1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Performance event support for the System z CPU-measurement Sampling Facility 4 * 5 * Copyright IBM Corp. 2013, 2018 6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com> 7 */ 8 #define KMSG_COMPONENT "cpum_sf" 9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/kernel_stat.h> 13 #include <linux/perf_event.h> 14 #include <linux/percpu.h> 15 #include <linux/pid.h> 16 #include <linux/notifier.h> 17 #include <linux/export.h> 18 #include <linux/slab.h> 19 #include <linux/mm.h> 20 #include <linux/moduleparam.h> 21 #include <asm/cpu_mf.h> 22 #include <asm/irq.h> 23 #include <asm/debug.h> 24 #include <asm/timex.h> 25 #include <asm-generic/io.h> 26 27 /* Minimum number of sample-data-block-tables: 28 * At least one table is required for the sampling buffer structure. 29 * A single table contains up to 511 pointers to sample-data-blocks. 30 */ 31 #define CPUM_SF_MIN_SDBT 1 32 33 /* Number of sample-data-blocks per sample-data-block-table (SDBT): 34 * A table contains SDB pointers (8 bytes) and one table-link entry 35 * that points to the origin of the next SDBT. 36 */ 37 #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8) 38 39 /* Maximum page offset for an SDBT table-link entry: 40 * If this page offset is reached, a table-link entry to the next SDBT 41 * must be added. 42 */ 43 #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8) 44 static inline int require_table_link(const void *sdbt) 45 { 46 return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET; 47 } 48 49 /* Minimum and maximum sampling buffer sizes: 50 * 51 * This number represents the maximum size of the sampling buffer taking 52 * the number of sample-data-block-tables into account. Note that these 53 * numbers apply to the basic-sampling function only. 54 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if 55 * the diagnostic-sampling function is active. 56 * 57 * Sampling buffer size Buffer characteristics 58 * --------------------------------------------------- 59 * 64KB == 16 pages (4KB per page) 60 * 1 page for SDB-tables 61 * 15 pages for SDBs 62 * 63 * 32MB == 8192 pages (4KB per page) 64 * 16 pages for SDB-tables 65 * 8176 pages for SDBs 66 */ 67 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15; 68 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176; 69 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1; 70 71 struct sf_buffer { 72 unsigned long *sdbt; /* Sample-data-block-table origin */ 73 /* buffer characteristics (required for buffer increments) */ 74 unsigned long num_sdb; /* Number of sample-data-blocks */ 75 unsigned long num_sdbt; /* Number of sample-data-block-tables */ 76 unsigned long *tail; /* last sample-data-block-table */ 77 }; 78 79 struct aux_buffer { 80 struct sf_buffer sfb; 81 unsigned long head; /* index of SDB of buffer head */ 82 unsigned long alert_mark; /* index of SDB of alert request position */ 83 unsigned long empty_mark; /* mark of SDB not marked full */ 84 unsigned long *sdb_index; /* SDB address for fast lookup */ 85 unsigned long *sdbt_index; /* SDBT address for fast lookup */ 86 }; 87 88 struct cpu_hw_sf { 89 /* CPU-measurement sampling information block */ 90 struct hws_qsi_info_block qsi; 91 /* CPU-measurement sampling control block */ 92 struct hws_lsctl_request_block lsctl; 93 struct sf_buffer sfb; /* Sampling buffer */ 94 unsigned int flags; /* Status flags */ 95 struct perf_event *event; /* Scheduled perf event */ 96 struct perf_output_handle handle; /* AUX buffer output handle */ 97 }; 98 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf); 99 100 /* Debug feature */ 101 static debug_info_t *sfdbg; 102 103 /* Sampling control helper functions */ 104 static inline unsigned long freq_to_sample_rate(struct hws_qsi_info_block *qsi, 105 unsigned long freq) 106 { 107 return (USEC_PER_SEC / freq) * qsi->cpu_speed; 108 } 109 110 static inline unsigned long sample_rate_to_freq(struct hws_qsi_info_block *qsi, 111 unsigned long rate) 112 { 113 return USEC_PER_SEC * qsi->cpu_speed / rate; 114 } 115 116 /* Return TOD timestamp contained in an trailer entry */ 117 static inline unsigned long long trailer_timestamp(struct hws_trailer_entry *te) 118 { 119 /* TOD in STCKE format */ 120 if (te->header.t) 121 return *((unsigned long long *)&te->timestamp[1]); 122 123 /* TOD in STCK format */ 124 return *((unsigned long long *)&te->timestamp[0]); 125 } 126 127 /* Return pointer to trailer entry of an sample data block */ 128 static inline struct hws_trailer_entry *trailer_entry_ptr(unsigned long v) 129 { 130 void *ret; 131 132 ret = (void *)v; 133 ret += PAGE_SIZE; 134 ret -= sizeof(struct hws_trailer_entry); 135 136 return ret; 137 } 138 139 /* 140 * Return true if the entry in the sample data block table (sdbt) 141 * is a link to the next sdbt 142 */ 143 static inline int is_link_entry(unsigned long *s) 144 { 145 return *s & 0x1UL ? 1 : 0; 146 } 147 148 /* Return pointer to the linked sdbt */ 149 static inline unsigned long *get_next_sdbt(unsigned long *s) 150 { 151 return phys_to_virt(*s & ~0x1UL); 152 } 153 154 /* 155 * sf_disable() - Switch off sampling facility 156 */ 157 static int sf_disable(void) 158 { 159 struct hws_lsctl_request_block sreq; 160 161 memset(&sreq, 0, sizeof(sreq)); 162 return lsctl(&sreq); 163 } 164 165 /* 166 * sf_buffer_available() - Check for an allocated sampling buffer 167 */ 168 static int sf_buffer_available(struct cpu_hw_sf *cpuhw) 169 { 170 return !!cpuhw->sfb.sdbt; 171 } 172 173 /* 174 * deallocate sampling facility buffer 175 */ 176 static void free_sampling_buffer(struct sf_buffer *sfb) 177 { 178 unsigned long *sdbt, *curr; 179 180 if (!sfb->sdbt) 181 return; 182 183 sdbt = sfb->sdbt; 184 curr = sdbt; 185 186 /* Free the SDBT after all SDBs are processed... */ 187 while (1) { 188 if (!*curr || !sdbt) 189 break; 190 191 /* Process table-link entries */ 192 if (is_link_entry(curr)) { 193 curr = get_next_sdbt(curr); 194 if (sdbt) 195 free_page((unsigned long) sdbt); 196 197 /* If the origin is reached, sampling buffer is freed */ 198 if (curr == sfb->sdbt) 199 break; 200 else 201 sdbt = curr; 202 } else { 203 /* Process SDB pointer */ 204 if (*curr) { 205 free_page((unsigned long)phys_to_virt(*curr)); 206 curr++; 207 } 208 } 209 } 210 211 debug_sprintf_event(sfdbg, 5, "%s: freed sdbt %#lx\n", __func__, 212 (unsigned long)sfb->sdbt); 213 memset(sfb, 0, sizeof(*sfb)); 214 } 215 216 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags) 217 { 218 struct hws_trailer_entry *te; 219 unsigned long sdb; 220 221 /* Allocate and initialize sample-data-block */ 222 sdb = get_zeroed_page(gfp_flags); 223 if (!sdb) 224 return -ENOMEM; 225 te = trailer_entry_ptr(sdb); 226 te->header.a = 1; 227 228 /* Link SDB into the sample-data-block-table */ 229 *sdbt = virt_to_phys((void *)sdb); 230 231 return 0; 232 } 233 234 /* 235 * realloc_sampling_buffer() - extend sampler memory 236 * 237 * Allocates new sample-data-blocks and adds them to the specified sampling 238 * buffer memory. 239 * 240 * Important: This modifies the sampling buffer and must be called when the 241 * sampling facility is disabled. 242 * 243 * Returns zero on success, non-zero otherwise. 244 */ 245 static int realloc_sampling_buffer(struct sf_buffer *sfb, 246 unsigned long num_sdb, gfp_t gfp_flags) 247 { 248 int i, rc; 249 unsigned long *new, *tail, *tail_prev = NULL; 250 251 if (!sfb->sdbt || !sfb->tail) 252 return -EINVAL; 253 254 if (!is_link_entry(sfb->tail)) 255 return -EINVAL; 256 257 /* Append to the existing sampling buffer, overwriting the table-link 258 * register. 259 * The tail variables always points to the "tail" (last and table-link) 260 * entry in an SDB-table. 261 */ 262 tail = sfb->tail; 263 264 /* Do a sanity check whether the table-link entry points to 265 * the sampling buffer origin. 266 */ 267 if (sfb->sdbt != get_next_sdbt(tail)) { 268 debug_sprintf_event(sfdbg, 3, "%s: " 269 "sampling buffer is not linked: origin %#lx" 270 " tail %#lx\n", __func__, 271 (unsigned long)sfb->sdbt, 272 (unsigned long)tail); 273 return -EINVAL; 274 } 275 276 /* Allocate remaining SDBs */ 277 rc = 0; 278 for (i = 0; i < num_sdb; i++) { 279 /* Allocate a new SDB-table if it is full. */ 280 if (require_table_link(tail)) { 281 new = (unsigned long *) get_zeroed_page(gfp_flags); 282 if (!new) { 283 rc = -ENOMEM; 284 break; 285 } 286 sfb->num_sdbt++; 287 /* Link current page to tail of chain */ 288 *tail = virt_to_phys((void *)new) + 1; 289 tail_prev = tail; 290 tail = new; 291 } 292 293 /* Allocate a new sample-data-block. 294 * If there is not enough memory, stop the realloc process 295 * and simply use what was allocated. If this is a temporary 296 * issue, a new realloc call (if required) might succeed. 297 */ 298 rc = alloc_sample_data_block(tail, gfp_flags); 299 if (rc) { 300 /* Undo last SDBT. An SDBT with no SDB at its first 301 * entry but with an SDBT entry instead can not be 302 * handled by the interrupt handler code. 303 * Avoid this situation. 