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         init_early_allocated_pages();
 122         /* Initialize dummy and failure stacks and link them to stack_list */
 123         dummy_stack.stack_record = __stack_depot_get_stack_record(dummy_handle);
 124         failure_stack.stack_record = __stack_depot_get_stack_record(failure_handle);
 125         if (dummy_stack.stack_record)
 126                 refcount_set(&dummy_stack.stack_record->count, 1);
 127         if (failure_stack.stack_record)
 128                 refcount_set(&failure_stack.stack_record->count, 1);
 129         dummy_stack.next = &failure_stack;
 130         stack_list = &dummy_stack;
 131         static_branch_enable(&page_owner_inited);
 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         set_current_in_page_owner();
 172         stack = kmalloc(sizeof(*stack), gfp_nested_mask(gfp_mask));
 173         if (!stack) {
 174                 unset_current_in_page_owner();
 175                 return;
 176         }
 177         unset_current_in_page_owner();
 178
 179         stack->stack_record = stack_record;
 180         stack->next = NULL;
 181
 182         spin_lock_irqsave(&stack_list_lock, flags);
 183         stack->next = stack_list;
 184         /*
 185          * This pairs with smp_load_acquire() from function
 186          * stack_start(). This guarantees that stack_start()
 187          * will see an updated stack_list before starting to
 188          * traverse the list.
 189          */
 190         smp_store_release(&stack_list, stack);
 191         spin_unlock_irqrestore(&stack_list_lock, flags);
 192 }
 193
 194 static void inc_stack_record_count(depot_stack_handle_t handle, gfp_t gfp_mask,
 195                                    int nr_base_pages)
 196 {
 197         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 198
 199         if (!stack_record)
 200                 return;
 201
 202         /*
 203          * New stack_record's that do not use STACK_DEPOT_FLAG_GET start
 204          * with REFCOUNT_SATURATED to catch spurious increments of their
 205          * refcount.
 206          * Since we do not use STACK_DEPOT_FLAG_GET API, let us
 207          * set a refcount of 1 ourselves.
 208          */
 209         if (refcount_read(&stack_record->count) == REFCOUNT_SATURATED) {
 210                 int old = REFCOUNT_SATURATED;
 211
 212                 if (atomic_try_cmpxchg_relaxed(&stack_record->count.refs, &old, 1))
 213                         /* Add the new stack_record to our list */
 214                         add_stack_record_to_list(stack_record, gfp_mask);
 215         }
 216         refcount_add(nr_base_pages, &stack_record->count);
 217 }
 218
 219 static void dec_stack_record_count(depot_stack_handle_t handle,
 220                                    int nr_base_pages)
 221 {
 222         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 223
 224         if (!stack_record)
 225                 return;
 226
 227         if (refcount_sub_and_test(nr_base_pages, &stack_record->count))
 228                 pr_warn("%s: refcount went to 0 for %u handle\n", __func__,
 229                         handle);
 230 }
 231
 232 static inline void __update_page_owner_handle(struct page_ext *page_ext,
 233                                               depot_stack_handle_t handle,
 234                                               unsigned short order,
 235                                               gfp_t gfp_mask,
 236                                               short last_migrate_reason, u64 ts_nsec,
 237                                               pid_t pid, pid_t tgid, char *comm)
 238 {
 239         int i;
 240         struct page_owner *page_owner;
 241
 242         for (i = 0; i < (1 << order); i++) {
 243                 page_owner = get_page_owner(page_ext);
 244                 page_owner->handle = handle;
 245                 page_owner->order = order;
 246                 page_owner->gfp_mask = gfp_mask;
 247                 page_owner->last_migrate_reason = last_migrate_reason;
 248                 page_owner->pid = pid;
 249                 page_owner->tgid = tgid;
 250                 page_owner->ts_nsec = ts_nsec;
 251                 strscpy(page_owner->comm, comm,
 252                         sizeof(page_owner->comm));
 253                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 254                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 255                 page_ext = page_ext_next(page_ext);
 256         }
 257 }
 258
 259 static inline void __update_page_owner_free_handle(struct page_ext *page_ext,
 260                                                    depot_stack_handle_t handle,
 261                                                    unsigned short order,
 262                                                    pid_t pid, pid_t tgid,
 263                                                    u64 free_ts_nsec)
 264 {
 265         int i;
 266         struct page_owner *page_owner;
 267
 268         for (i = 0; i < (1 << order); i++) {
