Merge branch 'i2c-for-ben' of git://git.kernel.org/pub/scm/linux/kernel/git/tmlind...
[linux-2.6] / fs / ocfs2 / buffer_head_io.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * io.c
5  *
6  * Buffer cache handling
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30
31 #include <cluster/masklog.h>
32
33 #include "ocfs2.h"
34
35 #include "alloc.h"
36 #include "inode.h"
37 #include "journal.h"
38 #include "uptodate.h"
39
40 #include "buffer_head_io.h"
41
42 int ocfs2_write_block(struct ocfs2_super *osb, struct buffer_head *bh,
43                       struct inode *inode)
44 {
45         int ret = 0;
46
47         mlog_entry("(bh->b_blocknr = %llu, inode=%p)\n",
48                    (unsigned long long)bh->b_blocknr, inode);
49
50         BUG_ON(bh->b_blocknr < OCFS2_SUPER_BLOCK_BLKNO);
51         BUG_ON(buffer_jbd(bh));
52
53         /* No need to check for a soft readonly file system here. non
54          * journalled writes are only ever done on system files which
55          * can get modified during recovery even if read-only. */
56         if (ocfs2_is_hard_readonly(osb)) {
57                 ret = -EROFS;
58                 goto out;
59         }
60
61         mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
62
63         lock_buffer(bh);
64         set_buffer_uptodate(bh);
65
66         /* remove from dirty list before I/O. */
67         clear_buffer_dirty(bh);
68
69         get_bh(bh); /* for end_buffer_write_sync() */
70         bh->b_end_io = end_buffer_write_sync;
71         submit_bh(WRITE, bh);
72
73         wait_on_buffer(bh);
74
75         if (buffer_uptodate(bh)) {
76                 ocfs2_set_buffer_uptodate(inode, bh);
77         } else {
78                 /* We don't need to remove the clustered uptodate
79                  * information for this bh as it's not marked locally
80                  * uptodate. */
81                 ret = -EIO;
82                 put_bh(bh);
83         }
84
85         mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
86 out:
87         mlog_exit(ret);
88         return ret;
89 }
90
91 int ocfs2_read_blocks_sync(struct ocfs2_super *osb, u64 block,
92                            unsigned int nr, struct buffer_head *bhs[])
93 {
94         int status = 0;
95         unsigned int i;
96         struct buffer_head *bh;
97
98         if (!nr) {
99                 mlog(ML_BH_IO, "No buffers will be read!\n");
100                 goto bail;
101         }
102
103         for (i = 0 ; i < nr ; i++) {
104                 if (bhs[i] == NULL) {
105                         bhs[i] = sb_getblk(osb->sb, block++);
106                         if (bhs[i] == NULL) {
107                                 status = -EIO;
108                                 mlog_errno(status);
109                                 goto bail;
110                         }
111                 }
112                 bh = bhs[i];
113
114                 if (buffer_jbd(bh)) {
115                         mlog(ML_BH_IO,
116                              "trying to sync read a jbd "
117                              "managed bh (blocknr = %llu), skipping\n",
118                              (unsigned long long)bh->b_blocknr);
119                         continue;
120                 }
121
122                 if (buffer_dirty(bh)) {
123                         /* This should probably be a BUG, or
124                          * at least return an error. */
125                         mlog(ML_ERROR,
126                              "trying to sync read a dirty "
127                              "buffer! (blocknr = %llu), skipping\n",
128                              (unsigned long long)bh->b_blocknr);
129                         continue;
130                 }
131
132                 lock_buffer(bh);
133                 if (buffer_jbd(bh)) {
134                         mlog(ML_ERROR,
135                              "block %llu had the JBD bit set "
136                              "while I was in lock_buffer!",
137                              (unsigned long long)bh->b_blocknr);
138                         BUG();
139                 }
140
141                 clear_buffer_uptodate(bh);
142                 get_bh(bh); /* for end_buffer_read_sync() */
143                 bh->b_end_io = end_buffer_read_sync;
144                 submit_bh(READ, bh);
145         }
146
147         for (i = nr; i > 0; i--) {
148                 bh = bhs[i - 1];
149
150                 /* No need to wait on the buffer if it's managed by JBD. */
151                 if (!buffer_jbd(bh))
152                         wait_on_buffer(bh);
153
154                 if (!buffer_uptodate(bh)) {
155                         /* Status won't be cleared from here on out,
156                          * so we can safely record this and loop back
157                          * to cleanup the other buffers. */
158                         status = -EIO;
159                         put_bh(bh);
160                         bhs[i - 1] = NULL;
161                 }
162         }
163
164 bail:
165         return status;
166 }
167
168 int ocfs2_read_blocks(struct inode *inode, u64 block, int nr,
