xref: /openbmc/linux/fs/ext4/page-io.c (revision 3381df09)
1  // SPDX-License-Identifier: GPL-2.0
2  /*
3   * linux/fs/ext4/page-io.c
4   *
5   * This contains the new page_io functions for ext4
6   *
7   * Written by Theodore Ts'o, 2010.
8   */
9  
10  #include <linux/fs.h>
11  #include <linux/time.h>
12  #include <linux/highuid.h>
13  #include <linux/pagemap.h>
14  #include <linux/quotaops.h>
15  #include <linux/string.h>
16  #include <linux/buffer_head.h>
17  #include <linux/writeback.h>
18  #include <linux/pagevec.h>
19  #include <linux/mpage.h>
20  #include <linux/namei.h>
21  #include <linux/uio.h>
22  #include <linux/bio.h>
23  #include <linux/workqueue.h>
24  #include <linux/kernel.h>
25  #include <linux/slab.h>
26  #include <linux/mm.h>
27  #include <linux/backing-dev.h>
28  
29  #include "ext4_jbd2.h"
30  #include "xattr.h"
31  #include "acl.h"
32  
33  static struct kmem_cache *io_end_cachep;
34  static struct kmem_cache *io_end_vec_cachep;
35  
36  int __init ext4_init_pageio(void)
37  {
38  	io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
39  	if (io_end_cachep == NULL)
40  		return -ENOMEM;
41  
42  	io_end_vec_cachep = KMEM_CACHE(ext4_io_end_vec, 0);
43  	if (io_end_vec_cachep == NULL) {
44  		kmem_cache_destroy(io_end_cachep);
45  		return -ENOMEM;
46  	}
47  	return 0;
48  }
49  
50  void ext4_exit_pageio(void)
51  {
52  	kmem_cache_destroy(io_end_cachep);
53  	kmem_cache_destroy(io_end_vec_cachep);
54  }
55  
56  struct ext4_io_end_vec *ext4_alloc_io_end_vec(ext4_io_end_t *io_end)
57  {
58  	struct ext4_io_end_vec *io_end_vec;
59  
60  	io_end_vec = kmem_cache_zalloc(io_end_vec_cachep, GFP_NOFS);
61  	if (!io_end_vec)
62  		return ERR_PTR(-ENOMEM);
63  	INIT_LIST_HEAD(&io_end_vec->list);
64  	list_add_tail(&io_end_vec->list, &io_end->list_vec);
65  	return io_end_vec;
66  }
67  
68  static void ext4_free_io_end_vec(ext4_io_end_t *io_end)
69  {
70  	struct ext4_io_end_vec *io_end_vec, *tmp;
71  
72  	if (list_empty(&io_end->list_vec))
73  		return;
74  	list_for_each_entry_safe(io_end_vec, tmp, &io_end->list_vec, list) {
75  		list_del(&io_end_vec->list);
76  		kmem_cache_free(io_end_vec_cachep, io_end_vec);
77  	}
78  }
79  
80  struct ext4_io_end_vec *ext4_last_io_end_vec(ext4_io_end_t *io_end)
81  {
82  	BUG_ON(list_empty(&io_end->list_vec));
83  	return list_last_entry(&io_end->list_vec, struct ext4_io_end_vec, list);
84  }
85  
86  /*
87   * Print an buffer I/O error compatible with the fs/buffer.c.  This
88   * provides compatibility with dmesg scrapers that look for a specific
89   * buffer I/O error message.  We really need a unified error reporting
90   * structure to userspace ala Digital Unix's uerf system, but it's
91   * probably not going to happen in my lifetime, due to LKML politics...
