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