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/swap.h>
14 #include <linux/swapops.h>
15 #include <linux/init.h>
16 #include <linux/pagemap.h>
17 #include <linux/backing-dev.h>
18 #include <linux/blkdev.h>
19 #include <linux/pagevec.h>
20 #include <linux/migrate.h>
21 #include <linux/vmalloc.h>
22 #include <linux/swap_slots.h>
23 #include <linux/huge_mm.h>
24 #include <linux/shmem_fs.h>
25 #include "internal.h"
26 #include "swap.h"
27 
28 /*
29  * swapper_space is a fiction, retained to simplify the path through
30  * vmscan's shrink_page_list.
31  */
32 static const struct address_space_operations swap_aops = {
33 	.writepage	= swap_writepage,
34 	.dirty_folio	= noop_dirty_folio,
35 #ifdef CONFIG_MIGRATION
36 	.migratepage	= migrate_page,
37 #endif
38 };
39 
40 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
41 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
42 static bool enable_vma_readahead __read_mostly = true;
43 
44 #define SWAP_RA_WIN_SHIFT	(PAGE_SHIFT / 2)
45 #define SWAP_RA_HITS_MASK	((1UL << SWAP_RA_WIN_SHIFT) - 1)
46 #define SWAP_RA_HITS_MAX	SWAP_RA_HITS_MASK
47 #define SWAP_RA_WIN_MASK	(~PAGE_MASK & ~SWAP_RA_HITS_MASK)
48 
49 #define SWAP_RA_HITS(v)		((v) & SWAP_RA_HITS_MASK)
50 #define SWAP_RA_WIN(v)		(((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
51 #define SWAP_RA_ADDR(v)		((v) & PAGE_MASK)
52 
53 #define SWAP_RA_VAL(addr, win, hits)				\
54 	(((addr) & PAGE_MASK) |					\
55 	 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |	\
56 	 ((hits) & SWAP_RA_HITS_MASK))
57 
58 /* Initial readahead hits is 4 to start up with a small window */
59 #define GET_SWAP_RA_VAL(vma)					\
60 	(atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
61 
62 #define INC_CACHE_INFO(x)	data_race(swap_cache_info.x++)
63 #define ADD_CACHE_INFO(x, nr)	data_race(swap_cache_info.x += (nr))
64 
65 static struct {
66 	unsigned long add_total;
67 	unsigned long del_total;
68 	unsigned long find_success;
69 	unsigned long find_total;
70 } swap_cache_info;
71 
72 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
73 
show_swap_cache_info(void)74 void show_swap_cache_info(void)
75 {
76 	printk("%lu pages in swap cache\n", total_swapcache_pages());
77 	printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
78 		swap_cache_info.add_total, swap_cache_info.del_total,
79 		swap_cache_info.find_success, swap_cache_info.find_total);
80 	printk("Free swap  = %ldkB\n",
81 		get_nr_swap_pages() << (PAGE_SHIFT - 10));
82 	printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
83 }
84 
get_shadow_from_swap_cache(swp_entry_t entry)85 void *get_shadow_from_swap_cache(swp_entry_t entry)
86 {
87 	struct address_space *address_space = swap_address_space(entry);
88 	pgoff_t idx = swp_offset(entry);
89 	struct page *page;
90 
91 	page = xa_load(&address_space->i_pages, idx);
92 	if (xa_is_value(page))
93 		return page;
94 	return NULL;
95 }
96 
97 /*
98  * add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
99  * but sets SwapCache flag and private instead of mapping and index.
