* Events CPU_CYLCES and INSTRUCTIONS can be submitted with two different
 829 * attribute::type values:
 830 * - PERF_TYPE_HARDWARE:
 831 * - pmu->type:
 832 * Handle both type of invocations identical. They address the same hardware.
 833 * The result is different when event modifiers exclude_kernel and/or
 834 * exclude_user are also set.
 835 */
 836static int cpumf_pmu_event_type(struct perf_event *event)
 837{
 838	u64 ev = event->attr.config;
 839
 840	if (cpumf_generic_events_basic[PERF_COUNT_HW_CPU_CYCLES] == ev ||
 841	    cpumf_generic_events_basic[PERF_COUNT_HW_INSTRUCTIONS] == ev ||
 842	    cpumf_generic_events_user[PERF_COUNT_HW_CPU_CYCLES] == ev ||
 843	    cpumf_generic_events_user[PERF_COUNT_HW_INSTRUCTIONS] == ev)
 844		return PERF_TYPE_HARDWARE;
 845	return PERF_TYPE_RAW;
 846}
 847
 848static int cpumf_pmu_event_init(struct perf_event *event)
 849{
 850	unsigned int type = event->attr.type;
 851	int err;
 852
 853	if (type == PERF_TYPE_HARDWARE || type == PERF_TYPE_RAW)
 854		err = __hw_perf_event_init(event, type);
 855	else if (event->pmu->type == type)
 856		/* Registered as unknown PMU */
 857		err = __hw_perf_event_init(event, cpumf_pmu_event_type(event));
 858	else
 859		return -ENOENT;
 860
 861	if (unlikely(err) && event->destroy)
 862		event->destroy(event);
 863
 864	return err;
 865}
 866
 867static int hw_perf_event_reset(struct perf_event *event)
 868{
 869	u64 prev, new;
 870	int err;
 871
 
 872	do {
 873		prev = local64_read(&event->hw.prev_count);
 874		err = ecctr(event->hw.config, &new);
 875		if (err) {
 876			if (err != 3)
 877				break;
 878			/* The counter is not (yet) available. This
 879			 * might happen if the counter set to which
 880			 * this counter belongs is in the disabled
 881			 * state.
 882			 */
 883			new = 0;
 884		}
 885	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
 886
 887	return err;
 888}
 889
 890static void hw_perf_event_update(struct perf_event *event)
 891{
 892	u64 prev, new, delta;
 893	int err;
 894
 
 895	do {
 896		prev = local64_read(&event->hw.prev_count);
 897		err = ecctr(event->hw.config, &new);
 898		if (err)
 899			return;
 900	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
 901
 902	delta = (prev <= new) ? new - prev
 903			      : (-1ULL - prev) + new + 1;	 /* overflow */
 904	local64_add(delta, &event->count);
 905}
 906
 907static void cpumf_pmu_read(struct perf_event *event)
 908{
 909	if (event->hw.state & PERF_HES_STOPPED)
 910		return;
 911
 912	hw_perf_event_update(event);
 913}
 914
 915static void cpumf_pmu_start(struct perf_event *event, int flags)
 916{
 917	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
 918	struct hw_perf_event *hwc = &event->hw;
 919	int i;
 920
 921	if (!(hwc->state & PERF_HES_STOPPED))
 922		return;
 923
 924	hwc->state = 0;
 925
 926	/* (Re-)enable and activate the counter set */
 927	ctr_set_enable(&cpuhw->state, hwc->config_base);
 928	ctr_set_start(&cpuhw->state, hwc->config_base);
 929
 930	/* The counter set to which this counter belongs can be already active.
 931	 * Because all counters in a set are active, the event->hw.prev_count
 932	 * needs to be synchronized.  At this point, the counter set can be in
 933	 * the inactive or disabled state.
 934	 */
 935	if (hwc->config == PERF_EVENT_CPUM_CF_DIAG) {
 936		cpuhw->usedss = cfdiag_getctr(cpuhw->start,
 937					      sizeof(cpuhw->start),
 938					      hwc->config_base, true);
 939	} else {
 940		hw_perf_event_reset(event);
 941	}
 942
 943	/* Increment refcount for counter sets */
 944	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i)
 945		if ((hwc->config_base & cpumf_ctr_ctl[i]))
 946			atomic_inc(&cpuhw->ctr_set[i]);
 947}
 948
 949/* Create perf event sample with the counter sets as raw data.	The sample
 950 * is then pushed to the event subsystem and the function checks for
 951 * possible event overflows. If an event overflow occurs, the PMU is
 952 * stopped.
 953 *
 954 * Return non-zero if an event overflow occurred.
 955 */
 956static int cfdiag_push_sample(struct perf_event *event,
 957			      struct cpu_cf_events *cpuhw)
 958{
 959	struct perf_sample_data data;
 960	struct perf_raw_record raw;
 961	struct pt_regs regs;
 962	int overflow;
 963
 964	/* Setup perf sample */
 965	perf_sample_data_init(&data, 0, event->hw.last_period);
 966	memset(&regs, 0, sizeof(regs));
 967	memset(&raw, 0, sizeof(raw));
 968
 969	if (event->attr.sample_type & PERF_SAMPLE_CPU)
 970		data.cpu_entry.cpu = event->cpu;
 971	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
 972		raw.frag.size = cpuhw->usedss;
 973		raw.frag.data = cpuhw->stop;
 974		perf_sample_save_raw_data(&data, &raw);
 975	}
 976
 977	overflow = perf_event_overflow(event, &data, &regs);
 978	if (overflow)
 979		event->pmu->stop(event, 0);
 980
 981	perf_event_update_userpage(event);
 982	return overflow;
 983}
 984
 985static void cpumf_pmu_stop(struct perf_event *event, int flags)
 986{
 987	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
 988	struct hw_perf_event *hwc = &event->hw;
 989	int i;
 990
 991	if (!(hwc->state & PERF_HES_STOPPED)) {
 992		/* Decrement reference count for this counter set and if this
 993		 * is the last used counter in the set, clear activation
 994		 * control and set the counter set state to inactive.
