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
30 static 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
52 static void end_swap_bio_write(struct bio *bio)
53 {
54 __end_swap_bio_write(bio);
55 bio_put(bio);
56 }
57
58 static 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
72 static void end_swap_bio_read(struct bio *bio)
73 {
74 __end_swap_bio_read(bio);
75 bio_put(bio);
76 }
77
78 int 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, &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);
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, page_no, 1, first_block);
152 if (ret < 0)
153 goto out;
154 nr_extents += ret;
155 page_no++;
156 probe_block += blocks_per_page;
157 reprobe:
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;
167 out:
168 return ret;
169 bad_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 */
179 int 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(folio);
193 if (ret) {
194 folio_mark_dirty(folio);
195 folio_unlock(folio);
196 return ret;
197 }
198 if (zswap_store(folio)) {
199 folio_unlock(folio);
200 return 0;
201 }
202 if (!mem_cgroup_zswap_writeback_enabled(folio_memcg(folio))) {
203 folio_mark_dirty(folio);
204 return AOP_WRITEPAGE_ACTIVATE;
205 }
206
207 __swap_writepage(folio, wbc);
208 return 0;
209 }
210
211 static inline void count_swpout_vm_event(struct folio *folio)
212 {
213 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
214 if (unlikely(folio_test_pmd_mappable(folio))) {
215 count_memcg_folio_events(folio, THP_SWPOUT, 1);
216 count_vm_event(THP_SWPOUT);
217 }
218 count_mthp_stat(folio_order(folio), MTHP_STAT_SWPOUT);
219 #endif
220 count_vm_events(PSWPOUT, folio_nr_pages(folio));
221 }
222
223 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
224 static void bio_associate_blkg_from_page(struct bio *bio, struct folio *folio)
225 {
226 struct cgroup_subsys_state *css;
227 struct mem_cgroup *memcg;
228
229 memcg = folio_memcg(folio);
230 if (!memcg)
231 return;
232
233 rcu_read_lock();
234 css = cgroup_e_css(memcg->css.cgroup, &io_cgrp_subsys);
235 bio_associate_blkg_from_css(bio, css);
236 rcu_read_unlock();
237 }
238 #else
239 #define bio_associate_blkg_from_page(bio, folio) do { } while (0)
240 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
241
242 struct swap_iocb {
243 struct kiocb iocb;
244 struct bio_vec bvec[SWAP_CLUSTER_MAX];
245 int pages;
246 int len;
247 };
248 static mempool_t *sio_pool;
249
250 int sio_pool_init(void)
251 {
252 if (!sio_pool) {
253 mempool_t *pool = mempool_create_kmalloc_pool(
254 SWAP_CLUSTER_MAX, sizeof(struct swap_iocb));
255 if (cmpxchg(&sio_pool, NULL, pool))
256 mempool_destroy(pool);
257 }
258 if (!sio_pool)
259 return -ENOMEM;
260 return 0;
261 }
262
263 static void sio_write_complete(struct kiocb *iocb, long ret)
264 {
265 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
266 struct page *page = sio->bvec[0].bv_page;
267 int p;
268
269 if (ret != sio->len) {
270 /*
271 * In the case of swap-over-nfs, this can be a
272 * temporary failure if the system has limited
273 * memory for allocating transmit buffers.
274 * Mark the page dirty and avoid
275 * folio_rotate_reclaimable but rate-limit the
276 * messages but do not flag PageError like
277 * the normal direct-to-bio case as it could
278 * be temporary.
