1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
8 * This code is issued under the GNU General Public License.
9 * See the file COPYING in this distribution
11 * by Peter Braam <braam@clusterfs.com>
12 * and Andreas Dilger <adilger@clusterfs.com>
16 #define DEBUG_SUBSYSTEM S_FILTER
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
20 #include <linux/lustre_dlm.h>
21 #include <linux/obd_filter.h>
22 #include <linux/ext3_jbd.h>
23 #include <linux/quotaops.h>
24 #include <linux/init.h>
28 #define FILTER_ROOTINO 2
31 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
33 [S_IFREG >> S_SHIFT] "R",
34 [S_IFDIR >> S_SHIFT] "D",
35 [S_IFCHR >> S_SHIFT] "C",
36 [S_IFBLK >> S_SHIFT] "B",
37 [S_IFIFO >> S_SHIFT] "F",
38 [S_IFSOCK >> S_SHIFT] "S",
39 [S_IFLNK >> S_SHIFT] "L"
42 static inline const char *obd_mode_to_type(int mode)
44 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
47 /* write the pathname into the string */
48 static int filter_id(char *buf, obd_id id, obd_mode mode)
50 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
53 static inline void f_dput(struct dentry *dentry)
55 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
56 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
57 LASSERT(atomic_read(&dentry->d_count) > 0);
61 /* setup the object store with correct subdirectories */
62 static int filter_prep(struct obd_device *obddev)
64 struct obd_run_ctxt saved;
65 struct filter_obd *filter = &obddev->u.filter;
66 struct dentry *dentry;
74 push_ctxt(&saved, &filter->fo_ctxt);
75 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
76 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
79 CERROR("cannot open/create O: rc = %d\n", rc);
82 filter->fo_dentry_O = dentry;
83 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
84 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
87 CERROR("cannot open/create P: rc = %d\n", rc);
90 CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
91 atomic_read(&dentry->d_count) - 1);
93 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
94 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
97 CERROR("cannot open/create D: rc = %d\n", rc);
100 CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
101 atomic_read(&dentry->d_count) - 1);
105 * Create directories and/or get dentries for each object type.
106 * This saves us from having to do multiple lookups for each one.
108 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
109 char *type = obd_type_by_mode[mode];
112 filter->fo_dentry_O_mode[mode] = NULL;
115 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
116 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
117 if (IS_ERR(dentry)) {
118 rc = PTR_ERR(dentry);
119 CERROR("cannot create O/%s: rc = %d\n", type, rc);
120 GOTO(out_O_mode, rc);
122 filter->fo_dentry_O_mode[mode] = dentry;
125 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
126 file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
129 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
131 GOTO(out_O_mode, rc);
135 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
136 if ( !file || IS_ERR(file) ) {
138 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
140 GOTO(out_O_mode, rc);
143 /* steal operations */
144 inode = file->f_dentry->d_inode;
145 filter->fo_fop = file->f_op;
146 filter->fo_iop = inode->i_op;
147 filter->fo_aops = inode->i_mapping->a_ops;
149 if (inode->i_size == 0) {
150 __u64 disk_lastino = cpu_to_le64(lastino);
151 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
152 sizeof(disk_lastino),
154 if (retval != sizeof(disk_lastino)) {
155 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
157 GOTO(out_O_mode, rc = -EIO);
161 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
162 sizeof(disk_lastino),
164 if (retval != sizeof(disk_lastino)) {
165 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
167 GOTO(out_O_mode, rc = -EIO);
169 lastino = le64_to_cpu(disk_lastino);
171 filter->fo_lastino = lastino;
182 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
185 filter->fo_dentry_O_mode[mode] = NULL;
189 f_dput(filter->fo_dentry_O);
190 filter->fo_dentry_O = NULL;
194 /* cleanup the filter: write last used object id to status file */
