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/uio.h>
22 #include <linux/bio.h>
23 #include <linux/workqueue.h>
24 #include <linux/kernel.h>
25 #include <linux/slab.h>
26 
27 #include "ext4_jbd2.h"
28 #include "xattr.h"
29 #include "acl.h"
30 #include "ext4_extents.h"
31 
32 static struct kmem_cache *io_page_cachep, *io_end_cachep;
33 
ext4_init_pageio(void)34 int __init ext4_init_pageio(void)
35 {
36 	io_page_cachep = KMEM_CACHE(ext4_io_page, SLAB_RECLAIM_ACCOUNT);
37 	if (io_page_cachep == NULL)
38 		return -ENOMEM;
39 	io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
40 	if (io_end_cachep == NULL) {
41 		kmem_cache_destroy(io_page_cachep);
42 		return -ENOMEM;
43 	}
44 	return 0;
45 }
46 
ext4_exit_pageio(void)47 void ext4_exit_pageio(void)
48 {
49 	kmem_cache_destroy(io_end_cachep);
50 	kmem_cache_destroy(io_page_cachep);
51 }
52 
ext4_ioend_wait(struct inode * inode)53 void ext4_ioend_wait(struct inode *inode)
54 {
55 	wait_queue_head_t *wq = ext4_ioend_wq(inode);
56 
57 	wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_ioend_count) == 0));
58 }
59 
put_io_page(struct ext4_io_page * io_page)60 static void put_io_page(struct ext4_io_page *io_page)
61 {
62 	if (atomic_dec_and_test(&io_page->p_count)) {
63 		end_page_writeback(io_page->p_page);
64 		put_page(io_page->p_page);
65 		kmem_cache_free(io_page_cachep, io_page);
66 	}
67 }
68 
ext4_free_io_end(ext4_io_end_t * io)69 void ext4_free_io_end(ext4_io_end_t *io)
70 {
71 	int i;
72 
73 	BUG_ON(!io);
74 	if (io->page)
75 		put_page(io->page);
76 	for (i = 0; i < io->num_io_pages; i++)
77 		put_io_page(io->pages[i]);
78 	io->num_io_pages = 0;
79 	if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count))
80 		wake_up_all(ext4_ioend_wq(io->inode));
81 	kmem_cache_free(io_end_cachep, io);
82 }
83 
84 /*
85  * check a range of space and convert unwritten extents to written.
86  *
87  * Called with inode->i_mutex; we depend on this when we manipulate
88  * io->flag, since we could otherwise race with ext4_flush_completed_IO()
89  */
ext4_end_io_nolock(ext4_io_end_t * io)90 int ext4_end_io_nolock(ext4_io_end_t *io)
91 {
92 	struct inode *inode = io->inode;
93 	loff_t offset = io->offset;
94 	ssize_t size = io->size;
95 	int ret = 0;
96 
97 	ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
98 		   "list->prev 0x%p\n",
99 		   io, inode->i_ino, io->list.next, io->list.prev);
100 
101 	ret = ext4_convert_unwritten_extents(inode, offset, size);
102 	if (ret < 0) {
103 		ext4_msg(inode->i_sb, KERN_EMERG,
104 			 "failed to convert unwritten extents to written "
105 			 "extents -- potential data loss!  "
106 			 "(inode %lu, offset %llu, size %zd, error %d)",
107 			 inode->i_ino, offset, size, ret);
108 	}
109 
110 	/* Wake up anyone waiting on unwritten extent conversion */
111 	if (atomic_dec_and_test(&EXT4_I(inode)->i_aiodio_unwritten))
112 		wake_up_all(ext4_ioend_wq(io->inode));
113 	if (io->flag & EXT4_IO_END_DIRECT)
114 		inode_dio_done(inode);
115 	if (io->iocb)
116 		aio_complete(io->iocb, io->result, 0);
117 	return ret;
118 }
119 
120 /*
121  * work on completed aio dio IO, to convert unwritten extents to extents
122  */
ext4_end_io_work(struct work_struct * work)123 static void ext4_end_io_work(struct work_struct *work)
124 {
125 	ext4_io_end_t		*io = container_of(work, ext4_io_end_t, work);
126 	struct inode		*inode = io->inode;
127 	struct ext4_inode_info	*ei = EXT4_I(inode);
128 	unsigned long		flags;
129 
130 	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
131 	if (io->flag & EXT4_IO_END_IN_FSYNC)
132 		goto requeue;
133 	if (list_empty(&io->list)) {
134 		spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
135 		goto free;
136 	}
137 
138 	if (!mutex_trylock(&inode->i_mutex)) {
139 		bool was_queued;
140 requeue:
141 		was_queued = !!(io->flag & EXT4_IO_END_QUEUED);
142 		io->flag |= EXT4_IO_END_QUEUED;
143 		spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
144 		/*
145 		 * Requeue the work instead of waiting so that the work
146 		 * items queued after this can be processed.
