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         /* Filter gfp_mask the same way stackdepot does, for consistency */
 172         gfp_mask &= ~GFP_ZONEMASK;
 173         gfp_mask &= (GFP_ATOMIC | GFP_KERNEL | __GFP_NOLOCKDEP);
 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                                    int nr_base_pages)
 201 {
 202         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 203
 204         if (!stack_record)
 205                 return;
 206
 207         /*
 208          * New stack_record's that do not use STACK_DEPOT_FLAG_GET start
 209          * with REFCOUNT_SATURATED to catch spurious increments of their
 210          * refcount.
 211          * Since we do not use STACK_DEPOT_FLAG_GET API, let us
 212          * set a refcount of 1 ourselves.
 213          */
 214         if (refcount_read(&stack_record->count) == REFCOUNT_SATURATED) {
 215                 int old = REFCOUNT_SATURATED;
 216
 217                 if (atomic_try_cmpxchg_relaxed(&stack_record->count.refs, &old, 1))
 218                         /* Add the new stack_record to our list */
 219                         add_stack_record_to_list(stack_record, gfp_mask);
 220         }
 221         refcount_add(nr_base_pages, &stack_record->count);
 222 }
 223
 224 static void dec_stack_record_count(depot_stack_handle_t handle,
 225                                    int nr_base_pages)
 226 {
 227         struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
 228
 229         if (!stack_record)
 230                 return;
 231
 232         if (refcount_sub_and_test(nr_base_pages, &stack_record->count))
 233                 pr_warn("%s: refcount went to 0 for %u handle\n", __func__,
 234                         handle);
 235 }
 236
 237 static inline void __update_page_owner_handle(struct page_ext *page_ext,
 238                                               depot_stack_handle_t handle,
 239                                               unsigned short order,
 240                                               gfp_t gfp_mask,
 241                                               short last_migrate_reason, u64 ts_nsec,
 242                                               pid_t pid, pid_t tgid, char *comm)
 243 {
 244         int i;
 245         struct page_owner *page_owner;
 246
 247         for (i = 0; i < (1 << order); i++) {
 248                 page_owner = get_page_owner(page_ext);
 249                 page_owner->handle = handle;
 250                 page_owner->order = order;
 251                 page_owner->gfp_mask = gfp_mask;
 252                 page_owner->last_migrate_reason = last_migrate_reason;
 253                 page_owner->pid = pid;
 254                 page_owner->tgid = tgid;
 255                 page_owner->ts_nsec = ts_nsec;
 256                 strscpy(page_owner->comm, comm,
 257                         sizeof(page_owner->comm));
 258                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 259                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 260                 page_ext = page_ext_next(page_ext);
 261         }
 262 }
 263
 264 static inline void __update_page_owner_free_handle(struct page_ext *page_ext,
 265                                                    depot_stack_handle_t handle,
 266                                                    unsigned short order,
 267                                                    pid_t pid, pid_t tgid,
 268                                                    u64 free_ts_nsec)
 269 {
 270         int i;
 271         struct page_owner *page_owner;
 272
 273         for (i = 0; i < (1 << order); i++) {
 274                 page_owner = get_page_owner(page_ext);
 275                 /* Only __reset_page_owner() wants to clear the bit */
 276                 if (handle) {
 277                         __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 278                         page_owner->free_handle = handle;
 279                 }
 280                 page_owner->free_ts_nsec = free_ts_nsec;
 281                 page_owner->free_pid = current->pid;
 282                 page_owner->free_tgid = current->tgid;
 283                 page_ext = page_ext_next(page_ext);
 284         }
 285 }
 286
 287 void __reset_page_owner(struct page *page, unsigned short order)
 288 {
 289         struct page_ext *page_ext;
 290         depot_stack_handle_t handle;
 291         depot_stack_handle_t alloc_handle;
 292         struct page_owner *page_owner;
 293         u64 free_ts_nsec = local_clock();
 294
 295         page_ext = page_ext_get(page);
 296         if (unlikely(!page_ext))
 297                 return;
 298
 299         page_owner = get_page_owner(page_ext);
 300         alloc_handle = page_owner->handle;
 301
 302         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 303         __update_page_owner_free_handle(page_ext, handle, order, current->pid,
 304                                         current->tgid, free_ts_nsec);
 305         page_ext_put(page_ext);
 306
 307         if (alloc_handle != early_handle)
 308                 /*
 309                  * early_handle is being set as a handle for all those
 310                  * early allocated pages. See init_pages_in_zone().
