mm/page_isolation.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/mm/page_isolation.c
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/page-isolation.h>
   8 #include <linux/pageblock-flags.h>
   9 #include <linux/memory.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/page_owner.h>
  12 #include <linux/migrate.h>
  13 #include "internal.h"
  14
  15 #define CREATE_TRACE_POINTS
  16 #include <trace/events/page_isolation.h>
  17
  18 static int set_migratetype_isolate(struct page *page,
  19                                 bool skip_hwpoisoned_pages)
  20 {
  21         struct zone *zone;
  22         unsigned long flags, pfn;
  23         struct memory_isolate_notify arg;
  24         int notifier_ret;
  25         int ret = -EBUSY;
  26
  27         zone = page_zone(page);
  28
  29         spin_lock_irqsave(&zone->lock, flags);
  30
  31         pfn = page_to_pfn(page);
  32         arg.start_pfn = pfn;
  33         arg.nr_pages = pageblock_nr_pages;
  34         arg.pages_found = 0;
  35
  36         /*
  37          * It may be possible to isolate a pageblock even if the
  38          * migratetype is not MIGRATE_MOVABLE. The memory isolation
  39          * notifier chain is used by balloon drivers to return the
  40          * number of pages in a range that are held by the balloon
  41          * driver to shrink memory. If all the pages are accounted for
  42          * by balloons, are free, or on the LRU, isolation can continue.
  43          * Later, for example, when memory hotplug notifier runs, these
  44          * pages reported as "can be isolated" should be isolated(freed)
  45          * by the balloon driver through the memory notifier chain.
  46          */
  47         notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
  48         notifier_ret = notifier_to_errno(notifier_ret);
  49         if (notifier_ret)
  50                 goto out;
  51         /*
  52          * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
  53          * We just check MOVABLE pages.
  54          */
  55         if (!has_unmovable_pages(zone, page, arg.pages_found,
  56                                  skip_hwpoisoned_pages))
  57                 ret = 0;
  58
  59         /*
  60          * immobile means "not-on-lru" pages. If immobile is larger than
  61          * removable-by-driver pages reported by notifier, we'll fail.
  62          */
  63
  64 out:
  65         if (!ret) {
  66                 unsigned long nr_pages;
  67                 int migratetype = get_pageblock_migratetype(page);
  68
  69                 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
  70                 zone->nr_isolate_pageblock++;
  71                 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE,
  72                                                                         NULL);
  73
  74                 __mod_zone_freepage_state(zone, -nr_pages, migratetype);
  75         }
  76
  77         spin_unlock_irqrestore(&zone->lock, flags);
  78         if (!ret)
  79                 drain_all_pages(zone);
  80         return ret;
  81 }
  82
  83 static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
  84 {
  85         struct zone *zone;
  86         unsigned long flags, nr_pages;
  87         bool isolated_page = false;
  88         unsigned int order;
  89         unsigned long pfn, buddy_pfn;
  90         struct page *buddy;
  91
  92         zone = page_zone(page);
  93         spin_lock_irqsave(&zone->lock, flags);
  94         if (!is_migrate_isolate_page(page))
  95                 goto out;
  96
  97         /*
  98          * Because freepage with more than pageblock_order on isolated
  99          * pageblock is restricted to merge due to freepage counting problem,
 100          * it is possible that there is free buddy page.
 101          * move_freepages_block() doesn't care of merge so we need other
 102          * approach in order to merge them. Isolation and free will make
 103          * these pages to be merged.
 104          */
 105         if (PageBuddy(page)) {
 106                 order = page_order(page);
 107                 if (order >= pageblock_order) {
 108                         pfn = page_to_pfn(page);
 109                         buddy_pfn = __find_buddy_pfn(pfn, order);
 110                         buddy = page + (buddy_pfn - pfn);
 111
 112                         if (pfn_valid_within(buddy_pfn) &&
 113                             !is_migrate_isolate_page(buddy)) {
 114                                 __isolate_free_page(page, order);
 115                                 isolated_page = true;
 116                         }
 117                 }
 118         }
 119
 120         /*
 121          * If we isolate freepage with more than pageblock_order, there
 122          * should be no freepage in the range, so we could avoid costly
 123          * pageblock scanning for freepage moving.
 124          */
 125         if (!isolated_page) {
 126                 nr_pages = move_freepages_block(zone, page, migratetype, NULL);
 127                 __mod_zone_freepage_state(zone, nr_pages, migratetype);
 128         }
 129         set_pageblock_migratetype(page, migratetype);
 130         zone->nr_isolate_pageblock--;
 131 out:
 132         spin_unlock_irqrestore(&zone->lock, flags);
 133         if (isolated_page) {
 134                 post_alloc_hook(page, order, __GFP_MOVABLE);
 135                 __free_pages(page, order);
 136         }
 137 }
 138
 139 static inline struct page *
 140 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
 141 {
 142         int i;
 143
 144         for (i = 0; i < nr_pages; i++) {
 145                 struct page *page;
 146
 147                 if (!pfn_valid_within(pfn + i))
 148                         continue;
 149                 page = pfn_to_online_page(pfn + i);
 150                 if (!page)
 151                         continue;
 152                 return page;
 153         }
 154         return NULL;
 155 }
 156
 157 /*
 158  * start_isolate_page_range() -- make page-allocation-type of range of pages
 159  * to be MIGRATE_ISOLATE.