92   */
93  static void buffer_io_error(struct buffer_head *bh)
94  {
95  	printk_ratelimited(KERN_ERR "Buffer I/O error on device %pg, logical block %llu\n",
96  		       bh->b_bdev,
97  			(unsigned long long)bh->b_blocknr);
98  }
99  
100  static void ext4_finish_bio(struct bio *bio)
101  {
102  	struct bio_vec *bvec;
103  	struct bvec_iter_all iter_all;
104  
105  	bio_for_each_segment_all(bvec, bio, iter_all) {
106  		struct page *page = bvec->bv_page;
107  		struct page *bounce_page = NULL;
108  		struct buffer_head *bh, *head;
109  		unsigned bio_start = bvec->bv_offset;
110  		unsigned bio_end = bio_start + bvec->bv_len;
111  		unsigned under_io = 0;
112  		unsigned long flags;
113  
114  		if (!page)
115  			continue;
116  
117  		if (fscrypt_is_bounce_page(page)) {
118  			bounce_page = page;
119  			page = fscrypt_pagecache_page(bounce_page);
120  		}
121  
122  		if (bio->bi_status) {
123  			SetPageError(page);
124  			mapping_set_error(page->mapping, -EIO);
125  		}
126  		bh = head = page_buffers(page);
127  		/*
128  		 * We check all buffers in the page under b_uptodate_lock
129  		 * to avoid races with other end io clearing async_write flags
130  		 */
131  		spin_lock_irqsave(&head->b_uptodate_lock, flags);
132  		do {
133  			if (bh_offset(bh) < bio_start ||
134  			    bh_offset(bh) + bh->b_size > bio_end) {
135  				if (buffer_async_write(bh))
136  					under_io++;
137  				continue;
138  			}
139  			clear_buffer_async_write(bh);
140  			if (bio->bi_status)
141  				buffer_io_error(bh);
142  		} while ((bh = bh->b_this_page) != head);
143  		spin_unlock_irqrestore(&head->b_uptodate_lock, flags);
144  		if (!under_io) {
145  			fscrypt_free_bounce_page(bounce_page);
146  			end_page_writeback(page);
147  		}
148  	}
149  }
150  
151  static void ext4_release_io_end(ext4_io_end_t *io_end)
152  {
153  	struct bio *bio, *next_bio;
154  
155  	BUG_ON(!list_empty(&io_end->list));
156  	BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
157  	WARN_ON(io_end->handle);
158  
159  	for (bio = io_end->bio; bio; bio = next_bio) {
160  		next_bio = bio->bi_private;
161  		ext4_finish_bio(bio);
162  		bio_put(bio);
163  	}
164  	ext4_free_io_end_vec(io_end);
165  	kmem_cache_free(io_end_cachep, io_end);
166  }
167  
168  /*
169   * Check a range of space and convert unwritten extents to written. Note that
170   * we are protected from truncate touching same part of extent tree by the
171   * fact that truncate code waits for all DIO to finish (thus exclusion from
172   * direct IO is achieved) and also waits for PageWriteback bits. Thus we
173   * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
174   * completed (happens from ext4_free_ioend()).
175   */
176  static int ext4_end_io_end(ext4_io_end_t *io_end)
177  {
178  	struct inode *inode = io_end->inode;
179  	handle_t *handle = io_end->handle;
180  	int ret = 0;
181  
182  	ext4_debug("ext4_end_io_nolock: io_end 0x%p from inode %lu,list->next 0x%p,"
183  		   "list->prev 0x%p\n",
184  		   io_end, inode->i_ino, io_end->list.next, io_end->list.prev);
185  
186  	io_end->handle = NULL;	/* Following call will use up the handle */
187  	ret = ext4_convert_unwritten_io_end_vec(handle, io_end);
188  	if (ret < 0 && !ext4_forced_shutdown(EXT4_SB(inode->i_sb))) {
189  		ext4_msg(inode->i_sb, KERN_EMERG,
190  			 "failed to convert unwritten extents to written "
191  			 "extents -- potential data loss!  "
192  			 "(inode %lu, error %d)", inode->i_ino, ret);
193  	}
194  	ext4_clear_io_unwritten_flag(io_end);
195  	ext4_release_io_end(io_end);
196  	return ret;
197  }
198  
199  static void dump_completed_IO(struct inode *inode, struct list_head *head)
200  {
201  #ifdef	EXT4FS_DEBUG
202  	struct list_head *cur, *before, *after;
203  	ext4_io_end_t *io_end, *io_end0, *io_end1;
204  
205  	if (list_empty(head))
206  		return;
207  
208  	ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
209  	list_for_each_entry(io_end, head, list) {
210  		cur = &io_end->list;
211  		before = cur->prev;
212  		io_end0 = container_of(before, ext4_io_end_t, list);
213  		after = cur->next;
214  		io_end1 = container_of(after, ext4_io_end_t, list);
215  
216  		ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
217  			    io_end, inode->i_ino, io_end0, io_end1);
218  	}
219  #endif
220  }
221  
222  /* Add the io_end to per-inode completed end_io list. */
