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 *);
56static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry,
57 unsigned int nr_pages);
58static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
59 unsigned int nr_entries);
60static bool folio_swapcache_freeable(struct folio *folio);
61static struct swap_cluster_info *lock_cluster_or_swap_info(
62 struct swap_info_struct *si, unsigned long offset);
63static void unlock_cluster_or_swap_info(struct swap_info_struct *si,
64 struct swap_cluster_info *ci);
65
66static DEFINE_SPINLOCK(swap_lock);
67static unsigned int nr_swapfiles;
68atomic_long_t nr_swap_pages;
69/*
70 * Some modules use swappable objects and may try to swap them out under
71 * memory pressure (via the shrinker). Before doing so, they may wish to
72 * check to see if any swap space is available.
73 */
74EXPORT_SYMBOL_GPL(nr_swap_pages);
75/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
76long total_swap_pages;
77static int least_priority = -1;
78unsigned long swapfile_maximum_size;
79#ifdef CONFIG_MIGRATION
80bool swap_migration_ad_supported;
81#endif /* CONFIG_MIGRATION */
82
83static const char Bad_file[] = "Bad swap file entry ";
84static const char Unused_file[] = "Unused swap file entry ";
85static const char Bad_offset[] = "Bad swap offset entry ";
86static const char Unused_offset[] = "Unused swap offset entry ";
87
88/*
89 * all active swap_info_structs
90 * protected with swap_lock, and ordered by priority.
91 */
92static PLIST_HEAD(swap_active_head);
93
94/*
95 * all available (active, not full) swap_info_structs
96 * protected with swap_avail_lock, ordered by priority.
97 * This is used by folio_alloc_swap() instead of swap_active_head
98 * because swap_active_head includes all swap_info_structs,
99 * but folio_alloc_swap() doesn't need to look at full ones.
100 * This uses its own lock instead of swap_lock because when a
101 * swap_info_struct changes between not-full/full, it needs to
102 * add/remove itself to/from this list, but the swap_info_struct->lock
103 * is held and the locking order requires swap_lock to be taken
104 * before any swap_info_struct->lock.
105 */
106static struct plist_head *swap_avail_heads;
107static DEFINE_SPINLOCK(swap_avail_lock);
108
109static struct swap_info_struct *swap_info[MAX_SWAPFILES];
110
111static DEFINE_MUTEX(swapon_mutex);
112
113static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
114/* Activity counter to indicate that a swapon or swapoff has occurred */
115static atomic_t proc_poll_event = ATOMIC_INIT(0);
116
117atomic_t nr_rotate_swap = ATOMIC_INIT(0);
118
119static struct swap_info_struct *swap_type_to_swap_info(int type)
120{
121 if (type >= MAX_SWAPFILES)
122 return NULL;
123
124 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
125}
126
127static inline unsigned char swap_count(unsigned char ent)
128{
129 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
130}
131
132/* Reclaim the swap entry anyway if possible */
133#define TTRS_ANYWAY 0x1
134/*
135 * Reclaim the swap entry if there are no more mappings of the
136 * corresponding page
137 */
138#define TTRS_UNMAPPED 0x2
139/* Reclaim the swap entry if swap is getting full */
140#define TTRS_FULL 0x4
141/* Reclaim directly, bypass the slot cache and don't touch device lock */
142#define TTRS_DIRECT 0x8
143
144static bool swap_is_has_cache(struct swap_info_struct *si,
145 unsigned long offset, int nr_pages)
146{
147 unsigned char *map = si->swap_map + offset;
148 unsigned char *map_end = map + nr_pages;
149
150 do {
151 VM_BUG_ON(!(*map & SWAP_HAS_CACHE));
152 if (*map != SWAP_HAS_CACHE)
153 return false;
154 } while (++map < map_end);
155
156 return true;
157}
158
159static bool swap_is_last_map(struct swap_info_struct *si,
160 unsigned long offset, int nr_pages, bool *has_cache)
161{
162 unsigned char *map = si->swap_map + offset;
163 unsigned char *map_end = map + nr_pages;
164 unsigned char count = *map;
165
166 if (swap_count(count) != 1)
167 return false;
168
169 while (++map < map_end) {
170 if (*map != count)
171 return false;
172 }
173
174 *has_cache = !!(count & SWAP_HAS_CACHE);
175 return true;
176}
177
178/*
179 * returns number of pages in the folio that backs the swap entry. If positive,
180 * the folio was reclaimed. If negative, the folio was not reclaimed. If 0, no
181 * folio was associated with the swap entry.
182 */
183static int __try_to_reclaim_swap(struct swap_info_struct *si,
184 unsigned long offset, unsigned long flags)
185{
186 swp_entry_t entry = swp_entry(si->type, offset);
187 struct address_space *address_space = swap_address_space(entry);
188 struct swap_cluster_info *ci;
189 struct folio *folio;
190 int ret, nr_pages;
191 bool need_reclaim;
192
193 folio = filemap_get_folio(address_space, swap_cache_index(entry));
194 if (IS_ERR(folio))
195 return 0;
196
197 nr_pages = folio_nr_pages(folio);
198 ret = -nr_pages;
199
200 /*
201 * When this function is called from scan_swap_map_slots() and it's
202 * called by vmscan.c at reclaiming folios. So we hold a folio lock
203 * here. We have to use trylock for avoiding deadlock. This is a special
204 * case and you should use folio_free_swap() with explicit folio_lock()
205 * in usual operations.
206 */
207 if (!folio_trylock(folio))
208 goto out;
209
210 /* offset could point to the middle of a large folio */
211 entry = folio->swap;
212 offset = swp_offset(entry);
213
214 need_reclaim = ((flags & TTRS_ANYWAY) ||
215 ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
216 ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)));
217 if (!need_reclaim || !folio_swapcache_freeable(folio))
218 goto out_unlock;
219
220 /*
221 * It's safe to delete the folio from swap cache only if the folio's
222 * swap_map is HAS_CACHE only, which means the slots have no page table
223 * reference or pending writeback, and can't be allocated to others.
224 */
225 ci = lock_cluster_or_swap_info(si, offset);
226 need_reclaim = swap_is_has_cache(si, offset, nr_pages);
227 unlock_cluster_or_swap_info(si, ci);
228 if (!need_reclaim)
229 goto out_unlock;
230
231 if (!(flags & TTRS_DIRECT)) {
232 /* Free through slot cache */
233 delete_from_swap_cache(folio);
234 folio_set_dirty(folio);
235 ret = nr_pages;
236 goto out_unlock;
237 }
238
239 xa_lock_irq(&address_space->i_pages);
240 __delete_from_swap_cache(folio, entry, NULL);
241 xa_unlock_irq(&address_space->i_pages);
242 folio_ref_sub(folio, nr_pages);
243 folio_set_dirty(folio);
244
245 spin_lock(&si->lock);
246 /* Only sinple page folio can be backed by zswap */
247 if (nr_pages == 1)
248 zswap_invalidate(entry);
249 swap_entry_range_free(si, entry, nr_pages);
250 spin_unlock(&si->lock);
251 ret = nr_pages;
252out_unlock:
253 folio_unlock(folio);
254out:
255 folio_put(folio);
256 return ret;
257}
258
259static inline struct swap_extent *first_se(struct swap_info_struct *sis)
260{
261 struct rb_node *rb = rb_first(&sis->swap_extent_root);
262 return rb_entry(rb, struct swap_extent, rb_node);
263}
264
265static inline struct swap_extent *next_se(struct swap_extent *se)
266{
267 struct rb_node *rb = rb_next(&se->rb_node);
268 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
269}
270
271/*
272 * swapon tell device that all the old swap contents can be discarded,
273 * to allow the swap device to optimize its wear-levelling.
274 */
275static int discard_swap(struct swap_info_struct *si)
276{
277 struct swap_extent *se;
278 sector_t start_block;
279 sector_t nr_blocks;
280 int err = 0;
281
282 /* Do not discard the swap header page! */
283 se = first_se(si);
284 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
285 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
286 if (nr_blocks) {
287 err = blkdev_issue_discard(si->bdev, start_block,
288 nr_blocks, GFP_KERNEL);
289 if (err)
290 return err;
291 cond_resched();
292 }
293
294 for (se = next_se(se); se; se = next_se(se)) {
295 start_block = se->start_block << (PAGE_SHIFT - 9);
296 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
297
298 err = blkdev_issue_discard(si->bdev, start_block,
299 nr_blocks, GFP_KERNEL);
300 if (err)
301 break;
302
303 cond_resched();
304 }
305 return err; /* That will often be -EOPNOTSUPP */
306}
307
308static struct swap_extent *
309offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
310{
311 struct swap_extent *se;
312 struct rb_node *rb;
313
314 rb = sis->swap_extent_root.rb_node;
315 while (rb) {
316 se = rb_entry(rb, struct swap_extent, rb_node);
317 if (offset < se->start_page)
318 rb = rb->rb_left;
319 else if (offset >= se->start_page + se->nr_pages)
320 rb = rb->rb_right;
321 else
322 return se;
323 }
324 /* It *must* be present */
325 BUG();
326}
327
328sector_t swap_folio_sector(struct folio *folio)
329{
330 struct swap_info_struct *sis = swp_swap_info(folio->swap);
331 struct swap_extent *se;
332 sector_t sector;
333 pgoff_t offset;
334
335 offset = swp_offset(folio->swap);
336 se = offset_to_swap_extent(sis, offset);
337 sector = se->start_block + (offset - se->start_page);
338 return sector << (PAGE_SHIFT - 9);
339}
340
341/*
342 * swap allocation tell device that a cluster of swap can now be discarded,
343 * to allow the swap device to optimize its wear-levelling.
344 */
345static void discard_swap_cluster(struct swap_info_struct *si,
346 pgoff_t start_page, pgoff_t nr_pages)
347{
348 struct swap_extent *se = offset_to_swap_extent(si, start_page);
349
350 while (nr_pages) {
351 pgoff_t offset = start_page - se->start_page;
352 sector_t start_block = se->start_block + offset;
353 sector_t nr_blocks = se->nr_pages - offset;
354
355 if (nr_blocks > nr_pages)
356 nr_blocks = nr_pages;
357 start_page += nr_blocks;
358 nr_pages -= nr_blocks;
359
360 start_block <<= PAGE_SHIFT - 9;
361 nr_blocks <<= PAGE_SHIFT - 9;
362 if (blkdev_issue_discard(si->bdev, start_block,
363 nr_blocks, GFP_NOIO))
364 break;
365
366 se = next_se(se);
367 }
368}
369
370#ifdef CONFIG_THP_SWAP
371#define SWAPFILE_CLUSTER HPAGE_PMD_NR
372
373#define swap_entry_order(order) (order)
374#else
375#define SWAPFILE_CLUSTER 256
376
377/*
378 * Define swap_entry_order() as constant to let compiler to optimize
379 * out some code if !CONFIG_THP_SWAP
380 */
381#define swap_entry_order(order) 0
382#endif
383#define LATENCY_LIMIT 256
384
385static inline bool cluster_is_free(struct swap_cluster_info *info)
386{
387 return info->flags & CLUSTER_FLAG_FREE;
388}
389
390static inline unsigned int cluster_index(struct swap_info_struct *si,
391 struct swap_cluster_info *ci)
392{
393 return ci - si->cluster_info;
394}
395
396static inline unsigned int cluster_offset(struct swap_info_struct *si,
397 struct swap_cluster_info *ci)
398{
399 return cluster_index(si, ci) * SWAPFILE_CLUSTER;
400}
401
402static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
403 unsigned long offset)
404{
405 struct swap_cluster_info *ci;
406
407 ci = si->cluster_info;
408 if (ci) {
409 ci += offset / SWAPFILE_CLUSTER;
410 spin_lock(&ci->lock);
411 }
412 return ci;
413}
414
415static inline void unlock_cluster(struct swap_cluster_info *ci)
416{
417 if (ci)
418 spin_unlock(&ci->lock);
419}
420
421/*
422 * Determine the locking method in use for this device. Return
423 * swap_cluster_info if SSD-style cluster-based locking is in place.
424 */
425static inline struct swap_cluster_info *lock_cluster_or_swap_info(
426 struct swap_info_struct *si, unsigned long offset)
427{
428 struct swap_cluster_info *ci;
429
430 /* Try to use fine-grained SSD-style locking if available: */
431 ci = lock_cluster(si, offset);
432 /* Otherwise, fall back to traditional, coarse locking: */
433 if (!ci)
434 spin_lock(&si->lock);
435
436 return ci;
437}
438
439static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
440 struct swap_cluster_info *ci)
441{
442 if (ci)
443 unlock_cluster(ci);
444 else
445 spin_unlock(&si->lock);
446}
447
448/* Add a cluster to discard list and schedule it to do discard */
449static void swap_cluster_schedule_discard(struct swap_info_struct *si,
450 struct swap_cluster_info *ci)
451{
452 unsigned int idx = cluster_index(si, ci);
453 /*
454 * If scan_swap_map_slots() can't find a free cluster, it will check
455 * si->swap_map directly. To make sure the discarding cluster isn't
456 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
457 * It will be cleared after discard
458 */
459 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
460 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
461
462 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE);
463 list_move_tail(&ci->list, &si->discard_clusters);
464 ci->flags = 0;
465 schedule_work(&si->discard_work);
466}
467
468static void __free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci)
469{
470 lockdep_assert_held(&si->lock);
471 lockdep_assert_held(&ci->lock);
472
473 if (ci->flags)
474 list_move_tail(&ci->list, &si->free_clusters);
475 else
476 list_add_tail(&ci->list, &si->free_clusters);
477 ci->flags = CLUSTER_FLAG_FREE;
478 ci->order = 0;
479}
480
481/*
482 * Doing discard actually. After a cluster discard is finished, the cluster
483 * will be added to free cluster list. caller should hold si->lock.
484*/
485static void swap_do_scheduled_discard(struct swap_info_struct *si)
486{
487 struct swap_cluster_info *ci;
488 unsigned int idx;
489
490 while (!list_empty(&si->discard_clusters)) {
491 ci = list_first_entry(&si->discard_clusters, struct swap_cluster_info, list);
492 list_del(&ci->list);
493 idx = cluster_index(si, ci);
494 spin_unlock(&si->lock);
495
496 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
497 SWAPFILE_CLUSTER);
498
499 spin_lock(&si->lock);
500 spin_lock(&ci->lock);
501 __free_cluster(si, ci);
502 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
503 0, SWAPFILE_CLUSTER);
504 spin_unlock(&ci->lock);
505 }
506}
507
508static void swap_discard_work(struct work_struct *work)
509{
510 struct swap_info_struct *si;
511
512 si = container_of(work, struct swap_info_struct, discard_work);
513
514 spin_lock(&si->lock);
515 swap_do_scheduled_discard(si);
516 spin_unlock(&si->lock);
517}
518
519static void swap_users_ref_free(struct percpu_ref *ref)
520{
521 struct swap_info_struct *si;
522
523 si = container_of(ref, struct swap_info_struct, users);
524 complete(&si->comp);
525}
526
527static void free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci)
528{
529 VM_BUG_ON(ci->count != 0);
530 lockdep_assert_held(&si->lock);
531 lockdep_assert_held(&ci->lock);
532
533 if (ci->flags & CLUSTER_FLAG_FRAG)
534 si->frag_cluster_nr[ci->order]--;
535
536 /*
537 * If the swap is discardable, prepare discard the cluster
538 * instead of free it immediately. The cluster will be freed
539 * after discard.
540 */
541 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
542 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
543 swap_cluster_schedule_discard(si, ci);
544 return;
545 }
546
547 __free_cluster(si, ci);
548}
549
550/*
551 * The cluster corresponding to page_nr will be used. The cluster will not be
552 * added to free cluster list and its usage counter will be increased by 1.
553 * Only used for initialization.
