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 <libcfs/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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,7))
56 # define lock_24kernel() lock_kernel()
57 # define unlock_24kernel() unlock_kernel()
59 # define lock_24kernel() do {} while (0)
60 # define unlock_24kernel() do {} while (0)
63 static kmem_cache_t *fcb_cache;
64 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
66 struct fsfilt_cb_data {
67 struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
68 fsfilt_cb_t cb_func; /* MDS/OBD completion function */
69 struct obd_device *cb_obd; /* MDS/OBD completion device */
70 __u64 cb_last_num; /* MDS/OST last committed operation */
71 void *cb_data; /* MDS/OST completion function data */
74 #ifndef EXT3_XATTR_INDEX_TRUSTED /* temporary until we hit l28 kernel */
75 #define EXT3_XATTR_INDEX_TRUSTED 4
78 #define XATTR_LUSTRE_MDS_LOV_EA "lov"
79 #define XATTR_LUSTRE_MDS_MEA_EA "mea"
80 #define XATTR_LUSTRE_MDS_MID_EA "mid"
81 #define XATTR_LUSTRE_MDS_SID_EA "sid"
84 * We don't currently need any additional blocks for rmdir and
85 * unlink transactions because we are storing the OST oa_id inside
86 * the inode (which we will be changing anyways as part of this
89 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
92 /* For updates to the last recieved file */
93 int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
97 if (current->journal_info) {
98 CDEBUG(D_INODE, "increasing refcount on %p\n",
99 current->journal_info);
104 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
105 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
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;
164 LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
166 handle = journal_start(EXT3_JOURNAL(inode), nblocks);
170 LASSERT(current->journal_info == handle);
172 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
173 op, nblocks, PTR_ERR(handle));
178 * Calculate the number of buffer credits needed to write multiple pages in
179 * a single ext3 transaction. No, this shouldn't be here, but as yet ext3
180 * doesn't have a nice API for calculating this sort of thing in advance.
182 * See comment above ext3_writepage_trans_blocks for details. We assume
183 * no data journaling is being done, but it does allow for all of the pages
184 * being non-contiguous. If we are guaranteed contiguous pages we could
185 * reduce the number of (d)indirect blocks a lot.
187 * With N blocks per page and P pages, for each inode we have at most:
189 * min(N*P, blocksize/4 + 1) dindirect blocks
192 * For the entire filesystem, we have at most:
193 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
194 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
195 * objcount inode blocks
197 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
199 * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
201 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
202 int niocount, struct niobuf_local *nb)
204 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
206 const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
207 int nbitmaps = 0, ngdblocks;
208 int needed = objcount + 1; /* inodes + superblock */
211 for (i = 0, j = 0; i < objcount; i++, fso++) {
212 /* two or more dindirect blocks in case we cross boundary */
213 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
215 sb->s_blocksize_bits) /
216 (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
217 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
219 /* leaf, indirect, tindirect blocks for first block */
220 nbitmaps += blockpp + 2;
222 j += fso->fso_bufcnt;
225 next_indir = nb[0].offset +
226 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
227 for (i = 1; i < niocount; i++) {
228 if (nb[i].offset >= next_indir) {
229 nbitmaps++; /* additional indirect */
230 next_indir = nb[i].offset +
231 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
232 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
233 nbitmaps++; /* additional indirect */
235 nbitmaps += blockpp; /* each leaf in different group? */
238 ngdblocks = nbitmaps;
239 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
240 nbitmaps = EXT3_SB(sb)->s_groups_count;
241 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
242 ngdblocks = EXT3_SB(sb)->s_gdb_count;
244 needed += nbitmaps + ngdblocks;
246 /* last_rcvd update */
247 needed += EXT3_DATA_TRANS_BLOCKS;
249 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
250 /* We assume that there will be 1 bit set in s_dquot.flags for each
251 * quota file that is active. This is at least true for now.