304 */ 305 if (tail_prev) { 306 sfb->num_sdbt--; 307 free_page((unsigned long) new); 308 tail = tail_prev; 309 } 310 break; 311 } 312 sfb->num_sdb++; 313 tail++; 314 tail_prev = new = NULL; /* Allocated at least one SBD */ 315 } 316 317 /* Link sampling buffer to its origin */ 318 *tail = virt_to_phys(sfb->sdbt) + 1; 319 sfb->tail = tail; 320 321 debug_sprintf_event(sfdbg, 4, "%s: new buffer" 322 " settings: sdbt %lu sdb %lu\n", __func__, 323 sfb->num_sdbt, sfb->num_sdb); 324 return rc; 325 } 326 327 /* 328 * allocate_sampling_buffer() - allocate sampler memory 329 * 330 * Allocates and initializes a sampling buffer structure using the 331 * specified number of sample-data-blocks (SDB). For each allocation, 332 * a 4K page is used. The number of sample-data-block-tables (SDBT) 333 * are calculated from SDBs. 334 * Also set the ALERT_REQ mask in each SDBs trailer. 335 * 336 * Returns zero on success, non-zero otherwise. 337 */ 338 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb) 339 { 340 int rc; 341 342 if (sfb->sdbt) 343 return -EINVAL; 344 345 /* Allocate the sample-data-block-table origin */ 346 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 347 if (!sfb->sdbt) 348 return -ENOMEM; 349 sfb->num_sdb = 0; 350 sfb->num_sdbt = 1; 351 352 /* Link the table origin to point to itself to prepare for 353 * realloc_sampling_buffer() invocation. 354 */ 355 sfb->tail = sfb->sdbt; 356 *sfb->tail = virt_to_phys((void *)sfb->sdbt) + 1; 357 358 /* Allocate requested number of sample-data-blocks */ 359 rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL); 360 if (rc) { 361 free_sampling_buffer(sfb); 362 debug_sprintf_event(sfdbg, 4, "%s: " 363 "realloc_sampling_buffer failed with rc %i\n", 364 __func__, rc); 365 } else 366 debug_sprintf_event(sfdbg, 4, 367 "%s: tear %#lx dear %#lx\n", __func__, 368 (unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt); 369 return rc; 370 } 371 372 static void sfb_set_limits(unsigned long min, unsigned long max) 373 { 374 struct hws_qsi_info_block si; 375 376 CPUM_SF_MIN_SDB = min; 377 CPUM_SF_MAX_SDB = max; 378 379 memset(&si, 0, sizeof(si)); 380 if (!qsi(&si)) 381 CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes); 382 } 383 384 static unsigned long sfb_max_limit(struct hw_perf_event *hwc) 385 { 386 return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR 387 : CPUM_SF_MAX_SDB; 388 } 389 390 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb, 391 struct hw_perf_event *hwc) 392 { 393 if (!sfb->sdbt) 394 return SFB_ALLOC_REG(hwc); 395 if (SFB_ALLOC_REG(hwc) > sfb->num_sdb) 396 return SFB_ALLOC_REG(hwc) - sfb->num_sdb; 397 return 0; 398 } 399 400 static int sfb_has_pending_allocs(struct sf_buffer *sfb, 401 struct hw_perf_event *hwc) 402 { 403 return sfb_pending_allocs(sfb, hwc) > 0; 404 } 405 406 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc) 407 { 408 /* Limit the number of SDBs to not exceed the maximum */ 409 num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc)); 410 if (num) 411 SFB_ALLOC_REG(hwc) += num; 412 } 413 414 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc) 415 { 416 SFB_ALLOC_REG(hwc) = 0; 417 sfb_account_allocs(num, hwc); 418 } 419 420 static void deallocate_buffers(struct cpu_hw_sf *cpuhw) 421 { 422 if (cpuhw->sfb.sdbt) 423 free_sampling_buffer(&cpuhw->sfb); 424 } 425 426 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc) 427 { 428 unsigned long n_sdb, freq; 429 size_t sample_size; 430 431 /* Calculate sampling buffers using 4K pages 432 * 433 * 1. The sampling size is 32 bytes for basic sampling. This size 434 * is the same for all machine types. Diagnostic 435 * sampling uses auxlilary data buffer setup which provides the 436 * memory for SDBs using linux common code auxiliary trace 437 * setup. 438 * 439 * 2. Function alloc_sampling_buffer() sets the Alert Request 440 * Control indicator to trigger a measurement-alert to harvest 441 * sample-data-blocks (SDB). This is done per SDB. This 442 * measurement alert interrupt fires quick enough to handle 443 * one SDB, on very high frequency and work loads there might 444 * be 2 to 3 SBDs available for sample processing. 445 * Currently there is no need for setup alert request on every 446 * n-th page. This is counterproductive as one IRQ triggers 447 * a very high number of samples to be processed at one IRQ. 448 * 449 * 3. Use the sampling frequency as input. 450 * Compute the number of SDBs and ensure a minimum 451 * of CPUM_SF_MIN_SDB. Depending on frequency add some more 452 * SDBs to handle a higher sampling rate. 453 * Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples 454 * (one SDB) for every 10000 HZ frequency increment. 455 * 456 * 4. Compute the number of sample-data-block-tables (SDBT) and 457 * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up 458 * to 511 SDBs). 459 */ 460 sample_size = sizeof(struct hws_basic_entry); 461 freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)); 462 n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000); 463 464 /* If there is already a sampling buffer allocated, it is very likely 465 * that the sampling facility is enabled too. If the event to be 466 * initialized requires a greater sampling buffer, the allocation must 467 * be postponed. Changing the sampling buffer requires the sampling 468 * facility to be in the disabled state. So, account the number of 469 * required SDBs and let cpumsf_pmu_enable() resize the buffer just 470 * before the event is started. 471 */ 472 sfb_init_allocs(n_sdb, hwc); 473 if (sf_buffer_available(cpuhw)) 474 return 0; 475 476 debug_sprintf_event(sfdbg, 3, 477 "%s: rate %lu f %lu sdb %lu/%lu" 478 " sample_size %lu cpuhw %p\n", __func__, 479 SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc), 480 sample_size, cpuhw); 481 482 return alloc_sampling_buffer(&cpuhw->sfb, 483 sfb_pending_allocs(&cpuhw->sfb, hwc)); 484 } 485 486 static unsigned long min_percent(unsigned int percent, unsigned long base, 487 unsigned long min) 488 { 489 return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100)); 490 } 491 492 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base) 493 { 494 /* Use a percentage-based approach to extend the sampling facility 495 * buffer. Accept up to 5% sample data loss. 496 * Vary the extents between 1% to 5% of the current number of 497 * sample-data-blocks. 498 */ 499 if (ratio <= 5) 500 return 0; 501 if (ratio <= 25) 502 return min_percent(1, base, 1); 503 if (ratio <= 50) 504 return min_percent(1, base, 1); 505 if (ratio <= 75) 506 return min_percent(2, base, 2); 507 if (ratio <= 100) 508 return min_percent(3, base, 3); 509 if (ratio <= 250) 510 return min_percent(4, base, 4); 511 512 return min_percent(5, base, 8); 513 } 514 515 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw, 516 struct hw_perf_event *hwc) 517 { 518 unsigned long ratio, num; 519 520 if (!OVERFLOW_REG(hwc)) 521 return; 522 523 /* The sample_overflow contains the average number of sample data 524 * that has been lost because sample-data-blocks were full. 525 * 526 * Calculate the total number of sample data entries that has been 527 * discarded. Then calculate the ratio of lost samples to total samples 528 * per second in percent. 529 */ 530 ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb, 531 sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc))); 532 533 /* Compute number of sample-data-blocks */ 534 num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb); 535 if (num) 536 sfb_account_allocs(num, hwc); 537 538 debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n", 539 __func__, OVERFLOW_REG(hwc), ratio, num); 540 OVERFLOW_REG(hwc) = 0; 541 } 542 543 /* extend_sampling_buffer() - Extend sampling buffer 544 * @sfb: Sampling buffer structure (for local CPU) 545 * @hwc: Perf event hardware structure 546 * 547 * Use this function to extend the sampling buffer based on the overflow counter 548 * and postponed allocation extents stored in the specified Perf event hardware. 549 * 550 * Important: This function disables the sampling facility in order to safely 551 * change the sampling buffer structure. Do not call this function 552 * when the PMU is active. 553 */ 554 static void extend_sampling_buffer(struct sf_buffer *sfb, 555 struct hw_perf_event *hwc) 556 { 557 unsigned long num, num_old; 558 int rc; 559 560 num = sfb_pending_allocs(sfb, hwc); 561 if (!num) 562 return; 563 num_old = sfb->num_sdb; 564 565 /* Disable the sampling facility to reset any states and also 566 * clear pending measurement alerts. 567 */ 568 sf_disable(); 569 570 /* Extend the sampling buffer. 571 * This memory allocation typically happens in an atomic context when 572 * called by perf. Because this is a reallocation, it is fine if the 573 * new SDB-request cannot be satisfied immediately. 