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"
82 #define XATTR_LUSTRE_MDS_KEY_EA "key"
85 * We don't currently need any additional blocks for rmdir and
86 * unlink transactions because we are storing the OST oa_id inside
87 * the inode (which we will be changing anyways as part of this
90 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
93 /* For updates to the last recieved file */
94 int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
98 if (current->journal_info) {
99 CDEBUG(D_INODE, "increasing refcount on %p\n",
100 current->journal_info);
105 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
106 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
109 case FSFILT_OP_RMDIR:
110 case FSFILT_OP_UNLINK:
111 /* delete one file + create/update logs for each stripe */
112 nblocks += EXT3_DELETE_TRANS_BLOCKS;
113 /*nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
114 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;*/
116 case FSFILT_OP_RENAME:
117 /* modify additional directory */
118 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
120 case FSFILT_OP_SYMLINK:
121 /* additional block + block bitmap + GDT for long symlink */
124 case FSFILT_OP_CREATE:
125 /* create/update logs for each stripe */
126 /*nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
127 EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;*/
129 case FSFILT_OP_MKDIR:
130 case FSFILT_OP_MKNOD:
131 /* modify one inode + block bitmap + GDT */
135 /* modify parent directory */
136 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
137 EXT3_DATA_TRANS_BLOCKS;
139 case FSFILT_OP_SETATTR:
140 /* Setattr on inode */
143 case FSFILT_OP_CANCEL_UNLINK:
144 /* blocks for log header bitmap update OR
145 * blocks for catalog header bitmap update + unlink of logs */
146 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
147 EXT3_DELETE_TRANS_BLOCKS * logs;
150 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
152 default: CERROR("unknown transaction start op %d\n", op);
156 LASSERT(current->journal_info == desc_private);
157 journal = EXT3_SB(inode->i_sb)->s_journal;
158 if (nblocks > journal->j_max_transaction_buffers) {
159 CERROR("too many credits %d for op %ux%u using %d instead\n",
160 nblocks, op, logs, journal->j_max_transaction_buffers);
161 nblocks = journal->j_max_transaction_buffers;
165 LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
167 handle = journal_start(EXT3_JOURNAL(inode), nblocks);
171 LASSERT(current->journal_info == handle);
173 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
174 op, nblocks, PTR_ERR(handle));
179 * Calculate the number of buffer credits needed to write multiple pages in
180 * a single ext3 transaction. No, this shouldn't be here, but as yet ext3
181 * doesn't have a nice API for calculating this sort of thing in advance.
183 * See comment above ext3_writepage_trans_blocks for details. We assume
184 * no data journaling is being done, but it does allow for all of the pages
185 * being non-contiguous. If we are guaranteed contiguous pages we could
186 * reduce the number of (d)indirect blocks a lot.
188 * With N blocks per page and P pages, for each inode we have at most:
190 * min(N*P, blocksize/4 + 1) dindirect blocks
193 * For the entire filesystem, we have at most:
194 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
195 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
196 * objcount inode blocks
198 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
200 * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
202 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
203 int niocount, struct niobuf_local *nb)
205 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
207 const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
208 int nbitmaps = 0, ngdblocks;
209 int needed = objcount + 1; /* inodes + superblock */
212 for (i = 0, j = 0; i < objcount; i++, fso++) {
213 /* two or more dindirect blocks in case we cross boundary */
214 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
216 sb->s_blocksize_bits) /
217 (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
218 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
220 /* leaf, indirect, tindirect blocks for first block */
221 nbitmaps += blockpp + 2;
223 j += fso->fso_bufcnt;
226 next_indir = nb[0].offset +
227 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
228 for (i = 1; i < niocount; i++) {
229 if (nb[i].offset >= next_indir) {
230 nbitmaps++; /* additional indirect */
231 next_indir = nb[i].offset +
232 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
233 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
234 nbitmaps++; /* additional indirect */
236 nbitmaps += blockpp; /* each leaf in different group? */
239 ngdblocks = nbitmaps;
240 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
241 nbitmaps = EXT3_SB(sb)->s_groups_count;
242 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
243 ngdblocks = EXT3_SB(sb)->s_gdb_count;
245 needed += nbitmaps + ngdblocks;
247 /* last_rcvd update */
248 needed += EXT3_DATA_TRANS_BLOCKS;
250 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
251 /* We assume that there will be 1 bit set in s_dquot.flags for each
252 * quota file that is active. This is at least true for now.
