ocfs2: Make cached block reads the common case.
[linux-2.6.git] / 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_ERROR,
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                 if (buffer_jbd(bh)) {
151                         mlog(ML_ERROR,
152                              "the journal got the buffer while it was "
153                              "locked for io! (blocknr = %llu)\n",
154                              (unsigned long long)bh->b_blocknr);
155                         BUG();
156                 }
157
158                 wait_on_buffer(bh);
159                 if (!buffer_uptodate(bh)) {
160                         /* Status won't be cleared from here on out,
161                          * so we can safely record this and loop back
162                          * to cleanup the other buffers. */
163                         status = -EIO;
164                         put_bh(bh);
165                         bhs[i - 1] = NULL;
166                 }
167         }
168
169 bail:
170         return status;
171 }
172
173 int ocfs2_read_blocks(struct inode *inode, u64 block, int nr,
174                       struct buffer_head *bhs[], int flags)
175 {
176         int status = 0;
177         int i, ignore_cache = 0;
178         struct buffer_head *bh;
179
180         mlog_entry("(inode=%p, block=(%llu), nr=(%d), flags=%d)\n",
181                    inode, (unsigned long long)block, nr, flags);
182
183         BUG_ON(!inode);
184         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
185                (flags & OCFS2_BH_IGNORE_CACHE));
186
187         if (bhs == NULL) {
188                 status = -EINVAL;
189                 mlog_errno(status);
190                 goto bail;
191         }
192
193         if (nr < 0) {
194                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
195                 status = -EINVAL;
196                 mlog_errno(status);
197                 goto bail;
198         }
199
200         if (nr == 0) {
201                 mlog(ML_BH_IO, "No buffers will be read!\n");
202                 status = 0;
203                 goto bail;
204         }
205
206         mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
207         for (i = 0 ; i < nr ; i++) {
208                 if (bhs[i] == NULL) {
209                         bhs[i] = sb_getblk(inode->i_sb, block++);
210                         if (bhs[i] == NULL) {
211                                 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
212                                 status = -EIO;
213                                 mlog_errno(status);
214                                 goto bail;
215                         }
216                 }
217                 bh = bhs[i];
218                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
219
220                 /* There are three read-ahead cases here which we need to
221                  * be concerned with. All three assume a buffer has
222                  * previously been submitted with OCFS2_BH_READAHEAD
223                  * and it hasn't yet completed I/O.
224                  *
225                  * 1) The current request is sync to disk. This rarely
226                  *    happens these days, and never when performance
227                  *    matters - the code can just wait on the buffer
228                  *    lock and re-submit.
229                  *
230                  * 2) The current request is cached, but not
231                  *    readahead. ocfs2_buffer_uptodate() will return
232                  *    false anyway, so we'll wind up waiting on the
233                  *    buffer lock to do I/O. We re-check the request
234                  *    with after getting the lock to avoid a re-submit.
235                  *
236                  * 3) The current request is readahead (and so must
237                  *    also be a caching one). We short circuit if the
238                  *    buffer is locked (under I/O) and if it's in the
239                  *    uptodate cache. The re-check from #2 catches the
240                  *    case that the previous read-ahead completes just
241                  *    before our is-it-in-flight check.
