1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Performance event support for the System z CPU-measurement Sampling Facility 4 * 5 * Copyright IBM Corp. 2013, 2018 6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com> 7 */ 8 #define KMSG_COMPONENT "cpum_sf" 9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/kernel_stat.h> 13 #include <linux/perf_event.h> 14 #include <linux/percpu.h> 15 #include <linux/pid.h> 16 #include <linux/notifier.h> 17 #include <linux/export.h> 18 #include <linux/slab.h> 19 #include <linux/mm.h> 20 #include <linux/moduleparam.h> 21 #include <asm/cpu_mf.h> 22 #include <asm/irq.h> 23 #include <asm/debug.h> 24 #include <asm/timex.h> 25 26 /* Minimum number of sample-data-block-tables: 27 * At least one table is required for the sampling buffer structure. 28 * A single table contains up to 511 pointers to sample-data-blocks. 29 */ 30 #define CPUM_SF_MIN_SDBT 1 31 32 /* Number of sample-data-blocks per sample-data-block-table (SDBT): 33 * A table contains SDB pointers (8 bytes) and one table-link entry 34 * that points to the origin of the next SDBT. 35 */ 36 #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8) 37 38 /* Maximum page offset for an SDBT table-link entry: 39 * If this page offset is reached, a table-link entry to the next SDBT 40 * must be added. 41 */ 42 #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8) 43 static inline int require_table_link(const void *sdbt) 44 { 45 return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET; 46 } 47 48 /* Minimum and maximum sampling buffer sizes: 49 * 50 * This number represents the maximum size of the sampling buffer taking 51 * the number of sample-data-block-tables into account. Note that these 52 * numbers apply to the basic-sampling function only. 53 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if 54 * the diagnostic-sampling function is active. 55 * 56 * Sampling buffer size Buffer characteristics 57 * --------------------------------------------------- 58 * 64KB == 16 pages (4KB per page) 59 * 1 page for SDB-tables 60 * 15 pages for SDBs 61 * 62 * 32MB == 8192 pages (4KB per page) 63 * 16 pages for SDB-tables 64 * 8176 pages for SDBs 65 */ 66 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15; 67 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176; 68 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1; 69 70 struct sf_buffer { 71 unsigned long *sdbt; /* Sample-data-block-table origin */ 72 /* buffer characteristics (required for buffer increments) */ 73 unsigned long num_sdb; /* Number of sample-data-blocks */ 74 unsigned long num_sdbt; /* Number of sample-data-block-tables */ 75 unsigned long *tail; /* last sample-data-block-table */ 76 }; 77 78 struct aux_buffer { 79 struct sf_buffer sfb; 80 unsigned long head; /* index of SDB of buffer head */ 81 unsigned long alert_mark; /* index of SDB of alert request position */ 82 unsigned long empty_mark; /* mark of SDB not marked full */ 83 unsigned long *sdb_index; /* SDB address for fast lookup */ 84 unsigned long *sdbt_index; /* SDBT address for fast lookup */ 85 }; 86 87 struct cpu_hw_sf { 88 /* CPU-measurement sampling information block */ 89 struct hws_qsi_info_block qsi; 90 /* CPU-measurement sampling control block */ 91 struct hws_lsctl_request_block lsctl; 92 struct sf_buffer sfb; /* Sampling buffer */ 93 unsigned int flags; /* Status flags */ 94 struct perf_event *event; /* Scheduled perf event */ 95 struct perf_output_handle handle; /* AUX buffer output handle */ 96 }; 97 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf); 98 99 /* Debug feature */ 100 static debug_info_t *sfdbg; 101 102 /* 103 * sf_disable() - Switch off sampling facility 104 */ 105 static int sf_disable(void) 106 { 107 struct hws_lsctl_request_block sreq; 108 109 memset(&sreq, 0, sizeof(sreq)); 110 return lsctl(&sreq); 111 } 112 113 /* 114 * sf_buffer_available() - Check for an allocated sampling buffer 115 */ 116 static int sf_buffer_available(struct cpu_hw_sf *cpuhw) 117 { 118 return !!cpuhw->sfb.sdbt; 119 } 120 121 /* 122 * deallocate sampling facility buffer 123 */ 124 static void free_sampling_buffer(struct sf_buffer *sfb) 125 { 126 unsigned long *sdbt, *curr; 127 128 if (!sfb->sdbt) 129 return; 130 131 sdbt = sfb->sdbt; 132 curr = sdbt; 133 134 /* Free the SDBT after all SDBs are processed... */ 135 while (1) { 136 if (!*curr || !sdbt) 137 break; 138 139 /* Process table-link entries */ 140 if (is_link_entry(curr)) { 141 curr = get_next_sdbt(curr); 142 if (sdbt) 143 free_page((unsigned long) sdbt); 144 145 /* If the origin is reached, sampling buffer is freed */ 146 if (curr == sfb->sdbt) 147 break; 148 else 149 sdbt = curr; 150 } else { 151 /* Process SDB pointer */ 152 if (*curr) { 153 free_page(*curr); 154 curr++; 155 } 156 } 157 } 158 159 debug_sprintf_event(sfdbg, 5, 160 "free_sampling_buffer: freed sdbt=%p\n", sfb->sdbt); 161 memset(sfb, 0, sizeof(*sfb)); 162 } 163 164 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags) 165 { 166 unsigned long sdb, *trailer; 167 168 /* Allocate and initialize sample-data-block */ 169 sdb = get_zeroed_page(gfp_flags); 170 if (!sdb) 171 return -ENOMEM; 172 trailer = trailer_entry_ptr(sdb); 173 *trailer = SDB_TE_ALERT_REQ_MASK; 174 175 /* Link SDB into the sample-data-block-table */ 176 *sdbt = sdb; 177 178 return 0; 179 } 180 181 /* 182 * realloc_sampling_buffer() - extend sampler memory 183 * 184 * Allocates new sample-data-blocks and adds them to the specified sampling 185 * buffer memory. 186 * 187 * Important: This modifies the sampling buffer and must be called when the 188 * sampling facility is disabled. 189 * 190 * Returns zero on success, non-zero otherwise. 191 */ 192 static int realloc_sampling_buffer(struct sf_buffer *sfb, 193 unsigned long num_sdb, gfp_t gfp_flags) 194 { 195 int i, rc; 196 unsigned long *new, *tail; 197 198 if (!sfb->sdbt || !sfb->tail) 199 return -EINVAL; 200 201 if (!is_link_entry(sfb->tail)) 202 return -EINVAL; 203 204 /* Append to the existing sampling buffer, overwriting the table-link 205 * register. 206 * The tail variables always points to the "tail" (last and table-link) 207 * entry in an SDB-table. 208 */ 209 tail = sfb->tail; 210 211 /* Do a sanity check whether the table-link entry points to 212 * the sampling buffer origin. 213 */ 214 if (sfb->sdbt != get_next_sdbt(tail)) { 215 debug_sprintf_event(sfdbg, 3, "realloc_sampling_buffer: " 216 "sampling buffer is not linked: origin=%p" 217 "tail=%p\n", 218 (void *) sfb->sdbt, (void *) tail); 219 return -EINVAL; 220 } 221 222 /* Allocate remaining SDBs */ 223 rc = 0; 224 for (i = 0; i < num_sdb; i++) { 225 /* Allocate a new SDB-table if it is full. */ 226 if (require_table_link(tail)) { 227 new = (unsigned long *) get_zeroed_page(gfp_flags); 228 if (!new) { 229 rc = -ENOMEM; 230 break; 231 } 232 sfb->num_sdbt++; 233 /* Link current page to tail of chain */ 234 *tail = (unsigned long)(void *) new + 1; 235 tail = new; 236 } 237 238 /* Allocate a new sample-data-block. 239 * If there is not enough memory, stop the realloc process 240 * and simply use what was allocated. If this is a temporary 241 * issue, a new realloc call (if required) might succeed. 242 */ 243 rc = alloc_sample_data_block(tail, gfp_flags); 244 if (rc) 245 break; 246 sfb->num_sdb++; 247 tail++; 248 } 249 250 /* Link sampling buffer to its origin */ 251 *tail = (unsigned long) sfb->sdbt + 1; 252 sfb->tail = tail; 253 254 debug_sprintf_event(sfdbg, 4, "realloc_sampling_buffer: new buffer" 255 " settings: sdbt=%lu sdb=%lu\n", 256 sfb->num_sdbt, sfb->num_sdb); 257 return rc; 258 } 259 260 /* 261 * allocate_sampling_buffer() - allocate sampler memory 262 * 263 * Allocates and initializes a sampling buffer structure using the 264 * specified number of sample-data-blocks (SDB). For each allocation, 265 * a 4K page is used. The number of sample-data-block-tables (SDBT) 266 * are calculated from SDBs. 267 * Also set the ALERT_REQ mask in each SDBs trailer. 268 * 269 * Returns zero on success, non-zero otherwise. 