253 needed += hweight32(sb_any_quota_enabled(sb)) *
254 EXT3_SINGLEDATA_TRANS_BLOCKS;
260 /* We have to start a huge journal transaction here to hold all of the
261 * metadata for the pages being written here. This is necessitated by
262 * the fact that we do lots of prepare_write operations before we do
263 * any of the matching commit_write operations, so even if we split
264 * up to use "smaller" transactions none of them could complete until
265 * all of them were opened. By having a single journal transaction,
266 * we eliminate duplicate reservations for common blocks like the
267 * superblock and group descriptors or bitmaps.
269 * We will start the transaction here, but each prepare_write will
270 * add a refcount to the transaction, and each commit_write will
271 * remove a refcount. The transaction will be closed when all of
272 * the pages have been written.
274 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
275 int niocount, struct niobuf_local *nb,
276 void *desc_private, int logs)
283 LASSERT(current->journal_info == desc_private);
284 journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
285 needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
287 /* The number of blocks we could _possibly_ dirty can very large.
288 * We reduce our request if it is absurd (and we couldn't get that
289 * many credits for a single handle anyways).
291 * At some point we have to limit the size of I/Os sent at one time,
292 * increase the size of the journal, or we have to calculate the
293 * actual journal requirements more carefully by checking all of
294 * the blocks instead of being maximally pessimistic. It remains to
295 * be seen if this is a real problem or not.
297 if (needed > journal->j_max_transaction_buffers) {
298 CERROR("want too many journal credits (%d) using %d instead\n",
299 needed, journal->j_max_transaction_buffers);
300 needed = journal->j_max_transaction_buffers;
303 LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
305 handle = journal_start(journal, needed);
307 if (IS_ERR(handle)) {
308 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
311 LASSERT(handle->h_buffer_credits >= needed);
312 LASSERT(current->journal_info == handle);
318 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
319 void *h, int force_sync)
322 handle_t *handle = h;
324 LASSERT(current->journal_info == handle);
326 handle->h_sync = 1; /* recovery likes this */
329 rc = journal_stop(handle);
335 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
339 transaction_t *transaction;
340 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
343 handle_t *handle = h;
347 LASSERT(current->journal_info == handle);
350 transaction = handle->h_transaction;
351 journal = transaction->t_journal;
352 tid = transaction->t_tid;
353 /* we don't want to be blocked */
355 rc = journal_stop(handle);
357 CERROR("error while stopping transaction: %d\n", rc);
361 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
362 rtid = log_start_commit(journal, transaction);
364 CERROR("strange race: %lu != %lu\n",
365 (unsigned long) tid, (unsigned long) rtid);
367 log_start_commit(journal, transaction->t_tid);
371 *wait_handle = (void *) tid;
372 CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
376 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
378 tid_t tid = (tid_t)(long)h;
380 CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
381 if (is_journal_aborted(EXT3_JOURNAL(inode)))
384 log_wait_commit(EXT3_JOURNAL(inode), tid);
389 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
390 struct iattr *iattr, int do_trunc)
392 struct inode *inode = dentry->d_inode;
397 /* A _really_ horrible hack to avoid removing the data stored
398 * in the block pointers; this is really the "small" stripe MD data.
