xref: /linux/mm/swap_state.c (revision b1d29ba82cf2bc784f4c963ddd6a2cf29e229b33)
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 
25 #include <asm/pgtable.h>
26 
27 /*
28  * swapper_space is a fiction, retained to simplify the path through
29  * vmscan's shrink_page_list.
30  */
31 static const struct address_space_operations swap_aops = {
32 	.writepage	= swap_writepage,
33 	.set_page_dirty	= swap_set_page_dirty,
34 #ifdef CONFIG_MIGRATION
35 	.migratepage	= migrate_page,
36 #endif
37 };
38 
39 struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
40 static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
41 static bool enable_vma_readahead __read_mostly = true;
42 
43 #define SWAP_RA_WIN_SHIFT	(PAGE_SHIFT / 2)
44 #define SWAP_RA_HITS_MASK	((1UL << SWAP_RA_WIN_SHIFT) - 1)
45 #define SWAP_RA_HITS_MAX	SWAP_RA_HITS_MASK
46 #define SWAP_RA_WIN_MASK	(~PAGE_MASK & ~SWAP_RA_HITS_MASK)
47 
48 #define SWAP_RA_HITS(v)		((v) & SWAP_RA_HITS_MASK)
49 #define SWAP_RA_WIN(v)		(((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
50 #define SWAP_RA_ADDR(v)		((v) & PAGE_MASK)
51 
52 #define SWAP_RA_VAL(addr, win, hits)				\
53 	(((addr) & PAGE_MASK) |					\
54 	 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |	\
55 	 ((hits) & SWAP_RA_HITS_MASK))
56 
57 /* Initial readahead hits is 4 to start up with a small window */
58 #define GET_SWAP_RA_VAL(vma)					\
59 	(atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
60 
61 #define INC_CACHE_INFO(x)	do { swap_cache_info.x++; } while (0)
62 #define ADD_CACHE_INFO(x, nr)	do { swap_cache_info.x += (nr); } while (0)
63 
64 static struct {
65 	unsigned long add_total;
66 	unsigned long del_total;
67 	unsigned long find_success;
68 	unsigned long find_total;
69 } swap_cache_info;
70 
71 unsigned long total_swapcache_pages(void)
72 {
73 	unsigned int i, j, nr;
74 	unsigned long ret = 0;
75 	struct address_space *spaces;
76 
77 	rcu_read_lock();
78 	for (i = 0; i < MAX_SWAPFILES; i++) {
79 		/*
80 		 * The corresponding entries in nr_swapper_spaces and
81 		 * swapper_spaces will be reused only after at least
82 		 * one grace period.  So it is impossible for them
83 		 * belongs to different usage.
84 		 */
85 		nr = nr_swapper_spaces[i];
86 		spaces = rcu_dereference(swapper_spaces[i]);
87 		if (!nr || !spaces)
88 			continue;
89 		for (j = 0; j < nr; j++)
90 			ret += spaces[j].nrpages;
91 	}
92 	rcu_read_unlock();
93 	return ret;
94 }
95 
96 static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
97 
98 void show_swap_cache_info(void)
99 {
100 	printk("%lu pages in swap cache\n", total_swapcache_pages());
101 	printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
102 		swap_cache_info.add_total, swap_cache_info.del_total,
103 		swap_cache_info.find_success, swap_cache_info.find_total);
104 	printk("Free swap  = %ldkB\n",
105 		get_nr_swap_pages() << (PAGE_SHIFT - 10));
106 	printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
107 }
108 
109 /*
110  * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
111  * but sets SwapCache flag and private instead of mapping and index.
