Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 *  linux/mm/swapfile.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 *  Swap reorganised 29.12.95, Stephen Tweedie
   6 */
   7
   8#include <linux/mm.h>
 
 
   9#include <linux/hugetlb.h>
  10#include <linux/mman.h>
  11#include <linux/slab.h>
  12#include <linux/kernel_stat.h>
  13#include <linux/swap.h>
  14#include <linux/vmalloc.h>
  15#include <linux/pagemap.h>
  16#include <linux/namei.h>
  17#include <linux/shmem_fs.h>
  18#include <linux/blkdev.h>
  19#include <linux/random.h>
  20#include <linux/writeback.h>
  21#include <linux/proc_fs.h>
  22#include <linux/seq_file.h>
  23#include <linux/init.h>
  24#include <linux/module.h>
  25#include <linux/ksm.h>
  26#include <linux/rmap.h>
  27#include <linux/security.h>
  28#include <linux/backing-dev.h>
  29#include <linux/mutex.h>
  30#include <linux/capability.h>
  31#include <linux/syscalls.h>
  32#include <linux/memcontrol.h>
  33#include <linux/poll.h>
  34#include <linux/oom.h>
 
 
 
 
 
  35
  36#include <asm/pgtable.h>
  37#include <asm/tlbflush.h>
  38#include <linux/swapops.h>
  39#include <linux/page_cgroup.h>
  40
  41static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
  42				 unsigned char);
  43static void free_swap_count_continuations(struct swap_info_struct *);
  44static sector_t map_swap_entry(swp_entry_t, struct block_device**);
  45
  46static DEFINE_SPINLOCK(swap_lock);
  47static unsigned int nr_swapfiles;
  48long nr_swap_pages;
 
 
 
 
 
 
 
  49long total_swap_pages;
  50static int least_priority;
  51
  52static const char Bad_file[] = "Bad swap file entry ";
  53static const char Unused_file[] = "Unused swap file entry ";
  54static const char Bad_offset[] = "Bad swap offset entry ";
  55static const char Unused_offset[] = "Unused swap offset entry ";
  56
  57static struct swap_list_t swap_list = {-1, -1};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  58
  59static struct swap_info_struct *swap_info[MAX_SWAPFILES];
  60
  61static DEFINE_MUTEX(swapon_mutex);
  62
  63static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
  64/* Activity counter to indicate that a swapon or swapoff has occurred */
  65static atomic_t proc_poll_event = ATOMIC_INIT(0);
  66
 
 
  67static inline unsigned char swap_count(unsigned char ent)
  68{
  69	return ent & ~SWAP_HAS_CACHE;	/* may include SWAP_HAS_CONT flag */
  70}
  71
  72/* returns 1 if swap entry is freed */
  73static int
  74__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
  75{
  76	swp_entry_t entry = swp_entry(si->type, offset);
  77	struct page *page;
  78	int ret = 0;
  79
  80	page = find_get_page(&swapper_space, entry.val);
  81	if (!page)
  82		return 0;
  83	/*
  84	 * This function is called from scan_swap_map() and it's called
  85	 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
  86	 * We have to use trylock for avoiding deadlock. This is a special
  87	 * case and you should use try_to_free_swap() with explicit lock_page()
  88	 * in usual operations.
  89	 */
  90	if (trylock_page(page)) {
  91		ret = try_to_free_swap(page);
  92		unlock_page(page);
  93	}
  94	page_cache_release(page);
  95	return ret;
  96}
  97
  98/*
  99 * swapon tell device that all the old swap contents can be discarded,
 100 * to allow the swap device to optimize its wear-levelling.
 101 */
 102static int discard_swap(struct swap_info_struct *si)
 103{
 104	struct swap_extent *se;
 105	sector_t start_block;
 106	sector_t nr_blocks;
 107	int err = 0;
 108
 109	/* Do not discard the swap header page! */
 110	se = &si->first_swap_extent;
 111	start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 112	nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 113	if (nr_blocks) {
 114		err = blkdev_issue_discard(si->bdev, start_block,
 115				nr_blocks, GFP_KERNEL, 0);
 116		if (err)
 117			return err;
 118		cond_resched();
 119	}
 120
 121	list_for_each_entry(se, &si->first_swap_extent.list, list) {
 122		start_block = se->start_block << (PAGE_SHIFT - 9);
 123		nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 124
 125		err = blkdev_issue_discard(si->bdev, start_block,
 126				nr_blocks, GFP_KERNEL, 0);
 127		if (err)
 128			break;
 129
 130		cond_resched();
 131	}
 132	return err;		/* That will often be -EOPNOTSUPP */
 133}
 134
 135/*
 136 * swap allocation tell device that a cluster of swap can now be discarded,
 137 * to allow the swap device to optimize its wear-levelling.
 138 */
 139static void discard_swap_cluster(struct swap_info_struct *si,
 140				 pgoff_t start_page, pgoff_t nr_pages)
 141{
 142	struct swap_extent *se = si->curr_swap_extent;
 143	int found_extent = 0;
 144
 145	while (nr_pages) {
 146		struct list_head *lh;
 147
 148		if (se->start_page <= start_page &&
 149		    start_page < se->start_page + se->nr_pages) {
 150			pgoff_t offset = start_page - se->start_page;
 151			sector_t start_block = se->start_block + offset;
 152			sector_t nr_blocks = se->nr_pages - offset;
 153
 154			if (nr_blocks > nr_pages)
 155				nr_blocks = nr_pages;
 156			start_page += nr_blocks;
 157			nr_pages -= nr_blocks;
 158
 159			if (!found_extent++)
 160				si->curr_swap_extent = se;
 161
 162			start_block <<= PAGE_SHIFT - 9;
 163			nr_blocks <<= PAGE_SHIFT - 9;
 164			if (blkdev_issue_discard(si->bdev, start_block,
 165				    nr_blocks, GFP_NOIO, 0))
 166				break;
 167		}
 168
 169		lh = se->list.next;
 170		se = list_entry(lh, struct swap_extent, list);
 171	}
 172}
 173
 174static int wait_for_discard(void *word)
 
 
 
 
 
 
 
 
 175{
 176	schedule();
 177	return 0;
 178}
 179
 180#define SWAPFILE_CLUSTER	256
 181#define LATENCY_LIMIT		256
 
 
 182
 183static unsigned long scan_swap_map(struct swap_info_struct *si,
 184				   unsigned char usage)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 185{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 186	unsigned long offset;
 187	unsigned long scan_base;
 188	unsigned long last_in_cluster = 0;
 189	int latency_ration = LATENCY_LIMIT;
 190	int found_free_cluster = 0;
 
 
 
 191
 192	/*
 193	 * We try to cluster swap pages by allocating them sequentially
 194	 * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 195	 * way, however, we resort to first-free allocation, starting
 196	 * a new cluster.  This prevents us from scattering swap pages
 197	 * all over the entire swap partition, so that we reduce
 198	 * overall disk seek times between swap pages.  -- sct
 199	 * But we do now try to find an empty cluster.  -Andrea
 200	 * And we let swap pages go all over an SSD partition.  Hugh
 201	 */
 202
 203	si->flags += SWP_SCANNING;
 204	scan_base = offset = si->cluster_next;
 205
 
 
 
 
 
 
 
 
 206	if (unlikely(!si->cluster_nr--)) {
 207		if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 208			si->cluster_nr = SWAPFILE_CLUSTER - 1;
 209			goto checks;
 210		}
 211		if (si->flags & SWP_DISCARDABLE) {
 212			/*
 213			 * Start range check on racing allocations, in case
 214			 * they overlap the cluster we eventually decide on
 215			 * (we scan without swap_lock to allow preemption).
 216			 * It's hardly conceivable that cluster_nr could be
 217			 * wrapped during our scan, but don't depend on it.
 218			 */
 219			if (si->lowest_alloc)
 220				goto checks;
 221			si->lowest_alloc = si->max;
 222			si->highest_alloc = 0;
 223		}
 224		spin_unlock(&swap_lock);
 225
 226		/*
 227		 * If seek is expensive, start searching for new cluster from
 228		 * start of partition, to minimize the span of allocated swap.
 229		 * But if seek is cheap, search from our current position, so
 230		 * that swap is allocated from all over the partition: if the
 231		 * Flash Translation Layer only remaps within limited zones,
 232		 * we don't want to wear out the first zone too quickly.
 233		 */
 234		if (!(si->flags & SWP_SOLIDSTATE))
 235			scan_base = offset = si->lowest_bit;
 236		last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 237
 238		/* Locate the first empty (unaligned) cluster */
 239		for (; last_in_cluster <= si->highest_bit; offset++) {
 240			if (si->swap_map[offset])
 241				last_in_cluster = offset + SWAPFILE_CLUSTER;
 242			else if (offset == last_in_cluster) {
 243				spin_lock(&swap_lock);
 244				offset -= SWAPFILE_CLUSTER - 1;
 245				si->cluster_next = offset;
 246				si->cluster_nr = SWAPFILE_CLUSTER - 1;
 247				found_free_cluster = 1;
 248				goto checks;
 249			}
 250			if (unlikely(--latency_ration < 0)) {
 251				cond_resched();
 252				latency_ration = LATENCY_LIMIT;
 253			}
 254		}
 255
 256		offset = si->lowest_bit;
 257		last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 258
 259		/* Locate the first empty (unaligned) cluster */
 260		for (; last_in_cluster < scan_base; offset++) {
 261			if (si->swap_map[offset])
 262				last_in_cluster = offset + SWAPFILE_CLUSTER;
 263			else if (offset == last_in_cluster) {
 264				spin_lock(&swap_lock);
 265				offset -= SWAPFILE_CLUSTER - 1;
 266				si->cluster_next = offset;
 267				si->cluster_nr = SWAPFILE_CLUSTER - 1;
 268				found_free_cluster = 1;
 269				goto checks;
 270			}
 271			if (unlikely(--latency_ration < 0)) {
 272				cond_resched();
 273				latency_ration = LATENCY_LIMIT;
 274			}
 275		}
 276
 277		offset = scan_base;
 278		spin_lock(&swap_lock);
 279		si->cluster_nr = SWAPFILE_CLUSTER - 1;
 280		si->lowest_alloc = 0;
 281	}
 282
 283checks:
 
 
 
 
 
 
 
 
 
 
 284	if (!(si->flags & SWP_WRITEOK))
 285		goto no_page;
 286	if (!si->highest_bit)
 287		goto no_page;
 288	if (offset > si->highest_bit)
 289		scan_base = offset = si->lowest_bit;
 290
 
 291	/* reuse swap entry of cache-only swap if not busy. */
 292	if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 293		int swap_was_freed;
 294		spin_unlock(&swap_lock);
 
 295		swap_was_freed = __try_to_reclaim_swap(si, offset);
 296		spin_lock(&swap_lock);
 297		/* entry was freed successfully, try to use this again */
 298		if (swap_was_freed)
 299			goto checks;
 300		goto scan; /* check next one */
 301	}
 302
 303	if (si->swap_map[offset])
 304		goto scan;
 305
 306	if (offset == si->lowest_bit)
 307		si->lowest_bit++;
 308	if (offset == si->highest_bit)
 309		si->highest_bit--;
 310	si->inuse_pages++;
 311	if (si->inuse_pages == si->pages) {
 312		si->lowest_bit = si->max;
 313		si->highest_bit = 0;
 314	}
 315	si->swap_map[offset] = usage;
 316	si->cluster_next = offset + 1;
 317	si->flags -= SWP_SCANNING;
 318
 319	if (si->lowest_alloc) {
 320		/*
 321		 * Only set when SWP_DISCARDABLE, and there's a scan
 322		 * for a free cluster in progress or just completed.
 323		 */
 324		if (found_free_cluster) {
 325			/*
 326			 * To optimize wear-levelling, discard the
 327			 * old data of the cluster, taking care not to
 328			 * discard any of its pages that have already
 329			 * been allocated by racing tasks (offset has
 330			 * already stepped over any at the beginning).
 331			 */
 332			if (offset < si->highest_alloc &&
 333			    si->lowest_alloc <= last_in_cluster)
 334				last_in_cluster = si->lowest_alloc - 1;
 335			si->flags |= SWP_DISCARDING;
 336			spin_unlock(&swap_lock);
 337
 338			if (offset < last_in_cluster)
 339				discard_swap_cluster(si, offset,
 340					last_in_cluster - offset + 1);
 341
 342			spin_lock(&swap_lock);
 343			si->lowest_alloc = 0;
 344			si->flags &= ~SWP_DISCARDING;
 
 
 
 
 
 
 
 
 345
 346			smp_mb();	/* wake_up_bit advises this */
 347			wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 
 
 
 
 
 
 
 348
 349		} else if (si->flags & SWP_DISCARDING) {
 350			/*
 351			 * Delay using pages allocated by racing tasks
 352			 * until the whole discard has been issued. We
 353			 * could defer that delay until swap_writepage,
 354			 * but it's easier to keep this self-contained.
 355			 */
 356			spin_unlock(&swap_lock);
 357			wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 358				wait_for_discard, TASK_UNINTERRUPTIBLE);
 359			spin_lock(&swap_lock);
 360		} else {
 361			/*
 362			 * Note pages allocated by racing tasks while
 363			 * scan for a free cluster is in progress, so
 364			 * that its final discard can exclude them.
 365			 */
 366			if (offset < si->lowest_alloc)
 367				si->lowest_alloc = offset;
 368			if (offset > si->highest_alloc)
 369				si->highest_alloc = offset;
 370		}
 371	}
 372	return offset;
 
 
 
 373
 374scan:
 375	spin_unlock(&swap_lock);
 376	while (++offset <= si->highest_bit) {
 377		if (!si->swap_map[offset]) {
 378			spin_lock(&swap_lock);
 379			goto checks;
 380		}
 381		if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 382			spin_lock(&swap_lock);
 383			goto checks;
 384		}
 385		if (unlikely(--latency_ration < 0)) {
 386			cond_resched();
 387			latency_ration = LATENCY_LIMIT;
 388		}
 389	}
 390	offset = si->lowest_bit;
 391	while (++offset < scan_base) {
 392		if (!si->swap_map[offset]) {
 393			spin_lock(&swap_lock);
 394			goto checks;
 395		}
 396		if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 397			spin_lock(&swap_lock);
 398			goto checks;
 399		}
 400		if (unlikely(--latency_ration < 0)) {
 401			cond_resched();
 402			latency_ration = LATENCY_LIMIT;
 403		}
 
 404	}
 405	spin_lock(&swap_lock);
 406
 407no_page:
 408	si->flags -= SWP_SCANNING;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 409	return 0;
 410}
 
 411
 412swp_entry_t get_swap_page(void)
 
 413{
 414	struct swap_info_struct *si;
 415	pgoff_t offset;
 416	int type, next;
 417	int wrapped = 0;
 418
 419	spin_lock(&swap_lock);
 420	if (nr_swap_pages <= 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 421		goto noswap;
 422	nr_swap_pages--;
 423
 424	for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 425		si = swap_info[type];
 426		next = si->next;
 427		if (next < 0 ||
 428		    (!wrapped && si->prio != swap_info[next]->prio)) {
 429			next = swap_list.head;
 430			wrapped++;
 431		}
 432
 433		if (!si->highest_bit)
 434			continue;
 435		if (!(si->flags & SWP_WRITEOK))
 436			continue;
 437
 438		swap_list.next = next;
 439		/* This is called for allocating swap entry for cache */
 440		offset = scan_swap_map(si, SWAP_HAS_CACHE);
 441		if (offset) {
 442			spin_unlock(&swap_lock);
 443			return swp_entry(type, offset);
 444		}
 445		next = swap_list.next;
 446	}
 447
 448	nr_swap_pages++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 449noswap:
 450	spin_unlock(&swap_lock);
 451	return (swp_entry_t) {0};
 452}
 453
 454/* The only caller of this function is now susupend routine */
 455swp_entry_t get_swap_page_of_type(int type)
 456{
 457	struct swap_info_struct *si;
 458	pgoff_t offset;
 459
 460	spin_lock(&swap_lock);
 461	si = swap_info[type];
 
 462	if (si && (si->flags & SWP_WRITEOK)) {
 463		nr_swap_pages--;
 464		/* This is called for allocating swap entry, not cache */
 465		offset = scan_swap_map(si, 1);
 466		if (offset) {
 467			spin_unlock(&swap_lock);
 468			return swp_entry(type, offset);
 469		}
 470		nr_swap_pages++;
 471	}
 472	spin_unlock(&swap_lock);
 473	return (swp_entry_t) {0};
 474}
 475
 476static struct swap_info_struct *swap_info_get(swp_entry_t entry)
 477{
 478	struct swap_info_struct *p;
 479	unsigned long offset, type;
 480
 481	if (!entry.val)
 482		goto out;
 483	type = swp_type(entry);
 484	if (type >= nr_swapfiles)
 485		goto bad_nofile;
 486	p = swap_info[type];
 487	if (!(p->flags & SWP_USED))
 488		goto bad_device;
 489	offset = swp_offset(entry);
 490	if (offset >= p->max)
 491		goto bad_offset;
 492	if (!p->swap_map[offset])
 493		goto bad_free;
 494	spin_lock(&swap_lock);
 495	return p;
 496
 497bad_free:
 498	printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 499	goto out;
 500bad_offset:
 501	printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 502	goto out;
 503bad_device:
 504	printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 505	goto out;
 506bad_nofile:
 507	printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 508out:
 509	return NULL;
 510}
 511
 512static unsigned char swap_entry_free(struct swap_info_struct *p,
 513				     swp_entry_t entry, unsigned char usage)
 514{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 515	unsigned long offset = swp_offset(entry);
 516	unsigned char count;
 517	unsigned char has_cache;
 518
 
 
 519	count = p->swap_map[offset];
 
 520	has_cache = count & SWAP_HAS_CACHE;
 521	count &= ~SWAP_HAS_CACHE;
 522
 523	if (usage == SWAP_HAS_CACHE) {
 524		VM_BUG_ON(!has_cache);
 525		has_cache = 0;
 526	} else if (count == SWAP_MAP_SHMEM) {
 527		/*
 528		 * Or we could insist on shmem.c using a special
 529		 * swap_shmem_free() and free_shmem_swap_and_cache()...
 530		 */
 531		count = 0;
 532	} else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
 533		if (count == COUNT_CONTINUED) {
 534			if (swap_count_continued(p, offset, count))
 535				count = SWAP_MAP_MAX | COUNT_CONTINUED;
 536			else
 537				count = SWAP_MAP_MAX;
 538		} else
 539			count--;
 540	}
 541
 542	if (!count)
 543		mem_cgroup_uncharge_swap(entry);
 544
 545	usage = count | has_cache;
 546	p->swap_map[offset] = usage;
 547
 548	/* free if no reference */
 549	if (!usage) {
 550		struct gendisk *disk = p->bdev->bd_disk;
 551		if (offset < p->lowest_bit)
 552			p->lowest_bit = offset;
 553		if (offset > p->highest_bit)
 554			p->highest_bit = offset;
 555		if (swap_list.next >= 0 &&
 556		    p->prio > swap_info[swap_list.next]->prio)
 557			swap_list.next = p->type;
 558		nr_swap_pages++;
 559		p->inuse_pages--;
 560		if ((p->flags & SWP_BLKDEV) &&
 561				disk->fops->swap_slot_free_notify)
 562			disk->fops->swap_slot_free_notify(p->bdev, offset);
 563	}
 564
 565	return usage;
 566}
 567
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 568/*
 569 * Caller has made sure that the swapdevice corresponding to entry
 570 * is still around or has not been recycled.
 571 */
 572void swap_free(swp_entry_t entry)
 573{
 574	struct swap_info_struct *p;
 575
 576	p = swap_info_get(entry);
 577	if (p) {
 578		swap_entry_free(p, entry, 1);
 579		spin_unlock(&swap_lock);
 580	}
 581}
 582
 583/*
 584 * Called after dropping swapcache to decrease refcnt to swap entries.
 585 */
 586void swapcache_free(swp_entry_t entry, struct page *page)
 587{
 588	struct swap_info_struct *p;
 589	unsigned char count;
 590
 591	p = swap_info_get(entry);
 592	if (p) {
 593		count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
 594		if (page)
 595			mem_cgroup_uncharge_swapcache(page, entry, count != 0);
 596		spin_unlock(&swap_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 597	}
 598}
 599
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 600/*
 601 * How many references to page are currently swapped out?
 602 * This does not give an exact answer when swap count is continued,
 603 * but does include the high COUNT_CONTINUED flag to allow for that.
 604 */
 605static inline int page_swapcount(struct page *page)
 606{
 607	int count = 0;
 608	struct swap_info_struct *p;
 
 609	swp_entry_t entry;
 
 610
 611	entry.val = page_private(page);
 612	p = swap_info_get(entry);
 613	if (p) {
 614		count = swap_count(p->swap_map[swp_offset(entry)]);
 615		spin_unlock(&swap_lock);
 
 
 616	}
 617	return count;
 618}
 619
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 620/*
 621 * We can write to an anon page without COW if there are no other references
 622 * to it.  And as a side-effect, free up its swap: because the old content
 623 * on disk will never be read, and seeking back there to write new content
 624 * later would only waste time away from clustering.
 
 
 
 
 625 */
 626int reuse_swap_page(struct page *page)
 627{
 628	int count;
 629
 630	VM_BUG_ON(!PageLocked(page));
 631	if (unlikely(PageKsm(page)))
 632		return 0;
 633	count = page_mapcount(page);
 634	if (count <= 1 && PageSwapCache(page)) {
 635		count += page_swapcount(page);
 636		if (count == 1 && !PageWriteback(page)) {
 
 
 
 
 
 
 637			delete_from_swap_cache(page);
 638			SetPageDirty(page);
 
 
 
 
 
 
 
 
 
 
 
 639		}
 640	}
 
 641	return count <= 1;
 642}
 643
 644/*
 645 * If swap is getting full, or if there are no more mappings of this page,
 646 * then try_to_free_swap is called to free its swap space.
 647 */
 648int try_to_free_swap(struct page *page)
 649{
 650	VM_BUG_ON(!PageLocked(page));
 651
 652	if (!PageSwapCache(page))
 653		return 0;
 654	if (PageWriteback(page))
 655		return 0;
 656	if (page_swapcount(page))
 657		return 0;
 658
 659	/*
 660	 * Once hibernation has begun to create its image of memory,
 661	 * there's a danger that one of the calls to try_to_free_swap()
 662	 * - most probably a call from __try_to_reclaim_swap() while
 663	 * hibernation is allocating its own swap pages for the image,
 664	 * but conceivably even a call from memory reclaim - will free
 665	 * the swap from a page which has already been recorded in the
 666	 * image as a clean swapcache page, and then reuse its swap for
 667	 * another page of the image.  On waking from hibernation, the
 668	 * original page might be freed under memory pressure, then
 669	 * later read back in from swap, now with the wrong data.
 670	 *
 671	 * Hibernation clears bits from gfp_allowed_mask to prevent
 672	 * memory reclaim from writing to disk, so check that here.
 673	 */
 674	if (!(gfp_allowed_mask & __GFP_IO))
 675		return 0;
 676
 
 677	delete_from_swap_cache(page);
 678	SetPageDirty(page);
 679	return 1;
 680}
 681
 682/*
 683 * Free the swap entry like above, but also try to
 684 * free the page cache entry if it is the last user.
 685 */
 686int free_swap_and_cache(swp_entry_t entry)
 687{
 688	struct swap_info_struct *p;
 689	struct page *page = NULL;
 
 690
 691	if (non_swap_entry(entry))
 692		return 1;
 693
 694	p = swap_info_get(entry);
 695	if (p) {
 696		if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
 697			page = find_get_page(&swapper_space, entry.val);
 
 
 
 698			if (page && !trylock_page(page)) {
 699				page_cache_release(page);
 700				page = NULL;
 701			}
 702		}
 703		spin_unlock(&swap_lock);
 704	}
 705	if (page) {
 706		/*
 707		 * Not mapped elsewhere, or swap space full? Free it!
 708		 * Also recheck PageSwapCache now page is locked (above).
 709		 */
 710		if (PageSwapCache(page) && !PageWriteback(page) &&
 711				(!page_mapped(page) || vm_swap_full())) {
 