554 */
555static void inc_cluster_info_page(struct swap_info_struct *si,
556 struct swap_cluster_info *cluster_info, unsigned long page_nr)
557{
558 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
559 struct swap_cluster_info *ci;
560
561 if (!cluster_info)
562 return;
563
564 ci = cluster_info + idx;
565 ci->count++;
566
567 VM_BUG_ON(ci->count > SWAPFILE_CLUSTER);
568 VM_BUG_ON(ci->flags);
569}
570
571/*
572 * The cluster ci decreases @nr_pages usage. If the usage counter becomes 0,
573 * which means no page in the cluster is in use, we can optionally discard
574 * the cluster and add it to free cluster list.
575 */
576static void dec_cluster_info_page(struct swap_info_struct *si,
577 struct swap_cluster_info *ci, int nr_pages)
578{
579 if (!si->cluster_info)
580 return;
581
582 VM_BUG_ON(ci->count < nr_pages);
583 VM_BUG_ON(cluster_is_free(ci));
584 lockdep_assert_held(&si->lock);
585 lockdep_assert_held(&ci->lock);
586 ci->count -= nr_pages;
587
588 if (!ci->count) {
589 free_cluster(si, ci);
590 return;
591 }
592
593 if (!(ci->flags & CLUSTER_FLAG_NONFULL)) {
594 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE);
595 if (ci->flags & CLUSTER_FLAG_FRAG)
596 si->frag_cluster_nr[ci->order]--;
597 list_move_tail(&ci->list, &si->nonfull_clusters[ci->order]);
598 ci->flags = CLUSTER_FLAG_NONFULL;
599 }
600}
601
602static bool cluster_reclaim_range(struct swap_info_struct *si,
603 struct swap_cluster_info *ci,
604 unsigned long start, unsigned long end)
605{
606 unsigned char *map = si->swap_map;
607 unsigned long offset;
608
609 spin_unlock(&ci->lock);
610 spin_unlock(&si->lock);
611
612 for (offset = start; offset < end; offset++) {
613 switch (READ_ONCE(map[offset])) {
614 case 0:
615 continue;
616 case SWAP_HAS_CACHE:
617 if (__try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT) > 0)
618 continue;
619 goto out;
620 default:
621 goto out;
622 }
623 }
624out:
625 spin_lock(&si->lock);
626 spin_lock(&ci->lock);
627
628 /*
629 * Recheck the range no matter reclaim succeeded or not, the slot
630 * could have been be freed while we are not holding the lock.
631 */
632 for (offset = start; offset < end; offset++)
633 if (READ_ONCE(map[offset]))
634 return false;
635
636 return true;
637}
638
639static bool cluster_scan_range(struct swap_info_struct *si,
640 struct swap_cluster_info *ci,
641 unsigned long start, unsigned int nr_pages)
642{
643 unsigned long offset, end = start + nr_pages;
644 unsigned char *map = si->swap_map;
645 bool need_reclaim = false;
646
647 for (offset = start; offset < end; offset++) {
648 switch (READ_ONCE(map[offset])) {
649 case 0:
650 continue;
651 case SWAP_HAS_CACHE:
652 if (!vm_swap_full())
653 return false;
654 need_reclaim = true;
655 continue;
656 default:
657 return false;
658 }
659 }
660
661 if (need_reclaim)
662 return cluster_reclaim_range(si, ci, start, end);
663
664 return true;
665}
666
667static bool cluster_alloc_range(struct swap_info_struct *si, struct swap_cluster_info *ci,
668 unsigned int start, unsigned char usage,
669 unsigned int order)
670{
671 unsigned int nr_pages = 1 << order;
672
673 if (!(si->flags & SWP_WRITEOK))
674 return false;
675
676 if (cluster_is_free(ci)) {
677 if (nr_pages < SWAPFILE_CLUSTER) {
678 list_move_tail(&ci->list, &si->nonfull_clusters[order]);
679 ci->flags = CLUSTER_FLAG_NONFULL;
680 }
681 ci->order = order;
682 }
683
684 memset(si->swap_map + start, usage, nr_pages);
685 swap_range_alloc(si, start, nr_pages);
686 ci->count += nr_pages;
687
688 if (ci->count == SWAPFILE_CLUSTER) {
689 VM_BUG_ON(!(ci->flags &
690 (CLUSTER_FLAG_FREE | CLUSTER_FLAG_NONFULL | CLUSTER_FLAG_FRAG)));
691 if (ci->flags & CLUSTER_FLAG_FRAG)
692 si->frag_cluster_nr[ci->order]--;
693 list_move_tail(&ci->list, &si->full_clusters);
694 ci->flags = CLUSTER_FLAG_FULL;
695 }
696
697 return true;
698}
699
700static unsigned int alloc_swap_scan_cluster(struct swap_info_struct *si, unsigned long offset,
701 unsigned int *foundp, unsigned int order,
702 unsigned char usage)
703{
704 unsigned long start = offset & ~(SWAPFILE_CLUSTER - 1);
705 unsigned long end = min(start + SWAPFILE_CLUSTER, si->max);
706 unsigned int nr_pages = 1 << order;
707 struct swap_cluster_info *ci;
708
709 if (end < nr_pages)
710 return SWAP_NEXT_INVALID;
711 end -= nr_pages;
712
713 ci = lock_cluster(si, offset);
714 if (ci->count + nr_pages > SWAPFILE_CLUSTER) {
715 offset = SWAP_NEXT_INVALID;
716 goto done;
717 }
718
719 while (offset <= end) {
720 if (cluster_scan_range(si, ci, offset, nr_pages)) {
721 if (!cluster_alloc_range(si, ci, offset, usage, order)) {
722 offset = SWAP_NEXT_INVALID;
723 goto done;
724 }
725 *foundp = offset;
726 if (ci->count == SWAPFILE_CLUSTER) {
727 offset = SWAP_NEXT_INVALID;
728 goto done;
729 }
730 offset += nr_pages;
731 break;
732 }
733 offset += nr_pages;
734 }
735 if (offset > end)
736 offset = SWAP_NEXT_INVALID;
737done:
738 unlock_cluster(ci);
739 return offset;
740}
741
742/* Return true if reclaimed a whole cluster */
743static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)
744{
745 long to_scan = 1;
746 unsigned long offset, end;
747 struct swap_cluster_info *ci;
748 unsigned char *map = si->swap_map;
749 int nr_reclaim;
750
751 if (force)
752 to_scan = si->inuse_pages / SWAPFILE_CLUSTER;
753
754 while (!list_empty(&si->full_clusters)) {
755 ci = list_first_entry(&si->full_clusters, struct swap_cluster_info, list);
756 list_move_tail(&ci->list, &si->full_clusters);
757 offset = cluster_offset(si, ci);
758 end = min(si->max, offset + SWAPFILE_CLUSTER);
759 to_scan--;
760
761 spin_unlock(&si->lock);
762 while (offset < end) {
763 if (READ_ONCE(map[offset]) == SWAP_HAS_CACHE) {
764 nr_reclaim = __try_to_reclaim_swap(si, offset,
765 TTRS_ANYWAY | TTRS_DIRECT);
766 if (nr_reclaim) {
767 offset += abs(nr_reclaim);
768 continue;
769 }
770 }
771 offset++;
772 }
773 spin_lock(&si->lock);
774
775 if (to_scan <= 0)
776 break;
777 }
778}
779
780static void swap_reclaim_work(struct work_struct *work)
781{
782 struct swap_info_struct *si;
783
784 si = container_of(work, struct swap_info_struct, reclaim_work);
785
786 spin_lock(&si->lock);
787 swap_reclaim_full_clusters(si, true);
788 spin_unlock(&si->lock);
789}
790
791/*
792 * Try to get swap entries with specified order from current cpu's swap entry
793 * pool (a cluster). This might involve allocating a new cluster for current CPU
794 * too.
795 */
796static unsigned long cluster_alloc_swap_entry(struct swap_info_struct *si, int order,
797 unsigned char usage)
798{
799 struct percpu_cluster *cluster;
800 struct swap_cluster_info *ci;
801 unsigned int offset, found = 0;
802
803new_cluster:
804 lockdep_assert_held(&si->lock);
805 cluster = this_cpu_ptr(si->percpu_cluster);
806 offset = cluster->next[order];
807 if (offset) {
808 offset = alloc_swap_scan_cluster(si, offset, &found, order, usage);
809 if (found)
810 goto done;
811 }
812
813 if (!list_empty(&si->free_clusters)) {
814 ci = list_first_entry(&si->free_clusters, struct swap_cluster_info, list);
815 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), &found, order, usage);
816 /*
817 * Either we didn't touch the cluster due to swapoff,
818 * or the allocation must success.
819 */
820 VM_BUG_ON((si->flags & SWP_WRITEOK) && !found);
821 goto done;
822 }
823
824 /* Try reclaim from full clusters if free clusters list is drained */
825 if (vm_swap_full())
826 swap_reclaim_full_clusters(si, false);
827
828 if (order < PMD_ORDER) {
829 unsigned int frags = 0;
830
831 while (!list_empty(&si->nonfull_clusters[order])) {
832 ci = list_first_entry(&si->nonfull_clusters[order],
833 struct swap_cluster_info, list);
834 list_move_tail(&ci->list, &si->frag_clusters[order]);
835 ci->flags = CLUSTER_FLAG_FRAG;
836 si->frag_cluster_nr[order]++;
837 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
838 &found, order, usage);
839 frags++;
840 if (found)
841 break;
842 }
843
844 if (!found) {
845 /*
846 * Nonfull clusters are moved to frag tail if we reached
847 * here, count them too, don't over scan the frag list.
848 */
849 while (frags < si->frag_cluster_nr[order]) {
850 ci = list_first_entry(&si->frag_clusters[order],
851 struct swap_cluster_info, list);
852 /*
853 * Rotate the frag list to iterate, they were all failing
854 * high order allocation or moved here due to per-CPU usage,
855 * this help keeping usable cluster ahead.
856 */
857 list_move_tail(&ci->list, &si->frag_clusters[order]);
858 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
859 &found, order, usage);
860 frags++;
861 if (found)
862 break;
863 }
864 }
865 }
866
867 if (found)
868 goto done;
869
870 if (!list_empty(&si->discard_clusters)) {
871 /*
872 * we don't have free cluster but have some clusters in
873 * discarding, do discard now and reclaim them, then
874 * reread cluster_next_cpu since we dropped si->lock
875 */
876 swap_do_scheduled_discard(si);
877 goto new_cluster;
878 }
879
880 if (order)
881 goto done;
882
883 /* Order 0 stealing from higher order */
884 for (int o = 1; o < SWAP_NR_ORDERS; o++) {
885 /*
886 * Clusters here have at least one usable slots and can't fail order 0
887 * allocation, but reclaim may drop si->lock and race with another user.
888 */
889 while (!list_empty(&si->frag_clusters[o])) {
890 ci = list_first_entry(&si->frag_clusters[o],
891 struct swap_cluster_info, list);
892 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
893 &found, 0, usage);
894 if (found)
895 goto done;
896 }
897
898 while (!list_empty(&si->nonfull_clusters[o])) {
899 ci = list_first_entry(&si->nonfull_clusters[o],
900 struct swap_cluster_info, list);
901 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
902 &found, 0, usage);
903 if (found)
904 goto done;
905 }
906 }
907
908done:
909 cluster->next[order] = offset;
910 return found;
911}
912
913static void __del_from_avail_list(struct swap_info_struct *si)
914{
915 int nid;
916
917 assert_spin_locked(&si->lock);
918 for_each_node(nid)
919 plist_del(&si->avail_lists[nid], &swap_avail_heads[nid]);
920}
921
922static void del_from_avail_list(struct swap_info_struct *si)
923{
924 spin_lock(&swap_avail_lock);
925 __del_from_avail_list(si);
926 spin_unlock(&swap_avail_lock);
927}
928
929static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
930 unsigned int nr_entries)
931{
932 unsigned int end = offset + nr_entries - 1;
933
934 if (offset == si->lowest_bit)
935 si->lowest_bit += nr_entries;
936 if (end == si->highest_bit)
937 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
938 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
939 if (si->inuse_pages == si->pages) {
940 si->lowest_bit = si->max;
941 si->highest_bit = 0;
942 del_from_avail_list(si);
943
944 if (si->cluster_info && vm_swap_full())
945 schedule_work(&si->reclaim_work);
946 }
947}
948
949static void add_to_avail_list(struct swap_info_struct *si)
950{
951 int nid;
952
953 spin_lock(&swap_avail_lock);
954 for_each_node(nid)
955 plist_add(&si->avail_lists[nid], &swap_avail_heads[nid]);
956 spin_unlock(&swap_avail_lock);
957}
958
959static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
960 unsigned int nr_entries)
961{
962 unsigned long begin = offset;
963 unsigned long end = offset + nr_entries - 1;
964 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
965 unsigned int i;
966
967 /*
968 * Use atomic clear_bit operations only on zeromap instead of non-atomic
969 * bitmap_clear to prevent adjacent bits corruption due to simultaneous writes.
970 */
971 for (i = 0; i < nr_entries; i++)
972 clear_bit(offset + i, si->zeromap);
973
974 if (offset < si->lowest_bit)
975 si->lowest_bit = offset;
976 if (end > si->highest_bit) {
977 bool was_full = !si->highest_bit;
978
979 WRITE_ONCE(si->highest_bit, end);
980 if (was_full && (si->flags & SWP_WRITEOK))
981 add_to_avail_list(si);
982 }
983 if (si->flags & SWP_BLKDEV)
984 swap_slot_free_notify =
985 si->bdev->bd_disk->fops->swap_slot_free_notify;
986 else
987 swap_slot_free_notify = NULL;
988 while (offset <= end) {
989 arch_swap_invalidate_page(si->type, offset);
990 if (swap_slot_free_notify)
991 swap_slot_free_notify(si->bdev, offset);
992 offset++;
993 }
994 clear_shadow_from_swap_cache(si->type, begin, end);
995
996 /*
997 * Make sure that try_to_unuse() observes si->inuse_pages reaching 0
998 * only after the above cleanups are done.
999 */
1000 smp_wmb();
1001 atomic_long_add(nr_entries, &nr_swap_pages);
1002 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
1003}
1004
1005static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
1006{
1007 unsigned long prev;
1008
1009 if (!(si->flags & SWP_SOLIDSTATE)) {
1010 si->cluster_next = next;
1011 return;
1012 }
1013
1014 prev = this_cpu_read(*si->cluster_next_cpu);
1015 /*
1016 * Cross the swap address space size aligned trunk, choose
1017 * another trunk randomly to avoid lock contention on swap
1018 * address space if possible.
1019 */
1020 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
1021 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
1022 /* No free swap slots available */
1023 if (si->highest_bit <= si->lowest_bit)
1024 return;
1025 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
1026 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
1027 next = max_t(unsigned int, next, si->lowest_bit);
1028 }
1029 this_cpu_write(*si->cluster_next_cpu, next);
1030}
1031
1032static bool swap_offset_available_and_locked(struct swap_info_struct *si,
1033 unsigned long offset)
1034{
1035 if (data_race(!si->swap_map[offset])) {
1036 spin_lock(&si->lock);
1037 return true;
1038 }
1039
1040 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
1041 spin_lock(&si->lock);
1042 return true;
1043 }
1044
1045 return false;
1046}
1047
1048static int cluster_alloc_swap(struct swap_info_struct *si,
1049 unsigned char usage, int nr,
1050 swp_entry_t slots[], int order)
1051{
1052 int n_ret = 0;
1053
1054 VM_BUG_ON(!si->cluster_info);
1055
1056 si->flags += SWP_SCANNING;
1057
1058 while (n_ret < nr) {
1059 unsigned long offset = cluster_alloc_swap_entry(si, order, usage);
1060
1061 if (!offset)
1062 break;
1063 slots[n_ret++] = swp_entry(si->type, offset);
1064 }
1065
1066 si->flags -= SWP_SCANNING;
1067
1068 return n_ret;
1069}
1070
1071static int scan_swap_map_slots(struct swap_info_struct *si,
1072 unsigned char usage, int nr,
1073 swp_entry_t slots[], int order)
1074{
1075 unsigned long offset;
1076 unsigned long scan_base;
1077 unsigned long last_in_cluster = 0;
1078 int latency_ration = LATENCY_LIMIT;
1079 unsigned int nr_pages = 1 << order;
1080 int n_ret = 0;
1081 bool scanned_many = false;
1082
1083 /*
1084 * We try to cluster swap pages by allocating them sequentially
1085 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
1086 * way, however, we resort to first-free allocation, starting
1087 * a new cluster. This prevents us from scattering swap pages
1088 * all over the entire swap partition, so that we reduce
1089 * overall disk seek times between swap pages. -- sct
1090 * But we do now try to find an empty cluster. -Andrea
1091 * And we let swap pages go all over an SSD partition. Hugh
1092 */
1093
1094 if (order > 0) {
1095 /*
1096 * Should not even be attempting large allocations when huge
1097 * page swap is disabled. Warn and fail the allocation.
