Merge branch 'hpet' into release
[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 /*
43  * Bits on bh->b_state used by ocfs2.
44  *
45  * These MUST be after the JBD2 bits.  Hence, we use BH_JBDPrivateStart.
46  */
47 enum ocfs2_state_bits {
48         BH_NeedsValidate = BH_JBDPrivateStart,
49 };
50
51 /* Expand the magic b_state functions */
52 BUFFER_FNS(NeedsValidate, needs_validate);
53
54 int ocfs2_write_block(struct ocfs2_super *osb, struct buffer_head *bh,
55                       struct inode *inode)
56 {
57         int ret = 0;
58
59         mlog_entry("(bh->b_blocknr = %llu, inode=%p)\n",
60                    (unsigned long long)bh->b_blocknr, inode);
61
62         BUG_ON(bh->b_blocknr < OCFS2_SUPER_BLOCK_BLKNO);
63         BUG_ON(buffer_jbd(bh));
64
65         /* No need to check for a soft readonly file system here. non
66          * journalled writes are only ever done on system files which
67          * can get modified during recovery even if read-only. */
68         if (ocfs2_is_hard_readonly(osb)) {
69                 ret = -EROFS;
70                 goto out;
71         }
72
73         mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
74
75         lock_buffer(bh);
76         set_buffer_uptodate(bh);
77
78         /* remove from dirty list before I/O. */
79         clear_buffer_dirty(bh);
80
81         get_bh(bh); /* for end_buffer_write_sync() */
82         bh->b_end_io = end_buffer_write_sync;
83         submit_bh(WRITE, bh);
84
85         wait_on_buffer(bh);
86
87         if (buffer_uptodate(bh)) {
88                 ocfs2_set_buffer_uptodate(inode, bh);
89         } else {
90                 /* We don't need to remove the clustered uptodate
91                  * information for this bh as it's not marked locally
92                  * uptodate. */
93                 ret = -EIO;
94                 put_bh(bh);
95         }
96
97         mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
98 out:
99         mlog_exit(ret);
100         return ret;
101 }
102
103 int ocfs2_read_blocks_sync(struct ocfs2_super *osb, u64 block,
104                            unsigned int nr, struct buffer_head *bhs[])
105 {
106         int status = 0;
107         unsigned int i;
108         struct buffer_head *bh;
109
110         if (!nr) {
111                 mlog(ML_BH_IO, "No buffers will be read!\n");
112                 goto bail;
113         }
114
115         for (i = 0 ; i < nr ; i++) {
116                 if (bhs[i] == NULL) {
117                         bhs[i] = sb_getblk(osb->sb, block++);
118                         if (bhs[i] == NULL) {
119                                 status = -EIO;
120                                 mlog_errno(status);
121                                 goto bail;
122                         }
123                 }
124                 bh = bhs[i];
125
126                 if (buffer_jbd(bh)) {
127                         mlog(ML_BH_IO,
128                              "trying to sync read a jbd "
129                              "managed bh (blocknr = %llu), skipping\n",
130                              (unsigned long long)bh->b_blocknr);
131                         continue;
132                 }
133
134                 if (buffer_dirty(bh)) {
135                         /* This should probably be a BUG, or
136                          * at least return an error. */
137                         mlog(ML_ERROR,
138                              "trying to sync read a dirty "
139                              "buffer! (blocknr = %llu), skipping\n",
140                              (unsigned long long)bh->b_blocknr);
141                         continue;
142                 }
143
144                 lock_buffer(bh);
145                 if (buffer_jbd(bh)) {
146                         mlog(ML_ERROR,
147                              "block %llu had the JBD bit set "
148                              "while I was in lock_buffer!",
149                              (unsigned long long)bh->b_blocknr);
150                         BUG();
151                 }
152
153                 clear_buffer_uptodate(bh);
154                 get_bh(bh); /* for end_buffer_read_sync() */
155                 bh->b_end_io = end_buffer_read_sync;
156                 submit_bh(READ, bh);
157         }
158
159         for (i = nr; i > 0; i--) {
160                 bh = bhs[i - 1];
161
162                 /* No need to wait on the buffer if it's managed by JBD. */
163                 if (!buffer_jbd(bh))
164                         wait_on_buffer(bh);
165
166                 if (!buffer_uptodate(bh)) {
167                         /* Status won't be cleared from here on out,
168                          * so we can safely record this and loop back
169                          * to cleanup the other buffers. */
170                         status = -EIO;
171                         put_bh(bh);
172                         bhs[i - 1] = NULL;
173                 }
174         }
175
176 bail:
177         return status;
178 }
179
180 int ocfs2_read_blocks(struct inode *inode, u64 block, int nr,
181                       struct buffer_head *bhs[], int flags,
182                       int (*validate)(struct super_block *sb,
183                                       struct buffer_head *bh))
184 {
185         int status = 0;
186         int i, ignore_cache = 0;
187         struct buffer_head *bh;
188
189         mlog_entry("(inode=%p, block=(%llu), nr=(%d), flags=%d)\n",
