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