 
 712			delete_from_swap_cache(page);
 713			SetPageDirty(page);
 714		}
 715		unlock_page(page);
 716		page_cache_release(page);
 717	}
 718	return p != NULL;
 719}
 720
 721#ifdef CONFIG_CGROUP_MEM_RES_CTLR
 722/**
 723 * mem_cgroup_count_swap_user - count the user of a swap entry
 724 * @ent: the swap entry to be checked
 725 * @pagep: the pointer for the swap cache page of the entry to be stored
 726 *
 727 * Returns the number of the user of the swap entry. The number is valid only
 728 * for swaps of anonymous pages.
 729 * If the entry is found on swap cache, the page is stored to pagep with
 730 * refcount of it being incremented.
 731 */
 732int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
 733{
 734	struct page *page;
 735	struct swap_info_struct *p;
 736	int count = 0;
 737
 738	page = find_get_page(&swapper_space, ent.val);
 739	if (page)
 740		count += page_mapcount(page);
 741	p = swap_info_get(ent);
 742	if (p) {
 743		count += swap_count(p->swap_map[swp_offset(ent)]);
 744		spin_unlock(&swap_lock);
 745	}
 746
 747	*pagep = page;
 748	return count;
 749}
 750#endif
 751
 752#ifdef CONFIG_HIBERNATION
 753/*
 754 * Find the swap type that corresponds to given device (if any).
 755 *
 756 * @offset - number of the PAGE_SIZE-sized block of the device, starting
 757 * from 0, in which the swap header is expected to be located.
 758 *
 759 * This is needed for the suspend to disk (aka swsusp).
 760 */
 761int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 762{
 763	struct block_device *bdev = NULL;
 764	int type;
 765
 766	if (device)
 767		bdev = bdget(device);
 768
 769	spin_lock(&swap_lock);
 770	for (type = 0; type < nr_swapfiles; type++) {
 771		struct swap_info_struct *sis = swap_info[type];
 772
 773		if (!(sis->flags & SWP_WRITEOK))
 774			continue;
 775
 776		if (!bdev) {
 777			if (bdev_p)
 778				*bdev_p = bdgrab(sis->bdev);
 779
 780			spin_unlock(&swap_lock);
 781			return type;
 782		}
 783		if (bdev == sis->bdev) {
 784			struct swap_extent *se = &sis->first_swap_extent;
 785
 786			if (se->start_block == offset) {
 787				if (bdev_p)
 788					*bdev_p = bdgrab(sis->bdev);
 789
 790				spin_unlock(&swap_lock);
 791				bdput(bdev);
 792				return type;
 793			}
 794		}
 795	}
 796	spin_unlock(&swap_lock);
 797	if (bdev)
 798		bdput(bdev);
 799
 800	return -ENODEV;
 801}
 802
 803/*
 804 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 805 * corresponding to given index in swap_info (swap type).
 806 */
 807sector_t swapdev_block(int type, pgoff_t offset)
 808{
 809	struct block_device *bdev;
 810
 811	if ((unsigned int)type >= nr_swapfiles)
 812		return 0;
 813	if (!(swap_info[type]->flags & SWP_WRITEOK))
 814		return 0;
 815	return map_swap_entry(swp_entry(type, offset), &bdev);
 816}
 817
 818/*
 819 * Return either the total number of swap pages of given type, or the number
 820 * of free pages of that type (depending on @free)
 821 *
 822 * This is needed for software suspend
 823 */
 824unsigned int count_swap_pages(int type, int free)
 825{
 826	unsigned int n = 0;
 827
 828	spin_lock(&swap_lock);
 829	if ((unsigned int)type < nr_swapfiles) {
 830		struct swap_info_struct *sis = swap_info[type];
 831
 
 832		if (sis->flags & SWP_WRITEOK) {
 833			n = sis->pages;
 834			if (free)
 835				n -= sis->inuse_pages;
 836		}
 
 837	}
 838	spin_unlock(&swap_lock);
 839	return n;
 840}
 841#endif /* CONFIG_HIBERNATION */
 842
 
 
 
 
 
 843/*
 844 * No need to decide whether this PTE shares the swap entry with others,
 845 * just let do_wp_page work it out if a write is requested later - to
 846 * force COW, vm_page_prot omits write permission from any private vma.
 847 */
 848static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 849		unsigned long addr, swp_entry_t entry, struct page *page)
 850{
 851	struct mem_cgroup *ptr;
 
 852	spinlock_t *ptl;
 853	pte_t *pte;
 854	int ret = 1;
 855
 856	if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
 
 
 
 
 
 
 857		ret = -ENOMEM;
 858		goto out_nolock;
 859	}
 860
 861	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 862	if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 863		if (ret > 0)
 864			mem_cgroup_cancel_charge_swapin(ptr);
 865		ret = 0;
 866		goto out;
 867	}
 868
 869	dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
 870	inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
 871	get_page(page);
 872	set_pte_at(vma->vm_mm, addr, pte,
 873		   pte_mkold(mk_pte(page, vma->vm_page_prot)));
 874	page_add_anon_rmap(page, vma, addr);
 875	mem_cgroup_commit_charge_swapin(page, ptr);
 
 
 
 
 
 
 876	swap_free(entry);
 877	/*
 878	 * Move the page to the active list so it is not
 879	 * immediately swapped out again after swapon.
 880	 */
 881	activate_page(page);
 882out:
 883	pte_unmap_unlock(pte, ptl);
 884out_nolock:
 
 
 
 
 885	return ret;
 886}
 887
 888static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 889				unsigned long addr, unsigned long end,
 890				swp_entry_t entry, struct page *page)
 891{
 892	pte_t swp_pte = swp_entry_to_pte(entry);
 893	pte_t *pte;
 894	int ret = 0;
 895
 896	/*
 897	 * We don't actually need pte lock while scanning for swp_pte: since
 898	 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 899	 * page table while we're scanning; though it could get zapped, and on
 900	 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 901	 * of unmatched parts which look like swp_pte, so unuse_pte must
 902	 * recheck under pte lock.  Scanning without pte lock lets it be
 903	 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 904	 */
 905	pte = pte_offset_map(pmd, addr);
 906	do {
 907		/*
 908		 * swapoff spends a _lot_ of time in this loop!
 909		 * Test inline before going to call unuse_pte.
 910		 */
 911		if (unlikely(pte_same(*pte, swp_pte))) {
 912			pte_unmap(pte);
 913			ret = unuse_pte(vma, pmd, addr, entry, page);
 914			if (ret)
 915				goto out;
 916			pte = pte_offset_map(pmd, addr);
 917		}
 918	} while (pte++, addr += PAGE_SIZE, addr != end);
 919	pte_unmap(pte - 1);
 920out:
 921	return ret;
 922}
 923
 924static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 925				unsigned long addr, unsigned long end,
 926				swp_entry_t entry, struct page *page)
 927{
 928	pmd_t *pmd;
 929	unsigned long next;
 930	int ret;
 931
 932	pmd = pmd_offset(pud, addr);
 933	do {
 
 934		next = pmd_addr_end(addr, end);
 935		if (unlikely(pmd_trans_huge(*pmd)))
 936			continue;
 937		if (pmd_none_or_clear_bad(pmd))
 938			continue;
 939		ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 940		if (ret)
 941			return ret;
 942	} while (pmd++, addr = next, addr != end);
 943	return 0;
 944}
 945
 946static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 947				unsigned long addr, unsigned long end,
 948				swp_entry_t entry, struct page *page)
 949{
 950	pud_t *pud;
 951	unsigned long next;
 952	int ret;
 953
 954	pud = pud_offset(pgd, addr);
 955	do {
 956		next = pud_addr_end(addr, end);
 957		if (pud_none_or_clear_bad(pud))
 958			continue;
 959		ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 960		if (ret)
 961			return ret;
 962	} while (pud++, addr = next, addr != end);
 963	return 0;
 964}
 965
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 966static int unuse_vma(struct vm_area_struct *vma,
 967				swp_entry_t entry, struct page *page)
 968{
 969	pgd_t *pgd;
 970	unsigned long addr, end, next;
 971	int ret;
 972
 973	if (page_anon_vma(page)) {
 974		addr = page_address_in_vma(page, vma);
 975		if (addr == -EFAULT)
 976			return 0;
 977		else
 978			end = addr + PAGE_SIZE;
 979	} else {
 980		addr = vma->vm_start;
 981		end = vma->vm_end;
 982	}
 983
 984	pgd = pgd_offset(vma->vm_mm, addr);
 985	do {
 986		next = pgd_addr_end(addr, end);
 987		if (pgd_none_or_clear_bad(pgd))
 988			continue;
 989		ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 990		if (ret)
 991			return ret;
 992	} while (pgd++, addr = next, addr != end);
 993	return 0;
 994}
 995
 996static int unuse_mm(struct mm_struct *mm,
 997				swp_entry_t entry, struct page *page)
 998{
 999	struct vm_area_struct *vma;
1000	int ret = 0;
1001
1002	if (!down_read_trylock(&mm->mmap_sem)) {
1003		/*
1004		 * Activate page so shrink_inactive_list is unlikely to unmap
1005		 * its ptes while lock is dropped, so swapoff can make progress.
1006		 */
1007		activate_page(page);
1008		unlock_page(page);
1009		down_read(&mm->mmap_sem);
1010		lock_page(page);
1011	}
1012	for (vma = mm->mmap; vma; vma = vma->vm_next) {
1013		if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1014			break;
 
1015	}
1016	up_read(&mm->mmap_sem);
1017	return (ret < 0)? ret: 0;
1018}
1019
1020/*
1021 * Scan swap_map from current position to next entry still in use.
 
1022 * Recycle to start on reaching the end, returning 0 when empty.
1023 */
1024static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1025					unsigned int prev)
1026{
1027	unsigned int max = si->max;
1028	unsigned int i = prev;
1029	unsigned char count;
1030
1031	/*
1032	 * No need for swap_lock here: we're just looking
1033	 * for whether an entry is in use, not modifying it; false
1034	 * hits are okay, and sys_swapoff() has already prevented new
1035	 * allocations from this area (while holding swap_lock).
1036	 */
1037	for (;;) {
1038		if (++i >= max) {
1039			if (!prev) {
1040				i = 0;
1041				break;
1042			}
1043			/*
1044			 * No entries in use at top of swap_map,
1045			 * loop back to start and recheck there.
1046			 */
1047			max = prev + 1;
1048			prev = 0;
1049			i = 1;
1050		}
1051		count = si->swap_map[i];
1052		if (count && swap_count(count) != SWAP_MAP_BAD)
1053			break;
 
 
 
1054	}
1055	return i;
1056}
1057
1058/*
1059 * We completely avoid races by reading each swap page in advance,
1060 * and then search for the process using it.  All the necessary
1061 * page table adjustments can then be made atomically.
 
 
 
1062 */
1063static int try_to_unuse(unsigned int type)
 
1064{
1065	struct swap_info_struct *si = swap_info[type];
1066	struct mm_struct *start_mm;
1067	unsigned char *swap_map;
 
 
 
 
1068	unsigned char swcount;
1069	struct page *page;
1070	swp_entry_t entry;
1071	unsigned int i = 0;
1072	int retval = 0;
1073
1074	/*
1075	 * When searching mms for an entry, a good strategy is to
1076	 * start at the first mm we freed the previous entry from
1077	 * (though actually we don't notice whether we or coincidence
1078	 * freed the entry).  Initialize this start_mm with a hold.
1079	 *
1080	 * A simpler strategy would be to start at the last mm we
1081	 * freed the previous entry from; but that would take less
1082	 * advantage of mmlist ordering, which clusters forked mms
1083	 * together, child after parent.  If we race with dup_mmap(), we
1084	 * prefer to resolve parent before child, lest we miss entries
1085	 * duplicated after we scanned child: using last mm would invert
1086	 * that.
1087	 */
1088	start_mm = &init_mm;
1089	atomic_inc(&init_mm.mm_users);
1090
1091	/*
1092	 * Keep on scanning until all entries have gone.  Usually,
1093	 * one pass through swap_map is enough, but not necessarily:
1094	 * there are races when an instance of an entry might be missed.
1095	 */
1096	while ((i = find_next_to_unuse(si, i)) != 0) {
1097		if (signal_pending(current)) {
1098			retval = -EINTR;
1099			break;
1100		}
1101
1102		/*
1103		 * Get a page for the entry, using the existing swap
1104		 * cache page if there is one.  Otherwise, get a clean
1105		 * page and read the swap into it.
1106		 */
1107		swap_map = &si->swap_map[i];
1108		entry = swp_entry(type, i);
1109		page = read_swap_cache_async(entry,
1110					GFP_HIGHUSER_MOVABLE, NULL, 0);
1111		if (!page) {
1112			/*
1113			 * Either swap_duplicate() failed because entry
1114			 * has been freed independently, and will not be
1115			 * reused since sys_swapoff() already disabled
1116			 * allocation from here, or alloc_page() failed.
1117			 */
1118			if (!*swap_map)
 
 
 
 
 
 
 
 
1119				continue;
1120			retval = -ENOMEM;
1121			break;
1122		}
1123
1124		/*
1125		 * Don't hold on to start_mm if it looks like exiting.
1126		 */
1127		if (atomic_read(&start_mm->mm_users) == 1) {
1128			mmput(start_mm);
1129			start_mm = &init_mm;
1130			atomic_inc(&init_mm.mm_users);
1131		}
1132
1133		/*
1134		 * Wait for and lock page.  When do_swap_page races with
1135		 * try_to_unuse, do_swap_page can handle the fault much
1136		 * faster than try_to_unuse can locate the entry.  This
1137		 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1138		 * defer to do_swap_page in such a case - in some tests,
1139		 * do_swap_page and try_to_unuse repeatedly compete.
1140		 */
1141		wait_on_page_locked(page);
1142		wait_on_page_writeback(page);
1143		lock_page(page);
1144		wait_on_page_writeback(page);
1145
1146		/*
1147		 * Remove all references to entry.
1148		 */
1149		swcount = *swap_map;
1150		if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1151			retval = shmem_unuse(entry, page);
1152			/* page has already been unlocked and released */
1153			if (retval < 0)
1154				break;
1155			continue;
1156		}
1157		if (swap_count(swcount) && start_mm != &init_mm)
1158			retval = unuse_mm(start_mm, entry, page);
1159
1160		if (swap_count(*swap_map)) {
1161			int set_start_mm = (*swap_map >= swcount);
1162			struct list_head *p = &start_mm->mmlist;
1163			struct mm_struct *new_start_mm = start_mm;
1164			struct mm_struct *prev_mm = start_mm;
1165			struct mm_struct *mm;
1166
1167			atomic_inc(&new_start_mm->mm_users);
1168			atomic_inc(&prev_mm->mm_users);
1169			spin_lock(&mmlist_lock);
1170			while (swap_count(*swap_map) && !retval &&
1171					(p = p->next) != &start_mm->mmlist) {
1172				mm = list_entry(p, struct mm_struct, mmlist);
1173				if (!atomic_inc_not_zero(&mm->mm_users))
1174					continue;
1175				spin_unlock(&mmlist_lock);
1176				mmput(prev_mm);
1177				prev_mm = mm;
1178
1179				cond_resched();
1180
1181				swcount = *swap_map;
1182				if (!swap_count(swcount)) /* any usage ? */
1183					;
1184				else if (mm == &init_mm)
1185					set_start_mm = 1;
1186				else
1187					retval = unuse_mm(mm, entry, page);
1188
1189				if (set_start_mm && *swap_map < swcount) {
1190					mmput(new_start_mm);
1191					atomic_inc(&mm->mm_users);
1192					new_start_mm = mm;
1193					set_start_mm = 0;
1194				}
1195				spin_lock(&mmlist_lock);
1196			}
1197			spin_unlock(&mmlist_lock);
1198			mmput(prev_mm);
1199			mmput(start_mm);
1200			start_mm = new_start_mm;
1201		}
1202		if (retval) {
1203			unlock_page(page);
1204			page_cache_release(page);
1205			break;
1206		}
1207
1208		/*
1209		 * If a reference remains (rare), we would like to leave
1210		 * the page in the swap cache; but try_to_unmap could
1211		 * then re-duplicate the entry once we drop page lock,
1212		 * so we might loop indefinitely; also, that page could
1213		 * not be swapped out to other storage meanwhile.  So:
1214		 * delete from cache even if there's another reference,
1215		 * after ensuring that the data has been saved to disk -
1216		 * since if the reference remains (rarer), it will be
1217		 * read from disk into another page.  Splitting into two
1218		 * pages would be incorrect if swap supported "shared
1219		 * private" pages, but they are handled by tmpfs files.
1220		 *
1221		 * Given how unuse_vma() targets one particular offset
1222		 * in an anon_vma, once the anon_vma has been determined,
1223		 * this splitting happens to be just what is needed to
1224		 * handle where KSM pages have been swapped out: re-reading
1225		 * is unnecessarily slow, but we can fix that later on.
1226		 */
1227		if (swap_count(*swap_map) &&
1228		     PageDirty(page) && PageSwapCache(page)) {
1229			struct writeback_control wbc = {
1230				.sync_mode = WB_SYNC_NONE,
1231			};
1232
1233			swap_writepage(page, &wbc);
1234			lock_page(page);
1235			wait_on_page_writeback(page);
1236		}
1237
1238		/*
1239		 * It is conceivable that a racing task removed this page from
1240		 * swap cache just before we acquired the page lock at the top,
1241		 * or while we dropped it in unuse_mm().  The page might even
1242		 * be back in swap cache on another swap area: that we must not
1243		 * delete, since it may not have been written out to swap yet.
1244		 */
1245		if (PageSwapCache(page) &&
1246		    likely(page_private(page) == entry.val))
1247			delete_from_swap_cache(page);
 
1248
1249		/*
1250		 * So we could skip searching mms once swap count went
1251		 * to 1, we did not mark any present ptes as dirty: must
1252		 * mark page dirty so shrink_page_list will preserve it.
1253		 */
1254		SetPageDirty(page);
1255		unlock_page(page);
1256		page_cache_release(page);
1257
1258		/*
1259		 * Make sure that we aren't completely killing
1260		 * interactive performance.
1261		 */
1262		cond_resched();
 
 
 
 
1263	}
1264
1265	mmput(start_mm);
1266	return retval;
1267}
1268
1269/*
1270 * After a successful try_to_unuse, if no swap is now in use, we know
1271 * we can empty the mmlist.  swap_lock must be held on entry and exit.
1272 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1273 * added to the mmlist just after page_duplicate - before would be racy.
1274 */
1275static void drain_mmlist(void)
1276{
1277	struct list_head *p, *next;
1278	unsigned int type;
1279
1280	for (type = 0; type < nr_swapfiles; type++)
1281		if (swap_info[type]->inuse_pages)
1282			return;
1283	spin_lock(&mmlist_lock);
1284	list_for_each_safe(p, next, &init_mm.mmlist)
1285		list_del_init(p);
1286	spin_unlock(&mmlist_lock);
1287}
1288
1289/*
1290 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1291 * corresponds to page offset for the specified swap entry.
1292 * Note that the type of this function is sector_t, but it returns page offset
1293 * into the bdev, not sector offset.
1294 */
1295static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1296{
1297	struct swap_info_struct *sis;
1298	struct swap_extent *start_se;
1299	struct swap_extent *se;
1300	pgoff_t offset;
1301
1302	sis = swap_info[swp_type(entry)];
1303	*bdev = sis->bdev;
1304
1305	offset = swp_offset(entry);
1306	start_se = sis->curr_swap_extent;
1307	se = start_se;
1308
1309	for ( ; ; ) {
1310		struct list_head *lh;
1311
1312		if (se->start_page <= offset &&
1313				offset < (se->start_page + se->nr_pages)) {
1314			return se->start_block + (offset - se->start_page);
1315		}
1316		lh = se->list.next;
1317		se = list_entry(lh, struct swap_extent, list);
1318		sis->curr_swap_extent = se;
1319		BUG_ON(se == start_se);		/* It *must* be present */
1320	}
1321}
1322
1323/*
1324 * Returns the page offset into bdev for the specified page's swap entry.
1325 */
1326sector_t map_swap_page(struct page *page, struct block_device **bdev)
1327{
1328	swp_entry_t entry;
1329	entry.val = page_private(page);
1330	return map_swap_entry(entry, bdev);
1331}
1332
1333/*
1334 * Free all of a swapdev's extent information
1335 */
1336static void destroy_swap_extents(struct swap_info_struct *sis)
1337{
1338	while (!list_empty(&sis->first_swap_extent.list)) {
1339		struct swap_extent *se;
1340
1341		se = list_entry(sis->first_swap_extent.list.next,
1342				struct swap_extent, list);
1343		list_del(&se->list);
1344		kfree(se);
1345	}
 
 
 
 
 
 
 
 
1346}
1347
1348/*
1349 * Add a block range (and the corresponding page range) into this swapdev's
1350 * extent list.  The extent list is kept sorted in page order.
1351 *
1352 * This function rather assumes that it is called in ascending page order.
1353 */
1354static int
1355add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1356		unsigned long nr_pages, sector_t start_block)
1357{
1358	struct swap_extent *se;
1359	struct swap_extent *new_se;
1360	struct list_head *lh;
1361
1362	if (start_page == 0) {
1363		se = &sis->first_swap_extent;
1364		sis->curr_swap_extent = se;
1365		se->start_page = 0;
1366		se->nr_pages = nr_pages;
1367		se->start_block = start_block;
1368		return 1;
1369	} else {
1370		lh = sis->first_swap_extent.list.prev;	/* Highest extent */
1371		se = list_entry(lh, struct swap_extent, list);
1372		BUG_ON(se->start_page + se->nr_pages != start_page);
1373		if (se->start_block + se->nr_pages == start_block) {
1374			/* Merge it */
1375			se->nr_pages += nr_pages;
1376			return 0;
1377		}
1378	}
1379
1380	/*
1381	 * No merge.  Insert a new extent, preserving ordering.
1382	 */
1383	new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1384	if (new_se == NULL)
1385		return -ENOMEM;
1386	new_se->start_page = start_page;
1387	new_se->nr_pages = nr_pages;
1388	new_se->start_block = start_block;
1389
1390	list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1391	return 1;
1392}
1393
1394/*
1395 * A `swap extent' is a simple thing which maps a contiguous range of pages
1396 * onto a contiguous range of disk blocks.  An ordered list of swap extents
1397 * is built at swapon time and is then used at swap_writepage/swap_readpage
1398 * time for locating where on disk a page belongs.
1399 *
1400 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1401 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1402 * swap files identically.
1403 *
1404 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1405 * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1406 * swapfiles are handled *identically* after swapon time.
1407 *
1408 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1409 * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1410 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1411 * requirements, they are simply tossed out - we will never use those blocks
1412 * for swapping.
1413 *
1414 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1415 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1416 * which will scribble on the fs.
1417 *
1418 * The amount of disk space which a single swap extent represents varies.
1419 * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1420 * extents in the list.  To avoid much list walking, we cache the previous
1421 * search location in `curr_swap_extent', and start new searches from there.
1422 * This is extremely effective.  The average number of iterations in
1423 * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1424 */
1425static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1426{
1427	struct inode *inode;
1428	unsigned blocks_per_page;
1429	unsigned long page_no;
1430	unsigned blkbits;
1431	sector_t probe_block;
1432	sector_t last_block;
1433	sector_t lowest_block = -1;
1434	sector_t highest_block = 0;
1435	int nr_extents = 0;
1436	int ret;
1437
1438	inode = sis->swap_file->f_mapping->host;
1439	if (S_ISBLK(inode->i_mode)) {
1440		ret = add_swap_extent(sis, 0, sis->max, 0);
1441		*span = sis->pages;
1442		goto out;
1443	}
1444
1445	blkbits = inode->i_blkbits;
1446	blocks_per_page = PAGE_SIZE >> blkbits;
1447
1448	/*
1449	 * Map all the blocks into the extent list.  This code doesn't try
1450	 * to be very smart.
1451	 */
1452	probe_block = 0;
1453	page_no = 0;
1454	last_block = i_size_read(inode) >> blkbits;
1455	while ((probe_block + blocks_per_page) <= last_block &&
1456			page_no < sis->max) {
1457		unsigned block_in_page;
1458		sector_t first_block;
1459
1460		first_block = bmap(inode, probe_block);
1461		if (first_block == 0)
1462			goto bad_bmap;
1463
1464		/*
1465		 * It must be PAGE_SIZE aligned on-disk
1466		 */
1467		if (first_block & (blocks_per_page - 1)) {
1468			probe_block++;
1469			goto reprobe;
1470		}
 
 
1471
1472		for (block_in_page = 1; block_in_page < blocks_per_page;
1473					block_in_page++) {
1474			sector_t block;
1475
1476			block = bmap(inode, probe_block + block_in_page);
1477			if (block == 0)
1478				goto bad_bmap;
1479			if (block != first_block + block_in_page) {
1480				/* Discontiguity */
1481				probe_block++;
1482				goto reprobe;
1483			}
1484		}
1485
1486		first_block >>= (PAGE_SHIFT - blkbits);
1487		if (page_no) {	/* exclude the header page */
1488			if (first_block < lowest_block)
1489				lowest_block = first_block;
1490			if (first_block > highest_block)
1491				highest_block = first_block;
1492		}
1493
1494		/*
1495		 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1496		 */
1497		ret = add_swap_extent(sis, page_no, 1, first_block);
1498		if (ret < 0)
1499			goto out;
1500		nr_extents += ret;
1501		page_no++;
1502		probe_block += blocks_per_page;
1503reprobe:
1504		continue;
1505	}
1506	ret = nr_extents;
1507	*span = 1 + highest_block - lowest_block;
1508	if (page_no == 0)
1509		page_no = 1;	/* force Empty message */
1510	sis->max = page_no;
1511	sis->pages = page_no - 1;
1512	sis->highest_bit = page_no - 1;
1513out:
1514	return ret;
1515bad_bmap:
1516	printk(KERN_ERR "swapon: swapfile has holes\n");
1517	ret = -EINVAL;
1518	goto out;
1519}
1520
1521static void enable_swap_info(struct swap_info_struct *p, int prio,
1522				unsigned char *swap_map)
 
1523{
1524	int i, prev;
1525
1526	spin_lock(&swap_lock);
1527	if (prio >= 0)
1528		p->prio = prio;
1529	else
1530		p->prio = --least_priority;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1531	p->swap_map = swap_map;
 
