mm/page_owner.c

   1 #include <linux/debugfs.h>
   2 #include <linux/mm.h>
   3 #include <linux/slab.h>
   4 #include <linux/uaccess.h>
   5 #include <linux/bootmem.h>
   6 #include <linux/stacktrace.h>
   7 #include <linux/page_owner.h>
   8 #include <linux/jump_label.h>
   9 #include <linux/migrate.h>
  10 #include <linux/stackdepot.h>
  11
  12 #include "internal.h"
  13
  14 /*
  15  * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
  16  * to use off stack temporal storage
  17  */
  18 #define PAGE_OWNER_STACK_DEPTH (16)
  19
  20 static bool page_owner_disabled = true;
  21 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  22
  23 static depot_stack_handle_t dummy_handle;
  24 static depot_stack_handle_t failure_handle;
  25
  26 static void init_early_allocated_pages(void);
  27
  28 static int early_page_owner_param(char *buf)
  29 {
  30         if (!buf)
  31                 return -EINVAL;
  32
  33         if (strcmp(buf, "on") == 0)
  34                 page_owner_disabled = false;
  35
  36         return 0;
  37 }
  38 early_param("page_owner", early_page_owner_param);
  39
  40 static bool need_page_owner(void)
  41 {
  42         if (page_owner_disabled)
  43                 return false;
  44
  45         return true;
  46 }
  47
  48 static noinline void register_dummy_stack(void)
  49 {
  50         unsigned long entries[4];
  51         struct stack_trace dummy;
  52
  53         dummy.nr_entries = 0;
  54         dummy.max_entries = ARRAY_SIZE(entries);
  55         dummy.entries = &entries[0];
  56         dummy.skip = 0;
  57
  58         save_stack_trace(&dummy);
  59         dummy_handle = depot_save_stack(&dummy, GFP_KERNEL);
  60 }
  61
  62 static noinline void register_failure_stack(void)
  63 {
  64         unsigned long entries[4];
  65         struct stack_trace failure;
  66
  67         failure.nr_entries = 0;
  68         failure.max_entries = ARRAY_SIZE(entries);
  69         failure.entries = &entries[0];
  70         failure.skip = 0;
  71
  72         save_stack_trace(&failure);
  73         failure_handle = depot_save_stack(&failure, GFP_KERNEL);
  74 }
  75
  76 static void init_page_owner(void)
  77 {
  78         if (page_owner_disabled)
  79                 return;
  80
  81         register_dummy_stack();
  82         register_failure_stack();
  83         static_branch_enable(&page_owner_inited);
  84         init_early_allocated_pages();
  85 }
  86
  87 struct page_ext_operations page_owner_ops = {
  88         .need = need_page_owner,
  89         .init = init_page_owner,
  90 };
  91
  92 void __reset_page_owner(struct page *page, unsigned int order)
  93 {
  94         int i;
  95         struct page_ext *page_ext;
  96
  97         for (i = 0; i < (1 << order); i++) {
  98                 page_ext = lookup_page_ext(page + i);
  99                 if (unlikely(!page_ext))
 100                         continue;
 101                 __clear_bit(PAGE_EXT_OWNER, &page_ext->flags);
 102         }
 103 }
 104
 105 static inline bool check_recursive_alloc(struct stack_trace *trace,
 106                                         unsigned long ip)
 107 {
 108         int i, count;
 109
 110         if (!trace->nr_entries)
 111                 return false;
 112
 113         for (i = 0, count = 0; i < trace->nr_entries; i++) {
 114                 if (trace->entries[i] == ip && ++count == 2)
 115                         return true;
 116         }
 117
 118         return false;
 119 }
 120
 121 static noinline depot_stack_handle_t save_stack(gfp_t flags)
 122 {
 123         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 124         struct stack_trace trace = {
 125                 .nr_entries = 0,
 126                 .entries = entries,
 127                 .max_entries = PAGE_OWNER_STACK_DEPTH,
 128                 .skip = 0
 129         };
 130         depot_stack_handle_t handle;
 131
 132         save_stack_trace(&trace);
 133         if (trace.nr_entries != 0 &&
 134             trace.entries[trace.nr_entries-1] == ULONG_MAX)
 135                 trace.nr_entries--;
 136
 137         /*
 138          * We need to check recursion here because our request to stackdepot
 139          * could trigger memory allocation to save new entry. New memory
 140          * allocation would reach here and call depot_save_stack() again
 141          * if we don't catch it. There is still not enough memory in stackdepot
 142          * so it would try to allocate memory again and loop forever.
