swap_state.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *  linux/mm/swap_state.c
  4 *
  5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  6 *  Swap reorganised 29.12.95, Stephen Tweedie
  7 *
  8 *  Rewritten to use page cache, (C) 1998 Stephen Tweedie
  9 */
 10#include <linux/mm.h>
 11#include <linux/gfp.h>
 12#include <linux/kernel_stat.h>
 13#include <linux/mempolicy.h>
 14#include <linux/swap.h>
 15#include <linux/leafops.h>
 16#include <linux/init.h>
 17#include <linux/pagemap.h>
 18#include <linux/pagevec.h>
 19#include <linux/backing-dev.h>
 20#include <linux/blkdev.h>
 21#include <linux/migrate.h>
 22#include <linux/vmalloc.h>
 23#include <linux/huge_mm.h>
 24#include <linux/shmem_fs.h>
 25#include "internal.h"
 26#include "swap_table.h"
 27#include "swap.h"
 28
 29/*
 30 * swapper_space is a fiction, retained to simplify the path through
 31 * vmscan's shrink_folio_list.
 32 */
 33static const struct address_space_operations swap_aops = {
 34	.dirty_folio	= noop_dirty_folio,
 35#ifdef CONFIG_MIGRATION
 36	.migrate_folio	= migrate_folio,
 37#endif
 38};
 39
 40struct address_space swap_space __read_mostly = {
 41	.a_ops = &swap_aops,
 42};
 43
 44static bool enable_vma_readahead __read_mostly = true;
 45
 46#define SWAP_RA_ORDER_CEILING	5
 47
 48#define SWAP_RA_WIN_SHIFT	(PAGE_SHIFT / 2)
 49#define SWAP_RA_HITS_MASK	((1UL << SWAP_RA_WIN_SHIFT) - 1)
 50#define SWAP_RA_HITS_MAX	SWAP_RA_HITS_MASK
 51#define SWAP_RA_WIN_MASK	(~PAGE_MASK & ~SWAP_RA_HITS_MASK)
 52
 53#define SWAP_RA_HITS(v)		((v) & SWAP_RA_HITS_MASK)
 54#define SWAP_RA_WIN(v)		(((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
 55#define SWAP_RA_ADDR(v)		((v) & PAGE_MASK)
 56
 57#define SWAP_RA_VAL(addr, win, hits)				\
 58	(((addr) & PAGE_MASK) |					\
 59	 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |	\
 60	 ((hits) & SWAP_RA_HITS_MASK))
 61
 62/* Initial readahead hits is 4 to start up with a small window */
 63#define GET_SWAP_RA_VAL(vma)					\
 64	(atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
 65
 66static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
 67
 68void show_swap_cache_info(void)
 69{
 70	printk("%lu pages in swap cache\n", total_swapcache_pages());
 71	printk("Free swap  = %ldkB\n", K(get_nr_swap_pages()));
 72	printk("Total swap = %lukB\n", K(total_swap_pages));
 73}
 74
 75/**
 76 * swap_cache_get_folio - Looks up a folio in the swap cache.
 77 * @entry: swap entry used for the lookup.
 78 *
 79 * A found folio will be returned unlocked and with its refcount increased.
 80 *
 81 * Context: Caller must ensure @entry is valid and protect the swap device
 82 * with reference count or locks.
 83 * Return: Returns the found folio on success, NULL otherwise. The caller
 84 * must lock nd check if the folio still matches the swap entry before
 85 * use (e.g., folio_matches_swap_entry).
 86 */
 87struct folio *swap_cache_get_folio(swp_entry_t entry)
 88{
 89	unsigned long swp_tb;
 90	struct folio *folio;
 91
 92	for (;;) {
 93		swp_tb = swap_table_get(__swap_entry_to_cluster(entry),
 94					swp_cluster_offset(entry));
 95		if (!swp_tb_is_folio(swp_tb))
 96			return NULL;
 97		folio = swp_tb_to_folio(swp_tb);
 98		if (likely(folio_try_get(folio)))
 99			return folio;
100	}
101
102	return NULL;
103}
104
105/**
106 * swap_cache_get_shadow - Looks up a shadow in the swap cache.
