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 #include <linux/memcontrol.h>
  14 #include <linux/sched/clock.h>
  15
  16 #include "internal.h"
  17
  18 /*
  19  * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
  20  * to use off stack temporal storage
  21  */
  22 #define PAGE_OWNER_STACK_DEPTH (16)
  23
  24 struct page_owner {
  25         unsigned short order;
  26         short last_migrate_reason;
  27         gfp_t gfp_mask;
  28         depot_stack_handle_t handle;
  29         depot_stack_handle_t free_handle;
  30         u64 ts_nsec;
  31         u64 free_ts_nsec;
  32         pid_t pid;
  33 };
  34
  35 static bool page_owner_enabled = false;
  36 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  37
  38 static depot_stack_handle_t dummy_handle;
  39 static depot_stack_handle_t failure_handle;
  40 static depot_stack_handle_t early_handle;
  41
  42 static void init_early_allocated_pages(void);
  43
  44 static int __init early_page_owner_param(char *buf)
  45 {
  46         return kstrtobool(buf, &page_owner_enabled);
  47 }
  48 early_param("page_owner", early_page_owner_param);
  49
  50 static __init bool need_page_owner(void)
  51 {
  52         return page_owner_enabled;
  53 }
  54
  55 static __always_inline depot_stack_handle_t create_dummy_stack(void)
  56 {
  57         unsigned long entries[4];
  58         unsigned int nr_entries;
  59
  60         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  61         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
  62 }
  63
  64 static noinline void register_dummy_stack(void)
  65 {
  66         dummy_handle = create_dummy_stack();
  67 }
  68
  69 static noinline void register_failure_stack(void)
  70 {
  71         failure_handle = create_dummy_stack();
  72 }
  73
  74 static noinline void register_early_stack(void)
  75 {
  76         early_handle = create_dummy_stack();
  77 }
  78
  79 static __init void init_page_owner(void)
  80 {
  81         if (!page_owner_enabled)
  82                 return;
  83
  84         stack_depot_init();
  85
  86         register_dummy_stack();
  87         register_failure_stack();
  88         register_early_stack();
  89         static_branch_enable(&page_owner_inited);
  90         init_early_allocated_pages();
  91 }
  92
  93 struct page_ext_operations page_owner_ops = {
  94         .size = sizeof(struct page_owner),
  95         .need = need_page_owner,
  96         .init = init_page_owner,
  97 };
  98
  99 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
 100 {
 101         return (void *)page_ext + page_owner_ops.offset;
 102 }
 103
 104 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 105 {
 106         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 107         depot_stack_handle_t handle;
 108         unsigned int nr_entries;
 109
 110         /*
 111          * Avoid recursion.
 112          *
 113          * Sometimes page metadata allocation tracking requires more
 114          * memory to be allocated:
 115          * - when new stack trace is saved to stack depot
 116          * - when backtrace itself is calculated (ia64)
 117          */
 118         if (current->in_page_owner)
 119                 return dummy_handle;
 120         current->in_page_owner = 1;
 121
 122         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 123         handle = stack_depot_save(entries, nr_entries, flags);
 124         if (!handle)
 125                 handle = failure_handle;
 126
 127         current->in_page_owner = 0;
 128         return handle;
 129 }
 130
 131 void __reset_page_owner(struct page *page, unsigned short order)
 132 {
 133         int i;
 134         struct page_ext *page_ext;
 135         depot_stack_handle_t handle;
 136         struct page_owner *page_owner;
 137         u64 free_ts_nsec = local_clock();
 138
 139         page_ext = lookup_page_ext(page);
 140         if (unlikely(!page_ext))
 141                 return;
 142
 143         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
 144         for (i = 0; i < (1 << order); i++) {
