1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/page_io.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
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
12 */
13
14#include <linux/mm.h>
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>
28#include "swap.h"
29
30static void __end_swap_bio_write(struct bio *bio)
31{
32 struct folio *folio = bio_first_folio_all(bio);
33
34 if (bio->bi_status) {
35 /*
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)
39 * very quickly.
40 *
41 * Also clear PG_reclaim to avoid folio_rotate_reclaimable()
42 */
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);
48 }
49 folio_end_writeback(folio);
50}
51
52static void end_swap_bio_write(struct bio *bio)
53{
54 __end_swap_bio_write(bio);
55 bio_put(bio);
56}
57
58static void __end_swap_bio_read(struct bio *bio)
59{
60 struct folio *folio = bio_first_folio_all(bio);
61
62 if (bio->bi_status) {
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);
66 } else {
67 folio_mark_uptodate(folio);
68 }
69 folio_unlock(folio);
70}
71
72static void end_swap_bio_read(struct bio *bio)
73{
74 __end_swap_bio_read(bio);
75 bio_put(bio);
76}
77
78int generic_swapfile_activate(struct swap_info_struct *sis,
79 struct file *swap_file,
80 sector_t *span)
81{
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;
86 unsigned blkbits;
87 sector_t probe_block;
88 sector_t last_block;
89 sector_t lowest_block = -1;
90 sector_t highest_block = 0;
91 int nr_extents = 0;
92 int ret;
93
94 blkbits = inode->i_blkbits;
95 blocks_per_page = PAGE_SIZE >> blkbits;
96
97 /*
98 * Map all the blocks into the extent tree. This code doesn't try
99 * to be very smart.
100 */
101 probe_block = 0;
102 page_no = 0;
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;
108
109 cond_resched();
110
111 first_block = probe_block;
112 ret = bmap(inode, block: &first_block);
113 if (ret || !first_block)
114 goto bad_bmap;
115
116 /*
117 * It must be PAGE_SIZE aligned on-disk
118 */
119 if (first_block & (blocks_per_page - 1)) {
120 probe_block++;
121 goto reprobe;
122 }
123
124 for (block_in_page = 1; block_in_page < blocks_per_page;
125 block_in_page++) {
126 sector_t block;
127
128 block = probe_block + block_in_page;
129 ret = bmap(inode, block: &block);
130 if (ret || !block)
131 goto bad_bmap;
132
133 if (block != first_block + block_in_page) {
134 /* Discontiguity */
135 probe_block++;
136 goto reprobe;
137 }
138 }
139
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;
146 }
147
148 /*
149 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
150 */
151 ret = add_swap_extent(sis, start_page: page_no, nr_pages: 1, start_block: first_block);
152 if (ret < 0)
153 goto out;
154 nr_extents += ret;
155 page_no++;
156 probe_block += blocks_per_page;
157reprobe:
158 continue;
159 }
160 ret = nr_extents;
161 *span = 1 + highest_block - lowest_block;
162 if (page_no == 0)
163 page_no = 1; /* force Empty message */
164 sis->max = page_no;
165 sis->pages = page_no - 1;
166 sis->highest_bit = page_no - 1;
167out:
168 return ret;
169bad_bmap:
170 pr_err("swapon: swapfile has holes\n");
171 ret = -EINVAL;
172 goto out;
173}
174
175/*
176 * We may have stale swap cache pages in memory: notice
177 * them here and get rid of the unnecessary final write.
178 */
179int swap_writepage(struct page *page, struct writeback_control *wbc)
180{
181 struct folio *folio = page_folio(page);
182 int ret;
183
184 if (folio_free_swap(folio)) {
185 folio_unlock(folio);
186 return 0;
187 }
188 /*
189 * Arch code may have to preserve more data than just the page
190 * contents, e.g. memory tags.