254 needed += hweight32(sb_any_quota_enabled(sb)) *
255 EXT3_SINGLEDATA_TRANS_BLOCKS;
261 /* We have to start a huge journal transaction here to hold all of the
262 * metadata for the pages being written here. This is necessitated by
263 * the fact that we do lots of prepare_write operations before we do
264 * any of the matching commit_write operations, so even if we split
265 * up to use "smaller" transactions none of them could complete until
266 * all of them were opened. By having a single journal transaction,
267 * we eliminate duplicate reservations for common blocks like the
268 * superblock and group descriptors or bitmaps.
270 * We will start the transaction here, but each prepare_write will
271 * add a refcount to the transaction, and each commit_write will
272 * remove a refcount. The transaction will be closed when all of
273 * the pages have been written.
275 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
276 int niocount, struct niobuf_local *nb,
277 void *desc_private, int logs)
284 LASSERT(current->journal_info == desc_private);
285 journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
286 needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
288 /* The number of blocks we could _possibly_ dirty can very large.
289 * We reduce our request if it is absurd (and we couldn't get that
290 * many credits for a single handle anyways).
292 * At some point we have to limit the size of I/Os sent at one time,
293 * increase the size of the journal, or we have to calculate the
294 * actual journal requirements more carefully by checking all of
295 * the blocks instead of being maximally pessimistic. It remains to
296 * be seen if this is a real problem or not.
298 if (needed > journal->j_max_transaction_buffers) {
299 CERROR("want too many journal credits (%d) using %d instead\n",
300 needed, journal->j_max_transaction_buffers);
301 needed = journal->j_max_transaction_buffers;
304 LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
306 handle = journal_start(journal, needed);
308 if (IS_ERR(handle)) {
309 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
312 LASSERT(handle->h_buffer_credits >= needed);
313 LASSERT(current->journal_info == handle);
319 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
320 void *h, int force_sync)
323 handle_t *handle = h;
325 LASSERT(current->journal_info == handle);
327 handle->h_sync = 1; /* recovery likes this */
330 rc = journal_stop(handle);
336 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
340 transaction_t *transaction;
341 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
344 handle_t *handle = h;
348 LASSERT(current->journal_info == handle);
351 transaction = handle->h_transaction;
352 journal = transaction->t_journal;
353 tid = transaction->t_tid;
354 /* we don't want to be blocked */
356 rc = journal_stop(handle);
358 CERROR("error while stopping transaction: %d\n", rc);
362 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
363 rtid = log_start_commit(journal, transaction);
365 CERROR("strange race: %lu != %lu\n",
366 (unsigned long) tid, (unsigned long) rtid);
368 log_start_commit(journal, tid);
372 *wait_handle = (void *) tid;
373 CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
377 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
379 tid_t tid = (tid_t)(long)h;
381 CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
382 if (is_journal_aborted(EXT3_JOURNAL(inode)))
385 log_wait_commit(EXT3_JOURNAL(inode), tid);
390 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
391 struct iattr *iattr, int do_trunc)
393 struct inode *inode = dentry->d_inode;
398 /* A _really_ horrible hack to avoid removing the data stored
399 * in the block pointers; this is really the "small" stripe MD data.
