mm/balloon_compaction.c

   1 /*
   2  * mm/balloon_compaction.c
   3  *
   4  * Common interface for making balloon pages movable to compaction.
   5  *
   6  * Copyright (C) 2012, Red Hat, Inc.  Rafael Aquini <aquini@redhat.com>
   7  */
   8 #include <linux/mm.h>
   9 #include <linux/slab.h>
  10 #include <linux/export.h>
  11 #include <linux/balloon_compaction.h>
  12
  13 /*
  14  * balloon_devinfo_alloc - allocates a balloon device information descriptor.
  15  * @balloon_dev_descriptor: pointer to reference the balloon device which
  16  *                          this struct balloon_dev_info will be servicing.
  17  *
  18  * Driver must call it to properly allocate and initialize an instance of
  19  * struct balloon_dev_info which will be used to reference a balloon device
  20  * as well as to keep track of the balloon device page list.
  21  */
  22 struct balloon_dev_info *balloon_devinfo_alloc(void *balloon_dev_descriptor)
  23 {
  24         struct balloon_dev_info *b_dev_info;
  25         b_dev_info = kmalloc(sizeof(*b_dev_info), GFP_KERNEL);
  26         if (!b_dev_info)
  27                 return ERR_PTR(-ENOMEM);
  28
  29         b_dev_info->balloon_device = balloon_dev_descriptor;
  30         b_dev_info->mapping = NULL;
  31         b_dev_info->isolated_pages = 0;
  32         spin_lock_init(&b_dev_info->pages_lock);
  33         INIT_LIST_HEAD(&b_dev_info->pages);
  34
  35         return b_dev_info;
  36 }
  37 EXPORT_SYMBOL_GPL(balloon_devinfo_alloc);
  38
  39 /*
  40  * balloon_page_enqueue - allocates a new page and inserts it into the balloon
  41  *                        page list.
  42  * @b_dev_info: balloon device decriptor where we will insert a new page to
  43  *
  44  * Driver must call it to properly allocate a new enlisted balloon page
  45  * before definetively removing it from the guest system.
  46  * This function returns the page address for the recently enqueued page or
  47  * NULL in the case we fail to allocate a new page this turn.
  48  */
  49 struct page *balloon_page_enqueue(struct balloon_dev_info *b_dev_info)
  50 {
  51         unsigned long flags;
  52         struct page *page = alloc_page(balloon_mapping_gfp_mask() |
  53                                         __GFP_NOMEMALLOC | __GFP_NORETRY);
  54         if (!page)
  55                 return NULL;
  56
  57         BUG_ON(!trylock_page(page));
  58         spin_lock_irqsave(&b_dev_info->pages_lock, flags);
  59         balloon_page_insert(page, b_dev_info->mapping, &b_dev_info->pages);
  60         spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
  61         unlock_page(page);
  62         return page;
  63 }
  64 EXPORT_SYMBOL_GPL(balloon_page_enqueue);
  65
  66 /*
  67  * balloon_page_dequeue - removes a page from balloon's page list and returns
  68  *                        the its address to allow the driver release the page.
  69  * @b_dev_info: balloon device decriptor where we will grab a page from.
  70  *
  71  * Driver must call it to properly de-allocate a previous enlisted balloon page
  72  * before definetively releasing it back to the guest system.
  73  * This function returns the page address for the recently dequeued page or
  74  * NULL in the case we find balloon's page list temporarely empty due to
  75  * compaction isolated pages.
  76  */
  77 struct page *balloon_page_dequeue(struct balloon_dev_info *b_dev_info)
  78 {
  79         struct page *page, *tmp;
  80         unsigned long flags;
  81         bool dequeued_page;
  82
  83         dequeued_page = false;
  84         list_for_each_entry_safe(page, tmp, &b_dev_info->pages, lru) {
  85                 if (trylock_page(page)) {
  86                         spin_lock_irqsave(&b_dev_info->pages_lock, flags);
  87                         balloon_page_delete(page);
  88                         spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
  89                         unlock_page(page);
  90                         dequeued_page = true;
  91                         break;
  92                 }
  93         }
  94
  95         if (!dequeued_page) {
  96                 spin_lock_irqsave(&b_dev_info->pages_lock, flags);
  97                 if (unlikely(list_empty(&b_dev_info->pages) &&
  98                              !b_dev_info->isolated_pages))
  99                         BUG();
 100                 spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
 101                 page = NULL;
 102         }
 103         return page;
 104 }
 105 EXPORT_SYMBOL_GPL(balloon_page_dequeue);
 106
 107 #ifdef CONFIG_BALLOON_COMPACTION
 108 /*
 109  * balloon_mapping_alloc - allocates a special ->mapping for ballooned pages.
 110  * @b_dev_info: holds the balloon device information descriptor.
 111  * @a_ops: balloon_mapping address_space_operations descriptor.
 112  *
 113  * Driver must call it to properly allocate and initialize an instance of
 114  * struct address_space which will be used as the special page->mapping for
 115  * balloon device enlisted page instances.
 116  */
 117 struct address_space *balloon_mapping_alloc(struct balloon_dev_info *b_dev_info,
 118                                 const struct address_space_operations *a_ops)
 119 {
 120         struct address_space *mapping;
 121
 122         mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
 123         if (!mapping)
 124                 return ERR_PTR(-ENOMEM);
 125
 126         /*
 127          * Give a clean 'zeroed' status to all elements of this special
 128          * balloon page->mapping struct address_space instance.