191 */
192 ret = arch_prepare_to_swap(page: &folio->page);
193 if (ret) {
194 folio_mark_dirty(folio);
195 folio_unlock(folio);
196 return ret;
197 }
198 if (zswap_store(folio)) {
199 folio_start_writeback(folio);
200 folio_unlock(folio);
201 folio_end_writeback(folio);
202 return 0;
203 }
204 __swap_writepage(page: &folio->page, wbc);
205 return 0;
206}
207
208static inline void count_swpout_vm_event(struct folio *folio)
209{
210#ifdef CONFIG_TRANSPARENT_HUGEPAGE
211 if (unlikely(folio_test_pmd_mappable(folio))) {
212 count_memcg_folio_events(folio, idx: THP_SWPOUT, nr: 1);
213 count_vm_event(item: THP_SWPOUT);
214 }
215#endif
216 count_vm_events(item: PSWPOUT, delta: folio_nr_pages(folio));
217}
218
219#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
220static void bio_associate_blkg_from_page(struct bio *bio, struct folio *folio)
221{
222 struct cgroup_subsys_state *css;
223 struct mem_cgroup *memcg;
224
225 memcg = folio_memcg(folio);
226 if (!memcg)
227 return;
228
229 rcu_read_lock();
230 css = cgroup_e_css(cgroup: memcg->css.cgroup, ss: &io_cgrp_subsys);
231 bio_associate_blkg_from_css(bio, css);
232 rcu_read_unlock();
233}
234#else
235#define bio_associate_blkg_from_page(bio, folio) do { } while (0)
236#endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
237
238struct swap_iocb {
239 struct kiocb iocb;
240 struct bio_vec bvec[SWAP_CLUSTER_MAX];
241 int pages;
242 int len;
243};
244static mempool_t *sio_pool;
245
246int sio_pool_init(void)
247{
248 if (!sio_pool) {
249 mempool_t *pool = mempool_create_kmalloc_pool(
250 SWAP_CLUSTER_MAX, size: sizeof(struct swap_iocb));
251 if (cmpxchg(&sio_pool, NULL, pool))
252 mempool_destroy(pool);
253 }
254 if (!sio_pool)
255 return -ENOMEM;
256 return 0;
257}
258
259static void sio_write_complete(struct kiocb *iocb, long ret)
260{
261 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
262 struct page *page = sio->bvec[0].bv_page;
263 int p;
264
265 if (ret != sio->len) {
266 /*
267 * In the case of swap-over-nfs, this can be a
268 * temporary failure if the system has limited
269 * memory for allocating transmit buffers.
270 * Mark the page dirty and avoid
271 * folio_rotate_reclaimable but rate-limit the
272 * messages but do not flag PageError like
273 * the normal direct-to-bio case as it could
274 * be temporary.
275 */
276 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n",
277 ret, page_file_offset(page));
278 for (p = 0; p < sio->pages; p++) {
279 page = sio->bvec[p].bv_page;
280 set_page_dirty(page);
281 ClearPageReclaim(page);
282 }
283 }
284
285 for (p = 0; p < sio->pages; p++)
286 end_page_writeback(page: sio->bvec[p].bv_page);
287
288 mempool_free(element: sio, pool: sio_pool);
289}
290
291static void swap_writepage_fs(struct page *page, struct writeback_control *wbc)
292{
293 struct swap_iocb *sio = NULL;
294 struct swap_info_struct *sis = page_swap_info(page);
295 struct file *swap_file = sis->swap_file;
296 loff_t pos = page_file_offset(page);
297
298 count_swpout_vm_event(page_folio(page));
299 set_page_writeback(page);
300 unlock_page(page);
301 if (wbc->swap_plug)
302 sio = *wbc->swap_plug;
303 if (sio) {
304 if (sio->iocb.ki_filp != swap_file ||
305 sio->iocb.ki_pos + sio->len != pos) {
306 swap_write_unplug(sio);
307 sio = NULL;
308 }
309 }
310 if (!sio) {
311 sio = mempool_alloc(pool: sio_pool, GFP_NOIO);
312 init_sync_kiocb(kiocb: &sio->iocb, filp: swap_file);
313 sio->iocb.ki_complete = sio_write_complete;
314 sio->iocb.ki_pos = pos;
315 sio->pages = 0;
316 sio->len = 0;
317 }
