1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Performance event support for the System z CPU-measurement Sampling Facility 4 * 5 * Copyright IBM Corp. 2013, 2018 6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com> 7 */ 8 #define KMSG_COMPONENT "cpum_sf" 9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/kernel_stat.h> 13 #include <linux/perf_event.h> 14 #include <linux/percpu.h> 15 #include <linux/pid.h> 16 #include <linux/notifier.h> 17 #include <linux/export.h> 18 #include <linux/slab.h> 19 #include <linux/mm.h> 20 #include <linux/moduleparam.h> 21 #include <asm/cpu_mf.h> 22 #include <asm/irq.h> 23 #include <asm/debug.h> 24 #include <asm/timex.h> 25 #include <asm-generic/io.h> 26 27 /* Minimum number of sample-data-block-tables: 28 * At least one table is required for the sampling buffer structure. 29 * A single table contains up to 511 pointers to sample-data-blocks. 30 */ 31 #define CPUM_SF_MIN_SDBT 1 32 33 /* Number of sample-data-blocks per sample-data-block-table (SDBT): 34 * A table contains SDB pointers (8 bytes) and one table-link entry 35 * that points to the origin of the next SDBT. 36 */ 37 #define CPUM_SF_SDB_PER_TABLE ((PAGE_SIZE - 8) / 8) 38 39 /* Maximum page offset for an SDBT table-link entry: 40 * If this page offset is reached, a table-link entry to the next SDBT 41 * must be added. 42 */ 43 #define CPUM_SF_SDBT_TL_OFFSET (CPUM_SF_SDB_PER_TABLE * 8) 44 static inline int require_table_link(const void *sdbt) 45 { 46 return ((unsigned long) sdbt & ~PAGE_MASK) == CPUM_SF_SDBT_TL_OFFSET; 47 } 48 49 /* Minimum and maximum sampling buffer sizes: 50 * 51 * This number represents the maximum size of the sampling buffer taking 52 * the number of sample-data-block-tables into account. Note that these 53 * numbers apply to the basic-sampling function only. 54 * The maximum number of SDBs is increased by CPUM_SF_SDB_DIAG_FACTOR if 55 * the diagnostic-sampling function is active. 56 * 57 * Sampling buffer size Buffer characteristics 58 * --------------------------------------------------- 59 * 64KB == 16 pages (4KB per page) 60 * 1 page for SDB-tables 61 * 15 pages for SDBs 62 * 63 * 32MB == 8192 pages (4KB per page) 64 * 16 pages for SDB-tables 65 * 8176 pages for SDBs 66 */ 67 static unsigned long __read_mostly CPUM_SF_MIN_SDB = 15; 68 static unsigned long __read_mostly CPUM_SF_MAX_SDB = 8176; 69 static unsigned long __read_mostly CPUM_SF_SDB_DIAG_FACTOR = 1; 70 71 struct sf_buffer { 72 unsigned long *sdbt; /* Sample-data-block-table origin */ 73 /* buffer characteristics (required for buffer increments) */ 74 unsigned long num_sdb; /* Number of sample-data-blocks */ 75 unsigned long num_sdbt; /* Number of sample-data-block-tables */ 76 unsigned long *tail; /* last sample-data-block-table */ 77 }; 78 79 struct aux_buffer { 80 struct sf_buffer sfb; 81 unsigned long head; /* index of SDB of buffer head */ 82 unsigned long alert_mark; /* index of SDB of alert request position */ 83 unsigned long empty_mark; /* mark of SDB not marked full */ 84 unsigned long *sdb_index; /* SDB address for fast lookup */ 85 unsigned long *sdbt_index; /* SDBT address for fast lookup */ 86 }; 87 88 struct cpu_hw_sf { 89 /* CPU-measurement sampling information block */ 90 struct hws_qsi_info_block qsi; 91 /* CPU-measurement sampling control block */ 92 struct hws_lsctl_request_block lsctl; 93 struct sf_buffer sfb; /* Sampling buffer */ 94 unsigned int flags; /* Status flags */ 95 struct perf_event *event; /* Scheduled perf event */ 96 struct perf_output_handle handle; /* AUX buffer output handle */ 97 }; 98 static DEFINE_PER_CPU(struct cpu_hw_sf, cpu_hw_sf); 99 100 /* Debug feature */ 101 static debug_info_t *sfdbg; 102 103 /* Sampling control helper functions */ 104 static inline unsigned long freq_to_sample_rate(struct hws_qsi_info_block *qsi, 105 unsigned long freq) 106 { 107 return (USEC_PER_SEC / freq) * qsi->cpu_speed; 108 } 109 110 static inline unsigned long sample_rate_to_freq(struct hws_qsi_info_block *qsi, 111 unsigned long rate) 112 { 113 return USEC_PER_SEC * qsi->cpu_speed / rate; 114 } 115 116 /* Return TOD timestamp contained in an trailer entry */ 117 static inline unsigned long long trailer_timestamp(struct hws_trailer_entry *te) 118 { 119 /* TOD in STCKE format */ 120 if (te->header.t) 121 return *((unsigned long long *)&te->timestamp[1]); 122 123 /* TOD in STCK format */ 124 return *((unsigned long long *)&te->timestamp[0]); 125 } 126 127 /* Return pointer to trailer entry of an sample data block */ 128 static inline struct hws_trailer_entry *trailer_entry_ptr(unsigned long v) 129 { 130 void *ret; 131 132 ret = (void *)v; 133 ret += PAGE_SIZE; 134 ret -= sizeof(struct hws_trailer_entry); 135 136 return ret; 137 } 138 139 /* 140 * Return true if the entry in the sample data block table (sdbt) 141 * is a link to the next sdbt 142 */ 143 static inline int is_link_entry(unsigned long *s) 144 { 145 return *s & 0x1UL ? 1 : 0; 146 } 147 148 /* Return pointer to the linked sdbt */ 149 static inline unsigned long *get_next_sdbt(unsigned long *s) 150 { 151 return phys_to_virt(*s & ~0x1UL); 152 } 153 154 /* 155 * sf_disable() - Switch off sampling facility 156 */ 157 static int sf_disable(void) 158 { 159 struct hws_lsctl_request_block sreq; 160 161 memset(&sreq, 0, sizeof(sreq)); 162 return lsctl(&sreq); 163 } 164 165 /* 166 * sf_buffer_available() - Check for an allocated sampling buffer 167 */ 168 static int sf_buffer_available(struct cpu_hw_sf *cpuhw) 169 { 170 return !!cpuhw->sfb.sdbt; 171 } 172 173 /* 174 * deallocate sampling facility buffer 175 */ 176 static void free_sampling_buffer(struct sf_buffer *sfb) 177 { 178 unsigned long *sdbt, *curr; 179 180 if (!sfb->sdbt) 181 return; 182 183 sdbt = sfb->sdbt; 184 curr = sdbt; 185 186 /* Free the SDBT after all SDBs are processed... */ 187 while (1) { 188 if (!*curr || !sdbt) 189 break; 190 191 /* Process table-link entries */ 192 if (is_link_entry(curr)) { 193 curr = get_next_sdbt(curr); 194 if (sdbt) 195 free_page((unsigned long) sdbt); 196 197 /* If the origin is reached, sampling buffer is freed */ 198 if (curr == sfb->sdbt) 199 break; 200 else 201 sdbt = curr; 202 } else { 203 /* Process SDB pointer */ 204 if (*curr) { 205 free_page((unsigned long)phys_to_virt(*curr)); 206 curr++; 207 } 208 } 209 } 210 211 debug_sprintf_event(sfdbg, 5, "%s: freed sdbt %#lx\n", __func__, 212 (unsigned long)sfb->sdbt); 213 memset(sfb, 0, sizeof(*sfb)); 214 } 215 216 static int alloc_sample_data_block(unsigned long *sdbt, gfp_t gfp_flags) 217 { 218 struct hws_trailer_entry *te; 219 unsigned long sdb; 220 221 /* Allocate and initialize sample-data-block */ 222 sdb = get_zeroed_page(gfp_flags); 223 if (!sdb) 224 return -ENOMEM; 225 te = trailer_entry_ptr(sdb); 226 te->header.a = 1; 227 228 /* Link SDB into the sample-data-block-table */ 229 *sdbt = virt_to_phys((void *)sdb); 230 231 return 0; 232 } 233 234 /* 235 * realloc_sampling_buffer() - extend sampler memory 236 * 237 * Allocates new sample-data-blocks and adds them to the specified sampling 238 * buffer memory. 239 * 240 * Important: This modifies the sampling buffer and must be called when the 241 * sampling facility is disabled. 242 * 243 * Returns zero on success, non-zero otherwise. 244 */ 245 static int realloc_sampling_buffer(struct sf_buffer *sfb, 246 unsigned long num_sdb, gfp_t gfp_flags) 247 { 248 int i, rc; 249 unsigned long *new, *tail, *tail_prev = NULL; 250 251 if (!sfb->sdbt || !sfb->tail) 252 return -EINVAL; 253 254 if (!is_link_entry(sfb->tail)) 255 return -EINVAL; 256 257 /* Append to the existing sampling buffer, overwriting the table-link 258 * register. 259 * The tail variables always points to the "tail" (last and table-link) 260 * entry in an SDB-table. 261 */ 262 tail = sfb->tail; 263 264 /* Do a sanity check whether the table-link entry points to 265 * the sampling buffer origin. 266 */ 267 if (sfb->sdbt != get_next_sdbt(tail)) { 268 debug_sprintf_event(sfdbg, 3, "%s: " 269 "sampling buffer is not linked: origin %#lx" 270 " tail %#lx\n", __func__, 271 (unsigned long)sfb->sdbt, 272 (unsigned long)tail); 273 return -EINVAL; 274 } 275 276 /* Allocate remaining SDBs */ 277 rc = 0; 278 for (i = 0; i < num_sdb; i++) { 279 /* Allocate a new SDB-table if it is full. */ 280 if (require_table_link(tail)) { 281 new = (unsigned long *) get_zeroed_page(gfp_flags); 282 if (!new) { 283 rc = -ENOMEM; 284 break; 285 } 286 sfb->num_sdbt++; 287 /* Link current page to tail of chain */ 288 *tail = virt_to_phys((void *)new) + 1; 289 tail_prev = tail; 290 tail = new; 291 } 292 293 /* Allocate a new sample-data-block. 294 * If there is not enough memory, stop the realloc process 295 * and simply use what was allocated. If this is a temporary 296 * issue, a new realloc call (if required) might succeed. 297 */ 298 rc = alloc_sample_data_block(tail, gfp_flags); 299 if (rc) { 300 /* Undo last SDBT. An SDBT with no SDB at its first 301 * entry but with an SDBT entry instead can not be 302 * handled by the interrupt handler code. 303 * Avoid this situation. 