107 * @entry: swap entry used for the lookup.
108 *
109 * Context: Caller must ensure @entry is valid and protect the swap device
110 * with reference count or locks.
111 * Return: Returns either NULL or an XA_VALUE (shadow).
112 */
113void *swap_cache_get_shadow(swp_entry_t entry)
114{
115	unsigned long swp_tb;
116
117	swp_tb = swap_table_get(__swap_entry_to_cluster(entry),
118				swp_cluster_offset(entry));
119	if (swp_tb_is_shadow(swp_tb))
120		return swp_tb_to_shadow(swp_tb);
121	return NULL;
122}
123
124/**
125 * swap_cache_add_folio - Add a folio into the swap cache.
126 * @folio: The folio to be added.
127 * @entry: The swap entry corresponding to the folio.
128 * @gfp: gfp_mask for XArray node allocation.
129 * @shadowp: If a shadow is found, return the shadow.
130 *
131 * Context: Caller must ensure @entry is valid and protect the swap device
132 * with reference count or locks.
133 * The caller also needs to update the corresponding swap_map slots with
134 * SWAP_HAS_CACHE bit to avoid race or conflict.
135 */
136void swap_cache_add_folio(struct folio *folio, swp_entry_t entry, void **shadowp)
137{
138	void *shadow = NULL;
139	unsigned long old_tb, new_tb;
140	struct swap_cluster_info *ci;
141	unsigned int ci_start, ci_off, ci_end;
142	unsigned long nr_pages = folio_nr_pages(folio);
143
144	VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
145	VM_WARN_ON_ONCE_FOLIO(folio_test_swapcache(folio), folio);
146	VM_WARN_ON_ONCE_FOLIO(!folio_test_swapbacked(folio), folio);
147
148	new_tb = folio_to_swp_tb(folio);
149	ci_start = swp_cluster_offset(entry);
150	ci_end = ci_start + nr_pages;
151	ci_off = ci_start;
152	ci = swap_cluster_lock(__swap_entry_to_info(entry), swp_offset(entry));
153	do {
154		old_tb = __swap_table_xchg(ci, ci_off, new_tb);
155		WARN_ON_ONCE(swp_tb_is_folio(old_tb));
156		if (swp_tb_is_shadow(old_tb))
157			shadow = swp_tb_to_shadow(old_tb);
158	} while (++ci_off < ci_end);
159
160	folio_ref_add(folio, nr_pages);
161	folio_set_swapcache(folio);
162	folio->swap = entry;
163	swap_cluster_unlock(ci);
164
165	node_stat_mod_folio(folio, NR_FILE_PAGES, nr_pages);
166	lruvec_stat_mod_folio(folio, NR_SWAPCACHE, nr_pages);
167
168	if (shadowp)
169		*shadowp = shadow;
170}
171
172/**
173 * __swap_cache_del_folio - Removes a folio from the swap cache.
174 * @ci: The locked swap cluster.
175 * @folio: The folio.
176 * @entry: The first swap entry that the folio corresponds to.
177 * @shadow: shadow value to be filled in the swap cache.
178 *
179 * Removes a folio from the swap cache and fills a shadow in place.
180 * This won't put the folio's refcount. The caller has to do that.
181 *
182 * Context: Caller must ensure the folio is locked and in the swap cache
183 * using the index of @entry, and lock the cluster that holds the entries.