242                  */
243
244                 if (!ignore_cache && !ocfs2_buffer_uptodate(inode, bh)) {
245                         mlog(ML_UPTODATE,
246                              "bh (%llu), inode %llu not uptodate\n",
247                              (unsigned long long)bh->b_blocknr,
248                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
249                         /* We're using ignore_cache here to say
250                          * "go to disk" */
251                         ignore_cache = 1;
252                 }
253
254                 /* XXX: Can we ever get this and *not* have the cached
255                  * flag set? */
256                 if (buffer_jbd(bh)) {
257                         if (ignore_cache)
258                                 mlog(ML_BH_IO, "trying to sync read a jbd "
259                                                "managed bh (blocknr = %llu)\n",
260                                      (unsigned long long)bh->b_blocknr);
261                         continue;
262                 }
263
264                 if (ignore_cache) {
265                         if (buffer_dirty(bh)) {
266                                 /* This should probably be a BUG, or
267                                  * at least return an error. */
268                                 mlog(ML_BH_IO, "asking me to sync read a dirty "
269                                                "buffer! (blocknr = %llu)\n",
270                                      (unsigned long long)bh->b_blocknr);
271                                 continue;
272                         }
273
274                         /* A read-ahead request was made - if the
275                          * buffer is already under read-ahead from a
276                          * previously submitted request than we are
277                          * done here. */
278                         if ((flags & OCFS2_BH_READAHEAD)
279                             && ocfs2_buffer_read_ahead(inode, bh))
280                                 continue;
281
282                         lock_buffer(bh);
283                         if (buffer_jbd(bh)) {
284 #ifdef CATCH_BH_JBD_RACES
285                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
286                                                "while I was in lock_buffer!",
287                                      (unsigned long long)bh->b_blocknr);
288                                 BUG();
289 #else
290                                 unlock_buffer(bh);
291                                 continue;
292 #endif
293                         }
294
295                         /* Re-check ocfs2_buffer_uptodate() as a
296                          * previously read-ahead buffer may have
297                          * completed I/O while we were waiting for the
298                          * buffer lock. */
299                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
300                             && !(flags & OCFS2_BH_READAHEAD)
301                             && ocfs2_buffer_uptodate(inode, bh)) {
302                                 unlock_buffer(bh);
303                                 continue;
304                         }
305
306                         clear_buffer_uptodate(bh);
307                         get_bh(bh); /* for end_buffer_read_sync() */
308                         bh->b_end_io = end_buffer_read_sync;
309                         submit_bh(READ, bh);
310                         continue;
311                 }
312         }
313
314         status = 0;
315
316         for (i = (nr - 1); i >= 0; i--) {
317                 bh = bhs[i];
318
319                 if (!(flags & OCFS2_BH_READAHEAD)) {
320                         /* We know this can't have changed as we hold the
321                          * inode sem. Avoid doing any work on the bh if the
322                          * journal has it. */
323                         if (!buffer_jbd(bh))
324                                 wait_on_buffer(bh);
325
326                         if (!buffer_uptodate(bh)) {
327                                 /* Status won't be cleared from here on out,
328                                  * so we can safely record this and loop back
329                                  * to cleanup the other buffers. Don't need to
330                                  * remove the clustered uptodate information
331                                  * for this bh as it's not marked locally
332                                  * uptodate. */
333                                 status = -EIO;
334                                 put_bh(bh);
335                                 bhs[i] = NULL;
336                                 continue;
337                         }
338                 }
339
340                 /* Always set the buffer in the cache, even if it was
341                  * a forced read, or read-ahead which hasn't yet
342                  * completed. */
343                 ocfs2_set_buffer_uptodate(inode, bh);
344         }
345         mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
346
347         mlog(ML_BH_IO, "block=(%llu), nr=(%d), cached=%s, flags=0x%x\n", 
348              (unsigned long long)block, nr,
349              ((flags & OCFS2_BH_IGNORE_CACHE) || ignore_cache) ? "no" : "yes",
350              flags);
351
352 bail:
353
354         mlog_exit(status);
355         return status;
356 }
357
358 /* Check whether the blkno is the super block or one of the backups. */
359 static void ocfs2_check_super_or_backup(struct super_block *sb,
360                                         sector_t blkno)
361 {
362         int i;
363         u64 backup_blkno;
364
365         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
366                 return;
367
368         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
369                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
370                 if (backup_blkno == blkno)
371                         return;
372         }
373
374         BUG();
375 }
376
377 /*
378  * Write super block and backups doesn't need to collaborate with journal,
379  * so we don't need to lock ip_io_mutex and inode doesn't need to bea passed
380  * into this function.
381  */
382 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
383                                 struct buffer_head *bh)
384 {
385         int ret = 0;
386
387         mlog_entry_void();
388
389         BUG_ON(buffer_jbd(bh));
390         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
391
392         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
393                 ret = -EROFS;
394                 goto out;
395         }
396
397         lock_buffer(bh);
398         set_buffer_uptodate(bh);
399
400         /* remove from dirty list before I/O. */
401         clear_buffer_dirty(bh);
402
403         get_bh(bh); /* for end_buffer_write_sync() */
404         bh->b_end_io = end_buffer_write_sync;
405         submit_bh(WRITE, bh);
406
407         wait_on_buffer(bh);
408
409         if (!buffer_uptodate(bh)) {
410                 ret = -EIO;
411                 put_bh(bh);
412         }
413
414 out:
415         mlog_exit(ret);
416         return ret;
417 }