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 #ifdef CONFIG_DEBUG_PAGEALLOC
  28         depot_stack_handle_t free_handle;
  29 #endif
  30 };
  31
  32 static bool page_owner_disabled = true;
  33 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  34
  35 static depot_stack_handle_t dummy_handle;
  36 static depot_stack_handle_t failure_handle;
  37 static depot_stack_handle_t early_handle;
  38
  39 static void init_early_allocated_pages(void);
  40
  41 static int __init early_page_owner_param(char *buf)
  42 {
  43         if (!buf)
  44                 return -EINVAL;
  45
  46         if (strcmp(buf, "on") == 0)
  47                 page_owner_disabled = false;
  48
  49         return 0;
  50 }
  51 early_param("page_owner", early_page_owner_param);
  52
  53 static bool need_page_owner(void)
  54 {
  55         if (page_owner_disabled)
  56                 return false;
  57
  58         return true;
  59 }
  60
  61 static __always_inline depot_stack_handle_t create_dummy_stack(void)
  62 {
  63         unsigned long entries[4];
  64         unsigned int nr_entries;
  65
  66         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  67         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
  68 }
  69
  70 static noinline void register_dummy_stack(void)
  71 {
  72         dummy_handle = create_dummy_stack();
  73 }
  74
  75 static noinline void register_failure_stack(void)
  76 {
  77         failure_handle = create_dummy_stack();
  78 }
  79
  80 static noinline void register_early_stack(void)
  81 {
  82         early_handle = create_dummy_stack();
  83 }
  84
  85 static void init_page_owner(void)
  86 {
  87         if (page_owner_disabled)
  88                 return;
  89
  90         register_dummy_stack();
  91         register_failure_stack();
  92         register_early_stack();
  93         static_branch_enable(&page_owner_inited);
  94         init_early_allocated_pages();
  95 }
  96
  97 struct page_ext_operations page_owner_ops = {
  98         .size = sizeof(struct page_owner),
  99         .need = need_page_owner,
 100         .init = init_page_owner,
 101 };
 102
 103 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
 104 {
 105         return (void *)page_ext + page_owner_ops.offset;
 106 }
 107
 108 static inline bool check_recursive_alloc(unsigned long *entries,
 109                                          unsigned int nr_entries,
 110                                          unsigned long ip)
 111 {
 112         unsigned int i;
 113
 114         for (i = 0; i < nr_entries; i++) {
 115                 if (entries[i] == ip)
 116                         return true;
 117         }
 118         return false;
 119 }
 120
 121 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 122 {
 123         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 124         depot_stack_handle_t handle;
 125         unsigned int nr_entries;
 126
 127         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 128
 129         /*
 130          * We need to check recursion here because our request to
 131          * stackdepot could trigger memory allocation to save new
 132          * entry. New memory allocation would reach here and call
 133          * stack_depot_save_entries() again if we don't catch it. There is
 134          * still not enough memory in stackdepot so it would try to
 135          * allocate memory again and loop forever.