195 static void filter_post(struct obd_device *obddev)
197 struct obd_run_ctxt saved;
198 struct filter_obd *filter = &obddev->u.filter;
204 push_ctxt(&saved, &filter->fo_ctxt);
205 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
207 CERROR("OBD filter: cannot create status file\n");
212 disk_lastino = cpu_to_le64(filter->fo_lastino);
213 rc = file->f_op->write(file, (char *)&disk_lastino,
214 sizeof(disk_lastino), &file->f_pos);
215 if (rc != sizeof(disk_lastino))
216 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
218 rc = filp_close(file, NULL);
220 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
222 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
223 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
226 filter->fo_dentry_O_mode[mode] = NULL;
229 f_dput(filter->fo_dentry_O);
235 static __u64 filter_next_id(struct obd_device *obddev)
239 spin_lock(&obddev->u.filter.fo_lock);
240 id = ++obddev->u.filter.fo_lastino;
241 spin_unlock(&obddev->u.filter.fo_lock);
243 /* FIXME: write the lastino to disk here */
247 /* how to get files, dentries, inodes from object id's */
248 /* parent i_sem is already held if needed for exclusivity */
249 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
250 struct dentry *dparent,
251 __u64 id, __u32 type)
253 struct super_block *sb = obddev->u.filter.fo_sb;
254 struct dentry *dchild;
259 if (!sb || !sb->s_dev) {
260 CERROR("fatal: device not initialized.\n");
261 RETURN(ERR_PTR(-ENXIO));
265 CERROR("fatal: invalid object #0\n");
267 RETURN(ERR_PTR(-ESTALE));
270 if (!(type & S_IFMT)) {
271 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
272 __FUNCTION__, id, type);
273 RETURN(ERR_PTR(-EINVAL));
276 len = sprintf(name, LPU64, id);
277 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
279 dchild = lookup_one_len(name, dparent, len);
280 if (IS_ERR(dchild)) {
281 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
285 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
286 obd_mode_to_type(type), name, dchild,
287 atomic_read(&dchild->d_count));
289 LASSERT(atomic_read(&dchild->d_count) > 0);
294 static struct file *filter_obj_open(struct obd_device *obddev,
295 __u64 id, __u32 type)
297 struct super_block *sb = obddev->u.filter.fo_sb;
298 struct obd_run_ctxt saved;
303 if (!sb || !sb->s_dev) {
304 CERROR("fatal: device not initialized.\n");
305 RETURN(ERR_PTR(-ENXIO));
309 CERROR("fatal: invalid obdo "LPU64"\n", id);
310 RETURN(ERR_PTR(-ESTALE));
313 if (!(type & S_IFMT)) {
314 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
315 __FUNCTION__, id, type);
316 RETURN(ERR_PTR(-EINVAL));
319 filter_id(name, id, type);
320 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
321 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
324 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
331 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
333 struct filter_obd *filter = &obddev->u.filter;
335 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
340 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
346 rc = class_connect(conn, obd, cluuid);
352 static int filter_disconnect(struct lustre_handle *conn)
357 rc = class_disconnect(conn);
361 /* XXX cleanup preallocated inodes */
365 /* mount the file system (secretly) */
366 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
368 struct obd_ioctl_data* data = buf;
369 struct filter_obd *filter;
370 struct vfsmount *mnt;
374 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
378 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
383 filter = &obddev->u.filter;;
384 filter->fo_vfsmnt = mnt;
385 filter->fo_fstype = strdup(data->ioc_inlbuf2);
386 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
387 /* XXX is this even possible if do_kern_mount succeeded? */
389 GOTO(err_kfree, err = -ENODEV);
391 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
392 filter->fo_ctxt.pwdmnt = mnt;
393 filter->fo_ctxt.pwd = mnt->mnt_root;
394 filter->fo_ctxt.fs = get_ds();
396 err = filter_prep(obddev);
398 GOTO(err_kfree, err);
399 spin_lock_init(&filter->fo_lock);
401 obddev->obd_namespace =
402 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
