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/buffer_head.h>
22 #include <linux/writeback.h>
23 #include <linux/frontswap.h>
24 #include <linux/blkdev.h>
25 #include <linux/psi.h>
26 #include <linux/uio.h>
27 #include <linux/sched/task.h>
28
get_swap_bio(gfp_t gfp_flags,struct page * page,bio_end_io_t end_io)29 static struct bio *get_swap_bio(gfp_t gfp_flags,
30 struct page *page, bio_end_io_t end_io)
31 {
32 struct bio *bio;
33
34 bio = bio_alloc(gfp_flags, 1);
35 if (bio) {
36 struct block_device *bdev;
37
38 bio->bi_iter.bi_sector = map_swap_page(page, &bdev);
39 bio_set_dev(bio, bdev);
40 bio->bi_iter.bi_sector <<= PAGE_SHIFT - 9;
41 bio->bi_end_io = end_io;
42
43 bio_add_page(bio, page, thp_size(page), 0);
44 }
45 return bio;
46 }
47
end_swap_bio_write(struct bio * bio)48 void end_swap_bio_write(struct bio *bio)
49 {
50 struct page *page = bio_first_page_all(bio);
51
52 if (bio->bi_status) {
53 SetPageError(page);
54 /*
55 * We failed to write the page out to swap-space.
56 * Re-dirty the page in order to avoid it being reclaimed.
57 * Also print a dire warning that things will go BAD (tm)
58 * very quickly.
59 *
60 * Also clear PG_reclaim to avoid rotate_reclaimable_page()
61 */
62 set_page_dirty(page);
63 pr_alert("Write-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 ClearPageReclaim(page);
67 }
68 end_page_writeback(page);
69 bio_put(bio);
70 }
71
swap_slot_free_notify(struct page * page)72 static void swap_slot_free_notify(struct page *page)
73 {
74 struct swap_info_struct *sis;
75 struct gendisk *disk;
76 swp_entry_t entry;
77
78 /*
79 * There is no guarantee that the page is in swap cache - the software
80 * suspend code (at least) uses end_swap_bio_read() against a non-
81 * swapcache page. So we must check PG_swapcache before proceeding with
82 * this optimization.
83 */
84 if (unlikely(!PageSwapCache(page)))
85 return;
86
87 sis = page_swap_info(page);
88 if (data_race(!(sis->flags & SWP_BLKDEV)))
89 return;
90
91 /*
92 * The swap subsystem performs lazy swap slot freeing,
93 * expecting that the page will be swapped out again.
94 * So we can avoid an unnecessary write if the page
95 * isn't redirtied.
96 * This is good for real swap storage because we can
97 * reduce unnecessary I/O and enhance wear-leveling
98 * if an SSD is used as the as swap device.
99 * But if in-memory swap device (eg zram) is used,
100 * this causes a duplicated copy between uncompressed
101 * data in VM-owned memory and compressed data in
102 * zram-owned memory. So let's free zram-owned memory
103 * and make the VM-owned decompressed page *dirty*,
104 * so the page should be swapped out somewhere again if
105 * we again wish to reclaim it.
106 */
107 disk = sis->bdev->bd_disk;
108 entry.val = page_private(page);
109 if (disk->fops->swap_slot_free_notify && __swap_count(entry) == 1) {
110 unsigned long offset;
111
112 offset = swp_offset(entry);
113
114 SetPageDirty(page);
115 disk->fops->swap_slot_free_notify(sis->bdev,
116 offset);
117 }
118 }
119
end_swap_bio_read(struct bio * bio)120 static void end_swap_bio_read(struct bio *bio)
121 {
122 struct page *page = bio_first_page_all(bio);
123 struct task_struct *waiter = bio->bi_private;
124
125 if (bio->bi_status) {
126 SetPageError(page);
127 ClearPageUptodate(page);
128 pr_alert("Read-error on swap-device (%u:%u:%llu)\n",
129 MAJOR(bio_dev(bio)), MINOR(bio_dev(bio)),
130 (unsigned long long)bio->bi_iter.bi_sector);
131 goto out;
132 }
133
134 SetPageUptodate(page);
135 swap_slot_free_notify(page);
136 out:
137 unlock_page(page);
138 WRITE_ONCE(bio->bi_private, NULL);
139 bio_put(bio);
140 if (waiter) {
141 blk_wake_io_task(waiter);
142 put_task_struct(waiter);
143 }
144 }
145
generic_swapfile_activate(struct swap_info_struct * sis,struct file * swap_file,sector_t * span)146 int generic_swapfile_activate(struct swap_info_struct *sis,
147 struct file *swap_file,
148 sector_t *span)
149 {
150 struct address_space *mapping = swap_file->f_mapping;
151 struct inode *inode = mapping->host;
152 unsigned blocks_per_page;
153 unsigned long page_no;
154 unsigned blkbits;
155 sector_t probe_block;
156 sector_t last_block;
157 sector_t lowest_block = -1;
158 sector_t highest_block = 0;
159 int nr_extents = 0;
160 int ret;
161
162 blkbits = inode->i_blkbits;
163 blocks_per_page = PAGE_SIZE >> blkbits;
164
165 /*
166 * Map all the blocks into the extent tree. This code doesn't try
167 * to be very smart.
