Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/swapfile.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 *  Swap reorganised 29.12.95, Stephen Tweedie
   7 */
   8
   9#include <linux/blkdev.h>
  10#include <linux/mm.h>
  11#include <linux/sched/mm.h>
  12#include <linux/sched/task.h>
  13#include <linux/hugetlb.h>
  14#include <linux/mman.h>
  15#include <linux/slab.h>
  16#include <linux/kernel_stat.h>
  17#include <linux/swap.h>
  18#include <linux/vmalloc.h>
  19#include <linux/pagemap.h>
  20#include <linux/namei.h>
  21#include <linux/shmem_fs.h>
  22#include <linux/blk-cgroup.h>
  23#include <linux/random.h>
  24#include <linux/writeback.h>
  25#include <linux/proc_fs.h>
  26#include <linux/seq_file.h>
  27#include <linux/init.h>
 
  28#include <linux/ksm.h>
  29#include <linux/rmap.h>
  30#include <linux/security.h>
  31#include <linux/backing-dev.h>
  32#include <linux/mutex.h>
  33#include <linux/capability.h>
  34#include <linux/syscalls.h>
  35#include <linux/memcontrol.h>
  36#include <linux/poll.h>
  37#include <linux/oom.h>
  38#include <linux/frontswap.h>
  39#include <linux/swapfile.h>
  40#include <linux/export.h>
  41#include <linux/swap_slots.h>
  42#include <linux/sort.h>
  43#include <linux/completion.h>
  44
 
  45#include <asm/tlbflush.h>
  46#include <linux/swapops.h>
  47#include <linux/swap_cgroup.h>
  48#include "swap.h"
  49
  50static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
  51				 unsigned char);
  52static void free_swap_count_continuations(struct swap_info_struct *);
 
  53
  54static DEFINE_SPINLOCK(swap_lock);
  55static unsigned int nr_swapfiles;
  56atomic_long_t nr_swap_pages;
  57/*
  58 * Some modules use swappable objects and may try to swap them out under
  59 * memory pressure (via the shrinker). Before doing so, they may wish to
  60 * check to see if any swap space is available.
  61 */
  62EXPORT_SYMBOL_GPL(nr_swap_pages);
  63/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
  64long total_swap_pages;
  65static int least_priority = -1;
  66unsigned long swapfile_maximum_size;
  67#ifdef CONFIG_MIGRATION
  68bool swap_migration_ad_supported;
  69#endif	/* CONFIG_MIGRATION */
  70
  71static const char Bad_file[] = "Bad swap file entry ";
  72static const char Unused_file[] = "Unused swap file entry ";
  73static const char Bad_offset[] = "Bad swap offset entry ";
  74static const char Unused_offset[] = "Unused swap offset entry ";
  75
  76/*
  77 * all active swap_info_structs
  78 * protected with swap_lock, and ordered by priority.
  79 */
  80static PLIST_HEAD(swap_active_head);
  81
  82/*
  83 * all available (active, not full) swap_info_structs
  84 * protected with swap_avail_lock, ordered by priority.
  85 * This is used by folio_alloc_swap() instead of swap_active_head
  86 * because swap_active_head includes all swap_info_structs,
  87 * but folio_alloc_swap() doesn't need to look at full ones.
  88 * This uses its own lock instead of swap_lock because when a
  89 * swap_info_struct changes between not-full/full, it needs to
  90 * add/remove itself to/from this list, but the swap_info_struct->lock
  91 * is held and the locking order requires swap_lock to be taken
  92 * before any swap_info_struct->lock.
  93 */
  94static struct plist_head *swap_avail_heads;
  95static DEFINE_SPINLOCK(swap_avail_lock);
  96
  97struct swap_info_struct *swap_info[MAX_SWAPFILES];
  98
  99static DEFINE_MUTEX(swapon_mutex);
 100
 101static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
 102/* Activity counter to indicate that a swapon or swapoff has occurred */
 103static atomic_t proc_poll_event = ATOMIC_INIT(0);
 104
 105atomic_t nr_rotate_swap = ATOMIC_INIT(0);
 106
 107static struct swap_info_struct *swap_type_to_swap_info(int type)
 108{
 109	if (type >= MAX_SWAPFILES)
 110		return NULL;
 111
 112	return READ_ONCE(swap_info[type]); /* rcu_dereference() */
 113}
 114
 115static inline unsigned char swap_count(unsigned char ent)
 116{
 117	return ent & ~SWAP_HAS_CACHE;	/* may include COUNT_CONTINUED flag */
 118}
 119
 120/* Reclaim the swap entry anyway if possible */
 121#define TTRS_ANYWAY		0x1
 122/*
 123 * Reclaim the swap entry if there are no more mappings of the
 124 * corresponding page
 125 */
 126#define TTRS_UNMAPPED		0x2
 127/* Reclaim the swap entry if swap is getting full*/
 128#define TTRS_FULL		0x4
 129
 130/* returns 1 if swap entry is freed */
 131static int __try_to_reclaim_swap(struct swap_info_struct *si,
 132				 unsigned long offset, unsigned long flags)
 133{
 134	swp_entry_t entry = swp_entry(si->type, offset);
 135	struct folio *folio;
 136	int ret = 0;
 137
 138	folio = filemap_get_folio(swap_address_space(entry), offset);
 139	if (!folio)
 140		return 0;
 141	/*
 142	 * When this function is called from scan_swap_map_slots() and it's
 143	 * called by vmscan.c at reclaiming folios. So we hold a folio lock
 144	 * here. We have to use trylock for avoiding deadlock. This is a special
 145	 * case and you should use folio_free_swap() with explicit folio_lock()
 146	 * in usual operations.
 147	 */
 148	if (folio_trylock(folio)) {
 149		if ((flags & TTRS_ANYWAY) ||
 150		    ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
 151		    ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)))
 152			ret = folio_free_swap(folio);
 153		folio_unlock(folio);
 154	}
 155	folio_put(folio);
 156	return ret;
 157}
 158
 159static inline struct swap_extent *first_se(struct swap_info_struct *sis)
 160{
 161	struct rb_node *rb = rb_first(&sis->swap_extent_root);
 162	return rb_entry(rb, struct swap_extent, rb_node);
 163}
 164
 165static inline struct swap_extent *next_se(struct swap_extent *se)
 166{
 167	struct rb_node *rb = rb_next(&se->rb_node);
 168	return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
 169}
 170
 171/*
 172 * swapon tell device that all the old swap contents can be discarded,
 173 * to allow the swap device to optimize its wear-levelling.
 174 */
 175static int discard_swap(struct swap_info_struct *si)
 176{
 177	struct swap_extent *se;
 178	sector_t start_block;
 179	sector_t nr_blocks;
 180	int err = 0;
 181
 182	/* Do not discard the swap header page! */
 183	se = first_se(si);
 184	start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 185	nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 186	if (nr_blocks) {
 187		err = blkdev_issue_discard(si->bdev, start_block,
 188				nr_blocks, GFP_KERNEL);
 189		if (err)
 190			return err;
 191		cond_resched();
 192	}
 193
 194	for (se = next_se(se); se; se = next_se(se)) {
 195		start_block = se->start_block << (PAGE_SHIFT - 9);
 196		nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 197
 198		err = blkdev_issue_discard(si->bdev, start_block,
 199				nr_blocks, GFP_KERNEL);
 200		if (err)
 201			break;
 202
 203		cond_resched();
 204	}
 205	return err;		/* That will often be -EOPNOTSUPP */
 206}
 207
 208static struct swap_extent *
 209offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
 210{
 211	struct swap_extent *se;
 212	struct rb_node *rb;
 213
 214	rb = sis->swap_extent_root.rb_node;
 215	while (rb) {
 216		se = rb_entry(rb, struct swap_extent, rb_node);
 217		if (offset < se->start_page)
 218			rb = rb->rb_left;
 219		else if (offset >= se->start_page + se->nr_pages)
 220			rb = rb->rb_right;
 221		else
 222			return se;
 223	}
 224	/* It *must* be present */
 225	BUG();
 226}
 227
 228sector_t swap_page_sector(struct page *page)
 229{
 230	struct swap_info_struct *sis = page_swap_info(page);
 231	struct swap_extent *se;
 232	sector_t sector;
 233	pgoff_t offset;
 234
 235	offset = __page_file_index(page);
 236	se = offset_to_swap_extent(sis, offset);
 237	sector = se->start_block + (offset - se->start_page);
 238	return sector << (PAGE_SHIFT - 9);
 239}
 240
 241/*
 242 * swap allocation tell device that a cluster of swap can now be discarded,
 243 * to allow the swap device to optimize its wear-levelling.
 244 */
 245static void discard_swap_cluster(struct swap_info_struct *si,
 246				 pgoff_t start_page, pgoff_t nr_pages)
 247{
 248	struct swap_extent *se = offset_to_swap_extent(si, start_page);
 