400 * We can avoid further hackery by virtue of the MDS file size being
401 * zero all the time (which doesn't invoke block truncate at unlink
402 * time), so we assert we never change the MDS file size from zero. */
403 if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
404 /* ATTR_SIZE would invoke truncate: clear it */
405 iattr->ia_valid &= ~ATTR_SIZE;
406 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
408 /* make sure _something_ gets set - so new inode
409 * goes to disk (probably won't work over XFS */
410 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
411 iattr->ia_valid |= ATTR_MODE;
412 iattr->ia_mode = inode->i_mode;
416 /* Don't allow setattr to change file type */
417 iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
419 /* We set these flags on the client, but have already checked perms
420 * so don't confuse inode_change_ok. */
421 iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
423 if (inode->i_op->setattr) {
424 rc = inode->i_op->setattr(dentry, iattr);
426 rc = inode_change_ok(inode, iattr);
428 rc = inode_setattr(inode, iattr);
436 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
437 unsigned int cmd, unsigned long arg)
442 if (inode->i_fop->ioctl)
443 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
450 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
451 void *buffer, int buffer_size)
457 rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
458 name, buffer, buffer_size, 0);
461 CERROR("set xattr %s from inode %lu: rc %d\n",
462 name, inode->i_ino, rc);
466 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
467 void *buffer, int buffer_size)
473 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
474 name, buffer, buffer_size);
478 return (rc == -ENODATA) ? 0 : rc;
480 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
481 name, inode->i_ino, rc);
482 memset(buffer, 0, buffer_size);
483 return (rc == -ENODATA) ? 0 : rc;
489 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
490 void *lmm, int lmm_size,
497 rc = fsfilt_ext3_set_xattr(inode, handle,
498 XATTR_LUSTRE_MDS_LOV_EA,
502 rc = fsfilt_ext3_set_xattr(inode, handle,
503 XATTR_LUSTRE_MDS_MEA_EA,
507 rc = fsfilt_ext3_set_xattr(inode, handle,
508 XATTR_LUSTRE_MDS_SID_EA,
512 rc = fsfilt_ext3_set_xattr(inode, handle,
513 XATTR_LUSTRE_MDS_MID_EA,
517 rc = fsfilt_ext3_set_xattr(inode, handle,
518 XATTR_LUSTRE_MDS_KEY_EA,
528 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm,
529 int lmm_size, enum ea_type type)
535 rc = fsfilt_ext3_get_xattr(inode,
536 XATTR_LUSTRE_MDS_LOV_EA,
540 rc = fsfilt_ext3_get_xattr(inode,
541 XATTR_LUSTRE_MDS_MEA_EA,
545 rc = fsfilt_ext3_get_xattr(inode,
546 XATTR_LUSTRE_MDS_SID_EA,
550 rc = fsfilt_ext3_get_xattr(inode,
551 XATTR_LUSTRE_MDS_MID_EA,
555 rc = fsfilt_ext3_get_xattr(inode,
556 XATTR_LUSTRE_MDS_KEY_EA,
566 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, void *bio)
569 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
570 submit_bio(rw, (struct bio *)bio);
572 struct bio *b = (struct kiobuf *)bio;
575 rc = brw_kiovec(rw, 1, &b, inode->i_dev,
576 b->blocks, 1 << inode->i_blkbits);
578 blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
580 if (rc != (1 << inode->i_blkbits) * b->nr_pages * blocks_per_page) {
581 CERROR("short write? expected %d, wrote %d\n",
582 (1 << inode->i_blkbits) * b->nr_pages *
583 blocks_per_page, rc);
589 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
594 page = grab_cache_page(inode->i_mapping, index);
596 return ERR_PTR(-ENOMEM);
598 if (PageUptodate(page)) {
603 rc = inode->i_mapping->a_ops->readpage(NULL, page);
605 page_cache_release(page);
612 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
615 struct inode *inode = file->f_dentry->d_inode;
618 if (S_ISREG(inode->i_mode))
619 rc = file->f_op->read(file, buf, count, off);
621 const int blkbits = inode->i_sb->s_blocksize_bits;
622 const int blksize = inode->i_sb->s_blocksize;
624 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
625 count, inode->i_ino, *off);
627 struct buffer_head *bh;
630 if (*off < inode->i_size) {
633 bh = ext3_bread(NULL, inode, *off >> blkbits,
636 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
639 memcpy(buf, bh->b_data, blksize);
642 /* XXX in theory we should just fake
643 * this buffer and continue like ext3,
644 * especially if this is a partial read
646 CERROR("error read dir %lu+%llu: %d\n",
647 inode->i_ino, *off, err);
652 struct ext3_dir_entry_2 *fake = (void *)buf;
654 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
655 memset(fake, 0, sizeof(*fake));
656 fake->rec_len = cpu_to_le32(blksize);
668 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
670 struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
672 fcb->cb_func(fcb->cb_obd, fcb->cb_last_num, fcb->cb_data, error);
674 OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
675 atomic_dec(&fcb_cache_count);
678 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
679 struct super_block *sb,
680 __u64 last_num, void *handle,
684 struct fsfilt_cb_data *fcb;
686 OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
690 atomic_inc(&fcb_cache_count);
691 fcb->cb_func = cb_func;
693 fcb->cb_last_num = last_num;
694 fcb->cb_data = cb_data;
696 CDEBUG(D_EXT2, "set callback for last_num: "LPD64"\n", last_num);
698 journal_callback_set(handle, fsfilt_ext3_cb_func,
699 (struct journal_callback *)fcb);
705 * We need to hack the return value for the free inode counts because
706 * the current EA code requires one filesystem block per inode with EAs,
707 * so it is possible to run out of blocks before we run out of inodes.
