1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * lustre/lib/fsfilt_ext3.c
5 * Lustre filesystem abstraction routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
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.
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.
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.
26 #define DEBUG_SUBSYSTEM S_FILTER
28 #include <linux/init.h>
29 #include <linux/module.h>
31 #include <linux/jbd.h>
32 #include <linux/slab.h>
33 #include <linux/pagemap.h>
34 #include <linux/quotaops.h>
35 #include <linux/ext3_fs.h>
36 #include <linux/ext3_jbd.h>
37 #include <linux/version.h>
38 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
39 #include <linux/ext3_xattr.h>
41 #include <ext3/xattr.h>
44 #include <linux/kp30.h>
45 #include <linux/lustre_fsfilt.h>
46 #include <linux/obd.h>
47 #include <linux/obd_class.h>
49 static kmem_cache_t *fcb_cache;
50 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
52 struct fsfilt_cb_data {
53 struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
54 fsfilt_cb_t cb_func; /* MDS/OBD completion function */
55 struct obd_device *cb_obd; /* MDS/OBD completion device */
56 __u64 cb_last_rcvd; /* MDS/OST last committed operation */
57 void *cb_data; /* MDS/OST completion function data */
60 #ifndef EXT3_XATTR_INDEX_TRUSTED /* temporary until we hit l28 kernel */
61 #define EXT3_XATTR_INDEX_TRUSTED 4
63 #define XATTR_LUSTRE_MDS_LOV_EA "lov"
65 #define EXT3_XATTR_INDEX_LUSTRE 5 /* old */
66 #define XATTR_LUSTRE_MDS_OBJID "system.lustre_mds_objid" /* old */
69 * We don't currently need any additional blocks for rmdir and
70 * unlink transactions because we are storing the OST oa_id inside
71 * the inode (which we will be changing anyways as part of this
74 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
77 /* For updates to the last recieved file */
78 int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
82 if (current->journal_info) {
83 CDEBUG(D_INODE, "increasing refcount on %p\n",
84 current->journal_info);
90 case FSFILT_OP_UNLINK:
91 /* delete one file + create/update logs for each stripe */
92 nblocks += EXT3_DELETE_TRANS_BLOCKS;
93 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
94 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
96 case FSFILT_OP_RENAME:
97 /* modify additional directory */
98 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
99 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
100 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
102 case FSFILT_OP_SYMLINK:
103 /* additional block + block bitmap + GDT for long symlink */
106 case FSFILT_OP_CREATE:
107 /* create/update logs for each stripe */
108 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
109 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
111 case FSFILT_OP_MKDIR:
112 case FSFILT_OP_MKNOD:
113 /* modify one inode + block bitmap + GDT */
117 /* modify parent directory */
118 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
119 EXT3_DATA_TRANS_BLOCKS;
121 case FSFILT_OP_SETATTR:
122 /* Setattr on inode */
125 case FSFILT_OP_CANCEL_UNLINK:
126 /* blocks for log header bitmap update OR
127 * blocks for catalog header bitmap update + unlink of logs */
128 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
129 EXT3_DELETE_TRANS_BLOCKS * logs;
131 default: CERROR("unknown transaction start op %d\n", op);
135 LASSERT(current->journal_info == desc_private);
136 journal = EXT3_SB(inode->i_sb)->s_journal;
137 if (nblocks > journal->j_max_transaction_buffers) {
138 CERROR("too many credits %d for op %ux%u using %d instead\n",
139 nblocks, op, logs, journal->j_max_transaction_buffers);
140 nblocks = journal->j_max_transaction_buffers;
145 LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
146 handle = journal_start(EXT3_JOURNAL(inode), nblocks);
150 LASSERT(current->journal_info == handle);
152 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
153 op, nblocks, PTR_ERR(handle));
158 * Calculate the number of buffer credits needed to write multiple pages in
159 * a single ext3 transaction. No, this shouldn't be here, but as yet ext3
160 * doesn't have a nice API for calculating this sort of thing in advance.
162 * See comment above ext3_writepage_trans_blocks for details. We assume
163 * no data journaling is being done, but it does allow for all of the pages
164 * being non-contiguous. If we are guaranteed contiguous pages we could
165 * reduce the number of (d)indirect blocks a lot.