168 */
169 probe_block = 0;
170 page_no = 0;
171 last_block = i_size_read(inode) >> blkbits;
172 while ((probe_block + blocks_per_page) <= last_block &&
173 page_no < sis->max) {
174 unsigned block_in_page;
175 sector_t first_block;
176
177 cond_resched();
178
179 first_block = probe_block;
180 ret = bmap(inode, &first_block);
181 if (ret || !first_block)
182 goto bad_bmap;
183
184 /*
185 * It must be PAGE_SIZE aligned on-disk
186 */
187 if (first_block & (blocks_per_page - 1)) {
188 probe_block++;
189 goto reprobe;
190 }
191
192 for (block_in_page = 1; block_in_page < blocks_per_page;
193 block_in_page++) {
194 sector_t block;
195
196 block = probe_block + block_in_page;
197 ret = bmap(inode, &block);
198 if (ret || !block)
199 goto bad_bmap;
200
201 if (block != first_block + block_in_page) {
202 /* Discontiguity */
203 probe_block++;
204 goto reprobe;
205 }
206 }
207
208 first_block >>= (PAGE_SHIFT - blkbits);
209 if (page_no) { /* exclude the header page */
210 if (first_block < lowest_block)
211 lowest_block = first_block;
212 if (first_block > highest_block)
213 highest_block = first_block;
214 }
215
216 /*
217 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
218 */
219 ret = add_swap_extent(sis, page_no, 1, first_block);
220 if (ret < 0)
221 goto out;
222 nr_extents += ret;
223 page_no++;
224 probe_block += blocks_per_page;
225 reprobe:
226 continue;
227 }
228 ret = nr_extents;
229 *span = 1 + highest_block - lowest_block;
230 if (page_no == 0)
231 page_no = 1; /* force Empty message */
232 sis->max = page_no;
233 sis->pages = page_no - 1;
234 sis->highest_bit = page_no - 1;
235 out:
236 return ret;
237 bad_bmap:
238 pr_err("swapon: swapfile has holes\n");
239 ret = -EINVAL;
240 goto out;
241 }
242
243 /*
244 * We may have stale swap cache pages in memory: notice
245 * them here and get rid of the unnecessary final write.
246 */
swap_writepage(struct page * page,struct writeback_control * wbc)247 int swap_writepage(struct page *page, struct writeback_control *wbc)
248 {
249 int ret = 0;
250
251 if (try_to_free_swap(page)) {
252 unlock_page(page);
253 goto out;
254 }
255 /*
256 * Arch code may have to preserve more data than just the page
257 * contents, e.g. memory tags.
258 */
259 ret = arch_prepare_to_swap(page);
260 if (ret) {
261 set_page_dirty(page);
262 unlock_page(page);
263 goto out;
264 }
265 if (frontswap_store(page) == 0) {
266 set_page_writeback(page);
267 unlock_page(page);
268 end_page_writeback(page);
269 goto out;
270 }
271 ret = __swap_writepage(page, wbc, end_swap_bio_write);
272 out:
273 return ret;
274 }
275
count_swpout_vm_event(struct page * page)276 static inline void count_swpout_vm_event(struct page *page)
277 {
278 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
279 if (unlikely(PageTransHuge(page)))
280 count_vm_event(THP_SWPOUT);
281 #endif
282 count_vm_events(PSWPOUT, thp_nr_pages(page));
283 }
284
285 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
bio_associate_blkg_from_page(struct bio * bio,struct page * page)286 static void bio_associate_blkg_from_page(struct bio *bio, struct page *page)
287 {
288 struct cgroup_subsys_state *css;
289
290 if (!page->mem_cgroup)
291 return;
292
293 rcu_read_lock();
294 css = cgroup_e_css(page->mem_cgroup->css.cgroup, &io_cgrp_subsys);
295 bio_associate_blkg_from_css(bio, css);
296 rcu_read_unlock();
297 }
298 #else
299 #define bio_associate_blkg_from_page(bio, page) do { } while (0)
300 #endif /* CONFIG_MEMCG && CONFIG_BLK_CGROUP */
301
__swap_writepage(struct page * page,struct writeback_control * wbc,bio_end_io_t end_write_func)302 int __swap_writepage(struct page *page, struct writeback_control *wbc,
303 bio_end_io_t end_write_func)
304 {
305 struct bio *bio;
306 int ret;
307 struct swap_info_struct *sis = page_swap_info(page);
308
309 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
310 if (data_race(sis->flags & SWP_FS_OPS)) {
311 struct kiocb kiocb;
312 struct file *swap_file = sis->swap_file;
313 struct address_space *mapping = swap_file->f_mapping;
314 struct bio_vec bv = {
315 .bv_page = page,
316 .bv_len = PAGE_SIZE,
317 .bv_offset = 0
318 };
319 struct iov_iter from;
320
321 iov_iter_bvec(&from, WRITE, &bv, 1, PAGE_SIZE);
322 init_sync_kiocb(&kiocb, swap_file);
323 kiocb.ki_pos = page_file_offset(page);
324
325 set_page_writeback(page);
326 unlock_page(page);
327 ret = mapping->a_ops->direct_IO(&kiocb, &from);
328 if (ret == PAGE_SIZE) {
329 count_vm_event(PSWPOUT);
330 ret = 0;
331 } else {
332 /*
333 * In the case of swap-over-nfs, this can be a
334 * temporary failure if the system has limited
335 * memory for allocating transmit buffers.