184 */
185void __swap_cache_del_folio(struct swap_cluster_info *ci, struct folio *folio,
186			    swp_entry_t entry, void *shadow)
187{
188	unsigned long old_tb, new_tb;
189	unsigned int ci_start, ci_off, ci_end;
190	unsigned long nr_pages = folio_nr_pages(folio);
191
192	VM_WARN_ON_ONCE(__swap_entry_to_cluster(entry) != ci);
193	VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
194	VM_WARN_ON_ONCE_FOLIO(!folio_test_swapcache(folio), folio);
195	VM_WARN_ON_ONCE_FOLIO(folio_test_writeback(folio), folio);
196
197	new_tb = shadow_swp_to_tb(shadow);
198	ci_start = swp_cluster_offset(entry);
199	ci_end = ci_start + nr_pages;
200	ci_off = ci_start;
201	do {
202		/* If shadow is NULL, we sets an empty shadow */
203		old_tb = __swap_table_xchg(ci, ci_off, new_tb);
204		WARN_ON_ONCE(!swp_tb_is_folio(old_tb) ||
205			     swp_tb_to_folio(old_tb) != folio);
206	} while (++ci_off < ci_end);
207
208	folio->swap.val = 0;
209	folio_clear_swapcache(folio);
210	node_stat_mod_folio(folio, NR_FILE_PAGES, -nr_pages);
211	lruvec_stat_mod_folio(folio, NR_SWAPCACHE, -nr_pages);
212}
213
214/**
215 * swap_cache_del_folio - Removes a folio from the swap cache.
216 * @folio: The folio.
217 *
218 * Same as __swap_cache_del_folio, but handles lock and refcount. The
219 * caller must ensure the folio is either clean or has a swap count
220 * equal to zero, or it may cause data loss.
221 *
222 * Context: Caller must ensure the folio is locked and in the swap cache.
223 */
224void swap_cache_del_folio(struct folio *folio)
225{
226	struct swap_cluster_info *ci;
227	swp_entry_t entry = folio->swap;
228
229	ci = swap_cluster_lock(__swap_entry_to_info(entry), swp_offset(entry));
230	__swap_cache_del_folio(ci, folio, entry, NULL);
231	swap_cluster_unlock(ci);
232
233	put_swap_folio(folio, entry);
234	folio_ref_sub(folio, folio_nr_pages(folio));
235}
236
237/**
238 * __swap_cache_replace_folio - Replace a folio in the swap cache.
239 * @ci: The locked swap cluster.
240 * @old: The old folio to be replaced.
241 * @new: The new folio.
242 *
243 * Replace an existing folio in the swap cache with a new folio. The
244 * caller is responsible for setting up the new folio's flag and swap
245 * entries. Replacement will take the new folio's swap entry value as
246 * the starting offset to override all slots covered by the new folio.
247 *
248 * Context: Caller must ensure both folios are locked, and lock the
249 * cluster that holds the old folio to be replaced.
250 */
251void __swap_cache_replace_folio(struct swap_cluster_info *ci,
252				struct folio *old, struct folio *new)
253{
254	swp_entry_t entry = new->swap;
255	unsigned long nr_pages = folio_nr_pages(new);
256	unsigned int ci_off = swp_cluster_offset(entry);
257	unsigned int ci_end = ci_off + nr_pages;
258	unsigned long old_tb, new_tb;
259
260	VM_WARN_ON_ONCE(!folio_test_swapcache(old) || !folio_test_swapcache(new));
261	VM_WARN_ON_ONCE(!folio_test_locked(old) || !folio_test_locked(new));
262	VM_WARN_ON_ONCE(!entry.val);
263
264	/* Swap cache still stores N entries instead of a high-order entry */
265	new_tb = folio_to_swp_tb(new);
266	do {
267		old_tb = __swap_table_xchg(ci, ci_off, new_tb);
268		WARN_ON_ONCE(!swp_tb_is_folio(old_tb) || swp_tb_to_folio(old_tb) != old);
269	} while (++ci_off < ci_end);
270
271	/*
272	 * If the old folio is partially replaced (e.g., splitting a large
273	 * folio, the old folio is shrunk, and new split sub folios replace
274	 * the shrunk part), ensure the new folio doesn't overlap it.
275	 */
276	if (IS_ENABLED(CONFIG_DEBUG_VM) &&
277	    folio_order(old) != folio_order(new)) {
278		ci_off = swp_cluster_offset(old->swap);
279		ci_end = ci_off + folio_nr_pages(old);
280		while (ci_off++ < ci_end)
281			WARN_ON_ONCE(swp_tb_to_folio(__swap_table_get(ci, ci_off)) != old);
282	}
283}
284
285/**
286 * swap_cache_clear_shadow - Clears a set of shadows in the swap cache.