304 */ 305 if (tail_prev) { 306 sfb->num_sdbt--; 307 free_page((unsigned long) new); 308 tail = tail_prev; 309 } 310 break; 311 } 312 sfb->num_sdb++; 313 tail++; 314 tail_prev = new = NULL; /* Allocated at least one SBD */ 315 } 316 317 /* Link sampling buffer to its origin */ 318 *tail = virt_to_phys(sfb->sdbt) + 1; 319 sfb->tail = tail; 320 321 debug_sprintf_event(sfdbg, 4, "%s: new buffer" 322 " settings: sdbt %lu sdb %lu\n", __func__, 323 sfb->num_sdbt, sfb->num_sdb); 324 return rc; 325 } 326 327 /* 328 * allocate_sampling_buffer() - allocate sampler memory 329 * 330 * Allocates and initializes a sampling buffer structure using the 331 * specified number of sample-data-blocks (SDB). For each allocation, 332 * a 4K page is used. The number of sample-data-block-tables (SDBT) 333 * are calculated from SDBs. 334 * Also set the ALERT_REQ mask in each SDBs trailer. 335 * 336 * Returns zero on success, non-zero otherwise. 337 */ 338 static int alloc_sampling_buffer(struct sf_buffer *sfb, unsigned long num_sdb) 339 { 340 int rc; 341 342 if (sfb->sdbt) 343 return -EINVAL; 344 345 /* Allocate the sample-data-block-table origin */ 346 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 347 if (!sfb->sdbt) 348 return -ENOMEM; 349 sfb->num_sdb = 0; 350 sfb->num_sdbt = 1; 351 352 /* Link the table origin to point to itself to prepare for 353 * realloc_sampling_buffer() invocation. 354 */ 355 sfb->tail = sfb->sdbt; 356 *sfb->tail = virt_to_phys((void *)sfb->sdbt) + 1; 357 358 /* Allocate requested number of sample-data-blocks */ 359 rc = realloc_sampling_buffer(sfb, num_sdb, GFP_KERNEL); 360 if (rc) { 361 free_sampling_buffer(sfb); 362 debug_sprintf_event(sfdbg, 4, "%s: " 363 "realloc_sampling_buffer failed with rc %i\n", 364 __func__, rc); 365 } else 366 debug_sprintf_event(sfdbg, 4, 367 "%s: tear %#lx dear %#lx\n", __func__, 368 (unsigned long)sfb->sdbt, (unsigned long)*sfb->sdbt); 369 return rc; 370 } 371 372 static void sfb_set_limits(unsigned long min, unsigned long max) 373 { 374 struct hws_qsi_info_block si; 375 376 CPUM_SF_MIN_SDB = min; 377 CPUM_SF_MAX_SDB = max; 378 379 memset(&si, 0, sizeof(si)); 380 if (!qsi(&si)) 381 CPUM_SF_SDB_DIAG_FACTOR = DIV_ROUND_UP(si.dsdes, si.bsdes); 382 } 383 384 static unsigned long sfb_max_limit(struct hw_perf_event *hwc) 385 { 386 return SAMPL_DIAG_MODE(hwc) ? CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR 387 : CPUM_SF_MAX_SDB; 388 } 389 390 static unsigned long sfb_pending_allocs(struct sf_buffer *sfb, 391 struct hw_perf_event *hwc) 392 { 393 if (!sfb->sdbt) 394 return SFB_ALLOC_REG(hwc); 395 if (SFB_ALLOC_REG(hwc) > sfb->num_sdb) 396 return SFB_ALLOC_REG(hwc) - sfb->num_sdb; 397 return 0; 398 } 399 400 static int sfb_has_pending_allocs(struct sf_buffer *sfb, 401 struct hw_perf_event *hwc) 402 { 403 return sfb_pending_allocs(sfb, hwc) > 0; 404 } 405 406 static void sfb_account_allocs(unsigned long num, struct hw_perf_event *hwc) 407 { 408 /* Limit the number of SDBs to not exceed the maximum */ 409 num = min_t(unsigned long, num, sfb_max_limit(hwc) - SFB_ALLOC_REG(hwc)); 410 if (num) 411 SFB_ALLOC_REG(hwc) += num; 412 } 413 414 static void sfb_init_allocs(unsigned long num, struct hw_perf_event *hwc) 415 { 416 SFB_ALLOC_REG(hwc) = 0; 417 sfb_account_allocs(num, hwc); 418 } 419 420 static void deallocate_buffers(struct cpu_hw_sf *cpuhw) 421 { 422 if (cpuhw->sfb.sdbt) 423 free_sampling_buffer(&cpuhw->sfb); 424 } 425 426 static int allocate_buffers(struct cpu_hw_sf *cpuhw, struct hw_perf_event *hwc) 427 { 428 unsigned long n_sdb, freq; 429 size_t sample_size; 430 431 /* Calculate sampling buffers using 4K pages 432 * 433 * 1. The sampling size is 32 bytes for basic sampling. This size 434 * is the same for all machine types. Diagnostic 435 * sampling uses auxlilary data buffer setup which provides the 436 * memory for SDBs using linux common code auxiliary trace 437 * setup. 438 * 439 * 2. Function alloc_sampling_buffer() sets the Alert Request 440 * Control indicator to trigger a measurement-alert to harvest 441 * sample-data-blocks (SDB). This is done per SDB. This 442 * measurement alert interrupt fires quick enough to handle 443 * one SDB, on very high frequency and work loads there might 444 * be 2 to 3 SBDs available for sample processing. 445 * Currently there is no need for setup alert request on every 446 * n-th page. This is counterproductive as one IRQ triggers 447 * a very high number of samples to be processed at one IRQ. 448 * 449 * 3. Use the sampling frequency as input. 450 * Compute the number of SDBs and ensure a minimum 451 * of CPUM_SF_MIN_SDB. Depending on frequency add some more 452 * SDBs to handle a higher sampling rate. 453 * Use a minimum of CPUM_SF_MIN_SDB and allow for 100 samples 454 * (one SDB) for every 10000 HZ frequency increment. 455 * 456 * 4. Compute the number of sample-data-block-tables (SDBT) and 457 * ensure a minimum of CPUM_SF_MIN_SDBT (one table can manage up 458 * to 511 SDBs). 459 */ 460 sample_size = sizeof(struct hws_basic_entry); 461 freq = sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc)); 462 n_sdb = CPUM_SF_MIN_SDB + DIV_ROUND_UP(freq, 10000); 463 464 /* If there is already a sampling buffer allocated, it is very likely 465 * that the sampling facility is enabled too. If the event to be 466 * initialized requires a greater sampling buffer, the allocation must 467 * be postponed. Changing the sampling buffer requires the sampling 468 * facility to be in the disabled state. So, account the number of 469 * required SDBs and let cpumsf_pmu_enable() resize the buffer just 470 * before the event is started. 471 */ 472 sfb_init_allocs(n_sdb, hwc); 473 if (sf_buffer_available(cpuhw)) 474 return 0; 475 476 debug_sprintf_event(sfdbg, 3, 477 "%s: rate %lu f %lu sdb %lu/%lu" 478 " sample_size %lu cpuhw %p\n", __func__, 479 SAMPL_RATE(hwc), freq, n_sdb, sfb_max_limit(hwc), 480 sample_size, cpuhw); 481 482 return alloc_sampling_buffer(&cpuhw->sfb, 483 sfb_pending_allocs(&cpuhw->sfb, hwc)); 484 } 485 486 static unsigned long min_percent(unsigned int percent, unsigned long base, 487 unsigned long min) 488 { 489 return min_t(unsigned long, min, DIV_ROUND_UP(percent * base, 100)); 490 } 491 492 static unsigned long compute_sfb_extent(unsigned long ratio, unsigned long base) 493 { 494 /* Use a percentage-based approach to extend the sampling facility 495 * buffer. Accept up to 5% sample data loss. 496 * Vary the extents between 1% to 5% of the current number of 497 * sample-data-blocks. 498 */ 499 if (ratio <= 5) 500 return 0; 501 if (ratio <= 25) 502 return min_percent(1, base, 1); 503 if (ratio <= 50) 504 return min_percent(1, base, 1); 505 if (ratio <= 75) 506 return min_percent(2, base, 2); 507 if (ratio <= 100) 508 return min_percent(3, base, 3); 509 if (ratio <= 250) 510 return min_percent(4, base, 4); 511 512 return min_percent(5, base, 8); 513 } 514 515 static void sfb_account_overflows(struct cpu_hw_sf *cpuhw, 516 struct hw_perf_event *hwc) 517 { 518 unsigned long ratio, num; 519 520 if (!OVERFLOW_REG(hwc)) 521 return; 522 523 /* The sample_overflow contains the average number of sample data 524 * that has been lost because sample-data-blocks were full. 525 * 526 * Calculate the total number of sample data entries that has been 527 * discarded. Then calculate the ratio of lost samples to total samples 528 * per second in percent. 529 */ 530 ratio = DIV_ROUND_UP(100 * OVERFLOW_REG(hwc) * cpuhw->sfb.num_sdb, 531 sample_rate_to_freq(&cpuhw->qsi, SAMPL_RATE(hwc))); 532 533 /* Compute number of sample-data-blocks */ 534 num = compute_sfb_extent(ratio, cpuhw->sfb.num_sdb); 535 if (num) 536 sfb_account_allocs(num, hwc); 537 538 debug_sprintf_event(sfdbg, 5, "%s: overflow %llu ratio %lu num %lu\n", 539 __func__, OVERFLOW_REG(hwc), ratio, num); 540 OVERFLOW_REG(hwc) = 0; 541 } 542 543 /* extend_sampling_buffer() - Extend sampling buffer 544 * @sfb: Sampling buffer structure (for local CPU) 545 * @hwc: Perf event hardware structure 546 * 547 * Use this function to extend the sampling buffer based on the overflow counter 548 * and postponed allocation extents stored in the specified Perf event hardware. 549 * 550 * Important: This function disables the sampling facility in order to safely 551 * change the sampling buffer structure. Do not call this function 552 * when the PMU is active. 553 */ 554 static void extend_sampling_buffer(struct sf_buffer *sfb, 555 struct hw_perf_event *hwc) 556 { 557 unsigned long num, num_old; 558 int rc; 559 560 num = sfb_pending_allocs(sfb, hwc); 561 if (!num) 562 return; 563 num_old = sfb->num_sdb; 564 565 /* Disable the sampling facility to reset any states and also 566 * clear pending measurement alerts. 567 */ 568 sf_disable(); 569 570 /* Extend the sampling buffer. 571 * This memory allocation typically happens in an atomic context when 572 * called by perf. Because this is a reallocation, it is fine if the 573 * new SDB-request cannot be satisfied immediately. 