112  */
113 int __add_to_swap_cache(struct page *page, swp_entry_t entry)
114 {
115 	int error, i, nr = hpage_nr_pages(page);
116 	struct address_space *address_space;
117 	pgoff_t idx = swp_offset(entry);
118 
119 	VM_BUG_ON_PAGE(!PageLocked(page), page);
120 	VM_BUG_ON_PAGE(PageSwapCache(page), page);
121 	VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
122 
123 	page_ref_add(page, nr);
124 	SetPageSwapCache(page);
125 
126 	address_space = swap_address_space(entry);
127 	xa_lock_irq(&address_space->i_pages);
128 	for (i = 0; i < nr; i++) {
129 		set_page_private(page + i, entry.val + i);
130 		error = radix_tree_insert(&address_space->i_pages,
131 					  idx + i, page + i);
132 		if (unlikely(error))
133 			break;
134 	}
135 	if (likely(!error)) {
136 		address_space->nrpages += nr;
137 		__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
138 		ADD_CACHE_INFO(add_total, nr);
139 	} else {
140 		/*
141 		 * Only the context which have set SWAP_HAS_CACHE flag
142 		 * would call add_to_swap_cache().
143 		 * So add_to_swap_cache() doesn't returns -EEXIST.
144 		 */
145 		VM_BUG_ON(error == -EEXIST);
146 		set_page_private(page + i, 0UL);
147 		while (i--) {
148 			radix_tree_delete(&address_space->i_pages, idx + i);
149 			set_page_private(page + i, 0UL);
150 		}
151 		ClearPageSwapCache(page);
152 		page_ref_sub(page, nr);
153 	}
154 	xa_unlock_irq(&address_space->i_pages);
155 
156 	return error;
157 }
158 
159 
160 int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
161 {
162 	int error;
163 
164 	error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
165 	if (!error) {
166 		error = __add_to_swap_cache(page, entry);
167 		radix_tree_preload_end();
168 	}
169 	return error;
170 }
171 
172 /*
173  * This must be called only on pages that have
174  * been verified to be in the swap cache.
175  */
176 void __delete_from_swap_cache(struct page *page)
177 {
178 	struct address_space *address_space;
179 	int i, nr = hpage_nr_pages(page);
180 	swp_entry_t entry;
181 	pgoff_t idx;
182 
183 	VM_BUG_ON_PAGE(!PageLocked(page), page);
184 	VM_BUG_ON_PAGE(!PageSwapCache(page), page);
185 	VM_BUG_ON_PAGE(PageWriteback(page), page);
186 
187 	entry.val = page_private(page);
188 	address_space = swap_address_space(entry);
189 	idx = swp_offset(entry);
190 	for (i = 0; i < nr; i++) {
191 		radix_tree_delete(&address_space->i_pages, idx + i);
192 		set_page_private(page + i, 0);
193 	}
194 	ClearPageSwapCache(page);
195 	address_space->nrpages -= nr;
196 	__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
197 	ADD_CACHE_INFO(del_total, nr);
198 }
199 
200 /**
201  * add_to_swap - allocate swap space for a page
202  * @page: page we want to move to swap
203  *
204  * Allocate swap space for the page and add the page to the
205  * swap cache.  Caller needs to hold the page lock.
206  */
207 int add_to_swap(struct page *page)
208 {
209 	swp_entry_t entry;
210 	int err;
211 
212 	VM_BUG_ON_PAGE(!PageLocked(page), page);
213 	VM_BUG_ON_PAGE(!PageUptodate(page), page);
214 
215 	entry = get_swap_page(page);
216 	if (!entry.val)
217 		return 0;
218 
219 	/*
220 	 * Radix-tree node allocations from PF_MEMALLOC contexts could
221 	 * completely exhaust the page allocator. __GFP_NOMEMALLOC
222 	 * stops emergency reserves from being allocated.
223 	 *
224 	 * TODO: this could cause a theoretical memory reclaim
225 	 * deadlock in the swap out path.
226 	 */
227 	/*
228 	 * Add it to the swap cache.
229 	 */
230 	err = add_to_swap_cache(page, entry,
231 			__GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
232 	/* -ENOMEM radix-tree allocation failure */
233 	if (err)
234 		/*
235 		 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
236 		 * clear SWAP_HAS_CACHE flag.