 145                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 146                 page_owner = get_page_owner(page_ext);
 147                 page_owner->free_handle = handle;
 148                 page_owner->free_ts_nsec = free_ts_nsec;
 149                 page_ext = page_ext_next(page_ext);
 150         }
 151 }
 152
 153 static inline void __set_page_owner_handle(struct page_ext *page_ext,
 154                                         depot_stack_handle_t handle,
 155                                         unsigned short order, gfp_t gfp_mask)
 156 {
 157         struct page_owner *page_owner;
 158         int i;
 159
 160         for (i = 0; i < (1 << order); i++) {
 161                 page_owner = get_page_owner(page_ext);
 162                 page_owner->handle = handle;
 163                 page_owner->order = order;
 164                 page_owner->gfp_mask = gfp_mask;
 165                 page_owner->last_migrate_reason = -1;
 166                 page_owner->pid = current->pid;
 167                 page_owner->ts_nsec = local_clock();
 168                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 169                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 170
 171                 page_ext = page_ext_next(page_ext);
 172         }
 173 }
 174
 175 noinline void __set_page_owner(struct page *page, unsigned short order,
 176                                         gfp_t gfp_mask)
 177 {
 178         struct page_ext *page_ext = lookup_page_ext(page);
 179         depot_stack_handle_t handle;
 180
 181         if (unlikely(!page_ext))
 182                 return;
 183
 184         handle = save_stack(gfp_mask);
 185         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 186 }
 187
 188 void __set_page_owner_migrate_reason(struct page *page, int reason)
 189 {
 190         struct page_ext *page_ext = lookup_page_ext(page);
 191         struct page_owner *page_owner;
 192
 193         if (unlikely(!page_ext))
 194                 return;
 195
 196         page_owner = get_page_owner(page_ext);
 197         page_owner->last_migrate_reason = reason;
 198 }
 199
 200 void __split_page_owner(struct page *page, unsigned int nr)
 201 {
 202         int i;
 203         struct page_ext *page_ext = lookup_page_ext(page);
 204         struct page_owner *page_owner;
 205
 206         if (unlikely(!page_ext))
 207                 return;
 208
 209         for (i = 0; i < nr; i++) {
 210                 page_owner = get_page_owner(page_ext);
 211                 page_owner->order = 0;
 212                 page_ext = page_ext_next(page_ext);
 213         }
 214 }
 215
 216 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
 217 {
 218         struct page_ext *old_ext = lookup_page_ext(&old->page);
 219         struct page_ext *new_ext = lookup_page_ext(&newfolio->page);
 220         struct page_owner *old_page_owner, *new_page_owner;
 221
 222         if (unlikely(!old_ext || !new_ext))
 223                 return;
 224
 225         old_page_owner = get_page_owner(old_ext);
 226         new_page_owner = get_page_owner(new_ext);
 227         new_page_owner->order = old_page_owner->order;
 228         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 229         new_page_owner->last_migrate_reason =
 230                 old_page_owner->last_migrate_reason;
 231         new_page_owner->handle = old_page_owner->handle;
 232         new_page_owner->pid = old_page_owner->pid;
 233         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
 234         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
 235
 236         /*
 237          * We don't clear the bit on the old folio as it's going to be freed
 238          * after migration. Until then, the info can be useful in case of
 239          * a bug, and the overall stats will be off a bit only temporarily.
 240          * Also, migrate_misplaced_transhuge_page() can still fail the
 241          * migration and then we want the old folio to retain the info. But
 242          * in that case we also don't need to explicitly clear the info from
 243          * the new page, which will be freed.