574 */ 575 rc = realloc_sampling_buffer(sfb, num, GFP_ATOMIC); 576 if (rc) 577 debug_sprintf_event(sfdbg, 5, "%s: realloc failed with rc %i\n", 578 __func__, rc); 579 580 if (sfb_has_pending_allocs(sfb, hwc)) 581 debug_sprintf_event(sfdbg, 5, "%s: " 582 "req %lu alloc %lu remaining %lu\n", 583 __func__, num, sfb->num_sdb - num_old, 584 sfb_pending_allocs(sfb, hwc)); 585 } 586 587 /* Number of perf events counting hardware events */ 588 static atomic_t num_events; 589 /* Used to avoid races in calling reserve/release_cpumf_hardware */ 590 static DEFINE_MUTEX(pmc_reserve_mutex); 591 592 #define PMC_INIT 0 593 #define PMC_RELEASE 1 594 #define PMC_FAILURE 2 595 static void setup_pmc_cpu(void *flags) 596 { 597 int err; 598 struct cpu_hw_sf *cpusf = this_cpu_ptr(&cpu_hw_sf); 599 600 err = 0; 601 switch (*((int *) flags)) { 602 case PMC_INIT: 603 memset(cpusf, 0, sizeof(*cpusf)); 604 err = qsi(&cpusf->qsi); 605 if (err) 606 break; 607 cpusf->flags |= PMU_F_RESERVED; 608 err = sf_disable(); 609 if (err) 610 pr_err("Switching off the sampling facility failed " 611 "with rc %i\n", err); 612 break; 613 case PMC_RELEASE: 614 cpusf->flags &= ~PMU_F_RESERVED; 615 err = sf_disable(); 616 if (err) { 617 pr_err("Switching off the sampling facility failed " 618 "with rc %i\n", err); 619 } else 620 deallocate_buffers(cpusf); 621 break; 622 } 623 if (err) 624 *((int *) flags) |= PMC_FAILURE; 625 } 626 627 static void release_pmc_hardware(void) 628 { 629 int flags = PMC_RELEASE; 630 631 irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT); 632 on_each_cpu(setup_pmc_cpu, &flags, 1); 633 } 634 635 static int reserve_pmc_hardware(void) 636 { 637 int flags = PMC_INIT; 638 639 on_each_cpu(setup_pmc_cpu, &flags, 1); 640 if (flags & PMC_FAILURE) { 641 release_pmc_hardware(); 642 return -ENODEV; 643 } 644 irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT); 645 646 return 0; 647 } 648 649 static void hw_perf_event_destroy(struct perf_event *event) 650 { 651 /* Release PMC if this is the last perf event */ 652 if (!atomic_add_unless(&num_events, -1, 1)) { 653 mutex_lock(&pmc_reserve_mutex); 654 if (atomic_dec_return(&num_events) == 0) 655 release_pmc_hardware(); 656 mutex_unlock(&pmc_reserve_mutex); 657 } 658 } 659 660 static void hw_init_period(struct hw_perf_event *hwc, u64 period) 661 { 662 hwc->sample_period = period; 663 hwc->last_period = hwc->sample_period; 664 local64_set(&hwc->period_left, hwc->sample_period); 665 } 666 667 static unsigned long hw_limit_rate(const struct hws_qsi_info_block *si, 668 unsigned long rate) 669 { 670 return clamp_t(unsigned long, rate, 671 si->min_sampl_rate, si->max_sampl_rate); 672 } 673 674 static u32 cpumsf_pid_type(struct perf_event *event, 675 u32 pid, enum pid_type type) 676 { 677 struct task_struct *tsk; 678 679 /* Idle process */ 680 if (!pid) 681 goto out; 682 683 tsk = find_task_by_pid_ns(pid, &init_pid_ns); 684 pid = -1; 685 if (tsk) { 686 /* 687 * Only top level events contain the pid namespace in which 688 * they are created. 689 */ 690 if (event->parent) 691 event = event->parent; 692 pid = __task_pid_nr_ns(tsk, type, event->ns); 693 /* 694 * See also 1d953111b648 695 * "perf/core: Don't report zero PIDs for exiting tasks". 696 */ 697 if (!pid && !pid_alive(tsk)) 698 pid = -1; 699 } 700 out: 701 return pid; 702 } 703 704 static void cpumsf_output_event_pid(struct perf_event *event, 705 struct perf_sample_data *data, 706 struct pt_regs *regs) 707 { 708 u32 pid; 709 struct perf_event_header header; 710 struct perf_output_handle handle; 711 712 /* 713 * Obtain the PID from the basic-sampling data entry and 714 * correct the data->tid_entry.pid value. 715 */ 716 pid = data->tid_entry.pid; 717 718 /* Protect callchain buffers, tasks */ 719 rcu_read_lock(); 720 721 perf_prepare_sample(data, event, regs); 722 perf_prepare_header(&header, data, event, regs); 723 if (perf_output_begin(&handle, data, event, header.size)) 724 goto out; 725 726 /* Update the process ID (see also kernel/events/core.c) */ 727 data->tid_entry.pid = cpumsf_pid_type(event, pid, PIDTYPE_TGID); 728 data->tid_entry.tid = cpumsf_pid_type(event, pid, PIDTYPE_PID); 729 730 perf_output_sample(&handle, &header, data, event); 731 perf_output_end(&handle); 732 out: 733 rcu_read_unlock(); 734 } 735 736 static unsigned long getrate(bool freq, unsigned long sample, 737 struct hws_qsi_info_block *si) 738 { 739 unsigned long rate; 740 741 if (freq) { 742 rate = freq_to_sample_rate(si, sample); 743 rate = hw_limit_rate(si, rate); 744 } else { 745 /* The min/max sampling rates specifies the valid range 746 * of sample periods. If the specified sample period is 747 * out of range, limit the period to the range boundary. 748 */ 749 rate = hw_limit_rate(si, sample); 750 751 /* The perf core maintains a maximum sample rate that is 752 * configurable through the sysctl interface. Ensure the 753 * sampling rate does not exceed this value. This also helps 754 * to avoid throttling when pushing samples with 755 * perf_event_overflow(). 756 */ 757 if (sample_rate_to_freq(si, rate) > 758 sysctl_perf_event_sample_rate) { 759 debug_sprintf_event(sfdbg, 1, "%s: " 760 "Sampling rate exceeds maximum " 761 "perf sample rate\n", __func__); 762 rate = 0; 763 } 764 } 765 return rate; 766 } 767 768 /* The sampling information (si) contains information about the 769 * min/max sampling intervals and the CPU speed. So calculate the 770 * correct sampling interval and avoid the whole period adjust 771 * feedback loop. 772 * 773 * Since the CPU Measurement sampling facility can not handle frequency 774 * calculate the sampling interval when frequency is specified using 775 * this formula: 776 * interval := cpu_speed * 1000000 / sample_freq 777 * 778 * Returns errno on bad input and zero on success with parameter interval 779 * set to the correct sampling rate. 780 * 781 * Note: This function turns off freq bit to avoid calling function 782 * perf_adjust_period(). This causes frequency adjustment in the common 783 * code part which causes tremendous variations in the counter values. 784 */ 785 static int __hw_perf_event_init_rate(struct perf_event *event, 786 struct hws_qsi_info_block *si) 787 { 788 struct perf_event_attr *attr = &event->attr; 789 struct hw_perf_event *hwc = &event->hw; 790 unsigned long rate; 791 792 if (attr->freq) { 793 if (!attr->sample_freq) 794 return -EINVAL; 795 rate = getrate(attr->freq, attr->sample_freq, si); 796 attr->freq = 0; /* Don't call perf_adjust_period() */ 797 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FREQ_MODE; 798 } else { 799 rate = getrate(attr->freq, attr->sample_period, si); 800 if (!rate) 801 return -EINVAL; 802 } 803 attr->sample_period = rate; 804 SAMPL_RATE(hwc) = rate; 805 hw_init_period(hwc, SAMPL_RATE(hwc)); 806 debug_sprintf_event(sfdbg, 4, "%s: cpu %d period %#llx freq %d,%#lx\n", 807 __func__, event->cpu, event->attr.sample_period, 808 event->attr.freq, SAMPLE_FREQ_MODE(hwc)); 809 return 0; 810 } 811 812 static int __hw_perf_event_init(struct perf_event *event) 813 { 814 struct cpu_hw_sf *cpuhw; 815 struct hws_qsi_info_block si; 816 struct perf_event_attr *attr = &event->attr; 817 struct hw_perf_event *hwc = &event->hw; 818 int cpu, err; 819 820 /* Reserve CPU-measurement sampling facility */ 821 err = 0; 822 if (!atomic_inc_not_zero(&num_events)) { 823 mutex_lock(&pmc_reserve_mutex); 824 if (atomic_read(&num_events) == 0 && reserve_pmc_hardware()) 825 err = -EBUSY; 826 else 827 atomic_inc(&num_events); 828 mutex_unlock(&pmc_reserve_mutex); 829 } 830 event->destroy = hw_perf_event_destroy; 831 832 if (err) 833 goto out; 834 835 /* Access per-CPU sampling information (query sampling info) */ 836 /* 837 * The event->cpu value can be -1 to count on every CPU, for example, 838 * when attaching to a task. If this is specified, use the query 839 * sampling info from the current CPU, otherwise use event->cpu to 840 * retrieve the per-CPU information. 841 * Later, cpuhw indicates whether to allocate sampling buffers for a 842 * particular CPU (cpuhw!=NULL) or each online CPU (cpuw==NULL). 843 */ 844 memset(&si, 0, sizeof(si)); 845 cpuhw = NULL; 846 if (event->cpu == -1) 847 qsi(&si); 848 else { 849 /* Event is pinned to a particular CPU, retrieve the per-CPU 850 * sampling structure for accessing the CPU-specific QSI. 851 */ 852 cpuhw = &per_cpu(cpu_hw_sf, event->cpu); 853 si = cpuhw->qsi; 854 } 855 856 /* Check sampling facility authorization and, if not authorized, 857 * fall back to other PMUs. It is safe to check any CPU because 858 * the authorization is identical for all configured CPUs. 859 */ 860 if (!si.as) { 861 err = -ENOENT; 862 goto out; 863 } 864 865 if (si.ribm & CPU_MF_SF_RIBM_NOTAV) { 866 pr_warn("CPU Measurement Facility sampling is temporarily not available\n"); 867 err = -EBUSY; 868 goto out; 869 } 870 871 /* Always enable basic sampling */ 872 SAMPL_FLAGS(hwc) = PERF_CPUM_SF_BASIC_MODE; 873 874 /* Check if diagnostic sampling is requested. Deny if the required 875 * sampling authorization is missing. 