 249
 250	while (nr_pages) {
 251		pgoff_t offset = start_page - se->start_page;
 252		sector_t start_block = se->start_block + offset;
 253		sector_t nr_blocks = se->nr_pages - offset;
 254
 255		if (nr_blocks > nr_pages)
 256			nr_blocks = nr_pages;
 257		start_page += nr_blocks;
 258		nr_pages -= nr_blocks;
 259
 260		start_block <<= PAGE_SHIFT - 9;
 261		nr_blocks <<= PAGE_SHIFT - 9;
 262		if (blkdev_issue_discard(si->bdev, start_block,
 263					nr_blocks, GFP_NOIO))
 264			break;
 265
 266		se = next_se(se);
 267	}
 268}
 269
 270#ifdef CONFIG_THP_SWAP
 271#define SWAPFILE_CLUSTER	HPAGE_PMD_NR
 272
 273#define swap_entry_size(size)	(size)
 274#else
 275#define SWAPFILE_CLUSTER	256
 276
 277/*
 278 * Define swap_entry_size() as constant to let compiler to optimize
 279 * out some code if !CONFIG_THP_SWAP
 280 */
 281#define swap_entry_size(size)	1
 282#endif
 283#define LATENCY_LIMIT		256
 284
 285static inline void cluster_set_flag(struct swap_cluster_info *info,
 286	unsigned int flag)
 287{
 288	info->flags = flag;
 289}
 290
 291static inline unsigned int cluster_count(struct swap_cluster_info *info)
 292{
 293	return info->data;
 294}
 295
 296static inline void cluster_set_count(struct swap_cluster_info *info,
 297				     unsigned int c)
 298{
 299	info->data = c;
 300}
 301
 302static inline void cluster_set_count_flag(struct swap_cluster_info *info,
 303					 unsigned int c, unsigned int f)
 304{
 305	info->flags = f;
 306	info->data = c;
 307}
 308
 309static inline unsigned int cluster_next(struct swap_cluster_info *info)
 310{
 311	return info->data;
 312}
 313
 314static inline void cluster_set_next(struct swap_cluster_info *info,
 315				    unsigned int n)
 316{
 317	info->data = n;
 318}
 319
 320static inline void cluster_set_next_flag(struct swap_cluster_info *info,
 321					 unsigned int n, unsigned int f)
 322{
 323	info->flags = f;
 324	info->data = n;
 325}
 326
 327static inline bool cluster_is_free(struct swap_cluster_info *info)
 328{
 329	return info->flags & CLUSTER_FLAG_FREE;
 330}
 331
 332static inline bool cluster_is_null(struct swap_cluster_info *info)
 333{
 334	return info->flags & CLUSTER_FLAG_NEXT_NULL;
 335}
 336
 337static inline void cluster_set_null(struct swap_cluster_info *info)
 338{
 339	info->flags = CLUSTER_FLAG_NEXT_NULL;
 340	info->data = 0;
 341}
 342
 343static inline bool cluster_is_huge(struct swap_cluster_info *info)
 344{
 345	if (IS_ENABLED(CONFIG_THP_SWAP))
 346		return info->flags & CLUSTER_FLAG_HUGE;
 347	return false;
 348}
 349
 350static inline void cluster_clear_huge(struct swap_cluster_info *info)
 351{
 352	info->flags &= ~CLUSTER_FLAG_HUGE;
 353}
 354
 355static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
 356						     unsigned long offset)
 357{
 358	struct swap_cluster_info *ci;
 359
 360	ci = si->cluster_info;
 361	if (ci) {
 362		ci += offset / SWAPFILE_CLUSTER;
 363		spin_lock(&ci->lock);
 364	}
 365	return ci;
 366}
 367
 368static inline void unlock_cluster(struct swap_cluster_info *ci)
 369{
 370	if (ci)
 371		spin_unlock(&ci->lock);
 372}
 373
 374/*
 375 * Determine the locking method in use for this device.  Return
 376 * swap_cluster_info if SSD-style cluster-based locking is in place.
 377 */
 378static inline struct swap_cluster_info *lock_cluster_or_swap_info(
 379		struct swap_info_struct *si, unsigned long offset)
 380{
 381	struct swap_cluster_info *ci;
 382
 383	/* Try to use fine-grained SSD-style locking if available: */
 384	ci = lock_cluster(si, offset);
 385	/* Otherwise, fall back to traditional, coarse locking: */
 386	if (!ci)
 387		spin_lock(&si->lock);
 388
 389	return ci;
 390}
 391
 392static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
 393					       struct swap_cluster_info *ci)
 394{
 395	if (ci)
 396		unlock_cluster(ci);
 397	else
 398		spin_unlock(&si->lock);
 399}
 400
 401static inline bool cluster_list_empty(struct swap_cluster_list *list)
 402{
 403	return cluster_is_null(&list->head);
 404}
 405
 406static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
 407{
 408	return cluster_next(&list->head);
 409}
 410
 411static void cluster_list_init(struct swap_cluster_list *list)
 412{
 413	cluster_set_null(&list->head);
 414	cluster_set_null(&list->tail);
 415}
 416
 417static void cluster_list_add_tail(struct swap_cluster_list *list,
 418				  struct swap_cluster_info *ci,
 419				  unsigned int idx)
 420{
 421	if (cluster_list_empty(list)) {
 422		cluster_set_next_flag(&list->head, idx, 0);
 423		cluster_set_next_flag(&list->tail, idx, 0);
 424	} else {
 425		struct swap_cluster_info *ci_tail;
 426		unsigned int tail = cluster_next(&list->tail);
 427
 428		/*
 429		 * Nested cluster lock, but both cluster locks are
 430		 * only acquired when we held swap_info_struct->lock
 431		 */
 432		ci_tail = ci + tail;
 433		spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
 434		cluster_set_next(ci_tail, idx);
 435		spin_unlock(&ci_tail->lock);
 436		cluster_set_next_flag(&list->tail, idx, 0);
 437	}
 438}
 439
 440static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
 441					   struct swap_cluster_info *ci)
 442{
 443	unsigned int idx;
 444
 445	idx = cluster_next(&list->head);
 446	if (cluster_next(&list->tail) == idx) {
 447		cluster_set_null(&list->head);
 448		cluster_set_null(&list->tail);
 449	} else
 450		cluster_set_next_flag(&list->head,
 451				      cluster_next(&ci[idx]), 0);
 452
 453	return idx;
 454}
 455
 456/* Add a cluster to discard list and schedule it to do discard */
 457static void swap_cluster_schedule_discard(struct swap_info_struct *si,
 458		unsigned int idx)
 459{
 460	/*
 461	 * If scan_swap_map_slots() can't find a free cluster, it will check
 462	 * si->swap_map directly. To make sure the discarding cluster isn't
 463	 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
 464	 * It will be cleared after discard
 465	 */
 466	memset(si->swap_map + idx * SWAPFILE_CLUSTER,
 467			SWAP_MAP_BAD, SWAPFILE_CLUSTER);
 468
 469	cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
 470
 471	schedule_work(&si->discard_work);
 472}
 473
 474static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
 475{
 476	struct swap_cluster_info *ci = si->cluster_info;
 477
 478	cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
 479	cluster_list_add_tail(&si->free_clusters, ci, idx);
 480}
 481
 482/*
 483 * Doing discard actually. After a cluster discard is finished, the cluster
 484 * will be added to free cluster list. caller should hold si->lock.
 485*/
 486static void swap_do_scheduled_discard(struct swap_info_struct *si)
 487{
 488	struct swap_cluster_info *info, *ci;
 489	unsigned int idx;
 490
 491	info = si->cluster_info;
 492
 493	while (!cluster_list_empty(&si->discard_clusters)) {
 494		idx = cluster_list_del_first(&si->discard_clusters, info);
 495		spin_unlock(&si->lock);
 496
 497		discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
 498				SWAPFILE_CLUSTER);
 499
 500		spin_lock(&si->lock);
 501		ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
 502		__free_cluster(si, idx);
 503		memset(si->swap_map + idx * SWAPFILE_CLUSTER,
 504				0, SWAPFILE_CLUSTER);
 505		unlock_cluster(ci);
 506	}
 507}
 508
 509static void swap_discard_work(struct work_struct *work)
 510{
 511	struct swap_info_struct *si;
 512
 513	si = container_of(work, struct swap_info_struct, discard_work);
 514
 515	spin_lock(&si->lock);
 516	swap_do_scheduled_discard(si);
 517	spin_unlock(&si->lock);
 518}
 519
 520static void swap_users_ref_free(struct percpu_ref *ref)
 521{
 522	struct swap_info_struct *si;
 523
 524	si = container_of(ref, struct swap_info_struct, users);
 525	complete(&si->comp);
 526}
 527
 528static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
 529{
 530	struct swap_cluster_info *ci = si->cluster_info;
 531
 532	VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
 533	cluster_list_del_first(&si->free_clusters, ci);
 534	cluster_set_count_flag(ci + idx, 0, 0);
 535}
 536
 537static void free_cluster(struct swap_info_struct *si, unsigned long idx)
 538{
 539	struct swap_cluster_info *ci = si->cluster_info + idx;
 540
 541	VM_BUG_ON(cluster_count(ci) != 0);
 542	/*
 543	 * If the swap is discardable, prepare discard the cluster
 544	 * instead of free it immediately. The cluster will be freed
 545	 * after discard.
 546	 */
 547	if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
 548	    (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
 549		swap_cluster_schedule_discard(si, idx);
 550		return;
 551	}
 552
 553	__free_cluster(si, idx);
 554}
 555
 556/*
 557 * The cluster corresponding to page_nr will be used. The cluster will be
 558 * removed from free cluster list and its usage counter will be increased.
 559 */
 560static void inc_cluster_info_page(struct swap_info_struct *p,
 561	struct swap_cluster_info *cluster_info, unsigned long page_nr)
 562{
 563	unsigned long idx = page_nr / SWAPFILE_CLUSTER;
 564
 565	if (!cluster_info)
 566		return;
 567	if (cluster_is_free(&cluster_info[idx]))
 568		alloc_cluster(p, idx);
 569
 570	VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
 571	cluster_set_count(&cluster_info[idx],
 572		cluster_count(&cluster_info[idx]) + 1);
 573}
 574
 575/*
 576 * The cluster corresponding to page_nr decreases one usage. If the usage
 577 * counter becomes 0, which means no page in the cluster is in using, we can
 578 * optionally discard the cluster and add it to free cluster list.
 579 */
 580static void dec_cluster_info_page(struct swap_info_struct *p,
 581	struct swap_cluster_info *cluster_info, unsigned long page_nr)
 582{
 583	unsigned long idx = page_nr / SWAPFILE_CLUSTER;
 584
 585	if (!cluster_info)
 586		return;
 587
 588	VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
 589	cluster_set_count(&cluster_info[idx],
 590		cluster_count(&cluster_info[idx]) - 1);
 591
 592	if (cluster_count(&cluster_info[idx]) == 0)
 593		free_cluster(p, idx);
 594}
 595
 596/*
 597 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
 598 * cluster list. Avoiding such abuse to avoid list corruption.
 599 */
 600static bool
 601scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
 602	unsigned long offset)
 603{
 604	struct percpu_cluster *percpu_cluster;
 605	bool conflict;
 606
 607	offset /= SWAPFILE_CLUSTER;
 608	conflict = !cluster_list_empty(&si->free_clusters) &&
 609		offset != cluster_list_first(&si->free_clusters) &&
 610		cluster_is_free(&si->cluster_info[offset]);
 611
 612	if (!conflict)
 613		return false;
 614
 615	percpu_cluster = this_cpu_ptr(si->percpu_cluster);
 616	cluster_set_null(&percpu_cluster->index);
 617	return true;
 618}
 619
 620/*
 621 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
 622 * might involve allocating a new cluster for current CPU too.
 623 */
 624static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
 625	unsigned long *offset, unsigned long *scan_base)
 626{
 627	struct percpu_cluster *cluster;
 628	struct swap_cluster_info *ci;
 629	unsigned long tmp, max;
 630
 631new_cluster:
 632	cluster = this_cpu_ptr(si->percpu_cluster);
 633	if (cluster_is_null(&cluster->index)) {
 634		if (!cluster_list_empty(&si->free_clusters)) {
 635			cluster->index = si->free_clusters.head;
 636			cluster->next = cluster_next(&cluster->index) *
 637					SWAPFILE_CLUSTER;
 638		} else if (!cluster_list_empty(&si->discard_clusters)) {
 639			/*
 640			 * we don't have free cluster but have some clusters in
 641			 * discarding, do discard now and reclaim them, then
 642			 * reread cluster_next_cpu since we dropped si->lock
 643			 */
 644			swap_do_scheduled_discard(si);
 645			*scan_base = this_cpu_read(*si->cluster_next_cpu);
 646			*offset = *scan_base;
 647			goto new_cluster;
 648		} else
 649			return false;
 650	}
 651
 652	/*
 653	 * Other CPUs can use our cluster if they can't find a free cluster,
 654	 * check if there is still free entry in the cluster
 655	 */
 656	tmp = cluster->next;
 657	max = min_t(unsigned long, si->max,
 658		    (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
 659	if (tmp < max) {
 660		ci = lock_cluster(si, tmp);
 661		while (tmp < max) {
 662			if (!si->swap_map[tmp])
 663				break;
 664			tmp++;
 665		}
 666		unlock_cluster(ci);
 667	}
 668	if (tmp >= max) {
 669		cluster_set_null(&cluster->index);
 670		goto new_cluster;
 671	}
 672	cluster->next = tmp + 1;
 673	*offset = tmp;
 674	*scan_base = tmp;
 675	return true;
 676}
 677
 678static void __del_from_avail_list(struct swap_info_struct *p)
 679{
 680	int nid;
 681
 682	for_each_node(nid)
 683		plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
 684}
 685
 686static void del_from_avail_list(struct swap_info_struct *p)
 687{
 688	spin_lock(&swap_avail_lock);
 689	__del_from_avail_list(p);
 690	spin_unlock(&swap_avail_lock);
 691}
 692
 693static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
 694			     unsigned int nr_entries)
 695{
 696	unsigned int end = offset + nr_entries - 1;
 697
 698	if (offset == si->lowest_bit)
 699		si->lowest_bit += nr_entries;
 700	if (end == si->highest_bit)
 701		WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
 702	WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
 703	if (si->inuse_pages == si->pages) {
 704		si->lowest_bit = si->max;
 705		si->highest_bit = 0;
 706		del_from_avail_list(si);
 707	}
 708}
 709
 710static void add_to_avail_list(struct swap_info_struct *p)
 711{
 712	int nid;
 713
 714	spin_lock(&swap_avail_lock);
 715	for_each_node(nid) {
 716		WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
 717		plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
 718	}
 719	spin_unlock(&swap_avail_lock);
 720}
 721
 722static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
 723			    unsigned int nr_entries)
 724{
 725	unsigned long begin = offset;
 726	unsigned long end = offset + nr_entries - 1;
 727	void (*swap_slot_free_notify)(struct block_device *, unsigned long);
 728
 729	if (offset < si->lowest_bit)
 730		si->lowest_bit = offset;
 731	if (end > si->highest_bit) {
 732		bool was_full = !si->highest_bit;
 733
 734		WRITE_ONCE(si->highest_bit, end);
 735		if (was_full && (si->flags & SWP_WRITEOK))
 736			add_to_avail_list(si);
 737	}
 738	atomic_long_add(nr_entries, &nr_swap_pages);
 739	WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
 740	if (si->flags & SWP_BLKDEV)
 741		swap_slot_free_notify =
 742			si->bdev->bd_disk->fops->swap_slot_free_notify;
 743	else
 744		swap_slot_free_notify = NULL;
 745	while (offset <= end) {
 746		arch_swap_invalidate_page(si->type, offset);
 747		frontswap_invalidate_page(si->type, offset);
 748		if (swap_slot_free_notify)
 749			swap_slot_free_notify(si->bdev, offset);
 750		offset++;
 751	}
 752	clear_shadow_from_swap_cache(si->type, begin, end);
 753}
 754
 755static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
 756{
 757	unsigned long prev;
 758
 759	if (!(si->flags & SWP_SOLIDSTATE)) {
 760		si->cluster_next = next;
 761		return;
 762	}
 763
 764	prev = this_cpu_read(*si->cluster_next_cpu);
 765	/*
 766	 * Cross the swap address space size aligned trunk, choose
 767	 * another trunk randomly to avoid lock contention on swap
 768	 * address space if possible.
 769	 */
 770	if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
 771	    (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
 772		/* No free swap slots available */
 773		if (si->highest_bit <= si->lowest_bit)
 774			return;
 775		next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
 776		next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
 777		next = max_t(unsigned int, next, si->lowest_bit);
 778	}
 779	this_cpu_write(*si->cluster_next_cpu, next);
 780}
 781
 782static bool swap_offset_available_and_locked(struct swap_info_struct *si,
 783					     unsigned long offset)
 784{
 785	if (data_race(!si->swap_map[offset])) {
 786		spin_lock(&si->lock);
 787		return true;
 788	}
 789
 790	if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
 791		spin_lock(&si->lock);
 792		return true;
 793	}
 794
 795	return false;
 796}
 797
 798static int scan_swap_map_slots(struct swap_info_struct *si,
 799			       unsigned char usage, int nr,
 800			       swp_entry_t slots[])
 801{
 802	struct swap_cluster_info *ci;
 803	unsigned long offset;
 804	unsigned long scan_base;
 805	unsigned long last_in_cluster = 0;
 806	int latency_ration = LATENCY_LIMIT;
 807	int n_ret = 0;
 808	bool scanned_many = false;
 809
 810	/*
 811	 * We try to cluster swap pages by allocating them sequentially
 812	 * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 813	 * way, however, we resort to first-free allocation, starting
 814	 * a new cluster.  This prevents us from scattering swap pages
 815	 * all over the entire swap partition, so that we reduce
 816	 * overall disk seek times between swap pages.  -- sct
 817	 * But we do now try to find an empty cluster.  -Andrea
 818	 * And we let swap pages go all over an SSD partition.  Hugh
 819	 */
 820
 821	si->flags += SWP_SCANNING;
 822	/*
 823	 * Use percpu scan base for SSD to reduce lock contention on
 824	 * cluster and swap cache.  For HDD, sequential access is more
 825	 * important.
 826	 */
 827	if (si->flags & SWP_SOLIDSTATE)
 828		scan_base = this_cpu_read(*si->cluster_next_cpu);
 829	else
 830		scan_base = si->cluster_next;
 831	offset = scan_base;
 832
 833	/* SSD algorithm */
 834	if (si->cluster_info) {
 835		if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
 836			goto scan;
 837	} else if (unlikely(!si->cluster_nr--)) {
 838		if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 839			si->cluster_nr = SWAPFILE_CLUSTER - 1;
 840			goto checks;
 841		}
 842
 843		spin_unlock(&si->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 844
 845		/*
 846		 * If seek is expensive, start searching for new cluster from
 847		 * start of partition, to minimize the span of allocated swap.
 848		 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
 849		 * case, just handled by scan_swap_map_try_ssd_cluster() above.
 
