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