 269                 page_owner = get_page_owner(page_ext);
 270                 /* Only __reset_page_owner() wants to clear the bit */
 271                 if (handle) {
 272                         __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 273                         page_owner->free_handle = handle;
 274                 }
 275                 page_owner->free_ts_nsec = free_ts_nsec;
 276                 page_owner->free_pid = current->pid;
 277                 page_owner->free_tgid = current->tgid;
 278                 page_ext = page_ext_next(page_ext);
 279         }
 280 }
 281
 282 void __reset_page_owner(struct page *page, unsigned short order)
 283 {
 284         struct page_ext *page_ext;
 285         depot_stack_handle_t handle;
 286         depot_stack_handle_t alloc_handle;
 287         struct page_owner *page_owner;
 288         u64 free_ts_nsec = local_clock();
 289
 290         page_ext = page_ext_get(page);
 291         if (unlikely(!page_ext))
 292                 return;
 293
 294         page_owner = get_page_owner(page_ext);
 295         alloc_handle = page_owner->handle;
 296
 297         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 298         __update_page_owner_free_handle(page_ext, handle, order, current->pid,
 299                                         current->tgid, free_ts_nsec);
 300         page_ext_put(page_ext);
 301
 302         if (alloc_handle != early_handle)
 303                 /*
 304                  * early_handle is being set as a handle for all those
 305                  * early allocated pages. See init_pages_in_zone().
 306                  * Since their refcount is not being incremented because
 307                  * the machinery is not ready yet, we cannot decrement
 308                  * their refcount either.
 309                  */
 310                 dec_stack_record_count(alloc_handle, 1 << order);
 311 }
 312
 313 noinline void __set_page_owner(struct page *page, unsigned short order,
 314                                         gfp_t gfp_mask)
 315 {
 316         struct page_ext *page_ext;
 317         u64 ts_nsec = local_clock();
 318         depot_stack_handle_t handle;
 319
 320         handle = save_stack(gfp_mask);
 321
 322         page_ext = page_ext_get(page);
 323         if (unlikely(!page_ext))
 324                 return;
 325         __update_page_owner_handle(page_ext, handle, order, gfp_mask, -1,
 326                                    ts_nsec, current->pid, current->tgid,
 327                                    current->comm);
 328         page_ext_put(page_ext);
 329         inc_stack_record_count(handle, gfp_mask, 1 << order);
 330 }
 331
 332 void __set_page_owner_migrate_reason(struct page *page, int reason)
 333 {
 334         struct page_ext *page_ext = page_ext_get(page);
 335         struct page_owner *page_owner;
 336
 337         if (unlikely(!page_ext))
 338                 return;
 339
 340         page_owner = get_page_owner(page_ext);
 341         page_owner->last_migrate_reason = reason;
 342         page_ext_put(page_ext);
 343 }
 344
 345 void __split_page_owner(struct page *page, int old_order, int new_order)
 346 {
 347         int i;
 348         struct page_ext *page_ext = page_ext_get(page);
 349         struct page_owner *page_owner;
 350
 351         if (unlikely(!page_ext))
 352                 return;
 353
 354         for (i = 0; i < (1 << old_order); i++) {
 355                 page_owner = get_page_owner(page_ext);
 356                 page_owner->order = new_order;
 357                 page_ext = page_ext_next(page_ext);
 358         }
 359         page_ext_put(page_ext);
 360 }
 361
 362 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 363 {
 364         int i;
 365         struct page_ext *old_ext;
 366         struct page_ext *new_ext;
 367         struct page_owner *old_page_owner;
 368         struct page_owner *new_page_owner;
 369         depot_stack_handle_t migrate_handle;
 370
 371         old_ext = page_ext_get(&old->page);
 372         if (unlikely(!old_ext))
 373                 return;
 374
 375         new_ext = page_ext_get(&newfolio->page);
 376         if (unlikely(!new_ext)) {
 377                 page_ext_put(old_ext);
 378                 return;
 379         }
 380
 381         old_page_owner = get_page_owner(old_ext);
 382         new_page_owner = get_page_owner(new_ext);
 383         migrate_handle = new_page_owner->handle;
 384         __update_page_owner_handle(new_ext, old_page_owner->handle,
 385                                    old_page_owner->order, old_page_owner->gfp_mask,
 386                                    old_page_owner->last_migrate_reason,
 387                                    old_page_owner->ts_nsec, old_page_owner->pid,
 388                                    old_page_owner->tgid, old_page_owner->comm);
 389         /*
 390          * Do not proactively clear PAGE_EXT_OWNER{_ALLOCATED} bits as the folio
 391          * will be freed after migration. Keep them until then as they may be
 392          * useful.
