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