167 * With N blocks per page and P pages, for each inode we have at most:
169 * min(N*P, blocksize/4 + 1) dindirect blocks
172 * For the entire filesystem, we have at most:
173 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
174 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
175 * objcount inode blocks
177 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
179 * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
181 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
182 int niocount, struct niobuf_local *nb)
184 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
186 const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
187 int nbitmaps = 0, ngdblocks;
188 int needed = objcount + 1; /* inodes + superblock */
191 for (i = 0, j = 0; i < objcount; i++, fso++) {
192 /* two or more dindirect blocks in case we cross boundary */
193 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
195 sb->s_blocksize_bits) /
196 (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
197 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
199 /* leaf, indirect, tindirect blocks for first block */
200 nbitmaps += blockpp + 2;
202 j += fso->fso_bufcnt;
205 next_indir = nb[0].offset +
206 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
207 for (i = 1; i < niocount; i++) {
208 if (nb[i].offset >= next_indir) {
209 nbitmaps++; /* additional indirect */
210 next_indir = nb[i].offset +
211 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
212 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
213 nbitmaps++; /* additional indirect */
215 nbitmaps += blockpp; /* each leaf in different group? */
218 ngdblocks = nbitmaps;
219 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
220 nbitmaps = EXT3_SB(sb)->s_groups_count;
221 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
222 ngdblocks = EXT3_SB(sb)->s_gdb_count;
224 needed += nbitmaps + ngdblocks;
226 /* last_rcvd update */
227 needed += EXT3_DATA_TRANS_BLOCKS;
229 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
230 /* We assume that there will be 1 bit set in s_dquot.flags for each
231 * quota file that is active. This is at least true for now.
233 needed += hweight32(sb_any_quota_enabled(sb)) *
234 EXT3_SINGLEDATA_TRANS_BLOCKS;
237 LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
241 /* We have to start a huge journal transaction here to hold all of the
242 * metadata for the pages being written here. This is necessitated by
243 * the fact that we do lots of prepare_write operations before we do
244 * any of the matching commit_write operations, so even if we split
245 * up to use "smaller" transactions none of them could complete until
246 * all of them were opened. By having a single journal transaction,
247 * we eliminate duplicate reservations for common blocks like the
248 * superblock and group descriptors or bitmaps.
250 * We will start the transaction here, but each prepare_write will
251 * add a refcount to the transaction, and each commit_write will
252 * remove a refcount. The transaction will be closed when all of
253 * the pages have been written.
255 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
256 int niocount, struct niobuf_local *nb,
257 void *desc_private, int logs)
264 LASSERT(current->journal_info == desc_private);
265 journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
266 needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
268 /* The number of blocks we could _possibly_ dirty can very large.
269 * We reduce our request if it is absurd (and we couldn't get that
270 * many credits for a single handle anyways).
272 * At some point we have to limit the size of I/Os sent at one time,
273 * increase the size of the journal, or we have to calculate the
274 * actual journal requirements more carefully by checking all of
275 * the blocks instead of being maximally pessimistic. It remains to
276 * be seen if this is a real problem or not.
278 if (needed > journal->j_max_transaction_buffers) {
279 CERROR("want too many journal credits (%d) using %d instead\n",
280 needed, journal->j_max_transaction_buffers);
281 needed = journal->j_max_transaction_buffers;
285 handle = journal_start(journal, needed);
287 if (IS_ERR(handle)) {
288 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
291 LASSERT(handle->h_buffer_credits >= needed);
292 LASSERT(current->journal_info == handle);
298 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
301 handle_t *handle = h;
303 LASSERT(current->journal_info == handle);
305 handle->h_sync = 1; /* recovery likes this */
308 rc = journal_stop(handle);
314 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
318 transaction_t *transaction;
319 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
322 handle_t *handle = h;
326 LASSERT(current->journal_info == handle);
329 transaction = handle->h_transaction;
330 journal = transaction->t_journal;
331 tid = transaction->t_tid;
332 /* we don't want to be blocked */
334 rc = journal_stop(handle);
336 CERROR("error while stopping transaction: %d\n", rc);
340 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
341 rtid = log_start_commit(journal, transaction);
343 CERROR("strange race: %lu != %lu\n",
344 (unsigned long) tid, (unsigned long) rtid);
346 log_start_commit(journal, transaction->t_tid);
350 *wait_handle = (void *) tid;
351 CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
355 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
357 tid_t tid = (tid_t)(long)h;
359 CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
360 if (is_journal_aborted(EXT3_JOURNAL(inode)))
363 log_wait_commit(EXT3_JOURNAL(inode), tid);
368 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
369 struct iattr *iattr, int do_trunc)
371 struct inode *inode = dentry->d_inode;
376 /* A _really_ horrible hack to avoid removing the data stored
377 * in the block pointers; this is really the "small" stripe MD data.
