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