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
29 #include <linux/jbd.h>
30 #include <linux/slab.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/ext3_fs.h>
34 #include <linux/ext3_jbd.h>
35 #include <linux/ext3_xattr.h>
36 #include <linux/kp30.h>
37 #include <linux/lustre_fsfilt.h>
38 #include <linux/obd.h>
39 #include <linux/obd_class.h>
40 #include <linux/module.h>
42 static kmem_cache_t *fcb_cache;
43 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
45 struct fsfilt_cb_data {
46 struct journal_callback cb_jcb; /* data private to jbd */
47 fsfilt_cb_t cb_func; /* MDS/OBD completion function */
48 struct obd_device *cb_obd; /* MDS/OBD completion device */
49 __u64 cb_last_rcvd; /* MDS/OST last committed operation */
52 #define EXT3_XATTR_INDEX_LUSTRE 5
53 #define XATTR_LUSTRE_MDS_OBJID "system.lustre_mds_objid"
56 * We don't currently need any additional blocks for rmdir and
57 * unlink transactions because we are storing the OST oa_id inside
58 * the inode (which we will be changing anyways as part of this
61 static void *fsfilt_ext3_start(struct inode *inode, int op)
63 /* For updates to the last recieved file */
64 int nblocks = EXT3_DATA_TRANS_BLOCKS;
69 case FSFILT_OP_UNLINK:
70 nblocks += EXT3_DELETE_TRANS_BLOCKS;
72 case FSFILT_OP_RENAME:
73 /* modify additional directory */
74 nblocks += EXT3_DATA_TRANS_BLOCKS;
76 case FSFILT_OP_SYMLINK:
77 /* additional block + block bitmap + GDT for long symlink */
80 case FSFILT_OP_CREATE:
83 /* modify one inode + block bitmap + GDT */
87 /* modify parent directory */
88 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
90 case FSFILT_OP_SETATTR:
91 /* Setattr on inode */
94 default: CERROR("unknown transaction start op %d\n", op);
98 LASSERT(!current->journal_info);
100 handle = journal_start(EXT3_JOURNAL(inode), nblocks);
107 * Calculate the number of buffer credits needed to write multiple pages in
108 * a single ext3 transaction. No, this shouldn't be here, but as yet ext3
109 * doesn't have a nice API for calculating this sort of thing in advance.
111 * See comment above ext3_writepage_trans_blocks for details. We assume
112 * no data journaling is being done, but it does allow for all of the pages
113 * being non-contiguous. If we are guaranteed contiguous pages we could
114 * reduce the number of (d)indirect blocks a lot.
116 * With N blocks per page and P pages, for each inode we have at most:
118 * min(N*P, blocksize/4 + 1) dindirect blocks
121 * For the entire filesystem, we have at most:
122 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
123 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
124 * objcount inode blocks
126 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
128 * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
130 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso)
132 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
133 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
134 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
137 int needed = objcount + 1;
140 for (i = 0; i < objcount; i++, fso++) {
141 int nblocks = fso->fso_bufcnt * blockpp;
142 int ndindirect = min(nblocks, addrpp + 1);
143 int nindir = nblocks + ndindirect + 1;
145 nbitmaps += nindir + nblocks;
146 ngdblocks += nindir + nblocks;
151 /* Assumes ext3 and ext3 have same sb_info layout at the start. */
152 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
153 nbitmaps = EXT3_SB(sb)->s_groups_count;
154 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
155 ngdblocks = EXT3_SB(sb)->s_gdb_count;
157 needed += nbitmaps + ngdblocks;
159 /* last_rcvd update */
160 needed += EXT3_DATA_TRANS_BLOCKS;
163 /* We assume that there will be 1 bit set in s_dquot.flags for each
164 * quota file that is active. This is at least true for now.
166 needed += hweight32(sb_any_quota_enabled(sb)) *
167 EXT3_SINGLEDATA_TRANS_BLOCKS;
173 /* We have to start a huge journal transaction here to hold all of the
174 * metadata for the pages being written here. This is necessitated by
175 * the fact that we do lots of prepare_write operations before we do
176 * any of the matching commit_write operations, so even if we split
177 * up to use "smaller" transactions none of them could complete until
178 * all of them were opened. By having a single journal transaction,
179 * we eliminate duplicate reservations for common blocks like the
180 * superblock and group descriptors or bitmaps.