574 */ 575 rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC); 576 if (rc) 577 debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n", 578 __func__, rc); 579 580 if (sfb_has_pending_allocs(sfb, hwc)) 581 debug_sprintf_event(sfdbg, 5, "%s: " 582 "req %lu alloc %lu remaining %lu\n", 583 __func__, num, sfb->num_sdb - num_old, 584 sfb_pending_allocs(sfb, hwc)); 585 } 586 587 /* Number of perf events counting hardware events */ 588 static atomic_t num_events; 589 /* Used to avoid races in calling reserve/release_cpumf_hardware */ 590 static DEFINE_MUTEX(pmc_reserve_mutex); 591 592 #define PMC_INIT 0 593 #define PMC_RELEASE 1 594 #define PMC_FAILURE 2 595 static void setup_pmc_cpu(void *flags) 596 { 597 int err; 598 struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf); 599 600 err = 0; 601 switch (*((int *) flags)) { 602 case PMC_INIT: 603 memset(cpusf, 0, sizeof(*cpusf)); 604 err = qsi(&cpusf->qsi); 605 if (err) 606 break; 607 cpusf->flags |= PMU_F_RESERVED; 608 err = sf_disable(); 609 if (err) 610 pr_err("Switching off the sampling facility failed " 611 "with rc %i\n", err); 612 break; 613 case PMC_RELEASE: 614 cpusf->flags &= ~PMU_F_RESERVED; 615 err = sf_disable(); 616 if (err) { 617 pr_err("Switching off the sampling facility failed " 618 "with rc %i\n", err); 619 } else 620 deallocate_buffers(cpusf); 621 break; 622 } 623 if (err) 624 *((int *) flags) |= PMC_FAILURE; 625 } 626 627 static void release_pmc_hardware(void) 628 { 629 int flags = PMC_RELEASE; 630 631 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT); 632 on_each_cpu(setup_pmc_cpu, &flags, 1); 633 } 634 635 static int reserve_pmc_hardware(void) 636 { 637 int flags = PMC_INIT; 638 639 on_each_cpu(setup_pmc_cpu, &flags, 1); 640 if (flags & PMC_FAILURE) { 641 release_pmc_hardware(); 642 return -ENODEV; 643 } 644 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT); 645 646 return 0; 647 } 648 649 static void hw_perf_event_destroy(struct perf_event *event) 650 { 651 /* Release PMC if this is the last perf event */ 652 if (!atomic_add_unless(&num_events, -1, 1)) { 653 mutex_lock(&pmc_reserve_mutex); 654 if (atomic_dec_return(&num_events) == 0) 655 release_pmc_hardware(); 656 mutex_unlock(&pmc_reserve_mutex); 657 } 658 } 659 660 static void hw_init_period(struct hw_perf_event *hwc, u64 period) 661 { 662 hwc->sample_period = period; 663 hwc->last_period = hwc->sample_period; 664 local64_set(&hwc->period_left, hwc->sample_period); 665 } 666 667 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si, 668 unsigned long rate) 669 { 670 return clamp_t(unsigned long, rate, 671 si->min_sampl_rate, si->max_sampl_rate); 672 } 673 674 static u32 cpumsf_pid_type(struct perf_event *event, 675 u32 pid, enum pid_type type) 676 { 677 struct task_struct *tsk; 678 679 /* Idle process */ 680 if (!pid) 681 goto out; 682 683 tsk = find_task_by_pid_ns(pid, &init_pid_ns); 684 pid = -1; 685 if (tsk) { 686 /* 687 * Only top level events contain the pid namespace in which 688 * they are created. 689 */ 690 if (event->parent) 691 event = event->parent; 692 pid = __task_pid_nr_ns(tsk, type, event->ns); 693 /* 694 * See also 1d953111b648 695 * "perf/core: Don't report zero PIDs for exiting tasks". 696 */ 697 if (!pid && !pid_alive(tsk)) 698 pid = -1; 699 } 700 out: 701 return pid; 702 } 703 704 static void cpumsf_output_event_pid(struct perf_event *event, 705 struct perf_sample_data *data, 706 struct pt_regs *regs) 707 { 708 u32 pid; 709 struct perf_event_header header; 710 struct perf_output_handle handle; 711 712 /* 713 * Obtain the PID from the basic-sampling data entry and 714 * correct the data->tid_entry.pid value. 715 */ 716 pid = data->tid_entry.pid; 717 718 /* Protect callchain buffers, tasks */ 719 rcu_read_lock(); 720 721 perf_prepare_sample(&header, data, event, regs); 722 if (perf_output_begin(&handle, data, event, header.size)) 723 goto out; 724 725 /* Update the process ID (see also kernel/events/core.c) */ 726 data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID); 727 data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID); 728 729 perf_output_sample(&handle, &header, data, event); 730 perf_output_end(&handle); 731 out: 732 rcu_read_unlock(); 733 } 734 735 static unsigned long getrate(bool freq, unsigned long sample, 736 struct hws_qsi_info_block *si) 737 { 738 unsigned long rate; 739 740 if (freq) { 741 rate = freq_to_sample_rate(si, sample); 742 rate = hw_limit_rate(si, rate); 743 } else { 744 /* The min/max sampling rates specifies the valid range 745 * of sample periods. If the specified sample period is 746 * out of range, limit the period to the range boundary. 747 */ 748 rate = hw_limit_rate(si, sample); 749 750 /* The perf core maintains a maximum sample rate that is 751 * configurable through the sysctl interface. Ensure the 752 * sampling rate does not exceed this value. This also helps 753 * to avoid throttling when pushing samples with 754 * perf_event_overflow(). 755 */ 756 if (sample_rate_to_freq(si, rate) > 757 sysctl_perf_event_sample_rate) { 758 debug_sprintf_event(sfdbg, 1, "%s: " 759 "Sampling rate exceeds maximum " 760 "perf sample rate\n", __func__); 761 rate = 0; 762 } 763 } 764 return rate; 765 } 766 767 /* The sampling information (si) contains information about the 768 * min/max sampling intervals and the CPU speed. So calculate the 769 * correct sampling interval and avoid the whole period adjust 770 * feedback loop. 771 * 772 * Since the CPU Measurement sampling facility can not handle frequency 773 * calculate the sampling interval when frequency is specified using 774 * this formula: 775 * interval := cpu_speed * 1000000 / sample_freq 776 * 777 * Returns errno on bad input and zero on success with parameter interval 778 * set to the correct sampling rate. 779 * 780 * Note: This function turns off freq bit to avoid calling function 781 * perf_adjust_period(). This causes frequency adjustment in the common 782 * code part which causes tremendous variations in the counter values. 783 */ 784 static int __hw_perf_event_init_rate(struct perf_event *event, 785 struct hws_qsi_info_block *si) 786 { 787 struct perf_event_attr *attr = &event->attr; 788 struct hw_perf_event *hwc = &event->hw; 789 unsigned long rate; 790 791 if (attr->freq) { 792 if (!attr->sample_freq) 793 return -EINVAL; 794 rate = getrate(attr->freq, attr->sample_freq, si); 795 attr->freq = 0; /* Don't call perf_adjust_period() */ 796 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE; 797 } else { 798 rate = getrate(attr->freq, attr->sample_period, si); 799 if (!rate) 800 return -EINVAL; 801 } 802 attr->sample_period = rate; 803 SAMPL_RATE(hwc) = rate; 804 hw_init_period(hwc, SAMPL_RATE(hwc)); 805 debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n", 806 __func__, event->cpu, event->attr.sample_period, 807 event->attr.freq, SAMPLE_FREQ_MODE(hwc)); 808 return 0; 809 } 810 811 static int __hw_perf_event_init(struct perf_event *event) 812 { 813 struct cpu_hw_sf *cpuhw; 814 struct hws_qsi_info_block si; 815 struct perf_event_attr *attr = &event->attr; 816 struct hw_perf_event *hwc = &event->hw; 817 int cpu, err; 818 819 /* Reserve CPU-measurement sampling facility */ 820 err = 0; 821 if (!atomic_inc_not_zero(&num_events)) { 822 mutex_lock(&pmc_reserve_mutex); 823 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware()) 824 err = -EBUSY; 825 else 826 atomic_inc(&num_events); 827 mutex_unlock(&pmc_reserve_mutex); 828 } 829 event->destroy = hw_perf_event_destroy; 830 831 if (err) 832 goto out; 833 834 /* Access per-CPU sampling information (query sampling info) */ 835 /* 836 * The event->cpu value can be -1 to count on every CPU, for example, 837 * when attaching to a task. If this is specified, use the query 838 * sampling info from the current CPU, otherwise use event->cpu to 839 * retrieve the per-CPU information. 840 * Later, cpuhw indicates whether to allocate sampling buffers for a 841 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL). 842 */ 843 memset(&si, 0, sizeof(si)); 844 cpuhw = NULL; 845 if (event->cpu == -1) 846 qsi(&si); 847 else { 848 /* Event is pinned to a particular CPU, retrieve the per-CPU 849 * sampling structure for accessing the CPU-specific QSI. 850 */ 851 cpuhw = &per_cpu(cpu_hw_sf, event->cpu); 852 si = cpuhw->qsi; 853 } 854 855 /* Check sampling facility authorization and, if not authorized, 856 * fall back to other PMUs. It is safe to check any CPU because 857 * the authorization is identical for all configured CPUs. 858 */ 859 if (!si.as) { 860 err = -ENOENT; 861 goto out; 862 } 863 864 if (si.ribm & CPU_MF_SF_RIBM_NOTAV) { 865 pr_warn("CPU Measurement Facility sampling is temporarily not available\n"); 866 err = -EBUSY; 867 goto out; 868 } 869 870 /* Always enable basic sampling */ 871 SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE; 872 873 /* Check if diagnostic sampling is requested. Deny if the required 874 * sampling authorization is missing. 