 136          */
 137         if (check_recursive_alloc(entries, nr_entries, _RET_IP_))
 138                 return dummy_handle;
 139
 140         handle = stack_depot_save(entries, nr_entries, flags);
 141         if (!handle)
 142                 handle = failure_handle;
 143
 144         return handle;
 145 }
 146
 147 void __reset_page_owner(struct page *page, unsigned int order)
 148 {
 149         int i;
 150         struct page_ext *page_ext;
 151 #ifdef CONFIG_DEBUG_PAGEALLOC
 152         depot_stack_handle_t handle = 0;
 153         struct page_owner *page_owner;
 154
 155         if (debug_pagealloc_enabled())
 156                 handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 157 #endif
 158
 159         page_ext = lookup_page_ext(page);
 160         if (unlikely(!page_ext))
 161                 return;
 162         for (i = 0; i < (1 << order); i++) {
 163                 __clear_bit(PAGE_EXT_OWNER_ACTIVE, &page_ext->flags);
 164 #ifdef CONFIG_DEBUG_PAGEALLOC
 165                 if (debug_pagealloc_enabled()) {
 166                         page_owner = get_page_owner(page_ext);
 167                         page_owner->free_handle = handle;
 168                 }
 169 #endif
 170                 page_ext = page_ext_next(page_ext);
 171         }
 172 }
 173
 174 static inline void __set_page_owner_handle(struct page *page,
 175         struct page_ext *page_ext, depot_stack_handle_t handle,
 176         unsigned int order, gfp_t gfp_mask)
 177 {
 178         struct page_owner *page_owner;
 179         int i;
 180
 181         for (i = 0; i < (1 << order); i++) {
 182                 page_owner = get_page_owner(page_ext);
 183                 page_owner->handle = handle;
 184                 page_owner->order = order;
 185                 page_owner->gfp_mask = gfp_mask;
 186                 page_owner->last_migrate_reason = -1;
 187                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 188                 __set_bit(PAGE_EXT_OWNER_ACTIVE, &page_ext->flags);
 189
 190                 page_ext = page_ext_next(page_ext);
 191         }
 192 }
 193
 194 noinline void __set_page_owner(struct page *page, unsigned int order,
 195                                         gfp_t gfp_mask)
 196 {
 197         struct page_ext *page_ext = lookup_page_ext(page);
 198         depot_stack_handle_t handle;
 199
 200         if (unlikely(!page_ext))
 201                 return;
 202
 203         handle = save_stack(gfp_mask);
 204         __set_page_owner_handle(page, page_ext, handle, order, gfp_mask);
 205 }
 206
 207 void __set_page_owner_migrate_reason(struct page *page, int reason)
 208 {
 209         struct page_ext *page_ext = lookup_page_ext(page);
 210         struct page_owner *page_owner;
 211
 212         if (unlikely(!page_ext))
 213                 return;
 214
 215         page_owner = get_page_owner(page_ext);
 216         page_owner->last_migrate_reason = reason;
 217 }
 218
 219 void __split_page_owner(struct page *page, unsigned int order)
 220 {
 221         int i;
 222         struct page_ext *page_ext = lookup_page_ext(page);
 223         struct page_owner *page_owner;
 224
 225         if (unlikely(!page_ext))
 226                 return;
 227
 228         for (i = 0; i < (1 << order); i++) {
 229                 page_owner = get_page_owner(page_ext);
 230                 page_owner->order = 0;
 231                 page_ext = page_ext_next(page_ext);
 232         }
 233 }
 234
 235 void __copy_page_owner(struct page *oldpage, struct page *newpage)
 236 {
 237         struct page_ext *old_ext = lookup_page_ext(oldpage);
 238         struct page_ext *new_ext = lookup_page_ext(newpage);
 239         struct page_owner *old_page_owner, *new_page_owner;
 240
 241         if (unlikely(!old_ext || !new_ext))
 242                 return;
 243
 244         old_page_owner = get_page_owner(old_ext);
 245         new_page_owner = get_page_owner(new_ext);
 246         new_page_owner->order = old_page_owner->order;
 247         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 248         new_page_owner->last_migrate_reason =
 249                 old_page_owner->last_migrate_reason;
 250         new_page_owner->handle = old_page_owner->handle;
 251
 252         /*
 253          * We don't clear the bit on the oldpage as it's going to be freed
 254          * after migration. Until then, the info can be useful in case of
 255          * a bug, and the overal stats will be off a bit only temporarily.
 256          * Also, migrate_misplaced_transhuge_page() can still fail the
 257          * migration and then we want the oldpage to retain the info. But
 258          * in that case we also don't need to explicitly clear the info from
 259          * the new page, which will be freed.