378 * We can avoid further hackery by virtue of the MDS file size being
379 * zero all the time (which doesn't invoke block truncate at unlink
380 * time), so we assert we never change the MDS file size from zero. */
381 if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
382 /* ATTR_SIZE would invoke truncate: clear it */
383 iattr->ia_valid &= ~ATTR_SIZE;
384 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
386 /* make sure _something_ gets set - so new inode
387 * goes to disk (probably won't work over XFS */
388 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
389 iattr->ia_valid |= ATTR_MODE;
390 iattr->ia_mode = inode->i_mode;
394 /* Don't allow setattr to change file type */
395 iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
397 /* We set these flags on the client, but have already checked perms
398 * so don't confuse inode_change_ok. */
399 iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
401 if (inode->i_op->setattr) {
402 rc = inode->i_op->setattr(dentry, iattr);
404 rc = inode_change_ok(inode, iattr);
406 rc = inode_setattr(inode, iattr);
414 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
415 unsigned int cmd, unsigned long arg)
420 if (inode->i_fop->ioctl)
421 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
430 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
431 void *lmm, int lmm_size)
435 LASSERT(down_trylock(&inode->i_sem) != 0);
437 #ifdef INLINE_EA /* can go away before 1.0 - just for testing bug 2097 now */
438 /* Nasty hack city - store stripe MD data in the block pointers if
439 * it will fit, because putting it in an EA currently kills the MDS
440 * performance. We'll fix this with "fast EAs" in the future.
442 if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
443 sizeof(EXT3_I(inode)->i_data[0])) {
444 unsigned old_size = EXT3_I(inode)->i_data[0];
446 LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
447 CERROR("setting EA on %lu/%u again... interesting\n",
448 inode->i_ino, inode->i_generation);
451 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
452 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
453 mark_inode_dirty(inode);
458 /* keep this when we get rid of OLD_EA (too noisy during conversion) */
459 if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
460 CWARN("setting EA on %lu/%u again... interesting\n",
461 inode->i_ino, inode->i_generation);
466 /* this can go away before 1.0. For bug 2097 testing only. */
467 rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
468 XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
471 rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
472 XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
474 /* This tries to delete the old-format LOV EA, but only as long as we
475 * have successfully saved the new-format LOV EA (we can always try
476 * the conversion again the next time the file is accessed). It is
477 * possible (although unlikely) that the new-format LOV EA couldn't be
478 * saved because it ran out of space but we would need a file striped
479 * over least 123 OSTs before the two EAs filled a 4kB block.
481 * This can be removed when all filesystems have converted to the
482 * new EA format, but otherwise adds little if any overhead. If we
483 * wanted backward compatibility for existing files, we could keep
484 * the old EA around for a while but we'd have to clean it up later. */
485 if (rc >= 0 && old_ea) {
486 int err = ext3_xattr_set_handle(handle, inode,
487 EXT3_XATTR_INDEX_LUSTRE,
488 XATTR_LUSTRE_MDS_OBJID,
491 CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
492 inode->i_ino, inode->i_generation, err);
498 CERROR("error adding MD data to inode %lu: rc = %d\n",
503 /* Must be called with i_sem held */
504 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
508 LASSERT(down_trylock(&inode->i_sem) != 0);
510 /* Keep support for reading "inline EAs" until we convert
511 * users over to new format entirely. See bug 841/2097. */
512 if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
513 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
516 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
518 if (size > lmm_size) {
519 CERROR("inline EA on %lu/%u bad size %u > %u\n",
520 inode->i_ino, inode->i_generation,
524 memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
528 /* migrate LOV EA data to external block - keep same format */
529 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
530 inode->i_ino, inode->i_generation);
532 handle = journal_start(EXT3_JOURNAL(inode),
