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 struct stack {
  40         struct stack_record *stack_record;
  41         struct stack *next;
  42 };
  43 static struct stack dummy_stack;
  44 static struct stack failure_stack;
  45 static struct stack *stack_list;
  46 static DEFINE_SPINLOCK(stack_list_lock);
  47
  48 static bool page_owner_enabled __initdata;
  49 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  50
  51 static depot_stack_handle_t dummy_handle;
  52 static depot_stack_handle_t failure_handle;
  53 static depot_stack_handle_t early_handle;
  54
  55 static void init_early_allocated_pages(void);
  56
  57 static inline void set_current_in_page_owner(void)
  58 {
  59         /*
  60          * Avoid recursion.
  61          *
  62          * We might need to allocate more memory from page_owner code, so make
  63          * sure to signal it in order to avoid recursion.
  64          */
  65         current->in_page_owner = 1;
  66 }
  67
  68 static inline void unset_current_in_page_owner(void)
  69 {
  70         current->in_page_owner = 0;
  71 }
  72
  73 static int __init early_page_owner_param(char *buf)
  74 {
  75         int ret = kstrtobool(buf, &page_owner_enabled);
  76
  77         if (page_owner_enabled)
  78                 stack_depot_request_early_init();
  79
  80         return ret;
  81 }
  82 early_param("page_owner", early_page_owner_param);
  83
  84 static __init bool need_page_owner(void)
  85 {
  86         return page_owner_enabled;
  87 }
  88
  89 static __always_inline depot_stack_handle_t create_dummy_stack(void)
  90 {
  91         unsigned long entries[4];
  92         unsigned int nr_entries;
  93
  94         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  95         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
  96 }
  97
  98 static noinline void register_dummy_stack(void)
  99 {
 100         dummy_handle = create_dummy_stack();
 101 }
 102
 103 static noinline void register_failure_stack(void)
 104 {
 105         failure_handle = create_dummy_stack();
 106 }
 107
 108 static noinline void register_early_stack(void)
 109 {
 110         early_handle = create_dummy_stack();
 111 }
 112
 113 static __init void init_page_owner(void)
 114 {
 115         if (!page_owner_enabled)
 116                 return;
 117
 118         register_dummy_stack();
 119         register_failure_stack();
 120         register_early_stack();
 121         static_branch_enable(&page_owner_inited);
 122         init_early_allocated_pages();
 123         /* Initialize dummy and failure stacks and link them to stack_list */
 124         dummy_stack.stack_record = __stack_depot_get_stack_record(dummy_handle);
 125         failure_stack.stack_record = __stack_depot_get_stack_record(failure_handle);
 126         if (dummy_stack.stack_record)
 127                 refcount_set(&dummy_stack.stack_record->count, 1);
 128         if (failure_stack.stack_record)
 129                 refcount_set(&failure_stack.stack_record->count, 1);
 130         dummy_stack.next = &failure_stack;
 131         stack_list = &dummy_stack;
 132 }
 133
 134 struct page_ext_operations page_owner_ops = {
 135         .size = sizeof(struct page_owner),
 136         .need = need_page_owner,
 137         .init = init_page_owner,
 138         .need_shared_flags = true,
 139 };
 140
 141 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
 142 {
 143         return page_ext_data(page_ext, &page_owner_ops);
 144 }
 145
 146 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 147 {
 148         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 149         depot_stack_handle_t handle;
 150         unsigned int nr_entries;
 151
 152         if (current->in_page_owner)
 153                 return dummy_handle;
 154
 155         set_current_in_page_owner();
 156         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 157         handle = stack_depot_save(entries, nr_entries, flags);
 158         if (!handle)
 159                 handle = failure_handle;
 160         unset_current_in_page_owner();
 161
 162         return handle;
 163 }
 164
 165 static void add_stack_record_to_list(struct stack_record *stack_record,
 166                                      gfp_t gfp_mask)
 167 {
 168         unsigned long flags;
 169         struct stack *stack;
 170
 171         /* Filter gfp_mask the same way stackdepot does, for consistency */
 172         gfp_mask &= ~GFP_ZONEMASK;
 173         gfp_mask &= (GFP_ATOMIC | GFP_KERNEL);
 174         gfp_mask |= __GFP_NOWARN;
 175
 176         set_current_in_page_owner();
 177         stack = kmalloc(sizeof(*stack), gfp_mask);
 178         if (!stack) {
 179                 unset_current_in_page_owner();
 180                 return;
 181         }
 182         unset_current_in_page_owner();
 183
 184         stack->stack_record = stack_record;
 185         stack->next = NULL;
 186
 187         spin_lock_irqsave(&stack_list_lock, flags);
 188         stack->next = stack_list;
 189         /*
 190          * This pairs with smp_load_acquire() from function
 191          * stack_start(). This guarantees that stack_start()
 192          * will see an updated stack_list before starting to
 193          * traverse the list.
