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/gfp.h> 12 #include <linux/kernel_stat.h> 13 #include <linux/swap.h> 14 #include <linux/swapops.h> 15 #include <linux/init.h> 16 #include <linux/pagemap.h> 17 #include <linux/buffer_head.h> 18 #include <linux/backing-dev.h> 19 #include <linux/pagevec.h> 20 #include <linux/migrate.h> 21 #include <linux/page_cgroup.h> 22 23 #include <asm/pgtable.h> 24 25 /* 26 * swapper_space is a fiction, retained to simplify the path through 27 * vmscan's shrink_page_list, to make sync_page look nicer, and to allow 28 * future use of radix_tree tags in the swap cache. 29 */ 30 static const struct address_space_operations swap_aops = { 31 .writepage = swap_writepage, 32 .sync_page = block_sync_page, 33 .set_page_dirty = __set_page_dirty_nobuffers, 34 .migratepage = migrate_page, 35 }; 36 37 static struct backing_dev_info swap_backing_dev_info = { 38 .name = "swap", 39 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_SWAP_BACKED, 40 .unplug_io_fn = swap_unplug_io_fn, 41 }; 42 43 struct address_space swapper_space = { 44 .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN), 45 .tree_lock = __SPIN_LOCK_UNLOCKED(swapper_space.tree_lock), 46 .a_ops = &swap_aops, 47 .i_mmap_nonlinear = LIST_HEAD_INIT(swapper_space.i_mmap_nonlinear), 48 .backing_dev_info = &swap_backing_dev_info, 49 }; 50 51 #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) 52 53 static struct { 54 unsigned long add_total; 55 unsigned long del_total; 56 unsigned long find_success; 57 unsigned long find_total; 58 } swap_cache_info; 59 60 void show_swap_cache_info(void) 61 { 62 printk("%lu pages in swap cache\n", total_swapcache_pages); 63 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", 64 swap_cache_info.add_total, swap_cache_info.del_total, 65 swap_cache_info.find_success, swap_cache_info.find_total); 66 printk("Free swap = %ldkB\n", nr_swap_pages << (PAGE_SHIFT - 10)); 67 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); 68 } 69 70 /* 71 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, 72 * but sets SwapCache flag and private instead of mapping and index. 73 */ 74 static int __add_to_swap_cache(struct page *page, swp_entry_t entry) 75 { 76 int error; 77 78 VM_BUG_ON(!PageLocked(page)); 79 VM_BUG_ON(PageSwapCache(page)); 80 VM_BUG_ON(!PageSwapBacked(page)); 81 82 page_cache_get(page); 83 SetPageSwapCache(page); 84 set_page_private(page, entry.val); 85 86 spin_lock_irq(&swapper_space.tree_lock); 87 error = radix_tree_insert(&swapper_space.page_tree, entry.val, page); 88 if (likely(!error)) { 89 total_swapcache_pages++; 90 __inc_zone_page_state(page, NR_FILE_PAGES); 91 INC_CACHE_INFO(add_total); 92 } 93 spin_unlock_irq(&swapper_space.tree_lock); 94 95 if (unlikely(error)) { 96 /* 97 * Only the context which have set SWAP_HAS_CACHE flag 98 * would call add_to_swap_cache(). 99 * So add_to_swap_cache() doesn't returns -EEXIST. 100 */ 101 VM_BUG_ON(error == -EEXIST); 102 set_page_private(page, 0UL); 103 ClearPageSwapCache(page); 104 page_cache_release(page); 105 } 106 107 return error; 108 } 109 110 111 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask) 112 { 113 int error; 114 115 error = radix_tree_preload(gfp_mask); 116 if (!error) { 117 error = __add_to_swap_cache(page, entry); 118 radix_tree_preload_end(); 119 } 120 return error; 121 } 122 123 /* 124 * This must be called only on pages that have 125 * been verified to be in the swap cache. 