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/memcontrol.h>
  14 #include <linux/sched/clock.h>
  15
  16 #include "internal.h"
  17
  18 /*
  19  * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
  20  * to use off stack temporal storage
  21  */
  22 #define PAGE_OWNER_STACK_DEPTH (16)
  23
  24 struct page_owner {
  25         unsigned short order;
  26         short last_migrate_reason;
  27         gfp_t gfp_mask;
  28         depot_stack_handle_t handle;
  29         depot_stack_handle_t free_handle;
  30         u64 ts_nsec;
  31         u64 free_ts_nsec;
  32         char comm[TASK_COMM_LEN];
  33         pid_t pid;
  34         pid_t tgid;
  35         pid_t free_pid;
  36         pid_t free_tgid;
  37 };
  38
  39 static bool page_owner_enabled __initdata;
  40 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  41
  42 static depot_stack_handle_t dummy_handle;
  43 static depot_stack_handle_t failure_handle;
  44 static depot_stack_handle_t early_handle;
  45
  46 static void init_early_allocated_pages(void);
  47
  48 static int __init early_page_owner_param(char *buf)
  49 {
  50         int ret = kstrtobool(buf, &page_owner_enabled);
  51
  52         if (page_owner_enabled)
  53                 stack_depot_request_early_init();
  54
  55         return ret;
  56 }
  57 early_param("page_owner", early_page_owner_param);
  58
  59 static __init bool need_page_owner(void)
  60 {
  61         return page_owner_enabled;
  62 }
  63
  64 static __always_inline depot_stack_handle_t create_dummy_stack(void)
  65 {
  66         unsigned long entries[4];
  67         unsigned int nr_entries;
  68
  69         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  70         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
  71 }
  72
  73 static noinline void register_dummy_stack(void)
  74 {
  75         dummy_handle = create_dummy_stack();
  76 }
  77
  78 static noinline void register_failure_stack(void)
  79 {
  80         failure_handle = create_dummy_stack();
  81 }
  82
  83 static noinline void register_early_stack(void)
  84 {
  85         early_handle = create_dummy_stack();
  86 }
  87
  88 static __init void init_page_owner(void)
  89 {
  90         if (!page_owner_enabled)
  91                 return;
  92
  93         register_dummy_stack();
  94         register_failure_stack();
  95         register_early_stack();
  96         static_branch_enable(&page_owner_inited);
  97         init_early_allocated_pages();
  98 }
  99
 100 struct page_ext_operations page_owner_ops = {
 101         .size = sizeof(struct page_owner),
 102         .need = need_page_owner,
 103         .init = init_page_owner,
 104         .need_shared_flags = true,
 105 };
 106
 107 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
 108 {
 109         return page_ext_data(page_ext, &page_owner_ops);
 110 }
 111
 112 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 113 {
 114         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 115         depot_stack_handle_t handle;
 116         unsigned int nr_entries;
 117
 118         /*
 119          * Avoid recursion.