1098 */
1099 if (!IS_ENABLED(CONFIG_THP_SWAP) ||
1100 nr_pages > SWAPFILE_CLUSTER) {
1101 VM_WARN_ON_ONCE(1);
1102 return 0;
1103 }
1104
1105 /*
1106 * Swapfile is not block device or not using clusters so unable
1107 * to allocate large entries.
1108 */
1109 if (!(si->flags & SWP_BLKDEV) || !si->cluster_info)
1110 return 0;
1111 }
1112
1113 if (si->cluster_info)
1114 return cluster_alloc_swap(si, usage, nr, slots, order);
1115
1116 si->flags += SWP_SCANNING;
1117
1118 /* For HDD, sequential access is more important. */
1119 scan_base = si->cluster_next;
1120 offset = scan_base;
1121
1122 if (unlikely(!si->cluster_nr--)) {
1123 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
1124 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1125 goto checks;
1126 }
1127
1128 spin_unlock(&si->lock);
1129
1130 /*
1131 * If seek is expensive, start searching for new cluster from
1132 * start of partition, to minimize the span of allocated swap.
1133 */
1134 scan_base = offset = si->lowest_bit;
1135 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
1136
1137 /* Locate the first empty (unaligned) cluster */
1138 for (; last_in_cluster <= READ_ONCE(si->highest_bit); offset++) {
1139 if (si->swap_map[offset])
1140 last_in_cluster = offset + SWAPFILE_CLUSTER;
1141 else if (offset == last_in_cluster) {
1142 spin_lock(&si->lock);
1143 offset -= SWAPFILE_CLUSTER - 1;
1144 si->cluster_next = offset;
1145 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1146 goto checks;
1147 }
1148 if (unlikely(--latency_ration < 0)) {
1149 cond_resched();
1150 latency_ration = LATENCY_LIMIT;
1151 }
1152 }
1153
1154 offset = scan_base;
1155 spin_lock(&si->lock);
1156 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1157 }
1158
1159checks:
1160 if (!(si->flags & SWP_WRITEOK))
1161 goto no_page;
1162 if (!si->highest_bit)
1163 goto no_page;
1164 if (offset > si->highest_bit)
1165 scan_base = offset = si->lowest_bit;
1166
1167 /* reuse swap entry of cache-only swap if not busy. */
1168 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
1169 int swap_was_freed;
1170 spin_unlock(&si->lock);
1171 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT);
1172 spin_lock(&si->lock);
1173 /* entry was freed successfully, try to use this again */
1174 if (swap_was_freed > 0)
1175 goto checks;
1176 goto scan; /* check next one */
1177 }
1178
1179 if (si->swap_map[offset]) {
1180 if (!n_ret)
1181 goto scan;
1182 else
1183 goto done;
1184 }
1185 memset(si->swap_map + offset, usage, nr_pages);
1186
1187 swap_range_alloc(si, offset, nr_pages);
1188 slots[n_ret++] = swp_entry(si->type, offset);
1189
1190 /* got enough slots or reach max slots? */
1191 if ((n_ret == nr) || (offset >= si->highest_bit))
1192 goto done;
1193
1194 /* search for next available slot */
1195
1196 /* time to take a break? */
1197 if (unlikely(--latency_ration < 0)) {
1198 if (n_ret)
1199 goto done;
1200 spin_unlock(&si->lock);
1201 cond_resched();
1202 spin_lock(&si->lock);
1203 latency_ration = LATENCY_LIMIT;
1204 }
1205
1206 if (si->cluster_nr && !si->swap_map[++offset]) {
1207 /* non-ssd case, still more slots in cluster? */
1208 --si->cluster_nr;
1209 goto checks;
1210 }
1211
1212 /*
1213 * Even if there's no free clusters available (fragmented),
1214 * try to scan a little more quickly with lock held unless we
1215 * have scanned too many slots already.
1216 */
1217 if (!scanned_many) {
1218 unsigned long scan_limit;
1219
1220 if (offset < scan_base)
1221 scan_limit = scan_base;
1222 else
1223 scan_limit = si->highest_bit;
1224 for (; offset <= scan_limit && --latency_ration > 0;
1225 offset++) {
1226 if (!si->swap_map[offset])
1227 goto checks;
1228 }
1229 }
1230
1231done:
1232 if (order == 0)
1233 set_cluster_next(si, offset + 1);
1234 si->flags -= SWP_SCANNING;
1235 return n_ret;
1236
1237scan:
1238 VM_WARN_ON(order > 0);
1239 spin_unlock(&si->lock);
1240 while (++offset <= READ_ONCE(si->highest_bit)) {
1241 if (unlikely(--latency_ration < 0)) {
1242 cond_resched();
1243 latency_ration = LATENCY_LIMIT;
1244 scanned_many = true;
1245 }
1246 if (swap_offset_available_and_locked(si, offset))
1247 goto checks;
1248 }
1249 offset = si->lowest_bit;
1250 while (offset < scan_base) {
1251 if (unlikely(--latency_ration < 0)) {
1252 cond_resched();
1253 latency_ration = LATENCY_LIMIT;
1254 scanned_many = true;
1255 }
1256 if (swap_offset_available_and_locked(si, offset))
1257 goto checks;
1258 offset++;
1259 }
1260 spin_lock(&si->lock);
1261
1262no_page:
1263 si->flags -= SWP_SCANNING;
1264 return n_ret;
1265}
1266
1267int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_order)
1268{
1269 int order = swap_entry_order(entry_order);
1270 unsigned long size = 1 << order;
1271 struct swap_info_struct *si, *next;
1272 long avail_pgs;
1273 int n_ret = 0;
1274 int node;
1275
1276 spin_lock(&swap_avail_lock);
1277
1278 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1279 if (avail_pgs <= 0) {
1280 spin_unlock(&swap_avail_lock);
1281 goto noswap;
1282 }
1283
1284 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1285
1286 atomic_long_sub(n_goal * size, &nr_swap_pages);
1287
1288start_over:
1289 node = numa_node_id();
1290 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1291 /* requeue si to after same-priority siblings */
1292 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1293 spin_unlock(&swap_avail_lock);
1294 spin_lock(&si->lock);
1295 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1296 spin_lock(&swap_avail_lock);
1297 if (plist_node_empty(&si->avail_lists[node])) {
1298 spin_unlock(&si->lock);
1299 goto nextsi;
1300 }
1301 WARN(!si->highest_bit,
1302 "swap_info %d in list but !highest_bit\n",
1303 si->type);
1304 WARN(!(si->flags & SWP_WRITEOK),
1305 "swap_info %d in list but !SWP_WRITEOK\n",
1306 si->type);
1307 __del_from_avail_list(si);
1308 spin_unlock(&si->lock);
1309 goto nextsi;
1310 }
1311 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1312 n_goal, swp_entries, order);
1313 spin_unlock(&si->lock);
1314 if (n_ret || size > 1)
1315 goto check_out;
1316 cond_resched();
1317
1318 spin_lock(&swap_avail_lock);
1319nextsi:
1320 /*
1321 * if we got here, it's likely that si was almost full before,
1322 * and since scan_swap_map_slots() can drop the si->lock,
1323 * multiple callers probably all tried to get a page from the
1324 * same si and it filled up before we could get one; or, the si
1325 * filled up between us dropping swap_avail_lock and taking
1326 * si->lock. Since we dropped the swap_avail_lock, the
1327 * swap_avail_head list may have been modified; so if next is
1328 * still in the swap_avail_head list then try it, otherwise
1329 * start over if we have not gotten any slots.
1330 */
1331 if (plist_node_empty(&next->avail_lists[node]))
1332 goto start_over;
1333 }
1334
1335 spin_unlock(&swap_avail_lock);
1336
1337check_out:
1338 if (n_ret < n_goal)
1339 atomic_long_add((long)(n_goal - n_ret) * size,
1340 &nr_swap_pages);
1341noswap:
1342 return n_ret;
1343}
1344
1345static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1346{
1347 struct swap_info_struct *si;
1348 unsigned long offset;
1349
1350 if (!entry.val)
1351 goto out;
1352 si = swp_swap_info(entry);
1353 if (!si)
1354 goto bad_nofile;
1355 if (data_race(!(si->flags & SWP_USED)))
1356 goto bad_device;
1357 offset = swp_offset(entry);
1358 if (offset >= si->max)
1359 goto bad_offset;
1360 if (data_race(!si->swap_map[swp_offset(entry)]))
1361 goto bad_free;
1362 return si;
1363
1364bad_free:
1365 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1366 goto out;
1367bad_offset:
1368 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1369 goto out;
1370bad_device:
1371 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1372 goto out;
1373bad_nofile:
1374 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1375out:
1376 return NULL;
1377}
1378
1379static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1380 struct swap_info_struct *q)
1381{
1382 struct swap_info_struct *p;
1383
1384 p = _swap_info_get(entry);
1385
1386 if (p != q) {
1387 if (q != NULL)
1388 spin_unlock(&q->lock);
1389 if (p != NULL)
1390 spin_lock(&p->lock);
1391 }
1392 return p;
1393}
1394
1395static unsigned char __swap_entry_free_locked(struct swap_info_struct *si,
1396 unsigned long offset,
1397 unsigned char usage)
1398{
1399 unsigned char count;
1400 unsigned char has_cache;
1401
1402 count = si->swap_map[offset];
1403
1404 has_cache = count & SWAP_HAS_CACHE;
1405 count &= ~SWAP_HAS_CACHE;
1406
1407 if (usage == SWAP_HAS_CACHE) {
1408 VM_BUG_ON(!has_cache);
1409 has_cache = 0;
1410 } else if (count == SWAP_MAP_SHMEM) {
1411 /*
1412 * Or we could insist on shmem.c using a special
1413 * swap_shmem_free() and free_shmem_swap_and_cache()...
1414 */
1415 count = 0;
1416 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1417 if (count == COUNT_CONTINUED) {
1418 if (swap_count_continued(si, offset, count))
1419 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1420 else
1421 count = SWAP_MAP_MAX;
1422 } else
1423 count--;
1424 }
1425
1426 usage = count | has_cache;
1427 if (usage)
1428 WRITE_ONCE(si->swap_map[offset], usage);
1429 else
1430 WRITE_ONCE(si->swap_map[offset], SWAP_HAS_CACHE);
1431
1432 return usage;
1433}
1434
1435/*
1436 * When we get a swap entry, if there aren't some other ways to
1437 * prevent swapoff, such as the folio in swap cache is locked, RCU
1438 * reader side is locked, etc., the swap entry may become invalid
1439 * because of swapoff. Then, we need to enclose all swap related
1440 * functions with get_swap_device() and put_swap_device(), unless the
1441 * swap functions call get/put_swap_device() by themselves.
1442 *
1443 * RCU reader side lock (including any spinlock) is sufficient to
1444 * prevent swapoff, because synchronize_rcu() is called in swapoff()
1445 * before freeing data structures.
1446 *
1447 * Check whether swap entry is valid in the swap device. If so,
1448 * return pointer to swap_info_struct, and keep the swap entry valid
1449 * via preventing the swap device from being swapoff, until
1450 * put_swap_device() is called. Otherwise return NULL.
1451 *
1452 * Notice that swapoff or swapoff+swapon can still happen before the
1453 * percpu_ref_tryget_live() in get_swap_device() or after the
1454 * percpu_ref_put() in put_swap_device() if there isn't any other way
1455 * to prevent swapoff. The caller must be prepared for that. For
1456 * example, the following situation is possible.
1457 *
1458 * CPU1 CPU2
1459 * do_swap_page()
1460 * ... swapoff+swapon
1461 * __read_swap_cache_async()
1462 * swapcache_prepare()
1463 * __swap_duplicate()
1464 * // check swap_map
1465 * // verify PTE not changed
1466 *
1467 * In __swap_duplicate(), the swap_map need to be checked before
1468 * changing partly because the specified swap entry may be for another
1469 * swap device which has been swapoff. And in do_swap_page(), after
1470 * the page is read from the swap device, the PTE is verified not
1471 * changed with the page table locked to check whether the swap device
1472 * has been swapoff or swapoff+swapon.
1473 */
1474struct swap_info_struct *get_swap_device(swp_entry_t entry)
1475{
1476 struct swap_info_struct *si;
1477 unsigned long offset;
1478
1479 if (!entry.val)
1480 goto out;
1481 si = swp_swap_info(entry);
1482 if (!si)
1483 goto bad_nofile;
1484 if (!percpu_ref_tryget_live(&si->users))
1485 goto out;
1486 /*
1487 * Guarantee the si->users are checked before accessing other
1488 * fields of swap_info_struct.
1489 *
1490 * Paired with the spin_unlock() after setup_swap_info() in
1491 * enable_swap_info().
1492 */
1493 smp_rmb();
1494 offset = swp_offset(entry);
1495 if (offset >= si->max)
1496 goto put_out;
1497
1498 return si;
1499bad_nofile:
1500 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1501out:
1502 return NULL;
1503put_out:
1504 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1505 percpu_ref_put(&si->users);
1506 return NULL;
1507}
1508
1509static unsigned char __swap_entry_free(struct swap_info_struct *si,
1510 swp_entry_t entry)
1511{
1512 struct swap_cluster_info *ci;
1513 unsigned long offset = swp_offset(entry);
1514 unsigned char usage;
1515
1516 ci = lock_cluster_or_swap_info(si, offset);
1517 usage = __swap_entry_free_locked(si, offset, 1);
1518 unlock_cluster_or_swap_info(si, ci);
1519 if (!usage)
1520 free_swap_slot(entry);
1521
1522 return usage;
1523}
1524
1525static bool __swap_entries_free(struct swap_info_struct *si,
1526 swp_entry_t entry, int nr)
1527{
1528 unsigned long offset = swp_offset(entry);
1529 unsigned int type = swp_type(entry);
1530 struct swap_cluster_info *ci;
1531 bool has_cache = false;
1532 unsigned char count;
1533 int i;
1534
1535 if (nr <= 1 || swap_count(data_race(si->swap_map[offset])) != 1)
1536 goto fallback;
1537 /* cross into another cluster */
1538 if (nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER)
1539 goto fallback;
1540
1541 ci = lock_cluster_or_swap_info(si, offset);
1542 if (!swap_is_last_map(si, offset, nr, &has_cache)) {
1543 unlock_cluster_or_swap_info(si, ci);
1544 goto fallback;
1545 }
1546 for (i = 0; i < nr; i++)
1547 WRITE_ONCE(si->swap_map[offset + i], SWAP_HAS_CACHE);
1548 unlock_cluster_or_swap_info(si, ci);
1549
1550 if (!has_cache) {
1551 for (i = 0; i < nr; i++)
1552 zswap_invalidate(swp_entry(si->type, offset + i));
1553 spin_lock(&si->lock);
1554 swap_entry_range_free(si, entry, nr);
1555 spin_unlock(&si->lock);
1556 }
1557 return has_cache;
1558
1559fallback:
1560 for (i = 0; i < nr; i++) {
1561 if (data_race(si->swap_map[offset + i])) {
1562 count = __swap_entry_free(si, swp_entry(type, offset + i));
1563 if (count == SWAP_HAS_CACHE)
1564 has_cache = true;
1565 } else {
1566 WARN_ON_ONCE(1);
1567 }
1568 }
1569 return has_cache;
1570}
1571
1572/*
1573 * Drop the last HAS_CACHE flag of swap entries, caller have to
1574 * ensure all entries belong to the same cgroup.
