| 1 | /* |
| 2 | * linux/mm/swap_state.c |
| 3 | * |
| 4 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
| 5 | * Swap reorganised 29.12.95, Stephen Tweedie |
| 6 | * |
| 7 | * Rewritten to use page cache, (C) 1998 Stephen Tweedie |
| 8 | */ |
| 9 | #include <linux/module.h> |
| 10 | #include <linux/mm.h> |
| 11 | #include <linux/kernel_stat.h> |
| 12 | #include <linux/swap.h> |
| 13 | #include <linux/swapops.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/pagemap.h> |
| 16 | #include <linux/buffer_head.h> |
| 17 | #include <linux/backing-dev.h> |
| 18 | #include <linux/pagevec.h> |
| 19 | #include <linux/migrate.h> |
| 20 | #include <linux/page_cgroup.h> |
| 21 | |
| 22 | #include <asm/pgtable.h> |
| 23 | |
| 24 | /* |
| 25 | * swapper_space is a fiction, retained to simplify the path through |
| 26 | * vmscan's shrink_page_list, to make sync_page look nicer, and to allow |
| 27 | * future use of radix_tree tags in the swap cache. |
| 28 | */ |
| 29 | static const struct address_space_operations swap_aops = { |
| 30 | .writepage = swap_writepage, |
| 31 | .sync_page = block_sync_page, |
| 32 | .set_page_dirty = __set_page_dirty_nobuffers, |
| 33 | .migratepage = migrate_page, |
| 34 | }; |
| 35 | |
| 36 | static struct backing_dev_info swap_backing_dev_info = { |
| 37 | .name = "swap", |
| 38 | .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED, |
| 39 | .unplug_io_fn = swap_unplug_io_fn, |
| 40 | }; |
| 41 | |
| 42 | struct address_space swapper_space = { |
| 43 | .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN), |
| 44 | .tree_lock = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock), |
| 45 | .a_ops = &swap_aops, |
| 46 | .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear), |
| 47 | .backing_dev_info = &swap_backing_dev_info, |
| 48 | }; |
| 49 | |
| 50 | #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) |
| 51 | |
| 52 | static struct { |
| 53 | unsigned long add_total; |
| 54 | unsigned long del_total; |
| 55 | unsigned long find_success; |
| 56 | unsigned long find_total; |
| 57 | } swap_cache_info; |
| 58 | |
| 59 | void show_swap_cache_info(void) |
| 60 | { |
| 61 | printk("%lu pages in swap cache\n", total_swapcache_pages); |
| 62 | printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", |
| 63 | swap_cache_info.add_total, swap_cache_info.del_total, |
| 64 | swap_cache_info.find_success, swap_cache_info.find_total); |
| 65 | printk("Free swap = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10)); |
| 66 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
| 67 | } |
| 68 | |
| 69 | /* |
| 70 | * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, |
| 71 | * but sets SwapCache flag and private instead of mapping and index. |
| 72 | */ |
| 73 | static int __add_to_swap_cache(struct page *page, swp_entry_t entry) |
| 74 | { |
| 75 | int error; |
| 76 | |
| 77 | VM_BUG_ON(!PageLocked(page)); |
| 78 | VM_BUG_ON(PageSwapCache(page)); |
| 79 | VM_BUG_ON(!PageSwapBacked(page)); |
| 80 | |
| 81 | page_cache_get(page); |
| 82 | SetPageSwapCache(page); |
| 83 | set_page_private(page, entry.val); |
| 84 | |
| 85 | spin_lock_irq(&swapper_space.tree_lock); |
| 86 | error = radix_tree_insert(&swapper_space.page_tree, entry.val, page); |
| 87 | if (likely(!error)) { |
| 88 | total_swapcache_pages++; |
| 89 | __inc_zone_page_state(page, NR_FILE_PAGES); |
| 90 | INC_CACHE_INFO(add_total); |
| 91 | } |
| 92 | spin_unlock_irq(&swapper_space.tree_lock); |
| 93 | |