399 * We can avoid further hackery by virtue of the MDS file size being
400 * zero all the time (which doesn't invoke block truncate at unlink
401 * time), so we assert we never change the MDS file size from zero. */
402 if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
403 /* ATTR_SIZE would invoke truncate: clear it */
404 iattr->ia_valid &= ~ATTR_SIZE;
405 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
407 /* make sure _something_ gets set - so new inode
408 * goes to disk (probably won't work over XFS */
409 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
410 iattr->ia_valid |= ATTR_MODE;
411 iattr->ia_mode = inode->i_mode;
415 /* Don't allow setattr to change file type */
416 iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
418 /* We set these flags on the client, but have already checked perms
419 * so don't confuse inode_change_ok. */
420 iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
422 if (inode->i_op->setattr) {
423 rc = inode->i_op->setattr(dentry, iattr);
425 rc = inode_change_ok(inode, iattr);
427 rc = inode_setattr(inode, iattr);
435 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
436 unsigned int cmd, unsigned long arg)
441 if (inode->i_fop->ioctl)
442 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
449 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
450 void *buffer, int buffer_size)
456 rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
457 name, buffer, buffer_size, 0);
460 CERROR("set xattr %s from inode %lu: rc %d\n",
461 name, inode->i_ino, rc);
465 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
466 void *buffer, int buffer_size)
472 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
473 name, buffer, buffer_size);
477 return (rc == -ENODATA) ? 0 : rc;
479 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
480 name, inode->i_ino, rc);
481 memset(buffer, 0, buffer_size);
482 return (rc == -ENODATA) ? 0 : rc;
488 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
489 void *lmm, int lmm_size,
496 rc = fsfilt_ext3_set_xattr(inode, handle,
497 XATTR_LUSTRE_MDS_LOV_EA,
501 rc = fsfilt_ext3_set_xattr(inode, handle,
502 XATTR_LUSTRE_MDS_MEA_EA,
506 rc = fsfilt_ext3_set_xattr(inode, handle,
507 XATTR_LUSTRE_MDS_SID_EA,
511 rc = fsfilt_ext3_set_xattr(inode, handle,
512 XATTR_LUSTRE_MDS_MID_EA,
522 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm,
523 int lmm_size, enum ea_type type)
529 rc = fsfilt_ext3_get_xattr(inode,
530 XATTR_LUSTRE_MDS_LOV_EA,
534 rc = fsfilt_ext3_get_xattr(inode,
535 XATTR_LUSTRE_MDS_MEA_EA,
539 rc = fsfilt_ext3_get_xattr(inode,
540 XATTR_LUSTRE_MDS_SID_EA,
544 rc = fsfilt_ext3_get_xattr(inode,
545 XATTR_LUSTRE_MDS_MID_EA,
555 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, void *bio)
558 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
559 submit_bio(rw, (struct bio *)bio);
561 struct bio *b = (struct kiobuf *)bio;
564 rc = brw_kiovec(rw, 1, &b, inode->i_dev,
565 b->blocks, 1 << inode->i_blkbits);
567 blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
569 if (rc != (1 << inode->i_blkbits) * b->nr_pages * blocks_per_page) {
570 CERROR("short write? expected %d, wrote %d\n",
571 (1 << inode->i_blkbits) * b->nr_pages *
572 blocks_per_page, rc);
578 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
583 page = grab_cache_page(inode->i_mapping, index);
585 return ERR_PTR(-ENOMEM);
587 if (PageUptodate(page)) {
592 rc = inode->i_mapping->a_ops->readpage(NULL, page);
594 page_cache_release(page);
601 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
604 struct inode *inode = file->f_dentry->d_inode;
607 if (S_ISREG(inode->i_mode))
608 rc = file->f_op->read(file, buf, count, off);
610 const int blkbits = inode->i_sb->s_blocksize_bits;
611 const int blksize = inode->i_sb->s_blocksize;
613 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
614 count, inode->i_ino, *off);
616 struct buffer_head *bh;
619 if (*off < inode->i_size) {
622 bh = ext3_bread(NULL, inode, *off >> blkbits,
625 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
628 memcpy(buf, bh->b_data, blksize);
631 /* XXX in theory we should just fake
632 * this buffer and continue like ext3,
633 * especially if this is a partial read
635 CERROR("error read dir %lu+%llu: %d\n",
636 inode->i_ino, *off, err);
641 struct ext3_dir_entry_2 *fake = (void *)buf;
643 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
644 memset(fake, 0, sizeof(*fake));
645 fake->rec_len = cpu_to_le32(blksize);
657 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
659 struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
661 fcb->cb_func(fcb->cb_obd, fcb->cb_last_num, fcb->cb_data, error);
663 OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
664 atomic_dec(&fcb_cache_count);
667 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
668 struct super_block *sb,
669 __u64 last_num, void *handle,
673 struct fsfilt_cb_data *fcb;
675 OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
679 atomic_inc(&fcb_cache_count);
680 fcb->cb_func = cb_func;
682 fcb->cb_last_num = last_num;
683 fcb->cb_data = cb_data;
685 CDEBUG(D_EXT2, "set callback for last_num: "LPD64"\n", last_num);
687 journal_callback_set(handle, fsfilt_ext3_cb_func,
688 (struct journal_callback *)fcb);
694 * We need to hack the return value for the free inode counts because
695 * the current EA code requires one filesystem block per inode with EAs,
696 * so it is possible to run out of blocks before we run out of inodes.
