Loading...
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
32static struct kmem_cache *io_page_cachep, *io_end_cachep;
33
34int __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
47void ext4_exit_pageio(void)
48{
49 kmem_cache_destroy(io_end_cachep);
50 kmem_cache_destroy(io_page_cachep);
51}
52
53void 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
60static 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
69void 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 */
90int 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 if (io->iocb)
111 aio_complete(io->iocb, io->result, 0);
112
113 if (io->flag & EXT4_IO_END_DIRECT)
114 inode_dio_done(inode);
115 /* Wake up anyone waiting on unwritten extent conversion */
116 if (atomic_dec_and_test(&EXT4_I(inode)->i_aiodio_unwritten))
117 wake_up_all(ext4_ioend_wq(io->inode));
118 return ret;
119}
120
121/*
122 * work on completed aio dio IO, to convert unwritten extents to extents
123 */
124static void ext4_end_io_work(struct work_struct *work)
125{
126 ext4_io_end_t *io = container_of(work, ext4_io_end_t, work);
127 struct inode *inode = io->inode;
128 struct ext4_inode_info *ei = EXT4_I(inode);
129 unsigned long flags;
130
131 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
132 if (io->flag & EXT4_IO_END_IN_FSYNC)
133 goto requeue;
134 if (list_empty(&io->list)) {
135 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
136 goto free;
137 }
138
139 if (!mutex_trylock(&inode->i_mutex)) {
140 bool was_queued;
141requeue:
142 was_queued = !!(io->flag & EXT4_IO_END_QUEUED);
143 io->flag |= EXT4_IO_END_QUEUED;
144 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
145 /*
146 * Requeue the work instead of waiting so that the work
147 * items queued after this can be processed.
148 */
149 queue_work(EXT4_SB(inode->i_sb)->dio_unwritten_wq, &io->work);
150 /*
151 * To prevent the ext4-dio-unwritten thread from keeping
152 * requeueing end_io requests and occupying cpu for too long,
153 * yield the cpu if it sees an end_io request that has already
154 * been requeued.
155 */
156 if (was_queued)
157 yield();
158 return;
159 }
160 list_del_init(&io->list);
161 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
162 (void) ext4_end_io_nolock(io);
163 mutex_unlock(&inode->i_mutex);
164free:
165 ext4_free_io_end(io);
166}
167
168ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
169{
170 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
171 if (io) {
172 atomic_inc(&EXT4_I(inode)->i_ioend_count);
173 io->inode = inode;
174 INIT_WORK(&io->work, ext4_end_io_work);
175 INIT_LIST_HEAD(&io->list);
176 }
177 return io;
178}
179
180/*
181 * Print an buffer I/O error compatible with the fs/buffer.c. This
182 * provides compatibility with dmesg scrapers that look for a specific
183 * buffer I/O error message. We really need a unified error reporting
184 * structure to userspace ala Digital Unix's uerf system, but it's
185 * probably not going to happen in my lifetime, due to LKML politics...
186 */
187static void buffer_io_error(struct buffer_head *bh)
188{
189 char b[BDEVNAME_SIZE];
190 printk(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
191 bdevname(bh->b_bdev, b),
192 (unsigned long long)bh->b_blocknr);
193}
194
195static void ext4_end_bio(struct bio *bio, int error)
196{
197 ext4_io_end_t *io_end = bio->bi_private;
198 struct workqueue_struct *wq;
199 struct inode *inode;
200 unsigned long flags;
201 int i;
202 sector_t bi_sector = bio->bi_sector;
203
204 BUG_ON(!io_end);
205 bio->bi_private = NULL;
206 bio->bi_end_io = NULL;