 995		 */
 996		for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
 997			if (!(hwc->config_base & cpumf_ctr_ctl[i]))
 998				continue;
 999			if (!atomic_dec_return(&cpuhw->ctr_set[i]))
1000				ctr_set_stop(&cpuhw->state, cpumf_ctr_ctl[i]);
1001		}
1002		hwc->state |= PERF_HES_STOPPED;
1003	}
1004
1005	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1006		if (hwc->config == PERF_EVENT_CPUM_CF_DIAG) {
1007			local64_inc(&event->count);
1008			cpuhw->usedss = cfdiag_getctr(cpuhw->stop,
1009						      sizeof(cpuhw->stop),
1010						      event->hw.config_base,
1011						      false);
1012			if (cfdiag_diffctr(cpuhw, event->hw.config_base))
1013				cfdiag_push_sample(event, cpuhw);
1014		} else {
1015			hw_perf_event_update(event);
1016		}
1017		hwc->state |= PERF_HES_UPTODATE;
1018	}
1019}
1020
1021static int cpumf_pmu_add(struct perf_event *event, int flags)
1022{
1023	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1024
1025	ctr_set_enable(&cpuhw->state, event->hw.config_base);
1026	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1027
1028	if (flags & PERF_EF_START)
1029		cpumf_pmu_start(event, PERF_EF_RELOAD);
1030
1031	return 0;
1032}
1033
1034static void cpumf_pmu_del(struct perf_event *event, int flags)
1035{
1036	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1037	int i;
1038
1039	cpumf_pmu_stop(event, PERF_EF_UPDATE);
1040
1041	/* Check if any counter in the counter set is still used.  If not used,
1042	 * change the counter set to the disabled state.  This also clears the
1043	 * content of all counters in the set.
1044	 *
1045	 * When a new perf event has been added but not yet started, this can
1046	 * clear enable control and resets all counters in a set.  Therefore,
1047	 * cpumf_pmu_start() always has to reenable a counter set.
1048	 */
1049	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i)
1050		if (!atomic_read(&cpuhw->ctr_set[i]))
1051			ctr_set_disable(&cpuhw->state, cpumf_ctr_ctl[i]);
1052}
1053
1054/* Performance monitoring unit for s390x */
1055static struct pmu cpumf_pmu = {
1056	.task_ctx_nr  = perf_sw_context,
1057	.capabilities = PERF_PMU_CAP_NO_INTERRUPT,
1058	.pmu_enable   = cpumf_pmu_enable,
1059	.pmu_disable  = cpumf_pmu_disable,
1060	.event_init   = cpumf_pmu_event_init,
1061	.add	      = cpumf_pmu_add,
1062	.del	      = cpumf_pmu_del,
1063	.start	      = cpumf_pmu_start,
1064	.stop	      = cpumf_pmu_stop,
1065	.read	      = cpumf_pmu_read,
1066};
1067
1068static struct cfset_session {		/* CPUs and counter set bit mask */
1069	struct list_head head;		/* Head of list of active processes */
1070} cfset_session = {
1071	.head = LIST_HEAD_INIT(cfset_session.head)
1072};
1073
1074static refcount_t cfset_opencnt = REFCOUNT_INIT(0);	/* Access count */
1075/*
1076 * Synchronize access to device /dev/hwc. This mutex protects against
1077 * concurrent access to functions cfset_open() and cfset_release().
1078 * Same for CPU hotplug add and remove events triggering
1079 * cpum_cf_online_cpu() and cpum_cf_offline_cpu().
1080 * It also serializes concurrent device ioctl access from multiple
1081 * processes accessing /dev/hwc.
1082 *
1083 * The mutex protects concurrent access to the /dev/hwctr session management
1084 * struct cfset_session and reference counting variable cfset_opencnt.
1085 */
1086static DEFINE_MUTEX(cfset_ctrset_mutex);
1087
1088/*
1089 * CPU hotplug handles only /dev/hwctr device.
1090 * For perf_event_open() the CPU hotplug handling is done on kernel common
1091 * code:
1092 * - CPU add: Nothing is done since a file descriptor can not be created
1093 *   and returned to the user.
1094 * - CPU delete: Handled by common code via pmu_disable(), pmu_stop() and
1095 *   pmu_delete(). The event itself is removed when the file descriptor is
1096 *   closed.
1097 */
1098static int cfset_online_cpu(unsigned int cpu);
1099
1100static int cpum_cf_online_cpu(unsigned int cpu)
1101{
1102	int rc = 0;
1103
1104	/*
1105	 * Ignore notification for perf_event_open().
1106	 * Handle only /dev/hwctr device sessions.
1107	 */
1108	mutex_lock(&cfset_ctrset_mutex);
1109	if (refcount_read(&cfset_opencnt)) {
1110		rc = cpum_cf_alloc_cpu(cpu);
1111		if (!rc)
1112			cfset_online_cpu(cpu);
1113	}
1114	mutex_unlock(&cfset_ctrset_mutex);
1115	return rc;
1116}
1117
1118static int cfset_offline_cpu(unsigned int cpu);
1119
1120static int cpum_cf_offline_cpu(unsigned int cpu)
1121{
1122	/*
1123	 * During task exit processing of grouped perf events triggered by CPU
1124	 * hotplug processing, pmu_disable() is called as part of perf context
1125	 * removal process. Therefore do not trigger event removal now for
1126	 * perf_event_open() created events. Perf common code triggers event
1127	 * destruction when the event file descriptor is closed.