 393          */
 394         __update_page_owner_free_handle(new_ext, 0, old_page_owner->order,
 395                                         old_page_owner->free_pid,
 396                                         old_page_owner->free_tgid,
 397                                         old_page_owner->free_ts_nsec);
 398         /*
 399          * We linked the original stack to the new folio, we need to do the same
 400          * for the new one and the old folio otherwise there will be an imbalance
 401          * when subtracting those pages from the stack.
 402          */
 403         for (i = 0; i < (1 << new_page_owner->order); i++) {
 404                 old_page_owner->handle = migrate_handle;
 405                 old_ext = page_ext_next(old_ext);
 406                 old_page_owner = get_page_owner(old_ext);
 407         }
 408
 409         page_ext_put(new_ext);
 410         page_ext_put(old_ext);
 411 }
 412
 413 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 414                                        pg_data_t *pgdat, struct zone *zone)
 415 {
 416         struct page *page;
 417         struct page_ext *page_ext;
 418         struct page_owner *page_owner;
 419         unsigned long pfn, block_end_pfn;
 420         unsigned long end_pfn = zone_end_pfn(zone);
 421         unsigned long count[MIGRATE_TYPES] = { 0, };
 422         int pageblock_mt, page_mt;
 423         int i;
 424
 425         /* Scan block by block. First and last block may be incomplete */
 426         pfn = zone->zone_start_pfn;
 427
 428         /*
 429          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 430          * a zone boundary, it will be double counted between zones. This does
 431          * not matter as the mixed block count will still be correct
 432          */
 433         for (; pfn < end_pfn; ) {
 434                 page = pfn_to_online_page(pfn);
 435                 if (!page) {
 436                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 437                         continue;
 438                 }
 439
 440                 block_end_pfn = pageblock_end_pfn(pfn);
 441                 block_end_pfn = min(block_end_pfn, end_pfn);
 442
 443                 pageblock_mt = get_pageblock_migratetype(page);
 444
 445                 for (; pfn < block_end_pfn; pfn++) {
 446                         /* The pageblock is online, no need to recheck. */
 447                         page = pfn_to_page(pfn);
 448
 449                         if (page_zone(page) != zone)
 450                                 continue;
 451
 452                         if (PageBuddy(page)) {
 453                                 unsigned long freepage_order;
 454
 455                                 freepage_order = buddy_order_unsafe(page);
 456                                 if (freepage_order <= MAX_PAGE_ORDER)
 457                                         pfn += (1UL << freepage_order) - 1;
 458                                 continue;
 459                         }
 460
 461                         if (PageReserved(page))
 462                                 continue;
 463
 464                         page_ext = page_ext_get(page);
 465                         if (unlikely(!page_ext))
 466                                 continue;
 467
 468                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 469                                 goto ext_put_continue;
 470
 471                         page_owner = get_page_owner(page_ext);
 472                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 473                         if (pageblock_mt != page_mt) {
 474                                 if (is_migrate_cma(pageblock_mt))
 475                                         count[MIGRATE_MOVABLE]++;
 476                                 else
 477                                         count[pageblock_mt]++;
 478
 479                                 pfn = block_end_pfn;
 480                                 page_ext_put(page_ext);
 481                                 break;
 482                         }
 483                         pfn += (1UL << page_owner->order) - 1;
 484 ext_put_continue:
 485                         page_ext_put(page_ext);
 486                 }
 487         }
 488
 489         /* Print counts */
 490         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 491         for (i = 0; i < MIGRATE_TYPES; i++)
 492                 seq_printf(m, "%12lu ", count[i]);