1575 */
1576static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry,
1577 unsigned int nr_pages)
1578{
1579 unsigned long offset = swp_offset(entry);
1580 unsigned char *map = si->swap_map + offset;
1581 unsigned char *map_end = map + nr_pages;
1582 struct swap_cluster_info *ci;
1583
1584 ci = lock_cluster(si, offset);
1585 do {
1586 VM_BUG_ON(*map != SWAP_HAS_CACHE);
1587 *map = 0;
1588 } while (++map < map_end);
1589 dec_cluster_info_page(si, ci, nr_pages);
1590 unlock_cluster(ci);
1591
1592 mem_cgroup_uncharge_swap(entry, nr_pages);
1593 swap_range_free(si, offset, nr_pages);
1594}
1595
1596static void cluster_swap_free_nr(struct swap_info_struct *si,
1597 unsigned long offset, int nr_pages,
1598 unsigned char usage)
1599{
1600 struct swap_cluster_info *ci;
1601 DECLARE_BITMAP(to_free, BITS_PER_LONG) = { 0 };
1602 int i, nr;
1603
1604 ci = lock_cluster_or_swap_info(si, offset);
1605 while (nr_pages) {
1606 nr = min(BITS_PER_LONG, nr_pages);
1607 for (i = 0; i < nr; i++) {
1608 if (!__swap_entry_free_locked(si, offset + i, usage))
1609 bitmap_set(to_free, i, 1);
1610 }
1611 if (!bitmap_empty(to_free, BITS_PER_LONG)) {
1612 unlock_cluster_or_swap_info(si, ci);
1613 for_each_set_bit(i, to_free, BITS_PER_LONG)
1614 free_swap_slot(swp_entry(si->type, offset + i));
1615 if (nr == nr_pages)
1616 return;
1617 bitmap_clear(to_free, 0, BITS_PER_LONG);
1618 ci = lock_cluster_or_swap_info(si, offset);
1619 }
1620 offset += nr;
1621 nr_pages -= nr;
1622 }
1623 unlock_cluster_or_swap_info(si, ci);
1624}
1625
1626/*
1627 * Caller has made sure that the swap device corresponding to entry
1628 * is still around or has not been recycled.
1629 */
1630void swap_free_nr(swp_entry_t entry, int nr_pages)
1631{
1632 int nr;
1633 struct swap_info_struct *sis;
1634 unsigned long offset = swp_offset(entry);
1635
1636 sis = _swap_info_get(entry);
1637 if (!sis)
1638 return;
1639
1640 while (nr_pages) {
1641 nr = min_t(int, nr_pages, SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER);
1642 cluster_swap_free_nr(sis, offset, nr, 1);
1643 offset += nr;
1644 nr_pages -= nr;
1645 }
1646}
1647
1648/*
1649 * Called after dropping swapcache to decrease refcnt to swap entries.
1650 */
1651void put_swap_folio(struct folio *folio, swp_entry_t entry)
1652{
1653 unsigned long offset = swp_offset(entry);
1654 struct swap_cluster_info *ci;
1655 struct swap_info_struct *si;
1656 int size = 1 << swap_entry_order(folio_order(folio));
1657
1658 si = _swap_info_get(entry);
1659 if (!si)
1660 return;
1661
1662 ci = lock_cluster_or_swap_info(si, offset);
1663 if (size > 1 && swap_is_has_cache(si, offset, size)) {
1664 unlock_cluster_or_swap_info(si, ci);
1665 spin_lock(&si->lock);
1666 swap_entry_range_free(si, entry, size);
1667 spin_unlock(&si->lock);
1668 return;
1669 }
1670 for (int i = 0; i < size; i++, entry.val++) {
1671 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1672 unlock_cluster_or_swap_info(si, ci);
1673 free_swap_slot(entry);
1674 if (i == size - 1)
1675 return;
1676 lock_cluster_or_swap_info(si, offset);
1677 }
1678 }
1679 unlock_cluster_or_swap_info(si, ci);
1680}
1681
1682static int swp_entry_cmp(const void *ent1, const void *ent2)
1683{
1684 const swp_entry_t *e1 = ent1, *e2 = ent2;
1685
1686 return (int)swp_type(*e1) - (int)swp_type(*e2);
1687}
1688
1689void swapcache_free_entries(swp_entry_t *entries, int n)
1690{
1691 struct swap_info_struct *p, *prev;
1692 int i;
1693
1694 if (n <= 0)
1695 return;
1696
1697 prev = NULL;
1698 p = NULL;
1699
1700 /*
1701 * Sort swap entries by swap device, so each lock is only taken once.
1702 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1703 * so low that it isn't necessary to optimize further.
1704 */
1705 if (nr_swapfiles > 1)
1706 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1707 for (i = 0; i < n; ++i) {
1708 p = swap_info_get_cont(entries[i], prev);
1709 if (p)
1710 swap_entry_range_free(p, entries[i], 1);
1711 prev = p;
1712 }
1713 if (p)
1714 spin_unlock(&p->lock);
1715}
1716
1717int __swap_count(swp_entry_t entry)
1718{
1719 struct swap_info_struct *si = swp_swap_info(entry);
1720 pgoff_t offset = swp_offset(entry);
1721
1722 return swap_count(si->swap_map[offset]);
1723}
1724
1725/*
1726 * How many references to @entry are currently swapped out?
1727 * This does not give an exact answer when swap count is continued,
1728 * but does include the high COUNT_CONTINUED flag to allow for that.
1729 */
1730int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1731{
1732 pgoff_t offset = swp_offset(entry);
1733 struct swap_cluster_info *ci;
1734 int count;
1735
1736 ci = lock_cluster_or_swap_info(si, offset);
1737 count = swap_count(si->swap_map[offset]);
1738 unlock_cluster_or_swap_info(si, ci);
1739 return count;
1740}
1741
1742/*
1743 * How many references to @entry are currently swapped out?
1744 * This considers COUNT_CONTINUED so it returns exact answer.
1745 */
1746int swp_swapcount(swp_entry_t entry)
1747{
1748 int count, tmp_count, n;
1749 struct swap_info_struct *si;
1750 struct swap_cluster_info *ci;
1751 struct page *page;
1752 pgoff_t offset;
1753 unsigned char *map;
1754
1755 si = _swap_info_get(entry);
1756 if (!si)
1757 return 0;
1758
1759 offset = swp_offset(entry);
1760
1761 ci = lock_cluster_or_swap_info(si, offset);
1762
1763 count = swap_count(si->swap_map[offset]);
1764 if (!(count & COUNT_CONTINUED))
1765 goto out;
1766
1767 count &= ~COUNT_CONTINUED;
1768 n = SWAP_MAP_MAX + 1;
1769
1770 page = vmalloc_to_page(si->swap_map + offset);
1771 offset &= ~PAGE_MASK;
1772 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1773
1774 do {
1775 page = list_next_entry(page, lru);
1776 map = kmap_local_page(page);
1777 tmp_count = map[offset];
1778 kunmap_local(map);
1779
1780 count += (tmp_count & ~COUNT_CONTINUED) * n;
1781 n *= (SWAP_CONT_MAX + 1);
1782 } while (tmp_count & COUNT_CONTINUED);
1783out:
1784 unlock_cluster_or_swap_info(si, ci);
1785 return count;
1786}
1787
1788static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1789 swp_entry_t entry, int order)
1790{
1791 struct swap_cluster_info *ci;
1792 unsigned char *map = si->swap_map;
1793 unsigned int nr_pages = 1 << order;
1794 unsigned long roffset = swp_offset(entry);
1795 unsigned long offset = round_down(roffset, nr_pages);
1796 int i;
1797 bool ret = false;
1798
1799 ci = lock_cluster_or_swap_info(si, offset);
1800 if (!ci || nr_pages == 1) {
1801 if (swap_count(map[roffset]))
1802 ret = true;
1803 goto unlock_out;
1804 }
1805 for (i = 0; i < nr_pages; i++) {
1806 if (swap_count(map[offset + i])) {
1807 ret = true;
1808 break;
1809 }
1810 }
1811unlock_out:
1812 unlock_cluster_or_swap_info(si, ci);
1813 return ret;
1814}
1815
1816static bool folio_swapped(struct folio *folio)
1817{
1818 swp_entry_t entry = folio->swap;
1819 struct swap_info_struct *si = _swap_info_get(entry);
1820
1821 if (!si)
1822 return false;
1823
1824 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1825 return swap_swapcount(si, entry) != 0;
1826
1827 return swap_page_trans_huge_swapped(si, entry, folio_order(folio));
1828}
1829
1830static bool folio_swapcache_freeable(struct folio *folio)
1831{
1832 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1833
1834 if (!folio_test_swapcache(folio))
1835 return false;
1836 if (folio_test_writeback(folio))
1837 return false;
1838
1839 /*
1840 * Once hibernation has begun to create its image of memory,
1841 * there's a danger that one of the calls to folio_free_swap()
1842 * - most probably a call from __try_to_reclaim_swap() while
1843 * hibernation is allocating its own swap pages for the image,
1844 * but conceivably even a call from memory reclaim - will free
1845 * the swap from a folio which has already been recorded in the
1846 * image as a clean swapcache folio, and then reuse its swap for
1847 * another page of the image. On waking from hibernation, the
1848 * original folio might be freed under memory pressure, then
1849 * later read back in from swap, now with the wrong data.
1850 *
1851 * Hibernation suspends storage while it is writing the image
1852 * to disk so check that here.
1853 */
1854 if (pm_suspended_storage())
1855 return false;
1856
1857 return true;
1858}
1859
1860/**
1861 * folio_free_swap() - Free the swap space used for this folio.
1862 * @folio: The folio to remove.
1863 *
1864 * If swap is getting full, or if there are no more mappings of this folio,
1865 * then call folio_free_swap to free its swap space.
1866 *
1867 * Return: true if we were able to release the swap space.
1868 */
1869bool folio_free_swap(struct folio *folio)
1870{
1871 if (!folio_swapcache_freeable(folio))
1872 return false;
1873 if (folio_swapped(folio))
1874 return false;
1875
1876 delete_from_swap_cache(folio);
1877 folio_set_dirty(folio);
1878 return true;
1879}
1880
1881/**
1882 * free_swap_and_cache_nr() - Release reference on range of swap entries and
1883 * reclaim their cache if no more references remain.
1884 * @entry: First entry of range.
1885 * @nr: Number of entries in range.
1886 *
1887 * For each swap entry in the contiguous range, release a reference. If any swap
1888 * entries become free, try to reclaim their underlying folios, if present. The
1889 * offset range is defined by [entry.offset, entry.offset + nr).
1890 */
1891void free_swap_and_cache_nr(swp_entry_t entry, int nr)
1892{
1893 const unsigned long start_offset = swp_offset(entry);
1894 const unsigned long end_offset = start_offset + nr;
1895 struct swap_info_struct *si;
1896 bool any_only_cache = false;
1897 unsigned long offset;
1898
1899 if (non_swap_entry(entry))
1900 return;
1901
1902 si = get_swap_device(entry);
1903 if (!si)
1904 return;
1905
1906 if (WARN_ON(end_offset > si->max))
1907 goto out;
1908
1909 /*
1910 * First free all entries in the range.
1911 */
1912 any_only_cache = __swap_entries_free(si, entry, nr);
1913
1914 /*
1915 * Short-circuit the below loop if none of the entries had their
1916 * reference drop to zero.
1917 */
1918 if (!any_only_cache)
1919 goto out;
1920
1921 /*
1922 * Now go back over the range trying to reclaim the swap cache. This is
1923 * more efficient for large folios because we will only try to reclaim
1924 * the swap once per folio in the common case. If we do
1925 * __swap_entry_free() and __try_to_reclaim_swap() in the same loop, the
1926 * latter will get a reference and lock the folio for every individual
1927 * page but will only succeed once the swap slot for every subpage is
1928 * zero.
1929 */
1930 for (offset = start_offset; offset < end_offset; offset += nr) {
1931 nr = 1;
1932 if (READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
1933 /*
1934 * Folios are always naturally aligned in swap so
1935 * advance forward to the next boundary. Zero means no
1936 * folio was found for the swap entry, so advance by 1
1937 * in this case. Negative value means folio was found
1938 * but could not be reclaimed. Here we can still advance
1939 * to the next boundary.
1940 */
1941 nr = __try_to_reclaim_swap(si, offset,
1942 TTRS_UNMAPPED | TTRS_FULL);
1943 if (nr == 0)
1944 nr = 1;
1945 else if (nr < 0)
1946 nr = -nr;
1947 nr = ALIGN(offset + 1, nr) - offset;
1948 }
1949 }
1950
1951out:
1952 put_swap_device(si);
1953}
1954
1955#ifdef CONFIG_HIBERNATION
1956
1957swp_entry_t get_swap_page_of_type(int type)
1958{
1959 struct swap_info_struct *si = swap_type_to_swap_info(type);
1960 swp_entry_t entry = {0};
1961
1962 if (!si)
1963 goto fail;
1964
1965 /* This is called for allocating swap entry, not cache */
1966 spin_lock(&si->lock);
1967 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry, 0))
1968 atomic_long_dec(&nr_swap_pages);
1969 spin_unlock(&si->lock);
1970fail:
1971 return entry;
1972}
1973
1974/*
1975 * Find the swap type that corresponds to given device (if any).
1976 *
1977 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1978 * from 0, in which the swap header is expected to be located.
1979 *
1980 * This is needed for the suspend to disk (aka swsusp).
1981 */
1982int swap_type_of(dev_t device, sector_t offset)
1983{
1984 int type;
1985
1986 if (!device)
1987 return -1;
1988
1989 spin_lock(&swap_lock);
1990 for (type = 0; type < nr_swapfiles; type++) {
1991 struct swap_info_struct *sis = swap_info[type];
1992
1993 if (!(sis->flags & SWP_WRITEOK))
1994 continue;
1995
1996 if (device == sis->bdev->bd_dev) {
1997 struct swap_extent *se = first_se(sis);
1998
1999 if (se->start_block == offset) {
2000 spin_unlock(&swap_lock);
2001 return type;
2002 }
2003 }
2004 }
2005 spin_unlock(&swap_lock);
2006 return -ENODEV;
2007}
2008
2009int find_first_swap(dev_t *device)
2010{
2011 int type;
2012
2013 spin_lock(&swap_lock);
2014 for (type = 0; type < nr_swapfiles; type++) {
2015 struct swap_info_struct *sis = swap_info[type];
2016
2017 if (!(sis->flags & SWP_WRITEOK))
2018 continue;
2019 *device = sis->bdev->bd_dev;
2020 spin_unlock(&swap_lock);
2021 return type;
2022 }
2023 spin_unlock(&swap_lock);
2024 return -ENODEV;
2025}
2026
2027/*
2028 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
2029 * corresponding to given index in swap_info (swap type).
2030 */
2031sector_t swapdev_block(int type, pgoff_t offset)
2032{
2033 struct swap_info_struct *si = swap_type_to_swap_info(type);
2034 struct swap_extent *se;
2035
2036 if (!si || !(si->flags & SWP_WRITEOK))
2037 return 0;
2038 se = offset_to_swap_extent(si, offset);
2039 return se->start_block + (offset - se->start_page);
2040}
2041
2042/*
2043 * Return either the total number of swap pages of given type, or the number
2044 * of free pages of that type (depending on @free)
2045 *
2046 * This is needed for software suspend
2047 */
2048unsigned int count_swap_pages(int type, int free)
2049{
2050 unsigned int n = 0;
2051
2052 spin_lock(&swap_lock);
2053 if ((unsigned int)type < nr_swapfiles) {
2054 struct swap_info_struct *sis = swap_info[type];
2055
2056 spin_lock(&sis->lock);
2057 if (sis->flags & SWP_WRITEOK) {
2058 n = sis->pages;
2059 if (free)
2060 n -= sis->inuse_pages;
2061 }
2062 spin_unlock(&sis->lock);
2063 }
2064 spin_unlock(&swap_lock);
2065 return n;
2066}
2067#endif /* CONFIG_HIBERNATION */
2068
2069static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
2070{
2071 return pte_same(pte_swp_clear_flags(pte), swp_pte);
2072}
2073
2074/*
2075 * No need to decide whether this PTE shares the swap entry with others,
2076 * just let do_wp_page work it out if a write is requested later - to
2077 * force COW, vm_page_prot omits write permission from any private vma.