207 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
208 error = 0;
209 bio_put(bio);
210
211 for (i = 0; i < io_end->num_io_pages; i++) {
212 struct page *page = io_end->pages[i]->p_page;
213 struct buffer_head *bh, *head;
214 loff_t offset;
215 loff_t io_end_offset;
216
217 if (error) {
218 SetPageError(page);
219 set_bit(AS_EIO, &page->mapping->flags);
220 head = page_buffers(page);
221 BUG_ON(!head);
222
223 io_end_offset = io_end->offset + io_end->size;
224
225 offset = (sector_t) page->index << PAGE_CACHE_SHIFT;
226 bh = head;
227 do {
228 if ((offset >= io_end->offset) &&
229 (offset+bh->b_size <= io_end_offset))
230 buffer_io_error(bh);
231
232 offset += bh->b_size;
233 bh = bh->b_this_page;
234 } while (bh != head);
235 }
236
237 put_io_page(io_end->pages[i]);
238 }
239 io_end->num_io_pages = 0;
240 inode = io_end->inode;
241
242 if (error) {
243 io_end->flag |= EXT4_IO_END_ERROR;
244 ext4_warning(inode->i_sb, "I/O error writing to inode %lu "
245 "(offset %llu size %ld starting block %llu)",
246 inode->i_ino,
247 (unsigned long long) io_end->offset,
248 (long) io_end->size,
249 (unsigned long long)
250 bi_sector >> (inode->i_blkbits - 9));
251 }
252
253 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
254 ext4_free_io_end(io_end);
255 return;
256 }
257
258 /* Add the io_end to per-inode completed io list*/
259 spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
260 list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
261 spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
262
263 wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
264 /* queue the work to convert unwritten extents to written */
265 queue_work(wq, &io_end->work);
266}
267
268void ext4_io_submit(struct ext4_io_submit *io)
269{
270 struct bio *bio = io->io_bio;
271
272 if (bio) {
273 bio_get(io->io_bio);
274 submit_bio(io->io_op, io->io_bio);
275 BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
276 bio_put(io->io_bio);
277 }
278 io->io_bio = NULL;
279 io->io_op = 0;
280 io->io_end = NULL;
281}
282
283static int io_submit_init(struct ext4_io_submit *io,
284 struct inode *inode,
285 struct writeback_control *wbc,
286 struct buffer_head *bh)
287{
288 ext4_io_end_t *io_end;
289 struct page *page = bh->b_page;
290 int nvecs = bio_get_nr_vecs(bh->b_bdev);
291 struct bio *bio;
292
293 io_end = ext4_init_io_end(inode, GFP_NOFS);
294 if (!io_end)
295 return -ENOMEM;
296 bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES));
297 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
298 bio->bi_bdev = bh->b_bdev;
299 bio->bi_private = io->io_end = io_end;
300 bio->bi_end_io = ext4_end_bio;
301
302 io_end->offset = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
303
304 io->io_bio = bio;
305 io->io_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
306 io->io_next_block = bh->b_blocknr;
307 return 0;
308}
309
310static int io_submit_add_bh(struct ext4_io_submit *io,
311 struct ext4_io_page *io_page,
312 struct inode *inode,
313 struct writeback_control *wbc,
314 struct buffer_head *bh)
315{
316 ext4_io_end_t *io_end;
317 int ret;
318
319 if (buffer_new(bh)) {
320 clear_buffer_new(bh);
321 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
322 }
323
324 if (!buffer_mapped(bh) || buffer_delay(bh)) {
325 if (!buffer_mapped(bh))
326 clear_buffer_dirty(bh);
327 if (io->io_bio)
328 ext4_io_submit(io);
329 return 0;
330 }
331
332 if (io->io_bio && bh->b_blocknr != io->io_next_block) {
333submit_and_retry:
334 ext4_io_submit(io);
335 }
336 if (io->io_bio == NULL) {
337 ret = io_submit_init(io, inode, wbc, bh);
338 if (ret)
339 return ret;