 194          */
 195         smp_store_release(&stack_list, stack);
 196         spin_unlock_irqrestore(&stack_list_lock, flags);
 197 }
 198
 199 static void inc_stack_record_count(depot_stack_handle_t handle, gfp_t gfp_mask)
 200 {
 201         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 202
 203         if (!stack_record)
 204                 return;
 205
 206         /*
 207          * New stack_record's that do not use STACK_DEPOT_FLAG_GET start
 208          * with REFCOUNT_SATURATED to catch spurious increments of their
 209          * refcount.
 210          * Since we do not use STACK_DEPOT_FLAG_GET API, let us
 211          * set a refcount of 1 ourselves.
 212          */
 213         if (refcount_read(&stack_record->count) == REFCOUNT_SATURATED) {
 214                 int old = REFCOUNT_SATURATED;
 215
 216                 if (atomic_try_cmpxchg_relaxed(&stack_record->count.refs, &old, 1))
 217                         /* Add the new stack_record to our list */
 218                         add_stack_record_to_list(stack_record, gfp_mask);
 219         }
 220         refcount_inc(&stack_record->count);
 221 }
 222
 223 static void dec_stack_record_count(depot_stack_handle_t handle)
 224 {
 225         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 226
 227         if (stack_record)
 228                 refcount_dec(&stack_record->count);
 229 }
 230
 231 void __reset_page_owner(struct page *page, unsigned short order)
 232 {
 233         int i;
 234         struct page_ext *page_ext;
 235         depot_stack_handle_t handle;
 236         depot_stack_handle_t alloc_handle;
 237         struct page_owner *page_owner;
 238         u64 free_ts_nsec = local_clock();
 239
 240         page_ext = page_ext_get(page);
 241         if (unlikely(!page_ext))
 242                 return;
 243
 244         page_owner = get_page_owner(page_ext);
 245         alloc_handle = page_owner->handle;
 246
 247         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 248         for (i = 0; i < (1 << order); i++) {
 249                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 250                 page_owner->free_handle = handle;
 251                 page_owner->free_ts_nsec = free_ts_nsec;
 252                 page_owner->free_pid = current->pid;
 253                 page_owner->free_tgid = current->tgid;
 254                 page_ext = page_ext_next(page_ext);
 255                 page_owner = get_page_owner(page_ext);
 256         }
 257         page_ext_put(page_ext);
 258         if (alloc_handle != early_handle)
 259                 /*
 260                  * early_handle is being set as a handle for all those
 261                  * early allocated pages. See init_pages_in_zone().
 262                  * Since their refcount is not being incremented because
 263                  * the machinery is not ready yet, we cannot decrement
 264                  * their refcount either.
