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 "internal.h"
  11
  12 static bool page_owner_disabled = true;
  13 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  14
  15 static void init_early_allocated_pages(void);
  16
  17 static int early_page_owner_param(char *buf)
  18 {
  19         if (!buf)
  20                 return -EINVAL;
  21
  22         if (strcmp(buf, "on") == 0)
  23                 page_owner_disabled = false;
  24
  25         return 0;
  26 }
  27 early_param("page_owner", early_page_owner_param);
  28
  29 static bool need_page_owner(void)
  30 {
  31         if (page_owner_disabled)
  32                 return false;
  33
  34         return true;
  35 }
  36
  37 static void init_page_owner(void)
  38 {
  39         if (page_owner_disabled)
  40                 return;
  41
  42         static_branch_enable(&page_owner_inited);
  43         init_early_allocated_pages();
  44 }
  45
  46 struct page_ext_operations page_owner_ops = {
  47         .need = need_page_owner,
  48         .init = init_page_owner,
  49 };
  50
  51 void __reset_page_owner(struct page *page, unsigned int order)
  52 {
  53         int i;
  54         struct page_ext *page_ext;
  55
  56         for (i = 0; i < (1 << order); i++) {
  57                 page_ext = lookup_page_ext(page + i);
  58                 if (unlikely(!page_ext))
  59                         continue;
  60                 __clear_bit(PAGE_EXT_OWNER, &page_ext->flags);
  61         }
  62 }
  63
  64 void __set_page_owner(struct page *page, unsigned int order, gfp_t gfp_mask)
  65 {
  66         struct page_ext *page_ext = lookup_page_ext(page);
  67
  68         struct stack_trace trace = {
  69                 .nr_entries = 0,
  70                 .max_entries = ARRAY_SIZE(page_ext->trace_entries),
  71                 .entries = &page_ext->trace_entries[0],
  72                 .skip = 3,
  73         };
  74
  75         if (unlikely(!page_ext))
  76                 return;
  77
  78         save_stack_trace(&trace);
  79
  80         page_ext->order = order;
  81         page_ext->gfp_mask = gfp_mask;
  82         page_ext->nr_entries = trace.nr_entries;
  83         page_ext->last_migrate_reason = -1;
  84
  85         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
  86 }
  87
  88 void __set_page_owner_migrate_reason(struct page *page, int reason)
  89 {
  90         struct page_ext *page_ext = lookup_page_ext(page);
  91         if (unlikely(!page_ext))
  92                 return;
  93
  94         page_ext->last_migrate_reason = reason;
  95 }
  96
  97 gfp_t __get_page_owner_gfp(struct page *page)
  98 {
  99         struct page_ext *page_ext = lookup_page_ext(page);
 100         if (unlikely(!page_ext))
 101                 /*
 102                  * The caller just returns 0 if no valid gfp
 103                  * So return 0 here too.
 104                  */
 105                 return 0;
 106
 107         return page_ext->gfp_mask;
 108 }
 109
 110 void __copy_page_owner(struct page *oldpage, struct page *newpage)
 111 {
 112         struct page_ext *old_ext = lookup_page_ext(oldpage);
 113         struct page_ext *new_ext = lookup_page_ext(newpage);
 114         int i;
 115
 116         if (unlikely(!old_ext || !new_ext))
 117                 return;
 118
 119         new_ext->order = old_ext->order;
 120         new_ext->gfp_mask = old_ext->gfp_mask;
 121         new_ext->nr_entries = old_ext->nr_entries;
 122
 123         for (i = 0; i < ARRAY_SIZE(new_ext->trace_entries); i++)
 124                 new_ext->trace_entries[i] = old_ext->trace_entries[i];
 125
 126         /*
 127          * We don't clear the bit on the oldpage as it's going to be freed
 128          * after migration. Until then, the info can be useful in case of
 129          * a bug, and the overal stats will be off a bit only temporarily.
 130          * Also, migrate_misplaced_transhuge_page() can still fail the
 131          * migration and then we want the oldpage to retain the info. But
 132          * in that case we also don't need to explicitly clear the info from
 133          * the new page, which will be freed.