100  */
add_to_swap_cache(struct page * page,swp_entry_t entry,gfp_t gfp,void ** shadowp)101 int add_to_swap_cache(struct page *page, swp_entry_t entry,
102 			gfp_t gfp, void **shadowp)
103 {
104 	struct address_space *address_space = swap_address_space(entry);
105 	pgoff_t idx = swp_offset(entry);
106 	XA_STATE_ORDER(xas, &address_space->i_pages, idx, compound_order(page));
107 	unsigned long i, nr = thp_nr_pages(page);
108 	void *old;
109 
110 	VM_BUG_ON_PAGE(!PageLocked(page), page);
111 	VM_BUG_ON_PAGE(PageSwapCache(page), page);
112 	VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
113 
114 	page_ref_add(page, nr);
115 	SetPageSwapCache(page);
116 
117 	do {
118 		xas_lock_irq(&xas);
119 		xas_create_range(&xas);
120 		if (xas_error(&xas))
121 			goto unlock;
122 		for (i = 0; i < nr; i++) {
123 			VM_BUG_ON_PAGE(xas.xa_index != idx + i, page);
124 			old = xas_load(&xas);
125 			if (xa_is_value(old)) {
126 				if (shadowp)
127 					*shadowp = old;
128 			}
129 			set_page_private(page + i, entry.val + i);
130 			xas_store(&xas, page);
131 			xas_next(&xas);
132 		}
133 		address_space->nrpages += nr;
134 		__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
135 		__mod_lruvec_page_state(page, NR_SWAPCACHE, nr);
136 		ADD_CACHE_INFO(add_total, nr);
137 unlock:
138 		xas_unlock_irq(&xas);
139 	} while (xas_nomem(&xas, gfp));
140 
141 	if (!xas_error(&xas))
142 		return 0;
143 
144 	ClearPageSwapCache(page);
145 	page_ref_sub(page, nr);
146 	return xas_error(&xas);
147 }
148 
149 /*
150  * This must be called only on pages that have
151  * been verified to be in the swap cache.
152  */
__delete_from_swap_cache(struct page * page,swp_entry_t entry,void * shadow)153 void __delete_from_swap_cache(struct page *page,
154 			swp_entry_t entry, void *shadow)
155 {
156 	struct address_space *address_space = swap_address_space(entry);
157 	int i, nr = thp_nr_pages(page);
158 	pgoff_t idx = swp_offset(entry);
159 	XA_STATE(xas, &address_space->i_pages, idx);
160 
161 	VM_BUG_ON_PAGE(!PageLocked(page), page);
162 	VM_BUG_ON_PAGE(!PageSwapCache(page), page);
163 	VM_BUG_ON_PAGE(PageWriteback(page), page);
164 
165 	for (i = 0; i < nr; i++) {
166 		void *entry = xas_store(&xas, shadow);
167 		VM_BUG_ON_PAGE(entry != page, entry);
168 		set_page_private(page + i, 0);
169 		xas_next(&xas);
170 	}
171 	ClearPageSwapCache(page);
172 	address_space->nrpages -= nr;
173 	__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
174 	__mod_lruvec_page_state(page, NR_SWAPCACHE, -nr);
175 	ADD_CACHE_INFO(del_total, nr);
176 }
177 
178 /**
179  * add_to_swap - allocate swap space for a folio
180  * @folio: folio we want to move to swap
181  *
182  * Allocate swap space for the folio and add the folio to the
183  * swap cache.
184  *
185  * Context: Caller needs to hold the folio lock.
186  * Return: Whether the folio was added to the swap cache.
187  */
add_to_swap(struct folio * folio)188 bool add_to_swap(struct folio *folio)
189 {
190 	swp_entry_t entry;
191 	int err;
192 
193 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
194 	VM_BUG_ON_FOLIO(!folio_test_uptodate(folio), folio);
195 
196 	entry = folio_alloc_swap(folio);
197 	if (!entry.val)
198 		return false;
199 
200 	/*
201 	 * XArray node allocations from PF_MEMALLOC contexts could
202 	 * completely exhaust the page allocator. __GFP_NOMEMALLOC
203 	 * stops emergency reserves from being allocated.
204 	 *
205 	 * TODO: this could cause a theoretical memory reclaim
206 	 * deadlock in the swap out path.
207 	 */
208 	/*
209 	 * Add it to the swap cache.
210 	 */
211 	err = add_to_swap_cache(&folio->page, entry,
212 			__GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN, NULL);
213 	if (err)
214 		/*
215 		 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
216 		 * clear SWAP_HAS_CACHE flag.
217 		 */
218 		goto fail;
219 	/*
220 	 * Normally the folio will be dirtied in unmap because its
221 	 * pte should be dirty. A special case is MADV_FREE page. The
222 	 * page's pte could have dirty bit cleared but the folio's
223 	 * SwapBacked flag is still set because clearing the dirty bit
224 	 * and SwapBacked flag has no lock protected. For such folio,
225 	 * unmap will not set dirty bit for it, so folio reclaim will
226 	 * not write the folio out. This can cause data corruption when
227 	 * the folio is swapped in later. Always setting the dirty flag
228 	 * for the folio solves the problem.
