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, int migratetype, int isol_flags)
  19 {
  20         struct zone *zone = page_zone(page);
  21         struct page *unmovable;
  22         unsigned long flags;
  23
  24         spin_lock_irqsave(&zone->lock, flags);
  25
  26         /*
  27          * We assume the caller intended to SET migrate type to isolate.
  28          * If it is already set, then someone else must have raced and
  29          * set it before us.
  30          */
  31         if (is_migrate_isolate_page(page)) {
  32                 spin_unlock_irqrestore(&zone->lock, flags);
  33                 return -EBUSY;
  34         }
  35
  36         /*
  37          * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
  38          * We just check MOVABLE pages.
  39          */
  40         unmovable = has_unmovable_pages(zone, page, migratetype, isol_flags);
  41         if (!unmovable) {
  42                 unsigned long nr_pages;
  43                 int mt = get_pageblock_migratetype(page);
  44
  45                 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
  46                 zone->nr_isolate_pageblock++;
  47                 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE,
  48                                                                         NULL);
  49
  50                 __mod_zone_freepage_state(zone, -nr_pages, mt);
  51                 spin_unlock_irqrestore(&zone->lock, flags);
  52                 return 0;
  53         }
  54
  55         spin_unlock_irqrestore(&zone->lock, flags);
  56         if (isol_flags & REPORT_FAILURE) {
  57                 /*
  58                  * printk() with zone->lock held will likely trigger a
  59                  * lockdep splat, so defer it here.
  60                  */
  61                 dump_page(unmovable, "unmovable page");
  62         }
  63
  64         return -EBUSY;
  65 }
  66
  67 static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
  68 {
  69         struct zone *zone;
  70         unsigned long flags, nr_pages;
  71         bool isolated_page = false;
  72         unsigned int order;
  73         unsigned long pfn, buddy_pfn;
  74         struct page *buddy;
  75
  76         zone = page_zone(page);
  77         spin_lock_irqsave(&zone->lock, flags);
  78         if (!is_migrate_isolate_page(page))
  79                 goto out;
  80
  81         /*
  82          * Because freepage with more than pageblock_order on isolated
  83          * pageblock is restricted to merge due to freepage counting problem,
  84          * it is possible that there is free buddy page.
  85          * move_freepages_block() doesn't care of merge so we need other
  86          * approach in order to merge them. Isolation and free will make
  87          * these pages to be merged.
  88          */
  89         if (PageBuddy(page)) {
  90                 order = buddy_order(page);
  91                 if (order >= pageblock_order && order < MAX_ORDER - 1) {
  92                         pfn = page_to_pfn(page);
  93                         buddy_pfn = __find_buddy_pfn(pfn, order);
  94                         buddy = page + (buddy_pfn - pfn);
  95
  96                         if (pfn_valid_within(buddy_pfn) &&
  97                             !is_migrate_isolate_page(buddy)) {
  98                                 __isolate_free_page(page, order);
  99                                 isolated_page = true;
 100                         }
 101                 }
 102         }
 103
 104         /*
 105          * If we isolate freepage with more than pageblock_order, there
 106          * should be no freepage in the range, so we could avoid costly
 107          * pageblock scanning for freepage moving.
 108          *
 109          * We didn't actually touch any of the isolated pages, so place them
 110          * to the tail of the freelist. This is an optimization for memory
 111          * onlining - just onlined memory won't immediately be considered for
 112          * allocation.
 113          */
 114         if (!isolated_page) {
 115                 nr_pages = move_freepages_block(zone, page, migratetype, NULL);
 116                 __mod_zone_freepage_state(zone, nr_pages, migratetype);
 117         }
 118         set_pageblock_migratetype(page, migratetype);
 119         if (isolated_page)
 120                 __putback_isolated_page(page, order, migratetype);
 121         zone->nr_isolate_pageblock--;
 122 out:
 123         spin_unlock_irqrestore(&zone->lock, flags);
 124 }
 125
 126 static inline struct page *
 127 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
 128 {
 129         int i;
 130
 131         for (i = 0; i < nr_pages; i++) {
 132                 struct page *page;
 133
 134                 page = pfn_to_online_page(pfn + i);
 135                 if (!page)
 136                         continue;
 137                 return page;
 138         }
 139         return NULL;
 140 }
 141
 142 /**
 143  * start_isolate_page_range() - make page-allocation-type of range of pages to
 144  * be MIGRATE_ISOLATE.
 145  * @start_pfn:          The lower PFN of the range to be isolated.
 146  * @end_pfn:            The upper PFN of the range to be isolated.
 147  *                      start_pfn/end_pfn must be aligned to pageblock_order.
 148  * @migratetype:        Migrate type to set in error recovery.
 149  * @flags:              The following flags are allowed (they can be combined in
 150  *                      a bit mask)
 151  *                      MEMORY_OFFLINE - isolate to offline (!allocate) memory
 152  *                                       e.g., skip over PageHWPoison() pages
 153  *                                       and PageOffline() pages.
 154  *                      REPORT_FAILURE - report details about the failure to
 155  *                      isolate the range
 156  *
 157  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
 158  * the range will never be allocated. Any free pages and pages freed in the
 159  * future will not be allocated again. If specified range includes migrate types
 160  * other than MOVABLE or CMA, this will fail with -EBUSY. For isolating all
 161  * pages in the range finally, the caller have to free all pages in the range.
