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, &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_ext *page_ext,
 158         depot_stack_handle_t handle, unsigned int order, gfp_t gfp_mask)
 159 {
 160         struct page_owner *page_owner;
 161
 162         page_owner = get_page_owner(page_ext);
 163         page_owner->handle = handle;
 164         page_owner->order = order;
 165         page_owner->gfp_mask = gfp_mask;
 166         page_owner->last_migrate_reason = -1;
 167
 168         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 169 }
 170
 171 noinline void __set_page_owner(struct page *page, unsigned int order,
 172                                         gfp_t gfp_mask)
 173 {
 174         struct page_ext *page_ext = lookup_page_ext(page);
 175         depot_stack_handle_t handle;
 176
 177         if (unlikely(!page_ext))
 178                 return;
 179
 180         handle = save_stack(gfp_mask);
 181         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
 182 }
 183
 184 void __set_page_owner_migrate_reason(struct page *page, int reason)
 185 {
 186         struct page_ext *page_ext = lookup_page_ext(page);
 187         struct page_owner *page_owner;
 188
 189         if (unlikely(!page_ext))
 190                 return;
 191
 192         page_owner = get_page_owner(page_ext);
 193         page_owner->last_migrate_reason = reason;
 194 }
 195
 196 void __split_page_owner(struct page *page, unsigned int order)
 197 {
 198         int i;
 199         struct page_ext *page_ext = lookup_page_ext(page);
 200         struct page_owner *page_owner;
 201
 202         if (unlikely(!page_ext))
 203                 return;
 204
 205         page_owner = get_page_owner(page_ext);
 206         page_owner->order = 0;
 207         for (i = 1; i < (1 << order); i++)
 208                 __copy_page_owner(page, page + i);
 209 }
 210
 211 void __copy_page_owner(struct page *oldpage, struct page *newpage)
 212 {
 213         struct page_ext *old_ext = lookup_page_ext(oldpage);
 214         struct page_ext *new_ext = lookup_page_ext(newpage);
 215         struct page_owner *old_page_owner, *new_page_owner;
 216
 217         if (unlikely(!old_ext || !new_ext))
 218                 return;
 219
 220         old_page_owner = get_page_owner(old_ext);
 221         new_page_owner = get_page_owner(new_ext);
 222         new_page_owner->order = old_page_owner->order;
 223         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
 224         new_page_owner->last_migrate_reason =
 225                 old_page_owner->last_migrate_reason;
 226         new_page_owner->handle = old_page_owner->handle;
 227
 228         /*
 229          * We don't clear the bit on the oldpage as it's going to be freed
 230          * after migration. Until then, the info can be useful in case of
 231          * a bug, and the overal stats will be off a bit only temporarily.
 232          * Also, migrate_misplaced_transhuge_page() can still fail the
 233          * migration and then we want the oldpage to retain the info. But
 234          * in that case we also don't need to explicitly clear the info from
 235          * the new page, which will be freed.
 236          */
 237         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 238 }
 239
 240 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
 241                                        pg_data_t *pgdat, struct zone *zone)
 242 {
 243         struct page *page;
 244         struct page_ext *page_ext;
 245         struct page_owner *page_owner;
 246         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 247         unsigned long end_pfn = pfn + zone->spanned_pages;
 248         unsigned long count[MIGRATE_TYPES] = { 0, };
 249         int pageblock_mt, page_mt;
 250         int i;
 251
 252         /* Scan block by block. First and last block may be incomplete */
 253         pfn = zone->zone_start_pfn;
 254
 255         /*
 256          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 257          * a zone boundary, it will be double counted between zones. This does
 258          * not matter as the mixed block count will still be correct
 259          */
 260         for (; pfn < end_pfn; ) {
 261                 page = pfn_to_online_page(pfn);
 262                 if (!page) {
 263                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 264                         continue;
 265                 }
 266
 267                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 268                 block_end_pfn = min(block_end_pfn, end_pfn);
 269
 270                 pageblock_mt = get_pageblock_migratetype(page);
 271
 272                 for (; pfn < block_end_pfn; pfn++) {
 273                         if (!pfn_valid_within(pfn))
 274                                 continue;
 275
 276                         /* The pageblock is online, no need to recheck. */
 277                         page = pfn_to_page(pfn);
 278
 279                         if (page_zone(page) != zone)
 280                                 continue;
 281
 282                         if (PageBuddy(page)) {
 283                                 unsigned long freepage_order;
 284
 285                                 freepage_order = page_order_unsafe(page);
 286                                 if (freepage_order < MAX_ORDER)
 287                                         pfn += (1UL << freepage_order) - 1;
 288                                 continue;
 289                         }
 290
 291                         if (PageReserved(page))
 292                                 continue;
 293
 294                         page_ext = lookup_page_ext(page);
 295                         if (unlikely(!page_ext))
 296                                 continue;
 297
 298                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 299                                 continue;
 300
 301                         page_owner = get_page_owner(page_ext);
 302                         page_mt = gfpflags_to_migratetype(
 303                                         page_owner->gfp_mask);
 304                         if (pageblock_mt != page_mt) {
 305                                 if (is_migrate_cma(pageblock_mt))
 306                                         count[MIGRATE_MOVABLE]++;
 307                                 else
 308                                         count[pageblock_mt]++;
 309
 310                                 pfn = block_end_pfn;
