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/highuid.h>
12#include <linux/pagemap.h>
13#include <linux/quotaops.h>
14#include <linux/string.h>
15#include <linux/buffer_head.h>
16#include <linux/writeback.h>
17#include <linux/pagevec.h>
18#include <linux/mpage.h>
19#include <linux/namei.h>
20#include <linux/uio.h>
21#include <linux/bio.h>
22#include <linux/workqueue.h>
23#include <linux/kernel.h>
24#include <linux/slab.h>
25#include <linux/mm.h>
26#include <linux/backing-dev.h>
27
28#include "ext4_jbd2.h"
29#include "xattr.h"
30#include "acl.h"
31
32static struct kmem_cache *io_end_cachep;
33
34int __init ext4_init_pageio(void)
35{
36 io_end_cachep = KMEM_CACHE(ext4_io_end, SLAB_RECLAIM_ACCOUNT);
37 if (io_end_cachep == NULL)
38 return -ENOMEM;
39 return 0;
40}
41
42void ext4_exit_pageio(void)
43{
44 kmem_cache_destroy(io_end_cachep);
45}
46
47/*
48 * Print an buffer I/O error compatible with the fs/buffer.c. This
49 * provides compatibility with dmesg scrapers that look for a specific
50 * buffer I/O error message. We really need a unified error reporting
51 * structure to userspace ala Digital Unix's uerf system, but it's
52 * probably not going to happen in my lifetime, due to LKML politics...
53 */
54static void buffer_io_error(struct buffer_head *bh)
55{
56 printk_ratelimited(KERN_ERR "Buffer I/O error on device %pg, logical block %llu\n",
57 bh->b_bdev,
58 (unsigned long long)bh->b_blocknr);
59}
60
61static void ext4_finish_bio(struct bio *bio)
62{
63 int i;
64 struct bio_vec *bvec;
65
66 bio_for_each_segment_all(bvec, bio, i) {
67 struct page *page = bvec->bv_page;
68#ifdef CONFIG_EXT4_FS_ENCRYPTION
69 struct page *data_page = NULL;
70 struct ext4_crypto_ctx *ctx = NULL;
71#endif
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#ifdef CONFIG_EXT4_FS_ENCRYPTION
82 if (!page->mapping) {
83 /* The bounce data pages are unmapped. */
84 data_page = page;
85 ctx = (struct ext4_crypto_ctx *)page_private(data_page);
86 page = ctx->w.control_page;
87 }
88#endif
89
90 if (bio->bi_error) {
91 SetPageError(page);
92 set_bit(AS_EIO, &page->mapping->flags);
93 }
94 bh = head = page_buffers(page);
95 /*
96 * We check all buffers in the page under BH_Uptodate_Lock
97 * to avoid races with other end io clearing async_write flags
98 */
99 local_irq_save(flags);
100 bit_spin_lock(BH_Uptodate_Lock, &head->b_state);
101 do {
102 if (bh_offset(bh) < bio_start ||
103 bh_offset(bh) + bh->b_size > bio_end) {
104 if (buffer_async_write(bh))
105 under_io++;
106 continue;
107 }
108 clear_buffer_async_write(bh);
109 if (bio->bi_error)
110 buffer_io_error(bh);
111 } while ((bh = bh->b_this_page) != head);
112 bit_spin_unlock(BH_Uptodate_Lock, &head->b_state);
113 local_irq_restore(flags);
114 if (!under_io) {
115#ifdef CONFIG_EXT4_FS_ENCRYPTION
116 if (ctx)
117 ext4_restore_control_page(data_page);
118#endif
119 end_page_writeback(page);
120 }
121 }
122}
123
124static void ext4_release_io_end(ext4_io_end_t *io_end)
125{
126 struct bio *bio, *next_bio;
127
128 BUG_ON(!list_empty(&io_end->list));
129 BUG_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
130 WARN_ON(io_end->handle);
131
132 for (bio = io_end->bio; bio; bio = next_bio) {
133 next_bio = bio->bi_private;
134 ext4_finish_bio(bio);
135 bio_put(bio);
136 }
137 kmem_cache_free(io_end_cachep, io_end);
138}
139
140/*
141 * Check a range of space and convert unwritten extents to written. Note that
142 * we are protected from truncate touching same part of extent tree by the
143 * fact that truncate code waits for all DIO to finish (thus exclusion from
144 * direct IO is achieved) and also waits for PageWriteback bits. Thus we
145 * cannot get to ext4_ext_truncate() before all IOs overlapping that range are
146 * completed (happens from ext4_free_ioend()).