229 	 */
230 	folio_mark_dirty(folio);
231 
232 	return true;
233 
234 fail:
235 	put_swap_page(&folio->page, entry);
236 	return false;
237 }
238 
239 /*
240  * This must be called only on pages that have
241  * been verified to be in the swap cache and locked.
242  * It will never put the page into the free list,
243  * the caller has a reference on the page.
244  */
delete_from_swap_cache(struct page * page)245 void delete_from_swap_cache(struct page *page)
246 {
247 	swp_entry_t entry = { .val = page_private(page) };
248 	struct address_space *address_space = swap_address_space(entry);
249 
250 	xa_lock_irq(&address_space->i_pages);
251 	__delete_from_swap_cache(page, entry, NULL);
252 	xa_unlock_irq(&address_space->i_pages);
253 
254 	put_swap_page(page, entry);
255 	page_ref_sub(page, thp_nr_pages(page));
256 }
257 
clear_shadow_from_swap_cache(int type,unsigned long begin,unsigned long end)258 void clear_shadow_from_swap_cache(int type, unsigned long begin,
259 				unsigned long end)
260 {
261 	unsigned long curr = begin;
262 	void *old;
263 
264 	for (;;) {
265 		swp_entry_t entry = swp_entry(type, curr);
266 		struct address_space *address_space = swap_address_space(entry);
267 		XA_STATE(xas, &address_space->i_pages, curr);
268 
269 		xa_lock_irq(&address_space->i_pages);
270 		xas_for_each(&xas, old, end) {
271 			if (!xa_is_value(old))
272 				continue;
273 			xas_store(&xas, NULL);
274 		}
275 		xa_unlock_irq(&address_space->i_pages);
276 
277 		/* search the next swapcache until we meet end */
278 		curr >>= SWAP_ADDRESS_SPACE_SHIFT;
279 		curr++;
280 		curr <<= SWAP_ADDRESS_SPACE_SHIFT;
281 		if (curr > end)
282 			break;
283 	}
284 }
285 
286 /*
287  * If we are the only user, then try to free up the swap cache.
288  *
289  * Its ok to check for PageSwapCache without the page lock
290  * here because we are going to recheck again inside
291  * try_to_free_swap() _with_ the lock.
292  * 					- Marcelo
293  */
free_swap_cache(struct page * page)294 void free_swap_cache(struct page *page)
295 {
296 	if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
297 		try_to_free_swap(page);
298 		unlock_page(page);
299 	}
300 }
301 
302 /*
303  * Perform a free_page(), also freeing any swap cache associated with
304  * this page if it is the last user of the page.
305  */
free_page_and_swap_cache(struct page * page)306 void free_page_and_swap_cache(struct page *page)
307 {
308 	free_swap_cache(page);
309 	if (!is_huge_zero_page(page))
310 		put_page(page);
311 }
312 
313 /*
314  * Passed an array of pages, drop them all from swapcache and then release
315  * them.  They are removed from the LRU and freed if this is their last use.
316  */
free_pages_and_swap_cache(struct page ** pages,int nr)317 void free_pages_and_swap_cache(struct page **pages, int nr)
318 {
319 	struct page **pagep = pages;
320 	int i;
321 
322 	lru_add_drain();
323 	for (i = 0; i < nr; i++)
324 		free_swap_cache(pagep[i]);
325 	release_pages(pagep, nr);
326 }
327 
swap_use_vma_readahead(void)328 static inline bool swap_use_vma_readahead(void)
329 {
330 	return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
331 }
332 
333 /*
334  * Lookup a swap entry in the swap cache. A found page will be returned
335  * unlocked and with its refcount incremented - we rely on the kernel
336  * lock getting page table operations atomic even if we drop the page
337  * lock before returning.