190                    inode, (unsigned long long)block, nr, flags);
191
192         BUG_ON(!inode);
193         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
194                (flags & OCFS2_BH_IGNORE_CACHE));
195
196         if (bhs == NULL) {
197                 status = -EINVAL;
198                 mlog_errno(status);
199                 goto bail;
200         }
201
202         if (nr < 0) {
203                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
204                 status = -EINVAL;
205                 mlog_errno(status);
206                 goto bail;
207         }
208
209         if (nr == 0) {
210                 mlog(ML_BH_IO, "No buffers will be read!\n");
211                 status = 0;
212                 goto bail;
213         }
214
215         mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
216         for (i = 0 ; i < nr ; i++) {
217                 if (bhs[i] == NULL) {
218                         bhs[i] = sb_getblk(inode->i_sb, block++);
219                         if (bhs[i] == NULL) {
220                                 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
221                                 status = -EIO;
222                                 mlog_errno(status);
223                                 goto bail;
224                         }
225                 }
226                 bh = bhs[i];
227                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
228
229                 /* There are three read-ahead cases here which we need to
230                  * be concerned with. All three assume a buffer has
231                  * previously been submitted with OCFS2_BH_READAHEAD
232                  * and it hasn't yet completed I/O.
233                  *
234                  * 1) The current request is sync to disk. This rarely
235                  *    happens these days, and never when performance
236                  *    matters - the code can just wait on the buffer
237                  *    lock and re-submit.
238                  *
239                  * 2) The current request is cached, but not
240                  *    readahead. ocfs2_buffer_uptodate() will return
241                  *    false anyway, so we'll wind up waiting on the
242                  *    buffer lock to do I/O. We re-check the request
243                  *    with after getting the lock to avoid a re-submit.
244                  *
245                  * 3) The current request is readahead (and so must
246                  *    also be a caching one). We short circuit if the
247                  *    buffer is locked (under I/O) and if it's in the
248                  *    uptodate cache. The re-check from #2 catches the
249                  *    case that the previous read-ahead completes just
250                  *    before our is-it-in-flight check.
251                  */
252
253                 if (!ignore_cache && !ocfs2_buffer_uptodate(inode, bh)) {
254                         mlog(ML_UPTODATE,
255                              "bh (%llu), inode %llu not uptodate\n",
256                              (unsigned long long)bh->b_blocknr,
257                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
258                         /* We're using ignore_cache here to say
259                          * "go to disk" */
260                         ignore_cache = 1;
261                 }
262
263                 if (buffer_jbd(bh)) {
264                         if (ignore_cache)
265                                 mlog(ML_BH_IO, "trying to sync read a jbd "
266                                                "managed bh (blocknr = %llu)\n",
267                                      (unsigned long long)bh->b_blocknr);
268                         continue;
269                 }
270
271                 if (ignore_cache) {
272                         if (buffer_dirty(bh)) {
273                                 /* This should probably be a BUG, or
274                                  * at least return an error. */
275                                 mlog(ML_BH_IO, "asking me to sync read a dirty "
276                                                "buffer! (blocknr = %llu)\n",
277                                      (unsigned long long)bh->b_blocknr);
278                                 continue;
279                         }
280
281                         /* A read-ahead request was made - if the
282                          * buffer is already under read-ahead from a
283                          * previously submitted request than we are
284                          * done here. */
285                         if ((flags & OCFS2_BH_READAHEAD)
286                             && ocfs2_buffer_read_ahead(inode, bh))
287                                 continue;
288
289                         lock_buffer(bh);
290                         if (buffer_jbd(bh)) {
291 #ifdef CATCH_BH_JBD_RACES
292                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
293                                                "while I was in lock_buffer!",
294                                      (unsigned long long)bh->b_blocknr);
295                                 BUG();
296 #else
297                                 unlock_buffer(bh);
298                                 continue;
299 #endif
300                         }
301
302                         /* Re-check ocfs2_buffer_uptodate() as a
