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                                                                         NULL);
  71
  72                 __mod_zone_freepage_state(zone, -nr_pages, migratetype);
  73         }
  74
  75         spin_unlock_irqrestore(&zone->lock, flags);
  76         if (!ret)
  77                 drain_all_pages(zone);
  78         return ret;
  79 }
  80
  81 static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
  82 {
  83         struct zone *zone;
  84         unsigned long flags, nr_pages;
  85         bool isolated_page = false;
  86         unsigned int order;
  87         unsigned long pfn, buddy_pfn;
  88         struct page *buddy;
  89
  90         zone = page_zone(page);
  91         spin_lock_irqsave(&zone->lock, flags);
  92         if (!is_migrate_isolate_page(page))
  93                 goto out;
  94
  95         /*
  96          * Because freepage with more than pageblock_order on isolated
  97          * pageblock is restricted to merge due to freepage counting problem,
  98          * it is possible that there is free buddy page.
  99          * move_freepages_block() doesn't care of merge so we need other
 100          * approach in order to merge them. Isolation and free will make
 101          * these pages to be merged.
 102          */
 103         if (PageBuddy(page)) {
 104                 order = page_order(page);
 105                 if (order >= pageblock_order) {
 106                         pfn = page_to_pfn(page);
 107                         buddy_pfn = __find_buddy_pfn(pfn, order);
 108                         buddy = page + (buddy_pfn - pfn);
 109
 110                         if (pfn_valid_within(buddy_pfn) &&
 111                             !is_migrate_isolate_page(buddy)) {
 112                                 __isolate_free_page(page, order);
 113                                 isolated_page = true;
 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, NULL);
 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                 post_alloc_hook(page, order, __GFP_MOVABLE);
 133                 __free_pages(page, order);
 134         }
 135 }
 136
 137 static inline struct page *
 138 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
 139 {
 140         int i;
 141
 142         for (i = 0; i < nr_pages; i++) {
 143                 struct page *page;
 144
 145                 if (!pfn_valid_within(pfn + i))
 146                         continue;
 147                 page = pfn_to_online_page(pfn + i);
 148                 if (!page)
 149                         continue;
 150                 return page;
 151         }
 152         return NULL;
 153 }
 154
 155 /*
 156  * start_isolate_page_range() -- make page-allocation-type of range of pages
 157  * to be MIGRATE_ISOLATE.
 158  * @start_pfn: The lower PFN of the range to be isolated.
 159  * @end_pfn: The upper PFN of the range to be isolated.
 160  * @migratetype: migrate type to set in error recovery.
 161  *
 162  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
 163  * the range will never be allocated. Any free pages and pages freed in the
 164  * future will not be allocated again.
 165  *
 166  * start_pfn/end_pfn must be aligned to pageblock_order.
 167  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
 168  */
 169 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 170                              unsigned migratetype, bool skip_hwpoisoned_pages)
 171 {
 172         unsigned long pfn;
 173         unsigned long undo_pfn;
 174         struct page *page;
 175
 176         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 177         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 178
 179         for (pfn = start_pfn;
 180              pfn < end_pfn;
 181              pfn += pageblock_nr_pages) {
 182                 page = __first_valid_page(pfn, pageblock_nr_pages);
 183                 if (page &&
 184                     set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
 185                         undo_pfn = pfn;
 186                         goto undo;
 187                 }
 188         }
 189         return 0;
 190 undo:
 191         for (pfn = start_pfn;
 192              pfn < undo_pfn;
 193              pfn += pageblock_nr_pages) {
 194                 struct page *page = pfn_to_online_page(pfn);
 195                 if (!page)
 196                         continue;
 197                 unset_migratetype_isolate(page, migratetype);
 198         }
 199
 200         return -EBUSY;
 201 }
 202
 203 /*
 204  * Make isolated pages available again.
 205  */
 206 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 207                             unsigned migratetype)
 208 {
 209         unsigned long pfn;
 210         struct page *page;
 211
 212         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 213         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 214
 215         for (pfn = start_pfn;
 216              pfn < end_pfn;
 217              pfn += pageblock_nr_pages) {
 218                 page = __first_valid_page(pfn, pageblock_nr_pages);
 219                 if (!page || !is_migrate_isolate_page(page))
 220                         continue;
 221                 unset_migratetype_isolate(page, migratetype);
 222         }
 223         return 0;
 224 }
 225 /*
 226  * Test all pages in the range is free(means isolated) or not.
 227  * all pages in [start_pfn...end_pfn) must be in the same zone.
 228  * zone->lock must be held before call this.
 229  *
 230  * Returns the last tested pfn.
 231  */
 232 static unsigned long
 233 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
 234                                   bool skip_hwpoisoned_pages)
 235 {
 236         struct page *page;
 237
 238         while (pfn < end_pfn) {
 239                 if (!pfn_valid_within(pfn)) {
 240                         pfn++;
 241                         continue;
 242                 }
 243                 page = pfn_to_page(pfn);
 244                 if (PageBuddy(page))
 245                         /*
 246                          * If the page is on a free list, it has to be on
 247                          * the correct MIGRATE_ISOLATE freelist. There is no
 248                          * simple way to verify that as VM_BUG_ON(), though.
 249                          */
 250                         pfn += 1 << page_order(page);
 251                 else if (skip_hwpoisoned_pages && PageHWPoison(page))
 252                         /* A HWPoisoned page cannot be also PageBuddy */
 253                         pfn++;
 254                 else
 255                         break;
 256         }
 257
 258         return pfn;
 259 }
 260
 261 /* Caller should ensure that requested range is in a single zone */
 262 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
 263                         bool skip_hwpoisoned_pages)
 264 {
 265         unsigned long pfn, flags;
 266         struct page *page;
 267         struct zone *zone;
 268
 269         /*
 270          * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
 271          * are not aligned to pageblock_nr_pages.
 272          * Then we just check migratetype first.
 273          */
 274         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 275                 page = __first_valid_page(pfn, pageblock_nr_pages);
 276                 if (page && !is_migrate_isolate_page(page))
 277                         break;
 278         }
 279         page = __first_valid_page(start_pfn, end_pfn - start_pfn);
 280         if ((pfn < end_pfn) || !page)
 281                 return -EBUSY;
 282         /* Check all pages are free or marked as ISOLATED */
 283         zone = page_zone(page);
 284         spin_lock_irqsave(&zone->lock, flags);
 285         pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
 286                                                 skip_hwpoisoned_pages);
 287         spin_unlock_irqrestore(&zone->lock, flags);
 288
 289         trace_test_pages_isolated(start_pfn, end_pfn, pfn);
 290
 291         return pfn < end_pfn ? -EBUSY : 0;
 292 }
 293
 294 struct page *alloc_migrate_target(struct page *page, unsigned long private,
 295                                   int **resultp)
 296 {
 297         gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
 298
 299         /*
 300          * TODO: allocate a destination hugepage from a nearest neighbor node,
 301          * accordance with memory policy of the user process if possible. For
 302          * now as a simple work-around, we use the next node for destination.
 303          */
 304         if (PageHuge(page))
 305                 return alloc_huge_page_node(page_hstate(compound_head(page)),
 306                                             next_node_in(page_to_nid(page),
 307                                                          node_online_map));
 308
 309         if (PageHighMem(page))
 310                 gfp_mask |= __GFP_HIGHMEM;
 311
 312         return alloc_page(gfp_mask);
 313 }