1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95,
8 * Asynchronous swapping added 30.12.95. Stephen Tweedie
9 * Removed race in async swapping. 14.4.1996. Bruno Haible
10 * Add swap of shared pages through the page cache. 20.2.1998. Stephen Tweedie
11 * Always use brw_page, life becomes simpler. 12 May 1998 Eric Biederman
15 #include <linux/kernel_stat.h>
16 #include <linux/gfp.h>
17 #include <linux/pagemap.h>
18 #include <linux/swap.h>
19 #include <linux/bio.h>
20 #include <linux/swapops.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/psi.h>
24 #include <linux/uio.h>
25 #include <linux/sched/task.h>
26 #include <linux/delayacct.h>
27 #include <linux/zswap.h>
30 static void __end_swap_bio_write(struct bio *bio)
32 struct folio *folio = bio_first_folio_all(bio);
36 * We failed to write the page out to swap-space.
37 * Re-dirty the page in order to avoid it being reclaimed.
38 * Also print a dire warning that things will go BAD (tm)
41 * Also clear PG_reclaim to avoid folio_rotate_reclaimable()
43 folio_mark_dirty(folio);
44 pr_alert_ratelimited("Write-error on swap-device (%u:%u:%llu)\n",
45 MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
46 (unsigned long long)bio->bi_iter.bi_sector);
47 folio_clear_reclaim(folio);
49 folio_end_writeback(folio);
52 static void end_swap_bio_write(struct bio *bio)
54 __end_swap_bio_write(bio);
58 static void __end_swap_bio_read(struct bio *bio)
60 struct folio *folio = bio_first_folio_all(bio);
63 pr_alert_ratelimited("Read-error on swap-device (%u:%u:%llu)\n",
64 MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
65 (unsigned long long)bio->bi_iter.bi_sector);
67 folio_mark_uptodate(folio);
72 static void end_swap_bio_read(struct bio *bio)
74 __end_swap_bio_read(bio);
78 int generic_swapfile_activate(struct swap_info_struct *sis,
79 struct file *swap_file,
82 struct address_space *mapping = swap_file->f_mapping;
83 struct inode *inode = mapping->host;
84 unsigned blocks_per_page;
85 unsigned long page_no;
89 sector_t lowest_block = -1;
90 sector_t highest_block = 0;
94 blkbits = inode->i_blkbits;
95 blocks_per_page = PAGE_SIZE >> blkbits;
98 * Map all the blocks into the extent tree. This code doesn't try
103 last_block = i_size_read(inode) >> blkbits;
104 while ((probe_block + blocks_per_page) <= last_block &&
105 page_no < sis->max) {
106 unsigned block_in_page;
107 sector_t first_block;
111 first_block = probe_block;
112 ret = bmap(inode, &first_block);
113 if (ret || !first_block)
117 * It must be PAGE_SIZE aligned on-disk
119 if (first_block & (blocks_per_page - 1)) {
124 for (block_in_page = 1; block_in_page < blocks_per_page;
128 block = probe_block + block_in_page;
129 ret = bmap(inode, &block);
133 if (block != first_block + block_in_page) {
140 first_block >>= (PAGE_SHIFT - blkbits);
141 if (page_no) { /* exclude the header page */
142 if (first_block < lowest_block)
143 lowest_block = first_block;
144 if (first_block > highest_block)
145 highest_block = first_block;
149 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
151 ret = add_swap_extent(sis, page_no, 1, first_block);
156 probe_block += blocks_per_page;
161 *span = 1 + highest_block - lowest_block;
163 page_no = 1; /* force Empty message */
165 sis->pages = page_no - 1;
166 sis->highest_bit = page_no - 1;
170 pr_err("swapon: swapfile has holes\n");
175 static bool is_folio_zero_filled(struct folio *folio)
177 unsigned int pos, last_pos;
181 last_pos = PAGE_SIZE / sizeof(*data) - 1;
182 for (i = 0; i < folio_nr_pages(folio); i++) {
183 data = kmap_local_folio(folio, i * PAGE_SIZE);
185 * Check last word first, incase the page is zero-filled at
186 * the start and has non-zero data at the end, which is common
187 * in real-world workloads.