 134          */
 135         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
 136 }
 137
 138 static ssize_t
 139 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 140                 struct page *page, struct page_ext *page_ext)
 141 {
 142         int ret;
 143         int pageblock_mt, page_mt;
 144         char *kbuf;
 145         struct stack_trace trace = {
 146                 .nr_entries = page_ext->nr_entries,
 147                 .entries = &page_ext->trace_entries[0],
 148         };
 149
 150         kbuf = kmalloc(count, GFP_KERNEL);
 151         if (!kbuf)
 152                 return -ENOMEM;
 153
 154         ret = snprintf(kbuf, count,
 155                         "Page allocated via order %u, mask %#x(%pGg)\n",
 156                         page_ext->order, page_ext->gfp_mask,
 157                         &page_ext->gfp_mask);
 158
 159         if (ret >= count)
 160                 goto err;
 161
 162         /* Print information relevant to grouping pages by mobility */
 163         pageblock_mt = get_pageblock_migratetype(page);
 164         page_mt  = gfpflags_to_migratetype(page_ext->gfp_mask);
 165         ret += snprintf(kbuf + ret, count - ret,
 166                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 167                         pfn,
 168                         migratetype_names[page_mt],
 169                         pfn >> pageblock_order,
 170                         migratetype_names[pageblock_mt],
 171                         page->flags, &page->flags);
 172
 173         if (ret >= count)
 174                 goto err;
 175
 176         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 177         if (ret >= count)
 178                 goto err;
 179
 180         if (page_ext->last_migrate_reason != -1) {
 181                 ret += snprintf(kbuf + ret, count - ret,
 182                         "Page has been migrated, last migrate reason: %s\n",
 183                         migrate_reason_names[page_ext->last_migrate_reason]);
 184                 if (ret >= count)
 185                         goto err;
 186         }
 187
 188         ret += snprintf(kbuf + ret, count - ret, "\n");
 189         if (ret >= count)
 190                 goto err;
 191
 192         if (copy_to_user(buf, kbuf, ret))
 193                 ret = -EFAULT;
 194
 195         kfree(kbuf);
 196         return ret;
 197
 198 err:
 199         kfree(kbuf);
 200         return -ENOMEM;
 201 }
 202
 203 void __dump_page_owner(struct page *page)
 204 {
 205         struct page_ext *page_ext = lookup_page_ext(page);
 206         struct stack_trace trace = {
 207                 .nr_entries = page_ext->nr_entries,
 208                 .entries = &page_ext->trace_entries[0],
 209         };
 210         gfp_t gfp_mask;
 211         int mt;
 212
 213         if (unlikely(!page_ext)) {
 214                 pr_alert("There is not page extension available.\n");
 215                 return;
 216         }
 217         gfp_mask = page_ext->gfp_mask;
 218         mt = gfpflags_to_migratetype(gfp_mask);
 219
 220         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
 221                 pr_alert("page_owner info is not active (free page?)\n");
 222                 return;
 223         }
 224
 225         pr_alert("page allocated via order %u, migratetype %s, gfp_mask %#x(%pGg)\n",
 226                  page_ext->order, migratetype_names[mt], gfp_mask, &gfp_mask);
 227         print_stack_trace(&trace, 0);
 228
 229         if (page_ext->last_migrate_reason != -1)
 230                 pr_alert("page has been migrated, last migrate reason: %s\n",
 231                         migrate_reason_names[page_ext->last_migrate_reason]);
 232 }
 233
 234 static ssize_t
 235 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 236 {
 237         unsigned long pfn;
 238         struct page *page;
 239         struct page_ext *page_ext;
 240
 241         if (!static_branch_unlikely(&page_owner_inited))
 242                 return -EINVAL;
 243
 244         page = NULL;
 245         pfn = min_low_pfn + *ppos;
 246
 247         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 248         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 249                 pfn++;
 250
 251         drain_all_pages(NULL);
 252
 253         /* Find an allocated page */
 254         for (; pfn < max_pfn; pfn++) {
 255                 /*
 256                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 257                  * validate the area as existing, skip it if not
 258                  */
 259                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 260                         pfn += MAX_ORDER_NR_PAGES - 1;
 261                         continue;
 262                 }
 263
 264                 /* Check for holes within a MAX_ORDER area */
 265                 if (!pfn_valid_within(pfn))
 266                         continue;
 267
 268                 page = pfn_to_page(pfn);
 269                 if (PageBuddy(page)) {
 270                         unsigned long freepage_order = page_order_unsafe(page);
 271
 272                         if (freepage_order < MAX_ORDER)
 273                                 pfn += (1UL << freepage_order) - 1;
 274                         continue;
 275                 }
 276
 277                 page_ext = lookup_page_ext(page);
 278                 if (unlikely(!page_ext))
 279                         continue;
 280
 281                 /*
 282                  * Some pages could be missed by concurrent allocation or free,
 283                  * because we don't hold the zone lock.
