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/ext3_extents.h>
38 #include <linux/version.h>
39 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
40 #include <linux/ext3_xattr.h>
42 #include <ext3/xattr.h>
45 #include <linux/kp30.h>
46 #include <linux/lustre_fsfilt.h>
47 #include <linux/obd.h>
48 #include <linux/obd_class.h>
49 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
50 #include <linux/module.h>
51 #include <linux/iobuf.h>
55 static kmem_cache_t *fcb_cache;
56 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
58 struct fsfilt_cb_data {
59 struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
60 fsfilt_cb_t cb_func; /* MDS/OBD completion function */
61 struct obd_device *cb_obd; /* MDS/OBD completion device */
62 __u64 cb_last_num; /* MDS/OST last committed operation */
63 void *cb_data; /* MDS/OST completion function data */
66 #ifndef EXT3_XATTR_INDEX_TRUSTED /* temporary until we hit l28 kernel */
67 #define EXT3_XATTR_INDEX_TRUSTED 4
70 #define XATTR_LUSTRE_MDS_LOV_EA "lov"
71 #define XATTR_LUSTRE_MDS_MID_EA "mid"
72 #define XATTR_LUSTRE_MDS_SID_EA "sid"
75 * We don't currently need any additional blocks for rmdir and
76 * unlink transactions because we are storing the OST oa_id inside
77 * the inode (which we will be changing anyways as part of this
80 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
83 /* For updates to the last recieved file */
84 int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
88 if (current->journal_info) {
89 CDEBUG(D_INODE, "increasing refcount on %p\n",
90 current->journal_info);
94 /* XXX BUG 3188 -- must return to one set of opcodes */
95 /* FIXME - cache hook */
97 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
103 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
108 case FSFILT_OP_RMDIR:
109 case FSFILT_OP_UNLINK:
110 /* delete one file + create/update logs for each stripe */
111 nblocks += EXT3_DELETE_TRANS_BLOCKS;
112 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
113 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
115 case FSFILT_OP_RENAME:
116 /* modify additional directory */
117 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
119 case FSFILT_OP_SYMLINK:
120 /* additional block + block bitmap + GDT for long symlink */
123 case FSFILT_OP_CREATE:
124 /* create/update logs for each stripe */
125 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
126 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
128 case FSFILT_OP_MKDIR:
129 case FSFILT_OP_MKNOD:
130 /* modify one inode + block bitmap + GDT */
134 /* modify parent directory */
135 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
136 EXT3_DATA_TRANS_BLOCKS;
138 case FSFILT_OP_SETATTR:
139 /* Setattr on inode */
142 case FSFILT_OP_CANCEL_UNLINK:
143 /* blocks for log header bitmap update OR
144 * blocks for catalog header bitmap update + unlink of logs */
145 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
146 EXT3_DELETE_TRANS_BLOCKS * logs;
149 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
151 default: CERROR("unknown transaction start op %d\n", op);
155 LASSERT(current->journal_info == desc_private);
156 journal = EXT3_SB(inode->i_sb)->s_journal;
157 if (nblocks > journal->j_max_transaction_buffers) {
158 CERROR("too many credits %d for op %ux%u using %d instead\n",
159 nblocks, op, logs, journal->j_max_transaction_buffers);
160 nblocks = journal->j_max_transaction_buffers;
165 handle = journal_start(EXT3_JOURNAL(inode), nblocks);
169 LASSERT(current->journal_info == handle);
171 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
172 op, nblocks, PTR_ERR(handle));
177 * Calculate the number of buffer credits needed to write multiple pages in
178 * a single ext3 transaction. No, this shouldn't be here, but as yet ext3
179 * doesn't have a nice API for calculating this sort of thing in advance.
181 * See comment above ext3_writepage_trans_blocks for details. We assume
182 * no data journaling is being done, but it does allow for all of the pages
183 * being non-contiguous. If we are guaranteed contiguous pages we could
184 * reduce the number of (d)indirect blocks a lot.