875 */ 876 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) { 877 if (!si.ad) { 878 err = -EPERM; 879 goto out; 880 } 881 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE; 882 } 883 884 /* Check and set other sampling flags */ 885 if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS) 886 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS; 887 888 err = __hw_perf_event_init_rate(event, &si); 889 if (err) 890 goto out; 891 892 /* Initialize sample data overflow accounting */ 893 hwc->extra_reg.reg = REG_OVERFLOW; 894 OVERFLOW_REG(hwc) = 0; 895 896 /* Use AUX buffer. No need to allocate it by ourself */ 897 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) 898 return 0; 899 900 /* Allocate the per-CPU sampling buffer using the CPU information 901 * from the event. If the event is not pinned to a particular 902 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling 903 * buffers for each online CPU. 904 */ 905 if (cpuhw) 906 /* Event is pinned to a particular CPU */ 907 err = allocate_buffers(cpuhw, hwc); 908 else { 909 /* Event is not pinned, allocate sampling buffer on 910 * each online CPU 911 */ 912 for_each_online_cpu(cpu) { 913 cpuhw = &per_cpu(cpu_hw_sf, cpu); 914 err = allocate_buffers(cpuhw, hwc); 915 if (err) 916 break; 917 } 918 } 919 920 /* If PID/TID sampling is active, replace the default overflow 921 * handler to extract and resolve the PIDs from the basic-sampling 922 * data entries. 923 */ 924 if (event->attr.sample_type & PERF_SAMPLE_TID) 925 if (is_default_overflow_handler(event)) 926 event->overflow_handler = cpumsf_output_event_pid; 927 out: 928 return err; 929 } 930 931 static bool is_callchain_event(struct perf_event *event) 932 { 933 u64 sample_type = event->attr.sample_type; 934 935 return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | 936 PERF_SAMPLE_STACK_USER); 937 } 938 939 static int cpumsf_pmu_event_init(struct perf_event *event) 940 { 941 int err; 942 943 /* No support for taken branch sampling */ 944 /* No support for callchain, stacks and registers */ 945 if (has_branch_stack(event) || is_callchain_event(event)) 946 return -EOPNOTSUPP; 947 948 switch (event->attr.type) { 949 case PERF_TYPE_RAW: 950 if ((event->attr.config != PERF_EVENT_CPUM_SF) && 951 (event->attr.config != PERF_EVENT_CPUM_SF_DIAG)) 952 return -ENOENT; 953 break; 954 case PERF_TYPE_HARDWARE: 955 /* Support sampling of CPU cycles in addition to the 956 * counter facility. However, the counter facility 957 * is more precise and, hence, restrict this PMU to 958 * sampling events only. 959 */ 960 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES) 961 return -ENOENT; 962 if (!is_sampling_event(event)) 963 return -ENOENT; 964 break; 965 default: 966 return -ENOENT; 967 } 968 969 /* Check online status of the CPU to which the event is pinned */ 970 if (event->cpu >= 0 && !cpu_online(event->cpu)) 971 return -ENODEV; 972 973 /* Force reset of idle/hv excludes regardless of what the 974 * user requested. 975 */ 976 if (event->attr.exclude_hv) 977 event->attr.exclude_hv = 0; 978 if (event->attr.exclude_idle) 979 event->attr.exclude_idle = 0; 980 981 err = __hw_perf_event_init(event); 982 if (unlikely(err)) 983 if (event->destroy) 984 event->destroy(event); 985 return err; 986 } 987 988 static void cpumsf_pmu_enable(struct pmu *pmu) 989 { 990 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 991 struct hw_perf_event *hwc; 992 int err; 993 994 if (cpuhw->flags & PMU_F_ENABLED) 995 return; 996 997 if (cpuhw->flags & PMU_F_ERR_MASK) 998 return; 999 1000 /* Check whether to extent the sampling buffer. 1001 * 1002 * Two conditions trigger an increase of the sampling buffer for a 1003 * perf event: 1004 * 1. Postponed buffer allocations from the event initialization. 1005 * 2. Sampling overflows that contribute to pending allocations. 1006 * 1007 * Note that the extend_sampling_buffer() function disables the sampling 1008 * facility, but it can be fully re-enabled using sampling controls that 1009 * have been saved in cpumsf_pmu_disable(). 1010 */ 1011 if (cpuhw->event) { 1012 hwc = &cpuhw->event->hw; 1013 if (!(SAMPL_DIAG_MODE(hwc))) { 1014 /* 1015 * Account number of overflow-designated 1016 * buffer extents 1017 */ 1018 sfb_account_overflows(cpuhw, hwc); 1019 extend_sampling_buffer(&cpuhw->sfb, hwc); 1020 } 1021 /* Rate may be adjusted with ioctl() */ 1022 cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw); 1023 } 1024 1025 /* (Re)enable the PMU and sampling facility */ 1026 cpuhw->flags |= PMU_F_ENABLED; 1027 barrier(); 1028 1029 err = lsctl(&cpuhw->lsctl); 1030 if (err) { 1031 cpuhw->flags &= ~PMU_F_ENABLED; 1032 pr_err("Loading sampling controls failed: op %i err %i\n", 1033 1, err); 1034 return; 1035 } 1036 1037 /* Load current program parameter */ 1038 lpp(&S390_lowcore.lpp); 1039 1040 debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i " 1041 "interval %#lx tear %#lx dear %#lx\n", __func__, 1042 cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed, 1043 cpuhw->lsctl.cd, cpuhw->lsctl.interval, 1044 cpuhw->lsctl.tear, cpuhw->lsctl.dear); 1045 } 1046 1047 static void cpumsf_pmu_disable(struct pmu *pmu) 1048 { 1049 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1050 struct hws_lsctl_request_block inactive; 1051 struct hws_qsi_info_block si; 1052 int err; 1053 1054 if (!(cpuhw->flags & PMU_F_ENABLED)) 1055 return; 1056 1057 if (cpuhw->flags & PMU_F_ERR_MASK) 1058 return; 1059 1060 /* Switch off sampling activation control */ 1061 inactive = cpuhw->lsctl; 1062 inactive.cs = 0; 1063 inactive.cd = 0; 1064 1065 err = lsctl(&inactive); 1066 if (err) { 1067 pr_err("Loading sampling controls failed: op %i err %i\n", 1068 2, err); 1069 return; 1070 } 1071 1072 /* Save state of TEAR and DEAR register contents */ 1073 err = qsi(&si); 1074 if (!err) { 1075 /* TEAR/DEAR values are valid only if the sampling facility is 1076 * enabled. Note that cpumsf_pmu_disable() might be called even 1077 * for a disabled sampling facility because cpumsf_pmu_enable() 1078 * controls the enable/disable state. 1079 */ 1080 if (si.es) { 1081 cpuhw->lsctl.tear = si.tear; 1082 cpuhw->lsctl.dear = si.dear; 1083 } 1084 } else 1085 debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n", 1086 __func__, err); 1087 1088 cpuhw->flags &= ~PMU_F_ENABLED; 1089 } 1090 1091 /* perf_exclude_event() - Filter event 1092 * @event: The perf event 1093 * @regs: pt_regs structure 1094 * @sde_regs: Sample-data-entry (sde) regs structure 1095 * 1096 * Filter perf events according to their exclude specification. 1097 * 1098 * Return non-zero if the event shall be excluded. 1099 */ 1100 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs, 1101 struct perf_sf_sde_regs *sde_regs) 1102 { 1103 if (event->attr.exclude_user && user_mode(regs)) 1104 return 1; 1105 if (event->attr.exclude_kernel && !user_mode(regs)) 1106 return 1; 1107 if (event->attr.exclude_guest && sde_regs->in_guest) 1108 return 1; 1109 if (event->attr.exclude_host && !sde_regs->in_guest) 1110 return 1; 1111 return 0; 1112 } 1113 1114 /* perf_push_sample() - Push samples to perf 1115 * @event: The perf event 1116 * @sample: Hardware sample data 1117 * 1118 * Use the hardware sample data to create perf event sample. The sample 1119 * is the pushed to the event subsystem and the function checks for 1120 * possible event overflows. If an event overflow occurs, the PMU is 1121 * stopped. 1122 * 1123 * Return non-zero if an event overflow occurred. 1124 */ 1125 static int perf_push_sample(struct perf_event *event, 1126 struct hws_basic_entry *basic) 1127 { 1128 int overflow; 1129 struct pt_regs regs; 1130 struct perf_sf_sde_regs *sde_regs; 1131 struct perf_sample_data data; 1132 1133 /* Setup perf sample */ 1134 perf_sample_data_init(&data, 0, event->hw.last_period); 1135 1136 /* Setup pt_regs to look like an CPU-measurement external interrupt 1137 * using the Program Request Alert code. The regs.int_parm_long 1138 * field which is unused contains additional sample-data-entry related 1139 * indicators. 1140 */ 1141 memset(®s, 0, sizeof(regs)); 1142 regs.int_code = 0x1407; 1143 regs.int_parm = CPU_MF_INT_SF_PRA; 1144 sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long; 1145 1146 psw_bits(regs.psw).ia = basic->ia; 1147 psw_bits(regs.psw).dat = basic->T; 1148 psw_bits(regs.psw).wait = basic->W; 1149 psw_bits(regs.psw).pstate = basic->P; 1150 psw_bits(regs.psw).as = basic->AS; 1151 1152 /* 1153 * Use the hardware provided configuration level to decide if the 1154 * sample belongs to a guest or host. If that is not available, 1155 * fall back to the following heuristics: 1156 * A non-zero guest program parameter always indicates a guest 1157 * sample. Some early samples or samples from guests without 1158 * lpp usage would be misaccounted to the host. We use the asn 1159 * value as an addon heuristic to detect most of these guest samples. 1160 * If the value differs from 0xffff (the host value), we assume to 1161 * be a KVM guest. 