 265                  */
 266                 dec_stack_record_count(alloc_handle);
 267 }
 268
 269 static inline void __set_page_owner_handle(struct page_ext *page_ext,
 270                                         depot_stack_handle_t handle,
 271                                         unsigned short order, gfp_t gfp_mask)
 272 {
 273         struct page_owner *page_owner;
 274         int i;
 275         u64 ts_nsec = local_clock();
 276
 277         for (i = 0; i < (1 << order); i++) {
 278                 page_owner = get_page_owner(page_ext);
 279                 page_owner->handle = handle;
 280                 page_owner->order = order;
 281                 page_owner->gfp_mask = gfp_mask;
 282                 page_owner->last_migrate_reason = -1;
 283                 page_owner->pid = current->pid;
 284                 page_owner->tgid = current->tgid;
 285                 page_owner->ts_nsec = ts_nsec;
 286                 strscpy(page_owner->comm, current->comm,
 287                         sizeof(page_owner->comm));
 288                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 289                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 290
 291                 page_ext = page_ext_next(page_ext);
 292         }
 293 }
 294
 295 noinline void __set_page_owner(struct page *page, unsigned short order,
 296                                         gfp_t gfp_mask)
 297 {
 298         struct page_ext *page_ext;
 299         depot_stack_handle_t handle;
 300
 301         handle = save_stack(gfp_mask);
 302
 303         page_ext = page_ext_get(page);
 304         if (unlikely(!page_ext))
 305                 return;
 306         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 307         page_ext_put(page_ext);
 308         inc_stack_record_count(handle, gfp_mask);
 309 }
 310
 311 void __set_page_owner_migrate_reason(struct page *page, int reason)
 312 {
 313         struct page_ext *page_ext = page_ext_get(page);
 314         struct page_owner *page_owner;
 315
 316         if (unlikely(!page_ext))
 317                 return;
 318
 319         page_owner = get_page_owner(page_ext);
 320         page_owner->last_migrate_reason = reason;
 321         page_ext_put(page_ext);
 322 }
 323
 324 void __split_page_owner(struct page *page, int old_order, int new_order)
 325 {
 326         int i;
 327         struct page_ext *page_ext = page_ext_get(page);
 328         struct page_owner *page_owner;
 329
 330         if (unlikely(!page_ext))
 331                 return;
 332
 333         for (i = 0; i < (1 << old_order); i++) {
 334                 page_owner = get_page_owner(page_ext);
 335                 page_owner->order = new_order;
 336                 page_ext = page_ext_next(page_ext);
 337         }
 338         page_ext_put(page_ext);
 339 }
 340
 341 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 342 {
 343         struct page_ext *old_ext;
 344         struct page_ext *new_ext;
 345         struct page_owner *old_page_owner, *new_page_owner;
 346
 347         old_ext = page_ext_get(&old->page);
 348         if (unlikely(!old_ext))
 349                 return;
 350
 351         new_ext = page_ext_get(&newfolio->page);
 352         if (unlikely(!new_ext)) {
 353                 page_ext_put(old_ext);
 354                 return;
 355         }
 356
 357         old_page_owner = get_page_owner(old_ext);
 358         new_page_owner = get_page_owner(new_ext);
 359         new_page_owner->order = old_page_owner->order;
 360         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 361         new_page_owner->last_migrate_reason =
 362                 old_page_owner->last_migrate_reason;
 363         new_page_owner->handle = old_page_owner->handle;
 364         new_page_owner->pid = old_page_owner->pid;
 365         new_page_owner->tgid = old_page_owner->tgid;
 366         new_page_owner->free_pid = old_page_owner->free_pid;
 367         new_page_owner->free_tgid = old_page_owner->free_tgid;
 368         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
 369         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
 370         strcpy(new_page_owner->comm, old_page_owner->comm);
 371
 372         /*
 373          * We don't clear the bit on the old folio as it's going to be freed
 374          * after migration. Until then, the info can be useful in case of
 375          * a bug, and the overall stats will be off a bit only temporarily.
 376          * Also, migrate_misplaced_transhuge_page() can still fail the
 377          * migration and then we want the old folio to retain the info. But
 378          * in that case we also don't need to explicitly clear the info from
 379          * the new page, which will be freed.