186 * With N blocks per page and P pages, for each inode we have at most:
188 * min(N*P, blocksize/4 + 1) dindirect blocks
191 * For the entire filesystem, we have at most:
192 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
193 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
194 * objcount inode blocks
196 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
198 * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
200 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
201 int niocount, struct niobuf_local *nb)
203 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
205 const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
206 int nbitmaps = 0, ngdblocks;
207 int needed = objcount + 1; /* inodes + superblock */
210 for (i = 0, j = 0; i < objcount; i++, fso++) {
211 /* two or more dindirect blocks in case we cross boundary */
212 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
214 sb->s_blocksize_bits) /
215 (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
216 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
218 /* leaf, indirect, tindirect blocks for first block */
219 nbitmaps += blockpp + 2;
221 j += fso->fso_bufcnt;
224 next_indir = nb[0].offset +
225 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
226 for (i = 1; i < niocount; i++) {
227 if (nb[i].offset >= next_indir) {
228 nbitmaps++; /* additional indirect */
229 next_indir = nb[i].offset +
230 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
231 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
232 nbitmaps++; /* additional indirect */
234 nbitmaps += blockpp; /* each leaf in different group? */
237 ngdblocks = nbitmaps;
238 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
239 nbitmaps = EXT3_SB(sb)->s_groups_count;
240 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
241 ngdblocks = EXT3_SB(sb)->s_gdb_count;
243 needed += nbitmaps + ngdblocks;
245 /* last_rcvd update */
246 needed += EXT3_DATA_TRANS_BLOCKS;
248 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
249 /* We assume that there will be 1 bit set in s_dquot.flags for each
250 * quota file that is active. This is at least true for now.
252 needed += hweight32(sb_any_quota_enabled(sb)) *
253 EXT3_SINGLEDATA_TRANS_BLOCKS;
259 /* We have to start a huge journal transaction here to hold all of the
260 * metadata for the pages being written here. This is necessitated by
261 * the fact that we do lots of prepare_write operations before we do
262 * any of the matching commit_write operations, so even if we split
263 * up to use "smaller" transactions none of them could complete until
264 * all of them were opened. By having a single journal transaction,
265 * we eliminate duplicate reservations for common blocks like the
266 * superblock and group descriptors or bitmaps.
268 * We will start the transaction here, but each prepare_write will
269 * add a refcount to the transaction, and each commit_write will
270 * remove a refcount. The transaction will be closed when all of
271 * the pages have been written.
273 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
274 int niocount, struct niobuf_local *nb,
275 void *desc_private, int logs)
282 LASSERT(current->journal_info == desc_private);
283 journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
284 needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
286 /* The number of blocks we could _possibly_ dirty can very large.
287 * We reduce our request if it is absurd (and we couldn't get that
288 * many credits for a single handle anyways).
290 * At some point we have to limit the size of I/Os sent at one time,
291 * increase the size of the journal, or we have to calculate the
292 * actual journal requirements more carefully by checking all of
293 * the blocks instead of being maximally pessimistic. It remains to
294 * be seen if this is a real problem or not.