 120          *
 121          * Sometimes page metadata allocation tracking requires more
 122          * memory to be allocated:
 123          * - when new stack trace is saved to stack depot
 124          * - when backtrace itself is calculated (ia64)
 125          */
 126         if (current->in_page_owner)
 127                 return dummy_handle;
 128         current->in_page_owner = 1;
 129
 130         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 131         handle = stack_depot_save(entries, nr_entries, flags);
 132         if (!handle)
 133                 handle = failure_handle;
 134
 135         current->in_page_owner = 0;
 136         return handle;
 137 }
 138
 139 void __reset_page_owner(struct page *page, unsigned short order)
 140 {
 141         int i;
 142         struct page_ext *page_ext;
 143         depot_stack_handle_t handle;
 144         struct page_owner *page_owner;
 145         u64 free_ts_nsec = local_clock();
 146
 147         page_ext = page_ext_get(page);
 148         if (unlikely(!page_ext))
 149                 return;
 150
 151         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 152         for (i = 0; i < (1 << order); i++) {
 153                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 154                 page_owner = get_page_owner(page_ext);
 155                 page_owner->free_handle = handle;
 156                 page_owner->free_ts_nsec = free_ts_nsec;
 157                 page_owner->free_pid = current->pid;
 158                 page_owner->free_tgid = current->tgid;
 159                 page_ext = page_ext_next(page_ext);
 160         }
 161         page_ext_put(page_ext);
 162 }
 163
 164 static inline void __set_page_owner_handle(struct page_ext *page_ext,
 165                                         depot_stack_handle_t handle,
 166                                         unsigned short order, gfp_t gfp_mask)
 167 {
 168         struct page_owner *page_owner;
 169         int i;
 170         u64 ts_nsec = local_clock();
 171
 172         for (i = 0; i < (1 << order); i++) {
 173                 page_owner = get_page_owner(page_ext);
 174                 page_owner->handle = handle;
 175                 page_owner->order = order;
 176                 page_owner->gfp_mask = gfp_mask;
 177                 page_owner->last_migrate_reason = -1;
 178                 page_owner->pid = current->pid;
 179                 page_owner->tgid = current->tgid;
 180                 page_owner->ts_nsec = ts_nsec;
 181                 strscpy(page_owner->comm, current->comm,
 182                         sizeof(page_owner->comm));
 183                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 184                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 185
 186                 page_ext = page_ext_next(page_ext);
 187         }
 188 }
 189
 190 noinline void __set_page_owner(struct page *page, unsigned short order,
 191                                         gfp_t gfp_mask)
 192 {
 193         struct page_ext *page_ext;
 194         depot_stack_handle_t handle;
 195
 196         handle = save_stack(gfp_mask);
 197
 198         page_ext = page_ext_get(page);
 199         if (unlikely(!page_ext))
 200                 return;
 201         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 202         page_ext_put(page_ext);
 203 }
 204
 205 void __set_page_owner_migrate_reason(struct page *page, int reason)
 206 {
 207         struct page_ext *page_ext = page_ext_get(page);
 208         struct page_owner *page_owner;
 209
 210         if (unlikely(!page_ext))
 211                 return;
 212
 213         page_owner = get_page_owner(page_ext);
 214         page_owner->last_migrate_reason = reason;
 215         page_ext_put(page_ext);
 216 }
 217
 218 void __split_page_owner(struct page *page, unsigned int nr)
 219 {
 220         int i;
 221         struct page_ext *page_ext = page_ext_get(page);
 222         struct page_owner *page_owner;
 223
 224         if (unlikely(!page_ext))
 225                 return;
 226
 227         for (i = 0; i < nr; i++) {
 228                 page_owner = get_page_owner(page_ext);
 229                 page_owner->order = 0;
 230                 page_ext = page_ext_next(page_ext);
 231         }
 232         page_ext_put(page_ext);
 233 }
 234
 235 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 236 {
 237         struct page_ext *old_ext;
 238         struct page_ext *new_ext;
 239         struct page_owner *old_page_owner, *new_page_owner;
 240
 241         old_ext = page_ext_get(&old->page);
 242         if (unlikely(!old_ext))
 243                 return;
 244
 245         new_ext = page_ext_get(&newfolio->page);
 246         if (unlikely(!new_ext)) {
 247                 page_ext_put(old_ext);
 248                 return;
 249         }
 250
 251         old_page_owner = get_page_owner(old_ext);
 252         new_page_owner = get_page_owner(new_ext);
 253         new_page_owner->order = old_page_owner->order;
 254         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 255         new_page_owner->last_migrate_reason =
 256                 old_page_owner->last_migrate_reason;
 257         new_page_owner->handle = old_page_owner->handle;
 258         new_page_owner->pid = old_page_owner->pid;
 259         new_page_owner->tgid = old_page_owner->tgid;
 260         new_page_owner->free_pid = old_page_owner->free_pid;
 261         new_page_owner->free_tgid = old_page_owner->free_tgid;
 262         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
 263         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
 264         strcpy(new_page_owner->comm, old_page_owner->comm);
 265
 266         /*
 267          * We don't clear the bit on the old folio as it's going to be freed
 268          * after migration. Until then, the info can be useful in case of
 269          * a bug, and the overall stats will be off a bit only temporarily.