 380          */
 381         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 382         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
 383         page_ext_put(new_ext);
 384         page_ext_put(old_ext);
 385 }
 386
 387 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 388                                        pg_data_t *pgdat, struct zone *zone)
 389 {
 390         struct page *page;
 391         struct page_ext *page_ext;
 392         struct page_owner *page_owner;
 393         unsigned long pfn, block_end_pfn;
 394         unsigned long end_pfn = zone_end_pfn(zone);
 395         unsigned long count[MIGRATE_TYPES] = { 0, };
 396         int pageblock_mt, page_mt;
 397         int i;
 398
 399         /* Scan block by block. First and last block may be incomplete */
 400         pfn = zone->zone_start_pfn;
 401
 402         /*
 403          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 404          * a zone boundary, it will be double counted between zones. This does
 405          * not matter as the mixed block count will still be correct
 406          */
 407         for (; pfn < end_pfn; ) {
 408                 page = pfn_to_online_page(pfn);
 409                 if (!page) {
 410                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 411                         continue;
 412                 }
 413
 414                 block_end_pfn = pageblock_end_pfn(pfn);
 415                 block_end_pfn = min(block_end_pfn, end_pfn);
 416
 417                 pageblock_mt = get_pageblock_migratetype(page);
 418
 419                 for (; pfn < block_end_pfn; pfn++) {
 420                         /* The pageblock is online, no need to recheck. */
 421                         page = pfn_to_page(pfn);
 422
 423                         if (page_zone(page) != zone)
 424                                 continue;
 425
 426                         if (PageBuddy(page)) {
 427                                 unsigned long freepage_order;
 428
 429                                 freepage_order = buddy_order_unsafe(page);
 430                                 if (freepage_order <= MAX_PAGE_ORDER)
 431                                         pfn += (1UL << freepage_order) - 1;
 432                                 continue;
 433                         }
 434
 435                         if (PageReserved(page))
 436                                 continue;
 437
 438                         page_ext = page_ext_get(page);
 439                         if (unlikely(!page_ext))
 440                                 continue;
 441
 442                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 443                                 goto ext_put_continue;
 444
 445                         page_owner = get_page_owner(page_ext);
 446                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 447                         if (pageblock_mt != page_mt) {
 448                                 if (is_migrate_cma(pageblock_mt))
 449                                         count[MIGRATE_MOVABLE]++;
 450                                 else
 451                                         count[pageblock_mt]++;
 452
 453                                 pfn = block_end_pfn;
 454                                 page_ext_put(page_ext);
 455                                 break;
 456                         }
 457                         pfn += (1UL << page_owner->order) - 1;
 458 ext_put_continue:
 459                         page_ext_put(page_ext);
 460                 }
 461         }
 462
 463         /* Print counts */
 464         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 465         for (i = 0; i < MIGRATE_TYPES; i++)
 466                 seq_printf(m, "%12lu ", count[i]);
 467         seq_putc(m, '\n');
 468 }
 469
 470 /*
 471  * Looking for memcg information and print it out
 472  */
 473 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
 474                                          struct page *page)
 475 {
 476 #ifdef CONFIG_MEMCG
 477         unsigned long memcg_data;
 478         struct mem_cgroup *memcg;
 479         bool online;
 480         char name[80];
 481
 482         rcu_read_lock();
 483         memcg_data = READ_ONCE(page->memcg_data);
 484         if (!memcg_data)
 485                 goto out_unlock;
 486
 487         if (memcg_data & MEMCG_DATA_OBJCGS)
 488                 ret += scnprintf(kbuf + ret, count - ret,
 489                                 "Slab cache page\n");
 490
 491         memcg = page_memcg_check(page);
 492         if (!memcg)
 493                 goto out_unlock;
 494
 495         online = (memcg->css.flags & CSS_ONLINE);
 496         cgroup_name(memcg->css.cgroup, name, sizeof(name));
 497         ret += scnprintf(kbuf + ret, count - ret,
 498                         "Charged %sto %smemcg %s\n",
 499                         PageMemcgKmem(page) ? "(via objcg) " : "",
 500                         online ? "" : "offline ",
 501                         name);
 502 out_unlock:
 503         rcu_read_unlock();
 504 #endif /* CONFIG_MEMCG */
 505
 506         return ret;
 507 }
 508
 509 static ssize_t
 510 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 511                 struct page *page, struct page_owner *page_owner,
 512                 depot_stack_handle_t handle)
 513 {
 514         int ret, pageblock_mt, page_mt;
 515         char *kbuf;
 516
 517         count = min_t(size_t, count, PAGE_SIZE);
 518         kbuf = kmalloc(count, GFP_KERNEL);
 519         if (!kbuf)
 520                 return -ENOMEM;
 521
 522         ret = scnprintf(kbuf, count,
 523                         "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
 524                         page_owner->order, page_owner->gfp_mask,
 525                         &page_owner->gfp_mask, page_owner->pid,
 526                         page_owner->tgid, page_owner->comm,
 527                         page_owner->ts_nsec);
 528
