mm/page_owner.c

   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 inline bool check_recursive_alloc(unsigned long *entries,
 102                                          unsigned int nr_entries,
 103                                          unsigned long ip)
 104 {
 105         unsigned int i;
 106
 107         for (i = 0; i < nr_entries; i++) {
 108                 if (entries[i] == ip)
 109                         return true;
 110         }
 111         return false;
 112 }
 113
 114 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 115 {
 116         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 117         depot_stack_handle_t handle;
 118         unsigned int nr_entries;
 119
 120         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 121
 122         /*
 123          * We need to check recursion here because our request to
 124          * stackdepot could trigger memory allocation to save new
 125          * entry. New memory allocation would reach here and call
 126          * stack_depot_save_entries() again if we don't catch it. There is
 127          * still not enough memory in stackdepot so it would try to
 128          * allocate memory again and loop forever.
 129          */
 130         if (check_recursive_alloc(entries, nr_entries, _RET_IP_))
 131                 return dummy_handle;
 132
 133         handle = stack_depot_save(entries, nr_entries, flags);
 134         if (!handle)
 135                 handle = failure_handle;
 136
 137         return handle;
 138 }
 139
 140 void __reset_page_owner(struct page *page, unsigned int order)
 141 {
 142         int i;
 143         struct page_ext *page_ext;
 144         depot_stack_handle_t handle;
 145         struct page_owner *page_owner;
 146         u64 free_ts_nsec = local_clock();
 147
 148         page_ext = lookup_page_ext(page);
 149         if (unlikely(!page_ext))
 150                 return;
 151
 152         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 153         for (i = 0; i < (1 << order); i++) {
 154                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 155                 page_owner = get_page_owner(page_ext);
 156                 page_owner->free_handle = handle;
 157                 page_owner->free_ts_nsec = free_ts_nsec;
 158                 page_ext = page_ext_next(page_ext);
 159         }
 160 }
 161
 162 static inline void __set_page_owner_handle(struct page_ext *page_ext,
 163                                         depot_stack_handle_t handle,
 164                                         unsigned int order, gfp_t gfp_mask)
 165 {
 166         struct page_owner *page_owner;
 167         int i;
 168
 169         for (i = 0; i < (1 << order); i++) {
 170                 page_owner = get_page_owner(page_ext);
 171                 page_owner->handle = handle;
 172                 page_owner->order = order;
 173                 page_owner->gfp_mask = gfp_mask;
 174                 page_owner->last_migrate_reason = -1;
 175                 page_owner->pid = current->pid;
 176                 page_owner->ts_nsec = local_clock();
 177                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 178                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 179
 180                 page_ext = page_ext_next(page_ext);
 181         }
 182 }
 183
 184 noinline void __set_page_owner(struct page *page, unsigned int order,
 185                                         gfp_t gfp_mask)
 186 {
 187         struct page_ext *page_ext = lookup_page_ext(page);
 188         depot_stack_handle_t handle;
 189
 190         if (unlikely(!page_ext))
 191                 return;
 192
 193         handle = save_stack(gfp_mask);
 194         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 195 }
 196
 197 void __set_page_owner_migrate_reason(struct page *page, int reason)
 198 {
 199         struct page_ext *page_ext = lookup_page_ext(page);
 200         struct page_owner *page_owner;
 201
 202         if (unlikely(!page_ext))
 203                 return;
 204
 205         page_owner = get_page_owner(page_ext);
 206         page_owner->last_migrate_reason = reason;
 207 }
 208
 209 void __split_page_owner(struct page *page, unsigned int nr)
 210 {
 211         int i;
 212         struct page_ext *page_ext = lookup_page_ext(page);
 213         struct page_owner *page_owner;
 214
 215         if (unlikely(!page_ext))
 216                 return;
 217
 218         for (i = 0; i < nr; i++) {
 219                 page_owner = get_page_owner(page_ext);
 220                 page_owner->order = 0;
 221                 page_ext = page_ext_next(page_ext);
 222         }
 223 }
 224
 225 void __copy_page_owner(struct page *oldpage, struct page *newpage)
 226 {
 227         struct page_ext *old_ext = lookup_page_ext(oldpage);
 228         struct page_ext *new_ext = lookup_page_ext(newpage);
 229         struct page_owner *old_page_owner, *new_page_owner;
 230
 231         if (unlikely(!old_ext || !new_ext))
 232                 return;
 233
 234         old_page_owner = get_page_owner(old_ext);
 235         new_page_owner = get_page_owner(new_ext);
 236         new_page_owner->order = old_page_owner->order;
 237         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 238         new_page_owner->last_migrate_reason =
 239                 old_page_owner->last_migrate_reason;
 240         new_page_owner->handle = old_page_owner->handle;
 241         new_page_owner->pid = old_page_owner->pid;
 242         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
 243         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
 244
 245         /*
 246          * We don't clear the bit on the oldpage as it's going to be freed
 247          * after migration. Until then, the info can be useful in case of
 248          * a bug, and the overal stats will be off a bit only temporarily.
