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         u64 free_ts_nsec;
  31         pid_t pid;
  32 };
  33
  34 static bool page_owner_enabled = false;
  35 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  36
  37 static depot_stack_handle_t dummy_handle;
  38 static depot_stack_handle_t failure_handle;
  39 static depot_stack_handle_t early_handle;
  40
  41 static void init_early_allocated_pages(void);
  42
  43 static int __init early_page_owner_param(char *buf)
  44 {
  45         return kstrtobool(buf, &page_owner_enabled);
  46 }
  47 early_param("page_owner", early_page_owner_param);
  48
  49 static __init bool need_page_owner(void)
  50 {
  51         return page_owner_enabled;
  52 }
  53
  54 static __always_inline depot_stack_handle_t create_dummy_stack(void)
  55 {
  56         unsigned long entries[4];
  57         unsigned int nr_entries;
  58
  59         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  60         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
  61 }
  62
  63 static noinline void register_dummy_stack(void)
  64 {
  65         dummy_handle = create_dummy_stack();
  66 }
  67
  68 static noinline void register_failure_stack(void)
  69 {
  70         failure_handle = create_dummy_stack();
  71 }
  72
  73 static noinline void register_early_stack(void)
  74 {
  75         early_handle = create_dummy_stack();
  76 }
  77
  78 static __init void init_page_owner(void)
  79 {
  80         if (!page_owner_enabled)
  81                 return;
  82
  83         stack_depot_init();
  84
  85         register_dummy_stack();
  86         register_failure_stack();
  87         register_early_stack();
  88         static_branch_enable(&page_owner_inited);
  89         init_early_allocated_pages();
  90 }
  91
  92 struct page_ext_operations page_owner_ops = {
  93         .size = sizeof(struct page_owner),
  94         .need = need_page_owner,
  95         .init = init_page_owner,
  96 };
  97
  98 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
  99 {
 100         return (void *)page_ext + page_owner_ops.offset;
 101 }
 102
 103 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 104 {
 105         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 106         depot_stack_handle_t handle;
 107         unsigned int nr_entries;
 108
 109         /*
 110          * Avoid recursion.
 111          *
 112          * Sometimes page metadata allocation tracking requires more
 113          * memory to be allocated:
 114          * - when new stack trace is saved to stack depot
 115          * - when backtrace itself is calculated (ia64)
 116          */
 117         if (current->in_page_owner)
 118                 return dummy_handle;
 119         current->in_page_owner = 1;
 120
 121         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 122         handle = stack_depot_save(entries, nr_entries, flags);
 123         if (!handle)
 124                 handle = failure_handle;
 125
 126         current->in_page_owner = 0;
 127         return handle;
 128 }
 129
 130 void __reset_page_owner(struct page *page, unsigned short order)
 131 {
 132         int i;
 133         struct page_ext *page_ext;
 134         depot_stack_handle_t handle;
 135         struct page_owner *page_owner;
 136         u64 free_ts_nsec = local_clock();
 137
 138         page_ext = lookup_page_ext(page);
 139         if (unlikely(!page_ext))
 140                 return;
 141
 142         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 143         for (i = 0; i < (1 << order); i++) {
 144                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 145                 page_owner = get_page_owner(page_ext);
 146                 page_owner->free_handle = handle;
 147                 page_owner->free_ts_nsec = free_ts_nsec;
 148                 page_ext = page_ext_next(page_ext);
 149         }
 150 }
 151
 152 static inline void __set_page_owner_handle(struct page_ext *page_ext,
 153                                         depot_stack_handle_t handle,
 154                                         unsigned short order, gfp_t gfp_mask)
 155 {
 156         struct page_owner *page_owner;
 157         int i;
 158
 159         for (i = 0; i < (1 << order); i++) {
 160                 page_owner = get_page_owner(page_ext);
 161                 page_owner->handle = handle;
 162                 page_owner->order = order;
 163                 page_owner->gfp_mask = gfp_mask;
 164                 page_owner->last_migrate_reason = -1;
 165                 page_owner->pid = current->pid;
 166                 page_owner->ts_nsec = local_clock();
 167                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 168                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 169
 170                 page_ext = page_ext_next(page_ext);
 171         }
 172 }
 173
 174 noinline void __set_page_owner(struct page *page, unsigned short order,
 175                                         gfp_t gfp_mask)
 176 {
 177         struct page_ext *page_ext = lookup_page_ext(page);
 178         depot_stack_handle_t handle;
 179
 180         if (unlikely(!page_ext))
 181                 return;
 182
 183         handle = save_stack(gfp_mask);
 184         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 185 }
 186
 187 void __set_page_owner_migrate_reason(struct page *page, int reason)
 188 {
 189         struct page_ext *page_ext = lookup_page_ext(page);
 190         struct page_owner *page_owner;
 191
 192         if (unlikely(!page_ext))
 193                 return;
 194
 195         page_owner = get_page_owner(page_ext);
 196         page_owner->last_migrate_reason = reason;
 197 }
 198
 199 void __split_page_owner(struct page *page, unsigned int nr)
 200 {
 201         int i;