 244          */
 245         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 246         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
 247 }
 248
 249 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 250                                        pg_data_t *pgdat, struct zone *zone)
 251 {
 252         struct page *page;
 253         struct page_ext *page_ext;
 254         struct page_owner *page_owner;
 255         unsigned long pfn, block_end_pfn;
 256         unsigned long end_pfn = zone_end_pfn(zone);
 257         unsigned long count[MIGRATE_TYPES] = { 0, };
 258         int pageblock_mt, page_mt;
 259         int i;
 260
 261         /* Scan block by block. First and last block may be incomplete */
 262         pfn = zone->zone_start_pfn;
 263
 264         /*
 265          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 266          * a zone boundary, it will be double counted between zones. This does
 267          * not matter as the mixed block count will still be correct
 268          */
 269         for (; pfn < end_pfn; ) {
 270                 page = pfn_to_online_page(pfn);
 271                 if (!page) {
 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                 pageblock_mt = get_pageblock_migratetype(page);
 280
 281                 for (; pfn < block_end_pfn; pfn++) {
 282                         /* The pageblock is online, no need to recheck. */
 283                         page = pfn_to_page(pfn);
 284
 285                         if (page_zone(page) != zone)
 286                                 continue;
 287
 288                         if (PageBuddy(page)) {
 289                                 unsigned long freepage_order;
 290
 291                                 freepage_order = buddy_order_unsafe(page);
 292                                 if (freepage_order < MAX_ORDER)
 293                                         pfn += (1UL << freepage_order) - 1;
 294                                 continue;
 295                         }
 296
 297                         if (PageReserved(page))
 298                                 continue;
 299
 300                         page_ext = lookup_page_ext(page);
 301                         if (unlikely(!page_ext))
 302                                 continue;
 303
 304                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 305                                 continue;
 306
 307                         page_owner = get_page_owner(page_ext);
 308                         page_mt = gfp_migratetype(page_owner->gfp_mask);
 309                         if (pageblock_mt != page_mt) {
 310                                 if (is_migrate_cma(pageblock_mt))
 311                                         count[MIGRATE_MOVABLE]++;
 312                                 else
 313                                         count[pageblock_mt]++;
 314
 315                                 pfn = block_end_pfn;
 316                                 break;
 317                         }
 318                         pfn += (1UL << page_owner->order) - 1;
 319                 }
 320         }
 321
 322         /* Print counts */
 323         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 324         for (i = 0; i < MIGRATE_TYPES; i++)
 325                 seq_printf(m, "%12lu ", count[i]);
 326         seq_putc(m, '\n');
 327 }
 328
 329 /*
 330  * Looking for memcg information and print it out
 331  */
 332 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
 333                                          struct page *page)
 334 {
 335 #ifdef CONFIG_MEMCG
 336         unsigned long memcg_data;
 337         struct mem_cgroup *memcg;
 338         bool online;
 339         char name[80];
 340
 341         rcu_read_lock();
 342         memcg_data = READ_ONCE(page->memcg_data);
 343         if (!memcg_data)
 344                 goto out_unlock;
 345
 346         if (memcg_data & MEMCG_DATA_OBJCGS)
 347                 ret += scnprintf(kbuf + ret, count - ret,
 348                                 "Slab cache page\n");
 349
 350         memcg = page_memcg_check(page);
 351         if (!memcg)
 352                 goto out_unlock;
 353
 354         online = (memcg->css.flags & CSS_ONLINE);
 355         cgroup_name(memcg->css.cgroup, name, sizeof(name));
 356         ret += scnprintf(kbuf + ret, count - ret,
 357                         "Charged %sto %smemcg %s\n",
 358                         PageMemcgKmem(page) ? "(via objcg) " : "",
 359                         online ? "" : "offline ",
 360                         name);
 361 out_unlock:
 362         rcu_read_unlock();
 363 #endif /* CONFIG_MEMCG */
 364
 365         return ret;
 366 }
 367
 368 static ssize_t
 369 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 370                 struct page *page, struct page_owner *page_owner,
 371                 depot_stack_handle_t handle)
 372 {
 373         int ret, pageblock_mt, page_mt;
 374         char *kbuf;
 375
 376         count = min_t(size_t, count, PAGE_SIZE);
 377         kbuf = kmalloc(count, GFP_KERNEL);
 378         if (!kbuf)
 379                 return -ENOMEM;
 380
 381         ret = scnprintf(kbuf, count,
 382                         "Page allocated via order %u, mask %#x(%pGg), pid %d, ts %llu ns, free_ts %llu ns\n",
 383                         page_owner->order, page_owner->gfp_mask,
 384                         &page_owner->gfp_mask, page_owner->pid,