876 */ 877 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) { 878 if (!si.ad) { 879 err = -EPERM; 880 goto out; 881 } 882 SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE; 883 } 884 885 err = __hw_perf_event_init_rate(event, &si); 886 if (err) 887 goto out; 888 889 /* Initialize sample data overflow accounting */ 890 hwc->extra_reg.reg = REG_OVERFLOW; 891 OVERFLOW_REG(hwc) = 0; 892 893 /* Use AUX buffer. No need to allocate it by ourself */ 894 if (attr->config == PERF_EVENT_CPUM_SF_DIAG) 895 return 0; 896 897 /* Allocate the per-CPU sampling buffer using the CPU information 898 * from the event. If the event is not pinned to a particular 899 * CPU (event->cpu == -1; or cpuhw == NULL), allocate sampling 900 * buffers for each online CPU. 901 */ 902 if (cpuhw) 903 /* Event is pinned to a particular CPU */ 904 err = allocate_buffers(cpuhw, hwc); 905 else { 906 /* Event is not pinned, allocate sampling buffer on 907 * each online CPU 908 */ 909 for_each_online_cpu(cpu) { 910 cpuhw = &per_cpu(cpu_hw_sf, cpu); 911 err = allocate_buffers(cpuhw, hwc); 912 if (err) 913 break; 914 } 915 } 916 917 /* If PID/TID sampling is active, replace the default overflow 918 * handler to extract and resolve the PIDs from the basic-sampling 919 * data entries. 920 */ 921 if (event->attr.sample_type & PERF_SAMPLE_TID) 922 if (is_default_overflow_handler(event)) 923 event->overflow_handler = cpumsf_output_event_pid; 924 out: 925 return err; 926 } 927 928 static bool is_callchain_event(struct perf_event *event) 929 { 930 u64 sample_type = event->attr.sample_type; 931 932 return sample_type & (PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | 933 PERF_SAMPLE_STACK_USER); 934 } 935 936 static int cpumsf_pmu_event_init(struct perf_event *event) 937 { 938 int err; 939 940 /* No support for taken branch sampling */ 941 /* No support for callchain, stacks and registers */ 942 if (has_branch_stack(event) || is_callchain_event(event)) 943 return -EOPNOTSUPP; 944 945 switch (event->attr.type) { 946 case PERF_TYPE_RAW: 947 if ((event->attr.config != PERF_EVENT_CPUM_SF) && 948 (event->attr.config != PERF_EVENT_CPUM_SF_DIAG)) 949 return -ENOENT; 950 break; 951 case PERF_TYPE_HARDWARE: 952 /* Support sampling of CPU cycles in addition to the 953 * counter facility. However, the counter facility 954 * is more precise and, hence, restrict this PMU to 955 * sampling events only. 956 */ 957 if (event->attr.config != PERF_COUNT_HW_CPU_CYCLES) 958 return -ENOENT; 959 if (!is_sampling_event(event)) 960 return -ENOENT; 961 break; 962 default: 963 return -ENOENT; 964 } 965 966 /* Check online status of the CPU to which the event is pinned */ 967 if (event->cpu >= 0 && !cpu_online(event->cpu)) 968 return -ENODEV; 969 970 /* Force reset of idle/hv excludes regardless of what the 971 * user requested. 972 */ 973 if (event->attr.exclude_hv) 974 event->attr.exclude_hv = 0; 975 if (event->attr.exclude_idle) 976 event->attr.exclude_idle = 0; 977 978 err = __hw_perf_event_init(event); 979 if (unlikely(err)) 980 if (event->destroy) 981 event->destroy(event); 982 return err; 983 } 984 985 static void cpumsf_pmu_enable(struct pmu *pmu) 986 { 987 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 988 struct hw_perf_event *hwc; 989 int err; 990 991 if (cpuhw->flags & PMU_F_ENABLED) 992 return; 993 994 if (cpuhw->flags & PMU_F_ERR_MASK) 995 return; 996 997 /* Check whether to extent the sampling buffer. 998 * 999 * Two conditions trigger an increase of the sampling buffer for a 1000 * perf event: 1001 * 1. Postponed buffer allocations from the event initialization. 1002 * 2. Sampling overflows that contribute to pending allocations. 1003 * 1004 * Note that the extend_sampling_buffer() function disables the sampling 1005 * facility, but it can be fully re-enabled using sampling controls that 1006 * have been saved in cpumsf_pmu_disable(). 1007 */ 1008 if (cpuhw->event) { 1009 hwc = &cpuhw->event->hw; 1010 if (!(SAMPL_DIAG_MODE(hwc))) { 1011 /* 1012 * Account number of overflow-designated 1013 * buffer extents 1014 */ 1015 sfb_account_overflows(cpuhw, hwc); 1016 extend_sampling_buffer(&cpuhw->sfb, hwc); 1017 } 1018 /* Rate may be adjusted with ioctl() */ 1019 cpuhw->lsctl.interval = SAMPL_RATE(&cpuhw->event->hw); 1020 } 1021 1022 /* (Re)enable the PMU and sampling facility */ 1023 cpuhw->flags |= PMU_F_ENABLED; 1024 barrier(); 1025 1026 err = lsctl(&cpuhw->lsctl); 1027 if (err) { 1028 cpuhw->flags &= ~PMU_F_ENABLED; 1029 pr_err("Loading sampling controls failed: op %i err %i\n", 1030 1, err); 1031 return; 1032 } 1033 1034 /* Load current program parameter */ 1035 lpp(&S390_lowcore.lpp); 1036 1037 debug_sprintf_event(sfdbg, 6, "%s: es %i cs %i ed %i cd %i " 1038 "interval %#lx tear %#lx dear %#lx\n", __func__, 1039 cpuhw->lsctl.es, cpuhw->lsctl.cs, cpuhw->lsctl.ed, 1040 cpuhw->lsctl.cd, cpuhw->lsctl.interval, 1041 cpuhw->lsctl.tear, cpuhw->lsctl.dear); 1042 } 1043 1044 static void cpumsf_pmu_disable(struct pmu *pmu) 1045 { 1046 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1047 struct hws_lsctl_request_block inactive; 1048 struct hws_qsi_info_block si; 1049 int err; 1050 1051 if (!(cpuhw->flags & PMU_F_ENABLED)) 1052 return; 1053 1054 if (cpuhw->flags & PMU_F_ERR_MASK) 1055 return; 1056 1057 /* Switch off sampling activation control */ 1058 inactive = cpuhw->lsctl; 1059 inactive.cs = 0; 1060 inactive.cd = 0; 1061 1062 err = lsctl(&inactive); 1063 if (err) { 1064 pr_err("Loading sampling controls failed: op %i err %i\n", 1065 2, err); 1066 return; 1067 } 1068 1069 /* Save state of TEAR and DEAR register contents */ 1070 err = qsi(&si); 1071 if (!err) { 1072 /* TEAR/DEAR values are valid only if the sampling facility is 1073 * enabled. Note that cpumsf_pmu_disable() might be called even 1074 * for a disabled sampling facility because cpumsf_pmu_enable() 1075 * controls the enable/disable state. 1076 */ 1077 if (si.es) { 1078 cpuhw->lsctl.tear = si.tear; 1079 cpuhw->lsctl.dear = si.dear; 1080 } 1081 } else 1082 debug_sprintf_event(sfdbg, 3, "%s: qsi() failed with err %i\n", 1083 __func__, err); 1084 1085 cpuhw->flags &= ~PMU_F_ENABLED; 1086 } 1087 1088 /* perf_exclude_event() - Filter event 1089 * @event: The perf event 1090 * @regs: pt_regs structure 1091 * @sde_regs: Sample-data-entry (sde) regs structure 1092 * 1093 * Filter perf events according to their exclude specification. 1094 * 1095 * Return non-zero if the event shall be excluded. 1096 */ 1097 static int perf_exclude_event(struct perf_event *event, struct pt_regs *regs, 1098 struct perf_sf_sde_regs *sde_regs) 1099 { 1100 if (event->attr.exclude_user && user_mode(regs)) 1101 return 1; 1102 if (event->attr.exclude_kernel && !user_mode(regs)) 1103 return 1; 1104 if (event->attr.exclude_guest && sde_regs->in_guest) 1105 return 1; 1106 if (event->attr.exclude_host && !sde_regs->in_guest) 1107 return 1; 1108 return 0; 1109 } 1110 1111 /* perf_push_sample() - Push samples to perf 1112 * @event: The perf event 1113 * @sample: Hardware sample data 1114 * 1115 * Use the hardware sample data to create perf event sample. The sample 1116 * is the pushed to the event subsystem and the function checks for 1117 * possible event overflows. If an event overflow occurs, the PMU is 1118 * stopped. 1119 * 1120 * Return non-zero if an event overflow occurred. 1121 */ 1122 static int perf_push_sample(struct perf_event *event, 1123 struct hws_basic_entry *basic) 1124 { 1125 int overflow; 1126 struct pt_regs regs; 1127 struct perf_sf_sde_regs *sde_regs; 1128 struct perf_sample_data data; 1129 1130 /* Setup perf sample */ 1131 perf_sample_data_init(&data, 0, event->hw.last_period); 1132 1133 /* Setup pt_regs to look like an CPU-measurement external interrupt 1134 * using the Program Request Alert code. The regs.int_parm_long 1135 * field which is unused contains additional sample-data-entry related 1136 * indicators. 1137 */ 1138 memset(®s, 0, sizeof(regs)); 1139 regs.int_code = 0x1407; 1140 regs.int_parm = CPU_MF_INT_SF_PRA; 1141 sde_regs = (struct perf_sf_sde_regs *) ®s.int_parm_long; 1142 1143 psw_bits(regs.psw).ia = basic->ia; 1144 psw_bits(regs.psw).dat = basic->T; 1145 psw_bits(regs.psw).wait = basic->W; 1146 psw_bits(regs.psw).pstate = basic->P; 1147 psw_bits(regs.psw).as = basic->AS; 1148 1149 /* 1150 * Use the hardware provided configuration level to decide if the 1151 * sample belongs to a guest or host. If that is not available, 1152 * fall back to the following heuristics: 1153 * A non-zero guest program parameter always indicates a guest 1154 * sample. Some early samples or samples from guests without 1155 * lpp usage would be misaccounted to the host. We use the asn 1156 * value as an addon heuristic to detect most of these guest samples. 1157 * If the value differs from 0xffff (the host value), we assume to 1158 * be a KVM guest. 