 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         for (i = 0; i < nr; i++) {
 209                 page_owner = get_page_owner(page_ext);
 210                 page_owner->order = 0;
 211                 page_ext = page_ext_next(page_ext);
 212         }
 213 }
 214
 215 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 216 {
 217         struct page_ext *old_ext = lookup_page_ext(&old->page);
 218         struct page_ext *new_ext = lookup_page_ext(&newfolio->page);
 219         struct page_owner *old_page_owner, *new_page_owner;
 220
 221         if (unlikely(!old_ext || !new_ext))
 222                 return;
 223
 224         old_page_owner = get_page_owner(old_ext);
 225         new_page_owner = get_page_owner(new_ext);
 226         new_page_owner->order = old_page_owner->order;
 227         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 228         new_page_owner->last_migrate_reason =
 229                 old_page_owner->last_migrate_reason;
 230         new_page_owner->handle = old_page_owner->handle;
 231         new_page_owner->pid = old_page_owner->pid;
 232         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
 233         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
 234
 235         /*
 236          * We don't clear the bit on the old folio as it's going to be freed
 237          * after migration. Until then, the info can be useful in case of
 238          * a bug, and the overall stats will be off a bit only temporarily.
 239          * Also, migrate_misplaced_transhuge_page() can still fail the
 240          * migration and then we want the old folio to retain the info. But
 241          * in that case we also don't need to explicitly clear the info from
 242          * the new page, which will be freed.
 243          */
 244         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 245         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
 246 }
 247
 248 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 249                                        pg_data_t *pgdat, struct zone *zone)
 250 {
 251         struct page *page;
 252         struct page_ext *page_ext;
 253         struct page_owner *page_owner;
 254         unsigned long pfn, block_end_pfn;
 255         unsigned long end_pfn = zone_end_pfn(zone);
 256         unsigned long count[MIGRATE_TYPES] = { 0, };
 257         int pageblock_mt, page_mt;
 258         int i;
 259
 260         /* Scan block by block. First and last block may be incomplete */
 261         pfn = zone->zone_start_pfn;
 262
 263         /*
 264          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 265          * a zone boundary, it will be double counted between zones. This does
 266          * not matter as the mixed block count will still be correct
 267          */
 268         for (; pfn < end_pfn; ) {
 269                 page = pfn_to_online_page(pfn);
 270                 if (!page) {
 271                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 272                         continue;
 273                 }
 274
 275                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 276                 block_end_pfn = min(block_end_pfn, end_pfn);
 277
 278                 pageblock_mt = get_pageblock_migratetype(page);
 279
 280                 for (; pfn < block_end_pfn; pfn++) {
 281                         /* The pageblock is online, no need to recheck. */
 282                         page = pfn_to_page(pfn);
 283
 284                         if (page_zone(page) != zone)
 285                                 continue;
 286
 287                         if (PageBuddy(page)) {
 288                                 unsigned long freepage_order;
 289
 290                                 freepage_order = buddy_order_unsafe(page);
 291                                 if (freepage_order < MAX_ORDER)
 292                                         pfn += (1UL << freepage_order) - 1;
 293                                 continue;
 294                         }
 295
 296                         if (PageReserved(page))
 297                                 continue;
 298
 299                         page_ext = lookup_page_ext(page);
 300                         if (unlikely(!page_ext))
 301                                 continue;
 302
 303                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 304                                 continue;
 305
 306                         page_owner = get_page_owner(page_ext);
 307                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 308                         if (pageblock_mt != page_mt) {
 309                                 if (is_migrate_cma(pageblock_mt))
 310                                         count[MIGRATE_MOVABLE]++;
 311                                 else
 312                                         count[pageblock_mt]++;
 313
 314                                 pfn = block_end_pfn;
 315                                 break;
 316                         }
 317                         pfn += (1UL << page_owner->order) - 1;
 318                 }
 319         }
 320
 321         /* Print counts */
 322         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 323         for (i = 0; i < MIGRATE_TYPES; i++)
 324                 seq_printf(m, "%12lu ", count[i]);
 325         seq_putc(m, '\n');
 326 }
 327
 328 static ssize_t
 329 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 330                 struct page *page, struct page_owner *page_owner,
 331                 depot_stack_handle_t handle)
 332 {
 333         int ret, pageblock_mt, page_mt;
 334         char *kbuf;
 335
 336         count = min_t(size_t, count, PAGE_SIZE);
 337         kbuf = kmalloc(count, GFP_KERNEL);
 338         if (!kbuf)
 339                 return -ENOMEM;
 340