2078 */
2079static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
2080 unsigned long addr, swp_entry_t entry, struct folio *folio)
2081{
2082 struct page *page;
2083 struct folio *swapcache;
2084 spinlock_t *ptl;
2085 pte_t *pte, new_pte, old_pte;
2086 bool hwpoisoned = false;
2087 int ret = 1;
2088
2089 swapcache = folio;
2090 folio = ksm_might_need_to_copy(folio, vma, addr);
2091 if (unlikely(!folio))
2092 return -ENOMEM;
2093 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
2094 hwpoisoned = true;
2095 folio = swapcache;
2096 }
2097
2098 page = folio_file_page(folio, swp_offset(entry));
2099 if (PageHWPoison(page))
2100 hwpoisoned = true;
2101
2102 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
2103 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte),
2104 swp_entry_to_pte(entry)))) {
2105 ret = 0;
2106 goto out;
2107 }
2108
2109 old_pte = ptep_get(pte);
2110
2111 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) {
2112 swp_entry_t swp_entry;
2113
2114 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
2115 if (hwpoisoned) {
2116 swp_entry = make_hwpoison_entry(page);
2117 } else {
2118 swp_entry = make_poisoned_swp_entry();
2119 }
2120 new_pte = swp_entry_to_pte(swp_entry);
2121 ret = 0;
2122 goto setpte;
2123 }
2124
2125 /*
2126 * Some architectures may have to restore extra metadata to the page
2127 * when reading from swap. This metadata may be indexed by swap entry
2128 * so this must be called before swap_free().
2129 */
2130 arch_swap_restore(folio_swap(entry, folio), folio);
2131
2132 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
2133 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
2134 folio_get(folio);
2135 if (folio == swapcache) {
2136 rmap_t rmap_flags = RMAP_NONE;
2137
2138 /*
2139 * See do_swap_page(): writeback would be problematic.
2140 * However, we do a folio_wait_writeback() just before this
2141 * call and have the folio locked.
2142 */
2143 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
2144 if (pte_swp_exclusive(old_pte))
2145 rmap_flags |= RMAP_EXCLUSIVE;
2146 /*
2147 * We currently only expect small !anon folios, which are either
2148 * fully exclusive or fully shared. If we ever get large folios
2149 * here, we have to be careful.
2150 */
2151 if (!folio_test_anon(folio)) {
2152 VM_WARN_ON_ONCE(folio_test_large(folio));
2153 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
2154 folio_add_new_anon_rmap(folio, vma, addr, rmap_flags);
2155 } else {
2156 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
2157 }
2158 } else { /* ksm created a completely new copy */
2159 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
2160 folio_add_lru_vma(folio, vma);
2161 }
2162 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
2163 if (pte_swp_soft_dirty(old_pte))
2164 new_pte = pte_mksoft_dirty(new_pte);
2165 if (pte_swp_uffd_wp(old_pte))
2166 new_pte = pte_mkuffd_wp(new_pte);
2167setpte:
2168 set_pte_at(vma->vm_mm, addr, pte, new_pte);
2169 swap_free(entry);
2170out:
2171 if (pte)
2172 pte_unmap_unlock(pte, ptl);
2173 if (folio != swapcache) {
2174 folio_unlock(folio);
2175 folio_put(folio);
2176 }
2177 return ret;
2178}
2179
2180static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
2181 unsigned long addr, unsigned long end,
2182 unsigned int type)
2183{
2184 pte_t *pte = NULL;
2185 struct swap_info_struct *si;
2186
2187 si = swap_info[type];
2188 do {
2189 struct folio *folio;
2190 unsigned long offset;
2191 unsigned char swp_count;
2192 swp_entry_t entry;
2193 int ret;
2194 pte_t ptent;
2195
2196 if (!pte++) {
2197 pte = pte_offset_map(pmd, addr);
2198 if (!pte)
2199 break;
2200 }
2201
2202 ptent = ptep_get_lockless(pte);
2203
2204 if (!is_swap_pte(ptent))
2205 continue;
2206
2207 entry = pte_to_swp_entry(ptent);
2208 if (swp_type(entry) != type)
2209 continue;
2210
2211 offset = swp_offset(entry);
2212 pte_unmap(pte);
2213 pte = NULL;
2214
2215 folio = swap_cache_get_folio(entry, vma, addr);
2216 if (!folio) {
2217 struct vm_fault vmf = {
2218 .vma = vma,
2219 .address = addr,
2220 .real_address = addr,
2221 .pmd = pmd,
2222 };
2223
2224 folio = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2225 &vmf);
2226 }
2227 if (!folio) {
2228 swp_count = READ_ONCE(si->swap_map[offset]);
2229 if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
2230 continue;
2231 return -ENOMEM;
2232 }
2233
2234 folio_lock(folio);
2235 folio_wait_writeback(folio);
2236 ret = unuse_pte(vma, pmd, addr, entry, folio);
2237 if (ret < 0) {
2238 folio_unlock(folio);
2239 folio_put(folio);
2240 return ret;
2241 }
2242
2243 folio_free_swap(folio);
2244 folio_unlock(folio);
2245 folio_put(folio);
2246 } while (addr += PAGE_SIZE, addr != end);
2247
2248 if (pte)
2249 pte_unmap(pte);
2250 return 0;
2251}
2252
2253static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2254 unsigned long addr, unsigned long end,
2255 unsigned int type)
2256{
2257 pmd_t *pmd;
2258 unsigned long next;
2259 int ret;
2260
2261 pmd = pmd_offset(pud, addr);
2262 do {
2263 cond_resched();
2264 next = pmd_addr_end(addr, end);
2265 ret = unuse_pte_range(vma, pmd, addr, next, type);
2266 if (ret)
2267 return ret;
2268 } while (pmd++, addr = next, addr != end);
2269 return 0;
2270}
2271
2272static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2273 unsigned long addr, unsigned long end,
2274 unsigned int type)
2275{
2276 pud_t *pud;
2277 unsigned long next;
2278 int ret;
2279
2280 pud = pud_offset(p4d, addr);
2281 do {
2282 next = pud_addr_end(addr, end);
2283 if (pud_none_or_clear_bad(pud))
2284 continue;
2285 ret = unuse_pmd_range(vma, pud, addr, next, type);
2286 if (ret)
2287 return ret;
2288 } while (pud++, addr = next, addr != end);
2289 return 0;
2290}
2291
2292static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2293 unsigned long addr, unsigned long end,
2294 unsigned int type)
2295{
2296 p4d_t *p4d;
2297 unsigned long next;
2298 int ret;
2299
2300 p4d = p4d_offset(pgd, addr);
2301 do {
2302 next = p4d_addr_end(addr, end);
2303 if (p4d_none_or_clear_bad(p4d))
2304 continue;
2305 ret = unuse_pud_range(vma, p4d, addr, next, type);
2306 if (ret)
2307 return ret;
2308 } while (p4d++, addr = next, addr != end);
2309 return 0;
2310}
2311
2312static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
2313{
2314 pgd_t *pgd;
2315 unsigned long addr, end, next;
2316 int ret;
2317
2318 addr = vma->vm_start;
2319 end = vma->vm_end;
2320
2321 pgd = pgd_offset(vma->vm_mm, addr);
2322 do {
2323 next = pgd_addr_end(addr, end);
2324 if (pgd_none_or_clear_bad(pgd))
2325 continue;
2326 ret = unuse_p4d_range(vma, pgd, addr, next, type);
2327 if (ret)
2328 return ret;
2329 } while (pgd++, addr = next, addr != end);
2330 return 0;
2331}
2332
2333static int unuse_mm(struct mm_struct *mm, unsigned int type)
2334{
2335 struct vm_area_struct *vma;
2336 int ret = 0;
2337 VMA_ITERATOR(vmi, mm, 0);
2338
2339 mmap_read_lock(mm);
2340 for_each_vma(vmi, vma) {
2341 if (vma->anon_vma && !is_vm_hugetlb_page(vma)) {
2342 ret = unuse_vma(vma, type);
2343 if (ret)
2344 break;
2345 }
2346
2347 cond_resched();
2348 }
2349 mmap_read_unlock(mm);
2350 return ret;
2351}
2352
2353/*
2354 * Scan swap_map from current position to next entry still in use.
2355 * Return 0 if there are no inuse entries after prev till end of
2356 * the map.
2357 */
2358static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2359 unsigned int prev)
2360{
2361 unsigned int i;
2362 unsigned char count;
2363
2364 /*
2365 * No need for swap_lock here: we're just looking
2366 * for whether an entry is in use, not modifying it; false
2367 * hits are okay, and sys_swapoff() has already prevented new
2368 * allocations from this area (while holding swap_lock).
2369 */
2370 for (i = prev + 1; i < si->max; i++) {
2371 count = READ_ONCE(si->swap_map[i]);
2372 if (count && swap_count(count) != SWAP_MAP_BAD)
2373 break;
2374 if ((i % LATENCY_LIMIT) == 0)
2375 cond_resched();
2376 }
2377
2378 if (i == si->max)
2379 i = 0;
2380
2381 return i;
2382}
2383
2384static int try_to_unuse(unsigned int type)
2385{
2386 struct mm_struct *prev_mm;
2387 struct mm_struct *mm;
2388 struct list_head *p;
2389 int retval = 0;
2390 struct swap_info_struct *si = swap_info[type];
2391 struct folio *folio;
2392 swp_entry_t entry;
2393 unsigned int i;
2394
2395 if (!READ_ONCE(si->inuse_pages))
2396 goto success;
2397
2398retry:
2399 retval = shmem_unuse(type);
2400 if (retval)
2401 return retval;
2402
2403 prev_mm = &init_mm;
2404 mmget(prev_mm);
2405
2406 spin_lock(&mmlist_lock);
2407 p = &init_mm.mmlist;
2408 while (READ_ONCE(si->inuse_pages) &&
2409 !signal_pending(current) &&
2410 (p = p->next) != &init_mm.mmlist) {
2411
2412 mm = list_entry(p, struct mm_struct, mmlist);
2413 if (!mmget_not_zero(mm))
2414 continue;
2415 spin_unlock(&mmlist_lock);
2416 mmput(prev_mm);
2417 prev_mm = mm;
2418 retval = unuse_mm(mm, type);
2419 if (retval) {
2420 mmput(prev_mm);
2421 return retval;
2422 }
2423
2424 /*
2425 * Make sure that we aren't completely killing
2426 * interactive performance.
2427 */
2428 cond_resched();
2429 spin_lock(&mmlist_lock);
2430 }
2431 spin_unlock(&mmlist_lock);
2432
2433 mmput(prev_mm);
2434
2435 i = 0;
2436 while (READ_ONCE(si->inuse_pages) &&
2437 !signal_pending(current) &&
2438 (i = find_next_to_unuse(si, i)) != 0) {
2439
2440 entry = swp_entry(type, i);
2441 folio = filemap_get_folio(swap_address_space(entry), swap_cache_index(entry));
2442 if (IS_ERR(folio))
2443 continue;
2444
2445 /*
2446 * It is conceivable that a racing task removed this folio from
2447 * swap cache just before we acquired the page lock. The folio
2448 * might even be back in swap cache on another swap area. But
2449 * that is okay, folio_free_swap() only removes stale folios.
2450 */
2451 folio_lock(folio);
2452 folio_wait_writeback(folio);
2453 folio_free_swap(folio);
2454 folio_unlock(folio);
2455 folio_put(folio);
2456 }
2457
2458 /*
2459 * Lets check again to see if there are still swap entries in the map.
2460 * If yes, we would need to do retry the unuse logic again.
2461 * Under global memory pressure, swap entries can be reinserted back
2462 * into process space after the mmlist loop above passes over them.
2463 *
2464 * Limit the number of retries? No: when mmget_not_zero()
2465 * above fails, that mm is likely to be freeing swap from
2466 * exit_mmap(), which proceeds at its own independent pace;
2467 * and even shmem_writepage() could have been preempted after
2468 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2469 * and robust (though cpu-intensive) just to keep retrying.
2470 */
2471 if (READ_ONCE(si->inuse_pages)) {
2472 if (!signal_pending(current))
2473 goto retry;
2474 return -EINTR;
2475 }
2476
2477success:
2478 /*
2479 * Make sure that further cleanups after try_to_unuse() returns happen
2480 * after swap_range_free() reduces si->inuse_pages to 0.
2481 */
2482 smp_mb();
2483 return 0;
2484}
2485
2486/*
2487 * After a successful try_to_unuse, if no swap is now in use, we know
2488 * we can empty the mmlist. swap_lock must be held on entry and exit.
2489 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2490 * added to the mmlist just after page_duplicate - before would be racy.
2491 */
2492static void drain_mmlist(void)
2493{
2494 struct list_head *p, *next;
2495 unsigned int type;
2496
2497 for (type = 0; type < nr_swapfiles; type++)
2498 if (swap_info[type]->inuse_pages)
2499 return;
2500 spin_lock(&mmlist_lock);
2501 list_for_each_safe(p, next, &init_mm.mmlist)
2502 list_del_init(p);
2503 spin_unlock(&mmlist_lock);
2504}
2505
2506/*
2507 * Free all of a swapdev's extent information
2508 */
2509static void destroy_swap_extents(struct swap_info_struct *sis)
2510{
2511 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2512 struct rb_node *rb = sis->swap_extent_root.rb_node;
2513 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2514
2515 rb_erase(rb, &sis->swap_extent_root);
2516 kfree(se);
2517 }
2518
2519 if (sis->flags & SWP_ACTIVATED) {
2520 struct file *swap_file = sis->swap_file;
2521 struct address_space *mapping = swap_file->f_mapping;
2522
2523 sis->flags &= ~SWP_ACTIVATED;
2524 if (mapping->a_ops->swap_deactivate)
2525 mapping->a_ops->swap_deactivate(swap_file);
2526 }
2527}
2528
2529/*
2530 * Add a block range (and the corresponding page range) into this swapdev's
2531 * extent tree.
2532 *
2533 * This function rather assumes that it is called in ascending page order.
2534 */
2535int
2536add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2537 unsigned long nr_pages, sector_t start_block)
2538{
2539 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2540 struct swap_extent *se;
2541 struct swap_extent *new_se;
2542
2543 /*
2544 * place the new node at the right most since the
2545 * function is called in ascending page order.
2546 */
2547 while (*link) {
2548 parent = *link;
2549 link = &parent->rb_right;
2550 }
2551
2552 if (parent) {
2553 se = rb_entry(parent, struct swap_extent, rb_node);
2554 BUG_ON(se->start_page + se->nr_pages != start_page);
2555 if (se->start_block + se->nr_pages == start_block) {
2556 /* Merge it */
2557 se->nr_pages += nr_pages;
2558 return 0;
2559 }
2560 }
2561
2562 /* No merge, insert a new extent. */
2563 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2564 if (new_se == NULL)
2565 return -ENOMEM;
2566 new_se->start_page = start_page;
2567 new_se->nr_pages = nr_pages;
2568 new_se->start_block = start_block;
2569
2570 rb_link_node(&new_se->rb_node, parent, link);
2571 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2572 return 1;
2573}
2574EXPORT_SYMBOL_GPL(add_swap_extent);
2575
2576/*
2577 * A `swap extent' is a simple thing which maps a contiguous range of pages
2578 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2579 * built at swapon time and is then used at swap_writepage/swap_read_folio
2580 * time for locating where on disk a page belongs.