126 */ 127 void __delete_from_swap_cache(struct page *page) 128 { 129 VM_BUG_ON(!PageLocked(page)); 130 VM_BUG_ON(!PageSwapCache(page)); 131 VM_BUG_ON(PageWriteback(page)); 132 133 radix_tree_delete(&swapper_space.page_tree, page_private(page)); 134 set_page_private(page, 0); 135 ClearPageSwapCache(page); 136 total_swapcache_pages--; 137 __dec_zone_page_state(page, NR_FILE_PAGES); 138 INC_CACHE_INFO(del_total); 139 } 140 141 /** 142 * add_to_swap - allocate swap space for a page 143 * @page: page we want to move to swap 144 * 145 * Allocate swap space for the page and add the page to the 146 * swap cache. Caller needs to hold the page lock. 147 */ 148 int add_to_swap(struct page *page) 149 { 150 swp_entry_t entry; 151 int err; 152 153 VM_BUG_ON(!PageLocked(page)); 154 VM_BUG_ON(!PageUptodate(page)); 155 156 entry = get_swap_page(); 157 if (!entry.val) 158 return 0; 159 160 if (unlikely(PageTransHuge(page))) 161 if (unlikely(split_huge_page(page))) { 162 swapcache_free(entry, NULL); 163 return 0; 164 } 165 166 /* 167 * Radix-tree node allocations from PF_MEMALLOC contexts could 168 * completely exhaust the page allocator. __GFP_NOMEMALLOC 169 * stops emergency reserves from being allocated. 170 * 171 * TODO: this could cause a theoretical memory reclaim 172 * deadlock in the swap out path. 173 */ 174 /* 175 * Add it to the swap cache and mark it dirty 176 */ 177 err = add_to_swap_cache(page, entry, 178 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); 179 180 if (!err) { /* Success */ 181 SetPageDirty(page); 182 return 1; 183 } else { /* -ENOMEM radix-tree allocation failure */ 184 /* 185 * add_to_swap_cache() doesn't return -EEXIST, so we can safely 186 * clear SWAP_HAS_CACHE flag. 187 */ 188 swapcache_free(entry, NULL); 189 return 0; 190 } 191 } 192 193 /* 194 * This must be called only on pages that have 195 * been verified to be in the swap cache and locked. 196 * It will never put the page into the free list, 197 * the caller has a reference on the page. 198 */ 199 void delete_from_swap_cache(struct page *page) 200 { 201 swp_entry_t entry; 202 203 entry.val = page_private(page); 204 205 spin_lock_irq(&swapper_space.tree_lock); 206 __delete_from_swap_cache(page); 207 spin_unlock_irq(&swapper_space.tree_lock); 208 209 swapcache_free(entry, page); 210 page_cache_release(page); 211 } 212 213 /* 214 * If we are the only user, then try to free up the swap cache. 215 * 216 * Its ok to check for PageSwapCache without the page lock 217 * here because we are going to recheck again inside 218 * try_to_free_swap() _with_ the lock. 219 * - Marcelo 220 */ 221 static inline void free_swap_cache(struct page *page) 222 { 223 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { 224 try_to_free_swap(page); 225 unlock_page(page); 226 } 227 } 228 229 /* 230 * Perform a free_page(), also freeing any swap cache associated with 231 * this page if it is the last user of the page. 232 */ 233 void free_page_and_swap_cache(struct page *page) 234 { 235 free_swap_cache(page); 236 page_cache_release(page); 237 } 238 239 /* 240 * Passed an array of pages, drop them all from swapcache and then release 241 * them. They are removed from the LRU and freed if this is their last use. 242 */ 243 void free_pages_and_swap_cache(struct page **pages, int nr) 244 { 245 struct page **pagep = pages; 246 247 lru_add_drain(); 248 while (nr) { 249 int todo = min(nr, PAGEVEC_SIZE); 250 int i; 251 252 for (i = 0; i < todo; i++) 253 free_swap_cache(pagep[i]); 254 release_pages(pagep, todo, 0); 255 pagep += todo; 256 nr -= todo; 257 } 258 } 259 260 /* 261 * Lookup a swap entry in the swap cache. A found page will be returned 262 * unlocked and with its refcount incremented - we rely on the kernel 263 * lock getting page table operations atomic even if we drop the page 264 * lock before returning. 265 */ 266 struct page * lookup_swap_cache(swp_entry_t entry) 267 { 268 struct page *page; 269 270 page = find_get_page(&swapper_space, entry.val); 271 272 if (page) 273 INC_CACHE_INFO(find_success); 274 275 INC_CACHE_INFO(find_total); 276 return page; 277 } 278 279 /* 280 * Locate a page of swap in physical memory, reserving swap cache space 281 * and reading the disk if it is not already cached. 