 143          */
 144         if (check_recursive_alloc(&trace, _RET_IP_))
 145                 return dummy_handle;
 146
 147         handle = depot_save_stack(&trace, flags);
 148         if (!handle)
 149                 handle = failure_handle;
 150
 151         return handle;
 152 }
 153
 154 noinline void __set_page_owner(struct page *page, unsigned int order,
 155                                         gfp_t gfp_mask)
 156 {
 157         struct page_ext *page_ext = lookup_page_ext(page);
 158
 159         if (unlikely(!page_ext))
 160                 return;
 161
 162         page_ext->handle = save_stack(gfp_mask);
 163         page_ext->order = order;
 164         page_ext->gfp_mask = gfp_mask;
 165         page_ext->last_migrate_reason = -1;
 166
 167         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
 168 }
 169
 170 void __set_page_owner_migrate_reason(struct page *page, int reason)
 171 {
 172         struct page_ext *page_ext = lookup_page_ext(page);
 173         if (unlikely(!page_ext))
 174                 return;
 175
 176         page_ext->last_migrate_reason = reason;
 177 }
 178
 179 void __split_page_owner(struct page *page, unsigned int order)
 180 {
 181         int i;
 182         struct page_ext *page_ext = lookup_page_ext(page);
 183
 184         if (unlikely(!page_ext))
 185                 return;
 186
 187         page_ext->order = 0;
 188         for (i = 1; i < (1 << order); i++)
 189                 __copy_page_owner(page, page + i);
 190 }
 191
 192 void __copy_page_owner(struct page *oldpage, struct page *newpage)
 193 {
 194         struct page_ext *old_ext = lookup_page_ext(oldpage);
 195         struct page_ext *new_ext = lookup_page_ext(newpage);
 196
 197         if (unlikely(!old_ext || !new_ext))
 198                 return;
 199
 200         new_ext->order = old_ext->order;
 201         new_ext->gfp_mask = old_ext->gfp_mask;
 202         new_ext->last_migrate_reason = old_ext->last_migrate_reason;
 203         new_ext->handle = old_ext->handle;
 204
 205         /*
 206          * We don't clear the bit on the oldpage as it's going to be freed
 207          * after migration. Until then, the info can be useful in case of
 208          * a bug, and the overal stats will be off a bit only temporarily.
 209          * Also, migrate_misplaced_transhuge_page() can still fail the
 210          * migration and then we want the oldpage to retain the info. But
 211          * in that case we also don't need to explicitly clear the info from
 212          * the new page, which will be freed.