318 bvec_set_page(bv: &sio->bvec[sio->pages], page, len: thp_size(page), offset: 0);
319 sio->len += thp_size(page);
320 sio->pages += 1;
321 if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) {
322 swap_write_unplug(sio);
323 sio = NULL;
324 }
325 if (wbc->swap_plug)
326 *wbc->swap_plug = sio;
327}
328
329static void swap_writepage_bdev_sync(struct page *page,
330 struct writeback_control *wbc, struct swap_info_struct *sis)
331{
332 struct bio_vec bv;
333 struct bio bio;
334 struct folio *folio = page_folio(page);
335
336 bio_init(bio: &bio, bdev: sis->bdev, table: &bv, max_vecs: 1,
337 opf: REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc));
338 bio.bi_iter.bi_sector = swap_page_sector(page);
339 __bio_add_page(bio: &bio, page, len: thp_size(page), off: 0);
340
341 bio_associate_blkg_from_page(bio: &bio, folio);
342 count_swpout_vm_event(folio);
343
344 folio_start_writeback(folio);
345 folio_unlock(folio);
346
347 submit_bio_wait(bio: &bio);
348 __end_swap_bio_write(bio: &bio);
349}
350
351static void swap_writepage_bdev_async(struct page *page,
352 struct writeback_control *wbc, struct swap_info_struct *sis)
353{
354 struct bio *bio;
355 struct folio *folio = page_folio(page);
356
357 bio = bio_alloc(bdev: sis->bdev, nr_vecs: 1,
358 opf: REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
359 GFP_NOIO);
360 bio->bi_iter.bi_sector = swap_page_sector(page);
361 bio->bi_end_io = end_swap_bio_write;
362 __bio_add_page(bio, page, len: thp_size(page), off: 0);
363
364 bio_associate_blkg_from_page(bio, folio);
365 count_swpout_vm_event(folio);
366 folio_start_writeback(folio);
367 folio_unlock(folio);
368 submit_bio(bio);
369}
370
371void __swap_writepage(struct page *page, struct writeback_control *wbc)
372{
373 struct swap_info_struct *sis = page_swap_info(page);
374
375 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
376 /*
377 * ->flags can be updated non-atomicially (scan_swap_map_slots),
378 * but that will never affect SWP_FS_OPS, so the data_race
379 * is safe.
380 */
381 if (data_race(sis->flags & SWP_FS_OPS))
382 swap_writepage_fs(page, wbc);
383 else if (sis->flags & SWP_SYNCHRONOUS_IO)
384 swap_writepage_bdev_sync(page, wbc, sis);
385 else
386 swap_writepage_bdev_async(page, wbc, sis);
387}
388
389void swap_write_unplug(struct swap_iocb *sio)
390{
391 struct iov_iter from;
392 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
393 int ret;
394
395 iov_iter_bvec(i: &from, ITER_SOURCE, bvec: sio->bvec, nr_segs: sio->pages, count: sio->len);
396 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
397 if (ret != -EIOCBQUEUED)
398 sio_write_complete(iocb: &sio->iocb, ret);
399}
400
401static void sio_read_complete(struct kiocb *iocb, long ret)
402{
403 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
404 int p;
405
406 if (ret == sio->len) {
407 for (p = 0; p < sio->pages; p++) {
408 struct folio *folio = page_folio(sio->bvec[p].bv_page);
409
410 folio_mark_uptodate(folio);
411 folio_unlock(folio);
412 }
413 count_vm_events(item: PSWPIN, delta: sio->pages);
414 } else {
415 for (p = 0; p < sio->pages; p++) {
416 struct folio *folio = page_folio(sio->bvec[p].bv_page);
417
418 folio_unlock(folio);
419 }
420 pr_alert_ratelimited("Read-error on swap-device\n");
421 }
422 mempool_free(element: sio, pool: sio_pool);
423}
424
425static void swap_readpage_fs(struct page *page,
426 struct swap_iocb **plug)
427{
428 struct swap_info_struct *sis = page_swap_info(page);
429 struct swap_iocb *sio = NULL;
430 loff_t pos = page_file_offset(page);
431
432 if (plug)
433 sio = *plug;
434 if (sio) {