296 if (needed > journal->j_max_transaction_buffers) {
297 CERROR("want too many journal credits (%d) using %d instead\n",
298 needed, journal->j_max_transaction_buffers);
299 needed = journal->j_max_transaction_buffers;
303 handle = journal_start(journal, needed);
305 if (IS_ERR(handle)) {
306 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
309 LASSERT(handle->h_buffer_credits >= needed);
310 LASSERT(current->journal_info == handle);
316 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
317 void *h, int force_sync)
320 handle_t *handle = h;
322 LASSERT(current->journal_info == handle);
324 handle->h_sync = 1; /* recovery likes this */
327 rc = journal_stop(handle);
333 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
337 transaction_t *transaction;
338 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
341 handle_t *handle = h;
345 LASSERT(current->journal_info == handle);
348 transaction = handle->h_transaction;
349 journal = transaction->t_journal;
350 tid = transaction->t_tid;
351 /* we don't want to be blocked */
353 rc = journal_stop(handle);
355 CERROR("error while stopping transaction: %d\n", rc);
359 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
360 rtid = log_start_commit(journal, transaction);
362 CERROR("strange race: %lu != %lu\n",
363 (unsigned long) tid, (unsigned long) rtid);
365 log_start_commit(journal, transaction->t_tid);
369 *wait_handle = (void *) tid;
370 CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
374 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
376 tid_t tid = (tid_t)(long)h;
378 CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
379 if (is_journal_aborted(EXT3_JOURNAL(inode)))
382 log_wait_commit(EXT3_JOURNAL(inode), tid);
387 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
388 struct iattr *iattr, int do_trunc)
390 struct inode *inode = dentry->d_inode;
395 /* A _really_ horrible hack to avoid removing the data stored
396 * in the block pointers; this is really the "small" stripe MD data.
397 * We can avoid further hackery by virtue of the MDS file size being
398 * zero all the time (which doesn't invoke block truncate at unlink
399 * time), so we assert we never change the MDS file size from zero. */
400 if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
401 /* ATTR_SIZE would invoke truncate: clear it */
402 iattr->ia_valid &= ~ATTR_SIZE;
403 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
405 /* make sure _something_ gets set - so new inode
406 * goes to disk (probably won't work over XFS */
407 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
408 iattr->ia_valid |= ATTR_MODE;
409 iattr->ia_mode = inode->i_mode;
413 /* Don't allow setattr to change file type */
414 iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
416 /* We set these flags on the client, but have already checked perms
417 * so don't confuse inode_change_ok. */
418 iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
420 if (inode->i_op->setattr) {
421 rc = inode->i_op->setattr(dentry, iattr);
423 rc = inode_change_ok(inode, iattr);
425 rc = inode_setattr(inode, iattr);
433 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
434 unsigned int cmd, unsigned long arg)
439 if (inode->i_fop->ioctl)
440 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
447 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
448 void *buffer, int buffer_size)
454 rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
455 name, buffer, buffer_size, 0);
458 CERROR("set xattr %s from inode %lu: rc %d\n",
459 name, inode->i_ino, rc);
463 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
464 void *buffer, int buffer_size)
470 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
471 name, buffer, buffer_size);
475 return (rc == -ENODATA) ? 0 : rc;
477 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
478 name, inode->i_ino, rc);
479 memset(buffer, 0, buffer_size);
480 return (rc == -ENODATA) ? 0 : rc;
486 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
487 void *lmm, int lmm_size)
491 LASSERT(down_trylock(&inode->i_sem) != 0);