 
 850		 */
 851		scan_base = offset = si->lowest_bit;
 
 852		last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 853
 854		/* Locate the first empty (unaligned) cluster */
 855		for (; last_in_cluster <= si->highest_bit; offset++) {
 856			if (si->swap_map[offset])
 857				last_in_cluster = offset + SWAPFILE_CLUSTER;
 858			else if (offset == last_in_cluster) {
 859				spin_lock(&si->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 860				offset -= SWAPFILE_CLUSTER - 1;
 861				si->cluster_next = offset;
 862				si->cluster_nr = SWAPFILE_CLUSTER - 1;
 
 863				goto checks;
 864			}
 865			if (unlikely(--latency_ration < 0)) {
 866				cond_resched();
 867				latency_ration = LATENCY_LIMIT;
 868			}
 869		}
 870
 871		offset = scan_base;
 872		spin_lock(&si->lock);
 873		si->cluster_nr = SWAPFILE_CLUSTER - 1;
 
 874	}
 875
 876checks:
 877	if (si->cluster_info) {
 878		while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
 879		/* take a break if we already got some slots */
 880			if (n_ret)
 881				goto done;
 882			if (!scan_swap_map_try_ssd_cluster(si, &offset,
 883							&scan_base))
 884				goto scan;
 885		}
 886	}
 887	if (!(si->flags & SWP_WRITEOK))
 888		goto no_page;
 889	if (!si->highest_bit)
 890		goto no_page;
 891	if (offset > si->highest_bit)
 892		scan_base = offset = si->lowest_bit;
 893
 894	ci = lock_cluster(si, offset);
 895	/* reuse swap entry of cache-only swap if not busy. */
 896	if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 897		int swap_was_freed;
 898		unlock_cluster(ci);
 899		spin_unlock(&si->lock);
 900		swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
 901		spin_lock(&si->lock);
 902		/* entry was freed successfully, try to use this again */
 903		if (swap_was_freed)
 904			goto checks;
 905		goto scan; /* check next one */
 906	}
 907
 908	if (si->swap_map[offset]) {
 909		unlock_cluster(ci);
 910		if (!n_ret)
 911			goto scan;
 912		else
 913			goto done;
 914	}
 915	WRITE_ONCE(si->swap_map[offset], usage);
 916	inc_cluster_info_page(si, si->cluster_info, offset);
 917	unlock_cluster(ci);
 918
 919	swap_range_alloc(si, offset, 1);
 920	slots[n_ret++] = swp_entry(si->type, offset);
 921
 922	/* got enough slots or reach max slots? */
 923	if ((n_ret == nr) || (offset >= si->highest_bit))
 924		goto done;
 925
 926	/* search for next available slot */
 927
 928	/* time to take a break? */
 929	if (unlikely(--latency_ration < 0)) {
 930		if (n_ret)
 931			goto done;
 932		spin_unlock(&si->lock);
 933		cond_resched();
 934		spin_lock(&si->lock);
 935		latency_ration = LATENCY_LIMIT;
 936	}
 937
 938	/* try to get more slots in cluster */
 939	if (si->cluster_info) {
 940		if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
 941			goto checks;
 942	} else if (si->cluster_nr && !si->swap_map[++offset]) {
 943		/* non-ssd case, still more slots in cluster? */
 944		--si->cluster_nr;
 945		goto checks;
 946	}
 
 
 
 947
 948	/*
 949	 * Even if there's no free clusters available (fragmented),
 950	 * try to scan a little more quickly with lock held unless we
 951	 * have scanned too many slots already.
 952	 */
 953	if (!scanned_many) {
 954		unsigned long scan_limit;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 955
 956		if (offset < scan_base)
 957			scan_limit = scan_base;
 958		else
 959			scan_limit = si->highest_bit;
 960		for (; offset <= scan_limit && --latency_ration > 0;
 961		     offset++) {
 962			if (!si->swap_map[offset])
 963				goto checks;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 964		}
 965	}
 966
 967done:
 968	set_cluster_next(si, offset + 1);
 969	si->flags -= SWP_SCANNING;
 970	return n_ret;
 971
 972scan:
 973	spin_unlock(&si->lock);
 974	while (++offset <= READ_ONCE(si->highest_bit)) {
 
 
 
 
 
 
 
 
 975		if (unlikely(--latency_ration < 0)) {
 976			cond_resched();
 977			latency_ration = LATENCY_LIMIT;
 978			scanned_many = true;
 979		}
 980		if (swap_offset_available_and_locked(si, offset))
 981			goto checks;
 982	}
 983	offset = si->lowest_bit;
 984	while (offset < scan_base) {
 
 
 
 
 
 
 
 
 985		if (unlikely(--latency_ration < 0)) {
 986			cond_resched();
 987			latency_ration = LATENCY_LIMIT;
 988			scanned_many = true;
 989		}
 990		if (swap_offset_available_and_locked(si, offset))
 991			goto checks;
 992		offset++;
 993	}
 994	spin_lock(&si->lock);
 995
 996no_page:
 997	si->flags -= SWP_SCANNING;
 998	return n_ret;
 999}
1000
1001static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
1002{
1003	unsigned long idx;
1004	struct swap_cluster_info *ci;
1005	unsigned long offset;
1006
1007	/*
1008	 * Should not even be attempting cluster allocations when huge
1009	 * page swap is disabled.  Warn and fail the allocation.
1010	 */
1011	if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1012		VM_WARN_ON_ONCE(1);
1013		return 0;
1014	}
1015
1016	if (cluster_list_empty(&si->free_clusters))
1017		return 0;
 
 
1018
1019	idx = cluster_list_first(&si->free_clusters);
1020	offset = idx * SWAPFILE_CLUSTER;
1021	ci = lock_cluster(si, offset);
1022	alloc_cluster(si, idx);
1023	cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1024
1025	memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1026	unlock_cluster(ci);
1027	swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1028	*slot = swp_entry(si->type, offset);
1029
1030	return 1;
1031}
 
 
1032
1033static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1034{
1035	unsigned long offset = idx * SWAPFILE_CLUSTER;
1036	struct swap_cluster_info *ci;
 
 
 
 
 
1037
1038	ci = lock_cluster(si, offset);
1039	memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1040	cluster_set_count_flag(ci, 0, 0);
1041	free_cluster(si, idx);
1042	unlock_cluster(ci);
1043	swap_range_free(si, offset, SWAPFILE_CLUSTER);
1044}
1045
1046int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
 
1047{
1048	unsigned long size = swap_entry_size(entry_size);
1049	struct swap_info_struct *si, *next;
1050	long avail_pgs;
1051	int n_ret = 0;
1052	int node;
1053
1054	/* Only single cluster request supported */
1055	WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1056
1057	spin_lock(&swap_avail_lock);
1058
1059	avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1060	if (avail_pgs <= 0) {
1061		spin_unlock(&swap_avail_lock);
1062		goto noswap;
 
 
 
 
 
 
 
1063	}
1064
1065	n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1066
1067	atomic_long_sub(n_goal * size, &nr_swap_pages);
1068
1069start_over:
1070	node = numa_node_id();
1071	plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1072		/* requeue si to after same-priority siblings */
1073		plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1074		spin_unlock(&swap_avail_lock);
1075		spin_lock(&si->lock);
1076		if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1077			spin_lock(&swap_avail_lock);
1078			if (plist_node_empty(&si->avail_lists[node])) {
1079				spin_unlock(&si->lock);
1080				goto nextsi;
1081			}
1082			WARN(!si->highest_bit,
1083			     "swap_info %d in list but !highest_bit\n",
1084			     si->type);
1085			WARN(!(si->flags & SWP_WRITEOK),
1086			     "swap_info %d in list but !SWP_WRITEOK\n",
1087			     si->type);
1088			__del_from_avail_list(si);
1089			spin_unlock(&si->lock);
1090			goto nextsi;
1091		}
1092		if (size == SWAPFILE_CLUSTER) {
1093			if (si->flags & SWP_BLKDEV)
1094				n_ret = swap_alloc_cluster(si, swp_entries);
1095		} else
1096			n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1097						    n_goal, swp_entries);
1098		spin_unlock(&si->lock);
1099		if (n_ret || size == SWAPFILE_CLUSTER)
1100			goto check_out;
1101		pr_debug("scan_swap_map of si %d failed to find offset\n",
1102			si->type);
1103		cond_resched();
1104
1105		spin_lock(&swap_avail_lock);
1106nextsi:
1107		/*
1108		 * if we got here, it's likely that si was almost full before,
1109		 * and since scan_swap_map_slots() can drop the si->lock,
1110		 * multiple callers probably all tried to get a page from the
1111		 * same si and it filled up before we could get one; or, the si
1112		 * filled up between us dropping swap_avail_lock and taking
1113		 * si->lock. Since we dropped the swap_avail_lock, the
1114		 * swap_avail_head list may have been modified; so if next is
1115		 * still in the swap_avail_head list then try it, otherwise
1116		 * start over if we have not gotten any slots.
1117		 */
1118		if (plist_node_empty(&next->avail_lists[node]))
1119			goto start_over;
1120	}
1121
1122	spin_unlock(&swap_avail_lock);
1123
1124check_out:
1125	if (n_ret < n_goal)
1126		atomic_long_add((long)(n_goal - n_ret) * size,
1127				&nr_swap_pages);
1128noswap:
1129	return n_ret;
1130}
1131
1132static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1133{
1134	struct swap_info_struct *p;
1135	unsigned long offset;
1136
1137	if (!entry.val)
1138		goto out;
1139	p = swp_swap_info(entry);
1140	if (!p)
1141		goto bad_nofile;
1142	if (data_race(!(p->flags & SWP_USED)))
 
1143		goto bad_device;
1144	offset = swp_offset(entry);
1145	if (offset >= p->max)
1146		goto bad_offset;
1147	if (data_race(!p->swap_map[swp_offset(entry)]))
1148		goto bad_free;
 
1149	return p;
1150
1151bad_free:
1152	pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1153	goto out;
1154bad_offset:
1155	pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1156	goto out;
1157bad_device:
1158	pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1159	goto out;
1160bad_nofile:
1161	pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1162out:
1163	return NULL;
1164}
1165
1166static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1167					struct swap_info_struct *q)
1168{
1169	struct swap_info_struct *p;
1170
1171	p = _swap_info_get(entry);
1172
1173	if (p != q) {
1174		if (q != NULL)
1175			spin_unlock(&q->lock);
1176		if (p != NULL)
1177			spin_lock(&p->lock);
1178	}
1179	return p;
1180}
1181
1182static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1183					      unsigned long offset,
1184					      unsigned char usage)
1185{
 
1186	unsigned char count;
1187	unsigned char has_cache;
1188
1189	count = p->swap_map[offset];
1190
1191	has_cache = count & SWAP_HAS_CACHE;
1192	count &= ~SWAP_HAS_CACHE;
1193
1194	if (usage == SWAP_HAS_CACHE) {
1195		VM_BUG_ON(!has_cache);
1196		has_cache = 0;
1197	} else if (count == SWAP_MAP_SHMEM) {
1198		/*
1199		 * Or we could insist on shmem.c using a special
1200		 * swap_shmem_free() and free_shmem_swap_and_cache()...
1201		 */
1202		count = 0;
1203	} else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1204		if (count == COUNT_CONTINUED) {
1205			if (swap_count_continued(p, offset, count))
1206				count = SWAP_MAP_MAX | COUNT_CONTINUED;
1207			else
1208				count = SWAP_MAP_MAX;
1209		} else
1210			count--;
1211	}
1212
1213	usage = count | has_cache;
1214	if (usage)
1215		WRITE_ONCE(p->swap_map[offset], usage);
1216	else
1217		WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1218
1219	return usage;
1220}
1221
1222/*
1223 * Check whether swap entry is valid in the swap device.  If so,
1224 * return pointer to swap_info_struct, and keep the swap entry valid
1225 * via preventing the swap device from being swapoff, until
1226 * put_swap_device() is called.  Otherwise return NULL.
1227 *
1228 * Notice that swapoff or swapoff+swapon can still happen before the
1229 * percpu_ref_tryget_live() in get_swap_device() or after the
1230 * percpu_ref_put() in put_swap_device() if there isn't any other way
1231 * to prevent swapoff, such as page lock, page table lock, etc.  The
1232 * caller must be prepared for that.  For example, the following
1233 * situation is possible.
1234 *
1235 *   CPU1				CPU2
1236 *   do_swap_page()
1237 *     ...				swapoff+swapon
1238 *     __read_swap_cache_async()
1239 *       swapcache_prepare()
1240 *         __swap_duplicate()
1241 *           // check swap_map
1242 *     // verify PTE not changed
1243 *
1244 * In __swap_duplicate(), the swap_map need to be checked before
1245 * changing partly because the specified swap entry may be for another
1246 * swap device which has been swapoff.  And in do_swap_page(), after
1247 * the page is read from the swap device, the PTE is verified not
1248 * changed with the page table locked to check whether the swap device
1249 * has been swapoff or swapoff+swapon.
1250 */
1251struct swap_info_struct *get_swap_device(swp_entry_t entry)
1252{
1253	struct swap_info_struct *si;
1254	unsigned long offset;
1255
1256	if (!entry.val)
1257		goto out;
1258	si = swp_swap_info(entry);
1259	if (!si)
1260		goto bad_nofile;
1261	if (!percpu_ref_tryget_live(&si->users))
1262		goto out;
1263	/*
1264	 * Guarantee the si->users are checked before accessing other
1265	 * fields of swap_info_struct.
1266	 *
1267	 * Paired with the spin_unlock() after setup_swap_info() in
1268	 * enable_swap_info().
1269	 */
1270	smp_rmb();
1271	offset = swp_offset(entry);
1272	if (offset >= si->max)
1273		goto put_out;
1274
1275	return si;
1276bad_nofile:
1277	pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1278out:
1279	return NULL;
1280put_out:
1281	pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1282	percpu_ref_put(&si->users);
1283	return NULL;
1284}
1285
1286static unsigned char __swap_entry_free(struct swap_info_struct *p,
1287				       swp_entry_t entry)
1288{
1289	struct swap_cluster_info *ci;
1290	unsigned long offset = swp_offset(entry);
1291	unsigned char usage;
1292
1293	ci = lock_cluster_or_swap_info(p, offset);
1294	usage = __swap_entry_free_locked(p, offset, 1);
1295	unlock_cluster_or_swap_info(p, ci);
1296	if (!usage)
1297		free_swap_slot(entry);
 
 
 
 
 
 
 
 
 
 
 
1298
1299	return usage;
1300}
1301
1302static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1303{
1304	struct swap_cluster_info *ci;
1305	unsigned long offset = swp_offset(entry);
1306	unsigned char count;
1307
1308	ci = lock_cluster(p, offset);
1309	count = p->swap_map[offset];
1310	VM_BUG_ON(count != SWAP_HAS_CACHE);
1311	p->swap_map[offset] = 0;
1312	dec_cluster_info_page(p, p->cluster_info, offset);
1313	unlock_cluster(ci);
1314
1315	mem_cgroup_uncharge_swap(entry, 1);
1316	swap_range_free(p, offset, 1);
1317}
1318
1319/*
1320 * Caller has made sure that the swap device corresponding to entry
1321 * is still around or has not been recycled.
1322 */
1323void swap_free(swp_entry_t entry)
1324{
1325	struct swap_info_struct *p;
1326
1327	p = _swap_info_get(entry);
1328	if (p)
1329		__swap_entry_free(p, entry);
 