 493         seq_putc(m, '\n');
 494 }
 495
 496 /*
 497  * Looking for memcg information and print it out
 498  */
 499 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
 500                                          struct page *page)
 501 {
 502 #ifdef CONFIG_MEMCG
 503         unsigned long memcg_data;
 504         struct mem_cgroup *memcg;
 505         bool online;
 506         char name[80];
 507
 508         rcu_read_lock();
 509         memcg_data = READ_ONCE(page->memcg_data);
 510         if (!memcg_data)
 511                 goto out_unlock;
 512
 513         if (memcg_data & MEMCG_DATA_OBJEXTS)
 514                 ret += scnprintf(kbuf + ret, count - ret,
 515                                 "Slab cache page\n");
 516
 517         memcg = page_memcg_check(page);
 518         if (!memcg)
 519                 goto out_unlock;
 520
 521         online = (memcg->css.flags & CSS_ONLINE);
 522         cgroup_name(memcg->css.cgroup, name, sizeof(name));
 523         ret += scnprintf(kbuf + ret, count - ret,
 524                         "Charged %sto %smemcg %s\n",
 525                         PageMemcgKmem(page) ? "(via objcg) " : "",
 526                         online ? "" : "offline ",
 527                         name);
 528 out_unlock:
 529         rcu_read_unlock();
 530 #endif /* CONFIG_MEMCG */
 531
 532         return ret;
 533 }
 534
 535 static ssize_t
 536 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 537                 struct page *page, struct page_owner *page_owner,
 538                 depot_stack_handle_t handle)
 539 {
 540         int ret, pageblock_mt, page_mt;
 541         char *kbuf;
 542
 543         count = min_t(size_t, count, PAGE_SIZE);
 544         kbuf = kmalloc(count, GFP_KERNEL);
 545         if (!kbuf)
 546                 return -ENOMEM;
 547
 548         ret = scnprintf(kbuf, count,
 549                         "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
 550                         page_owner->order, page_owner->gfp_mask,
 551                         &page_owner->gfp_mask, page_owner->pid,
 552                         page_owner->tgid, page_owner->comm,
 553                         page_owner->ts_nsec);
 554
 555         /* Print information relevant to grouping pages by mobility */
 556         pageblock_mt = get_pageblock_migratetype(page);
 557         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 558         ret += scnprintf(kbuf + ret, count - ret,
 559                         "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
 560                         pfn,
 561                         migratetype_names[page_mt],
 562                         pfn >> pageblock_order,
 563                         migratetype_names[pageblock_mt],
 564                         &page->flags);
 565
 566         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 567         if (ret >= count)
 568                 goto err;
 569
 570         if (page_owner->last_migrate_reason != -1) {
 571                 ret += scnprintf(kbuf + ret, count - ret,
 572                         "Page has been migrated, last migrate reason: %s\n",
 573                         migrate_reason_names[page_owner->last_migrate_reason]);
 574         }
 575
 576         ret = print_page_owner_memcg(kbuf, count, ret, page);
 577
 578         ret += snprintf(kbuf + ret, count - ret, "\n");
 579         if (ret >= count)
 580                 goto err;
 581
 582         if (copy_to_user(buf, kbuf, ret))
 583                 ret = -EFAULT;
 584
 585         kfree(kbuf);
 586         return ret;
 587
 588 err:
 589         kfree(kbuf);
 590         return -ENOMEM;
 591 }
 592
 593 void __dump_page_owner(const struct page *page)
 594 {
 595         struct page_ext *page_ext = page_ext_get((void *)page);
 596         struct page_owner *page_owner;
 597         depot_stack_handle_t handle;
 598         gfp_t gfp_mask;
 599         int mt;
 600
 601         if (unlikely(!page_ext)) {
 602                 pr_alert("There is not page extension available.\n");
 603                 return;
 604         }
 605
 606         page_owner = get_page_owner(page_ext);
 607         gfp_mask = page_owner->gfp_mask;
 608         mt = gfp_migratetype(gfp_mask);
 609
 610         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 611                 pr_alert("page_owner info is not present (never set?)\n");
 612                 page_ext_put(page_ext);