270 */ 271 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb) 272 { 273 int rc; 274 275 if (sfb->sdbt) 276 return -EINVAL; 277 278 /* Allocate the sample-data-block-table origin */ 279 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 280 if (!sfb->sdbt) 281 return -ENOMEM; 282 sfb->num_sdb = 0; 283 sfb->num_sdbt = 1; 284 285 /* Link the table origin to point to itself to prepare for 286 * realloc_sampling_buffer() invocation. 287 */ 288 sfb->tail = sfb->sdbt; 289 *sfb->tail = (unsigned long)(void *) sfb->sdbt + 1; 290 291 /* Allocate requested number of sample-data-blocks */ 292 rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL); 293 if (rc) { 294 free_sampling_buffer(sfb); 295 debug_sprintf_event(sfdbg, 4, "alloc_sampling_buffer: " 296 "realloc_sampling_buffer failed with rc=%i\n", rc); 297 } else 298 debug_sprintf_event(sfdbg, 4, 299 "alloc_sampling_buffer: tear=%p dear=%p\n", 300 sfb->sdbt, (void *) *sfb->sdbt); 301 return rc; 302 } 303 304 static void sfb_set_limits(unsigned long min, unsigned long max) 305 { 306 struct hws_qsi_info_block si; 307 308 CPUM_SF_MIN_SDB = min; 309 CPUM_SF_MAX_SDB = max; 310 311 memset(&si, 0, sizeof(si)); 312 if (!qsi(&si)) 313 CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes); 314 } 315 316 static unsigned long sfb_max_limit(struct hw_perf_event *hwc) 317 { 318 return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR 319 : CPUM_SF_MAX_SDB; 320 } 321 322 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb, 323 struct hw_perf_event *hwc) 324 { 325 if (!sfb->sdbt) 326 return SFB_ALLOC_REG(hwc); 327 if (SFB_ALLOC_REG(hwc) > sfb->num_sdb) 328 return SFB_ALLOC_REG(hwc) - sfb->num_sdb; 329 return 0; 330 } 331 332 static int sfb_has_pending_allocs(struct sf_buffer *sfb, 333 struct hw_perf_event *hwc) 334 { 335 return sfb_pending_allocs(sfb, hwc) > 0; 336 } 337 338 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc) 339 { 340 /* Limit the number of SDBs to not exceed the maximum */ 341 num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc)); 342 if (num) 343 SFB_ALLOC_REG(hwc) += num; 344 } 345 346 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc) 347 { 348 SFB_ALLOC_REG(hwc) = 0; 349 sfb_account_allocs(num, hwc); 350 } 351 352 static void deallocate_buffers(struct cpu_hw_sf *cpuhw) 353 { 354 if (cpuhw->sfb.sdbt) 355 free_sampling_buffer(&cpuhw->sfb); 356 } 357 358 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc) 359 { 360 unsigned long n_sdb, freq, factor; 361 size_t sample_size; 362 363 /* Calculate sampling buffers using 4K pages 364 * 365 * 1. Determine the sample data size which depends on the used 366 * sampling functions, for example, basic-sampling or 367 * basic-sampling with diagnostic-sampling. 368 * 369 * 2. Use the sampling frequency as input. The sampling buffer is 370 * designed for almost one second. This can be adjusted through 371 * the "factor" variable. 372 * In any case, alloc_sampling_buffer() sets the Alert Request 373 * Control indicator to trigger a measurement-alert to harvest 374 * sample-data-blocks (sdb). 375 * 376 * 3. Compute the number of sample-data-blocks and ensure a minimum 377 * of CPUM_SF_MIN_SDB. Also ensure the upper limit does not 378 * exceed a "calculated" maximum. The symbolic maximum is 379 * designed for basic-sampling only and needs to be increased if 380 * diagnostic-sampling is active. 381 * See also the remarks for these symbolic constants. 382 * 383 * 4. Compute the number of sample-data-block-tables (SDBT) and 384 * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up 385 * to 511 SDBs). 386 */ 387 sample_size = sizeof(struct hws_basic_entry); 388 freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)); 389 factor = 1; 390 n_sdb = DIV_ROUND_UP(freq, factor * ((PAGE_SIZE-64) / sample_size)); 391 if (n_sdb < CPUM_SF_MIN_SDB) 392 n_sdb = CPUM_SF_MIN_SDB; 393 394 /* If there is already a sampling buffer allocated, it is very likely 395 * that the sampling facility is enabled too. If the event to be 396 * initialized requires a greater sampling buffer, the allocation must 397 * be postponed. Changing the sampling buffer requires the sampling 398 * facility to be in the disabled state. So, account the number of 399 * required SDBs and let cpumsf_pmu_enable() resize the buffer just 400 * before the event is started. 401 */ 402 sfb_init_allocs(n_sdb, hwc); 403 if (sf_buffer_available(cpuhw)) 404 return 0; 405 406 debug_sprintf_event(sfdbg, 3, 407 "allocate_buffers: rate=%lu f=%lu sdb=%lu/%lu" 408 " sample_size=%lu cpuhw=%p\n", 409 SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc), 410 sample_size, cpuhw); 411 412 return alloc_sampling_buffer(&cpuhw->sfb, 413 sfb_pending_allocs(&cpuhw->sfb, hwc)); 414 } 415 416 static unsigned long min_percent(unsigned int percent, unsigned long base, 417 unsigned long min) 418 { 419 return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100)); 420 } 421 422 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base) 423 { 424 /* Use a percentage-based approach to extend the sampling facility 425 * buffer. Accept up to 5% sample data loss. 426 * Vary the extents between 1% to 5% of the current number of 427 * sample-data-blocks. 428 */ 429 if (ratio <= 5) 430 return 0; 431 if (ratio <= 25) 432 return min_percent(1, base, 1); 433 if (ratio <= 50) 434 return min_percent(1, base, 1); 435 if (ratio <= 75) 436 return min_percent(2, base, 2); 437 if (ratio <= 100) 438 return min_percent(3, base, 3); 439 if (ratio <= 250) 440 return min_percent(4, base, 4); 441 442 return min_percent(5, base, 8); 443 } 444 445 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw, 446 struct hw_perf_event *hwc) 447 { 448 unsigned long ratio, num; 449 450 if (!OVERFLOW_REG(hwc)) 451 return; 452 453 /* The sample_overflow contains the average number of sample data 454 * that has been lost because sample-data-blocks were full. 455 * 456 * Calculate the total number of sample data entries that has been 457 * discarded. Then calculate the ratio of lost samples to total samples 458 * per second in percent. 459 */ 460 ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb, 461 sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc))); 462 463 /* Compute number of sample-data-blocks */ 464 num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb); 465 if (num) 466 sfb_account_allocs(num, hwc); 467 468 debug_sprintf_event(sfdbg, 5, "sfb: overflow: overflow=%llu ratio=%lu" 469 " num=%lu\n", OVERFLOW_REG(hwc), ratio, num); 470 OVERFLOW_REG(hwc) = 0; 471 } 472 473 /* extend_sampling_buffer() - Extend sampling buffer 474 * @sfb: Sampling buffer structure (for local CPU) 475 * @hwc: Perf event hardware structure 476 * 477 * Use this function to extend the sampling buffer based on the overflow counter 478 * and postponed allocation extents stored in the specified Perf event hardware. 479 * 480 * Important: This function disables the sampling facility in order to safely 481 * change the sampling buffer structure. Do not call this function 482 * when the PMU is active. 483 */ 484 static void extend_sampling_buffer(struct sf_buffer *sfb, 485 struct hw_perf_event *hwc) 486 { 487 unsigned long num, num_old; 488 int rc; 489 490 num = sfb_pending_allocs(sfb, hwc); 491 if (!num) 492 return; 493 num_old = sfb->num_sdb; 494 495 /* Disable the sampling facility to reset any states and also 496 * clear pending measurement alerts. 497 */ 498 sf_disable(); 499 500 /* Extend the sampling buffer. 501 * This memory allocation typically happens in an atomic context when 502 * called by perf. Because this is a reallocation, it is fine if the 503 * new SDB-request cannot be satisfied immediately. 