533 EXT3_XATTR_TRANS_BLOCKS);
534 if (!IS_ERR(handle)) {
536 rc = fsfilt_ext3_set_md(inode, handle,
537 &EXT3_I(inode)->i_data[1],size);
539 memset(EXT3_I(inode)->i_data, 0,
540 sizeof(EXT3_I(inode)->i_data));
541 mark_inode_dirty(inode);
543 err = journal_stop(handle);
547 rc = PTR_ERR(handle);
554 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
555 XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
556 /* try old EA type if new one failed - MDS will convert it for us */
557 if (rc == -ENODATA) {
558 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
559 EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
562 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
563 XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
567 /* This gives us the MD size */
569 return (rc == -ENODATA) ? 0 : rc;
572 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
573 EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
575 memset(lmm, 0, lmm_size);
576 return (rc == -ENODATA) ? 0 : rc;
582 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
585 struct inode *inode = file->f_dentry->d_inode;
588 if (S_ISREG(inode->i_mode))
589 rc = file->f_op->read(file, buf, count, off);
591 const int blkbits = inode->i_sb->s_blocksize_bits;
592 const int blksize = inode->i_sb->s_blocksize;
594 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
595 count, inode->i_ino, *off);
597 struct buffer_head *bh;
600 if (*off < inode->i_size) {
603 bh = ext3_bread(NULL, inode, *off >> blkbits,
606 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
609 memcpy(buf, bh->b_data, blksize);
612 /* XXX in theory we should just fake
613 * this buffer and continue like ext3,
614 * especially if this is a partial read
616 CERROR("error read dir %lu+%llu: %d\n",
617 inode->i_ino, *off, err);
622 struct ext3_dir_entry_2 *fake = (void *)buf;
624 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
625 memset(fake, 0, sizeof(*fake));
626 fake->rec_len = cpu_to_le32(blksize);
638 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
640 struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
642 fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
644 OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
645 atomic_dec(&fcb_cache_count);
648 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
649 void *handle, fsfilt_cb_t cb_func,
652 struct fsfilt_cb_data *fcb;
654 OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
658 atomic_inc(&fcb_cache_count);
659 fcb->cb_func = cb_func;
661 fcb->cb_last_rcvd = last_rcvd;
662 fcb->cb_data = cb_data;
664 CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
666 journal_callback_set(handle, fsfilt_ext3_cb_func,
667 (struct journal_callback *)fcb);
674 * We need to hack the return value for the free inode counts because
675 * the current EA code requires one filesystem block per inode with EAs,
676 * so it is possible to run out of blocks before we run out of inodes.
678 * This can be removed when the ext3 EA code is fixed.
680 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
685 memset(&sfs, 0, sizeof(sfs));
687 rc = sb->s_op->statfs(sb, &sfs);
689 if (!rc && sfs.f_bfree < sfs.f_ffree) {
690 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
691 sfs.f_ffree = sfs.f_bfree;
694 statfs_pack(osfs, &sfs);
698 static int fsfilt_ext3_sync(struct super_block *sb)
700 return ext3_force_commit(sb);
703 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
704 unsigned long *blocks, int *created, int create);
705 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
706 unsigned long *blocks, int *created, int create)
708 return ext3_map_inode_page(inode, page, blocks, created, create);
711 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
712 int nblocks, loff_t newsize);
713 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
714 int nblocks, loff_t newsize)
716 return ext3_prep_san_write(inode, blocks, nblocks, newsize);
719 static int fsfilt_ext3_read_record(struct file * file, void *buf,
720 int size, loff_t *offs)
722 struct inode *inode = file->f_dentry->d_inode;
724 struct buffer_head *bh;
725 int err, blocksize, csize, boffs;
727 /* prevent reading after eof */
729 if (inode->i_size < *offs + size) {
730 size = inode->i_size - *offs;
733 CERROR("size %llu is too short for read %u@%llu\n",
734 inode->i_size, size, *offs);
736 } else if (size == 0) {
743 blocksize = 1 << inode->i_blkbits;
746 block = *offs >> inode->i_blkbits;
747 boffs = *offs & (blocksize - 1);
748 csize = min(blocksize - boffs, size);