169                       struct buffer_head *bhs[], int flags)
170 {
171         int status = 0;
172         int i, ignore_cache = 0;
173         struct buffer_head *bh;
174
175         mlog_entry("(inode=%p, block=(%llu), nr=(%d), flags=%d)\n",
176                    inode, (unsigned long long)block, nr, flags);
177
178         BUG_ON(!inode);
179         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
180                (flags & OCFS2_BH_IGNORE_CACHE));
181
182         if (bhs == NULL) {
183                 status = -EINVAL;
184                 mlog_errno(status);
185                 goto bail;
186         }
187
188         if (nr < 0) {
189                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
190                 status = -EINVAL;
191                 mlog_errno(status);
192                 goto bail;
193         }
194
195         if (nr == 0) {
196                 mlog(ML_BH_IO, "No buffers will be read!\n");
197                 status = 0;
198                 goto bail;
199         }
200
201         mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
202         for (i = 0 ; i < nr ; i++) {
203                 if (bhs[i] == NULL) {
204                         bhs[i] = sb_getblk(inode->i_sb, block++);
205                         if (bhs[i] == NULL) {
206                                 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
207                                 status = -EIO;
208                                 mlog_errno(status);
209                                 goto bail;
210                         }
211                 }
212                 bh = bhs[i];
213                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
214
215                 /* There are three read-ahead cases here which we need to
216                  * be concerned with. All three assume a buffer has
217                  * previously been submitted with OCFS2_BH_READAHEAD
218                  * and it hasn't yet completed I/O.
219                  *
220                  * 1) The current request is sync to disk. This rarely
221                  *    happens these days, and never when performance
222                  *    matters - the code can just wait on the buffer
223                  *    lock and re-submit.
224                  *
225                  * 2) The current request is cached, but not
226                  *    readahead. ocfs2_buffer_uptodate() will return
227                  *    false anyway, so we'll wind up waiting on the
228                  *    buffer lock to do I/O. We re-check the request
229                  *    with after getting the lock to avoid a re-submit.
230                  *
231                  * 3) The current request is readahead (and so must
232                  *    also be a caching one). We short circuit if the
233                  *    buffer is locked (under I/O) and if it's in the
234                  *    uptodate cache. The re-check from #2 catches the
235                  *    case that the previous read-ahead completes just
236                  *    before our is-it-in-flight check.
237                  */
238
239                 if (!ignore_cache && !ocfs2_buffer_uptodate(inode, bh)) {
240                         mlog(ML_UPTODATE,
241                              "bh (%llu), inode %llu not uptodate\n",
242                              (unsigned long long)bh->b_blocknr,
243                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
244                         /* We're using ignore_cache here to say
245                          * "go to disk" */
246                         ignore_cache = 1;
247                 }
248
249                 if (buffer_jbd(bh)) {
250                         if (ignore_cache)
251                                 mlog(ML_BH_IO, "trying to sync read a jbd "
252                                                "managed bh (blocknr = %llu)\n",
253                                      (unsigned long long)bh->b_blocknr);
254                         continue;
255                 }
256
257                 if (ignore_cache) {
258                         if (buffer_dirty(bh)) {
259                                 /* This should probably be a BUG, or
260                                  * at least return an error. */
261                                 mlog(ML_BH_IO, "asking me to sync read a dirty "
262                                                "buffer! (blocknr = %llu)\n",
263                                      (unsigned long long)bh->b_blocknr);
264                                 continue;
265                         }
266
267                         /* A read-ahead request was made - if the
268                          * buffer is already under read-ahead from a
269                          * previously submitted request than we are
270                          * done here. */
271                         if ((flags & OCFS2_BH_READAHEAD)
272                             && ocfs2_buffer_read_ahead(inode, bh))
273                                 continue;
274
275                         lock_buffer(bh);
