Whamcloud - gitweb
b=2306
[fs/lustre-release.git] / lustre / lvfs / fsfilt_ext3.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  lustre/lib/fsfilt_ext3.c
5  *  Lustre filesystem abstraction routines
6  *
7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8  *   Author: Andreas Dilger <adilger@clusterfs.com>
9  *
10  *   This file is part of Lustre, http://www.lustre.org.
11  *
12  *   Lustre is free software; you can redistribute it and/or
13  *   modify it under the terms of version 2 of the GNU General Public
14  *   License as published by the Free Software Foundation.
15  *
16  *   Lustre is distributed in the hope that it will be useful,
17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
19  *   GNU General Public License for more details.
20  *
21  *   You should have received a copy of the GNU General Public License
22  *   along with Lustre; if not, write to the Free Software
23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24  */
25
26 #define DEBUG_SUBSYSTEM S_FILTER
27
28 #include <linux/fs.h>
29 #include <linux/jbd.h>
30 #include <linux/slab.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/ext3_fs.h>
34 #include <linux/ext3_jbd.h>
35 #include <linux/version.h>
36 /* XXX ugh */
37 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
38  #include <linux/ext3_xattr.h>
39 #else
40  #include <linux/../../fs/ext3/xattr.h>
41 #endif
42 #include <linux/kp30.h>
43 #include <linux/lustre_fsfilt.h>
44 #include <linux/obd.h>
45 #include <linux/obd_class.h>
46 #include <linux/module.h>
47
48 static kmem_cache_t *fcb_cache;
49 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
50
51 struct fsfilt_cb_data {
52         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
53         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
54         struct obd_device *cb_obd;      /* MDS/OBD completion device */
55         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
56         void *cb_data;                  /* MDS/OST completion function data */
57 };
58
59 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
60 #define EXT3_XATTR_INDEX_TRUSTED        4
61 #endif
62 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
63
64 #define EXT3_XATTR_INDEX_LUSTRE         5                         /* old */
65 #define XATTR_LUSTRE_MDS_OBJID          "system.lustre_mds_objid" /* old */
66
67 /*
68  * We don't currently need any additional blocks for rmdir and
69  * unlink transactions because we are storing the OST oa_id inside
70  * the inode (which we will be changing anyways as part of this
71  * transaction).
72  */
73 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private)
74 {
75         /* For updates to the last recieved file */
76         int nblocks = EXT3_DATA_TRANS_BLOCKS;
77         void *handle;
78
79         if (current->journal_info) {
80                 CDEBUG(D_INODE, "increasing refcount on %p\n", current->journal_info);
81                 goto journal_start;
82         }
83
84         switch(op) {
85         case FSFILT_OP_CREATE_LOG:
86                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
87                 op = FSFILT_OP_CREATE;
88                 break;
89         case FSFILT_OP_UNLINK_LOG:
90                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
91                 op = FSFILT_OP_UNLINK;
92                 break;
93         }
94
95         switch(op) {
96         case FSFILT_OP_RMDIR:
97         case FSFILT_OP_UNLINK:
98                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
99                 break;
100         case FSFILT_OP_RENAME:
101                 /* modify additional directory */
102                 nblocks += EXT3_DATA_TRANS_BLOCKS;
103                 /* no break */
104         case FSFILT_OP_SYMLINK:
105                 /* additional block + block bitmap + GDT for long symlink */
106                 nblocks += 3;
107                 /* no break */
108         case FSFILT_OP_CREATE:
109         case FSFILT_OP_MKDIR:
110         case FSFILT_OP_MKNOD:
111                 /* modify one inode + block bitmap + GDT */
112                 nblocks += 3;
113                 /* no break */
114         case FSFILT_OP_LINK:
115                 /* modify parent directory */
116                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
117                 break;
118         case FSFILT_OP_SETATTR:
119                 /* Setattr on inode */
120                 nblocks += 1;
121                 break;
122         default: CERROR("unknown transaction start op %d\n", op);
123                  LBUG();
124         }
125
126         LASSERT(current->journal_info == desc_private);
127
128  journal_start:
129         lock_kernel();
130         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
131         unlock_kernel();
132
133         if (!IS_ERR(handle))
134                 LASSERT(current->journal_info == handle);
135         return handle;
136 }
137
138 /*
139  * Calculate the number of buffer credits needed to write multiple pages in
140  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
141  * doesn't have a nice API for calculating this sort of thing in advance.
