| 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/mm.h> |
| 10 | #include <linux/gfp.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/backing-dev.h> |
| 17 | #include <linux/blkdev.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. |
| 27 | */ |
| 28 | static const struct address_space_operations swap_aops = { |
| 29 | .writepage = swap_writepage, |
| 30 | .set_page_dirty = swap_set_page_dirty, |
| 31 | .migratepage = migrate_page, |
| 32 | }; |
| 33 | |
| 34 | static struct backing_dev_info swap_backing_dev_info = { |
| 35 | .name = "swap", |
| 36 | .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED, |
| 37 | }; |
| 38 | |
| 39 | struct address_space swapper_space = { |
| 40 | .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN), |
| 41 | .tree_lock = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock), |
| 42 | .a_ops = &swap_aops, |
| 43 | .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear), |
| 44 | .backing_dev_info = &swap_backing_dev_info, |
| 45 | }; |
| 46 | |
| 47 | #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) |
| 48 | |
| 49 | static struct { |
| 50 | unsigned long add_total; |
| 51 | unsigned long del_total; |
| 52 | unsigned long find_success; |
| 53 | unsigned long find_total; |
| 54 | } swap_cache_info; |
| 55 | |
| 56 | void show_swap_cache_info(void) |
| 57 | { |
| 58 | printk("%lu pages in swap cache\n", total_swapcache_pages); |
| 59 | printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", |
| 60 | swap_cache_info.add_total, swap_cache_info.del_total, |
| 61 | swap_cache_info.find_success, swap_cache_info.find_total); |
| 62 | printk("Free swap = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10)); |
| 63 | printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); |
| 64 | } |
| 65 | |
| 66 | /* |
| 67 | * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, |
| 68 | * but sets SwapCache flag and private instead of mapping and index. |
| 69 | */ |
| 70 | static int __add_to_swap_cache(struct page *page, swp_entry_t entry) |
| 71 | { |
| 72 | int error; |
| 73 | |
| 74 | VM_BUG_ON(!PageLocked(page)); |
| 75 | VM_BUG_ON(PageSwapCache(page)); |
| 76 | VM_BUG_ON(!PageSwapBacked(page)); |
| 77 | |
| 78 | page_cache_get(page); |
| 79 | SetPageSwapCache(page); |
| 80 | set_page_private(page, entry.val); |
| 81 | |
| 82 | spin_lock_irq(&swapper_space.tree_lock); |
| 83 | error = radix_tree_insert(&swapper_space.page_tree, entry.val, page); |
| 84 | if (likely(!error)) { |
| 85 | total_swapcache_pages++; |
| 86 | __inc_zone_page_state(page, NR_FILE_PAGES); |
| 87 | INC_CACHE_INFO(add_total); |
| 88 | } |
| 89 | spin_unlock_irq(&swapper_space.tree_lock); |
| 90 | |
| 91 | if (unlikely(error)) { |
| 92 | /* |
| 93 | * Only the context which have set SWAP_HAS_CACHE flag |
| 94 | * would call add_to_swap_cache(). |
| 95 | * So add_to_swap_cache() doesn't returns -EEXIST. |
| 96 | */ |
| 97 | VM_BUG_ON(error == -EEXIST); |
| 98 | set_page_private(page, 0UL); |
| 99 | ClearPageSwapCache(page); |
| 100 | page_cache_release(page); |
| 101 | } |
| 102 | |
| 103 | return error; |
| 104 | } |
| 105 | |
| 106 | |
| 107 | int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask) |
| 108 | { |
| 109 | int error; |
| 110 | |
| 111 | error = radix_tree_preload(gfp_mask); |
| 112 | if (!error) { |