340 }
341 io_end = io->io_end;
342 if ((io_end->num_io_pages >= MAX_IO_PAGES) &&
343 (io_end->pages[io_end->num_io_pages-1] != io_page))
344 goto submit_and_retry;
345 if (buffer_uninit(bh))
346 ext4_set_io_unwritten_flag(inode, io_end);
347 io->io_end->size += bh->b_size;
348 io->io_next_block++;
349 ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
350 if (ret != bh->b_size)
351 goto submit_and_retry;
352 if ((io_end->num_io_pages == 0) ||
353 (io_end->pages[io_end->num_io_pages-1] != io_page)) {
354 io_end->pages[io_end->num_io_pages++] = io_page;
355 atomic_inc(&io_page->p_count);
356 }
357 return 0;
358}
359
360int ext4_bio_write_page(struct ext4_io_submit *io,
361 struct page *page,
362 int len,
363 struct writeback_control *wbc)
364{
365 struct inode *inode = page->mapping->host;
366 unsigned block_start, block_end, blocksize;
367 struct ext4_io_page *io_page;
368 struct buffer_head *bh, *head;
369 int ret = 0;
370
371 blocksize = 1 << inode->i_blkbits;
372
373 BUG_ON(!PageLocked(page));
374 BUG_ON(PageWriteback(page));
375
376 io_page = kmem_cache_alloc(io_page_cachep, GFP_NOFS);
377 if (!io_page) {
378 set_page_dirty(page);
379 unlock_page(page);
380 return -ENOMEM;
381 }
382 io_page->p_page = page;
383 atomic_set(&io_page->p_count, 1);
384 get_page(page);
385 set_page_writeback(page);
386 ClearPageError(page);
387
388 for (bh = head = page_buffers(page), block_start = 0;
389 bh != head || !block_start;
390 block_start = block_end, bh = bh->b_this_page) {
391
392 block_end = block_start + blocksize;
393 if (block_start >= len) {
394 /*
395 * Comments copied from block_write_full_page_endio:
396 *
397 * The page straddles i_size. It must be zeroed out on
398 * each and every writepage invocation because it may
399 * be mmapped. "A file is mapped in multiples of the
400 * page size. For a file that is not a multiple of
401 * the page size, the remaining memory is zeroed when
402 * mapped, and writes to that region are not written
403 * out to the file."
404 */
405 zero_user_segment(page, block_start, block_end);
406 clear_buffer_dirty(bh);
407 set_buffer_uptodate(bh);
408 continue;
409 }
410 clear_buffer_dirty(bh);
411 ret = io_submit_add_bh(io, io_page, inode, wbc, bh);
412 if (ret) {
413 /*
414 * We only get here on ENOMEM. Not much else
415 * we can do but mark the page as dirty, and
416 * better luck next time.
417 */
418 set_page_dirty(page);
419 break;
420 }
421 }
422 unlock_page(page);
423 /*
424 * If the page was truncated before we could do the writeback,
425 * or we had a memory allocation error while trying to write
426 * the first buffer head, we won't have submitted any pages for
427 * I/O. In that case we need to make sure we've cleared the
428 * PageWriteback bit from the page to prevent the system from
429 * wedging later on.
430 */
431 put_io_page(io_page);
432 return ret;
433}
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#include <linux/ratelimit.h>
29
30#include "ext4_jbd2.h"
31#include "xattr.h"
32#include "acl.h"
33
34static struct kmem_cache *io_end_cachep;
35
36int __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 return 0;
42}
43
44void ext4_exit_pageio(void)
45{
46 kmem_cache_destroy(io_end_cachep);
47}
48
49/*
50 * Print an buffer I/O error compatible with the fs/buffer.c. This
51 * provides compatibility with dmesg scrapers that look for a specific
52 * buffer I/O error message. We really need a unified error reporting
53 * structure to userspace ala Digital Unix's uerf system, but it's
54 * probably not going to happen in my lifetime, due to LKML politics...