698 * This can be removed when the ext3 EA code is fixed.
700 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
705 memset(&sfs, 0, sizeof(sfs));
707 rc = sb->s_op->statfs(sb, &sfs);
709 if (!rc && sfs.f_bfree < sfs.f_ffree) {
710 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
711 sfs.f_ffree = sfs.f_bfree;
714 statfs_pack(osfs, &sfs);
718 static int fsfilt_ext3_sync(struct super_block *sb)
720 return ext3_force_commit(sb);
723 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
724 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
725 #define ext3_up_truncate_sem(inode) up_write(&EXT3_I(inode)->truncate_sem);
726 #define ext3_down_truncate_sem(inode) down_write(&EXT3_I(inode)->truncate_sem);
728 #define ext3_up_truncate_sem(inode) up(&EXT3_I(inode)->truncate_sem);
729 #define ext3_down_truncate_sem(inode) down(&EXT3_I(inode)->truncate_sem);
732 #include <linux/lustre_version.h>
733 #if EXT3_EXT_MAGIC == 0xf301
734 #define ee_start e_start
735 #define ee_block e_block
738 #ifndef EXT3_BB_MAX_BLOCKS
739 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
740 ext3_new_blocks(handle, inode, count, goal, err)
744 unsigned long *blocks;
751 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
752 unsigned long block, int *aflags)
754 struct ext3_inode_info *ei = EXT3_I(inode);
755 unsigned long bg_start;
756 unsigned long colour;
760 struct ext3_extent *ex;
761 depth = path->p_depth;
763 /* try to predict block placement */
764 if ((ex = path[depth].p_ext)) {
765 if (ex->ee_block + ex->ee_len == block)
767 return ex->ee_start + (block - ex->ee_block);
770 /* it looks index is empty
771 * try to find starting from index itself */
772 if (path[depth].p_bh)
773 return path[depth].p_bh->b_blocknr;
776 /* OK. use inode's group */
777 bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
778 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
779 colour = (current->pid % 16) *
780 (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
781 return bg_start + colour + block;
784 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
785 struct ext3_ext_path *path,
786 struct ext3_extent *newex, int exist)
788 struct inode *inode = tree->inode;
789 struct bpointers *bp = tree->private;
790 int count, err, goal;
791 unsigned long pblock;
798 EXT_ASSERT(i == path->p_depth);
799 EXT_ASSERT(path[i].p_hdr);
806 if (bp->create == 0) {
808 if (newex->ee_block < bp->start)
809 i = bp->start - newex->ee_block;
810 if (i >= newex->ee_len)
811 CERROR("nothing to do?! i = %d, e_num = %u\n",
813 for (; i < newex->ee_len && bp->num; i++) {
824 tgen = EXT_GENERATION(tree);
825 count = ext3_ext_calc_credits_for_insert(tree, path);
826 ext3_up_truncate_sem(inode);
828 handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
830 if (IS_ERR(handle)) {
831 ext3_down_truncate_sem(inode);
832 return PTR_ERR(handle);
835 if (tgen != EXT_GENERATION(tree)) {
836 /* the tree has changed. so path can be invalid at moment */
838 journal_stop(handle);
840 ext3_down_truncate_sem(inode);
843 ext3_down_truncate_sem(inode);
844 count = newex->ee_len;
845 goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
846 aflags |= 2; /* block have been already reserved */
847 pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
850 EXT_ASSERT(count <= newex->ee_len);
852 /* insert new extent */
853 newex->ee_start = pblock;
854 newex->ee_len = count;
855 err = ext3_ext_insert_extent(handle, tree, path, newex);
859 /* correct on-disk inode size */
860 if (newex->ee_len > 0) {
861 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
862 new_i_size = new_i_size << inode->i_blkbits;
863 if (new_i_size > EXT3_I(inode)->i_disksize) {
864 EXT3_I(inode)->i_disksize = new_i_size;
865 err = ext3_mark_inode_dirty(handle, inode);
870 journal_stop(handle);
876 CERROR("hmm. why do we find this extent?\n");
877 CERROR("initial space: %lu:%u\n",
878 bp->start, bp->init_num);
879 CERROR("current extent: %u/%u/%u %d\n",
880 newex->ee_block, newex->ee_len,
881 newex->ee_start, exist);
884 if (newex->ee_block < bp->start)
885 i = bp->start - newex->ee_block;
886 if (i >= newex->ee_len)