493 /* keep this when we get rid of OLD_EA (too noisy during conversion) */
494 if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
495 CWARN("setting EA on %lu/%u again... interesting\n",
496 inode->i_ino, inode->i_generation);
498 rc = fsfilt_ext3_set_xattr(inode, handle, XATTR_LUSTRE_MDS_LOV_EA,
503 /* Must be called with i_sem held */
504 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
508 rc = fsfilt_ext3_get_xattr(inode, XATTR_LUSTRE_MDS_LOV_EA,
513 static int fsfilt_ext3_set_mid(struct inode *inode, void *handle,
514 void *mid, int mid_size)
518 rc = fsfilt_ext3_set_xattr(inode, handle, XATTR_LUSTRE_MDS_MID_EA,
523 /* Must be called with i_sem held */
524 static int fsfilt_ext3_get_mid(struct inode *inode, void *mid, int mid_size)
528 rc = fsfilt_ext3_get_xattr(inode, XATTR_LUSTRE_MDS_MID_EA,
533 static int fsfilt_ext3_set_sid(struct inode *inode, void *handle,
534 void *sid, int sid_size)
538 rc = fsfilt_ext3_set_xattr(inode, handle, XATTR_LUSTRE_MDS_SID_EA,
543 /* Must be called with i_sem held */
544 static int fsfilt_ext3_get_sid(struct inode *inode, void *sid, int sid_size)
548 rc = fsfilt_ext3_get_xattr(inode, XATTR_LUSTRE_MDS_SID_EA,
553 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
554 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct bio *bio)
560 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct kiobuf *bio)
562 int rc, blocks_per_page;
564 rc = brw_kiovec(rw, 1, &bio, inode->i_dev,
565 bio->blocks, 1 << inode->i_blkbits);
567 blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
569 if (rc != (1 << inode->i_blkbits) * bio->nr_pages * blocks_per_page) {
570 CERROR("short write? expected %d, wrote %d\n",
571 (1 << inode->i_blkbits) * bio->nr_pages *
572 blocks_per_page, rc);
579 /* FIXME-UMKA: This should be used in 2.6.x io code later. */
580 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
585 page = grab_cache_page(inode->i_mapping, index);
587 return ERR_PTR(-ENOMEM);
589 if (PageUptodate(page)) {
594 rc = inode->i_mapping->a_ops->readpage(NULL, page);
596 page_cache_release(page);
603 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
606 struct inode *inode = file->f_dentry->d_inode;
609 if (S_ISREG(inode->i_mode))
610 rc = file->f_op->read(file, buf, count, off);
612 const int blkbits = inode->i_sb->s_blocksize_bits;
613 const int blksize = inode->i_sb->s_blocksize;
615 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
616 count, inode->i_ino, *off);
618 struct buffer_head *bh;
621 if (*off < inode->i_size) {
624 bh = ext3_bread(NULL, inode, *off >> blkbits,
627 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
630 memcpy(buf, bh->b_data, blksize);
633 /* XXX in theory we should just fake
634 * this buffer and continue like ext3,
635 * especially if this is a partial read
637 CERROR("error read dir %lu+%llu: %d\n",
638 inode->i_ino, *off, err);
643 struct ext3_dir_entry_2 *fake = (void *)buf;
645 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
646 memset(fake, 0, sizeof(*fake));
647 fake->rec_len = cpu_to_le32(blksize);
659 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
661 struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
663 fcb->cb_func(fcb->cb_obd, fcb->cb_last_num, fcb->cb_data, error);
665 OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
666 atomic_dec(&fcb_cache_count);
669 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
670 struct super_block *sb,
671 __u64 last_num, void *handle,
675 struct fsfilt_cb_data *fcb;
677 OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
681 atomic_inc(&fcb_cache_count);
682 fcb->cb_func = cb_func;
684 fcb->cb_last_num = last_num;
685 fcb->cb_data = cb_data;
687 CDEBUG(D_EXT2, "set callback for last_num: "LPD64"\n", last_num);
690 journal_callback_set(handle, fsfilt_ext3_cb_func,
691 (struct journal_callback *)fcb);
698 * We need to hack the return value for the free inode counts because
699 * the current EA code requires one filesystem block per inode with EAs,
700 * so it is possible to run out of blocks before we run out of inodes.
702 * This can be removed when the ext3 EA code is fixed.