147 */
148static int ext4_end_io(ext4_io_end_t *io)
149{
150 struct inode *inode = io->inode;
151 loff_t offset = io->offset;
152 ssize_t size = io->size;
153 handle_t *handle = io->handle;
154 int ret = 0;
155
156 ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
157 "list->prev 0x%p\n",
158 io, inode->i_ino, io->list.next, io->list.prev);
159
160 io->handle = NULL; /* Following call will use up the handle */
161 ret = ext4_convert_unwritten_extents(handle, inode, offset, size);
162 if (ret < 0) {
163 ext4_msg(inode->i_sb, KERN_EMERG,
164 "failed to convert unwritten extents to written "
165 "extents -- potential data loss! "
166 "(inode %lu, offset %llu, size %zd, error %d)",
167 inode->i_ino, offset, size, ret);
168 }
169 ext4_clear_io_unwritten_flag(io);
170 ext4_release_io_end(io);
171 return ret;
172}
173
174static void dump_completed_IO(struct inode *inode, struct list_head *head)
175{
176#ifdef EXT4FS_DEBUG
177 struct list_head *cur, *before, *after;
178 ext4_io_end_t *io, *io0, *io1;
179
180 if (list_empty(head))
181 return;
182
183 ext4_debug("Dump inode %lu completed io list\n", inode->i_ino);
184 list_for_each_entry(io, head, list) {
185 cur = &io->list;
186 before = cur->prev;
187 io0 = container_of(before, ext4_io_end_t, list);
188 after = cur->next;
189 io1 = container_of(after, ext4_io_end_t, list);
190
191 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
192 io, inode->i_ino, io0, io1);
193 }
194#endif
195}
196
197/* Add the io_end to per-inode completed end_io list. */
198static void ext4_add_complete_io(ext4_io_end_t *io_end)
199{
200 struct ext4_inode_info *ei = EXT4_I(io_end->inode);
201 struct ext4_sb_info *sbi = EXT4_SB(io_end->inode->i_sb);
202 struct workqueue_struct *wq;
203 unsigned long flags;
204
205 /* Only reserved conversions from writeback should enter here */
206 WARN_ON(!(io_end->flag & EXT4_IO_END_UNWRITTEN));
207 WARN_ON(!io_end->handle && sbi->s_journal);
208 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
209 wq = sbi->rsv_conversion_wq;
210 if (list_empty(&ei->i_rsv_conversion_list))
211 queue_work(wq, &ei->i_rsv_conversion_work);
212 list_add_tail(&io_end->list, &ei->i_rsv_conversion_list);
213 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
214}
215
216static int ext4_do_flush_completed_IO(struct inode *inode,
217 struct list_head *head)
218{
219 ext4_io_end_t *io;
220 struct list_head unwritten;
221 unsigned long flags;
222 struct ext4_inode_info *ei = EXT4_I(inode);
223 int err, ret = 0;
224
225 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
226 dump_completed_IO(inode, head);
227 list_replace_init(head, &unwritten);
228 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
229
230 while (!list_empty(&unwritten)) {
231 io = list_entry(unwritten.next, ext4_io_end_t, list);
232 BUG_ON(!(io->flag & EXT4_IO_END_UNWRITTEN));
233 list_del_init(&io->list);
234
235 err = ext4_end_io(io);
236 if (unlikely(!ret && err))
237 ret = err;
238 }
239 return ret;
240}
241
242/*
243 * work on completed IO, to convert unwritten extents to extents
244 */
245void ext4_end_io_rsv_work(struct work_struct *work)
246{
247 struct ext4_inode_info *ei = container_of(work, struct ext4_inode_info,
248 i_rsv_conversion_work);
249 ext4_do_flush_completed_IO(&ei->vfs_inode, &ei->i_rsv_conversion_list);
250}
251
252ext4_io_end_t *ext4_init_io_end(struct inode *inode, gfp_t flags)
253{
254 ext4_io_end_t *io = kmem_cache_zalloc(io_end_cachep, flags);