142  *
143  * See comment above ext3_writepage_trans_blocks for details.  We assume
144  * no data journaling is being done, but it does allow for all of the pages
145  * being non-contiguous.  If we are guaranteed contiguous pages we could
146  * reduce the number of (d)indirect blocks a lot.
147  *
148  * With N blocks per page and P pages, for each inode we have at most:
149  * N*P indirect
150  * min(N*P, blocksize/4 + 1) dindirect blocks
151  * niocount tindirect
152  *
153  * For the entire filesystem, we have at most:
154  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
155  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
156  * objcount inode blocks
157  * 1 superblock
158  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
159  *
160  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
161  */
162 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso)
163 {
164         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
165         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
166         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
167         int nbitmaps = 0;
168         int ngdblocks = 0;
169         int needed = objcount + 1;
170         int i;
171
172         for (i = 0; i < objcount; i++, fso++) {
173                 int nblocks = fso->fso_bufcnt * blockpp;
174                 int ndindirect = min(nblocks, addrpp + 1);
175                 int nindir = nblocks + ndindirect + 1;
176
177                 nbitmaps += nindir + nblocks;
178                 ngdblocks += nindir + nblocks;
179
180                 needed += nindir;
181         }
182
183         /* Assumes ext3 and ext3 have same sb_info layout at the start. */
184         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
185                 nbitmaps = EXT3_SB(sb)->s_groups_count;
186         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
187                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
188
189         needed += nbitmaps + ngdblocks;
190
191         /* last_rcvd update */
192         needed += EXT3_DATA_TRANS_BLOCKS;
193
194 #ifdef CONFIG_QUOTA
195         /* We assume that there will be 1 bit set in s_dquot.flags for each
196          * quota file that is active.  This is at least true for now.
197          */
198         needed += hweight32(sb_any_quota_enabled(sb)) *
199                 EXT3_SINGLEDATA_TRANS_BLOCKS;
200 #endif
201
202         return needed;
203 }
204
205 /* We have to start a huge journal transaction here to hold all of the
206  * metadata for the pages being written here.  This is necessitated by
207  * the fact that we do lots of prepare_write operations before we do
208  * any of the matching commit_write operations, so even if we split
209  * up to use "smaller" transactions none of them could complete until
210  * all of them were opened.  By having a single journal transaction,
211  * we eliminate duplicate reservations for common blocks like the
212  * superblock and group descriptors or bitmaps.
213  *
214  * We will start the transaction here, but each prepare_write will
215  * add a refcount to the transaction, and each commit_write will
216  * remove a refcount.  The transaction will be closed when all of
217  * the pages have been written.
218  */
219 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
220                                    int niocount, void *desc_private)
221 {
222         journal_t *journal;
223         handle_t *handle;
224         int needed;
225         ENTRY;
226
227         LASSERT(current->journal_info == desc_private);
228         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
229         needed = fsfilt_ext3_credits_needed(objcount, fso);
230
231         /* The number of blocks we could _possibly_ dirty can very large.
232          * We reduce our request if it is absurd (and we couldn't get that
233          * many credits for a single handle anyways).
234          *
235          * At some point we have to limit the size of I/Os sent at one time,
236          * increase the size of the journal, or we have to calculate the
237          * actual journal requirements more carefully by checking all of
238          * the blocks instead of being maximally pessimistic.  It remains to
239          * be seen if this is a real problem or not.