704 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
709 memset(&sfs, 0, sizeof(sfs));
711 rc = sb->s_op->statfs(sb, &sfs);
713 if (!rc && sfs.f_bfree < sfs.f_ffree) {
714 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
715 sfs.f_ffree = sfs.f_bfree;
718 statfs_pack(osfs, &sfs);
722 static int fsfilt_ext3_sync(struct super_block *sb)
724 return ext3_force_commit(sb);
727 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
728 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
729 #define ext3_up_truncate_sem(inode) up_write(&EXT3_I(inode)->truncate_sem);
730 #define ext3_down_truncate_sem(inode) down_write(&EXT3_I(inode)->truncate_sem);
732 #define ext3_up_truncate_sem(inode) up(&EXT3_I(inode)->truncate_sem);
733 #define ext3_down_truncate_sem(inode) down(&EXT3_I(inode)->truncate_sem);
736 #include <linux/lustre_version.h>
737 #if EXT3_EXT_MAGIC == 0xf301
738 #define ee_start e_start
739 #define ee_block e_block
742 #ifndef EXT3_BB_MAX_BLOCKS
743 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
744 ext3_new_blocks(handle, inode, count, goal, err)
748 unsigned long *blocks;
755 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
756 unsigned long block, int *aflags)
758 struct ext3_inode_info *ei = EXT3_I(inode);
759 unsigned long bg_start;
760 unsigned long colour;
764 struct ext3_extent *ex;
765 depth = path->p_depth;
767 /* try to predict block placement */
768 if ((ex = path[depth].p_ext)) {
769 if (ex->ee_block + ex->ee_len == block)
771 return ex->ee_start + (block - ex->ee_block);
774 /* it looks index is empty
775 * try to find starting from index itself */
776 if (path[depth].p_bh)
777 return path[depth].p_bh->b_blocknr;
780 /* OK. use inode's group */
781 bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
782 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
783 colour = (current->pid % 16) *
784 (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
785 return bg_start + colour + block;
788 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
789 struct ext3_ext_path *path,
790 struct ext3_extent *newex, int exist)
792 struct inode *inode = tree->inode;
793 struct bpointers *bp = tree->private;
794 int count, err, goal;
795 unsigned long pblock;
802 EXT_ASSERT(i == path->p_depth);
803 EXT_ASSERT(path[i].p_hdr);
810 if (bp->create == 0) {
812 if (newex->ee_block < bp->start)
813 i = bp->start - newex->ee_block;
814 if (i >= newex->ee_len)
815 CERROR("nothing to do?! i = %d, e_num = %u\n",
817 for (; i < newex->ee_len && bp->num; i++) {
828 tgen = EXT_GENERATION(tree);
829 count = ext3_ext_calc_credits_for_insert(tree, path);
830 ext3_up_truncate_sem(inode);
833 handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
835 if (IS_ERR(handle)) {
836 ext3_down_truncate_sem(inode);
837 return PTR_ERR(handle);
840 if (tgen != EXT_GENERATION(tree)) {
841 /* the tree has changed. so path can be invalid at moment */
843 journal_stop(handle);
845 ext3_down_truncate_sem(inode);
849 ext3_down_truncate_sem(inode);
850 count = newex->ee_len;
851 goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
852 aflags |= 2; /* block have been already reserved */
853 pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
856 EXT_ASSERT(count <= newex->ee_len);
858 /* insert new extent */
859 newex->ee_start = pblock;
860 newex->ee_len = count;
861 err = ext3_ext_insert_extent(handle, tree, path, newex);
865 /* correct on-disk inode size */
866 if (newex->ee_len > 0) {
867 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
868 new_i_size = new_i_size << inode->i_blkbits;
869 if (new_i_size > EXT3_I(inode)->i_disksize) {
870 EXT3_I(inode)->i_disksize = new_i_size;
871 err = ext3_mark_inode_dirty(handle, inode);
876 journal_stop(handle);
882 CERROR("hmm. why do we find this extent?\n");
883 CERROR("initial space: %lu:%u\n",
884 bp->start, bp->init_num);
885 CERROR("current extent: %u/%u/%u %d\n",
886 newex->ee_block, newex->ee_len,
887 newex->ee_start, exist);
890 if (newex->ee_block < bp->start)
891 i = bp->start - newex->ee_block;
892 if (i >= newex->ee_len)
893 CERROR("nothing to do?! i = %d, e_num = %u\n",
895 for (; i < newex->ee_len && bp->num; i++) {