749 bh = ext3_bread(NULL, inode, block, 0, &err);
751 CERROR("can't read block: %d\n", err);
755 memcpy(buf, bh->b_data + boffs, csize);
765 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
766 loff_t *offs, int force_sync)
768 struct buffer_head *bh = NULL;
770 struct inode *inode = file->f_dentry->d_inode;
771 loff_t old_size = inode->i_size, offset = *offs;
772 loff_t new_size = inode->i_size;
775 int err, block_count = 0, blocksize, size, boffs;
777 /* Determine how many transaction credits are needed */
778 blocksize = 1 << inode->i_blkbits;
779 block_count = (*offs & (blocksize - 1)) + bufsize;
780 block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
782 journal = EXT3_SB(inode->i_sb)->s_journal;
784 handle = journal_start(journal,
785 block_count * EXT3_DATA_TRANS_BLOCKS + 2);
787 if (IS_ERR(handle)) {
788 CERROR("can't start transaction\n");
789 return PTR_ERR(handle);
792 while (bufsize > 0) {
796 block = offset >> inode->i_blkbits;
797 boffs = offset & (blocksize - 1);
798 size = min(blocksize - boffs, bufsize);
799 bh = ext3_bread(handle, inode, block, 1, &err);
801 CERROR("can't read/create block: %d\n", err);
805 err = ext3_journal_get_write_access(handle, bh);
807 CERROR("journal_get_write_access() returned error %d\n",
811 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
812 memcpy(bh->b_data + boffs, buf, size);
813 err = ext3_journal_dirty_metadata(handle, bh);
815 CERROR("journal_dirty_metadata() returned error %d\n",
819 if (offset + size > new_size)
820 new_size = offset + size;
827 handle->h_sync = 1; /* recovery likes this */
832 /* correct in-core and on-disk sizes */
833 if (new_size > inode->i_size) {
835 if (new_size > inode->i_size)
836 inode->i_size = new_size;
837 if (inode->i_size > EXT3_I(inode)->i_disksize)
838 EXT3_I(inode)->i_disksize = inode->i_size;
839 if (inode->i_size > old_size)
840 mark_inode_dirty(inode);
845 journal_stop(handle);
853 static int fsfilt_ext3_setup(struct super_block *sb)
856 EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
857 EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
860 CWARN("Enabling PDIROPS\n");
861 set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
862 sb->s_flags |= S_PDIROPS;
867 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
868 objects. Logs is number of logfiles to update */
869 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
873 case FSFILT_OP_CREATE:
874 /* directory leaf, index & indirect & EA*/
876 case FSFILT_OP_UNLINK:
882 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
883 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
884 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
885 for (i = 0; i < op; i++, fso++) {
886 int nblocks = fso->fso_bufcnt * blockpp;
887 int ndindirect = min(nblocks, addrpp + 1);
888 int nindir = nblocks + ndindirect + 1;
892 return needed + 3 * logs;
898 static struct fsfilt_operations fsfilt_ext3_ops = {
900 .fs_owner = THIS_MODULE,
901 .fs_start = fsfilt_ext3_start,
902 .fs_brw_start = fsfilt_ext3_brw_start,
903 .fs_commit = fsfilt_ext3_commit,
904 .fs_commit_async = fsfilt_ext3_commit_async,
905 .fs_commit_wait = fsfilt_ext3_commit_wait,
906 .fs_setattr = fsfilt_ext3_setattr,
907 .fs_iocontrol = fsfilt_ext3_iocontrol,
908 .fs_set_md = fsfilt_ext3_set_md,
909 .fs_get_md = fsfilt_ext3_get_md,
910 .fs_readpage = fsfilt_ext3_readpage,
911 .fs_add_journal_cb = fsfilt_ext3_add_journal_cb,
912 .fs_statfs = fsfilt_ext3_statfs,
913 .fs_sync = fsfilt_ext3_sync,
914 .fs_map_inode_page = fsfilt_ext3_map_inode_page,
915 .fs_prep_san_write = fsfilt_ext3_prep_san_write,
916 .fs_write_record = fsfilt_ext3_write_record,
917 .fs_read_record = fsfilt_ext3_read_record,
918 .fs_setup = fsfilt_ext3_setup,
919 .fs_get_op_len = fsfilt_ext3_get_op_len,
922 static int __init fsfilt_ext3_init(void)
926 fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
927 sizeof(struct fsfilt_cb_data), 0,
930 CERROR("error allocating fsfilt journal callback cache\n");
931 GOTO(out, rc = -ENOMEM);
934 rc = fsfilt_register_ops(&fsfilt_ext3_ops);
937 kmem_cache_destroy(fcb_cache);
942 static void __exit fsfilt_ext3_exit(void)
944 fsfilt_unregister_ops(&fsfilt_ext3_ops);
945 LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
946 "can't free fsfilt callback cache: count %d\n",
947 atomic_read(&fcb_cache_count));
950 module_init(fsfilt_ext3_init);
951 module_exit(fsfilt_ext3_exit);
953 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
954 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
955 MODULE_LICENSE("GPL");