709 * This can be removed when the ext3 EA code is fixed.
711 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
716 memset(&sfs, 0, sizeof(sfs));
718 rc = sb->s_op->statfs(sb, &sfs);
720 if (!rc && sfs.f_bfree < sfs.f_ffree) {
721 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
722 sfs.f_ffree = sfs.f_bfree;
725 statfs_pack(osfs, &sfs);
729 static int fsfilt_ext3_sync(struct super_block *sb)
731 return ext3_force_commit(sb);
734 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
735 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
736 #define ext3_up_truncate_sem(inode) up_write(&EXT3_I(inode)->truncate_sem);
737 #define ext3_down_truncate_sem(inode) down_write(&EXT3_I(inode)->truncate_sem);
739 #define ext3_up_truncate_sem(inode) up(&EXT3_I(inode)->truncate_sem);
740 #define ext3_down_truncate_sem(inode) down(&EXT3_I(inode)->truncate_sem);
743 #include <linux/lustre_version.h>
744 #if EXT3_EXT_MAGIC == 0xf301
745 #define ee_start e_start
746 #define ee_block e_block
749 #ifndef EXT3_BB_MAX_BLOCKS
750 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
751 ext3_new_blocks(handle, inode, count, goal, err)
755 unsigned long *blocks;
762 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
763 unsigned long block, int *aflags)
765 struct ext3_inode_info *ei = EXT3_I(inode);
766 unsigned long bg_start;
767 unsigned long colour;
771 struct ext3_extent *ex;
772 depth = path->p_depth;
774 /* try to predict block placement */
775 if ((ex = path[depth].p_ext)) {
777 /* This prefers to eat into a contiguous extent
778 * rather than find an extent that the whole
779 * request will fit into. This can fragment data
780 * block allocation and prevents our lovely 1M I/Os
781 * from reaching the disk intact. */
783 if (ex->ee_block + ex->ee_len == block)
786 return ex->ee_start + (block - ex->ee_block);
789 /* it looks index is empty
790 * try to find starting from index itself */
791 if (path[depth].p_bh)
792 return path[depth].p_bh->b_blocknr;
795 /* OK. use inode's group */
796 bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
797 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
798 colour = (current->pid % 16) *
799 (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
800 return bg_start + colour + block;
803 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
804 struct ext3_ext_path *path,
805 struct ext3_extent *newex, int exist)
807 struct inode *inode = tree->inode;
808 struct bpointers *bp = tree->private;
809 int count, err, goal;
810 unsigned long pblock;
817 EXT_ASSERT(i == path->p_depth);
818 EXT_ASSERT(path[i].p_hdr);
825 if (bp->create == 0) {
827 if (newex->ee_block < bp->start)
828 i = bp->start - newex->ee_block;
829 if (i >= newex->ee_len)
830 CERROR("nothing to do?! i = %d, e_num = %u\n",
832 for (; i < newex->ee_len && bp->num; i++) {
843 tgen = EXT_GENERATION(tree);
844 count = ext3_ext_calc_credits_for_insert(tree, path);
845 ext3_up_truncate_sem(inode);
847 handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
849 if (IS_ERR(handle)) {
850 ext3_down_truncate_sem(inode);
851 return PTR_ERR(handle);
854 if (tgen != EXT_GENERATION(tree)) {
855 /* the tree has changed. so path can be invalid at moment */
857 journal_stop(handle);
859 ext3_down_truncate_sem(inode);
862 ext3_down_truncate_sem(inode);
863 count = newex->ee_len;
864 goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
865 aflags |= 2; /* block have been already reserved */
866 pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
869 EXT_ASSERT(count <= newex->ee_len);
871 /* insert new extent */
872 newex->ee_start = pblock;
873 newex->ee_len = count;
874 err = ext3_ext_insert_extent(handle, tree, path, newex);
878 /* correct on-disk inode size */
879 if (newex->ee_len > 0) {
880 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
881 new_i_size = new_i_size << inode->i_blkbits;
882 if (new_i_size > EXT3_I(inode)->i_disksize) {
883 EXT3_I(inode)->i_disksize = new_i_size;
884 err = ext3_mark_inode_dirty(handle, inode);
889 journal_stop(handle);
895 CERROR("hmm. why do we find this extent?\n");
896 CERROR("initial space: %lu:%u\n",
897 bp->start, bp->init_num);
898 CERROR("current extent: %u/%u/%u %d\n",
899 newex->ee_block, newex->ee_len,