1532	p->flags |= SWP_WRITEOK;
1533	nr_swap_pages += p->pages;
1534	total_swap_pages += p->pages;
1535
1536	/* insert swap space into swap_list: */
1537	prev = -1;
1538	for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1539		if (p->prio >= swap_info[i]->prio)
1540			break;
1541		prev = i;
1542	}
1543	p->next = i;
1544	if (prev < 0)
1545		swap_list.head = swap_list.next = p->type;
1546	else
1547		swap_info[prev]->next = p->type;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1548	spin_unlock(&swap_lock);
 
1549}
1550
1551SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1552{
1553	struct swap_info_struct *p = NULL;
1554	unsigned char *swap_map;
 
 
1555	struct file *swap_file, *victim;
1556	struct address_space *mapping;
1557	struct inode *inode;
1558	char *pathname;
1559	int oom_score_adj;
1560	int i, type, prev;
1561	int err;
1562
1563	if (!capable(CAP_SYS_ADMIN))
1564		return -EPERM;
1565
 
 
1566	pathname = getname(specialfile);
1567	err = PTR_ERR(pathname);
1568	if (IS_ERR(pathname))
1569		goto out;
1570
1571	victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1572	putname(pathname);
1573	err = PTR_ERR(victim);
1574	if (IS_ERR(victim))
1575		goto out;
1576
1577	mapping = victim->f_mapping;
1578	prev = -1;
1579	spin_lock(&swap_lock);
1580	for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1581		p = swap_info[type];
1582		if (p->flags & SWP_WRITEOK) {
1583			if (p->swap_file->f_mapping == mapping)
 
1584				break;
 
1585		}
1586		prev = type;
1587	}
1588	if (type < 0) {
1589		err = -EINVAL;
1590		spin_unlock(&swap_lock);
1591		goto out_dput;
1592	}
1593	if (!security_vm_enough_memory(p->pages))
1594		vm_unacct_memory(p->pages);
1595	else {
1596		err = -ENOMEM;
1597		spin_unlock(&swap_lock);
1598		goto out_dput;
1599	}
1600	if (prev < 0)
1601		swap_list.head = p->next;
1602	else
1603		swap_info[prev]->next = p->next;
1604	if (type == swap_list.next) {
1605		/* just pick something that's safe... */
1606		swap_list.next = swap_list.head;
1607	}
1608	if (p->prio < 0) {
1609		for (i = p->next; i >= 0; i = swap_info[i]->next)
1610			swap_info[i]->prio = p->prio--;
 
 
 
 
 
 
 
 
 
1611		least_priority++;
1612	}
1613	nr_swap_pages -= p->pages;
 
1614	total_swap_pages -= p->pages;
1615	p->flags &= ~SWP_WRITEOK;
 
1616	spin_unlock(&swap_lock);
1617
1618	oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1619	err = try_to_unuse(type);
1620	test_set_oom_score_adj(oom_score_adj);
 
 
1621
1622	if (err) {
1623		/*
1624		 * reading p->prio and p->swap_map outside the lock is
1625		 * safe here because only sys_swapon and sys_swapoff
1626		 * change them, and there can be no other sys_swapon or
1627		 * sys_swapoff for this swap_info_struct at this point.
1628		 */
1629		/* re-insert swap space back into swap_list */
1630		enable_swap_info(p, p->prio, p->swap_map);
 
1631		goto out_dput;
1632	}
1633
 
 
 
 
1634	destroy_swap_extents(p);
1635	if (p->flags & SWP_CONTINUED)
1636		free_swap_count_continuations(p);
1637
 
 
 
1638	mutex_lock(&swapon_mutex);
1639	spin_lock(&swap_lock);
 
1640	drain_mmlist();
1641
1642	/* wait for anyone still in scan_swap_map */
1643	p->highest_bit = 0;		/* cuts scans short */
1644	while (p->flags >= SWP_SCANNING) {
 
1645		spin_unlock(&swap_lock);
1646		schedule_timeout_uninterruptible(1);
1647		spin_lock(&swap_lock);
 
1648	}
1649
1650	swap_file = p->swap_file;
 
1651	p->swap_file = NULL;
1652	p->max = 0;
1653	swap_map = p->swap_map;
1654	p->swap_map = NULL;
1655	p->flags = 0;
 
 
 
1656	spin_unlock(&swap_lock);
 
 
1657	mutex_unlock(&swapon_mutex);
 
 
1658	vfree(swap_map);
1659	/* Destroy swap account informatin */
1660	swap_cgroup_swapoff(type);
 
 
 
1661
1662	inode = mapping->host;
1663	if (S_ISBLK(inode->i_mode)) {
1664		struct block_device *bdev = I_BDEV(inode);
1665		set_blocksize(bdev, p->old_block_size);
1666		blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1667	} else {
1668		mutex_lock(&inode->i_mutex);
1669		inode->i_flags &= ~S_SWAPFILE;
1670		mutex_unlock(&inode->i_mutex);
1671	}
1672	filp_close(swap_file, NULL);
 
 
 
 
 
 
 
 
 
 
1673	err = 0;
1674	atomic_inc(&proc_poll_event);
1675	wake_up_interruptible(&proc_poll_wait);
1676
1677out_dput:
1678	filp_close(victim, NULL);
1679out:
 
1680	return err;
1681}
1682
1683#ifdef CONFIG_PROC_FS
1684static unsigned swaps_poll(struct file *file, poll_table *wait)
1685{
1686	struct seq_file *seq = file->private_data;
1687
1688	poll_wait(file, &proc_poll_wait, wait);
1689
1690	if (seq->poll_event != atomic_read(&proc_poll_event)) {
1691		seq->poll_event = atomic_read(&proc_poll_event);
1692		return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1693	}
1694
1695	return POLLIN | POLLRDNORM;
1696}
1697
1698/* iterator */
1699static void *swap_start(struct seq_file *swap, loff_t *pos)
1700{
1701	struct swap_info_struct *si;
1702	int type;
1703	loff_t l = *pos;
1704
1705	mutex_lock(&swapon_mutex);
1706
1707	if (!l)
1708		return SEQ_START_TOKEN;
1709
1710	for (type = 0; type < nr_swapfiles; type++) {
1711		smp_rmb();	/* read nr_swapfiles before swap_info[type] */
1712		si = swap_info[type];
1713		if (!(si->flags & SWP_USED) || !si->swap_map)
1714			continue;
1715		if (!--l)
1716			return si;
1717	}
1718
1719	return NULL;
1720}
1721
1722static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1723{
1724	struct swap_info_struct *si = v;
1725	int type;
1726
1727	if (v == SEQ_START_TOKEN)
1728		type = 0;
1729	else
1730		type = si->type + 1;
1731
1732	for (; type < nr_swapfiles; type++) {
1733		smp_rmb();	/* read nr_swapfiles before swap_info[type] */
1734		si = swap_info[type];
1735		if (!(si->flags & SWP_USED) || !si->swap_map)
1736			continue;
1737		++*pos;
1738		return si;
1739	}
1740
1741	return NULL;
1742}
1743
1744static void swap_stop(struct seq_file *swap, void *v)
1745{
1746	mutex_unlock(&swapon_mutex);
1747}
1748
1749static int swap_show(struct seq_file *swap, void *v)
1750{
1751	struct swap_info_struct *si = v;
1752	struct file *file;
1753	int len;
1754
1755	if (si == SEQ_START_TOKEN) {
1756		seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1757		return 0;
1758	}
1759
1760	file = si->swap_file;
1761	len = seq_path(swap, &file->f_path, " \t\n\\");
1762	seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1763			len < 40 ? 40 - len : 1, " ",
1764			S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1765				"partition" : "file\t",
1766			si->pages << (PAGE_SHIFT - 10),
1767			si->inuse_pages << (PAGE_SHIFT - 10),
1768			si->prio);
1769	return 0;
1770}
1771
1772static const struct seq_operations swaps_op = {
1773	.start =	swap_start,
1774	.next =		swap_next,
1775	.stop =		swap_stop,
1776	.show =		swap_show
1777};
1778
1779static int swaps_open(struct inode *inode, struct file *file)
1780{
1781	struct seq_file *seq;
1782	int ret;
1783
1784	ret = seq_open(file, &swaps_op);
1785	if (ret)
1786		return ret;
1787
1788	seq = file->private_data;
1789	seq->poll_event = atomic_read(&proc_poll_event);
1790	return 0;
1791}
1792
1793static const struct file_operations proc_swaps_operations = {
1794	.open		= swaps_open,
1795	.read		= seq_read,
1796	.llseek		= seq_lseek,
1797	.release	= seq_release,
1798	.poll		= swaps_poll,
1799};
1800
1801static int __init procswaps_init(void)
1802{
1803	proc_create("swaps", 0, NULL, &proc_swaps_operations);
1804	return 0;
1805}
1806__initcall(procswaps_init);
1807#endif /* CONFIG_PROC_FS */
1808
1809#ifdef MAX_SWAPFILES_CHECK
1810static int __init max_swapfiles_check(void)
1811{
1812	MAX_SWAPFILES_CHECK();
1813	return 0;
1814}
1815late_initcall(max_swapfiles_check);
1816#endif
1817
1818static struct swap_info_struct *alloc_swap_info(void)
1819{
1820	struct swap_info_struct *p;
1821	unsigned int type;
 
1822
1823	p = kzalloc(sizeof(*p), GFP_KERNEL);
1824	if (!p)
1825		return ERR_PTR(-ENOMEM);
1826
1827	spin_lock(&swap_lock);
1828	for (type = 0; type < nr_swapfiles; type++) {
1829		if (!(swap_info[type]->flags & SWP_USED))
1830			break;
1831	}
1832	if (type >= MAX_SWAPFILES) {
1833		spin_unlock(&swap_lock);
1834		kfree(p);
1835		return ERR_PTR(-EPERM);
1836	}
1837	if (type >= nr_swapfiles) {
1838		p->type = type;
1839		swap_info[type] = p;
1840		/*
1841		 * Write swap_info[type] before nr_swapfiles, in case a
1842		 * racing procfs swap_start() or swap_next() is reading them.
1843		 * (We never shrink nr_swapfiles, we never free this entry.)
1844		 */
1845		smp_wmb();
1846		nr_swapfiles++;
1847	} else {
1848		kfree(p);
1849		p = swap_info[type];
1850		/*
1851		 * Do not memset this entry: a racing procfs swap_next()
1852		 * would be relying on p->type to remain valid.
1853		 */
1854	}
1855	INIT_LIST_HEAD(&p->first_swap_extent.list);
 
 
 
1856	p->flags = SWP_USED;
1857	p->next = -1;
1858	spin_unlock(&swap_lock);
 
 
1859
1860	return p;
1861}
1862
1863static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1864{
1865	int error;
1866
1867	if (S_ISBLK(inode->i_mode)) {
1868		p->bdev = bdgrab(I_BDEV(inode));
1869		error = blkdev_get(p->bdev,
1870				   FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1871				   sys_swapon);
1872		if (error < 0) {
1873			p->bdev = NULL;
1874			return -EINVAL;
1875		}
1876		p->old_block_size = block_size(p->bdev);
1877		error = set_blocksize(p->bdev, PAGE_SIZE);
1878		if (error < 0)
1879			return error;
1880		p->flags |= SWP_BLKDEV;
1881	} else if (S_ISREG(inode->i_mode)) {
1882		p->bdev = inode->i_sb->s_bdev;
1883		mutex_lock(&inode->i_mutex);
1884		if (IS_SWAPFILE(inode))
1885			return -EBUSY;
1886	} else
1887		return -EINVAL;
1888
1889	return 0;
1890}
1891
1892static unsigned long read_swap_header(struct swap_info_struct *p,
1893					union swap_header *swap_header,
1894					struct inode *inode)
1895{
1896	int i;
1897	unsigned long maxpages;
1898	unsigned long swapfilepages;
 
1899
1900	if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1901		printk(KERN_ERR "Unable to find swap-space signature\n");
1902		return 0;
1903	}
1904
1905	/* swap partition endianess hack... */
1906	if (swab32(swap_header->info.version) == 1) {
1907		swab32s(&swap_header->info.version);
1908		swab32s(&swap_header->info.last_page);
1909		swab32s(&swap_header->info.nr_badpages);
 
 
1910		for (i = 0; i < swap_header->info.nr_badpages; i++)
1911			swab32s(&swap_header->info.badpages[i]);
1912	}
1913	/* Check the swap header's sub-version */
1914	if (swap_header->info.version != 1) {
1915		printk(KERN_WARNING
1916		       "Unable to handle swap header version %d\n",
1917		       swap_header->info.version);
1918		return 0;
1919	}
1920
1921	p->lowest_bit  = 1;
1922	p->cluster_next = 1;
1923	p->cluster_nr = 0;
1924
1925	/*
1926	 * Find out how many pages are allowed for a single swap
1927	 * device. There are three limiting factors: 1) the number
1928	 * of bits for the swap offset in the swp_entry_t type, and
1929	 * 2) the number of bits in the swap pte as defined by the
1930	 * the different architectures, and 3) the number of free bits
1931	 * in an exceptional radix_tree entry. In order to find the
1932	 * largest possible bit mask, a swap entry with swap type 0
1933	 * and swap offset ~0UL is created, encoded to a swap pte,
1934	 * decoded to a swp_entry_t again, and finally the swap
1935	 * offset is extracted. This will mask all the bits from
1936	 * the initial ~0UL mask that can't be encoded in either
1937	 * the swp_entry_t or the architecture definition of a
1938	 * swap pte.  Then the same is done for a radix_tree entry.
1939	 */
1940	maxpages = swp_offset(pte_to_swp_entry(
1941			swp_entry_to_pte(swp_entry(0, ~0UL))));
1942	maxpages = swp_offset(radix_to_swp_entry(
1943			swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1944
1945	if (maxpages > swap_header->info.last_page) {
1946		maxpages = swap_header->info.last_page + 1;
 
 
 
 
 
 
 
1947		/* p->max is an unsigned int: don't overflow it */
1948		if ((unsigned int)maxpages == 0)
1949			maxpages = UINT_MAX;
1950	}
1951	p->highest_bit = maxpages - 1;
1952
1953	if (!maxpages)
1954		return 0;
1955	swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1956	if (swapfilepages && maxpages > swapfilepages) {
1957		printk(KERN_WARNING
1958		       "Swap area shorter than signature indicates\n");
1959		return 0;
1960	}
1961	if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1962		return 0;
1963	if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1964		return 0;
1965
1966	return maxpages;
1967}
1968
 
 
 
 
 
 
 
1969static int setup_swap_map_and_extents(struct swap_info_struct *p,
1970					union swap_header *swap_header,
1971					unsigned char *swap_map,
 
1972					unsigned long maxpages,
1973					sector_t *span)
1974{
1975	int i;
1976	unsigned int nr_good_pages;
1977	int nr_extents;
 
 
 
1978
1979	nr_good_pages = maxpages - 1;	/* omit header page */
1980
 
 
 
1981	for (i = 0; i < swap_header->info.nr_badpages; i++) {
1982		unsigned int page_nr = swap_header->info.badpages[i];
1983		if (page_nr == 0 || page_nr > swap_header->info.last_page)
1984			return -EINVAL;
1985		if (page_nr < maxpages) {
1986			swap_map[page_nr] = SWAP_MAP_BAD;
1987			nr_good_pages--;
 
 
 
 
 
1988		}
1989	}
1990
 
 
 
 
1991	if (nr_good_pages) {
1992		swap_map[0] = SWAP_MAP_BAD;
 
 
 
 
 
1993		p->max = maxpages;
1994		p->pages = nr_good_pages;
1995		nr_extents = setup_swap_extents(p, span);
1996		if (nr_extents < 0)
1997			return nr_extents;
1998		nr_good_pages = p->pages;
1999	}
2000	if (!nr_good_pages) {
2001		printk(KERN_WARNING "Empty swap-file\n");
2002		return -EINVAL;
2003	}
2004
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2005	return nr_extents;
2006}
2007
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2008SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2009{
2010	struct swap_info_struct *p;
2011	char *name;
2012	struct file *swap_file = NULL;
2013	struct address_space *mapping;
2014	int i;
2015	int prio;
2016	int error;
2017	union swap_header *swap_header;
2018	int nr_extents;
2019	sector_t span;
2020	unsigned long maxpages;
2021	unsigned char *swap_map = NULL;
 
 
2022	struct page *page = NULL;
2023	struct inode *inode = NULL;
 
 
 
 
2024
2025	if (!capable(CAP_SYS_ADMIN))
2026		return -EPERM;
2027
 
 
 
2028	p = alloc_swap_info();
2029	if (IS_ERR(p))
2030		return PTR_ERR(p);
2031
 
 
2032	name = getname(specialfile);
2033	if (IS_ERR(name)) {
2034		error = PTR_ERR(name);
2035		name = NULL;
2036		goto bad_swap;
2037	}
2038	swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2039	if (IS_ERR(swap_file)) {
2040		error = PTR_ERR(swap_file);
2041		swap_file = NULL;
2042		goto bad_swap;
2043	}
2044
2045	p->swap_file = swap_file;
2046	mapping = swap_file->f_mapping;
2047
2048	for (i = 0; i < nr_swapfiles; i++) {
2049		struct swap_info_struct *q = swap_info[i];
2050
2051		if (q == p || !q->swap_file)
2052			continue;
2053		if (mapping == q->swap_file->f_mapping) {
2054			error = -EBUSY;
2055			goto bad_swap;
2056		}
2057	}
2058
2059	inode = mapping->host;
2060	/* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
 
2061	error = claim_swapfile(p, inode);
2062	if (unlikely(error))
2063		goto bad_swap;
2064
2065	/*
2066	 * Read the swap header.
2067	 */
2068	if (!mapping->a_ops->readpage) {
2069		error = -EINVAL;
2070		goto bad_swap;
2071	}
2072	page = read_mapping_page(mapping, 0, swap_file);
2073	if (IS_ERR(page)) {
2074		error = PTR_ERR(page);
2075		goto bad_swap;
2076	}
2077	swap_header = kmap(page);
2078
2079	maxpages = read_swap_header(p, swap_header, inode);
2080	if (unlikely(!maxpages)) {
2081		error = -EINVAL;
2082		goto bad_swap;
2083	}
2084
2085	/* OK, set up the swap map and apply the bad block list */
2086	swap_map = vzalloc(maxpages);
2087	if (!swap_map) {
2088		error = -ENOMEM;
2089		goto bad_swap;
2090	}
2091
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2092	error = swap_cgroup_swapon(p->type, maxpages);
2093	if (error)
2094		goto bad_swap;
2095
2096	nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2097		maxpages, &span);
2098	if (unlikely(nr_extents < 0)) {
2099		error = nr_extents;
2100		goto bad_swap;
2101	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2102
2103	if (p->bdev) {
2104		if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2105			p->flags |= SWP_SOLIDSTATE;
2106			p->cluster_next = 1 + (random32() % p->highest_bit);
 
 
2107		}
2108		if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2109			p->flags |= SWP_DISCARDABLE;
2110	}
2111
 
 
 
 
2112	mutex_lock(&swapon_mutex);
2113	prio = -1;
2114	if (swap_flags & SWAP_FLAG_PREFER)
2115		prio =
2116		  (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2117	enable_swap_info(p, prio, swap_map);
2118
2119	printk(KERN_INFO "Adding %uk swap on %s.  "
2120			"Priority:%d extents:%d across:%lluk %s%s\n",
2121		p->pages<<(PAGE_SHIFT-10), name, p->prio,
2122		nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123		(p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124		(p->flags & SWP_DISCARDABLE) ? "D" : "");
 
 
 
2125
2126	mutex_unlock(&swapon_mutex);
2127	atomic_inc(&proc_poll_event);
2128	wake_up_interruptible(&proc_poll_wait);
2129
2130	if (S_ISREG(inode->i_mode))
2131		inode->i_flags |= S_SWAPFILE;
2132	error = 0;
2133	goto out;
2134bad_swap:
 
 
2135	if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2136		set_blocksize(p->bdev, p->old_block_size);
2137		blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2138	}
2139	destroy_swap_extents(p);
2140	swap_cgroup_swapoff(p->type);
2141	spin_lock(&swap_lock);
2142	p->swap_file = NULL;
2143	p->flags = 0;
2144	spin_unlock(&swap_lock);
2145	vfree(swap_map);
 
 
 
 
2146	if (swap_file) {
2147		if (inode && S_ISREG(inode->i_mode)) {
2148			mutex_unlock(&inode->i_mutex);
2149			inode = NULL;
2150		}
2151		filp_close(swap_file, NULL);
2152	}
2153out:
2154	if (page && !IS_ERR(page)) {
2155		kunmap(page);
2156		page_cache_release(page);
2157	}
2158	if (name)
2159		putname(name);
2160	if (inode && S_ISREG(inode->i_mode))
2161		mutex_unlock(&inode->i_mutex);
 
 
2162	return error;
2163}
2164
2165void si_swapinfo(struct sysinfo *val)
2166{
2167	unsigned int type;
2168	unsigned long nr_to_be_unused = 0;
2169
2170	spin_lock(&swap_lock);
2171	for (type = 0; type < nr_swapfiles; type++) {
2172		struct swap_info_struct *si = swap_info[type];
2173
2174		if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2175			nr_to_be_unused += si->inuse_pages;
2176	}
2177	val->freeswap = nr_swap_pages + nr_to_be_unused;
2178	val->totalswap = total_swap_pages + nr_to_be_unused;
2179	spin_unlock(&swap_lock);
2180}
2181
2182/*
2183 * Verify that a swap entry is valid and increment its swap map count.
2184 *
2185 * Returns error code in following case.
2186 * - success -> 0
2187 * - swp_entry is invalid -> EINVAL
2188 * - swp_entry is migration entry -> EINVAL
2189 * - swap-cache reference is requested but there is already one. -> EEXIST
2190 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2191 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2192 */
2193static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2194{
2195	struct swap_info_struct *p;
 
2196	unsigned long offset, type;
2197	unsigned char count;
2198	unsigned char has_cache;
2199	int err = -EINVAL;
2200
2201	if (non_swap_entry(entry))
2202		goto out;
2203
2204	type = swp_type(entry);
2205	if (type >= nr_swapfiles)
2206		goto bad_file;
2207	p = swap_info[type];
2208	offset = swp_offset(entry);
2209
2210	spin_lock(&swap_lock);
2211	if (unlikely(offset >= p->max))
2212		goto unlock_out;
 
 
2213
2214	count = p->swap_map[offset];
 
 
 
 
 
 
 
 
 
 
2215	has_cache = count & SWAP_HAS_CACHE;
2216	count &= ~SWAP_HAS_CACHE;
2217	err = 0;
2218
2219	if (usage == SWAP_HAS_CACHE) {
2220
2221		/* set SWAP_HAS_CACHE if there is no cache and entry is used */
2222		if (!has_cache && count)
2223			has_cache = SWAP_HAS_CACHE;
2224		else if (has_cache)		/* someone else added cache */
2225			err = -EEXIST;
2226		else				/* no users remaining */
2227			err = -ENOENT;
2228
2229	} else if (count || has_cache) {
2230
2231		if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2232			count += usage;
2233		else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2234			err = -EINVAL;
2235		else if (swap_count_continued(p, offset, count))
2236			count = COUNT_CONTINUED;
2237		else
2238			err = -ENOMEM;
2239	} else
2240		err = -ENOENT;			/* unused swap entry */
2241
2242	p->swap_map[offset] = count | has_cache;
2243
2244unlock_out:
2245	spin_unlock(&swap_lock);
2246out:
2247	return err;
2248
2249bad_file:
2250	printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2251	goto out;
2252}
2253
2254/*
2255 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2256 * (in which case its reference count is never incremented).
2257 */
2258void swap_shmem_alloc(swp_entry_t entry)
2259{
2260	__swap_duplicate(entry, SWAP_MAP_SHMEM);
2261}
2262
2263/*
2264 * Increase reference count of swap entry by 1.
2265 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2266 * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2267 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2268 * might occur if a page table entry has got corrupted.
2269 */
2270int swap_duplicate(swp_entry_t entry)
2271{
2272	int err = 0;
2273
2274	while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2275		err = add_swap_count_continuation(entry, GFP_ATOMIC);
2276	return err;
2277}
2278
2279/*
2280 * @entry: swap entry for which we allocate swap cache.
2281 *
2282 * Called when allocating swap cache for existing swap entry,
2283 * This can return error codes. Returns 0 at success.
2284 * -EBUSY means there is a swap cache.
2285 * Note: return code is different from swap_duplicate().
2286 */
2287int swapcache_prepare(swp_entry_t entry)
2288{
2289	return __swap_duplicate(entry, SWAP_HAS_CACHE);
2290}
2291
2292/*
2293 * swap_lock prevents swap_map being freed. Don't grab an extra
2294 * reference on the swaphandle, it doesn't matter if it becomes unused.
2295 */
2296int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2297{
2298	struct swap_info_struct *si;
2299	int our_page_cluster = page_cluster;
2300	pgoff_t target, toff;
2301	pgoff_t base, end;
2302	int nr_pages = 0;
2303
2304	if (!our_page_cluster)	/* no readahead */
2305		return 0;
2306
2307	si = swap_info[swp_type(entry)];
2308	target = swp_offset(entry);
2309	base = (target >> our_page_cluster) << our_page_cluster;
2310	end = base + (1 << our_page_cluster);
2311	if (!base)		/* first page is swap header */
2312		base++;
2313
2314	spin_lock(&swap_lock);
2315	if (end > si->max)	/* don't go beyond end of map */
2316		end = si->max;
 
