1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/debugfs.h> 3 #include <linux/mm.h> 4 #include <linux/slab.h> 5 #include <linux/uaccess.h> 6 #include <linux/memblock.h> 7 #include <linux/stacktrace.h> 8 #include <linux/page_owner.h> 9 #include <linux/jump_label.h> 10 #include <linux/migrate.h> 11 #include <linux/stackdepot.h> 12 #include <linux/seq_file.h> 13 #include <linux/sched/clock.h> 14 15 #include "internal.h" 16 17 /* 18 * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack) 19 * to use off stack temporal storage 20 */ 21 #define PAGE_OWNER_STACK_DEPTH (16) 22 23 struct page_owner { 24 unsigned short order; 25 short last_migrate_reason; 26 gfp_t gfp_mask; 27 depot_stack_handle_t handle; 28 depot_stack_handle_t free_handle; 29 u64 ts_nsec; 30 u64 free_ts_nsec; 31 pid_t pid; 32 }; 33 34 static bool page_owner_enabled = false; 35 DEFINE_STATIC_KEY_FALSE(page_owner_inited); 36 37 static depot_stack_handle_t dummy_handle; 38 static depot_stack_handle_t failure_handle; 39 static depot_stack_handle_t early_handle; 40 41 static void init_early_allocated_pages(void); 42 43 static int __init early_page_owner_param(char *buf) 44 { 45 return kstrtobool(buf, &page_owner_enabled); 46 } 47 early_param("page_owner", early_page_owner_param); 48 49 static bool need_page_owner(void) 50 { 51 return page_owner_enabled; 52 } 53 54 static __always_inline depot_stack_handle_t create_dummy_stack(void) 55 { 56 unsigned long entries[4]; 57 unsigned int nr_entries; 58 59 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0); 60 return stack_depot_save(entries, nr_entries, GFP_KERNEL); 61 } 62 63 static noinline void register_dummy_stack(void) 64 { 65 dummy_handle = create_dummy_stack(); 66 } 67 68 static noinline void register_failure_stack(void) 69 { 70 failure_handle = create_dummy_stack(); 71 } 72 73 static noinline void register_early_stack(void) 74 { 75 early_handle = create_dummy_stack(); 76 } 77 78 static void init_page_owner(void) 79 { 80 if (!page_owner_enabled) 81 return; 82 83 register_dummy_stack(); 84 register_failure_stack(); 85 register_early_stack(); 86 static_branch_enable(&page_owner_inited); 87 init_early_allocated_pages(); 88 } 89 90 struct page_ext_operations page_owner_ops = { 91 .size = sizeof(struct page_owner), 92 .need = need_page_owner, 93 .init = init_page_owner, 94 }; 95 96 static inline struct page_owner *get_page_owner(struct page_ext *page_ext) 97 { 98 return (void *)page_ext + page_owner_ops.offset; 99 } 100 101 static noinline depot_stack_handle_t save_stack(gfp_t flags) 102 { 103 unsigned long entries[PAGE_OWNER_STACK_DEPTH]; 104 depot_stack_handle_t handle; 105 unsigned int nr_entries; 106 107 /* 108 * Avoid recursion. 109 * 110 * Sometimes page metadata allocation tracking requires more 111 * memory to be allocated: 112 * - when new stack trace is saved to stack depot 113 * - when backtrace itself is calculated (ia64) 114 */ 115 if (current->in_page_owner) 116 return dummy_handle; 117 current->in_page_owner = 1; 118 119 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2); 120 handle = stack_depot_save(entries, nr_entries, flags); 121 if (!handle) 122 handle = failure_handle; 123 124 current->in_page_owner = 0; 125 return handle; 126 } 127 128 void __reset_page_owner(struct page *page, unsigned int order) 129 { 130 int i; 131 struct page_ext *page_ext; 132 depot_stack_handle_t handle; 133 struct page_owner *page_owner; 134 u64 free_ts_nsec = local_clock(); 135 136 page_ext = lookup_page_ext(page); 137 if (unlikely(!page_ext)) 138 return; 139 140 handle = save_stack(GFP_NOWAIT | __GFP_NOWARN); 