504 */ 505 rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC); 506 if (rc) 507 debug_sprintf_event(sfdbg, 5, "sfb: extend: realloc " 508 "failed with rc=%i\n", rc); 509 510 if (sfb_has_pending_allocs(sfb, hwc)) 511 debug_sprintf_event(sfdbg, 5, "sfb: extend: " 512 "req=%lu alloc=%lu remaining=%lu\n", 513 num, sfb->num_sdb - num_old, 514 sfb_pending_allocs(sfb, hwc)); 515 } 516 517 /* Number of perf events counting hardware events */ 518 static atomic_t num_events; 519 /* Used to avoid races in calling reserve/release_cpumf_hardware */ 520 static DEFINE_MUTEX(pmc_reserve_mutex); 521 522 #define PMC_INIT 0 523 #define PMC_RELEASE 1 524 #define PMC_FAILURE 2 525 static void setup_pmc_cpu(void *flags) 526 { 527 int err; 528 struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf); 529 530 err = 0; 531 switch (*((int *) flags)) { 532 case PMC_INIT: 533 memset(cpusf, 0, sizeof(*cpusf)); 534 err = qsi(&cpusf->qsi); 535 if (err) 536 break; 537 cpusf->flags |= PMU_F_RESERVED; 538 err = sf_disable(); 539 if (err) 540 pr_err("Switching off the sampling facility failed " 541 "with rc=%i\n", err); 542 debug_sprintf_event(sfdbg, 5, 543 "setup_pmc_cpu: initialized: cpuhw=%p\n", cpusf); 544 break; 545 case PMC_RELEASE: 546 cpusf->flags &= ~PMU_F_RESERVED; 547 err = sf_disable(); 548 if (err) { 549 pr_err("Switching off the sampling facility failed " 550 "with rc=%i\n", err); 551 } else 552 deallocate_buffers(cpusf); 553 debug_sprintf_event(sfdbg, 5, 554 "setup_pmc_cpu: released: cpuhw=%p\n", cpusf); 555 break; 556 } 557 if (err) 558 *((int *) flags) |= PMC_FAILURE; 559 } 560 561 static void release_pmc_hardware(void) 562 { 563 int flags = PMC_RELEASE; 564 565 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT); 566 on_each_cpu(setup_pmc_cpu, &flags, 1); 567 } 568 569 static int reserve_pmc_hardware(void) 570 { 571 int flags = PMC_INIT; 572 573 on_each_cpu(setup_pmc_cpu, &flags, 1); 574 if (flags & PMC_FAILURE) { 575 release_pmc_hardware(); 576 return -ENODEV; 577 } 578 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT); 579 580 return 0; 581 } 582 583 static void hw_perf_event_destroy(struct perf_event *event) 584 { 585 /* Release PMC if this is the last perf event */ 586 if (!atomic_add_unless(&num_events, -1, 1)) { 587 mutex_lock(&pmc_reserve_mutex); 588 if (atomic_dec_return(&num_events) == 0) 589 release_pmc_hardware(); 590 mutex_unlock(&pmc_reserve_mutex); 591 } 592 } 593 594 static void hw_init_period(struct hw_perf_event *hwc, u64 period) 595 { 596 hwc->sample_period = period; 597 hwc->last_period = hwc->sample_period; 598 local64_set(&hwc->period_left, hwc->sample_period); 599 } 600 601 static void hw_reset_registers(struct hw_perf_event *hwc, 602 unsigned long *sdbt_origin) 603 { 604 /* (Re)set to first sample-data-block-table */ 605 TEAR_REG(hwc) = (unsigned long) sdbt_origin; 606 } 607 608 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si, 609 unsigned long rate) 610 { 611 return clamp_t(unsigned long, rate, 612 si->min_sampl_rate, si->max_sampl_rate); 613 } 614 615 static u32 cpumsf_pid_type(struct perf_event *event, 616 u32 pid, enum pid_type type) 617 { 618 struct task_struct *tsk; 619 620 /* Idle process */ 621 if (!pid) 622 goto out; 623 624 tsk = find_task_by_pid_ns(pid, &init_pid_ns); 625 pid = -1; 626 if (tsk) { 627 /* 628 * Only top level events contain the pid namespace in which 629 * they are created. 630 */ 631 if (event->parent) 632 event = event->parent; 633 pid = __task_pid_nr_ns(tsk, type, event->ns); 634 /* 635 * See also 1d953111b648 636 * "perf/core: Don't report zero PIDs for exiting tasks". 637 */ 638 if (!pid && !pid_alive(tsk)) 639 pid = -1; 640 } 641 out: 642 return pid; 643 } 644 645 static void cpumsf_output_event_pid(struct perf_event *event, 646 struct perf_sample_data *data, 647 struct pt_regs *regs) 648 { 649 u32 pid; 650 struct perf_event_header header; 651 struct perf_output_handle handle; 652 653 /* 654 * Obtain the PID from the basic-sampling data entry and 655 * correct the data->tid_entry.pid value. 656 */ 657 pid = data->tid_entry.pid; 658 659 /* Protect callchain buffers, tasks */ 660 rcu_read_lock(); 661 662 perf_prepare_sample(&header, data, event, regs); 663 if (perf_output_begin(&handle, event, header.size)) 664 goto out; 665 666 /* Update the process ID (see also kernel/events/core.c) */ 667 data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID); 668 data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID); 669 670 perf_output_sample(&handle, &header, data, event); 671 perf_output_end(&handle); 672 out: 673 rcu_read_unlock(); 674 } 675 676 static int __hw_perf_event_init(struct perf_event *event) 677 { 678 struct cpu_hw_sf *cpuhw; 679 struct hws_qsi_info_block si; 680 struct perf_event_attr *attr = &event->attr; 681 struct hw_perf_event *hwc = &event->hw; 682 unsigned long rate; 683 int cpu, err; 684 685 /* Reserve CPU-measurement sampling facility */ 686 err = 0; 687 if (!atomic_inc_not_zero(&num_events)) { 688 mutex_lock(&pmc_reserve_mutex); 689 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware()) 690 err = -EBUSY; 691 else 692 atomic_inc(&num_events); 693 mutex_unlock(&pmc_reserve_mutex); 694 } 695 event->destroy = hw_perf_event_destroy; 696 697 if (err) 698 goto out; 699 700 /* Access per-CPU sampling information (query sampling info) */ 701 /* 702 * The event->cpu value can be -1 to count on every CPU, for example, 703 * when attaching to a task. If this is specified, use the query 704 * sampling info from the current CPU, otherwise use event->cpu to 705 * retrieve the per-CPU information. 706 * Later, cpuhw indicates whether to allocate sampling buffers for a 707 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL). 708 */ 709 memset(&si, 0, sizeof(si)); 710 cpuhw = NULL; 711 if (event->cpu == -1) 712 qsi(&si); 713 else { 714 /* Event is pinned to a particular CPU, retrieve the per-CPU 715 * sampling structure for accessing the CPU-specific QSI. 716 */ 717 cpuhw = &per_cpu(cpu_hw_sf, event->cpu); 718 si = cpuhw->qsi; 719 } 720 721 /* Check sampling facility authorization and, if not authorized, 722 * fall back to other PMUs. It is safe to check any CPU because 723 * the authorization is identical for all configured CPUs. 724 */ 725 if (!si.as) { 726 err = -ENOENT; 727 goto out; 728 } 729 730 /* Always enable basic sampling */ 731 SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE; 732 733 /* Check if diagnostic sampling is requested. Deny if the required 734 * sampling authorization is missing. 735 */ 736 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) { 737 if (!si.ad) { 738 err = -EPERM; 739 goto out; 740 } 741 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE; 742 } 743 744 /* Check and set other sampling flags */ 745 if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS) 746 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS; 747 748 /* The sampling information (si) contains information about the 749 * min/max sampling intervals and the CPU speed. So calculate the 750 * correct sampling interval and avoid the whole period adjust 751 * feedback loop. 752 */ 753 rate = 0; 754 if (attr->freq) { 755 if (!attr->sample_freq) { 756 err = -EINVAL; 757 goto out; 758 } 759 rate = freq_to_sample_rate(&si, attr->sample_freq); 760 rate = hw_limit_rate(&si, rate); 761 attr->freq = 0; 762 attr->sample_period = rate; 763 } else { 764 /* The min/max sampling rates specifies the valid range 765 * of sample periods. If the specified sample period is 766 * out of range, limit the period to the range boundary. 767 */ 768 rate = hw_limit_rate(&si, hwc->sample_period); 769 770 /* The perf core maintains a maximum sample rate that is 771 * configurable through the sysctl interface. Ensure the 772 * sampling rate does not exceed this value. This also helps 773 * to avoid throttling when pushing samples with 774 * perf_event_overflow(). 775 */ 776 if (sample_rate_to_freq(&si, rate) > 777 sysctl_perf_event_sample_rate) { 778 err = -EINVAL; 779 debug_sprintf_event(sfdbg, 1, "Sampling rate exceeds maximum perf sample rate\n"); 780 goto out; 781 } 782 } 783 SAMPL_RATE(hwc) = rate; 784 hw_init_period(hwc, SAMPL_RATE(hwc)); 785 786 /* Initialize sample data overflow accounting */ 787 hwc->extra_reg.reg = REG_OVERFLOW; 788 OVERFLOW_REG(hwc) = 0; 789 790 /* Use AUX buffer. No need to allocate it by ourself */ 791 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) 792 return 0; 793 794 /* Allocate the per-CPU sampling buffer using the CPU information 795 * from the event. If the event is not pinned to a particular 796 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling 797 * buffers for each online CPU. 798 */ 799 if (cpuhw) 800 /* Event is pinned to a particular CPU */ 801 err = allocate_buffers(cpuhw, hwc); 802 else { 803 /* Event is not pinned, allocate sampling buffer on 804 * each online CPU 805 */ 806 for_each_online_cpu(cpu) { 807 cpuhw = &per_cpu(cpu_hw_sf, cpu); 808 err = allocate_buffers(cpuhw, hwc); 809 if (err) 810 break; 811 } 812 } 813 814 /* If PID/TID sampling is active, replace the default overflow 815 * handler to extract and resolve the PIDs from the basic-sampling 816 * data entries. 