403 if (obddev->obd_namespace == NULL)
406 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
407 "filter_ldlm_client", &obddev->obd_ldlm_client);
412 kfree(filter->fo_fstype);
414 mntput(filter->fo_vfsmnt);
424 static int filter_cleanup(struct obd_device * obddev)
426 struct super_block *sb;
429 if (!list_empty(&obddev->obd_exports)) {
430 CERROR("still has clients!\n");
431 class_disconnect_all(obddev);
432 if (!list_empty(&obddev->obd_exports)) {
433 CERROR("still has exports after forced cleanup?\n");
438 ldlm_namespace_free(obddev->obd_namespace);
440 sb = obddev->u.filter.fo_sb;
441 if (!obddev->u.filter.fo_sb)
446 shrink_dcache_parent(sb->s_root);
448 mntput(obddev->u.filter.fo_vfsmnt);
449 obddev->u.filter.fo_sb = 0;
450 kfree(obddev->u.filter.fo_fstype);
459 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
462 int type = oa->o_mode & S_IFMT;
465 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
466 inode->i_ino, inode, (long)oa->o_id, valid);
467 /* Don't copy the inode number in place of the object ID */
468 obdo_from_inode(oa, inode, valid);
469 oa->o_mode &= ~S_IFMT;
472 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
473 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
475 oa->o_valid |= OBD_MD_FLRDEV;
481 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
482 struct lov_stripe_md *md)
484 struct obd_device *obddev = class_conn2obd(conn);
485 struct dentry *dentry;
488 if (!class_conn2export(conn)) {
489 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
493 obddev = class_conn2obd(conn);
494 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
495 oa->o_id, oa->o_mode);
497 RETURN(PTR_ERR(dentry));
499 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
505 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
506 struct lov_stripe_md *md)
508 struct obd_run_ctxt saved;
509 struct obd_device *obd = class_conn2obd(conn);
510 struct dentry *dentry;
516 iattr_from_obdo(&iattr, oa, oa->o_valid);
517 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
518 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
519 oa->o_id, iattr.ia_mode);
521 RETURN(PTR_ERR(dentry));
523 inode = dentry->d_inode;
525 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
526 GOTO(out_setattr, rc = -ENOENT);
530 if (iattr.ia_valid & ATTR_SIZE)
532 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
533 if (inode->i_op->setattr)
534 rc = inode->i_op->setattr(dentry, &iattr);
536 rc = inode_setattr(inode, &iattr);
538 if (iattr.ia_valid & ATTR_SIZE) {
540 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
541 obdo_from_inode(oa, inode, oa->o_valid);
550 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
551 struct lov_stripe_md *ea)
553 struct obd_export *export;
554 struct obd_device *obd;
555 struct dentry *dentry;
558 export = class_conn2export(conn);
560 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
564 obd = export->exp_obd;
565 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
566 oa->o_id, oa->o_mode);
568 RETURN(PTR_ERR(dentry));
570 if (!dentry->d_inode) {
571 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
575 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
576 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
581 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
582 struct lov_stripe_md *ea)
584 struct obd_device *obd;
585 struct dentry *dentry;
588 obd = class_conn2obd(conn);
590 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
594 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
595 oa->o_id, oa->o_mode);
597 RETURN(PTR_ERR(dentry));
598 LASSERT(atomic_read(&dentry->d_count) > 1);
600 f_dput(dentry); /* for the close */
601 f_dput(dentry); /* for this call */
605 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
606 struct lov_stripe_md **ea)
609 struct obd_run_ctxt saved;
612 struct obd_device *obd = class_conn2obd(conn);
613 struct filter_obd *filter = &obd->u.filter;
618 CERROR("invalid client "LPX64"\n", conn->addr);
622 if (!(oa->o_mode && S_IFMT)) {
623 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
624 __FUNCTION__, oa->o_id, oa->o_mode);
628 oa->o_id = filter_next_id(obd);
630 //filter_id(name, oa->o_id, oa->o_mode);