 311                  * Since their refcount is not being incremented because
 312                  * the machinery is not ready yet, we cannot decrement
 313                  * their refcount either.
 314                  */
 315                 dec_stack_record_count(alloc_handle, 1 << order);
 316 }
 317
 318 noinline void __set_page_owner(struct page *page, unsigned short order,
 319                                         gfp_t gfp_mask)
 320 {
 321         struct page_ext *page_ext;
 322         u64 ts_nsec = local_clock();
 323         depot_stack_handle_t handle;
 324
 325         handle = save_stack(gfp_mask);
 326
 327         page_ext = page_ext_get(page);
 328         if (unlikely(!page_ext))
 329                 return;
 330         __update_page_owner_handle(page_ext, handle, order, gfp_mask, -1,
 331                                    ts_nsec, current->pid, current->tgid,
 332                                    current->comm);
 333         page_ext_put(page_ext);
 334         inc_stack_record_count(handle, gfp_mask, 1 << order);
 335 }
 336
 337 void __set_page_owner_migrate_reason(struct page *page, int reason)
 338 {
 339         struct page_ext *page_ext = page_ext_get(page);
 340         struct page_owner *page_owner;
 341
 342         if (unlikely(!page_ext))
 343                 return;
 344
 345         page_owner = get_page_owner(page_ext);
 346         page_owner->last_migrate_reason = reason;
 347         page_ext_put(page_ext);
 348 }
 349
 350 void __split_page_owner(struct page *page, int old_order, int new_order)
 351 {
 352         int i;
 353         struct page_ext *page_ext = page_ext_get(page);
 354         struct page_owner *page_owner;
 355
 356         if (unlikely(!page_ext))
 357                 return;
 358
 359         for (i = 0; i < (1 << old_order); i++) {
 360                 page_owner = get_page_owner(page_ext);
 361                 page_owner->order = new_order;
 362                 page_ext = page_ext_next(page_ext);
 363         }
 364         page_ext_put(page_ext);
 365 }
 366
 367 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 368 {
 369         int i;
 370         struct page_ext *old_ext;
 371         struct page_ext *new_ext;
 372         struct page_owner *old_page_owner;
 373         struct page_owner *new_page_owner;
 374         depot_stack_handle_t migrate_handle;
 375
 376         old_ext = page_ext_get(&old->page);
 377         if (unlikely(!old_ext))
 378                 return;
 379
 380         new_ext = page_ext_get(&newfolio->page);
 381         if (unlikely(!new_ext)) {
 382                 page_ext_put(old_ext);
 383                 return;
 384         }
 385
 386         old_page_owner = get_page_owner(old_ext);
 387         new_page_owner = get_page_owner(new_ext);
 388         migrate_handle = new_page_owner->handle;
 389         __update_page_owner_handle(new_ext, old_page_owner->handle,
 390                                    old_page_owner->order, old_page_owner->gfp_mask,
 391                                    old_page_owner->last_migrate_reason,
 392                                    old_page_owner->ts_nsec, old_page_owner->pid,
 393                                    old_page_owner->tgid, old_page_owner->comm);
 394         /*
 395          * Do not proactively clear PAGE_EXT_OWNER{_ALLOCATED} bits as the folio
 396          * will be freed after migration. Keep them until then as they may be
 397          * useful.
 398          */
 399         __update_page_owner_free_handle(new_ext, 0, old_page_owner->order,
 400                                         old_page_owner->free_pid,
 401                                         old_page_owner->free_tgid,
 402                                         old_page_owner->free_ts_nsec);
 403         /*
 404          * We linked the original stack to the new folio, we need to do the same
 405          * for the new one and the old folio otherwise there will be an imbalance
 406          * when subtracting those pages from the stack.