 613                 return;
 614         }
 615
 616         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 617                 pr_alert("page_owner tracks the page as allocated\n");
 618         else
 619                 pr_alert("page_owner tracks the page as freed\n");
 620
 621         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",
 622                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 623                  page_owner->pid, page_owner->tgid, page_owner->comm,
 624                  page_owner->ts_nsec, page_owner->free_ts_nsec);
 625
 626         handle = READ_ONCE(page_owner->handle);
 627         if (!handle)
 628                 pr_alert("page_owner allocation stack trace missing\n");
 629         else
 630                 stack_depot_print(handle);
 631
 632         handle = READ_ONCE(page_owner->free_handle);
 633         if (!handle) {
 634                 pr_alert("page_owner free stack trace missing\n");
 635         } else {
 636                 pr_alert("page last free pid %d tgid %d stack trace:\n",
 637                           page_owner->free_pid, page_owner->free_tgid);
 638                 stack_depot_print(handle);
 639         }
 640
 641         if (page_owner->last_migrate_reason != -1)
 642                 pr_alert("page has been migrated, last migrate reason: %s\n",
 643                         migrate_reason_names[page_owner->last_migrate_reason]);
 644         page_ext_put(page_ext);
 645 }
 646
 647 static ssize_t
 648 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 649 {
 650         unsigned long pfn;
 651         struct page *page;
 652         struct page_ext *page_ext;
 653         struct page_owner *page_owner;
 654         depot_stack_handle_t handle;
 655
 656         if (!static_branch_unlikely(&page_owner_inited))
 657                 return -EINVAL;
 658
 659         page = NULL;
 660         if (*ppos == 0)
 661                 pfn = min_low_pfn;
 662         else
 663                 pfn = *ppos;
 664         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 665         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 666                 pfn++;
 667
 668         /* Find an allocated page */
 669         for (; pfn < max_pfn; pfn++) {
 670                 /*
 671                  * This temporary page_owner is required so
 672                  * that we can avoid the context switches while holding
 673                  * the rcu lock and copying the page owner information to
 674                  * user through copy_to_user() or GFP_KERNEL allocations.
 675                  */
 676                 struct page_owner page_owner_tmp;
 677
 678                 /*
 679                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 680                  * validate the area as existing, skip it if not
 681                  */
 682                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 683                         pfn += MAX_ORDER_NR_PAGES - 1;
 684                         continue;
 685                 }
 686
 687                 page = pfn_to_page(pfn);
 688                 if (PageBuddy(page)) {
 689                         unsigned long freepage_order = buddy_order_unsafe(page);
 690
 691                         if (freepage_order <= MAX_PAGE_ORDER)
 692                                 pfn += (1UL << freepage_order) - 1;
 693                         continue;
 694                 }
 695
 696                 page_ext = page_ext_get(page);
 697                 if (unlikely(!page_ext))
 698                         continue;
 699
 700                 /*
 701                  * Some pages could be missed by concurrent allocation or free,
 702                  * because we don't hold the zone lock.
 703                  */
 704                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 705                         goto ext_put_continue;
 706
 707                 /*
 708                  * Although we do have the info about past allocation of free
 709                  * pages, it's not relevant for current memory usage.
 710                  */
 711                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 712                         goto ext_put_continue;
 713
 714                 page_owner = get_page_owner(page_ext);
 715
 716                 /*
 717                  * Don't print "tail" pages of high-order allocations as that
 718                  * would inflate the stats.