1128	 *
1129	 * Handle only /dev/hwctr device sessions.
1130	 */
1131	mutex_lock(&cfset_ctrset_mutex);
1132	if (refcount_read(&cfset_opencnt)) {
1133		cfset_offline_cpu(cpu);
1134		cpum_cf_free_cpu(cpu);
1135	}
1136	mutex_unlock(&cfset_ctrset_mutex);
1137	return 0;
1138}
1139
1140/* Return true if store counter set multiple instruction is available */
1141static inline int stccm_avail(void)
1142{
1143	return test_facility(142);
1144}
1145
1146/* CPU-measurement alerts for the counter facility */
1147static void cpumf_measurement_alert(struct ext_code ext_code,
1148				    unsigned int alert, unsigned long unused)
1149{
1150	struct cpu_cf_events *cpuhw;
1151
1152	if (!(alert & CPU_MF_INT_CF_MASK))
1153		return;
1154
1155	inc_irq_stat(IRQEXT_CMC);
1156
1157	/*
1158	 * Measurement alerts are shared and might happen when the PMU
1159	 * is not reserved.  Ignore these alerts in this case.
1160	 */
1161	cpuhw = this_cpu_cfhw();
1162	if (!cpuhw)
1163		return;
1164
1165	/* counter authorization change alert */
1166	if (alert & CPU_MF_INT_CF_CACA)
1167		qctri(&cpumf_ctr_info);
1168
1169	/* loss of counter data alert */
1170	if (alert & CPU_MF_INT_CF_LCDA)
1171		pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
1172
1173	/* loss of MT counter data alert */
1174	if (alert & CPU_MF_INT_CF_MTDA)
1175		pr_warn("CPU[%i] MT counter data was lost\n",
1176			smp_processor_id());
1177}
1178
1179static int cfset_init(void);
1180static int __init cpumf_pmu_init(void)
1181{
1182	int rc;
1183
1184	/* Extract counter measurement facility information */
1185	if (!cpum_cf_avail() || qctri(&cpumf_ctr_info))
1186		return -ENODEV;
1187
1188	/* Determine and store counter set sizes for later reference */
1189	for (rc = CPUMF_CTR_SET_BASIC; rc < CPUMF_CTR_SET_MAX; ++rc)
1190		cpum_cf_make_setsize(rc);
1191
1192	/*
1193	 * Clear bit 15 of cr0 to unauthorize problem-state to
1194	 * extract measurement counters
1195	 */
1196	system_ctl_clear_bit(0, CR0_CPUMF_EXTRACTION_AUTH_BIT);
1197
1198	/* register handler for measurement-alert interruptions */
1199	rc = register_external_irq(EXT_IRQ_MEASURE_ALERT,
1200				   cpumf_measurement_alert);
1201	if (rc) {
1202		pr_err("Registering for CPU-measurement alerts failed with rc=%i\n", rc);
1203		return rc;
1204	}
1205
1206	/* Setup s390dbf facility */
1207	cf_dbg = debug_register(KMSG_COMPONENT, 2, 1, 128);
1208	if (!cf_dbg) {
1209		pr_err("Registration of s390dbf(cpum_cf) failed\n");
1210		rc = -ENOMEM;
1211		goto out1;
1212	}
1213	debug_register_view(cf_dbg, &debug_sprintf_view);
1214
1215	cpumf_pmu.attr_groups = cpumf_cf_event_group();
1216	rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", -1);
1217	if (rc) {
1218		pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
1219		goto out2;
1220	} else if (stccm_avail()) {	/* Setup counter set device */
1221		cfset_init();
1222	}
1223
1224	rc = cpuhp_setup_state(CPUHP_AP_PERF_S390_CF_ONLINE,
1225			       "perf/s390/cf:online",
1226			       cpum_cf_online_cpu, cpum_cf_offline_cpu);
1227	return rc;
1228
1229out2:
1230	debug_unregister_view(cf_dbg, &debug_sprintf_view);
1231	debug_unregister(cf_dbg);
1232out1:
1233	unregister_external_irq(EXT_IRQ_MEASURE_ALERT, cpumf_measurement_alert);
1234	return rc;
1235}
1236
1237/* Support for the CPU Measurement Facility counter set extraction using
1238 * device /dev/hwctr. This allows user space programs to extract complete
1239 * counter set via normal file operations.
1240 */
1241
1242struct cfset_call_on_cpu_parm {		/* Parm struct for smp_call_on_cpu */
1243	unsigned int sets;		/* Counter set bit mask */
1244	atomic_t cpus_ack;		/* # CPUs successfully executed func */
1245};
1246
1247struct cfset_request {			/* CPUs and counter set bit mask */
1248	unsigned long ctrset;		/* Bit mask of counter set to read */
1249	cpumask_t mask;			/* CPU mask to read from */
1250	struct list_head node;		/* Chain to cfset_session.head */
1251};
1252
1253static void cfset_session_init(void)
1254{
1255	INIT_LIST_HEAD(&cfset_session.head);
1256}
1257
1258/* Remove current request from global bookkeeping. Maintain a counter set bit
1259 * mask on a per CPU basis.
1260 * Done in process context under mutex protection.
