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