631 sprintf(name, LPU64, oa->o_id);
632 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
633 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
634 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name, oa->o_mode);
637 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
641 /* Set flags for fields we have set in the inode struct */
642 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
643 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
644 filter_from_inode(oa, new->d_inode, oa->o_valid);
650 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
651 struct lov_stripe_md *ea)
653 struct obd_device *obd;
654 struct filter_obd *filter;
655 struct obd_run_ctxt saved;
657 struct dentry *dir_dentry, *object_dentry;
661 obd = class_conn2obd(conn);
663 CERROR("invalid client "LPX64"\n", conn->addr);
667 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
669 dir_dentry = filter_parent(obd, oa->o_mode);
670 down(&dir_dentry->d_inode->i_sem);
672 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
674 if (IS_ERR(object_dentry))
675 GOTO(out, rc = -ENOENT);
677 inode = object_dentry->d_inode;
679 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
680 GOTO(out, rc = -ENOENT);
683 if (inode->i_nlink != 1) {
684 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
688 inode->i_mode = S_IFREG;
690 filter = &obd->u.filter;
691 push_ctxt(&saved, &filter->fo_ctxt);
693 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
695 f_dput(object_dentry);
699 up(&dir_dentry->d_inode->i_sem);
703 /* NB count and offset are used for punch, but not truncate */
704 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
705 struct lov_stripe_md *md,
706 obd_off start, obd_off end)
711 if (end != OBD_PUNCH_EOF)
712 CERROR("PUNCH not supported, only truncate works\n");
714 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
715 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
717 error = filter_setattr(conn, oa, NULL);
721 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
722 struct lov_stripe_md *lsm, obd_count oa_bufs,
723 struct brw_page *pga, brw_callback_t callback,
724 struct io_cb_data *data)
726 struct obd_run_ctxt saved;
727 struct super_block *sb;
728 int pnum; /* index to pages (bufs) */
729 unsigned long retval;
732 struct obd_device *obd = class_conn2obd(conn);
737 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
741 sb = obd->u.filter.fo_sb;
742 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
743 pnum = 0; /* pnum indexes buf 0..num_pages */
745 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
747 GOTO(out, retval = PTR_ERR(file));
749 /* count doubles as retval */
750 for (pg = 0; pg < oa_bufs; pg++) {
751 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
752 ") off count ("LPU64",%d)\n",
753 cmd, pnum, file->f_dentry->d_inode->i_ino,
754 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
755 (int)pga[pnum].count);
756 if (cmd & OBD_BRW_WRITE) {
760 buffer = kmap(pga[pnum].pg);
761 retval = file->f_op->write(file, buffer,
764 kunmap(pga[pnum].pg);
765 CDEBUG(D_INODE, "retval %ld\n", retval);
767 loff_t off = pga[pnum].off;
768 char *buffer = kmap(pga[pnum].pg);
770 if (off >= file->f_dentry->d_inode->i_size) {
771 memset(buffer, 0, pga[pnum].count);
772 retval = pga[pnum].count;
774 retval = file->f_op->read(file, buffer,
775 pga[pnum].count, &off);
777 kunmap(pga[pnum].pg);
779 if (retval != pga[pnum].count) {
781 GOTO(out, retval = -EIO);
783 CDEBUG(D_INODE, "retval %ld\n", retval);
787 /* sizes and blocks are set by generic_file_write */
788 /* ctimes/mtimes will follow with a setattr call */
791 /* XXX: do something with callback if it is set? */
796 error = (retval >= 0) ? 0 : retval;
801 * Calculate the number of buffer credits needed to write multiple pages in
802 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
803 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
805 * See comment above ext3_writepage_trans_blocks for details. We assume
806 * no data journaling is being done, but it does allow for all of the pages
807 * being non-contiguous. If we are guaranteed contiguous pages we could
808 * reduce the number of (d)indirect blocks a lot.