 270          * Also, migrate_misplaced_transhuge_page() can still fail the
 271          * migration and then we want the old folio to retain the info. But
 272          * in that case we also don't need to explicitly clear the info from
 273          * the new page, which will be freed.
 274          */
 275         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 276         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
 277         page_ext_put(new_ext);
 278         page_ext_put(old_ext);
 279 }
 280
 281 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 282                                        pg_data_t *pgdat, struct zone *zone)
 283 {
 284         struct page *page;
 285         struct page_ext *page_ext;
 286         struct page_owner *page_owner;
 287         unsigned long pfn, block_end_pfn;
 288         unsigned long end_pfn = zone_end_pfn(zone);
 289         unsigned long count[MIGRATE_TYPES] = { 0, };
 290         int pageblock_mt, page_mt;
 291         int i;
 292
 293         /* Scan block by block. First and last block may be incomplete */
 294         pfn = zone->zone_start_pfn;
 295
 296         /*
 297          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 298          * a zone boundary, it will be double counted between zones. This does
 299          * not matter as the mixed block count will still be correct
 300          */
 301         for (; pfn < end_pfn; ) {
 302                 page = pfn_to_online_page(pfn);
 303                 if (!page) {
 304                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 305                         continue;
 306                 }
 307
 308                 block_end_pfn = pageblock_end_pfn(pfn);
 309                 block_end_pfn = min(block_end_pfn, end_pfn);
 310
 311                 pageblock_mt = get_pageblock_migratetype(page);
 312
 313                 for (; pfn < block_end_pfn; pfn++) {
 314                         /* The pageblock is online, no need to recheck. */
 315                         page = pfn_to_page(pfn);
 316
 317                         if (page_zone(page) != zone)
 318                                 continue;
 319
 320                         if (PageBuddy(page)) {
 321                                 unsigned long freepage_order;
 322
 323                                 freepage_order = buddy_order_unsafe(page);
 324                                 if (freepage_order <= MAX_ORDER)
 325                                         pfn += (1UL << freepage_order) - 1;
 326                                 continue;
 327                         }
 328
 329                         if (PageReserved(page))
 330                                 continue;
 331
 332                         page_ext = page_ext_get(page);
 333                         if (unlikely(!page_ext))
 334                                 continue;
 335
 336                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 337                                 goto ext_put_continue;
 338
 339                         page_owner = get_page_owner(page_ext);
 340                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 341                         if (pageblock_mt != page_mt) {
 342                                 if (is_migrate_cma(pageblock_mt))
 343                                         count[MIGRATE_MOVABLE]++;
 344                                 else
 345                                         count[pageblock_mt]++;
 346
 347                                 pfn = block_end_pfn;
 348                                 page_ext_put(page_ext);
 349                                 break;
 350                         }
 351                         pfn += (1UL << page_owner->order) - 1;
 352 ext_put_continue:
 353                         page_ext_put(page_ext);
 354                 }
 355         }
 356
 357         /* Print counts */
 358         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 359         for (i = 0; i < MIGRATE_TYPES; i++)
 360                 seq_printf(m, "%12lu ", count[i]);
 361         seq_putc(m, '\n');
 362 }
 363
 364 /*
 365  * Looking for memcg information and print it out
 366  */
 367 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
 368                                          struct page *page)
 369 {
 370 #ifdef CONFIG_MEMCG
 371         unsigned long memcg_data;
 372         struct mem_cgroup *memcg;
 373         bool online;