287 * @entry: The starting index entry.
288 * @nr_ents: How many slots need to be cleared.
289 *
290 * Context: Caller must ensure the range is valid, all in one single cluster,
291 * not occupied by any folio, and lock the cluster.
292 */
293void __swap_cache_clear_shadow(swp_entry_t entry, int nr_ents)
294{
295	struct swap_cluster_info *ci = __swap_entry_to_cluster(entry);
296	unsigned int ci_off = swp_cluster_offset(entry), ci_end;
297	unsigned long old;
298
299	ci_end = ci_off + nr_ents;
300	do {
301		old = __swap_table_xchg(ci, ci_off, null_to_swp_tb());
302		WARN_ON_ONCE(swp_tb_is_folio(old));
303	} while (++ci_off < ci_end);
304}
305
306/*
307 * If we are the only user, then try to free up the swap cache.
308 *
309 * Its ok to check the swapcache flag without the folio lock
310 * here because we are going to recheck again inside
311 * folio_free_swap() _with_ the lock.
312 * 					- Marcelo
313 */
314void free_swap_cache(struct folio *folio)
315{
316	if (folio_test_swapcache(folio) && !folio_mapped(folio) &&
317	    folio_trylock(folio)) {
318		folio_free_swap(folio);
319		folio_unlock(folio);
320	}
321}
322
323/*
324 * Freeing a folio and also freeing any swap cache associated with
325 * this folio if it is the last user.
326 */
327void free_folio_and_swap_cache(struct folio *folio)
328{
329	free_swap_cache(folio);
330	if (!is_huge_zero_folio(folio))
331		folio_put(folio);
332}
333
334/*
335 * Passed an array of pages, drop them all from swapcache and then release
336 * them.  They are removed from the LRU and freed if this is their last use.
337 */
338void free_pages_and_swap_cache(struct encoded_page **pages, int nr)
339{
340	struct folio_batch folios;
341	unsigned int refs[PAGEVEC_SIZE];
342
343	folio_batch_init(&folios);
344	for (int i = 0; i < nr; i++) {
345		struct folio *folio = page_folio(encoded_page_ptr(pages[i]));
346
347		free_swap_cache(folio);
348		refs[folios.nr] = 1;
349		if (unlikely(encoded_page_flags(pages[i]) &
350			     ENCODED_PAGE_BIT_NR_PAGES_NEXT))
351			refs[folios.nr] = encoded_nr_pages(pages[++i]);
352
353		if (folio_batch_add(&folios, folio) == 0)
354			folios_put_refs(&folios, refs);
355	}
356	if (folios.nr)
357		folios_put_refs(&folios, refs);
358}
359
360static inline bool swap_use_vma_readahead(void)
361{
362	return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
363}
364
365/**
366 * swap_update_readahead - Update the readahead statistics of VMA or globally.
367 * @folio: the swap cache folio that just got hit.
368 * @vma: the VMA that should be updated, could be NULL for global update.
369 * @addr: the addr that triggered the swapin, ignored if @vma is NULL.
370 */
371void swap_update_readahead(struct folio *folio, struct vm_area_struct *vma,
372			   unsigned long addr)
373{
374	bool readahead, vma_ra = swap_use_vma_readahead();
375
376	/*
377	 * At the moment, we don't support PG_readahead for anon THP
378	 * so let's bail out rather than confusing the readahead stat.
379	 */
380	if (unlikely(folio_test_large(folio)))
381		return;
382
383	readahead = folio_test_clear_readahead(folio);
384	if (vma && vma_ra) {
385		unsigned long ra_val;
386		int win, hits;
387
388		ra_val = GET_SWAP_RA_VAL(vma);
389		win = SWAP_RA_WIN(ra_val);
390		hits = SWAP_RA_HITS(ra_val);
391		if (readahead)
392			hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
393		atomic_long_set(&vma->swap_readahead_info,
394				SWAP_RA_VAL(addr, win, hits));
395	}
396
397	if (readahead) {
398		count_vm_event(SWAP_RA_HIT);
399		if (!vma || !vma_ra)
400			atomic_inc(&swapin_readahead_hits);
401	}
402}
403
404struct folio *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
405		struct mempolicy *mpol, pgoff_t ilx, bool *new_page_allocated,
406		bool skip_if_exists)
407{
408	struct swap_info_struct *si = __swap_entry_to_info(entry);
409	struct folio *folio;
410	struct folio *new_folio = NULL;
411	struct folio *result = NULL;
412	void *shadow = NULL;
413
414	*new_page_allocated = false;
415	for (;;) {
416		int err;
417
418		/*
419		 * Check the swap cache first, if a cached folio is found,
420		 * return it unlocked. The caller will lock and check it.