 
1330}
1331
1332/*
1333 * Called after dropping swapcache to decrease refcnt to swap entries.
1334 */
1335void put_swap_folio(struct folio *folio, swp_entry_t entry)
1336{
1337	unsigned long offset = swp_offset(entry);
1338	unsigned long idx = offset / SWAPFILE_CLUSTER;
1339	struct swap_cluster_info *ci;
1340	struct swap_info_struct *si;
1341	unsigned char *map;
1342	unsigned int i, free_entries = 0;
1343	unsigned char val;
1344	int size = swap_entry_size(folio_nr_pages(folio));
1345
1346	si = _swap_info_get(entry);
1347	if (!si)
1348		return;
1349
1350	ci = lock_cluster_or_swap_info(si, offset);
1351	if (size == SWAPFILE_CLUSTER) {
1352		VM_BUG_ON(!cluster_is_huge(ci));
1353		map = si->swap_map + offset;
1354		for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1355			val = map[i];
1356			VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1357			if (val == SWAP_HAS_CACHE)
1358				free_entries++;
1359		}
1360		cluster_clear_huge(ci);
1361		if (free_entries == SWAPFILE_CLUSTER) {
1362			unlock_cluster_or_swap_info(si, ci);
1363			spin_lock(&si->lock);
1364			mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1365			swap_free_cluster(si, idx);
1366			spin_unlock(&si->lock);
1367			return;
1368		}
1369	}
1370	for (i = 0; i < size; i++, entry.val++) {
1371		if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1372			unlock_cluster_or_swap_info(si, ci);
1373			free_swap_slot(entry);
1374			if (i == size - 1)
1375				return;
1376			lock_cluster_or_swap_info(si, offset);
1377		}
1378	}
1379	unlock_cluster_or_swap_info(si, ci);
1380}
1381
1382#ifdef CONFIG_THP_SWAP
1383int split_swap_cluster(swp_entry_t entry)
1384{
1385	struct swap_info_struct *si;
1386	struct swap_cluster_info *ci;
1387	unsigned long offset = swp_offset(entry);
1388
1389	si = _swap_info_get(entry);
1390	if (!si)
1391		return -EBUSY;
1392	ci = lock_cluster(si, offset);
1393	cluster_clear_huge(ci);
1394	unlock_cluster(ci);
1395	return 0;
1396}
1397#endif
1398
1399static int swp_entry_cmp(const void *ent1, const void *ent2)
1400{
1401	const swp_entry_t *e1 = ent1, *e2 = ent2;
1402
1403	return (int)swp_type(*e1) - (int)swp_type(*e2);
1404}
1405
1406void swapcache_free_entries(swp_entry_t *entries, int n)
1407{
1408	struct swap_info_struct *p, *prev;
1409	int i;
1410
1411	if (n <= 0)
1412		return;
1413
1414	prev = NULL;
1415	p = NULL;
1416
1417	/*
1418	 * Sort swap entries by swap device, so each lock is only taken once.
1419	 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1420	 * so low that it isn't necessary to optimize further.
1421	 */
1422	if (nr_swapfiles > 1)
1423		sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1424	for (i = 0; i < n; ++i) {
1425		p = swap_info_get_cont(entries[i], prev);
1426		if (p)
1427			swap_entry_free(p, entries[i]);
1428		prev = p;
1429	}
1430	if (p)
1431		spin_unlock(&p->lock);
1432}
1433
1434int __swap_count(swp_entry_t entry)
1435{
1436	struct swap_info_struct *si;
1437	pgoff_t offset = swp_offset(entry);
1438	int count = 0;
1439
1440	si = get_swap_device(entry);
1441	if (si) {
1442		count = swap_count(si->swap_map[offset]);
1443		put_swap_device(si);
 
 
1444	}
1445	return count;
1446}
1447
1448/*
1449 * How many references to @entry are currently swapped out?
1450 * This does not give an exact answer when swap count is continued,
1451 * but does include the high COUNT_CONTINUED flag to allow for that.
1452 */
1453static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1454{
1455	pgoff_t offset = swp_offset(entry);
1456	struct swap_cluster_info *ci;
1457	int count;
1458
1459	ci = lock_cluster_or_swap_info(si, offset);
1460	count = swap_count(si->swap_map[offset]);
1461	unlock_cluster_or_swap_info(si, ci);
1462	return count;
1463}
1464
1465/*
1466 * How many references to @entry are currently swapped out?
1467 * This does not give an exact answer when swap count is continued,
1468 * but does include the high COUNT_CONTINUED flag to allow for that.
1469 */
1470int __swp_swapcount(swp_entry_t entry)
1471{
1472	int count = 0;
1473	struct swap_info_struct *si;
 
1474
1475	si = get_swap_device(entry);
1476	if (si) {
1477		count = swap_swapcount(si, entry);
1478		put_swap_device(si);
 
1479	}
1480	return count;
1481}
1482
1483/*
1484 * How many references to @entry are currently swapped out?
1485 * This considers COUNT_CONTINUED so it returns exact answer.
 
 
1486 */
1487int swp_swapcount(swp_entry_t entry)
1488{
1489	int count, tmp_count, n;
1490	struct swap_info_struct *p;
1491	struct swap_cluster_info *ci;
1492	struct page *page;
1493	pgoff_t offset;
1494	unsigned char *map;
1495
1496	p = _swap_info_get(entry);
1497	if (!p)
1498		return 0;
1499
1500	offset = swp_offset(entry);
1501
1502	ci = lock_cluster_or_swap_info(p, offset);
1503
1504	count = swap_count(p->swap_map[offset]);
1505	if (!(count & COUNT_CONTINUED))
1506		goto out;
1507
1508	count &= ~COUNT_CONTINUED;
1509	n = SWAP_MAP_MAX + 1;
1510
1511	page = vmalloc_to_page(p->swap_map + offset);
1512	offset &= ~PAGE_MASK;
1513	VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1514
1515	do {
1516		page = list_next_entry(page, lru);
1517		map = kmap_atomic(page);
1518		tmp_count = map[offset];
1519		kunmap_atomic(map);
1520
1521		count += (tmp_count & ~COUNT_CONTINUED) * n;
1522		n *= (SWAP_CONT_MAX + 1);
1523	} while (tmp_count & COUNT_CONTINUED);
1524out:
1525	unlock_cluster_or_swap_info(p, ci);
1526	return count;
1527}
1528
1529static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1530					 swp_entry_t entry)
1531{
1532	struct swap_cluster_info *ci;
1533	unsigned char *map = si->swap_map;
1534	unsigned long roffset = swp_offset(entry);
1535	unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1536	int i;
1537	bool ret = false;
1538
1539	ci = lock_cluster_or_swap_info(si, offset);
1540	if (!ci || !cluster_is_huge(ci)) {
1541		if (swap_count(map[roffset]))
1542			ret = true;
1543		goto unlock_out;
1544	}
1545	for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1546		if (swap_count(map[offset + i])) {
1547			ret = true;
1548			break;
1549		}
1550	}
1551unlock_out:
1552	unlock_cluster_or_swap_info(si, ci);
1553	return ret;
1554}
1555
1556static bool folio_swapped(struct folio *folio)
1557{
1558	swp_entry_t entry = folio_swap_entry(folio);
1559	struct swap_info_struct *si = _swap_info_get(entry);
1560
1561	if (!si)
1562		return false;
1563
1564	if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1565		return swap_swapcount(si, entry) != 0;
1566
1567	return swap_page_trans_huge_swapped(si, entry);
1568}
1569
1570/**
1571 * folio_free_swap() - Free the swap space used for this folio.
1572 * @folio: The folio to remove.
1573 *
1574 * If swap is getting full, or if there are no more mappings of this folio,
1575 * then call folio_free_swap to free its swap space.
1576 *
1577 * Return: true if we were able to release the swap space.
1578 */
1579bool folio_free_swap(struct folio *folio)
1580{
1581	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1582
1583	if (!folio_test_swapcache(folio))
1584		return false;
1585	if (folio_test_writeback(folio))
1586		return false;
1587	if (folio_swapped(folio))
1588		return false;
1589
1590	/*
1591	 * Once hibernation has begun to create its image of memory,
1592	 * there's a danger that one of the calls to folio_free_swap()
1593	 * - most probably a call from __try_to_reclaim_swap() while
1594	 * hibernation is allocating its own swap pages for the image,
1595	 * but conceivably even a call from memory reclaim - will free
1596	 * the swap from a folio which has already been recorded in the
1597	 * image as a clean swapcache folio, and then reuse its swap for
1598	 * another page of the image.  On waking from hibernation, the
1599	 * original folio might be freed under memory pressure, then
1600	 * later read back in from swap, now with the wrong data.
1601	 *
1602	 * Hibernation suspends storage while it is writing the image
1603	 * to disk so check that here.
1604	 */
1605	if (pm_suspended_storage())
1606		return false;
1607
1608	delete_from_swap_cache(folio);
1609	folio_set_dirty(folio);
1610	return true;
1611}
1612
1613/*
1614 * Free the swap entry like above, but also try to
1615 * free the page cache entry if it is the last user.
1616 */
1617int free_swap_and_cache(swp_entry_t entry)
1618{
1619	struct swap_info_struct *p;
1620	unsigned char count;
1621
1622	if (non_swap_entry(entry))
1623		return 1;
1624
1625	p = _swap_info_get(entry);
1626	if (p) {
1627		count = __swap_entry_free(p, entry);
1628		if (count == SWAP_HAS_CACHE &&
1629		    !swap_page_trans_huge_swapped(p, entry))
1630			__try_to_reclaim_swap(p, swp_offset(entry),
1631					      TTRS_UNMAPPED | TTRS_FULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1632	}
1633	return p != NULL;
1634}
1635
1636#ifdef CONFIG_HIBERNATION
1637
1638swp_entry_t get_swap_page_of_type(int type)
 
 
 
 
 
 
 
 
 
1639{
1640	struct swap_info_struct *si = swap_type_to_swap_info(type);
1641	swp_entry_t entry = {0};
 
1642
1643	if (!si)
1644		goto fail;
 
 
 
 
 
 
1645
1646	/* This is called for allocating swap entry, not cache */
1647	spin_lock(&si->lock);
1648	if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1649		atomic_long_dec(&nr_swap_pages);
1650	spin_unlock(&si->lock);
1651fail:
1652	return entry;
1653}
 
1654
 
1655/*
1656 * Find the swap type that corresponds to given device (if any).
1657 *
1658 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1659 * from 0, in which the swap header is expected to be located.
1660 *
1661 * This is needed for the suspend to disk (aka swsusp).
1662 */
1663int swap_type_of(dev_t device, sector_t offset)
1664{
 
1665	int type;
1666
1667	if (!device)
1668		return -1;
1669
1670	spin_lock(&swap_lock);
1671	for (type = 0; type < nr_swapfiles; type++) {
1672		struct swap_info_struct *sis = swap_info[type];
1673
1674		if (!(sis->flags & SWP_WRITEOK))
1675			continue;
1676
1677		if (device == sis->bdev->bd_dev) {
1678			struct swap_extent *se = first_se(sis);
 
 
 
 
 
 
 
1679
1680			if (se->start_block == offset) {
 
 
 
1681				spin_unlock(&swap_lock);
 
1682				return type;
1683			}
1684		}
1685	}
1686	spin_unlock(&swap_lock);
1687	return -ENODEV;
1688}
1689
1690int find_first_swap(dev_t *device)
1691{
1692	int type;
1693
1694	spin_lock(&swap_lock);
1695	for (type = 0; type < nr_swapfiles; type++) {
1696		struct swap_info_struct *sis = swap_info[type];
1697
1698		if (!(sis->flags & SWP_WRITEOK))
1699			continue;
1700		*device = sis->bdev->bd_dev;
1701		spin_unlock(&swap_lock);
1702		return type;
1703	}
1704	spin_unlock(&swap_lock);
1705	return -ENODEV;
1706}
1707
1708/*
1709 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1710 * corresponding to given index in swap_info (swap type).
1711 */
1712sector_t swapdev_block(int type, pgoff_t offset)
1713{
1714	struct swap_info_struct *si = swap_type_to_swap_info(type);
1715	struct swap_extent *se;
1716
1717	if (!si || !(si->flags & SWP_WRITEOK))
 
 
1718		return 0;
1719	se = offset_to_swap_extent(si, offset);
1720	return se->start_block + (offset - se->start_page);
1721}
1722
1723/*
1724 * Return either the total number of swap pages of given type, or the number
1725 * of free pages of that type (depending on @free)
1726 *
1727 * This is needed for software suspend
1728 */
1729unsigned int count_swap_pages(int type, int free)
1730{
1731	unsigned int n = 0;
1732
1733	spin_lock(&swap_lock);
1734	if ((unsigned int)type < nr_swapfiles) {
1735		struct swap_info_struct *sis = swap_info[type];
1736
1737		spin_lock(&sis->lock);
1738		if (sis->flags & SWP_WRITEOK) {
1739			n = sis->pages;
1740			if (free)
1741				n -= sis->inuse_pages;
1742		}
1743		spin_unlock(&sis->lock);
1744	}
1745	spin_unlock(&swap_lock);
1746	return n;
1747}
1748#endif /* CONFIG_HIBERNATION */
1749
1750static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1751{
1752	return pte_same(pte_swp_clear_flags(pte), swp_pte);
1753}
1754
1755/*
1756 * No need to decide whether this PTE shares the swap entry with others,
1757 * just let do_wp_page work it out if a write is requested later - to
1758 * force COW, vm_page_prot omits write permission from any private vma.
1759 */
1760static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1761		unsigned long addr, swp_entry_t entry, struct folio *folio)
1762{
1763	struct page *page = folio_file_page(folio, swp_offset(entry));
1764	struct page *swapcache;
1765	spinlock_t *ptl;
1766	pte_t *pte, new_pte;
1767	bool hwposioned = false;
1768	int ret = 1;
1769
1770	swapcache = page;
1771	page = ksm_might_need_to_copy(page, vma, addr);
1772	if (unlikely(!page))
1773		return -ENOMEM;
1774	else if (unlikely(PTR_ERR(page) == -EHWPOISON))
1775		hwposioned = true;
1776
1777	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1778	if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
 
 
1779		ret = 0;
1780		goto out;
1781	}
1782
1783	if (unlikely(hwposioned || !PageUptodate(page))) {
1784		swp_entry_t swp_entry;
1785
1786		dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1787		if (hwposioned) {
1788			swp_entry = make_hwpoison_entry(swapcache);
1789			page = swapcache;
1790		} else {
1791			swp_entry = make_swapin_error_entry();
1792		}
1793		new_pte = swp_entry_to_pte(swp_entry);
1794		ret = 0;
1795		goto setpte;
1796	}
1797
1798	/* See do_swap_page() */
1799	BUG_ON(!PageAnon(page) && PageMappedToDisk(page));
1800	BUG_ON(PageAnon(page) && PageAnonExclusive(page));
1801
1802	dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1803	inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1804	get_page(page);
1805	if (page == swapcache) {
1806		rmap_t rmap_flags = RMAP_NONE;
1807
1808		/*
1809		 * See do_swap_page(): PageWriteback() would be problematic.
1810		 * However, we do a wait_on_page_writeback() just before this
1811		 * call and have the page locked.
1812		 */
1813		VM_BUG_ON_PAGE(PageWriteback(page), page);
1814		if (pte_swp_exclusive(*pte))
1815			rmap_flags |= RMAP_EXCLUSIVE;
1816
1817		page_add_anon_rmap(page, vma, addr, rmap_flags);
1818	} else { /* ksm created a completely new copy */
1819		page_add_new_anon_rmap(page, vma, addr);
1820		lru_cache_add_inactive_or_unevictable(page, vma);
1821	}
1822	new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
1823	if (pte_swp_soft_dirty(*pte))
1824		new_pte = pte_mksoft_dirty(new_pte);
1825	if (pte_swp_uffd_wp(*pte))
1826		new_pte = pte_mkuffd_wp(new_pte);
1827setpte:
1828	set_pte_at(vma->vm_mm, addr, pte, new_pte);
1829	swap_free(entry);
 
 
 
 
 