240          */
241         if (needed > journal->j_max_transaction_buffers) {
242                 CERROR("want too many journal credits (%d) using %d instead\n",
243                        needed, journal->j_max_transaction_buffers);
244                 needed = journal->j_max_transaction_buffers;
245         }
246
247         lock_kernel();
248         handle = journal_start(journal, needed);
249         unlock_kernel();
250         if (IS_ERR(handle)) {
251                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
252                        PTR_ERR(handle));
253         } else {
254                 LASSERT(handle->h_buffer_credits >= needed);
255                 LASSERT(current->journal_info == handle);
256         }
257
258         RETURN(handle);
259 }
260
261 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
262 {
263         int rc;
264         handle_t *handle = h;
265
266         LASSERT(current->journal_info == handle);
267         if (force_sync)
268                 handle->h_sync = 1; /* recovery likes this */
269
270         lock_kernel();
271         rc = journal_stop(handle);
272         unlock_kernel();
273
274         // LASSERT(current->journal_info == NULL);
275         return rc;
276 }
277
278 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
279                                         void **wait_handle)
280 {
281         transaction_t *transaction;
282         unsigned long tid, rtid;
283         handle_t *handle = h;
284         journal_t *journal;
285         int rc;
286
287         LASSERT(current->journal_info == handle);
288
289         lock_kernel();
290         transaction = handle->h_transaction;
291         journal = transaction->t_journal;
292         tid = transaction->t_tid;
293         /* we don't want to be blocked */
294         handle->h_sync = 0;
295         rc = journal_stop(handle);
296         if (rc) {
297                 CERROR("error while stopping transaction: %d\n", rc);
298                 unlock_kernel();
299                 return rc;
300         }
301
302         rtid = log_start_commit(journal, transaction);
303         if (rtid != tid)
304                 CERROR("strange race: %lu != %lu\n",
305                        (unsigned long) tid, (unsigned long) rtid);
306         unlock_kernel();
307
308         *wait_handle = (void *) tid;
309         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
310         return 0;
311 }
312
313 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
314 {
315         tid_t tid = (tid_t)(long)h;
316
317         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
318         if (is_journal_aborted(EXT3_JOURNAL(inode)))
319                 return -EIO;
320
321         log_wait_commit(EXT3_JOURNAL(inode), tid);
322
323         return 0;
324 }
325
326 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
327                                struct iattr *iattr, int do_trunc)
328 {
329         struct inode *inode = dentry->d_inode;
330         int rc;
331
332         lock_kernel();
333
334         /* A _really_ horrible hack to avoid removing the data stored
335          * in the block pointers; this is really the "small" stripe MD data.
336          * We can avoid further hackery by virtue of the MDS file size being
337          * zero all the time (which doesn't invoke block truncate at unlink
338          * time), so we assert we never change the MDS file size from zero. */
339         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
340                 /* ATTR_SIZE would invoke truncate: clear it */
341                 iattr->ia_valid &= ~ATTR_SIZE;
342                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
343
344                 /* make sure _something_ gets set - so new inode
345                  * goes to disk (probably won't work over XFS */
346                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
347                         iattr->ia_valid |= ATTR_MODE;
348                         iattr->ia_mode = inode->i_mode;
349                 }
350         }
351
352         /* Don't allow setattr to change file type */
353         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
354
355         /* We set these flags on the client, but have already checked perms
356          * so don't confuse inode_change_ok. */
357         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
358
359         if (inode->i_op->setattr) {
360                 rc = inode->i_op->setattr(dentry, iattr);
361         } else {
362                 rc = inode_change_ok(inode, iattr);
363                 if (!rc)
364                         rc = inode_setattr(inode, iattr);
365         }
366
367         unlock_kernel();
368
369         return rc;
370 }
371
372 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
373                                  unsigned int cmd, unsigned long arg)
374 {
375         int rc = 0;
376         ENTRY;
377
378         if (inode->i_fop->ioctl)
379                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
380         else
381                 RETURN(-ENOTTY);
382
383         RETURN(rc);
384 }
385
386 #undef INLINE_EA
387 #undef OLD_EA
388 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
389                               void *lmm, int lmm_size)
390 {
391         int rc, old_ea = 0;
392
393 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
394         /* Nasty hack city - store stripe MD data in the block pointers if
395          * it will fit, because putting it in an EA currently kills the MDS
396          * performance.  We'll fix this with "fast EAs" in the future.