 385                         page_owner->ts_nsec, page_owner->free_ts_nsec);
 386
 387         /* Print information relevant to grouping pages by mobility */
 388         pageblock_mt = get_pageblock_migratetype(page);
 389         page_mt  = gfp_migratetype(page_owner->gfp_mask);
 390         ret += scnprintf(kbuf + ret, count - ret,
 391                         "PFN %lu type %s Block %lu type %s Flags %pGp\n",
 392                         pfn,
 393                         migratetype_names[page_mt],
 394                         pfn >> pageblock_order,
 395                         migratetype_names[pageblock_mt],
 396                         &page->flags);
 397
 398         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 399         if (ret >= count)
 400                 goto err;
 401
 402         if (page_owner->last_migrate_reason != -1) {
 403                 ret += scnprintf(kbuf + ret, count - ret,
 404                         "Page has been migrated, last migrate reason: %s\n",
 405                         migrate_reason_names[page_owner->last_migrate_reason]);
 406         }
 407
 408         ret = print_page_owner_memcg(kbuf, count, ret, page);
 409
 410         ret += snprintf(kbuf + ret, count - ret, "\n");
 411         if (ret >= count)
 412                 goto err;
 413
 414         if (copy_to_user(buf, kbuf, ret))
 415                 ret = -EFAULT;
 416
 417         kfree(kbuf);
 418         return ret;
 419
 420 err:
 421         kfree(kbuf);
 422         return -ENOMEM;
 423 }
 424
 425 void __dump_page_owner(const struct page *page)
 426 {
 427         struct page_ext *page_ext = lookup_page_ext(page);
 428         struct page_owner *page_owner;
 429         depot_stack_handle_t handle;
 430         gfp_t gfp_mask;
 431         int mt;
 432
 433         if (unlikely(!page_ext)) {
 434                 pr_alert("There is not page extension available.\n");
 435                 return;
 436         }
 437
 438         page_owner = get_page_owner(page_ext);
 439         gfp_mask = page_owner->gfp_mask;
 440         mt = gfp_migratetype(gfp_mask);
 441
 442         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 443                 pr_alert("page_owner info is not present (never set?)\n");
 444                 return;
 445         }
 446
 447         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 448                 pr_alert("page_owner tracks the page as allocated\n");
 449         else
 450                 pr_alert("page_owner tracks the page as freed\n");
 451
 452         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, ts %llu, free_ts %llu\n",
 453                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
 454                  page_owner->pid, page_owner->ts_nsec, page_owner->free_ts_nsec);
 455
 456         handle = READ_ONCE(page_owner->handle);
 457         if (!handle)
 458                 pr_alert("page_owner allocation stack trace missing\n");
 459         else
 460                 stack_depot_print(handle);
 461
 462         handle = READ_ONCE(page_owner->free_handle);
 463         if (!handle) {
 464                 pr_alert("page_owner free stack trace missing\n");
 465         } else {
 466                 pr_alert("page last free stack trace:\n");
 467                 stack_depot_print(handle);
 468         }
 469
 470         if (page_owner->last_migrate_reason != -1)
 471                 pr_alert("page has been migrated, last migrate reason: %s\n",
 472                         migrate_reason_names[page_owner->last_migrate_reason]);
 473 }
 474
 475 static ssize_t
 476 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 477 {
 478         unsigned long pfn;
 479         struct page *page;
 480         struct page_ext *page_ext;
 481         struct page_owner *page_owner;
 482         depot_stack_handle_t handle;
 483
 484         if (!static_branch_unlikely(&page_owner_inited))
 485                 return -EINVAL;
 486
 487         page = NULL;
 488         pfn = min_low_pfn + *ppos;
 489
 490         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 491         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 492                 pfn++;
 493
 494         drain_all_pages(NULL);
 495
 496         /* Find an allocated page */
 497         for (; pfn < max_pfn; pfn++) {
 498                 /*
 499                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 500                  * validate the area as existing, skip it if not
 501                  */
 502                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 503                         pfn += MAX_ORDER_NR_PAGES - 1;
 504                         continue;
 505                 }
 506
 507                 page = pfn_to_page(pfn);
 508                 if (PageBuddy(page)) {
 509                         unsigned long freepage_order = buddy_order_unsafe(page);
 510
 511                         if (freepage_order < MAX_ORDER)
 512                                 pfn += (1UL << freepage_order) - 1;
 513                         continue;
 514                 }
 515
 516                 page_ext = lookup_page_ext(page);
 517                 if (unlikely(!page_ext))
 518                         continue;
 519
 520                 /*
 521                  * Some pages could be missed by concurrent allocation or free,
 522                  * because we don't hold the zone lock.