896 *(bp->created) = (exist == 0 ? 1 : 0);
898 *(bp->blocks) = newex->ee_start + i;
907 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
908 unsigned long num, unsigned long *blocks,
909 int *created, int create)
911 struct ext3_extents_tree tree;
915 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
916 block, block + num, (unsigned) inode->i_ino);
918 ext3_init_tree_desc(&tree, inode);
921 bp.created = created;
923 bp.init_num = bp.num = num;
926 ext3_down_truncate_sem(inode);
927 err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
928 ext3_ext_invalidate_cache(&tree);
929 ext3_up_truncate_sem(inode);
934 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
935 int pages, unsigned long *blocks,
936 int *created, int create)
938 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
940 struct page *fp = NULL;
943 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
944 inode->i_ino, pages, (*page)->index);
946 /* pages are sorted already. so, we just have to find
947 * contig. space and process them properly */
950 /* start new extent */
955 } else if (fp->index + clen == (*page)->index) {
956 /* continue the extent */
963 /* process found extent */
964 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
965 clen * blocks_per_page, blocks,
970 /* look for next extent */
972 blocks += blocks_per_page * clen;
973 created += blocks_per_page * clen;
977 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
978 clen * blocks_per_page, blocks,
985 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
986 unsigned long *blocks, int *created, int create);
987 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
988 int pages, unsigned long *blocks,
989 int *created, int create)
991 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
995 for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
996 rc = ext3_map_inode_page(inode, *page, b, cr, create);
998 CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
999 inode->i_ino, *b, *cr, create, rc);
1003 b += blocks_per_page;
1004 cr += blocks_per_page;
1009 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
1010 int pages, unsigned long *blocks,
1011 int *created, int create,
1012 struct semaphore *optional_sem)
1015 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
1016 if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
1017 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
1018 blocks, created, create);
1022 if (optional_sem != NULL)
1024 rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
1026 if (optional_sem != NULL)
1032 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
1033 int nblocks, loff_t newsize);
1034 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
1035 int nblocks, loff_t newsize)
1037 return ext3_prep_san_write(inode, blocks, nblocks, newsize);
1040 static int fsfilt_ext3_read_record(struct file * file, void *buf,
1041 int size, loff_t *offs)
1043 struct inode *inode = file->f_dentry->d_inode;
1044 unsigned long block;
1045 struct buffer_head *bh;
1046 int err, blocksize, csize, boffs;
1048 /* prevent reading after eof */
1050 if (inode->i_size < *offs + size) {
1051 size = inode->i_size - *offs;
1054 CERROR("size %llu is too short for read %u@%llu\n",
1055 inode->i_size, size, *offs);
1057 } else if (size == 0) {
1064 blocksize = 1 << inode->i_blkbits;
1067 block = *offs >> inode->i_blkbits;
1068 boffs = *offs & (blocksize - 1);
1069 csize = min(blocksize - boffs, size);
1070 bh = ext3_bread(NULL, inode, block, 0, &err);
1072 CERROR("can't read block: %d\n", err);
1076 memcpy(buf, bh->b_data + boffs, csize);
1086 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1087 loff_t *offs, int force_sync)
1089 struct buffer_head *bh = NULL;
1090 unsigned long block;
1091 struct inode *inode = file->f_dentry->d_inode;
1092 loff_t old_size = inode->i_size, offset = *offs;
1093 loff_t new_size = inode->i_size;
1096 int err = 0, block_count = 0, blocksize, size, boffs;
1098 /* Determine how many transaction credits are needed */
1099 blocksize = 1 << inode->i_blkbits;
1100 block_count = (*offs & (blocksize - 1)) + bufsize;