421		 */
422		folio = swap_cache_get_folio(entry);
423		if (folio)
424			goto got_folio;
425
426		/*
427		 * Just skip read ahead for unused swap slot.
428		 */
429		if (!swap_entry_swapped(si, entry))
430			goto put_and_return;
431
432		/*
433		 * Get a new folio to read into from swap.  Allocate it now if
434		 * new_folio not exist, before marking swap_map SWAP_HAS_CACHE,
435		 * when -EEXIST will cause any racers to loop around until we
436		 * add it to cache.
437		 */
438		if (!new_folio) {
439			new_folio = folio_alloc_mpol(gfp_mask, 0, mpol, ilx, numa_node_id());
440			if (!new_folio)
441				goto put_and_return;
442		}
443
444		/*
445		 * Swap entry may have been freed since our caller observed it.
446		 */
447		err = swapcache_prepare(entry, 1);
448		if (!err)
449			break;
450		else if (err != -EEXIST)
451			goto put_and_return;
452
453		/*
454		 * Protect against a recursive call to __read_swap_cache_async()
455		 * on the same entry waiting forever here because SWAP_HAS_CACHE
456		 * is set but the folio is not the swap cache yet. This can
457		 * happen today if mem_cgroup_swapin_charge_folio() below
458		 * triggers reclaim through zswap, which may call
459		 * __read_swap_cache_async() in the writeback path.
460		 */
461		if (skip_if_exists)
462			goto put_and_return;
463
464		/*
465		 * We might race against __swap_cache_del_folio(), and
466		 * stumble across a swap_map entry whose SWAP_HAS_CACHE
467		 * has not yet been cleared.  Or race against another
468		 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
469		 * in swap_map, but not yet added its folio to swap cache.
470		 */
471		schedule_timeout_uninterruptible(1);
472	}
473
474	/*
475	 * The swap entry is ours to swap in. Prepare the new folio.
476	 */
477	__folio_set_locked(new_folio);
478	__folio_set_swapbacked(new_folio);
479
480	if (mem_cgroup_swapin_charge_folio(new_folio, NULL, gfp_mask, entry))
481		goto fail_unlock;
482
483	swap_cache_add_folio(new_folio, entry, &shadow);
484	memcg1_swapin(entry, 1);
485
486	if (shadow)
487		workingset_refault(new_folio, shadow);
488
489	/* Caller will initiate read into locked new_folio */
490	folio_add_lru(new_folio);
491	*new_page_allocated = true;
492	folio = new_folio;
493got_folio:
494	result = folio;
495	goto put_and_return;
496
497fail_unlock:
498	put_swap_folio(new_folio, entry);
499	folio_unlock(new_folio);
500put_and_return:
501	if (!(*new_page_allocated) && new_folio)
502		folio_put(new_folio);
503	return result;
504}
505
506/*
507 * Locate a page of swap in physical memory, reserving swap cache space
508 * and reading the disk if it is not already cached.
509 * A failure return means that either the page allocation failed or that
510 * the swap entry is no longer in use.