| 113 | error = __add_to_swap_cache(page, entry); |
| 114 | radix_tree_preload_end(); |
| 115 | } |
| 116 | return error; |
| 117 | } |
| 118 | |
| 119 | /* |
| 120 | * This must be called only on pages that have |
| 121 | * been verified to be in the swap cache. |
| 122 | */ |
| 123 | void __delete_from_swap_cache(struct page *page) |
| 124 | { |
| 125 | VM_BUG_ON(!PageLocked(page)); |
| 126 | VM_BUG_ON(!PageSwapCache(page)); |
| 127 | VM_BUG_ON(PageWriteback(page)); |
| 128 | |
| 129 | radix_tree_delete(&swapper_space.page_tree, page_private(page)); |
| 130 | set_page_private(page, 0); |
| 131 | ClearPageSwapCache(page); |
| 132 | total_swapcache_pages--; |
| 133 | __dec_zone_page_state(page, NR_FILE_PAGES); |
| 134 | INC_CACHE_INFO(del_total); |
| 135 | } |
| 136 | |
| 137 | /** |
| 138 | * add_to_swap - allocate swap space for a page |
| 139 | * @page: page we want to move to swap |
| 140 | * |
| 141 | * Allocate swap space for the page and add the page to the |
| 142 | * swap cache. Caller needs to hold the page lock. |
| 143 | */ |
| 144 | int add_to_swap(struct page *page) |
| 145 | { |
| 146 | swp_entry_t entry; |
| 147 | int err; |
| 148 | |
| 149 | VM_BUG_ON(!PageLocked(page)); |
| 150 | VM_BUG_ON(!PageUptodate(page)); |
| 151 | |
| 152 | entry = get_swap_page(); |
| 153 | if (!entry.val) |
| 154 | return 0; |
| 155 | |
| 156 | if (unlikely(PageTransHuge(page))) |
| 157 | if (unlikely(split_huge_page(page))) { |
| 158 | swapcache_free(entry, NULL); |
| 159 | return 0; |
| 160 | } |
| 161 | |
| 162 | /* |
| 163 | * Radix-tree node allocations from PF_MEMALLOC contexts could |
| 164 | * completely exhaust the page allocator. __GFP_NOMEMALLOC |
| 165 | * stops emergency reserves from being allocated. |
| 166 | * |
| 167 | * TODO: this could cause a theoretical memory reclaim |
| 168 | * deadlock in the swap out path. |
| 169 | */ |
| 170 | /* |
| 171 | * Add it to the swap cache and mark it dirty |
| 172 | */ |
| 173 | err = add_to_swap_cache(page, entry, |
| 174 | __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); |
| 175 | |
| 176 | if (!err) { /* Success */ |
| 177 | SetPageDirty(page); |
| 178 | return 1; |
| 179 | } else { /* -ENOMEM radix-tree allocation failure */ |
| 180 | /* |
| 181 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
| 182 | * clear SWAP_HAS_CACHE flag. |
| 183 | */ |
| 184 | swapcache_free(entry, NULL); |
| 185 | return 0; |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | /* |
| 190 | * This must be called only on pages that have |
| 191 | * been verified to be in the swap cache and locked. |
| 192 | * It will never put the page into the free list, |
| 193 | * the caller has a reference on the page. |
| 194 | */ |
| 195 | void delete_from_swap_cache(struct page *page) |
| 196 | { |
| 197 | swp_entry_t entry; |
| 198 | |
| 199 | entry.val = page_private(page); |
| 200 | |
| 201 | spin_lock_irq(&swapper_space.tree_lock); |
| 202 | __delete_from_swap_cache(page); |
| 203 | spin_unlock_irq(&swapper_space.tree_lock); |
| 204 | |
| 205 | swapcache_free(entry, page); |
| 206 | page_cache_release(page); |
| 207 | } |
| 208 | |
| 209 | /* |
| 210 | * If we are the only user, then try to free up the swap cache. |
| 211 | * |
| 212 | * Its ok to check for PageSwapCache without the page lock |
| 213 | * here because we are going to recheck again inside |