147 		 */
148 		queue_work(EXT4_SB(inode->i_sb)->dio_unwritten_wq, &io->work);
149 		/*
150 		 * To prevent the ext4-dio-unwritten thread from keeping
151 		 * requeueing end_io requests and occupying cpu for too long,
152 		 * yield the cpu if it sees an end_io request that has already
153 		 * been requeued.
154 		 */
155 		if (was_queued)
156 			yield();
157 		return;
158 	}
159 	list_del_init(&io->list);
160 	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
161 	(void) ext4_end_io_nolock(io);
162 	mutex_unlock(&inode->i_mutex);
163 free:
164 	ext4_free_io_end(io);
165 }
166 
ext4_init_io_end(struct inode * inode,gfp_t flags)167 ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
168 {
169 	ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
170 	if (io) {
171 		atomic_inc(&EXT4_I(inode)->i_ioend_count);
172 		io->inode = inode;
173 		INIT_WORK(&io->work, ext4_end_io_work);
174 		INIT_LIST_HEAD(&io->list);
175 	}
176 	return io;
177 }
178 
179 /*
180  * Print an buffer I/O error compatible with the fs/buffer.c.  This
181  * provides compatibility with dmesg scrapers that look for a specific
182  * buffer I/O error message.  We really need a unified error reporting
183  * structure to userspace ala Digital Unix's uerf system, but it's
184  * probably not going to happen in my lifetime, due to LKML politics...
185  */
buffer_io_error(struct buffer_head * bh)186 static void buffer_io_error(struct buffer_head *bh)
187 {
188 	char b[BDEVNAME_SIZE];
189 	printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
190 			bdevname(bh->b_bdev, b),
191 			(unsigned long long)bh->b_blocknr);
192 }
193 
ext4_end_bio(struct bio * bio,int error)194 static void ext4_end_bio(struct bio *bio, int error)
195 {
196 	ext4_io_end_t *io_end = bio->bi_private;
197 	struct workqueue_struct *wq;
198 	struct inode *inode;
199 	unsigned long flags;
200 	int i;
201 	sector_t bi_sector = bio->bi_sector;
202 
203 	BUG_ON(!io_end);
204 	bio->bi_private = NULL;
205 	bio->bi_end_io = NULL;
206 	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
207 		error = 0;
208 	bio_put(bio);
209 
210 	for (i = 0; i < io_end->num_io_pages; i++) {
211 		struct page *page = io_end->pages[i]->p_page;
212 		struct buffer_head *bh, *head;
213 		loff_t offset;
214 		loff_t io_end_offset;
215 
216 		if (error) {
217 			SetPageError(page);
218 			set_bit(AS_EIO, &page->mapping->flags);
219 			head = page_buffers(page);
220 			BUG_ON(!head);
221 
222 			io_end_offset = io_end->offset + io_end->size;
223 
224 			offset = (sector_t) page->index << PAGE_CACHE_SHIFT;
225 			bh = head;
226 			do {
227 				if ((offset >= io_end->offset) &&
228 				    (offset+bh->b_size <= io_end_offset))
229 					buffer_io_error(bh);
230 
231 				offset += bh->b_size;
232 				bh = bh->b_this_page;
233 			} while (bh != head);
234 		}
235 
236 		put_io_page(io_end->pages[i]);
237 	}
238 	io_end->num_io_pages = 0;
239 	inode = io_end->inode;
240 
241 	if (error) {
242 		io_end->flag |= EXT4_IO_END_ERROR;
243 		ext4_warning(inode->i_sb, "I/O error writing to inode %lu "
244 			     "(offset %llu size %ld starting block %llu)",