2581 *
2582 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2583 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2584 * swap files identically.
2585 *
2586 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2587 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2588 * swapfiles are handled *identically* after swapon time.
2589 *
2590 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2591 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2592 * blocks are found which do not fall within the PAGE_SIZE alignment
2593 * requirements, they are simply tossed out - we will never use those blocks
2594 * for swapping.
2595 *
2596 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2597 * prevents users from writing to the swap device, which will corrupt memory.
2598 *
2599 * The amount of disk space which a single swap extent represents varies.
2600 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2601 * extents in the rbtree. - akpm.
2602 */
2603static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2604{
2605 struct file *swap_file = sis->swap_file;
2606 struct address_space *mapping = swap_file->f_mapping;
2607 struct inode *inode = mapping->host;
2608 int ret;
2609
2610 if (S_ISBLK(inode->i_mode)) {
2611 ret = add_swap_extent(sis, 0, sis->max, 0);
2612 *span = sis->pages;
2613 return ret;
2614 }
2615
2616 if (mapping->a_ops->swap_activate) {
2617 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2618 if (ret < 0)
2619 return ret;
2620 sis->flags |= SWP_ACTIVATED;
2621 if ((sis->flags & SWP_FS_OPS) &&
2622 sio_pool_init() != 0) {
2623 destroy_swap_extents(sis);
2624 return -ENOMEM;
2625 }
2626 return ret;
2627 }
2628
2629 return generic_swapfile_activate(sis, swap_file, span);
2630}
2631
2632static int swap_node(struct swap_info_struct *si)
2633{
2634 struct block_device *bdev;
2635
2636 if (si->bdev)
2637 bdev = si->bdev;
2638 else
2639 bdev = si->swap_file->f_inode->i_sb->s_bdev;
2640
2641 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2642}
2643
2644static void setup_swap_info(struct swap_info_struct *si, int prio,
2645 unsigned char *swap_map,
2646 struct swap_cluster_info *cluster_info,
2647 unsigned long *zeromap)
2648{
2649 int i;
2650
2651 if (prio >= 0)
2652 si->prio = prio;
2653 else
2654 si->prio = --least_priority;
2655 /*
2656 * the plist prio is negated because plist ordering is
2657 * low-to-high, while swap ordering is high-to-low
2658 */
2659 si->list.prio = -si->prio;
2660 for_each_node(i) {
2661 if (si->prio >= 0)
2662 si->avail_lists[i].prio = -si->prio;
2663 else {
2664 if (swap_node(si) == i)
2665 si->avail_lists[i].prio = 1;
2666 else
2667 si->avail_lists[i].prio = -si->prio;
2668 }
2669 }
2670 si->swap_map = swap_map;
2671 si->cluster_info = cluster_info;
2672 si->zeromap = zeromap;
2673}
2674
2675static void _enable_swap_info(struct swap_info_struct *si)
2676{
2677 si->flags |= SWP_WRITEOK;
2678 atomic_long_add(si->pages, &nr_swap_pages);
2679 total_swap_pages += si->pages;
2680
2681 assert_spin_locked(&swap_lock);
2682 /*
2683 * both lists are plists, and thus priority ordered.
2684 * swap_active_head needs to be priority ordered for swapoff(),
2685 * which on removal of any swap_info_struct with an auto-assigned
2686 * (i.e. negative) priority increments the auto-assigned priority
2687 * of any lower-priority swap_info_structs.
2688 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2689 * which allocates swap pages from the highest available priority
2690 * swap_info_struct.
2691 */
2692 plist_add(&si->list, &swap_active_head);
2693
2694 /* add to available list iff swap device is not full */
2695 if (si->highest_bit)
2696 add_to_avail_list(si);
2697}
2698
2699static void enable_swap_info(struct swap_info_struct *si, int prio,
2700 unsigned char *swap_map,
2701 struct swap_cluster_info *cluster_info,
2702 unsigned long *zeromap)
2703{
2704 spin_lock(&swap_lock);
2705 spin_lock(&si->lock);
2706 setup_swap_info(si, prio, swap_map, cluster_info, zeromap);
2707 spin_unlock(&si->lock);
2708 spin_unlock(&swap_lock);
2709 /*
2710 * Finished initializing swap device, now it's safe to reference it.
2711 */
2712 percpu_ref_resurrect(&si->users);
2713 spin_lock(&swap_lock);
2714 spin_lock(&si->lock);
2715 _enable_swap_info(si);
2716 spin_unlock(&si->lock);
2717 spin_unlock(&swap_lock);
2718}
2719
2720static void reinsert_swap_info(struct swap_info_struct *si)
2721{
2722 spin_lock(&swap_lock);
2723 spin_lock(&si->lock);
2724 setup_swap_info(si, si->prio, si->swap_map, si->cluster_info, si->zeromap);
2725 _enable_swap_info(si);
2726 spin_unlock(&si->lock);
2727 spin_unlock(&swap_lock);
2728}
2729
2730static bool __has_usable_swap(void)
2731{
2732 return !plist_head_empty(&swap_active_head);
2733}
2734
2735bool has_usable_swap(void)
2736{
2737 bool ret;
2738
2739 spin_lock(&swap_lock);
2740 ret = __has_usable_swap();
2741 spin_unlock(&swap_lock);
2742 return ret;
2743}
2744
2745SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2746{
2747 struct swap_info_struct *p = NULL;
2748 unsigned char *swap_map;
2749 unsigned long *zeromap;
2750 struct swap_cluster_info *cluster_info;
2751 struct file *swap_file, *victim;
2752 struct address_space *mapping;
2753 struct inode *inode;
2754 struct filename *pathname;
2755 int err, found = 0;
2756
2757 if (!capable(CAP_SYS_ADMIN))
2758 return -EPERM;
2759
2760 BUG_ON(!current->mm);
2761
2762 pathname = getname(specialfile);
2763 if (IS_ERR(pathname))
2764 return PTR_ERR(pathname);
2765
2766 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2767 err = PTR_ERR(victim);
2768 if (IS_ERR(victim))
2769 goto out;
2770
2771 mapping = victim->f_mapping;
2772 spin_lock(&swap_lock);
2773 plist_for_each_entry(p, &swap_active_head, list) {
2774 if (p->flags & SWP_WRITEOK) {
2775 if (p->swap_file->f_mapping == mapping) {
2776 found = 1;
2777 break;
2778 }
2779 }
2780 }
2781 if (!found) {
2782 err = -EINVAL;
2783 spin_unlock(&swap_lock);
2784 goto out_dput;
2785 }
2786 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2787 vm_unacct_memory(p->pages);
2788 else {
2789 err = -ENOMEM;
2790 spin_unlock(&swap_lock);
2791 goto out_dput;
2792 }
2793 spin_lock(&p->lock);
2794 del_from_avail_list(p);
2795 if (p->prio < 0) {
2796 struct swap_info_struct *si = p;
2797 int nid;
2798
2799 plist_for_each_entry_continue(si, &swap_active_head, list) {
2800 si->prio++;
2801 si->list.prio--;
2802 for_each_node(nid) {
2803 if (si->avail_lists[nid].prio != 1)
2804 si->avail_lists[nid].prio--;
2805 }
2806 }
2807 least_priority++;
2808 }
2809 plist_del(&p->list, &swap_active_head);
2810 atomic_long_sub(p->pages, &nr_swap_pages);
2811 total_swap_pages -= p->pages;
2812 p->flags &= ~SWP_WRITEOK;
2813 spin_unlock(&p->lock);
2814 spin_unlock(&swap_lock);
2815
2816 disable_swap_slots_cache_lock();
2817
2818 set_current_oom_origin();
2819 err = try_to_unuse(p->type);
2820 clear_current_oom_origin();
2821
2822 if (err) {
2823 /* re-insert swap space back into swap_list */
2824 reinsert_swap_info(p);
2825 reenable_swap_slots_cache_unlock();
2826 goto out_dput;
2827 }
2828
2829 reenable_swap_slots_cache_unlock();
2830
2831 /*
2832 * Wait for swap operations protected by get/put_swap_device()
2833 * to complete. Because of synchronize_rcu() here, all swap
2834 * operations protected by RCU reader side lock (including any
2835 * spinlock) will be waited too. This makes it easy to
2836 * prevent folio_test_swapcache() and the following swap cache
2837 * operations from racing with swapoff.
2838 */
2839 percpu_ref_kill(&p->users);
2840 synchronize_rcu();
2841 wait_for_completion(&p->comp);
2842
2843 flush_work(&p->discard_work);
2844 flush_work(&p->reclaim_work);
2845
2846 destroy_swap_extents(p);
2847 if (p->flags & SWP_CONTINUED)
2848 free_swap_count_continuations(p);
2849
2850 if (!p->bdev || !bdev_nonrot(p->bdev))
2851 atomic_dec(&nr_rotate_swap);
2852
2853 mutex_lock(&swapon_mutex);
2854 spin_lock(&swap_lock);
2855 spin_lock(&p->lock);
2856 drain_mmlist();
2857
2858 /* wait for anyone still in scan_swap_map_slots */
2859 p->highest_bit = 0; /* cuts scans short */
2860 while (p->flags >= SWP_SCANNING) {
2861 spin_unlock(&p->lock);
2862 spin_unlock(&swap_lock);
2863 schedule_timeout_uninterruptible(1);
2864 spin_lock(&swap_lock);
2865 spin_lock(&p->lock);
2866 }
2867
2868 swap_file = p->swap_file;
2869 p->swap_file = NULL;
2870 p->max = 0;
2871 swap_map = p->swap_map;
2872 p->swap_map = NULL;
2873 zeromap = p->zeromap;
2874 p->zeromap = NULL;
2875 cluster_info = p->cluster_info;
2876 p->cluster_info = NULL;
2877 spin_unlock(&p->lock);
2878 spin_unlock(&swap_lock);
2879 arch_swap_invalidate_area(p->type);
2880 zswap_swapoff(p->type);
2881 mutex_unlock(&swapon_mutex);
2882 free_percpu(p->percpu_cluster);
2883 p->percpu_cluster = NULL;
2884 free_percpu(p->cluster_next_cpu);
2885 p->cluster_next_cpu = NULL;
2886 vfree(swap_map);
2887 kvfree(zeromap);
2888 kvfree(cluster_info);
2889 /* Destroy swap account information */
2890 swap_cgroup_swapoff(p->type);
2891 exit_swap_address_space(p->type);
2892
2893 inode = mapping->host;
2894
2895 inode_lock(inode);
2896 inode->i_flags &= ~S_SWAPFILE;
2897 inode_unlock(inode);
2898 filp_close(swap_file, NULL);
2899
2900 /*
2901 * Clear the SWP_USED flag after all resources are freed so that swapon
2902 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2903 * not hold p->lock after we cleared its SWP_WRITEOK.
2904 */
2905 spin_lock(&swap_lock);
2906 p->flags = 0;
2907 spin_unlock(&swap_lock);
2908
2909 err = 0;
2910 atomic_inc(&proc_poll_event);
2911 wake_up_interruptible(&proc_poll_wait);
2912
2913out_dput:
2914 filp_close(victim, NULL);
2915out:
2916 putname(pathname);
2917 return err;
2918}
2919
2920#ifdef CONFIG_PROC_FS
2921static __poll_t swaps_poll(struct file *file, poll_table *wait)
2922{
2923 struct seq_file *seq = file->private_data;
2924
2925 poll_wait(file, &proc_poll_wait, wait);
2926
2927 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2928 seq->poll_event = atomic_read(&proc_poll_event);
2929 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2930 }
2931
2932 return EPOLLIN | EPOLLRDNORM;
2933}
2934
2935/* iterator */
2936static void *swap_start(struct seq_file *swap, loff_t *pos)
2937{
2938 struct swap_info_struct *si;
2939 int type;
2940 loff_t l = *pos;
2941
2942 mutex_lock(&swapon_mutex);
2943
2944 if (!l)
2945 return SEQ_START_TOKEN;
2946
2947 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2948 if (!(si->flags & SWP_USED) || !si->swap_map)
2949 continue;
2950 if (!--l)
2951 return si;
2952 }
2953
2954 return NULL;
2955}
2956
2957static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2958{
2959 struct swap_info_struct *si = v;
2960 int type;
2961
2962 if (v == SEQ_START_TOKEN)
2963 type = 0;
2964 else
2965 type = si->type + 1;
2966
2967 ++(*pos);
2968 for (; (si = swap_type_to_swap_info(type)); type++) {
2969 if (!(si->flags & SWP_USED) || !si->swap_map)
2970 continue;
2971 return si;
2972 }
2973
2974 return NULL;
2975}
2976
2977static void swap_stop(struct seq_file *swap, void *v)
2978{
2979 mutex_unlock(&swapon_mutex);
2980}
2981
2982static int swap_show(struct seq_file *swap, void *v)
2983{
2984 struct swap_info_struct *si = v;
2985 struct file *file;
2986 int len;
2987 unsigned long bytes, inuse;
2988
2989 if (si == SEQ_START_TOKEN) {
2990 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2991 return 0;
2992 }
2993
2994 bytes = K(si->pages);
2995 inuse = K(READ_ONCE(si->inuse_pages));
2996
2997 file = si->swap_file;
2998 len = seq_file_path(swap, file, " \t\n\\");
2999 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
3000 len < 40 ? 40 - len : 1, " ",
3001 S_ISBLK(file_inode(file)->i_mode) ?
3002 "partition" : "file\t",
3003 bytes, bytes < 10000000 ? "\t" : "",
3004 inuse, inuse < 10000000 ? "\t" : "",
3005 si->prio);
3006 return 0;
3007}
3008
3009static const struct seq_operations swaps_op = {
3010 .start = swap_start,
3011 .next = swap_next,
3012 .stop = swap_stop,
3013 .show = swap_show
3014};
3015
3016static int swaps_open(struct inode *inode, struct file *file)
3017{
3018 struct seq_file *seq;
3019 int ret;
3020
3021 ret = seq_open(file, &swaps_op);
3022 if (ret)
3023 return ret;
3024
3025 seq = file->private_data;
3026 seq->poll_event = atomic_read(&proc_poll_event);
3027 return 0;
3028}
3029
3030static const struct proc_ops swaps_proc_ops = {
3031 .proc_flags = PROC_ENTRY_PERMANENT,
3032 .proc_open = swaps_open,
3033 .proc_read = seq_read,
3034 .proc_lseek = seq_lseek,
3035 .proc_release = seq_release,
3036 .proc_poll = swaps_poll,
3037};
3038
3039static int __init procswaps_init(void)
3040{
3041 proc_create("swaps", 0, NULL, &swaps_proc_ops);
3042 return 0;
3043}
3044__initcall(procswaps_init);
3045#endif /* CONFIG_PROC_FS */
3046
3047#ifdef MAX_SWAPFILES_CHECK
3048static int __init max_swapfiles_check(void)
3049{
3050 MAX_SWAPFILES_CHECK();
3051 return 0;
3052}
3053late_initcall(max_swapfiles_check);
3054#endif
3055
3056static struct swap_info_struct *alloc_swap_info(void)
3057{
3058 struct swap_info_struct *p;
3059 struct swap_info_struct *defer = NULL;
3060 unsigned int type;
3061 int i;
3062
3063 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
3064 if (!p)
3065 return ERR_PTR(-ENOMEM);
3066
3067 if (percpu_ref_init(&p->users, swap_users_ref_free,
3068 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
3069 kvfree(p);
3070 return ERR_PTR(-ENOMEM);
3071 }
3072
3073 spin_lock(&swap_lock);
3074 for (type = 0; type < nr_swapfiles; type++) {
3075 if (!(swap_info[type]->flags & SWP_USED))
3076 break;
3077 }
3078 if (type >= MAX_SWAPFILES) {
3079 spin_unlock(&swap_lock);
3080 percpu_ref_exit(&p->users);
3081 kvfree(p);
3082 return ERR_PTR(-EPERM);
3083 }
3084 if (type >= nr_swapfiles) {
3085 p->type = type;
3086 /*
3087 * Publish the swap_info_struct after initializing it.