279 */
280 pr_err_ratelimited("Write error %ld on dio swapfile (%llu)\n",
281 ret, swap_dev_pos(page_swap_entry(page)));
282 for (p = 0; p < sio->pages; p++) {
283 page = sio->bvec[p].bv_page;
284 set_page_dirty(page);
285 ClearPageReclaim(page);
286 }
287 }
288
289 for (p = 0; p < sio->pages; p++)
290 end_page_writeback(sio->bvec[p].bv_page);
291
292 mempool_free(sio, sio_pool);
293 }
294
295 static void swap_writepage_fs(struct folio *folio, struct writeback_control *wbc)
296 {
297 struct swap_iocb *sio = NULL;
298 struct swap_info_struct *sis = swp_swap_info(folio->swap);
299 struct file *swap_file = sis->swap_file;
300 loff_t pos = swap_dev_pos(folio->swap);
301
302 count_swpout_vm_event(folio);
303 folio_start_writeback(folio);
304 folio_unlock(folio);
305 if (wbc->swap_plug)
306 sio = *wbc->swap_plug;
307 if (sio) {
308 if (sio->iocb.ki_filp != swap_file ||
309 sio->iocb.ki_pos + sio->len != pos) {
310 swap_write_unplug(sio);
311 sio = NULL;
312 }
313 }
314 if (!sio) {
315 sio = mempool_alloc(sio_pool, GFP_NOIO);
316 init_sync_kiocb(&sio->iocb, swap_file);
317 sio->iocb.ki_complete = sio_write_complete;
318 sio->iocb.ki_pos = pos;
319 sio->pages = 0;
320 sio->len = 0;
321 }
322 bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
323 sio->len += folio_size(folio);
324 sio->pages += 1;
325 if (sio->pages == ARRAY_SIZE(sio->bvec) || !wbc->swap_plug) {
326 swap_write_unplug(sio);
327 sio = NULL;
328 }
329 if (wbc->swap_plug)
330 *wbc->swap_plug = sio;
331 }
332
333 static void swap_writepage_bdev_sync(struct folio *folio,
334 struct writeback_control *wbc, struct swap_info_struct *sis)
335 {
336 struct bio_vec bv;
337 struct bio bio;
338
339 bio_init(&bio, sis->bdev, &bv, 1,
340 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc));
341 bio.bi_iter.bi_sector = swap_folio_sector(folio);
342 bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
343
344 bio_associate_blkg_from_page(&bio, folio);
345 count_swpout_vm_event(folio);
346
347 folio_start_writeback(folio);
348 folio_unlock(folio);
349
350 submit_bio_wait(&bio);
351 __end_swap_bio_write(&bio);
352 }
353
354 static void swap_writepage_bdev_async(struct folio *folio,
355 struct writeback_control *wbc, struct swap_info_struct *sis)
356 {
357 struct bio *bio;
358
359 bio = bio_alloc(sis->bdev, 1,
360 REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc),
361 GFP_NOIO);
362 bio->bi_iter.bi_sector = swap_folio_sector(folio);
363 bio->bi_end_io = end_swap_bio_write;
364 bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
365
366 bio_associate_blkg_from_page(bio, folio);
367 count_swpout_vm_event(folio);
368 folio_start_writeback(folio);
369 folio_unlock(folio);
370 submit_bio(bio);
371 }
372
373 void __swap_writepage(struct folio *folio, struct writeback_control *wbc)
374 {
375 struct swap_info_struct *sis = swp_swap_info(folio->swap);
376
377 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio), folio);
378 /*
379 * ->flags can be updated non-atomicially (scan_swap_map_slots),
380 * but that will never affect SWP_FS_OPS, so the data_race
381 * is safe.
382 */
383 if (data_race(sis->flags & SWP_FS_OPS))
384 swap_writepage_fs(folio, wbc);
385 /*
386 * ->flags can be updated non-atomicially (scan_swap_map_slots),
387 * but that will never affect SWP_SYNCHRONOUS_IO, so the data_race
388 * is safe.
389 */
390 else if (data_race(sis->flags & SWP_SYNCHRONOUS_IO))
391 swap_writepage_bdev_sync(folio, wbc, sis);
392 else
393 swap_writepage_bdev_async(folio, wbc, sis);
394 }
395
396 void swap_write_unplug(struct swap_iocb *sio)
397 {
398 struct iov_iter from;
399 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
400 int ret;
401
402 iov_iter_bvec(&from, ITER_SOURCE, sio->bvec, sio->pages, sio->len);
403 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
404 if (ret != -EIOCBQUEUED)
405 sio_write_complete(&sio->iocb, ret);
406 }
407
408 static void sio_read_complete(struct kiocb *iocb, long ret)
409 {
410 struct swap_iocb *sio = container_of(iocb, struct swap_iocb, iocb);
411 int p;
412
413 if (ret == sio->len) {
414 for (p = 0; p < sio->pages; p++) {