338  */
lookup_swap_cache(swp_entry_t entry,struct vm_area_struct * vma,unsigned long addr)339 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
340 			       unsigned long addr)
341 {
342 	struct page *page;
343 	struct swap_info_struct *si;
344 
345 	si = get_swap_device(entry);
346 	if (!si)
347 		return NULL;
348 	page = find_get_page(swap_address_space(entry), swp_offset(entry));
349 	put_swap_device(si);
350 
351 	INC_CACHE_INFO(find_total);
352 	if (page) {
353 		bool vma_ra = swap_use_vma_readahead();
354 		bool readahead;
355 
356 		INC_CACHE_INFO(find_success);
357 		/*
358 		 * At the moment, we don't support PG_readahead for anon THP
359 		 * so let's bail out rather than confusing the readahead stat.
360 		 */
361 		if (unlikely(PageTransCompound(page)))
362 			return page;
363 
364 		readahead = TestClearPageReadahead(page);
365 		if (vma && vma_ra) {
366 			unsigned long ra_val;
367 			int win, hits;
368 
369 			ra_val = GET_SWAP_RA_VAL(vma);
370 			win = SWAP_RA_WIN(ra_val);
371 			hits = SWAP_RA_HITS(ra_val);
372 			if (readahead)
373 				hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
374 			atomic_long_set(&vma->swap_readahead_info,
375 					SWAP_RA_VAL(addr, win, hits));
376 		}
377 
378 		if (readahead) {
379 			count_vm_event(SWAP_RA_HIT);
380 			if (!vma || !vma_ra)
381 				atomic_inc(&swapin_readahead_hits);
382 		}
383 	}
384 
385 	return page;
386 }
387 
388 /**
389  * find_get_incore_page - Find and get a page from the page or swap caches.
390  * @mapping: The address_space to search.
391  * @index: The page cache index.
392  *
393  * This differs from find_get_page() in that it will also look for the
394  * page in the swap cache.
395  *
396  * Return: The found page or %NULL.
397  */
find_get_incore_page(struct address_space * mapping,pgoff_t index)398 struct page *find_get_incore_page(struct address_space *mapping, pgoff_t index)
399 {
400 	swp_entry_t swp;
401 	struct swap_info_struct *si;
402 	struct page *page = pagecache_get_page(mapping, index,
403 						FGP_ENTRY | FGP_HEAD, 0);
404 
405 	if (!page)
406 		return page;
407 	if (!xa_is_value(page))
408 		return find_subpage(page, index);
409 	if (!shmem_mapping(mapping))
410 		return NULL;
411 
412 	swp = radix_to_swp_entry(page);
413 	/* There might be swapin error entries in shmem mapping. */
414 	if (non_swap_entry(swp))
415 		return NULL;
416 	/* Prevent swapoff from happening to us */
417 	si = get_swap_device(swp);
418 	if (!si)
419 		return NULL;
420 	page = find_get_page(swap_address_space(swp), swp_offset(swp));
421 	put_swap_device(si);
422 	return page;
423 }
424 
__read_swap_cache_async(swp_entry_t entry,gfp_t gfp_mask,struct vm_area_struct * vma,unsigned long addr,bool * new_page_allocated)425 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
426 			struct vm_area_struct *vma, unsigned long addr,
427 			bool *new_page_allocated)
428 {
429 	struct swap_info_struct *si;
430 	struct page *page;
431 	void *shadow = NULL;
432 
433 	*new_page_allocated = false;
434 
435 	for (;;) {
436 		int err;
437 		/*
438 		 * First check the swap cache.  Since this is normally
439 		 * called after lookup_swap_cache() failed, re-calling
440 		 * that would confuse statistics.
441 		 */
442 		si = get_swap_device(entry);
443 		if (!si)
444 			return NULL;
445 		page = find_get_page(swap_address_space(entry),
446 				     swp_offset(entry));
447 		put_swap_device(si);
448 		if (page)
449 			return page;
450 
451 		/*
452 		 * Just skip read ahead for unused swap slot.
453 		 * During swap_off when swap_slot_cache is disabled,
454 		 * we have to handle the race between putting
455 		 * swap entry in swap cache and marking swap slot
456 		 * as SWAP_HAS_CACHE.  That's done in later part of code or
457 		 * else swap_off will be aborted if we return NULL.