3088 * Note that kvzalloc() above zeroes all its fields.
3089 */
3090 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
3091 nr_swapfiles++;
3092 } else {
3093 defer = p;
3094 p = swap_info[type];
3095 /*
3096 * Do not memset this entry: a racing procfs swap_next()
3097 * would be relying on p->type to remain valid.
3098 */
3099 }
3100 p->swap_extent_root = RB_ROOT;
3101 plist_node_init(&p->list, 0);
3102 for_each_node(i)
3103 plist_node_init(&p->avail_lists[i], 0);
3104 p->flags = SWP_USED;
3105 spin_unlock(&swap_lock);
3106 if (defer) {
3107 percpu_ref_exit(&defer->users);
3108 kvfree(defer);
3109 }
3110 spin_lock_init(&p->lock);
3111 spin_lock_init(&p->cont_lock);
3112 init_completion(&p->comp);
3113
3114 return p;
3115}
3116
3117static int claim_swapfile(struct swap_info_struct *si, struct inode *inode)
3118{
3119 if (S_ISBLK(inode->i_mode)) {
3120 si->bdev = I_BDEV(inode);
3121 /*
3122 * Zoned block devices contain zones that have a sequential
3123 * write only restriction. Hence zoned block devices are not
3124 * suitable for swapping. Disallow them here.
3125 */
3126 if (bdev_is_zoned(si->bdev))
3127 return -EINVAL;
3128 si->flags |= SWP_BLKDEV;
3129 } else if (S_ISREG(inode->i_mode)) {
3130 si->bdev = inode->i_sb->s_bdev;
3131 }
3132
3133 return 0;
3134}
3135
3136
3137/*
3138 * Find out how many pages are allowed for a single swap device. There
3139 * are two limiting factors:
3140 * 1) the number of bits for the swap offset in the swp_entry_t type, and
3141 * 2) the number of bits in the swap pte, as defined by the different
3142 * architectures.
3143 *
3144 * In order to find the largest possible bit mask, a swap entry with
3145 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
3146 * decoded to a swp_entry_t again, and finally the swap offset is
3147 * extracted.
3148 *
3149 * This will mask all the bits from the initial ~0UL mask that can't
3150 * be encoded in either the swp_entry_t or the architecture definition
3151 * of a swap pte.
3152 */
3153unsigned long generic_max_swapfile_size(void)
3154{
3155 return swp_offset(pte_to_swp_entry(
3156 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
3157}
3158
3159/* Can be overridden by an architecture for additional checks. */
3160__weak unsigned long arch_max_swapfile_size(void)
3161{
3162 return generic_max_swapfile_size();
3163}
3164
3165static unsigned long read_swap_header(struct swap_info_struct *si,
3166 union swap_header *swap_header,
3167 struct inode *inode)
3168{
3169 int i;
3170 unsigned long maxpages;
3171 unsigned long swapfilepages;
3172 unsigned long last_page;
3173
3174 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
3175 pr_err("Unable to find swap-space signature\n");
3176 return 0;
3177 }
3178
3179 /* swap partition endianness hack... */
3180 if (swab32(swap_header->info.version) == 1) {
3181 swab32s(&swap_header->info.version);
3182 swab32s(&swap_header->info.last_page);
3183 swab32s(&swap_header->info.nr_badpages);
3184 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3185 return 0;
3186 for (i = 0; i < swap_header->info.nr_badpages; i++)
3187 swab32s(&swap_header->info.badpages[i]);
3188 }
3189 /* Check the swap header's sub-version */
3190 if (swap_header->info.version != 1) {
3191 pr_warn("Unable to handle swap header version %d\n",
3192 swap_header->info.version);
3193 return 0;
3194 }
3195
3196 si->lowest_bit = 1;
3197 si->cluster_next = 1;
3198 si->cluster_nr = 0;
3199
3200 maxpages = swapfile_maximum_size;
3201 last_page = swap_header->info.last_page;
3202 if (!last_page) {
3203 pr_warn("Empty swap-file\n");
3204 return 0;
3205 }
3206 if (last_page > maxpages) {
3207 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3208 K(maxpages), K(last_page));
3209 }
3210 if (maxpages > last_page) {
3211 maxpages = last_page + 1;
3212 /* p->max is an unsigned int: don't overflow it */
3213 if ((unsigned int)maxpages == 0)
3214 maxpages = UINT_MAX;
3215 }
3216 si->highest_bit = maxpages - 1;
3217
3218 if (!maxpages)
3219 return 0;
3220 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3221 if (swapfilepages && maxpages > swapfilepages) {
3222 pr_warn("Swap area shorter than signature indicates\n");
3223 return 0;
3224 }
3225 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3226 return 0;
3227 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3228 return 0;
3229
3230 return maxpages;
3231}
3232
3233#define SWAP_CLUSTER_INFO_COLS \
3234 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3235#define SWAP_CLUSTER_SPACE_COLS \
3236 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3237#define SWAP_CLUSTER_COLS \
3238 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3239
3240static int setup_swap_map_and_extents(struct swap_info_struct *si,
3241 union swap_header *swap_header,
3242 unsigned char *swap_map,
3243 unsigned long maxpages,
3244 sector_t *span)
3245{
3246 unsigned int nr_good_pages;
3247 unsigned long i;
3248 int nr_extents;
3249
3250 nr_good_pages = maxpages - 1; /* omit header page */
3251
3252 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3253 unsigned int page_nr = swap_header->info.badpages[i];
3254 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3255 return -EINVAL;
3256 if (page_nr < maxpages) {
3257 swap_map[page_nr] = SWAP_MAP_BAD;
3258 nr_good_pages--;
3259 }
3260 }
3261
3262 if (nr_good_pages) {
3263 swap_map[0] = SWAP_MAP_BAD;
3264 si->max = maxpages;
3265 si->pages = nr_good_pages;
3266 nr_extents = setup_swap_extents(si, span);
3267 if (nr_extents < 0)
3268 return nr_extents;
3269 nr_good_pages = si->pages;
3270 }
3271 if (!nr_good_pages) {
3272 pr_warn("Empty swap-file\n");
3273 return -EINVAL;
3274 }
3275
3276 return nr_extents;
3277}
3278
3279static struct swap_cluster_info *setup_clusters(struct swap_info_struct *si,
3280 union swap_header *swap_header,
3281 unsigned long maxpages)
3282{
3283 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3284 unsigned long col = si->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3285 struct swap_cluster_info *cluster_info;
3286 unsigned long i, j, k, idx;
3287 int cpu, err = -ENOMEM;
3288
3289 cluster_info = kvcalloc(nr_clusters, sizeof(*cluster_info), GFP_KERNEL);
3290 if (!cluster_info)
3291 goto err;
3292
3293 for (i = 0; i < nr_clusters; i++)
3294 spin_lock_init(&cluster_info[i].lock);
3295
3296 si->cluster_next_cpu = alloc_percpu(unsigned int);
3297 if (!si->cluster_next_cpu)
3298 goto err_free;
3299
3300 /* Random start position to help with wear leveling */
3301 for_each_possible_cpu(cpu)
3302 per_cpu(*si->cluster_next_cpu, cpu) =
3303 get_random_u32_inclusive(1, si->highest_bit);
3304
3305 si->percpu_cluster = alloc_percpu(struct percpu_cluster);
3306 if (!si->percpu_cluster)
3307 goto err_free;
3308
3309 for_each_possible_cpu(cpu) {
3310 struct percpu_cluster *cluster;
3311
3312 cluster = per_cpu_ptr(si->percpu_cluster, cpu);
3313 for (i = 0; i < SWAP_NR_ORDERS; i++)
3314 cluster->next[i] = SWAP_NEXT_INVALID;
3315 }
3316
3317 /*
3318 * Mark unusable pages as unavailable. The clusters aren't
3319 * marked free yet, so no list operations are involved yet.
3320 *
3321 * See setup_swap_map_and_extents(): header page, bad pages,
3322 * and the EOF part of the last cluster.
3323 */
3324 inc_cluster_info_page(si, cluster_info, 0);
3325 for (i = 0; i < swap_header->info.nr_badpages; i++)
3326 inc_cluster_info_page(si, cluster_info,
3327 swap_header->info.badpages[i]);
3328 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3329 inc_cluster_info_page(si, cluster_info, i);
3330
3331 INIT_LIST_HEAD(&si->free_clusters);
3332 INIT_LIST_HEAD(&si->full_clusters);
3333 INIT_LIST_HEAD(&si->discard_clusters);
3334
3335 for (i = 0; i < SWAP_NR_ORDERS; i++) {
3336 INIT_LIST_HEAD(&si->nonfull_clusters[i]);
3337 INIT_LIST_HEAD(&si->frag_clusters[i]);
3338 si->frag_cluster_nr[i] = 0;
3339 }
3340
3341 /*
3342 * Reduce false cache line sharing between cluster_info and
3343 * sharing same address space.
3344 */
3345 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3346 j = (k + col) % SWAP_CLUSTER_COLS;
3347 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3348 struct swap_cluster_info *ci;
3349 idx = i * SWAP_CLUSTER_COLS + j;
3350 ci = cluster_info + idx;
3351 if (idx >= nr_clusters)
3352 continue;
3353 if (ci->count) {
3354 ci->flags = CLUSTER_FLAG_NONFULL;
3355 list_add_tail(&ci->list, &si->nonfull_clusters[0]);
3356 continue;
3357 }
3358 ci->flags = CLUSTER_FLAG_FREE;
3359 list_add_tail(&ci->list, &si->free_clusters);
3360 }
3361 }
3362
3363 return cluster_info;
3364
3365err_free:
3366 kvfree(cluster_info);
3367err:
3368 return ERR_PTR(err);
3369}
3370
3371SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3372{
3373 struct swap_info_struct *si;
3374 struct filename *name;
3375 struct file *swap_file = NULL;
3376 struct address_space *mapping;
3377 struct dentry *dentry;
3378 int prio;
3379 int error;
3380 union swap_header *swap_header;
3381 int nr_extents;
3382 sector_t span;
3383 unsigned long maxpages;
3384 unsigned char *swap_map = NULL;
3385 unsigned long *zeromap = NULL;
3386 struct swap_cluster_info *cluster_info = NULL;
3387 struct folio *folio = NULL;
3388 struct inode *inode = NULL;
3389 bool inced_nr_rotate_swap = false;
3390
3391 if (swap_flags & ~SWAP_FLAGS_VALID)
3392 return -EINVAL;
3393
3394 if (!capable(CAP_SYS_ADMIN))
3395 return -EPERM;
3396
3397 if (!swap_avail_heads)
3398 return -ENOMEM;
3399
3400 si = alloc_swap_info();
3401 if (IS_ERR(si))
3402 return PTR_ERR(si);
3403
3404 INIT_WORK(&si->discard_work, swap_discard_work);
3405 INIT_WORK(&si->reclaim_work, swap_reclaim_work);
3406
3407 name = getname(specialfile);
3408 if (IS_ERR(name)) {
3409 error = PTR_ERR(name);
3410 name = NULL;
3411 goto bad_swap;
3412 }
3413 swap_file = file_open_name(name, O_RDWR | O_LARGEFILE | O_EXCL, 0);
3414 if (IS_ERR(swap_file)) {
3415 error = PTR_ERR(swap_file);
3416 swap_file = NULL;
3417 goto bad_swap;
3418 }
3419
3420 si->swap_file = swap_file;
3421 mapping = swap_file->f_mapping;
3422 dentry = swap_file->f_path.dentry;
3423 inode = mapping->host;
3424
3425 error = claim_swapfile(si, inode);
3426 if (unlikely(error))
3427 goto bad_swap;
3428
3429 inode_lock(inode);
3430 if (d_unlinked(dentry) || cant_mount(dentry)) {
3431 error = -ENOENT;
3432 goto bad_swap_unlock_inode;
3433 }
3434 if (IS_SWAPFILE(inode)) {
3435 error = -EBUSY;
3436 goto bad_swap_unlock_inode;
3437 }
3438
3439 /*
3440 * Read the swap header.
3441 */
3442 if (!mapping->a_ops->read_folio) {
3443 error = -EINVAL;
3444 goto bad_swap_unlock_inode;
3445 }
3446 folio = read_mapping_folio(mapping, 0, swap_file);
3447 if (IS_ERR(folio)) {
3448 error = PTR_ERR(folio);
3449 goto bad_swap_unlock_inode;
3450 }
3451 swap_header = kmap_local_folio(folio, 0);
3452
3453 maxpages = read_swap_header(si, swap_header, inode);
3454 if (unlikely(!maxpages)) {
3455 error = -EINVAL;
3456 goto bad_swap_unlock_inode;
3457 }
3458
3459 /* OK, set up the swap map and apply the bad block list */
3460 swap_map = vzalloc(maxpages);
3461 if (!swap_map) {
3462 error = -ENOMEM;
3463 goto bad_swap_unlock_inode;
3464 }
3465
3466 error = swap_cgroup_swapon(si->type, maxpages);
3467 if (error)
3468 goto bad_swap_unlock_inode;
3469
3470 nr_extents = setup_swap_map_and_extents(si, swap_header, swap_map,
3471 maxpages, &span);
3472 if (unlikely(nr_extents < 0)) {
3473 error = nr_extents;
3474 goto bad_swap_unlock_inode;
3475 }
3476
3477 /*
3478 * Use kvmalloc_array instead of bitmap_zalloc as the allocation order might
3479 * be above MAX_PAGE_ORDER incase of a large swap file.
3480 */
3481 zeromap = kvmalloc_array(BITS_TO_LONGS(maxpages), sizeof(long),
3482 GFP_KERNEL | __GFP_ZERO);
3483 if (!zeromap) {
3484 error = -ENOMEM;
3485 goto bad_swap_unlock_inode;
3486 }
3487
3488 if (si->bdev && bdev_stable_writes(si->bdev))
3489 si->flags |= SWP_STABLE_WRITES;
3490
3491 if (si->bdev && bdev_synchronous(si->bdev))
3492 si->flags |= SWP_SYNCHRONOUS_IO;
3493
3494 if (si->bdev && bdev_nonrot(si->bdev)) {
3495 si->flags |= SWP_SOLIDSTATE;
3496
3497 cluster_info = setup_clusters(si, swap_header, maxpages);
3498 if (IS_ERR(cluster_info)) {
3499 error = PTR_ERR(cluster_info);
3500 cluster_info = NULL;
3501 goto bad_swap_unlock_inode;
3502 }
3503 } else {
3504 atomic_inc(&nr_rotate_swap);
3505 inced_nr_rotate_swap = true;
3506 }
3507
3508 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3509 si->bdev && bdev_max_discard_sectors(si->bdev)) {
3510 /*
3511 * When discard is enabled for swap with no particular
3512 * policy flagged, we set all swap discard flags here in
3513 * order to sustain backward compatibility with older
3514 * swapon(8) releases.
3515 */
3516 si->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3517 SWP_PAGE_DISCARD);
3518
3519 /*
3520 * By flagging sys_swapon, a sysadmin can tell us to
3521 * either do single-time area discards only, or to just
3522 * perform discards for released swap page-clusters.
3523 * Now it's time to adjust the p->flags accordingly.
3524 */
3525 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3526 si->flags &= ~SWP_PAGE_DISCARD;
3527 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3528 si->flags &= ~SWP_AREA_DISCARD;
3529
3530 /* issue a swapon-time discard if it's still required */
3531 if (si->flags & SWP_AREA_DISCARD) {
3532 int err = discard_swap(si);
3533 if (unlikely(err))
3534 pr_err("swapon: discard_swap(%p): %d\n",
3535 si, err);
3536 }
3537 }
3538
3539 error = init_swap_address_space(si->type, maxpages);
3540 if (error)
3541 goto bad_swap_unlock_inode;
3542
3543 error = zswap_swapon(si->type, maxpages);
3544 if (error)
3545 goto free_swap_address_space;
3546
3547 /*
3548 * Flush any pending IO and dirty mappings before we start using this
3549 * swap device.