1261 */
1262static void cfset_session_del(struct cfset_request *p)
1263{
1264	list_del(&p->node);
1265}
1266
1267/* Add current request to global bookkeeping. Maintain a counter set bit mask
1268 * on a per CPU basis.
1269 * Done in process context under mutex protection.
1270 */
1271static void cfset_session_add(struct cfset_request *p)
1272{
1273	list_add(&p->node, &cfset_session.head);
1274}
1275
1276/* The /dev/hwctr device access uses PMU_F_IN_USE to mark the device access
1277 * path is currently used.
1278 * The cpu_cf_events::dev_state is used to denote counter sets in use by this
1279 * interface. It is always or'ed in. If this interface is not active, its
1280 * value is zero and no additional counter sets will be included.
1281 *
1282 * The cpu_cf_events::state is used by the perf_event_open SVC and remains
1283 * unchanged.
1284 *
1285 * perf_pmu_enable() and perf_pmu_enable() and its call backs
1286 * cpumf_pmu_enable() and  cpumf_pmu_disable() are called by the
1287 * performance measurement subsystem to enable per process
1288 * CPU Measurement counter facility.
1289 * The XXX_enable() and XXX_disable functions are used to turn off
1290 * x86 performance monitoring interrupt (PMI) during scheduling.
1291 * s390 uses these calls to temporarily stop and resume the active CPU
1292 * counters sets during scheduling.
1293 *
1294 * We do allow concurrent access of perf_event_open() SVC and /dev/hwctr
1295 * device access.  The perf_event_open() SVC interface makes a lot of effort
1296 * to only run the counters while the calling process is actively scheduled
1297 * to run.
1298 * When /dev/hwctr interface is also used at the same time, the counter sets
1299 * will keep running, even when the process is scheduled off a CPU.
1300 * However this is not a problem and does not lead to wrong counter values
1301 * for the perf_event_open() SVC. The current counter value will be recorded
1302 * during schedule-in. At schedule-out time the current counter value is
1303 * extracted again and the delta is calculated and added to the event.
1304 */
1305/* Stop all counter sets via ioctl interface */
1306static void cfset_ioctl_off(void *parm)
1307{
1308	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1309	struct cfset_call_on_cpu_parm *p = parm;
1310	int rc;
1311
1312	/* Check if any counter set used by /dev/hwctr */
1313	for (rc = CPUMF_CTR_SET_BASIC; rc < CPUMF_CTR_SET_MAX; ++rc)
1314		if ((p->sets & cpumf_ctr_ctl[rc])) {
1315			if (!atomic_dec_return(&cpuhw->ctr_set[rc])) {
1316				ctr_set_disable(&cpuhw->dev_state,
1317						cpumf_ctr_ctl[rc]);
1318				ctr_set_stop(&cpuhw->dev_state,
1319					     cpumf_ctr_ctl[rc]);
1320			}
1321		}
1322	/* Keep perf_event_open counter sets */
1323	rc = lcctl(cpuhw->dev_state | cpuhw->state);
1324	if (rc)
1325		pr_err("Counter set stop %#llx of /dev/%s failed rc=%i\n",
1326		       cpuhw->state, S390_HWCTR_DEVICE, rc);
1327	if (!cpuhw->dev_state)
1328		cpuhw->flags &= ~PMU_F_IN_USE;
1329}
1330
1331/* Start counter sets on particular CPU */
1332static void cfset_ioctl_on(void *parm)
1333{
1334	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1335	struct cfset_call_on_cpu_parm *p = parm;
1336	int rc;
1337
1338	cpuhw->flags |= PMU_F_IN_USE;
1339	ctr_set_enable(&cpuhw->dev_state, p->sets);
1340	ctr_set_start(&cpuhw->dev_state, p->sets);
1341	for (rc = CPUMF_CTR_SET_BASIC; rc < CPUMF_CTR_SET_MAX; ++rc)
1342		if ((p->sets & cpumf_ctr_ctl[rc]))
1343			atomic_inc(&cpuhw->ctr_set[rc]);
1344	rc = lcctl(cpuhw->dev_state | cpuhw->state);	/* Start counter sets */
1345	if (!rc)
1346		atomic_inc(&p->cpus_ack);
1347	else
1348		pr_err("Counter set start %#llx of /dev/%s failed rc=%i\n",
1349		       cpuhw->dev_state | cpuhw->state, S390_HWCTR_DEVICE, rc);
1350}
1351
1352static void cfset_release_cpu(void *p)
1353{
1354	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1355	int rc;
1356
1357	cpuhw->dev_state = 0;
1358	rc = lcctl(cpuhw->state);	/* Keep perf_event_open counter sets */
1359	if (rc)
1360		pr_err("Counter set release %#llx of /dev/%s failed rc=%i\n",
1361		       cpuhw->state, S390_HWCTR_DEVICE, rc);
1362}
1363
1364/* This modifies the process CPU mask to adopt it to the currently online
1365 * CPUs. Offline CPUs can not be addresses. This call terminates the access
1366 * and is usually followed by close() or a new iotcl(..., START, ...) which
1367 * creates a new request structure.
1368 */
1369static void cfset_all_stop(struct cfset_request *req)
1370{
1371	struct cfset_call_on_cpu_parm p = {
1372		.sets = req->ctrset,
1373	};
1374
1375	cpumask_and(&req->mask, &req->mask, cpu_online_mask);
1376	on_each_cpu_mask(&req->mask, cfset_ioctl_off, &p, 1);
1377}
1378
1379/* Release function is also called when application gets terminated without
1380 * doing a proper ioctl(..., S390_HWCTR_STOP, ...) command.