817 */ 818 if (event->attr.sample_type & PERF_SAMPLE_TID) 819 if (is_default_overflow_handler(event)) 820 event->overflow_handler = cpumsf_output_event_pid; 821 out: 822 return err; 823 } 824 825 static int cpumsf_pmu_event_init(struct perf_event *event) 826 { 827 int err; 828 829 /* No support for taken branch sampling */ 830 if (has_branch_stack(event)) 831 return -EOPNOTSUPP; 832 833 switch (event->attr.type) { 834 case PERF_TYPE_RAW: 835 if ((event->attr.config != PERF_EVENT_CPUM_SF) && 836 (event->attr.config != PERF_EVENT_CPUM_SF_DIAG)) 837 return -ENOENT; 838 break; 839 case PERF_TYPE_HARDWARE: 840 /* Support sampling of CPU cycles in addition to the 841 * counter facility. However, the counter facility 842 * is more precise and, hence, restrict this PMU to 843 * sampling events only. 844 */ 845 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES) 846 return -ENOENT; 847 if (!is_sampling_event(event)) 848 return -ENOENT; 849 break; 850 default: 851 return -ENOENT; 852 } 853 854 /* Check online status of the CPU to which the event is pinned */ 855 if (event->cpu >= 0 && !cpu_online(event->cpu)) 856 return -ENODEV; 857 858 /* Force reset of idle/hv excludes regardless of what the 859 * user requested. 860 */ 861 if (event->attr.exclude_hv) 862 event->attr.exclude_hv = 0; 863 if (event->attr.exclude_idle) 864 event->attr.exclude_idle = 0; 865 866 err = __hw_perf_event_init(event); 867 if (unlikely(err)) 868 if (event->destroy) 869 event->destroy(event); 870 return err; 871 } 872 873 static void cpumsf_pmu_enable(struct pmu *pmu) 874 { 875 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 876 struct hw_perf_event *hwc; 877 int err; 878 879 if (cpuhw->flags & PMU_F_ENABLED) 880 return; 881 882 if (cpuhw->flags & PMU_F_ERR_MASK) 883 return; 884 885 /* Check whether to extent the sampling buffer. 886 * 887 * Two conditions trigger an increase of the sampling buffer for a 888 * perf event: 889 * 1. Postponed buffer allocations from the event initialization. 890 * 2. Sampling overflows that contribute to pending allocations. 891 * 892 * Note that the extend_sampling_buffer() function disables the sampling 893 * facility, but it can be fully re-enabled using sampling controls that 894 * have been saved in cpumsf_pmu_disable(). 895 */ 896 if (cpuhw->event) { 897 hwc = &cpuhw->event->hw; 898 if (!(SAMPL_DIAG_MODE(hwc))) { 899 /* 900 * Account number of overflow-designated 901 * buffer extents 902 */ 903 sfb_account_overflows(cpuhw, hwc); 904 if (sfb_has_pending_allocs(&cpuhw->sfb, hwc)) 905 extend_sampling_buffer(&cpuhw->sfb, hwc); 906 } 907 } 908 909 /* (Re)enable the PMU and sampling facility */ 910 cpuhw->flags |= PMU_F_ENABLED; 911 barrier(); 912 913 err = lsctl(&cpuhw->lsctl); 914 if (err) { 915 cpuhw->flags &= ~PMU_F_ENABLED; 916 pr_err("Loading sampling controls failed: op=%i err=%i\n", 917 1, err); 918 return; 919 } 920 921 /* Load current program parameter */ 922 lpp(&S390_lowcore.lpp); 923 924 debug_sprintf_event(sfdbg, 6, "pmu_enable: es=%i cs=%i ed=%i cd=%i " 925 "tear=%p dear=%p\n", cpuhw->lsctl.es, 926 cpuhw->lsctl.cs, cpuhw->lsctl.ed, cpuhw->lsctl.cd, 927 (void *) cpuhw->lsctl.tear, 928 (void *) cpuhw->lsctl.dear); 929 } 930 931 static void cpumsf_pmu_disable(struct pmu *pmu) 932 { 933 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 934 struct hws_lsctl_request_block inactive; 935 struct hws_qsi_info_block si; 936 int err; 937 938 if (!(cpuhw->flags & PMU_F_ENABLED)) 939 return; 940 941 if (cpuhw->flags & PMU_F_ERR_MASK) 942 return; 943 944 /* Switch off sampling activation control */ 945 inactive = cpuhw->lsctl; 946 inactive.cs = 0; 947 inactive.cd = 0; 948 949 err = lsctl(&inactive); 950 if (err) { 951 pr_err("Loading sampling controls failed: op=%i err=%i\n", 952 2, err); 953 return; 954 } 955 956 /* Save state of TEAR and DEAR register contents */ 957 if (!qsi(&si)) { 958 /* TEAR/DEAR values are valid only if the sampling facility is 959 * enabled. Note that cpumsf_pmu_disable() might be called even 960 * for a disabled sampling facility because cpumsf_pmu_enable() 961 * controls the enable/disable state. 962 */ 963 if (si.es) { 964 cpuhw->lsctl.tear = si.tear; 965 cpuhw->lsctl.dear = si.dear; 966 } 967 } else 968 debug_sprintf_event(sfdbg, 3, "cpumsf_pmu_disable: " 969 "qsi() failed with err=%i\n", err); 970 971 cpuhw->flags &= ~PMU_F_ENABLED; 972 } 973 974 /* perf_exclude_event() - Filter event 975 * @event: The perf event 976 * @regs: pt_regs structure 977 * @sde_regs: Sample-data-entry (sde) regs structure 978 * 979 * Filter perf events according to their exclude specification. 980 * 981 * Return non-zero if the event shall be excluded. 982 */ 983 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs, 984 struct perf_sf_sde_regs *sde_regs) 985 { 986 if (event->attr.exclude_user && user_mode(regs)) 987 return 1; 988 if (event->attr.exclude_kernel && !user_mode(regs)) 989 return 1; 990 if (event->attr.exclude_guest && sde_regs->in_guest) 991 return 1; 992 if (event->attr.exclude_host && !sde_regs->in_guest) 993 return 1; 994 return 0; 995 } 996 997 /* perf_push_sample() - Push samples to perf 998 * @event: The perf event 999 * @sample: Hardware sample data 1000 * 1001 * Use the hardware sample data to create perf event sample. The sample 1002 * is the pushed to the event subsystem and the function checks for 1003 * possible event overflows. If an event overflow occurs, the PMU is 1004 * stopped. 1005 * 1006 * Return non-zero if an event overflow occurred. 1007 */ 1008 static int perf_push_sample(struct perf_event *event, 1009 struct hws_basic_entry *basic) 1010 { 1011 int overflow; 1012 struct pt_regs regs; 1013 struct perf_sf_sde_regs *sde_regs; 1014 struct perf_sample_data data; 1015 1016 /* Setup perf sample */ 1017 perf_sample_data_init(&data, 0, event->hw.last_period); 1018 1019 /* Setup pt_regs to look like an CPU-measurement external interrupt 1020 * using the Program Request Alert code. The regs.int_parm_long 1021 * field which is unused contains additional sample-data-entry related 1022 * indicators. 1023 */ 1024 memset(®s, 0, sizeof(regs)); 1025 regs.int_code = 0x1407; 1026 regs.int_parm = CPU_MF_INT_SF_PRA; 1027 sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long; 1028 1029 psw_bits(regs.psw).ia = basic->ia; 1030 psw_bits(regs.psw).dat = basic->T; 1031 psw_bits(regs.psw).wait = basic->W; 1032 psw_bits(regs.psw).pstate = basic->P; 1033 psw_bits(regs.psw).as = basic->AS; 1034 1035 /* 1036 * Use the hardware provided configuration level to decide if the 1037 * sample belongs to a guest or host. If that is not available, 1038 * fall back to the following heuristics: 1039 * A non-zero guest program parameter always indicates a guest 1040 * sample. Some early samples or samples from guests without 1041 * lpp usage would be misaccounted to the host. We use the asn 1042 * value as an addon heuristic to detect most of these guest samples. 1043 * If the value differs from 0xffff (the host value), we assume to 1044 * be a KVM guest. 1045 */ 1046 switch (basic->CL) { 1047 case 1: /* logical partition */ 1048 sde_regs->in_guest = 0; 1049 break; 1050 case 2: /* virtual machine */ 1051 sde_regs->in_guest = 1; 1052 break; 1053 default: /* old machine, use heuristics */ 1054 if (basic->gpp || basic->prim_asn != 0xffff) 1055 sde_regs->in_guest = 1; 1056 break; 1057 } 1058 1059 /* 1060 * Store the PID value from the sample-data-entry to be 1061 * processed and resolved by cpumsf_output_event_pid(). 