255 if (io) {
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)
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
306 if (bio->bi_error) {
307 struct inode *inode = io_end->inode;
308
309 ext4_warning(inode->i_sb, "I/O error %d writing to inode %lu "
310 "(offset %llu size %ld starting block %llu)",
311 bio->bi_error, inode->i_ino,
312 (unsigned long long) io_end->offset,
313 (long) io_end->size,
314 (unsigned long long)
315 bi_sector >> (inode->i_blkbits - 9));
316 mapping_set_error(inode->i_mapping, bio->bi_error);
317 }
318
319 if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
320 /*
321 * Link bio into list hanging from io_end. We have to do it
322 * atomically as bio completions can be racing against each
323 * other.
324 */
325 bio->bi_private = xchg(&io_end->bio, bio);
326 ext4_put_io_end_defer(io_end);
327 } else {
328 /*
329 * Drop io_end reference early. Inode can get freed once
330 * we finish the bio.
331 */
332 ext4_put_io_end_defer(io_end);
333 ext4_finish_bio(bio);
334 bio_put(bio);
335 }
336}
337
338void ext4_io_submit(struct ext4_io_submit *io)
339{
340 struct bio *bio = io->io_bio;
341
342 if (bio) {
343 int io_op = io->io_wbc->sync_mode == WB_SYNC_ALL ?
344 WRITE_SYNC : WRITE;
345 bio_get(io->io_bio);
346 submit_bio(io_op, io->io_bio);
347 bio_put(io->io_bio);
348 }
349 io->io_bio = NULL;
350}
351
352void ext4_io_submit_init(struct ext4_io_submit *io,
353 struct writeback_control *wbc)
354{
355 io->io_wbc = wbc;
356 io->io_bio = NULL;
357 io->io_end = NULL;
358}
359
360static int io_submit_init_bio(struct ext4_io_submit *io,
361 struct buffer_head *bh)
362{
363 struct bio *bio;
364
365 bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
366 if (!bio)
367 return -ENOMEM;
368 wbc_init_bio(io->io_wbc, bio);
369 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
370 bio->bi_bdev = bh->b_bdev;
371 bio->bi_end_io = ext4_end_bio;
372 bio->bi_private = ext4_get_io_end(io->io_end);
373 io->io_bio = bio;
374 io->io_next_block = bh->b_blocknr;
375 return 0;
376}
377
378static int io_submit_add_bh(struct ext4_io_submit *io,
379 struct inode *inode,
380 struct page *page,
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, page, bh->b_size, bh_offset(bh));
395 if (ret != bh->b_size)
396 goto submit_and_retry;
397 wbc_account_io(io->io_wbc, page, bh->b_size);
398 io->io_next_block++;
399 return 0;
400}
401
402int ext4_bio_write_page(struct ext4_io_submit *io,
403 struct page *page,
404 int len,
405 struct writeback_control *wbc,
406 bool keep_towrite)
407{
408 struct page *data_page = NULL;
409 struct inode *inode = page->mapping->host;
410 unsigned block_start, blocksize;
411 struct buffer_head *bh, *head;
412 int ret = 0;
413 int nr_submitted = 0;
414 int nr_to_submit = 0;
415
416 blocksize = 1 << inode->i_blkbits;
417
418 BUG_ON(!PageLocked(page));
419 BUG_ON(PageWriteback(page));
420
421 if (keep_towrite)
422 set_page_writeback_keepwrite(page);
423 else
424 set_page_writeback(page);
425 ClearPageError(page);
426
427 /*
428 * Comments copied from block_write_full_page:
429 *
430 * The page straddles i_size. It must be zeroed out on each and every
431 * writepage invocation because it may be mmapped. "A file is mapped
432 * in multiples of the page size. For a file that is not a multiple of
433 * the page size, the remaining memory is zeroed when mapped, and
434 * writes to that region are not written out to the file."