1162 */ 1163 switch (basic->CL) { 1164 case 1: /* logical partition */ 1165 sde_regs->in_guest = 0; 1166 break; 1167 case 2: /* virtual machine */ 1168 sde_regs->in_guest = 1; 1169 break; 1170 default: /* old machine, use heuristics */ 1171 if (basic->gpp || basic->prim_asn != 0xffff) 1172 sde_regs->in_guest = 1; 1173 break; 1174 } 1175 1176 /* 1177 * Store the PID value from the sample-data-entry to be 1178 * processed and resolved by cpumsf_output_event_pid(). 1179 */ 1180 data.tid_entry.pid = basic->hpp & LPP_PID_MASK; 1181 1182 overflow = 0; 1183 if (perf_exclude_event(event, ®s, sde_regs)) 1184 goto out; 1185 if (perf_event_overflow(event, &data, ®s)) { 1186 overflow = 1; 1187 event->pmu->stop(event, 0); 1188 } 1189 perf_event_update_userpage(event); 1190 out: 1191 return overflow; 1192 } 1193 1194 static void perf_event_count_update(struct perf_event *event, u64 count) 1195 { 1196 local64_add(count, &event->count); 1197 } 1198 1199 /* hw_collect_samples() - Walk through a sample-data-block and collect samples 1200 * @event: The perf event 1201 * @sdbt: Sample-data-block table 1202 * @overflow: Event overflow counter 1203 * 1204 * Walks through a sample-data-block and collects sampling data entries that are 1205 * then pushed to the perf event subsystem. Depending on the sampling function, 1206 * there can be either basic-sampling or combined-sampling data entries. A 1207 * combined-sampling data entry consists of a basic- and a diagnostic-sampling 1208 * data entry. The sampling function is determined by the flags in the perf 1209 * event hardware structure. The function always works with a combined-sampling 1210 * data entry but ignores the the diagnostic portion if it is not available. 1211 * 1212 * Note that the implementation focuses on basic-sampling data entries and, if 1213 * such an entry is not valid, the entire combined-sampling data entry is 1214 * ignored. 1215 * 1216 * The overflow variables counts the number of samples that has been discarded 1217 * due to a perf event overflow. 1218 */ 1219 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt, 1220 unsigned long long *overflow) 1221 { 1222 struct hws_trailer_entry *te; 1223 struct hws_basic_entry *sample; 1224 1225 te = trailer_entry_ptr((unsigned long)sdbt); 1226 sample = (struct hws_basic_entry *)sdbt; 1227 while ((unsigned long *) sample < (unsigned long *) te) { 1228 /* Check for an empty sample */ 1229 if (!sample->def || sample->LS) 1230 break; 1231 1232 /* Update perf event period */ 1233 perf_event_count_update(event, SAMPL_RATE(&event->hw)); 1234 1235 /* Check whether sample is valid */ 1236 if (sample->def == 0x0001) { 1237 /* If an event overflow occurred, the PMU is stopped to 1238 * throttle event delivery. Remaining sample data is 1239 * discarded. 1240 */ 1241 if (!*overflow) { 1242 /* Check whether sample is consistent */ 1243 if (sample->I == 0 && sample->W == 0) { 1244 /* Deliver sample data to perf */ 1245 *overflow = perf_push_sample(event, 1246 sample); 1247 } 1248 } else 1249 /* Count discarded samples */ 1250 *overflow += 1; 1251 } else { 1252 debug_sprintf_event(sfdbg, 4, 1253 "%s: Found unknown" 1254 " sampling data entry: te->f %i" 1255 " basic.def %#4x (%p)\n", __func__, 1256 te->header.f, sample->def, sample); 1257 /* Sample slot is not yet written or other record. 1258 * 1259 * This condition can occur if the buffer was reused 1260 * from a combined basic- and diagnostic-sampling. 1261 * If only basic-sampling is then active, entries are 1262 * written into the larger diagnostic entries. 1263 * This is typically the case for sample-data-blocks 1264 * that are not full. Stop processing if the first 1265 * invalid format was detected. 1266 */ 1267 if (!te->header.f) 1268 break; 1269 } 1270 1271 /* Reset sample slot and advance to next sample */ 1272 sample->def = 0; 1273 sample++; 1274 } 1275 } 1276 1277 static inline __uint128_t __cdsg(__uint128_t *ptr, __uint128_t old, __uint128_t new) 1278 { 1279 asm volatile( 1280 " cdsg %[old],%[new],%[ptr]\n" 1281 : [old] "+d" (old), [ptr] "+QS" (*ptr) 1282 : [new] "d" (new) 1283 : "memory", "cc"); 1284 return old; 1285 } 1286 1287 /* hw_perf_event_update() - Process sampling buffer 1288 * @event: The perf event 1289 * @flush_all: Flag to also flush partially filled sample-data-blocks 1290 * 1291 * Processes the sampling buffer and create perf event samples. 1292 * The sampling buffer position are retrieved and saved in the TEAR_REG 1293 * register of the specified perf event. 1294 * 1295 * Only full sample-data-blocks are processed. Specify the flash_all flag 1296 * to also walk through partially filled sample-data-blocks. It is ignored 1297 * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag 1298 * enforces the processing of full sample-data-blocks only (trailer entries 1299 * with the block-full-indicator bit set). 1300 */ 1301 static void hw_perf_event_update(struct perf_event *event, int flush_all) 1302 { 1303 unsigned long long event_overflow, sampl_overflow, num_sdb; 1304 union hws_trailer_header old, prev, new; 1305 struct hw_perf_event *hwc = &event->hw; 1306 struct hws_trailer_entry *te; 1307 unsigned long *sdbt, sdb; 1308 int done; 1309 1310 /* 1311 * AUX buffer is used when in diagnostic sampling mode. 1312 * No perf events/samples are created. 1313 */ 1314 if (SAMPL_DIAG_MODE(&event->hw)) 1315 return; 1316 1317 if (flush_all && SDB_FULL_BLOCKS(hwc)) 1318 flush_all = 0; 1319 1320 sdbt = (unsigned long *) TEAR_REG(hwc); 1321 done = event_overflow = sampl_overflow = num_sdb = 0; 1322 while (!done) { 1323 /* Get the trailer entry of the sample-data-block */ 1324 sdb = (unsigned long)phys_to_virt(*sdbt); 1325 te = trailer_entry_ptr(sdb); 1326 1327 /* Leave loop if no more work to do (block full indicator) */ 1328 if (!te->header.f) { 1329 done = 1; 1330 if (!flush_all) 1331 break; 1332 } 1333 1334 /* Check the sample overflow count */ 1335 if (te->header.overflow) 1336 /* Account sample overflows and, if a particular limit 1337 * is reached, extend the sampling buffer. 1338 * For details, see sfb_account_overflows(). 1339 */ 1340 sampl_overflow += te->header.overflow; 1341 1342 /* Timestamps are valid for full sample-data-blocks only */ 1343 debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx/%#lx " 1344 "overflow %llu timestamp %#llx\n", 1345 __func__, sdb, (unsigned long)sdbt, 1346 te->header.overflow, 1347 (te->header.f) ? trailer_timestamp(te) : 0ULL); 1348 1349 /* Collect all samples from a single sample-data-block and 1350 * flag if an (perf) event overflow happened. If so, the PMU 1351 * is stopped and remaining samples will be discarded. 1352 */ 1353 hw_collect_samples(event, (unsigned long *)sdb, &event_overflow); 1354 num_sdb++; 1355 1356 /* Reset trailer (using compare-double-and-swap) */ 1357 /* READ_ONCE() 16 byte header */ 1358 prev.val = __cdsg(&te->header.val, 0, 0); 1359 do { 1360 old.val = prev.val; 1361 new.val = prev.val; 1362 new.f = 0; 1363 new.a = 1; 1364 new.overflow = 0; 1365 prev.val = __cdsg(&te->header.val, old.val, new.val); 1366 } while (prev.val != old.val); 1367 1368 /* Advance to next sample-data-block */ 1369 sdbt++; 1370 if (is_link_entry(sdbt)) 1371 sdbt = get_next_sdbt(sdbt); 1372 1373 /* Update event hardware registers */ 1374 TEAR_REG(hwc) = (unsigned long) sdbt; 1375 1376 /* Stop processing sample-data if all samples of the current 1377 * sample-data-block were flushed even if it was not full. 1378 */ 1379 if (flush_all && done) 1380 break; 1381 } 1382 1383 /* Account sample overflows in the event hardware structure */ 1384 if (sampl_overflow) 1385 OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) + 1386 sampl_overflow, 1 + num_sdb); 1387 1388 /* Perf_event_overflow() and perf_event_account_interrupt() limit 1389 * the interrupt rate to an upper limit. Roughly 1000 samples per 1390 * task tick. 1391 * Hitting this limit results in a large number 1392 * of throttled REF_REPORT_THROTTLE entries and the samples 1393 * are dropped. 1394 * Slightly increase the interval to avoid hitting this limit. 1395 */ 1396 if (event_overflow) { 1397 SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10); 1398 debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n", 1399 __func__, 1400 DIV_ROUND_UP(SAMPL_RATE(hwc), 10)); 1401 } 1402 1403 if (sampl_overflow || event_overflow) 1404 debug_sprintf_event(sfdbg, 4, "%s: " 1405 "overflows: sample %llu event %llu" 1406 " total %llu num_sdb %llu\n", 1407 __func__, sampl_overflow, event_overflow, 1408 OVERFLOW_REG(hwc), num_sdb); 1409 } 1410 1411 #define AUX_SDB_INDEX(aux, i) ((i) % aux->sfb.num_sdb) 1412 #define AUX_SDB_NUM(aux, start, end) (end >= start ? end - start + 1 : 0) 1413 #define AUX_SDB_NUM_ALERT(aux) AUX_SDB_NUM(aux, aux->head, aux->alert_mark) 1414 #define AUX_SDB_NUM_EMPTY(aux) AUX_SDB_NUM(aux, aux->head, aux->empty_mark) 1415 1416 /* 1417 * Get trailer entry by index of SDB. 1418 */ 1419 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux, 1420 unsigned long index) 1421 { 1422 unsigned long sdb; 1423 1424 index = AUX_SDB_INDEX(aux, index); 1425 sdb = aux->sdb_index[index]; 1426 return trailer_entry_ptr(sdb); 1427 } 1428 1429 /* 1430 * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu 1431 * disabled. Collect the full SDBs in AUX buffer which have not reached 1432 * the point of alert indicator. And ignore the SDBs which are not 1433 * full. 1434 * 1435 * 1. Scan SDBs to see how much data is there and consume them. 1436 * 2. Remove alert indicator in the buffer. 