 374         char name[80];
 375
 376         rcu_read_lock();
 377         memcg_data = READ_ONCE(page->memcg_data);
 378         if (!memcg_data)
 379                 goto out_unlock;
 380
 381         if (memcg_data & MEMCG_DATA_OBJCGS)
 382                 ret += scnprintf(kbuf + ret, count - ret,
 383                                 "Slab cache page\n");
 384
 385         memcg = page_memcg_check(page);
 386         if (!memcg)
 387                 goto out_unlock;
 388
 389         online = (memcg->css.flags & CSS_ONLINE);
 390         cgroup_name(memcg->css.cgroup, name, sizeof(name));
 391         ret += scnprintf(kbuf + ret, count - ret,
 392                         "Charged %sto %smemcg %s\n",
 393                         PageMemcgKmem(page) ? "(via objcg) " : "",
 394                         online ? "" : "offline ",
 395                         name);
 396 out_unlock:
 397         rcu_read_unlock();
 398 #endif /* CONFIG_MEMCG */
 399
 400         return ret;
 401 }
 402
 403 static ssize_t
 404 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 405                 struct page *page, struct page_owner *page_owner,
 406                 depot_stack_handle_t handle)
 407 {
 408         int ret, pageblock_mt, page_mt;
 409         char *kbuf;
 410
 411         count = min_t(size_t, count, PAGE_SIZE);
 412         kbuf = kmalloc(count, GFP_KERNEL);
 413         if (!kbuf)
 414                 return -ENOMEM;
 415
 416         ret = scnprintf(kbuf, count,
 417                         "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
 418                         page_owner->order, page_owner->gfp_mask,
 419                         &page_owner->gfp_mask, page_owner->pid,
 420                         page_owner->tgid, page_owner->comm,
 421                         page_owner->ts_nsec);
 422
 423         /* Print information relevant to grouping pages by mobility */
 424         pageblock_mt = get_pageblock_migratetype(page);
 425         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 426         ret += scnprintf(kbuf + ret, count - ret,
 427                         "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
 428                         pfn,
 429                         migratetype_names[page_mt],
 430                         pfn >> pageblock_order,
 431                         migratetype_names[pageblock_mt],
 432                         &page->flags);
 433
 434         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 435         if (ret >= count)
 436                 goto err;
 437
 438         if (page_owner->last_migrate_reason != -1) {
 439                 ret += scnprintf(kbuf + ret, count - ret,
 440                         "Page has been migrated, last migrate reason: %s\n",
 441                         migrate_reason_names[page_owner->last_migrate_reason]);
 442         }
 443
 444         ret = print_page_owner_memcg(kbuf, count, ret, page);
 445
 446         ret += snprintf(kbuf + ret, count - ret, "\n");
 447         if (ret >= count)
 448                 goto err;
 449
 450         if (copy_to_user(buf, kbuf, ret))
 451                 ret = -EFAULT;
 452
 453         kfree(kbuf);
 454         return ret;
 455
 456 err:
 457         kfree(kbuf);
 458         return -ENOMEM;
 459 }
 460
 461 void __dump_page_owner(const struct page *page)
 462 {
 463         struct page_ext *page_ext = page_ext_get((void *)page);
 464         struct page_owner *page_owner;
 465         depot_stack_handle_t handle;
 466         gfp_t gfp_mask;
 467         int mt;
 468
 469         if (unlikely(!page_ext)) {
 470                 pr_alert("There is not page extension available.\n");
 471                 return;
 472         }
 473
 474         page_owner = get_page_owner(page_ext);
 475         gfp_mask = page_owner->gfp_mask;
 476         mt = gfp_migratetype(gfp_mask);
 477
 478         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 479                 pr_alert("page_owner info is not present (never set?)\n");
 480                 page_ext_put(page_ext);
 481                 return;
 482         }
 483
 484         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 485                 pr_alert("page_owner tracks the page as allocated\n");
 486         else
 487                 pr_alert("page_owner tracks the page as freed\n");
 488
 489         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu, free_ts %llu\n",