397          */
398         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
399                                             sizeof(EXT3_I(inode)->i_data[0])) {
400                 unsigned old_size = EXT3_I(inode)->i_data[0];
401                 if (old_size != 0) {
402                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
403                         CERROR("setting EA on %lu/%u again... interesting\n",
404                                inode->i_ino, inode->i_generation);
405                 }
406
407                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
408                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
409                 mark_inode_dirty(inode);
410                 return 0;
411         }
412 #endif
413 #ifdef OLD_EA
414         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
415         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
416                 CWARN("setting EA on %lu/%u again... interesting\n",
417                        inode->i_ino, inode->i_generation);
418                 old_ea = 1;
419         }
420
421         lock_kernel();
422         /* this can go away before 1.0.  For bug 2097 testing only. */
423         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
424                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
425 #else
426         lock_kernel();
427         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
428                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
429
430         /* This tries to delete the old-format LOV EA, but only as long as we
431          * have successfully saved the new-format LOV EA (we can always try
432          * the conversion again the next time the file is accessed).  It is
433          * possible (although unlikely) that the new-format LOV EA couldn't be
434          * saved because it ran out of space but we would need a file striped
435          * over least 123 OSTs before the two EAs filled a 4kB block.
436          *
437          * This can be removed when all filesystems have converted to the
438          * new EA format, but otherwise adds little if any overhead.  If we
439          * wanted backward compatibility for existing files, we could keep
440          * the old EA around for a while but we'd have to clean it up later. */
441         if (rc >= 0 && old_ea) {
442                 int err = ext3_xattr_set_handle(handle, inode,
443                                                 EXT3_XATTR_INDEX_LUSTRE,
444                                                 XATTR_LUSTRE_MDS_OBJID,
445                                                 NULL, 0, 0);
446                 if (err)
447                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
448                                inode->i_ino, inode->i_generation, err);
449         }
450 #endif
451         unlock_kernel();
452
453         if (rc)
454                 CERROR("error adding MD data to inode %lu: rc = %d\n",
455                        inode->i_ino, rc);
456         return rc;
457 }
458
459 /* Must be called with i_sem held */
460 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
461 {
462         int rc;
463
464         LASSERT(down_trylock(&inode->i_sem) != 0);
465         lock_kernel();
466         /* Keep support for reading "inline EAs" until we convert
467          * users over to new format entirely.  See bug 841/2097. */
468         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
469                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
470                 void *handle;
471
472                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
473                 if (lmm) {
474                         if (size > lmm_size) {
475                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
476                                        inode->i_ino, inode->i_generation,
477                                        size, lmm_size);
478                                 return -ERANGE;
479                         }
480                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
481                 }
482
483 #ifndef INLINE_EA
484                 /* migrate LOV EA data to external block - keep same format */
485                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
486                       inode->i_ino, inode->i_generation);
487
488                 handle = journal_start(EXT3_JOURNAL(inode),
489                                        EXT3_XATTR_TRANS_BLOCKS);
490                 if (!IS_ERR(handle)) {
491                         int err;
492                         rc = fsfilt_ext3_set_md(inode, handle,
493                                                 &EXT3_I(inode)->i_data[1],size);
494                         if (rc == 0) {
495                                 memset(EXT3_I(inode)->i_data, 0,
496                                        sizeof(EXT3_I(inode)->i_data));
497                                 mark_inode_dirty(inode);
498                         }
499                         err = journal_stop(handle);
500                         if (err && rc == 0)
501                                 rc = err;
502                 } else {
503                         rc = PTR_ERR(handle);
