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 "internal.h"
   9
  10 static bool page_owner_disabled = true;
  11 bool page_owner_inited __read_mostly;
  12
  13 static void init_early_allocated_pages(void);
  14
  15 static int early_page_owner_param(char *buf)
  16 {
  17         if (!buf)
  18                 return -EINVAL;
  19
  20         if (strcmp(buf, "on") == 0)
  21                 page_owner_disabled = false;
  22
  23         return 0;
  24 }
  25 early_param("page_owner", early_page_owner_param);
  26
  27 static bool need_page_owner(void)
  28 {
  29         if (page_owner_disabled)
  30                 return false;
  31
  32         return true;
  33 }
  34
  35 static void init_page_owner(void)
  36 {
  37         if (page_owner_disabled)
  38                 return;
  39
  40         page_owner_inited = true;
  41         init_early_allocated_pages();
  42 }
  43
  44 struct page_ext_operations page_owner_ops = {
  45         .need = need_page_owner,
  46         .init = init_page_owner,
  47 };
  48
  49 void __reset_page_owner(struct page *page, unsigned int order)
  50 {
  51         int i;
  52         struct page_ext *page_ext;
  53
  54         for (i = 0; i < (1 << order); i++) {
  55                 page_ext = lookup_page_ext(page + i);
  56                 if (unlikely(!page_ext))
  57                         continue;
  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
  66         struct stack_trace trace = {
  67                 .nr_entries = 0,
  68                 .max_entries = ARRAY_SIZE(page_ext->trace_entries),
  69                 .entries = &page_ext->trace_entries[0],
  70                 .skip = 3,
  71         };
  72
  73         if (unlikely(!page_ext))
  74                 return;
  75
  76         save_stack_trace(&trace);
  77
  78         page_ext->order = order;
  79         page_ext->gfp_mask = gfp_mask;
  80         page_ext->nr_entries = trace.nr_entries;
  81
  82         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
  83 }
  84
  85 gfp_t __get_page_owner_gfp(struct page *page)
  86 {
  87         struct page_ext *page_ext = lookup_page_ext(page);
  88         if (unlikely(!page_ext))
  89                 /*
  90                  * The caller just returns 0 if no valid gfp
  91                  * So return 0 here too.
  92                  */
  93                 return 0;
  94
  95         return page_ext->gfp_mask;
  96 }
  97
  98 static ssize_t
  99 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
 100                 struct page *page, struct page_ext *page_ext)
 101 {
 102         int ret;
 103         int pageblock_mt, page_mt;
 104         char *kbuf;
 105         struct stack_trace trace = {
 106                 .nr_entries = page_ext->nr_entries,
 107                 .entries = &page_ext->trace_entries[0],
 108         };
 109
 110         kbuf = kmalloc(count, GFP_KERNEL);
 111         if (!kbuf)
 112                 return -ENOMEM;
 113
 114         ret = snprintf(kbuf, count,
 115                         "Page allocated via order %u, mask 0x%x\n",
 116                         page_ext->order, page_ext->gfp_mask);
 117
 118         if (ret >= count)
 119                 goto err;
 120
 121         /* Print information relevant to grouping pages by mobility */
 122         pageblock_mt = get_pfnblock_migratetype(page, pfn);
 123         page_mt  = gfpflags_to_migratetype(page_ext->gfp_mask);
 124         ret += snprintf(kbuf + ret, count - ret,
 125                         "PFN %lu Block %lu type %d %s Flags %s%s%s%s%s%s%s%s%s%s%s%s\n",
 126                         pfn,
 127                         pfn >> pageblock_order,
 128                         pageblock_mt,
 129                         pageblock_mt != page_mt ? "Fallback" : "        ",
 130                         PageLocked(page)        ? "K" : " ",
 131                         PageError(page)         ? "E" : " ",
 132                         PageReferenced(page)    ? "R" : " ",
 133                         PageUptodate(page)      ? "U" : " ",
 134                         PageDirty(page)         ? "D" : " ",
 135                         PageLRU(page)           ? "L" : " ",
 136                         PageActive(page)        ? "A" : " ",
 137                         PageSlab(page)          ? "S" : " ",
 138                         PageWriteback(page)     ? "W" : " ",
 139                         PageCompound(page)      ? "C" : " ",
 140                         PageSwapCache(page)     ? "B" : " ",
 141                         PageMappedToDisk(page)  ? "M" : " ");
 142
 143         if (ret >= count)
 144                 goto err;
 145
 146         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 147         if (ret >= count)
 148                 goto err;
 149
 150         ret += snprintf(kbuf + ret, count - ret, "\n");
 151         if (ret >= count)
 152                 goto err;
 153
 154         if (copy_to_user(buf, kbuf, ret))
 155                 ret = -EFAULT;
 156
 157         kfree(kbuf);
 158         return ret;
 159
 160 err:
 161         kfree(kbuf);
 162         return -ENOMEM;
 163 }
 164
 165 static ssize_t
 166 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 167 {
 168         unsigned long pfn;
 169         struct page *page;
 170         struct page_ext *page_ext;
 171
 172         if (!page_owner_inited)
 173                 return -EINVAL;
 174
 175         page = NULL;
 176         pfn = min_low_pfn + *ppos;
 177
 178         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 179         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 180                 pfn++;
 181
 182         drain_all_pages(NULL);
 183
 184         /* Find an allocated page */
 185         for (; pfn < max_pfn; pfn++) {
 186                 /*
 187                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 188                  * validate the area as existing, skip it if not
 189                  */
 190                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 191                         pfn += MAX_ORDER_NR_PAGES - 1;
 192                         continue;
 193                 }
 194
 195                 /* Check for holes within a MAX_ORDER area */
 196                 if (!pfn_valid_within(pfn))
 197                         continue;
 198
 199                 page = pfn_to_page(pfn);
 200                 if (PageBuddy(page)) {
 201                         unsigned long freepage_order = page_order_unsafe(page);
 202
 203                         if (freepage_order < MAX_ORDER)
 204                                 pfn += (1UL << freepage_order) - 1;
 205                         continue;
 206                 }
 207
 208                 page_ext = lookup_page_ext(page);
 209                 if (unlikely(!page_ext))
 210                         continue;
 211
 212                 /*
 213                  * Some pages could be missed by concurrent allocation or free,
 214                  * because we don't hold the zone lock.