141 for (i = 0; i < (1 << order); i++) { 142 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags); 143 page_owner = get_page_owner(page_ext); 144 page_owner->free_handle = handle; 145 page_owner->free_ts_nsec = free_ts_nsec; 146 page_ext = page_ext_next(page_ext); 147 } 148 } 149 150 static inline void __set_page_owner_handle(struct page_ext *page_ext, 151 depot_stack_handle_t handle, 152 unsigned int order, gfp_t gfp_mask) 153 { 154 struct page_owner *page_owner; 155 int i; 156 157 for (i = 0; i < (1 << order); i++) { 158 page_owner = get_page_owner(page_ext); 159 page_owner->handle = handle; 160 page_owner->order = order; 161 page_owner->gfp_mask = gfp_mask; 162 page_owner->last_migrate_reason = -1; 163 page_owner->pid = current->pid; 164 page_owner->ts_nsec = local_clock(); 165 __set_bit(PAGE_EXT_OWNER, &page_ext->flags); 166 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags); 167 168 page_ext = page_ext_next(page_ext); 169 } 170 } 171 172 noinline void __set_page_owner(struct page *page, unsigned int order, 173 gfp_t gfp_mask) 174 { 175 struct page_ext *page_ext = lookup_page_ext(page); 176 depot_stack_handle_t handle; 177 178 if (unlikely(!page_ext)) 179 return; 180 181 handle = save_stack(gfp_mask); 182 __set_page_owner_handle(page_ext, handle, order, gfp_mask); 183 } 184 185 void __set_page_owner_migrate_reason(struct page *page, int reason) 186 { 187 struct page_ext *page_ext = lookup_page_ext(page); 188 struct page_owner *page_owner; 189 190 if (unlikely(!page_ext)) 191 return; 192 193 page_owner = get_page_owner(page_ext); 194 page_owner->last_migrate_reason = reason; 195 } 196 197 void __split_page_owner(struct page *page, unsigned int nr) 198 { 199 int i; 200 struct page_ext *page_ext = lookup_page_ext(page); 201 struct page_owner *page_owner; 202 203 if (unlikely(!page_ext)) 204 return; 205 206 for (i = 0; i < nr; i++) { 207 page_owner = get_page_owner(page_ext); 208 page_owner->order = 0; 209 page_ext = page_ext_next(page_ext); 210 } 211 } 212 213 void __copy_page_owner(struct page *oldpage, struct page *newpage) 214 { 215 struct page_ext *old_ext = lookup_page_ext(oldpage); 216 struct page_ext *new_ext = lookup_page_ext(newpage); 217 struct page_owner *old_page_owner, *new_page_owner; 218 219 if (unlikely(!old_ext || !new_ext)) 220 return; 221 222 old_page_owner = get_page_owner(old_ext); 223 new_page_owner = get_page_owner(new_ext); 224 new_page_owner->order = old_page_owner->order; 225 new_page_owner->gfp_mask = old_page_owner->gfp_mask; 226 new_page_owner->last_migrate_reason = 227 old_page_owner->last_migrate_reason; 228 new_page_owner->handle = old_page_owner->handle; 229 new_page_owner->pid = old_page_owner->pid; 230 new_page_owner->ts_nsec = old_page_owner->ts_nsec; 231 new_page_owner->free_ts_nsec = old_page_owner->ts_nsec; 232 233 /* 234 * We don't clear the bit on the oldpage as it's going to be freed 235 * after migration. Until then, the info can be useful in case of 236 * a bug, and the overall stats will be off a bit only temporarily. 237 * Also, migrate_misplaced_transhuge_page() can still fail the 238 * migration and then we want the oldpage to retain the info. But 239 * in that case we also don't need to explicitly clear the info from 240 * the new page, which will be freed. 