 260          */
 261         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 262         __set_bit(PAGE_EXT_OWNER_ACTIVE, &new_ext->flags);
 263 }
 264
 265 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 266                                        pg_data_t *pgdat, struct zone *zone)
 267 {
 268         struct page *page;
 269         struct page_ext *page_ext;
 270         struct page_owner *page_owner;
 271         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 272         unsigned long end_pfn = pfn + zone->spanned_pages;
 273         unsigned long count[MIGRATE_TYPES] = { 0, };
 274         int pageblock_mt, page_mt;
 275         int i;
 276
 277         /* Scan block by block. First and last block may be incomplete */
 278         pfn = zone->zone_start_pfn;
 279
 280         /*
 281          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 282          * a zone boundary, it will be double counted between zones. This does
 283          * not matter as the mixed block count will still be correct
 284          */
 285         for (; pfn < end_pfn; ) {
 286                 if (!pfn_valid(pfn)) {
 287                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 288                         continue;
 289                 }
 290
 291                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 292                 block_end_pfn = min(block_end_pfn, end_pfn);
 293
 294                 page = pfn_to_page(pfn);
 295                 pageblock_mt = get_pageblock_migratetype(page);
 296
 297                 for (; pfn < block_end_pfn; pfn++) {
 298                         if (!pfn_valid_within(pfn))
 299                                 continue;
 300
 301                         page = pfn_to_page(pfn);
 302
 303                         if (page_zone(page) != zone)
 304                                 continue;
 305
 306                         if (PageBuddy(page)) {
 307                                 unsigned long freepage_order;
 308
 309                                 freepage_order = page_order_unsafe(page);
 310                                 if (freepage_order < MAX_ORDER)
 311                                         pfn += (1UL << freepage_order) - 1;
 312                                 continue;
 313                         }
 314
 315                         if (PageReserved(page))
 316                                 continue;
 317
 318                         page_ext = lookup_page_ext(page);
 319                         if (unlikely(!page_ext))
 320                                 continue;
 321
 322                         if (!test_bit(PAGE_EXT_OWNER_ACTIVE, &page_ext->flags))
 323                                 continue;
 324
 325                         page_owner = get_page_owner(page_ext);
 326                         page_mt = gfpflags_to_migratetype(
 327                                         page_owner->gfp_mask);
 328                         if (pageblock_mt != page_mt) {
 329                                 if (is_migrate_cma(pageblock_mt))
 330                                         count[MIGRATE_MOVABLE]++;
 331                                 else
 332                                         count[pageblock_mt]++;
 333
 334                                 pfn = block_end_pfn;
 335                                 break;
 336                         }
 337                         pfn += (1UL << page_owner->order) - 1;
 338                 }
 339         }
 340
 341         /* Print counts */
 342         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 343         for (i = 0; i < MIGRATE_TYPES; i++)
 344                 seq_printf(m, "%12lu ", count[i]);
 345         seq_putc(m, '\n');
 346 }
 347
 348 static ssize_t
 349 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 350                 struct page *page, struct page_owner *page_owner,
 351                 depot_stack_handle_t handle)
 352 {
 353         int ret, pageblock_mt, page_mt;
 354         unsigned long *entries;
 355         unsigned int nr_entries;
 356         char *kbuf;
 357
 358         count = min_t(size_t, count, PAGE_SIZE);
 359         kbuf = kmalloc(count, GFP_KERNEL);
 360         if (!kbuf)
 361                 return -ENOMEM;
 362
 363         ret = snprintf(kbuf, count,
 364                         "Page allocated via order %u, mask %#x(%pGg)\n",
 365                         page_owner->order, page_owner->gfp_mask,
 366                         &page_owner->gfp_mask);
 367
 368         if (ret >= count)
 369                 goto err;
 370
 371         /* Print information relevant to grouping pages by mobility */
 372         pageblock_mt = get_pageblock_migratetype(page);
 373         page_mt  = gfpflags_to_migratetype(page_owner->gfp_mask);
 374         ret += snprintf(kbuf + ret, count - ret,
 375                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 376                         pfn,
 377                         migratetype_names[page_mt],
 378                         pfn >> pageblock_order,
 379                         migratetype_names[pageblock_mt],
 380                         page->flags, &page->flags);
 381
 382         if (ret >= count)
 383                 goto err;
 384
 385         nr_entries = stack_depot_fetch(handle, &entries);
 386         ret += stack_trace_snprint(kbuf + ret, count - ret, entries, nr_entries, 0);