336 * Mark the page dirty and avoid
337 * rotate_reclaimable_page but rate-limit the
338 * messages but do not flag PageError like
339 * the normal direct-to-bio case as it could
340 * be temporary.
341 */
342 set_page_dirty(page);
343 ClearPageReclaim(page);
344 pr_err_ratelimited("Write error on dio swapfile (%llu)\n",
345 page_file_offset(page));
346 }
347 end_page_writeback(page);
348 return ret;
349 }
350
351 ret = bdev_write_page(sis->bdev, swap_page_sector(page), page, wbc);
352 if (!ret) {
353 count_swpout_vm_event(page);
354 return 0;
355 }
356
357 bio = get_swap_bio(GFP_NOIO, page, end_write_func);
358 if (bio == NULL) {
359 set_page_dirty(page);
360 unlock_page(page);
361 return -ENOMEM;
362 }
363 bio->bi_opf = REQ_OP_WRITE | REQ_SWAP | wbc_to_write_flags(wbc);
364 bio_associate_blkg_from_page(bio, page);
365 count_swpout_vm_event(page);
366 set_page_writeback(page);
367 unlock_page(page);
368 submit_bio(bio);
369
370 return 0;
371 }
372
swap_readpage(struct page * page,bool synchronous)373 int swap_readpage(struct page *page, bool synchronous)
374 {
375 struct bio *bio;
376 int ret = 0;
377 struct swap_info_struct *sis = page_swap_info(page);
378 blk_qc_t qc;
379 struct gendisk *disk;
380 unsigned long pflags;
381
382 VM_BUG_ON_PAGE(!PageSwapCache(page) && !synchronous, page);
383 VM_BUG_ON_PAGE(!PageLocked(page), page);
384 VM_BUG_ON_PAGE(PageUptodate(page), page);
385
386 /*
387 * Count submission time as memory stall. When the device is congested,
388 * or the submitting cgroup IO-throttled, submission can be a
389 * significant part of overall IO time.
390 */
391 psi_memstall_enter(&pflags);
392
393 if (frontswap_load(page) == 0) {
394 SetPageUptodate(page);
395 unlock_page(page);
396 goto out;
397 }
398
399 if (data_race(sis->flags & SWP_FS_OPS)) {
400 struct file *swap_file = sis->swap_file;
401 struct address_space *mapping = swap_file->f_mapping;
402
403 ret = mapping->a_ops->readpage(swap_file, page);
404 if (!ret)
405 count_vm_event(PSWPIN);
406 goto out;
407 }
408
409 if (sis->flags & SWP_SYNCHRONOUS_IO) {
410 ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
411 if (!ret) {
412 if (trylock_page(page)) {
413 swap_slot_free_notify(page);
414 unlock_page(page);
415 }
416
417 count_vm_event(PSWPIN);
418 goto out;
419 }
420 }
421
422 ret = 0;
423 bio = get_swap_bio(GFP_KERNEL, page, end_swap_bio_read);
424 if (bio == NULL) {
425 unlock_page(page);
426 ret = -ENOMEM;
427 goto out;
428 }
429 disk = bio->bi_disk;
430 /*
431 * Keep this task valid during swap readpage because the oom killer may
432 * attempt to access it in the page fault retry time check.
433 */
434 bio_set_op_attrs(bio, REQ_OP_READ, 0);
435 if (synchronous) {
436 bio->bi_opf |= REQ_HIPRI;
437 get_task_struct(current);
438 bio->bi_private = current;
439 }
440 count_vm_event(PSWPIN);
441 bio_get(bio);
442 qc = submit_bio(bio);
443 while (synchronous) {
444 set_current_state(TASK_UNINTERRUPTIBLE);
445 if (!READ_ONCE(bio->bi_private))
446 break;
447
448 if (!blk_poll(disk->queue, qc, true))
449 blk_io_schedule();
450 }
451 __set_current_state(TASK_RUNNING);
452 bio_put(bio);
453
454 out:
455 psi_memstall_leave(&pflags);
456 return ret;
457 }
458
swap_set_page_dirty(struct page * page)459 int swap_set_page_dirty(struct page *page)
460 {
461 struct swap_info_struct *sis = page_swap_info(page);
462
463 if (data_race(sis->flags & SWP_FS_OPS)) {
464 struct address_space *mapping = sis->swap_file->f_mapping;
465
466 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
467 return mapping->a_ops->set_page_dirty(page);
468 } else {
469 return __set_page_dirty_no_writeback(page);
470 }
471 }
472