 490                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 491                  page_owner->pid, page_owner->tgid, page_owner->comm,
 492                  page_owner->ts_nsec, page_owner->free_ts_nsec);
 493
 494         handle = READ_ONCE(page_owner->handle);
 495         if (!handle)
 496                 pr_alert("page_owner allocation stack trace missing\n");
 497         else
 498                 stack_depot_print(handle);
 499
 500         handle = READ_ONCE(page_owner->free_handle);
 501         if (!handle) {
 502                 pr_alert("page_owner free stack trace missing\n");
 503         } else {
 504                 pr_alert("page last free pid %d tgid %d stack trace:\n",
 505                           page_owner->free_pid, page_owner->free_tgid);
 506                 stack_depot_print(handle);
 507         }
 508
 509         if (page_owner->last_migrate_reason != -1)
 510                 pr_alert("page has been migrated, last migrate reason: %s\n",
 511                         migrate_reason_names[page_owner->last_migrate_reason]);
 512         page_ext_put(page_ext);
 513 }
 514
 515 static ssize_t
 516 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 517 {
 518         unsigned long pfn;
 519         struct page *page;
 520         struct page_ext *page_ext;
 521         struct page_owner *page_owner;
 522         depot_stack_handle_t handle;
 523
 524         if (!static_branch_unlikely(&page_owner_inited))
 525                 return -EINVAL;
 526
 527         page = NULL;
 528         if (*ppos == 0)
 529                 pfn = min_low_pfn;
 530         else
 531                 pfn = *ppos;
 532         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 533         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 534                 pfn++;
 535
 536         /* Find an allocated page */
 537         for (; pfn < max_pfn; pfn++) {
 538                 /*
 539                  * This temporary page_owner is required so
 540                  * that we can avoid the context switches while holding
 541                  * the rcu lock and copying the page owner information to
 542                  * user through copy_to_user() or GFP_KERNEL allocations.
 543                  */
 544                 struct page_owner page_owner_tmp;
 545
 546                 /*
 547                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 548                  * validate the area as existing, skip it if not
 549                  */
 550                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 551                         pfn += MAX_ORDER_NR_PAGES - 1;
 552                         continue;
 553                 }
 554
 555                 page = pfn_to_page(pfn);
 556                 if (PageBuddy(page)) {
 557                         unsigned long freepage_order = buddy_order_unsafe(page);
 558
 559                         if (freepage_order <= MAX_ORDER)
 560                                 pfn += (1UL << freepage_order) - 1;
 561                         continue;
 562                 }
 563
 564                 page_ext = page_ext_get(page);
 565                 if (unlikely(!page_ext))
 566                         continue;
 567
 568                 /*
 569                  * Some pages could be missed by concurrent allocation or free,
 570                  * because we don't hold the zone lock.
 571                  */
 572                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 573                         goto ext_put_continue;
 574
 575                 /*
 576                  * Although we do have the info about past allocation of free
 577                  * pages, it's not relevant for current memory usage.
 578                  */
 579                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 580                         goto ext_put_continue;
 581
 582                 page_owner = get_page_owner(page_ext);
 583
 584                 /*
 585                  * Don't print "tail" pages of high-order allocations as that
 586                  * would inflate the stats.
 587                  */
 588                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 589                         goto ext_put_continue;
 590
 591                 /*
 592                  * Access to page_ext->handle isn't synchronous so we should
 593                  * be careful to access it.