435 if (sio->iocb.ki_filp != sis->swap_file ||
436 sio->iocb.ki_pos + sio->len != pos) {
437 swap_read_unplug(plug: sio);
438 sio = NULL;
439 }
440 }
441 if (!sio) {
442 sio = mempool_alloc(pool: sio_pool, GFP_KERNEL);
443 init_sync_kiocb(kiocb: &sio->iocb, filp: sis->swap_file);
444 sio->iocb.ki_pos = pos;
445 sio->iocb.ki_complete = sio_read_complete;
446 sio->pages = 0;
447 sio->len = 0;
448 }
449 bvec_set_page(bv: &sio->bvec[sio->pages], page, len: thp_size(page), offset: 0);
450 sio->len += thp_size(page);
451 sio->pages += 1;
452 if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) {
453 swap_read_unplug(plug: sio);
454 sio = NULL;
455 }
456 if (plug)
457 *plug = sio;
458}
459
460static void swap_readpage_bdev_sync(struct page *page,
461 struct swap_info_struct *sis)
462{
463 struct bio_vec bv;
464 struct bio bio;
465
466 bio_init(bio: &bio, bdev: sis->bdev, table: &bv, max_vecs: 1, opf: REQ_OP_READ);
467 bio.bi_iter.bi_sector = swap_page_sector(page);
468 __bio_add_page(bio: &bio, page, len: thp_size(page), off: 0);
469 /*
470 * Keep this task valid during swap readpage because the oom killer may
471 * attempt to access it in the page fault retry time check.
472 */
473 get_task_struct(current);
474 count_vm_event(item: PSWPIN);
475 submit_bio_wait(bio: &bio);
476 __end_swap_bio_read(bio: &bio);
477 put_task_struct(current);
478}
479
480static void swap_readpage_bdev_async(struct page *page,
481 struct swap_info_struct *sis)
482{
483 struct bio *bio;
484
485 bio = bio_alloc(bdev: sis->bdev, nr_vecs: 1, opf: REQ_OP_READ, GFP_KERNEL);
486 bio->bi_iter.bi_sector = swap_page_sector(page);
487 bio->bi_end_io = end_swap_bio_read;
488 __bio_add_page(bio, page, len: thp_size(page), off: 0);
489 count_vm_event(item: PSWPIN);
490 submit_bio(bio);
491}
492
493void swap_readpage(struct page *page, bool synchronous, struct swap_iocb **plug)
494{
495 struct folio *folio = page_folio(page);
496 struct swap_info_struct *sis = page_swap_info(page);
497 bool workingset = folio_test_workingset(folio);
498 unsigned long pflags;
499 bool in_thrashing;
500
501 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio) && !synchronous, folio);
502 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
503 VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
504
505 /*
506 * Count submission time as memory stall and delay. When the device
507 * is congested, or the submitting cgroup IO-throttled, submission
508 * can be a significant part of overall IO time.
509 */
510 if (workingset) {
511 delayacct_thrashing_start(in_thrashing: &in_thrashing);
512 psi_memstall_enter(flags: &pflags);
513 }
514 delayacct_swapin_start();
515
516 if (zswap_load(folio)) {
517 folio_mark_uptodate(folio);
518 folio_unlock(folio);
519 } else if (data_race(sis->flags & SWP_FS_OPS)) {
520 swap_readpage_fs(page, plug);
521 } else if (synchronous || (sis->flags & SWP_SYNCHRONOUS_IO)) {
522 swap_readpage_bdev_sync(page, sis);
523 } else {
524 swap_readpage_bdev_async(page, sis);
525 }
526
527 if (workingset) {
528 delayacct_thrashing_end(in_thrashing: &in_thrashing);
529 psi_memstall_leave(flags: &pflags);
530 }
531 delayacct_swapin_end();
532}
533
534void __swap_read_unplug(struct swap_iocb *sio)
535{
536 struct iov_iter from;
537 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
538 int ret;
539
540 iov_iter_bvec(i: &from, ITER_DEST, bvec: sio->bvec, nr_segs: sio->pages, count: sio->len);
541 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
542 if (ret != -EIOCBQUEUED)
543 sio_read_complete(iocb: &sio->iocb, ret);
544}
545

source code of linux/mm/page_io.c