| 214 | * try_to_free_swap() _with_ the lock. |
| 215 | * - Marcelo |
| 216 | */ |
| 217 | static inline void free_swap_cache(struct page *page) |
| 218 | { |
| 219 | if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { |
| 220 | try_to_free_swap(page); |
| 221 | unlock_page(page); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | /* |
| 226 | * Perform a free_page(), also freeing any swap cache associated with |
| 227 | * this page if it is the last user of the page. |
| 228 | */ |
| 229 | void free_page_and_swap_cache(struct page *page) |
| 230 | { |
| 231 | free_swap_cache(page); |
| 232 | page_cache_release(page); |
| 233 | } |
| 234 | |
| 235 | /* |
| 236 | * Passed an array of pages, drop them all from swapcache and then release |
| 237 | * them. They are removed from the LRU and freed if this is their last use. |
| 238 | */ |
| 239 | void free_pages_and_swap_cache(struct page **pages, int nr) |
| 240 | { |
| 241 | struct page **pagep = pages; |
| 242 | |
| 243 | lru_add_drain(); |
| 244 | while (nr) { |
| 245 | int todo = min(nr, PAGEVEC_SIZE); |
| 246 | int i; |
| 247 | |
| 248 | for (i = 0; i < todo; i++) |
| 249 | free_swap_cache(pagep[i]); |
| 250 | release_pages(pagep, todo, 0); |
| 251 | pagep += todo; |
| 252 | nr -= todo; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | /* |
| 257 | * Lookup a swap entry in the swap cache. A found page will be returned |
| 258 | * unlocked and with its refcount incremented - we rely on the kernel |
| 259 | * lock getting page table operations atomic even if we drop the page |
| 260 | * lock before returning. |
| 261 | */ |
| 262 | struct page * lookup_swap_cache(swp_entry_t entry) |
| 263 | { |
| 264 | struct page *page; |
| 265 | |
| 266 | page = find_get_page(&swapper_space, entry.val); |
| 267 | |
| 268 | if (page) |
| 269 | INC_CACHE_INFO(find_success); |
| 270 | |
| 271 | INC_CACHE_INFO(find_total); |
| 272 | return page; |
| 273 | } |
| 274 | |
| 275 | /* |
| 276 | * Locate a page of swap in physical memory, reserving swap cache space |
| 277 | * and reading the disk if it is not already cached. |
| 278 | * A failure return means that either the page allocation failed or that |
| 279 | * the swap entry is no longer in use. |
| 280 | */ |
| 281 | struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, |
| 282 | struct vm_area_struct *vma, unsigned long addr) |
| 283 | { |
| 284 | struct page *found_page, *new_page = NULL; |
| 285 | int err; |
| 286 | |
| 287 | do { |
| 288 | /* |
| 289 | * First check the swap cache. Since this is normally |
| 290 | * called after lookup_swap_cache() failed, re-calling |
| 291 | * that would confuse statistics. |
| 292 | */ |
| 293 | found_page = find_get_page(&swapper_space, entry.val); |
| 294 | if (found_page) |
| 295 | break; |
| 296 | |
| 297 | /* |
| 298 | * Get a new page to read into from swap. |
| 299 | */ |
| 300 | if (!new_page) { |
| 301 | new_page = alloc_page_vma(gfp_mask, vma, addr); |
| 302 | if (!new_page) |
| 303 | break; /* Out of memory */ |
| 304 | } |
| 305 | |
| 306 | /* |
| 307 | * call radix_tree_preload() while we can wait. |
| 308 | */ |
| 309 | err = radix_tree_preload(gfp_mask & GFP_KERNEL); |
| 310 | if (err) |
| 311 | break; |
| 312 | |
| 313 | /* |
| 314 | * Swap entry may have been freed since our caller observed it. |