237 		 */
238 		goto fail;
239 	/*
240 	 * Normally the page will be dirtied in unmap because its pte should be
241 	 * dirty. A special case is MADV_FREE page. The page'e pte could have
242 	 * dirty bit cleared but the page's SwapBacked bit is still set because
243 	 * clearing the dirty bit and SwapBacked bit has no lock protected. For
244 	 * such page, unmap will not set dirty bit for it, so page reclaim will
245 	 * not write the page out. This can cause data corruption when the page
246 	 * is swap in later. Always setting the dirty bit for the page solves
247 	 * the problem.
248 	 */
249 	set_page_dirty(page);
250 
251 	return 1;
252 
253 fail:
254 	put_swap_page(page, entry);
255 	return 0;
256 }
257 
258 /*
259  * This must be called only on pages that have
260  * been verified to be in the swap cache and locked.
261  * It will never put the page into the free list,
262  * the caller has a reference on the page.
263  */
264 void delete_from_swap_cache(struct page *page)
265 {
266 	swp_entry_t entry;
267 	struct address_space *address_space;
268 
269 	entry.val = page_private(page);
270 
271 	address_space = swap_address_space(entry);
272 	xa_lock_irq(&address_space->i_pages);
273 	__delete_from_swap_cache(page);
274 	xa_unlock_irq(&address_space->i_pages);
275 
276 	put_swap_page(page, entry);
277 	page_ref_sub(page, hpage_nr_pages(page));
278 }
279 
280 /*
281  * If we are the only user, then try to free up the swap cache.
282  *
283  * Its ok to check for PageSwapCache without the page lock
284  * here because we are going to recheck again inside
285  * try_to_free_swap() _with_ the lock.
286  * 					- Marcelo
287  */
288 static inline void free_swap_cache(struct page *page)
289 {
290 	if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
291 		try_to_free_swap(page);
292 		unlock_page(page);
293 	}
294 }
295 
296 /*
297  * Perform a free_page(), also freeing any swap cache associated with
298  * this page if it is the last user of the page.
299  */
300 void free_page_and_swap_cache(struct page *page)
301 {
302 	free_swap_cache(page);
303 	if (!is_huge_zero_page(page))
304 		put_page(page);
305 }
306 
307 /*
308  * Passed an array of pages, drop them all from swapcache and then release
309  * them.  They are removed from the LRU and freed if this is their last use.
310  */
311 void free_pages_and_swap_cache(struct page **pages, int nr)
312 {
313 	struct page **pagep = pages;
314 	int i;
315 
316 	lru_add_drain();
317 	for (i = 0; i < nr; i++)
318 		free_swap_cache(pagep[i]);
319 	release_pages(pagep, nr);
320 }
321 
322 static inline bool swap_use_vma_readahead(void)
323 {
324 	return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
325 }
326 
327 /*
328  * Lookup a swap entry in the swap cache. A found page will be returned
329  * unlocked and with its refcount incremented - we rely on the kernel
330  * lock getting page table operations atomic even if we drop the page
331  * lock before returning.
332  */
333 struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
334 			       unsigned long addr)
335 {
336 	struct page *page;
337 
338 	page = find_get_page(swap_address_space(entry), swp_offset(entry));
339 
340 	INC_CACHE_INFO(find_total);
341 	if (page) {
342 		bool vma_ra = swap_use_vma_readahead();
343 		bool readahead;
344 
345 		INC_CACHE_INFO(find_success);
346 		/*
347 		 * At the moment, we don't support PG_readahead for anon THP
348 		 * so let's bail out rather than confusing the readahead stat.
349 		 */
350 		if (unlikely(PageTransCompound(page)))
351 			return page;
352 
353 		readahead = TestClearPageReadahead(page);
354 		if (vma && vma_ra) {
355 			unsigned long ra_val;
356 			int win, hits;
357 
358 			ra_val = GET_SWAP_RA_VAL(vma);
359 			win = SWAP_RA_WIN(ra_val);
360 			hits = SWAP_RA_HITS(ra_val);
361 			if (readahead)
362 				hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
363 			atomic_long_set(&vma->swap_readahead_info,
364 					SWAP_RA_VAL(addr, win, hits));
365 		}
366 
367 		if (readahead) {
368 			count_vm_event(SWAP_RA_HIT);
369 			if (!vma || !vma_ra)
370 				atomic_inc(&swapin_readahead_hits);
371 		}
372 	}
373 
374 	return page;
375 }
376 
377 struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
378 			struct vm_area_struct *vma, unsigned long addr,
379 			bool *new_page_allocated)
380 {
381 	struct page *found_page, *new_page = NULL;
382 	struct address_space *swapper_space = swap_address_space(entry);
383 	int err;
384 	*new_page_allocated = false;
385 
386 	do {
387 		/*
388 		 * First check the swap cache.  Since this is normally
389 		 * called after lookup_swap_cache() failed, re-calling
390 		 * that would confuse statistics.