887 CERROR("nothing to do?! i = %d, e_num = %u\n",
889 for (; i < newex->ee_len && bp->num; i++) {
890 *(bp->created) = (exist == 0 ? 1 : 0);
892 *(bp->blocks) = newex->ee_start + i;
901 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
902 unsigned long num, unsigned long *blocks,
903 int *created, int create)
905 struct ext3_extents_tree tree;
909 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
910 block, block + num, (unsigned) inode->i_ino);
912 ext3_init_tree_desc(&tree, inode);
915 bp.created = created;
917 bp.init_num = bp.num = num;
920 ext3_down_truncate_sem(inode);
921 err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
922 ext3_ext_invalidate_cache(&tree);
923 ext3_up_truncate_sem(inode);
927 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
928 int pages, unsigned long *blocks,
929 int *created, int create)
931 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
933 struct page *fp = NULL;
936 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
937 inode->i_ino, pages, (*page)->index);
939 /* pages are sorted already. so, we just have to find
940 * contig. space and process them properly */
943 /* start new extent */
948 } else if (fp->index + clen == (*page)->index) {
949 /* continue the extent */
956 /* process found extent */
957 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
958 clen * blocks_per_page, blocks,
963 /* look for next extent */
965 blocks += blocks_per_page * clen;
966 created += blocks_per_page * clen;
970 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
971 clen * blocks_per_page, blocks,
978 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
979 unsigned long *blocks, int *created, int create);
980 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
981 int pages, unsigned long *blocks,
982 int *created, int create)
984 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
988 for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
989 rc = ext3_map_inode_page(inode, *page, b, cr, create);
991 CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
992 inode->i_ino, *b, *cr, create, rc);
996 b += blocks_per_page;
997 cr += blocks_per_page;
1002 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
1003 int pages, unsigned long *blocks,
1004 int *created, int create,
1005 struct semaphore *optional_sem)
1008 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
1009 if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
1010 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
1011 blocks, created, create);
1015 if (optional_sem != NULL)
1017 rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
1019 if (optional_sem != NULL)
1025 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
1026 int nblocks, loff_t newsize);
1027 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
1028 int nblocks, loff_t newsize)
1030 return ext3_prep_san_write(inode, blocks, nblocks, newsize);
1033 static int fsfilt_ext3_read_record(struct file * file, void *buf,
1034 int size, loff_t *offs)
1036 struct inode *inode = file->f_dentry->d_inode;
1037 unsigned long block;
1038 struct buffer_head *bh;
1039 int err, blocksize, csize, boffs;
1041 /* prevent reading after eof */
1043 if (inode->i_size < *offs + size) {
1044 size = inode->i_size - *offs;
1047 CERROR("size %llu is too short for read %u@%llu\n",
1048 inode->i_size, size, *offs);
1050 } else if (size == 0) {
1057 blocksize = 1 << inode->i_blkbits;
1060 block = *offs >> inode->i_blkbits;
1061 boffs = *offs & (blocksize - 1);
1062 csize = min(blocksize - boffs, size);
1063 bh = ext3_bread(NULL, inode, block, 0, &err);
1065 CERROR("can't read block: %d\n", err);
1069 memcpy(buf, bh->b_data + boffs, csize);
1079 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1080 loff_t *offs, int force_sync)
1082 struct buffer_head *bh = NULL;
1083 unsigned long block;
1084 struct inode *inode = file->f_dentry->d_inode;
1085 loff_t old_size = inode->i_size, offset = *offs;
1086 loff_t new_size = inode->i_size;
1089 int err = 0, block_count = 0, blocksize, size, boffs;
1091 /* Determine how many transaction credits are needed */
1092 blocksize = 1 << inode->i_blkbits;