1062 */ 1063 data.tid_entry.pid = basic->hpp & LPP_PID_MASK; 1064 1065 overflow = 0; 1066 if (perf_exclude_event(event, ®s, sde_regs)) 1067 goto out; 1068 if (perf_event_overflow(event, &data, ®s)) { 1069 overflow = 1; 1070 event->pmu->stop(event, 0); 1071 } 1072 perf_event_update_userpage(event); 1073 out: 1074 return overflow; 1075 } 1076 1077 static void perf_event_count_update(struct perf_event *event, u64 count) 1078 { 1079 local64_add(count, &event->count); 1080 } 1081 1082 static void debug_sample_entry(struct hws_basic_entry *sample, 1083 struct hws_trailer_entry *te) 1084 { 1085 debug_sprintf_event(sfdbg, 4, "hw_collect_samples: Found unknown " 1086 "sampling data entry: te->f=%i basic.def=%04x " 1087 "(%p)\n", 1088 te->f, sample->def, sample); 1089 } 1090 1091 /* hw_collect_samples() - Walk through a sample-data-block and collect samples 1092 * @event: The perf event 1093 * @sdbt: Sample-data-block table 1094 * @overflow: Event overflow counter 1095 * 1096 * Walks through a sample-data-block and collects sampling data entries that are 1097 * then pushed to the perf event subsystem. Depending on the sampling function, 1098 * there can be either basic-sampling or combined-sampling data entries. A 1099 * combined-sampling data entry consists of a basic- and a diagnostic-sampling 1100 * data entry. The sampling function is determined by the flags in the perf 1101 * event hardware structure. The function always works with a combined-sampling 1102 * data entry but ignores the the diagnostic portion if it is not available. 1103 * 1104 * Note that the implementation focuses on basic-sampling data entries and, if 1105 * such an entry is not valid, the entire combined-sampling data entry is 1106 * ignored. 1107 * 1108 * The overflow variables counts the number of samples that has been discarded 1109 * due to a perf event overflow. 1110 */ 1111 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt, 1112 unsigned long long *overflow) 1113 { 1114 struct hws_trailer_entry *te; 1115 struct hws_basic_entry *sample; 1116 1117 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt); 1118 sample = (struct hws_basic_entry *) *sdbt; 1119 while ((unsigned long *) sample < (unsigned long *) te) { 1120 /* Check for an empty sample */ 1121 if (!sample->def) 1122 break; 1123 1124 /* Update perf event period */ 1125 perf_event_count_update(event, SAMPL_RATE(&event->hw)); 1126 1127 /* Check whether sample is valid */ 1128 if (sample->def == 0x0001) { 1129 /* If an event overflow occurred, the PMU is stopped to 1130 * throttle event delivery. Remaining sample data is 1131 * discarded. 1132 */ 1133 if (!*overflow) { 1134 /* Check whether sample is consistent */ 1135 if (sample->I == 0 && sample->W == 0) { 1136 /* Deliver sample data to perf */ 1137 *overflow = perf_push_sample(event, 1138 sample); 1139 } 1140 } else 1141 /* Count discarded samples */ 1142 *overflow += 1; 1143 } else { 1144 debug_sample_entry(sample, te); 1145 /* Sample slot is not yet written or other record. 1146 * 1147 * This condition can occur if the buffer was reused 1148 * from a combined basic- and diagnostic-sampling. 1149 * If only basic-sampling is then active, entries are 1150 * written into the larger diagnostic entries. 1151 * This is typically the case for sample-data-blocks 1152 * that are not full. Stop processing if the first 1153 * invalid format was detected. 1154 */ 1155 if (!te->f) 1156 break; 1157 } 1158 1159 /* Reset sample slot and advance to next sample */ 1160 sample->def = 0; 1161 sample++; 1162 } 1163 } 1164 1165 /* hw_perf_event_update() - Process sampling buffer 1166 * @event: The perf event 1167 * @flush_all: Flag to also flush partially filled sample-data-blocks 1168 * 1169 * Processes the sampling buffer and create perf event samples. 1170 * The sampling buffer position are retrieved and saved in the TEAR_REG 1171 * register of the specified perf event. 1172 * 1173 * Only full sample-data-blocks are processed. Specify the flash_all flag 1174 * to also walk through partially filled sample-data-blocks. It is ignored 1175 * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag 1176 * enforces the processing of full sample-data-blocks only (trailer entries 1177 * with the block-full-indicator bit set). 1178 */ 1179 static void hw_perf_event_update(struct perf_event *event, int flush_all) 1180 { 1181 struct hw_perf_event *hwc = &event->hw; 1182 struct hws_trailer_entry *te; 1183 unsigned long *sdbt; 1184 unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags; 1185 int done; 1186 1187 /* 1188 * AUX buffer is used when in diagnostic sampling mode. 1189 * No perf events/samples are created. 1190 */ 1191 if (SAMPL_DIAG_MODE(&event->hw)) 1192 return; 1193 1194 if (flush_all && SDB_FULL_BLOCKS(hwc)) 1195 flush_all = 0; 1196 1197 sdbt = (unsigned long *) TEAR_REG(hwc); 1198 done = event_overflow = sampl_overflow = num_sdb = 0; 1199 while (!done) { 1200 /* Get the trailer entry of the sample-data-block */ 1201 te = (struct hws_trailer_entry *) trailer_entry_ptr(*sdbt); 1202 1203 /* Leave loop if no more work to do (block full indicator) */ 1204 if (!te->f) { 1205 done = 1; 1206 if (!flush_all) 1207 break; 1208 } 1209 1210 /* Check the sample overflow count */ 1211 if (te->overflow) 1212 /* Account sample overflows and, if a particular limit 1213 * is reached, extend the sampling buffer. 1214 * For details, see sfb_account_overflows(). 1215 */ 1216 sampl_overflow += te->overflow; 1217 1218 /* Timestamps are valid for full sample-data-blocks only */ 1219 debug_sprintf_event(sfdbg, 6, "hw_perf_event_update: sdbt=%p " 1220 "overflow=%llu timestamp=%#llx\n", 1221 sdbt, te->overflow, 1222 (te->f) ? trailer_timestamp(te) : 0ULL); 1223 1224 /* Collect all samples from a single sample-data-block and 1225 * flag if an (perf) event overflow happened. If so, the PMU 1226 * is stopped and remaining samples will be discarded. 1227 */ 1228 hw_collect_samples(event, sdbt, &event_overflow); 1229 num_sdb++; 1230 1231 /* Reset trailer (using compare-double-and-swap) */ 1232 do { 1233 te_flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK; 1234 te_flags |= SDB_TE_ALERT_REQ_MASK; 1235 } while (!cmpxchg_double(&te->flags, &te->overflow, 1236 te->flags, te->overflow, 1237 te_flags, 0ULL)); 1238 1239 /* Advance to next sample-data-block */ 1240 sdbt++; 1241 if (is_link_entry(sdbt)) 1242 sdbt = get_next_sdbt(sdbt); 1243 1244 /* Update event hardware registers */ 1245 TEAR_REG(hwc) = (unsigned long) sdbt; 1246 1247 /* Stop processing sample-data if all samples of the current 1248 * sample-data-block were flushed even if it was not full. 1249 */ 1250 if (flush_all && done) 1251 break; 1252 1253 /* If an event overflow happened, discard samples by 1254 * processing any remaining sample-data-blocks. 1255 */ 1256 if (event_overflow) 1257 flush_all = 1; 1258 } 1259 1260 /* Account sample overflows in the event hardware structure */ 1261 if (sampl_overflow) 1262 OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) + 1263 sampl_overflow, 1 + num_sdb); 1264 if (sampl_overflow || event_overflow) 1265 debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: " 1266 "overflow stats: sample=%llu event=%llu\n", 1267 sampl_overflow, event_overflow); 1268 } 1269 1270 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb) 1271 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0) 1272 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark) 1273 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark) 1274 1275 /* 1276 * Get trailer entry by index of SDB. 1277 */ 1278 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux, 1279 unsigned long index) 1280 { 1281 unsigned long sdb; 1282 1283 index = AUX_SDB_INDEX(aux, index); 1284 sdb = aux->sdb_index[index]; 1285 return (struct hws_trailer_entry *)trailer_entry_ptr(sdb); 1286 } 1287 1288 /* 1289 * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu 1290 * disabled. Collect the full SDBs in AUX buffer which have not reached 1291 * the point of alert indicator. And ignore the SDBs which are not 1292 * full. 1293 * 1294 * 1. Scan SDBs to see how much data is there and consume them. 1295 * 2. Remove alert indicator in the buffer. 1296 */ 1297 static void aux_output_end(struct perf_output_handle *handle) 1298 { 1299 unsigned long i, range_scan, idx; 1300 struct aux_buffer *aux; 1301 struct hws_trailer_entry *te; 1302 1303 aux = perf_get_aux(handle); 1304 if (!aux) 1305 return; 1306 1307 range_scan = AUX_SDB_NUM_ALERT(aux); 1308 for (i = 0, idx = aux->head; i < range_scan; i++, idx++) { 1309 te = aux_sdb_trailer(aux, idx); 1310 if (!