391 		 */
392 		found_page = find_get_page(swapper_space, swp_offset(entry));
393 		if (found_page)
394 			break;
395 
396 		/*
397 		 * Just skip read ahead for unused swap slot.
398 		 * During swap_off when swap_slot_cache is disabled,
399 		 * we have to handle the race between putting
400 		 * swap entry in swap cache and marking swap slot
401 		 * as SWAP_HAS_CACHE.  That's done in later part of code or
402 		 * else swap_off will be aborted if we return NULL.
403 		 */
404 		if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
405 			break;
406 
407 		/*
408 		 * Get a new page to read into from swap.
409 		 */
410 		if (!new_page) {
411 			new_page = alloc_page_vma(gfp_mask, vma, addr);
412 			if (!new_page)
413 				break;		/* Out of memory */
414 		}
415 
416 		/*
417 		 * call radix_tree_preload() while we can wait.
418 		 */
419 		err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
420 		if (err)
421 			break;
422 
423 		/*
424 		 * Swap entry may have been freed since our caller observed it.
425 		 */
426 		err = swapcache_prepare(entry);
427 		if (err == -EEXIST) {
428 			radix_tree_preload_end();
429 			/*
430 			 * We might race against get_swap_page() and stumble
431 			 * across a SWAP_HAS_CACHE swap_map entry whose page
432 			 * has not been brought into the swapcache yet.
433 			 */
434 			cond_resched();
435 			continue;
436 		}
437 		if (err) {		/* swp entry is obsolete ? */
438 			radix_tree_preload_end();
439 			break;
440 		}
441 
442 		/* May fail (-ENOMEM) if radix-tree node allocation failed. */
443 		__SetPageLocked(new_page);
444 		__SetPageSwapBacked(new_page);
445 		err = __add_to_swap_cache(new_page, entry);
446 		if (likely(!err)) {
447 			radix_tree_preload_end();
448 			/*
449 			 * Initiate read into locked page and return.
450 			 */
451 			SetPageWorkingset(new_page);
452 			lru_cache_add_anon(new_page);
453 			*new_page_allocated = true;
454 			return new_page;
455 		}
456 		radix_tree_preload_end();
457 		__ClearPageLocked(new_page);
458 		/*
459 		 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
460 		 * clear SWAP_HAS_CACHE flag.
461 		 */
462 		put_swap_page(new_page, entry);
463 	} while (err != -ENOMEM);
464 
465 	if (new_page)
466 		put_page(new_page);
467 	return found_page;
468 }
469 
470 /*
471  * Locate a page of swap in physical memory, reserving swap cache space
472  * and reading the disk if it is not already cached.
473  * A failure return means that either the page allocation failed or that
474  * the swap entry is no longer in use.
475  */
476 struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
477 		struct vm_area_struct *vma, unsigned long addr, bool do_poll)
478 {
479 	bool page_was_allocated;
480 	struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
481 			vma, addr, &page_was_allocated);
482 
483 	if (page_was_allocated)
484 		swap_readpage(retpage, do_poll);
485 
486 	return retpage;
487 }
488 
489 static unsigned int __swapin_nr_pages(unsigned long prev_offset,
490 				      unsigned long offset,
491 				      int hits,
492 				      int max_pages,
493 				      int prev_win)
494 {
495 	unsigned int pages, last_ra;
496 
497 	/*
498 	 * This heuristic has been found to work well on both sequential and
499 	 * random loads, swapping to hard disk or to SSD: please don't ask
500 	 * what the "+ 2" means, it just happens to work well, that's all.