 594                  */
 595                 handle = READ_ONCE(page_owner->handle);
 596                 if (!handle)
 597                         goto ext_put_continue;
 598
 599                 /* Record the next PFN to read in the file offset */
 600                 *ppos = pfn + 1;
 601
 602                 page_owner_tmp = *page_owner;
 603                 page_ext_put(page_ext);
 604                 return print_page_owner(buf, count, pfn, page,
 605                                 &page_owner_tmp, handle);
 606 ext_put_continue:
 607                 page_ext_put(page_ext);
 608         }
 609
 610         return 0;
 611 }
 612
 613 static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
 614 {
 615         switch (orig) {
 616         case SEEK_SET:
 617                 file->f_pos = offset;
 618                 break;
 619         case SEEK_CUR:
 620                 file->f_pos += offset;
 621                 break;
 622         default:
 623                 return -EINVAL;
 624         }
 625         return file->f_pos;
 626 }
 627
 628 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 629 {
 630         unsigned long pfn = zone->zone_start_pfn;
 631         unsigned long end_pfn = zone_end_pfn(zone);
 632         unsigned long count = 0;
 633
 634         /*
 635          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 636          * a zone boundary, it will be double counted between zones. This does
 637          * not matter as the mixed block count will still be correct
 638          */
 639         for (; pfn < end_pfn; ) {
 640                 unsigned long block_end_pfn;
 641
 642                 if (!pfn_valid(pfn)) {
 643                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 644                         continue;
 645                 }
 646
 647                 block_end_pfn = pageblock_end_pfn(pfn);
 648                 block_end_pfn = min(block_end_pfn, end_pfn);
 649
 650                 for (; pfn < block_end_pfn; pfn++) {
 651                         struct page *page = pfn_to_page(pfn);
 652                         struct page_ext *page_ext;
 653
 654                         if (page_zone(page) != zone)
 655                                 continue;
 656
 657                         /*
 658                          * To avoid having to grab zone->lock, be a little
 659                          * careful when reading buddy page order. The only
 660                          * danger is that we skip too much and potentially miss
 661                          * some early allocated pages, which is better than
 662                          * heavy lock contention.
 663                          */
 664                         if (PageBuddy(page)) {
 665                                 unsigned long order = buddy_order_unsafe(page);
 666
 667                                 if (order > 0 && order <= MAX_ORDER)
 668                                         pfn += (1UL << order) - 1;
 669                                 continue;
 670                         }
 671
 672                         if (PageReserved(page))
 673                                 continue;
 674
 675                         page_ext = page_ext_get(page);
 676                         if (unlikely(!page_ext))
 677                                 continue;
 678
 679                         /* Maybe overlapping zone */
 680                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 681                                 goto ext_put_continue;
 682
 683                         /* Found early allocated page */
 684                         __set_page_owner_handle(page_ext, early_handle,
 685                                                 0, 0);
 686                         count++;
 687 ext_put_continue:
 688                         page_ext_put(page_ext);
 689                 }
 690                 cond_resched();
 691         }
 692
 693         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 694                 pgdat->node_id, zone->name, count);
 695 }
 696
 697 static void init_zones_in_node(pg_data_t *pgdat)
 698 {
 699         struct zone *zone;
 700         struct zone *node_zones = pgdat->node_zones;
 701
 702         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 703                 if (!populated_zone(zone))
 704                         continue;
 705
 706                 init_pages_in_zone(pgdat, zone);
 707         }
 708 }
 709
 710 static void init_early_allocated_pages(void)
 711 {
 712         pg_data_t *pgdat;
 713
 714         for_each_online_pgdat(pgdat)
 715                 init_zones_in_node(pgdat);
 716 }
 717
 718 static const struct file_operations proc_page_owner_operations = {
 719         .read           = read_page_owner,
 720         .llseek         = lseek_page_owner,
 721 };
 722
 723 static int __init pageowner_init(void)
 724 {
 725         if (!static_branch_unlikely(&page_owner_inited)) {
 726                 pr_info("page_owner is disabled\n");
 727                 return 0;
 728         }
 729
 730         debugfs_create_file("page_owner", 0400, NULL, NULL,
 731                             &proc_page_owner_operations);
 732
 733         return 0;
 734 }
 735 late_initcall(pageowner_init)