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