 719                  */
 720                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 721                         goto ext_put_continue;
 722
 723                 /*
 724                  * Access to page_ext->handle isn't synchronous so we should
 725                  * be careful to access it.
 726                  */
 727                 handle = READ_ONCE(page_owner->handle);
 728                 if (!handle)
 729                         goto ext_put_continue;
 730
 731                 /* Record the next PFN to read in the file offset */
 732                 *ppos = pfn + 1;
 733
 734                 page_owner_tmp = *page_owner;
 735                 page_ext_put(page_ext);
 736                 return print_page_owner(buf, count, pfn, page,
 737                                 &page_owner_tmp, handle);
 738 ext_put_continue:
 739                 page_ext_put(page_ext);
 740         }
 741
 742         return 0;
 743 }
 744
 745 static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
 746 {
 747         switch (orig) {
 748         case SEEK_SET:
 749                 file->f_pos = offset;
 750                 break;
 751         case SEEK_CUR:
 752                 file->f_pos += offset;
 753                 break;
 754         default:
 755                 return -EINVAL;
 756         }
 757         return file->f_pos;
 758 }
 759
 760 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 761 {
 762         unsigned long pfn = zone->zone_start_pfn;
 763         unsigned long end_pfn = zone_end_pfn(zone);
 764         unsigned long count = 0;
 765
 766         /*
 767          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 768          * a zone boundary, it will be double counted between zones. This does
 769          * not matter as the mixed block count will still be correct
 770          */
 771         for (; pfn < end_pfn; ) {
 772                 unsigned long block_end_pfn;
 773
 774                 if (!pfn_valid(pfn)) {
 775                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 776                         continue;
 777                 }
 778
 779                 block_end_pfn = pageblock_end_pfn(pfn);
 780                 block_end_pfn = min(block_end_pfn, end_pfn);
 781
 782                 for (; pfn < block_end_pfn; pfn++) {
 783                         struct page *page = pfn_to_page(pfn);
 784                         struct page_ext *page_ext;
 785
 786                         if (page_zone(page) != zone)
 787                                 continue;
 788
 789                         /*
 790                          * To avoid having to grab zone->lock, be a little
 791                          * careful when reading buddy page order. The only
 792                          * danger is that we skip too much and potentially miss
 793                          * some early allocated pages, which is better than
 794                          * heavy lock contention.
 795                          */
 796                         if (PageBuddy(page)) {
 797                                 unsigned long order = buddy_order_unsafe(page);
 798
 799                                 if (order > 0 && order <= MAX_PAGE_ORDER)
 800                                         pfn += (1UL << order) - 1;
 801                                 continue;
 802                         }
 803
 804                         if (PageReserved(page))
 805                                 continue;
 806
 807                         page_ext = page_ext_get(page);
 808                         if (unlikely(!page_ext))
 809                                 continue;
 810
 811                         /* Maybe overlapping zone */
 812                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 813                                 goto ext_put_continue;
 814
 815                         /* Found early allocated page */
 816                         __update_page_owner_handle(page_ext, early_handle, 0, 0,
 817                                                    -1, local_clock(), current->pid,
 818                                                    current->tgid, current->comm);
 819                         count++;
 820 ext_put_continue:
 821                         page_ext_put(page_ext);
 822                 }
 823                 cond_resched();
 824         }
 825
 826         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 827                 pgdat->node_id, zone->name, count);
 828 }
 829
 830 static void init_zones_in_node(pg_data_t *pgdat)
 831 {
 832         struct zone *zone;
 833         struct zone *node_zones = pgdat->node_zones;
 834
 835         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 836                 if (!populated_zone(zone))
 837                         continue;
 838
 839                 init_pages_in_zone(pgdat, zone);
 840         }
 841 }
 842
 843 static void init_early_allocated_pages(void)
 844 {
 845         pg_data_t *pgdat;
 846
 847         for_each_online_pgdat(pgdat)
 848                 init_zones_in_node(pgdat);
 849 }
 850
 851 static const struct file_operations proc_page_owner_operations = {
 852         .read           = read_page_owner,
 853         .llseek         = lseek_page_owner,
 854 };
 855
 856 static void *stack_start(struct seq_file *m, loff_t *ppos)
 857 {
 858         struct stack *stack;
 859
 860         if (*ppos == -1UL)
 861                 return NULL;
 862
 863         if (!*ppos) {
 864                 /*
 865                  * This pairs with smp_store_release() from function
 866                  * add_stack_record_to_list(), so we get a consistent
 867                  * value of stack_list.