511 */
512struct folio *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
513		struct vm_area_struct *vma, unsigned long addr,
514		struct swap_iocb **plug)
515{
516	struct swap_info_struct *si;
517	bool page_allocated;
518	struct mempolicy *mpol;
519	pgoff_t ilx;
520	struct folio *folio;
521
522	si = get_swap_device(entry);
523	if (!si)
524		return NULL;
525
526	mpol = get_vma_policy(vma, addr, 0, &ilx);
527	folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
528					&page_allocated, false);
529	mpol_cond_put(mpol);
530
531	if (page_allocated)
532		swap_read_folio(folio, plug);
533
534	put_swap_device(si);
535	return folio;
536}
537
538static unsigned int __swapin_nr_pages(unsigned long prev_offset,
539				      unsigned long offset,
540				      int hits,
541				      int max_pages,
542				      int prev_win)
543{
544	unsigned int pages, last_ra;
545
546	/*
547	 * This heuristic has been found to work well on both sequential and
548	 * random loads, swapping to hard disk or to SSD: please don't ask
549	 * what the "+ 2" means, it just happens to work well, that's all.
550	 */
551	pages = hits + 2;
552	if (pages == 2) {
553		/*
554		 * We can have no readahead hits to judge by: but must not get
555		 * stuck here forever, so check for an adjacent offset instead
556		 * (and don't even bother to check whether swap type is same).
557		 */
558		if (offset != prev_offset + 1 && offset != prev_offset - 1)
559			pages = 1;
560	} else {
561		unsigned int roundup = 4;
562		while (roundup < pages)
563			roundup <<= 1;
564		pages = roundup;
565	}
566
567	if (pages > max_pages)
568		pages = max_pages;
569
570	/* Don't shrink readahead too fast */
571	last_ra = prev_win / 2;
572	if (pages < last_ra)
573		pages = last_ra;
574
575	return pages;
576}
577
578static unsigned long swapin_nr_pages(unsigned long offset)
579{
580	static unsigned long prev_offset;
581	unsigned int hits, pages, max_pages;
582	static atomic_t last_readahead_pages;
583
584	max_pages = 1 << READ_ONCE(page_cluster);
585	if (max_pages <= 1)
586		return 1;
587
588	hits = atomic_xchg(&swapin_readahead_hits, 0);
589	pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
590				  max_pages,
591				  atomic_read(&last_readahead_pages));
592	if (!hits)
593		WRITE_ONCE(prev_offset, offset);
594	atomic_set(&last_readahead_pages, pages);
595
596	return pages;
597}
598
599/**
600 * swap_cluster_readahead - swap in pages in hope we need them soon
601 * @entry: swap entry of this memory
602 * @gfp_mask: memory allocation flags
603 * @mpol: NUMA memory allocation policy to be applied
604 * @ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
605 *
606 * Returns the struct folio for entry and addr, after queueing swapin.
607 *
608 * Primitive swap readahead code. We simply read an aligned block of
609 * (1 << page_cluster) entries in the swap area. This method is chosen
610 * because it doesn't cost us any seek time.  We also make sure to queue
611 * the 'original' request together with the readahead ones...
612 *
613 * Note: it is intentional that the same NUMA policy and interleave index
614 * are used for every page of the readahead: neighbouring pages on swap
615 * are fairly likely to have been swapped out from the same node.
