xref: /openbmc/linux/fs/ext4/page-io.c (revision cdfce539) 
1 /*
2  * linux/fs/ext4/page-io.c
3  *
4  * This contains the new page_io functions for ext4
5  *
6  * Written by Theodore Ts'o, 2010.
7  */
8 
9 #include <linux/fs.h>
10 #include <linux/time.h>
11 #include <linux/jbd2.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/aio.h>
22 #include <linux/uio.h>
23 #include <linux/bio.h>
24 #include <linux/workqueue.h>
25 #include <linux/kernel.h>
26 #include <linux/slab.h>
27 #include <linux/mm.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 
35 int __init ext4_init_pageio(void)
36 {
37 	io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
38 	if (io_end_cachep == NULL)
39 		return -ENOMEM;
40 	return 0;
41 }
42 
43 void ext4_exit_pageio(void)
44 {
45 	kmem_cache_destroy(io_end_cachep);
46 }
47 
48 /*
49  * This function is called by ext4_evict_inode() to make sure there is
50  * no more pending I/O completion work left to do.
51  */
52 void ext4_ioend_shutdown(struct inode *inode)
53 {
54 	wait_queue_head_t *wq = ext4_ioend_wq(inode);
55 
56 	wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_ioend_count) == 0));
57 	/*
58 	 * We need to make sure the work structure is finished being
59 	 * used before we let the inode get destroyed.
60 	 */
61 	if (work_pending(&EXT4_I(inode)->i_unwritten_work))
62 		cancel_work_sync(&EXT4_I(inode)->i_unwritten_work);
63 }
64 
65 void ext4_free_io_end(ext4_io_end_t *io)
66 {
67 	BUG_ON(!io);
68 	BUG_ON(!list_empty(&io->list));
69 	BUG_ON(io->flag & EXT4_IO_END_UNWRITTEN);
70 
71 	if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count))
72 		wake_up_all(ext4_ioend_wq(io->inode));
73 	kmem_cache_free(io_end_cachep, io);
74 }
75 
76 /* check a range of space and convert unwritten extents to written. */
77 static int ext4_end_io(ext4_io_end_t *io)
78 {
79 	struct inode *inode = io->inode;
80 	loff_t offset = io->offset;
81 	ssize_t size = io->size;
82 	int ret = 0;
83 
84 	ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
85 		   "list->prev 0x%p\n",
86 		   io, inode->i_ino, io->list.next, io->list.prev);
87 
88 	ret = ext4_convert_unwritten_extents(inode, offset, size);
89 	if (ret < 0) {
90 		ext4_msg(inode->i_sb, KERN_EMERG,
91 			 "failed to convert unwritten extents to written "
92 			 "extents -- potential data loss!  "
93 			 "(inode %lu, offset %llu, size %zd, error %d)",
94 			 inode->i_ino, offset, size, ret);
95 	}
96 	/* Wake up anyone waiting on unwritten extent conversion */
97 	if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
98 		wake_up_all(ext4_ioend_wq(inode));
99 	if (io->flag & EXT4_IO_END_DIRECT)
100 		inode_dio_done(inode);
101 	if (io->iocb)
102 		aio_complete(io->iocb, io->result, 0);
103 	return ret;
104 }
105 
106 static void dump_completed_IO(struct inode *inode)
107 {
108 #ifdef	EXT4FS_DEBUG
109 	struct list_head *cur, *before, *after;
110 	ext4_io_end_t *io, *io0, *io1;
111 
112 	if (list_empty(&EXT4_I(inode)->i_completed_io_list)) {
113 		ext4_debug("inode %lu completed_io list is empty\n",
114 			   inode->i_ino);
115 		return;
116 	}
117 
118 	ext4_debug("Dump inode %lu completed_io list\n", inode->i_ino);