 129          */
 130         address_space_init_once(mapping);
 131
 132         /*
 133          * Set mapping->flags appropriately, to allow balloon pages
 134          * ->mapping identification.
 135          */
 136         mapping_set_balloon(mapping);
 137         mapping_set_gfp_mask(mapping, balloon_mapping_gfp_mask());
 138
 139         /* balloon's page->mapping->a_ops callback descriptor */
 140         mapping->a_ops = a_ops;
 141
 142         /*
 143          * Establish a pointer reference back to the balloon device descriptor
 144          * this particular page->mapping will be servicing.
 145          * This is used by compaction / migration procedures to identify and
 146          * access the balloon device pageset while isolating / migrating pages.
 147          *
 148          * As some balloon drivers can register multiple balloon devices
 149          * for a single guest, this also helps compaction / migration to
 150          * properly deal with multiple balloon pagesets, when required.
 151          */
 152         mapping->private_data = b_dev_info;
 153         b_dev_info->mapping = mapping;
 154
 155         return mapping;
 156 }
 157 EXPORT_SYMBOL_GPL(balloon_mapping_alloc);
 158
 159 static inline void __isolate_balloon_page(struct page *page)
 160 {
 161         struct balloon_dev_info *b_dev_info = page->mapping->private_data;
 162         unsigned long flags;
 163         spin_lock_irqsave(&b_dev_info->pages_lock, flags);
 164         list_del(&page->lru);
 165         b_dev_info->isolated_pages++;
 166         spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
 167 }
 168
 169 static inline void __putback_balloon_page(struct page *page)
 170 {
 171         struct balloon_dev_info *b_dev_info = page->mapping->private_data;
 172         unsigned long flags;
 173         spin_lock_irqsave(&b_dev_info->pages_lock, flags);
 174         list_add(&page->lru, &b_dev_info->pages);
 175         b_dev_info->isolated_pages--;
 176         spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
 177 }
 178
 179 static inline int __migrate_balloon_page(struct address_space *mapping,
 180                 struct page *newpage, struct page *page, enum migrate_mode mode)
 181 {
 182         return page->mapping->a_ops->migratepage(mapping, newpage, page, mode);
 183 }
 184
 185 /* __isolate_lru_page() counterpart for a ballooned page */
 186 bool balloon_page_isolate(struct page *page)
 187 {
 188         /*
 189          * Avoid burning cycles with pages that are yet under __free_pages(),
 190          * or just got freed under us.
 191          *
 192          * In case we 'win' a race for a balloon page being freed under us and
 193          * raise its refcount preventing __free_pages() from doing its job
 194          * the put_page() at the end of this block will take care of
 195          * release this page, thus avoiding a nasty leakage.
 196          */
 197         if (likely(get_page_unless_zero(page))) {
 198                 /*
 199                  * As balloon pages are not isolated from LRU lists, concurrent
 200                  * compaction threads can race against page migration functions
 201                  * as well as race against the balloon driver releasing a page.
 202                  *
 203                  * In order to avoid having an already isolated balloon page
 204                  * being (wrongly) re-isolated while it is under migration,
 205                  * or to avoid attempting to isolate pages being released by
 206                  * the balloon driver, lets be sure we have the page lock
 207                  * before proceeding with the balloon page isolation steps.
 208                  */
 209                 if (likely(trylock_page(page))) {
 210                         /*
 211                          * A ballooned page, by default, has just one refcount.
 212                          * Prevent concurrent compaction threads from isolating
 213                          * an already isolated balloon page by refcount check.
 214                          */
 215                         if (__is_movable_balloon_page(page) &&
 216                             page_count(page) == 2) {
 217                                 __isolate_balloon_page(page);
 218                                 balloon_event_count(COMPACTBALLOONISOLATED);
 219                                 unlock_page(page);
 220                                 return true;
 221                         }
 222                         unlock_page(page);
 223                 }
 224                 put_page(page);
 225         }
 226         return false;
 227 }
 228
 229 /* putback_lru_page() counterpart for a ballooned page */
 230 void balloon_page_putback(struct page *page)
 231 {
 232         /*
 233          * 'lock_page()' stabilizes the page and prevents races against
 234          * concurrent isolation threads attempting to re-isolate it.
 235          */
 236         lock_page(page);
 237
 238         if (__is_movable_balloon_page(page)) {
 239                 __putback_balloon_page(page);
 240                 put_page(page);
 241                 balloon_event_count(COMPACTBALLOONRETURNED);
 242         } else {
 243                 WARN_ON(1);
 244                 dump_page(page);
 245         }
 246         unlock_page(page);
 247 }
 248
 249 /* move_to_new_page() counterpart for a ballooned page */
 250 int balloon_page_migrate(struct page *newpage,
 251                          struct page *page, enum migrate_mode mode)
 252 {
 253         struct address_space *mapping;
 254         int rc = -EAGAIN;
 255
 256         BUG_ON(!trylock_page(newpage));
 257
 258         if (WARN_ON(!__is_movable_balloon_page(page))) {
 259                 dump_page(page);
 260                 unlock_page(newpage);
 261                 return rc;
 262         }
 263
 264         mapping = page->mapping;
 265         if (mapping)
 266                 rc = __migrate_balloon_page(mapping, newpage, page, mode);
 267
 268         unlock_page(newpage);
 269         return rc;
 270 }
 271 #endif /* CONFIG_BALLOON_COMPACTION */