| 94 | if (unlikely(error)) { |
| 95 | /* |
| 96 | * Only the context which have set SWAP_HAS_CACHE flag |
| 97 | * would call add_to_swap_cache(). |
| 98 | * So add_to_swap_cache() doesn't returns -EEXIST. |
| 99 | */ |
| 100 | VM_BUG_ON(error == -EEXIST); |
| 101 | set_page_private(page, 0UL); |
| 102 | ClearPageSwapCache(page); |
| 103 | page_cache_release(page); |
| 104 | } |
| 105 | |
| 106 | return error; |
| 107 | } |
| 108 | |
| 109 | |
| 110 | int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask) |
| 111 | { |
| 112 | int error; |
| 113 | |
| 114 | error = radix_tree_preload(gfp_mask); |
| 115 | if (!error) { |
| 116 | error = __add_to_swap_cache(page, entry); |
| 117 | radix_tree_preload_end(); |
| 118 | } |
| 119 | return error; |
| 120 | } |
| 121 | |
| 122 | /* |
| 123 | * This must be called only on pages that have |
| 124 | * been verified to be in the swap cache. |
| 125 | */ |
| 126 | void __delete_from_swap_cache(struct page *page) |
| 127 | { |
| 128 | VM_BUG_ON(!PageLocked(page)); |
| 129 | VM_BUG_ON(!PageSwapCache(page)); |
| 130 | VM_BUG_ON(PageWriteback(page)); |
| 131 | |
| 132 | radix_tree_delete(&swapper_space.page_tree, page_private(page)); |
| 133 | set_page_private(page, 0); |
| 134 | ClearPageSwapCache(page); |
| 135 | total_swapcache_pages--; |
| 136 | __dec_zone_page_state(page, NR_FILE_PAGES); |
| 137 | INC_CACHE_INFO(del_total); |
| 138 | } |
| 139 | |
| 140 | /** |
| 141 | * add_to_swap - allocate swap space for a page |
| 142 | * @page: page we want to move to swap |
| 143 | * |
| 144 | * Allocate swap space for the page and add the page to the |
| 145 | * swap cache. Caller needs to hold the page lock. |
| 146 | */ |
| 147 | int add_to_swap(struct page *page) |
| 148 | { |
| 149 | swp_entry_t entry; |
| 150 | int err; |
| 151 | |
| 152 | VM_BUG_ON(!PageLocked(page)); |
| 153 | VM_BUG_ON(!PageUptodate(page)); |
| 154 | |
| 155 | entry = get_swap_page(); |
| 156 | if (!entry.val) |
| 157 | return 0; |
| 158 | |
| 159 | /* |
| 160 | * Radix-tree node allocations from PF_MEMALLOC contexts could |
| 161 | * completely exhaust the page allocator. __GFP_NOMEMALLOC |
| 162 | * stops emergency reserves from being allocated. |
| 163 | * |
| 164 | * TODO: this could cause a theoretical memory reclaim |
| 165 | * deadlock in the swap out path. |
| 166 | */ |
| 167 | /* |
| 168 | * Add it to the swap cache and mark it dirty |
| 169 | */ |
| 170 | err = add_to_swap_cache(page, entry, |
| 171 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); |
| 172 | |
| 173 | if (!err) { /* Success */ |
| 174 | SetPageDirty(page); |
| 175 | return 1; |
| 176 | } else { /* -ENOMEM radix-tree allocation failure */ |
| 177 | /* |
| 178 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
| 179 | * clear SWAP_HAS_CACHE flag. |
| 180 | */ |
| 181 | swapcache_free(entry, NULL); |
| 182 | return 0; |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | /* |
| 187 | * This must be called only on pages that have |
| 188 | * been verified to be in the swap cache and locked. |
| 189 | * It will never put the page into the free list, |
| 190 | * the caller has a reference on the page. |
| 191 | */ |
| 192 | void delete_from_swap_cache(struct page *page) |
| 193 | { |
| 194 | swp_entry_t entry; |
| 195 | |
| 196 | entry.val = page_private(page); |
| 197 | |
| 198 | spin_lock_irq(&swapper_space.tree_lock); |
| 199 | __delete_from_swap_cache(page); |
| 200 | spin_unlock_irq(&swapper_space.tree_lock); |