1381 */
1382static int cfset_release(struct inode *inode, struct file *file)
1383{
1384	mutex_lock(&cfset_ctrset_mutex);
1385	/* Open followed by close/exit has no private_data */
1386	if (file->private_data) {
1387		cfset_all_stop(file->private_data);
1388		cfset_session_del(file->private_data);
1389		kfree(file->private_data);
1390		file->private_data = NULL;
1391	}
1392	if (refcount_dec_and_test(&cfset_opencnt)) {	/* Last close */
1393		on_each_cpu(cfset_release_cpu, NULL, 1);
1394		cpum_cf_free(-1);
1395	}
1396	mutex_unlock(&cfset_ctrset_mutex);
1397	return 0;
1398}
1399
1400/*
1401 * Open via /dev/hwctr device. Allocate all per CPU resources on the first
1402 * open of the device. The last close releases all per CPU resources.
1403 * Parallel perf_event_open system calls also use per CPU resources.
1404 * These invocations are handled via reference counting on the per CPU data
1405 * structures.
1406 */
1407static int cfset_open(struct inode *inode, struct file *file)
1408{
1409	int rc = 0;
1410
1411	if (!perfmon_capable())
1412		return -EPERM;
1413	file->private_data = NULL;
1414
1415	mutex_lock(&cfset_ctrset_mutex);
1416	if (!refcount_inc_not_zero(&cfset_opencnt)) {	/* First open */
1417		rc = cpum_cf_alloc(-1);
1418		if (!rc) {
1419			cfset_session_init();
1420			refcount_set(&cfset_opencnt, 1);
1421		}
1422	}
1423	mutex_unlock(&cfset_ctrset_mutex);
1424
1425	/* nonseekable_open() never fails */
1426	return rc ?: nonseekable_open(inode, file);
1427}
1428
1429static int cfset_all_start(struct cfset_request *req)
1430{
1431	struct cfset_call_on_cpu_parm p = {
1432		.sets = req->ctrset,
1433		.cpus_ack = ATOMIC_INIT(0),
1434	};
1435	cpumask_var_t mask;
1436	int rc = 0;
1437
1438	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
1439		return -ENOMEM;
1440	cpumask_and(mask, &req->mask, cpu_online_mask);
1441	on_each_cpu_mask(mask, cfset_ioctl_on, &p, 1);
1442	if (atomic_read(&p.cpus_ack) != cpumask_weight(mask)) {
1443		on_each_cpu_mask(mask, cfset_ioctl_off, &p, 1);
1444		rc = -EIO;
1445	}
1446	free_cpumask_var(mask);
1447	return rc;
1448}
1449
1450/* Return the maximum required space for all possible CPUs in case one
1451 * CPU will be onlined during the START, READ, STOP cycles.
1452 * To find out the size of the counter sets, any one CPU will do. They
1453 * all have the same counter sets.
1454 */
1455static size_t cfset_needspace(unsigned int sets)
1456{
1457	size_t bytes = 0;
1458	int i;
1459
1460	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1461		if (!(sets & cpumf_ctr_ctl[i]))
1462			continue;
1463		bytes += cpum_cf_read_setsize(i) * sizeof(u64) +
1464			 sizeof(((struct s390_ctrset_setdata *)0)->set) +
1465			 sizeof(((struct s390_ctrset_setdata *)0)->no_cnts);
1466	}
1467	bytes = sizeof(((struct s390_ctrset_read *)0)->no_cpus) + nr_cpu_ids *
1468		(bytes + sizeof(((struct s390_ctrset_cpudata *)0)->cpu_nr) +
1469		     sizeof(((struct s390_ctrset_cpudata *)0)->no_sets));
1470	return bytes;
1471}
1472
1473static int cfset_all_copy(unsigned long arg, cpumask_t *mask)
1474{
1475	struct s390_ctrset_read __user *ctrset_read;
1476	unsigned int cpu, cpus, rc = 0;
1477	void __user *uptr;
1478
1479	ctrset_read = (struct s390_ctrset_read __user *)arg;
1480	uptr = ctrset_read->data;
1481	for_each_cpu(cpu, mask) {
1482		struct cpu_cf_events *cpuhw = get_cpu_cfhw(cpu);
1483		struct s390_ctrset_cpudata __user *ctrset_cpudata;
1484
1485		ctrset_cpudata = uptr;
1486		rc  = put_user(cpu, &ctrset_cpudata->cpu_nr);
1487		rc |= put_user(cpuhw->sets, &ctrset_cpudata->no_sets);
1488		rc |= copy_to_user(ctrset_cpudata->data, cpuhw->data,
1489				   cpuhw->used);
1490		if (rc) {
1491			rc = -EFAULT;
1492			goto out;
1493		}
1494		uptr += sizeof(struct s390_ctrset_cpudata) + cpuhw->used;
1495		cond_resched();
1496	}
1497	cpus = cpumask_weight(mask);
1498	if (put_user(cpus, &ctrset_read->no_cpus))
1499		rc = -EFAULT;
1500out:
1501	return rc;
1502}
1503
1504static size_t cfset_cpuset_read(struct s390_ctrset_setdata *p, int ctrset,
1505				int ctrset_size, size_t room)
1506{
1507	size_t need = 0;
1508	int rc = -1;
1509
1510	need = sizeof(*p) + sizeof(u64) * ctrset_size;
1511	if (need <= room) {
1512		p->set = cpumf_ctr_ctl[ctrset];
1513		p->no_cnts = ctrset_size;