1159 */ 1160 switch (basic->CL) { 1161 case 1: /* logical partition */ 1162 sde_regs->in_guest = 0; 1163 break; 1164 case 2: /* virtual machine */ 1165 sde_regs->in_guest = 1; 1166 break; 1167 default: /* old machine, use heuristics */ 1168 if (basic->gpp || basic->prim_asn != 0xffff) 1169 sde_regs->in_guest = 1; 1170 break; 1171 } 1172 1173 /* 1174 * Store the PID value from the sample-data-entry to be 1175 * processed and resolved by cpumsf_output_event_pid(). 1176 */ 1177 data.tid_entry.pid = basic->hpp & LPP_PID_MASK; 1178 1179 overflow = 0; 1180 if (perf_exclude_event(event, ®s, sde_regs)) 1181 goto out; 1182 if (perf_event_overflow(event, &data, ®s)) { 1183 overflow = 1; 1184 event->pmu->stop(event, 0); 1185 } 1186 perf_event_update_userpage(event); 1187 out: 1188 return overflow; 1189 } 1190 1191 static void perf_event_count_update(struct perf_event *event, u64 count) 1192 { 1193 local64_add(count, &event->count); 1194 } 1195 1196 /* hw_collect_samples() - Walk through a sample-data-block and collect samples 1197 * @event: The perf event 1198 * @sdbt: Sample-data-block table 1199 * @overflow: Event overflow counter 1200 * 1201 * Walks through a sample-data-block and collects sampling data entries that are 1202 * then pushed to the perf event subsystem. Depending on the sampling function, 1203 * there can be either basic-sampling or combined-sampling data entries. A 1204 * combined-sampling data entry consists of a basic- and a diagnostic-sampling 1205 * data entry. The sampling function is determined by the flags in the perf 1206 * event hardware structure. The function always works with a combined-sampling 1207 * data entry but ignores the the diagnostic portion if it is not available. 1208 * 1209 * Note that the implementation focuses on basic-sampling data entries and, if 1210 * such an entry is not valid, the entire combined-sampling data entry is 1211 * ignored. 1212 * 1213 * The overflow variables counts the number of samples that has been discarded 1214 * due to a perf event overflow. 1215 */ 1216 static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt, 1217 unsigned long long *overflow) 1218 { 1219 struct hws_trailer_entry *te; 1220 struct hws_basic_entry *sample; 1221 1222 te = trailer_entry_ptr((unsigned long)sdbt); 1223 sample = (struct hws_basic_entry *)sdbt; 1224 while ((unsigned long *) sample < (unsigned long *) te) { 1225 /* Check for an empty sample */ 1226 if (!sample->def || sample->LS) 1227 break; 1228 1229 /* Update perf event period */ 1230 perf_event_count_update(event, SAMPL_RATE(&event->hw)); 1231 1232 /* Check whether sample is valid */ 1233 if (sample->def == 0x0001) { 1234 /* If an event overflow occurred, the PMU is stopped to 1235 * throttle event delivery. Remaining sample data is 1236 * discarded. 1237 */ 1238 if (!*overflow) { 1239 /* Check whether sample is consistent */ 1240 if (sample->I == 0 && sample->W == 0) { 1241 /* Deliver sample data to perf */ 1242 *overflow = perf_push_sample(event, 1243 sample); 1244 } 1245 } else 1246 /* Count discarded samples */ 1247 *overflow += 1; 1248 } else { 1249 debug_sprintf_event(sfdbg, 4, 1250 "%s: Found unknown" 1251 " sampling data entry: te->f %i" 1252 " basic.def %#4x (%p)\n", __func__, 1253 te->header.f, sample->def, sample); 1254 /* Sample slot is not yet written or other record. 1255 * 1256 * This condition can occur if the buffer was reused 1257 * from a combined basic- and diagnostic-sampling. 1258 * If only basic-sampling is then active, entries are 1259 * written into the larger diagnostic entries. 1260 * This is typically the case for sample-data-blocks 1261 * that are not full. Stop processing if the first 1262 * invalid format was detected. 1263 */ 1264 if (!te->header.f) 1265 break; 1266 } 1267 1268 /* Reset sample slot and advance to next sample */ 1269 sample->def = 0; 1270 sample++; 1271 } 1272 } 1273 1274 static inline __uint128_t __cdsg(__uint128_t *ptr, __uint128_t old, __uint128_t new) 1275 { 1276 asm volatile( 1277 " cdsg %[old],%[new],%[ptr]\n" 1278 : [old] "+d" (old), [ptr] "+QS" (*ptr) 1279 : [new] "d" (new) 1280 : "memory", "cc"); 1281 return old; 1282 } 1283 1284 /* hw_perf_event_update() - Process sampling buffer 1285 * @event: The perf event 1286 * @flush_all: Flag to also flush partially filled sample-data-blocks 1287 * 1288 * Processes the sampling buffer and create perf event samples. 1289 * The sampling buffer position are retrieved and saved in the TEAR_REG 1290 * register of the specified perf event. 1291 * 1292 * Only full sample-data-blocks are processed. Specify the flush_all flag 1293 * to also walk through partially filled sample-data-blocks. 1294 */ 1295 static void hw_perf_event_update(struct perf_event *event, int flush_all) 1296 { 1297 unsigned long long event_overflow, sampl_overflow, num_sdb; 1298 union hws_trailer_header old, prev, new; 1299 struct hw_perf_event *hwc = &event->hw; 1300 struct hws_trailer_entry *te; 1301 unsigned long *sdbt, sdb; 1302 int done; 1303 1304 /* 1305 * AUX buffer is used when in diagnostic sampling mode. 1306 * No perf events/samples are created. 1307 */ 1308 if (SAMPL_DIAG_MODE(&event->hw)) 1309 return; 1310 1311 sdbt = (unsigned long *) TEAR_REG(hwc); 1312 done = event_overflow = sampl_overflow = num_sdb = 0; 1313 while (!done) { 1314 /* Get the trailer entry of the sample-data-block */ 1315 sdb = (unsigned long)phys_to_virt(*sdbt); 1316 te = trailer_entry_ptr(sdb); 1317 1318 /* Leave loop if no more work to do (block full indicator) */ 1319 if (!te->header.f) { 1320 done = 1; 1321 if (!flush_all) 1322 break; 1323 } 1324 1325 /* Check the sample overflow count */ 1326 if (te->header.overflow) 1327 /* Account sample overflows and, if a particular limit 1328 * is reached, extend the sampling buffer. 1329 * For details, see sfb_account_overflows(). 1330 */ 1331 sampl_overflow += te->header.overflow; 1332 1333 /* Timestamps are valid for full sample-data-blocks only */ 1334 debug_sprintf_event(sfdbg, 6, "%s: sdbt %#lx/%#lx " 1335 "overflow %llu timestamp %#llx\n", 1336 __func__, sdb, (unsigned long)sdbt, 1337 te->header.overflow, 1338 (te->header.f) ? trailer_timestamp(te) : 0ULL); 1339 1340 /* Collect all samples from a single sample-data-block and 1341 * flag if an (perf) event overflow happened. If so, the PMU 1342 * is stopped and remaining samples will be discarded. 1343 */ 1344 hw_collect_samples(event, (unsigned long *)sdb, &event_overflow); 1345 num_sdb++; 1346 1347 /* Reset trailer (using compare-double-and-swap) */ 1348 prev.val = READ_ONCE_ALIGNED_128(te->header.val); 1349 do { 1350 old.val = prev.val; 1351 new.val = prev.val; 1352 new.f = 0; 1353 new.a = 1; 1354 new.overflow = 0; 1355 prev.val = __cdsg(&te->header.val, old.val, new.val); 1356 } while (prev.val != old.val); 1357 1358 /* Advance to next sample-data-block */ 1359 sdbt++; 1360 if (is_link_entry(sdbt)) 1361 sdbt = get_next_sdbt(sdbt); 1362 1363 /* Update event hardware registers */ 1364 TEAR_REG(hwc) = (unsigned long) sdbt; 1365 1366 /* Stop processing sample-data if all samples of the current 1367 * sample-data-block were flushed even if it was not full. 1368 */ 1369 if (flush_all && done) 1370 break; 1371 } 1372 1373 /* Account sample overflows in the event hardware structure */ 1374 if (sampl_overflow) 1375 OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) + 1376 sampl_overflow, 1 + num_sdb); 1377 1378 /* Perf_event_overflow() and perf_event_account_interrupt() limit 1379 * the interrupt rate to an upper limit. Roughly 1000 samples per 1380 * task tick. 1381 * Hitting this limit results in a large number 1382 * of throttled REF_REPORT_THROTTLE entries and the samples 1383 * are dropped. 1384 * Slightly increase the interval to avoid hitting this limit. 1385 */ 1386 if (event_overflow) { 1387 SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10); 1388 debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n", 1389 __func__, 1390 DIV_ROUND_UP(SAMPL_RATE(hwc), 10)); 1391 } 1392 1393 if (sampl_overflow || event_overflow) 1394 debug_sprintf_event(sfdbg, 4, "%s: " 1395 "overflows: sample %llu event %llu" 1396 " total %llu num_sdb %llu\n", 1397 __func__, sampl_overflow, event_overflow, 1398 OVERFLOW_REG(hwc), num_sdb); 1399 } 1400 1401 static inline unsigned long aux_sdb_index(struct aux_buffer *aux, 1402 unsigned long i) 1403 { 1404 return i % aux->sfb.num_sdb; 1405 } 1406 1407 static inline unsigned long aux_sdb_num(unsigned long start, unsigned long end) 1408 { 1409 return end >= start ? end - start + 1 : 0; 1410 } 1411 1412 static inline unsigned long aux_sdb_num_alert(struct aux_buffer *aux) 1413 { 1414 return aux_sdb_num(aux->head, aux->alert_mark); 1415 } 1416 1417 static inline unsigned long aux_sdb_num_empty(struct aux_buffer *aux) 1418 { 1419 return aux_sdb_num(aux->head, aux->empty_mark); 1420 } 1421 1422 /* 1423 * Get trailer entry by index of SDB. 1424 */ 1425 static struct hws_trailer_entry *aux_sdb_trailer(struct aux_buffer *aux, 1426 unsigned long index) 1427 { 1428 unsigned long sdb; 1429 1430 index = aux_sdb_index(aux, index); 1431 sdb = aux->sdb_index[index]; 1432 return trailer_entry_ptr(sdb); 1433 } 1434 1435 /* 1436 * Finish sampling on the cpu. Called by cpumsf_pmu_del() with pmu 1437 * disabled. Collect the full SDBs in AUX buffer which have not reached 1438 * the point of alert indicator. And ignore the SDBs which are not 1439 * full. 1440 * 1441 * 1. Scan SDBs to see how much data is there and consume them. 1442 * 2. Remove alert indicator in the buffer. 