(te->flags & SDB_TE_BUFFER_FULL_MASK)) 1311 break; 1312 } 1313 /* i is num of SDBs which are full */ 1314 perf_aux_output_end(handle, i << PAGE_SHIFT); 1315 1316 /* Remove alert indicators in the buffer */ 1317 te = aux_sdb_trailer(aux, aux->alert_mark); 1318 te->flags &= ~SDB_TE_ALERT_REQ_MASK; 1319 1320 debug_sprintf_event(sfdbg, 6, "aux_output_end: collect %lx SDBs\n", i); 1321 } 1322 1323 /* 1324 * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event 1325 * is first added to the CPU or rescheduled again to the CPU. It is called 1326 * with pmu disabled. 1327 * 1328 * 1. Reset the trailer of SDBs to get ready for new data. 1329 * 2. Tell the hardware where to put the data by reset the SDBs buffer 1330 * head(tear/dear). 1331 */ 1332 static int aux_output_begin(struct perf_output_handle *handle, 1333 struct aux_buffer *aux, 1334 struct cpu_hw_sf *cpuhw) 1335 { 1336 unsigned long range; 1337 unsigned long i, range_scan, idx; 1338 unsigned long head, base, offset; 1339 struct hws_trailer_entry *te; 1340 1341 if (WARN_ON_ONCE(handle->head & ~PAGE_MASK)) 1342 return -EINVAL; 1343 1344 aux->head = handle->head >> PAGE_SHIFT; 1345 range = (handle->size + 1) >> PAGE_SHIFT; 1346 if (range <= 1) 1347 return -ENOMEM; 1348 1349 /* 1350 * SDBs between aux->head and aux->empty_mark are already ready 1351 * for new data. range_scan is num of SDBs not within them. 1352 */ 1353 if (range > AUX_SDB_NUM_EMPTY(aux)) { 1354 range_scan = range - AUX_SDB_NUM_EMPTY(aux); 1355 idx = aux->empty_mark + 1; 1356 for (i = 0; i < range_scan; i++, idx++) { 1357 te = aux_sdb_trailer(aux, idx); 1358 te->flags = te->flags & ~SDB_TE_BUFFER_FULL_MASK; 1359 te->flags = te->flags & ~SDB_TE_ALERT_REQ_MASK; 1360 te->overflow = 0; 1361 } 1362 /* Save the position of empty SDBs */ 1363 aux->empty_mark = aux->head + range - 1; 1364 } 1365 1366 /* Set alert indicator */ 1367 aux->alert_mark = aux->head + range/2 - 1; 1368 te = aux_sdb_trailer(aux, aux->alert_mark); 1369 te->flags = te->flags | SDB_TE_ALERT_REQ_MASK; 1370 1371 /* Reset hardware buffer head */ 1372 head = AUX_SDB_INDEX(aux, aux->head); 1373 base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE]; 1374 offset = head % CPUM_SF_SDB_PER_TABLE; 1375 cpuhw->lsctl.tear = base + offset * sizeof(unsigned long); 1376 cpuhw->lsctl.dear = aux->sdb_index[head]; 1377 1378 debug_sprintf_event(sfdbg, 6, "aux_output_begin: " 1379 "head->alert_mark->empty_mark (num_alert, range)" 1380 "[%lx -> %lx -> %lx] (%lx, %lx) " 1381 "tear index %lx, tear %lx dear %lx\n", 1382 aux->head, aux->alert_mark, aux->empty_mark, 1383 AUX_SDB_NUM_ALERT(aux), range, 1384 head / CPUM_SF_SDB_PER_TABLE, 1385 cpuhw->lsctl.tear, 1386 cpuhw->lsctl.dear); 1387 1388 return 0; 1389 } 1390 1391 /* 1392 * Set alert indicator on SDB at index @alert_index while sampler is running. 1393 * 1394 * Return true if successfully. 1395 * Return false if full indicator is already set by hardware sampler. 1396 */ 1397 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index, 1398 unsigned long long *overflow) 1399 { 1400 unsigned long long orig_overflow, orig_flags, new_flags; 1401 struct hws_trailer_entry *te; 1402 1403 te = aux_sdb_trailer(aux, alert_index); 1404 do { 1405 orig_flags = te->flags; 1406 orig_overflow = te->overflow; 1407 *overflow = orig_overflow; 1408 if (orig_flags & SDB_TE_BUFFER_FULL_MASK) { 1409 /* 1410 * SDB is already set by hardware. 1411 * Abort and try to set somewhere 1412 * behind. 1413 */ 1414 return false; 1415 } 1416 new_flags = orig_flags | SDB_TE_ALERT_REQ_MASK; 1417 } while (!cmpxchg_double(&te->flags, &te->overflow, 1418 orig_flags, orig_overflow, 1419 new_flags, 0ULL)); 1420 return true; 1421 } 1422 1423 /* 1424 * aux_reset_buffer() - Scan and setup SDBs for new samples 1425 * @aux: The AUX buffer to set 1426 * @range: The range of SDBs to scan started from aux->head 1427 * @overflow: Set to overflow count 1428 * 1429 * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is 1430 * marked as empty, check if it is already set full by the hardware sampler. 1431 * If yes, that means new data is already there before we can set an alert 1432 * indicator. Caller should try to set alert indicator to some position behind. 1433 * 1434 * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used 1435 * previously and have already been consumed by user space. Reset these SDBs 1436 * (clear full indicator and alert indicator) for new data. 1437 * If aux->alert_mark fall in this area, just set it. Overflow count is 1438 * recorded while scanning. 1439 * 1440 * SDBs between aux->head and aux->empty_mark are already reset at last time. 1441 * and ready for new samples. So scanning on this area could be skipped. 1442 * 1443 * Return true if alert indicator is set successfully and false if not. 1444 */ 1445 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range, 1446 unsigned long long *overflow) 1447 { 1448 unsigned long long orig_overflow, orig_flags, new_flags; 1449 unsigned long i, range_scan, idx; 1450 struct hws_trailer_entry *te; 1451 1452 if (range <= AUX_SDB_NUM_EMPTY(aux)) 1453 /* 1454 * No need to scan. All SDBs in range are marked as empty. 1455 * Just set alert indicator. Should check race with hardware 1456 * sampler. 1457 */ 1458 return aux_set_alert(aux, aux->alert_mark, overflow); 1459 1460 if (aux->alert_mark <= aux->empty_mark) 1461 /* 1462 * Set alert indicator on empty SDB. Should check race 1463 * with hardware sampler. 1464 */ 1465 if (!aux_set_alert(aux, aux->alert_mark, overflow)) 1466 return false; 1467 1468 /* 1469 * Scan the SDBs to clear full and alert indicator used previously. 1470 * Start scanning from one SDB behind empty_mark. If the new alert 1471 * indicator fall into this range, set it. 1472 */ 1473 range_scan = range - AUX_SDB_NUM_EMPTY(aux); 1474 idx = aux->empty_mark + 1; 1475 for (i = 0; i < range_scan; i++, idx++) { 1476 te = aux_sdb_trailer(aux, idx); 1477 do { 1478 orig_flags = te->flags; 1479 orig_overflow = te->overflow; 1480 new_flags = orig_flags & ~SDB_TE_BUFFER_FULL_MASK; 1481 if (idx == aux->alert_mark) 1482 new_flags |= SDB_TE_ALERT_REQ_MASK; 1483 else 1484 new_flags &= ~SDB_TE_ALERT_REQ_MASK; 1485 } while (!cmpxchg_double(&te->flags, &te->overflow, 1486 orig_flags, orig_overflow, 1487 new_flags, 0ULL)); 1488 *overflow += orig_overflow; 1489 } 1490 1491 /* Update empty_mark to new position */ 1492 aux->empty_mark = aux->head + range - 1; 1493 1494 return true; 1495 } 1496 1497 /* 1498 * Measurement alert handler for diagnostic mode sampling. 1499 */ 1500 static void hw_collect_aux(struct cpu_hw_sf *cpuhw) 1501 { 1502 struct aux_buffer *aux; 1503 int done = 0; 1504 unsigned long range = 0, size; 1505 unsigned long long overflow = 0; 1506 struct perf_output_handle *handle = &cpuhw->handle; 1507 unsigned long num_sdb; 1508 1509 aux = perf_get_aux(handle); 1510 if (WARN_ON_ONCE(!aux)) 1511 return; 1512 1513 /* Inform user space new data arrived */ 1514 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; 1515 perf_aux_output_end(handle, size); 1516 num_sdb = aux->sfb.num_sdb; 1517 1518 while (!done) { 1519 /* Get an output handle */ 1520 aux = perf_aux_output_begin(handle, cpuhw->event); 1521 if (handle->size == 0) { 1522 pr_err("The AUX buffer with %lu pages for the " 1523 "diagnostic-sampling mode is full\n", 1524 num_sdb); 1525 debug_sprintf_event(sfdbg, 1, "AUX buffer used up\n"); 1526 break; 1527 } 1528 if (WARN_ON_ONCE(!aux)) 1529 return; 1530 1531 /* Update head and alert_mark to new position */ 1532 aux->head = handle->head >> PAGE_SHIFT; 1533 range = (handle->size + 1) >> PAGE_SHIFT; 1534 if (range == 1) 1535 aux->alert_mark = aux->head; 1536 else 1537 aux->alert_mark = aux->head + range/2 - 1; 1538 1539 if (aux_reset_buffer(aux, range, &overflow)) { 1540 if (!overflow) { 1541 done = 1; 1542 break; 1543 } 1544 size = range << PAGE_SHIFT; 1545 perf_aux_output_end(&cpuhw->handle, size); 1546 pr_err("Sample data caused the AUX buffer with %lu " 1547 "pages to overflow\n", num_sdb); 1548 debug_sprintf_event(sfdbg, 1, "head %lx range %lx " 1549 "overflow %llx\n", 1550 aux->head, range, overflow); 1551 } else { 1552 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; 1553 perf_aux_output_end(&cpuhw->handle, size); 1554 debug_sprintf_event(sfdbg, 6, "head %lx alert %lx " 1555 "already full, try another\n", 1556 aux->head, aux->alert_mark); 1557 } 1558 } 1559 1560 if (done) 1561 debug_sprintf_event(sfdbg, 6, "aux_reset_buffer: " 1562 "[%lx -> %lx -> %lx] (%lx, %lx)\n", 1563 aux->head, aux->alert_mark, aux->empty_mark, 1564 AUX_SDB_NUM_ALERT(aux), range); 1565 } 1566 1567 /* 1568 * Callback when freeing AUX buffers. 