 249          * Also, migrate_misplaced_transhuge_page() can still fail the
 250          * migration and then we want the oldpage to retain the info. But
 251          * in that case we also don't need to explicitly clear the info from
 252          * the new page, which will be freed.
 253          */
 254         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 255         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
 256 }
 257
 258 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 259                                        pg_data_t *pgdat, struct zone *zone)
 260 {
 261         struct page *page;
 262         struct page_ext *page_ext;
 263         struct page_owner *page_owner;
 264         unsigned long pfn, block_end_pfn;
 265         unsigned long end_pfn = zone_end_pfn(zone);
 266         unsigned long count[MIGRATE_TYPES] = { 0, };
 267         int pageblock_mt, page_mt;
 268         int i;
 269
 270         /* Scan block by block. First and last block may be incomplete */
 271         pfn = zone->zone_start_pfn;
 272
 273         /*
 274          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 275          * a zone boundary, it will be double counted between zones. This does
 276          * not matter as the mixed block count will still be correct
 277          */
 278         for (; pfn < end_pfn; ) {
 279                 page = pfn_to_online_page(pfn);
 280                 if (!page) {
 281                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 282                         continue;
 283                 }
 284
 285                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 286                 block_end_pfn = min(block_end_pfn, end_pfn);
 287
 288                 pageblock_mt = get_pageblock_migratetype(page);
 289
 290                 for (; pfn < block_end_pfn; pfn++) {
 291                         if (!pfn_valid_within(pfn))
 292                                 continue;
 293
 294                         /* The pageblock is online, no need to recheck. */
 295                         page = pfn_to_page(pfn);
 296
 297                         if (page_zone(page) != zone)
 298                                 continue;
 299
 300                         if (PageBuddy(page)) {
 301                                 unsigned long freepage_order;
 302
 303                                 freepage_order = buddy_order_unsafe(page);
 304                                 if (freepage_order < MAX_ORDER)
 305                                         pfn += (1UL << freepage_order) - 1;
 306                                 continue;
 307                         }
 308
 309                         if (PageReserved(page))
 310                                 continue;
 311
 312                         page_ext = lookup_page_ext(page);
 313                         if (unlikely(!page_ext))
 314                                 continue;
 315
 316                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 317                                 continue;
 318
 319                         page_owner = get_page_owner(page_ext);
 320                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 321                         if (pageblock_mt != page_mt) {
 322                                 if (is_migrate_cma(pageblock_mt))
 323                                         count[MIGRATE_MOVABLE]++;
 324                                 else
 325                                         count[pageblock_mt]++;
 326
 327                                 pfn = block_end_pfn;
 328                                 break;
 329                         }
 330                         pfn += (1UL << page_owner->order) - 1;
 331                 }
 332         }
 333
 334         /* Print counts */
 335         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 336         for (i = 0; i < MIGRATE_TYPES; i++)
 337                 seq_printf(m, "%12lu ", count[i]);
 338         seq_putc(m, '\n');
 339 }
 340
 341 static ssize_t
 342 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 343                 struct page *page, struct page_owner *page_owner,
 344                 depot_stack_handle_t handle)
 345 {
 346         int ret, pageblock_mt, page_mt;
 347         unsigned long *entries;
 348         unsigned int nr_entries;
 349         char *kbuf;
 350
 351         count = min_t(size_t, count, PAGE_SIZE);
 352         kbuf = kmalloc(count, GFP_KERNEL);
 353         if (!kbuf)
 354                 return -ENOMEM;
 355
 356         ret = snprintf(kbuf, count,
 357                         "Page allocated via order %u, mask %#x(%pGg), pid %d, ts %llu ns, free_ts %llu ns\n",
 358                         page_owner->order, page_owner->gfp_mask,
 359                         &page_owner->gfp_mask, page_owner->pid,
 360                         page_owner->ts_nsec, page_owner->free_ts_nsec);
 361
 362         if (ret >= count)
 363                 goto err;
 364
 365         /* Print information relevant to grouping pages by mobility */
 366         pageblock_mt = get_pageblock_migratetype(page);
 367         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 368         ret += snprintf(kbuf + ret, count - ret,
 369                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 370                         pfn,
 371                         migratetype_names[page_mt],
 372                         pfn >> pageblock_order,
