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/bootmem.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 int order;
  24         gfp_t gfp_mask;
  25         int last_migrate_reason;
  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 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                 page = pfn_to_online_page(pfn);
 277                 if (!page) {
 278                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 279                         continue;
 280                 }
 281
 282                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 283                 block_end_pfn = min(block_end_pfn, end_pfn);
 284
 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                         /* The pageblock is online, no need to recheck. */
 292                         page = pfn_to_page(pfn);
 293
 294                         if (page_zone(page) != zone)
 295                                 continue;
 296
 297                         if (PageBuddy(page)) {
 298                                 unsigned long freepage_order;
 299
 300                                 freepage_order = page_order_unsafe(page);
 301                                 if (freepage_order < MAX_ORDER)
 302                                         pfn += (1UL << freepage_order) - 1;
 303                                 continue;
 304                         }
 305
 306                         if (PageReserved(page))
 307                                 continue;
 308
 309                         page_ext = lookup_page_ext(page);
 310                         if (unlikely(!page_ext))
 311                                 continue;
 312
 313                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 314                                 continue;
 315
 316                         page_owner = get_page_owner(page_ext);
 317                         page_mt = gfpflags_to_migratetype(
 318                                         page_owner->gfp_mask);
 319                         if (pageblock_mt != page_mt) {
 320                                 if (is_migrate_cma(pageblock_mt))
 321                                         count[MIGRATE_MOVABLE]++;
 322                                 else
 323                                         count[pageblock_mt]++;
 324
 325                                 pfn = block_end_pfn;
 326                                 break;
 327                         }
 328                         pfn += (1UL << page_owner->order) - 1;
 329                 }
 330         }
 331
 332         /* Print counts */
 333         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 334         for (i = 0; i < MIGRATE_TYPES; i++)
 335                 seq_printf(m, "%12lu ", count[i]);
 336         seq_putc(m, '\n');
 337 }
 338
 339 static ssize_t
 340 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 341                 struct page *page, struct page_owner *page_owner,
 342                 depot_stack_handle_t handle)
 343 {
 344         int ret;
 345         int pageblock_mt, page_mt;
 346         char *kbuf;
 347         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 348         struct stack_trace trace = {
 349                 .nr_entries = 0,
 350                 .entries = entries,
 351                 .max_entries = PAGE_OWNER_STACK_DEPTH,
 352                 .skip = 0
 353         };
 354
 355         kbuf = kmalloc(count, GFP_KERNEL);
 356         if (!kbuf)
 357                 return -ENOMEM;
 358
 359         ret = snprintf(kbuf, count,
 360                         "Page allocated via order %u, mask %#x(%pGg)\n",
 361                         page_owner->order, page_owner->gfp_mask,
 362                         &page_owner->gfp_mask);
 363
 364         if (ret >= count)
 365                 goto err;
 366
 367         /* Print information relevant to grouping pages by mobility */
 368         pageblock_mt = get_pageblock_migratetype(page);
 369         page_mt  = gfpflags_to_migratetype(page_owner->gfp_mask);
 370         ret += snprintf(kbuf + ret, count - ret,
 371                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 372                         pfn,
 373                         migratetype_names[page_mt],
 374                         pfn >> pageblock_order,
 375                         migratetype_names[pageblock_mt],
 376                         page->flags, &page->flags);
 377
 378         if (ret >= count)
 379                 goto err;
 380
 381         depot_fetch_stack(handle, &trace);
 382         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 383         if (ret >= count)
 384                 goto err;
 385
 386         if (page_owner->last_migrate_reason != -1) {
 387                 ret += snprintf(kbuf + ret, count - ret,
 388                         "Page has been migrated, last migrate reason: %s\n",
 389                         migrate_reason_names[page_owner->last_migrate_reason]);
 390                 if (ret >= count)
 391                         goto err;
 392         }
 393
 394         ret += snprintf(kbuf + ret, count - ret, "\n");
 395         if (ret >= count)
 396                 goto err;
 397
 398         if (copy_to_user(buf, kbuf, ret))
 399                 ret = -EFAULT;
 400
 401         kfree(kbuf);
 402         return ret;
 403
 404 err:
 405         kfree(kbuf);
 406         return -ENOMEM;
 407 }
 408
 409 void __dump_page_owner(struct page *page)
 410 {
 411         struct page_ext *page_ext = lookup_page_ext(page);
 412         struct page_owner *page_owner;
 413         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 414         struct stack_trace trace = {
 415                 .nr_entries = 0,
 416                 .entries = entries,
 417                 .max_entries = PAGE_OWNER_STACK_DEPTH,
 418                 .skip = 0
 419         };
 420         depot_stack_handle_t handle;
 421         gfp_t gfp_mask;
 422         int mt;
 423
 424         if (unlikely(!page_ext)) {
 425                 pr_alert("There is not page extension available.\n");
 426                 return;
 427         }
 428
 429         page_owner = get_page_owner(page_ext);
 430         gfp_mask = page_owner->gfp_mask;
 431         mt = gfpflags_to_migratetype(gfp_mask);
 432
 433         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 434                 pr_alert("page_owner info is not active (free page?)\n");
 435                 return;
 436         }
 437
 438         handle = READ_ONCE(page_owner->handle);
 439         if (!handle) {
 440                 pr_alert("page_owner info is not active (free page?)\n");
 441                 return;
 442         }
 443
 444         depot_fetch_stack(handle, &trace);
 445         pr_alert("page allocated via order %u, migratetype %s, gfp_mask %#x(%pGg)\n",
 446                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask);
 447         print_stack_trace(&trace, 0);
 448
 449         if (page_owner->last_migrate_reason != -1)
 450                 pr_alert("page has been migrated, last migrate reason: %s\n",