 213          */
 214         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 215 }
 216
 217 static ssize_t
 218 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 219                 struct page *page, struct page_ext *page_ext,
 220                 depot_stack_handle_t handle)
 221 {
 222         int ret;
 223         int pageblock_mt, page_mt;
 224         char *kbuf;
 225         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 226         struct stack_trace trace = {
 227                 .nr_entries = 0,
 228                 .entries = entries,
 229                 .max_entries = PAGE_OWNER_STACK_DEPTH,
 230                 .skip = 0
 231         };
 232
 233         kbuf = kmalloc(count, GFP_KERNEL);
 234         if (!kbuf)
 235                 return -ENOMEM;
 236
 237         ret = snprintf(kbuf, count,
 238                         "Page allocated via order %u, mask %#x(%pGg)\n",
 239                         page_ext->order, page_ext->gfp_mask,
 240                         &page_ext->gfp_mask);
 241
 242         if (ret >= count)
 243                 goto err;
 244
 245         /* Print information relevant to grouping pages by mobility */
 246         pageblock_mt = get_pageblock_migratetype(page);
 247         page_mt  = gfpflags_to_migratetype(page_ext->gfp_mask);
 248         ret += snprintf(kbuf + ret, count - ret,
 249                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 250                         pfn,
 251                         migratetype_names[page_mt],
 252                         pfn >> pageblock_order,
 253                         migratetype_names[pageblock_mt],
 254                         page->flags, &page->flags);
 255
 256         if (ret >= count)
 257                 goto err;
 258
 259         depot_fetch_stack(handle, &trace);
 260         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 261         if (ret >= count)
 262                 goto err;
 263
 264         if (page_ext->last_migrate_reason != -1) {
 265                 ret += snprintf(kbuf + ret, count - ret,
 266                         "Page has been migrated, last migrate reason: %s\n",
 267                         migrate_reason_names[page_ext->last_migrate_reason]);
 268                 if (ret >= count)
 269                         goto err;
 270         }
 271
 272         ret += snprintf(kbuf + ret, count - ret, "\n");
 273         if (ret >= count)
 274                 goto err;
 275
 276         if (copy_to_user(buf, kbuf, ret))
 277                 ret = -EFAULT;
 278
 279         kfree(kbuf);
 280         return ret;
 281
 282 err:
 283         kfree(kbuf);
 284         return -ENOMEM;
 285 }
 286
 287 void __dump_page_owner(struct page *page)
 288 {
 289         struct page_ext *page_ext = lookup_page_ext(page);
 290         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 291         struct stack_trace trace = {
 292                 .nr_entries = 0,
 293                 .entries = entries,
 294                 .max_entries = PAGE_OWNER_STACK_DEPTH,
 295                 .skip = 0
 296         };
 297         depot_stack_handle_t handle;
 298         gfp_t gfp_mask;
 299         int mt;
 300
 301         if (unlikely(!page_ext)) {
 302                 pr_alert("There is not page extension available.\n");
 303                 return;
 304         }
 305         gfp_mask = page_ext->gfp_mask;
 306         mt = gfpflags_to_migratetype(gfp_mask);
 307
 308         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 309                 pr_alert("page_owner info is not active (free page?)\n");
 310                 return;
 311         }
 312
 313         handle = READ_ONCE(page_ext->handle);
 314         if (!handle) {
 315                 pr_alert("page_owner info is not active (free page?)\n");
 316                 return;
 317         }
 318
 319         depot_fetch_stack(handle, &trace);
 320         pr_alert("page allocated via order %u, migratetype %s, gfp_mask %#x(%pGg)\n",
 321                  page_ext->order, migratetype_names[mt], gfp_mask, &gfp_mask);
 322         print_stack_trace(&trace, 0);
 323
 324         if (page_ext->last_migrate_reason != -1)
 325                 pr_alert("page has been migrated, last migrate reason: %s\n",
 326                         migrate_reason_names[page_ext->last_migrate_reason]);
 327 }
 328
 329 static ssize_t
 330 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 331 {
 332         unsigned long pfn;
 333         struct page *page;
 334         struct page_ext *page_ext;
 335         depot_stack_handle_t handle;
 336
 337         if (!static_branch_unlikely(&page_owner_inited))
 338                 return -EINVAL;
 339
 340         page = NULL;
 341         pfn = min_low_pfn + *ppos;
 342
 343         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 344         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 345                 pfn++;
 346
 347         drain_all_pages(NULL);
 348
 349         /* Find an allocated page */
 350         for (; pfn < max_pfn; pfn++) {
 351                 /*
 352                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 353                  * validate the area as existing, skip it if not
 354                  */
 355                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 356                         pfn += MAX_ORDER_NR_PAGES - 1;
 357                         continue;
 358                 }
 359
 360                 /* Check for holes within a MAX_ORDER area */
 361                 if (!pfn_valid_within(pfn))
 362                         continue;
 363
 364                 page = pfn_to_page(pfn);
 365                 if (PageBuddy(page)) {
 366                         unsigned long freepage_order = page_order_unsafe(page);
 367
 368                         if (freepage_order < MAX_ORDER)
 369                                 pfn += (1UL << freepage_order) - 1;
 370                         continue;
 371                 }
 372
 373                 page_ext = lookup_page_ext(page);
 374                 if (unlikely(!page_ext))
 375                         continue;
 376
 377                 /*
 378                  * Some pages could be missed by concurrent allocation or free,
 379                  * because we don't hold the zone lock.
