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_spaces[MAX_SWAPFILES] = { 40 [0 ... MAX_SWAPFILES - 1] = { 41 .page_tree = RADIX_TREE_INIT(GFP_ATOMIC|__GFP_NOWARN), 42 .i_mmap_writable = ATOMIC_INIT(0), 43 .a_ops = &swap_aops, 44 .backing_dev_info = &swap_backing_dev_info, 45 } 46 }; 47 48 #define INC_CACHE_INFO(x) do { swap_cache_info.x++; } while (0) 49 50 static struct { 51 unsigned long add_total; 52 unsigned long del_total; 53 unsigned long find_success; 54 unsigned long find_total; 55 } swap_cache_info; 56 57 unsigned long total_swapcache_pages(void) 58 { 59 int i; 60 unsigned long ret = 0; 61 62 for (i = 0; i < MAX_SWAPFILES; i++) 63 ret += swapper_spaces[i].nrpages; 64 return ret; 65 } 66 67 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4); 68 69 void show_swap_cache_info(void) 70 { 71 printk("%lu pages in swap cache\n", total_swapcache_pages()); 72 printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n", 73 swap_cache_info.add_total, swap_cache_info.del_total, 74 swap_cache_info.find_success, swap_cache_info.find_total); 75 printk("Free swap = %ldkB\n", 76 get_nr_swap_pages() << (PAGE_SHIFT - 10)); 77 printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10)); 78 } 79 80 /* 81 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space, 82 * but sets SwapCache flag and private instead of mapping and index. 83 */ 84 int __add_to_swap_cache(struct page *page, swp_entry_t entry) 85 { 86 int error; 87 struct address_space *address_space; 88 89 VM_BUG_ON_PAGE(!PageLocked(page), page); 90 VM_BUG_ON_PAGE(PageSwapCache(page), page); 91 VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 92 93 page_cache_get(page); 94 SetPageSwapCache(page); 95 set_page_private(page, entry.val); 96 97 address_space = swap_address_space(entry); 98 spin_lock_irq(&address_space->tree_lock); 99 error = radix_tree_insert(&address_space->page_tree, 100 entry.val, page); 101 if (likely(!error)) { 102 address_space->nrpages++; 103 __inc_zone_page_state(page, NR_FILE_PAGES); 104 INC_CACHE_INFO(add_total); 105 } 106 spin_unlock_irq(&address_space->tree_lock); 107 108 if (unlikely(error)) { 109 /* 110 * Only the context which have set SWAP_HAS_CACHE flag 111 * would call add_to_swap_cache(). 112 * So add_to_swap_cache() doesn't returns -EEXIST. 113 */ 114 VM_BUG_ON(error == -EEXIST); 115 set_page_private(page, 0UL); 116 ClearPageSwapCache(page); 117 page_cache_release(page); 118 } 119 120 return error; 121 } 122 123 124 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask) 125 { 126 int error; 127 128 error = radix_tree_maybe_preload(gfp_mask); 129 if (!error) { 130 error = __add_to_swap_cache(page, entry); 131 radix_tree_preload_end(); 132 } 133 return error; 134 } 135 136 /* 137 * This must be called only on pages that have 138 * been verified to be in the swap cache. 139 */ 140 void __delete_from_swap_cache(struct page *page) 141 { 142 swp_entry_t entry; 143 struct address_space *address_space; 144 145 VM_BUG_ON_PAGE(!PageLocked(page), page); 146 VM_BUG_ON_PAGE(!PageSwapCache(page), page); 147 VM_BUG_ON_PAGE(PageWriteback(page), page); 148 149 entry.val = page_private(page); 150 address_space = swap_address_space(entry); 151 radix_tree_delete(&address_space->page_tree, page_private(page)); 152 set_page_private(page, 0); 153 ClearPageSwapCache(page); 154 address_space->nrpages--; 155 __dec_zone_page_state(page, NR_FILE_PAGES); 156 INC_CACHE_INFO(del_total); 157 } 158 159 /** 160 * add_to_swap - allocate swap space for a page 161 * @page: page we want to move to swap 162 * 163 * Allocate swap space for the page and add the page to the 164 * swap cache. Caller needs to hold the page lock. 165 */ 166 int add_to_swap(struct page *page, struct list_head *list) 167 { 168 swp_entry_t entry; 169 int err; 170 171 VM_BUG_ON_PAGE(!PageLocked(page), page); 172 VM_BUG_ON_PAGE(!PageUptodate(page), page); 173 174 entry = get_swap_page(); 175 if (!entry.val) 176 return 0; 177 178 if (unlikely(PageTransHuge(page))) 179 if (unlikely(split_huge_page_to_list(page, list))) { 180 swapcache_free(entry); 181 return 0; 182 } 183 184 /* 185 * Radix-tree node allocations from PF_MEMALLOC contexts could 186 * completely exhaust the page allocator. __GFP_NOMEMALLOC 187 * stops emergency reserves from being allocated. 