458 		 */
459 		if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
460 			return NULL;
461 
462 		/*
463 		 * Get a new page to read into from swap.  Allocate it now,
464 		 * before marking swap_map SWAP_HAS_CACHE, when -EEXIST will
465 		 * cause any racers to loop around until we add it to cache.
466 		 */
467 		page = alloc_page_vma(gfp_mask, vma, addr);
468 		if (!page)
469 			return NULL;
470 
471 		/*
472 		 * Swap entry may have been freed since our caller observed it.
473 		 */
474 		err = swapcache_prepare(entry);
475 		if (!err)
476 			break;
477 
478 		put_page(page);
479 		if (err != -EEXIST)
480 			return NULL;
481 
482 		/*
483 		 * We might race against __delete_from_swap_cache(), and
484 		 * stumble across a swap_map entry whose SWAP_HAS_CACHE
485 		 * has not yet been cleared.  Or race against another
486 		 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
487 		 * in swap_map, but not yet added its page to swap cache.
488 		 */
489 		schedule_timeout_uninterruptible(1);
490 	}
491 
492 	/*
493 	 * The swap entry is ours to swap in. Prepare the new page.
494 	 */
495 
496 	__SetPageLocked(page);
497 	__SetPageSwapBacked(page);
498 
499 	if (mem_cgroup_swapin_charge_page(page, NULL, gfp_mask, entry))
500 		goto fail_unlock;
501 
502 	/* May fail (-ENOMEM) if XArray node allocation failed. */
503 	if (add_to_swap_cache(page, entry, gfp_mask & GFP_RECLAIM_MASK, &shadow))
504 		goto fail_unlock;
505 
506 	mem_cgroup_swapin_uncharge_swap(entry);
507 
508 	if (shadow)
509 		workingset_refault(page_folio(page), shadow);
510 
511 	/* Caller will initiate read into locked page */
512 	lru_cache_add(page);
513 	*new_page_allocated = true;
514 	return page;
515 
516 fail_unlock:
517 	put_swap_page(page, entry);
518 	unlock_page(page);
519 	put_page(page);
520 	return NULL;
521 }
522 
523 /*
524  * Locate a page of swap in physical memory, reserving swap cache space
525  * and reading the disk if it is not already cached.
526  * A failure return means that either the page allocation failed or that
527  * the swap entry is no longer in use.
528  */
read_swap_cache_async(swp_entry_t entry,gfp_t gfp_mask,struct vm_area_struct * vma,unsigned long addr,bool do_poll,struct swap_iocb ** plug)529 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
530 				   struct vm_area_struct *vma,
531 				   unsigned long addr, bool do_poll,
532 				   struct swap_iocb **plug)
533 {
534 	bool page_was_allocated;
535 	struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
536 			vma, addr, &page_was_allocated);
537 
538 	if (page_was_allocated)
539 		swap_readpage(retpage, do_poll, plug);
540 
541 	return retpage;
542 }
543 
__swapin_nr_pages(unsigned long prev_offset,unsigned long offset,int hits,int max_pages,int prev_win)544 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
545 				      unsigned long offset,
546 				      int hits,
547 				      int max_pages,
548 				      int prev_win)
549 {
550 	unsigned int pages, last_ra;
551 
552 	/*
553 	 * This heuristic has been found to work well on both sequential and
554 	 * random loads, swapping to hard disk or to SSD: please don't ask
555 	 * what the "+ 2" means, it just happens to work well, that's all.
556 	 */
557 	pages = hits + 2;
558 	if (pages == 2) {
559 		/*
560 		 * We can have no readahead hits to judge by: but must not get
561 		 * stuck here forever, so check for an adjacent offset instead
562 		 * (and don't even bother to check whether swap type is same).