1093 block_count = (*offs & (blocksize - 1)) + bufsize;
1094 block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1096 journal = EXT3_SB(inode->i_sb)->s_journal;
1098 handle = journal_start(journal,
1099 block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1101 if (IS_ERR(handle)) {
1102 CERROR("can't start transaction\n");
1103 return PTR_ERR(handle);
1106 while (bufsize > 0) {
1110 block = offset >> inode->i_blkbits;
1111 boffs = offset & (blocksize - 1);
1112 size = min(blocksize - boffs, bufsize);
1113 bh = ext3_bread(handle, inode, block, 1, &err);
1115 CERROR("can't read/create block: %d\n", err);
1119 err = ext3_journal_get_write_access(handle, bh);
1121 CERROR("journal_get_write_access() returned error %d\n",
1125 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1126 memcpy(bh->b_data + boffs, buf, size);
1127 err = ext3_journal_dirty_metadata(handle, bh);
1129 CERROR("journal_dirty_metadata() returned error %d\n",
1133 if (offset + size > new_size)
1134 new_size = offset + size;
1141 handle->h_sync = 1; /* recovery likes this */
1146 /* correct in-core and on-disk sizes */
1147 if (new_size > inode->i_size) {
1149 if (new_size > inode->i_size)
1150 inode->i_size = new_size;
1151 if (inode->i_size > EXT3_I(inode)->i_disksize)
1152 EXT3_I(inode)->i_disksize = inode->i_size;
1153 if (inode->i_size > old_size)
1154 mark_inode_dirty(inode);
1159 journal_stop(handle);
1167 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1169 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1170 struct mds_obd *mds = &obd->u.mds;
1173 EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1174 EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1177 CWARN("Enabling PDIROPS\n");
1178 set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1179 sb->s_flags |= S_PDIROPS;
1181 /* setup mdsnum in underlying fs */
1182 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1183 if (mds->mds_md_obd) {
1184 struct ext3_sb_info *sbi = EXT3_SB(sb);
1185 struct ext3_super_block *es = sbi->s_es;
1189 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1190 CWARN("%s: set mdsnum %d in ext3\n",
1191 obd->obd_name, mds->mds_num);
1193 handle = journal_start(sbi->s_journal, 1);
1195 LASSERT(!IS_ERR(handle));
1196 err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1198 EXT3_SET_INCOMPAT_FEATURE(sb,
1199 EXT3_FEATURE_INCOMPAT_MDSNUM);
1200 es->s_mdsnum = mds->mds_num;
1201 err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1204 journal_stop(handle);
1207 CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1208 obd->obd_name, es->s_mdsnum);
1210 sbi->s_mdsnum = es->s_mdsnum;
1216 extern int ext3_add_dir_entry(struct dentry *dentry);
1217 extern int ext3_del_dir_entry(struct dentry *dentry);
1219 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1220 struct dentry *parent,
1221 char *name, int namelen,
1223 unsigned long generation,
1227 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1228 struct dentry *dentry;
1231 LASSERT(namelen != 0);
1232 dentry = ll_lookup_one_len(name, parent, namelen);
1233 if (IS_ERR(dentry)) {
1234 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1235 dentry->d_name.name, dentry->d_inode->i_ino,
1236 (unsigned long) dentry->d_inode->i_generation,
1237 (int) PTR_ERR(dentry));
1238 RETURN(PTR_ERR(dentry));
1240 if (dentry->d_inode != NULL || dentry->d_flags & DCACHE_CROSS_REF) {
1241 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1242 dentry->d_name.name, dentry);
1247 /* mds_reint_rename() may use this method to add dir entry
1248 * that points onto local inode. and we don't want to find
1249 * it cross-ref by subsequent lookups */
1252 dentry->d_flags |= DCACHE_CROSS_REF;
1253 dentry->d_inum = ino;
1254 dentry->d_mdsnum = mds;
1255 dentry->d_generation = generation;
1256 dentry->d_fid = fid;
1258 err = ext3_add_dir_entry(dentry);
1265 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1270 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1271 struct dentry *dentry)
1273 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1276 err = ext3_del_dir_entry(dentry);