1830out:
1831	pte_unmap_unlock(pte, ptl);
1832	if (page != swapcache) {
1833		unlock_page(page);
1834		put_page(page);
1835	}
1836	return ret;
1837}
1838
1839static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1840			unsigned long addr, unsigned long end,
1841			unsigned int type)
1842{
1843	swp_entry_t entry;
1844	pte_t *pte;
1845	struct swap_info_struct *si;
1846	int ret = 0;
1847	volatile unsigned char *swap_map;
1848
1849	si = swap_info[type];
 
 
 
 
 
 
 
 
1850	pte = pte_offset_map(pmd, addr);
1851	do {
1852		struct folio *folio;
1853		unsigned long offset;
1854
1855		if (!is_swap_pte(*pte))
1856			continue;
1857
1858		entry = pte_to_swp_entry(*pte);
1859		if (swp_type(entry) != type)
1860			continue;
1861
1862		offset = swp_offset(entry);
1863		pte_unmap(pte);
1864		swap_map = &si->swap_map[offset];
1865		folio = swap_cache_get_folio(entry, vma, addr);
1866		if (!folio) {
1867			struct page *page;
1868			struct vm_fault vmf = {
1869				.vma = vma,
1870				.address = addr,
1871				.real_address = addr,
1872				.pmd = pmd,
1873			};
1874
1875			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1876						&vmf);
1877			if (page)
1878				folio = page_folio(page);
1879		}
1880		if (!folio) {
1881			if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1882				goto try_next;
1883			return -ENOMEM;
1884		}
1885
1886		folio_lock(folio);
1887		folio_wait_writeback(folio);
1888		ret = unuse_pte(vma, pmd, addr, entry, folio);
1889		if (ret < 0) {
1890			folio_unlock(folio);
1891			folio_put(folio);
1892			goto out;
1893		}
1894
1895		folio_free_swap(folio);
1896		folio_unlock(folio);
1897		folio_put(folio);
1898try_next:
1899		pte = pte_offset_map(pmd, addr);
1900	} while (pte++, addr += PAGE_SIZE, addr != end);
1901	pte_unmap(pte - 1);
1902
1903	ret = 0;
1904out:
1905	return ret;
1906}
1907
1908static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1909				unsigned long addr, unsigned long end,
1910				unsigned int type)
1911{
1912	pmd_t *pmd;
1913	unsigned long next;
1914	int ret;
1915
1916	pmd = pmd_offset(pud, addr);
1917	do {
1918		cond_resched();
1919		next = pmd_addr_end(addr, end);
1920		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1921			continue;
1922		ret = unuse_pte_range(vma, pmd, addr, next, type);
 
 
1923		if (ret)
1924			return ret;
1925	} while (pmd++, addr = next, addr != end);
1926	return 0;
1927}
1928
1929static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1930				unsigned long addr, unsigned long end,
1931				unsigned int type)
1932{
1933	pud_t *pud;
1934	unsigned long next;
1935	int ret;
1936
1937	pud = pud_offset(p4d, addr);
1938	do {
1939		next = pud_addr_end(addr, end);
1940		if (pud_none_or_clear_bad(pud))
1941			continue;
1942		ret = unuse_pmd_range(vma, pud, addr, next, type);
1943		if (ret)
1944			return ret;
1945	} while (pud++, addr = next, addr != end);
1946	return 0;
1947}
1948
1949static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1950				unsigned long addr, unsigned long end,
1951				unsigned int type)
1952{
1953	p4d_t *p4d;
1954	unsigned long next;
1955	int ret;
1956
1957	p4d = p4d_offset(pgd, addr);
1958	do {
1959		next = p4d_addr_end(addr, end);
1960		if (p4d_none_or_clear_bad(p4d))
1961			continue;
1962		ret = unuse_pud_range(vma, p4d, addr, next, type);
1963		if (ret)
1964			return ret;
1965	} while (p4d++, addr = next, addr != end);
1966	return 0;
1967}
1968
1969static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
1970{
1971	pgd_t *pgd;
1972	unsigned long addr, end, next;
1973	int ret;
1974
1975	addr = vma->vm_start;
1976	end = vma->vm_end;
 
 
 
 
 
 
 
 
1977
1978	pgd = pgd_offset(vma->vm_mm, addr);
1979	do {
1980		next = pgd_addr_end(addr, end);
1981		if (pgd_none_or_clear_bad(pgd))
1982			continue;
1983		ret = unuse_p4d_range(vma, pgd, addr, next, type);
1984		if (ret)
1985			return ret;
1986	} while (pgd++, addr = next, addr != end);
1987	return 0;
1988}
1989
1990static int unuse_mm(struct mm_struct *mm, unsigned int type)
 
1991{
1992	struct vm_area_struct *vma;
1993	int ret = 0;
1994	VMA_ITERATOR(vmi, mm, 0);
1995
1996	mmap_read_lock(mm);
1997	for_each_vma(vmi, vma) {
1998		if (vma->anon_vma) {
1999			ret = unuse_vma(vma, type);
2000			if (ret)
2001				break;
2002		}
2003
2004		cond_resched();
 
 
 
 
 
 
 
 
2005	}
2006	mmap_read_unlock(mm);
2007	return ret;
 
 
 
 
2008}
2009
2010/*
2011 * Scan swap_map from current position to next entry still in use.
2012 * Return 0 if there are no inuse entries after prev till end of
2013 * the map.
2014 */
2015static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2016					unsigned int prev)
2017{
2018	unsigned int i;
 
2019	unsigned char count;
2020
2021	/*
2022	 * No need for swap_lock here: we're just looking
2023	 * for whether an entry is in use, not modifying it; false
2024	 * hits are okay, and sys_swapoff() has already prevented new
2025	 * allocations from this area (while holding swap_lock).
2026	 */
2027	for (i = prev + 1; i < si->max; i++) {
2028		count = READ_ONCE(si->swap_map[i]);
 
 
 
 
 
 
 
 
 
 
 
 
 
2029		if (count && swap_count(count) != SWAP_MAP_BAD)
2030			break;
2031		if ((i % LATENCY_LIMIT) == 0)
2032			cond_resched();
2033	}
2034
2035	if (i == si->max)
2036		i = 0;
2037
2038	return i;
2039}
2040
 
 
 
 
 
2041static int try_to_unuse(unsigned int type)
2042{
2043	struct mm_struct *prev_mm;
2044	struct mm_struct *mm;
2045	struct list_head *p;
2046	int retval = 0;
2047	struct swap_info_struct *si = swap_info[type];
2048	struct folio *folio;
 
 
 
2049	swp_entry_t entry;
2050	unsigned int i;
 
2051
2052	if (!READ_ONCE(si->inuse_pages))
2053		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2054
2055retry:
2056	retval = shmem_unuse(type);
2057	if (retval)
2058		return retval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2059
2060	prev_mm = &init_mm;
2061	mmget(prev_mm);
 
 
 
 
 
 
2062
2063	spin_lock(&mmlist_lock);
2064	p = &init_mm.mmlist;
2065	while (READ_ONCE(si->inuse_pages) &&
2066	       !signal_pending(current) &&
2067	       (p = p->next) != &init_mm.mmlist) {
 
 
 
 
 
 
 
2068
2069		mm = list_entry(p, struct mm_struct, mmlist);
2070		if (!mmget_not_zero(mm))
 
 
 
 
 
 
 
2071			continue;
2072		spin_unlock(&mmlist_lock);
2073		mmput(prev_mm);
2074		prev_mm = mm;
2075		retval = unuse_mm(mm, type);
2076		if (retval) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2077			mmput(prev_mm);
2078			return retval;
 
 
 
 
 
 
2079		}
2080
2081		/*
2082		 * Make sure that we aren't completely killing
2083		 * interactive performance.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2084		 */
2085		cond_resched();
2086		spin_lock(&mmlist_lock);
2087	}
2088	spin_unlock(&mmlist_lock);
2089
2090	mmput(prev_mm);
2091
2092	i = 0;
2093	while (READ_ONCE(si->inuse_pages) &&
2094	       !signal_pending(current) &&
2095	       (i = find_next_to_unuse(si, i)) != 0) {
2096
2097		entry = swp_entry(type, i);
2098		folio = filemap_get_folio(swap_address_space(entry), i);
2099		if (!folio)
2100			continue;
 
 
 
 
 
 
2101
2102		/*
2103		 * It is conceivable that a racing task removed this folio from
2104		 * swap cache just before we acquired the page lock. The folio
2105		 * might even be back in swap cache on another swap area. But
2106		 * that is okay, folio_free_swap() only removes stale folios.
2107		 */
2108		folio_lock(folio);
2109		folio_wait_writeback(folio);
2110		folio_free_swap(folio);
2111		folio_unlock(folio);
2112		folio_put(folio);
2113	}
2114
2115	/*
2116	 * Lets check again to see if there are still swap entries in the map.
2117	 * If yes, we would need to do retry the unuse logic again.
2118	 * Under global memory pressure, swap entries can be reinserted back
2119	 * into process space after the mmlist loop above passes over them.
2120	 *
2121	 * Limit the number of retries? No: when mmget_not_zero()
2122	 * above fails, that mm is likely to be freeing swap from
2123	 * exit_mmap(), which proceeds at its own independent pace;
2124	 * and even shmem_writepage() could have been preempted after
2125	 * folio_alloc_swap(), temporarily hiding that swap.  It's easy
2126	 * and robust (though cpu-intensive) just to keep retrying.
2127	 */
2128	if (READ_ONCE(si->inuse_pages)) {
2129		if (!signal_pending(current))
2130			goto retry;
2131		return -EINTR;
2132	}
2133
2134	return 0;
 
2135}
2136
2137/*
2138 * After a successful try_to_unuse, if no swap is now in use, we know
2139 * we can empty the mmlist.  swap_lock must be held on entry and exit.
2140 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2141 * added to the mmlist just after page_duplicate - before would be racy.
2142 */
2143static void drain_mmlist(void)
2144{
2145	struct list_head *p, *next;
2146	unsigned int type;
2147
2148	for (type = 0; type < nr_swapfiles; type++)
2149		if (swap_info[type]->inuse_pages)
2150			return;
2151	spin_lock(&mmlist_lock);
2152	list_for_each_safe(p, next, &init_mm.mmlist)
2153		list_del_init(p);
2154	spin_unlock(&mmlist_lock);
2155}
2156
2157/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2158 * Free all of a swapdev's extent information
2159 */
2160static void destroy_swap_extents(struct swap_info_struct *sis)
2161{
2162	while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2163		struct rb_node *rb = sis->swap_extent_root.rb_node;
2164		struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2165
2166		rb_erase(rb, &sis->swap_extent_root);
 
 
2167		kfree(se);
2168	}
2169
2170	if (sis->flags & SWP_ACTIVATED) {
2171		struct file *swap_file = sis->swap_file;
2172		struct address_space *mapping = swap_file->f_mapping;
2173
2174		sis->flags &= ~SWP_ACTIVATED;
2175		if (mapping->a_ops->swap_deactivate)
2176			mapping->a_ops->swap_deactivate(swap_file);
2177	}
2178}
2179
2180/*
2181 * Add a block range (and the corresponding page range) into this swapdev's
2182 * extent tree.
2183 *
2184 * This function rather assumes that it is called in ascending page order.
2185 */
2186int
2187add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2188		unsigned long nr_pages, sector_t start_block)
2189{
2190	struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2191	struct swap_extent *se;
2192	struct swap_extent *new_se;
 
2193
2194	/*
2195	 * place the new node at the right most since the
2196	 * function is called in ascending page order.
2197	 */
2198	while (*link) {
2199		parent = *link;
2200		link = &parent->rb_right;
2201	}
2202
2203	if (parent) {
2204		se = rb_entry(parent, struct swap_extent, rb_node);
2205		BUG_ON(se->start_page + se->nr_pages != start_page);
2206		if (se->start_block + se->nr_pages == start_block) {
2207			/* Merge it */
2208			se->nr_pages += nr_pages;
2209			return 0;
2210		}
2211	}
2212
2213	/* No merge, insert a new extent. */
 
 
2214	new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2215	if (new_se == NULL)
2216		return -ENOMEM;
2217	new_se->start_page = start_page;
2218	new_se->nr_pages = nr_pages;
2219	new_se->start_block = start_block;
2220
2221	rb_link_node(&new_se->rb_node, parent, link);
2222	rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2223	return 1;
2224}
2225EXPORT_SYMBOL_GPL(add_swap_extent);
2226
2227/*
2228 * A `swap extent' is a simple thing which maps a contiguous range of pages
2229 * onto a contiguous range of disk blocks.  A rbtree of swap extents is
2230 * built at swapon time and is then used at swap_writepage/swap_readpage
2231 * time for locating where on disk a page belongs.
2232 *
2233 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2234 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2235 * swap files identically.
2236 *
2237 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2238 * extent rbtree operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2239 * swapfiles are handled *identically* after swapon time.
2240 *
2241 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2242 * and will parse them into a rbtree, in PAGE_SIZE chunks.  If some stray
2243 * blocks are found which do not fall within the PAGE_SIZE alignment
2244 * requirements, they are simply tossed out - we will never use those blocks
2245 * for swapping.
2246 *
2247 * For all swap devices we set S_SWAPFILE across the life of the swapon.  This
2248 * prevents users from writing to the swap device, which will corrupt memory.
 
2249 *
2250 * The amount of disk space which a single swap extent represents varies.
2251 * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2252 * extents in the rbtree. - akpm.
 
