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