245 			     inode->i_ino,
246 			     (unsigned long long) io_end->offset,
247 			     (long) io_end->size,
248 			     (unsigned long long)
249 			     bi_sector >> (inode->i_blkbits - 9));
250 	}
251 
252 	if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
253 		ext4_free_io_end(io_end);
254 		return;
255 	}
256 
257 	/* Add the io_end to per-inode completed io list*/
258 	spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
259 	list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
260 	spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
261 
262 	wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
263 	/* queue the work to convert unwritten extents to written */
264 	queue_work(wq, &io_end->work);
265 }
266 
ext4_io_submit(struct ext4_io_submit * io)267 void ext4_io_submit(struct ext4_io_submit *io)
268 {
269 	struct bio *bio = io->io_bio;
270 
271 	if (bio) {
272 		bio_get(io->io_bio);
273 		submit_bio(io->io_op, io->io_bio);
274 		BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
275 		bio_put(io->io_bio);
276 	}
277 	io->io_bio = NULL;
278 	io->io_op = 0;
279 	io->io_end = NULL;
280 }
281 
io_submit_init(struct ext4_io_submit * io,struct inode * inode,struct writeback_control * wbc,struct buffer_head * bh)282 static int io_submit_init(struct ext4_io_submit *io,
283 			  struct inode *inode,
284 			  struct writeback_control *wbc,
285 			  struct buffer_head *bh)
286 {
287 	ext4_io_end_t *io_end;
288 	struct page *page = bh->b_page;
289 	int nvecs = bio_get_nr_vecs(bh->b_bdev);
290 	struct bio *bio;
291 
292 	io_end = ext4_init_io_end(inode, GFP_NOFS);
293 	if (!io_end)
294 		return -ENOMEM;
295 	bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES));
296 	bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
297 	bio->bi_bdev = bh->b_bdev;
298 	bio->bi_private = io->io_end = io_end;
299 	bio->bi_end_io = ext4_end_bio;
300 
301 	io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
302 
303 	io->io_bio = bio;
304 	io->io_op = (wbc->sync_mode == WB_SYNC_ALL ?  WRITE_SYNC : WRITE);
305 	io->io_next_block = bh->b_blocknr;
306 	return 0;
307 }
308 
io_submit_add_bh(struct ext4_io_submit * io,struct ext4_io_page * io_page,struct inode * inode,struct writeback_control * wbc,struct buffer_head * bh)309 static int io_submit_add_bh(struct ext4_io_submit *io,
310 			    struct ext4_io_page *io_page,
311 			    struct inode *inode,
312 			    struct writeback_control *wbc,
313 			    struct buffer_head *bh)
314 {
315 	ext4_io_end_t *io_end;
316 	int ret;
317 
318 	if (buffer_new(bh)) {
319 		clear_buffer_new(bh);
320 		unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
321 	}
322 
323 	if (!buffer_mapped(bh) || buffer_delay(bh)) {
324 		if (!buffer_mapped(bh))
325 			clear_buffer_dirty(bh);
326 		if (io->io_bio)
327 			ext4_io_submit(io);
328 		return 0;
329 	}
330 
331 	if (io->io_bio && bh->b_blocknr != io->io_next_block) {
332 submit_and_retry:
333 		ext4_io_submit(io);
334 	}
335 	if (io->io_bio == NULL) {
336 		ret = io_submit_init(io, inode, wbc, bh);
337 		if (ret)
338 			return ret;
339 	}
340 	io_end = io->io_end;