241 */ 242 __set_bit(PAGE_EXT_OWNER, &new_ext->flags); 243 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags); 244 } 245 246 void pagetypeinfo_showmixedcount_print(struct seq_file *m, 247 pg_data_t *pgdat, struct zone *zone) 248 { 249 struct page *page; 250 struct page_ext *page_ext; 251 struct page_owner *page_owner; 252 unsigned long pfn, block_end_pfn; 253 unsigned long end_pfn = zone_end_pfn(zone); 254 unsigned long count[MIGRATE_TYPES] = { 0, }; 255 int pageblock_mt, page_mt; 256 int i; 257 258 /* Scan block by block. First and last block may be incomplete */ 259 pfn = zone->zone_start_pfn; 260 261 /* 262 * Walk the zone in pageblock_nr_pages steps. If a page block spans 263 * a zone boundary, it will be double counted between zones. This does 264 * not matter as the mixed block count will still be correct 265 */ 266 for (; pfn < end_pfn; ) { 267 page = pfn_to_online_page(pfn); 268 if (!page) { 269 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES); 270 continue; 271 } 272 273 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); 274 block_end_pfn = min(block_end_pfn, end_pfn); 275 276 pageblock_mt = get_pageblock_migratetype(page); 277 278 for (; pfn < block_end_pfn; pfn++) { 279 if (!pfn_valid_within(pfn)) 280 continue; 281 282 /* The pageblock is online, no need to recheck. */ 283 page = pfn_to_page(pfn); 284 285 if (page_zone(page) != zone) 286 continue; 287 288 if (PageBuddy(page)) { 289 unsigned long freepage_order; 290 291 freepage_order = buddy_order_unsafe(page); 292 if (freepage_order < MAX_ORDER) 293 pfn += (1UL << freepage_order) - 1; 294 continue; 295 } 296 297 if (PageReserved(page)) 298 continue; 299 300 page_ext = lookup_page_ext(page); 301 if (unlikely(!page_ext)) 302 continue; 303 304 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags)) 305 continue; 306 307 page_owner = get_page_owner(page_ext); 308 page_mt = gfp_migratetype(page_owner->gfp_mask); 309 if (pageblock_mt != page_mt) { 310 if (is_migrate_cma(pageblock_mt)) 311 count[MIGRATE_MOVABLE]++; 312 else 313 count[pageblock_mt]++; 314 315 pfn = block_end_pfn; 316 break; 317 } 318 pfn += (1UL << page_owner->order) - 1; 319 } 320 } 321 322 /* Print counts */ 323 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name); 324 for (i = 0; i < MIGRATE_TYPES; i++) 325 seq_printf(m, "%12lu ", count[i]); 326 seq_putc(m, '\n'); 327 } 328 329 static ssize_t 330 print_page_owner(char __user *buf, size_t count, unsigned long pfn, 331 struct page *page, struct page_owner *page_owner, 332 depot_stack_handle_t handle) 333 { 334 int ret, pageblock_mt, page_mt; 335 unsigned long *entries; 336 unsigned int nr_entries; 337 char *kbuf; 338 339 count = min_t(size_t, count, PAGE_SIZE); 340 kbuf = kmalloc(count, GFP_KERNEL); 341 if (!kbuf) 342 return -ENOMEM; 343 344 ret = snprintf(kbuf, count, 345 "Page allocated via order %u, mask %#x(%pGg), pid %d, ts %llu ns, free_ts %llu ns\n", 346 page_owner->order, page_owner->gfp_mask, 347 &page_owner->gfp_mask, page_owner->pid, 348 page_owner->ts_nsec, page_owner->free_ts_nsec); 349 350 if (ret >= count) 351 goto err; 352 353 /* Print information relevant to grouping pages by mobility */ 354 pageblock_mt = get_pageblock_migratetype(page); 355 page_mt = gfp_migratetype(page_owner->gfp_mask); 356 ret += snprintf(kbuf + ret, count - ret, 357 "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n", 358 pfn, 359 migratetype_names[page_mt], 360 pfn >> pageblock_order, 361 migratetype_names[pageblock_mt], 362 page->flags, &page->flags); 363 364 if (ret >= count) 365 goto err; 366 367 nr_entries = stack_depot_fetch(handle, &entries); 368 ret += stack_trace_snprint(kbuf + ret, count - ret, entries, nr_entries, 0); 369 if (ret >= count) 370 goto err; 371 372 if (page_owner->last_migrate_reason != -1) { 373 ret += snprintf(kbuf + ret, count - ret, 374 "Page has been migrated, last migrate reason: %s\n", 375 