 407          */
 408         for (i = 0; i < (1 << new_page_owner->order); i++) {
 409                 old_page_owner->handle = migrate_handle;
 410                 old_ext = page_ext_next(old_ext);
 411                 old_page_owner = get_page_owner(old_ext);
 412         }
 413
 414         page_ext_put(new_ext);
 415         page_ext_put(old_ext);
 416 }
 417
 418 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 419                                        pg_data_t *pgdat, struct zone *zone)
 420 {
 421         struct page *page;
 422         struct page_ext *page_ext;
 423         struct page_owner *page_owner;
 424         unsigned long pfn, block_end_pfn;
 425         unsigned long end_pfn = zone_end_pfn(zone);
 426         unsigned long count[MIGRATE_TYPES] = { 0, };
 427         int pageblock_mt, page_mt;
 428         int i;
 429
 430         /* Scan block by block. First and last block may be incomplete */
 431         pfn = zone->zone_start_pfn;
 432
 433         /*
 434          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 435          * a zone boundary, it will be double counted between zones. This does
 436          * not matter as the mixed block count will still be correct
 437          */
 438         for (; pfn < end_pfn; ) {
 439                 page = pfn_to_online_page(pfn);
 440                 if (!page) {
 441                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 442                         continue;
 443                 }
 444
 445                 block_end_pfn = pageblock_end_pfn(pfn);
 446                 block_end_pfn = min(block_end_pfn, end_pfn);
 447
 448                 pageblock_mt = get_pageblock_migratetype(page);
 449
 450                 for (; pfn < block_end_pfn; pfn++) {
 451                         /* The pageblock is online, no need to recheck. */
 452                         page = pfn_to_page(pfn);
 453
 454                         if (page_zone(page) != zone)
 455                                 continue;
 456
 457                         if (PageBuddy(page)) {
 458                                 unsigned long freepage_order;
 459
 460                                 freepage_order = buddy_order_unsafe(page);
 461                                 if (freepage_order <= MAX_PAGE_ORDER)
 462                                         pfn += (1UL << freepage_order) - 1;
 463                                 continue;
 464                         }
 465
 466                         if (PageReserved(page))
 467                                 continue;
 468
 469                         page_ext = page_ext_get(page);
 470                         if (unlikely(!page_ext))
 471                                 continue;
 472
 473                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 474                                 goto ext_put_continue;
 475
 476                         page_owner = get_page_owner(page_ext);
 477                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 478                         if (pageblock_mt != page_mt) {
 479                                 if (is_migrate_cma(pageblock_mt))
 480                                         count[MIGRATE_MOVABLE]++;
 481                                 else
 482                                         count[pageblock_mt]++;
 483
 484                                 pfn = block_end_pfn;
 485                                 page_ext_put(page_ext);
 486                                 break;
 487                         }
 488                         pfn += (1UL << page_owner->order) - 1;
 489 ext_put_continue:
 490                         page_ext_put(page_ext);
 491                 }
 492         }
 493
 494         /* Print counts */
 495         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 496         for (i = 0; i < MIGRATE_TYPES; i++)
 497                 seq_printf(m, "%12lu ", count[i]);
 498         seq_putc(m, '\n');
 499 }
 500
 501 /*
 502  * Looking for memcg information and print it out
 503  */
 504 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
 505                                          struct page *page)
 506 {
 507 #ifdef CONFIG_MEMCG
 508         unsigned long memcg_data;
 509         struct mem_cgroup *memcg;
 510         bool online;
 511         char name[80];
 512
 513         rcu_read_lock();
 514         memcg_data = READ_ONCE(page->memcg_data);
 515         if (!memcg_data)
 516                 goto out_unlock;
 517
 518         if (memcg_data & MEMCG_DATA_OBJCGS)
 519                 ret += scnprintf(kbuf + ret, count - ret,
 520                                 "Slab cache page\n");
 521
 522         memcg = page_memcg_check(page);
 523         if (!memcg)
 524                 goto out_unlock;
 525
 526         online = (memcg->css.flags & CSS_ONLINE);
 527         cgroup_name(memcg->css.cgroup, name, sizeof(name));
 528         ret += scnprintf(kbuf + ret, count - ret,
 529                         "Charged %sto %smemcg %s\n",
 530                         PageMemcgKmem(page) ? "(via objcg) " : "",
 531                         online ? "" : "offline ",
 532                         name);
 533 out_unlock:
 534         rcu_read_unlock();
 535 #endif /* CONFIG_MEMCG */
 536
 537         return ret;
 538 }
 539
 540 static ssize_t
 541 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 542                 struct page *page, struct page_owner *page_owner,
 543                 depot_stack_handle_t handle)