189 if (data[last_pos]) {
193 for (pos = 0; pos < last_pos; pos++) {
205 static void swap_zeromap_folio_set(struct folio *folio)
207 struct swap_info_struct *sis = swp_swap_info(folio->swap);
211 for (i = 0; i < folio_nr_pages(folio); i++) {
212 entry = page_swap_entry(folio_page(folio, i));
213 set_bit(swp_offset(entry), sis->zeromap);
217 static void swap_zeromap_folio_clear(struct folio *folio)
219 struct swap_info_struct *sis = swp_swap_info(folio->swap);
223 for (i = 0; i < folio_nr_pages(folio); i++) {
224 entry = page_swap_entry(folio_page(folio, i));
225 clear_bit(swp_offset(entry), sis->zeromap);
230 * We may have stale swap cache pages in memory: notice
231 * them here and get rid of the unnecessary final write.
233 int swap_writepage(struct page *page, struct writeback_control *wbc)
235 struct folio *folio = page_folio(page);
238 if (folio_free_swap(folio)) {
243 * Arch code may have to preserve more data than just the page
244 * contents, e.g. memory tags.
246 ret = arch_prepare_to_swap(folio);
248 folio_mark_dirty(folio);
254 * Use a bitmap (zeromap) to avoid doing IO for zero-filled pages.
255 * The bits in zeromap are protected by the locked swapcache folio
256 * and atomic updates are used to protect against read-modify-write
257 * corruption due to other zero swap entries seeing concurrent updates.
259 if (is_folio_zero_filled(folio)) {
260 swap_zeromap_folio_set(folio);
265 * Clear bits this folio occupies in the zeromap to prevent
266 * zero data being read in from any previous zero writes that
267 * occupied the same swap entries.
269 swap_zeromap_folio_clear(folio);
271 if (zswap_store(folio)) {
275 if (!mem_cgroup_zswap_writeback_enabled(folio_memcg(folio))) {
276 folio_mark_dirty(folio);
277 return AOP_WRITEPAGE_ACTIVATE;
280 __swap_writepage(folio, wbc);
284 static inline void count_swpout_vm_event(struct folio *folio)
286 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
287 if (unlikely(folio_test_pmd_mappable(folio))) {
288 count_memcg_folio_events(folio, THP_SWPOUT, 1);
289 count_vm_event(THP_SWPOUT);
291 count_mthp_stat(folio_order(folio), MTHP_STAT_SWPOUT);
293 count_vm_events(PSWPOUT, folio_nr_pages(folio));
296 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
297 static void bio_associate_blkg_from_page(struct bio *bio, struct folio *folio)
299 struct cgroup_subsys_state *css;
300 struct mem_cgroup *memcg;
302 memcg = folio_memcg(folio);
307 css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
308 bio_associate_blkg_from_css(bio, css);
312 #define bio_associate_blkg_from_page(bio, folio) do { } while (0)
313 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
317 struct bio_vec bvec[SWAP_CLUSTER_MAX];
321 static mempool_t *sio_pool;
323 int sio_pool_init(void)
326 mempool_t *pool = mempool_create_kmalloc_pool(
327 SWAP_CLUSTER_MAX, sizeof(struct swap_iocb));
328 if (cmpxchg(&sio_pool, NULL, pool))
329 mempool_destroy(pool);
336 static void sio_write_complete(struct kiocb *iocb, long ret)
338 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
339 struct page *page = sio->bvec[0].bv_page;
342 if (ret != sio->len) {
344 * In the case of swap-over-nfs, this can be a
345 * temporary failure if the system has limited
346 * memory for allocating transmit buffers.