1514		rc = ctr_stcctm(ctrset, ctrset_size, (u64 *)p->cv);
1515		if (rc == 3)		/* Nothing stored */
1516			need = 0;
1517	}
1518	return need;
1519}
1520
1521/* Read all counter sets. */
1522static void cfset_cpu_read(void *parm)
1523{
1524	struct cpu_cf_events *cpuhw = this_cpu_cfhw();
1525	struct cfset_call_on_cpu_parm *p = parm;
1526	int set, set_size;
1527	size_t space;
1528
1529	/* No data saved yet */
1530	cpuhw->used = 0;
1531	cpuhw->sets = 0;
1532	memset(cpuhw->data, 0, sizeof(cpuhw->data));
1533
1534	/* Scan the counter sets */
1535	for (set = CPUMF_CTR_SET_BASIC; set < CPUMF_CTR_SET_MAX; ++set) {
1536		struct s390_ctrset_setdata *sp = (void *)cpuhw->data +
1537						 cpuhw->used;
1538
1539		if (!(p->sets & cpumf_ctr_ctl[set]))
1540			continue;	/* Counter set not in list */
1541		set_size = cpum_cf_read_setsize(set);
1542		space = sizeof(cpuhw->data) - cpuhw->used;
1543		space = cfset_cpuset_read(sp, set, set_size, space);
1544		if (space) {
1545			cpuhw->used += space;
1546			cpuhw->sets += 1;
1547		}
1548	}
1549}
1550
1551static int cfset_all_read(unsigned long arg, struct cfset_request *req)
1552{
1553	struct cfset_call_on_cpu_parm p;
1554	cpumask_var_t mask;
1555	int rc;
1556
1557	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
1558		return -ENOMEM;
1559
1560	p.sets = req->ctrset;
1561	cpumask_and(mask, &req->mask, cpu_online_mask);
1562	on_each_cpu_mask(mask, cfset_cpu_read, &p, 1);
1563	rc = cfset_all_copy(arg, mask);
1564	free_cpumask_var(mask);
1565	return rc;
1566}
1567
1568static long cfset_ioctl_read(unsigned long arg, struct cfset_request *req)
1569{
1570	int ret = -ENODATA;
1571
1572	if (req && req->ctrset)
1573		ret = cfset_all_read(arg, req);
1574	return ret;
1575}
1576
1577static long cfset_ioctl_stop(struct file *file)
1578{
1579	struct cfset_request *req = file->private_data;
1580	int ret = -ENXIO;
1581
1582	if (req) {
1583		cfset_all_stop(req);
1584		cfset_session_del(req);
1585		kfree(req);
1586		file->private_data = NULL;
1587		ret = 0;
1588	}
1589	return ret;
1590}
1591
1592static long cfset_ioctl_start(unsigned long arg, struct file *file)
1593{
1594	struct s390_ctrset_start __user *ustart;
1595	struct s390_ctrset_start start;
1596	struct cfset_request *preq;
1597	void __user *umask;
1598	unsigned int len;
1599	int ret = 0;
1600	size_t need;
1601
1602	if (file->private_data)
1603		return -EBUSY;
1604	ustart = (struct s390_ctrset_start __user *)arg;
1605	if (copy_from_user(&start, ustart, sizeof(start)))
1606		return -EFAULT;
1607	if (start.version != S390_HWCTR_START_VERSION)
1608		return -EINVAL;
1609	if (start.counter_sets & ~(cpumf_ctr_ctl[CPUMF_CTR_SET_BASIC] |
1610				   cpumf_ctr_ctl[CPUMF_CTR_SET_USER] |
1611				   cpumf_ctr_ctl[CPUMF_CTR_SET_CRYPTO] |
1612				   cpumf_ctr_ctl[CPUMF_CTR_SET_EXT] |
1613				   cpumf_ctr_ctl[CPUMF_CTR_SET_MT_DIAG]))
1614		return -EINVAL;		/* Invalid counter set */
1615	if (!start.counter_sets)
1616		return -EINVAL;		/* No counter set at all? */
1617
1618	preq = kzalloc(sizeof(*preq), GFP_KERNEL);
1619	if (!preq)
1620		return -ENOMEM;
1621	cpumask_clear(&preq->mask);
1622	len = min_t(u64, start.cpumask_len, cpumask_size());
1623	umask = (void __user *)start.cpumask;
1624	if (copy_from_user(&preq->mask, umask, len)) {
1625		kfree(preq);
1626		return -EFAULT;
1627	}
1628	if (cpumask_empty(&preq->mask)) {
1629		kfree(preq);
1630		return -EINVAL;
1631	}
1632	need = cfset_needspace(start.counter_sets);
1633	if (put_user(need, &ustart->data_bytes)) {
1634		kfree(preq);
1635		return -EFAULT;
1636	}
1637	preq->ctrset = start.counter_sets;
1638	ret = cfset_all_start(preq);
1639	if (!ret) {
1640		cfset_session_add(preq);
1641		file->private_data = preq;
1642	} else {
1643		kfree(preq);
1644	}
1645	return ret;
1646}
1647
1648/* Entry point to the /dev/hwctr device interface.
1649 * The ioctl system call supports three subcommands:
1650 * S390_HWCTR_START: Start the specified counter sets on a CPU list. The
1651 *    counter set keeps running until explicitly stopped. Returns the number
1652 *    of bytes needed to store the counter values. If another S390_HWCTR_START
1653 *    ioctl subcommand is called without a previous S390_HWCTR_STOP stop
1654 *    command on the same file descriptor, -EBUSY is returned.