188 * 189 * TODO: this could cause a theoretical memory reclaim 190 * deadlock in the swap out path. 191 */ 192 /* 193 * Add it to the swap cache and mark it dirty 194 */ 195 err = add_to_swap_cache(page, entry, 196 __GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN); 197 198 if (!err) { /* Success */ 199 SetPageDirty(page); 200 return 1; 201 } else { /* -ENOMEM radix-tree allocation failure */ 202 /* 203 * add_to_swap_cache() doesn't return -EEXIST, so we can safely 204 * clear SWAP_HAS_CACHE flag. 205 */ 206 swapcache_free(entry); 207 return 0; 208 } 209 } 210 211 /* 212 * This must be called only on pages that have 213 * been verified to be in the swap cache and locked. 214 * It will never put the page into the free list, 215 * the caller has a reference on the page. 216 */ 217 void delete_from_swap_cache(struct page *page) 218 { 219 swp_entry_t entry; 220 struct address_space *address_space; 221 222 entry.val = page_private(page); 223 224 address_space = swap_address_space(entry); 225 spin_lock_irq(&address_space->tree_lock); 226 __delete_from_swap_cache(page); 227 spin_unlock_irq(&address_space->tree_lock); 228 229 swapcache_free(entry); 230 page_cache_release(page); 231 } 232 233 /* 234 * If we are the only user, then try to free up the swap cache. 235 * 236 * Its ok to check for PageSwapCache without the page lock 237 * here because we are going to recheck again inside 238 * try_to_free_swap() _with_ the lock. 239 * - Marcelo 240 */ 241 static inline void free_swap_cache(struct page *page) 242 { 243 if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) { 244 try_to_free_swap(page); 245 unlock_page(page); 246 } 247 } 248 249 /* 250 * Perform a free_page(), also freeing any swap cache associated with 251 * this page if it is the last user of the page. 252 */ 253 void free_page_and_swap_cache(struct page *page) 254 { 255 free_swap_cache(page); 256 page_cache_release(page); 257 } 258 259 /* 260 * Passed an array of pages, drop them all from swapcache and then release 261 * them. They are removed from the LRU and freed if this is their last use. 262 */ 263 void free_pages_and_swap_cache(struct page **pages, int nr) 264 { 265 struct page **pagep = pages; 266 267 lru_add_drain(); 268 while (nr) { 269 int todo = min(nr, PAGEVEC_SIZE); 270 int i; 271 272 for (i = 0; i < todo; i++) 273 free_swap_cache(pagep[i]); 274 release_pages(pagep, todo, false); 275 pagep += todo; 276 nr -= todo; 277 } 278 } 279 280 /* 281 * Lookup a swap entry in the swap cache. A found page will be returned 282 * unlocked and with its refcount incremented - we rely on the kernel 283 * lock getting page table operations atomic even if we drop the page 284 * lock before returning. 285 */ 286 struct page * lookup_swap_cache(swp_entry_t entry) 287 { 288 struct page *page; 289 290 page = find_get_page(swap_address_space(entry), entry.val); 291 292 if (page) { 293 INC_CACHE_INFO(find_success); 294 if (TestClearPageReadahead(page)) 295 atomic_inc(&swapin_readahead_hits); 296 } 297 298 INC_CACHE_INFO(find_total); 299 return page; 300 } 301 302 /* 303 * Locate a page of swap in physical memory, reserving swap cache space 304 * and reading the disk if it is not already cached. 305 * A failure return means that either the page allocation failed or that 306 * the swap entry is no longer in use. 307 */ 308 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask, 309 struct vm_area_struct *vma, unsigned long addr) 310 { 311 struct page *found_page, *new_page = NULL; 312 int err; 313 314 do { 315 /* 316 * First check the swap cache. Since this is normally 317 * called after lookup_swap_cache() failed, re-calling 318 * that would confuse statistics. 319 */ 320 found_page = find_get_page(swap_address_space(entry), 321 entry.val); 322 if (found_page) 323 break; 324 325 /* 326 * Get a new page to read into from swap. 327 */ 328 if (!new_page) { 329 new_page = alloc_page_vma(gfp_mask, vma, addr); 330 if (!new_page) 331 break; /* Out of memory */ 332 } 333 334 /* 335 * call radix_tree_preload() while we can wait. 336 */ 337 err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL); 338 if (err) 339 break; 340 341 /* 342 * Swap entry may have been freed since our caller observed it. 343 */ 344 err = swapcache_prepare(entry); 345 if (err == -EEXIST) { 346 radix_tree_preload_end(); 347 /* 348 * We might race against get_swap_page() and stumble 349 * across a SWAP_HAS_CACHE swap_map entry whose page 350 * has not been brought into the swapcache yet, while 351 * the other end is scheduled away waiting on discard 352 * I/O completion at scan_swap_map(). 