 544 {
 545         int ret, pageblock_mt, page_mt;
 546         char *kbuf;
 547
 548         count = min_t(size_t, count, PAGE_SIZE);
 549         kbuf = kmalloc(count, GFP_KERNEL);
 550         if (!kbuf)
 551                 return -ENOMEM;
 552
 553         ret = scnprintf(kbuf, count,
 554                         "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
 555                         page_owner->order, page_owner->gfp_mask,
 556                         &page_owner->gfp_mask, page_owner->pid,
 557                         page_owner->tgid, page_owner->comm,
 558                         page_owner->ts_nsec);
 559
 560         /* Print information relevant to grouping pages by mobility */
 561         pageblock_mt = get_pageblock_migratetype(page);
 562         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 563         ret += scnprintf(kbuf + ret, count - ret,
 564                         "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
 565                         pfn,
 566                         migratetype_names[page_mt],
 567                         pfn >> pageblock_order,
 568                         migratetype_names[pageblock_mt],
 569                         &page->flags);
 570
 571         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 572         if (ret >= count)
 573                 goto err;
 574
 575         if (page_owner->last_migrate_reason != -1) {
 576                 ret += scnprintf(kbuf + ret, count - ret,
 577                         "Page has been migrated, last migrate reason: %s\n",
 578                         migrate_reason_names[page_owner->last_migrate_reason]);
 579         }
 580
 581         ret = print_page_owner_memcg(kbuf, count, ret, page);
 582
 583         ret += snprintf(kbuf + ret, count - ret, "\n");
 584         if (ret >= count)
 585                 goto err;
 586
 587         if (copy_to_user(buf, kbuf, ret))
 588                 ret = -EFAULT;
 589
 590         kfree(kbuf);
 591         return ret;
 592
 593 err:
 594         kfree(kbuf);
 595         return -ENOMEM;
 596 }
 597
 598 void __dump_page_owner(const struct page *page)
 599 {
 600         struct page_ext *page_ext = page_ext_get((void *)page);
 601         struct page_owner *page_owner;
 602         depot_stack_handle_t handle;
 603         gfp_t gfp_mask;
 604         int mt;
 605
 606         if (unlikely(!page_ext)) {
 607                 pr_alert("There is not page extension available.\n");
 608                 return;
 609         }
 610
 611         page_owner = get_page_owner(page_ext);
 612         gfp_mask = page_owner->gfp_mask;
 613         mt = gfp_migratetype(gfp_mask);
 614
 615         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 616                 pr_alert("page_owner info is not present (never set?)\n");
 617                 page_ext_put(page_ext);
 618                 return;
 619         }
 620
 621         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 622                 pr_alert("page_owner tracks the page as allocated\n");
 623         else
 624                 pr_alert("page_owner tracks the page as freed\n");
 625
 626         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",
 627                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 628                  page_owner->pid, page_owner->tgid, page_owner->comm,
 629                  page_owner->ts_nsec, page_owner->free_ts_nsec);
 630
 631         handle = READ_ONCE(page_owner->handle);
 632         if (!handle)
 633                 pr_alert("page_owner allocation stack trace missing\n");
 634         else
 635                 stack_depot_print(handle);
 636
 637         handle = READ_ONCE(page_owner->free_handle);
 638         if (!handle) {
 639                 pr_alert("page_owner free stack trace missing\n");
 640         } else {
 641                 pr_alert("page last free pid %d tgid %d stack trace:\n",
 642                           page_owner->free_pid, page_owner->free_tgid);
 643                 stack_depot_print(handle);
 644         }
 645
 646         if (page_owner->last_migrate_reason != -1)
 647                 pr_alert("page has been migrated, last migrate reason: %s\n",
 648                         migrate_reason_names[page_owner->last_migrate_reason]);
 649         page_ext_put(page_ext);
 650 }
 651
 652 static ssize_t
 653 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 654 {
 655         unsigned long pfn;
 656         struct page *page;
 657         struct page_ext *page_ext;
 658         struct page_owner *page_owner;
 659         depot_stack_handle_t handle;
 660
 661         if (!static_branch_unlikely(&page_owner_inited))
 662                 return -EINVAL;
 663
 664         page = NULL;
 665         if (*ppos == 0)
 666                 pfn = min_low_pfn;
 667         else
 668                 pfn = *ppos;
 669         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 670         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 671                 pfn++;
 672
 673         /* Find an allocated page */
 674         for (; pfn < max_pfn; pfn++) {
 675                 /*
 676                  * This temporary page_owner is required so
 677                  * that we can avoid the context switches while holding
 678                  * the rcu lock and copying the page owner information to
 679                  * user through copy_to_user() or GFP_KERNEL allocations.