 341         ret = snprintf(kbuf, count,
 342                         "Page allocated via order %u, mask %#x(%pGg), pid %d, ts %llu ns, free_ts %llu ns\n",
 343                         page_owner->order, page_owner->gfp_mask,
 344                         &page_owner->gfp_mask, page_owner->pid,
 345                         page_owner->ts_nsec, page_owner->free_ts_nsec);
 346
 347         if (ret >= count)
 348                 goto err;
 349
 350         /* Print information relevant to grouping pages by mobility */
 351         pageblock_mt = get_pageblock_migratetype(page);
 352         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 353         ret += snprintf(kbuf + ret, count - ret,
 354                         "PFN %lu type %s Block %lu type %s Flags %pGp\n",
 355                         pfn,
 356                         migratetype_names[page_mt],
 357                         pfn >> pageblock_order,
 358                         migratetype_names[pageblock_mt],
 359                         &page->flags);
 360
 361         if (ret >= count)
 362                 goto err;
 363
 364         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 365         if (ret >= count)
 366                 goto err;
 367
 368         if (page_owner->last_migrate_reason != -1) {
 369                 ret += snprintf(kbuf + ret, count - ret,
 370                         "Page has been migrated, last migrate reason: %s\n",
 371                         migrate_reason_names[page_owner->last_migrate_reason]);
 372                 if (ret >= count)
 373                         goto err;
 374         }
 375
 376         ret += snprintf(kbuf + ret, count - ret, "\n");
 377         if (ret >= count)
 378                 goto err;
 379
 380         if (copy_to_user(buf, kbuf, ret))
 381                 ret = -EFAULT;
 382
 383         kfree(kbuf);
 384         return ret;
 385
 386 err:
 387         kfree(kbuf);
 388         return -ENOMEM;
 389 }
 390
 391 void __dump_page_owner(const struct page *page)
 392 {
 393         struct page_ext *page_ext = lookup_page_ext(page);
 394         struct page_owner *page_owner;
 395         depot_stack_handle_t handle;
 396         gfp_t gfp_mask;
 397         int mt;
 398
 399         if (unlikely(!page_ext)) {
 400                 pr_alert("There is not page extension available.\n");
 401                 return;
 402         }
 403
 404         page_owner = get_page_owner(page_ext);
 405         gfp_mask = page_owner->gfp_mask;
 406         mt = gfp_migratetype(gfp_mask);
 407
 408         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 409                 pr_alert("page_owner info is not present (never set?)\n");
 410                 return;
 411         }
 412
 413         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 414                 pr_alert("page_owner tracks the page as allocated\n");
 415         else
 416                 pr_alert("page_owner tracks the page as freed\n");
 417
 418         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, ts %llu, free_ts %llu\n",
 419                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 420                  page_owner->pid, page_owner->ts_nsec, page_owner->free_ts_nsec);
 421
 422         handle = READ_ONCE(page_owner->handle);
 423         if (!handle)
 424                 pr_alert("page_owner allocation stack trace missing\n");
 425         else
 426                 stack_depot_print(handle);
 427
 428         handle = READ_ONCE(page_owner->free_handle);
 429         if (!handle) {
 430                 pr_alert("page_owner free stack trace missing\n");
 431         } else {
 432                 pr_alert("page last free stack trace:\n");
 433                 stack_depot_print(handle);
 434         }
 435
 436         if (page_owner->last_migrate_reason != -1)
 437                 pr_alert("page has been migrated, last migrate reason: %s\n",
 438                         migrate_reason_names[page_owner->last_migrate_reason]);
 439 }
 440
 441 static ssize_t
 442 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 443 {
 444         unsigned long pfn;
 445         struct page *page;
 446         struct page_ext *page_ext;
 447         struct page_owner *page_owner;
 448         depot_stack_handle_t handle;
 449
 450         if (!static_branch_unlikely(&page_owner_inited))
 451                 return -EINVAL;
 452
 453         page = NULL;
 454         pfn = min_low_pfn + *ppos;
 455
 456         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 457         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 458                 pfn++;
 459
 460         drain_all_pages(NULL);
 461
 462         /* Find an allocated page */
 463         for (; pfn < max_pfn; pfn++) {
 464                 /*
 465                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 466                  * validate the area as existing, skip it if not
 467                  */
 468                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 469                         pfn += MAX_ORDER_NR_PAGES - 1;
 470                         continue;
 471                 }
 472
 473                 page = pfn_to_page(pfn);
 474                 if (PageBuddy(page)) {
 475                         unsigned long freepage_order = buddy_order_unsafe(page);
 476
 477                         if (freepage_order < MAX_ORDER)
 478                                 pfn += (1UL << freepage_order) - 1;
 479                         continue;
 480                 }
 481
 482                 page_ext = lookup_page_ext(page);
 483                 if (unlikely(!page_ext))
 484                         continue;
 485
 486                 /*
 487                  * Some pages could be missed by concurrent allocation or free,
 488                  * because we don't hold the zone lock.