 162  * test_page_isolated() can be used for test it.
 163  *
 164  * There is no high level synchronization mechanism that prevents two threads
 165  * from trying to isolate overlapping ranges. If this happens, one thread
 166  * will notice pageblocks in the overlapping range already set to isolate.
 167  * This happens in set_migratetype_isolate, and set_migratetype_isolate
 168  * returns an error. We then clean up by restoring the migration type on
 169  * pageblocks we may have modified and return -EBUSY to caller. This
 170  * prevents two threads from simultaneously working on overlapping ranges.
 171  *
 172  * Please note that there is no strong synchronization with the page allocator
 173  * either. Pages might be freed while their page blocks are marked ISOLATED.
 174  * A call to drain_all_pages() after isolation can flush most of them. However
 175  * in some cases pages might still end up on pcp lists and that would allow
 176  * for their allocation even when they are in fact isolated already. Depending
 177  * on how strong of a guarantee the caller needs, zone_pcp_disable/enable()
 178  * might be used to flush and disable pcplist before isolation and enable after
 179  * unisolation.
 180  *
 181  * Return: 0 on success and -EBUSY if any part of range cannot be isolated.
 182  */
 183 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 184                              unsigned migratetype, int flags)
 185 {
 186         unsigned long pfn;
 187         unsigned long undo_pfn;
 188         struct page *page;
 189
 190         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 191         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 192
 193         for (pfn = start_pfn;
 194              pfn < end_pfn;
 195              pfn += pageblock_nr_pages) {
 196                 page = __first_valid_page(pfn, pageblock_nr_pages);
 197                 if (page) {
 198                         if (set_migratetype_isolate(page, migratetype, flags)) {
 199                                 undo_pfn = pfn;
 200                                 goto undo;
 201                         }
 202                 }
 203         }
 204         return 0;
 205 undo:
 206         for (pfn = start_pfn;
 207              pfn < undo_pfn;
 208              pfn += pageblock_nr_pages) {
 209                 struct page *page = pfn_to_online_page(pfn);
 210                 if (!page)
 211                         continue;
 212                 unset_migratetype_isolate(page, migratetype);
 213         }
 214
 215         return -EBUSY;
 216 }
 217
 218 /*
 219  * Make isolated pages available again.
 220  */
 221 void undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
 222                             unsigned migratetype)
 223 {
 224         unsigned long pfn;
 225         struct page *page;
 226
 227         BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
 228         BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
 229
 230         for (pfn = start_pfn;
 231              pfn < end_pfn;
 232              pfn += pageblock_nr_pages) {
 233                 page = __first_valid_page(pfn, pageblock_nr_pages);
 234                 if (!page || !is_migrate_isolate_page(page))
 235                         continue;
 236                 unset_migratetype_isolate(page, migratetype);
 237         }
 238 }
 239 /*
 240  * Test all pages in the range is free(means isolated) or not.
 241  * all pages in [start_pfn...end_pfn) must be in the same zone.
 242  * zone->lock must be held before call this.
 243  *
 244  * Returns the last tested pfn.
 245  */
 246 static unsigned long
 247 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
 248                                   int flags)
 249 {
 250         struct page *page;
 251
 252         while (pfn < end_pfn) {
 253                 if (!pfn_valid_within(pfn)) {
 254                         pfn++;
 255                         continue;
 256                 }
 257                 page = pfn_to_page(pfn);
 258                 if (PageBuddy(page))
 259                         /*
 260                          * If the page is on a free list, it has to be on
 261                          * the correct MIGRATE_ISOLATE freelist. There is no
 262                          * simple way to verify that as VM_BUG_ON(), though.
 263                          */
 264                         pfn += 1 << buddy_order(page);
 265                 else if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
 266                         /* A HWPoisoned page cannot be also PageBuddy */
 267                         pfn++;
 268                 else if ((flags & MEMORY_OFFLINE) && PageOffline(page) &&
 269                          !page_count(page))
 270                         /*
 271                          * The responsible driver agreed to skip PageOffline()
 272                          * pages when offlining memory by dropping its
 273                          * reference in MEM_GOING_OFFLINE.
 274                          */
 275                         pfn++;
 276                 else
 277                         break;
 278         }
 279
 280         return pfn;
 281 }
 282
 283 /* Caller should ensure that requested range is in a single zone */
 284 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
 285                         int isol_flags)
 286 {
 287         unsigned long pfn, flags;
 288         struct page *page;
 289         struct zone *zone;
 290
 291         /*
 292          * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
 293          * are not aligned to pageblock_nr_pages.
 294          * Then we just check migratetype first.
 295          */
 296         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 297                 page = __first_valid_page(pfn, pageblock_nr_pages);
 298                 if (page && !is_migrate_isolate_page(page))
 299                         break;
 300         }
 301         page = __first_valid_page(start_pfn, end_pfn - start_pfn);
 302         if ((pfn < end_pfn) || !page)
 303                 return -EBUSY;
 304         /* Check all pages are free or marked as ISOLATED */
 305         zone = page_zone(page);
 306         spin_lock_irqsave(&zone->lock, flags);
 307         pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn, isol_flags);
 308         spin_unlock_irqrestore(&zone->lock, flags);
 309
 310         trace_test_pages_isolated(start_pfn, end_pfn, pfn);
 311
 312         return pfn < end_pfn ? -EBUSY : 0;
 313 }