563 		 */
564 		if (offset != prev_offset + 1 && offset != prev_offset - 1)
565 			pages = 1;
566 	} else {
567 		unsigned int roundup = 4;
568 		while (roundup < pages)
569 			roundup <<= 1;
570 		pages = roundup;
571 	}
572 
573 	if (pages > max_pages)
574 		pages = max_pages;
575 
576 	/* Don't shrink readahead too fast */
577 	last_ra = prev_win / 2;
578 	if (pages < last_ra)
579 		pages = last_ra;
580 
581 	return pages;
582 }
583 
swapin_nr_pages(unsigned long offset)584 static unsigned long swapin_nr_pages(unsigned long offset)
585 {
586 	static unsigned long prev_offset;
587 	unsigned int hits, pages, max_pages;
588 	static atomic_t last_readahead_pages;
589 
590 	max_pages = 1 << READ_ONCE(page_cluster);
591 	if (max_pages <= 1)
592 		return 1;
593 
594 	hits = atomic_xchg(&swapin_readahead_hits, 0);
595 	pages = __swapin_nr_pages(READ_ONCE(prev_offset), offset, hits,
596 				  max_pages,
597 				  atomic_read(&last_readahead_pages));
598 	if (!hits)
599 		WRITE_ONCE(prev_offset, offset);
600 	atomic_set(&last_readahead_pages, pages);
601 
602 	return pages;
603 }
604 
605 /**
606  * swap_cluster_readahead - swap in pages in hope we need them soon
607  * @entry: swap entry of this memory
608  * @gfp_mask: memory allocation flags
609  * @vmf: fault information
610  *
611  * Returns the struct page for entry and addr, after queueing swapin.
612  *
613  * Primitive swap readahead code. We simply read an aligned block of
614  * (1 << page_cluster) entries in the swap area. This method is chosen
615  * because it doesn't cost us any seek time.  We also make sure to queue
616  * the 'original' request together with the readahead ones...
617  *
618  * This has been extended to use the NUMA policies from the mm triggering
619  * the readahead.
620  *
621  * Caller must hold read mmap_lock if vmf->vma is not NULL.
622  */
swap_cluster_readahead(swp_entry_t entry,gfp_t gfp_mask,struct vm_fault * vmf)623 struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
624 				struct vm_fault *vmf)
625 {
626 	struct page *page;
627 	unsigned long entry_offset = swp_offset(entry);
628 	unsigned long offset = entry_offset;
629 	unsigned long start_offset, end_offset;
630 	unsigned long mask;
631 	struct swap_info_struct *si = swp_swap_info(entry);
632 	struct blk_plug plug;
633 	struct swap_iocb *splug = NULL;
634 	bool do_poll = true, page_allocated;
635 	struct vm_area_struct *vma = vmf->vma;
636 	unsigned long addr = vmf->address;
637 
638 	mask = swapin_nr_pages(offset) - 1;
639 	if (!mask)
640 		goto skip;
641 
642 	do_poll = false;
643 	/* Read a page_cluster sized and aligned cluster around offset. */
644 	start_offset = offset & ~mask;
645 	end_offset = offset | mask;
646 	if (!start_offset)	/* First page is swap header. */
647 		start_offset++;
648 	if (end_offset >= si->max)
649 		end_offset = si->max - 1;
650 
651 	blk_start_plug(&plug);
652 	for (offset = start_offset; offset <= end_offset ; offset++) {
653 		/* Ok, do the async read-ahead now */
654 		page = __read_swap_cache_async(
655 			swp_entry(swp_type(entry), offset),
656 			gfp_mask, vma, addr, &page_allocated);
657 		if (!page)
658 			continue;
659 		if (page_allocated) {
660 			swap_readpage(page, false, &splug);
661 			if (offset != entry_offset) {
662 				SetPageReadahead(page);
663 				count_vm_event(SWAP_RA);
664 			}
665 		}
666 		put_page(page);
667 	}
668 	blk_finish_plug(&plug);
669 	swap_read_unplug(splug);
670 
671 	lru_add_drain();	/* Push any new pages onto the LRU now */
672 skip:
673 	/* The page was likely read above, so no need for plugging here */
674 	return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll, NULL);
675 }
676 
init_swap_address_space(unsigned int type,unsigned long nr_pages)677 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
678 {
679 	struct address_space *spaces, *space;
680 	unsigned int i, nr;
681 