55 */
56static void buffer_io_error(struct buffer_head *bh)
57{
58 char b[BDEVNAME_SIZE];
59 printk_ratelimited(KERN_ERR "Buffer I/O error on device %s, logical block %llu\n",
60 bdevname(bh->b_bdev, b),
61 (unsigned long long)bh->b_blocknr);
62}
63
64static void ext4_finish_bio(struct bio *bio)
65{
66 int i;
67 int error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
68 struct bio_vec *bvec;
69
70 bio_for_each_segment_all(bvec, bio, i) {
71 struct page *page = bvec->bv_page;
72 struct buffer_head *bh, *head;
73 unsigned bio_start = bvec->bv_offset;
74 unsigned bio_end = bio_start + bvec->bv_len;
75 unsigned under_io = 0;
76 unsigned long flags;
77
78 if (!page)
79 continue;
80
81 if (error) {
82 SetPageError(page);
83 set_bit(AS_EIO, &page->mapping->flags);
84 }
85 bh = head = page_buffers(page);
86 /*
87 * We check all buffers in the page under BH_Uptodate_Lock
88 * to avoid races with other end io clearing async_write flags
89 */
90 local_irq_save(flags);
91 bit_spin_lock(BH_Uptodate_Lock, &head->b_state);
92 do {
93 if (bh_offset(bh) < bio_start ||
94 bh_offset(bh) + bh->b_size > bio_end) {
95 if (buffer_async_write(bh))
96 under_io++;
97 continue;
98 }
99 clear_buffer_async_write(bh);
100 if (error)
101 buffer_io_error(bh);
102 } while ((bh = bh->b_this_page) != head);
103 bit_spin_unlock(BH_Uptodate_Lock, &head->b_state);
104 local_irq_restore(flags);
105 if (!under_io)
106 end_page_writeback(page);
107 }
108}
109
110static void ext4_release_io_end(ext4_io_end_t *io_end)
111{
112 struct bio *bio, *next_bio;
113
114 BUG_ON(!list_empty(&io_end->list));
115 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
116 WARN_ON(io_end->handle);
117
118 if (atomic_dec_and_test(&EXT4_I(io_end->inode)->i_ioend_count))
119 wake_up_all(ext4_ioend_wq(io_end->inode));
120
121 for (bio = io_end->bio; bio; bio = next_bio) {
122 next_bio = bio->bi_private;
123 ext4_finish_bio(bio);
124 bio_put(bio);
125 }
126 kmem_cache_free(io_end_cachep, io_end);
127}
128
129static void ext4_clear_io_unwritten_flag(ext4_io_end_t *io_end)
130{
131 struct inode *inode = io_end->inode;
132
133 io_end->flag &= ~EXT4_IO_END_UNWRITTEN;
134 /* Wake up anyone waiting on unwritten extent conversion */
135 if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
136 wake_up_all(ext4_ioend_wq(inode));
137}
138
139/*
140 * Check a range of space and convert unwritten extents to written. Note that
141 * we are protected from truncate touching same part of extent tree by the
142 * fact that truncate code waits for all DIO to finish (thus exclusion from
143 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
144 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
145 * completed (happens from ext4_free_ioend()).
146 */
147static int ext4_end_io(ext4_io_end_t *io)
148{
149 struct inode *inode = io->inode;
150 loff_t offset = io->offset;
151 ssize_t size = io->size;
152 handle_t *handle = io->handle;
153 int ret = 0;
154
155 ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
156 "list->prev 0x%p\n",
157 io, inode->i_ino, io->list.next, io->list.prev);
158
159 io->handle = NULL; /* Following call will use up the handle */
160 ret = ext4_convert_unwritten_extents(handle, inode, offset, size);
161 if (ret < 0) {
162 ext4_msg(inode->i_sb, KERN_EMERG,
163 "failed to convert unwritten extents to written "
164 "extents -- potential data loss! "
165 "(inode %lu, offset %llu, size %zd, error %d)",
166 inode->i_ino, offset, size, ret);
167 }
168 ext4_clear_io_unwritten_flag(io);
169 ext4_release_io_end(io);
170 return ret;
171}
172
173static void dump_completed_IO(struct inode *inode, struct list_head *head)
174{
175#ifdef EXT4FS_DEBUG
176 struct list_head *cur, *before, *after;
177 ext4_io_end_t *io, *io0, *io1;
178
179 if (list_empty(head))
180 return;
181
182 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
183 list_for_each_entry(io, head, list) {
184 cur = &io->list;
185 before = cur->prev;
186 io0 = container_of(before, ext4_io_end_t, list);
187 after = cur->next;
188 io1 = container_of(after, ext4_io_end_t, list);
189
190 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