 
2317
2318	/* Count contiguous allocated slots above our target */
2319	for (toff = target; ++toff < end; nr_pages++) {
2320		/* Don't read in free or bad pages */
2321		if (!si->swap_map[toff])
2322			break;
2323		if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2324			break;
2325	}
2326	/* Count contiguous allocated slots below our target */
2327	for (toff = target; --toff >= base; nr_pages++) {
2328		/* Don't read in free or bad pages */
2329		if (!si->swap_map[toff])
2330			break;
2331		if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2332			break;
2333	}
2334	spin_unlock(&swap_lock);
2335
2336	/*
2337	 * Indicate starting offset, and return number of pages to get:
2338	 * if only 1, say 0, since there's then no readahead to be done.
2339	 */
2340	*offset = ++toff;
2341	return nr_pages? ++nr_pages: 0;
2342}
 
2343
2344/*
2345 * add_swap_count_continuation - called when a swap count is duplicated
2346 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2347 * page of the original vmalloc'ed swap_map, to hold the continuation count
2348 * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2349 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2350 *
2351 * These continuation pages are seldom referenced: the common paths all work
2352 * on the original swap_map, only referring to a continuation page when the
2353 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2354 *
2355 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2356 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2357 * can be called after dropping locks.
2358 */
2359int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2360{
2361	struct swap_info_struct *si;
 
2362	struct page *head;
2363	struct page *page;
2364	struct page *list_page;
2365	pgoff_t offset;
2366	unsigned char count;
2367
2368	/*
2369	 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2370	 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2371	 */
2372	page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2373
2374	si = swap_info_get(entry);
2375	if (!si) {
2376		/*
2377		 * An acceptable race has occurred since the failing
2378		 * __swap_duplicate(): the swap entry has been freed,
2379		 * perhaps even the whole swap_map cleared for swapoff.
2380		 */
2381		goto outer;
2382	}
2383
2384	offset = swp_offset(entry);
 
 
 
2385	count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2386
2387	if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2388		/*
2389		 * The higher the swap count, the more likely it is that tasks
2390		 * will race to add swap count continuation: we need to avoid
2391		 * over-provisioning.
2392		 */
2393		goto out;
2394	}
2395
2396	if (!page) {
2397		spin_unlock(&swap_lock);
 
2398		return -ENOMEM;
2399	}
2400
2401	/*
2402	 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2403	 * no architecture is using highmem pages for kernel pagetables: so it
2404	 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2405	 */
2406	head = vmalloc_to_page(si->swap_map + offset);
2407	offset &= ~PAGE_MASK;
2408
 
2409	/*
2410	 * Page allocation does not initialize the page's lru field,
2411	 * but it does always reset its private field.
2412	 */
2413	if (!page_private(head)) {
2414		BUG_ON(count & COUNT_CONTINUED);
2415		INIT_LIST_HEAD(&head->lru);
2416		set_page_private(head, SWP_CONTINUED);
2417		si->flags |= SWP_CONTINUED;
2418	}
2419
2420	list_for_each_entry(list_page, &head->lru, lru) {
2421		unsigned char *map;
2422
2423		/*
2424		 * If the previous map said no continuation, but we've found
2425		 * a continuation page, free our allocation and use this one.
2426		 */
2427		if (!(count & COUNT_CONTINUED))
2428			goto out;
2429
2430		map = kmap_atomic(list_page, KM_USER0) + offset;
2431		count = *map;
2432		kunmap_atomic(map, KM_USER0);
2433
2434		/*
2435		 * If this continuation count now has some space in it,
2436		 * free our allocation and use this one.
2437		 */
2438		if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2439			goto out;
2440	}
2441
2442	list_add_tail(&page->lru, &head->lru);
2443	page = NULL;			/* now it's attached, don't free it */
 
 
2444out:
2445	spin_unlock(&swap_lock);
 
2446outer:
2447	if (page)
2448		__free_page(page);
2449	return 0;
2450}
2451
2452/*
2453 * swap_count_continued - when the original swap_map count is incremented
2454 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2455 * into, carry if so, or else fail until a new continuation page is allocated;
2456 * when the original swap_map count is decremented from 0 with continuation,
2457 * borrow from the continuation and report whether it still holds more.
2458 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
 
2459 */
2460static bool swap_count_continued(struct swap_info_struct *si,
2461				 pgoff_t offset, unsigned char count)
2462{
2463	struct page *head;
2464	struct page *page;
2465	unsigned char *map;
 
2466
2467	head = vmalloc_to_page(si->swap_map + offset);
2468	if (page_private(head) != SWP_CONTINUED) {
2469		BUG_ON(count & COUNT_CONTINUED);
2470		return false;		/* need to add count continuation */
2471	}
2472
 
2473	offset &= ~PAGE_MASK;
2474	page = list_entry(head->lru.next, struct page, lru);
2475	map = kmap_atomic(page, KM_USER0) + offset;
2476
2477	if (count == SWAP_MAP_MAX)	/* initial increment from swap_map */
2478		goto init_map;		/* jump over SWAP_CONT_MAX checks */
2479
2480	if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2481		/*
2482		 * Think of how you add 1 to 999
2483		 */
2484		while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2485			kunmap_atomic(map, KM_USER0);
2486			page = list_entry(page->lru.next, struct page, lru);
2487			BUG_ON(page == head);
2488			map = kmap_atomic(page, KM_USER0) + offset;
2489		}
2490		if (*map == SWAP_CONT_MAX) {
2491			kunmap_atomic(map, KM_USER0);
2492			page = list_entry(page->lru.next, struct page, lru);
2493			if (page == head)
2494				return false;	/* add count continuation */
2495			map = kmap_atomic(page, KM_USER0) + offset;
 
 
2496init_map:		*map = 0;		/* we didn't zero the page */
2497		}
2498		*map += 1;
2499		kunmap_atomic(map, KM_USER0);
2500		page = list_entry(page->lru.prev, struct page, lru);
2501		while (page != head) {
2502			map = kmap_atomic(page, KM_USER0) + offset;
2503			*map = COUNT_CONTINUED;
2504			kunmap_atomic(map, KM_USER0);
2505			page = list_entry(page->lru.prev, struct page, lru);
2506		}
2507		return true;			/* incremented */
2508
2509	} else {				/* decrementing */
2510		/*
2511		 * Think of how you subtract 1 from 1000
2512		 */
2513		BUG_ON(count != COUNT_CONTINUED);
2514		while (*map == COUNT_CONTINUED) {
2515			kunmap_atomic(map, KM_USER0);
2516			page = list_entry(page->lru.next, struct page, lru);
2517			BUG_ON(page == head);
2518			map = kmap_atomic(page, KM_USER0) + offset;
2519		}
2520		BUG_ON(*map == 0);
2521		*map -= 1;
2522		if (*map == 0)
2523			count = 0;
2524		kunmap_atomic(map, KM_USER0);
2525		page = list_entry(page->lru.prev, struct page, lru);
2526		while (page != head) {
2527			map = kmap_atomic(page, KM_USER0) + offset;
2528			*map = SWAP_CONT_MAX | count;
2529			count = COUNT_CONTINUED;
2530			kunmap_atomic(map, KM_USER0);
2531			page = list_entry(page->lru.prev, struct page, lru);
2532		}
2533		return count == COUNT_CONTINUED;
2534	}
 
 
 
2535}
2536
2537/*
2538 * free_swap_count_continuations - swapoff free all the continuation pages
2539 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2540 */
2541static void free_swap_count_continuations(struct swap_info_struct *si)
2542{
2543	pgoff_t offset;
2544
2545	for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2546		struct page *head;
2547		head = vmalloc_to_page(si->swap_map + offset);
2548		if (page_private(head)) {
2549			struct list_head *this, *next;
2550			list_for_each_safe(this, next, &head->lru) {
2551				struct page *page;
2552				page = list_entry(this, struct page, lru);
2553				list_del(this);
2554				__free_page(page);
2555			}
2556		}
2557	}
2558}
v4.17
   1/*
   2 *  linux/mm/swapfile.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 *  Swap reorganised 29.12.95, Stephen Tweedie
   6 */
   7
   8#include <linux/mm.h>
   9#include <linux/sched/mm.h>
  10#include <linux/sched/task.h>
  11#include <linux/hugetlb.h>
  12#include <linux/mman.h>
  13#include <linux/slab.h>
  14#include <linux/kernel_stat.h>
  15#include <linux/swap.h>
  16#include <linux/vmalloc.h>
  17#include <linux/pagemap.h>
  18#include <linux/namei.h>
  19#include <linux/shmem_fs.h>
  20#include <linux/blkdev.h>
  21#include <linux/random.h>
  22#include <linux/writeback.h>
  23#include <linux/proc_fs.h>
  24#include <linux/seq_file.h>
  25#include <linux/init.h>
 
  26#include <linux/ksm.h>
  27#include <linux/rmap.h>
  28#include <linux/security.h>
  29#include <linux/backing-dev.h>
  30#include <linux/mutex.h>
  31#include <linux/capability.h>
  32#include <linux/syscalls.h>
  33#include <linux/memcontrol.h>
  34#include <linux/poll.h>
  35#include <linux/oom.h>
  36#include <linux/frontswap.h>
  37#include <linux/swapfile.h>
  38#include <linux/export.h>
  39#include <linux/swap_slots.h>
  40#include <linux/sort.h>
  41
  42#include <asm/pgtable.h>
  43#include <asm/tlbflush.h>
  44#include <linux/swapops.h>
  45#include <linux/swap_cgroup.h>
  46
  47static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
  48				 unsigned char);
  49static void free_swap_count_continuations(struct swap_info_struct *);
  50static sector_t map_swap_entry(swp_entry_t, struct block_device**);
  51
  52DEFINE_SPINLOCK(swap_lock);
  53static unsigned int nr_swapfiles;
  54atomic_long_t nr_swap_pages;
  55/*
  56 * Some modules use swappable objects and may try to swap them out under
  57 * memory pressure (via the shrinker). Before doing so, they may wish to
  58 * check to see if any swap space is available.
  59 */
  60EXPORT_SYMBOL_GPL(nr_swap_pages);
  61/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
  62long total_swap_pages;
  63static int least_priority = -1;
  64
  65static const char Bad_file[] = "Bad swap file entry ";
  66static const char Unused_file[] = "Unused swap file entry ";
  67static const char Bad_offset[] = "Bad swap offset entry ";
  68static const char Unused_offset[] = "Unused swap offset entry ";
  69
  70/*
  71 * all active swap_info_structs
  72 * protected with swap_lock, and ordered by priority.
  73 */
  74PLIST_HEAD(swap_active_head);
  75
  76/*
  77 * all available (active, not full) swap_info_structs
  78 * protected with swap_avail_lock, ordered by priority.
  79 * This is used by get_swap_page() instead of swap_active_head
  80 * because swap_active_head includes all swap_info_structs,
  81 * but get_swap_page() doesn't need to look at full ones.
  82 * This uses its own lock instead of swap_lock because when a
  83 * swap_info_struct changes between not-full/full, it needs to
  84 * add/remove itself to/from this list, but the swap_info_struct->lock
  85 * is held and the locking order requires swap_lock to be taken
  86 * before any swap_info_struct->lock.
  87 */
  88static struct plist_head *swap_avail_heads;
  89static DEFINE_SPINLOCK(swap_avail_lock);
  90
  91struct swap_info_struct *swap_info[MAX_SWAPFILES];
  92
  93static DEFINE_MUTEX(swapon_mutex);
  94
  95static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
  96/* Activity counter to indicate that a swapon or swapoff has occurred */
  97static atomic_t proc_poll_event = ATOMIC_INIT(0);
  98
  99atomic_t nr_rotate_swap = ATOMIC_INIT(0);
 100
 101static inline unsigned char swap_count(unsigned char ent)
 102{
 103	return ent & ~SWAP_HAS_CACHE;	/* may include SWAP_HAS_CONT flag */
 104}
 105
 106/* returns 1 if swap entry is freed */
 107static int
 108__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
 109{
 110	swp_entry_t entry = swp_entry(si->type, offset);
 111	struct page *page;
 112	int ret = 0;
 113
 114	page = find_get_page(swap_address_space(entry), swp_offset(entry));
 115	if (!page)
 116		return 0;
 117	/*
 118	 * This function is called from scan_swap_map() and it's called
 119	 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
 120	 * We have to use trylock for avoiding deadlock. This is a special
 121	 * case and you should use try_to_free_swap() with explicit lock_page()
 122	 * in usual operations.
 123	 */
 124	if (trylock_page(page)) {
 125		ret = try_to_free_swap(page);
 126		unlock_page(page);
 127	}
 128	put_page(page);
 129	return ret;
 130}
 131
 132/*
 133 * swapon tell device that all the old swap contents can be discarded,
 134 * to allow the swap device to optimize its wear-levelling.
 135 */
 136static int discard_swap(struct swap_info_struct *si)
 137{
 138	struct swap_extent *se;
 139	sector_t start_block;
 140	sector_t nr_blocks;
 141	int err = 0;
 142
 143	/* Do not discard the swap header page! */
 144	se = &si->first_swap_extent;
 145	start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 146	nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 147	if (nr_blocks) {
 148		err = blkdev_issue_discard(si->bdev, start_block,
 149				nr_blocks, GFP_KERNEL, 0);
 150		if (err)
 151			return err;
 152		cond_resched();
 153	}
 154
 155	list_for_each_entry(se, &si->first_swap_extent.list, list) {
 156		start_block = se->start_block << (PAGE_SHIFT - 9);
 157		nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 158
 159		err = blkdev_issue_discard(si->bdev, start_block,
 160				nr_blocks, GFP_KERNEL, 0);
 161		if (err)
 162			break;
 163
 164		cond_resched();
 165	}
 166	return err;		/* That will often be -EOPNOTSUPP */
 167}
 168
 169/*
 170 * swap allocation tell device that a cluster of swap can now be discarded,
 171 * to allow the swap device to optimize its wear-levelling.
 172 */
 173static void discard_swap_cluster(struct swap_info_struct *si,
 174				 pgoff_t start_page, pgoff_t nr_pages)
 175{
 176	struct swap_extent *se = si->curr_swap_extent;
 177	int found_extent = 0;
 178
 179	while (nr_pages) {
 
 
 180		if (se->start_page <= start_page &&
 181		    start_page < se->start_page + se->nr_pages) {
 182			pgoff_t offset = start_page - se->start_page;
 183			sector_t start_block = se->start_block + offset;
 184			sector_t nr_blocks = se->nr_pages - offset;
 185
 186			if (nr_blocks > nr_pages)
 187				nr_blocks = nr_pages;
 188			start_page += nr_blocks;
 189			nr_pages -= nr_blocks;
 190
 191			if (!found_extent++)
 192				si->curr_swap_extent = se;
 193
 194			start_block <<= PAGE_SHIFT - 9;
 195			nr_blocks <<= PAGE_SHIFT - 9;
 196			if (blkdev_issue_discard(si->bdev, start_block,
 197				    nr_blocks, GFP_NOIO, 0))
 198				break;
 199		}
 200
 201		se = list_next_entry(se, list);
 
 202	}
 203}
 204
 205#ifdef CONFIG_THP_SWAP
 206#define SWAPFILE_CLUSTER	HPAGE_PMD_NR
 207#else
 208#define SWAPFILE_CLUSTER	256
 209#endif
 210#define LATENCY_LIMIT		256
 211
 212static inline void cluster_set_flag(struct swap_cluster_info *info,
 213	unsigned int flag)
 214{
 215	info->flags = flag;
 
 216}
 217
 218static inline unsigned int cluster_count(struct swap_cluster_info *info)
 219{
 220	return info->data;
 221}
 222
 223static inline void cluster_set_count(struct swap_cluster_info *info,
 224				     unsigned int c)
 225{
 226	info->data = c;
 227}
 228
 229static inline void cluster_set_count_flag(struct swap_cluster_info *info,
 230					 unsigned int c, unsigned int f)
 231{
 232	info->flags = f;
 233	info->data = c;
 234}
 235
 236static inline unsigned int cluster_next(struct swap_cluster_info *info)
 237{
 238	return info->data;
 239}
 240
 241static inline void cluster_set_next(struct swap_cluster_info *info,
 242				    unsigned int n)
 243{
 244	info->data = n;
 245}
 246
 247static inline void cluster_set_next_flag(struct swap_cluster_info *info,
 248					 unsigned int n, unsigned int f)
 249{
 250	info->flags = f;
 251	info->data = n;
 252}
 253
 254static inline bool cluster_is_free(struct swap_cluster_info *info)
 255{
 256	return info->flags & CLUSTER_FLAG_FREE;
 257}
 258
 259static inline bool cluster_is_null(struct swap_cluster_info *info)
 260{
 261	return info->flags & CLUSTER_FLAG_NEXT_NULL;
 262}
 263
 264static inline void cluster_set_null(struct swap_cluster_info *info)
 265{
 266	info->flags = CLUSTER_FLAG_NEXT_NULL;
 267	info->data = 0;
 268}
 269
 270static inline bool cluster_is_huge(struct swap_cluster_info *info)
 271{
 272	return info->flags & CLUSTER_FLAG_HUGE;
 273}
 274
 275static inline void cluster_clear_huge(struct swap_cluster_info *info)
 276{
 277	info->flags &= ~CLUSTER_FLAG_HUGE;
 278}
 279
 280static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
 281						     unsigned long offset)
 282{
 283	struct swap_cluster_info *ci;
 284
 285	ci = si->cluster_info;
 286	if (ci) {
 287		ci += offset / SWAPFILE_CLUSTER;
 288		spin_lock(&ci->lock);
 289	}
 290	return ci;
 291}
 292
 293static inline void unlock_cluster(struct swap_cluster_info *ci)
 294{
 295	if (ci)
 296		spin_unlock(&ci->lock);
 297}
 298
 299static inline struct swap_cluster_info *lock_cluster_or_swap_info(
 300	struct swap_info_struct *si,
 301	unsigned long offset)
 302{
 303	struct swap_cluster_info *ci;
 304
 305	ci = lock_cluster(si, offset);
 306	if (!ci)
 307		spin_lock(&si->lock);
 308
 309	return ci;
 310}
 311
 312static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
 313					       struct swap_cluster_info *ci)
 314{
 315	if (ci)
 316		unlock_cluster(ci);
 317	else
 318		spin_unlock(&si->lock);
 319}
 320
 321static inline bool cluster_list_empty(struct swap_cluster_list *list)
 322{
 323	return cluster_is_null(&list->head);
 324}
 325
 326static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
 327{
 328	return cluster_next(&list->head);
 329}
 330
 331static void cluster_list_init(struct swap_cluster_list *list)
 332{
 333	cluster_set_null(&list->head);
 334	cluster_set_null(&list->tail);
 335}
 336
 337static void cluster_list_add_tail(struct swap_cluster_list *list,
 338				  struct swap_cluster_info *ci,
 339				  unsigned int idx)
 340{
 341	if (cluster_list_empty(list)) {
 342		cluster_set_next_flag(&list->head, idx, 0);
 343		cluster_set_next_flag(&list->tail, idx, 0);
 344	} else {
 345		struct swap_cluster_info *ci_tail;
 346		unsigned int tail = cluster_next(&list->tail);
 347
 348		/*
 349		 * Nested cluster lock, but both cluster locks are
 350		 * only acquired when we held swap_info_struct->lock
 351		 */
 352		ci_tail = ci + tail;
 353		spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
 354		cluster_set_next(ci_tail, idx);
 355		spin_unlock(&ci_tail->lock);
 356		cluster_set_next_flag(&list->tail, idx, 0);
 357	}
 358}
 359
 360static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
 361					   struct swap_cluster_info *ci)
 362{
 363	unsigned int idx;
 364
 365	idx = cluster_next(&list->head);
 366	if (cluster_next(&list->tail) == idx) {
 367		cluster_set_null(&list->head);
 368		cluster_set_null(&list->tail);
 369	} else
 370		cluster_set_next_flag(&list->head,
 371				      cluster_next(&ci[idx]), 0);
 372
 373	return idx;
 374}
 375
 376/* Add a cluster to discard list and schedule it to do discard */
 377static void swap_cluster_schedule_discard(struct swap_info_struct *si,
 378		unsigned int idx)
 379{
 380	/*
 381	 * If scan_swap_map() can't find a free cluster, it will check
 382	 * si->swap_map directly. To make sure the discarding cluster isn't
 383	 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
 384	 * will be cleared after discard
 385	 */
 386	memset(si->swap_map + idx * SWAPFILE_CLUSTER,
 387			SWAP_MAP_BAD, SWAPFILE_CLUSTER);
 388
 389	cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
 390
 391	schedule_work(&si->discard_work);
 392}
 393
 394static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
 395{
 396	struct swap_cluster_info *ci = si->cluster_info;
 397
 398	cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
 399	cluster_list_add_tail(&si->free_clusters, ci, idx);
 400}
 401
 402/*
 403 * Doing discard actually. After a cluster discard is finished, the cluster
 404 * will be added to free cluster list. caller should hold si->lock.
 405*/
 406static void swap_do_scheduled_discard(struct swap_info_struct *si)
 407{
 408	struct swap_cluster_info *info, *ci;
 409	unsigned int idx;
 410
 411	info = si->cluster_info;
 412
 413	while (!cluster_list_empty(&si->discard_clusters)) {
 414		idx = cluster_list_del_first(&si->discard_clusters, info);
 415		spin_unlock(&si->lock);
 416
 417		discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
 418				SWAPFILE_CLUSTER);
 419
 420		spin_lock(&si->lock);
 421		ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
 422		__free_cluster(si, idx);
 423		memset(si->swap_map + idx * SWAPFILE_CLUSTER,
 424				0, SWAPFILE_CLUSTER);
 425		unlock_cluster(ci);
 426	}
 427}
 428
 429static void swap_discard_work(struct work_struct *work)
 430{
 431	struct swap_info_struct *si;
 432
 433	si = container_of(work, struct swap_info_struct, discard_work);
 434
 435	spin_lock(&si->lock);
 436	swap_do_scheduled_discard(si);
 437	spin_unlock(&si->lock);
 438}
 439
 440static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
 441{
 442	struct swap_cluster_info *ci = si->cluster_info;
 443
 444	VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
 445	cluster_list_del_first(&si->free_clusters, ci);
 446	cluster_set_count_flag(ci + idx, 0, 0);
 447}
 448
 449static void free_cluster(struct swap_info_struct *si, unsigned long idx)
 450{
 451	struct swap_cluster_info *ci = si->cluster_info + idx;
 452
 453	VM_BUG_ON(cluster_count(ci) != 0);
 454	/*
 455	 * If the swap is discardable, prepare discard the cluster
 456	 * instead of free it immediately. The cluster will be freed
 457	 * after discard.
 458	 */
 459	if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
 460	    (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
 461		swap_cluster_schedule_discard(si, idx);
 462		return;
 463	}
 464
 465	__free_cluster(si, idx);
 466}
 467
 468/*
 469 * The cluster corresponding to page_nr will be used. The cluster will be
 470 * removed from free cluster list and its usage counter will be increased.
 471 */
 472static void inc_cluster_info_page(struct swap_info_struct *p,
 473	struct swap_cluster_info *cluster_info, unsigned long page_nr)
 474{
 475	unsigned long idx = page_nr / SWAPFILE_CLUSTER;
 476
 477	if (!cluster_info)
 478		return;
 479	if (cluster_is_free(&cluster_info[idx]))
 480		alloc_cluster(p, idx);
 481
 482	VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
 483	cluster_set_count(&cluster_info[idx],
 484		cluster_count(&cluster_info[idx]) + 1);
 485}
 486
 487/*
 488 * The cluster corresponding to page_nr decreases one usage. If the usage
 489 * counter becomes 0, which means no page in the cluster is in using, we can
 490 * optionally discard the cluster and add it to free cluster list.
 491 */
 492static void dec_cluster_info_page(struct swap_info_struct *p,
 493	struct swap_cluster_info *cluster_info, unsigned long page_nr)
 494{
 495	unsigned long idx = page_nr / SWAPFILE_CLUSTER;
 496
 497	if (!cluster_info)
 498		return;
 499
 500	VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
 501	cluster_set_count(&cluster_info[idx],
 502		cluster_count(&cluster_info[idx]) - 1);
 503
 504	if (cluster_count(&cluster_info[idx]) == 0)
 505		free_cluster(p, idx);
 506}
 507
 508/*
 509 * It's possible scan_swap_map() uses a free cluster in the middle of free
 510 * cluster list. Avoiding such abuse to avoid list corruption.
 511 */
 512static bool
 513scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
 514	unsigned long offset)
 515{
 516	struct percpu_cluster *percpu_cluster;
 517	bool conflict;
 518
 519	offset /= SWAPFILE_CLUSTER;
 520	conflict = !cluster_list_empty(&si->free_clusters) &&
 521		offset != cluster_list_first(&si->free_clusters) &&
 522		cluster_is_free(&si->cluster_info[offset]);
 523
 524	if (!conflict)
 525		return false;
 526
 527	percpu_cluster = this_cpu_ptr(si->percpu_cluster);
 528	cluster_set_null(&percpu_cluster->index);
 529	return true;
 530}
 531
 532/*
 533 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
 534 * might involve allocating a new cluster for current CPU too.
 535 */
 536static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
 537	unsigned long *offset, unsigned long *scan_base)
 538{
 539	struct percpu_cluster *cluster;
 540	struct swap_cluster_info *ci;
 541	bool found_free;
 542	unsigned long tmp, max;
 543
 544new_cluster:
 545	cluster = this_cpu_ptr(si->percpu_cluster);
 546	if (cluster_is_null(&cluster->index)) {
 547		if (!cluster_list_empty(&si->free_clusters)) {
 548			cluster->index = si->free_clusters.head;
 549			cluster->next = cluster_next(&cluster->index) *
 550					SWAPFILE_CLUSTER;
 551		} else if (!cluster_list_empty(&si->discard_clusters)) {
 552			/*
 553			 * we don't have free cluster but have some clusters in
 554			 * discarding, do discard now and reclaim them
 555			 */
 556			swap_do_scheduled_discard(si);
 557			*scan_base = *offset = si->cluster_next;
 558			goto new_cluster;
 559		} else
 560			return false;
 561	}
 562
 563	found_free = false;
 564
 565	/*
 566	 * Other CPUs can use our cluster if they can't find a free cluster,
 567	 * check if there is still free entry in the cluster
 568	 */
 569	tmp = cluster->next;
 570	max = min_t(unsigned long, si->max,
 571		    (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
 572	if (tmp >= max) {
 573		cluster_set_null(&cluster->index);
 574		goto new_cluster;
 575	}
 576	ci = lock_cluster(si, tmp);
 577	while (tmp < max) {
 578		if (!si->swap_map[tmp]) {
 579			found_free = true;
 580			break;
 581		}
 582		tmp++;
 583	}
 584	unlock_cluster(ci);
 585	if (!found_free) {
 586		cluster_set_null(&cluster->index);
 587		goto new_cluster;
 588	}
 589	cluster->next = tmp + 1;
 590	*offset = tmp;
 591	*scan_base = tmp;
 592	return found_free;
 593}
 594
 595static void __del_from_avail_list(struct swap_info_struct *p)
 596{
 597	int nid;
 598
 599	for_each_node(nid)
 600		plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
 601}
 602
 603static void del_from_avail_list(struct swap_info_struct *p)
 604{
 605	spin_lock(&swap_avail_lock);
 606	__del_from_avail_list(p);
 607	spin_unlock(&swap_avail_lock);
 608}
 609
 610static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
 611			     unsigned int nr_entries)
 612{
 613	unsigned int end = offset + nr_entries - 1;
 614
 615	if (offset == si->lowest_bit)
 616		si->lowest_bit += nr_entries;
 617	if (end == si->highest_bit)
 618		si->highest_bit -= nr_entries;
 619	si->inuse_pages += nr_entries;
 620	if (si->inuse_pages == si->pages) {
 621		si->lowest_bit = si->max;
 622		si->highest_bit = 0;
 623		del_from_avail_list(si);
 624	}
 625}
 626
 627static void add_to_avail_list(struct swap_info_struct *p)
 628{
 629	int nid;
 630
 631	spin_lock(&swap_avail_lock);
 632	for_each_node(nid) {
 633		WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
 634		plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
 635	}
 636	spin_unlock(&swap_avail_lock);
 637}
 638
 639static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
 640			    unsigned int nr_entries)
 641{
 642	unsigned long end = offset + nr_entries - 1;
 643	void (*swap_slot_free_notify)(struct block_device *, unsigned long);
 644
 645	if (offset < si->lowest_bit)
 646		si->lowest_bit = offset;
 647	if (end > si->highest_bit) {
 648		bool was_full = !si->highest_bit;
 649
 650		si->highest_bit = end;
 651		if (was_full && (si->flags & SWP_WRITEOK))
 652			add_to_avail_list(si);
 653	}
 654	atomic_long_add(nr_entries, &nr_swap_pages);
 655	si->inuse_pages -= nr_entries;
 656	if (si->flags & SWP_BLKDEV)
 657		swap_slot_free_notify =
 658			si->bdev->bd_disk->fops->swap_slot_free_notify;
 659	else
 660		swap_slot_free_notify = NULL;
 661	while (offset <= end) {
 662		frontswap_invalidate_page(si->type, offset);
 663		if (swap_slot_free_notify)
 664			swap_slot_free_notify(si->bdev, offset);
 665		offset++;
 666	}
 667}
 668
 669static int scan_swap_map_slots(struct swap_info_struct *si,
 670			       unsigned char usage, int nr,
 671			       swp_entry_t slots[])
 672{
 673	struct swap_cluster_info *ci;
 674	unsigned long offset;
 675	unsigned long scan_base;
 676	unsigned long last_in_cluster = 0;
 677	int latency_ration = LATENCY_LIMIT;
 678	int n_ret = 0;
 679
 680	if (nr > SWAP_BATCH)
 681		nr = SWAP_BATCH;
 682
 683	/*
 684	 * We try to cluster swap pages by allocating them sequentially
 685	 * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 686	 * way, however, we resort to first-free allocation, starting
 687	 * a new cluster.  This prevents us from scattering swap pages
 688	 * all over the entire swap partition, so that we reduce
 689	 * overall disk seek times between swap pages.  -- sct
 690	 * But we do now try to find an empty cluster.  -Andrea
 691	 * And we let swap pages go all over an SSD partition.  Hugh
 692	 */
 693
 694	si->flags += SWP_SCANNING;
 695	scan_base = offset = si->cluster_next;
 696
 697	/* SSD algorithm */
 698	if (si->cluster_info) {
 699		if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
 700			goto checks;
 701		else
 702			goto scan;
 703	}
 704
 705	if (unlikely(!si->cluster_nr--)) {
 706		if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 707			si->cluster_nr = SWAPFILE_CLUSTER - 1;
 708			goto checks;
 709		}
 710
 711		spin_unlock(&si->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 712
 713		/*
 714		 * If seek is expensive, start searching for new cluster from
 715		 * start of partition, to minimize the span of allocated swap.
 716		 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
 717		 * case, just handled by scan_swap_map_try_ssd_cluster() above.
 