347 * Mark the page dirty and avoid
348 * folio_rotate_reclaimable but rate-limit the
351 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n",
352 ret, swap_dev_pos(page_swap_entry(page)));
353 for (p = 0; p < sio->pages; p++) {
354 page = sio->bvec[p].bv_page;
355 set_page_dirty(page);
356 ClearPageReclaim(page);
360 for (p = 0; p < sio->pages; p++)
361 end_page_writeback(sio->bvec[p].bv_page);
363 mempool_free(sio, sio_pool);
366 static void swap_writepage_fs(struct folio *folio, struct writeback_control *wbc)
368 struct swap_iocb *sio = NULL;
369 struct swap_info_struct *sis = swp_swap_info(folio->swap);
370 struct file *swap_file = sis->swap_file;
371 loff_t pos = swap_dev_pos(folio->swap);
373 count_swpout_vm_event(folio);
374 folio_start_writeback(folio);
377 sio = *wbc->swap_plug;
379 if (sio->iocb.ki_filp != swap_file ||
380 sio->iocb.ki_pos + sio->len != pos) {
381 swap_write_unplug(sio);
386 sio = mempool_alloc(sio_pool, GFP_NOIO);
387 init_sync_kiocb(&sio->iocb, swap_file);
388 sio->iocb.ki_complete = sio_write_complete;
389 sio->iocb.ki_pos = pos;
393 bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
394 sio->len += folio_size(folio);
396 if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) {
397 swap_write_unplug(sio);
401 *wbc->swap_plug = sio;
404 static void swap_writepage_bdev_sync(struct folio *folio,
405 struct writeback_control *wbc, struct swap_info_struct *sis)
410 bio_init(&bio, sis->bdev, &bv, 1,
411 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc));
412 bio.bi_iter.bi_sector = swap_folio_sector(folio);
413 bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
415 bio_associate_blkg_from_page(&bio, folio);
416 count_swpout_vm_event(folio);
418 folio_start_writeback(folio);
421 submit_bio_wait(&bio);
422 __end_swap_bio_write(&bio);
425 static void swap_writepage_bdev_async(struct folio *folio,
426 struct writeback_control *wbc, struct swap_info_struct *sis)
430 bio = bio_alloc(sis->bdev, 1,
431 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
433 bio->bi_iter.bi_sector = swap_folio_sector(folio);
434 bio->bi_end_io = end_swap_bio_write;
435 bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
437 bio_associate_blkg_from_page(bio, folio);
438 count_swpout_vm_event(folio);
439 folio_start_writeback(folio);
444 void __swap_writepage(struct folio *folio, struct writeback_control *wbc)
446 struct swap_info_struct *sis = swp_swap_info(folio->swap);
448 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
450 * ->flags can be updated non-atomicially (scan_swap_map_slots),
451 * but that will never affect SWP_FS_OPS, so the data_race
454 if (data_race(sis->flags & SWP_FS_OPS))
455 swap_writepage_fs(folio, wbc);
457 * ->flags can be updated non-atomicially (scan_swap_map_slots),
458 * but that will never affect SWP_SYNCHRONOUS_IO, so the data_race
461 else if (data_race(sis->flags & SWP_SYNCHRONOUS_IO))
462 swap_writepage_bdev_sync(folio, wbc, sis);
464 swap_writepage_bdev_async(folio, wbc, sis);
467 void swap_write_unplug(struct swap_iocb *sio)
469 struct iov_iter from;
470 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
473 iov_iter_bvec(&from, ITER_SOURCE, sio->bvec, sio->pages, sio->len);
474 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
475 if (ret != -EIOCBQUEUED)
476 sio_write_complete(&sio->iocb, ret);
479 static void sio_read_complete(struct kiocb *iocb, long ret)
481 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
484 if (ret == sio->len) {
485 for (p = 0; p < sio->pages; p++) {
486 struct folio *folio = page_folio(sio->bvec[p].bv_page);
488 folio_mark_uptodate(folio);
491 count_vm_events(PSWPIN, sio->pages);
493 for (p = 0; p < sio->pages; p++) {
494 struct folio *folio = page_folio(sio->bvec[p].bv_page);
498 pr_alert_ratelimited("Read-error on swap-device\n");
500 mempool_free(sio, sio_pool);
503 static bool swap_read_folio_zeromap(struct folio *folio)
505 int nr_pages = folio_nr_pages(folio);
509 * Swapping in a large folio that is partially in the zeromap is not
510 * currently handled. Return true without marking the folio uptodate so
511 * that an IO error is emitted (e.g. do_swap_page() will sigbus).