1443 */ 1444 static void aux_output_end(struct perf_output_handle *handle) 1445 { 1446 unsigned long i, range_scan, idx; 1447 struct aux_buffer *aux; 1448 struct hws_trailer_entry *te; 1449 1450 aux = perf_get_aux(handle); 1451 if (!aux) 1452 return; 1453 1454 range_scan = aux_sdb_num_alert(aux); 1455 for (i = 0, idx = aux->head; i < range_scan; i++, idx++) { 1456 te = aux_sdb_trailer(aux, idx); 1457 if (!te->header.f) 1458 break; 1459 } 1460 /* i is num of SDBs which are full */ 1461 perf_aux_output_end(handle, i << PAGE_SHIFT); 1462 1463 /* Remove alert indicators in the buffer */ 1464 te = aux_sdb_trailer(aux, aux->alert_mark); 1465 te->header.a = 0; 1466 1467 debug_sprintf_event(sfdbg, 6, "%s: SDBs %ld range %ld head %ld\n", 1468 __func__, i, range_scan, aux->head); 1469 } 1470 1471 /* 1472 * Start sampling on the CPU. Called by cpumsf_pmu_add() when an event 1473 * is first added to the CPU or rescheduled again to the CPU. It is called 1474 * with pmu disabled. 1475 * 1476 * 1. Reset the trailer of SDBs to get ready for new data. 1477 * 2. Tell the hardware where to put the data by reset the SDBs buffer 1478 * head(tear/dear). 1479 */ 1480 static int aux_output_begin(struct perf_output_handle *handle, 1481 struct aux_buffer *aux, 1482 struct cpu_hw_sf *cpuhw) 1483 { 1484 unsigned long range, i, range_scan, idx, head, base, offset; 1485 struct hws_trailer_entry *te; 1486 1487 if (WARN_ON_ONCE(handle->head & ~PAGE_MASK)) 1488 return -EINVAL; 1489 1490 aux->head = handle->head >> PAGE_SHIFT; 1491 range = (handle->size + 1) >> PAGE_SHIFT; 1492 if (range <= 1) 1493 return -ENOMEM; 1494 1495 /* 1496 * SDBs between aux->head and aux->empty_mark are already ready 1497 * for new data. range_scan is num of SDBs not within them. 1498 */ 1499 debug_sprintf_event(sfdbg, 6, 1500 "%s: range %ld head %ld alert %ld empty %ld\n", 1501 __func__, range, aux->head, aux->alert_mark, 1502 aux->empty_mark); 1503 if (range > aux_sdb_num_empty(aux)) { 1504 range_scan = range - aux_sdb_num_empty(aux); 1505 idx = aux->empty_mark + 1; 1506 for (i = 0; i < range_scan; i++, idx++) { 1507 te = aux_sdb_trailer(aux, idx); 1508 te->header.f = 0; 1509 te->header.a = 0; 1510 te->header.overflow = 0; 1511 } 1512 /* Save the position of empty SDBs */ 1513 aux->empty_mark = aux->head + range - 1; 1514 } 1515 1516 /* Set alert indicator */ 1517 aux->alert_mark = aux->head + range/2 - 1; 1518 te = aux_sdb_trailer(aux, aux->alert_mark); 1519 te->header.a = 1; 1520 1521 /* Reset hardware buffer head */ 1522 head = aux_sdb_index(aux, aux->head); 1523 base = aux->sdbt_index[head / CPUM_SF_SDB_PER_TABLE]; 1524 offset = head % CPUM_SF_SDB_PER_TABLE; 1525 cpuhw->lsctl.tear = virt_to_phys((void *)base) + offset * sizeof(unsigned long); 1526 cpuhw->lsctl.dear = virt_to_phys((void *)aux->sdb_index[head]); 1527 1528 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld empty %ld " 1529 "index %ld tear %#lx dear %#lx\n", __func__, 1530 aux->head, aux->alert_mark, aux->empty_mark, 1531 head / CPUM_SF_SDB_PER_TABLE, 1532 cpuhw->lsctl.tear, cpuhw->lsctl.dear); 1533 1534 return 0; 1535 } 1536 1537 /* 1538 * Set alert indicator on SDB at index @alert_index while sampler is running. 1539 * 1540 * Return true if successfully. 1541 * Return false if full indicator is already set by hardware sampler. 1542 */ 1543 static bool aux_set_alert(struct aux_buffer *aux, unsigned long alert_index, 1544 unsigned long long *overflow) 1545 { 1546 union hws_trailer_header old, prev, new; 1547 struct hws_trailer_entry *te; 1548 1549 te = aux_sdb_trailer(aux, alert_index); 1550 prev.val = READ_ONCE_ALIGNED_128(te->header.val); 1551 do { 1552 old.val = prev.val; 1553 new.val = prev.val; 1554 *overflow = old.overflow; 1555 if (old.f) { 1556 /* 1557 * SDB is already set by hardware. 1558 * Abort and try to set somewhere 1559 * behind. 1560 */ 1561 return false; 1562 } 1563 new.a = 1; 1564 new.overflow = 0; 1565 prev.val = __cdsg(&te->header.val, old.val, new.val); 1566 } while (prev.val != old.val); 1567 return true; 1568 } 1569 1570 /* 1571 * aux_reset_buffer() - Scan and setup SDBs for new samples 1572 * @aux: The AUX buffer to set 1573 * @range: The range of SDBs to scan started from aux->head 1574 * @overflow: Set to overflow count 1575 * 1576 * Set alert indicator on the SDB at index of aux->alert_mark. If this SDB is 1577 * marked as empty, check if it is already set full by the hardware sampler. 1578 * If yes, that means new data is already there before we can set an alert 1579 * indicator. Caller should try to set alert indicator to some position behind. 1580 * 1581 * Scan the SDBs in AUX buffer from behind aux->empty_mark. They are used 1582 * previously and have already been consumed by user space. Reset these SDBs 1583 * (clear full indicator and alert indicator) for new data. 1584 * If aux->alert_mark fall in this area, just set it. Overflow count is 1585 * recorded while scanning. 1586 * 1587 * SDBs between aux->head and aux->empty_mark are already reset at last time. 1588 * and ready for new samples. So scanning on this area could be skipped. 1589 * 1590 * Return true if alert indicator is set successfully and false if not. 1591 */ 1592 static bool aux_reset_buffer(struct aux_buffer *aux, unsigned long range, 1593 unsigned long long *overflow) 1594 { 1595 unsigned long i, range_scan, idx, idx_old; 1596 union hws_trailer_header old, prev, new; 1597 unsigned long long orig_overflow; 1598 struct hws_trailer_entry *te; 1599 1600 debug_sprintf_event(sfdbg, 6, "%s: range %ld head %ld alert %ld " 1601 "empty %ld\n", __func__, range, aux->head, 1602 aux->alert_mark, aux->empty_mark); 1603 if (range <= aux_sdb_num_empty(aux)) 1604 /* 1605 * No need to scan. All SDBs in range are marked as empty. 1606 * Just set alert indicator. Should check race with hardware 1607 * sampler. 1608 */ 1609 return aux_set_alert(aux, aux->alert_mark, overflow); 1610 1611 if (aux->alert_mark <= aux->empty_mark) 1612 /* 1613 * Set alert indicator on empty SDB. Should check race 1614 * with hardware sampler. 1615 */ 1616 if (!aux_set_alert(aux, aux->alert_mark, overflow)) 1617 return false; 1618 1619 /* 1620 * Scan the SDBs to clear full and alert indicator used previously. 1621 * Start scanning from one SDB behind empty_mark. If the new alert 1622 * indicator fall into this range, set it. 1623 */ 1624 range_scan = range - aux_sdb_num_empty(aux); 1625 idx_old = idx = aux->empty_mark + 1; 1626 for (i = 0; i < range_scan; i++, idx++) { 1627 te = aux_sdb_trailer(aux, idx); 1628 prev.val = READ_ONCE_ALIGNED_128(te->header.val); 1629 do { 1630 old.val = prev.val; 1631 new.val = prev.val; 1632 orig_overflow = old.overflow; 1633 new.f = 0; 1634 new.overflow = 0; 1635 if (idx == aux->alert_mark) 1636 new.a = 1; 1637 else 1638 new.a = 0; 1639 prev.val = __cdsg(&te->header.val, old.val, new.val); 1640 } while (prev.val != old.val); 1641 *overflow += orig_overflow; 1642 } 1643 1644 /* Update empty_mark to new position */ 1645 aux->empty_mark = aux->head + range - 1; 1646 1647 debug_sprintf_event(sfdbg, 6, "%s: range_scan %ld idx %ld..%ld " 1648 "empty %ld\n", __func__, range_scan, idx_old, 1649 idx - 1, aux->empty_mark); 1650 return true; 1651 } 1652 1653 /* 1654 * Measurement alert handler for diagnostic mode sampling. 1655 */ 1656 static void hw_collect_aux(struct cpu_hw_sf *cpuhw) 1657 { 1658 struct aux_buffer *aux; 1659 int done = 0; 1660 unsigned long range = 0, size; 1661 unsigned long long overflow = 0; 1662 struct perf_output_handle *handle = &cpuhw->handle; 1663 unsigned long num_sdb; 1664 1665 aux = perf_get_aux(handle); 1666 if (WARN_ON_ONCE(!aux)) 1667 return; 1668 1669 /* Inform user space new data arrived */ 1670 size = aux_sdb_num_alert(aux) << PAGE_SHIFT; 1671 debug_sprintf_event(sfdbg, 6, "%s: #alert %ld\n", __func__, 1672 size >> PAGE_SHIFT); 1673 perf_aux_output_end(handle, size); 1674 1675 num_sdb = aux->sfb.num_sdb; 1676 while (!done) { 1677 /* Get an output handle */ 1678 aux = perf_aux_output_begin(handle, cpuhw->event); 1679 if (handle->size == 0) { 1680 pr_err("The AUX buffer with %lu pages for the " 1681 "diagnostic-sampling mode is full\n", 1682 num_sdb); 1683 debug_sprintf_event(sfdbg, 1, 1684 "%s: AUX buffer used up\n", 1685 __func__); 1686 break; 1687 } 1688 if (WARN_ON_ONCE(!aux)) 1689 return; 1690 1691 /* Update head and alert_mark to new position */ 1692 aux->head = handle->head >> PAGE_SHIFT; 1693 range = (handle->size + 1) >> PAGE_SHIFT; 1694 if (range == 1) 1695 aux->alert_mark = aux->head; 1696 else 1697 aux->alert_mark = aux->head + range/2 - 1; 1698 1699 if (aux_reset_buffer(aux, range, &overflow)) { 1700 if (!overflow) { 1701 done = 1; 1702 break; 1703 } 1704 size = range << PAGE_SHIFT; 1705 perf_aux_output_end(&cpuhw->handle, size); 1706 pr_err("Sample data caused the AUX buffer with %lu " 1707 "pages to overflow\n", aux->sfb.num_sdb); 1708 debug_sprintf_event(sfdbg, 1, "%s: head %ld range %ld " 1709 "overflow %lld\n", __func__, 1710 aux->head, range, overflow); 1711 } else { 1712 size = aux_sdb_num_alert(aux) << PAGE_SHIFT; 1713 perf_aux_output_end(&cpuhw->handle, size); 1714 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " 1715 "already full, try another\n", 1716 __func__, 1717 aux->head, aux->alert_mark); 1718 } 1719 } 1720 1721 if (done) 1722 debug_sprintf_event(sfdbg, 6, "%s: head %ld alert %ld " 1723 "empty %ld\n", __func__, aux->head, 1724 aux->alert_mark, aux->empty_mark); 1725 } 1726 1727 /* 1728 * Callback when freeing AUX buffers. 