 
 718		 */
 719		scan_base = offset = si->lowest_bit;
 
 720		last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 721
 722		/* Locate the first empty (unaligned) cluster */
 723		for (; last_in_cluster <= si->highest_bit; offset++) {
 724			if (si->swap_map[offset])
 725				last_in_cluster = offset + SWAPFILE_CLUSTER;
 726			else if (offset == last_in_cluster) {
 727				spin_lock(&si->lock);
 728				offset -= SWAPFILE_CLUSTER - 1;
 729				si->cluster_next = offset;
 730				si->cluster_nr = SWAPFILE_CLUSTER - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 731				goto checks;
 732			}
 733			if (unlikely(--latency_ration < 0)) {
 734				cond_resched();
 735				latency_ration = LATENCY_LIMIT;
 736			}
 737		}
 738
 739		offset = scan_base;
 740		spin_lock(&si->lock);
 741		si->cluster_nr = SWAPFILE_CLUSTER - 1;
 
 742	}
 743
 744checks:
 745	if (si->cluster_info) {
 746		while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
 747		/* take a break if we already got some slots */
 748			if (n_ret)
 749				goto done;
 750			if (!scan_swap_map_try_ssd_cluster(si, &offset,
 751							&scan_base))
 752				goto scan;
 753		}
 754	}
 755	if (!(si->flags & SWP_WRITEOK))
 756		goto no_page;
 757	if (!si->highest_bit)
 758		goto no_page;
 759	if (offset > si->highest_bit)
 760		scan_base = offset = si->lowest_bit;
 761
 762	ci = lock_cluster(si, offset);
 763	/* reuse swap entry of cache-only swap if not busy. */
 764	if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 765		int swap_was_freed;
 766		unlock_cluster(ci);
 767		spin_unlock(&si->lock);
 768		swap_was_freed = __try_to_reclaim_swap(si, offset);
 769		spin_lock(&si->lock);
 770		/* entry was freed successfully, try to use this again */
 771		if (swap_was_freed)
 772			goto checks;
 773		goto scan; /* check next one */
 774	}
 775
 776	if (si->swap_map[offset]) {
 777		unlock_cluster(ci);
 778		if (!n_ret)
 779			goto scan;
 780		else
 781			goto done;
 
 
 
 
 
 782	}
 783	si->swap_map[offset] = usage;
 784	inc_cluster_info_page(si, si->cluster_info, offset);
 785	unlock_cluster(ci);
 786
 787	swap_range_alloc(si, offset, 1);
 788	si->cluster_next = offset + 1;
 789	slots[n_ret++] = swp_entry(si->type, offset);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 790
 791	/* got enough slots or reach max slots? */
 792	if ((n_ret == nr) || (offset >= si->highest_bit))
 793		goto done;
 794
 795	/* search for next available slot */
 796
 797	/* time to take a break? */
 798	if (unlikely(--latency_ration < 0)) {
 799		if (n_ret)
 800			goto done;
 801		spin_unlock(&si->lock);
 802		cond_resched();
 803		spin_lock(&si->lock);
 804		latency_ration = LATENCY_LIMIT;
 805	}
 806
 807	/* try to get more slots in cluster */
 808	if (si->cluster_info) {
 809		if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
 810			goto checks;
 811		else
 812			goto done;
 813	}
 814	/* non-ssd case */
 815	++offset;
 816
 817	/* non-ssd case, still more slots in cluster? */
 818	if (si->cluster_nr && !si->swap_map[offset]) {
 819		--si->cluster_nr;
 820		goto checks;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 821	}
 822
 823done:
 824	si->flags -= SWP_SCANNING;
 825	return n_ret;
 826
 827scan:
 828	spin_unlock(&si->lock);
 829	while (++offset <= si->highest_bit) {
 830		if (!si->swap_map[offset]) {
 831			spin_lock(&si->lock);
 832			goto checks;
 833		}
 834		if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 835			spin_lock(&si->lock);
 836			goto checks;
 837		}
 838		if (unlikely(--latency_ration < 0)) {
 839			cond_resched();
 840			latency_ration = LATENCY_LIMIT;
 841		}
 842	}
 843	offset = si->lowest_bit;
 844	while (offset < scan_base) {
 845		if (!si->swap_map[offset]) {
 846			spin_lock(&si->lock);
 847			goto checks;
 848		}
 849		if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 850			spin_lock(&si->lock);
 851			goto checks;
 852		}
 853		if (unlikely(--latency_ration < 0)) {
 854			cond_resched();
 855			latency_ration = LATENCY_LIMIT;
 856		}
 857		offset++;
 858	}
 859	spin_lock(&si->lock);
 860
 861no_page:
 862	si->flags -= SWP_SCANNING;
 863	return n_ret;
 864}
 865
 866#ifdef CONFIG_THP_SWAP
 867static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
 868{
 869	unsigned long idx;
 870	struct swap_cluster_info *ci;
 871	unsigned long offset, i;
 872	unsigned char *map;
 873
 874	if (cluster_list_empty(&si->free_clusters))
 875		return 0;
 876
 877	idx = cluster_list_first(&si->free_clusters);
 878	offset = idx * SWAPFILE_CLUSTER;
 879	ci = lock_cluster(si, offset);
 880	alloc_cluster(si, idx);
 881	cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
 882
 883	map = si->swap_map + offset;
 884	for (i = 0; i < SWAPFILE_CLUSTER; i++)
 885		map[i] = SWAP_HAS_CACHE;
 886	unlock_cluster(ci);
 887	swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
 888	*slot = swp_entry(si->type, offset);
 889
 890	return 1;
 891}
 892
 893static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
 894{
 895	unsigned long offset = idx * SWAPFILE_CLUSTER;
 896	struct swap_cluster_info *ci;
 897
 898	ci = lock_cluster(si, offset);
 899	cluster_set_count_flag(ci, 0, 0);
 900	free_cluster(si, idx);
 901	unlock_cluster(ci);
 902	swap_range_free(si, offset, SWAPFILE_CLUSTER);
 903}
 904#else
 905static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
 906{
 907	VM_WARN_ON_ONCE(1);
 908	return 0;
 909}
 910#endif /* CONFIG_THP_SWAP */
 911
 912static unsigned long scan_swap_map(struct swap_info_struct *si,
 913				   unsigned char usage)
 914{
 915	swp_entry_t entry;
 916	int n_ret;
 
 
 917
 918	n_ret = scan_swap_map_slots(si, usage, 1, &entry);
 919
 920	if (n_ret)
 921		return swp_offset(entry);
 922	else
 923		return 0;
 924
 925}
 926
 927int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
 928{
 929	unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
 930	struct swap_info_struct *si, *next;
 931	long avail_pgs;
 932	int n_ret = 0;
 933	int node;
 934
 935	/* Only single cluster request supported */
 936	WARN_ON_ONCE(n_goal > 1 && cluster);
 937
 938	avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
 939	if (avail_pgs <= 0)
 940		goto noswap;
 
 941
 942	if (n_goal > SWAP_BATCH)
 943		n_goal = SWAP_BATCH;
 
 
 
 
 
 
 944
 945	if (n_goal > avail_pgs)
 946		n_goal = avail_pgs;
 
 
 947
 948	atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
 
 
 
 
 
 
 
 
 949
 950	spin_lock(&swap_avail_lock);
 951
 952start_over:
 953	node = numa_node_id();
 954	plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
 955		/* requeue si to after same-priority siblings */
 956		plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
 957		spin_unlock(&swap_avail_lock);
 958		spin_lock(&si->lock);
 959		if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
 960			spin_lock(&swap_avail_lock);
 961			if (plist_node_empty(&si->avail_lists[node])) {
 962				spin_unlock(&si->lock);
 963				goto nextsi;
 964			}
 965			WARN(!si->highest_bit,
 966			     "swap_info %d in list but !highest_bit\n",
 967			     si->type);
 968			WARN(!(si->flags & SWP_WRITEOK),
 969			     "swap_info %d in list but !SWP_WRITEOK\n",
 970			     si->type);
 971			__del_from_avail_list(si);
 972			spin_unlock(&si->lock);
 973			goto nextsi;
 974		}
 975		if (cluster) {
 976			if (!(si->flags & SWP_FILE))
 977				n_ret = swap_alloc_cluster(si, swp_entries);
 978		} else
 979			n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
 980						    n_goal, swp_entries);
 981		spin_unlock(&si->lock);
 982		if (n_ret || cluster)
 983			goto check_out;
 984		pr_debug("scan_swap_map of si %d failed to find offset\n",
 985			si->type);
 986
 987		spin_lock(&swap_avail_lock);
 988nextsi:
 989		/*
 990		 * if we got here, it's likely that si was almost full before,
 991		 * and since scan_swap_map() can drop the si->lock, multiple
 992		 * callers probably all tried to get a page from the same si
 993		 * and it filled up before we could get one; or, the si filled
 994		 * up between us dropping swap_avail_lock and taking si->lock.
 995		 * Since we dropped the swap_avail_lock, the swap_avail_head
 996		 * list may have been modified; so if next is still in the
 997		 * swap_avail_head list then try it, otherwise start over
 998		 * if we have not gotten any slots.
 999		 */
1000		if (plist_node_empty(&next->avail_lists[node]))
1001			goto start_over;
1002	}
1003
1004	spin_unlock(&swap_avail_lock);
1005
1006check_out:
1007	if (n_ret < n_goal)
1008		atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1009				&nr_swap_pages);
1010noswap:
1011	return n_ret;
 
1012}
1013
1014/* The only caller of this function is now suspend routine */
1015swp_entry_t get_swap_page_of_type(int type)
1016{
1017	struct swap_info_struct *si;
1018	pgoff_t offset;
1019
 
1020	si = swap_info[type];
1021	spin_lock(&si->lock);
1022	if (si && (si->flags & SWP_WRITEOK)) {
1023		atomic_long_dec(&nr_swap_pages);
1024		/* This is called for allocating swap entry, not cache */
1025		offset = scan_swap_map(si, 1);
1026		if (offset) {
1027			spin_unlock(&si->lock);
1028			return swp_entry(type, offset);
1029		}
1030		atomic_long_inc(&nr_swap_pages);
1031	}
1032	spin_unlock(&si->lock);
1033	return (swp_entry_t) {0};
1034}
1035
1036static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1037{
1038	struct swap_info_struct *p;
1039	unsigned long offset, type;
1040
1041	if (!entry.val)
1042		goto out;
1043	type = swp_type(entry);
1044	if (type >= nr_swapfiles)
1045		goto bad_nofile;
1046	p = swap_info[type];
1047	if (!(p->flags & SWP_USED))
1048		goto bad_device;
1049	offset = swp_offset(entry);
1050	if (offset >= p->max)
1051		goto bad_offset;
 
 
 
1052	return p;
1053
 
 
 
1054bad_offset:
1055	pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1056	goto out;
1057bad_device:
1058	pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1059	goto out;
1060bad_nofile:
1061	pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1062out:
1063	return NULL;
1064}
1065
1066static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
 
1067{
1068	struct swap_info_struct *p;
1069
1070	p = __swap_info_get(entry);
1071	if (!p)
1072		goto out;
1073	if (!p->swap_map[swp_offset(entry)])
1074		goto bad_free;
1075	return p;
1076
1077bad_free:
1078	pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1079	goto out;
1080out:
1081	return NULL;
1082}
1083
1084static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1085{
1086	struct swap_info_struct *p;
1087
1088	p = _swap_info_get(entry);
1089	if (p)
1090		spin_lock(&p->lock);
1091	return p;
1092}
1093
1094static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1095					struct swap_info_struct *q)
1096{
1097	struct swap_info_struct *p;
1098
1099	p = _swap_info_get(entry);
1100
1101	if (p != q) {
1102		if (q != NULL)
1103			spin_unlock(&q->lock);
1104		if (p != NULL)
1105			spin_lock(&p->lock);
1106	}
1107	return p;
1108}
1109
1110static unsigned char __swap_entry_free(struct swap_info_struct *p,
1111				       swp_entry_t entry, unsigned char usage)
1112{
1113	struct swap_cluster_info *ci;
1114	unsigned long offset = swp_offset(entry);
1115	unsigned char count;
1116	unsigned char has_cache;
1117
1118	ci = lock_cluster_or_swap_info(p, offset);
1119
1120	count = p->swap_map[offset];
1121
1122	has_cache = count & SWAP_HAS_CACHE;
1123	count &= ~SWAP_HAS_CACHE;
1124
1125	if (usage == SWAP_HAS_CACHE) {
1126		VM_BUG_ON(!has_cache);
1127		has_cache = 0;
1128	} else if (count == SWAP_MAP_SHMEM) {
1129		/*
1130		 * Or we could insist on shmem.c using a special
1131		 * swap_shmem_free() and free_shmem_swap_and_cache()...
1132		 */
1133		count = 0;
1134	} else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1135		if (count == COUNT_CONTINUED) {
1136			if (swap_count_continued(p, offset, count))
1137				count = SWAP_MAP_MAX | COUNT_CONTINUED;
1138			else
1139				count = SWAP_MAP_MAX;
1140		} else
1141			count--;
1142	}
1143
 
 
 
1144	usage = count | has_cache;
1145	p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1146
1147	unlock_cluster_or_swap_info(p, ci);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1148
1149	return usage;
1150}
1151
1152static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1153{
1154	struct swap_cluster_info *ci;
1155	unsigned long offset = swp_offset(entry);
1156	unsigned char count;
1157
1158	ci = lock_cluster(p, offset);
1159	count = p->swap_map[offset];
1160	VM_BUG_ON(count != SWAP_HAS_CACHE);
1161	p->swap_map[offset] = 0;
1162	dec_cluster_info_page(p, p->cluster_info, offset);
1163	unlock_cluster(ci);
1164
1165	mem_cgroup_uncharge_swap(entry, 1);
1166	swap_range_free(p, offset, 1);
1167}
1168
1169/*
1170 * Caller has made sure that the swap device corresponding to entry
1171 * is still around or has not been recycled.
1172 */
1173void swap_free(swp_entry_t entry)
1174{
1175	struct swap_info_struct *p;
1176
1177	p = _swap_info_get(entry);
1178	if (p) {
1179		if (!__swap_entry_free(p, entry, 1))
1180			free_swap_slot(entry);
1181	}
1182}
1183
1184/*
1185 * Called after dropping swapcache to decrease refcnt to swap entries.
1186 */
1187static void swapcache_free(swp_entry_t entry)
1188{
1189	struct swap_info_struct *p;
 