 
 
2253 */
2254static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2255{
2256	struct file *swap_file = sis->swap_file;
2257	struct address_space *mapping = swap_file->f_mapping;
2258	struct inode *inode = mapping->host;
 
 
 
 
 
 
2259	int ret;
2260
 
2261	if (S_ISBLK(inode->i_mode)) {
2262		ret = add_swap_extent(sis, 0, sis->max, 0);
2263		*span = sis->pages;
2264		return ret;
2265	}
2266
2267	if (mapping->a_ops->swap_activate) {
2268		ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2269		if (ret < 0)
2270			return ret;
2271		sis->flags |= SWP_ACTIVATED;
2272		if ((sis->flags & SWP_FS_OPS) &&
2273		    sio_pool_init() != 0) {
2274			destroy_swap_extents(sis);
2275			return -ENOMEM;
2276		}
2277		return ret;
2278	}
2279
2280	return generic_swapfile_activate(sis, swap_file, span);
2281}
 
 
 
 
 
 
 
 
 
 
 
 
 
2282
2283static int swap_node(struct swap_info_struct *p)
2284{
2285	struct block_device *bdev;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2286
2287	if (p->bdev)
2288		bdev = p->bdev;
2289	else
2290		bdev = p->swap_file->f_inode->i_sb->s_bdev;
 
 
 
2291
2292	return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2293}
2294
2295static void setup_swap_info(struct swap_info_struct *p, int prio,
2296			    unsigned char *swap_map,
2297			    struct swap_cluster_info *cluster_info)
2298{
2299	int i;
2300
 
2301	if (prio >= 0)
2302		p->prio = prio;
2303	else
2304		p->prio = --least_priority;
2305	/*
2306	 * the plist prio is negated because plist ordering is
2307	 * low-to-high, while swap ordering is high-to-low
2308	 */
2309	p->list.prio = -p->prio;
2310	for_each_node(i) {
2311		if (p->prio >= 0)
2312			p->avail_lists[i].prio = -p->prio;
2313		else {
2314			if (swap_node(p) == i)
2315				p->avail_lists[i].prio = 1;
2316			else
2317				p->avail_lists[i].prio = -p->prio;
2318		}
2319	}
2320	p->swap_map = swap_map;
2321	p->cluster_info = cluster_info;
2322}
2323
2324static void _enable_swap_info(struct swap_info_struct *p)
2325{
2326	p->flags |= SWP_WRITEOK;
2327	atomic_long_add(p->pages, &nr_swap_pages);
2328	total_swap_pages += p->pages;
2329
2330	assert_spin_locked(&swap_lock);
2331	/*
2332	 * both lists are plists, and thus priority ordered.
2333	 * swap_active_head needs to be priority ordered for swapoff(),
2334	 * which on removal of any swap_info_struct with an auto-assigned
2335	 * (i.e. negative) priority increments the auto-assigned priority
2336	 * of any lower-priority swap_info_structs.
2337	 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2338	 * which allocates swap pages from the highest available priority
2339	 * swap_info_struct.
2340	 */
2341	plist_add(&p->list, &swap_active_head);
2342	add_to_avail_list(p);
2343}
2344
2345static void enable_swap_info(struct swap_info_struct *p, int prio,
2346				unsigned char *swap_map,
2347				struct swap_cluster_info *cluster_info,
2348				unsigned long *frontswap_map)
2349{
2350	if (IS_ENABLED(CONFIG_FRONTSWAP))
2351		frontswap_init(p->type, frontswap_map);
2352	spin_lock(&swap_lock);
2353	spin_lock(&p->lock);
2354	setup_swap_info(p, prio, swap_map, cluster_info);
2355	spin_unlock(&p->lock);
2356	spin_unlock(&swap_lock);
2357	/*
2358	 * Finished initializing swap device, now it's safe to reference it.
2359	 */
2360	percpu_ref_resurrect(&p->users);
2361	spin_lock(&swap_lock);
2362	spin_lock(&p->lock);
2363	_enable_swap_info(p);
2364	spin_unlock(&p->lock);
2365	spin_unlock(&swap_lock);
2366}
2367
2368static void reinsert_swap_info(struct swap_info_struct *p)
2369{
2370	spin_lock(&swap_lock);
2371	spin_lock(&p->lock);
2372	setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2373	_enable_swap_info(p);
2374	spin_unlock(&p->lock);
2375	spin_unlock(&swap_lock);
2376}
2377
2378bool has_usable_swap(void)
2379{
2380	bool ret = true;
2381
2382	spin_lock(&swap_lock);
2383	if (plist_head_empty(&swap_active_head))
2384		ret = false;
2385	spin_unlock(&swap_lock);
2386	return ret;
2387}
2388
2389SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2390{
2391	struct swap_info_struct *p = NULL;
2392	unsigned char *swap_map;
2393	struct swap_cluster_info *cluster_info;
2394	unsigned long *frontswap_map;
2395	struct file *swap_file, *victim;
2396	struct address_space *mapping;
2397	struct inode *inode;
2398	struct filename *pathname;
2399	int err, found = 0;
2400	unsigned int old_block_size;
 
2401
2402	if (!capable(CAP_SYS_ADMIN))
2403		return -EPERM;
2404
2405	BUG_ON(!current->mm);
2406
2407	pathname = getname(specialfile);
 
2408	if (IS_ERR(pathname))
2409		return PTR_ERR(pathname);
2410
2411	victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
 
2412	err = PTR_ERR(victim);
2413	if (IS_ERR(victim))
2414		goto out;
2415
2416	mapping = victim->f_mapping;
 
2417	spin_lock(&swap_lock);
2418	plist_for_each_entry(p, &swap_active_head, list) {
 
2419		if (p->flags & SWP_WRITEOK) {
2420			if (p->swap_file->f_mapping == mapping) {
2421				found = 1;
2422				break;
2423			}
2424		}
 
2425	}
2426	if (!found) {
2427		err = -EINVAL;
2428		spin_unlock(&swap_lock);
2429		goto out_dput;
2430	}
2431	if (!security_vm_enough_memory_mm(current->mm, p->pages))
2432		vm_unacct_memory(p->pages);
2433	else {
2434		err = -ENOMEM;
2435		spin_unlock(&swap_lock);
2436		goto out_dput;
2437	}
2438	del_from_avail_list(p);
2439	spin_lock(&p->lock);
 
 
 
 
 
 
2440	if (p->prio < 0) {
2441		struct swap_info_struct *si = p;
2442		int nid;
2443
2444		plist_for_each_entry_continue(si, &swap_active_head, list) {
2445			si->prio++;
2446			si->list.prio--;
2447			for_each_node(nid) {
2448				if (si->avail_lists[nid].prio != 1)
2449					si->avail_lists[nid].prio--;
2450			}
2451		}
2452		least_priority++;
2453	}
2454	plist_del(&p->list, &swap_active_head);
2455	atomic_long_sub(p->pages, &nr_swap_pages);
2456	total_swap_pages -= p->pages;
2457	p->flags &= ~SWP_WRITEOK;
2458	spin_unlock(&p->lock);
2459	spin_unlock(&swap_lock);
2460
2461	disable_swap_slots_cache_lock();
2462
2463	set_current_oom_origin();
2464	err = try_to_unuse(p->type);
2465	clear_current_oom_origin();
2466
2467	if (err) {
 
 
 
 
 
 
2468		/* re-insert swap space back into swap_list */
2469		reinsert_swap_info(p);
2470		reenable_swap_slots_cache_unlock();
2471		goto out_dput;
2472	}
2473
2474	reenable_swap_slots_cache_unlock();
2475
2476	/*
2477	 * Wait for swap operations protected by get/put_swap_device()
2478	 * to complete.
2479	 *
2480	 * We need synchronize_rcu() here to protect the accessing to
2481	 * the swap cache data structure.
2482	 */
2483	percpu_ref_kill(&p->users);
2484	synchronize_rcu();
2485	wait_for_completion(&p->comp);
2486
2487	flush_work(&p->discard_work);
2488
2489	destroy_swap_extents(p);
2490	if (p->flags & SWP_CONTINUED)
2491		free_swap_count_continuations(p);
2492
2493	if (!p->bdev || !bdev_nonrot(p->bdev))
2494		atomic_dec(&nr_rotate_swap);
2495
2496	mutex_lock(&swapon_mutex);
2497	spin_lock(&swap_lock);
2498	spin_lock(&p->lock);
2499	drain_mmlist();
2500
2501	/* wait for anyone still in scan_swap_map_slots */
2502	p->highest_bit = 0;		/* cuts scans short */
2503	while (p->flags >= SWP_SCANNING) {
2504		spin_unlock(&p->lock);
2505		spin_unlock(&swap_lock);
2506		schedule_timeout_uninterruptible(1);
2507		spin_lock(&swap_lock);
2508		spin_lock(&p->lock);
2509	}
2510
2511	swap_file = p->swap_file;
2512	old_block_size = p->old_block_size;
2513	p->swap_file = NULL;
2514	p->max = 0;
2515	swap_map = p->swap_map;
2516	p->swap_map = NULL;
2517	cluster_info = p->cluster_info;
2518	p->cluster_info = NULL;
2519	frontswap_map = frontswap_map_get(p);
2520	spin_unlock(&p->lock);
2521	spin_unlock(&swap_lock);
2522	arch_swap_invalidate_area(p->type);
2523	frontswap_invalidate_area(p->type);
2524	frontswap_map_set(p, NULL);
2525	mutex_unlock(&swapon_mutex);
2526	free_percpu(p->percpu_cluster);
2527	p->percpu_cluster = NULL;
2528	free_percpu(p->cluster_next_cpu);
2529	p->cluster_next_cpu = NULL;
2530	vfree(swap_map);
2531	kvfree(cluster_info);
2532	kvfree(frontswap_map);
2533	/* Destroy swap account information */
2534	swap_cgroup_swapoff(p->type);
2535	exit_swap_address_space(p->type);
2536
2537	inode = mapping->host;
2538	if (S_ISBLK(inode->i_mode)) {
2539		struct block_device *bdev = I_BDEV(inode);
2540
2541		set_blocksize(bdev, old_block_size);
2542		blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
 
 
 
 
2543	}
2544
2545	inode_lock(inode);
2546	inode->i_flags &= ~S_SWAPFILE;
2547	inode_unlock(inode);
2548	filp_close(swap_file, NULL);
2549
2550	/*
2551	 * Clear the SWP_USED flag after all resources are freed so that swapon
2552	 * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2553	 * not hold p->lock after we cleared its SWP_WRITEOK.
2554	 */
2555	spin_lock(&swap_lock);
2556	p->flags = 0;
2557	spin_unlock(&swap_lock);
2558
2559	err = 0;
2560	atomic_inc(&proc_poll_event);
2561	wake_up_interruptible(&proc_poll_wait);
2562
2563out_dput:
2564	filp_close(victim, NULL);
2565out:
2566	putname(pathname);
2567	return err;
2568}
2569
2570#ifdef CONFIG_PROC_FS
2571static __poll_t swaps_poll(struct file *file, poll_table *wait)
2572{
2573	struct seq_file *seq = file->private_data;
2574
2575	poll_wait(file, &proc_poll_wait, wait);
2576
2577	if (seq->poll_event != atomic_read(&proc_poll_event)) {
2578		seq->poll_event = atomic_read(&proc_poll_event);
2579		return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2580	}
2581
2582	return EPOLLIN | EPOLLRDNORM;
2583}
2584
2585/* iterator */
2586static void *swap_start(struct seq_file *swap, loff_t *pos)
2587{
2588	struct swap_info_struct *si;
2589	int type;
2590	loff_t l = *pos;
2591
2592	mutex_lock(&swapon_mutex);
2593
2594	if (!l)
2595		return SEQ_START_TOKEN;
2596
2597	for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
 
 
2598		if (!(si->flags & SWP_USED) || !si->swap_map)
2599			continue;
2600		if (!--l)
2601			return si;
2602	}
2603
2604	return NULL;
2605}
2606
2607static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2608{
2609	struct swap_info_struct *si = v;
2610	int type;
2611
2612	if (v == SEQ_START_TOKEN)
2613		type = 0;
2614	else
2615		type = si->type + 1;
2616
2617	++(*pos);
2618	for (; (si = swap_type_to_swap_info(type)); type++) {
 
2619		if (!(si->flags & SWP_USED) || !si->swap_map)
2620			continue;
 
2621		return si;
2622	}
2623
2624	return NULL;
2625}
2626
2627static void swap_stop(struct seq_file *swap, void *v)
2628{
2629	mutex_unlock(&swapon_mutex);
2630}
2631
2632static int swap_show(struct seq_file *swap, void *v)
2633{
2634	struct swap_info_struct *si = v;
2635	struct file *file;
2636	int len;
2637	unsigned long bytes, inuse;
2638
2639	if (si == SEQ_START_TOKEN) {
2640		seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2641		return 0;
2642	}
2643
2644	bytes = si->pages << (PAGE_SHIFT - 10);
2645	inuse = READ_ONCE(si->inuse_pages) << (PAGE_SHIFT - 10);
2646
2647	file = si->swap_file;
2648	len = seq_file_path(swap, file, " \t\n\\");
2649	seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
2650			len < 40 ? 40 - len : 1, " ",
2651			S_ISBLK(file_inode(file)->i_mode) ?
2652				"partition" : "file\t",
2653			bytes, bytes < 10000000 ? "\t" : "",
2654			inuse, inuse < 10000000 ? "\t" : "",
2655			si->prio);
2656	return 0;
2657}
2658
2659static const struct seq_operations swaps_op = {
2660	.start =	swap_start,
2661	.next =		swap_next,
2662	.stop =		swap_stop,
2663	.show =		swap_show
2664};
2665
2666static int swaps_open(struct inode *inode, struct file *file)
2667{
2668	struct seq_file *seq;
2669	int ret;
2670
2671	ret = seq_open(file, &swaps_op);
2672	if (ret)
2673		return ret;
2674
2675	seq = file->private_data;
2676	seq->poll_event = atomic_read(&proc_poll_event);
2677	return 0;
2678}
2679
2680static const struct proc_ops swaps_proc_ops = {
2681	.proc_flags	= PROC_ENTRY_PERMANENT,
2682	.proc_open	= swaps_open,
2683	.proc_read	= seq_read,
2684	.proc_lseek	= seq_lseek,
2685	.proc_release	= seq_release,
2686	.proc_poll	= swaps_poll,
2687};
2688
2689static int __init procswaps_init(void)
2690{
2691	proc_create("swaps", 0, NULL, &swaps_proc_ops);
2692	return 0;
2693}
2694__initcall(procswaps_init);
2695#endif /* CONFIG_PROC_FS */
2696
2697#ifdef MAX_SWAPFILES_CHECK
2698static int __init max_swapfiles_check(void)
2699{
2700	MAX_SWAPFILES_CHECK();
2701	return 0;
2702}
2703late_initcall(max_swapfiles_check);
2704#endif
2705
2706static struct swap_info_struct *alloc_swap_info(void)
2707{
2708	struct swap_info_struct *p;
2709	struct swap_info_struct *defer = NULL;
2710	unsigned int type;
2711	int i;
2712
2713	p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2714	if (!p)
2715		return ERR_PTR(-ENOMEM);
2716
2717	if (percpu_ref_init(&p->users, swap_users_ref_free,
2718			    PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2719		kvfree(p);
2720		return ERR_PTR(-ENOMEM);
2721	}
2722
2723	spin_lock(&swap_lock);
2724	for (type = 0; type < nr_swapfiles; type++) {
2725		if (!(swap_info[type]->flags & SWP_USED))
2726			break;
2727	}
2728	if (type >= MAX_SWAPFILES) {
2729		spin_unlock(&swap_lock);
2730		percpu_ref_exit(&p->users);
2731		kvfree(p);
2732		return ERR_PTR(-EPERM);
2733	}
2734	if (type >= nr_swapfiles) {
2735		p->type = type;
 
2736		/*
2737		 * Publish the swap_info_struct after initializing it.
2738		 * Note that kvzalloc() above zeroes all its fields.
 
