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 "internal.h"
  10
  11 static bool page_owner_disabled = true;
  12 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
  13
  14 static void init_early_allocated_pages(void);
  15
  16 static int early_page_owner_param(char *buf)
  17 {
  18         if (!buf)
  19                 return -EINVAL;
  20
  21         if (strcmp(buf, "on") == 0)
  22                 page_owner_disabled = false;
  23
  24         return 0;
  25 }
  26 early_param("page_owner", early_page_owner_param);
  27
  28 static bool need_page_owner(void)
  29 {
  30         if (page_owner_disabled)
  31                 return false;
  32
  33         return true;
  34 }
  35
  36 static void init_page_owner(void)
  37 {
  38         if (page_owner_disabled)
  39                 return;
  40
  41         static_branch_enable(&page_owner_inited);
  42         init_early_allocated_pages();
  43 }
  44
  45 struct page_ext_operations page_owner_ops = {
  46         .need = need_page_owner,
  47         .init = init_page_owner,
  48 };
  49
  50 void __reset_page_owner(struct page *page, unsigned int order)
  51 {
  52         int i;
  53         struct page_ext *page_ext;
  54
  55         for (i = 0; i < (1 << order); i++) {
  56                 page_ext = lookup_page_ext(page + i);
  57                 __clear_bit(PAGE_EXT_OWNER, &page_ext->flags);
  58         }
  59 }
  60
  61 void __set_page_owner(struct page *page, unsigned int order, gfp_t gfp_mask)
  62 {
  63         struct page_ext *page_ext = lookup_page_ext(page);
  64         struct stack_trace trace = {
  65                 .nr_entries = 0,
  66                 .max_entries = ARRAY_SIZE(page_ext->trace_entries),
  67                 .entries = &page_ext->trace_entries[0],
  68                 .skip = 3,
  69         };
  70
  71         save_stack_trace(&trace);
  72
  73         page_ext->order = order;
  74         page_ext->gfp_mask = gfp_mask;
  75         page_ext->nr_entries = trace.nr_entries;
  76
  77         __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
  78 }
  79
  80 gfp_t __get_page_owner_gfp(struct page *page)
  81 {
  82         struct page_ext *page_ext = lookup_page_ext(page);
  83
  84         return page_ext->gfp_mask;
  85 }
  86
  87 static ssize_t
  88 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
  89                 struct page *page, struct page_ext *page_ext)
  90 {
  91         int ret;
  92         int pageblock_mt, page_mt;
  93         char *kbuf;
  94         struct stack_trace trace = {
  95                 .nr_entries = page_ext->nr_entries,
  96                 .entries = &page_ext->trace_entries[0],
  97         };
  98
  99         kbuf = kmalloc(count, GFP_KERNEL);
 100         if (!kbuf)
 101                 return -ENOMEM;
 102
 103         ret = snprintf(kbuf, count,
 104                         "Page allocated via order %u, mask %#x(%pGg)\n",
 105                         page_ext->order, page_ext->gfp_mask,
 106                         &page_ext->gfp_mask);
 107
 108         if (ret >= count)
 109                 goto err;
 110
 111         /* Print information relevant to grouping pages by mobility */
 112         pageblock_mt = get_pfnblock_migratetype(page, pfn);
 113         page_mt  = gfpflags_to_migratetype(page_ext->gfp_mask);
 114         ret += snprintf(kbuf + ret, count - ret,
 115                         "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
 116                         pfn,
 117                         migratetype_names[page_mt],
 118                         pfn >> pageblock_order,
 119                         migratetype_names[pageblock_mt],
 120                         page->flags, &page->flags);
 121
 122         if (ret >= count)
 123                 goto err;
 124
 125         ret += snprint_stack_trace(kbuf + ret, count - ret, &trace, 0);
 126         if (ret >= count)
 127                 goto err;
 128
 129         ret += snprintf(kbuf + ret, count - ret, "\n");
 130         if (ret >= count)
 131                 goto err;
 132
 133         if (copy_to_user(buf, kbuf, ret))
 134                 ret = -EFAULT;
 135
 136         kfree(kbuf);
 137         return ret;
 138
 139 err:
 140         kfree(kbuf);
 141         return -ENOMEM;
 142 }
 143
 144 static ssize_t
 145 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 146 {
 147         unsigned long pfn;
 148         struct page *page;
 149         struct page_ext *page_ext;
 150
 151         if (!static_branch_unlikely(&page_owner_inited))
 152                 return -EINVAL;
 153
 154         page = NULL;
 155         pfn = min_low_pfn + *ppos;
 156
 157         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 158         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
 159                 pfn++;
 160
 161         drain_all_pages(NULL);
 162
 163         /* Find an allocated page */
 164         for (; pfn < max_pfn; pfn++) {
 165                 /*
 166                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
 167                  * validate the area as existing, skip it if not
 168                  */
 169                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
 170                         pfn += MAX_ORDER_NR_PAGES - 1;
 171                         continue;
 172                 }
 173
 174                 /* Check for holes within a MAX_ORDER area */
 175                 if (!pfn_valid_within(pfn))
 176                         continue;
 177
 178                 page = pfn_to_page(pfn);
 179                 if (PageBuddy(page)) {
 180                         unsigned long freepage_order = page_order_unsafe(page);
 181
 182                         if (freepage_order < MAX_ORDER)
 183                                 pfn += (1UL << freepage_order) - 1;
 184                         continue;
 185                 }
 186
 187                 page_ext = lookup_page_ext(page);
 188
 189                 /*
 190                  * Some pages could be missed by concurrent allocation or free,
 191                  * because we don't hold the zone lock.