 380                  */
 381                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 382                         continue;
 383
 384                 /*
 385                  * Access to page_ext->handle isn't synchronous so we should
 386                  * be careful to access it.
 387                  */
 388                 handle = READ_ONCE(page_ext->handle);
 389                 if (!handle)
 390                         continue;
 391
 392                 /* Record the next PFN to read in the file offset */
 393                 *ppos = (pfn - min_low_pfn) + 1;
 394
 395                 return print_page_owner(buf, count, pfn, page,
 396                                 page_ext, handle);
 397         }
 398
 399         return 0;
 400 }
 401
 402 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 403 {
 404         struct page *page;
 405         struct page_ext *page_ext;
 406         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 407         unsigned long end_pfn = pfn + zone->spanned_pages;
 408         unsigned long count = 0;
 409
 410         /* Scan block by block. First and last block may be incomplete */
 411         pfn = zone->zone_start_pfn;
 412
 413         /*
 414          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 415          * a zone boundary, it will be double counted between zones. This does
 416          * not matter as the mixed block count will still be correct
 417          */
 418         for (; pfn < end_pfn; ) {
 419                 if (!pfn_valid(pfn)) {
 420                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 421                         continue;
 422                 }
 423
 424                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 425                 block_end_pfn = min(block_end_pfn, end_pfn);
 426
 427                 page = pfn_to_page(pfn);
 428
 429                 for (; pfn < block_end_pfn; pfn++) {
 430                         if (!pfn_valid_within(pfn))
 431                                 continue;
 432
 433                         page = pfn_to_page(pfn);
 434
 435                         if (page_zone(page) != zone)
 436                                 continue;
 437
 438                         /*
 439                          * We are safe to check buddy flag and order, because
 440                          * this is init stage and only single thread runs.
 441                          */
 442                         if (PageBuddy(page)) {
 443                                 pfn += (1UL << page_order(page)) - 1;
 444                                 continue;
 445                         }
 446
 447                         if (PageReserved(page))
 448                                 continue;
 449
 450                         page_ext = lookup_page_ext(page);
 451                         if (unlikely(!page_ext))
 452                                 continue;
 453
 454                         /* Maybe overraping zone */
 455                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 456                                 continue;
 457
 458                         /* Found early allocated page */
 459                         set_page_owner(page, 0, 0);
 460                         count++;
 461                 }
 462         }
 463
 464         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 465                 pgdat->node_id, zone->name, count);
 466 }
 467
 468 static void init_zones_in_node(pg_data_t *pgdat)
 469 {
 470         struct zone *zone;
 471         struct zone *node_zones = pgdat->node_zones;
 472         unsigned long flags;
 473
 474         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 475                 if (!populated_zone(zone))
 476                         continue;
 477
 478                 spin_lock_irqsave(&zone->lock, flags);
 479                 init_pages_in_zone(pgdat, zone);
 480                 spin_unlock_irqrestore(&zone->lock, flags);
 481         }
 482 }
 483
 484 static void init_early_allocated_pages(void)
 485 {
 486         pg_data_t *pgdat;
 487
 488         drain_all_pages(NULL);
 489         for_each_online_pgdat(pgdat)
 490                 init_zones_in_node(pgdat);
 491 }
 492
 493 static const struct file_operations proc_page_owner_operations = {
 494         .read           = read_page_owner,
 495 };
 496
 497 static int __init pageowner_init(void)
 498 {
 499         struct dentry *dentry;
 500
 501         if (!static_branch_unlikely(&page_owner_inited)) {
 502                 pr_info("page_owner is disabled\n");
 503                 return 0;
 504         }
 505
 506         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 507                         NULL, &proc_page_owner_operations);
 508         if (IS_ERR(dentry))
 509                 return PTR_ERR(dentry);
 510
 511         return 0;
 512 }
 513 late_initcall(pageowner_init)