 311                                 break;
 312                         }
 313                         pfn += (1UL << page_owner->order) - 1;
 314                 }
 315         }
 316
 317         /* Print counts */
 318         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 319         for (i = 0; i < MIGRATE_TYPES; i++)
 320                 seq_printf(m, "%12lu ", count[i]);
 321         seq_putc(m, '\n');
 322 }
 323
 324 static ssize_t
 325 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 326                 struct page *page, struct page_owner *page_owner,
 327                 depot_stack_handle_t handle)
 328 {
 329         int ret, pageblock_mt, page_mt;
 330         unsigned long *entries;
 331         unsigned int nr_entries;
 332         char *kbuf;
 333
 334         count = min_t(size_t, count, PAGE_SIZE);
 335         kbuf = kmalloc(count, GFP_KERNEL);
 336         if (!kbuf)
 337                 return -ENOMEM;
 338
 339         ret = snprintf(kbuf, count,
 340                         "Page allocated via order %u, mask %#x(%pGg)\n",
 341                         page_owner->order, page_owner->gfp_mask,
 342                         &page_owner->gfp_mask);
 343
 344         if (ret >= count)
 345                 goto err;
 346
 347         /* Print information relevant to grouping pages by mobility */
 348         pageblock_mt = get_pageblock_migratetype(page);
 349         page_mt  = gfpflags_to_migratetype(page_owner->gfp_mask);
 350         ret += snprintf(kbuf + ret, count - ret,
 351                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 352                         pfn,
 353                         migratetype_names[page_mt],
 354                         pfn >> pageblock_order,
 355                         migratetype_names[pageblock_mt],
 356                         page->flags, &page->flags);
 357
 358         if (ret >= count)
 359                 goto err;
 360
 361         nr_entries = stack_depot_fetch(handle, &entries);
 362         ret += stack_trace_snprint(kbuf + ret, count - ret, entries, nr_entries, 0);
 363         if (ret >= count)
 364                 goto err;
 365
 366         if (page_owner->last_migrate_reason != -1) {
 367                 ret += snprintf(kbuf + ret, count - ret,
 368                         "Page has been migrated, last migrate reason: %s\n",
 369                         migrate_reason_names[page_owner->last_migrate_reason]);
 370                 if (ret >= count)
 371                         goto err;
 372         }
 373
 374         ret += snprintf(kbuf + ret, count - ret, "\n");
 375         if (ret >= count)
 376                 goto err;
 377
 378         if (copy_to_user(buf, kbuf, ret))
 379                 ret = -EFAULT;
 380
 381         kfree(kbuf);
 382         return ret;
 383
 384 err:
 385         kfree(kbuf);
 386         return -ENOMEM;
 387 }
 388
 389 void __dump_page_owner(struct page *page)
 390 {
 391         struct page_ext *page_ext = lookup_page_ext(page);
 392         struct page_owner *page_owner;
 393         depot_stack_handle_t handle;
 394         unsigned long *entries;
 395         unsigned int nr_entries;
 396         gfp_t gfp_mask;
 397         int mt;
 398
 399         if (unlikely(!page_ext)) {
 400                 pr_alert("There is not page extension available.\n");
 401                 return;
 402         }
 403
 404         page_owner = get_page_owner(page_ext);
 405         gfp_mask = page_owner->gfp_mask;
 406         mt = gfpflags_to_migratetype(gfp_mask);
 407
 408         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 409                 pr_alert("page_owner info is not active (free page?)\n");
 410                 return;
 411         }
 412
 413         handle = READ_ONCE(page_owner->handle);
 414         if (!handle) {
 415                 pr_alert("page_owner info is not active (free page?)\n");
 416                 return;
 417         }
 418
 419         nr_entries = stack_depot_fetch(handle, &entries);
 420         pr_alert("page allocated via order %u, migratetype %s, gfp_mask %#x(%pGg)\n",
 421                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask);
 422         stack_trace_print(entries, nr_entries, 0);
 423
 424         if (page_owner->last_migrate_reason != -1)
 425                 pr_alert("page has been migrated, last migrate reason: %s\n",
 426                         migrate_reason_names[page_owner->last_migrate_reason]);
 427 }
 428
 429 static ssize_t
 430 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 431 {
 432         unsigned long pfn;
 433         struct page *page;
 434         struct page_ext *page_ext;
 435         struct page_owner *page_owner;
 436         depot_stack_handle_t handle;
 437
 438         if (!static_branch_unlikely(&page_owner_inited))
 439                 return -EINVAL;
 440
 441         page = NULL;
 442         pfn = min_low_pfn + *ppos;
 443
 444         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 445         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 446                 pfn++;
 447
 448         drain_all_pages(NULL);
 449
 450         /* Find an allocated page */
 451         for (; pfn < max_pfn; pfn++) {
 452                 /*
 453                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 454                  * validate the area as existing, skip it if not
 455                  */
 456                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 457                         pfn += MAX_ORDER_NR_PAGES - 1;
 458                         continue;
 459                 }
 460
 461                 /* Check for holes within a MAX_ORDER area */
 462                 if (!pfn_valid_within(pfn))
 463                         continue;
 464
 465                 page = pfn_to_page(pfn);
 466                 if (PageBuddy(page)) {
 467                         unsigned long freepage_order = page_order_unsafe(page);
 468
 469                         if (freepage_order < MAX_ORDER)
 470                                 pfn += (1UL << freepage_order) - 1;
 471                         continue;
 472                 }
 473
 474                 page_ext = lookup_page_ext(page);
 475                 if (unlikely(!page_ext))
 476                         continue;
 477
 478                 /*
 479                  * Some pages could be missed by concurrent allocation or free,
 480                  * because we don't hold the zone lock.