303                          * previously read-ahead buffer may have
304                          * completed I/O while we were waiting for the
305                          * buffer lock. */
306                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
307                             && !(flags & OCFS2_BH_READAHEAD)
308                             && ocfs2_buffer_uptodate(inode, bh)) {
309                                 unlock_buffer(bh);
310                                 continue;
311                         }
312
313                         clear_buffer_uptodate(bh);
314                         get_bh(bh); /* for end_buffer_read_sync() */
315                         if (validate)
316                                 set_buffer_needs_validate(bh);
317                         bh->b_end_io = end_buffer_read_sync;
318                         submit_bh(READ, bh);
319                         continue;
320                 }
321         }
322
323         status = 0;
324
325         for (i = (nr - 1); i >= 0; i--) {
326                 bh = bhs[i];
327
328                 if (!(flags & OCFS2_BH_READAHEAD)) {
329                         /* We know this can't have changed as we hold the
330                          * inode sem. Avoid doing any work on the bh if the
331                          * journal has it. */
332                         if (!buffer_jbd(bh))
333                                 wait_on_buffer(bh);
334
335                         if (!buffer_uptodate(bh)) {
336                                 /* Status won't be cleared from here on out,
337                                  * so we can safely record this and loop back
338                                  * to cleanup the other buffers. Don't need to
339                                  * remove the clustered uptodate information
340                                  * for this bh as it's not marked locally
341                                  * uptodate. */
342                                 status = -EIO;
343                                 put_bh(bh);
344                                 bhs[i] = NULL;
345                                 continue;
346                         }
347
348                         if (buffer_needs_validate(bh)) {
349                                 /* We never set NeedsValidate if the
350                                  * buffer was held by the journal, so
351                                  * that better not have changed */
352                                 BUG_ON(buffer_jbd(bh));
353                                 clear_buffer_needs_validate(bh);
354                                 status = validate(inode->i_sb, bh);
355                                 if (status) {
356                                         put_bh(bh);
357                                         bhs[i] = NULL;
358                                         continue;
359                                 }
360                         }
361                 }
362
363                 /* Always set the buffer in the cache, even if it was
364                  * a forced read, or read-ahead which hasn't yet
365                  * completed. */
366                 ocfs2_set_buffer_uptodate(inode, bh);
367         }
368         mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
369
370         mlog(ML_BH_IO, "block=(%llu), nr=(%d), cached=%s, flags=0x%x\n", 
371              (unsigned long long)block, nr,
372              ((flags & OCFS2_BH_IGNORE_CACHE) || ignore_cache) ? "no" : "yes",
373              flags);
374
375 bail:
376
377         mlog_exit(status);
378         return status;
379 }
380
381 /* Check whether the blkno is the super block or one of the backups. */
382 static void ocfs2_check_super_or_backup(struct super_block *sb,
383                                         sector_t blkno)
384 {
385         int i;
386         u64 backup_blkno;
387
388         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
389                 return;
390
391         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
392                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
393                 if (backup_blkno == blkno)
394                         return;
395         }
396
397         BUG();
398 }
399
400 /*
401  * Write super block and backups doesn't need to collaborate with journal,
402  * so we don't need to lock ip_io_mutex and inode doesn't need to bea passed
403  * into this function.
404  */
405 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
406                                 struct buffer_head *bh)
407 {
408         int ret = 0;
409
410         mlog_entry_void();
411
412         BUG_ON(buffer_jbd(bh));
413         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
414
415         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
416                 ret = -EROFS;
417                 goto out;
418         }
419
420         lock_buffer(bh);
421         set_buffer_uptodate(bh);
422
423         /* remove from dirty list before I/O. */
424         clear_buffer_dirty(bh);
425
426         get_bh(bh); /* for end_buffer_write_sync() */
427         bh->b_end_io = end_buffer_write_sync;
428         submit_bh(WRITE, bh);
429
430         wait_on_buffer(bh);
431
432         if (!buffer_uptodate(bh)) {
433                 ret = -EIO;
434                 put_bh(bh);
435         }
436
437 out:
438         mlog_exit(ret);
439         return ret;
440 }