415 struct folio *folio = page_folio(sio->bvec[p].bv_page);
416
417 folio_mark_uptodate(folio);
418 folio_unlock(folio);
419 }
420 count_vm_events(PSWPIN, sio->pages);
421 } else {
422 for (p = 0; p < sio->pages; p++) {
423 struct folio *folio = page_folio(sio->bvec[p].bv_page);
424
425 folio_unlock(folio);
426 }
427 pr_alert_ratelimited("Read-error on swap-device\n");
428 }
429 mempool_free(sio, sio_pool);
430 }
431
432 static void swap_read_folio_fs(struct folio *folio, struct swap_iocb **plug)
433 {
434 struct swap_info_struct *sis = swp_swap_info(folio->swap);
435 struct swap_iocb *sio = NULL;
436 loff_t pos = swap_dev_pos(folio->swap);
437
438 if (plug)
439 sio = *plug;
440 if (sio) {
441 if (sio->iocb.ki_filp != sis->swap_file ||
442 sio->iocb.ki_pos + sio->len != pos) {
443 swap_read_unplug(sio);
444 sio = NULL;
445 }
446 }
447 if (!sio) {
448 sio = mempool_alloc(sio_pool, GFP_KERNEL);
449 init_sync_kiocb(&sio->iocb, sis->swap_file);
450 sio->iocb.ki_pos = pos;
451 sio->iocb.ki_complete = sio_read_complete;
452 sio->pages = 0;
453 sio->len = 0;
454 }
455 bvec_set_folio(&sio->bvec[sio->pages], folio, folio_size(folio), 0);
456 sio->len += folio_size(folio);
457 sio->pages += 1;
458 if (sio->pages == ARRAY_SIZE(sio->bvec) || !plug) {
459 swap_read_unplug(sio);
460 sio = NULL;
461 }
462 if (plug)
463 *plug = sio;
464 }
465
466 static void swap_read_folio_bdev_sync(struct folio *folio,
467 struct swap_info_struct *sis)
468 {
469 struct bio_vec bv;
470 struct bio bio;
471
472 bio_init(&bio, sis->bdev, &bv, 1, REQ_OP_READ);
473 bio.bi_iter.bi_sector = swap_folio_sector(folio);
474 bio_add_folio_nofail(&bio, folio, folio_size(folio), 0);
475 /*
476 * Keep this task valid during swap readpage because the oom killer may
477 * attempt to access it in the page fault retry time check.
478 */
479 get_task_struct(current);
480 count_vm_event(PSWPIN);
481 submit_bio_wait(&bio);
482 __end_swap_bio_read(&bio);
483 put_task_struct(current);
484 }
485
486 static void swap_read_folio_bdev_async(struct folio *folio,
487 struct swap_info_struct *sis)
488 {
489 struct bio *bio;
490
491 bio = bio_alloc(sis->bdev, 1, REQ_OP_READ, GFP_KERNEL);
492 bio->bi_iter.bi_sector = swap_folio_sector(folio);
493 bio->bi_end_io = end_swap_bio_read;
494 bio_add_folio_nofail(bio, folio, folio_size(folio), 0);
495 count_vm_event(PSWPIN);
496 submit_bio(bio);
497 }
498
499 void swap_read_folio(struct folio *folio, struct swap_iocb **plug)
500 {
501 struct swap_info_struct *sis = swp_swap_info(folio->swap);
502 bool synchronous = sis->flags & SWP_SYNCHRONOUS_IO;
503 bool workingset = folio_test_workingset(folio);
504 unsigned long pflags;
505 bool in_thrashing;
506
507 VM_BUG_ON_FOLIO(!folio_test_swapcache(folio) && !synchronous, folio);
508 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
509 VM_BUG_ON_FOLIO(folio_test_uptodate(folio), folio);
510
511 /*
512 * Count submission time as memory stall and delay. When the device
513 * is congested, or the submitting cgroup IO-throttled, submission
514 * can be a significant part of overall IO time.
515 */
516 if (workingset) {
517 delayacct_thrashing_start(&in_thrashing);
518 psi_memstall_enter(&pflags);
519 }
520 delayacct_swapin_start();
521
522 if (zswap_load(folio)) {
523 folio_unlock(folio);
524 } else if (data_race(sis->flags & SWP_FS_OPS)) {
525 swap_read_folio_fs(folio, plug);
526 } else if (synchronous) {
527 swap_read_folio_bdev_sync(folio, sis);
528 } else {
529 swap_read_folio_bdev_async(folio, sis);
530 }
531
532 if (workingset) {
533 delayacct_thrashing_end(&in_thrashing);
534 psi_memstall_leave(&pflags);
535 }
536 delayacct_swapin_end();
537 }
538
539 void __swap_read_unplug(struct swap_iocb *sio)
540 {
541 struct iov_iter from;
542 struct address_space *mapping = sio->iocb.ki_filp->f_mapping;
543 int ret;
544
545 iov_iter_bvec(&from, ITER_DEST, sio->bvec, sio->pages, sio->len);
546 ret = mapping->a_ops->swap_rw(&sio->iocb, &from);
547 if (ret != -EIOCBQUEUED)
548 sio_read_complete(&sio->iocb, ret);
549 }
550
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.