 529         /* Print information relevant to grouping pages by mobility */
 530         pageblock_mt = get_pageblock_migratetype(page);
 531         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 532         ret += scnprintf(kbuf + ret, count - ret,
 533                         "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
 534                         pfn,
 535                         migratetype_names[page_mt],
 536                         pfn >> pageblock_order,
 537                         migratetype_names[pageblock_mt],
 538                         &page->flags);
 539
 540         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 541         if (ret >= count)
 542                 goto err;
 543
 544         if (page_owner->last_migrate_reason != -1) {
 545                 ret += scnprintf(kbuf + ret, count - ret,
 546                         "Page has been migrated, last migrate reason: %s\n",
 547                         migrate_reason_names[page_owner->last_migrate_reason]);
 548         }
 549
 550         ret = print_page_owner_memcg(kbuf, count, ret, page);
 551
 552         ret += snprintf(kbuf + ret, count - ret, "\n");
 553         if (ret >= count)
 554                 goto err;
 555
 556         if (copy_to_user(buf, kbuf, ret))
 557                 ret = -EFAULT;
 558
 559         kfree(kbuf);
 560         return ret;
 561
 562 err:
 563         kfree(kbuf);
 564         return -ENOMEM;
 565 }
 566
 567 void __dump_page_owner(const struct page *page)
 568 {
 569         struct page_ext *page_ext = page_ext_get((void *)page);
 570         struct page_owner *page_owner;
 571         depot_stack_handle_t handle;
 572         gfp_t gfp_mask;
 573         int mt;
 574
 575         if (unlikely(!page_ext)) {
 576                 pr_alert("There is not page extension available.\n");
 577                 return;
 578         }
 579
 580         page_owner = get_page_owner(page_ext);
 581         gfp_mask = page_owner->gfp_mask;
 582         mt = gfp_migratetype(gfp_mask);
 583
 584         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 585                 pr_alert("page_owner info is not present (never set?)\n");
 586                 page_ext_put(page_ext);
 587                 return;
 588         }
 589
 590         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 591                 pr_alert("page_owner tracks the page as allocated\n");
 592         else
 593                 pr_alert("page_owner tracks the page as freed\n");
 594
 595         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",
 596                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 597                  page_owner->pid, page_owner->tgid, page_owner->comm,
 598                  page_owner->ts_nsec, page_owner->free_ts_nsec);
 599
 600         handle = READ_ONCE(page_owner->handle);
 601         if (!handle)
 602                 pr_alert("page_owner allocation stack trace missing\n");
 603         else
 604                 stack_depot_print(handle);
 605
 606         handle = READ_ONCE(page_owner->free_handle);
 607         if (!handle) {
 608                 pr_alert("page_owner free stack trace missing\n");
 609         } else {
 610                 pr_alert("page last free pid %d tgid %d stack trace:\n",
 611                           page_owner->free_pid, page_owner->free_tgid);
 612                 stack_depot_print(handle);
 613         }
 614
 615         if (page_owner->last_migrate_reason != -1)
 616                 pr_alert("page has been migrated, last migrate reason: %s\n",
 617                         migrate_reason_names[page_owner->last_migrate_reason]);
 618         page_ext_put(page_ext);
 619 }
 620
 621 static ssize_t
 622 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 623 {
 624         unsigned long pfn;
 625         struct page *page;
 626         struct page_ext *page_ext;
 627         struct page_owner *page_owner;
 628         depot_stack_handle_t handle;
 629
 630         if (!static_branch_unlikely(&page_owner_inited))
 631                 return -EINVAL;
 632
 633         page = NULL;
 634         if (*ppos == 0)
 635                 pfn = min_low_pfn;
 636         else
 637                 pfn = *ppos;
 638         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 639         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 640                 pfn++;
 641
 642         /* Find an allocated page */
 643         for (; pfn < max_pfn; pfn++) {
 644                 /*
 645                  * This temporary page_owner is required so
 646                  * that we can avoid the context switches while holding
 647                  * the rcu lock and copying the page owner information to
 648                  * user through copy_to_user() or GFP_KERNEL allocations.
 649                  */
 650                 struct page_owner page_owner_tmp;
 651
 652                 /*
 653                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 654                  * validate the area as existing, skip it if not
 655                  */
 656                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 657                         pfn += MAX_ORDER_NR_PAGES - 1;
 658                         continue;
 659                 }
 660
 661                 page = pfn_to_page(pfn);
 662                 if (PageBuddy(page)) {
 663                         unsigned long freepage_order = buddy_order_unsafe(page);
 664
 665                         if (freepage_order <= MAX_PAGE_ORDER)
 666                                 pfn += (1UL << freepage_order) - 1;
 667                         continue;
 668                 }
 669
 670                 page_ext = page_ext_get(page);
 671                 if (unlikely(!page_ext))
 672                         continue;
 673
 674                 /*
 675                  * Some pages could be missed by concurrent allocation or free,
 676                  * because we don't hold the zone lock.
