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