1437 */ 1438 static void aux_output_end(struct perf_output_handle *handle) 1439 { 1440 unsigned long i, range_scan, idx; 1441 struct aux_buffer *aux; 1442 struct hws_trailer_entry *te; 1443 1444 aux = perf_get_aux(handle); 1445 if (!aux) 1446 return; 1447 1448 range_scan = AUX_SDB_NUM_ALERT(aux); 1449 for (i = 0, idx = aux->head; i < range_scan; i++, idx++) { 1450 te = aux_sdb_trailer(aux, idx); 1451 if (!te->header.f) 1452 break; 1453 } 1454 /* i is num of SDBs which are full */ 1455 perf_aux_output_end(handle, i << PAGE_SHIFT); 1456 1457 /* Remove alert indicators in the buffer */ 1458 te = aux_sdb_trailer(aux, aux->alert_mark); 1459 te->header.a = 0; 1460 1461 debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n", 1462 __func__, i, range_scan, aux->head); 1463 } 1464 1465 /* 1466 * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event 1467 * is first added to the CPU or rescheduled again to the CPU. It is called 1468 * with pmu disabled. 1469 * 1470 * 1. Reset the trailer of SDBs to get ready for new data. 1471 * 2. Tell the hardware where to put the data by reset the SDBs buffer 1472 * head(tear/dear). 1473 */ 1474 static int aux_output_begin(struct perf_output_handle *handle, 1475 struct aux_buffer *aux, 1476 struct cpu_hw_sf *cpuhw) 1477 { 1478 unsigned long range; 1479 unsigned long i, range_scan, idx; 1480 unsigned long head, base, offset; 1481 struct hws_trailer_entry *te; 1482 1483 if (WARN_ON_ONCE(handle->head & ~PAGE_MASK)) 1484 return -EINVAL; 1485 1486 aux->head = handle->head >> PAGE_SHIFT; 1487 range = (handle->size + 1) >> PAGE_SHIFT; 1488 if (range <= 1) 1489 return -ENOMEM; 1490 1491 /* 1492 * SDBs between aux->head and aux->empty_mark are already ready 1493 * for new data. range_scan is num of SDBs not within them. 1494 */ 1495 debug_sprintf_event(sfdbg, 6, 1496 "%s: range %ld head %ld alert %ld empty %ld\n", 1497 __func__, range, aux->head, aux->alert_mark, 1498 aux->empty_mark); 1499 if (range > AUX_SDB_NUM_EMPTY(aux)) { 1500 range_scan = range - AUX_SDB_NUM_EMPTY(aux); 1501 idx = aux->empty_mark + 1; 1502 for (i = 0; i < range_scan; i++, idx++) { 1503 te = aux_sdb_trailer(aux, idx); 1504 te->header.f = 0; 1505 te->header.a = 0; 1506 te->header.overflow = 0; 1507 } 1508 /* Save the position of empty SDBs */ 1509 aux->empty_mark = aux->head + range - 1; 1510 } 1511 1512 /* Set alert indicator */ 1513 aux->alert_mark = aux->head + range/2 - 1; 1514 te = aux_sdb_trailer(aux, aux->alert_mark); 1515 te->header.a = 1; 1516 1517 /* Reset hardware buffer head */ 1518 head = AUX_SDB_INDEX(aux, aux->head); 1519 base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE]; 1520 offset = head % CPUM_SF_SDB_PER_TABLE; 1521 cpuhw->lsctl.tear = base + offset * sizeof(unsigned long); 1522 cpuhw->lsctl.dear = aux->sdb_index[head]; 1523 1524 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld " 1525 "index %ld tear %#lx dear %#lx\n", __func__, 1526 aux->head, aux->alert_mark, aux->empty_mark, 1527 head / CPUM_SF_SDB_PER_TABLE, 1528 cpuhw->lsctl.tear, cpuhw->lsctl.dear); 1529 1530 return 0; 1531 } 1532 1533 /* 1534 * Set alert indicator on SDB at index @alert_index while sampler is running. 1535 * 1536 * Return true if successfully. 1537 * Return false if full indicator is already set by hardware sampler. 1538 */ 1539 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index, 1540 unsigned long long *overflow) 1541 { 1542 union hws_trailer_header old, prev, new; 1543 struct hws_trailer_entry *te; 1544 1545 te = aux_sdb_trailer(aux, alert_index); 1546 /* READ_ONCE() 16 byte header */ 1547 prev.val = __cdsg(&te->header.val, 0, 0); 1548 do { 1549 old.val = prev.val; 1550 new.val = prev.val; 1551 *overflow = old.overflow; 1552 if (old.f) { 1553 /* 1554 * SDB is already set by hardware. 1555 * Abort and try to set somewhere 1556 * behind. 1557 */ 1558 return false; 1559 } 1560 new.a = 1; 1561 new.overflow = 0; 1562 prev.val = __cdsg(&te->header.val, old.val, new.val); 1563 } while (prev.val != old.val); 1564 return true; 1565 } 1566 1567 /* 1568 * aux_reset_buffer() - Scan and setup SDBs for new samples 1569 * @aux: The AUX buffer to set 1570 * @range: The range of SDBs to scan started from aux->head 1571 * @overflow: Set to overflow count 1572 * 1573 * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is 1574 * marked as empty, check if it is already set full by the hardware sampler. 1575 * If yes, that means new data is already there before we can set an alert 1576 * indicator. Caller should try to set alert indicator to some position behind. 1577 * 1578 * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used 1579 * previously and have already been consumed by user space. Reset these SDBs 1580 * (clear full indicator and alert indicator) for new data. 1581 * If aux->alert_mark fall in this area, just set it. Overflow count is 1582 * recorded while scanning. 1583 * 1584 * SDBs between aux->head and aux->empty_mark are already reset at last time. 1585 * and ready for new samples. So scanning on this area could be skipped. 1586 * 1587 * Return true if alert indicator is set successfully and false if not. 1588 */ 1589 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range, 1590 unsigned long long *overflow) 1591 { 1592 unsigned long i, range_scan, idx, idx_old; 1593 union hws_trailer_header old, prev, new; 1594 unsigned long long orig_overflow; 1595 struct hws_trailer_entry *te; 1596 1597 debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld " 1598 "empty %ld\n", __func__, range, aux->head, 1599 aux->alert_mark, aux->empty_mark); 1600 if (range <= AUX_SDB_NUM_EMPTY(aux)) 1601 /* 1602 * No need to scan. All SDBs in range are marked as empty. 1603 * Just set alert indicator. Should check race with hardware 1604 * sampler. 1605 */ 1606 return aux_set_alert(aux, aux->alert_mark, overflow); 1607 1608 if (aux->alert_mark <= aux->empty_mark) 1609 /* 1610 * Set alert indicator on empty SDB. Should check race 1611 * with hardware sampler. 1612 */ 1613 if (!aux_set_alert(aux, aux->alert_mark, overflow)) 1614 return false; 1615 1616 /* 1617 * Scan the SDBs to clear full and alert indicator used previously. 1618 * Start scanning from one SDB behind empty_mark. If the new alert 1619 * indicator fall into this range, set it. 1620 */ 1621 range_scan = range - AUX_SDB_NUM_EMPTY(aux); 1622 idx_old = idx = aux->empty_mark + 1; 1623 for (i = 0; i < range_scan; i++, idx++) { 1624 te = aux_sdb_trailer(aux, idx); 1625 /* READ_ONCE() 16 byte header */ 1626 prev.val = __cdsg(&te->header.val, 0, 0); 1627 do { 1628 old.val = prev.val; 1629 new.val = prev.val; 1630 orig_overflow = old.overflow; 1631 new.f = 0; 1632 new.overflow = 0; 1633 if (idx == aux->alert_mark) 1634 new.a = 1; 1635 else 1636 new.a = 0; 1637 prev.val = __cdsg(&te->header.val, old.val, new.val); 1638 } while (prev.val != old.val); 1639 *overflow += orig_overflow; 1640 } 1641 1642 /* Update empty_mark to new position */ 1643 aux->empty_mark = aux->head + range - 1; 1644 1645 debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld " 1646 "empty %ld\n", __func__, range_scan, idx_old, 1647 idx - 1, aux->empty_mark); 1648 return true; 1649 } 1650 1651 /* 1652 * Measurement alert handler for diagnostic mode sampling. 1653 */ 1654 static void hw_collect_aux(struct cpu_hw_sf *cpuhw) 1655 { 1656 struct aux_buffer *aux; 1657 int done = 0; 1658 unsigned long range = 0, size; 1659 unsigned long long overflow = 0; 1660 struct perf_output_handle *handle = &cpuhw->handle; 1661 unsigned long num_sdb; 1662 1663 aux = perf_get_aux(handle); 1664 if (WARN_ON_ONCE(!aux)) 1665 return; 1666 1667 /* Inform user space new data arrived */ 1668 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; 1669 debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__, 1670 size >> PAGE_SHIFT); 1671 perf_aux_output_end(handle, size); 1672 1673 num_sdb = aux->sfb.num_sdb; 1674 while (!done) { 1675 /* Get an output handle */ 1676 aux = perf_aux_output_begin(handle, cpuhw->event); 1677 if (handle->size == 0) { 1678 pr_err("The AUX buffer with %lu pages for the " 1679 "diagnostic-sampling mode is full\n", 1680 num_sdb); 1681 debug_sprintf_event(sfdbg, 1, 1682 "%s: AUX buffer used up\n", 1683 __func__); 1684 break; 1685 } 1686 if (WARN_ON_ONCE(!aux)) 1687 return; 1688 1689 /* Update head and alert_mark to new position */ 1690 aux->head = handle->head >> PAGE_SHIFT; 1691 range = (handle->size + 1) >> PAGE_SHIFT; 1692 if (range == 1) 1693 aux->alert_mark = aux->head; 1694 else 1695 aux->alert_mark = aux->head + range/2 - 1; 1696 1697 if (aux_reset_buffer(aux, range, &overflow)) { 1698 if (!overflow) { 1699 done = 1; 1700 break; 1701 } 1702 size = range << PAGE_SHIFT; 1703 perf_aux_output_end(&cpuhw->handle, size); 1704 pr_err("Sample data caused the AUX buffer with %lu " 1705 "pages to overflow\n", aux->sfb.num_sdb); 1706 debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld " 1707 "overflow %lld\n", __func__, 1708 aux->head, range, overflow); 1709 } else { 1710 size = AUX_SDB_NUM_ALERT(aux) << PAGE_SHIFT; 1711 perf_aux_output_end(&cpuhw->handle, size); 1712 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " 1713 "already full, try another\n", 1714 __func__, 1715 aux->head, aux->alert_mark); 1716 } 1717 } 1718 1719 if (done) 1720 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " 1721 "empty %ld\n", __func__, aux->head, 1722 aux->alert_mark, aux->empty_mark); 1723 } 1724 1725 /* 1726 * Callback when freeing AUX buffers. 