191 io, inode->i_ino, io0, io1);
192 }
193#endif
194}
195
196/* Add the io_end to per-inode completed end_io list. */
197static void ext4_add_complete_io(ext4_io_end_t *io_end)
198{
199 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
200 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
201 struct workqueue_struct *wq;
202 unsigned long flags;
203
204 /* Only reserved conversions from writeback should enter here */
205 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
206 WARN_ON(!io_end->handle && sbi->s_journal);
207 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
208 wq = sbi->rsv_conversion_wq;
209 if (list_empty(&ei->i_rsv_conversion_list))
210 queue_work(wq, &ei->i_rsv_conversion_work);
211 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
212 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
213}
214
215static int ext4_do_flush_completed_IO(struct inode *inode,
216 struct list_head *head)
217{
218 ext4_io_end_t *io;
219 struct list_head unwritten;
220 unsigned long flags;
221 struct ext4_inode_info *ei = EXT4_I(inode);
222 int err, ret = 0;
223
224 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
225 dump_completed_IO(inode, head);
226 list_replace_init(head, &unwritten);
227 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
228
229 while (!list_empty(&unwritten)) {
230 io = list_entry(unwritten.next, ext4_io_end_t, list);
231 BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
232 list_del_init(&io->list);
233
234 err = ext4_end_io(io);
235 if (unlikely(!ret && err))
236 ret = err;
237 }
238 return ret;
239}
240
241/*
242 * work on completed IO, to convert unwritten extents to extents
243 */
244void ext4_end_io_rsv_work(struct work_struct *work)
245{
246 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
247 i_rsv_conversion_work);
248 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
249}
250
251ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
252{
253 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
254 if (io) {
255 atomic_inc(&EXT4_I(inode)->i_ioend_count);
256 io->inode = inode;
257 INIT_LIST_HEAD(&io->list);
258 atomic_set(&io->count, 1);
259 }
260 return io;
261}
262
263void ext4_put_io_end_defer(ext4_io_end_t *io_end)
264{
265 if (atomic_dec_and_test(&io_end->count)) {
266 if (!(io_end->flag & EXT4_IO_END_UNWRITTEN) || !io_end->size) {
267 ext4_release_io_end(io_end);
268 return;
269 }
270 ext4_add_complete_io(io_end);
271 }
272}
273
274int ext4_put_io_end(ext4_io_end_t *io_end)
275{
276 int err = 0;
277
278 if (atomic_dec_and_test(&io_end->count)) {
279 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
280 err = ext4_convert_unwritten_extents(io_end->handle,
281 io_end->inode, io_end->offset,
282 io_end->size);
283 io_end->handle = NULL;
284 ext4_clear_io_unwritten_flag(io_end);
285 }
286 ext4_release_io_end(io_end);
287 }
288 return err;
289}
290
291ext4_io_end_t *ext4_get_io_end(ext4_io_end_t *io_end)
292{
293 atomic_inc(&io_end->count);
294 return io_end;
295}
296
297/* BIO completion function for page writeback */
298static void ext4_end_bio(struct bio *bio, int error)
299{
300 ext4_io_end_t *io_end = bio->bi_private;
301 sector_t bi_sector = bio->bi_iter.bi_sector;
302
303 BUG_ON(!io_end);
304 bio->bi_end_io = NULL;
305 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
306 error = 0;
307
308 if (error) {
309 struct inode *inode = io_end->inode;
310
311 ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
312 "(offset %llu size %ld starting block %llu)",
313 error, inode->i_ino,
314 (unsigned long long) io_end->offset,
315 (long) io_end->size,
316 (unsigned long long)
317 bi_sector >> (inode->i_blkbits - 9));
318 mapping_set_error(inode->i_mapping, error);
319 }
320
321 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
322 /*
323 * Link bio into list hanging from io_end. We have to do it
324 * atomically as bio completions can be racing against each
325 * other.
326 */
327 bio->bi_private = xchg(&io_end->bio, bio);
328 ext4_put_io_end_defer(io_end);
329 } else {
330 /*
331 * Drop io_end reference early. Inode can get freed once
332 * we finish the bio.
