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