501 	 */
502 	pages = hits + 2;
503 	if (pages == 2) {
504 		/*
505 		 * We can have no readahead hits to judge by: but must not get
506 		 * stuck here forever, so check for an adjacent offset instead
507 		 * (and don't even bother to check whether swap type is same).
508 		 */
509 		if (offset != prev_offset + 1 && offset != prev_offset - 1)
510 			pages = 1;
511 	} else {
512 		unsigned int roundup = 4;
513 		while (roundup < pages)
514 			roundup <<= 1;
515 		pages = roundup;
516 	}
517 
518 	if (pages > max_pages)
519 		pages = max_pages;
520 
521 	/* Don't shrink readahead too fast */
522 	last_ra = prev_win / 2;
523 	if (pages < last_ra)
524 		pages = last_ra;
525 
526 	return pages;
527 }
528 
529 static unsigned long swapin_nr_pages(unsigned long offset)
530 {
531 	static unsigned long prev_offset;
532 	unsigned int hits, pages, max_pages;
533 	static atomic_t last_readahead_pages;
534 
535 	max_pages = 1 << READ_ONCE(page_cluster);
536 	if (max_pages <= 1)
537 		return 1;
538 
539 	hits = atomic_xchg(&swapin_readahead_hits, 0);
540 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
541 				  atomic_read(&last_readahead_pages));
542 	if (!hits)
543 		prev_offset = offset;
544 	atomic_set(&last_readahead_pages, pages);
545 
546 	return pages;
547 }
548 
549 /**
550  * swap_cluster_readahead - swap in pages in hope we need them soon
551  * @entry: swap entry of this memory
552  * @gfp_mask: memory allocation flags
553  * @vmf: fault information
554  *
555  * Returns the struct page for entry and addr, after queueing swapin.
556  *
557  * Primitive swap readahead code. We simply read an aligned block of
558  * (1 << page_cluster) entries in the swap area. This method is chosen
559  * because it doesn't cost us any seek time.  We also make sure to queue
560  * the 'original' request together with the readahead ones...
561  *
562  * This has been extended to use the NUMA policies from the mm triggering
563  * the readahead.
564  *
565  * Caller must hold down_read on the vma->vm_mm if vmf->vma is not NULL.
566  */
567 struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
568 				struct vm_fault *vmf)
569 {
570 	struct page *page;
571 	unsigned long entry_offset = swp_offset(entry);
572 	unsigned long offset = entry_offset;
573 	unsigned long start_offset, end_offset;
574 	unsigned long mask;
575 	struct swap_info_struct *si = swp_swap_info(entry);
576 	struct blk_plug plug;
577 	bool do_poll = true, page_allocated;
578 	struct vm_area_struct *vma = vmf->vma;
579 	unsigned long addr = vmf->address;
580 
581 	mask = swapin_nr_pages(offset) - 1;
582 	if (!mask)
583 		goto skip;
584 
585 	do_poll = false;
586 	/* Read a page_cluster sized and aligned cluster around offset. */
587 	start_offset = offset & ~mask;
588 	end_offset = offset | mask;
589 	if (!start_offset)	/* First page is swap header. */
590 		start_offset++;
591 	if (end_offset >= si->max)
592 		end_offset = si->max - 1;
593 
594 	blk_start_plug(&plug);
595 	for (offset = start_offset; offset <= end_offset ; offset++) {
596 		/* Ok, do the async read-ahead now */
597 		page = __read_swap_cache_async(
598 			swp_entry(swp_type(entry), offset),
599 			gfp_mask, vma, addr, &page_allocated);
600 		if (!page)
601 			continue;
602 		if (page_allocated) {
603 			swap_readpage(page, false);
604 			if (offset != entry_offset) {
605 				SetPageReadahead(page);
606 				count_vm_event(SWAP_RA);
607 			}
608 		}
609 		put_page(page);
610 	}
611 	blk_finish_plug(&plug);
612 
613 	lru_add_drain();	/* Push any new pages onto the LRU now */