333 */
334 ext4_put_io_end_defer(io_end);
335 ext4_finish_bio(bio);
336 bio_put(bio);
337 }
338}
339
340void ext4_io_submit(struct ext4_io_submit *io)
341{
342 struct bio *bio = io->io_bio;
343
344 if (bio) {
345 bio_get(io->io_bio);
346 submit_bio(io->io_op, io->io_bio);
347 BUG_ON(bio_flagged(io->io_bio, BIO_EOPNOTSUPP));
348 bio_put(io->io_bio);
349 }
350 io->io_bio = NULL;
351}
352
353void ext4_io_submit_init(struct ext4_io_submit *io,
354 struct writeback_control *wbc)
355{
356 io->io_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
357 io->io_bio = NULL;
358 io->io_end = NULL;
359}
360
361static int io_submit_init_bio(struct ext4_io_submit *io,
362 struct buffer_head *bh)
363{
364 int nvecs = bio_get_nr_vecs(bh->b_bdev);
365 struct bio *bio;
366
367 bio = bio_alloc(GFP_NOIO, min(nvecs, BIO_MAX_PAGES));
368 if (!bio)
369 return -ENOMEM;
370 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
371 bio->bi_bdev = bh->b_bdev;
372 bio->bi_end_io = ext4_end_bio;
373 bio->bi_private = ext4_get_io_end(io->io_end);
374 io->io_bio = bio;
375 io->io_next_block = bh->b_blocknr;
376 return 0;
377}
378
379static int io_submit_add_bh(struct ext4_io_submit *io,
380 struct inode *inode,
381 struct buffer_head *bh)
382{
383 int ret;
384
385 if (io->io_bio && bh->b_blocknr != io->io_next_block) {
386submit_and_retry:
387 ext4_io_submit(io);
388 }
389 if (io->io_bio == NULL) {
390 ret = io_submit_init_bio(io, bh);
391 if (ret)
392 return ret;
393 }
394 ret = bio_add_page(io->io_bio, bh->b_page, bh->b_size, bh_offset(bh));
395 if (ret != bh->b_size)
396 goto submit_and_retry;
397 io->io_next_block++;
398 return 0;
399}
400
401int ext4_bio_write_page(struct ext4_io_submit *io,
402 struct page *page,
403 int len,
404 struct writeback_control *wbc)
405{
406 struct inode *inode = page->mapping->host;
407 unsigned block_start, blocksize;
408 struct buffer_head *bh, *head;
409 int ret = 0;
410 int nr_submitted = 0;
411
412 blocksize = 1 << inode->i_blkbits;
413
414 BUG_ON(!PageLocked(page));
415 BUG_ON(PageWriteback(page));
416
417 set_page_writeback(page);
418 ClearPageError(page);
419
420 /*
421 * In the first loop we prepare and mark buffers to submit. We have to
422 * mark all buffers in the page before submitting so that
423 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
424 * on the first buffer finishes and we are still working on submitting
425 * the second buffer.
426 */
427 bh = head = page_buffers(page);
428 do {
429 block_start = bh_offset(bh);
430 if (block_start >= len) {
431 /*
432 * Comments copied from block_write_full_page_endio:
433 *
434 * The page straddles i_size. It must be zeroed out on
435 * each and every writepage invocation because it may
436 * be mmapped. "A file is mapped in multiples of the
437 * page size. For a file that is not a multiple of
438 * the page size, the remaining memory is zeroed when
439 * mapped, and writes to that region are not written
440 * out to the file."
441 */
442 zero_user_segment(page, block_start,
443 block_start + blocksize);
444 clear_buffer_dirty(bh);
445 set_buffer_uptodate(bh);
446 continue;
447 }
448 if (!buffer_dirty(bh) || buffer_delay(bh) ||
449 !buffer_mapped(bh) || buffer_unwritten(bh)) {
450 /* A hole? We can safely clear the dirty bit */
451 if (!buffer_mapped(bh))
452 clear_buffer_dirty(bh);
453 if (io->io_bio)
454 ext4_io_submit(io);
455 continue;
456 }
457 if (buffer_new(bh)) {
458 clear_buffer_new(bh);
459 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
460 }
461 set_buffer_async_write(bh);
462 } while ((bh = bh->b_this_page) != head);
463
464 /* Now submit buffers to write */
465 bh = head = page_buffers(page);
466 do {
467 if (!buffer_async_write(bh))
468 continue;
469 ret = io_submit_add_bh(io, inode, bh);
470 if (ret) {
471 /*
472 * We only get here on ENOMEM. Not much else
473 * we can do but mark the page as dirty, and
474 * better luck next time.
475 */
476 redirty_page_for_writepage(wbc, page);
477 break;
478 }
479 nr_submitted++;
480 clear_buffer_dirty(bh);
481 } while ((bh = bh->b_this_page) != head);
482
483 /* Error stopped previous loop? Clean up buffers... */
484 if (ret) {
485 do {
486 clear_buffer_async_write(bh);
487 bh = bh->b_this_page;
488 } while (bh != head);
489 }
490 unlock_page(page);
491 /* Nothing submitted - we have to end page writeback */
492 if (!nr_submitted)
493 end_page_writeback(page);
494 return ret;
495}