3550 */
3551 inode->i_flags |= S_SWAPFILE;
3552 error = inode_drain_writes(inode);
3553 if (error) {
3554 inode->i_flags &= ~S_SWAPFILE;
3555 goto free_swap_zswap;
3556 }
3557
3558 mutex_lock(&swapon_mutex);
3559 prio = -1;
3560 if (swap_flags & SWAP_FLAG_PREFER)
3561 prio =
3562 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3563 enable_swap_info(si, prio, swap_map, cluster_info, zeromap);
3564
3565 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n",
3566 K(si->pages), name->name, si->prio, nr_extents,
3567 K((unsigned long long)span),
3568 (si->flags & SWP_SOLIDSTATE) ? "SS" : "",
3569 (si->flags & SWP_DISCARDABLE) ? "D" : "",
3570 (si->flags & SWP_AREA_DISCARD) ? "s" : "",
3571 (si->flags & SWP_PAGE_DISCARD) ? "c" : "");
3572
3573 mutex_unlock(&swapon_mutex);
3574 atomic_inc(&proc_poll_event);
3575 wake_up_interruptible(&proc_poll_wait);
3576
3577 error = 0;
3578 goto out;
3579free_swap_zswap:
3580 zswap_swapoff(si->type);
3581free_swap_address_space:
3582 exit_swap_address_space(si->type);
3583bad_swap_unlock_inode:
3584 inode_unlock(inode);
3585bad_swap:
3586 free_percpu(si->percpu_cluster);
3587 si->percpu_cluster = NULL;
3588 free_percpu(si->cluster_next_cpu);
3589 si->cluster_next_cpu = NULL;
3590 inode = NULL;
3591 destroy_swap_extents(si);
3592 swap_cgroup_swapoff(si->type);
3593 spin_lock(&swap_lock);
3594 si->swap_file = NULL;
3595 si->flags = 0;
3596 spin_unlock(&swap_lock);
3597 vfree(swap_map);
3598 kvfree(zeromap);
3599 kvfree(cluster_info);
3600 if (inced_nr_rotate_swap)
3601 atomic_dec(&nr_rotate_swap);
3602 if (swap_file)
3603 filp_close(swap_file, NULL);
3604out:
3605 if (!IS_ERR_OR_NULL(folio))
3606 folio_release_kmap(folio, swap_header);
3607 if (name)
3608 putname(name);
3609 if (inode)
3610 inode_unlock(inode);
3611 if (!error)
3612 enable_swap_slots_cache();
3613 return error;
3614}
3615
3616void si_swapinfo(struct sysinfo *val)
3617{
3618 unsigned int type;
3619 unsigned long nr_to_be_unused = 0;
3620
3621 spin_lock(&swap_lock);
3622 for (type = 0; type < nr_swapfiles; type++) {
3623 struct swap_info_struct *si = swap_info[type];
3624
3625 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3626 nr_to_be_unused += READ_ONCE(si->inuse_pages);
3627 }
3628 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3629 val->totalswap = total_swap_pages + nr_to_be_unused;
3630 spin_unlock(&swap_lock);
3631}
3632
3633/*
3634 * Verify that nr swap entries are valid and increment their swap map counts.
3635 *
3636 * Returns error code in following case.
3637 * - success -> 0
3638 * - swp_entry is invalid -> EINVAL
3639 * - swp_entry is migration entry -> EINVAL
3640 * - swap-cache reference is requested but there is already one. -> EEXIST
3641 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3642 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3643 */
3644static int __swap_duplicate(swp_entry_t entry, unsigned char usage, int nr)
3645{
3646 struct swap_info_struct *si;
3647 struct swap_cluster_info *ci;
3648 unsigned long offset;
3649 unsigned char count;
3650 unsigned char has_cache;
3651 int err, i;
3652
3653 si = swp_swap_info(entry);
3654
3655 offset = swp_offset(entry);
3656 VM_WARN_ON(nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER);
3657 VM_WARN_ON(usage == 1 && nr > 1);
3658 ci = lock_cluster_or_swap_info(si, offset);
3659
3660 err = 0;
3661 for (i = 0; i < nr; i++) {
3662 count = si->swap_map[offset + i];
3663
3664 /*
3665 * swapin_readahead() doesn't check if a swap entry is valid, so the
3666 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3667 */
3668 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3669 err = -ENOENT;
3670 goto unlock_out;
3671 }
3672
3673 has_cache = count & SWAP_HAS_CACHE;
3674 count &= ~SWAP_HAS_CACHE;
3675
3676 if (!count && !has_cache) {
3677 err = -ENOENT;
3678 } else if (usage == SWAP_HAS_CACHE) {
3679 if (has_cache)
3680 err = -EEXIST;
3681 } else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX) {
3682 err = -EINVAL;
3683 }
3684
3685 if (err)
3686 goto unlock_out;
3687 }
3688
3689 for (i = 0; i < nr; i++) {
3690 count = si->swap_map[offset + i];
3691 has_cache = count & SWAP_HAS_CACHE;
3692 count &= ~SWAP_HAS_CACHE;
3693
3694 if (usage == SWAP_HAS_CACHE)
3695 has_cache = SWAP_HAS_CACHE;
3696 else if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3697 count += usage;
3698 else if (swap_count_continued(si, offset + i, count))
3699 count = COUNT_CONTINUED;
3700 else {
3701 /*
3702 * Don't need to rollback changes, because if
3703 * usage == 1, there must be nr == 1.
3704 */
3705 err = -ENOMEM;
3706 goto unlock_out;
3707 }
3708
3709 WRITE_ONCE(si->swap_map[offset + i], count | has_cache);
3710 }
3711
3712unlock_out:
3713 unlock_cluster_or_swap_info(si, ci);
3714 return err;
3715}
3716
3717/*
3718 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3719 * (in which case its reference count is never incremented).
3720 */
3721void swap_shmem_alloc(swp_entry_t entry, int nr)
3722{
3723 __swap_duplicate(entry, SWAP_MAP_SHMEM, nr);
3724}
3725
3726/*
3727 * Increase reference count of swap entry by 1.
3728 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3729 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3730 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3731 * might occur if a page table entry has got corrupted.
3732 */
3733int swap_duplicate(swp_entry_t entry)
3734{
3735 int err = 0;
3736
3737 while (!err && __swap_duplicate(entry, 1, 1) == -ENOMEM)
3738 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3739 return err;
3740}
3741
3742/*
3743 * @entry: first swap entry from which we allocate nr swap cache.
3744 *
3745 * Called when allocating swap cache for existing swap entries,
3746 * This can return error codes. Returns 0 at success.
3747 * -EEXIST means there is a swap cache.
3748 * Note: return code is different from swap_duplicate().
3749 */
3750int swapcache_prepare(swp_entry_t entry, int nr)
3751{
3752 return __swap_duplicate(entry, SWAP_HAS_CACHE, nr);
3753}
3754
3755void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry, int nr)
3756{
3757 unsigned long offset = swp_offset(entry);
3758
3759 cluster_swap_free_nr(si, offset, nr, SWAP_HAS_CACHE);
3760}
3761
3762struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3763{
3764 return swap_type_to_swap_info(swp_type(entry));
3765}
3766
3767/*
3768 * out-of-line methods to avoid include hell.
3769 */
3770struct address_space *swapcache_mapping(struct folio *folio)
3771{
3772 return swp_swap_info(folio->swap)->swap_file->f_mapping;
3773}
3774EXPORT_SYMBOL_GPL(swapcache_mapping);
3775
3776pgoff_t __folio_swap_cache_index(struct folio *folio)
3777{
3778 return swap_cache_index(folio->swap);
3779}
3780EXPORT_SYMBOL_GPL(__folio_swap_cache_index);
3781
3782/*
3783 * add_swap_count_continuation - called when a swap count is duplicated
3784 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3785 * page of the original vmalloc'ed swap_map, to hold the continuation count
3786 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3787 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3788 *
3789 * These continuation pages are seldom referenced: the common paths all work
3790 * on the original swap_map, only referring to a continuation page when the
3791 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3792 *
3793 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3794 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3795 * can be called after dropping locks.
3796 */
3797int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3798{
3799 struct swap_info_struct *si;
3800 struct swap_cluster_info *ci;
3801 struct page *head;
3802 struct page *page;
3803 struct page *list_page;
3804 pgoff_t offset;
3805 unsigned char count;
3806 int ret = 0;
3807
3808 /*
3809 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3810 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3811 */
3812 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3813
3814 si = get_swap_device(entry);
3815 if (!si) {
3816 /*
3817 * An acceptable race has occurred since the failing
3818 * __swap_duplicate(): the swap device may be swapoff
3819 */
3820 goto outer;
3821 }
3822 spin_lock(&si->lock);
3823
3824 offset = swp_offset(entry);
3825
3826 ci = lock_cluster(si, offset);
3827
3828 count = swap_count(si->swap_map[offset]);
3829
3830 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3831 /*
3832 * The higher the swap count, the more likely it is that tasks
3833 * will race to add swap count continuation: we need to avoid
3834 * over-provisioning.
3835 */
3836 goto out;
3837 }
3838
3839 if (!page) {
3840 ret = -ENOMEM;
3841 goto out;
3842 }
3843
3844 head = vmalloc_to_page(si->swap_map + offset);
3845 offset &= ~PAGE_MASK;
3846
3847 spin_lock(&si->cont_lock);
3848 /*
3849 * Page allocation does not initialize the page's lru field,
3850 * but it does always reset its private field.
3851 */
3852 if (!page_private(head)) {
3853 BUG_ON(count & COUNT_CONTINUED);
3854 INIT_LIST_HEAD(&head->lru);
3855 set_page_private(head, SWP_CONTINUED);
3856 si->flags |= SWP_CONTINUED;
3857 }
3858
3859 list_for_each_entry(list_page, &head->lru, lru) {
3860 unsigned char *map;
3861
3862 /*
3863 * If the previous map said no continuation, but we've found
3864 * a continuation page, free our allocation and use this one.
3865 */
3866 if (!(count & COUNT_CONTINUED))
3867 goto out_unlock_cont;
3868
3869 map = kmap_local_page(list_page) + offset;
3870 count = *map;
3871 kunmap_local(map);
3872
3873 /*
3874 * If this continuation count now has some space in it,
3875 * free our allocation and use this one.
3876 */
3877 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3878 goto out_unlock_cont;
3879 }
3880
3881 list_add_tail(&page->lru, &head->lru);
3882 page = NULL; /* now it's attached, don't free it */
3883out_unlock_cont:
3884 spin_unlock(&si->cont_lock);
3885out:
3886 unlock_cluster(ci);
3887 spin_unlock(&si->lock);
3888 put_swap_device(si);
3889outer:
3890 if (page)
3891 __free_page(page);
3892 return ret;
3893}
3894
3895/*
3896 * swap_count_continued - when the original swap_map count is incremented
3897 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3898 * into, carry if so, or else fail until a new continuation page is allocated;
3899 * when the original swap_map count is decremented from 0 with continuation,
3900 * borrow from the continuation and report whether it still holds more.
3901 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3902 * lock.
3903 */
3904static bool swap_count_continued(struct swap_info_struct *si,
3905 pgoff_t offset, unsigned char count)
3906{
3907 struct page *head;
3908 struct page *page;
3909 unsigned char *map;
3910 bool ret;
3911
3912 head = vmalloc_to_page(si->swap_map + offset);
3913 if (page_private(head) != SWP_CONTINUED) {
3914 BUG_ON(count & COUNT_CONTINUED);
3915 return false; /* need to add count continuation */
3916 }
3917
3918 spin_lock(&si->cont_lock);
3919 offset &= ~PAGE_MASK;
3920 page = list_next_entry(head, lru);
3921 map = kmap_local_page(page) + offset;
3922
3923 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3924 goto init_map; /* jump over SWAP_CONT_MAX checks */
3925
3926 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3927 /*
3928 * Think of how you add 1 to 999
3929 */
3930 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3931 kunmap_local(map);
3932 page = list_next_entry(page, lru);
3933 BUG_ON(page == head);
3934 map = kmap_local_page(page) + offset;
3935 }
3936 if (*map == SWAP_CONT_MAX) {
3937 kunmap_local(map);
3938 page = list_next_entry(page, lru);
3939 if (page == head) {
3940 ret = false; /* add count continuation */
3941 goto out;
3942 }
3943 map = kmap_local_page(page) + offset;
3944init_map: *map = 0; /* we didn't zero the page */
3945 }
3946 *map += 1;
3947 kunmap_local(map);
3948 while ((page = list_prev_entry(page, lru)) != head) {
3949 map = kmap_local_page(page) + offset;
3950 *map = COUNT_CONTINUED;
3951 kunmap_local(map);
3952 }
3953 ret = true; /* incremented */
3954
3955 } else { /* decrementing */
3956 /*
3957 * Think of how you subtract 1 from 1000
3958 */
3959 BUG_ON(count != COUNT_CONTINUED);
3960 while (*map == COUNT_CONTINUED) {
3961 kunmap_local(map);
3962 page = list_next_entry(page, lru);
3963 BUG_ON(page == head);
3964 map = kmap_local_page(page) + offset;
3965 }
3966 BUG_ON(*map == 0);
3967 *map -= 1;
3968 if (*map == 0)
3969 count = 0;
3970 kunmap_local(map);
3971 while ((page = list_prev_entry(page, lru)) != head) {
3972 map = kmap_local_page(page) + offset;
3973 *map = SWAP_CONT_MAX | count;
3974 count = COUNT_CONTINUED;
3975 kunmap_local(map);
3976 }
3977 ret = count == COUNT_CONTINUED;
3978 }
3979out:
3980 spin_unlock(&si->cont_lock);
3981 return ret;
3982}
3983
3984/*
3985 * free_swap_count_continuations - swapoff free all the continuation pages
3986 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3987 */
3988static void free_swap_count_continuations(struct swap_info_struct *si)
3989{
3990 pgoff_t offset;
3991
3992 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3993 struct page *head;
3994 head = vmalloc_to_page(si->swap_map + offset);
3995 if (page_private(head)) {
3996 struct page *page, *next;
3997
3998 list_for_each_entry_safe(page, next, &head->lru, lru) {
3999 list_del(&page->lru);
4000 __free_page(page);
4001 }
4002 }
4003 }
4004}
4005
4006#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
4007void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp)
4008{
4009 struct swap_info_struct *si, *next;
4010 int nid = folio_nid(folio);
4011
4012 if (!(gfp & __GFP_IO))
4013 return;
4014
4015 if (!__has_usable_swap())
4016 return;
4017
4018 if (!blk_cgroup_congested())
4019 return;
4020
4021 /*
4022 * We've already scheduled a throttle, avoid taking the global swap
4023 * lock.
4024 */
4025 if (current->throttle_disk)
4026 return;
4027
4028 spin_lock(&swap_avail_lock);
4029 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
4030 avail_lists[nid]) {
4031 if (si->bdev) {
4032 blkcg_schedule_throttle(si->bdev->bd_disk, true);
4033 break;
4034 }
4035 }
4036 spin_unlock(&swap_avail_lock);
4037}
4038#endif
4039
4040static int __init swapfile_init(void)
4041{
4042 int nid;
4043
4044 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
4045 GFP_KERNEL);
4046 if (!swap_avail_heads) {
4047 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
4048 return -ENOMEM;
4049 }
4050
4051 for_each_node(nid)
4052 plist_head_init(&swap_avail_heads[nid]);
4053
4054 swapfile_maximum_size = arch_max_swapfile_size();
4055
4056#ifdef CONFIG_MIGRATION
4057 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
4058 swap_migration_ad_supported = true;
4059#endif /* CONFIG_MIGRATION */
4060
4061 return 0;
4062}
4063subsys_initcall(swapfile_init);