900 newex->ee_start, exist);
903 if (newex->ee_block < bp->start)
904 i = bp->start - newex->ee_block;
905 if (i >= newex->ee_len)
906 CERROR("nothing to do?! i = %d, e_num = %u\n",
908 for (; i < newex->ee_len && bp->num; i++) {
909 *(bp->created) = (exist == 0 ? 1 : 0);
911 *(bp->blocks) = newex->ee_start + i;
920 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
921 unsigned long num, unsigned long *blocks,
922 int *created, int create)
924 struct ext3_extents_tree tree;
928 CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
929 block, block + num, (unsigned) inode->i_ino);
931 ext3_init_tree_desc(&tree, inode);
934 bp.created = created;
936 bp.init_num = bp.num = num;
939 ext3_down_truncate_sem(inode);
940 err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
941 ext3_ext_invalidate_cache(&tree);
942 ext3_up_truncate_sem(inode);
944 /* unmap underlying pages/buffers from blockdevice mapping */
946 struct block_device *bdev = inode->i_sb->s_bdev;
947 for (i = 0; i < num; i++) {
950 unmap_underlying_metadata(bdev, blocks[i]);
956 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
957 int pages, unsigned long *blocks,
958 int *created, int create)
960 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
962 struct page *fp = NULL;
965 CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
966 inode->i_ino, pages, (*page)->index);
968 /* pages are sorted already. so, we just have to find
969 * contig. space and process them properly */
972 /* start new extent */
977 } else if (fp->index + clen == (*page)->index) {
978 /* continue the extent */
985 /* process found extent */
986 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
987 clen * blocks_per_page, blocks,
992 /* look for next extent */
994 blocks += blocks_per_page * clen;
995 created += blocks_per_page * clen;
999 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
1000 clen * blocks_per_page, blocks,
1007 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
1008 unsigned long *blocks, int *created, int create);
1009 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
1010 int pages, unsigned long *blocks,
1011 int *created, int create)
1013 int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1017 for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
1018 rc = ext3_map_inode_page(inode, *page, b, cr, create);
1020 CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
1021 inode->i_ino, *b, *cr, create, rc);
1025 b += blocks_per_page;
1026 cr += blocks_per_page;
1031 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
1032 int pages, unsigned long *blocks,
1033 int *created, int create,
1034 struct semaphore *optional_sem)
1037 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
1038 if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
1039 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
1040 blocks, created, create);
1044 if (optional_sem != NULL)
1046 rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
1048 if (optional_sem != NULL)
1054 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
1055 int nblocks, loff_t newsize);
1056 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
1057 int nblocks, loff_t newsize)
1059 return ext3_prep_san_write(inode, blocks, nblocks, newsize);
1062 static int fsfilt_ext3_read_record(struct file * file, void *buf,
1063 int size, loff_t *offs)
1065 struct inode *inode = file->f_dentry->d_inode;
1066 unsigned long block;
1067 struct buffer_head *bh;
1068 int err, blocksize, csize, boffs;
1070 /* prevent reading after eof */
1072 if (inode->i_size < *offs + size) {
1073 size = inode->i_size - *offs;
1076 CERROR("size %llu is too short for read %u@%llu\n",
1077 inode->i_size, size, *offs);
1079 } else if (size == 0) {
1086 blocksize = 1 << inode->i_blkbits;
1089 block = *offs >> inode->i_blkbits;
1090 boffs = *offs & (blocksize - 1);
1091 csize = min(blocksize - boffs, size);
1092 bh = ext3_bread(NULL, inode, block, 0, &err);
1094 CERROR("can't read block: %d\n", err);
1098 memcpy(buf, bh->b_data + boffs, csize);
1108 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1109 loff_t *offs, int force_sync)
1111 struct buffer_head *bh = NULL;
1112 unsigned long block;
1113 struct inode *inode = file->f_dentry->d_inode;