282 * A failure return means that either the page allocation failed or that 283 * the swap entry is no longer in use. 284 */ 285 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, 286 struct vm_area_struct *vma, unsigned long addr) 287 { 288 struct page *found_page, *new_page = NULL; 289 int err; 290 291 do { 292 /* 293 * First check the swap cache. Since this is normally 294 * called after lookup_swap_cache() failed, re-calling 295 * that would confuse statistics. 296 */ 297 found_page = find_get_page(&swapper_space, entry.val); 298 if (found_page) 299 break; 300 301 /* 302 * Get a new page to read into from swap. 303 */ 304 if (!new_page) { 305 new_page = alloc_page_vma(gfp_mask, vma, addr); 306 if (!new_page) 307 break; /* Out of memory */ 308 } 309 310 /* 311 * call radix_tree_preload() while we can wait. 312 */ 313 err = radix_tree_preload(gfp_mask & GFP_KERNEL); 314 if (err) 315 break; 316 317 /* 318 * Swap entry may have been freed since our caller observed it. 319 */ 320 err = swapcache_prepare(entry); 321 if (err == -EEXIST) { /* seems racy */ 322 radix_tree_preload_end(); 323 continue; 324 } 325 if (err) { /* swp entry is obsolete ? */ 326 radix_tree_preload_end(); 327 break; 328 } 329 330 /* May fail (-ENOMEM) if radix-tree node allocation failed. */ 331 __set_page_locked(new_page); 332 SetPageSwapBacked(new_page); 333 err = __add_to_swap_cache(new_page, entry); 334 if (likely(!err)) { 335 radix_tree_preload_end(); 336 /* 337 * Initiate read into locked page and return. 338 */ 339 lru_cache_add_anon(new_page); 340 swap_readpage(new_page); 341 return new_page; 342 } 343 radix_tree_preload_end(); 344 ClearPageSwapBacked(new_page); 345 __clear_page_locked(new_page); 346 /* 347 * add_to_swap_cache() doesn't return -EEXIST, so we can safely 348 * clear SWAP_HAS_CACHE flag. 349 */ 350 swapcache_free(entry, NULL); 351 } while (err != -ENOMEM); 352 353 if (new_page) 354 page_cache_release(new_page); 355 return found_page; 356 } 357 358 /** 359 * swapin_readahead - swap in pages in hope we need them soon 360 * @entry: swap entry of this memory 361 * @gfp_mask: memory allocation flags 362 * @vma: user vma this address belongs to 363 * @addr: target address for mempolicy 364 * 365 * Returns the struct page for entry and addr, after queueing swapin. 366 * 367 * Primitive swap readahead code. We simply read an aligned block of 368 * (1 << page_cluster) entries in the swap area. This method is chosen 369 * because it doesn't cost us any seek time. We also make sure to queue 370 * the 'original' request together with the readahead ones... 371 * 372 * This has been extended to use the NUMA policies from the mm triggering 373 * the readahead. 374 * 375 * Caller must hold down_read on the vma->vm_mm if vma is not NULL. 376 */ 377 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, 378 struct vm_area_struct *vma, unsigned long addr) 379 { 380 int nr_pages; 381 struct page *page; 382 unsigned long offset; 383 unsigned long end_offset; 384 385 /* 386 * Get starting offset for readaround, and number of pages to read. 387 * Adjust starting address by readbehind (for NUMA interleave case)? 388 * No, it's very unlikely that swap layout would follow vma layout, 389 * more likely that neighbouring swap pages came from the same node: 390 * so use the same "addr" to choose the same node for each swap read. 391 */ 392 nr_pages = valid_swaphandles(entry, &offset); 393 for (end_offset = offset + nr_pages; offset < end_offset; offset++) { 394 /* Ok, do the async read-ahead now */ 395 page = read_swap_cache_async(swp_entry(swp_type(entry), offset), 396 gfp_mask, vma, addr); 397 if (!page) 398 break; 399 page_cache_release(page); 400 } 401 lru_add_drain(); /* Push any new pages onto the LRU now */ 402 return read_swap_cache_async(entry, gfp_mask, vma, addr); 403 } 404