 373                         migratetype_names[pageblock_mt],
 374                         page->flags, &page->flags);
 375
 376         if (ret >= count)
 377                 goto err;
 378
 379         nr_entries = stack_depot_fetch(handle, &entries);
 380         ret += stack_trace_snprint(kbuf + ret, count - ret, entries, nr_entries, 0);
 381         if (ret >= count)
 382                 goto err;
 383
 384         if (page_owner->last_migrate_reason != -1) {
 385                 ret += snprintf(kbuf + ret, count - ret,
 386                         "Page has been migrated, last migrate reason: %s\n",
 387                         migrate_reason_names[page_owner->last_migrate_reason]);
 388                 if (ret >= count)
 389                         goto err;
 390         }
 391
 392         ret += snprintf(kbuf + ret, count - ret, "\n");
 393         if (ret >= count)
 394                 goto err;
 395
 396         if (copy_to_user(buf, kbuf, ret))
 397                 ret = -EFAULT;
 398
 399         kfree(kbuf);
 400         return ret;
 401
 402 err:
 403         kfree(kbuf);
 404         return -ENOMEM;
 405 }
 406
 407 void __dump_page_owner(struct page *page)
 408 {
 409         struct page_ext *page_ext = lookup_page_ext(page);
 410         struct page_owner *page_owner;
 411         depot_stack_handle_t handle;
 412         unsigned long *entries;
 413         unsigned int nr_entries;
 414         gfp_t gfp_mask;
 415         int mt;
 416
 417         if (unlikely(!page_ext)) {
 418                 pr_alert("There is not page extension available.\n");
 419                 return;
 420         }
 421
 422         page_owner = get_page_owner(page_ext);
 423         gfp_mask = page_owner->gfp_mask;
 424         mt = gfp_migratetype(gfp_mask);
 425
 426         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 427                 pr_alert("page_owner info is not present (never set?)\n");
 428                 return;
 429         }
 430
 431         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 432                 pr_alert("page_owner tracks the page as allocated\n");
 433         else
 434                 pr_alert("page_owner tracks the page as freed\n");
 435
 436         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, ts %llu, free_ts %llu\n",
 437                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 438                  page_owner->pid, page_owner->ts_nsec, page_owner->free_ts_nsec);
 439
 440         handle = READ_ONCE(page_owner->handle);
 441         if (!handle) {
 442                 pr_alert("page_owner allocation stack trace missing\n");
 443         } else {
 444                 nr_entries = stack_depot_fetch(handle, &entries);
 445                 stack_trace_print(entries, nr_entries, 0);
 446         }
 447
 448         handle = READ_ONCE(page_owner->free_handle);
 449         if (!handle) {
 450                 pr_alert("page_owner free stack trace missing\n");
 451         } else {
 452                 nr_entries = stack_depot_fetch(handle, &entries);
 453                 pr_alert("page last free stack trace:\n");
 454                 stack_trace_print(entries, nr_entries, 0);
 455         }
 456
 457         if (page_owner->last_migrate_reason != -1)
 458                 pr_alert("page has been migrated, last migrate reason: %s\n",
 459                         migrate_reason_names[page_owner->last_migrate_reason]);
 460 }
 461
 462 static ssize_t
 463 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 464 {
 465         unsigned long pfn;
 466         struct page *page;
 467         struct page_ext *page_ext;
 468         struct page_owner *page_owner;
 469         depot_stack_handle_t handle;
 470
 471         if (!static_branch_unlikely(&page_owner_inited))
 472                 return -EINVAL;
 473
 474         page = NULL;
 475         pfn = min_low_pfn + *ppos;
 476
 477         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 478         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 479                 pfn++;
 480
 481         drain_all_pages(NULL);
 482
 483         /* Find an allocated page */
 484         for (; pfn < max_pfn; pfn++) {
 485                 /*
 486                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 487                  * validate the area as existing, skip it if not
 488                  */
 489                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 490                         pfn += MAX_ORDER_NR_PAGES - 1;
 491                         continue;
 492                 }
 493
 494                 /* Check for holes within a MAX_ORDER area */
 495                 if (!pfn_valid_within(pfn))
 496                         continue;
 497
 498                 page = pfn_to_page(pfn);
 499                 if (PageBuddy(page)) {
 500                         unsigned long freepage_order = buddy_order_unsafe(page);
 501
 502                         if (freepage_order < MAX_ORDER)
 503                                 pfn += (1UL << freepage_order) - 1;
 504                         continue;
 505                 }
 506
 507                 page_ext = lookup_page_ext(page);
 508                 if (unlikely(!page_ext))
 509                         continue;
 510
 511                 /*
 512                  * Some pages could be missed by concurrent allocation or free,
 513                  * because we don't hold the zone lock.