1569 */ 1570 static void aux_buffer_free(void *data) 1571 { 1572 struct aux_buffer *aux = data; 1573 unsigned long i, num_sdbt; 1574 1575 if (!aux) 1576 return; 1577 1578 /* Free SDBT. SDB is freed by the caller */ 1579 num_sdbt = aux->sfb.num_sdbt; 1580 for (i = 0; i < num_sdbt; i++) 1581 free_page(aux->sdbt_index[i]); 1582 1583 kfree(aux->sdbt_index); 1584 kfree(aux->sdb_index); 1585 kfree(aux); 1586 1587 debug_sprintf_event(sfdbg, 4, "aux_buffer_free: free " 1588 "%lu SDBTs\n", num_sdbt); 1589 } 1590 1591 static void aux_sdb_init(unsigned long sdb) 1592 { 1593 struct hws_trailer_entry *te; 1594 1595 te = (struct hws_trailer_entry *)trailer_entry_ptr(sdb); 1596 1597 /* Save clock base */ 1598 te->clock_base = 1; 1599 memcpy(&te->progusage2, &tod_clock_base[1], 8); 1600 } 1601 1602 /* 1603 * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling 1604 * @event: Event the buffer is setup for, event->cpu == -1 means current 1605 * @pages: Array of pointers to buffer pages passed from perf core 1606 * @nr_pages: Total pages 1607 * @snapshot: Flag for snapshot mode 1608 * 1609 * This is the callback when setup an event using AUX buffer. Perf tool can 1610 * trigger this by an additional mmap() call on the event. Unlike the buffer 1611 * for basic samples, AUX buffer belongs to the event. It is scheduled with 1612 * the task among online cpus when it is a per-thread event. 1613 * 1614 * Return the private AUX buffer structure if success or NULL if fails. 1615 */ 1616 static void *aux_buffer_setup(struct perf_event *event, void **pages, 1617 int nr_pages, bool snapshot) 1618 { 1619 struct sf_buffer *sfb; 1620 struct aux_buffer *aux; 1621 unsigned long *new, *tail; 1622 int i, n_sdbt; 1623 1624 if (!nr_pages || !pages) 1625 return NULL; 1626 1627 if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1628 pr_err("AUX buffer size (%i pages) is larger than the " 1629 "maximum sampling buffer limit\n", 1630 nr_pages); 1631 return NULL; 1632 } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1633 pr_err("AUX buffer size (%i pages) is less than the " 1634 "minimum sampling buffer limit\n", 1635 nr_pages); 1636 return NULL; 1637 } 1638 1639 /* Allocate aux_buffer struct for the event */ 1640 aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL); 1641 if (!aux) 1642 goto no_aux; 1643 sfb = &aux->sfb; 1644 1645 /* Allocate sdbt_index for fast reference */ 1646 n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE; 1647 aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL); 1648 if (!aux->sdbt_index) 1649 goto no_sdbt_index; 1650 1651 /* Allocate sdb_index for fast reference */ 1652 aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL); 1653 if (!aux->sdb_index) 1654 goto no_sdb_index; 1655 1656 /* Allocate the first SDBT */ 1657 sfb->num_sdbt = 0; 1658 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1659 if (!sfb->sdbt) 1660 goto no_sdbt; 1661 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt; 1662 tail = sfb->tail = sfb->sdbt; 1663 1664 /* 1665 * Link the provided pages of AUX buffer to SDBT. 1666 * Allocate SDBT if needed. 1667 */ 1668 for (i = 0; i < nr_pages; i++, tail++) { 1669 if (require_table_link(tail)) { 1670 new = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1671 if (!new) 1672 goto no_sdbt; 1673 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new; 1674 /* Link current page to tail of chain */ 1675 *tail = (unsigned long)(void *) new + 1; 1676 tail = new; 1677 } 1678 /* Tail is the entry in a SDBT */ 1679 *tail = (unsigned long)pages[i]; 1680 aux->sdb_index[i] = (unsigned long)pages[i]; 1681 aux_sdb_init((unsigned long)pages[i]); 1682 } 1683 sfb->num_sdb = nr_pages; 1684 1685 /* Link the last entry in the SDBT to the first SDBT */ 1686 *tail = (unsigned long) sfb->sdbt + 1; 1687 sfb->tail = tail; 1688 1689 /* 1690 * Initial all SDBs are zeroed. Mark it as empty. 1691 * So there is no need to clear the full indicator 1692 * when this event is first added. 1693 */ 1694 aux->empty_mark = sfb->num_sdb - 1; 1695 1696 debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs" 1697 " and %lu SDBs\n", 1698 sfb->num_sdbt, sfb->num_sdb); 1699 1700 return aux; 1701 1702 no_sdbt: 1703 /* SDBs (AUX buffer pages) are freed by caller */ 1704 for (i = 0; i < sfb->num_sdbt; i++) 1705 free_page(aux->sdbt_index[i]); 1706 kfree(aux->sdb_index); 1707 no_sdb_index: 1708 kfree(aux->sdbt_index); 1709 no_sdbt_index: 1710 kfree(aux); 1711 no_aux: 1712 return NULL; 1713 } 1714 1715 static void cpumsf_pmu_read(struct perf_event *event) 1716 { 1717 /* Nothing to do ... updates are interrupt-driven */ 1718 } 1719 1720 /* Activate sampling control. 1721 * Next call of pmu_enable() starts sampling. 1722 */ 1723 static void cpumsf_pmu_start(struct perf_event *event, int flags) 1724 { 1725 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1726 1727 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) 1728 return; 1729 1730 if (flags & PERF_EF_RELOAD) 1731 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); 1732 1733 perf_pmu_disable(event->pmu); 1734 event->hw.state = 0; 1735 cpuhw->lsctl.cs = 1; 1736 if (SAMPL_DIAG_MODE(&event->hw)) 1737 cpuhw->lsctl.cd = 1; 1738 perf_pmu_enable(event->pmu); 1739 } 1740 1741 /* Deactivate sampling control. 1742 * Next call of pmu_enable() stops sampling. 1743 */ 1744 static void cpumsf_pmu_stop(struct perf_event *event, int flags) 1745 { 1746 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1747 1748 if (event->hw.state & PERF_HES_STOPPED) 1749 return; 1750 1751 perf_pmu_disable(event->pmu); 1752 cpuhw->lsctl.cs = 0; 1753 cpuhw->lsctl.cd = 0; 1754 event->hw.state |= PERF_HES_STOPPED; 1755 1756 if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) { 1757 hw_perf_event_update(event, 1); 1758 event->hw.state |= PERF_HES_UPTODATE; 1759 } 1760 perf_pmu_enable(event->pmu); 1761 } 1762 1763 static int cpumsf_pmu_add(struct perf_event *event, int flags) 1764 { 1765 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1766 struct aux_buffer *aux; 1767 int err; 1768 1769 if (cpuhw->flags & PMU_F_IN_USE) 1770 return -EAGAIN; 1771 1772 if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt) 1773 return -EINVAL; 1774 1775 err = 0; 1776 perf_pmu_disable(event->pmu); 1777 1778 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED; 1779 1780 /* Set up sampling controls. Always program the sampling register 1781 * using the SDB-table start. Reset TEAR_REG event hardware register 1782 * that is used by hw_perf_event_update() to store the sampling buffer 1783 * position after samples have been flushed. 1784 */ 1785 cpuhw->lsctl.s = 0; 1786 cpuhw->lsctl.h = 1; 1787 cpuhw->lsctl.interval = SAMPL_RATE(&event->hw); 1788 if (!SAMPL_DIAG_MODE(&event->hw)) { 1789 cpuhw->lsctl.tear = (unsigned long) cpuhw->sfb.sdbt; 1790 cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt; 1791 hw_reset_registers(&event->hw, cpuhw->sfb.sdbt); 1792 } 1793 1794 /* Ensure sampling functions are in the disabled state. If disabled, 1795 * switch on sampling enable control. */ 1796 if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) { 1797 err = -EAGAIN; 1798 goto out; 1799 } 1800 if (SAMPL_DIAG_MODE(&event->hw)) { 1801 aux = perf_aux_output_begin(&cpuhw->handle, event); 1802 if (!aux) { 1803 err = -EINVAL; 1804 goto out; 1805 } 1806 err = aux_output_begin(&cpuhw->handle, aux, cpuhw); 1807 if (err) 1808 goto out; 1809 cpuhw->lsctl.ed = 1; 1810 } 1811 cpuhw->lsctl.es = 1; 1812 1813 /* Set in_use flag and store event */ 1814 cpuhw->event = event; 1815 cpuhw->flags |= PMU_F_IN_USE; 1816 1817 if (flags & PERF_EF_START) 1818 cpumsf_pmu_start(event, PERF_EF_RELOAD); 1819 out: 1820 perf_event_update_userpage(event); 1821 perf_pmu_enable(event->pmu); 1822 return err; 1823 } 1824 1825 static void cpumsf_pmu_del(struct perf_event *event, int flags) 1826 { 1827 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1828 1829 perf_pmu_disable(event->pmu); 1830 cpumsf_pmu_stop(event, PERF_EF_UPDATE); 1831 1832 cpuhw->lsctl.es = 0; 1833 cpuhw->lsctl.ed = 0; 1834 cpuhw->flags &= ~PMU_F_IN_USE; 1835 cpuhw->event = NULL; 1836 1837 if (SAMPL_DIAG_MODE(&event->hw)) 1838 aux_output_end(&cpuhw->handle); 1839 perf_event_update_userpage(event); 1840 perf_pmu_enable(event->pmu); 1841 } 1842 1843 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF); 1844 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG); 1845 1846 /* Attribute list for CPU_SF. 1847 * 1848 * The availablitiy depends on the CPU_MF sampling facility authorization 1849 * for basic + diagnositic samples. This is determined at initialization 1850 * time by the sampling facility device driver. 