mm/page_isolation.c

   1 /*
   2  * linux/mm/page_isolation.c
   3  */
   4
   5 #include <linux/mm.h>
   6 #include <linux/page-isolation.h>
   7 #include <linux/pageblock-flags.h>
   8 #include <linux/memory.h>
   9 #include <linux/hugetlb.h>
  10 #include "internal.h"
  11
  12 int set_migratetype_isolate(struct page *page, bool skip_hwpoisoned_pages)
  13 {
  14         struct zone *zone;
  15         unsigned long flags, pfn;
  16         struct memory_isolate_notify arg;
  17         int notifier_ret;
  18         int ret = -EBUSY;
  19
  20         zone = page_zone(page);
  21
  22         spin_lock_irqsave(&zone->lock, flags);
  23
  24         pfn = page_to_pfn(page);
  25         arg.start_pfn = pfn;
  26         arg.nr_pages = pageblock_nr_pages;
  27         arg.pages_found = 0;
  28
  29         /*
  30          * It may be possible to isolate a pageblock even if the
  31          * migratetype is not MIGRATE_MOVABLE. The memory isolation
  32          * notifier chain is used by balloon drivers to return the
  33          * number of pages in a range that are held by the balloon
  34          * driver to shrink memory. If all the pages are accounted for
  35          * by balloons, are free, or on the LRU, isolation can continue.
  36          * Later, for example, when memory hotplug notifier runs, these
  37          * pages reported as "can be isolated" should be isolated(freed)
  38          * by the balloon driver through the memory notifier chain.
  39          */
  40         notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
  41         notifier_ret = notifier_to_errno(notifier_ret);
  42         if (notifier_ret)
  43                 goto out;
  44         /*
  45          * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
  46          * We just check MOVABLE pages.
  47          */
  48         if (!has_unmovable_pages(zone, page, arg.pages_found,
  49                                  skip_hwpoisoned_pages))
  50                 ret = 0;
  51
  52         /*
  53          * immobile means "not-on-lru" paes. If immobile is larger than
  54          * removable-by-driver pages reported by notifier, we'll fail.
  55          */
  56
  57 out:
  58         if (!ret) {
  59                 unsigned long nr_pages;
  60                 int migratetype = get_pageblock_migratetype(page);
  61
  62                 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
  63                 zone->nr_isolate_pageblock++;
  64                 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE);
  65
  66                 __mod_zone_freepage_state(zone, -nr_pages, migratetype);
  67         }
  68
  69         spin_unlock_irqrestore(&zone->lock, flags);
  70         if (!ret)
  71                 drain_all_pages(zone);
  72         return ret;
  73 }
  74
  75 void unset_migratetype_isolate(struct page *page, unsigned migratetype)
  76 {
  77         struct zone *zone;
  78         unsigned long flags, nr_pages;
  79         struct page *isolated_page = NULL;
  80         unsigned int order;
  81         unsigned long page_idx, buddy_idx;
  82         struct page *buddy;
  83
  84         zone = page_zone(page);
  85         spin_lock_irqsave(&zone->lock, flags);
  86         if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
  87                 goto out;
  88
  89         /*
  90          * Because freepage with more than pageblock_order on isolated
  91          * pageblock is restricted to merge due to freepage counting problem,
  92          * it is possible that there is free buddy page.
  93          * move_freepages_block() doesn't care of merge so we need other
  94          * approach in order to merge them. Isolation and free will make
  95          * these pages to be merged.
  96          */
  97         if (PageBuddy(page)) {
  98                 order = page_order(page);
  99                 if (order >= pageblock_order) {
 100                         page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 101                         buddy_idx = __find_buddy_index(page_idx, order);
 102                         buddy = page + (buddy_idx - page_idx);
 103
 104                         if (!is_migrate_isolate_page(buddy)) {
 105                                 __isolate_free_page(page, order);
 106                                 set_page_refcounted(page);
 107                                 isolated_page = page;
 108                         }
 109                 }
 110         }
 111
 112         /*
 113          * If we isolate freepage with more than pageblock_order, there
 114          * should be no freepage in the range, so we could avoid costly
 115          * pageblock scanning for freepage moving.
 116          */
 117         if (!isolated_page) {
 118                 nr_pages = move_freepages_block(zone, page, migratetype);
 119                 __mod_zone_freepage_state(zone, nr_pages, migratetype);
 120         }
 121         set_pageblock_migratetype(page, migratetype);
 122         zone->nr_isolate_pageblock--;
 123 out:
 124         spin_unlock_irqrestore(&zone->lock, flags);
 125         if (isolated_page)
 126                 __free_pages(isolated_page, order);
 127 }
 128
 129 static inline struct page *
 130 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
 131 {
 132         int i;
 133         for (i = 0; i < nr_pages; i++)
 134                 if (pfn_valid_within(pfn + i))
 135                         break;
 136         if (unlikely(i == nr_pages))
 137                 return NULL;
 138         return pfn_to_page(pfn + i);
 139 }
 140
 141 /*
 142  * start_isolate_page_range() -- make page-allocation-type of range of pages
 143  * to be MIGRATE_ISOLATE.
 144  * @start_pfn: The lower PFN of the range to be isolated.
 145  * @end_pfn: The upper PFN of the range to be isolated.
 146  * @migratetype: migrate type to set in error recovery.
 147  *
 148  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
 149  * the range will never be allocated. Any free pages and pages freed in the
 150  * future will not be allocated again.