 489                  */
 490                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 491                         continue;
 492
 493                 /*
 494                  * Although we do have the info about past allocation of free
 495                  * pages, it's not relevant for current memory usage.
 496                  */
 497                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 498                         continue;
 499
 500                 page_owner = get_page_owner(page_ext);
 501
 502                 /*
 503                  * Don't print "tail" pages of high-order allocations as that
 504                  * would inflate the stats.
 505                  */
 506                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 507                         continue;
 508
 509                 /*
 510                  * Access to page_ext->handle isn't synchronous so we should
 511                  * be careful to access it.
 512                  */
 513                 handle = READ_ONCE(page_owner->handle);
 514                 if (!handle)
 515                         continue;
 516
 517                 /* Record the next PFN to read in the file offset */
 518                 *ppos = (pfn - min_low_pfn) + 1;
 519
 520                 return print_page_owner(buf, count, pfn, page,
 521                                 page_owner, handle);
 522         }
 523
 524         return 0;
 525 }
 526
 527 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 528 {
 529         unsigned long pfn = zone->zone_start_pfn;
 530         unsigned long end_pfn = zone_end_pfn(zone);
 531         unsigned long count = 0;
 532
 533         /*
 534          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 535          * a zone boundary, it will be double counted between zones. This does
 536          * not matter as the mixed block count will still be correct
 537          */
 538         for (; pfn < end_pfn; ) {
 539                 unsigned long block_end_pfn;
 540
 541                 if (!pfn_valid(pfn)) {
 542                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 543                         continue;
 544                 }
 545
 546                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 547                 block_end_pfn = min(block_end_pfn, end_pfn);
 548
 549                 for (; pfn < block_end_pfn; pfn++) {
 550                         struct page *page = pfn_to_page(pfn);
 551                         struct page_ext *page_ext;
 552
 553                         if (page_zone(page) != zone)
 554                                 continue;
 555
 556                         /*
 557                          * To avoid having to grab zone->lock, be a little
 558                          * careful when reading buddy page order. The only
 559                          * danger is that we skip too much and potentially miss
 560                          * some early allocated pages, which is better than
 561                          * heavy lock contention.
 562                          */
 563                         if (PageBuddy(page)) {
 564                                 unsigned long order = buddy_order_unsafe(page);
 565
 566                                 if (order > 0 && order < MAX_ORDER)
 567                                         pfn += (1UL << order) - 1;
 568                                 continue;
 569                         }
 570
 571                         if (PageReserved(page))
 572                                 continue;
 573
 574                         page_ext = lookup_page_ext(page);
 575                         if (unlikely(!page_ext))
 576                                 continue;
 577
 578                         /* Maybe overlapping zone */
 579                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 580                                 continue;
 581
 582                         /* Found early allocated page */
 583                         __set_page_owner_handle(page_ext, early_handle,
 584                                                 0, 0);
 585                         count++;
 586                 }
 587                 cond_resched();
 588         }
 589
 590         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 591                 pgdat->node_id, zone->name, count);
 592 }
 593
 594 static void init_zones_in_node(pg_data_t *pgdat)
 595 {
 596         struct zone *zone;
 597         struct zone *node_zones = pgdat->node_zones;
 598
 599         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 600                 if (!populated_zone(zone))
 601                         continue;
 602
 603                 init_pages_in_zone(pgdat, zone);
 604         }
 605 }
 606
 607 static void init_early_allocated_pages(void)
 608 {
 609         pg_data_t *pgdat;
 610
 611         for_each_online_pgdat(pgdat)
 612                 init_zones_in_node(pgdat);
 613 }
 614
 615 static const struct file_operations proc_page_owner_operations = {
 616         .read           = read_page_owner,
 617 };
 618
 619 static int __init pageowner_init(void)
 620 {
 621         if (!static_branch_unlikely(&page_owner_inited)) {
 622                 pr_info("page_owner is disabled\n");
 623                 return 0;
 624         }
 625
 626         debugfs_create_file("page_owner", 0400, NULL, NULL,
 627                             &proc_page_owner_operations);
 628
 629         return 0;
 630 }
 631 late_initcall(pageowner_init)