341 	if ((io_end->num_io_pages >= MAX_IO_PAGES) &&
342 	    (io_end->pages[io_end->num_io_pages-1] != io_page))
343 		goto submit_and_retry;
344 	if (buffer_uninit(bh))
345 		ext4_set_io_unwritten_flag(inode, io_end);
346 	io->io_end->size += bh->b_size;
347 	io->io_next_block++;
348 	ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
349 	if (ret != bh->b_size)
350 		goto submit_and_retry;
351 	if ((io_end->num_io_pages == 0) ||
352 	    (io_end->pages[io_end->num_io_pages-1] != io_page)) {
353 		io_end->pages[io_end->num_io_pages++] = io_page;
354 		atomic_inc(&io_page->p_count);
355 	}
356 	return 0;
357 }
358 
ext4_bio_write_page(struct ext4_io_submit * io,struct page * page,int len,struct writeback_control * wbc)359 int ext4_bio_write_page(struct ext4_io_submit *io,
360 			struct page *page,
361 			int len,
362 			struct writeback_control *wbc)
363 {
364 	struct inode *inode = page->mapping->host;
365 	unsigned block_start, block_end, blocksize;
366 	struct ext4_io_page *io_page;
367 	struct buffer_head *bh, *head;
368 	int ret = 0;
369 
370 	blocksize = 1 << inode->i_blkbits;
371 
372 	BUG_ON(!PageLocked(page));
373 	BUG_ON(PageWriteback(page));
374 
375 	io_page = kmem_cache_alloc(io_page_cachep, GFP_NOFS);
376 	if (!io_page) {
377 		set_page_dirty(page);
378 		unlock_page(page);
379 		return -ENOMEM;
380 	}
381 	io_page->p_page = page;
382 	atomic_set(&io_page->p_count, 1);
383 	get_page(page);
384 	set_page_writeback(page);
385 	ClearPageError(page);
386 
387 	for (bh = head = page_buffers(page), block_start = 0;
388 	     bh != head || !block_start;
389 	     block_start = block_end, bh = bh->b_this_page) {
390 
391 		block_end = block_start + blocksize;
392 		if (block_start >= len) {
393 			/*
394 			 * Comments copied from block_write_full_page_endio:
395 			 *
396 			 * The page straddles i_size.  It must be zeroed out on
397 			 * each and every writepage invocation because it may
398 			 * be mmapped.  "A file is mapped in multiples of the
399 			 * page size.  For a file that is not a multiple of
400 			 * the  page size, the remaining memory is zeroed when
401 			 * mapped, and writes to that region are not written
402 			 * out to the file."
403 			 */
404 			zero_user_segment(page, block_start, block_end);
405 			clear_buffer_dirty(bh);
406 			set_buffer_uptodate(bh);
407 			continue;
408 		}
409 		clear_buffer_dirty(bh);
410 		ret = io_submit_add_bh(io, io_page, inode, wbc, bh);
411 		if (ret) {
412 			/*
413 			 * We only get here on ENOMEM.  Not much else
414 			 * we can do but mark the page as dirty, and
415 			 * better luck next time.
416 			 */
417 			set_page_dirty(page);
418 			break;
419 		}
420 	}
421 	unlock_page(page);
422 	/*
423 	 * If the page was truncated before we could do the writeback,
424 	 * or we had a memory allocation error while trying to write
425 	 * the first buffer head, we won't have submitted any pages for
426 	 * I/O.  In that case we need to make sure we've cleared the
427 	 * PageWriteback bit from the page to prevent the system from
428 	 * wedging later on.
429 	 */
430 	put_io_page(io_page);
431 	return ret;
432 }
433