 677                  */
 678                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 679                         goto ext_put_continue;
 680
 681                 /*
 682                  * Although we do have the info about past allocation of free
 683                  * pages, it's not relevant for current memory usage.
 684                  */
 685                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 686                         goto ext_put_continue;
 687
 688                 page_owner = get_page_owner(page_ext);
 689
 690                 /*
 691                  * Don't print "tail" pages of high-order allocations as that
 692                  * would inflate the stats.
 693                  */
 694                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 695                         goto ext_put_continue;
 696
 697                 /*
 698                  * Access to page_ext->handle isn't synchronous so we should
 699                  * be careful to access it.
 700                  */
 701                 handle = READ_ONCE(page_owner->handle);
 702                 if (!handle)
 703                         goto ext_put_continue;
 704
 705                 /* Record the next PFN to read in the file offset */
 706                 *ppos = pfn + 1;
 707
 708                 page_owner_tmp = *page_owner;
 709                 page_ext_put(page_ext);
 710                 return print_page_owner(buf, count, pfn, page,
 711                                 &page_owner_tmp, handle);
 712 ext_put_continue:
 713                 page_ext_put(page_ext);
 714         }
 715
 716         return 0;
 717 }
 718
 719 static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
 720 {
 721         switch (orig) {
 722         case SEEK_SET:
 723                 file->f_pos = offset;
 724                 break;
 725         case SEEK_CUR:
 726                 file->f_pos += offset;
 727                 break;
 728         default:
 729                 return -EINVAL;
 730         }
 731         return file->f_pos;
 732 }
 733
 734 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 735 {
 736         unsigned long pfn = zone->zone_start_pfn;
 737         unsigned long end_pfn = zone_end_pfn(zone);
 738         unsigned long count = 0;
 739
 740         /*
 741          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 742          * a zone boundary, it will be double counted between zones. This does
 743          * not matter as the mixed block count will still be correct
 744          */
 745         for (; pfn < end_pfn; ) {
 746                 unsigned long block_end_pfn;
 747
 748                 if (!pfn_valid(pfn)) {
 749                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 750                         continue;
 751                 }
 752
 753                 block_end_pfn = pageblock_end_pfn(pfn);
 754                 block_end_pfn = min(block_end_pfn, end_pfn);
 755
 756                 for (; pfn < block_end_pfn; pfn++) {
 757                         struct page *page = pfn_to_page(pfn);
 758                         struct page_ext *page_ext;
 759
 760                         if (page_zone(page) != zone)
 761                                 continue;
 762
 763                         /*
 764                          * To avoid having to grab zone->lock, be a little
 765                          * careful when reading buddy page order. The only
 766                          * danger is that we skip too much and potentially miss
 767                          * some early allocated pages, which is better than
 768                          * heavy lock contention.
 769                          */
 770                         if (PageBuddy(page)) {
 771                                 unsigned long order = buddy_order_unsafe(page);
 772
 773                                 if (order > 0 && order <= MAX_PAGE_ORDER)
 774                                         pfn += (1UL << order) - 1;
 775                                 continue;
 776                         }
 777
 778                         if (PageReserved(page))
 779                                 continue;
 780
 781                         page_ext = page_ext_get(page);
 782                         if (unlikely(!page_ext))
 783                                 continue;
 784
 785                         /* Maybe overlapping zone */
 786                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 787                                 goto ext_put_continue;
 788
 789                         /* Found early allocated page */
 790                         __set_page_owner_handle(page_ext, early_handle,
 791                                                 0, 0);
 792                         count++;
 793 ext_put_continue:
 794                         page_ext_put(page_ext);
 795                 }
 796                 cond_resched();
 797         }
 798
 799         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 800                 pgdat->node_id, zone->name, count);
 801 }
 802
 803 static void init_zones_in_node(pg_data_t *pgdat)
 804 {
 805         struct zone *zone;
 806         struct zone *node_zones = pgdat->node_zones;
 807
 808         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 809                 if (!populated_zone(zone))
 810                         continue;
 811
 812                 init_pages_in_zone(pgdat, zone);
 813         }
 814 }
 815
 816 static void init_early_allocated_pages(void)
 817 {
 818         pg_data_t *pgdat;
 819
 820         for_each_online_pgdat(pgdat)
 821                 init_zones_in_node(pgdat);
 822 }
 823
 824 static const struct file_operations proc_page_owner_operations = {
 825         .read           = read_page_owner,
 826         .llseek         = lseek_page_owner,
 827 };
 828
 829 static void *stack_start(struct seq_file *m, loff_t *ppos)
 830 {
 831         struct stack *stack;
 832
 833         if (*ppos == -1UL)
 834                 return NULL;
 835
 836         if (!*ppos) {
 837                 /*
 838                  * This pairs with smp_store_release() from function
 839                  * add_stack_record_to_list(), so we get a consistent
 840                  * value of stack_list.