504                 }
505 #endif
506                 unlock_kernel();
507                 return size;
508         }
509
510         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
511                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
512         /* try old EA type if new one failed - MDS will convert it for us */
513         if (rc == -ENODATA) {
514                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
515                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
516                        inode->i_ino, rc);
517
518                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
519                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
520         }
521         unlock_kernel();
522
523         /* This gives us the MD size */
524         if (lmm == NULL)
525                 return (rc == -ENODATA) ? 0 : rc;
526
527         if (rc < 0) {
528                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
529                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
530                        inode->i_ino, rc);
531                 memset(lmm, 0, lmm_size);
532                 return (rc == -ENODATA) ? 0 : rc;
533         }
534
535         return rc;
536 }
537
538 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
539                                     loff_t *off)
540 {
541         struct inode *inode = file->f_dentry->d_inode;
542         int rc = 0;
543
544         if (S_ISREG(inode->i_mode))
545                 rc = file->f_op->read(file, buf, count, off);
546         else {
547                 const int blkbits = inode->i_sb->s_blocksize_bits;
548                 const int blksize = inode->i_sb->s_blocksize;
549
550                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
551                        count, inode->i_ino, *off);
552                 while (count > 0) {
553                         struct buffer_head *bh;
554
555                         bh = NULL;
556                         if (*off < inode->i_size) {
557                                 int err = 0;
558
559                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
560                                                 0, &err);
561
562                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
563
564                                 if (bh) {
565                                         memcpy(buf, bh->b_data, blksize);
566                                         brelse(bh);
567                                 } else if (err) {
568                                         /* XXX in theory we should just fake
569                                          * this buffer and continue like ext3,
570                                          * especially if this is a partial read
571                                          */
572                                         CERROR("error read dir %lu+%llu: %d\n",
573                                                inode->i_ino, *off, err);
574                                         RETURN(err);
575                                 }
576                         }
577                         if (!bh) {
578                                 struct ext3_dir_entry_2 *fake = (void *)buf;
579
580                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
581                                 memset(fake, 0, sizeof(*fake));
582                                 fake->rec_len = cpu_to_le32(blksize);
583                         }
584                         count -= blksize;
585                         buf += blksize;
586                         *off += blksize;
587                         rc += blksize;
588                 }
589         }
590
591         return rc;
592 }
593
594 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
595 {
596         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
597
598         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
599
600         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
601         atomic_dec(&fcb_cache_count);
602 }
603
604 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
605                                       void *handle, fsfilt_cb_t cb_func,
606                                       void *cb_data)
607 {
608         struct fsfilt_cb_data *fcb;
609
610         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
611         if (fcb == NULL)
612                 RETURN(-ENOMEM);
613
614         atomic_inc(&fcb_cache_count);
615         fcb->cb_func = cb_func;
616         fcb->cb_obd = obd;
617         fcb->cb_last_rcvd = last_rcvd;
618         fcb->cb_data = cb_data;
619
620         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
621         lock_kernel();
622         journal_callback_set(handle, fsfilt_ext3_cb_func,
623                              (struct journal_callback *)fcb);
624         unlock_kernel();
625
626         return 0;
627 }
628
629 /*
630  * We need to hack the return value for the free inode counts because
631  * the current EA code requires one filesystem block per inode with EAs,
632  * so it is possible to run out of blocks before we run out of inodes.
633  *
634  * This can be removed when the ext3 EA code is fixed.
