mm/page_owner.c

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