1655 * S390_HWCTR_READ: Read the counter set values from specified CPU list given
1656 *    with the S390_HWCTR_START command.
1657 * S390_HWCTR_STOP: Stops the counter sets on the CPU list given with the
1658 *    previous S390_HWCTR_START subcommand.
1659 */
1660static long cfset_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1661{
1662	int ret;
1663
1664	cpus_read_lock();
1665	mutex_lock(&cfset_ctrset_mutex);
1666	switch (cmd) {
1667	case S390_HWCTR_START:
1668		ret = cfset_ioctl_start(arg, file);
1669		break;
1670	case S390_HWCTR_STOP:
1671		ret = cfset_ioctl_stop(file);
1672		break;
1673	case S390_HWCTR_READ:
1674		ret = cfset_ioctl_read(arg, file->private_data);
1675		break;
1676	default:
1677		ret = -ENOTTY;
1678		break;
1679	}
1680	mutex_unlock(&cfset_ctrset_mutex);
1681	cpus_read_unlock();
1682	return ret;
1683}
1684
1685static const struct file_operations cfset_fops = {
1686	.owner = THIS_MODULE,
1687	.open = cfset_open,
1688	.release = cfset_release,
1689	.unlocked_ioctl	= cfset_ioctl,
1690	.compat_ioctl = cfset_ioctl,
1691	.llseek = no_llseek
1692};
1693
1694static struct miscdevice cfset_dev = {
1695	.name	= S390_HWCTR_DEVICE,
1696	.minor	= MISC_DYNAMIC_MINOR,
1697	.fops	= &cfset_fops,
1698	.mode	= 0666,
1699};
1700
1701/* Hotplug add of a CPU. Scan through all active processes and add
1702 * that CPU to the list of CPUs supplied with ioctl(..., START, ...).
1703 */
1704static int cfset_online_cpu(unsigned int cpu)
1705{
1706	struct cfset_call_on_cpu_parm p;
1707	struct cfset_request *rp;
1708
1709	if (!list_empty(&cfset_session.head)) {
1710		list_for_each_entry(rp, &cfset_session.head, node) {
1711			p.sets = rp->ctrset;
1712			cfset_ioctl_on(&p);
1713			cpumask_set_cpu(cpu, &rp->mask);
1714		}
1715	}
1716	return 0;
1717}
1718
1719/* Hotplug remove of a CPU. Scan through all active processes and clear
1720 * that CPU from the list of CPUs supplied with ioctl(..., START, ...).
1721 * Adjust reference counts.
1722 */
1723static int cfset_offline_cpu(unsigned int cpu)
1724{
1725	struct cfset_call_on_cpu_parm p;
1726	struct cfset_request *rp;
1727
1728	if (!list_empty(&cfset_session.head)) {
1729		list_for_each_entry(rp, &cfset_session.head, node) {
1730			p.sets = rp->ctrset;
1731			cfset_ioctl_off(&p);
1732			cpumask_clear_cpu(cpu, &rp->mask);
1733		}
1734	}
1735	return 0;
1736}
1737
1738static void cfdiag_read(struct perf_event *event)
1739{
1740}
1741
1742static int get_authctrsets(void)
1743{
1744	unsigned long auth = 0;
1745	enum cpumf_ctr_set i;
1746
1747	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1748		if (cpumf_ctr_info.auth_ctl & cpumf_ctr_ctl[i])
1749			auth |= cpumf_ctr_ctl[i];
1750	}
1751	return auth;
1752}
1753
1754/* Setup the event. Test for authorized counter sets and only include counter
1755 * sets which are authorized at the time of the setup. Including unauthorized
1756 * counter sets result in specification exception (and panic).
1757 */
1758static int cfdiag_event_init2(struct perf_event *event)
1759{
1760	struct perf_event_attr *attr = &event->attr;
1761	int err = 0;
1762
1763	/* Set sample_period to indicate sampling */
1764	event->hw.config = attr->config;
1765	event->hw.sample_period = attr->sample_period;
1766	local64_set(&event->hw.period_left, event->hw.sample_period);
1767	local64_set(&event->count, 0);
1768	event->hw.last_period = event->hw.sample_period;
1769
1770	/* Add all authorized counter sets to config_base. The
1771	 * the hardware init function is either called per-cpu or just once
1772	 * for all CPUS (event->cpu == -1).  This depends on the whether
1773	 * counting is started for all CPUs or on a per workload base where
1774	 * the perf event moves from one CPU to another CPU.
1775	 * Checking the authorization on any CPU is fine as the hardware
1776	 * applies the same authorization settings to all CPUs.
1777	 */
1778	event->hw.config_base = get_authctrsets();
1779
1780	/* No authorized counter sets, nothing to count/sample */
1781	if (!event->hw.config_base)
1782		err = -EINVAL;
1783
1784	return err;
1785}
1786
1787static int cfdiag_event_init(struct perf_event *event)
1788{
1789	struct perf_event_attr *attr = &event->attr;
1790	int err = -ENOENT;
1791
1792	if (event->attr.config != PERF_EVENT_CPUM_CF_DIAG ||
1793	    event->attr.type != event->pmu->type)
1794		goto out;
1795
1796	/* Raw events are used to access counters directly,
1797	 * hence do not permit excludes.
1798	 * This event is useless without PERF_SAMPLE_RAW to return counter set
1799	 * values as raw data.