 284                  */
 285                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 286                         continue;
 287
 288                 /* Record the next PFN to read in the file offset */
 289                 *ppos = (pfn - min_low_pfn) + 1;
 290
 291                 return print_page_owner(buf, count, pfn, page, page_ext);
 292         }
 293
 294         return 0;
 295 }
 296
 297 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 298 {
 299         struct page *page;
 300         struct page_ext *page_ext;
 301         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 302         unsigned long end_pfn = pfn + zone->spanned_pages;
 303         unsigned long count = 0;
 304
 305         /* Scan block by block. First and last block may be incomplete */
 306         pfn = zone->zone_start_pfn;
 307
 308         /*
 309          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 310          * a zone boundary, it will be double counted between zones. This does
 311          * not matter as the mixed block count will still be correct
 312          */
 313         for (; pfn < end_pfn; ) {
 314                 if (!pfn_valid(pfn)) {
 315                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 316                         continue;
 317                 }
 318
 319                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 320                 block_end_pfn = min(block_end_pfn, end_pfn);
 321
 322                 page = pfn_to_page(pfn);
 323
 324                 for (; pfn < block_end_pfn; pfn++) {
 325                         if (!pfn_valid_within(pfn))
 326                                 continue;
 327
 328                         page = pfn_to_page(pfn);
 329
 330                         if (page_zone(page) != zone)
 331                                 continue;
 332
 333                         /*
 334                          * We are safe to check buddy flag and order, because
 335                          * this is init stage and only single thread runs.
 336                          */
 337                         if (PageBuddy(page)) {
 338                                 pfn += (1UL << page_order(page)) - 1;
 339                                 continue;
 340                         }
 341
 342                         if (PageReserved(page))
 343                                 continue;
 344
 345                         page_ext = lookup_page_ext(page);
 346                         if (unlikely(!page_ext))
 347                                 continue;
 348
 349                         /* Maybe overraping zone */
 350                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 351                                 continue;
 352
 353                         /* Found early allocated page */
 354                         set_page_owner(page, 0, 0);
 355                         count++;
 356                 }
 357         }
 358
 359         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 360                 pgdat->node_id, zone->name, count);
 361 }
 362
 363 static void init_zones_in_node(pg_data_t *pgdat)
 364 {
 365         struct zone *zone;
 366         struct zone *node_zones = pgdat->node_zones;
 367         unsigned long flags;
 368
 369         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 370                 if (!populated_zone(zone))
 371                         continue;
 372
 373                 spin_lock_irqsave(&zone->lock, flags);
 374                 init_pages_in_zone(pgdat, zone);
 375                 spin_unlock_irqrestore(&zone->lock, flags);
 376         }
 377 }
 378
 379 static void init_early_allocated_pages(void)
 380 {
 381         pg_data_t *pgdat;
 382
 383         drain_all_pages(NULL);
 384         for_each_online_pgdat(pgdat)
 385                 init_zones_in_node(pgdat);
 386 }
 387
 388 static const struct file_operations proc_page_owner_operations = {
 389         .read           = read_page_owner,
 390 };
 391
 392 static int __init pageowner_init(void)
 393 {
 394         struct dentry *dentry;
 395
 396         if (!static_branch_unlikely(&page_owner_inited)) {
 397                 pr_info("page_owner is disabled\n");
 398                 return 0;
 399         }
 400
 401         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 402                         NULL, &proc_page_owner_operations);
 403         if (IS_ERR(dentry))
 404                 return PTR_ERR(dentry);
 405
 406         return 0;
 407 }
 408 late_initcall(pageowner_init)