223  static void ext4_add_complete_io(ext4_io_end_t *io_end)
224  {
225  	struct ext4_inode_info *ei = EXT4_I(io_end->inode);
226  	struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
227  	struct workqueue_struct *wq;
228  	unsigned long flags;
229  
230  	/* Only reserved conversions from writeback should enter here */
231  	WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
232  	WARN_ON(!io_end->handle && sbi->s_journal);
233  	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
234  	wq = sbi->rsv_conversion_wq;
235  	if (list_empty(&ei->i_rsv_conversion_list))
236  		queue_work(wq, &ei->i_rsv_conversion_work);
237  	list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
238  	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
239  }
240  
241  static int ext4_do_flush_completed_IO(struct inode *inode,
242  				      struct list_head *head)
243  {
244  	ext4_io_end_t *io_end;
245  	struct list_head unwritten;
246  	unsigned long flags;
247  	struct ext4_inode_info *ei = EXT4_I(inode);
248  	int err, ret = 0;
249  
250  	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
251  	dump_completed_IO(inode, head);
252  	list_replace_init(head, &unwritten);
253  	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
254  
255  	while (!list_empty(&unwritten)) {
256  		io_end = list_entry(unwritten.next, ext4_io_end_t, list);
257  		BUG_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
258  		list_del_init(&io_end->list);
259  
260  		err = ext4_end_io_end(io_end);
261  		if (unlikely(!ret && err))
262  			ret = err;
263  	}
264  	return ret;
265  }
266  
267  /*
268   * work on completed IO, to convert unwritten extents to extents
269   */
270  void ext4_end_io_rsv_work(struct work_struct *work)
271  {
272  	struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
273  						  i_rsv_conversion_work);
274  	ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
275  }
276  
277  ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
278  {
279  	ext4_io_end_t *io_end = kmem_cache_zalloc(io_end_cachep, flags);
280  
281  	if (io_end) {
282  		io_end->inode = inode;
283  		INIT_LIST_HEAD(&io_end->list);
284  		INIT_LIST_HEAD(&io_end->list_vec);
285  		atomic_set(&io_end->count, 1);
286  	}
287  	return io_end;
288  }
289  
290  void ext4_put_io_end_defer(ext4_io_end_t *io_end)
291  {
292  	if (atomic_dec_and_test(&io_end->count)) {
293  		if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) ||
294  				list_empty(&io_end->list_vec)) {
295  			ext4_release_io_end(io_end);
296  			return;
297  		}
298  		ext4_add_complete_io(io_end);
299  	}
300  }
301  
302  int ext4_put_io_end(ext4_io_end_t *io_end)
303  {
304  	int err = 0;
305  
306  	if (atomic_dec_and_test(&io_end->count)) {
307  		if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
308  			err = ext4_convert_unwritten_io_end_vec(io_end->handle,
309  								io_end);
310  			io_end->handle = NULL;
311  			ext4_clear_io_unwritten_flag(io_end);
312  		}
313  		ext4_release_io_end(io_end);
314  	}
315  	return err;
316  }
317  
318  ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
319  {
320  	atomic_inc(&io_end->count);
321  	return io_end;
322  }
323  
324  /* BIO completion function for page writeback */
325  static void ext4_end_bio(struct bio *bio)
326  {
327  	ext4_io_end_t *io_end = bio->bi_private;
328  	sector_t bi_sector = bio->bi_iter.bi_sector;
329  	char b[BDEVNAME_SIZE];
330  
331  	if (WARN_ONCE(!io_end, "io_end is NULL: %s: sector %Lu len %u err %d\n",
332  		      bio_devname(bio, b),
333  		      (long long) bio->bi_iter.bi_sector,
334  		      (unsigned) bio_sectors(bio),
335  		      bio->bi_status)) {
336  		ext4_finish_bio(bio);
337  		bio_put(bio);
338  		return;
339  	}
340  	bio->bi_end_io = NULL;
341  
342  	if (bio->bi_status) {
343  		struct inode *inode = io_end->inode;
344  
345  		ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
346  			     "starting block %llu)",
347  			     bio->bi_status, inode->i_ino,
348  			     (unsigned long long)
349  			     bi_sector >> (inode->i_blkbits - 9));
350  		mapping_set_error(inode->i_mapping,
351  				blk_status_to_errno(bio->bi_status));
352  	}
353  
354  	if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
355  		/*
356  		 * Link bio into list hanging from io_end. We have to do it
357  		 * atomically as bio completions can be racing against each
358  		 * other.