1101 block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1103 journal = EXT3_SB(inode->i_sb)->s_journal;
1105 handle = journal_start(journal,
1106 block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1108 if (IS_ERR(handle)) {
1109 CERROR("can't start transaction\n");
1110 return PTR_ERR(handle);
1113 while (bufsize > 0) {
1117 block = offset >> inode->i_blkbits;
1118 boffs = offset & (blocksize - 1);
1119 size = min(blocksize - boffs, bufsize);
1120 bh = ext3_bread(handle, inode, block, 1, &err);
1122 CERROR("can't read/create block: %d\n", err);
1126 err = ext3_journal_get_write_access(handle, bh);
1128 CERROR("journal_get_write_access() returned error %d\n",
1132 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1133 memcpy(bh->b_data + boffs, buf, size);
1134 err = ext3_journal_dirty_metadata(handle, bh);
1136 CERROR("journal_dirty_metadata() returned error %d\n",
1140 if (offset + size > new_size)
1141 new_size = offset + size;
1148 handle->h_sync = 1; /* recovery likes this */
1153 /* correct in-core and on-disk sizes */
1154 if (new_size > inode->i_size) {
1156 if (new_size > inode->i_size)
1157 inode->i_size = new_size;
1158 if (inode->i_size > EXT3_I(inode)->i_disksize)
1159 EXT3_I(inode)->i_disksize = inode->i_size;
1160 if (inode->i_size > old_size)
1161 mark_inode_dirty(inode);
1166 journal_stop(handle);
1174 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1176 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1177 struct mds_obd *mds = &obd->u.mds;
1180 EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1181 EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1184 CWARN("Enabling PDIROPS\n");
1185 set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1186 sb->s_flags |= S_PDIROPS;
1188 /* setup mdsnum in underlying fs */
1189 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1190 if (mds->mds_lmv_obd) {
1191 struct ext3_sb_info *sbi = EXT3_SB(sb);
1192 struct ext3_super_block *es = sbi->s_es;
1196 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1197 CWARN("%s: set mdsnum %d in ext3fs\n",
1198 obd->obd_name, mds->mds_num);
1200 handle = journal_start(sbi->s_journal, 1);
1202 LASSERT(!IS_ERR(handle));
1203 err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1205 EXT3_SET_INCOMPAT_FEATURE(sb,
1206 EXT3_FEATURE_INCOMPAT_MDSNUM);
1207 es->s_mdsnum = mds->mds_num;
1208 err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1211 journal_stop(handle);
1214 CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1215 obd->obd_name, es->s_mdsnum);
1217 sbi->s_mdsnum = es->s_mdsnum;
1223 extern int ext3_add_dir_entry(struct dentry *dentry);
1224 extern int ext3_del_dir_entry(struct dentry *dentry);
1226 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1227 struct dentry *parent,
1228 char *name, int namelen,
1230 unsigned long generation,
1234 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1235 struct dentry *dentry;
1238 LASSERT(namelen != 0);
1239 dentry = ll_lookup_one_len(name, parent, namelen);
1240 if (IS_ERR(dentry)) {
1241 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1242 dentry->d_name.name, dentry->d_inode->i_ino,
1243 (unsigned long) dentry->d_inode->i_generation,
1244 (int) PTR_ERR(dentry));
1245 RETURN(PTR_ERR(dentry));
1247 if (dentry->d_inode != NULL) {
1248 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1249 dentry->d_name.name, dentry);
1254 /* mds_reint_rename() may use this method to add dir entry
1255 * that points onto local inode. and we don't want to find
1256 * it cross-ref by subsequent lookups */
1259 dentry->d_flags |= DCACHE_CROSS_REF;
1260 dentry->d_inum = ino;
1261 dentry->d_mdsnum = mds;
1262 dentry->d_generation = generation;
1263 dentry->d_fid = fid;
1265 err = ext3_add_dir_entry(dentry);
1272 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1277 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1278 struct dentry *dentry)
1280 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1283 err = ext3_del_dir_entry(dentry);
1289 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1294 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1295 objects. Logs is number of logfiles to update */