2739		 */
2740		smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2741		nr_swapfiles++;
2742	} else {
2743		defer = p;
2744		p = swap_info[type];
2745		/*
2746		 * Do not memset this entry: a racing procfs swap_next()
2747		 * would be relying on p->type to remain valid.
2748		 */
2749	}
2750	p->swap_extent_root = RB_ROOT;
2751	plist_node_init(&p->list, 0);
2752	for_each_node(i)
2753		plist_node_init(&p->avail_lists[i], 0);
2754	p->flags = SWP_USED;
 
2755	spin_unlock(&swap_lock);
2756	if (defer) {
2757		percpu_ref_exit(&defer->users);
2758		kvfree(defer);
2759	}
2760	spin_lock_init(&p->lock);
2761	spin_lock_init(&p->cont_lock);
2762	init_completion(&p->comp);
2763
2764	return p;
2765}
2766
2767static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2768{
2769	int error;
2770
2771	if (S_ISBLK(inode->i_mode)) {
2772		p->bdev = blkdev_get_by_dev(inode->i_rdev,
2773				   FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2774		if (IS_ERR(p->bdev)) {
2775			error = PTR_ERR(p->bdev);
 
2776			p->bdev = NULL;
2777			return error;
2778		}
2779		p->old_block_size = block_size(p->bdev);
2780		error = set_blocksize(p->bdev, PAGE_SIZE);
2781		if (error < 0)
2782			return error;
2783		/*
2784		 * Zoned block devices contain zones that have a sequential
2785		 * write only restriction.  Hence zoned block devices are not
2786		 * suitable for swapping.  Disallow them here.
2787		 */
2788		if (bdev_is_zoned(p->bdev))
2789			return -EINVAL;
2790		p->flags |= SWP_BLKDEV;
2791	} else if (S_ISREG(inode->i_mode)) {
2792		p->bdev = inode->i_sb->s_bdev;
2793	}
 
 
 
 
2794
2795	return 0;
2796}
2797
2798
2799/*
2800 * Find out how many pages are allowed for a single swap device. There
2801 * are two limiting factors:
2802 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2803 * 2) the number of bits in the swap pte, as defined by the different
2804 * architectures.
2805 *
2806 * In order to find the largest possible bit mask, a swap entry with
2807 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2808 * decoded to a swp_entry_t again, and finally the swap offset is
2809 * extracted.
2810 *
2811 * This will mask all the bits from the initial ~0UL mask that can't
2812 * be encoded in either the swp_entry_t or the architecture definition
2813 * of a swap pte.
2814 */
2815unsigned long generic_max_swapfile_size(void)
2816{
2817	return swp_offset(pte_to_swp_entry(
2818			swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2819}
2820
2821/* Can be overridden by an architecture for additional checks. */
2822__weak unsigned long arch_max_swapfile_size(void)
2823{
2824	return generic_max_swapfile_size();
2825}
2826
2827static unsigned long read_swap_header(struct swap_info_struct *p,
2828					union swap_header *swap_header,
2829					struct inode *inode)
2830{
2831	int i;
2832	unsigned long maxpages;
2833	unsigned long swapfilepages;
2834	unsigned long last_page;
2835
2836	if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2837		pr_err("Unable to find swap-space signature\n");
2838		return 0;
2839	}
2840
2841	/* swap partition endianness hack... */
2842	if (swab32(swap_header->info.version) == 1) {
2843		swab32s(&swap_header->info.version);
2844		swab32s(&swap_header->info.last_page);
2845		swab32s(&swap_header->info.nr_badpages);
2846		if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2847			return 0;
2848		for (i = 0; i < swap_header->info.nr_badpages; i++)
2849			swab32s(&swap_header->info.badpages[i]);
2850	}
2851	/* Check the swap header's sub-version */
2852	if (swap_header->info.version != 1) {
2853		pr_warn("Unable to handle swap header version %d\n",
2854			swap_header->info.version);
 
2855		return 0;
2856	}
2857
2858	p->lowest_bit  = 1;
2859	p->cluster_next = 1;
2860	p->cluster_nr = 0;
2861
2862	maxpages = swapfile_maximum_size;
2863	last_page = swap_header->info.last_page;
2864	if (!last_page) {
2865		pr_warn("Empty swap-file\n");
2866		return 0;
2867	}
2868	if (last_page > maxpages) {
2869		pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2870			maxpages << (PAGE_SHIFT - 10),
2871			last_page << (PAGE_SHIFT - 10));
2872	}
2873	if (maxpages > last_page) {
2874		maxpages = last_page + 1;
 
 
 
 
 
 
 
 
 
2875		/* p->max is an unsigned int: don't overflow it */
2876		if ((unsigned int)maxpages == 0)
2877			maxpages = UINT_MAX;
2878	}
2879	p->highest_bit = maxpages - 1;
2880
2881	if (!maxpages)
2882		return 0;
2883	swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2884	if (swapfilepages && maxpages > swapfilepages) {
2885		pr_warn("Swap area shorter than signature indicates\n");
 
2886		return 0;
2887	}
2888	if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2889		return 0;
2890	if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2891		return 0;
2892
2893	return maxpages;
2894}
2895
2896#define SWAP_CLUSTER_INFO_COLS						\
2897	DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2898#define SWAP_CLUSTER_SPACE_COLS						\
2899	DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2900#define SWAP_CLUSTER_COLS						\
2901	max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2902
2903static int setup_swap_map_and_extents(struct swap_info_struct *p,
2904					union swap_header *swap_header,
2905					unsigned char *swap_map,
2906					struct swap_cluster_info *cluster_info,
2907					unsigned long maxpages,
2908					sector_t *span)
2909{
2910	unsigned int j, k;
2911	unsigned int nr_good_pages;
2912	int nr_extents;
2913	unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2914	unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2915	unsigned long i, idx;
2916
2917	nr_good_pages = maxpages - 1;	/* omit header page */
2918
2919	cluster_list_init(&p->free_clusters);
2920	cluster_list_init(&p->discard_clusters);
2921
2922	for (i = 0; i < swap_header->info.nr_badpages; i++) {
2923		unsigned int page_nr = swap_header->info.badpages[i];
2924		if (page_nr == 0 || page_nr > swap_header->info.last_page)
2925			return -EINVAL;
2926		if (page_nr < maxpages) {
2927			swap_map[page_nr] = SWAP_MAP_BAD;
2928			nr_good_pages--;
2929			/*
2930			 * Haven't marked the cluster free yet, no list
2931			 * operation involved
2932			 */
2933			inc_cluster_info_page(p, cluster_info, page_nr);
2934		}
2935	}
2936
2937	/* Haven't marked the cluster free yet, no list operation involved */
2938	for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2939		inc_cluster_info_page(p, cluster_info, i);
2940
2941	if (nr_good_pages) {
2942		swap_map[0] = SWAP_MAP_BAD;
2943		/*
2944		 * Not mark the cluster free yet, no list
2945		 * operation involved
2946		 */
2947		inc_cluster_info_page(p, cluster_info, 0);
2948		p->max = maxpages;
2949		p->pages = nr_good_pages;
2950		nr_extents = setup_swap_extents(p, span);
2951		if (nr_extents < 0)
2952			return nr_extents;
2953		nr_good_pages = p->pages;
2954	}
2955	if (!nr_good_pages) {
2956		pr_warn("Empty swap-file\n");
2957		return -EINVAL;
2958	}
2959
2960	if (!cluster_info)
2961		return nr_extents;
2962
2963
2964	/*
2965	 * Reduce false cache line sharing between cluster_info and
2966	 * sharing same address space.
2967	 */
2968	for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2969		j = (k + col) % SWAP_CLUSTER_COLS;
2970		for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2971			idx = i * SWAP_CLUSTER_COLS + j;
2972			if (idx >= nr_clusters)
2973				continue;
2974			if (cluster_count(&cluster_info[idx]))
2975				continue;
2976			cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2977			cluster_list_add_tail(&p->free_clusters, cluster_info,
2978					      idx);
2979		}
2980	}
2981	return nr_extents;
2982}
2983
2984SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2985{
2986	struct swap_info_struct *p;
2987	struct filename *name;
2988	struct file *swap_file = NULL;
2989	struct address_space *mapping;
2990	struct dentry *dentry;
2991	int prio;
2992	int error;
2993	union swap_header *swap_header;
2994	int nr_extents;
2995	sector_t span;
2996	unsigned long maxpages;
2997	unsigned char *swap_map = NULL;
2998	struct swap_cluster_info *cluster_info = NULL;
2999	unsigned long *frontswap_map = NULL;
3000	struct page *page = NULL;
3001	struct inode *inode = NULL;
3002	bool inced_nr_rotate_swap = false;
3003
3004	if (swap_flags & ~SWAP_FLAGS_VALID)
3005		return -EINVAL;
3006
3007	if (!capable(CAP_SYS_ADMIN))
3008		return -EPERM;
3009
3010	if (!swap_avail_heads)
3011		return -ENOMEM;
3012
3013	p = alloc_swap_info();
3014	if (IS_ERR(p))
3015		return PTR_ERR(p);
3016
3017	INIT_WORK(&p->discard_work, swap_discard_work);
3018
3019	name = getname(specialfile);
3020	if (IS_ERR(name)) {
3021		error = PTR_ERR(name);
3022		name = NULL;
3023		goto bad_swap;
3024	}
3025	swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3026	if (IS_ERR(swap_file)) {
3027		error = PTR_ERR(swap_file);
3028		swap_file = NULL;
3029		goto bad_swap;
3030	}
3031
3032	p->swap_file = swap_file;
3033	mapping = swap_file->f_mapping;
3034	dentry = swap_file->f_path.dentry;
3035	inode = mapping->host;
3036
 
 
 
 
 
 
 
 
 
 
 
 
 
3037	error = claim_swapfile(p, inode);
3038	if (unlikely(error))
3039		goto bad_swap;
3040
3041	inode_lock(inode);
3042	if (d_unlinked(dentry) || cant_mount(dentry)) {
3043		error = -ENOENT;
3044		goto bad_swap_unlock_inode;
3045	}
3046	if (IS_SWAPFILE(inode)) {
3047		error = -EBUSY;
3048		goto bad_swap_unlock_inode;
3049	}
3050
3051	/*
3052	 * Read the swap header.
3053	 */
3054	if (!mapping->a_ops->read_folio) {
3055		error = -EINVAL;
3056		goto bad_swap_unlock_inode;
3057	}
3058	page = read_mapping_page(mapping, 0, swap_file);
3059	if (IS_ERR(page)) {
3060		error = PTR_ERR(page);
3061		goto bad_swap_unlock_inode;
3062	}
3063	swap_header = kmap(page);
3064
3065	maxpages = read_swap_header(p, swap_header, inode);
3066	if (unlikely(!maxpages)) {
3067		error = -EINVAL;
3068		goto bad_swap_unlock_inode;
3069	}
3070
3071	/* OK, set up the swap map and apply the bad block list */
3072	swap_map = vzalloc(maxpages);
3073	if (!swap_map) {
3074		error = -ENOMEM;
3075		goto bad_swap_unlock_inode;
3076	}
3077
3078	if (p->bdev && bdev_stable_writes(p->bdev))
3079		p->flags |= SWP_STABLE_WRITES;
3080
3081	if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3082		p->flags |= SWP_SYNCHRONOUS_IO;
3083
3084	if (p->bdev && bdev_nonrot(p->bdev)) {
3085		int cpu;
3086		unsigned long ci, nr_cluster;
3087
3088		p->flags |= SWP_SOLIDSTATE;
3089		p->cluster_next_cpu = alloc_percpu(unsigned int);
3090		if (!p->cluster_next_cpu) {
3091			error = -ENOMEM;
3092			goto bad_swap_unlock_inode;
3093		}
3094		/*
3095		 * select a random position to start with to help wear leveling
3096		 * SSD
3097		 */
3098		for_each_possible_cpu(cpu) {
3099			per_cpu(*p->cluster_next_cpu, cpu) =
3100				get_random_u32_inclusive(1, p->highest_bit);
3101		}
3102		nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3103
3104		cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3105					GFP_KERNEL);
3106		if (!cluster_info) {
3107			error = -ENOMEM;
3108			goto bad_swap_unlock_inode;
3109		}
3110
3111		for (ci = 0; ci < nr_cluster; ci++)
3112			spin_lock_init(&((cluster_info + ci)->lock));
3113
3114		p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3115		if (!p->percpu_cluster) {
3116			error = -ENOMEM;
3117			goto bad_swap_unlock_inode;
3118		}
3119		for_each_possible_cpu(cpu) {
3120			struct percpu_cluster *cluster;
3121			cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3122			cluster_set_null(&cluster->index);
3123		}
3124	} else {
3125		atomic_inc(&nr_rotate_swap);
3126		inced_nr_rotate_swap = true;
3127	}
3128
3129	error = swap_cgroup_swapon(p->type, maxpages);
3130	if (error)
3131		goto bad_swap_unlock_inode;
3132
3133	nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3134		cluster_info, maxpages, &span);
3135	if (unlikely(nr_extents < 0)) {
3136		error = nr_extents;
3137		goto bad_swap_unlock_inode;
3138	}
3139	/* frontswap enabled? set up bit-per-page map for frontswap */
3140	if (IS_ENABLED(CONFIG_FRONTSWAP))
3141		frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3142					 sizeof(long),
3143					 GFP_KERNEL);
3144
3145	if ((swap_flags & SWAP_FLAG_DISCARD) &&
3146	    p->bdev && bdev_max_discard_sectors(p->bdev)) {
3147		/*
3148		 * When discard is enabled for swap with no particular
3149		 * policy flagged, we set all swap discard flags here in
3150		 * order to sustain backward compatibility with older
3151		 * swapon(8) releases.
3152		 */
3153		p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3154			     SWP_PAGE_DISCARD);
3155
3156		/*
3157		 * By flagging sys_swapon, a sysadmin can tell us to
3158		 * either do single-time area discards only, or to just
3159		 * perform discards for released swap page-clusters.
3160		 * Now it's time to adjust the p->flags accordingly.
3161		 */
3162		if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3163			p->flags &= ~SWP_PAGE_DISCARD;
3164		else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3165			p->flags &= ~SWP_AREA_DISCARD;
3166
3167		/* issue a swapon-time discard if it's still required */
3168		if (p->flags & SWP_AREA_DISCARD) {
3169			int err = discard_swap(p);
3170			if (unlikely(err))
3171				pr_err("swapon: discard_swap(%p): %d\n",
3172					p, err);
3173		}
3174	}
3175
3176	error = init_swap_address_space(p->type, maxpages);
3177	if (error)
3178		goto bad_swap_unlock_inode;
3179
3180	/*
3181	 * Flush any pending IO and dirty mappings before we start using this
3182	 * swap device.
3183	 */
3184	inode->i_flags |= S_SWAPFILE;
3185	error = inode_drain_writes(inode);
3186	if (error) {
3187		inode->i_flags &= ~S_SWAPFILE;
3188		goto free_swap_address_space;
3189	}
3190
3191	mutex_lock(&swapon_mutex);
3192	prio = -1;
3193	if (swap_flags & SWAP_FLAG_PREFER)
3194		prio =
3195		  (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3196	enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3197
3198	pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3199		p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
 