1190
1191	p = _swap_info_get(entry);
1192	if (p) {
1193		if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1194			free_swap_slot(entry);
1195	}
1196}
1197
1198#ifdef CONFIG_THP_SWAP
1199static void swapcache_free_cluster(swp_entry_t entry)
1200{
1201	unsigned long offset = swp_offset(entry);
1202	unsigned long idx = offset / SWAPFILE_CLUSTER;
1203	struct swap_cluster_info *ci;
1204	struct swap_info_struct *si;
1205	unsigned char *map;
1206	unsigned int i, free_entries = 0;
1207	unsigned char val;
1208
1209	si = _swap_info_get(entry);
1210	if (!si)
1211		return;
1212
1213	ci = lock_cluster(si, offset);
1214	VM_BUG_ON(!cluster_is_huge(ci));
1215	map = si->swap_map + offset;
1216	for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1217		val = map[i];
1218		VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1219		if (val == SWAP_HAS_CACHE)
1220			free_entries++;
1221	}
1222	if (!free_entries) {
1223		for (i = 0; i < SWAPFILE_CLUSTER; i++)
1224			map[i] &= ~SWAP_HAS_CACHE;
1225	}
1226	cluster_clear_huge(ci);
1227	unlock_cluster(ci);
1228	if (free_entries == SWAPFILE_CLUSTER) {
1229		spin_lock(&si->lock);
1230		ci = lock_cluster(si, offset);
1231		memset(map, 0, SWAPFILE_CLUSTER);
1232		unlock_cluster(ci);
1233		mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1234		swap_free_cluster(si, idx);
1235		spin_unlock(&si->lock);
1236	} else if (free_entries) {
1237		for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1238			if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1239				free_swap_slot(entry);
1240		}
1241	}
1242}
1243
1244int split_swap_cluster(swp_entry_t entry)
1245{
1246	struct swap_info_struct *si;
1247	struct swap_cluster_info *ci;
1248	unsigned long offset = swp_offset(entry);
1249
1250	si = _swap_info_get(entry);
1251	if (!si)
1252		return -EBUSY;
1253	ci = lock_cluster(si, offset);
1254	cluster_clear_huge(ci);
1255	unlock_cluster(ci);
1256	return 0;
1257}
1258#else
1259static inline void swapcache_free_cluster(swp_entry_t entry)
1260{
1261}
1262#endif /* CONFIG_THP_SWAP */
1263
1264void put_swap_page(struct page *page, swp_entry_t entry)
1265{
1266	if (!PageTransHuge(page))
1267		swapcache_free(entry);
1268	else
1269		swapcache_free_cluster(entry);
1270}
1271
1272static int swp_entry_cmp(const void *ent1, const void *ent2)
1273{
1274	const swp_entry_t *e1 = ent1, *e2 = ent2;
1275
1276	return (int)swp_type(*e1) - (int)swp_type(*e2);
1277}
1278
1279void swapcache_free_entries(swp_entry_t *entries, int n)
1280{
1281	struct swap_info_struct *p, *prev;
1282	int i;
1283
1284	if (n <= 0)
1285		return;
1286
1287	prev = NULL;
1288	p = NULL;
1289
1290	/*
1291	 * Sort swap entries by swap device, so each lock is only taken once.
1292	 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1293	 * so low that it isn't necessary to optimize further.
1294	 */
1295	if (nr_swapfiles > 1)
1296		sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1297	for (i = 0; i < n; ++i) {
1298		p = swap_info_get_cont(entries[i], prev);
1299		if (p)
1300			swap_entry_free(p, entries[i]);
1301		prev = p;
1302	}
1303	if (p)
1304		spin_unlock(&p->lock);
1305}
1306
1307/*
1308 * How many references to page are currently swapped out?
1309 * This does not give an exact answer when swap count is continued,
1310 * but does include the high COUNT_CONTINUED flag to allow for that.
1311 */
1312int page_swapcount(struct page *page)
1313{
1314	int count = 0;
1315	struct swap_info_struct *p;
1316	struct swap_cluster_info *ci;
1317	swp_entry_t entry;
1318	unsigned long offset;
1319
1320	entry.val = page_private(page);
1321	p = _swap_info_get(entry);
1322	if (p) {
1323		offset = swp_offset(entry);
1324		ci = lock_cluster_or_swap_info(p, offset);
1325		count = swap_count(p->swap_map[offset]);
1326		unlock_cluster_or_swap_info(p, ci);
1327	}
1328	return count;
1329}
1330
1331int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1332{
1333	pgoff_t offset = swp_offset(entry);
1334
1335	return swap_count(si->swap_map[offset]);
1336}
1337
1338static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1339{
1340	int count = 0;
1341	pgoff_t offset = swp_offset(entry);
1342	struct swap_cluster_info *ci;
1343
1344	ci = lock_cluster_or_swap_info(si, offset);
1345	count = swap_count(si->swap_map[offset]);
1346	unlock_cluster_or_swap_info(si, ci);
1347	return count;
1348}
1349
1350/*
1351 * How many references to @entry are currently swapped out?
1352 * This does not give an exact answer when swap count is continued,
1353 * but does include the high COUNT_CONTINUED flag to allow for that.
1354 */
1355int __swp_swapcount(swp_entry_t entry)
1356{
1357	int count = 0;
1358	struct swap_info_struct *si;
1359
1360	si = __swap_info_get(entry);
1361	if (si)
1362		count = swap_swapcount(si, entry);
1363	return count;
1364}
1365
1366/*
1367 * How many references to @entry are currently swapped out?
1368 * This considers COUNT_CONTINUED so it returns exact answer.
1369 */
1370int swp_swapcount(swp_entry_t entry)
1371{
1372	int count, tmp_count, n;
1373	struct swap_info_struct *p;
1374	struct swap_cluster_info *ci;
1375	struct page *page;
1376	pgoff_t offset;
1377	unsigned char *map;
1378
1379	p = _swap_info_get(entry);
1380	if (!p)
1381		return 0;
1382
1383	offset = swp_offset(entry);
1384
1385	ci = lock_cluster_or_swap_info(p, offset);
1386
1387	count = swap_count(p->swap_map[offset]);
1388	if (!(count & COUNT_CONTINUED))
1389		goto out;
1390
1391	count &= ~COUNT_CONTINUED;
1392	n = SWAP_MAP_MAX + 1;
1393
1394	page = vmalloc_to_page(p->swap_map + offset);
1395	offset &= ~PAGE_MASK;
1396	VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1397
1398	do {
1399		page = list_next_entry(page, lru);
1400		map = kmap_atomic(page);
1401		tmp_count = map[offset];
1402		kunmap_atomic(map);
1403
1404		count += (tmp_count & ~COUNT_CONTINUED) * n;
1405		n *= (SWAP_CONT_MAX + 1);
1406	} while (tmp_count & COUNT_CONTINUED);
1407out:
1408	unlock_cluster_or_swap_info(p, ci);
1409	return count;
1410}
1411
1412#ifdef CONFIG_THP_SWAP
1413static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1414					 swp_entry_t entry)
1415{
1416	struct swap_cluster_info *ci;
1417	unsigned char *map = si->swap_map;
1418	unsigned long roffset = swp_offset(entry);
1419	unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1420	int i;
1421	bool ret = false;
1422
1423	ci = lock_cluster_or_swap_info(si, offset);
1424	if (!ci || !cluster_is_huge(ci)) {
1425		if (map[roffset] != SWAP_HAS_CACHE)
1426			ret = true;
1427		goto unlock_out;
1428	}
1429	for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1430		if (map[offset + i] != SWAP_HAS_CACHE) {
1431			ret = true;
1432			break;
1433		}
1434	}
1435unlock_out:
1436	unlock_cluster_or_swap_info(si, ci);
1437	return ret;
1438}
1439
1440static bool page_swapped(struct page *page)
1441{
1442	swp_entry_t entry;
1443	struct swap_info_struct *si;
1444
1445	if (likely(!PageTransCompound(page)))
1446		return page_swapcount(page) != 0;
1447
1448	page = compound_head(page);
1449	entry.val = page_private(page);
1450	si = _swap_info_get(entry);
1451	if (si)
1452		return swap_page_trans_huge_swapped(si, entry);
1453	return false;
1454}
1455
1456static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1457					 int *total_swapcount)
1458{
1459	int i, map_swapcount, _total_mapcount, _total_swapcount;
1460	unsigned long offset = 0;
1461	struct swap_info_struct *si;
1462	struct swap_cluster_info *ci = NULL;
1463	unsigned char *map = NULL;
1464	int mapcount, swapcount = 0;
1465
1466	/* hugetlbfs shouldn't call it */
1467	VM_BUG_ON_PAGE(PageHuge(page), page);
1468
1469	if (likely(!PageTransCompound(page))) {
1470		mapcount = atomic_read(&page->_mapcount) + 1;
1471		if (total_mapcount)
1472			*total_mapcount = mapcount;
1473		if (PageSwapCache(page))
1474			swapcount = page_swapcount(page);
1475		if (total_swapcount)
1476			*total_swapcount = swapcount;
1477		return mapcount + swapcount;
1478	}
1479
1480	page = compound_head(page);
1481
1482	_total_mapcount = _total_swapcount = map_swapcount = 0;
1483	if (PageSwapCache(page)) {
1484		swp_entry_t entry;
1485
1486		entry.val = page_private(page);
1487		si = _swap_info_get(entry);
1488		if (si) {
1489			map = si->swap_map;
1490			offset = swp_offset(entry);
1491		}
1492	}
1493	if (map)
1494		ci = lock_cluster(si, offset);
1495	for (i = 0; i < HPAGE_PMD_NR; i++) {
1496		mapcount = atomic_read(&page[i]._mapcount) + 1;
1497		_total_mapcount += mapcount;
1498		if (map) {
1499			swapcount = swap_count(map[offset + i]);
1500			_total_swapcount += swapcount;
1501		}
1502		map_swapcount = max(map_swapcount, mapcount + swapcount);
1503	}
1504	unlock_cluster(ci);
1505	if (PageDoubleMap(page)) {
1506		map_swapcount -= 1;
1507		_total_mapcount -= HPAGE_PMD_NR;
1508	}
1509	mapcount = compound_mapcount(page);
1510	map_swapcount += mapcount;
1511	_total_mapcount += mapcount;
1512	if (total_mapcount)
1513		*total_mapcount = _total_mapcount;
1514	if (total_swapcount)
1515		*total_swapcount = _total_swapcount;
1516
1517	return map_swapcount;
1518}
1519#else
1520#define swap_page_trans_huge_swapped(si, entry)	swap_swapcount(si, entry)
1521#define page_swapped(page)			(page_swapcount(page) != 0)
1522
1523static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1524					 int *total_swapcount)
1525{
1526	int mapcount, swapcount = 0;
1527
1528	/* hugetlbfs shouldn't call it */
1529	VM_BUG_ON_PAGE(PageHuge(page), page);
1530
1531	mapcount = page_trans_huge_mapcount(page, total_mapcount);
1532	if (PageSwapCache(page))
1533		swapcount = page_swapcount(page);
1534	if (total_swapcount)
1535		*total_swapcount = swapcount;
1536	return mapcount + swapcount;
1537}
1538#endif
1539
1540/*
1541 * We can write to an anon page without COW if there are no other references
1542 * to it.  And as a side-effect, free up its swap: because the old content
1543 * on disk will never be read, and seeking back there to write new content
1544 * later would only waste time away from clustering.
1545 *
1546 * NOTE: total_map_swapcount should not be relied upon by the caller if
1547 * reuse_swap_page() returns false, but it may be always overwritten
1548 * (see the other implementation for CONFIG_SWAP=n).
1549 */
1550bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1551{
1552	int count, total_mapcount, total_swapcount;
1553
1554	VM_BUG_ON_PAGE(!PageLocked(page), page);
1555	if (unlikely(PageKsm(page)))
1556		return false;
1557	count = page_trans_huge_map_swapcount(page, &total_mapcount,
1558					      &total_swapcount);
1559	if (total_map_swapcount)
1560		*total_map_swapcount = total_mapcount + total_swapcount;
1561	if (count == 1 && PageSwapCache(page) &&
1562	    (likely(!PageTransCompound(page)) ||
1563	     /* The remaining swap count will be freed soon */
1564	     total_swapcount == page_swapcount(page))) {
1565		if (!PageWriteback(page)) {
1566			page = compound_head(page);
1567			delete_from_swap_cache(page);
1568			SetPageDirty(page);
1569		} else {
1570			swp_entry_t entry;
1571			struct swap_info_struct *p;
1572
1573			entry.val = page_private(page);
1574			p = swap_info_get(entry);
1575			if (p->flags & SWP_STABLE_WRITES) {
1576				spin_unlock(&p->lock);
1577				return false;
1578			}
1579			spin_unlock(&p->lock);
1580		}
1581	}
1582
1583	return count <= 1;
1584}
1585
1586/*
1587 * If swap is getting full, or if there are no more mappings of this page,
1588 * then try_to_free_swap is called to free its swap space.
1589 */
1590int try_to_free_swap(struct page *page)
1591{
1592	VM_BUG_ON_PAGE(!PageLocked(page), page);
1593
1594	if (!PageSwapCache(page))
1595		return 0;
1596	if (PageWriteback(page))
1597		return 0;
1598	if (page_swapped(page))
1599		return 0;
1600
1601	/*
1602	 * Once hibernation has begun to create its image of memory,
1603	 * there's a danger that one of the calls to try_to_free_swap()
1604	 * - most probably a call from __try_to_reclaim_swap() while
1605	 * hibernation is allocating its own swap pages for the image,
1606	 * but conceivably even a call from memory reclaim - will free
1607	 * the swap from a page which has already been recorded in the
1608	 * image as a clean swapcache page, and then reuse its swap for
1609	 * another page of the image.  On waking from hibernation, the
1610	 * original page might be freed under memory pressure, then
1611	 * later read back in from swap, now with the wrong data.
1612	 *
1613	 * Hibernation suspends storage while it is writing the image
1614	 * to disk so check that here.
1615	 */
1616	if (pm_suspended_storage())
1617		return 0;
1618
1619	page = compound_head(page);
1620	delete_from_swap_cache(page);
1621	SetPageDirty(page);
1622	return 1;
1623}
1624
1625/*
1626 * Free the swap entry like above, but also try to
1627 * free the page cache entry if it is the last user.
1628 */
1629int free_swap_and_cache(swp_entry_t entry)
1630{
1631	struct swap_info_struct *p;
1632	struct page *page = NULL;
1633	unsigned char count;
1634
1635	if (non_swap_entry(entry))
1636		return 1;
1637
1638	p = _swap_info_get(entry);
1639	if (p) {
1640		count = __swap_entry_free(p, entry, 1);
1641		if (count == SWAP_HAS_CACHE &&
1642		    !swap_page_trans_huge_swapped(p, entry)) {
1643			page = find_get_page(swap_address_space(entry),
1644					     swp_offset(entry));
1645			if (page && !trylock_page(page)) {
1646				put_page(page);
1647				page = NULL;
1648			}
1649		} else if (!count)
1650			free_swap_slot(entry);
1651	}
1652	if (page) {
1653		/*
1654		 * Not mapped elsewhere, or swap space full? Free it!
1655		 * Also recheck PageSwapCache now page is locked (above).
1656		 */
1657		if (PageSwapCache(page) && !PageWriteback(page) &&
1658		    (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1659		    !swap_page_trans_huge_swapped(p, entry)) {
1660			page = compound_head(page);
1661			delete_from_swap_cache(page);
1662			SetPageDirty(page);
1663		}
1664		unlock_page(page);
1665		put_page(page);
1666	}
1667	return p != NULL;
1668}
1669
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1670#ifdef CONFIG_HIBERNATION
1671/*
1672 * Find the swap type that corresponds to given device (if any).
1673 *
1674 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1675 * from 0, in which the swap header is expected to be located.
1676 *
1677 * This is needed for the suspend to disk (aka swsusp).
1678 */
1679int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1680{
1681	struct block_device *bdev = NULL;
1682	int type;
1683
1684	if (device)
1685		bdev = bdget(device);
1686
1687	spin_lock(&swap_lock);
1688	for (type = 0; type < nr_swapfiles; type++) {
1689		struct swap_info_struct *sis = swap_info[type];
1690
1691		if (!(sis->flags & SWP_WRITEOK))
1692			continue;
1693
1694		if (!bdev) {
1695			if (bdev_p)
1696				*bdev_p = bdgrab(sis->bdev);
1697
1698			spin_unlock(&swap_lock);
1699			return type;
1700		}
1701		if (bdev == sis->bdev) {
1702			struct swap_extent *se = &sis->first_swap_extent;
1703
1704			if (se->start_block == offset) {
1705				if (bdev_p)
1706					*bdev_p = bdgrab(sis->bdev);
1707
1708				spin_unlock(&swap_lock);
1709				bdput(bdev);
1710				return type;
1711			}
1712		}
1713	}
1714	spin_unlock(&swap_lock);
1715	if (bdev)
1716		bdput(bdev);
1717
1718	return -ENODEV;
1719}
1720
1721/*
1722 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1723 * corresponding to given index in swap_info (swap type).
1724 */
1725sector_t swapdev_block(int type, pgoff_t offset)
1726{
1727	struct block_device *bdev;
1728
1729	if ((unsigned int)type >= nr_swapfiles)
1730		return 0;
1731	if (!(swap_info[type]->flags & SWP_WRITEOK))
1732		return 0;
1733	return map_swap_entry(swp_entry(type, offset), &bdev);
1734}
1735
1736/*
1737 * Return either the total number of swap pages of given type, or the number
1738 * of free pages of that type (depending on @free)
1739 *
1740 * This is needed for software suspend
1741 */
1742unsigned int count_swap_pages(int type, int free)
1743{
1744	unsigned int n = 0;
1745
1746	spin_lock(&swap_lock);
1747	if ((unsigned int)type < nr_swapfiles) {
1748		struct swap_info_struct *sis = swap_info[type];
1749
1750		spin_lock(&sis->lock);
1751		if (sis->flags & SWP_WRITEOK) {
1752			n = sis->pages;
1753			if (free)
1754				n -= sis->inuse_pages;
1755		}
1756		spin_unlock(&sis->lock);
1757	}
1758	spin_unlock(&swap_lock);
1759	return n;
1760}
1761#endif /* CONFIG_HIBERNATION */
1762
1763static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1764{
1765	return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1766}
1767
1768/*
1769 * No need to decide whether this PTE shares the swap entry with others,
1770 * just let do_wp_page work it out if a write is requested later - to
1771 * force COW, vm_page_prot omits write permission from any private vma.
1772 */
1773static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1774		unsigned long addr, swp_entry_t entry, struct page *page)
1775{
1776	struct page *swapcache;
1777	struct mem_cgroup *memcg;
1778	spinlock_t *ptl;
1779	pte_t *pte;
1780	int ret = 1;
1781
1782	swapcache = page;
1783	page = ksm_might_need_to_copy(page, vma, addr);
1784	if (unlikely(!page))
1785		return -ENOMEM;
1786
1787	if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1788				&memcg, false)) {
1789		ret = -ENOMEM;
1790		goto out_nolock;
1791	}
1792
1793	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1794	if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1795		mem_cgroup_cancel_charge(page, memcg, false);
 
1796		ret = 0;
1797		goto out;
1798	}
1799
1800	dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1801	inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1802	get_page(page);
1803	set_pte_at(vma->vm_mm, addr, pte,
1804		   pte_mkold(mk_pte(page, vma->vm_page_prot)));
1805	if (page == swapcache) {
1806		page_add_anon_rmap(page, vma, addr, false);
1807		mem_cgroup_commit_charge(page, memcg, true, false);
1808	} else { /* ksm created a completely new copy */
1809		page_add_new_anon_rmap(page, vma, addr, false);
1810		mem_cgroup_commit_charge(page, memcg, false, false);
1811		lru_cache_add_active_or_unevictable(page, vma);
1812	}
1813	swap_free(entry);
1814	/*
1815	 * Move the page to the active list so it is not
1816	 * immediately swapped out again after swapon.
1817	 */
1818	activate_page(page);
1819out:
1820	pte_unmap_unlock(pte, ptl);
1821out_nolock:
1822	if (page != swapcache) {
1823		unlock_page(page);
1824		put_page(page);
1825	}
1826	return ret;
1827}
1828
1829static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1830				unsigned long addr, unsigned long end,
1831				swp_entry_t entry, struct page *page)
1832{
1833	pte_t swp_pte = swp_entry_to_pte(entry);
1834	pte_t *pte;
1835	int ret = 0;
1836
1837	/*
1838	 * We don't actually need pte lock while scanning for swp_pte: since
1839	 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1840	 * page table while we're scanning; though it could get zapped, and on
1841	 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1842	 * of unmatched parts which look like swp_pte, so unuse_pte must
1843	 * recheck under pte lock.  Scanning without pte lock lets it be
1844	 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1845	 */
1846	pte = pte_offset_map(pmd, addr);
1847	do {
1848		/*
1849		 * swapoff spends a _lot_ of time in this loop!
1850		 * Test inline before going to call unuse_pte.
1851		 */
1852		if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1853			pte_unmap(pte);
1854			ret = unuse_pte(vma, pmd, addr, entry, page);
1855			if (ret)
1856				goto out;
1857			pte = pte_offset_map(pmd, addr);
1858		}
1859	} while (pte++, addr += PAGE_SIZE, addr != end);
1860	pte_unmap(pte - 1);
1861out:
1862	return ret;
1863}
1864
1865static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1866				unsigned long addr, unsigned long end,
1867				swp_entry_t entry, struct page *page)
1868{
1869	pmd_t *pmd;
1870	unsigned long next;
1871	int ret;
1872
1873	pmd = pmd_offset(pud, addr);
1874	do {
1875		cond_resched();
1876		next = pmd_addr_end(addr, end);
1877		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 
 
1878			continue;
1879		ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1880		if (ret)
1881			return ret;
1882	} while (pmd++, addr = next, addr != end);
1883	return 0;
1884}
1885
1886static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1887				unsigned long addr, unsigned long end,
1888				swp_entry_t entry, struct page *page)
1889{
1890	pud_t *pud;
1891	unsigned long next;
1892	int ret;
1893
1894	pud = pud_offset(p4d, addr);
1895	do {
1896		next = pud_addr_end(addr, end);
1897		if (pud_none_or_clear_bad(pud))
1898			continue;
1899		ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1900		if (ret)
1901			return ret;
1902	} while (pud++, addr = next, addr != end);
1903	return 0;
1904}
1905
1906static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1907				unsigned long addr, unsigned long end,
1908				swp_entry_t entry, struct page *page)
1909{
1910	p4d_t *p4d;
1911	unsigned long next;
1912	int ret;
1913
1914	p4d = p4d_offset(pgd, addr);
1915	do {
1916		next = p4d_addr_end(addr, end);
1917		if (p4d_none_or_clear_bad(p4d))
1918			continue;
1919		ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1920		if (ret)
1921			return ret;
1922	} while (p4d++, addr = next, addr != end);
1923	return 0;
1924}
1925
1926static int unuse_vma(struct vm_area_struct *vma,
1927				swp_entry_t entry, struct page *page)
1928{
1929	pgd_t *pgd;
1930	unsigned long addr, end, next;
1931	int ret;
1932
1933	if (page_anon_vma(page)) {
1934		addr = page_address_in_vma(page, vma);
1935		if (addr == -EFAULT)
1936			return 0;
1937		else
1938			end = addr + PAGE_SIZE;
1939	} else {
1940		addr = vma->vm_start;
1941		end = vma->vm_end;
1942	}
1943
1944	pgd = pgd_offset(vma->vm_mm, addr);
1945	do {
1946		next = pgd_addr_end(addr, end);
1947		if (pgd_none_or_clear_bad(pgd))
1948			continue;
1949		ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1950		if (ret)
1951			return ret;
1952	} while (pgd++, addr = next, addr != end);
1953	return 0;
1954}
1955
1956static int unuse_mm(struct mm_struct *mm,
1957				swp_entry_t entry, struct page *page)
1958{
1959	struct vm_area_struct *vma;
1960	int ret = 0;
1961
1962	if (!down_read_trylock(&mm->mmap_sem)) {
1963		/*
1964		 * Activate page so shrink_inactive_list is unlikely to unmap
1965		 * its ptes while lock is dropped, so swapoff can make progress.
1966		 */
1967		activate_page(page);
1968		unlock_page(page);
1969		down_read(&mm->mmap_sem);
1970		lock_page(page);
1971	}
1972	for (vma = mm->mmap; vma; vma = vma->vm_next) {
1973		if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1974			break;
1975		cond_resched();
1976	}
1977	up_read(&mm->mmap_sem);
1978	return (ret < 0)? ret: 0;
1979}
1980
1981/*
1982 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1983 * from current position to next entry still in use.
1984 * Recycle to start on reaching the end, returning 0 when empty.
1985 */
1986static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1987					unsigned int prev, bool frontswap)
1988{
1989	unsigned int max = si->max;
1990	unsigned int i = prev;
1991	unsigned char count;
1992
1993	/*
1994	 * No need for swap_lock here: we're just looking
1995	 * for whether an entry is in use, not modifying it; false
1996	 * hits are okay, and sys_swapoff() has already prevented new
1997	 * allocations from this area (while holding swap_lock).
1998	 */
1999	for (;;) {
2000		if (++i >= max) {
2001			if (!prev) {
2002				i = 0;
2003				break;
2004			}
2005			/*
2006			 * No entries in use at top of swap_map,
2007			 * loop back to start and recheck there.
2008			 */
2009			max = prev + 1;
2010			prev = 0;
2011			i = 1;
2012		}
2013		count = READ_ONCE(si->swap_map[i]);
2014		if (count && swap_count(count) != SWAP_MAP_BAD)
2015			if (!frontswap || frontswap_test(si, i))
2016				break;
2017		if ((i % LATENCY_LIMIT) == 0)
2018			cond_resched();
2019	}
2020	return i;
2021}
2022
2023/*
2024 * We completely avoid races by reading each swap page in advance,
2025 * and then search for the process using it.  All the necessary
2026 * page table adjustments can then be made atomically.
2027 *
2028 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2029 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2030 */
2031int try_to_unuse(unsigned int type, bool frontswap,
2032		 unsigned long pages_to_unuse)
2033{
2034	struct swap_info_struct *si = swap_info[type];
2035	struct mm_struct *start_mm;
2036	volatile unsigned char *swap_map; /* swap_map is accessed without
2037					   * locking. Mark it as volatile
2038					   * to prevent compiler doing
2039					   * something odd.
2040					   */
2041	unsigned char swcount;
2042	struct page *page;
2043	swp_entry_t entry;
2044	unsigned int i = 0;
2045	int retval = 0;
2046
2047	/*
2048	 * When searching mms for an entry, a good strategy is to
2049	 * start at the first mm we freed the previous entry from
2050	 * (though actually we don't notice whether we or coincidence
2051	 * freed the entry).  Initialize this start_mm with a hold.
2052	 *
2053	 * A simpler strategy would be to start at the last mm we
2054	 * freed the previous entry from; but that would take less
2055	 * advantage of mmlist ordering, which clusters forked mms
2056	 * together, child after parent.  If we race with dup_mmap(), we
2057	 * prefer to resolve parent before child, lest we miss entries
2058	 * duplicated after we scanned child: using last mm would invert
2059	 * that.
2060	 */
2061	start_mm = &init_mm;
2062	mmget(&init_mm);
2063
2064	/*
2065	 * Keep on scanning until all entries have gone.  Usually,
2066	 * one pass through swap_map is enough, but not necessarily:
2067	 * there are races when an instance of an entry might be missed.
2068	 */
2069	while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2070		if (signal_pending(current)) {
2071			retval = -EINTR;
2072			break;
2073		}
2074
2075		/*
2076		 * Get a page for the entry, using the existing swap
2077		 * cache page if there is one.  Otherwise, get a clean
2078		 * page and read the swap into it.
2079		 */
2080		swap_map = &si->swap_map[i];
2081		entry = swp_entry(type, i);
2082		page = read_swap_cache_async(entry,
2083					GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2084		if (!page) {
2085			/*
2086			 * Either swap_duplicate() failed because entry
2087			 * has been freed independently, and will not be
2088			 * reused since sys_swapoff() already disabled
2089			 * allocation from here, or alloc_page() failed.
2090			 */
2091			swcount = *swap_map;
2092			/*
2093			 * We don't hold lock here, so the swap entry could be
2094			 * SWAP_MAP_BAD (when the cluster is discarding).
2095			 * Instead of fail out, We can just skip the swap
2096			 * entry because swapoff will wait for discarding
2097			 * finish anyway.
2098			 */
2099			if (!swcount || swcount == SWAP_MAP_BAD)
2100				continue;
2101			retval = -ENOMEM;
2102			break;
2103		}
2104
2105		/*
2106		 * Don't hold on to start_mm if it looks like exiting.
2107		 */
2108		if (atomic_read(&start_mm->mm_users) == 1) {
2109			mmput(start_mm);
2110			start_mm = &init_mm;
2111			mmget(&init_mm);
2112		}
2113
2114		/*
2115		 * Wait for and lock page.  When do_swap_page races with
2116		 * try_to_unuse, do_swap_page can handle the fault much
2117		 * faster than try_to_unuse can locate the entry.  This
2118		 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2119		 * defer to do_swap_page in such a case - in some tests,
2120		 * do_swap_page and try_to_unuse repeatedly compete.
2121		 */
2122		wait_on_page_locked(page);
2123		wait_on_page_writeback(page);
2124		lock_page(page);
2125		wait_on_page_writeback(page);
2126
2127		/*
2128		 * Remove all references to entry.
2129		 */
2130		swcount = *swap_map;
2131		if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2132			retval = shmem_unuse(entry, page);
2133			/* page has already been unlocked and released */
2134			if (retval < 0)
2135				break;
2136			continue;
2137		}
2138		if (swap_count(swcount) && start_mm != &init_mm)
2139			retval = unuse_mm(start_mm, entry, page);
2140
2141		if (swap_count(*swap_map)) {
2142			int set_start_mm = (*swap_map >= swcount);
2143			struct list_head *p = &start_mm->mmlist;
2144			struct mm_struct *new_start_mm = start_mm;
2145			struct mm_struct *prev_mm = start_mm;
2146			struct mm_struct *mm;
2147
2148			mmget(new_start_mm);
2149			mmget(prev_mm);
2150			spin_lock(&mmlist_lock);
2151			while (swap_count(*swap_map) && !retval &&
2152					(p = p->next) != &start_mm->mmlist) {
2153				mm = list_entry(p, struct mm_struct, mmlist);
2154				if (!mmget_not_zero(mm))
2155					continue;
2156				spin_unlock(&mmlist_lock);
2157				mmput(prev_mm);
2158				prev_mm = mm;
2159
2160				cond_resched();
2161
2162				swcount = *swap_map;
2163				if (!swap_count(swcount)) /* any usage ? */
2164					;
2165				else if (mm == &init_mm)
2166					set_start_mm = 1;
2167				else
2168					retval = unuse_mm(mm, entry, page);
2169
2170				if (set_start_mm && *swap_map < swcount) {
2171					mmput(new_start_mm);
2172					mmget(mm);
2173					new_start_mm = mm;
2174					set_start_mm = 0;
2175				}
2176				spin_lock(&mmlist_lock);
2177			}
2178			spin_unlock(&mmlist_lock);
2179			mmput(prev_mm);
2180			mmput(start_mm);
2181			start_mm = new_start_mm;
2182		}
2183		if (retval) {
2184			unlock_page(page);
2185			put_page(page);
2186			break;
2187		}
2188
2189		/*
2190		 * If a reference remains (rare), we would like to leave
2191		 * the page in the swap cache; but try_to_unmap could
2192		 * then re-duplicate the entry once we drop page lock,
2193		 * so we might loop indefinitely; also, that page could
2194		 * not be swapped out to other storage meanwhile.  So:
2195		 * delete from cache even if there's another reference,
2196		 * after ensuring that the data has been saved to disk -
2197		 * since if the reference remains (rarer), it will be
2198		 * read from disk into another page.  Splitting into two
2199		 * pages would be incorrect if swap supported "shared
2200		 * private" pages, but they are handled by tmpfs files.
2201		 *
2202		 * Given how unuse_vma() targets one particular offset
2203		 * in an anon_vma, once the anon_vma has been determined,
2204		 * this splitting happens to be just what is needed to
2205		 * handle where KSM pages have been swapped out: re-reading
2206		 * is unnecessarily slow, but we can fix that later on.
2207		 */
2208		if (swap_count(*swap_map) &&
2209		     PageDirty(page) && PageSwapCache(page)) {
2210			struct writeback_control wbc = {
2211				.sync_mode = WB_SYNC_NONE,
2212			};
2213
2214			swap_writepage(compound_head(page), &wbc);
2215			lock_page(page);
2216			wait_on_page_writeback(page);
2217		}
2218
2219		/*
2220		 * It is conceivable that a racing task removed this page from
2221		 * swap cache just before we acquired the page lock at the top,
2222		 * or while we dropped it in unuse_mm().  The page might even
2223		 * be back in swap cache on another swap area: that we must not
2224		 * delete, since it may not have been written out to swap yet.
2225		 */
2226		if (PageSwapCache(page) &&
2227		    likely(page_private(page) == entry.val) &&
2228		    !page_swapped(page))
2229			delete_from_swap_cache(compound_head(page));
2230
2231		/*
2232		 * So we could skip searching mms once swap count went
2233		 * to 1, we did not mark any present ptes as dirty: must
2234		 * mark page dirty so shrink_page_list will preserve it.
2235		 */
2236		SetPageDirty(page);
2237		unlock_page(page);
2238		put_page(page);
2239
2240		/*
2241		 * Make sure that we aren't completely killing
2242		 * interactive performance.
2243		 */
2244		cond_resched();
2245		if (frontswap && pages_to_unuse > 0) {
2246			if (!--pages_to_unuse)
2247				break;
2248		}
2249	}
2250
2251	mmput(start_mm);
2252	return retval;
2253}
2254
2255/*
2256 * After a successful try_to_unuse, if no swap is now in use, we know
2257 * we can empty the mmlist.  swap_lock must be held on entry and exit.
2258 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2259 * added to the mmlist just after page_duplicate - before would be racy.
2260 */
2261static void drain_mmlist(void)
2262{
2263	struct list_head *p, *next;
2264	unsigned int type;
2265
2266	for (type = 0; type < nr_swapfiles; type++)
2267		if (swap_info[type]->inuse_pages)
2268			return;
2269	spin_lock(&mmlist_lock);
2270	list_for_each_safe(p, next, &init_mm.mmlist)
2271		list_del_init(p);
2272	spin_unlock(&mmlist_lock);
2273}
2274
2275/*
2276 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2277 * corresponds to page offset for the specified swap entry.
2278 * Note that the type of this function is sector_t, but it returns page offset
2279 * into the bdev, not sector offset.
2280 */
2281static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2282{
2283	struct swap_info_struct *sis;
2284	struct swap_extent *start_se;
2285	struct swap_extent *se;
2286	pgoff_t offset;
2287
2288	sis = swap_info[swp_type(entry)];
2289	*bdev = sis->bdev;
2290
2291	offset = swp_offset(entry);
2292	start_se = sis->curr_swap_extent;
2293	se = start_se;
2294
2295	for ( ; ; ) {
 