 387         if (ret >= count)
 388                 goto err;
 389
 390         if (page_owner->last_migrate_reason != -1) {
 391                 ret += snprintf(kbuf + ret, count - ret,
 392                         "Page has been migrated, last migrate reason: %s\n",
 393                         migrate_reason_names[page_owner->last_migrate_reason]);
 394                 if (ret >= count)
 395                         goto err;
 396         }
 397
 398         ret += snprintf(kbuf + ret, count - ret, "\n");
 399         if (ret >= count)
 400                 goto err;
 401
 402         if (copy_to_user(buf, kbuf, ret))
 403                 ret = -EFAULT;
 404
 405         kfree(kbuf);
 406         return ret;
 407
 408 err:
 409         kfree(kbuf);
 410         return -ENOMEM;
 411 }
 412
 413 void __dump_page_owner(struct page *page)
 414 {
 415         struct page_ext *page_ext = lookup_page_ext(page);
 416         struct page_owner *page_owner;
 417         depot_stack_handle_t handle;
 418         unsigned long *entries;
 419         unsigned int nr_entries;
 420         gfp_t gfp_mask;
 421         int mt;
 422
 423         if (unlikely(!page_ext)) {
 424                 pr_alert("There is not page extension available.\n");
 425                 return;
 426         }
 427
 428         page_owner = get_page_owner(page_ext);
 429         gfp_mask = page_owner->gfp_mask;
 430         mt = gfpflags_to_migratetype(gfp_mask);
 431
 432         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 433                 pr_alert("page_owner info is not present (never set?)\n");
 434                 return;
 435         }
 436
 437         if (test_bit(PAGE_EXT_OWNER_ACTIVE, &page_ext->flags))
 438                 pr_alert("page_owner tracks the page as allocated\n");
 439         else
 440                 pr_alert("page_owner tracks the page as freed\n");
 441
 442         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg)\n",
 443                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask);
 444
 445         handle = READ_ONCE(page_owner->handle);
 446         if (!handle) {
 447                 pr_alert("page_owner allocation stack trace missing\n");
 448         } else {
 449                 nr_entries = stack_depot_fetch(handle, &entries);
 450                 stack_trace_print(entries, nr_entries, 0);
 451         }
 452
 453 #ifdef CONFIG_DEBUG_PAGEALLOC
 454         handle = READ_ONCE(page_owner->free_handle);
 455         if (!handle) {
 456                 pr_alert("page_owner free stack trace missing\n");
 457         } else {
 458                 nr_entries = stack_depot_fetch(handle, &entries);
 459                 pr_alert("page last free stack trace:\n");
 460                 stack_trace_print(entries, nr_entries, 0);
 461         }
 462 #endif
 463
 464         if (page_owner->last_migrate_reason != -1)
 465                 pr_alert("page has been migrated, last migrate reason: %s\n",
 466                         migrate_reason_names[page_owner->last_migrate_reason]);
 467 }
 468
 469 static ssize_t
 470 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 471 {
 472         unsigned long pfn;
 473         struct page *page;
 474         struct page_ext *page_ext;
 475         struct page_owner *page_owner;
 476         depot_stack_handle_t handle;
 477
 478         if (!static_branch_unlikely(&page_owner_inited))
 479                 return -EINVAL;
 480
 481         page = NULL;
 482         pfn = min_low_pfn + *ppos;
 483
 484         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 485         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 486                 pfn++;
 487
 488         drain_all_pages(NULL);
 489
 490         /* Find an allocated page */
 491         for (; pfn < max_pfn; pfn++) {
 492                 /*
 493                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 494                  * validate the area as existing, skip it if not
 495                  */
 496                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 497                         pfn += MAX_ORDER_NR_PAGES - 1;
 498                         continue;
 499                 }
 500
 501                 /* Check for holes within a MAX_ORDER area */
 502                 if (!pfn_valid_within(pfn))
 503                         continue;
 504
 505                 page = pfn_to_page(pfn);
 506                 if (PageBuddy(page)) {
 507                         unsigned long freepage_order = page_order_unsafe(page);
 508
 509                         if (freepage_order < MAX_ORDER)
 510                                 pfn += (1UL << freepage_order) - 1;
 511                         continue;
 512                 }
 513
 514                 page_ext = lookup_page_ext(page);
 515                 if (unlikely(!page_ext))
 516                         continue;
 517
 518                 /*
 519                  * Some pages could be missed by concurrent allocation or free,
 520                  * because we don't hold the zone lock.
 521                  */
 522                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 523                         continue;
 524
 525                 /*
 526                  * Although we do have the info about past allocation of free
 527                  * pages, it's not relevant for current memory usage.