359  		 */
360  		bio->bi_private = xchg(&io_end->bio, bio);
361  		ext4_put_io_end_defer(io_end);
362  	} else {
363  		/*
364  		 * Drop io_end reference early. Inode can get freed once
365  		 * we finish the bio.
366  		 */
367  		ext4_put_io_end_defer(io_end);
368  		ext4_finish_bio(bio);
369  		bio_put(bio);
370  	}
371  }
372  
373  void ext4_io_submit(struct ext4_io_submit *io)
374  {
375  	struct bio *bio = io->io_bio;
376  
377  	if (bio) {
378  		int io_op_flags = io->io_wbc->sync_mode == WB_SYNC_ALL ?
379  				  REQ_SYNC : 0;
380  		io->io_bio->bi_write_hint = io->io_end->inode->i_write_hint;
381  		bio_set_op_attrs(io->io_bio, REQ_OP_WRITE, io_op_flags);
382  		submit_bio(io->io_bio);
383  	}
384  	io->io_bio = NULL;
385  }
386  
387  void ext4_io_submit_init(struct ext4_io_submit *io,
388  			 struct writeback_control *wbc)
389  {
390  	io->io_wbc = wbc;
391  	io->io_bio = NULL;
392  	io->io_end = NULL;
393  }
394  
395  static void io_submit_init_bio(struct ext4_io_submit *io,
396  			       struct buffer_head *bh)
397  {
398  	struct bio *bio;
399  
400  	/*
401  	 * bio_alloc will _always_ be able to allocate a bio if
402  	 * __GFP_DIRECT_RECLAIM is set, see comments for bio_alloc_bioset().
403  	 */
404  	bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
405  	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
406  	bio_set_dev(bio, bh->b_bdev);
407  	bio->bi_end_io = ext4_end_bio;
408  	bio->bi_private = ext4_get_io_end(io->io_end);
409  	io->io_bio = bio;
410  	io->io_next_block = bh->b_blocknr;
411  	wbc_init_bio(io->io_wbc, bio);
412  }
413  
414  static void io_submit_add_bh(struct ext4_io_submit *io,
415  			     struct inode *inode,
416  			     struct page *page,
417  			     struct buffer_head *bh)
418  {
419  	int ret;
420  
421  	if (io->io_bio && bh->b_blocknr != io->io_next_block) {
422  submit_and_retry:
423  		ext4_io_submit(io);
424  	}
425  	if (io->io_bio == NULL) {
426  		io_submit_init_bio(io, bh);
427  		io->io_bio->bi_write_hint = inode->i_write_hint;
428  	}
429  	ret = bio_add_page(io->io_bio, page, bh->b_size, bh_offset(bh));
430  	if (ret != bh->b_size)
431  		goto submit_and_retry;
432  	wbc_account_cgroup_owner(io->io_wbc, page, bh->b_size);
433  	io->io_next_block++;
434  }
435  
436  int ext4_bio_write_page(struct ext4_io_submit *io,
437  			struct page *page,
438  			int len,
439  			struct writeback_control *wbc,
440  			bool keep_towrite)
441  {
442  	struct page *bounce_page = NULL;
443  	struct inode *inode = page->mapping->host;
444  	unsigned block_start;
445  	struct buffer_head *bh, *head;
446  	int ret = 0;
447  	int nr_submitted = 0;
448  	int nr_to_submit = 0;
449  
450  	BUG_ON(!PageLocked(page));
451  	BUG_ON(PageWriteback(page));
452  
453  	if (keep_towrite)
454  		set_page_writeback_keepwrite(page);
455  	else
456  		set_page_writeback(page);
457  	ClearPageError(page);
458  
459  	/*
460  	 * Comments copied from block_write_full_page:
461  	 *
462  	 * The page straddles i_size.  It must be zeroed out on each and every
463  	 * writepage invocation because it may be mmapped.  "A file is mapped
464  	 * in multiples of the page size.  For a file that is not a multiple of
465  	 * the page size, the remaining memory is zeroed when mapped, and
466  	 * writes to that region are not written out to the file."