810 * With N blocks per page and P pages, for each inode we have at most:
812 * min(N*P, blocksize/4 + 1) dindirect blocks
815 * For the entire filesystem, we have at most:
816 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
817 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
820 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
822 static int ext3_credits_needed(struct super_block *sb, int objcount,
823 struct obd_ioobj *obj)
825 struct obd_ioobj *o = obj;
826 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
827 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
830 int needed = objcount + 1;
833 for (i = 0; i < objcount; i++, o++) {
834 int nblocks = o->ioo_bufcnt * blockpp;
835 int ndindirect = min(nblocks, addrpp + 1);
836 int nindir = nblocks + ndindirect + 1;
838 nbitmaps += nindir + nblocks;
839 ngdblocks += nindir + nblocks;
844 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
845 nbitmaps = EXT3_SB(sb)->s_groups_count;
846 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
847 ngdblocks = EXT3_SB(sb)->s_gdb_count;
849 needed += nbitmaps + ngdblocks;
852 /* We assume that there will be 1 bit set in s_dquot.flags for each
853 * quota file that is active. This is at least true for now.
855 needed += hweight32(sb_any_quota_enabled(sb)) *
856 EXT3_SINGLEDATA_TRANS_BLOCKS;
862 /* We have to start a huge journal transaction here to hold all of the
863 * metadata for the pages being written here. This is necessitated by
864 * the fact that we do lots of prepare_write operations before we do
865 * any of the matching commit_write operations, so even if we split
866 * up to use "smaller" transactions none of them could complete until
867 * all of them were opened. By having a single journal transaction,
868 * we eliminate duplicate reservations for common blocks like the
869 * superblock and group descriptors or bitmaps.
871 * We will start the transaction here, but each prepare_write will
872 * add a refcount to the transaction, and each commit_write will
873 * remove a refcount. The transaction will be closed when all of
874 * the pages have been written.
876 static void *ext3_filter_journal_start(struct filter_obd *filter,
877 int objcount, struct obd_ioobj *obj,
878 int niocount, struct niobuf_remote *nb)
880 journal_t *journal = NULL;
881 handle_t *handle = NULL;
884 /* Assumes ext3 and extN have same sb_info layout, but avoids issues
885 * with having extN built properly before filterobd for now.
887 journal = EXT3_SB(filter->fo_sb)->s_journal;
888 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
890 /* The number of blocks we could _possibly_ dirty can very large.
891 * We reduce our request if it is absurd (and we couldn't get that
892 * many credits for a single handle anyways).
894 * At some point we have to limit the size of I/Os sent at one time,
895 * increase the size of the journal, or we have to calculate the
896 * actual journal requirements more carefully by checking all of
897 * the blocks instead of being maximally pessimistic. It remains to
898 * be seen if this is a real problem or not.
900 if (needed > journal->j_max_transaction_buffers) {
901 CERROR("want too many journal credits (%d) using %d instead\n",
902 needed, journal->j_max_transaction_buffers);
903 needed = journal->j_max_transaction_buffers;
906 handle = journal_start(journal, needed);
908 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
914 static void *filter_journal_start(void **journal_save,
915 struct filter_obd *filter,
916 int objcount, struct obd_ioobj *obj,
917 int niocount, struct niobuf_remote *nb)
921 /* This may not be necessary - we probably never have a
922 * transaction started when we enter here, so we can
923 * remove the saving of the journal state entirely.
924 * For now leave it in just to see if it ever happens.
926 *journal_save = current->journal_info;
928 CERROR("Already have handle %p???\n", *journal_save);
930 current->journal_info = NULL;
933 if (!strcmp(filter->fo_fstype, "ext3") ||
934 !strcmp(filter->fo_fstype, "extN"))
935 handle = ext3_filter_journal_start(filter, objcount, obj,
940 static int ext3_filter_journal_stop(void *handle)
944 /* We got a refcount on the handle for each call to prepare_write,
945 * so we can drop the "parent" handle here to avoid the need for
946 * osc to call back into filterobd to close the handle. The
947 * remaining references will be dropped in commit_write.