| 201 | |
| 202 | swapcache_free(entry, page); |
| 203 | page_cache_release(page); |
| 204 | } |
| 205 | |
| 206 | /* |
| 207 | * If we are the only user, then try to free up the swap cache. |
| 208 | * |
| 209 | * Its ok to check for PageSwapCache without the page lock |
| 210 | * here because we are going to recheck again inside |
| 211 | * try_to_free_swap() _with_ the lock. |
| 212 | * - Marcelo |
| 213 | */ |
| 214 | static inline void free_swap_cache(struct page *page) |
| 215 | { |
| 216 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
| 217 | try_to_free_swap(page); |
| 218 | unlock_page(page); |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | /* |
| 223 | * Perform a free_page(), also freeing any swap cache associated with |
| 224 | * this page if it is the last user of the page. |
| 225 | */ |
| 226 | void free_page_and_swap_cache(struct page *page) |
| 227 | { |
| 228 | free_swap_cache(page); |
| 229 | page_cache_release(page); |
| 230 | } |
| 231 | |
| 232 | /* |
| 233 | * Passed an array of pages, drop them all from swapcache and then release |
| 234 | * them. They are removed from the LRU and freed if this is their last use. |
| 235 | */ |
| 236 | void free_pages_and_swap_cache(struct page **pages, int nr) |
| 237 | { |
| 238 | struct page **pagep = pages; |
| 239 | |
| 240 | lru_add_drain(); |
| 241 | while (nr) { |
| 242 | int todo = min(nr, PAGEVEC_SIZE); |
| 243 | int i; |
| 244 | |
| 245 | for (i = 0; i < todo; i++) |
| 246 | free_swap_cache(pagep[i]); |
| 247 | release_pages(pagep, todo, 0); |
| 248 | pagep += todo; |
| 249 | nr -= todo; |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | /* |
| 254 | * Lookup a swap entry in the swap cache. A found page will be returned |
| 255 | * unlocked and with its refcount incremented - we rely on the kernel |
| 256 | * lock getting page table operations atomic even if we drop the page |
| 257 | * lock before returning. |
| 258 | */ |
| 259 | struct page * lookup_swap_cache(swp_entry_t entry) |
| 260 | { |
| 261 | struct page *page; |
| 262 | |
| 263 | page = find_get_page(&swapper_space, entry.val); |
| 264 | |
| 265 | if (page) |
| 266 | INC_CACHE_INFO(find_success); |
| 267 | |
| 268 | INC_CACHE_INFO(find_total); |
| 269 | return page; |
| 270 | } |
| 271 | |
| 272 | /* |
| 273 | * Locate a page of swap in physical memory, reserving swap cache space |
| 274 | * and reading the disk if it is not already cached. |
| 275 | * A failure return means that either the page allocation failed or that |
| 276 | * the swap entry is no longer in use. |
| 277 | */ |
| 278 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
| 279 | struct vm_area_struct *vma, unsigned long addr) |
| 280 | { |
| 281 | struct page *found_page, *new_page = NULL; |
| 282 | int err; |
| 283 | |
| 284 | do { |
| 285 | /* |
| 286 | * First check the swap cache. Since this is normally |
| 287 | * called after lookup_swap_cache() failed, re-calling |
| 288 | * that would confuse statistics. |
| 289 | */ |
| 290 | found_page = find_get_page(&swapper_space, entry.val); |
| 291 | if (found_page) |
| 292 | break; |
| 293 | |
| 294 | /* |
| 295 | * Get a new page to read into from swap. |
| 296 | */ |
| 297 | if (!new_page) { |
| 298 | new_page = alloc_page_vma(gfp_mask, vma, addr); |
| 299 | if (!new_page) |
| 300 | break; /* Out of memory */ |
| 301 | } |
| 302 | |
| 303 | /* |
| 304 | * call radix_tree_preload() while we can wait. |
| 305 | */ |
| 306 | err = radix_tree_preload(gfp_mask & GFP_KERNEL); |