 451                         migrate_reason_names[page_owner->last_migrate_reason]);
 452 }
 453
 454 static ssize_t
 455 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 456 {
 457         unsigned long pfn;
 458         struct page *page;
 459         struct page_ext *page_ext;
 460         struct page_owner *page_owner;
 461         depot_stack_handle_t handle;
 462
 463         if (!static_branch_unlikely(&page_owner_inited))
 464                 return -EINVAL;
 465
 466         page = NULL;
 467         pfn = min_low_pfn + *ppos;
 468
 469         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 470         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 471                 pfn++;
 472
 473         drain_all_pages(NULL);
 474
 475         /* Find an allocated page */
 476         for (; pfn < max_pfn; pfn++) {
 477                 /*
 478                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 479                  * validate the area as existing, skip it if not
 480                  */
 481                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 482                         pfn += MAX_ORDER_NR_PAGES - 1;
 483                         continue;
 484                 }
 485
 486                 /* Check for holes within a MAX_ORDER area */
 487                 if (!pfn_valid_within(pfn))
 488                         continue;
 489
 490                 page = pfn_to_page(pfn);
 491                 if (PageBuddy(page)) {
 492                         unsigned long freepage_order = page_order_unsafe(page);
 493
 494                         if (freepage_order < MAX_ORDER)
 495                                 pfn += (1UL << freepage_order) - 1;
 496                         continue;
 497                 }
 498
 499                 page_ext = lookup_page_ext(page);
 500                 if (unlikely(!page_ext))
 501                         continue;
 502
 503                 /*
 504                  * Some pages could be missed by concurrent allocation or free,
 505                  * because we don't hold the zone lock.
 506                  */
 507                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 508                         continue;
 509
 510                 page_owner = get_page_owner(page_ext);
 511
 512                 /*
 513                  * Access to page_ext->handle isn't synchronous so we should
 514                  * be careful to access it.
 515                  */
 516                 handle = READ_ONCE(page_owner->handle);
 517                 if (!handle)
 518                         continue;
 519
 520                 /* Record the next PFN to read in the file offset */
 521                 *ppos = (pfn - min_low_pfn) + 1;
 522
 523                 return print_page_owner(buf, count, pfn, page,
 524                                 page_owner, handle);
 525         }
 526
 527         return 0;
 528 }
 529
 530 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 531 {
 532         struct page *page;
 533         struct page_ext *page_ext;
 534         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 535         unsigned long end_pfn = pfn + zone->spanned_pages;
 536         unsigned long count = 0;
 537
 538         /* Scan block by block. First and last block may be incomplete */
 539         pfn = zone->zone_start_pfn;
 540
 541         /*
 542          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 543          * a zone boundary, it will be double counted between zones. This does
 544          * not matter as the mixed block count will still be correct
 545          */
 546         for (; pfn < end_pfn; ) {
 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                 page = pfn_to_page(pfn);
 556
 557                 for (; pfn < block_end_pfn; pfn++) {
 558                         if (!pfn_valid_within(pfn))
 559                                 continue;
 560
 561                         page = pfn_to_page(pfn);
 562
 563                         if (page_zone(page) != zone)
 564                                 continue;
 565
 566                         /*
 567                          * To avoid having to grab zone->lock, be a little
 568                          * careful when reading buddy page order. The only
 569                          * danger is that we skip too much and potentially miss
 570                          * some early allocated pages, which is better than
 571                          * heavy lock contention.
 572                          */
 573                         if (PageBuddy(page)) {
 574                                 unsigned long order = page_order_unsafe(page);
 575
 576                                 if (order > 0 && order < MAX_ORDER)
 577                                         pfn += (1UL << order) - 1;
 578                                 continue;
 579                         }
 580
 581                         if (PageReserved(page))
 582                                 continue;
 583
 584                         page_ext = lookup_page_ext(page);
 585                         if (unlikely(!page_ext))
 586                                 continue;
 587
 588                         /* Maybe overlapping zone */
 589                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 590                                 continue;
 591
 592                         /* Found early allocated page */
 593                         __set_page_owner_handle(page_ext, early_handle, 0, 0);
 594                         count++;
 595                 }
 596                 cond_resched();
 597         }
 598
 599         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 600                 pgdat->node_id, zone->name, count);
 601 }
 602
 603 static void init_zones_in_node(pg_data_t *pgdat)
 604 {
 605         struct zone *zone;
 606         struct zone *node_zones = pgdat->node_zones;
 607
 608         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 609                 if (!populated_zone(zone))
 610                         continue;
 611
 612                 init_pages_in_zone(pgdat, zone);
 613         }
 614 }
 615
 616 static void init_early_allocated_pages(void)
 617 {
 618         pg_data_t *pgdat;
 619
 620         for_each_online_pgdat(pgdat)
 621                 init_zones_in_node(pgdat);
 622 }
 623
 624 static const struct file_operations proc_page_owner_operations = {
 625         .read           = read_page_owner,
 626 };
 627
 628 static int __init pageowner_init(void)
 629 {
 630         struct dentry *dentry;
 631
 632         if (!static_branch_unlikely(&page_owner_inited)) {
 633                 pr_info("page_owner is disabled\n");
 634                 return 0;
 635         }
 636
 637         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 638                         NULL, &proc_page_owner_operations);
 639         if (IS_ERR(dentry))
 640                 return PTR_ERR(dentry);
 641
 642         return 0;
 643 }
 644 late_initcall(pageowner_init)