1296 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1300 case FSFILT_OP_CREATE:
1301 /* directory leaf, index & indirect & EA*/
1302 return 4 + 3 * logs;
1303 case FSFILT_OP_UNLINK:
1309 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1310 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1311 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1312 for (i = 0; i < op; i++, fso++) {
1313 int nblocks = fso->fso_bufcnt * blockpp;
1314 int ndindirect = min(nblocks, addrpp + 1);
1315 int nindir = nblocks + ndindirect + 1;
1319 return needed + 3 * logs;
1326 #define EXTENTS_EA "write_extents"
1327 #define EXTENTS_EA_SIZE 64
1329 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1330 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1331 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1332 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1334 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1340 rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1341 EXTENTS_EA, from, num);
1345 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1351 rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1352 EXTENTS_EA, from, num);
1356 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1357 const char *eaname, int size);
1359 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1363 rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1368 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1369 char **pbuf, int *size)
1373 rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1374 EXTENTS_EA, pbuf, size);
1378 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1382 rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1387 static struct fsfilt_operations fsfilt_ext3_ops = {
1389 .fs_owner = THIS_MODULE,
1390 .fs_start = fsfilt_ext3_start,
1391 .fs_brw_start = fsfilt_ext3_brw_start,
1392 .fs_commit = fsfilt_ext3_commit,
1393 .fs_commit_async = fsfilt_ext3_commit_async,
1394 .fs_commit_wait = fsfilt_ext3_commit_wait,
1395 .fs_setattr = fsfilt_ext3_setattr,
1396 .fs_iocontrol = fsfilt_ext3_iocontrol,
1397 .fs_set_md = fsfilt_ext3_set_md,
1398 .fs_get_md = fsfilt_ext3_get_md,
1399 .fs_set_mid = fsfilt_ext3_set_mid,
1400 .fs_get_mid = fsfilt_ext3_get_mid,
1401 .fs_set_sid = fsfilt_ext3_set_sid,
1402 .fs_get_sid = fsfilt_ext3_get_sid,
1403 .fs_readpage = fsfilt_ext3_readpage,
1404 .fs_add_journal_cb = fsfilt_ext3_add_journal_cb,
1405 .fs_statfs = fsfilt_ext3_statfs,
1406 .fs_sync = fsfilt_ext3_sync,
1407 .fs_map_inode_pages = fsfilt_ext3_map_inode_pages,
1408 .fs_prep_san_write = fsfilt_ext3_prep_san_write,
1409 .fs_write_record = fsfilt_ext3_write_record,
1410 .fs_read_record = fsfilt_ext3_read_record,
1411 .fs_setup = fsfilt_ext3_setup,
1412 .fs_getpage = fsfilt_ext3_getpage,
1413 .fs_send_bio = fsfilt_ext3_send_bio,
1414 .fs_set_xattr = fsfilt_ext3_set_xattr,
1415 .fs_get_xattr = fsfilt_ext3_get_xattr,
1416 .fs_get_op_len = fsfilt_ext3_get_op_len,
1417 .fs_add_dir_entry = fsfilt_ext3_add_dir_entry,
1418 .fs_del_dir_entry = fsfilt_ext3_del_dir_entry,
1419 .fs_init_extents_ea = fsfilt_ext3_init_extents_ea,
1420 .fs_insert_extents_ea = fsfilt_ext3_insert_extents_ea,
1421 .fs_remove_extents_ea = fsfilt_ext3_remove_extents_ea,
1422 .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1423 .fs_get_write_extents_num = fsfilt_ext3_get_write_extents_num,
1426 static int __init fsfilt_ext3_init(void)
1430 fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1431 sizeof(struct fsfilt_cb_data), 0,
1434 CERROR("error allocating fsfilt journal callback cache\n");
1435 GOTO(out, rc = -ENOMEM);
1438 rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1441 kmem_cache_destroy(fcb_cache);
1446 static void __exit fsfilt_ext3_exit(void)
1448 fsfilt_unregister_ops(&fsfilt_ext3_ops);
1449 LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1450 "can't free fsfilt callback cache: count %d\n",
1451 atomic_read(&fcb_cache_count));
1454 module_init(fsfilt_ext3_init);
1455 module_exit(fsfilt_ext3_exit);
1457 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1458 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1459 MODULE_LICENSE("GPL");