353 * 354 * In order to avoid turning this transitory state 355 * into a permanent loop around this -EEXIST case 356 * if !CONFIG_PREEMPT and the I/O completion happens 357 * to be waiting on the CPU waitqueue where we are now 358 * busy looping, we just conditionally invoke the 359 * scheduler here, if there are some more important 360 * tasks to run. 361 */ 362 cond_resched(); 363 continue; 364 } 365 if (err) { /* swp entry is obsolete ? */ 366 radix_tree_preload_end(); 367 break; 368 } 369 370 /* May fail (-ENOMEM) if radix-tree node allocation failed. */ 371 __set_page_locked(new_page); 372 SetPageSwapBacked(new_page); 373 err = __add_to_swap_cache(new_page, entry); 374 if (likely(!err)) { 375 radix_tree_preload_end(); 376 /* 377 * Initiate read into locked page and return. 378 */ 379 lru_cache_add_anon(new_page); 380 swap_readpage(new_page); 381 return new_page; 382 } 383 radix_tree_preload_end(); 384 ClearPageSwapBacked(new_page); 385 __clear_page_locked(new_page); 386 /* 387 * add_to_swap_cache() doesn't return -EEXIST, so we can safely 388 * clear SWAP_HAS_CACHE flag. 389 */ 390 swapcache_free(entry); 391 } while (err != -ENOMEM); 392 393 if (new_page) 394 page_cache_release(new_page); 395 return found_page; 396 } 397 398 static unsigned long swapin_nr_pages(unsigned long offset) 399 { 400 static unsigned long prev_offset; 401 unsigned int pages, max_pages, last_ra; 402 static atomic_t last_readahead_pages; 403 404 max_pages = 1 << ACCESS_ONCE(page_cluster); 405 if (max_pages <= 1) 406 return 1; 407 408 /* 409 * This heuristic has been found to work well on both sequential and 410 * random loads, swapping to hard disk or to SSD: please don't ask 411 * what the "+ 2" means, it just happens to work well, that's all. 412 */ 413 pages = atomic_xchg(&swapin_readahead_hits, 0) + 2; 414 if (pages == 2) { 415 /* 416 * We can have no readahead hits to judge by: but must not get 417 * stuck here forever, so check for an adjacent offset instead 418 * (and don't even bother to check whether swap type is same). 419 */ 420 if (offset != prev_offset + 1 && offset != prev_offset - 1) 421 pages = 1; 422 prev_offset = offset; 423 } else { 424 unsigned int roundup = 4; 425 while (roundup < pages) 426 roundup <<= 1; 427 pages = roundup; 428 } 429 430 if (pages > max_pages) 431 pages = max_pages; 432 433 /* Don't shrink readahead too fast */ 434 last_ra = atomic_read(&last_readahead_pages) / 2; 435 if (pages < last_ra) 436 pages = last_ra; 437 atomic_set(&last_readahead_pages, pages); 438 439 return pages; 440 } 441 442 /** 443 * swapin_readahead - swap in pages in hope we need them soon 444 * @entry: swap entry of this memory 445 * @gfp_mask: memory allocation flags 446 * @vma: user vma this address belongs to 447 * @addr: target address for mempolicy 448 * 449 * Returns the struct page for entry and addr, after queueing swapin. 450 * 451 * Primitive swap readahead code. We simply read an aligned block of 452 * (1 << page_cluster) entries in the swap area. This method is chosen 453 * because it doesn't cost us any seek time. We also make sure to queue 454 * the 'original' request together with the readahead ones... 455 * 456 * This has been extended to use the NUMA policies from the mm triggering 457 * the readahead. 458 * 459 * Caller must hold down_read on the vma->vm_mm if vma is not NULL. 460 */ 461 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask, 462 struct vm_area_struct *vma, unsigned long addr) 463 { 464 struct page *page; 465 unsigned long entry_offset = swp_offset(entry); 466 unsigned long offset = entry_offset; 467 unsigned long start_offset, end_offset; 468 unsigned long mask; 469 struct blk_plug plug; 470 471 mask = swapin_nr_pages(offset) - 1; 472 if (!mask) 473 goto skip; 474 475 /* Read a page_cluster sized and aligned cluster around offset. */ 476 start_offset = offset & ~mask; 477 end_offset = offset | mask; 478 if (!start_offset) /* First page is swap header. */ 479 start_offset++; 480 481 blk_start_plug(&plug); 482 for (offset = start_offset; offset <= end_offset ; offset++) { 483 /* Ok, do the async read-ahead now */ 484 page = read_swap_cache_async(swp_entry(swp_type(entry), offset), 485 gfp_mask, vma, addr); 486 if (!page) 487 continue; 488 if (offset != entry_offset) 489 SetPageReadahead(page); 490 page_cache_release(page); 491 } 492 blk_finish_plug(&plug); 493 494 lru_add_drain(); /* Push any new pages onto the LRU now */ 495 skip: 496 return read_swap_cache_async(entry, gfp_mask, vma, addr); 497 } 498