616 */
617struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
618				    struct mempolicy *mpol, pgoff_t ilx)
619{
620	struct folio *folio;
621	unsigned long entry_offset = swp_offset(entry);
622	unsigned long offset = entry_offset;
623	unsigned long start_offset, end_offset;
624	unsigned long mask;
625	struct swap_info_struct *si = __swap_entry_to_info(entry);
626	struct blk_plug plug;
627	struct swap_iocb *splug = NULL;
628	bool page_allocated;
629
630	mask = swapin_nr_pages(offset) - 1;
631	if (!mask)
632		goto skip;
633
634	/* Read a page_cluster sized and aligned cluster around offset. */
635	start_offset = offset & ~mask;
636	end_offset = offset | mask;
637	if (!start_offset)	/* First page is swap header. */
638		start_offset++;
639	if (end_offset >= si->max)
640		end_offset = si->max - 1;
641
642	blk_start_plug(&plug);
643	for (offset = start_offset; offset <= end_offset ; offset++) {
644		/* Ok, do the async read-ahead now */
645		folio = __read_swap_cache_async(
646				swp_entry(swp_type(entry), offset),
647				gfp_mask, mpol, ilx, &page_allocated, false);
648		if (!folio)
649			continue;
650		if (page_allocated) {
651			swap_read_folio(folio, &splug);
652			if (offset != entry_offset) {
653				folio_set_readahead(folio);
654				count_vm_event(SWAP_RA);
655			}
656		}
657		folio_put(folio);
658	}
659	blk_finish_plug(&plug);
660	swap_read_unplug(splug);
661	lru_add_drain();	/* Push any new pages onto the LRU now */
662skip:
663	/* The page was likely read above, so no need for plugging here */
664	folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
665					&page_allocated, false);
666	if (unlikely(page_allocated))
667		swap_read_folio(folio, NULL);
668	return folio;
669}
670
671static int swap_vma_ra_win(struct vm_fault *vmf, unsigned long *start,
672			   unsigned long *end)
673{
674	struct vm_area_struct *vma = vmf->vma;
675	unsigned long ra_val;
676	unsigned long faddr, prev_faddr, left, right;
677	unsigned int max_win, hits, prev_win, win;
678
679	max_win = 1 << min(READ_ONCE(page_cluster), SWAP_RA_ORDER_CEILING);
680	if (max_win == 1)
681		return 1;
682
683	faddr = vmf->address;
684	ra_val = GET_SWAP_RA_VAL(vma);
685	prev_faddr = SWAP_RA_ADDR(ra_val);
686	prev_win = SWAP_RA_WIN(ra_val);
687	hits = SWAP_RA_HITS(ra_val);
688	win = __swapin_nr_pages(PFN_DOWN(prev_faddr), PFN_DOWN(faddr), hits,
689				max_win, prev_win);
690	atomic_long_set(&vma->swap_readahead_info, SWAP_RA_VAL(faddr, win, 0));
691	if (win == 1)
692		return 1;
693
694	if (faddr == prev_faddr + PAGE_SIZE)
695		left = faddr;
696	else if (prev_faddr == faddr + PAGE_SIZE)
697		left = faddr - (win << PAGE_SHIFT) + PAGE_SIZE;
698	else
699		left = faddr - (((win - 1) / 2) << PAGE_SHIFT);
700	right = left + (win << PAGE_SHIFT);
701	if ((long)left < 0)
702		left = 0;
703	*start = max3(left, vma->vm_start, faddr & PMD_MASK);
704	*end = min3(right, vma->vm_end, (faddr & PMD_MASK) + PMD_SIZE);
705
706	return win;
707}
708
709/**
710 * swap_vma_readahead - swap in pages in hope we need them soon
711 * @targ_entry: swap entry of the targeted memory
712 * @gfp_mask: memory allocation flags
713 * @mpol: NUMA memory allocation policy to be applied
714 * @targ_ilx: NUMA interleave index, for use only when MPOL_INTERLEAVE
715 * @vmf: fault information
716 *
717 * Returns the struct folio for entry and addr, after queueing swapin.
718 *
719 * Primitive swap readahead code. We simply read in a few pages whose
720 * virtual addresses are around the fault address in the same vma.
721 *
722 * Caller must hold read mmap_lock if vmf->vma is not NULL.
723 *
724 */
725static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask,
726		struct mempolicy *mpol, pgoff_t targ_ilx, struct vm_fault *vmf)
727{
728	struct blk_plug plug;
729	struct swap_iocb *splug = NULL;
730	struct folio *folio;
731	pte_t *pte = NULL, pentry;
732	int win;
733	unsigned long start, end, addr;
734	pgoff_t ilx;
735	bool page_allocated;
736
737	win = swap_vma_ra_win(vmf, &start, &end);
738	if (win == 1)
739		goto skip;
740
741	ilx = targ_ilx - PFN_DOWN(vmf->address - start);
742
743	blk_start_plug(&plug);
744	for (addr = start; addr < end; ilx++, addr += PAGE_SIZE) {
745		struct swap_info_struct *si = NULL;
746		softleaf_t entry;
747
748		if (!pte++) {
749			pte = pte_offset_map(vmf->pmd, addr);
750			if (!pte)
751				break;
752		}
753		pentry = ptep_get_lockless(pte);
754		entry = softleaf_from_pte(pentry);
755
756		if (!softleaf_is_swap(entry))
757			continue;
758		pte_unmap(pte);
759		pte = NULL;
760		/*
761		 * Readahead entry may come from a device that we are not
762		 * holding a reference to, try to grab a reference, or skip.