1729 */ 1730 static void aux_buffer_free(void *data) 1731 { 1732 struct aux_buffer *aux = data; 1733 unsigned long i, num_sdbt; 1734 1735 if (!aux) 1736 return; 1737 1738 /* Free SDBT. SDB is freed by the caller */ 1739 num_sdbt = aux->sfb.num_sdbt; 1740 for (i = 0; i < num_sdbt; i++) 1741 free_page(aux->sdbt_index[i]); 1742 1743 kfree(aux->sdbt_index); 1744 kfree(aux->sdb_index); 1745 kfree(aux); 1746 1747 debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu\n", __func__, num_sdbt); 1748 } 1749 1750 static void aux_sdb_init(unsigned long sdb) 1751 { 1752 struct hws_trailer_entry *te; 1753 1754 te = trailer_entry_ptr(sdb); 1755 1756 /* Save clock base */ 1757 te->clock_base = 1; 1758 te->progusage2 = tod_clock_base.tod; 1759 } 1760 1761 /* 1762 * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling 1763 * @event: Event the buffer is setup for, event->cpu == -1 means current 1764 * @pages: Array of pointers to buffer pages passed from perf core 1765 * @nr_pages: Total pages 1766 * @snapshot: Flag for snapshot mode 1767 * 1768 * This is the callback when setup an event using AUX buffer. Perf tool can 1769 * trigger this by an additional mmap() call on the event. Unlike the buffer 1770 * for basic samples, AUX buffer belongs to the event. It is scheduled with 1771 * the task among online cpus when it is a per-thread event. 1772 * 1773 * Return the private AUX buffer structure if success or NULL if fails. 1774 */ 1775 static void *aux_buffer_setup(struct perf_event *event, void **pages, 1776 int nr_pages, bool snapshot) 1777 { 1778 struct sf_buffer *sfb; 1779 struct aux_buffer *aux; 1780 unsigned long *new, *tail; 1781 int i, n_sdbt; 1782 1783 if (!nr_pages || !pages) 1784 return NULL; 1785 1786 if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1787 pr_err("AUX buffer size (%i pages) is larger than the " 1788 "maximum sampling buffer limit\n", 1789 nr_pages); 1790 return NULL; 1791 } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) { 1792 pr_err("AUX buffer size (%i pages) is less than the " 1793 "minimum sampling buffer limit\n", 1794 nr_pages); 1795 return NULL; 1796 } 1797 1798 /* Allocate aux_buffer struct for the event */ 1799 aux = kzalloc(sizeof(struct aux_buffer), GFP_KERNEL); 1800 if (!aux) 1801 goto no_aux; 1802 sfb = &aux->sfb; 1803 1804 /* Allocate sdbt_index for fast reference */ 1805 n_sdbt = DIV_ROUND_UP(nr_pages, CPUM_SF_SDB_PER_TABLE); 1806 aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL); 1807 if (!aux->sdbt_index) 1808 goto no_sdbt_index; 1809 1810 /* Allocate sdb_index for fast reference */ 1811 aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL); 1812 if (!aux->sdb_index) 1813 goto no_sdb_index; 1814 1815 /* Allocate the first SDBT */ 1816 sfb->num_sdbt = 0; 1817 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1818 if (!sfb->sdbt) 1819 goto no_sdbt; 1820 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt; 1821 tail = sfb->tail = sfb->sdbt; 1822 1823 /* 1824 * Link the provided pages of AUX buffer to SDBT. 1825 * Allocate SDBT if needed. 1826 */ 1827 for (i = 0; i < nr_pages; i++, tail++) { 1828 if (require_table_link(tail)) { 1829 new = (unsigned long *) get_zeroed_page(GFP_KERNEL); 1830 if (!new) 1831 goto no_sdbt; 1832 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new; 1833 /* Link current page to tail of chain */ 1834 *tail = virt_to_phys(new) + 1; 1835 tail = new; 1836 } 1837 /* Tail is the entry in a SDBT */ 1838 *tail = virt_to_phys(pages[i]); 1839 aux->sdb_index[i] = (unsigned long)pages[i]; 1840 aux_sdb_init((unsigned long)pages[i]); 1841 } 1842 sfb->num_sdb = nr_pages; 1843 1844 /* Link the last entry in the SDBT to the first SDBT */ 1845 *tail = virt_to_phys(sfb->sdbt) + 1; 1846 sfb->tail = tail; 1847 1848 /* 1849 * Initial all SDBs are zeroed. Mark it as empty. 1850 * So there is no need to clear the full indicator 1851 * when this event is first added. 1852 */ 1853 aux->empty_mark = sfb->num_sdb - 1; 1854 1855 debug_sprintf_event(sfdbg, 4, "%s: SDBTs %lu SDBs %lu\n", __func__, 1856 sfb->num_sdbt, sfb->num_sdb); 1857 1858 return aux; 1859 1860 no_sdbt: 1861 /* SDBs (AUX buffer pages) are freed by caller */ 1862 for (i = 0; i < sfb->num_sdbt; i++) 1863 free_page(aux->sdbt_index[i]); 1864 kfree(aux->sdb_index); 1865 no_sdb_index: 1866 kfree(aux->sdbt_index); 1867 no_sdbt_index: 1868 kfree(aux); 1869 no_aux: 1870 return NULL; 1871 } 1872 1873 static void cpumsf_pmu_read(struct perf_event *event) 1874 { 1875 /* Nothing to do ... updates are interrupt-driven */ 1876 } 1877 1878 /* Check if the new sampling period/freqeuncy is appropriate. 1879 * 1880 * Return non-zero on error and zero on passed checks. 1881 */ 1882 static int cpumsf_pmu_check_period(struct perf_event *event, u64 value) 1883 { 1884 struct hws_qsi_info_block si; 1885 unsigned long rate; 1886 bool do_freq; 1887 1888 memset(&si, 0, sizeof(si)); 1889 if (event->cpu == -1) { 1890 if (qsi(&si)) 1891 return -ENODEV; 1892 } else { 1893 /* Event is pinned to a particular CPU, retrieve the per-CPU 1894 * sampling structure for accessing the CPU-specific QSI. 1895 */ 1896 struct cpu_hw_sf *cpuhw = &per_cpu(cpu_hw_sf, event->cpu); 1897 1898 si = cpuhw->qsi; 1899 } 1900 1901 do_freq = !!SAMPLE_FREQ_MODE(&event->hw); 1902 rate = getrate(do_freq, value, &si); 1903 if (!rate) 1904 return -EINVAL; 1905 1906 event->attr.sample_period = rate; 1907 SAMPL_RATE(&event->hw) = rate; 1908 hw_init_period(&event->hw, SAMPL_RATE(&event->hw)); 1909 debug_sprintf_event(sfdbg, 4, "%s:" 1910 " cpu %d value %#llx period %#llx freq %d\n", 1911 __func__, event->cpu, value, 1912 event->attr.sample_period, do_freq); 1913 return 0; 1914 } 1915 1916 /* Activate sampling control. 1917 * Next call of pmu_enable() starts sampling. 1918 */ 1919 static void cpumsf_pmu_start(struct perf_event *event, int flags) 1920 { 1921 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1922 1923 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) 1924 return; 1925 1926 if (flags & PERF_EF_RELOAD) 1927 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); 1928 1929 perf_pmu_disable(event->pmu); 1930 event->hw.state = 0; 1931 cpuhw->lsctl.cs = 1; 1932 if (SAMPL_DIAG_MODE(&event->hw)) 1933 cpuhw->lsctl.cd = 1; 1934 perf_pmu_enable(event->pmu); 1935 } 1936 1937 /* Deactivate sampling control. 1938 * Next call of pmu_enable() stops sampling. 1939 */ 1940 static void cpumsf_pmu_stop(struct perf_event *event, int flags) 1941 { 1942 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1943 1944 if (event->hw.state & PERF_HES_STOPPED) 1945 return; 1946 1947 perf_pmu_disable(event->pmu); 1948 cpuhw->lsctl.cs = 0; 1949 cpuhw->lsctl.cd = 0; 1950 event->hw.state |= PERF_HES_STOPPED; 1951 1952 if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) { 1953 hw_perf_event_update(event, 1); 1954 event->hw.state |= PERF_HES_UPTODATE; 1955 } 1956 perf_pmu_enable(event->pmu); 1957 } 1958 1959 static int cpumsf_pmu_add(struct perf_event *event, int flags) 1960 { 1961 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 1962 struct aux_buffer *aux; 1963 int err; 1964 1965 if (cpuhw->flags & PMU_F_IN_USE) 1966 return -EAGAIN; 1967 1968 if (!SAMPL_DIAG_MODE(&event->hw) && !cpuhw->sfb.sdbt) 1969 return -EINVAL; 1970 1971 err = 0; 1972 perf_pmu_disable(event->pmu); 1973 1974 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED; 1975 1976 /* Set up sampling controls. Always program the sampling register 1977 * using the SDB-table start. Reset TEAR_REG event hardware register 1978 * that is used by hw_perf_event_update() to store the sampling buffer 1979 * position after samples have been flushed. 