1282 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1287 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1288 objects. Logs is number of logfiles to update */
1289 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1293 case FSFILT_OP_CREATE:
1294 /* directory leaf, index & indirect & EA*/
1295 return 4 + 3 * logs;
1296 case FSFILT_OP_UNLINK:
1302 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1303 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1304 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1305 for (i = 0; i < op; i++, fso++) {
1306 int nblocks = fso->fso_bufcnt * blockpp;
1307 int ndindirect = min(nblocks, addrpp + 1);
1308 int nindir = nblocks + ndindirect + 1;
1312 return needed + 3 * logs;
1319 #define EXTENTS_EA "write_extents"
1320 #define EXTENTS_EA_SIZE 64
1322 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1323 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1324 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1325 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1327 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1333 rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1334 EXTENTS_EA, from, num);
1338 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1344 rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1345 EXTENTS_EA, from, num);
1349 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1350 const char *eaname, int size);
1352 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1356 rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1361 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1362 char **pbuf, int *size)
1366 rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1367 EXTENTS_EA, pbuf, size);
1371 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1375 rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1380 static struct fsfilt_operations fsfilt_ext3_ops = {
1382 .fs_owner = THIS_MODULE,
1383 .fs_start = fsfilt_ext3_start,
1384 .fs_brw_start = fsfilt_ext3_brw_start,
1385 .fs_commit = fsfilt_ext3_commit,
1386 .fs_commit_async = fsfilt_ext3_commit_async,
1387 .fs_commit_wait = fsfilt_ext3_commit_wait,
1388 .fs_setattr = fsfilt_ext3_setattr,
1389 .fs_iocontrol = fsfilt_ext3_iocontrol,
1390 .fs_set_md = fsfilt_ext3_set_md,
1391 .fs_get_md = fsfilt_ext3_get_md,
1392 .fs_readpage = fsfilt_ext3_readpage,
1393 .fs_add_journal_cb = fsfilt_ext3_add_journal_cb,
1394 .fs_statfs = fsfilt_ext3_statfs,
1395 .fs_sync = fsfilt_ext3_sync,
1396 .fs_map_inode_pages = fsfilt_ext3_map_inode_pages,
1397 .fs_prep_san_write = fsfilt_ext3_prep_san_write,
1398 .fs_write_record = fsfilt_ext3_write_record,
1399 .fs_read_record = fsfilt_ext3_read_record,
1400 .fs_setup = fsfilt_ext3_setup,
1401 .fs_getpage = fsfilt_ext3_getpage,
1402 .fs_send_bio = fsfilt_ext3_send_bio,
1403 .fs_set_xattr = fsfilt_ext3_set_xattr,
1404 .fs_get_xattr = fsfilt_ext3_get_xattr,
1405 .fs_get_op_len = fsfilt_ext3_get_op_len,
1406 .fs_add_dir_entry = fsfilt_ext3_add_dir_entry,
1407 .fs_del_dir_entry = fsfilt_ext3_del_dir_entry,
1408 .fs_init_extents_ea = fsfilt_ext3_init_extents_ea,
1409 .fs_insert_extents_ea = fsfilt_ext3_insert_extents_ea,
1410 .fs_remove_extents_ea = fsfilt_ext3_remove_extents_ea,
1411 .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1412 .fs_get_write_extents_num = fsfilt_ext3_get_write_extents_num,
1415 static int __init fsfilt_ext3_init(void)
1419 fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1420 sizeof(struct fsfilt_cb_data), 0,
1423 CERROR("error allocating fsfilt journal callback cache\n");
1424 GOTO(out, rc = -ENOMEM);
1427 rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1430 kmem_cache_destroy(fcb_cache);
1435 static void __exit fsfilt_ext3_exit(void)
1437 fsfilt_unregister_ops(&fsfilt_ext3_ops);
1438 LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1439 "can't free fsfilt callback cache: count %d\n",
1440 atomic_read(&fcb_cache_count));
1443 module_init(fsfilt_ext3_init);
1444 module_exit(fsfilt_ext3_exit);
1446 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1447 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1448 MODULE_LICENSE("GPL");