 
2296		if (se->start_page <= offset &&
2297				offset < (se->start_page + se->nr_pages)) {
2298			return se->start_block + (offset - se->start_page);
2299		}
2300		se = list_next_entry(se, list);
 
2301		sis->curr_swap_extent = se;
2302		BUG_ON(se == start_se);		/* It *must* be present */
2303	}
2304}
2305
2306/*
2307 * Returns the page offset into bdev for the specified page's swap entry.
2308 */
2309sector_t map_swap_page(struct page *page, struct block_device **bdev)
2310{
2311	swp_entry_t entry;
2312	entry.val = page_private(page);
2313	return map_swap_entry(entry, bdev);
2314}
2315
2316/*
2317 * Free all of a swapdev's extent information
2318 */
2319static void destroy_swap_extents(struct swap_info_struct *sis)
2320{
2321	while (!list_empty(&sis->first_swap_extent.list)) {
2322		struct swap_extent *se;
2323
2324		se = list_first_entry(&sis->first_swap_extent.list,
2325				struct swap_extent, list);
2326		list_del(&se->list);
2327		kfree(se);
2328	}
2329
2330	if (sis->flags & SWP_FILE) {
2331		struct file *swap_file = sis->swap_file;
2332		struct address_space *mapping = swap_file->f_mapping;
2333
2334		sis->flags &= ~SWP_FILE;
2335		mapping->a_ops->swap_deactivate(swap_file);
2336	}
2337}
2338
2339/*
2340 * Add a block range (and the corresponding page range) into this swapdev's
2341 * extent list.  The extent list is kept sorted in page order.
2342 *
2343 * This function rather assumes that it is called in ascending page order.
2344 */
2345int
2346add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2347		unsigned long nr_pages, sector_t start_block)
2348{
2349	struct swap_extent *se;
2350	struct swap_extent *new_se;
2351	struct list_head *lh;
2352
2353	if (start_page == 0) {
2354		se = &sis->first_swap_extent;
2355		sis->curr_swap_extent = se;
2356		se->start_page = 0;
2357		se->nr_pages = nr_pages;
2358		se->start_block = start_block;
2359		return 1;
2360	} else {
2361		lh = sis->first_swap_extent.list.prev;	/* Highest extent */
2362		se = list_entry(lh, struct swap_extent, list);
2363		BUG_ON(se->start_page + se->nr_pages != start_page);
2364		if (se->start_block + se->nr_pages == start_block) {
2365			/* Merge it */
2366			se->nr_pages += nr_pages;
2367			return 0;
2368		}
2369	}
2370
2371	/*
2372	 * No merge.  Insert a new extent, preserving ordering.
2373	 */
2374	new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2375	if (new_se == NULL)
2376		return -ENOMEM;
2377	new_se->start_page = start_page;
2378	new_se->nr_pages = nr_pages;
2379	new_se->start_block = start_block;
2380
2381	list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2382	return 1;
2383}
2384
2385/*
2386 * A `swap extent' is a simple thing which maps a contiguous range of pages
2387 * onto a contiguous range of disk blocks.  An ordered list of swap extents
2388 * is built at swapon time and is then used at swap_writepage/swap_readpage
2389 * time for locating where on disk a page belongs.
2390 *
2391 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2392 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2393 * swap files identically.
2394 *
2395 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2396 * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2397 * swapfiles are handled *identically* after swapon time.
2398 *
2399 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2400 * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
2401 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2402 * requirements, they are simply tossed out - we will never use those blocks
2403 * for swapping.
2404 *
2405 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
2406 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2407 * which will scribble on the fs.
2408 *
2409 * The amount of disk space which a single swap extent represents varies.
2410 * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2411 * extents in the list.  To avoid much list walking, we cache the previous
2412 * search location in `curr_swap_extent', and start new searches from there.
2413 * This is extremely effective.  The average number of iterations in
2414 * map_swap_page() has been measured at about 0.3 per page.  - akpm.
2415 */
2416static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2417{
2418	struct file *swap_file = sis->swap_file;
2419	struct address_space *mapping = swap_file->f_mapping;
2420	struct inode *inode = mapping->host;
 
 
 
 
 
 
2421	int ret;
2422
 
2423	if (S_ISBLK(inode->i_mode)) {
2424		ret = add_swap_extent(sis, 0, sis->max, 0);
2425		*span = sis->pages;
2426		return ret;
2427	}
2428
2429	if (mapping->a_ops->swap_activate) {
2430		ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2431		if (!ret) {
2432			sis->flags |= SWP_FILE;
2433			ret = add_swap_extent(sis, 0, sis->max, 0);
2434			*span = sis->pages;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2435		}
2436		return ret;
2437	}
2438
2439	return generic_swapfile_activate(sis, swap_file, span);
2440}
 
2441
2442static int swap_node(struct swap_info_struct *p)
2443{
2444	struct block_device *bdev;
 
 
 
 
 
 
2445
2446	if (p->bdev)
2447		bdev = p->bdev;
2448	else
2449		bdev = p->swap_file->f_inode->i_sb->s_bdev;
 
 
 
2450
2451	return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2452}
2453
2454static void _enable_swap_info(struct swap_info_struct *p, int prio,
2455				unsigned char *swap_map,
2456				struct swap_cluster_info *cluster_info)
2457{
2458	int i;
2459
 
2460	if (prio >= 0)
2461		p->prio = prio;
2462	else
2463		p->prio = --least_priority;
2464	/*
2465	 * the plist prio is negated because plist ordering is
2466	 * low-to-high, while swap ordering is high-to-low
2467	 */
2468	p->list.prio = -p->prio;
2469	for_each_node(i) {
2470		if (p->prio >= 0)
2471			p->avail_lists[i].prio = -p->prio;
2472		else {
2473			if (swap_node(p) == i)
2474				p->avail_lists[i].prio = 1;
2475			else
2476				p->avail_lists[i].prio = -p->prio;
2477		}
2478	}
2479	p->swap_map = swap_map;
2480	p->cluster_info = cluster_info;
2481	p->flags |= SWP_WRITEOK;
2482	atomic_long_add(p->pages, &nr_swap_pages);
2483	total_swap_pages += p->pages;
2484
2485	assert_spin_locked(&swap_lock);
2486	/*
2487	 * both lists are plists, and thus priority ordered.
2488	 * swap_active_head needs to be priority ordered for swapoff(),
2489	 * which on removal of any swap_info_struct with an auto-assigned
2490	 * (i.e. negative) priority increments the auto-assigned priority
2491	 * of any lower-priority swap_info_structs.
2492	 * swap_avail_head needs to be priority ordered for get_swap_page(),
2493	 * which allocates swap pages from the highest available priority
2494	 * swap_info_struct.
2495	 */
2496	plist_add(&p->list, &swap_active_head);
2497	add_to_avail_list(p);
2498}
2499
2500static void enable_swap_info(struct swap_info_struct *p, int prio,
2501				unsigned char *swap_map,
2502				struct swap_cluster_info *cluster_info,
2503				unsigned long *frontswap_map)
2504{
2505	frontswap_init(p->type, frontswap_map);
2506	spin_lock(&swap_lock);
2507	spin_lock(&p->lock);
2508	 _enable_swap_info(p, prio, swap_map, cluster_info);
2509	spin_unlock(&p->lock);
2510	spin_unlock(&swap_lock);
2511}
2512
2513static void reinsert_swap_info(struct swap_info_struct *p)
2514{
2515	spin_lock(&swap_lock);
2516	spin_lock(&p->lock);
2517	_enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2518	spin_unlock(&p->lock);
2519	spin_unlock(&swap_lock);
2520}
2521
2522bool has_usable_swap(void)
2523{
2524	bool ret = true;
2525
2526	spin_lock(&swap_lock);
2527	if (plist_head_empty(&swap_active_head))
2528		ret = false;
2529	spin_unlock(&swap_lock);
2530	return ret;
2531}
2532
2533SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2534{
2535	struct swap_info_struct *p = NULL;
2536	unsigned char *swap_map;
2537	struct swap_cluster_info *cluster_info;
2538	unsigned long *frontswap_map;
2539	struct file *swap_file, *victim;
2540	struct address_space *mapping;
2541	struct inode *inode;
2542	struct filename *pathname;
2543	int err, found = 0;
2544	unsigned int old_block_size;
 
2545
2546	if (!capable(CAP_SYS_ADMIN))
2547		return -EPERM;
2548
2549	BUG_ON(!current->mm);
2550
2551	pathname = getname(specialfile);
 
2552	if (IS_ERR(pathname))
2553		return PTR_ERR(pathname);
2554
2555	victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
 
2556	err = PTR_ERR(victim);
2557	if (IS_ERR(victim))
2558		goto out;
2559
2560	mapping = victim->f_mapping;
 
2561	spin_lock(&swap_lock);
2562	plist_for_each_entry(p, &swap_active_head, list) {
 
2563		if (p->flags & SWP_WRITEOK) {
2564			if (p->swap_file->f_mapping == mapping) {
2565				found = 1;
2566				break;
2567			}
2568		}
 
2569	}
2570	if (!found) {
2571		err = -EINVAL;
2572		spin_unlock(&swap_lock);
2573		goto out_dput;
2574	}
2575	if (!security_vm_enough_memory_mm(current->mm, p->pages))
2576		vm_unacct_memory(p->pages);
2577	else {
2578		err = -ENOMEM;
2579		spin_unlock(&swap_lock);
2580		goto out_dput;
2581	}
2582	del_from_avail_list(p);
2583	spin_lock(&p->lock);
 
 
 
 
 
 
2584	if (p->prio < 0) {
2585		struct swap_info_struct *si = p;
2586		int nid;
2587
2588		plist_for_each_entry_continue(si, &swap_active_head, list) {
2589			si->prio++;
2590			si->list.prio--;
2591			for_each_node(nid) {
2592				if (si->avail_lists[nid].prio != 1)
2593					si->avail_lists[nid].prio--;
2594			}
2595		}
2596		least_priority++;
2597	}
2598	plist_del(&p->list, &swap_active_head);
2599	atomic_long_sub(p->pages, &nr_swap_pages);
2600	total_swap_pages -= p->pages;
2601	p->flags &= ~SWP_WRITEOK;
2602	spin_unlock(&p->lock);
2603	spin_unlock(&swap_lock);
2604
2605	disable_swap_slots_cache_lock();
2606
2607	set_current_oom_origin();
2608	err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2609	clear_current_oom_origin();
2610
2611	if (err) {
 
 
 
 
 
 
2612		/* re-insert swap space back into swap_list */
2613		reinsert_swap_info(p);
2614		reenable_swap_slots_cache_unlock();
2615		goto out_dput;
2616	}
2617
2618	reenable_swap_slots_cache_unlock();
2619
2620	flush_work(&p->discard_work);
2621
2622	destroy_swap_extents(p);
2623	if (p->flags & SWP_CONTINUED)
2624		free_swap_count_continuations(p);
2625
2626	if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2627		atomic_dec(&nr_rotate_swap);
2628
2629	mutex_lock(&swapon_mutex);
2630	spin_lock(&swap_lock);
2631	spin_lock(&p->lock);
2632	drain_mmlist();
2633
2634	/* wait for anyone still in scan_swap_map */
2635	p->highest_bit = 0;		/* cuts scans short */
2636	while (p->flags >= SWP_SCANNING) {
2637		spin_unlock(&p->lock);
2638		spin_unlock(&swap_lock);
2639		schedule_timeout_uninterruptible(1);
2640		spin_lock(&swap_lock);
2641		spin_lock(&p->lock);
2642	}
2643
2644	swap_file = p->swap_file;
2645	old_block_size = p->old_block_size;
2646	p->swap_file = NULL;
2647	p->max = 0;
2648	swap_map = p->swap_map;
2649	p->swap_map = NULL;
2650	cluster_info = p->cluster_info;
2651	p->cluster_info = NULL;
2652	frontswap_map = frontswap_map_get(p);
2653	spin_unlock(&p->lock);
2654	spin_unlock(&swap_lock);
2655	frontswap_invalidate_area(p->type);
2656	frontswap_map_set(p, NULL);
2657	mutex_unlock(&swapon_mutex);
2658	free_percpu(p->percpu_cluster);
2659	p->percpu_cluster = NULL;
2660	vfree(swap_map);
2661	kvfree(cluster_info);
2662	kvfree(frontswap_map);
2663	/* Destroy swap account information */
2664	swap_cgroup_swapoff(p->type);
2665	exit_swap_address_space(p->type);
2666
2667	inode = mapping->host;
2668	if (S_ISBLK(inode->i_mode)) {
2669		struct block_device *bdev = I_BDEV(inode);
2670		set_blocksize(bdev, old_block_size);
2671		blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2672	} else {
2673		inode_lock(inode);
2674		inode->i_flags &= ~S_SWAPFILE;
2675		inode_unlock(inode);
2676	}
2677	filp_close(swap_file, NULL);
2678
2679	/*
2680	 * Clear the SWP_USED flag after all resources are freed so that swapon
2681	 * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2682	 * not hold p->lock after we cleared its SWP_WRITEOK.
2683	 */
2684	spin_lock(&swap_lock);
2685	p->flags = 0;
2686	spin_unlock(&swap_lock);
2687
2688	err = 0;
2689	atomic_inc(&proc_poll_event);
2690	wake_up_interruptible(&proc_poll_wait);
2691
2692out_dput:
2693	filp_close(victim, NULL);
2694out:
2695	putname(pathname);
2696	return err;
2697}
2698
2699#ifdef CONFIG_PROC_FS
2700static __poll_t swaps_poll(struct file *file, poll_table *wait)
2701{
2702	struct seq_file *seq = file->private_data;
2703
2704	poll_wait(file, &proc_poll_wait, wait);
2705
2706	if (seq->poll_event != atomic_read(&proc_poll_event)) {
2707		seq->poll_event = atomic_read(&proc_poll_event);
2708		return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2709	}
2710
2711	return EPOLLIN | EPOLLRDNORM;
2712}
2713
2714/* iterator */
2715static void *swap_start(struct seq_file *swap, loff_t *pos)
2716{
2717	struct swap_info_struct *si;
2718	int type;
2719	loff_t l = *pos;
2720
2721	mutex_lock(&swapon_mutex);
2722
2723	if (!l)
2724		return SEQ_START_TOKEN;
2725
2726	for (type = 0; type < nr_swapfiles; type++) {
2727		smp_rmb();	/* read nr_swapfiles before swap_info[type] */
2728		si = swap_info[type];
2729		if (!(si->flags & SWP_USED) || !si->swap_map)
2730			continue;
2731		if (!--l)
2732			return si;
2733	}
2734
2735	return NULL;
2736}
2737
2738static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2739{
2740	struct swap_info_struct *si = v;
2741	int type;
2742
2743	if (v == SEQ_START_TOKEN)
2744		type = 0;
2745	else
2746		type = si->type + 1;
2747
2748	for (; type < nr_swapfiles; type++) {
2749		smp_rmb();	/* read nr_swapfiles before swap_info[type] */
2750		si = swap_info[type];
2751		if (!(si->flags & SWP_USED) || !si->swap_map)
2752			continue;
2753		++*pos;
2754		return si;
2755	}
2756
2757	return NULL;
2758}
2759
2760static void swap_stop(struct seq_file *swap, void *v)
2761{
2762	mutex_unlock(&swapon_mutex);
2763}
2764
2765static int swap_show(struct seq_file *swap, void *v)
2766{
2767	struct swap_info_struct *si = v;
2768	struct file *file;
2769	int len;
2770
2771	if (si == SEQ_START_TOKEN) {
2772		seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2773		return 0;
2774	}
2775
2776	file = si->swap_file;
2777	len = seq_file_path(swap, file, " \t\n\\");
2778	seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2779			len < 40 ? 40 - len : 1, " ",
2780			S_ISBLK(file_inode(file)->i_mode) ?
2781				"partition" : "file\t",
2782			si->pages << (PAGE_SHIFT - 10),
2783			si->inuse_pages << (PAGE_SHIFT - 10),
2784			si->prio);
2785	return 0;
2786}
2787
2788static const struct seq_operations swaps_op = {
2789	.start =	swap_start,
2790	.next =		swap_next,
2791	.stop =		swap_stop,
2792	.show =		swap_show
2793};
2794
2795static int swaps_open(struct inode *inode, struct file *file)
2796{
2797	struct seq_file *seq;
2798	int ret;
2799
2800	ret = seq_open(file, &swaps_op);
2801	if (ret)
2802		return ret;
2803
2804	seq = file->private_data;
2805	seq->poll_event = atomic_read(&proc_poll_event);
2806	return 0;
2807}
2808
2809static const struct file_operations proc_swaps_operations = {
2810	.open		= swaps_open,
2811	.read		= seq_read,
2812	.llseek		= seq_lseek,
2813	.release	= seq_release,
2814	.poll		= swaps_poll,
2815};
2816
2817static int __init procswaps_init(void)
2818{
2819	proc_create("swaps", 0, NULL, &proc_swaps_operations);
2820	return 0;
2821}
2822__initcall(procswaps_init);
2823#endif /* CONFIG_PROC_FS */
2824
2825#ifdef MAX_SWAPFILES_CHECK
2826static int __init max_swapfiles_check(void)
2827{
2828	MAX_SWAPFILES_CHECK();
2829	return 0;
2830}
2831late_initcall(max_swapfiles_check);
2832#endif
2833
2834static struct swap_info_struct *alloc_swap_info(void)
2835{
2836	struct swap_info_struct *p;
2837	unsigned int type;
2838	int i;
2839
2840	p = kzalloc(sizeof(*p), GFP_KERNEL);
2841	if (!p)
2842		return ERR_PTR(-ENOMEM);
2843
2844	spin_lock(&swap_lock);
2845	for (type = 0; type < nr_swapfiles; type++) {
2846		if (!(swap_info[type]->flags & SWP_USED))
2847			break;
2848	}
2849	if (type >= MAX_SWAPFILES) {
2850		spin_unlock(&swap_lock);
2851		kfree(p);
2852		return ERR_PTR(-EPERM);
2853	}
2854	if (type >= nr_swapfiles) {
2855		p->type = type;
2856		swap_info[type] = p;
2857		/*
2858		 * Write swap_info[type] before nr_swapfiles, in case a
2859		 * racing procfs swap_start() or swap_next() is reading them.
2860		 * (We never shrink nr_swapfiles, we never free this entry.)
2861		 */
2862		smp_wmb();
2863		nr_swapfiles++;
2864	} else {
2865		kfree(p);
2866		p = swap_info[type];
2867		/*
2868		 * Do not memset this entry: a racing procfs swap_next()
2869		 * would be relying on p->type to remain valid.
2870		 */
2871	}
2872	INIT_LIST_HEAD(&p->first_swap_extent.list);
2873	plist_node_init(&p->list, 0);
2874	for_each_node(i)
2875		plist_node_init(&p->avail_lists[i], 0);
2876	p->flags = SWP_USED;
 
2877	spin_unlock(&swap_lock);
2878	spin_lock_init(&p->lock);
2879	spin_lock_init(&p->cont_lock);
2880
2881	return p;
2882}
2883
2884static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2885{
2886	int error;
2887
2888	if (S_ISBLK(inode->i_mode)) {
2889		p->bdev = bdgrab(I_BDEV(inode));
2890		error = blkdev_get(p->bdev,
2891				   FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
 
2892		if (error < 0) {
2893			p->bdev = NULL;
2894			return error;
2895		}
2896		p->old_block_size = block_size(p->bdev);
2897		error = set_blocksize(p->bdev, PAGE_SIZE);
2898		if (error < 0)
2899			return error;
2900		p->flags |= SWP_BLKDEV;
2901	} else if (S_ISREG(inode->i_mode)) {
2902		p->bdev = inode->i_sb->s_bdev;
2903		inode_lock(inode);
2904		if (IS_SWAPFILE(inode))
2905			return -EBUSY;
2906	} else
2907		return -EINVAL;
2908
2909	return 0;
2910}
2911
2912static unsigned long read_swap_header(struct swap_info_struct *p,
2913					union swap_header *swap_header,
2914					struct inode *inode)
2915{
2916	int i;
2917	unsigned long maxpages;
2918	unsigned long swapfilepages;
2919	unsigned long last_page;
2920
2921	if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2922		pr_err("Unable to find swap-space signature\n");
2923		return 0;
2924	}
2925
2926	/* swap partition endianess hack... */
2927	if (swab32(swap_header->info.version) == 1) {
2928		swab32s(&swap_header->info.version);
2929		swab32s(&swap_header->info.last_page);
2930		swab32s(&swap_header->info.nr_badpages);
2931		if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2932			return 0;
2933		for (i = 0; i < swap_header->info.nr_badpages; i++)
2934			swab32s(&swap_header->info.badpages[i]);
2935	}
2936	/* Check the swap header's sub-version */
2937	if (swap_header->info.version != 1) {
2938		pr_warn("Unable to handle swap header version %d\n",
2939			swap_header->info.version);
 
