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