 481                  */
 482                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 483                         continue;
 484
 485                 page_owner = get_page_owner(page_ext);
 486
 487                 /*
 488                  * Access to page_ext->handle isn't synchronous so we should
 489                  * be careful to access it.
 490                  */
 491                 handle = READ_ONCE(page_owner->handle);
 492                 if (!handle)
 493                         continue;
 494
 495                 /* Record the next PFN to read in the file offset */
 496                 *ppos = (pfn - min_low_pfn) + 1;
 497
 498                 return print_page_owner(buf, count, pfn, page,
 499                                 page_owner, handle);
 500         }
 501
 502         return 0;
 503 }
 504
 505 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 506 {
 507         unsigned long pfn = zone->zone_start_pfn;
 508         unsigned long end_pfn = zone_end_pfn(zone);
 509         unsigned long count = 0;
 510
 511         /*
 512          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 513          * a zone boundary, it will be double counted between zones. This does
 514          * not matter as the mixed block count will still be correct
 515          */
 516         for (; pfn < end_pfn; ) {
 517                 unsigned long block_end_pfn;
 518
 519                 if (!pfn_valid(pfn)) {
 520                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 521                         continue;
 522                 }
 523
 524                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 525                 block_end_pfn = min(block_end_pfn, end_pfn);
 526
 527                 for (; pfn < block_end_pfn; pfn++) {
 528                         struct page *page;
 529                         struct page_ext *page_ext;
 530
 531                         if (!pfn_valid_within(pfn))
 532                                 continue;
 533
 534                         page = pfn_to_page(pfn);
 535
 536                         if (page_zone(page) != zone)
 537                                 continue;
 538
 539                         /*
 540                          * To avoid having to grab zone->lock, be a little
 541                          * careful when reading buddy page order. The only
 542                          * danger is that we skip too much and potentially miss
 543                          * some early allocated pages, which is better than
 544                          * heavy lock contention.
 545                          */
 546                         if (PageBuddy(page)) {
 547                                 unsigned long order = page_order_unsafe(page);
 548
 549                                 if (order > 0 && order < MAX_ORDER)
 550                                         pfn += (1UL << order) - 1;
 551                                 continue;
 552                         }
 553
 554                         if (PageReserved(page))
 555                                 continue;
 556
 557                         page_ext = lookup_page_ext(page);
 558                         if (unlikely(!page_ext))
 559                                 continue;
 560
 561                         /* Maybe overlapping zone */
 562                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 563                                 continue;
 564
 565                         /* Found early allocated page */
 566                         __set_page_owner_handle(page_ext, early_handle, 0, 0);
 567                         count++;
 568                 }
 569                 cond_resched();
 570         }
 571
 572         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 573                 pgdat->node_id, zone->name, count);
 574 }
 575
 576 static void init_zones_in_node(pg_data_t *pgdat)
 577 {
 578         struct zone *zone;
 579         struct zone *node_zones = pgdat->node_zones;
 580
 581         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 582                 if (!populated_zone(zone))
 583                         continue;
 584
 585                 init_pages_in_zone(pgdat, zone);
 586         }
 587 }
 588
 589 static void init_early_allocated_pages(void)
 590 {
 591         pg_data_t *pgdat;
 592
 593         for_each_online_pgdat(pgdat)
 594                 init_zones_in_node(pgdat);
 595 }
 596
 597 static const struct file_operations proc_page_owner_operations = {
 598         .read           = read_page_owner,
 599 };
 600
 601 static int __init pageowner_init(void)
 602 {
 603         if (!static_branch_unlikely(&page_owner_inited)) {
 604                 pr_info("page_owner is disabled\n");
 605                 return 0;
 606         }
 607
 608         debugfs_create_file("page_owner", 0400, NULL, NULL,
 609                             &proc_page_owner_operations);
 610
 611         return 0;
 612 }
 613 late_initcall(pageowner_init)