 160  * @start_pfn: The lower PFN of the range to be isolated.
 161  * @end_pfn: The upper PFN of the range to be isolated.
 162  * @migratetype: migrate type to set in error recovery.
 163  *
 164  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
 165  * the range will never be allocated. Any free pages and pages freed in the
 166  * future will not be allocated again.
 167  *
 168  * start_pfn/end_pfn must be aligned to pageblock_order.
 169  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
 170  */
 171 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 172                              unsigned migratetype, bool skip_hwpoisoned_pages)
 173 {
 174         unsigned long pfn;
 175         unsigned long undo_pfn;
 176         struct page *page;
 177
 178         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 179         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 180
 181         for (pfn = start_pfn;
 182              pfn < end_pfn;
 183              pfn += pageblock_nr_pages) {
 184                 page = __first_valid_page(pfn, pageblock_nr_pages);
 185                 if (page &&
 186                     set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
 187                         undo_pfn = pfn;
 188                         goto undo;
 189                 }
 190         }
 191         return 0;
 192 undo:
 193         for (pfn = start_pfn;
 194              pfn < undo_pfn;
 195              pfn += pageblock_nr_pages) {
 196                 struct page *page = pfn_to_online_page(pfn);
 197                 if (!page)
 198                         continue;
 199                 unset_migratetype_isolate(page, migratetype);
 200         }
 201
 202         return -EBUSY;
 203 }
 204
 205 /*
 206  * Make isolated pages available again.
 207  */
 208 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 209                             unsigned migratetype)
 210 {
 211         unsigned long pfn;
 212         struct page *page;
 213
 214         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 215         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 216
 217         for (pfn = start_pfn;
 218              pfn < end_pfn;
 219              pfn += pageblock_nr_pages) {
 220                 page = __first_valid_page(pfn, pageblock_nr_pages);
 221                 if (!page || !is_migrate_isolate_page(page))
 222                         continue;
 223                 unset_migratetype_isolate(page, migratetype);
 224         }
 225         return 0;
 226 }
 227 /*
 228  * Test all pages in the range is free(means isolated) or not.
 229  * all pages in [start_pfn...end_pfn) must be in the same zone.
 230  * zone->lock must be held before call this.
 231  *
 232  * Returns the last tested pfn.
 233  */
 234 static unsigned long
 235 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
 236                                   bool skip_hwpoisoned_pages)
 237 {
 238         struct page *page;
 239
 240         while (pfn < end_pfn) {
 241                 if (!pfn_valid_within(pfn)) {
 242                         pfn++;
 243                         continue;
 244                 }
 245                 page = pfn_to_page(pfn);
 246                 if (PageBuddy(page))
 247                         /*
 248                          * If the page is on a free list, it has to be on
 249                          * the correct MIGRATE_ISOLATE freelist. There is no
 250                          * simple way to verify that as VM_BUG_ON(), though.
 251                          */
 252                         pfn += 1 << page_order(page);
 253                 else if (skip_hwpoisoned_pages && PageHWPoison(page))
 254                         /* A HWPoisoned page cannot be also PageBuddy */
 255                         pfn++;
 256                 else
 257                         break;
 258         }
 259
 260         return pfn;
 261 }
 262
 263 /* Caller should ensure that requested range is in a single zone */
 264 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
 265                         bool skip_hwpoisoned_pages)
 266 {
 267         unsigned long pfn, flags;
 268         struct page *page;
 269         struct zone *zone;
 270
 271         /*
 272          * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
 273          * are not aligned to pageblock_nr_pages.
 274          * Then we just check migratetype first.
 275          */
 276         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 277                 page = __first_valid_page(pfn, pageblock_nr_pages);
 278                 if (page && !is_migrate_isolate_page(page))
 279                         break;
 280         }
 281         page = __first_valid_page(start_pfn, end_pfn - start_pfn);
 282         if ((pfn < end_pfn) || !page)
 283                 return -EBUSY;
 284         /* Check all pages are free or marked as ISOLATED */
 285         zone = page_zone(page);
 286         spin_lock_irqsave(&zone->lock, flags);
 287         pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
 288                                                 skip_hwpoisoned_pages);
 289
 290         if (pfn < end_pfn) {
 291                 pr_info_ratelimited("%s: page of pfn %lu is not isolated\n", __func__, pfn);
 292                 /* __dump_page(pfn_to_page(pfn), "isolation failure"); */
 293         }
 294
 295         spin_unlock_irqrestore(&zone->lock, flags);
 296
 297         trace_test_pages_isolated(start_pfn, end_pfn, pfn);
 298
 299         return pfn < end_pfn ? -EBUSY : 0;
 300 }
 301
 302 struct page *alloc_migrate_target(struct page *page, unsigned long private,
 303                                   int **resultp)
 304 {
 305         return new_page_nodemask(page, numa_node_id(), &node_states[N_MEMORY]);
 306 }