 192                  */
 193                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 194                         continue;
 195
 196                 /* Record the next PFN to read in the file offset */
 197                 *ppos = (pfn - min_low_pfn) + 1;
 198
 199                 return print_page_owner(buf, count, pfn, page, page_ext);
 200         }
 201
 202         return 0;
 203 }
 204
 205 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
 206 {
 207         struct page *page;
 208         struct page_ext *page_ext;
 209         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
 210         unsigned long end_pfn = pfn + zone->spanned_pages;
 211         unsigned long count = 0;
 212
 213         /* Scan block by block. First and last block may be incomplete */
 214         pfn = zone->zone_start_pfn;
 215
 216         /*
 217          * Walk the zone in pageblock_nr_pages steps. If a page block spans
 218          * a zone boundary, it will be double counted between zones. This does
 219          * not matter as the mixed block count will still be correct
 220          */
 221         for (; pfn < end_pfn; ) {
 222                 if (!pfn_valid(pfn)) {
 223                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
 224                         continue;
 225                 }
 226
 227                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
 228                 block_end_pfn = min(block_end_pfn, end_pfn);
 229
 230                 page = pfn_to_page(pfn);
 231
 232                 for (; pfn < block_end_pfn; pfn++) {
 233                         if (!pfn_valid_within(pfn))
 234                                 continue;
 235
 236                         page = pfn_to_page(pfn);
 237
 238                         /*
 239                          * We are safe to check buddy flag and order, because
 240                          * this is init stage and only single thread runs.
 241                          */
 242                         if (PageBuddy(page)) {
 243                                 pfn += (1UL << page_order(page)) - 1;
 244                                 continue;
 245                         }
 246
 247                         if (PageReserved(page))
 248                                 continue;
 249
 250                         page_ext = lookup_page_ext(page);
 251
 252                         /* Maybe overraping zone */
 253                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
 254                                 continue;
 255
 256                         /* Found early allocated page */
 257                         set_page_owner(page, 0, 0);
 258                         count++;
 259                 }
 260         }
 261
 262         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
 263                 pgdat->node_id, zone->name, count);
 264 }
 265
 266 static void init_zones_in_node(pg_data_t *pgdat)
 267 {
 268         struct zone *zone;
 269         struct zone *node_zones = pgdat->node_zones;
 270         unsigned long flags;
 271
 272         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 273                 if (!populated_zone(zone))
 274                         continue;
 275
 276                 spin_lock_irqsave(&zone->lock, flags);
 277                 init_pages_in_zone(pgdat, zone);
 278                 spin_unlock_irqrestore(&zone->lock, flags);
 279         }
 280 }
 281
 282 static void init_early_allocated_pages(void)
 283 {
 284         pg_data_t *pgdat;
 285
 286         drain_all_pages(NULL);
 287         for_each_online_pgdat(pgdat)
 288                 init_zones_in_node(pgdat);
 289 }
 290
 291 static const struct file_operations proc_page_owner_operations = {
 292         .read           = read_page_owner,
 293 };
 294
 295 static int __init pageowner_init(void)
 296 {
 297         struct dentry *dentry;
 298
 299         if (!static_branch_unlikely(&page_owner_inited)) {
 300                 pr_info("page_owner is disabled\n");
 301                 return 0;
 302         }
 303
 304         dentry = debugfs_create_file("page_owner", S_IRUSR, NULL,
 305                         NULL, &proc_page_owner_operations);
 306         if (IS_ERR(dentry))
 307                 return PTR_ERR(dentry);
 308
 309         return 0;
 310 }
 311 late_initcall(pageowner_init)