1727 */ 1728 static void aux_buffer_free(void *data) 1729 { 1730 struct aux_buffer *aux = data; 1731 unsigned long i, num_sdbt; 1732 1733 if (!aux) 1734 return; 1735 1736 /* Free SDBT. SDB is freed by the caller */ 1737 num_sdbt = aux->sfb.num_sdbt; 1738 for (i = 0; i < num_sdbt; i++) 1739 free_page(aux->sdbt_index[i]); 1740 1741 kfree(aux->sdbt_index); 1742 kfree(aux->sdb_index); 1743 kfree(aux); 1744 1745 debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt); 1746 } 1747 1748 static void aux_sdb_init(unsigned long sdb) 1749 { 1750 struct hws_trailer_entry *te; 1751 1752 te = trailer_entry_ptr(sdb); 1753 1754 /* Save clock base */ 1755 te->clock_base = 1; 1756 te->progusage2 = tod_clock_base.tod; 1757 } 1758 1759 /* 1760 * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling 1761 * @event: Event the buffer is setup for, event->cpu == -1 means current 1762 * @pages: Array of pointers to buffer pages passed from perf core 1763 * @nr_pages: Total pages 1764 * @snapshot: Flag for snapshot mode 1765 * 1766 * This is the callback when setup an event using AUX buffer. Perf tool can 1767 * trigger this by an additional mmap() call on the event. Unlike the buffer 1768 * for basic samples, AUX buffer belongs to the event. It is scheduled with 1769 * the task among online cpus when it is a per-thread event. 1770 * 1771 * Return the private AUX buffer structure if success or NULL if fails. 1772 */ 1773 static void *aux_buffer_setup(struct perf_event *event, void **pages, 1774 int nr_pages, bool snapshot) 1775 { 1776 struct sf_buffer *sfb; 1777 struct aux_buffer *aux; 1778 unsigned long *new, *tail; 1779 int i, n_sdbt; 1780 1781 if (!nr_pages || !pages) 1782 return NULL; 1783 1784 if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1785 pr_err("AUX buffer size (%i pages) is larger than the " 1786 "maximum sampling buffer limit\n", 1787 nr_pages); 1788 return NULL; 1789 } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1790 pr_err("AUX buffer size (%i pages) is less than the " 1791 "minimum sampling buffer limit\n", 1792 nr_pages); 1793 return NULL; 1794 } 1795 1796 /* Allocate aux_buffer struct for the event */ 1797 aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL); 1798 if (!aux) 1799 goto no_aux; 1800 sfb = &aux->sfb; 1801 1802 /* Allocate sdbt_index for fast reference */ 1803 n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE); 1804 aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL); 1805 if (!aux->sdbt_index) 1806 goto no_sdbt_index; 1807 1808 /* Allocate sdb_index for fast reference */ 1809 aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL); 1810 if (!aux->sdb_index) 1811 goto no_sdb_index; 1812 1813 /* Allocate the first SDBT */ 1814 sfb->num_sdbt = 0; 1815 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1816 if (!sfb->sdbt) 1817 goto no_sdbt; 1818 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt; 1819 tail = sfb->tail = sfb->sdbt; 1820 1821 /* 1822 * Link the provided pages of AUX buffer to SDBT. 1823 * Allocate SDBT if needed. 1824 */ 1825 for (i = 0; i < nr_pages; i++, tail++) { 1826 if (require_table_link(tail)) { 1827 new = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1828 if (!new) 1829 goto no_sdbt; 1830 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new; 1831 /* Link current page to tail of chain */ 1832 *tail = (unsigned long)(void *) new + 1; 1833 tail = new; 1834 } 1835 /* Tail is the entry in a SDBT */ 1836 *tail = (unsigned long)pages[i]; 1837 aux->sdb_index[i] = (unsigned long)pages[i]; 1838 aux_sdb_init((unsigned long)pages[i]); 1839 } 1840 sfb->num_sdb = nr_pages; 1841 1842 /* Link the last entry in the SDBT to the first SDBT */ 1843 *tail = (unsigned long) sfb->sdbt + 1; 1844 sfb->tail = tail; 1845 1846 /* 1847 * Initial all SDBs are zeroed. Mark it as empty. 1848 * So there is no need to clear the full indicator 1849 * when this event is first added. 1850 */ 1851 aux->empty_mark = sfb->num_sdb - 1; 1852 1853 debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__, 1854 sfb->num_sdbt, sfb->num_sdb); 1855 1856 return aux; 1857 1858 no_sdbt: 1859 /* SDBs (AUX buffer pages) are freed by caller */ 1860 for (i = 0; i < sfb->num_sdbt; i++) 1861 free_page(aux->sdbt_index[i]); 1862 kfree(aux->sdb_index); 1863 no_sdb_index: 1864 kfree(aux->sdbt_index); 1865 no_sdbt_index: 1866 kfree(aux); 1867 no_aux: 1868 return NULL; 1869 } 1870 1871 static void cpumsf_pmu_read(struct perf_event *event) 1872 { 1873 /* Nothing to do ... updates are interrupt-driven */ 1874 } 1875 1876 /* Check if the new sampling period/freqeuncy is appropriate. 1877 * 1878 * Return non-zero on error and zero on passed checks. 1879 */ 1880 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value) 1881 { 1882 struct hws_qsi_info_block si; 1883 unsigned long rate; 1884 bool do_freq; 1885 1886 memset(&si, 0, sizeof(si)); 1887 if (event->cpu == -1) { 1888 if (qsi(&si)) 1889 return -ENODEV; 1890 } else { 1891 /* Event is pinned to a particular CPU, retrieve the per-CPU 1892 * sampling structure for accessing the CPU-specific QSI. 1893 */ 1894 struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu); 1895 1896 si = cpuhw->qsi; 1897 } 1898 1899 do_freq = !!SAMPLE_FREQ_MODE(&event->hw); 1900 rate = getrate(do_freq, value, &si); 1901 if (!rate) 1902 return -EINVAL; 1903 1904 event->attr.sample_period = rate; 1905 SAMPL_RATE(&event->hw) = rate; 1906 hw_init_period(&event->hw, SAMPL_RATE(&event->hw)); 1907 debug_sprintf_event(sfdbg, 4, "%s:" 1908 " cpu %d value %#llx period %#llx freq %d\n", 1909 __func__, event->cpu, value, 1910 event->attr.sample_period, do_freq); 1911 return 0; 1912 } 1913 1914 /* Activate sampling control. 1915 * Next call of pmu_enable() starts sampling. 1916 */ 1917 static void cpumsf_pmu_start(struct perf_event *event, int flags) 1918 { 1919 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1920 1921 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) 1922 return; 1923 1924 if (flags & PERF_EF_RELOAD) 1925 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); 1926 1927 perf_pmu_disable(event->pmu); 1928 event->hw.state = 0; 1929 cpuhw->lsctl.cs = 1; 1930 if (SAMPL_DIAG_MODE(&event->hw)) 1931 cpuhw->lsctl.cd = 1; 1932 perf_pmu_enable(event->pmu); 1933 } 1934 1935 /* Deactivate sampling control. 1936 * Next call of pmu_enable() stops sampling. 1937 */ 1938 static void cpumsf_pmu_stop(struct perf_event *event, int flags) 1939 { 1940 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1941 1942 if (event->hw.state & PERF_HES_STOPPED) 1943 return; 1944 1945 perf_pmu_disable(event->pmu); 1946 cpuhw->lsctl.cs = 0; 1947 cpuhw->lsctl.cd = 0; 1948 event->hw.state |= PERF_HES_STOPPED; 1949 1950 if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) { 1951 hw_perf_event_update(event, 1); 1952 event->hw.state |= PERF_HES_UPTODATE; 1953 } 1954 perf_pmu_enable(event->pmu); 1955 } 1956 1957 static int cpumsf_pmu_add(struct perf_event *event, int flags) 1958 { 1959 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1960 struct aux_buffer *aux; 1961 int err; 1962 1963 if (cpuhw->flags & PMU_F_IN_USE) 1964 return -EAGAIN; 1965 1966 if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt) 1967 return -EINVAL; 1968 1969 err = 0; 1970 perf_pmu_disable(event->pmu); 1971 1972 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED; 1973 1974 /* Set up sampling controls. Always program the sampling register 1975 * using the SDB-table start. Reset TEAR_REG event hardware register 1976 * that is used by hw_perf_event_update() to store the sampling buffer 1977 * position after samples have been flushed. 