 514                  */
 515                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 516                         continue;
 517
 518                 /*
 519                  * Although we do have the info about past allocation of free
 520                  * pages, it's not relevant for current memory usage.
 521                  */
 522                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 523                         continue;
 524
 525                 page_owner = get_page_owner(page_ext);
 526
 527                 /*
 528                  * Don't print "tail" pages of high-order allocations as that
 529                  * would inflate the stats.
 530                  */
 531                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 532                         continue;
 533
 534                 /*
 535                  * Access to page_ext->handle isn't synchronous so we should
 536                  * be careful to access it.
 537                  */
 538                 handle = READ_ONCE(page_owner->handle);
 539                 if (!handle)
 540                         continue;
 541
 542                 /* Record the next PFN to read in the file offset */
 543                 *ppos = (pfn - min_low_pfn) + 1;
 544
 545                 return print_page_owner(buf, count, pfn, page,
 546                                 page_owner, handle);
 547         }
 548
 549         return 0;
 550 }
 551
 552 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 553 {
 554         unsigned long pfn = zone->zone_start_pfn;
 555         unsigned long end_pfn = zone_end_pfn(zone);
 556         unsigned long count = 0;
 557
 558         /*
 559          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 560          * a zone boundary, it will be double counted between zones. This does
 561          * not matter as the mixed block count will still be correct
 562          */
 563         for (; pfn < end_pfn; ) {
 564                 unsigned long block_end_pfn;
 565
 566                 if (!pfn_valid(pfn)) {
 567                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 568                         continue;
 569                 }
 570
 571                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 572                 block_end_pfn = min(block_end_pfn, end_pfn);
 573
 574                 for (; pfn < block_end_pfn; pfn++) {
 575                         struct page *page;
 576                         struct page_ext *page_ext;
 577
 578                         if (!pfn_valid_within(pfn))
 579                                 continue;
 580
 581                         page = pfn_to_page(pfn);
 582
 583                         if (page_zone(page) != zone)
 584                                 continue;
 585
 586                         /*
 587                          * To avoid having to grab zone->lock, be a little
 588                          * careful when reading buddy page order. The only
 589                          * danger is that we skip too much and potentially miss
 590                          * some early allocated pages, which is better than
 591                          * heavy lock contention.
 592                          */
 593                         if (PageBuddy(page)) {
 594                                 unsigned long order = buddy_order_unsafe(page);
 595
 596                                 if (order > 0 && order < MAX_ORDER)
 597                                         pfn += (1UL << order) - 1;
 598                                 continue;
 599                         }
 600
 601                         if (PageReserved(page))
 602                                 continue;
 603
 604                         page_ext = lookup_page_ext(page);
 605                         if (unlikely(!page_ext))
 606                                 continue;
 607
 608                         /* Maybe overlapping zone */
 609                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 610                                 continue;
 611
 612                         /* Found early allocated page */
 613                         __set_page_owner_handle(page_ext, early_handle,
 614                                                 0, 0);
 615                         count++;
 616                 }
 617                 cond_resched();
 618         }
 619
 620         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 621                 pgdat->node_id, zone->name, count);
 622 }
 623
 624 static void init_zones_in_node(pg_data_t *pgdat)
 625 {
 626         struct zone *zone;
 627         struct zone *node_zones = pgdat->node_zones;
 628
 629         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 630                 if (!populated_zone(zone))
 631                         continue;
 632
 633                 init_pages_in_zone(pgdat, zone);
 634         }
 635 }
 636
 637 static void init_early_allocated_pages(void)
 638 {
 639         pg_data_t *pgdat;
 640
 641         for_each_online_pgdat(pgdat)
 642                 init_zones_in_node(pgdat);
 643 }
 644
 645 static const struct file_operations proc_page_owner_operations = {
 646         .read           = read_page_owner,
 647 };
 648
 649 static int __init pageowner_init(void)
 650 {
 651         if (!static_branch_unlikely(&page_owner_inited)) {
 652                 pr_info("page_owner is disabled\n");
 653                 return 0;
 654         }
 655
 656         debugfs_create_file("page_owner", 0400, NULL, NULL,
 657                             &proc_page_owner_operations);
 658
 659         return 0;
 660 }
 661 late_initcall(pageowner_init)