1851 * If the authorization for basic samples is turned off, it should be 1852 * also turned off for diagnostic sampling. 1853 * 1854 * During initialization of the device driver, check the authorization 1855 * level for diagnostic sampling and installs the attribute 1856 * file for diagnostic sampling if necessary. 1857 * 1858 * For now install a placeholder to reference all possible attributes: 1859 * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG. 1860 * Add another entry for the final NULL pointer. 1861 */ 1862 enum { 1863 SF_CYCLES_BASIC_ATTR_IDX = 0, 1864 SF_CYCLES_BASIC_DIAG_ATTR_IDX, 1865 SF_CYCLES_ATTR_MAX 1866 }; 1867 1868 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = { 1869 [SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC) 1870 }; 1871 1872 PMU_FORMAT_ATTR(event, "config:0-63"); 1873 1874 static struct attribute *cpumsf_pmu_format_attr[] = { 1875 &format_attr_event.attr, 1876 NULL, 1877 }; 1878 1879 static struct attribute_group cpumsf_pmu_events_group = { 1880 .name = "events", 1881 .attrs = cpumsf_pmu_events_attr, 1882 }; 1883 1884 static struct attribute_group cpumsf_pmu_format_group = { 1885 .name = "format", 1886 .attrs = cpumsf_pmu_format_attr, 1887 }; 1888 1889 static const struct attribute_group *cpumsf_pmu_attr_groups[] = { 1890 &cpumsf_pmu_events_group, 1891 &cpumsf_pmu_format_group, 1892 NULL, 1893 }; 1894 1895 static struct pmu cpumf_sampling = { 1896 .pmu_enable = cpumsf_pmu_enable, 1897 .pmu_disable = cpumsf_pmu_disable, 1898 1899 .event_init = cpumsf_pmu_event_init, 1900 .add = cpumsf_pmu_add, 1901 .del = cpumsf_pmu_del, 1902 1903 .start = cpumsf_pmu_start, 1904 .stop = cpumsf_pmu_stop, 1905 .read = cpumsf_pmu_read, 1906 1907 .attr_groups = cpumsf_pmu_attr_groups, 1908 1909 .setup_aux = aux_buffer_setup, 1910 .free_aux = aux_buffer_free, 1911 }; 1912 1913 static void cpumf_measurement_alert(struct ext_code ext_code, 1914 unsigned int alert, unsigned long unused) 1915 { 1916 struct cpu_hw_sf *cpuhw; 1917 1918 if (!(alert & CPU_MF_INT_SF_MASK)) 1919 return; 1920 inc_irq_stat(IRQEXT_CMS); 1921 cpuhw = this_cpu_ptr(&cpu_hw_sf); 1922 1923 /* Measurement alerts are shared and might happen when the PMU 1924 * is not reserved. Ignore these alerts in this case. */ 1925 if (!(cpuhw->flags & PMU_F_RESERVED)) 1926 return; 1927 1928 /* The processing below must take care of multiple alert events that 1929 * might be indicated concurrently. */ 1930 1931 /* Program alert request */ 1932 if (alert & CPU_MF_INT_SF_PRA) { 1933 if (cpuhw->flags & PMU_F_IN_USE) 1934 if (SAMPL_DIAG_MODE(&cpuhw->event->hw)) 1935 hw_collect_aux(cpuhw); 1936 else 1937 hw_perf_event_update(cpuhw->event, 0); 1938 else 1939 WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE)); 1940 } 1941 1942 /* Report measurement alerts only for non-PRA codes */ 1943 if (alert != CPU_MF_INT_SF_PRA) 1944 debug_sprintf_event(sfdbg, 6, "measurement alert: %#x\n", 1945 alert); 1946 1947 /* Sampling authorization change request */ 1948 if (alert & CPU_MF_INT_SF_SACA) 1949 qsi(&cpuhw->qsi); 1950 1951 /* Loss of sample data due to high-priority machine activities */ 1952 if (alert & CPU_MF_INT_SF_LSDA) { 1953 pr_err("Sample data was lost\n"); 1954 cpuhw->flags |= PMU_F_ERR_LSDA; 1955 sf_disable(); 1956 } 1957 1958 /* Invalid sampling buffer entry */ 1959 if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) { 1960 pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n", 1961 alert); 1962 cpuhw->flags |= PMU_F_ERR_IBE; 1963 sf_disable(); 1964 } 1965 } 1966 1967 static int cpusf_pmu_setup(unsigned int cpu, int flags) 1968 { 1969 /* Ignore the notification if no events are scheduled on the PMU. 1970 * This might be racy... 1971 */ 1972 if (!atomic_read(&num_events)) 1973 return 0; 1974 1975 local_irq_disable(); 1976 setup_pmc_cpu(&flags); 1977 local_irq_enable(); 1978 return 0; 1979 } 1980 1981 static int s390_pmu_sf_online_cpu(unsigned int cpu) 1982 { 1983 return cpusf_pmu_setup(cpu, PMC_INIT); 1984 } 1985 1986 static int s390_pmu_sf_offline_cpu(unsigned int cpu) 1987 { 1988 return cpusf_pmu_setup(cpu, PMC_RELEASE); 1989 } 1990 1991 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp) 1992 { 1993 if (!cpum_sf_avail()) 1994 return -ENODEV; 1995 return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 1996 } 1997 1998 static int param_set_sfb_size(const char *val, const struct kernel_param *kp) 1999 { 2000 int rc; 2001 unsigned long min, max; 2002 2003 if (!cpum_sf_avail()) 2004 return -ENODEV; 2005 if (!val || !strlen(val)) 2006 return -EINVAL; 2007 2008 /* Valid parameter values: "min,max" or "max" */ 2009 min = CPUM_SF_MIN_SDB; 2010 max = CPUM_SF_MAX_SDB; 2011 if (strchr(val, ',')) 2012 rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL; 2013 else 2014 rc = kstrtoul(val, 10, &max); 2015 2016 if (min < 2 || min >= max || max > get_num_physpages()) 2017 rc = -EINVAL; 2018 if (rc) 2019 return rc; 2020 2021 sfb_set_limits(min, max); 2022 pr_info("The sampling buffer limits have changed to: " 2023 "min=%lu max=%lu (diag=x%lu)\n", 2024 CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR); 2025 return 0; 2026 } 2027 2028 #define param_check_sfb_size(name, p) __param_check(name, p, void) 2029 static const struct kernel_param_ops param_ops_sfb_size = { 2030 .set = param_set_sfb_size, 2031 .get = param_get_sfb_size, 2032 }; 2033 2034 #define RS_INIT_FAILURE_QSI 0x0001 2035 #define RS_INIT_FAILURE_BSDES 0x0002 2036 #define RS_INIT_FAILURE_ALRT 0x0003 2037 #define RS_INIT_FAILURE_PERF 0x0004 2038 static void __init pr_cpumsf_err(unsigned int reason) 2039 { 2040 pr_err("Sampling facility support for perf is not available: " 2041 "reason=%04x\n", reason); 2042 } 2043 2044 static int __init init_cpum_sampling_pmu(void) 2045 { 2046 struct hws_qsi_info_block si; 2047 int err; 2048 2049 if (!cpum_sf_avail()) 2050 return -ENODEV; 2051 2052 memset(&si, 0, sizeof(si)); 2053 if (qsi(&si)) { 2054 pr_cpumsf_err(RS_INIT_FAILURE_QSI); 2055 return -ENODEV; 2056 } 2057 2058 if (!si.as && !si.ad) 2059 return -ENODEV; 2060 2061 if (si.bsdes != sizeof(struct hws_basic_entry)) { 2062 pr_cpumsf_err(RS_INIT_FAILURE_BSDES); 2063 return -EINVAL; 2064 } 2065 2066 if (si.ad) { 2067 sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 2068 /* Sampling of diagnostic data authorized, 2069 * install event into attribute list of PMU device. 2070 */ 2071 cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] = 2072 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG); 2073 } 2074 2075 sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80); 2076 if (!sfdbg) { 2077 pr_err("Registering for s390dbf failed\n"); 2078 return -ENOMEM; 2079 } 2080 debug_register_view(sfdbg, &debug_sprintf_view); 2081 2082 err = register_external_irq(EXT_IRQ_MEASURE_ALERT, 2083 cpumf_measurement_alert); 2084 if (err) { 2085 pr_cpumsf_err(RS_INIT_FAILURE_ALRT); 2086 debug_unregister(sfdbg); 2087 goto out; 2088 } 2089 2090 err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW); 2091 if (err) { 2092 pr_cpumsf_err(RS_INIT_FAILURE_PERF); 2093 unregister_external_irq(EXT_IRQ_MEASURE_ALERT, 2094 cpumf_measurement_alert); 2095 debug_unregister(sfdbg); 2096 goto out; 2097 } 2098 2099 cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online", 2100 s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu); 2101 out: 2102 return err; 2103 } 2104 2105 arch_initcall(init_cpum_sampling_pmu); 2106 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0640); 2107