migrate_reason_names[page_owner->last_migrate_reason]); 376 if (ret >= count) 377 goto err; 378 } 379 380 ret += snprintf(kbuf + ret, count - ret, "\n"); 381 if (ret >= count) 382 goto err; 383 384 if (copy_to_user(buf, kbuf, ret)) 385 ret = -EFAULT; 386 387 kfree(kbuf); 388 return ret; 389 390 err: 391 kfree(kbuf); 392 return -ENOMEM; 393 } 394 395 void __dump_page_owner(const struct page *page) 396 { 397 struct page_ext *page_ext = lookup_page_ext(page); 398 struct page_owner *page_owner; 399 depot_stack_handle_t handle; 400 unsigned long *entries; 401 unsigned int nr_entries; 402 gfp_t gfp_mask; 403 int mt; 404 405 if (unlikely(!page_ext)) { 406 pr_alert("There is not page extension available.\n"); 407 return; 408 } 409 410 page_owner = get_page_owner(page_ext); 411 gfp_mask = page_owner->gfp_mask; 412 mt = gfp_migratetype(gfp_mask); 413 414 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) { 415 pr_alert("page_owner info is not present (never set?)\n"); 416 return; 417 } 418 419 if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags)) 420 pr_alert("page_owner tracks the page as allocated\n"); 421 else 422 pr_alert("page_owner tracks the page as freed\n"); 423 424 pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, ts %llu, free_ts %llu\n", 425 page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask, 426 page_owner->pid, page_owner->ts_nsec, page_owner->free_ts_nsec); 427 428 handle = READ_ONCE(page_owner->handle); 429 if (!handle) { 430 pr_alert("page_owner allocation stack trace missing\n"); 431 } else { 432 nr_entries = stack_depot_fetch(handle, &entries); 433 stack_trace_print(entries, nr_entries, 0); 434 } 435 436 handle = READ_ONCE(page_owner->free_handle); 437 if (!handle) { 438 pr_alert("page_owner free stack trace missing\n"); 439 } else { 440 nr_entries = stack_depot_fetch(handle, &entries); 441 pr_alert("page last free stack trace:\n"); 442 stack_trace_print(entries, nr_entries, 0); 443 } 444 445 if (page_owner->last_migrate_reason != -1) 446 pr_alert("page has been migrated, last migrate reason: %s\n", 447 migrate_reason_names[page_owner->last_migrate_reason]); 448 } 449 450 static ssize_t 451 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos) 452 { 453 unsigned long pfn; 454 struct page *page; 455 struct page_ext *page_ext; 456 struct page_owner *page_owner; 457 depot_stack_handle_t handle; 458 459 if (!static_branch_unlikely(&page_owner_inited)) 460 return -EINVAL; 461 462 page = NULL; 463 pfn = min_low_pfn + *ppos; 464 465 /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */ 466 while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) 467 pfn++; 468 469 drain_all_pages(NULL); 470 471 /* Find an allocated page */ 472 for (; pfn < max_pfn; pfn++) { 473 /* 474 * If the new page is in a new MAX_ORDER_NR_PAGES area, 475 * validate the area as existing, skip it if not 476 */ 477 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) { 478 pfn += MAX_ORDER_NR_PAGES - 1; 479 continue; 480 } 481 482 /* Check for holes within a MAX_ORDER area */ 483 if (!pfn_valid_within(pfn)) 484 continue; 485 486 page = pfn_to_page(pfn); 487 if (PageBuddy(page)) { 488 unsigned long freepage_order = buddy_order_unsafe(page); 489 490 if (freepage_order < MAX_ORDER) 491 pfn += (1UL << freepage_order) - 1; 492 continue; 493 } 494 495 page_ext = lookup_page_ext(page); 496 if (unlikely(!page_ext)) 497 continue; 498 499 /* 500 * Some pages could be missed by concurrent allocation or free, 501 * because we don't hold the zone lock. 502 */ 503 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) 504 continue; 505 506 /* 507 * Although we do have the info about past allocation of free 508 * pages, it's not relevant for current memory usage. 