182 * We will start the transaction here, but each prepare_write will
183 * add a refcount to the transaction, and each commit_write will
184 * remove a refcount. The transaction will be closed when all of
185 * the pages have been written.
187 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
188 int niocount, struct niobuf_remote *nb)
195 LASSERT(!current->journal_info);
196 journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
197 needed = fsfilt_ext3_credits_needed(objcount, fso);
199 /* The number of blocks we could _possibly_ dirty can very large.
200 * We reduce our request if it is absurd (and we couldn't get that
201 * many credits for a single handle anyways).
203 * At some point we have to limit the size of I/Os sent at one time,
204 * increase the size of the journal, or we have to calculate the
205 * actual journal requirements more carefully by checking all of
206 * the blocks instead of being maximally pessimistic. It remains to
207 * be seen if this is a real problem or not.
209 if (needed > journal->j_max_transaction_buffers) {
210 CERROR("want too many journal credits (%d) using %d instead\n",
211 needed, journal->j_max_transaction_buffers);
212 needed = journal->j_max_transaction_buffers;
216 handle = journal_start(journal, needed);
219 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
225 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
228 handle_t *handle = h;
231 handle->h_sync = 1; /* recovery likes this */
234 rc = journal_stop(handle);
240 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
243 struct inode *inode = dentry->d_inode;
248 /* A _really_ horrible hack to avoid removing the data stored
249 * in the block pointers; this is really the "small" stripe MD data.
250 * We can avoid further hackery by virtue of the MDS file size being
251 * zero all the time (which doesn't invoke block truncate at unlink
252 * time), so we assert we never change the MDS file size from zero.
254 if (iattr->ia_valid & ATTR_SIZE) {
255 CERROR("hmm, setting %*s file size to %lld\n",
256 dentry->d_name.len, dentry->d_name.name, iattr->ia_size);
257 LASSERT(iattr->ia_size == 0);
259 /* ATTR_SIZE would invoke truncate: clear it */
260 iattr->ia_valid &= ~ATTR_SIZE;
261 inode->i_size = iattr->ia_size;
263 /* make sure _something_ gets set - so new inode
264 * goes to disk (probably won't work over XFS
266 if (!iattr->ia_valid & ATTR_MODE) {
267 iattr->ia_valid |= ATTR_MODE;
268 iattr->ia_mode = inode->i_mode;
272 if (inode->i_op->setattr)
273 rc = inode->i_op->setattr(dentry, iattr);
275 rc = inode_change_ok(inode, iattr);
277 rc = inode_setattr(inode, iattr);
285 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
286 void *lmm, int lmm_size)
290 /* Nasty hack city - store stripe MD data in the block pointers if
291 * it will fit, because putting it in an EA currently kills the MDS
292 * performance. We'll fix this with "fast EAs" in the future.
294 if (lmm_size <= sizeof(EXT3_I(inode)->i_data) -
295 sizeof(EXT3_I(inode)->i_data[0])) {
296 /* XXX old_size is debugging only */
297 int old_size = EXT3_I(inode)->i_data[0];
299 LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
300 CERROR("setting EA on %lu again... interesting\n",
304 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
305 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
306 mark_inode_dirty(inode);
311 rc = ext3_xattr_set(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
312 XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
318 CERROR("error adding MD data to inode %lu: rc = %d\n",
323 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
327 if (EXT3_I(inode)->i_data[0]) {
328 int size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
329 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
333 memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
340 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
341 XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
345 /* This gives us the MD size */
347 return (rc == -ENODATA) ? 0 : rc;
350 CDEBUG(D_INFO, "error getting EA %s from inode %lu: "
351 "rc = %d\n", XATTR_LUSTRE_MDS_OBJID, inode->i_ino, rc);
352 memset(lmm, 0, lmm_size);
353 return (rc == -ENODATA) ? 0 : rc;