1114 loff_t old_size = inode->i_size, offset = *offs;
1115 loff_t new_size = inode->i_size;
1118 int err = 0, block_count = 0, blocksize, size, boffs;
1120 /* Determine how many transaction credits are needed */
1121 blocksize = 1 << inode->i_blkbits;
1122 block_count = (*offs & (blocksize - 1)) + bufsize;
1123 block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1125 journal = EXT3_SB(inode->i_sb)->s_journal;
1127 handle = journal_start(journal,
1128 block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1130 if (IS_ERR(handle)) {
1131 CERROR("can't start transaction\n");
1132 return PTR_ERR(handle);
1135 while (bufsize > 0) {
1139 block = offset >> inode->i_blkbits;
1140 boffs = offset & (blocksize - 1);
1141 size = min(blocksize - boffs, bufsize);
1142 bh = ext3_bread(handle, inode, block, 1, &err);
1144 CERROR("can't read/create block: %d\n", err);
1148 err = ext3_journal_get_write_access(handle, bh);
1150 CERROR("journal_get_write_access() returned error %d\n",
1154 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1155 memcpy(bh->b_data + boffs, buf, size);
1156 err = ext3_journal_dirty_metadata(handle, bh);
1158 CERROR("journal_dirty_metadata() returned error %d\n",
1162 if (offset + size > new_size)
1163 new_size = offset + size;
1170 handle->h_sync = 1; /* recovery likes this */
1175 /* correct in-core and on-disk sizes */
1176 if (new_size > inode->i_size) {
1178 if (new_size > inode->i_size)
1179 inode->i_size = new_size;
1180 if (inode->i_size > EXT3_I(inode)->i_disksize)
1181 EXT3_I(inode)->i_disksize = inode->i_size;
1182 if (inode->i_size > old_size)
1183 mark_inode_dirty(inode);
1188 journal_stop(handle);
1196 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1198 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1199 struct mds_obd *mds = &obd->u.mds;
1202 EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1203 EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1206 CWARN("Enabling PDIROPS\n");
1207 set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1208 sb->s_flags |= S_PDIROPS;
1210 /* setup mdsnum in underlying fs */
1211 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1212 if (mds->mds_md_obd) {
1213 struct ext3_sb_info *sbi = EXT3_SB(sb);
1214 struct ext3_super_block *es = sbi->s_es;
1218 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1219 CWARN("%s: set mdsnum %d in ext3\n",
1220 obd->obd_name, mds->mds_num);
1222 handle = journal_start(sbi->s_journal, 1);
1224 LASSERT(!IS_ERR(handle));
1225 err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1227 EXT3_SET_INCOMPAT_FEATURE(sb,
1228 EXT3_FEATURE_INCOMPAT_MDSNUM);
1229 es->s_mdsnum = mds->mds_num;
1230 err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1233 journal_stop(handle);
1236 CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1237 obd->obd_name, es->s_mdsnum);
1239 sbi->s_mdsnum = es->s_mdsnum;
1245 extern int ext3_add_dir_entry(struct dentry *dentry);
1246 extern int ext3_del_dir_entry(struct dentry *dentry);
1248 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1249 struct dentry *parent,
1250 char *name, int namelen,
1252 unsigned long generation,
1256 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1257 struct dentry *dentry;
1260 LASSERT(namelen != 0);
1261 dentry = ll_lookup_one_len(name, parent, namelen);
1262 if (IS_ERR(dentry)) {
1263 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1264 dentry->d_name.name, dentry->d_inode->i_ino,
1265 (unsigned long) dentry->d_inode->i_generation,
1266 (int) PTR_ERR(dentry));
1267 RETURN(PTR_ERR(dentry));
1269 if (dentry->d_inode != NULL || dentry->d_flags & DCACHE_CROSS_REF) {
1270 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1271 dentry->d_name.name, dentry);
1276 /* mds_reint_rename() may use this method to add dir entry
1277 * that points onto local inode. and we don't want to find
1278 * it cross-ref by subsequent lookups */
1281 dentry->d_flags |= DCACHE_CROSS_REF;
1282 dentry->d_inum = ino;
1283 dentry->d_mdsnum = mds;
1284 dentry->d_generation = generation;
1285 dentry->d_fid = fid;
1287 err = ext3_add_dir_entry(dentry);
1294 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1299 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1300 struct dentry *dentry)