635  */
636 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
637 {
638         struct kstatfs sfs;
639         int rc = vfs_statfs(sb, &sfs);
640
641         if (!rc && sfs.f_bfree < sfs.f_ffree) {
642                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
643                 sfs.f_ffree = sfs.f_bfree;
644         }
645
646         statfs_pack(osfs, &sfs);
647         return rc;
648 }
649
650 static int fsfilt_ext3_sync(struct super_block *sb)
651 {
652         return ext3_force_commit(sb);
653 }
654
655 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
656                                unsigned long *blocks, int *created, int create);
657 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
658                                unsigned long *blocks, int *created, int create)
659 {
660         return ext3_map_inode_page(inode, page, blocks, created, create);
661 }
662
663 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
664                                int nblocks, loff_t newsize);
665 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
666                                       int nblocks, loff_t newsize)
667 {
668         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
669 }
670
671 static int fsfilt_ext3_read_record(struct file * file, void *buf,
672                                    int size, loff_t *offs)
673 {
674         struct inode *inode = file->f_dentry->d_inode;
675         unsigned long block;
676         struct buffer_head *bh;
677         int err, blocksize, csize, boffs;
678
679         /* prevent reading after eof */
680         lock_kernel();
681         if (inode->i_size < *offs + size) {
682                 size = inode->i_size - *offs;
683                 unlock_kernel();
684                 if (size < 0) {
685                         CERROR("size %llu is too short for read %u@%llu\n",
686                                inode->i_size, size, *offs);
687                         return -EIO;
688                 } else if (size == 0) {
689                         return 0;
690                 }
691         } else {
692                 unlock_kernel();
693         }
694
695         blocksize = 1 << inode->i_blkbits;
696
697         while (size > 0) {
698                 block = *offs >> inode->i_blkbits;
699                 boffs = *offs & (blocksize - 1);
700                 csize = min(blocksize - boffs, size);
701                 bh = ext3_bread(NULL, inode, block, 0, &err);
702                 if (!bh) {
703                         CERROR("can't read block: %d\n", err);
704                         return err;
705                 }
706
707                 memcpy(buf, bh->b_data + boffs, csize);
708                 brelse(bh);
709
710                 *offs += csize;
711                 buf += csize;
712                 size -= csize;
713         }
714         return 0;
715 }
716
717 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
718                                     loff_t *offs, int force_sync)
719 {
720         struct buffer_head *bh = NULL;
721         unsigned long block;
722         struct inode *inode = file->f_dentry->d_inode;
723         loff_t old_size = inode->i_size, offset = *offs;
724         loff_t new_size = inode->i_size;
725         journal_t *journal;
726         handle_t *handle;
727         int err, block_count = 0, blocksize, size, boffs;
728
729         /* Determine how many transaction credits are needed */
730         blocksize = 1 << inode->i_blkbits;
731         block_count = (*offs & (blocksize - 1)) + bufsize;
732         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
733
734         journal = EXT3_SB(inode->i_sb)->s_journal;
735         lock_kernel();
736         handle = journal_start(journal,
737                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
738         unlock_kernel();
739         if (IS_ERR(handle)) {
740                 CERROR("can't start transaction\n");
741                 return PTR_ERR(handle);
742         }
743
744         while (bufsize > 0) {
745                 if (bh != NULL)
746                         brelse(bh);
747
748                 block = offset >> inode->i_blkbits;
749                 boffs = offset & (blocksize - 1);
750                 size = min(blocksize - boffs, bufsize);
751                 bh = ext3_bread(handle, inode, block, 1, &err);
752                 if (!bh) {
753                         CERROR("can't read/create block: %d\n", err);
754                         goto out;
755                 }
756
757                 err = ext3_journal_get_write_access(handle, bh);
758                 if (err) {