1980 */ 1981 cpuhw->lsctl.s = 0; 1982 cpuhw->lsctl.h = 1; 1983 cpuhw->lsctl.interval = SAMPL_RATE(&event->hw); 1984 if (!SAMPL_DIAG_MODE(&event->hw)) { 1985 cpuhw->lsctl.tear = virt_to_phys(cpuhw->sfb.sdbt); 1986 cpuhw->lsctl.dear = *(unsigned long *) cpuhw->sfb.sdbt; 1987 TEAR_REG(&event->hw) = (unsigned long) cpuhw->sfb.sdbt; 1988 } 1989 1990 /* Ensure sampling functions are in the disabled state. If disabled, 1991 * switch on sampling enable control. */ 1992 if (WARN_ON_ONCE(cpuhw->lsctl.es == 1 || cpuhw->lsctl.ed == 1)) { 1993 err = -EAGAIN; 1994 goto out; 1995 } 1996 if (SAMPL_DIAG_MODE(&event->hw)) { 1997 aux = perf_aux_output_begin(&cpuhw->handle, event); 1998 if (!aux) { 1999 err = -EINVAL; 2000 goto out; 2001 } 2002 err = aux_output_begin(&cpuhw->handle, aux, cpuhw); 2003 if (err) 2004 goto out; 2005 cpuhw->lsctl.ed = 1; 2006 } 2007 cpuhw->lsctl.es = 1; 2008 2009 /* Set in_use flag and store event */ 2010 cpuhw->event = event; 2011 cpuhw->flags |= PMU_F_IN_USE; 2012 2013 if (flags & PERF_EF_START) 2014 cpumsf_pmu_start(event, PERF_EF_RELOAD); 2015 out: 2016 perf_event_update_userpage(event); 2017 perf_pmu_enable(event->pmu); 2018 return err; 2019 } 2020 2021 static void cpumsf_pmu_del(struct perf_event *event, int flags) 2022 { 2023 struct cpu_hw_sf *cpuhw = this_cpu_ptr(&cpu_hw_sf); 2024 2025 perf_pmu_disable(event->pmu); 2026 cpumsf_pmu_stop(event, PERF_EF_UPDATE); 2027 2028 cpuhw->lsctl.es = 0; 2029 cpuhw->lsctl.ed = 0; 2030 cpuhw->flags &= ~PMU_F_IN_USE; 2031 cpuhw->event = NULL; 2032 2033 if (SAMPL_DIAG_MODE(&event->hw)) 2034 aux_output_end(&cpuhw->handle); 2035 perf_event_update_userpage(event); 2036 perf_pmu_enable(event->pmu); 2037 } 2038 2039 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC, PERF_EVENT_CPUM_SF); 2040 CPUMF_EVENT_ATTR(SF, SF_CYCLES_BASIC_DIAG, PERF_EVENT_CPUM_SF_DIAG); 2041 2042 /* Attribute list for CPU_SF. 2043 * 2044 * The availablitiy depends on the CPU_MF sampling facility authorization 2045 * for basic + diagnositic samples. This is determined at initialization 2046 * time by the sampling facility device driver. 2047 * If the authorization for basic samples is turned off, it should be 2048 * also turned off for diagnostic sampling. 2049 * 2050 * During initialization of the device driver, check the authorization 2051 * level for diagnostic sampling and installs the attribute 2052 * file for diagnostic sampling if necessary. 2053 * 2054 * For now install a placeholder to reference all possible attributes: 2055 * SF_CYCLES_BASIC and SF_CYCLES_BASIC_DIAG. 2056 * Add another entry for the final NULL pointer. 2057 */ 2058 enum { 2059 SF_CYCLES_BASIC_ATTR_IDX = 0, 2060 SF_CYCLES_BASIC_DIAG_ATTR_IDX, 2061 SF_CYCLES_ATTR_MAX 2062 }; 2063 2064 static struct attribute *cpumsf_pmu_events_attr[SF_CYCLES_ATTR_MAX + 1] = { 2065 [SF_CYCLES_BASIC_ATTR_IDX] = CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC) 2066 }; 2067 2068 PMU_FORMAT_ATTR(event, "config:0-63"); 2069 2070 static struct attribute *cpumsf_pmu_format_attr[] = { 2071 &format_attr_event.attr, 2072 NULL, 2073 }; 2074 2075 static struct attribute_group cpumsf_pmu_events_group = { 2076 .name = "events", 2077 .attrs = cpumsf_pmu_events_attr, 2078 }; 2079 2080 static struct attribute_group cpumsf_pmu_format_group = { 2081 .name = "format", 2082 .attrs = cpumsf_pmu_format_attr, 2083 }; 2084 2085 static const struct attribute_group *cpumsf_pmu_attr_groups[] = { 2086 &cpumsf_pmu_events_group, 2087 &cpumsf_pmu_format_group, 2088 NULL, 2089 }; 2090 2091 static struct pmu cpumf_sampling = { 2092 .pmu_enable = cpumsf_pmu_enable, 2093 .pmu_disable = cpumsf_pmu_disable, 2094 2095 .event_init = cpumsf_pmu_event_init, 2096 .add = cpumsf_pmu_add, 2097 .del = cpumsf_pmu_del, 2098 2099 .start = cpumsf_pmu_start, 2100 .stop = cpumsf_pmu_stop, 2101 .read = cpumsf_pmu_read, 2102 2103 .attr_groups = cpumsf_pmu_attr_groups, 2104 2105 .setup_aux = aux_buffer_setup, 2106 .free_aux = aux_buffer_free, 2107 2108 .check_period = cpumsf_pmu_check_period, 2109 }; 2110 2111 static void cpumf_measurement_alert(struct ext_code ext_code, 2112 unsigned int alert, unsigned long unused) 2113 { 2114 struct cpu_hw_sf *cpuhw; 2115 2116 if (!(alert & CPU_MF_INT_SF_MASK)) 2117 return; 2118 inc_irq_stat(IRQEXT_CMS); 2119 cpuhw = this_cpu_ptr(&cpu_hw_sf); 2120 2121 /* Measurement alerts are shared and might happen when the PMU 2122 * is not reserved. Ignore these alerts in this case. */ 2123 if (!(cpuhw->flags & PMU_F_RESERVED)) 2124 return; 2125 2126 /* The processing below must take care of multiple alert events that 2127 * might be indicated concurrently. */ 2128 2129 /* Program alert request */ 2130 if (alert & CPU_MF_INT_SF_PRA) { 2131 if (cpuhw->flags & PMU_F_IN_USE) 2132 if (SAMPL_DIAG_MODE(&cpuhw->event->hw)) 2133 hw_collect_aux(cpuhw); 2134 else 2135 hw_perf_event_update(cpuhw->event, 0); 2136 else 2137 WARN_ON_ONCE(!(cpuhw->flags & PMU_F_IN_USE)); 2138 } 2139 2140 /* Report measurement alerts only for non-PRA codes */ 2141 if (alert != CPU_MF_INT_SF_PRA) 2142 debug_sprintf_event(sfdbg, 6, "%s: alert %#x\n", __func__, 2143 alert); 2144 2145 /* Sampling authorization change request */ 2146 if (alert & CPU_MF_INT_SF_SACA) 2147 qsi(&cpuhw->qsi); 2148 2149 /* Loss of sample data due to high-priority machine activities */ 2150 if (alert & CPU_MF_INT_SF_LSDA) { 2151 pr_err("Sample data was lost\n"); 2152 cpuhw->flags |= PMU_F_ERR_LSDA; 2153 sf_disable(); 2154 } 2155 2156 /* Invalid sampling buffer entry */ 2157 if (alert & (CPU_MF_INT_SF_IAE|CPU_MF_INT_SF_ISE)) { 2158 pr_err("A sampling buffer entry is incorrect (alert=0x%x)\n", 2159 alert); 2160 cpuhw->flags |= PMU_F_ERR_IBE; 2161 sf_disable(); 2162 } 2163 } 2164 2165 static int cpusf_pmu_setup(unsigned int cpu, int flags) 2166 { 2167 /* Ignore the notification if no events are scheduled on the PMU. 2168 * This might be racy... 2169 */ 2170 if (!atomic_read(&num_events)) 2171 return 0; 2172 2173 local_irq_disable(); 2174 setup_pmc_cpu(&flags); 2175 local_irq_enable(); 2176 return 0; 2177 } 2178 2179 static int s390_pmu_sf_online_cpu(unsigned int cpu) 2180 { 2181 return cpusf_pmu_setup(cpu, PMC_INIT); 2182 } 2183 2184 static int s390_pmu_sf_offline_cpu(unsigned int cpu) 2185 { 2186 return cpusf_pmu_setup(cpu, PMC_RELEASE); 2187 } 2188 2189 static int param_get_sfb_size(char *buffer, const struct kernel_param *kp) 2190 { 2191 if (!cpum_sf_avail()) 2192 return -ENODEV; 2193 return sprintf(buffer, "%lu,%lu", CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 2194 } 2195 2196 static int param_set_sfb_size(const char *val, const struct kernel_param *kp) 2197 { 2198 int rc; 2199 unsigned long min, max; 2200 2201 if (!cpum_sf_avail()) 2202 return -ENODEV; 2203 if (!val || !strlen(val)) 2204 return -EINVAL; 2205 2206 /* Valid parameter values: "min,max" or "max" */ 2207 min = CPUM_SF_MIN_SDB; 2208 max = CPUM_SF_MAX_SDB; 2209 if (strchr(val, ',')) 2210 rc = (sscanf(val, "%lu,%lu", &min, &max) == 2) ? 0 : -EINVAL; 2211 else 2212 rc = kstrtoul(val, 10, &max); 2213 2214 if (min < 2 || min >= max || max > get_num_physpages()) 2215 rc = -EINVAL; 2216 if (rc) 2217 return rc; 2218 2219 sfb_set_limits(min, max); 2220 pr_info("The sampling buffer limits have changed to: " 2221 "min %lu max %lu (diag %lu)\n", 2222 CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB, CPUM_SF_SDB_DIAG_FACTOR); 2223 return 0; 2224 } 2225 2226 #define param_check_sfb_size(name, p) __param_check(name, p, void) 2227 static const struct kernel_param_ops param_ops_sfb_size = { 2228 .set = param_set_sfb_size, 2229 .get = param_get_sfb_size, 2230 }; 2231 2232 #define RS_INIT_FAILURE_QSI 0x0001 2233 #define RS_INIT_FAILURE_BSDES 0x0002 2234 #define RS_INIT_FAILURE_ALRT 0x0003 2235 #define RS_INIT_FAILURE_PERF 0x0004 2236 static void __init pr_cpumsf_err(unsigned int reason) 2237 { 2238 pr_err("Sampling facility support for perf is not available: " 2239 "reason %#x\n", reason); 2240 } 2241 2242 static int __init init_cpum_sampling_pmu(void) 2243 { 2244 struct hws_qsi_info_block si; 2245 int err; 2246 2247 if (!cpum_sf_avail()) 2248 return -ENODEV; 2249 2250 memset(&si, 0, sizeof(si)); 2251 if (qsi(&si)) { 2252 pr_cpumsf_err(RS_INIT_FAILURE_QSI); 2253 return -ENODEV; 2254 } 2255 2256 if (!si.as && !si.ad) 2257 return -ENODEV; 2258 2259 if (si.bsdes != sizeof(struct hws_basic_entry)) { 2260 pr_cpumsf_err(RS_INIT_FAILURE_BSDES); 2261 return -EINVAL; 2262 } 2263 2264 if (si.ad) { 2265 sfb_set_limits(CPUM_SF_MIN_SDB, CPUM_SF_MAX_SDB); 2266 /* Sampling of diagnostic data authorized, 2267 * install event into attribute list of PMU device. 2268 */ 2269 cpumsf_pmu_events_attr[SF_CYCLES_BASIC_DIAG_ATTR_IDX] = 2270 CPUMF_EVENT_PTR(SF, SF_CYCLES_BASIC_DIAG); 2271 } 2272 2273 sfdbg = debug_register(KMSG_COMPONENT, 2, 1, 80); 2274 if (!sfdbg) { 2275 pr_err("Registering for s390dbf failed\n"); 2276 return -ENOMEM; 2277 } 2278 debug_register_view(sfdbg, &debug_sprintf_view); 2279 2280 err = register_external_irq(EXT_IRQ_MEASURE_ALERT, 2281 cpumf_measurement_alert); 2282 if (err) { 2283 pr_cpumsf_err(RS_INIT_FAILURE_ALRT); 2284 debug_unregister(sfdbg); 2285 goto out; 2286 } 2287 2288 err = perf_pmu_register(&cpumf_sampling, "cpum_sf", PERF_TYPE_RAW); 2289 if (err) { 2290 pr_cpumsf_err(RS_INIT_FAILURE_PERF); 2291 unregister_external_irq(EXT_IRQ_MEASURE_ALERT, 2292 cpumf_measurement_alert); 2293 debug_unregister(sfdbg); 2294 goto out; 2295 } 2296 2297 cpuhp_setup_state(CPUHP_AP_PERF_S390_SF_ONLINE, "perf/s390/sf:online", 2298 s390_pmu_sf_online_cpu, s390_pmu_sf_offline_cpu); 2299 out: 2300 return err; 2301 } 2302 2303 arch_initcall(init_cpum_sampling_pmu); 2304 core_param(cpum_sfb_size, CPUM_SF_MAX_SDB, sfb_size, 0644); 2305