509 */ 510 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags)) 511 continue; 512 513 page_owner = get_page_owner(page_ext); 514 515 /* 516 * Don't print "tail" pages of high-order allocations as that 517 * would inflate the stats. 518 */ 519 if (!IS_ALIGNED(pfn, 1 << page_owner->order)) 520 continue; 521 522 /* 523 * Access to page_ext->handle isn't synchronous so we should 524 * be careful to access it. 525 */ 526 handle = READ_ONCE(page_owner->handle); 527 if (!handle) 528 continue; 529 530 /* Record the next PFN to read in the file offset */ 531 *ppos = (pfn - min_low_pfn) + 1; 532 533 return print_page_owner(buf, count, pfn, page, 534 page_owner, handle); 535 } 536 537 return 0; 538 } 539 540 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone) 541 { 542 unsigned long pfn = zone->zone_start_pfn; 543 unsigned long end_pfn = zone_end_pfn(zone); 544 unsigned long count = 0; 545 546 /* 547 * Walk the zone in pageblock_nr_pages steps. If a page block spans 548 * a zone boundary, it will be double counted between zones. This does 549 * not matter as the mixed block count will still be correct 550 */ 551 for (; pfn < end_pfn; ) { 552 unsigned long block_end_pfn; 553 554 if (!pfn_valid(pfn)) { 555 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES); 556 continue; 557 } 558 559 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages); 560 block_end_pfn = min(block_end_pfn, end_pfn); 561 562 for (; pfn < block_end_pfn; pfn++) { 563 struct page *page; 564 struct page_ext *page_ext; 565 566 if (!pfn_valid_within(pfn)) 567 continue; 568 569 page = pfn_to_page(pfn); 570 571 if (page_zone(page) != zone) 572 continue; 573 574 /* 575 * To avoid having to grab zone->lock, be a little 576 * careful when reading buddy page order. The only 577 * danger is that we skip too much and potentially miss 578 * some early allocated pages, which is better than 579 * heavy lock contention. 580 */ 581 if (PageBuddy(page)) { 582 unsigned long order = buddy_order_unsafe(page); 583 584 if (order > 0 && order < MAX_ORDER) 585 pfn += (1UL << order) - 1; 586 continue; 587 } 588 589 if (PageReserved(page)) 590 continue; 591 592 page_ext = lookup_page_ext(page); 593 if (unlikely(!page_ext)) 594 continue; 595 596 /* Maybe overlapping zone */ 597 if (test_bit(PAGE_EXT_OWNER, &page_ext->flags)) 598 continue; 599 600 /* Found early allocated page */ 601 __set_page_owner_handle(page_ext, early_handle, 602 0, 0); 603 count++; 604 } 605 cond_resched(); 606 } 607 608 pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n", 609 pgdat->node_id, zone->name, count); 610 } 611 612 static void init_zones_in_node(pg_data_t *pgdat) 613 { 614 struct zone *zone; 615 struct zone *node_zones = pgdat->node_zones; 616 617 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 618 if (!populated_zone(zone)) 619 continue; 620 621 init_pages_in_zone(pgdat, zone); 622 } 623 } 624 625 static void init_early_allocated_pages(void) 626 { 627 pg_data_t *pgdat; 628 629 for_each_online_pgdat(pgdat) 630 init_zones_in_node(pgdat); 631 } 632 633 static const struct file_operations proc_page_owner_operations = { 634 .read = read_page_owner, 635 }; 636 637 static int __init pageowner_init(void) 638 { 639 if (!static_branch_unlikely(&page_owner_inited)) { 640 pr_info("page_owner is disabled\n"); 641 return 0; 642 } 643 644 debugfs_create_file("page_owner", 0400, NULL, NULL, 645 &proc_page_owner_operations); 646 647 return 0; 648 } 649 late_initcall(pageowner_init) 650