763		 */
764		if (swp_type(entry) != swp_type(targ_entry)) {
765			si = get_swap_device(entry);
766			if (!si)
767				continue;
768		}
769		folio = __read_swap_cache_async(entry, gfp_mask, mpol, ilx,
770						&page_allocated, false);
771		if (si)
772			put_swap_device(si);
773		if (!folio)
774			continue;
775		if (page_allocated) {
776			swap_read_folio(folio, &splug);
777			if (addr != vmf->address) {
778				folio_set_readahead(folio);
779				count_vm_event(SWAP_RA);
780			}
781		}
782		folio_put(folio);
783	}
784	if (pte)
785		pte_unmap(pte);
786	blk_finish_plug(&plug);
787	swap_read_unplug(splug);
788	lru_add_drain();
789skip:
790	/* The folio was likely read above, so no need for plugging here */
791	folio = __read_swap_cache_async(targ_entry, gfp_mask, mpol, targ_ilx,
792					&page_allocated, false);
793	if (unlikely(page_allocated))
794		swap_read_folio(folio, NULL);
795	return folio;
796}
797
798/**
799 * swapin_readahead - swap in pages in hope we need them soon
800 * @entry: swap entry of this memory
801 * @gfp_mask: memory allocation flags
802 * @vmf: fault information
803 *
804 * Returns the struct folio for entry and addr, after queueing swapin.
805 *
806 * It's a main entry function for swap readahead. By the configuration,
807 * it will read ahead blocks by cluster-based(ie, physical disk based)
808 * or vma-based(ie, virtual address based on faulty address) readahead.
809 */
810struct folio *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
811				struct vm_fault *vmf)
812{
813	struct mempolicy *mpol;
814	pgoff_t ilx;
815	struct folio *folio;
816
817	mpol = get_vma_policy(vmf->vma, vmf->address, 0, &ilx);
818	folio = swap_use_vma_readahead() ?
819		swap_vma_readahead(entry, gfp_mask, mpol, ilx, vmf) :
820		swap_cluster_readahead(entry, gfp_mask, mpol, ilx);
821	mpol_cond_put(mpol);
822
823	return folio;
824}
825
826#ifdef CONFIG_SYSFS
827static ssize_t vma_ra_enabled_show(struct kobject *kobj,
828				     struct kobj_attribute *attr, char *buf)
829{
830	return sysfs_emit(buf, "%s\n", str_true_false(enable_vma_readahead));
831}
832static ssize_t vma_ra_enabled_store(struct kobject *kobj,
833				      struct kobj_attribute *attr,
834				      const char *buf, size_t count)
835{
836	ssize_t ret;
837
838	ret = kstrtobool(buf, &enable_vma_readahead);
839	if (ret)
840		return ret;
841
842	return count;
843}
844static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled);
845
846static struct attribute *swap_attrs[] = {
847	&vma_ra_enabled_attr.attr,
848	NULL,
849};
850
851static const struct attribute_group swap_attr_group = {
852	.attrs = swap_attrs,
853};
854
855static int __init swap_init(void)
856{
857	int err;
858	struct kobject *swap_kobj;
859
860	swap_kobj = kobject_create_and_add("swap", mm_kobj);
861	if (!swap_kobj) {
862		pr_err("failed to create swap kobject\n");
863		return -ENOMEM;
864	}
865	err = sysfs_create_group(swap_kobj, &swap_attr_group);
866	if (err) {
867		pr_err("failed to register swap group\n");
868		goto delete_obj;
869	}
870	/* Swap cache writeback is LRU based, no tags for it */
871	mapping_set_no_writeback_tags(&swap_space);
872	return 0;
873
874delete_obj:
875	kobject_put(swap_kobj);
876	return err;
877}
878subsys_initcall(swap_init);
879#endif