1978 */ 1979 cpuhw->lsctl.s = 0; 1980 cpuhw->lsctl.h = 1; 1981 cpuhw->lsctl.interval = SAMPL_RATE(&event->hw); 1982 if (!SAMPL_DIAG_MODE(&event->hw)) { 1983 cpuhw->lsctl.tear = virt_to_phys(cpuhw->sfb.sdbt); 1984 cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt; 1985 TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt; 1986 } 1987 1988 /* Ensure sampling functions are in the disabled state. If disabled, 1989 * switch on sampling enable control. */ 1990 if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) { 1991 err = -EAGAIN; 1992 goto out; 1993 } 1994 if (SAMPL_DIAG_MODE(&event->hw)) { 1995 aux = perf_aux_output_begin(&cpuhw->handle, event); 1996 if (!aux) { 1997 err = -EINVAL; 1998 goto out; 1999 } 2000 err = aux_output_begin(&cpuhw->handle, aux, cpuhw); 2001 if (err) 2002 goto out; 2003 cpuhw->lsctl.ed = 1; 2004 } 2005 cpuhw->lsctl.es = 1; 2006 2007 /* Set in_use flag and store event */ 2008 cpuhw->event = event; 2009 cpuhw->flags |= PMU_F_IN_USE; 2010 2011 if (flags & PERF_EF_START) 2012 cpumsf_pmu_start(event, PERF_EF_RELOAD); 2013 out: 2014 perf_event_update_userpage(event); 2015 perf_pmu_enable(event->pmu); 2016 return err; 2017 } 2018 2019 static void cpumsf_pmu_del(struct perf_event *event, int flags) 2020 { 2021 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 2022 2023 perf_pmu_disable(event->pmu); 2024 cpumsf_pmu_stop(event, PERF_EF_UPDATE); 2025 2026 cpuhw->lsctl.es = 0; 2027 cpuhw->lsctl.ed = 0; 2028 cpuhw->flags &= ~PMU_F_IN_USE; 2029 cpuhw->event = NULL; 2030 2031 if (SAMPL_DIAG_MODE(&event->hw)) 2032 aux_output_end(&cpuhw->handle); 2033 perf_event_update_userpage(event); 2034 perf_pmu_enable(event->pmu); 2035 } 2036 2037 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF); 2038 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG); 2039 2040 /* Attribute list for CPU_SF. 2041 * 2042 * The availablitiy depends on the CPU_MF sampling facility authorization 2043 * for basic + diagnositic samples. This is determined at initialization 2044 * time by the sampling facility device driver. 2045 * If the authorization for basic samples is turned off, it should be 2046 * also turned off for diagnostic sampling. 2047 * 2048 * During initialization of the device driver, check the authorization 2049 * level for diagnostic sampling and installs the attribute 2050 * file for diagnostic sampling if necessary. 2051 * 2052 * For now install a placeholder to reference all possible attributes: 2053 * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG. 2054 * Add another entry for the final NULL pointer. 2055 */ 2056 enum { 2057 SF_CYCLES_BASIC_ATTR_IDX = 0, 2058 SF_CYCLES_BASIC_DIAG_ATTR_IDX, 2059 SF_CYCLES_ATTR_MAX 2060 }; 2061 2062 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = { 2063 [SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC) 2064 }; 2065 2066 PMU_FORMAT_ATTR(event, "config:0-63"); 2067 2068 static struct attribute *cpumsf_pmu_format_attr[] = { 2069 &format_attr_event.attr, 2070 NULL, 2071 }; 2072 2073 static struct attribute_group cpumsf_pmu_events_group = { 2074 .name = "events", 2075 .attrs = cpumsf_pmu_events_attr, 2076 }; 2077 2078 static struct attribute_group cpumsf_pmu_format_group = { 2079 .name = "format", 2080 .attrs = cpumsf_pmu_format_attr, 2081 }; 2082 2083 static const struct attribute_group *cpumsf_pmu_attr_groups[] = { 2084 &cpumsf_pmu_events_group, 2085 &cpumsf_pmu_format_group, 2086 NULL, 2087 }; 2088 2089 static struct pmu cpumf_sampling = { 2090 .pmu_enable = cpumsf_pmu_enable, 2091 .pmu_disable = cpumsf_pmu_disable, 2092 2093 .event_init = cpumsf_pmu_event_init, 2094 .add = cpumsf_pmu_add, 2095 .del = cpumsf_pmu_del, 2096 2097 .start = cpumsf_pmu_start, 2098 .stop = cpumsf_pmu_stop, 2099 .read = cpumsf_pmu_read, 2100 2101 .attr_groups = cpumsf_pmu_attr_groups, 2102 2103 .setup_aux = aux_buffer_setup, 2104 .free_aux = aux_buffer_free, 2105 2106 .check_period = cpumsf_pmu_check_period, 2107 }; 2108 2109 static void cpumf_measurement_alert(struct ext_code ext_code, 2110 unsigned int alert, unsigned long unused) 2111 { 2112 struct cpu_hw_sf *cpuhw; 2113 2114 if (!(alert & CPU_MF_INT_SF_MASK)) 2115 return; 2116 inc_irq_stat(IRQEXT_CMS); 2117 cpuhw = this_cpu_ptr(&cpu_hw_sf); 2118 2119 /* Measurement alerts are shared and might happen when the PMU 2120 * is not reserved. Ignore these alerts in this case. */ 2121 if (!(cpuhw->flags & PMU_F_RESERVED)) 2122 return; 2123 2124 /* The processing below must take care of multiple alert events that 2125 * might be indicated concurrently. */ 2126 2127 /* Program alert request */ 2128 if (alert & CPU_MF_INT_SF_PRA) { 2129 if (cpuhw->flags & PMU_F_IN_USE) 2130 if (SAMPL_DIAG_MODE(&cpuhw->event->hw)) 2131 hw_collect_aux(cpuhw); 2132 else 2133 hw_perf_event_update(cpuhw->event, 0); 2134 else 2135 WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE)); 2136 } 2137 2138 /* Report measurement alerts only for non-PRA codes */ 2139 if (alert != CPU_MF_INT_SF_PRA) 2140 debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__, 2141 alert); 2142 2143 /* Sampling authorization change request */ 2144 if (alert & CPU_MF_INT_SF_SACA) 2145 qsi(&cpuhw->qsi); 2146 2147 /* Loss of sample data due to high-priority machine activities */ 2148 if (alert & CPU_MF_INT_SF_LSDA) { 2149 pr_err("Sample data was lost\n"); 2150 cpuhw->flags |= PMU_F_ERR_LSDA; 2151 sf_disable(); 2152 } 2153 2154 /* Invalid sampling buffer entry */ 2155 if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) { 2156 pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n", 2157 alert); 2158 cpuhw->flags |= PMU_F_ERR_IBE; 2159 sf_disable(); 2160 } 2161 } 2162 2163 static int cpusf_pmu_setup(unsigned int cpu, int flags) 2164 { 2165 /* Ignore the notification if no events are scheduled on the PMU. 2166 * This might be racy... 2167 */ 2168 if (!atomic_read(&num_events)) 2169 return 0; 2170 2171 local_irq_disable(); 2172 setup_pmc_cpu(&flags); 2173 local_irq_enable(); 2174 return 0; 2175 } 2176 2177 static int s390_pmu_sf_online_cpu(unsigned int cpu) 2178 { 2179 return cpusf_pmu_setup(cpu, PMC_INIT); 2180 } 2181 2182 static int s390_pmu_sf_offline_cpu(unsigned int cpu) 2183 { 2184 return cpusf_pmu_setup(cpu, PMC_RELEASE); 2185 } 2186 2187 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp) 2188 { 2189 if (!cpum_sf_avail()) 2190 return -ENODEV; 2191 return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 2192 } 2193 2194 static int param_set_sfb_size(const char *val, const struct kernel_param *kp) 2195 { 2196 int rc; 2197 unsigned long min, max; 2198 2199 if (!cpum_sf_avail()) 2200 return -ENODEV; 2201 if (!val || !strlen(val)) 2202 return -EINVAL; 2203 2204 /* Valid parameter values: "min,max" or "max" */ 2205 min = CPUM_SF_MIN_SDB; 2206 max = CPUM_SF_MAX_SDB; 2207 if (strchr(val, ',')) 2208 rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL; 2209 else 2210 rc = kstrtoul(val, 10, &max); 2211 2212 if (min < 2 || min >= max || max > get_num_physpages()) 2213 rc = -EINVAL; 2214 if (rc) 2215 return rc; 2216 2217 sfb_set_limits(min, max); 2218 pr_info("The sampling buffer limits have changed to: " 2219 "min %lu max %lu (diag %lu)\n", 2220 CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR); 2221 return 0; 2222 } 2223 2224 #define param_check_sfb_size(name, p) __param_check(name, p, void) 2225 static const struct kernel_param_ops param_ops_sfb_size = { 2226 .set = param_set_sfb_size, 2227 .get = param_get_sfb_size, 2228 }; 2229 2230 #define RS_INIT_FAILURE_QSI 0x0001 2231 #define RS_INIT_FAILURE_BSDES 0x0002 2232 #define RS_INIT_FAILURE_ALRT 0x0003 2233 #define RS_INIT_FAILURE_PERF 0x0004 2234 static void __init pr_cpumsf_err(unsigned int reason) 2235 { 2236 pr_err("Sampling facility support for perf is not available: " 2237 "reason %#x\n", reason); 2238 } 2239 2240 static int __init init_cpum_sampling_pmu(void) 2241 { 2242 struct hws_qsi_info_block si; 2243 int err; 2244 2245 if (!cpum_sf_avail()) 2246 return -ENODEV; 2247 2248 memset(&si, 0, sizeof(si)); 2249 if (qsi(&si)) { 2250 pr_cpumsf_err(RS_INIT_FAILURE_QSI); 2251 return -ENODEV; 2252 } 2253 2254 if (!si.as && !si.ad) 2255 return -ENODEV; 2256 2257 if (si.bsdes != sizeof(struct hws_basic_entry)) { 2258 pr_cpumsf_err(RS_INIT_FAILURE_BSDES); 2259 return -EINVAL; 2260 } 2261 2262 if (si.ad) { 2263 sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 2264 /* Sampling of diagnostic data authorized, 2265 * install event into attribute list of PMU device. 2266 */ 2267 cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] = 2268 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG); 2269 } 2270 2271 sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80); 2272 if (!sfdbg) { 2273 pr_err("Registering for s390dbf failed\n"); 2274 return -ENOMEM; 2275 } 2276 debug_register_view(sfdbg, &debug_sprintf_view); 2277 2278 err = register_external_irq(EXT_IRQ_MEASURE_ALERT, 2279 cpumf_measurement_alert); 2280 if (err) { 2281 pr_cpumsf_err(RS_INIT_FAILURE_ALRT); 2282 debug_unregister(sfdbg); 2283 goto out; 2284 } 2285 2286 err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW); 2287 if (err) { 2288 pr_cpumsf_err(RS_INIT_FAILURE_PERF); 2289 unregister_external_irq(EXT_IRQ_MEASURE_ALERT, 2290 cpumf_measurement_alert); 2291 debug_unregister(sfdbg); 2292 goto out; 2293 } 2294 2295 cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online", 2296 s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu); 2297 out: 2298 return err; 2299 } 2300 2301 arch_initcall(init_cpum_sampling_pmu); 2302 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644); 2303