 528                  */
 529                 if (!test_bit(PAGE_EXT_OWNER_ACTIVE, &page_ext->flags))
 530                         continue;
 531
 532                 page_owner = get_page_owner(page_ext);
 533
 534                 /*
 535                  * Don't print "tail" pages of high-order allocations as that
 536                  * would inflate the stats.
 537                  */
 538                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 539                         continue;
 540
 541                 /*
 542                  * Access to page_ext->handle isn't synchronous so we should
 543                  * be careful to access it.
 544                  */
 545                 handle = READ_ONCE(page_owner->handle);
 546                 if (!handle)
 547                         continue;
 548
 549                 /* Record the next PFN to read in the file offset */
 550                 *ppos = (pfn - min_low_pfn) + 1;
 551
 552                 return print_page_owner(buf, count, pfn, page,
 553                                 page_owner, handle);
 554         }
 555
 556         return 0;
 557 }
 558
 559 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 560 {
 561         unsigned long pfn = zone->zone_start_pfn;
 562         unsigned long end_pfn = zone_end_pfn(zone);
 563         unsigned long count = 0;
 564
 565         /*
 566          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 567          * a zone boundary, it will be double counted between zones. This does
 568          * not matter as the mixed block count will still be correct
 569          */
 570         for (; pfn < end_pfn; ) {
 571                 unsigned long block_end_pfn;
 572
 573                 if (!pfn_valid(pfn)) {
 574                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 575                         continue;
 576                 }
 577
 578                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 579                 block_end_pfn = min(block_end_pfn, end_pfn);
 580
 581                 for (; pfn < block_end_pfn; pfn++) {
 582                         struct page *page;
 583                         struct page_ext *page_ext;
 584
 585                         if (!pfn_valid_within(pfn))
 586                                 continue;
 587
 588                         page = pfn_to_page(pfn);
 589
 590                         if (page_zone(page) != zone)
 591                                 continue;
 592
 593                         /*
 594                          * To avoid having to grab zone->lock, be a little
 595                          * careful when reading buddy page order. The only
 596                          * danger is that we skip too much and potentially miss
 597                          * some early allocated pages, which is better than
 598                          * heavy lock contention.
 599                          */
 600                         if (PageBuddy(page)) {
 601                                 unsigned long order = page_order_unsafe(page);
 602
 603                                 if (order > 0 && order < MAX_ORDER)
 604                                         pfn += (1UL << order) - 1;
 605                                 continue;
 606                         }
 607
 608                         if (PageReserved(page))
 609                                 continue;
 610
 611                         page_ext = lookup_page_ext(page);
 612                         if (unlikely(!page_ext))
 613                                 continue;
 614
 615                         /* Maybe overlapping zone */
 616                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 617                                 continue;
 618
 619                         /* Found early allocated page */
 620                         __set_page_owner_handle(page, page_ext, early_handle,
 621                                                 0, 0);
 622                         count++;
 623                 }
 624                 cond_resched();
 625         }
 626
 627         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 628                 pgdat->node_id, zone->name, count);
 629 }
 630
 631 static void init_zones_in_node(pg_data_t *pgdat)
 632 {
 633         struct zone *zone;
 634         struct zone *node_zones = pgdat->node_zones;
 635
 636         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 637                 if (!populated_zone(zone))
 638                         continue;
 639
 640                 init_pages_in_zone(pgdat, zone);
 641         }
 642 }
 643
 644 static void init_early_allocated_pages(void)
 645 {
 646         pg_data_t *pgdat;
 647
 648         for_each_online_pgdat(pgdat)
 649                 init_zones_in_node(pgdat);
 650 }
 651
 652 static const struct file_operations proc_page_owner_operations = {
 653         .read           = read_page_owner,
 654 };
 655
 656 static int __init pageowner_init(void)
 657 {
 658         if (!static_branch_unlikely(&page_owner_inited)) {
 659                 pr_info("page_owner is disabled\n");
 660                 return 0;
 661         }
 662
 663         debugfs_create_file("page_owner", 0400, NULL, NULL,
 664                             &proc_page_owner_operations);
 665
 666         return 0;
 667 }
 668 late_initcall(pageowner_init)