 215                  */
 216                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 217                         continue;
 218
 219                 /* Record the next PFN to read in the file offset */
 220                 *ppos = (pfn - min_low_pfn) + 1;
 221
 222                 return print_page_owner(buf, count, pfn, page, page_ext);
 223         }
 224
 225         return 0;
 226 }
 227
 228 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 229 {
 230         struct page *page;
 231         struct page_ext *page_ext;
 232         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 233         unsigned long end_pfn = pfn + zone->spanned_pages;
 234         unsigned long count = 0;
 235
 236         /* Scan block by block. First and last block may be incomplete */
 237         pfn = zone->zone_start_pfn;
 238
 239         /*
 240          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 241          * a zone boundary, it will be double counted between zones. This does
 242          * not matter as the mixed block count will still be correct
 243          */
 244         for (; pfn < end_pfn; ) {
 245                 if (!pfn_valid(pfn)) {
 246                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 247                         continue;
 248                 }
 249
 250                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 251                 block_end_pfn = min(block_end_pfn, end_pfn);
 252
 253                 page = pfn_to_page(pfn);
 254
 255                 for (; pfn < block_end_pfn; pfn++) {
 256                         if (!pfn_valid_within(pfn))
 257                                 continue;
 258
 259                         page = pfn_to_page(pfn);
 260
 261                         /*
 262                          * We are safe to check buddy flag and order, because
 263                          * this is init stage and only single thread runs.
 264                          */
 265                         if (PageBuddy(page)) {
 266                                 pfn += (1UL << page_order(page)) - 1;
 267                                 continue;
 268                         }
 269
 270                         if (PageReserved(page))
 271                                 continue;
 272
 273                         page_ext = lookup_page_ext(page);
 274                         if (unlikely(!page_ext))
 275                                 continue;
 276
 277                         /* Maybe overraping zone */
 278                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 279                                 continue;
 280
 281                         /* Found early allocated page */
 282                         set_page_owner(page, 0, 0);
 283                         count++;
 284                 }
 285         }
 286
 287         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 288                 pgdat->node_id, zone->name, count);
 289 }
 290
 291 static void init_zones_in_node(pg_data_t *pgdat)
 292 {
 293         struct zone *zone;
 294         struct zone *node_zones = pgdat->node_zones;
 295         unsigned long flags;
 296
 297         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 298                 if (!populated_zone(zone))
 299                         continue;
 300
 301                 spin_lock_irqsave(&zone->lock, flags);
 302                 init_pages_in_zone(pgdat, zone);
 303                 spin_unlock_irqrestore(&zone->lock, flags);
 304         }
 305 }
 306
 307 static void init_early_allocated_pages(void)
 308 {
 309         pg_data_t *pgdat;
 310
 311         drain_all_pages(NULL);
 312         for_each_online_pgdat(pgdat)
 313                 init_zones_in_node(pgdat);
 314 }
 315
 316 static const struct file_operations proc_page_owner_operations = {
 317         .read           = read_page_owner,
 318 };
 319
 320 static int __init pageowner_init(void)
 321 {
 322         struct dentry *dentry;
 323
 324         if (!page_owner_inited) {
 325                 pr_info("page_owner is disabled\n");
 326                 return 0;
 327         }
 328
 329         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 330                         NULL, &proc_page_owner_operations);
 331         if (IS_ERR(dentry))
 332                 return PTR_ERR(dentry);
 333
 334         return 0;
 335 }
 336 late_initcall(pageowner_init)