614 skip:
615 	return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
616 }
617 
618 int init_swap_address_space(unsigned int type, unsigned long nr_pages)
619 {
620 	struct address_space *spaces, *space;
621 	unsigned int i, nr;
622 
623 	nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
624 	spaces = kvcalloc(nr, sizeof(struct address_space), GFP_KERNEL);
625 	if (!spaces)
626 		return -ENOMEM;
627 	for (i = 0; i < nr; i++) {
628 		space = spaces + i;
629 		INIT_RADIX_TREE(&space->i_pages, GFP_ATOMIC|__GFP_NOWARN);
630 		atomic_set(&space->i_mmap_writable, 0);
631 		space->a_ops = &swap_aops;
632 		/* swap cache doesn't use writeback related tags */
633 		mapping_set_no_writeback_tags(space);
634 	}
635 	nr_swapper_spaces[type] = nr;
636 	rcu_assign_pointer(swapper_spaces[type], spaces);
637 
638 	return 0;
639 }
640 
641 void exit_swap_address_space(unsigned int type)
642 {
643 	struct address_space *spaces;
644 
645 	spaces = swapper_spaces[type];
646 	nr_swapper_spaces[type] = 0;
647 	rcu_assign_pointer(swapper_spaces[type], NULL);
648 	synchronize_rcu();
649 	kvfree(spaces);
650 }
651 
652 static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
653 				     unsigned long faddr,
654 				     unsigned long lpfn,
655 				     unsigned long rpfn,
656 				     unsigned long *start,
657 				     unsigned long *end)
658 {
659 	*start = max3(lpfn, PFN_DOWN(vma->vm_start),
660 		      PFN_DOWN(faddr & PMD_MASK));
661 	*end = min3(rpfn, PFN_DOWN(vma->vm_end),
662 		    PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
663 }
664 
665 static void swap_ra_info(struct vm_fault *vmf,
666 			struct vma_swap_readahead *ra_info)
667 {
668 	struct vm_area_struct *vma = vmf->vma;
669 	unsigned long ra_val;
670 	swp_entry_t entry;
671 	unsigned long faddr, pfn, fpfn;
672 	unsigned long start, end;
673 	pte_t *pte, *orig_pte;
674 	unsigned int max_win, hits, prev_win, win, left;
675 #ifndef CONFIG_64BIT
676 	pte_t *tpte;
677 #endif
678 
679 	max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
680 			     SWAP_RA_ORDER_CEILING);
681 	if (max_win == 1) {
682 		ra_info->win = 1;
683 		return;
684 	}
685 
686 	faddr = vmf->address;
687 	orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
688 	entry = pte_to_swp_entry(*pte);
689 	if ((unlikely(non_swap_entry(entry)))) {
690 		pte_unmap(orig_pte);
691 		return;
692 	}
693 
694 	fpfn = PFN_DOWN(faddr);
695 	ra_val = GET_SWAP_RA_VAL(vma);
696 	pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
697 	prev_win = SWAP_RA_WIN(ra_val);
698 	hits = SWAP_RA_HITS(ra_val);
699 	ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
700 					       max_win, prev_win);
701 	atomic_long_set(&vma->swap_readahead_info,
702 			SWAP_RA_VAL(faddr, win, 0));
703 
704 	if (win == 1) {
705 		pte_unmap(orig_pte);
706 		return;
707 	}
708 
709 	/* Copy the PTEs because the page table may be unmapped */
710 	if (fpfn == pfn + 1)
711 		swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
712 	else if (pfn == fpfn + 1)
713 		swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
714 				  &start, &end);
715 	else {
716 		left = (win - 1) / 2;
717 		swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
718 				  &start, &end);
719 	}
720 	ra_info->nr_pte = end - start;
721 	ra_info->offset = fpfn - start;
722 	pte -= ra_info->offset;
723 #ifdef CONFIG_64BIT
724 	ra_info->ptes = pte;
725 #else
726 	tpte = ra_info->ptes;
727 	for (pfn = start; pfn != end; pfn++)
728 		*tpte++ = *pte++;
729 #endif
730 	pte_unmap(orig_pte);
731 }
732 