| 307 | if (err) |
| 308 | break; |
| 309 | |
| 310 | /* |
| 311 | * Swap entry may have been freed since our caller observed it. |
| 312 | */ |
| 313 | err = swapcache_prepare(entry); |
| 314 | if (err == -EEXIST) { /* seems racy */ |
| 315 | radix_tree_preload_end(); |
| 316 | continue; |
| 317 | } |
| 318 | if (err) { /* swp entry is obsolete ? */ |
| 319 | radix_tree_preload_end(); |
| 320 | break; |
| 321 | } |
| 322 | |
| 323 | /* May fail (-ENOMEM) if radix-tree node allocation failed. */ |
| 324 | __set_page_locked(new_page); |
| 325 | SetPageSwapBacked(new_page); |
| 326 | err = __add_to_swap_cache(new_page, entry); |
| 327 | if (likely(!err)) { |
| 328 | radix_tree_preload_end(); |
| 329 | /* |
| 330 | * Initiate read into locked page and return. |
| 331 | */ |
| 332 | lru_cache_add_anon(new_page); |
| 333 | swap_readpage(new_page); |
| 334 | return new_page; |
| 335 | } |
| 336 | radix_tree_preload_end(); |
| 337 | ClearPageSwapBacked(new_page); |
| 338 | __clear_page_locked(new_page); |
| 339 | /* |
| 340 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
| 341 | * clear SWAP_HAS_CACHE flag. |
| 342 | */ |
| 343 | swapcache_free(entry, NULL); |
| 344 | } while (err != -ENOMEM); |
| 345 | |
| 346 | if (new_page) |
| 347 | page_cache_release(new_page); |
| 348 | return found_page; |
| 349 | } |
| 350 | |
| 351 | /** |
| 352 | * swapin_readahead - swap in pages in hope we need them soon |
| 353 | * @entry: swap entry of this memory |
| 354 | * @gfp_mask: memory allocation flags |
| 355 | * @vma: user vma this address belongs to |
| 356 | * @addr: target address for mempolicy |
| 357 | * |
| 358 | * Returns the struct page for entry and addr, after queueing swapin. |
| 359 | * |
| 360 | * Primitive swap readahead code. We simply read an aligned block of |
| 361 | * (1 << page_cluster) entries in the swap area. This method is chosen |
| 362 | * because it doesn't cost us any seek time. We also make sure to queue |
| 363 | * the 'original' request together with the readahead ones... |
| 364 | * |
| 365 | * This has been extended to use the NUMA policies from the mm triggering |
| 366 | * the readahead. |
| 367 | * |
| 368 | * Caller must hold down_read on the vma->vm_mm if vma is not NULL. |
| 369 | */ |
| 370 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, |
| 371 | struct vm_area_struct *vma, unsigned long addr) |
| 372 | { |
| 373 | int nr_pages; |
| 374 | struct page *page; |
| 375 | unsigned long offset; |
| 376 | unsigned long end_offset; |
| 377 | |
| 378 | /* |
| 379 | * Get starting offset for readaround, and number of pages to read. |
| 380 | * Adjust starting address by readbehind (for NUMA interleave case)? |
| 381 | * No, it's very unlikely that swap layout would follow vma layout, |
| 382 | * more likely that neighbouring swap pages came from the same node: |
| 383 | * so use the same "addr" to choose the same node for each swap read. |
| 384 | */ |
| 385 | nr_pages = valid_swaphandles(entry, &offset); |
| 386 | for (end_offset = offset + nr_pages; offset < end_offset; offset++) { |
| 387 | /* Ok, do the async read-ahead now */ |
| 388 | page = read_swap_cache_async(swp_entry(swp_type(entry), offset), |
| 389 | gfp_mask, vma, addr); |
| 390 | if (!page) |
| 391 | break; |
| 392 | page_cache_release(page); |
| 393 | } |
| 394 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 395 | return read_swap_cache_async(entry, gfp_mask, vma, addr); |
| 396 | } |