682 	nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
683 	spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
684 	if (!spaces)
685 		return -ENOMEM;
686 	for (i = 0; i < nr; i++) {
687 		space = spaces + i;
688 		xa_init_flags(&space->i_pages, XA_FLAGS_LOCK_IRQ);
689 		atomic_set(&space->i_mmap_writable, 0);
690 		space->a_ops = &swap_aops;
691 		/* swap cache doesn't use writeback related tags */
692 		mapping_set_no_writeback_tags(space);
693 	}
694 	nr_swapper_spaces[type] = nr;
695 	swapper_spaces[type] = spaces;
696 
697 	return 0;
698 }
699 
exit_swap_address_space(unsigned int type)700 void exit_swap_address_space(unsigned int type)
701 {
702 	int i;
703 	struct address_space *spaces = swapper_spaces[type];
704 
705 	for (i = 0; i < nr_swapper_spaces[type]; i++)
706 		VM_WARN_ON_ONCE(!mapping_empty(&spaces[i]));
707 	kvfree(spaces);
708 	nr_swapper_spaces[type] = 0;
709 	swapper_spaces[type] = NULL;
710 }
711 
swap_ra_clamp_pfn(struct vm_area_struct * vma,unsigned long faddr,unsigned long lpfn,unsigned long rpfn,unsigned long * start,unsigned long * end)712 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
713 				     unsigned long faddr,
714 				     unsigned long lpfn,
715 				     unsigned long rpfn,
716 				     unsigned long *start,
717 				     unsigned long *end)
718 {
719 	*start = max3(lpfn, PFN_DOWN(vma->vm_start),
720 		      PFN_DOWN(faddr & PMD_MASK));
721 	*end = min3(rpfn, PFN_DOWN(vma->vm_end),
722 		    PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
723 }
724 
swap_ra_info(struct vm_fault * vmf,struct vma_swap_readahead * ra_info)725 static void swap_ra_info(struct vm_fault *vmf,
726 			struct vma_swap_readahead *ra_info)
727 {
728 	struct vm_area_struct *vma = vmf->vma;
729 	unsigned long ra_val;
730 	unsigned long faddr, pfn, fpfn;
731 	unsigned long start, end;
732 	pte_t *pte, *orig_pte;
733 	unsigned int max_win, hits, prev_win, win, left;
734 #ifndef CONFIG_64BIT
735 	pte_t *tpte;
736 #endif
737 
738 	max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
739 			     SWAP_RA_ORDER_CEILING);
740 	if (max_win == 1) {
741 		ra_info->win = 1;
742 		return;
743 	}
744 
745 	faddr = vmf->address;
746 	orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
747 
748 	fpfn = PFN_DOWN(faddr);
749 	ra_val = GET_SWAP_RA_VAL(vma);
750 	pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
751 	prev_win = SWAP_RA_WIN(ra_val);
752 	hits = SWAP_RA_HITS(ra_val);
753 	ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
754 					       max_win, prev_win);
755 	atomic_long_set(&vma->swap_readahead_info,
756 			SWAP_RA_VAL(faddr, win, 0));
757 
758 	if (win == 1) {
759 		pte_unmap(orig_pte);
760 		return;
761 	}
762 
763 	/* Copy the PTEs because the page table may be unmapped */
764 	if (fpfn == pfn + 1)
765 		swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
766 	else if (pfn == fpfn + 1)
767 		swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
768 				  &start, &end);
769 	else {
770 		left = (win - 1) / 2;
771 		swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
772 				  &start, &end);
773 	}
774 	ra_info->nr_pte = end - start;
775 	ra_info->offset = fpfn - start;
776 	pte -= ra_info->offset;
777 #ifdef CONFIG_64BIT
778 	ra_info->ptes = pte;
779 #else
780 	tpte = ra_info->ptes;
781 	for (pfn = start; pfn != end; pfn++)
782 		*tpte++ = *pte++;
783 #endif
784 	pte_unmap(orig_pte);
785 }
786 
787 /**
788  * swap_vma_readahead - swap in pages in hope we need them soon
789  * @fentry: swap entry of this memory
790  * @gfp_mask: memory allocation flags
791  * @vmf: fault information
792  *
793  * Returns the struct page for entry and addr, after queueing swapin.
794  *
795  * Primitive swap readahead code. We simply read in a few pages whose
796  * virtual addresses are around the fault address in the same vma.
797  *
798  * Caller must hold read mmap_lock if vmf->vma is not NULL.