3200		nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3201		(p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3202		(p->flags & SWP_DISCARDABLE) ? "D" : "",
3203		(p->flags & SWP_AREA_DISCARD) ? "s" : "",
3204		(p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3205		(frontswap_map) ? "FS" : "");
3206
3207	mutex_unlock(&swapon_mutex);
3208	atomic_inc(&proc_poll_event);
3209	wake_up_interruptible(&proc_poll_wait);
3210
 
 
3211	error = 0;
3212	goto out;
3213free_swap_address_space:
3214	exit_swap_address_space(p->type);
3215bad_swap_unlock_inode:
3216	inode_unlock(inode);
3217bad_swap:
3218	free_percpu(p->percpu_cluster);
3219	p->percpu_cluster = NULL;
3220	free_percpu(p->cluster_next_cpu);
3221	p->cluster_next_cpu = NULL;
3222	if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3223		set_blocksize(p->bdev, p->old_block_size);
3224		blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3225	}
3226	inode = NULL;
3227	destroy_swap_extents(p);
3228	swap_cgroup_swapoff(p->type);
3229	spin_lock(&swap_lock);
3230	p->swap_file = NULL;
3231	p->flags = 0;
3232	spin_unlock(&swap_lock);
3233	vfree(swap_map);
3234	kvfree(cluster_info);
3235	kvfree(frontswap_map);
3236	if (inced_nr_rotate_swap)
3237		atomic_dec(&nr_rotate_swap);
3238	if (swap_file)
3239		filp_close(swap_file, NULL);
 
3240out:
3241	if (page && !IS_ERR(page)) {
3242		kunmap(page);
3243		put_page(page);
3244	}
3245	if (name)
3246		putname(name);
3247	if (inode)
3248		inode_unlock(inode);
3249	if (!error)
3250		enable_swap_slots_cache();
3251	return error;
3252}
3253
3254void si_swapinfo(struct sysinfo *val)
3255{
3256	unsigned int type;
3257	unsigned long nr_to_be_unused = 0;
3258
3259	spin_lock(&swap_lock);
3260	for (type = 0; type < nr_swapfiles; type++) {
3261		struct swap_info_struct *si = swap_info[type];
3262
3263		if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3264			nr_to_be_unused += READ_ONCE(si->inuse_pages);
3265	}
3266	val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3267	val->totalswap = total_swap_pages + nr_to_be_unused;
3268	spin_unlock(&swap_lock);
3269}
3270
3271/*
3272 * Verify that a swap entry is valid and increment its swap map count.
3273 *
3274 * Returns error code in following case.
3275 * - success -> 0
3276 * - swp_entry is invalid -> EINVAL
3277 * - swp_entry is migration entry -> EINVAL
3278 * - swap-cache reference is requested but there is already one. -> EEXIST
3279 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3280 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3281 */
3282static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3283{
3284	struct swap_info_struct *p;
3285	struct swap_cluster_info *ci;
3286	unsigned long offset;
3287	unsigned char count;
3288	unsigned char has_cache;
3289	int err;
3290
3291	p = get_swap_device(entry);
3292	if (!p)
3293		return -EINVAL;
3294
 
 
 
 
3295	offset = swp_offset(entry);
3296	ci = lock_cluster_or_swap_info(p, offset);
3297
3298	count = p->swap_map[offset];
3299
3300	/*
3301	 * swapin_readahead() doesn't check if a swap entry is valid, so the
3302	 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3303	 */
3304	if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3305		err = -ENOENT;
3306		goto unlock_out;
3307	}
3308
 
3309	has_cache = count & SWAP_HAS_CACHE;
3310	count &= ~SWAP_HAS_CACHE;
3311	err = 0;
3312
3313	if (usage == SWAP_HAS_CACHE) {
3314
3315		/* set SWAP_HAS_CACHE if there is no cache and entry is used */
3316		if (!has_cache && count)
3317			has_cache = SWAP_HAS_CACHE;
3318		else if (has_cache)		/* someone else added cache */
3319			err = -EEXIST;
3320		else				/* no users remaining */
3321			err = -ENOENT;
3322
3323	} else if (count || has_cache) {
3324
3325		if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3326			count += usage;
3327		else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3328			err = -EINVAL;
3329		else if (swap_count_continued(p, offset, count))
3330			count = COUNT_CONTINUED;
3331		else
3332			err = -ENOMEM;
3333	} else
3334		err = -ENOENT;			/* unused swap entry */
3335
3336	WRITE_ONCE(p->swap_map[offset], count | has_cache);
3337
3338unlock_out:
3339	unlock_cluster_or_swap_info(p, ci);
3340	put_swap_device(p);
3341	return err;
 
 
 
 
3342}
3343
3344/*
3345 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3346 * (in which case its reference count is never incremented).
3347 */
3348void swap_shmem_alloc(swp_entry_t entry)
3349{
3350	__swap_duplicate(entry, SWAP_MAP_SHMEM);
3351}
3352
3353/*
3354 * Increase reference count of swap entry by 1.
3355 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3356 * but could not be atomically allocated.  Returns 0, just as if it succeeded,
3357 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3358 * might occur if a page table entry has got corrupted.
3359 */
3360int swap_duplicate(swp_entry_t entry)
3361{
3362	int err = 0;
3363
3364	while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3365		err = add_swap_count_continuation(entry, GFP_ATOMIC);
3366	return err;
3367}
3368
3369/*
3370 * @entry: swap entry for which we allocate swap cache.
3371 *
3372 * Called when allocating swap cache for existing swap entry,
3373 * This can return error codes. Returns 0 at success.
3374 * -EEXIST means there is a swap cache.
3375 * Note: return code is different from swap_duplicate().
3376 */
3377int swapcache_prepare(swp_entry_t entry)
3378{
3379	return __swap_duplicate(entry, SWAP_HAS_CACHE);
3380}
3381
3382struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3383{
3384	return swap_type_to_swap_info(swp_type(entry));
3385}
3386
3387struct swap_info_struct *page_swap_info(struct page *page)
3388{
3389	swp_entry_t entry = { .val = page_private(page) };
3390	return swp_swap_info(entry);
3391}
3392
3393/*
3394 * out-of-line methods to avoid include hell.
 
3395 */
3396struct address_space *swapcache_mapping(struct folio *folio)
3397{
3398	return page_swap_info(&folio->page)->swap_file->f_mapping;
3399}
3400EXPORT_SYMBOL_GPL(swapcache_mapping);
 
 
3401
3402pgoff_t __page_file_index(struct page *page)
3403{
3404	swp_entry_t swap = { .val = page_private(page) };
3405	return swp_offset(swap);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3406}
3407EXPORT_SYMBOL_GPL(__page_file_index);
3408
3409/*
3410 * add_swap_count_continuation - called when a swap count is duplicated
3411 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3412 * page of the original vmalloc'ed swap_map, to hold the continuation count
3413 * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
3414 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3415 *
3416 * These continuation pages are seldom referenced: the common paths all work
3417 * on the original swap_map, only referring to a continuation page when the
3418 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3419 *
3420 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3421 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3422 * can be called after dropping locks.
3423 */
3424int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3425{
3426	struct swap_info_struct *si;
3427	struct swap_cluster_info *ci;
3428	struct page *head;
3429	struct page *page;
3430	struct page *list_page;
3431	pgoff_t offset;
3432	unsigned char count;
3433	int ret = 0;
3434
3435	/*
3436	 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3437	 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3438	 */
3439	page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3440
3441	si = get_swap_device(entry);
3442	if (!si) {
3443		/*
3444		 * An acceptable race has occurred since the failing
3445		 * __swap_duplicate(): the swap device may be swapoff
 
3446		 */
3447		goto outer;
3448	}
3449	spin_lock(&si->lock);
3450
3451	offset = swp_offset(entry);
3452
3453	ci = lock_cluster(si, offset);
3454
3455	count = swap_count(si->swap_map[offset]);
3456
3457	if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3458		/*
3459		 * The higher the swap count, the more likely it is that tasks
3460		 * will race to add swap count continuation: we need to avoid
3461		 * over-provisioning.
3462		 */
3463		goto out;
3464	}
3465
3466	if (!page) {
3467		ret = -ENOMEM;
3468		goto out;
3469	}
3470
3471	/*
3472	 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3473	 * no architecture is using highmem pages for kernel page tables: so it
3474	 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3475	 */
3476	head = vmalloc_to_page(si->swap_map + offset);
3477	offset &= ~PAGE_MASK;
3478
3479	spin_lock(&si->cont_lock);
3480	/*
3481	 * Page allocation does not initialize the page's lru field,
3482	 * but it does always reset its private field.
3483	 */
3484	if (!page_private(head)) {
3485		BUG_ON(count & COUNT_CONTINUED);
3486		INIT_LIST_HEAD(&head->lru);
3487		set_page_private(head, SWP_CONTINUED);
3488		si->flags |= SWP_CONTINUED;
3489	}
3490
3491	list_for_each_entry(list_page, &head->lru, lru) {
3492		unsigned char *map;
3493
3494		/*
3495		 * If the previous map said no continuation, but we've found
3496		 * a continuation page, free our allocation and use this one.
3497		 */
3498		if (!(count & COUNT_CONTINUED))
3499			goto out_unlock_cont;
3500
3501		map = kmap_atomic(list_page) + offset;
3502		count = *map;
3503		kunmap_atomic(map);
3504
3505		/*
3506		 * If this continuation count now has some space in it,
3507		 * free our allocation and use this one.
3508		 */
3509		if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3510			goto out_unlock_cont;
3511	}
3512
3513	list_add_tail(&page->lru, &head->lru);
3514	page = NULL;			/* now it's attached, don't free it */
3515out_unlock_cont:
3516	spin_unlock(&si->cont_lock);
3517out:
3518	unlock_cluster(ci);
3519	spin_unlock(&si->lock);
3520	put_swap_device(si);
3521outer:
3522	if (page)
3523		__free_page(page);
3524	return ret;
3525}
3526
3527/*
3528 * swap_count_continued - when the original swap_map count is incremented
3529 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3530 * into, carry if so, or else fail until a new continuation page is allocated;
3531 * when the original swap_map count is decremented from 0 with continuation,
3532 * borrow from the continuation and report whether it still holds more.
3533 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3534 * lock.
3535 */
3536static bool swap_count_continued(struct swap_info_struct *si,
3537				 pgoff_t offset, unsigned char count)
3538{
3539	struct page *head;
3540	struct page *page;
3541	unsigned char *map;
3542	bool ret;
3543
3544	head = vmalloc_to_page(si->swap_map + offset);
3545	if (page_private(head) != SWP_CONTINUED) {
3546		BUG_ON(count & COUNT_CONTINUED);
3547		return false;		/* need to add count continuation */
3548	}
3549
3550	spin_lock(&si->cont_lock);
3551	offset &= ~PAGE_MASK;
3552	page = list_next_entry(head, lru);
3553	map = kmap_atomic(page) + offset;
3554
3555	if (count == SWAP_MAP_MAX)	/* initial increment from swap_map */
3556		goto init_map;		/* jump over SWAP_CONT_MAX checks */
3557
3558	if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3559		/*
3560		 * Think of how you add 1 to 999
3561		 */
3562		while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3563			kunmap_atomic(map);
3564			page = list_next_entry(page, lru);
3565			BUG_ON(page == head);
3566			map = kmap_atomic(page) + offset;
3567		}
3568		if (*map == SWAP_CONT_MAX) {
3569			kunmap_atomic(map);
3570			page = list_next_entry(page, lru);
3571			if (page == head) {
3572				ret = false;	/* add count continuation */
3573				goto out;
3574			}
3575			map = kmap_atomic(page) + offset;
3576init_map:		*map = 0;		/* we didn't zero the page */
3577		}
3578		*map += 1;
3579		kunmap_atomic(map);
3580		while ((page = list_prev_entry(page, lru)) != head) {
3581			map = kmap_atomic(page) + offset;
 
3582			*map = COUNT_CONTINUED;
3583			kunmap_atomic(map);
 
3584		}
3585		ret = true;			/* incremented */
3586
3587	} else {				/* decrementing */
3588		/*
3589		 * Think of how you subtract 1 from 1000
3590		 */
3591		BUG_ON(count != COUNT_CONTINUED);
3592		while (*map == COUNT_CONTINUED) {
3593			kunmap_atomic(map);
3594			page = list_next_entry(page, lru);
3595			BUG_ON(page == head);
3596			map = kmap_atomic(page) + offset;
3597		}
3598		BUG_ON(*map == 0);
3599		*map -= 1;
3600		if (*map == 0)
3601			count = 0;
3602		kunmap_atomic(map);
3603		while ((page = list_prev_entry(page, lru)) != head) {
3604			map = kmap_atomic(page) + offset;
 
3605			*map = SWAP_CONT_MAX | count;
3606			count = COUNT_CONTINUED;
3607			kunmap_atomic(map);
 
3608		}
3609		ret = count == COUNT_CONTINUED;
3610	}
3611out:
3612	spin_unlock(&si->cont_lock);
3613	return ret;
3614}
3615
3616/*
3617 * free_swap_count_continuations - swapoff free all the continuation pages
3618 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3619 */
3620static void free_swap_count_continuations(struct swap_info_struct *si)
3621{
3622	pgoff_t offset;
3623
3624	for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3625		struct page *head;
3626		head = vmalloc_to_page(si->swap_map + offset);
3627		if (page_private(head)) {
3628			struct page *page, *next;
3629
3630			list_for_each_entry_safe(page, next, &head->lru, lru) {
3631				list_del(&page->lru);
 
3632				__free_page(page);
3633			}
3634		}
3635	}
3636}
3637
3638#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3639void __cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3640{
3641	struct swap_info_struct *si, *next;
3642	int nid = page_to_nid(page);
3643
3644	if (!(gfp_mask & __GFP_IO))
3645		return;
3646
3647	if (!blk_cgroup_congested())
3648		return;
3649
3650	/*
3651	 * We've already scheduled a throttle, avoid taking the global swap
3652	 * lock.
3653	 */
3654	if (current->throttle_queue)
3655		return;
3656
3657	spin_lock(&swap_avail_lock);
3658	plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3659				  avail_lists[nid]) {
3660		if (si->bdev) {
3661			blkcg_schedule_throttle(si->bdev->bd_disk, true);
3662			break;
3663		}
3664	}
3665	spin_unlock(&swap_avail_lock);
3666}
3667#endif
3668
3669static int __init swapfile_init(void)
3670{
3671	int nid;
3672
3673	swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3674					 GFP_KERNEL);
3675	if (!swap_avail_heads) {
3676		pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3677		return -ENOMEM;
3678	}
3679
3680	for_each_node(nid)
3681		plist_head_init(&swap_avail_heads[nid]);
3682
3683	swapfile_maximum_size = arch_max_swapfile_size();
3684
3685#ifdef CONFIG_MIGRATION
3686	if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
3687		swap_migration_ad_supported = true;
3688#endif	/* CONFIG_MIGRATION */
3689
3690	return 0;
3691}
3692subsys_initcall(swapfile_init);
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}