1800	 */
1801	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv ||
1802	    !(attr->sample_type & (PERF_SAMPLE_CPU | PERF_SAMPLE_RAW))) {
1803		err = -EOPNOTSUPP;
1804		goto out;
1805	}
1806
1807	/* Initialize for using the CPU-measurement counter facility */
1808	if (cpum_cf_alloc(event->cpu))
1809		return -ENOMEM;
1810	event->destroy = hw_perf_event_destroy;
1811
1812	err = cfdiag_event_init2(event);
1813	if (unlikely(err))
1814		event->destroy(event);
1815out:
1816	return err;
1817}
1818
1819/* Create cf_diag/events/CF_DIAG event sysfs file. This counter is used
1820 * to collect the complete counter sets for a scheduled process. Target
1821 * are complete counter sets attached as raw data to the artificial event.
1822 * This results in complete counter sets available when a process is
1823 * scheduled. Contains the delta of every counter while the process was
1824 * running.
1825 */
1826CPUMF_EVENT_ATTR(CF_DIAG, CF_DIAG, PERF_EVENT_CPUM_CF_DIAG);
1827
1828static struct attribute *cfdiag_events_attr[] = {
1829	CPUMF_EVENT_PTR(CF_DIAG, CF_DIAG),
1830	NULL,
1831};
1832
1833PMU_FORMAT_ATTR(event, "config:0-63");
1834
1835static struct attribute *cfdiag_format_attr[] = {
1836	&format_attr_event.attr,
1837	NULL,
1838};
1839
1840static struct attribute_group cfdiag_events_group = {
1841	.name = "events",
1842	.attrs = cfdiag_events_attr,
1843};
1844static struct attribute_group cfdiag_format_group = {
1845	.name = "format",
1846	.attrs = cfdiag_format_attr,
1847};
1848static const struct attribute_group *cfdiag_attr_groups[] = {
1849	&cfdiag_events_group,
1850	&cfdiag_format_group,
1851	NULL,
1852};
1853
1854/* Performance monitoring unit for event CF_DIAG. Since this event
1855 * is also started and stopped via the perf_event_open() system call, use
1856 * the same event enable/disable call back functions. They do not
1857 * have a pointer to the perf_event strcture as first parameter.
1858 *
1859 * The functions XXX_add, XXX_del, XXX_start and XXX_stop are also common.
1860 * Reuse them and distinguish the event (always first parameter) via
1861 * 'config' member.
1862 */
1863static struct pmu cf_diag = {
1864	.task_ctx_nr  = perf_sw_context,
1865	.event_init   = cfdiag_event_init,
1866	.pmu_enable   = cpumf_pmu_enable,
1867	.pmu_disable  = cpumf_pmu_disable,
1868	.add	      = cpumf_pmu_add,
1869	.del	      = cpumf_pmu_del,
1870	.start	      = cpumf_pmu_start,
1871	.stop	      = cpumf_pmu_stop,
1872	.read	      = cfdiag_read,
1873
1874	.attr_groups  = cfdiag_attr_groups
1875};
1876
1877/* Calculate memory needed to store all counter sets together with header and
1878 * trailer data. This is independent of the counter set authorization which
1879 * can vary depending on the configuration.
1880 */
1881static size_t cfdiag_maxsize(struct cpumf_ctr_info *info)
1882{
1883	size_t max_size = sizeof(struct cf_trailer_entry);
1884	enum cpumf_ctr_set i;
1885
1886	for (i = CPUMF_CTR_SET_BASIC; i < CPUMF_CTR_SET_MAX; ++i) {
1887		size_t size = cpum_cf_read_setsize(i);
1888
1889		if (size)
1890			max_size += size * sizeof(u64) +
1891				    sizeof(struct cf_ctrset_entry);
1892	}
1893	return max_size;
1894}
1895
1896/* Get the CPU speed, try sampling facility first and CPU attributes second. */
1897static void cfdiag_get_cpu_speed(void)
1898{
1899	unsigned long mhz;
1900
1901	if (cpum_sf_avail()) {			/* Sampling facility first */
1902		struct hws_qsi_info_block si;
1903
1904		memset(&si, 0, sizeof(si));
1905		if (!qsi(&si)) {
1906			cfdiag_cpu_speed = si.cpu_speed;
1907			return;
1908		}
1909	}
1910
1911	/* Fallback: CPU speed extract static part. Used in case
1912	 * CPU Measurement Sampling Facility is turned off.
1913	 */
1914	mhz = __ecag(ECAG_CPU_ATTRIBUTE, 0);
1915	if (mhz != -1UL)
1916		cfdiag_cpu_speed = mhz & 0xffffffff;
1917}
1918
1919static int cfset_init(void)
1920{
1921	size_t need;
1922	int rc;
1923
1924	cfdiag_get_cpu_speed();
1925	/* Make sure the counter set data fits into predefined buffer. */
1926	need = cfdiag_maxsize(&cpumf_ctr_info);
1927	if (need > sizeof(((struct cpu_cf_events *)0)->start)) {
1928		pr_err("Insufficient memory for PMU(cpum_cf_diag) need=%zu\n",
1929		       need);
1930		return -ENOMEM;
1931	}
1932
1933	rc = misc_register(&cfset_dev);
1934	if (rc) {
1935		pr_err("Registration of /dev/%s failed rc=%i\n",
1936		       cfset_dev.name, rc);
1937		goto out;
1938	}
1939
1940	rc = perf_pmu_register(&cf_diag, "cpum_cf_diag", -1);
1941	if (rc) {
1942		misc_deregister(&cfset_dev);
1943		pr_err("Registration of PMU(cpum_cf_diag) failed with rc=%i\n",
1944		       rc);
1945	}
1946out:
1947	return rc;
1948}
1949
1950device_initcall(cpumf_pmu_init);