799  *
800  */
swap_vma_readahead(swp_entry_t fentry,gfp_t gfp_mask,struct vm_fault * vmf)801 static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
802 				       struct vm_fault *vmf)
803 {
804 	struct blk_plug plug;
805 	struct swap_iocb *splug = NULL;
806 	struct vm_area_struct *vma = vmf->vma;
807 	struct page *page;
808 	pte_t *pte, pentry;
809 	swp_entry_t entry;
810 	unsigned int i;
811 	bool page_allocated;
812 	struct vma_swap_readahead ra_info = {
813 		.win = 1,
814 	};
815 
816 	swap_ra_info(vmf, &ra_info);
817 	if (ra_info.win == 1)
818 		goto skip;
819 
820 	blk_start_plug(&plug);
821 	for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
822 	     i++, pte++) {
823 		pentry = *pte;
824 		if (!is_swap_pte(pentry))
825 			continue;
826 		entry = pte_to_swp_entry(pentry);
827 		if (unlikely(non_swap_entry(entry)))
828 			continue;
829 		page = __read_swap_cache_async(entry, gfp_mask, vma,
830 					       vmf->address, &page_allocated);
831 		if (!page)
832 			continue;
833 		if (page_allocated) {
834 			swap_readpage(page, false, &splug);
835 			if (i != ra_info.offset) {
836 				SetPageReadahead(page);
837 				count_vm_event(SWAP_RA);
838 			}
839 		}
840 		put_page(page);
841 	}
842 	blk_finish_plug(&plug);
843 	swap_read_unplug(splug);
844 	lru_add_drain();
845 skip:
846 	/* The page was likely read above, so no need for plugging here */
847 	return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
848 				     ra_info.win == 1, NULL);
849 }
850 
851 /**
852  * swapin_readahead - swap in pages in hope we need them soon
853  * @entry: swap entry of this memory
854  * @gfp_mask: memory allocation flags
855  * @vmf: fault information
856  *
857  * Returns the struct page for entry and addr, after queueing swapin.
858  *
859  * It's a main entry function for swap readahead. By the configuration,
860  * it will read ahead blocks by cluster-based(ie, physical disk based)
861  * or vma-based(ie, virtual address based on faulty address) readahead.
862  */
swapin_readahead(swp_entry_t entry,gfp_t gfp_mask,struct vm_fault * vmf)863 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
864 				struct vm_fault *vmf)
865 {
866 	return swap_use_vma_readahead() ?
867 			swap_vma_readahead(entry, gfp_mask, vmf) :
868 			swap_cluster_readahead(entry, gfp_mask, vmf);
869 }
870 
871 #ifdef CONFIG_SYSFS
vma_ra_enabled_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)872 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
873 				     struct kobj_attribute *attr, char *buf)
874 {
875 	return sysfs_emit(buf, "%s\n",
876 			  enable_vma_readahead ? "true" : "false");
877 }
vma_ra_enabled_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)878 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
879 				      struct kobj_attribute *attr,
880 				      const char *buf, size_t count)
881 {
882 	ssize_t ret;
883 
884 	ret = kstrtobool(buf, &enable_vma_readahead);
885 	if (ret)
886 		return ret;
887 
888 	return count;
889 }
890 static struct kobj_attribute vma_ra_enabled_attr = __ATTR_RW(vma_ra_enabled);
891 
892 static struct attribute *swap_attrs[] = {
893 	&vma_ra_enabled_attr.attr,
894 	NULL,
895 };
896 
897 static const struct attribute_group swap_attr_group = {
898 	.attrs = swap_attrs,
899 };
900 
swap_init_sysfs(void)901 static int __init swap_init_sysfs(void)
902 {
903 	int err;
904 	struct kobject *swap_kobj;
905 
906 	swap_kobj = kobject_create_and_add("swap", mm_kobj);
907 	if (!swap_kobj) {
908 		pr_err("failed to create swap kobject\n");
909 		return -ENOMEM;
910 	}
911 	err = sysfs_create_group(swap_kobj, &swap_attr_group);
912 	if (err) {
913 		pr_err("failed to register swap group\n");
914 		goto delete_obj;
915 	}
916 	return 0;
917 
918 delete_obj:
919 	kobject_put(swap_kobj);
920 	return err;
921 }
922 subsys_initcall(swap_init_sysfs);
923 #endif
924