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                 __clear_bit(PAGE_EXT_OWNER, &page_ext->flags);
  57         }
  58 }
  59
  60 void __set_page_owner(struct page *page, unsigned int order, gfp_t gfp_mask)
  61 {
  62         struct page_ext *page_ext = lookup_page_ext(page);
  63         struct stack_trace trace = {
  64                 .nr_entries = 0,
  65                 .max_entries = ARRAY_SIZE(page_ext->trace_entries),
  66                 .entries = &page_ext->trace_entries[0],
  67                 .skip = 3,
  68         };
  69
  70         save_stack_trace(&trace);
  71
  72         page_ext->order = order;
  73         page_ext->gfp_mask = gfp_mask;
  74         page_ext->nr_entries = trace.nr_entries;
  75
  76         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
  77 }
  78
  79 gfp_t __get_page_owner_gfp(struct page *page)
  80 {
  81         struct page_ext *page_ext = lookup_page_ext(page);
  82
  83         return page_ext->gfp_mask;
  84 }
  85
  86 static ssize_t
  87 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
  88                 struct page *page, struct page_ext *page_ext)
  89 {
  90         int ret;
  91         int pageblock_mt, page_mt;
  92         char *kbuf;
  93         struct stack_trace trace = {
  94                 .nr_entries = page_ext->nr_entries,
  95                 .entries = &page_ext->trace_entries[0],
  96         };
  97
  98         kbuf = kmalloc(count, GFP_KERNEL);
  99         if (!kbuf)
 100                 return -ENOMEM;
 101
 102         ret = snprintf(kbuf, count,
 103                         "Page allocated via order %u, mask %#x(%pGg)\n",
 104                         page_ext->order, page_ext->gfp_mask,
 105                         &page_ext->gfp_mask);
 106
 107         if (ret >= count)
 108                 goto err;
 109
 110         /* Print information relevant to grouping pages by mobility */
 111         pageblock_mt = get_pfnblock_migratetype(page, pfn);
 112         page_mt  = gfpflags_to_migratetype(page_ext->gfp_mask);
 113         ret += snprintf(kbuf + ret, count - ret,
 114                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 115                         pfn,
 116                         migratetype_names[page_mt],
 117                         pfn >> pageblock_order,
 118                         migratetype_names[pageblock_mt],
 119                         page->flags, &page->flags);
 120
 121         if (ret >= count)
 122                 goto err;
 123
 124         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 125         if (ret >= count)
 126                 goto err;
 127
 128         ret += snprintf(kbuf + ret, count - ret, "\n");
 129         if (ret >= count)
 130                 goto err;
 131
 132         if (copy_to_user(buf, kbuf, ret))
 133                 ret = -EFAULT;
 134
 135         kfree(kbuf);
 136         return ret;
 137
 138 err:
 139         kfree(kbuf);
 140         return -ENOMEM;
 141 }
 142
 143 static ssize_t
 144 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 145 {
 146         unsigned long pfn;
 147         struct page *page;
 148         struct page_ext *page_ext;
 149
 150         if (!page_owner_inited)
 151                 return -EINVAL;
 152
 153         page = NULL;
 154         pfn = min_low_pfn + *ppos;
 155
 156         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 157         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 158                 pfn++;
 159
 160         drain_all_pages(NULL);
 161
 162         /* Find an allocated page */
 163         for (; pfn < max_pfn; pfn++) {
 164                 /*
 165                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 166                  * validate the area as existing, skip it if not
 167                  */
 168                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 169                         pfn += MAX_ORDER_NR_PAGES - 1;
 170                         continue;
 171                 }
 172
 173                 /* Check for holes within a MAX_ORDER area */
 174                 if (!pfn_valid_within(pfn))
 175                         continue;
 176
 177                 page = pfn_to_page(pfn);
 178                 if (PageBuddy(page)) {
 179                         unsigned long freepage_order = page_order_unsafe(page);
 180
 181                         if (freepage_order < MAX_ORDER)
 182                                 pfn += (1UL << freepage_order) - 1;
 183                         continue;
 184                 }
 185
 186                 page_ext = lookup_page_ext(page);
 187
 188                 /*
 189                  * Some pages could be missed by concurrent allocation or free,
 190                  * because we don't hold the zone lock.
 191                  */
 192                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 193                         continue;
 194
 195                 /* Record the next PFN to read in the file offset */
 196                 *ppos = (pfn - min_low_pfn) + 1;
 197
 198                 return print_page_owner(buf, count, pfn, page, page_ext);
 199         }
 200
 201         return 0;
 202 }
 203
 204 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 205 {
 206         struct page *page;
 207         struct page_ext *page_ext;
 208         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 209         unsigned long end_pfn = pfn + zone->spanned_pages;
 210         unsigned long count = 0;
 211
 212         /* Scan block by block. First and last block may be incomplete */
 213         pfn = zone->zone_start_pfn;
 214
 215         /*
 216          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 217          * a zone boundary, it will be double counted between zones. This does
 218          * not matter as the mixed block count will still be correct
 219          */
 220         for (; pfn < end_pfn; ) {
 221                 if (!pfn_valid(pfn)) {
 222                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 223                         continue;
 224                 }
 225
 226                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 227                 block_end_pfn = min(block_end_pfn, end_pfn);
 228
 229                 page = pfn_to_page(pfn);
 230
 231                 for (; pfn < block_end_pfn; pfn++) {
 232                         if (!pfn_valid_within(pfn))
 233                                 continue;
 234
 235                         page = pfn_to_page(pfn);
 236
 237                         /*
 238                          * We are safe to check buddy flag and order, because
 239                          * this is init stage and only single thread runs.
 240                          */
 241                         if (PageBuddy(page)) {
 242                                 pfn += (1UL << page_order(page)) - 1;
 243                                 continue;
 244                         }
 245
 246                         if (PageReserved(page))
 247                                 continue;
 248
 249                         page_ext = lookup_page_ext(page);
 250
 251                         /* Maybe overraping zone */
 252                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 253                                 continue;
 254
 255                         /* Found early allocated page */
 256                         set_page_owner(page, 0, 0);
 257                         count++;
 258                 }
 259         }
 260
 261         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 262                 pgdat->node_id, zone->name, count);
 263 }
 264
 265 static void init_zones_in_node(pg_data_t *pgdat)
 266 {
 267         struct zone *zone;
 268         struct zone *node_zones = pgdat->node_zones;
 269         unsigned long flags;
 270
 271         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 272                 if (!populated_zone(zone))
 273                         continue;
 274
 275                 spin_lock_irqsave(&zone->lock, flags);
 276                 init_pages_in_zone(pgdat, zone);
 277                 spin_unlock_irqrestore(&zone->lock, flags);
 278         }
 279 }
 280
 281 static void init_early_allocated_pages(void)
 282 {
 283         pg_data_t *pgdat;
 284
 285         drain_all_pages(NULL);
 286         for_each_online_pgdat(pgdat)
 287                 init_zones_in_node(pgdat);
 288 }
 289
 290 static const struct file_operations proc_page_owner_operations = {
 291         .read           = read_page_owner,
 292 };
 293
 294 static int __init pageowner_init(void)
 295 {
 296         struct dentry *dentry;
 297
 298         if (!page_owner_inited) {
 299                 pr_info("page_owner is disabled\n");
 300                 return 0;
 301         }
 302
 303         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 304                         NULL, &proc_page_owner_operations);
 305         if (IS_ERR(dentry))
 306                 return PTR_ERR(dentry);
 307
 308         return 0;
 309 }
 310 late_initcall(pageowner_init)