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