276                         if (buffer_jbd(bh)) {
277 #ifdef CATCH_BH_JBD_RACES
278                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
279                                                "while I was in lock_buffer!",
280                                      (unsigned long long)bh->b_blocknr);
281                                 BUG();
282 #else
283                                 unlock_buffer(bh);
284                                 continue;
285 #endif
286                         }
287
288                         /* Re-check ocfs2_buffer_uptodate() as a
289                          * previously read-ahead buffer may have
290                          * completed I/O while we were waiting for the
291                          * buffer lock. */
292                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
293                             && !(flags & OCFS2_BH_READAHEAD)
294                             && ocfs2_buffer_uptodate(inode, bh)) {
295                                 unlock_buffer(bh);
296                                 continue;
297                         }
298
299                         clear_buffer_uptodate(bh);
300                         get_bh(bh); /* for end_buffer_read_sync() */
301                         bh->b_end_io = end_buffer_read_sync;
302                         submit_bh(READ, bh);
303                         continue;
304                 }
305         }
306
307         status = 0;
308
309         for (i = (nr - 1); i >= 0; i--) {
310                 bh = bhs[i];
311
312                 if (!(flags & OCFS2_BH_READAHEAD)) {
313                         /* We know this can't have changed as we hold the
314                          * inode sem. Avoid doing any work on the bh if the
315                          * journal has it. */
316                         if (!buffer_jbd(bh))
317                                 wait_on_buffer(bh);
318
319                         if (!buffer_uptodate(bh)) {
320                                 /* Status won't be cleared from here on out,
321                                  * so we can safely record this and loop back
322                                  * to cleanup the other buffers. Don't need to
323                                  * remove the clustered uptodate information
324                                  * for this bh as it's not marked locally
325                                  * uptodate. */
326                                 status = -EIO;
327                                 put_bh(bh);
328                                 bhs[i] = NULL;
329                                 continue;
330                         }
331                 }
332
333                 /* Always set the buffer in the cache, even if it was
334                  * a forced read, or read-ahead which hasn't yet
335                  * completed. */
336                 ocfs2_set_buffer_uptodate(inode, bh);
337         }
338         mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
339
340         mlog(ML_BH_IO, "block=(%llu), nr=(%d), cached=%s, flags=0x%x\n", 
341              (unsigned long long)block, nr,
342              ((flags & OCFS2_BH_IGNORE_CACHE) || ignore_cache) ? "no" : "yes",
343              flags);
344
345 bail:
346
347         mlog_exit(status);
348         return status;
349 }
350
351 /* Check whether the blkno is the super block or one of the backups. */
352 static void ocfs2_check_super_or_backup(struct super_block *sb,
353                                         sector_t blkno)
354 {
355         int i;
356         u64 backup_blkno;
357
358         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
359                 return;
360
361         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
362                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
363                 if (backup_blkno == blkno)
364                         return;
365         }
366
367         BUG();
368 }
369
370 /*
371  * Write super block and backups doesn't need to collaborate with journal,
372  * so we don't need to lock ip_io_mutex and inode doesn't need to bea passed
373  * into this function.
374  */
375 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
376                                 struct buffer_head *bh)
377 {
378         int ret = 0;
379
380         mlog_entry_void();
381
382         BUG_ON(buffer_jbd(bh));
383         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
384
385         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
386                 ret = -EROFS;
387                 goto out;
388         }
389
390         lock_buffer(bh);
391         set_buffer_uptodate(bh);
392
393         /* remove from dirty list before I/O. */
394         clear_buffer_dirty(bh);
395
396         get_bh(bh); /* for end_buffer_write_sync() */
397         bh->b_end_io = end_buffer_write_sync;
398         submit_bh(WRITE, bh);
399
400         wait_on_buffer(bh);
401
402         if (!buffer_uptodate(bh)) {
403                 ret = -EIO;
404                 put_bh(bh);
405         }
406
407 out:
408         mlog_exit(ret);
409         return ret;
410 }