513 if (WARN_ON_ONCE(swap_zeromap_batch(folio->swap, nr_pages,
514 &is_zeromap) != nr_pages))
520 folio_zero_range(folio, 0, folio_size(folio));
521 folio_mark_uptodate(folio);
525 static void swap_read_folio_fs(struct folio *folio, struct swap_iocb **plug)
527 struct swap_info_struct *sis = swp_swap_info(folio->swap);
528 struct swap_iocb *sio = NULL;
529 loff_t pos = swap_dev_pos(folio->swap);
534 if (sio->iocb.ki_filp != sis->swap_file ||
535 sio->iocb.ki_pos + sio->len != pos) {
536 swap_read_unplug(sio);
541 sio = mempool_alloc(sio_pool, GFP_KERNEL);
542 init_sync_kiocb(&sio->iocb, sis->swap_file);
543 sio->iocb.ki_pos = pos;
544 sio->iocb.ki_complete = sio_read_complete;
548 bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
549 sio->len += folio_size(folio);
551 if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) {
552 swap_read_unplug(sio);
559 static void swap_read_folio_bdev_sync(struct folio *folio,
560 struct swap_info_struct *sis)
565 bio_init(&bio, sis->bdev, &bv, 1, REQ_OP_READ);
566 bio.bi_iter.bi_sector = swap_folio_sector(folio);
567 bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
569 * Keep this task valid during swap readpage because the oom killer may
570 * attempt to access it in the page fault retry time check.
572 get_task_struct(current);
573 count_vm_event(PSWPIN);
574 submit_bio_wait(&bio);
575 __end_swap_bio_read(&bio);
576 put_task_struct(current);
579 static void swap_read_folio_bdev_async(struct folio *folio,
580 struct swap_info_struct *sis)
584 bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL);
585 bio->bi_iter.bi_sector = swap_folio_sector(folio);
586 bio->bi_end_io = end_swap_bio_read;
587 bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
588 count_vm_event(PSWPIN);
592 void swap_read_folio(struct folio *folio, struct swap_iocb **plug)
594 struct swap_info_struct *sis = swp_swap_info(folio->swap);
595 bool synchronous = sis->flags & SWP_SYNCHRONOUS_IO;
596 bool workingset = folio_test_workingset(folio);
597 unsigned long pflags;
600 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio) && !synchronous, folio);
601 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
602 VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
605 * Count submission time as memory stall and delay. When the device
606 * is congested, or the submitting cgroup IO-throttled, submission
607 * can be a significant part of overall IO time.
610 delayacct_thrashing_start(&in_thrashing);
611 psi_memstall_enter(&pflags);
613 delayacct_swapin_start();
615 if (swap_read_folio_zeromap(folio)) {
618 } else if (zswap_load(folio)) {
623 /* We have to read from slower devices. Increase zswap protection. */
624 zswap_folio_swapin(folio);
626 if (data_race(sis->flags & SWP_FS_OPS)) {
627 swap_read_folio_fs(folio, plug);
628 } else if (synchronous) {
629 swap_read_folio_bdev_sync(folio, sis);
631 swap_read_folio_bdev_async(folio, sis);
636 delayacct_thrashing_end(&in_thrashing);
637 psi_memstall_leave(&pflags);
639 delayacct_swapin_end();
642 void __swap_read_unplug(struct swap_iocb *sio)
644 struct iov_iter from;
645 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
648 iov_iter_bvec(&from, ITER_DEST, sio->bvec, sio->pages, sio->len);
649 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
650 if (ret != -EIOCBQUEUED)
651 sio_read_complete(&sio->iocb, ret);