 868                  */
 869                 stack = smp_load_acquire(&stack_list);
 870                 m->private = stack;
 871         } else {
 872                 stack = m->private;
 873         }
 874
 875         return stack;
 876 }
 877
 878 static void *stack_next(struct seq_file *m, void *v, loff_t *ppos)
 879 {
 880         struct stack *stack = v;
 881
 882         stack = stack->next;
 883         *ppos = stack ? *ppos + 1 : -1UL;
 884         m->private = stack;
 885
 886         return stack;
 887 }
 888
 889 static unsigned long page_owner_pages_threshold;
 890
 891 static int stack_print(struct seq_file *m, void *v)
 892 {
 893         int i, nr_base_pages;
 894         struct stack *stack = v;
 895         unsigned long *entries;
 896         unsigned long nr_entries;
 897         struct stack_record *stack_record = stack->stack_record;
 898
 899         if (!stack->stack_record)
 900                 return 0;
 901
 902         nr_entries = stack_record->size;
 903         entries = stack_record->entries;
 904         nr_base_pages = refcount_read(&stack_record->count) - 1;
 905
 906         if (nr_base_pages < 1 || nr_base_pages < page_owner_pages_threshold)
 907                 return 0;
 908
 909         for (i = 0; i < nr_entries; i++)
 910                 seq_printf(m, " %pS\n", (void *)entries[i]);
 911         seq_printf(m, "nr_base_pages: %d\n\n", nr_base_pages);
 912
 913         return 0;
 914 }
 915
 916 static void stack_stop(struct seq_file *m, void *v)
 917 {
 918 }
 919
 920 static const struct seq_operations page_owner_stack_op = {
 921         .start  = stack_start,
 922         .next   = stack_next,
 923         .stop   = stack_stop,
 924         .show   = stack_print
 925 };
 926
 927 static int page_owner_stack_open(struct inode *inode, struct file *file)
 928 {
 929         return seq_open_private(file, &page_owner_stack_op, 0);
 930 }
 931
 932 static const struct file_operations page_owner_stack_operations = {
 933         .open           = page_owner_stack_open,
 934         .read           = seq_read,
 935         .llseek         = seq_lseek,
 936         .release        = seq_release,
 937 };
 938
 939 static int page_owner_threshold_get(void *data, u64 *val)
 940 {
 941         *val = READ_ONCE(page_owner_pages_threshold);
 942         return 0;
 943 }
 944
 945 static int page_owner_threshold_set(void *data, u64 val)
 946 {
 947         WRITE_ONCE(page_owner_pages_threshold, val);
 948         return 0;
 949 }
 950
 951 DEFINE_SIMPLE_ATTRIBUTE(proc_page_owner_threshold, &page_owner_threshold_get,
 952                         &page_owner_threshold_set, "%llu");
 953
 954
 955 static int __init pageowner_init(void)
 956 {
 957         struct dentry *dir;
 958
 959         if (!static_branch_unlikely(&page_owner_inited)) {
 960                 pr_info("page_owner is disabled\n");
 961                 return 0;
 962         }
 963
 964         debugfs_create_file("page_owner", 0400, NULL, NULL,
 965                             &proc_page_owner_operations);
 966         dir = debugfs_create_dir("page_owner_stacks", NULL);
 967         debugfs_create_file("show_stacks", 0400, dir, NULL,
 968                             &page_owner_stack_operations);
 969         debugfs_create_file("count_threshold", 0600, dir, NULL,
 970                             &proc_page_owner_threshold);
 971
 972         return 0;
 973 }
 974 late_initcall(pageowner_init)