2940		return 0;
2941	}
2942
2943	p->lowest_bit  = 1;
2944	p->cluster_next = 1;
2945	p->cluster_nr = 0;
2946
2947	/*
2948	 * Find out how many pages are allowed for a single swap
2949	 * device. There are two limiting factors: 1) the number
2950	 * of bits for the swap offset in the swp_entry_t type, and
2951	 * 2) the number of bits in the swap pte as defined by the
2952	 * different architectures. In order to find the
 
2953	 * largest possible bit mask, a swap entry with swap type 0
2954	 * and swap offset ~0UL is created, encoded to a swap pte,
2955	 * decoded to a swp_entry_t again, and finally the swap
2956	 * offset is extracted. This will mask all the bits from
2957	 * the initial ~0UL mask that can't be encoded in either
2958	 * the swp_entry_t or the architecture definition of a
2959	 * swap pte.
2960	 */
2961	maxpages = swp_offset(pte_to_swp_entry(
2962			swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2963	last_page = swap_header->info.last_page;
2964	if (!last_page) {
2965		pr_warn("Empty swap-file\n");
2966		return 0;
2967	}
2968	if (last_page > maxpages) {
2969		pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2970			maxpages << (PAGE_SHIFT - 10),
2971			last_page << (PAGE_SHIFT - 10));
2972	}
2973	if (maxpages > last_page) {
2974		maxpages = last_page + 1;
2975		/* p->max is an unsigned int: don't overflow it */
2976		if ((unsigned int)maxpages == 0)
2977			maxpages = UINT_MAX;
2978	}
2979	p->highest_bit = maxpages - 1;
2980
2981	if (!maxpages)
2982		return 0;
2983	swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2984	if (swapfilepages && maxpages > swapfilepages) {
2985		pr_warn("Swap area shorter than signature indicates\n");
 
2986		return 0;
2987	}
2988	if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2989		return 0;
2990	if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2991		return 0;
2992
2993	return maxpages;
2994}
2995
2996#define SWAP_CLUSTER_INFO_COLS						\
2997	DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2998#define SWAP_CLUSTER_SPACE_COLS						\
2999	DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3000#define SWAP_CLUSTER_COLS						\
3001	max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3002
3003static int setup_swap_map_and_extents(struct swap_info_struct *p,
3004					union swap_header *swap_header,
3005					unsigned char *swap_map,
3006					struct swap_cluster_info *cluster_info,
3007					unsigned long maxpages,
3008					sector_t *span)
3009{
3010	unsigned int j, k;
3011	unsigned int nr_good_pages;
3012	int nr_extents;
3013	unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3014	unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3015	unsigned long i, idx;
3016
3017	nr_good_pages = maxpages - 1;	/* omit header page */
3018
3019	cluster_list_init(&p->free_clusters);
3020	cluster_list_init(&p->discard_clusters);
3021
3022	for (i = 0; i < swap_header->info.nr_badpages; i++) {
3023		unsigned int page_nr = swap_header->info.badpages[i];
3024		if (page_nr == 0 || page_nr > swap_header->info.last_page)
3025			return -EINVAL;
3026		if (page_nr < maxpages) {
3027			swap_map[page_nr] = SWAP_MAP_BAD;
3028			nr_good_pages--;
3029			/*
3030			 * Haven't marked the cluster free yet, no list
3031			 * operation involved
3032			 */
3033			inc_cluster_info_page(p, cluster_info, page_nr);
3034		}
3035	}
3036
3037	/* Haven't marked the cluster free yet, no list operation involved */
3038	for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3039		inc_cluster_info_page(p, cluster_info, i);
3040
3041	if (nr_good_pages) {
3042		swap_map[0] = SWAP_MAP_BAD;
3043		/*
3044		 * Not mark the cluster free yet, no list
3045		 * operation involved
3046		 */
3047		inc_cluster_info_page(p, cluster_info, 0);
3048		p->max = maxpages;
3049		p->pages = nr_good_pages;
3050		nr_extents = setup_swap_extents(p, span);
3051		if (nr_extents < 0)
3052			return nr_extents;
3053		nr_good_pages = p->pages;
3054	}
3055	if (!nr_good_pages) {
3056		pr_warn("Empty swap-file\n");
3057		return -EINVAL;
3058	}
3059
3060	if (!cluster_info)
3061		return nr_extents;
3062
3063
3064	/*
3065	 * Reduce false cache line sharing between cluster_info and
3066	 * sharing same address space.
3067	 */
3068	for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3069		j = (k + col) % SWAP_CLUSTER_COLS;
3070		for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3071			idx = i * SWAP_CLUSTER_COLS + j;
3072			if (idx >= nr_clusters)
3073				continue;
3074			if (cluster_count(&cluster_info[idx]))
3075				continue;
3076			cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3077			cluster_list_add_tail(&p->free_clusters, cluster_info,
3078					      idx);
3079		}
3080	}
3081	return nr_extents;
3082}
3083
3084/*
3085 * Helper to sys_swapon determining if a given swap
3086 * backing device queue supports DISCARD operations.
3087 */
3088static bool swap_discardable(struct swap_info_struct *si)
3089{
3090	struct request_queue *q = bdev_get_queue(si->bdev);
3091
3092	if (!q || !blk_queue_discard(q))
3093		return false;
3094
3095	return true;
3096}
3097
3098SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3099{
3100	struct swap_info_struct *p;
3101	struct filename *name;
3102	struct file *swap_file = NULL;
3103	struct address_space *mapping;
 
3104	int prio;
3105	int error;
3106	union swap_header *swap_header;
3107	int nr_extents;
3108	sector_t span;
3109	unsigned long maxpages;
3110	unsigned char *swap_map = NULL;
3111	struct swap_cluster_info *cluster_info = NULL;
3112	unsigned long *frontswap_map = NULL;
3113	struct page *page = NULL;
3114	struct inode *inode = NULL;
3115	bool inced_nr_rotate_swap = false;
3116
3117	if (swap_flags & ~SWAP_FLAGS_VALID)
3118		return -EINVAL;
3119
3120	if (!capable(CAP_SYS_ADMIN))
3121		return -EPERM;
3122
3123	if (!swap_avail_heads)
3124		return -ENOMEM;
3125
3126	p = alloc_swap_info();
3127	if (IS_ERR(p))
3128		return PTR_ERR(p);
3129
3130	INIT_WORK(&p->discard_work, swap_discard_work);
3131
3132	name = getname(specialfile);
3133	if (IS_ERR(name)) {
3134		error = PTR_ERR(name);
3135		name = NULL;
3136		goto bad_swap;
3137	}
3138	swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3139	if (IS_ERR(swap_file)) {
3140		error = PTR_ERR(swap_file);
3141		swap_file = NULL;
3142		goto bad_swap;
3143	}
3144
3145	p->swap_file = swap_file;
3146	mapping = swap_file->f_mapping;
 
 
 
 
 
 
 
 
 
 
 
 
3147	inode = mapping->host;
3148
3149	/* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3150	error = claim_swapfile(p, inode);
3151	if (unlikely(error))
3152		goto bad_swap;
3153
3154	/*
3155	 * Read the swap header.
3156	 */
3157	if (!mapping->a_ops->readpage) {
3158		error = -EINVAL;
3159		goto bad_swap;
3160	}
3161	page = read_mapping_page(mapping, 0, swap_file);
3162	if (IS_ERR(page)) {
3163		error = PTR_ERR(page);
3164		goto bad_swap;
3165	}
3166	swap_header = kmap(page);
3167
3168	maxpages = read_swap_header(p, swap_header, inode);
3169	if (unlikely(!maxpages)) {
3170		error = -EINVAL;
3171		goto bad_swap;
3172	}
3173
3174	/* OK, set up the swap map and apply the bad block list */
3175	swap_map = vzalloc(maxpages);
3176	if (!swap_map) {
3177		error = -ENOMEM;
3178		goto bad_swap;
3179	}
3180
3181	if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3182		p->flags |= SWP_STABLE_WRITES;
3183
3184	if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3185		p->flags |= SWP_SYNCHRONOUS_IO;
3186
3187	if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3188		int cpu;
3189		unsigned long ci, nr_cluster;
3190
3191		p->flags |= SWP_SOLIDSTATE;
3192		/*
3193		 * select a random position to start with to help wear leveling
3194		 * SSD
3195		 */
3196		p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3197		nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3198
3199		cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3200					GFP_KERNEL);
3201		if (!cluster_info) {
3202			error = -ENOMEM;
3203			goto bad_swap;
3204		}
3205
3206		for (ci = 0; ci < nr_cluster; ci++)
3207			spin_lock_init(&((cluster_info + ci)->lock));
3208
3209		p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3210		if (!p->percpu_cluster) {
3211			error = -ENOMEM;
3212			goto bad_swap;
3213		}
3214		for_each_possible_cpu(cpu) {
3215			struct percpu_cluster *cluster;
3216			cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3217			cluster_set_null(&cluster->index);
3218		}
3219	} else {
3220		atomic_inc(&nr_rotate_swap);
3221		inced_nr_rotate_swap = true;
3222	}
3223
3224	error = swap_cgroup_swapon(p->type, maxpages);
3225	if (error)
3226		goto bad_swap;
3227
3228	nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3229		cluster_info, maxpages, &span);
3230	if (unlikely(nr_extents < 0)) {
3231		error = nr_extents;
3232		goto bad_swap;
3233	}
3234	/* frontswap enabled? set up bit-per-page map for frontswap */
3235	if (IS_ENABLED(CONFIG_FRONTSWAP))
3236		frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3237					 GFP_KERNEL);
3238
3239	if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3240		/*
3241		 * When discard is enabled for swap with no particular
3242		 * policy flagged, we set all swap discard flags here in
3243		 * order to sustain backward compatibility with older
3244		 * swapon(8) releases.
3245		 */
3246		p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3247			     SWP_PAGE_DISCARD);
3248
3249		/*
3250		 * By flagging sys_swapon, a sysadmin can tell us to
3251		 * either do single-time area discards only, or to just
3252		 * perform discards for released swap page-clusters.
3253		 * Now it's time to adjust the p->flags accordingly.
3254		 */
3255		if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3256			p->flags &= ~SWP_PAGE_DISCARD;
3257		else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3258			p->flags &= ~SWP_AREA_DISCARD;
3259
3260		/* issue a swapon-time discard if it's still required */
3261		if (p->flags & SWP_AREA_DISCARD) {
3262			int err = discard_swap(p);
3263			if (unlikely(err))
3264				pr_err("swapon: discard_swap(%p): %d\n",
3265					p, err);
3266		}
 
 
3267	}
3268
3269	error = init_swap_address_space(p->type, maxpages);
3270	if (error)
3271		goto bad_swap;
3272
3273	mutex_lock(&swapon_mutex);
3274	prio = -1;
3275	if (swap_flags & SWAP_FLAG_PREFER)
3276		prio =
3277		  (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3278	enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3279
3280	pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3281		p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
 
3282		nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3283		(p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3284		(p->flags & SWP_DISCARDABLE) ? "D" : "",
3285		(p->flags & SWP_AREA_DISCARD) ? "s" : "",
3286		(p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3287		(frontswap_map) ? "FS" : "");
3288
3289	mutex_unlock(&swapon_mutex);
3290	atomic_inc(&proc_poll_event);
3291	wake_up_interruptible(&proc_poll_wait);
3292
3293	if (S_ISREG(inode->i_mode))
3294		inode->i_flags |= S_SWAPFILE;
3295	error = 0;
3296	goto out;
3297bad_swap:
3298	free_percpu(p->percpu_cluster);
3299	p->percpu_cluster = NULL;
3300	if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3301		set_blocksize(p->bdev, p->old_block_size);
3302		blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3303	}
3304	destroy_swap_extents(p);
3305	swap_cgroup_swapoff(p->type);
3306	spin_lock(&swap_lock);
3307	p->swap_file = NULL;
3308	p->flags = 0;
3309	spin_unlock(&swap_lock);
3310	vfree(swap_map);
3311	kvfree(cluster_info);
3312	kvfree(frontswap_map);
3313	if (inced_nr_rotate_swap)
3314		atomic_dec(&nr_rotate_swap);
3315	if (swap_file) {
3316		if (inode && S_ISREG(inode->i_mode)) {
3317			inode_unlock(inode);
3318			inode = NULL;
3319		}
3320		filp_close(swap_file, NULL);
3321	}
3322out:
3323	if (page && !IS_ERR(page)) {
3324		kunmap(page);
3325		put_page(page);
3326	}
3327	if (name)
3328		putname(name);
3329	if (inode && S_ISREG(inode->i_mode))
3330		inode_unlock(inode);
3331	if (!error)
3332		enable_swap_slots_cache();
3333	return error;
3334}
3335
3336void si_swapinfo(struct sysinfo *val)
3337{
3338	unsigned int type;
3339	unsigned long nr_to_be_unused = 0;
3340
3341	spin_lock(&swap_lock);
3342	for (type = 0; type < nr_swapfiles; type++) {
3343		struct swap_info_struct *si = swap_info[type];
3344
3345		if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3346			nr_to_be_unused += si->inuse_pages;
3347	}
3348	val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3349	val->totalswap = total_swap_pages + nr_to_be_unused;
3350	spin_unlock(&swap_lock);
3351}
3352
3353/*
3354 * Verify that a swap entry is valid and increment its swap map count.
3355 *
3356 * Returns error code in following case.
3357 * - success -> 0
3358 * - swp_entry is invalid -> EINVAL
3359 * - swp_entry is migration entry -> EINVAL
3360 * - swap-cache reference is requested but there is already one. -> EEXIST
3361 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3362 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3363 */
3364static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3365{
3366	struct swap_info_struct *p;
3367	struct swap_cluster_info *ci;
3368	unsigned long offset, type;
3369	unsigned char count;
3370	unsigned char has_cache;
3371	int err = -EINVAL;
3372
3373	if (non_swap_entry(entry))
3374		goto out;
3375
3376	type = swp_type(entry);
3377	if (type >= nr_swapfiles)
3378		goto bad_file;
3379	p = swap_info[type];
3380	offset = swp_offset(entry);
 
 
3381	if (unlikely(offset >= p->max))
3382		goto out;
3383
3384	ci = lock_cluster_or_swap_info(p, offset);
3385
3386	count = p->swap_map[offset];
3387
3388	/*
3389	 * swapin_readahead() doesn't check if a swap entry is valid, so the
3390	 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3391	 */
3392	if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3393		err = -ENOENT;
3394		goto unlock_out;
3395	}
3396
3397	has_cache = count & SWAP_HAS_CACHE;
3398	count &= ~SWAP_HAS_CACHE;
3399	err = 0;
3400
3401	if (usage == SWAP_HAS_CACHE) {
3402
3403		/* set SWAP_HAS_CACHE if there is no cache and entry is used */
3404		if (!has_cache && count)
3405			has_cache = SWAP_HAS_CACHE;
3406		else if (has_cache)		/* someone else added cache */
3407			err = -EEXIST;
3408		else				/* no users remaining */
3409			err = -ENOENT;
3410
3411	} else if (count || has_cache) {
3412
3413		if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3414			count += usage;
3415		else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3416			err = -EINVAL;
3417		else if (swap_count_continued(p, offset, count))
3418			count = COUNT_CONTINUED;
3419		else
3420			err = -ENOMEM;
3421	} else
3422		err = -ENOENT;			/* unused swap entry */
3423
3424	p->swap_map[offset] = count | has_cache;
3425
3426unlock_out:
3427	unlock_cluster_or_swap_info(p, ci);
3428out:
3429	return err;
3430
3431bad_file:
3432	pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3433	goto out;
3434}
3435
3436/*
3437 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3438 * (in which case its reference count is never incremented).
3439 */
3440void swap_shmem_alloc(swp_entry_t entry)
3441{
3442	__swap_duplicate(entry, SWAP_MAP_SHMEM);
3443}
3444
3445/*
3446 * Increase reference count of swap entry by 1.
3447 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3448 * but could not be atomically allocated.  Returns 0, just as if it succeeded,
3449 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3450 * might occur if a page table entry has got corrupted.
3451 */
3452int swap_duplicate(swp_entry_t entry)
3453{
3454	int err = 0;
3455
3456	while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3457		err = add_swap_count_continuation(entry, GFP_ATOMIC);
3458	return err;
3459}
3460
3461/*
3462 * @entry: swap entry for which we allocate swap cache.
3463 *
3464 * Called when allocating swap cache for existing swap entry,
3465 * This can return error codes. Returns 0 at success.
3466 * -EBUSY means there is a swap cache.
3467 * Note: return code is different from swap_duplicate().
3468 */
3469int swapcache_prepare(swp_entry_t entry)
3470{
3471	return __swap_duplicate(entry, SWAP_HAS_CACHE);
3472}
3473
3474struct swap_info_struct *swp_swap_info(swp_entry_t entry)
 
 
 
 
3475{
3476	return swap_info[swp_type(entry)];
3477}
 
 
 
 
 
 
 
 
 
 
 
 
 
3478
3479struct swap_info_struct *page_swap_info(struct page *page)
3480{
3481	swp_entry_t entry = { .val = page_private(page) };
3482	return swp_swap_info(entry);
3483}
3484
3485/*
3486 * out-of-line __page_file_ methods to avoid include hell.
3487 */
3488struct address_space *__page_file_mapping(struct page *page)
3489{
3490	return page_swap_info(page)->swap_file->f_mapping;
3491}
3492EXPORT_SYMBOL_GPL(__page_file_mapping);
 
 
 
 
 
 
 
 
 
3493
3494pgoff_t __page_file_index(struct page *page)
3495{
3496	swp_entry_t swap = { .val = page_private(page) };
3497	return swp_offset(swap);
 
 
3498}
3499EXPORT_SYMBOL_GPL(__page_file_index);
3500
3501/*
3502 * add_swap_count_continuation - called when a swap count is duplicated
3503 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3504 * page of the original vmalloc'ed swap_map, to hold the continuation count
3505 * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
3506 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3507 *
3508 * These continuation pages are seldom referenced: the common paths all work
3509 * on the original swap_map, only referring to a continuation page when the
3510 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3511 *
3512 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3513 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3514 * can be called after dropping locks.
3515 */
3516int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3517{
3518	struct swap_info_struct *si;
3519	struct swap_cluster_info *ci;
3520	struct page *head;
3521	struct page *page;
3522	struct page *list_page;
3523	pgoff_t offset;
3524	unsigned char count;
3525
3526	/*
3527	 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3528	 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3529	 */
3530	page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3531
3532	si = swap_info_get(entry);
3533	if (!si) {
3534		/*
3535		 * An acceptable race has occurred since the failing
3536		 * __swap_duplicate(): the swap entry has been freed,
3537		 * perhaps even the whole swap_map cleared for swapoff.
3538		 */
3539		goto outer;
3540	}
3541
3542	offset = swp_offset(entry);
3543
3544	ci = lock_cluster(si, offset);
3545
3546	count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3547
3548	if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3549		/*
3550		 * The higher the swap count, the more likely it is that tasks
3551		 * will race to add swap count continuation: we need to avoid
3552		 * over-provisioning.
3553		 */
3554		goto out;
3555	}
3556
3557	if (!page) {
3558		unlock_cluster(ci);
3559		spin_unlock(&si->lock);
3560		return -ENOMEM;
3561	}
3562
3563	/*
3564	 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3565	 * no architecture is using highmem pages for kernel page tables: so it
3566	 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3567	 */
3568	head = vmalloc_to_page(si->swap_map + offset);
3569	offset &= ~PAGE_MASK;
3570
3571	spin_lock(&si->cont_lock);
3572	/*
3573	 * Page allocation does not initialize the page's lru field,
3574	 * but it does always reset its private field.
3575	 */
3576	if (!page_private(head)) {
3577		BUG_ON(count & COUNT_CONTINUED);
3578		INIT_LIST_HEAD(&head->lru);
3579		set_page_private(head, SWP_CONTINUED);
3580		si->flags |= SWP_CONTINUED;
3581	}
3582
3583	list_for_each_entry(list_page, &head->lru, lru) {
3584		unsigned char *map;
3585
3586		/*
3587		 * If the previous map said no continuation, but we've found
3588		 * a continuation page, free our allocation and use this one.
3589		 */
3590		if (!(count & COUNT_CONTINUED))
3591			goto out_unlock_cont;
3592
3593		map = kmap_atomic(list_page) + offset;
3594		count = *map;
3595		kunmap_atomic(map);
3596
3597		/*
3598		 * If this continuation count now has some space in it,
3599		 * free our allocation and use this one.
3600		 */
3601		if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3602			goto out_unlock_cont;
3603	}
3604
3605	list_add_tail(&page->lru, &head->lru);
3606	page = NULL;			/* now it's attached, don't free it */
3607out_unlock_cont:
3608	spin_unlock(&si->cont_lock);
3609out:
3610	unlock_cluster(ci);
3611	spin_unlock(&si->lock);
3612outer:
3613	if (page)
3614		__free_page(page);
3615	return 0;
3616}
3617
3618/*
3619 * swap_count_continued - when the original swap_map count is incremented
3620 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3621 * into, carry if so, or else fail until a new continuation page is allocated;
3622 * when the original swap_map count is decremented from 0 with continuation,
3623 * borrow from the continuation and report whether it still holds more.
3624 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3625 * lock.
3626 */
3627static bool swap_count_continued(struct swap_info_struct *si,
3628				 pgoff_t offset, unsigned char count)
3629{
3630	struct page *head;
3631	struct page *page;
3632	unsigned char *map;
3633	bool ret;
3634
3635	head = vmalloc_to_page(si->swap_map + offset);
3636	if (page_private(head) != SWP_CONTINUED) {
3637		BUG_ON(count & COUNT_CONTINUED);
3638		return false;		/* need to add count continuation */
3639	}
3640
3641	spin_lock(&si->cont_lock);
3642	offset &= ~PAGE_MASK;
3643	page = list_entry(head->lru.next, struct page, lru);
3644	map = kmap_atomic(page) + offset;
3645
3646	if (count == SWAP_MAP_MAX)	/* initial increment from swap_map */
3647		goto init_map;		/* jump over SWAP_CONT_MAX checks */
3648
3649	if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3650		/*
3651		 * Think of how you add 1 to 999
3652		 */
3653		while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3654			kunmap_atomic(map);
3655			page = list_entry(page->lru.next, struct page, lru);
3656			BUG_ON(page == head);
3657			map = kmap_atomic(page) + offset;
3658		}
3659		if (*map == SWAP_CONT_MAX) {
3660			kunmap_atomic(map);
3661			page = list_entry(page->lru.next, struct page, lru);
3662			if (page == head) {
3663				ret = false;	/* add count continuation */
3664				goto out;
3665			}
3666			map = kmap_atomic(page) + offset;
3667init_map:		*map = 0;		/* we didn't zero the page */
3668		}
3669		*map += 1;
3670		kunmap_atomic(map);
3671		page = list_entry(page->lru.prev, struct page, lru);
3672		while (page != head) {
3673			map = kmap_atomic(page) + offset;
3674			*map = COUNT_CONTINUED;
3675			kunmap_atomic(map);
3676			page = list_entry(page->lru.prev, struct page, lru);
3677		}
3678		ret = true;			/* incremented */
3679
3680	} else {				/* decrementing */
3681		/*
3682		 * Think of how you subtract 1 from 1000
3683		 */
3684		BUG_ON(count != COUNT_CONTINUED);
3685		while (*map == COUNT_CONTINUED) {
3686			kunmap_atomic(map);
3687			page = list_entry(page->lru.next, struct page, lru);
3688			BUG_ON(page == head);
3689			map = kmap_atomic(page) + offset;
3690		}
3691		BUG_ON(*map == 0);
3692		*map -= 1;
3693		if (*map == 0)
3694			count = 0;
3695		kunmap_atomic(map);
3696		page = list_entry(page->lru.prev, struct page, lru);
3697		while (page != head) {
3698			map = kmap_atomic(page) + offset;
3699			*map = SWAP_CONT_MAX | count;
3700			count = COUNT_CONTINUED;
3701			kunmap_atomic(map);
3702			page = list_entry(page->lru.prev, struct page, lru);
3703		}
3704		ret = count == COUNT_CONTINUED;
3705	}
3706out:
3707	spin_unlock(&si->cont_lock);
3708	return ret;
3709}
3710
3711/*
3712 * free_swap_count_continuations - swapoff free all the continuation pages
3713 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3714 */
3715static void free_swap_count_continuations(struct swap_info_struct *si)
3716{
3717	pgoff_t offset;
3718
3719	for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3720		struct page *head;
3721		head = vmalloc_to_page(si->swap_map + offset);
3722		if (page_private(head)) {
3723			struct page *page, *next;
3724
3725			list_for_each_entry_safe(page, next, &head->lru, lru) {
3726				list_del(&page->lru);
 
3727				__free_page(page);
3728			}
3729		}
3730	}
3731}
3732
3733static int __init swapfile_init(void)
3734{
3735	int nid;
3736
3737	swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3738					 GFP_KERNEL);
3739	if (!swap_avail_heads) {
3740		pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3741		return -ENOMEM;
3742	}
3743
3744	for_each_node(nid)
3745		plist_head_init(&swap_avail_heads[nid]);
3746
3747	return 0;
3748}
3749subsys_initcall(swapfile_init);