 151  *
 152  * start_pfn/end_pfn must be aligned to pageblock_order.
 153  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
 154  */
 155 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 156                              unsigned migratetype, bool skip_hwpoisoned_pages)
 157 {
 158         unsigned long pfn;
 159         unsigned long undo_pfn;
 160         struct page *page;
 161
 162         BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
 163         BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
 164
 165         for (pfn = start_pfn;
 166              pfn < end_pfn;
 167              pfn += pageblock_nr_pages) {
 168                 page = __first_valid_page(pfn, pageblock_nr_pages);
 169                 if (page &&
 170                     set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
 171                         undo_pfn = pfn;
 172                         goto undo;
 173                 }
 174         }
 175         return 0;
 176 undo:
 177         for (pfn = start_pfn;
 178              pfn < undo_pfn;
 179              pfn += pageblock_nr_pages)
 180                 unset_migratetype_isolate(pfn_to_page(pfn), migratetype);
 181
 182         return -EBUSY;
 183 }
 184
 185 /*
 186  * Make isolated pages available again.
 187  */
 188 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 189                             unsigned migratetype)
 190 {
 191         unsigned long pfn;
 192         struct page *page;
 193         BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
 194         BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
 195         for (pfn = start_pfn;
 196              pfn < end_pfn;
 197              pfn += pageblock_nr_pages) {
 198                 page = __first_valid_page(pfn, pageblock_nr_pages);
 199                 if (!page || get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
 200                         continue;
 201                 unset_migratetype_isolate(page, migratetype);
 202         }
 203         return 0;
 204 }
 205 /*
 206  * Test all pages in the range is free(means isolated) or not.
 207  * all pages in [start_pfn...end_pfn) must be in the same zone.
 208  * zone->lock must be held before call this.
 209  *
 210  * Returns 1 if all pages in the range are isolated.
 211  */
 212 static int
 213 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
 214                                   bool skip_hwpoisoned_pages)
 215 {
 216         struct page *page;
 217
 218         while (pfn < end_pfn) {
 219                 if (!pfn_valid_within(pfn)) {
 220                         pfn++;
 221                         continue;
 222                 }
 223                 page = pfn_to_page(pfn);
 224                 if (PageBuddy(page)) {
 225                         /*
 226                          * If race between isolatation and allocation happens,
 227                          * some free pages could be in MIGRATE_MOVABLE list
 228                          * although pageblock's migratation type of the page
 229                          * is MIGRATE_ISOLATE. Catch it and move the page into
 230                          * MIGRATE_ISOLATE list.
 231                          */
 232                         if (get_freepage_migratetype(page) != MIGRATE_ISOLATE) {
 233                                 struct page *end_page;
 234
 235                                 end_page = page + (1 << page_order(page)) - 1;
 236                                 move_freepages(page_zone(page), page, end_page,
 237                                                 MIGRATE_ISOLATE);
 238                         }
 239                         pfn += 1 << page_order(page);
 240                 }
 241                 else if (page_count(page) == 0 &&
 242                         get_freepage_migratetype(page) == MIGRATE_ISOLATE)
 243                         pfn += 1;
 244                 else if (skip_hwpoisoned_pages && PageHWPoison(page)) {
 245                         /*
 246                          * The HWPoisoned page may be not in buddy
 247                          * system, and page_count() is not 0.
 248                          */
 249                         pfn++;
 250                         continue;
 251                 }
 252                 else
 253                         break;
 254         }
 255         if (pfn < end_pfn)
 256                 return 0;
 257         return 1;
 258 }
 259
 260 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
 261                         bool skip_hwpoisoned_pages)
 262 {
 263         unsigned long pfn, flags;
 264         struct page *page;
 265         struct zone *zone;
 266         int ret;
 267
 268         /*
 269          * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
 270          * are not aligned to pageblock_nr_pages.
 271          * Then we just check migratetype first.
 272          */
 273         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 274                 page = __first_valid_page(pfn, pageblock_nr_pages);
 275                 if (page && get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
 276                         break;
 277         }
 278         page = __first_valid_page(start_pfn, end_pfn - start_pfn);
 279         if ((pfn < end_pfn) || !page)
 280                 return -EBUSY;
 281         /* Check all pages are free or marked as ISOLATED */
 282         zone = page_zone(page);
 283         spin_lock_irqsave(&zone->lock, flags);
 284         ret = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
 285                                                 skip_hwpoisoned_pages);
 286         spin_unlock_irqrestore(&zone->lock, flags);
 287         return ret ? 0 : -EBUSY;
 288 }
 289
 290 struct page *alloc_migrate_target(struct page *page, unsigned long private,
 291                                   int **resultp)
 292 {
 293         gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
 294
 295         /*
 296          * TODO: allocate a destination hugepage from a nearest neighbor node,
 297          * accordance with memory policy of the user process if possible. For
 298          * now as a simple work-around, we use the next node for destination.
 299          */
 300         if (PageHuge(page)) {
 301                 nodemask_t src = nodemask_of_node(page_to_nid(page));
 302                 nodemask_t dst;
 303                 nodes_complement(dst, src);
 304                 return alloc_huge_page_node(page_hstate(compound_head(page)),
 305                                             next_node(page_to_nid(page), dst));
 306         }
 307
 308         if (PageHighMem(page))
 309                 gfp_mask |= __GFP_HIGHMEM;
 310
 311         return alloc_page(gfp_mask);
 312 }