359 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
362 struct inode *inode = file->f_dentry->d_inode;
365 if (S_ISREG(inode->i_mode))
366 rc = file->f_op->read(file, buf, count, off);
368 const int blkbits = inode->i_sb->s_blocksize_bits;
369 const int blksize = inode->i_sb->s_blocksize;
371 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
372 count, inode->i_ino, *off);
374 struct buffer_head *bh;
377 if (*off < inode->i_size) {
380 bh = ext3_bread(NULL, inode, *off >> blkbits,
383 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
386 memcpy(buf, bh->b_data, blksize);
389 /* XXX in theory we should just fake
390 * this buffer and continue like ext3,
391 * especially if this is a partial read
393 CERROR("error read dir %lu+%llu: %d\n",
394 inode->i_ino, *off, err);
399 struct ext3_dir_entry_2 *fake = (void *)buf;
401 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
402 memset(fake, 0, sizeof(*fake));
403 fake->rec_len = cpu_to_le32(blksize);
415 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
417 struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
419 fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, error);
421 OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
422 atomic_dec(&fcb_cache_count);
425 static int fsfilt_ext3_set_last_rcvd(struct obd_device *obd, __u64 last_rcvd,
426 void *handle, fsfilt_cb_t cb_func)
428 struct fsfilt_cb_data *fcb;
430 OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
434 atomic_inc(&fcb_cache_count);
435 fcb->cb_func = cb_func;
437 fcb->cb_last_rcvd = last_rcvd;
439 CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
441 /* Note that an "incompatible pointer" warning here is OK for now */
442 journal_callback_set(handle, fsfilt_ext3_cb_func,
443 (struct journal_callback *)fcb);
449 static int fsfilt_ext3_journal_data(struct file *filp)
451 struct inode *inode = filp->f_dentry->d_inode;
453 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
459 * We need to hack the return value for the free inode counts because
460 * the current EA code requires one filesystem block per inode with EAs,
461 * so it is possible to run out of blocks before we run out of inodes.
463 * This can be removed when the ext3 EA code is fixed.
465 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
468 int rc = vfs_statfs(sb, &sfs);
470 if (!rc && sfs.f_bfree < sfs.f_ffree) {
471 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
472 sfs.f_ffree = sfs.f_bfree;
475 statfs_pack(osfs, &sfs);
479 static int fsfilt_ext3_sync(struct super_block *sb)
481 return ext3_force_commit(sb);
484 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
485 int nblocks, loff_t newsize);
486 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
487 int nblocks, loff_t newsize)
489 return ext3_prep_san_write(inode, blocks, nblocks, newsize);
492 static struct fsfilt_operations fsfilt_ext3_ops = {
494 fs_owner: THIS_MODULE,
495 fs_start: fsfilt_ext3_start,
496 fs_brw_start: fsfilt_ext3_brw_start,
497 fs_commit: fsfilt_ext3_commit,
498 fs_setattr: fsfilt_ext3_setattr,
499 fs_set_md: fsfilt_ext3_set_md,
500 fs_get_md: fsfilt_ext3_get_md,
501 fs_readpage: fsfilt_ext3_readpage,
502 fs_journal_data: fsfilt_ext3_journal_data,
503 fs_set_last_rcvd: fsfilt_ext3_set_last_rcvd,
504 fs_statfs: fsfilt_ext3_statfs,
505 fs_sync: fsfilt_ext3_sync,
506 fs_prep_san_write: fsfilt_ext3_prep_san_write,
509 static int __init fsfilt_ext3_init(void)
513 //rc = ext3_xattr_register();
514 fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
515 sizeof(struct fsfilt_cb_data), 0,
518 CERROR("error allocating fsfilt journal callback cache\n");
519 GOTO(out, rc = -ENOMEM);
522 rc = fsfilt_register_ops(&fsfilt_ext3_ops);
525 kmem_cache_destroy(fcb_cache);
530 static void __exit fsfilt_ext3_exit(void)
534 fsfilt_unregister_ops(&fsfilt_ext3_ops);
535 rc = kmem_cache_destroy(fcb_cache);
537 if (rc || atomic_read(&fcb_cache_count)) {
538 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
539 atomic_read(&fcb_cache_count), rc);
542 //rc = ext3_xattr_unregister();
545 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
546 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
547 MODULE_LICENSE("GPL");
549 module_init(fsfilt_ext3_init);
550 module_exit(fsfilt_ext3_exit);