119 	list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list) {
120 		cur = &io->list;
121 		before = cur->prev;
122 		io0 = container_of(before, ext4_io_end_t, list);
123 		after = cur->next;
124 		io1 = container_of(after, ext4_io_end_t, list);
125 
126 		ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
127 			    io, inode->i_ino, io0, io1);
128 	}
129 #endif
130 }
131 
132 /* Add the io_end to per-inode completed end_io list. */
133 void ext4_add_complete_io(ext4_io_end_t *io_end)
134 {
135 	struct ext4_inode_info *ei = EXT4_I(io_end->inode);
136 	struct workqueue_struct *wq;
137 	unsigned long flags;
138 
139 	BUG_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
140 	wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
141 
142 	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
143 	if (list_empty(&ei->i_completed_io_list))
144 		queue_work(wq, &ei->i_unwritten_work);
145 	list_add_tail(&io_end->list, &ei->i_completed_io_list);
146 	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
147 }
148 
149 static int ext4_do_flush_completed_IO(struct inode *inode)
150 {
151 	ext4_io_end_t *io;
152 	struct list_head unwritten;
153 	unsigned long flags;
154 	struct ext4_inode_info *ei = EXT4_I(inode);
155 	int err, ret = 0;
156 
157 	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
158 	dump_completed_IO(inode);
159 	list_replace_init(&ei->i_completed_io_list, &unwritten);
160 	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
161 
162 	while (!list_empty(&unwritten)) {
163 		io = list_entry(unwritten.next, ext4_io_end_t, list);
164 		BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
165 		list_del_init(&io->list);
166 
167 		err = ext4_end_io(io);
168 		if (unlikely(!ret && err))
169 			ret = err;
170 		io->flag &= ~EXT4_IO_END_UNWRITTEN;
171 		ext4_free_io_end(io);
172 	}
173 	return ret;
174 }
175 
176 /*
177  * work on completed aio dio IO, to convert unwritten extents to extents
178  */
179 void ext4_end_io_work(struct work_struct *work)
180 {
181 	struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
182 						  i_unwritten_work);
183 	ext4_do_flush_completed_IO(&ei->vfs_inode);
184 }
185 
186 int ext4_flush_unwritten_io(struct inode *inode)
187 {
188 	int ret;
189 	WARN_ON_ONCE(!mutex_is_locked(&inode->i_mutex) &&
190 		     !(inode->i_state & I_FREEING));
191 	ret = ext4_do_flush_completed_IO(inode);
192 	ext4_unwritten_wait(inode);
193 	return ret;
194 }
195 
196 ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
197 {
198 	ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
199 	if (io) {
200 		atomic_inc(&EXT4_I(inode)->i_ioend_count);
201 		io->inode = inode;
202 		INIT_LIST_HEAD(&io->list);
203 	}
204 	return io;
205 }
206 
207 /*
208  * Print an buffer I/O error compatible with the fs/buffer.c.  This
209  * provides compatibility with dmesg scrapers that look for a specific
210  * buffer I/O error message.  We really need a unified error reporting
211  * structure to userspace ala Digital Unix's uerf system, but it's
212  * probably not going to happen in my lifetime, due to LKML politics...