 523                  */
 524                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 525                         continue;
 526
 527                 /*
 528                  * Although we do have the info about past allocation of free
 529                  * pages, it's not relevant for current memory usage.
 530                  */
 531                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
 532                         continue;
 533
 534                 page_owner = get_page_owner(page_ext);
 535
 536                 /*
 537                  * Don't print "tail" pages of high-order allocations as that
 538                  * would inflate the stats.
 539                  */
 540                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
 541                         continue;
 542
 543                 /*
 544                  * Access to page_ext->handle isn't synchronous so we should
 545                  * be careful to access it.
 546                  */
 547                 handle = READ_ONCE(page_owner->handle);
 548                 if (!handle)
 549                         continue;
 550
 551                 /* Record the next PFN to read in the file offset */
 552                 *ppos = (pfn - min_low_pfn) + 1;
 553
 554                 return print_page_owner(buf, count, pfn, page,
 555                                 page_owner, handle);
 556         }
 557
 558         return 0;
 559 }
 560
 561 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 562 {
 563         unsigned long pfn = zone->zone_start_pfn;
 564         unsigned long end_pfn = zone_end_pfn(zone);
 565         unsigned long count = 0;
 566
 567         /*
 568          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 569          * a zone boundary, it will be double counted between zones. This does
 570          * not matter as the mixed block count will still be correct
 571          */
 572         for (; pfn < end_pfn; ) {
 573                 unsigned long block_end_pfn;
 574
 575                 if (!pfn_valid(pfn)) {
 576                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 577                         continue;
 578                 }
 579
 580                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 581                 block_end_pfn = min(block_end_pfn, end_pfn);
 582
 583                 for (; pfn < block_end_pfn; pfn++) {
 584                         struct page *page = pfn_to_page(pfn);
 585                         struct page_ext *page_ext;
 586
 587                         if (page_zone(page) != zone)
 588                                 continue;
 589
 590                         /*
 591                          * To avoid having to grab zone->lock, be a little
 592                          * careful when reading buddy page order. The only
 593                          * danger is that we skip too much and potentially miss
 594                          * some early allocated pages, which is better than
 595                          * heavy lock contention.
 596                          */
 597                         if (PageBuddy(page)) {
 598                                 unsigned long order = buddy_order_unsafe(page);
 599
 600                                 if (order > 0 && order < MAX_ORDER)
 601                                         pfn += (1UL << order) - 1;
 602                                 continue;
 603                         }
 604
 605                         if (PageReserved(page))
 606                                 continue;
 607
 608                         page_ext = lookup_page_ext(page);
 609                         if (unlikely(!page_ext))
 610                                 continue;
 611
 612                         /* Maybe overlapping zone */
 613                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 614                                 continue;
 615
 616                         /* Found early allocated page */
 617                         __set_page_owner_handle(page_ext, early_handle,
 618                                                 0, 0);
 619                         count++;
 620                 }
 621                 cond_resched();
 622         }
 623
 624         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 625                 pgdat->node_id, zone->name, count);
 626 }
 627
 628 static void init_zones_in_node(pg_data_t *pgdat)
 629 {
 630         struct zone *zone;
 631         struct zone *node_zones = pgdat->node_zones;
 632
 633         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 634                 if (!populated_zone(zone))
 635                         continue;
 636
 637                 init_pages_in_zone(pgdat, zone);
 638         }
 639 }
 640
 641 static void init_early_allocated_pages(void)
 642 {
 643         pg_data_t *pgdat;
 644
 645         for_each_online_pgdat(pgdat)
 646                 init_zones_in_node(pgdat);
 647 }
 648
 649 static const struct file_operations proc_page_owner_operations = {
 650         .read           = read_page_owner,
 651 };
 652
 653 static int __init pageowner_init(void)
 654 {
 655         if (!static_branch_unlikely(&page_owner_inited)) {
 656                 pr_info("page_owner is disabled\n");
 657                 return 0;
 658         }
 659
 660         debugfs_create_file("page_owner", 0400, NULL, NULL,
 661                             &proc_page_owner_operations);
 662
 663         return 0;
 664 }
 665 late_initcall(pageowner_init)