1302 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1305 err = ext3_del_dir_entry(dentry);
1311 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1316 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1317 objects. Logs is number of logfiles to update */
1318 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1322 case FSFILT_OP_CREATE:
1323 /* directory leaf, index & indirect & EA*/
1324 return 4 + 3 * logs;
1325 case FSFILT_OP_UNLINK:
1331 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1332 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1333 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1334 for (i = 0; i < op; i++, fso++) {
1335 int nblocks = fso->fso_bufcnt * blockpp;
1336 int ndindirect = min(nblocks, addrpp + 1);
1337 int nindir = nblocks + ndindirect + 1;
1341 return needed + 3 * logs;
1348 #define EXTENTS_EA "write_extents"
1349 #define EXTENTS_EA_SIZE 64
1351 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1352 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1353 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1354 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1356 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1362 rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1363 EXTENTS_EA, from, num);
1367 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1373 rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1374 EXTENTS_EA, from, num);
1378 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1379 const char *eaname, int size);
1381 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1385 rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1390 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1391 char **pbuf, int *size)
1395 rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1396 EXTENTS_EA, pbuf, size);
1400 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1404 rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1409 static struct fsfilt_operations fsfilt_ext3_ops = {
1411 .fs_owner = THIS_MODULE,
1412 .fs_start = fsfilt_ext3_start,
1413 .fs_brw_start = fsfilt_ext3_brw_start,
1414 .fs_commit = fsfilt_ext3_commit,
1415 .fs_commit_async = fsfilt_ext3_commit_async,
1416 .fs_commit_wait = fsfilt_ext3_commit_wait,
1417 .fs_setattr = fsfilt_ext3_setattr,
1418 .fs_iocontrol = fsfilt_ext3_iocontrol,
1419 .fs_set_md = fsfilt_ext3_set_md,
1420 .fs_get_md = fsfilt_ext3_get_md,
1421 .fs_readpage = fsfilt_ext3_readpage,
1422 .fs_add_journal_cb = fsfilt_ext3_add_journal_cb,
1423 .fs_statfs = fsfilt_ext3_statfs,
1424 .fs_sync = fsfilt_ext3_sync,
1425 .fs_map_inode_pages = fsfilt_ext3_map_inode_pages,
1426 .fs_prep_san_write = fsfilt_ext3_prep_san_write,
1427 .fs_write_record = fsfilt_ext3_write_record,
1428 .fs_read_record = fsfilt_ext3_read_record,
1429 .fs_setup = fsfilt_ext3_setup,
1430 .fs_getpage = fsfilt_ext3_getpage,
1431 .fs_send_bio = fsfilt_ext3_send_bio,
1432 .fs_set_xattr = fsfilt_ext3_set_xattr,
1433 .fs_get_xattr = fsfilt_ext3_get_xattr,
1434 .fs_get_op_len = fsfilt_ext3_get_op_len,
1435 .fs_add_dir_entry = fsfilt_ext3_add_dir_entry,
1436 .fs_del_dir_entry = fsfilt_ext3_del_dir_entry,
1437 .fs_init_extents_ea = fsfilt_ext3_init_extents_ea,
1438 .fs_insert_extents_ea = fsfilt_ext3_insert_extents_ea,
1439 .fs_remove_extents_ea = fsfilt_ext3_remove_extents_ea,
1440 .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1441 .fs_get_write_extents_num = fsfilt_ext3_get_write_extents_num,
1444 static int __init fsfilt_ext3_init(void)
1448 fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1449 sizeof(struct fsfilt_cb_data), 0,
1452 CERROR("error allocating fsfilt journal callback cache\n");
1453 GOTO(out, rc = -ENOMEM);
1456 rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1459 kmem_cache_destroy(fcb_cache);
1464 static void __exit fsfilt_ext3_exit(void)
1466 fsfilt_unregister_ops(&fsfilt_ext3_ops);
1467 LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1468 "can't free fsfilt callback cache: count %d\n",
1469 atomic_read(&fcb_cache_count));
1472 module_init(fsfilt_ext3_init);
1473 module_exit(fsfilt_ext3_exit);
1475 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1476 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1477 MODULE_LICENSE("GPL");