 680                  */
 681                 struct page_owner page_owner_tmp;
 682
 683                 /*
 684                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 685                  * validate the area as existing, skip it if not
 686                  */
 687                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 688                         pfn += MAX_ORDER_NR_PAGES - 1;
 689                         continue;
 690                 }
 691
 692                 page = pfn_to_page(pfn);
 693                 if (PageBuddy(page)) {
 694                         unsigned long freepage_order = buddy_order_unsafe(page);
 695
 696                         if (freepage_order <= MAX_PAGE_ORDER)
 697                                 pfn += (1UL << freepage_order) - 1;
 698                         continue;
 699                 }
 700
 701                 page_ext = page_ext_get(page);
 702                 if (unlikely(!page_ext))
 703                         continue;
 704
 705                 /*
 706                  * Some pages could be missed by concurrent allocation or free,
 707                  * because we don't hold the zone lock.
 708                  */
 709                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 710                         goto ext_put_continue;
 711
 712                 /*
 713                  * Although we do have the info about past allocation of free
 714                  * pages, it's not relevant for current memory usage.
 715                  */
 716                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 717                         goto ext_put_continue;
 718
 719                 page_owner = get_page_owner(page_ext);
 720
 721                 /*
 722                  * Don't print "tail" pages of high-order allocations as that
 723                  * would inflate the stats.
 724                  */
 725                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 726                         goto ext_put_continue;
 727
 728                 /*
 729                  * Access to page_ext->handle isn't synchronous so we should
 730                  * be careful to access it.
 731                  */
 732                 handle = READ_ONCE(page_owner->handle);
 733                 if (!handle)
 734                         goto ext_put_continue;
 735
 736                 /* Record the next PFN to read in the file offset */
 737                 *ppos = pfn + 1;
 738
 739                 page_owner_tmp = *page_owner;
 740                 page_ext_put(page_ext);
 741                 return print_page_owner(buf, count, pfn, page,
 742                                 &page_owner_tmp, handle);
 743 ext_put_continue:
 744                 page_ext_put(page_ext);
 745         }
 746
 747         return 0;
 748 }
 749
 750 static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
 751 {
 752         switch (orig) {
 753         case SEEK_SET:
 754                 file->f_pos = offset;
 755                 break;
 756         case SEEK_CUR:
 757                 file->f_pos += offset;
 758                 break;
 759         default:
 760                 return -EINVAL;
 761         }
 762         return file->f_pos;
 763 }
 764
 765 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 766 {
 767         unsigned long pfn = zone->zone_start_pfn;
 768         unsigned long end_pfn = zone_end_pfn(zone);
 769         unsigned long count = 0;
 770
 771         /*
 772          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 773          * a zone boundary, it will be double counted between zones. This does
 774          * not matter as the mixed block count will still be correct
 775          */
 776         for (; pfn < end_pfn; ) {
 777                 unsigned long block_end_pfn;
 778
 779                 if (!pfn_valid(pfn)) {
 780                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 781                         continue;
 782                 }
 783
 784                 block_end_pfn = pageblock_end_pfn(pfn);
 785                 block_end_pfn = min(block_end_pfn, end_pfn);
 786
 787                 for (; pfn < block_end_pfn; pfn++) {
 788                         struct page *page = pfn_to_page(pfn);
 789                         struct page_ext *page_ext;
 790
 791                         if (page_zone(page) != zone)
 792                                 continue;
 793
 794                         /*
 795                          * To avoid having to grab zone->lock, be a little
 796                          * careful when reading buddy page order. The only
 797                          * danger is that we skip too much and potentially miss
 798                          * some early allocated pages, which is better than
 799                          * heavy lock contention.
 800                          */
 801                         if (PageBuddy(page)) {
 802                                 unsigned long order = buddy_order_unsafe(page);
 803
 804                                 if (order > 0 && order <= MAX_PAGE_ORDER)
 805                                         pfn += (1UL << order) - 1;
 806                                 continue;
 807                         }
 808
 809                         if (PageReserved(page))
 810                                 continue;
 811
 812                         page_ext = page_ext_get(page);
 813                         if (unlikely(!page_ext))
 814                                 continue;
 815
 816                         /* Maybe overlapping zone */
 817                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 818                                 goto ext_put_continue;
 819
 820                         /* Found early allocated page */
 821                         __update_page_owner_handle(page_ext, early_handle, 0, 0,
 822                                                    -1, local_clock(), current->pid,
 823                                                    current->tgid, current->comm);
 824                         count++;
 825 ext_put_continue:
 826                         page_ext_put(page_ext);
 827                 }
 828                 cond_resched();
 829         }
 830
 831         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 832                 pgdat->node_id, zone->name, count);
 833 }
 834
 835 static void init_zones_in_node(pg_data_t *pgdat)
 836 {
 837         struct zone *zone;
 838         struct zone *node_zones = pgdat->node_zones;
 839
 840         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 841                 if (!populated_zone(zone))
 842                         continue;
 843
 844                 init_pages_in_zone(pgdat, zone);
 845         }
 846 }
 847
 848 static void init_early_allocated_pages(void)
 849 {
 850         pg_data_t *pgdat;
 851
 852         for_each_online_pgdat(pgdat)
 853                 init_zones_in_node(pgdat);
 854 }
 855
 856 static const struct file_operations proc_page_owner_operations = {
 857         .read           = read_page_owner,
 858         .llseek         = lseek_page_owner,
 859 };
 860
 861 static void *stack_start(struct seq_file *m, loff_t *ppos)
 862 {
 863         struct stack *stack;
 864
 865         if (*ppos == -1UL)
 866                 return NULL;
 867
 868         if (!*ppos) {
 869                 /*
 870                  * This pairs with smp_store_release() from function
 871                  * add_stack_record_to_list(), so we get a consistent
 872                  * value of stack_list.