467  	 */
468  	if (len < PAGE_SIZE)
469  		zero_user_segment(page, len, PAGE_SIZE);
470  	/*
471  	 * In the first loop we prepare and mark buffers to submit. We have to
472  	 * mark all buffers in the page before submitting so that
473  	 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
474  	 * on the first buffer finishes and we are still working on submitting
475  	 * the second buffer.
476  	 */
477  	bh = head = page_buffers(page);
478  	do {
479  		block_start = bh_offset(bh);
480  		if (block_start >= len) {
481  			clear_buffer_dirty(bh);
482  			set_buffer_uptodate(bh);
483  			continue;
484  		}
485  		if (!buffer_dirty(bh) || buffer_delay(bh) ||
486  		    !buffer_mapped(bh) || buffer_unwritten(bh)) {
487  			/* A hole? We can safely clear the dirty bit */
488  			if (!buffer_mapped(bh))
489  				clear_buffer_dirty(bh);
490  			if (io->io_bio)
491  				ext4_io_submit(io);
492  			continue;
493  		}
494  		if (buffer_new(bh))
495  			clear_buffer_new(bh);
496  		set_buffer_async_write(bh);
497  		nr_to_submit++;
498  	} while ((bh = bh->b_this_page) != head);
499  
500  	bh = head = page_buffers(page);
501  
502  	/*
503  	 * If any blocks are being written to an encrypted file, encrypt them
504  	 * into a bounce page.  For simplicity, just encrypt until the last
505  	 * block which might be needed.  This may cause some unneeded blocks
506  	 * (e.g. holes) to be unnecessarily encrypted, but this is rare and
507  	 * can't happen in the common case of blocksize == PAGE_SIZE.
508  	 */
509  	if (IS_ENCRYPTED(inode) && S_ISREG(inode->i_mode) && nr_to_submit) {
510  		gfp_t gfp_flags = GFP_NOFS;
511  		unsigned int enc_bytes = round_up(len, i_blocksize(inode));
512  
513  		/*
514  		 * Since bounce page allocation uses a mempool, we can only use
515  		 * a waiting mask (i.e. request guaranteed allocation) on the
516  		 * first page of the bio.  Otherwise it can deadlock.
517  		 */
518  		if (io->io_bio)
519  			gfp_flags = GFP_NOWAIT | __GFP_NOWARN;
520  	retry_encrypt:
521  		bounce_page = fscrypt_encrypt_pagecache_blocks(page, enc_bytes,
522  							       0, gfp_flags);
523  		if (IS_ERR(bounce_page)) {
524  			ret = PTR_ERR(bounce_page);
525  			if (ret == -ENOMEM &&
526  			    (io->io_bio || wbc->sync_mode == WB_SYNC_ALL)) {
527  				gfp_flags = GFP_NOFS;
528  				if (io->io_bio)
529  					ext4_io_submit(io);
530  				else
531  					gfp_flags |= __GFP_NOFAIL;
532  				congestion_wait(BLK_RW_ASYNC, HZ/50);
533  				goto retry_encrypt;
534  			}
535  
536  			printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
537  			redirty_page_for_writepage(wbc, page);
538  			do {
539  				clear_buffer_async_write(bh);
540  				bh = bh->b_this_page;
541  			} while (bh != head);
542  			goto unlock;
543  		}
544  	}
545  
546  	/* Now submit buffers to write */
547  	do {
548  		if (!buffer_async_write(bh))
549  			continue;
550  		io_submit_add_bh(io, inode,
551  				 bounce_page ? bounce_page : page, bh);
552  		nr_submitted++;
553  		clear_buffer_dirty(bh);
554  	} while ((bh = bh->b_this_page) != head);
555  
556  unlock:
557  	unlock_page(page);
558  	/* Nothing submitted - we have to end page writeback */
559  	if (!nr_submitted)
560  		end_page_writeback(page);
561  	return ret;
562  }
563