 841                  */
 842                 stack = smp_load_acquire(&stack_list);
 843         } else {
 844                 stack = m->private;
 845                 stack = stack->next;
 846         }
 847
 848         m->private = stack;
 849
 850         return stack;
 851 }
 852
 853 static void *stack_next(struct seq_file *m, void *v, loff_t *ppos)
 854 {
 855         struct stack *stack = v;
 856
 857         stack = stack->next;
 858         *ppos = stack ? *ppos + 1 : -1UL;
 859         m->private = stack;
 860
 861         return stack;
 862 }
 863
 864 static unsigned long page_owner_stack_threshold;
 865
 866 static int stack_print(struct seq_file *m, void *v)
 867 {
 868         int i, stack_count;
 869         struct stack *stack = v;
 870         unsigned long *entries;
 871         unsigned long nr_entries;
 872         struct stack_record *stack_record = stack->stack_record;
 873
 874         if (!stack->stack_record)
 875                 return 0;
 876
 877         nr_entries = stack_record->size;
 878         entries = stack_record->entries;
 879         stack_count = refcount_read(&stack_record->count) - 1;
 880
 881         if (stack_count < 1 || stack_count < page_owner_stack_threshold)
 882                 return 0;
 883
 884         for (i = 0; i < nr_entries; i++)
 885                 seq_printf(m, " %pS\n", (void *)entries[i]);
 886         seq_printf(m, "stack_count: %d\n\n", stack_count);
 887
 888         return 0;
 889 }
 890
 891 static void stack_stop(struct seq_file *m, void *v)
 892 {
 893 }
 894
 895 static const struct seq_operations page_owner_stack_op = {
 896         .start  = stack_start,
 897         .next   = stack_next,
 898         .stop   = stack_stop,
 899         .show   = stack_print
 900 };
 901
 902 static int page_owner_stack_open(struct inode *inode, struct file *file)
 903 {
 904         return seq_open_private(file, &page_owner_stack_op, 0);
 905 }
 906
 907 static const struct file_operations page_owner_stack_operations = {
 908         .open           = page_owner_stack_open,
 909         .read           = seq_read,
 910         .llseek         = seq_lseek,
 911         .release        = seq_release,
 912 };
 913
 914 static int page_owner_threshold_get(void *data, u64 *val)
 915 {
 916         *val = READ_ONCE(page_owner_stack_threshold);
 917         return 0;
 918 }
 919
 920 static int page_owner_threshold_set(void *data, u64 val)
 921 {
 922         WRITE_ONCE(page_owner_stack_threshold, val);
 923         return 0;
 924 }
 925
 926 DEFINE_SIMPLE_ATTRIBUTE(proc_page_owner_threshold, &page_owner_threshold_get,
 927                         &page_owner_threshold_set, "%llu");
 928
 929
 930 static int __init pageowner_init(void)
 931 {
 932         struct dentry *dir;
 933
 934         if (!static_branch_unlikely(&page_owner_inited)) {
 935                 pr_info("page_owner is disabled\n");
 936                 return 0;
 937         }
 938
 939         debugfs_create_file("page_owner", 0400, NULL, NULL,
 940                             &proc_page_owner_operations);
 941         dir = debugfs_create_dir("page_owner_stacks", NULL);
 942         debugfs_create_file("show_stacks", 0400, dir, NULL,
 943                             &page_owner_stack_operations);
 944         debugfs_create_file("count_threshold", 0600, dir, NULL,
 945                             &proc_page_owner_threshold);
 946
 947         return 0;
 948 }
 949 late_initcall(pageowner_init)