435 */
436 if (len < PAGE_SIZE)
437 zero_user_segment(page, len, PAGE_SIZE);
438 /*
439 * In the first loop we prepare and mark buffers to submit. We have to
440 * mark all buffers in the page before submitting so that
441 * end_page_writeback() cannot be called from ext4_bio_end_io() when IO
442 * on the first buffer finishes and we are still working on submitting
443 * the second buffer.
444 */
445 bh = head = page_buffers(page);
446 do {
447 block_start = bh_offset(bh);
448 if (block_start >= len) {
449 clear_buffer_dirty(bh);
450 set_buffer_uptodate(bh);
451 continue;
452 }
453 if (!buffer_dirty(bh) || buffer_delay(bh) ||
454 !buffer_mapped(bh) || buffer_unwritten(bh)) {
455 /* A hole? We can safely clear the dirty bit */
456 if (!buffer_mapped(bh))
457 clear_buffer_dirty(bh);
458 if (io->io_bio)
459 ext4_io_submit(io);
460 continue;
461 }
462 if (buffer_new(bh)) {
463 clear_buffer_new(bh);
464 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
465 }
466 set_buffer_async_write(bh);
467 nr_to_submit++;
468 } while ((bh = bh->b_this_page) != head);
469
470 bh = head = page_buffers(page);
471
472 if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode) &&
473 nr_to_submit) {
474 gfp_t gfp_flags = GFP_NOFS;
475
476 retry_encrypt:
477 data_page = ext4_encrypt(inode, page, gfp_flags);
478 if (IS_ERR(data_page)) {
479 ret = PTR_ERR(data_page);
480 if (ret == -ENOMEM && wbc->sync_mode == WB_SYNC_ALL) {
481 if (io->io_bio) {
482 ext4_io_submit(io);
483 congestion_wait(BLK_RW_ASYNC, HZ/50);
484 }
485 gfp_flags |= __GFP_NOFAIL;
486 goto retry_encrypt;
487 }
488 data_page = NULL;
489 goto out;
490 }
491 }
492
493 /* Now submit buffers to write */
494 do {
495 if (!buffer_async_write(bh))
496 continue;
497 ret = io_submit_add_bh(io, inode,
498 data_page ? data_page : page, bh);
499 if (ret) {
500 /*
501 * We only get here on ENOMEM. Not much else
502 * we can do but mark the page as dirty, and
503 * better luck next time.
504 */
505 break;
506 }
507 nr_submitted++;
508 clear_buffer_dirty(bh);
509 } while ((bh = bh->b_this_page) != head);
510
511 /* Error stopped previous loop? Clean up buffers... */
512 if (ret) {
513 out:
514 if (data_page)
515 ext4_restore_control_page(data_page);
516 printk_ratelimited(KERN_ERR "%s: ret = %d\n", __func__, ret);
517 redirty_page_for_writepage(wbc, page);
518 do {
519 clear_buffer_async_write(bh);
520 bh = bh->b_this_page;
521 } while (bh != head);
522 }
523 unlock_page(page);
524 /* Nothing submitted - we have to end page writeback */
525 if (!nr_submitted)
526 end_page_writeback(page);
527 return ret;
528}