213  */
214 static void buffer_io_error(struct buffer_head *bh)
215 {
216 	char b[BDEVNAME_SIZE];
217 	printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
218 			bdevname(bh->b_bdev, b),
219 			(unsigned long long)bh->b_blocknr);
220 }
221 
222 static void ext4_end_bio(struct bio *bio, int error)
223 {
224 	ext4_io_end_t *io_end = bio->bi_private;
225 	struct inode *inode;
226 	int i;
227 	int blocksize;
228 	sector_t bi_sector = bio->bi_sector;
229 
230 	BUG_ON(!io_end);
231 	inode = io_end->inode;
232 	blocksize = 1 << inode->i_blkbits;
233 	bio->bi_private = NULL;
234 	bio->bi_end_io = NULL;
235 	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
236 		error = 0;
237 	for (i = 0; i < bio->bi_vcnt; i++) {
238 		struct bio_vec *bvec = &bio->bi_io_vec[i];
239 		struct page *page = bvec->bv_page;
240 		struct buffer_head *bh, *head;
241 		unsigned bio_start = bvec->bv_offset;
242 		unsigned bio_end = bio_start + bvec->bv_len;
243 		unsigned under_io = 0;
244 		unsigned long flags;
245 
246 		if (!page)
247 			continue;
248 
249 		if (error) {
250 			SetPageError(page);
251 			set_bit(AS_EIO, &page->mapping->flags);
252 		}
253 		bh = head = page_buffers(page);
254 		/*
255 		 * We check all buffers in the page under BH_Uptodate_Lock
256 		 * to avoid races with other end io clearing async_write flags
257 		 */
258 		local_irq_save(flags);
259 		bit_spin_lock(BH_Uptodate_Lock, &head->b_state);
260 		do {
261 			if (bh_offset(bh) < bio_start ||
262 			    bh_offset(bh) + blocksize > bio_end) {
263 				if (buffer_async_write(bh))
264 					under_io++;
265 				continue;
266 			}
267 			clear_buffer_async_write(bh);
268 			if (error)
269 				buffer_io_error(bh);
270 		} while ((bh = bh->b_this_page) != head);
271 		bit_spin_unlock(BH_Uptodate_Lock, &head->b_state);
272 		local_irq_restore(flags);
273 		if (!under_io)
274 			end_page_writeback(page);
275 	}
276 	bio_put(bio);
277 
278 	if (error) {
279 		io_end->flag |= EXT4_IO_END_ERROR;
280 		ext4_warning(inode->i_sb, "I/O error writing to inode %lu "
281 			     "(offset %llu size %ld starting block %llu)",
282 			     inode->i_ino,
283 			     (unsigned long long) io_end->offset,
284 			     (long) io_end->size,
285 			     (unsigned long long)
286 			     bi_sector >> (inode->i_blkbits - 9));
287 	}
288 
289 	if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
290 		ext4_free_io_end(io_end);
291 		return;
292 	}
293 
294 	ext4_add_complete_io(io_end);
295 }
296 
297 void ext4_io_submit(struct ext4_io_submit *io)
298 {
299 	struct bio *bio = io->io_bio;
300 
301 	if (bio) {
302 		bio_get(io->io_bio);
303 		submit_bio(io->io_op, io->io_bio);
304 		BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
305 		bio_put(io->io_bio);
306 	}
307 	io->io_bio = NULL;
308 	io->io_op = 0;
309 	io->io_end = NULL;
310 }
311 
312 static int io_submit_init(struct ext4_io_submit *io,
313 			  struct inode *inode,
314 			  struct writeback_control *wbc,
315 			  struct buffer_head *bh)
316 {
317 	ext4_io_end_t *io_end;
318 	struct page *page = bh->b_page;
319 	int nvecs = bio_get_nr_vecs(bh->b_bdev);
320 	struct bio *bio;
321 
322 	io_end = ext4_init_io_end(inode, GFP_NOFS);
323 	if (!io_end)
324 		return -ENOMEM;
325 	bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES));
326 	bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
327 	bio->bi_bdev = bh->b_bdev;
328 	bio->bi_private = io->io_end = io_end;
329 	bio->bi_end_io = ext4_end_bio;
330 
331 	io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
332 
333 	io->io_bio = bio;
334 	io->io_op = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
335 	io->io_next_block = bh->b_blocknr;
336 	return 0;
337 }
338 
339 static int io_submit_add_bh(struct ext4_io_submit *io,