 873                  */
 874                 stack = smp_load_acquire(&stack_list);
 875                 m->private = stack;
 876         } else {
 877                 stack = m->private;
 878         }
 879
 880         return stack;
 881 }
 882
 883 static void *stack_next(struct seq_file *m, void *v, loff_t *ppos)
 884 {
 885         struct stack *stack = v;
 886
 887         stack = stack->next;
 888         *ppos = stack ? *ppos + 1 : -1UL;
 889         m->private = stack;
 890
 891         return stack;
 892 }
 893
 894 static unsigned long page_owner_pages_threshold;
 895
 896 static int stack_print(struct seq_file *m, void *v)
 897 {
 898         int i, nr_base_pages;
 899         struct stack *stack = v;
 900         unsigned long *entries;
 901         unsigned long nr_entries;
 902         struct stack_record *stack_record = stack->stack_record;
 903
 904         if (!stack->stack_record)
 905                 return 0;
 906
 907         nr_entries = stack_record->size;
 908         entries = stack_record->entries;
 909         nr_base_pages = refcount_read(&stack_record->count) - 1;
 910
 911         if (nr_base_pages < 1 || nr_base_pages < page_owner_pages_threshold)
 912                 return 0;
 913
 914         for (i = 0; i < nr_entries; i++)
 915                 seq_printf(m, " %pS\n", (void *)entries[i]);
 916         seq_printf(m, "nr_base_pages: %d\n\n", nr_base_pages);
 917
 918         return 0;
 919 }
 920
 921 static void stack_stop(struct seq_file *m, void *v)
 922 {
 923 }
 924
 925 static const struct seq_operations page_owner_stack_op = {
 926         .start  = stack_start,
 927         .next   = stack_next,
 928         .stop   = stack_stop,
 929         .show   = stack_print
 930 };
 931
 932 static int page_owner_stack_open(struct inode *inode, struct file *file)
 933 {
 934         return seq_open_private(file, &page_owner_stack_op, 0);
 935 }
 936
 937 static const struct file_operations page_owner_stack_operations = {
 938         .open           = page_owner_stack_open,
 939         .read           = seq_read,
 940         .llseek         = seq_lseek,
 941         .release        = seq_release,
 942 };
 943
 944 static int page_owner_threshold_get(void *data, u64 *val)
 945 {
 946         *val = READ_ONCE(page_owner_pages_threshold);
 947         return 0;
 948 }
 949
 950 static int page_owner_threshold_set(void *data, u64 val)
 951 {
 952         WRITE_ONCE(page_owner_pages_threshold, val);
 953         return 0;
 954 }
 955
 956 DEFINE_SIMPLE_ATTRIBUTE(proc_page_owner_threshold, &page_owner_threshold_get,
 957                         &page_owner_threshold_set, "%llu");
 958
 959
 960 static int __init pageowner_init(void)
 961 {
 962         struct dentry *dir;
 963
 964         if (!static_branch_unlikely(&page_owner_inited)) {
 965                 pr_info("page_owner is disabled\n");
 966                 return 0;
 967         }
 968
 969         debugfs_create_file("page_owner", 0400, NULL, NULL,
 970                             &proc_page_owner_operations);
 971         dir = debugfs_create_dir("page_owner_stacks", NULL);
 972         debugfs_create_file("show_stacks", 0400, dir, NULL,
 973                             &page_owner_stack_operations);
 974         debugfs_create_file("count_threshold", 0600, dir, NULL,
 975                             &proc_page_owner_threshold);
 976
 977         return 0;
 978 }
 979 late_initcall(pageowner_init)