733 static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
734 				       struct vm_fault *vmf)
735 {
736 	struct blk_plug plug;
737 	struct vm_area_struct *vma = vmf->vma;
738 	struct page *page;
739 	pte_t *pte, pentry;
740 	swp_entry_t entry;
741 	unsigned int i;
742 	bool page_allocated;
743 	struct vma_swap_readahead ra_info = {0,};
744 
745 	swap_ra_info(vmf, &ra_info);
746 	if (ra_info.win == 1)
747 		goto skip;
748 
749 	blk_start_plug(&plug);
750 	for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
751 	     i++, pte++) {
752 		pentry = *pte;
753 		if (pte_none(pentry))
754 			continue;
755 		if (pte_present(pentry))
756 			continue;
757 		entry = pte_to_swp_entry(pentry);
758 		if (unlikely(non_swap_entry(entry)))
759 			continue;
760 		page = __read_swap_cache_async(entry, gfp_mask, vma,
761 					       vmf->address, &page_allocated);
762 		if (!page)
763 			continue;
764 		if (page_allocated) {
765 			swap_readpage(page, false);
766 			if (i != ra_info.offset) {
767 				SetPageReadahead(page);
768 				count_vm_event(SWAP_RA);
769 			}
770 		}
771 		put_page(page);
772 	}
773 	blk_finish_plug(&plug);
774 	lru_add_drain();
775 skip:
776 	return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
777 				     ra_info.win == 1);
778 }
779 
780 /**
781  * swapin_readahead - swap in pages in hope we need them soon
782  * @entry: swap entry of this memory
783  * @gfp_mask: memory allocation flags
784  * @vmf: fault information
785  *
786  * Returns the struct page for entry and addr, after queueing swapin.
787  *
788  * It's a main entry function for swap readahead. By the configuration,
789  * it will read ahead blocks by cluster-based(ie, physical disk based)
790  * or vma-based(ie, virtual address based on faulty address) readahead.
791  */
792 struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
793 				struct vm_fault *vmf)
794 {
795 	return swap_use_vma_readahead() ?
796 			swap_vma_readahead(entry, gfp_mask, vmf) :
797 			swap_cluster_readahead(entry, gfp_mask, vmf);
798 }
799 
800 #ifdef CONFIG_SYSFS
801 static ssize_t vma_ra_enabled_show(struct kobject *kobj,
802 				     struct kobj_attribute *attr, char *buf)
803 {
804 	return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
805 }
806 static ssize_t vma_ra_enabled_store(struct kobject *kobj,
807 				      struct kobj_attribute *attr,
808 				      const char *buf, size_t count)
809 {
810 	if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
811 		enable_vma_readahead = true;
812 	else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
813 		enable_vma_readahead = false;
814 	else
815 		return -EINVAL;
816 
817 	return count;
818 }
819 static struct kobj_attribute vma_ra_enabled_attr =
820 	__ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
821 	       vma_ra_enabled_store);
822 
823 static struct attribute *swap_attrs[] = {
824 	&vma_ra_enabled_attr.attr,
825 	NULL,
826 };
827 
828 static struct attribute_group swap_attr_group = {
829 	.attrs = swap_attrs,
830 };
831 
832 static int __init swap_init_sysfs(void)
833 {
834 	int err;
835 	struct kobject *swap_kobj;
836 
837 	swap_kobj = kobject_create_and_add("swap", mm_kobj);
838 	if (!swap_kobj) {
839 		pr_err("failed to create swap kobject\n");
840 		return -ENOMEM;
841 	}
842 	err = sysfs_create_group(swap_kobj, &swap_attr_group);
843 	if (err) {
844 		pr_err("failed to register swap group\n");
845 		goto delete_obj;
846 	}
847 	return 0;
848 
849 delete_obj:
850 	kobject_put(swap_kobj);
851 	return err;
852 }
853 subsys_initcall(swap_init_sysfs);
854 #endif
855