340 			    struct inode *inode,
341 			    struct writeback_control *wbc,
342 			    struct buffer_head *bh)
343 {
344 	ext4_io_end_t *io_end;
345 	int ret;
346 
347 	if (io->io_bio && bh->b_blocknr != io->io_next_block) {
348 submit_and_retry:
349 		ext4_io_submit(io);
350 	}
351 	if (io->io_bio == NULL) {
352 		ret = io_submit_init(io, inode, wbc, bh);
353 		if (ret)
354 			return ret;
355 	}
356 	io_end = io->io_end;
357 	if (test_clear_buffer_uninit(bh))
358 		ext4_set_io_unwritten_flag(inode, io_end);
359 	io->io_end->size += bh->b_size;
360 	io->io_next_block++;
361 	ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
362 	if (ret != bh->b_size)
363 		goto submit_and_retry;
364 	return 0;
365 }
366 
367 int ext4_bio_write_page(struct ext4_io_submit *io,
368 			struct page *page,
369 			int len,
370 			struct writeback_control *wbc)
371 {
372 	struct inode *inode = page->mapping->host;
373 	unsigned block_start, blocksize;
374 	struct buffer_head *bh, *head;
375 	int ret = 0;
376 	int nr_submitted = 0;
377 
378 	blocksize = 1 << inode->i_blkbits;
379 
380 	BUG_ON(!PageLocked(page));
381 	BUG_ON(PageWriteback(page));
382 
383 	set_page_writeback(page);
384 	ClearPageError(page);
385 
386 	/*
387 	 * In the first loop we prepare and mark buffers to submit. We have to
388 	 * mark all buffers in the page before submitting so that
389 	 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
390 	 * on the first buffer finishes and we are still working on submitting
391 	 * the second buffer.
392 	 */
393 	bh = head = page_buffers(page);
394 	do {
395 		block_start = bh_offset(bh);
396 		if (block_start >= len) {
397 			/*
398 			 * Comments copied from block_write_full_page_endio:
399 			 *
400 			 * The page straddles i_size.  It must be zeroed out on
401 			 * each and every writepage invocation because it may
402 			 * be mmapped.  "A file is mapped in multiples of the
403 			 * page size.  For a file that is not a multiple of
404 			 * the  page size, the remaining memory is zeroed when
405 			 * mapped, and writes to that region are not written
406 			 * out to the file."
407 			 */
408 			zero_user_segment(page, block_start,
409 					  block_start + blocksize);
410 			clear_buffer_dirty(bh);
411 			set_buffer_uptodate(bh);
412 			continue;
413 		}
414 		if (!buffer_dirty(bh) || buffer_delay(bh) ||
415 		    !buffer_mapped(bh) || buffer_unwritten(bh)) {
416 			/* A hole? We can safely clear the dirty bit */
417 			if (!buffer_mapped(bh))
418 				clear_buffer_dirty(bh);
419 			if (io->io_bio)
420 				ext4_io_submit(io);
421 			continue;
422 		}
423 		if (buffer_new(bh)) {
424 			clear_buffer_new(bh);
425 			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
426 		}
427 		set_buffer_async_write(bh);
428 	} while ((bh = bh->b_this_page) != head);
429 
430 	/* Now submit buffers to write */
431 	bh = head = page_buffers(page);
432 	do {
433 		if (!buffer_async_write(bh))
434 			continue;
435 		ret = io_submit_add_bh(io, inode, wbc, bh);
436 		if (ret) {
437 			/*
438 			 * We only get here on ENOMEM.  Not much else
439 			 * we can do but mark the page as dirty, and
440 			 * better luck next time.
441 			 */
442 			redirty_page_for_writepage(wbc, page);
443 			break;
444 		}
445 		nr_submitted++;
446 		clear_buffer_dirty(bh);
447 	} while ((bh = bh->b_this_page) != head);
448 
449 	/* Error stopped previous loop? Clean up buffers... */
450 	if (ret) {
451 		do {
452 			clear_buffer_async_write(bh);
453 			bh = bh->b_this_page;
454 		} while (bh != head);
455 	}
456 	unlock_page(page);
457 	/* Nothing submitted - we have to end page writeback */
458 	if (!nr_submitted)
459 		end_page_writeback(page);
460 	return ret;
461 }
462