| 315 | */ |
| 316 | err = swapcache_prepare(entry); |
| 317 | if (err == -EEXIST) { /* seems racy */ |
| 318 | radix_tree_preload_end(); |
| 319 | continue; |
| 320 | } |
| 321 | if (err) { /* swp entry is obsolete ? */ |
| 322 | radix_tree_preload_end(); |
| 323 | break; |
| 324 | } |
| 325 | |
| 326 | /* May fail (-ENOMEM) if radix-tree node allocation failed. */ |
| 327 | __set_page_locked(new_page); |
| 328 | SetPageSwapBacked(new_page); |
| 329 | err = __add_to_swap_cache(new_page, entry); |
| 330 | if (likely(!err)) { |
| 331 | radix_tree_preload_end(); |
| 332 | /* |
| 333 | * Initiate read into locked page and return. |
| 334 | */ |
| 335 | lru_cache_add_anon(new_page); |
| 336 | swap_readpage(new_page); |
| 337 | return new_page; |
| 338 | } |
| 339 | radix_tree_preload_end(); |
| 340 | ClearPageSwapBacked(new_page); |
| 341 | __clear_page_locked(new_page); |
| 342 | /* |
| 343 | * add_to_swap_cache() doesn't return -EEXIST, so we can safely |
| 344 | * clear SWAP_HAS_CACHE flag. |
| 345 | */ |
| 346 | swapcache_free(entry, NULL); |
| 347 | } while (err != -ENOMEM); |
| 348 | |
| 349 | if (new_page) |
| 350 | page_cache_release(new_page); |
| 351 | return found_page; |
| 352 | } |
| 353 | |
| 354 | /** |
| 355 | * swapin_readahead - swap in pages in hope we need them soon |
| 356 | * @entry: swap entry of this memory |
| 357 | * @gfp_mask: memory allocation flags |
| 358 | * @vma: user vma this address belongs to |
| 359 | * @addr: target address for mempolicy |
| 360 | * |
| 361 | * Returns the struct page for entry and addr, after queueing swapin. |
| 362 | * |
| 363 | * Primitive swap readahead code. We simply read an aligned block of |
| 364 | * (1 << page_cluster) entries in the swap area. This method is chosen |
| 365 | * because it doesn't cost us any seek time. We also make sure to queue |
| 366 | * the 'original' request together with the readahead ones... |
| 367 | * |
| 368 | * This has been extended to use the NUMA policies from the mm triggering |
| 369 | * the readahead. |
| 370 | * |
| 371 | * Caller must hold down_read on the vma->vm_mm if vma is not NULL. |
| 372 | */ |
| 373 | struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, |
| 374 | struct vm_area_struct *vma, unsigned long addr) |
| 375 | { |
| 376 | struct page *page; |
| 377 | unsigned long offset = swp_offset(entry); |
| 378 | unsigned long start_offset, end_offset; |
| 379 | unsigned long mask = (1UL << page_cluster) - 1; |
| 380 | struct blk_plug plug; |
| 381 | |
| 382 | /* Read a page_cluster sized and aligned cluster around offset. */ |
| 383 | start_offset = offset & ~mask; |
| 384 | end_offset = offset | mask; |
| 385 | if (!start_offset) /* First page is swap header. */ |
| 386 | start_offset++; |
| 387 | |
| 388 | blk_start_plug(&plug); |
| 389 | for (offset = start_offset; offset <= end_offset ; offset++) { |
| 390 | /* Ok, do the async read-ahead now */ |
| 391 | page = read_swap_cache_async(swp_entry(swp_type(entry), offset), |
| 392 | gfp_mask, vma, addr); |
| 393 | if (!page) |
| 394 | continue; |
| 395 | page_cache_release(page); |
| 396 | } |
| 397 | blk_finish_plug(&plug); |
| 398 | |
| 399 | lru_add_drain(); /* Push any new pages onto the LRU now */ |
| 400 | return read_swap_cache_async(entry, gfp_mask, vma, addr); |
| 401 | } |