759                         CERROR("journal_get_write_access() returned error %d\n",
760                                err);
761                         goto out;
762                 }
763                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
764                 memcpy(bh->b_data + boffs, buf, size);
765                 err = ext3_journal_dirty_metadata(handle, bh);
766                 if (err) {
767                         CERROR("journal_dirty_metadata() returned error %d\n",
768                                err);
769                         goto out;
770                 }
771                 if (offset + size > new_size)
772                         new_size = offset + size;
773                 offset += size;
774                 bufsize -= size;
775                 buf += size;
776         }
777
778         if (force_sync)
779                 handle->h_sync = 1; /* recovery likes this */
780 out:
781         if (bh)
782                 brelse(bh);
783
784         /* correct in-core and on-disk sizes */
785         if (new_size > inode->i_size) {
786                 lock_kernel();
787                 if (new_size > inode->i_size)
788                         inode->i_size = new_size;
789                 if (inode->i_size > EXT3_I(inode)->i_disksize)
790                         EXT3_I(inode)->i_disksize = inode->i_size;
791                 if (inode->i_size > old_size)
792                         mark_inode_dirty(inode);
793                 unlock_kernel();
794         }
795
796         lock_kernel();
797         journal_stop(handle);
798         unlock_kernel();
799
800         if (err == 0)
801                 *offs = offset;
802         return err;
803 }
804
805 static int fsfilt_ext3_setup(struct super_block *sb)
806 {
807 #if 0
808         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
809         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
810 #endif
811 #ifdef S_PDIROPS
812         CWARN("Enabling PDIROPS\n");
813         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
814         sb->s_flags |= S_PDIROPS;
815 #endif
816         return 0;
817 }
818
819 static struct fsfilt_operations fsfilt_ext3_ops = {
820         fs_type:                "ext3",
821         fs_owner:               THIS_MODULE,
822         fs_start:               fsfilt_ext3_start,
823         fs_brw_start:           fsfilt_ext3_brw_start,
824         fs_commit:              fsfilt_ext3_commit,
825         fs_commit_async:        fsfilt_ext3_commit_async,
826         fs_commit_wait:         fsfilt_ext3_commit_wait,
827         fs_setattr:             fsfilt_ext3_setattr,
828         fs_iocontrol:           fsfilt_ext3_iocontrol,
829         fs_set_md:              fsfilt_ext3_set_md,
830         fs_get_md:              fsfilt_ext3_get_md,
831         fs_readpage:            fsfilt_ext3_readpage,
832         fs_add_journal_cb:      fsfilt_ext3_add_journal_cb,
833         fs_statfs:              fsfilt_ext3_statfs,
834         fs_sync:                fsfilt_ext3_sync,
835         fs_map_inode_page:      fsfilt_ext3_map_inode_page,
836         fs_prep_san_write:      fsfilt_ext3_prep_san_write,
837         fs_write_record:        fsfilt_ext3_write_record,
838         fs_read_record:         fsfilt_ext3_read_record,
839         fs_setup:               fsfilt_ext3_setup,
840 };
841
842 static int __init fsfilt_ext3_init(void)
843 {
844         int rc;
845
846         //rc = ext3_xattr_register();
847         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
848                                       sizeof(struct fsfilt_cb_data), 0,
849                                       0, NULL, NULL);
850         if (!fcb_cache) {
851                 CERROR("error allocating fsfilt journal callback cache\n");
852                 GOTO(out, rc = -ENOMEM);
853         }
854
855         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
856
857         if (rc)
858                 kmem_cache_destroy(fcb_cache);
859 out:
860         return rc;
861 }
862
863 static void __exit fsfilt_ext3_exit(void)
864 {
865         int rc;
866
867         fsfilt_unregister_ops(&fsfilt_ext3_ops);
868         rc = kmem_cache_destroy(fcb_cache);
869
870         if (rc || atomic_read(&fcb_cache_count)) {
871                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
872                        atomic_read(&fcb_cache_count), rc);
873         }
874
875         //rc = ext3_xattr_unregister();
876 }
877
878 module_init(fsfilt_ext3_init);
879 module_exit(fsfilt_ext3_exit);
880
881 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
882 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
883 MODULE_LICENSE("GPL");