949 rc = journal_stop((handle_t *)handle);
954 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
959 if (!strcmp(filter->fo_fstype, "ext3") ||
960 !strcmp(filter->fo_fstype, "extN"))
961 rc = ext3_filter_journal_stop(handle);
964 CERROR("error on journal stop: rc = %d\n", rc);
966 current->journal_info = journal_save;
971 static inline void lustre_put_page(struct page *page)
974 page_cache_release(page);
978 lustre_get_page_read(struct inode *inode, unsigned long index)
980 struct address_space *mapping = inode->i_mapping;
984 page = read_cache_page(mapping, index,
985 (filler_t*)mapping->a_ops->readpage, NULL);
989 if (!Page_Uptodate(page)) {
990 CERROR("page index %lu not uptodate\n", index);
991 GOTO(err_page, rc = -EIO);
993 if (PageError(page)) {
994 CERROR("page index %lu has error\n", index);
995 GOTO(err_page, rc = -EIO);
1001 lustre_put_page(page);
1005 static struct page *
1006 lustre_get_page_write(struct inode *inode, unsigned long index)
1008 struct address_space *mapping = inode->i_mapping;
1012 page = grab_cache_page(mapping, index); /* locked page */
1014 if (!IS_ERR(page)) {
1016 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1017 * a no-op for most filesystems, because we write the whole
1018 * page. For partial-page I/O this will read in the page.
1020 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1022 CERROR("page index %lu, rc = %d\n", index, rc);
1025 GOTO(err_unlock, rc);
1027 /* XXX not sure if we need this if we are overwriting page */
1028 if (PageError(page)) {
1029 CERROR("error on page index %lu, rc = %d\n", index, rc);
1031 GOTO(err_unlock, rc = -EIO);
1038 lustre_put_page(page);
1042 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1044 struct inode *inode = page->mapping->host;
1047 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1048 if (!err && IS_SYNC(inode))
1049 err = waitfor_one_page(page);
1051 //SetPageUptodate(page); // the client commit_write will do this
1053 SetPageReferenced(page);
1055 lustre_put_page(page);
1059 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1060 struct niobuf_local *lnb, int *pglocked)
1062 struct address_space *mapping = inode->i_mapping;
1066 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1068 page = grab_cache_page_nowait(mapping, index); /* locked page */
1070 page = grab_cache_page(mapping, index); /* locked page */
1073 /* This page is currently locked, so get a temporary page instead. */
1074 /* XXX I believe this is a very dangerous thing to do - consider if
1075 * we had multiple writers for the same file (definitely the case
1076 * if we are using this codepath). If writer A locks the page,
1077 * writer B writes to a copy (as here), writer A drops the page
1078 * lock, and writer C grabs the lock before B does, then B will
1079 * later overwrite the data from C, even if C had LDLM locked
1080 * and initiated the write after B did.
1084 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1085 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1087 CERROR("no memory for a temp page\n");
1089 GOTO(err, rc = -ENOMEM);
1091 page = virt_to_page(addr);
1093 page->index = index;
1094 lnb->flags |= N_LOCAL_TEMP_PAGE;
1095 } else if (!IS_ERR(page)) {
1099 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1100 * a no-op for most filesystems, because we write the whole
1101 * page. For partial-page I/O this will read in the page.
1103 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1105 CERROR("page index %lu, rc = %d\n", index, rc);
1108 GOTO(err_unlock, rc);
1110 /* XXX not sure if we need this if we are overwriting page */
1111 if (PageError(page)) {
1112 CERROR("error on page index %lu, rc = %d\n", index, rc);
1114 GOTO(err_unlock, rc = -EIO);
1121 lustre_put_page(page);
1127 * We need to balance prepare_write() calls with commit_write() calls.
1128 * If the page has been prepared, but we have no data for it, we don't
1129 * want to overwrite valid data on disk, but we still need to zero out
1130 * data for space which was newly allocated. Like part of what happens
1131 * in __block_prepare_write() for newly allocated blocks.
1133 * XXX currently __block_prepare_write() creates buffers for all the
1134 * pages, and the filesystems mark these buffers as BH_New if they
1135 * were newly allocated from disk. We use the BH_New flag similarly.
1137 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1141 unsigned block_start, block_end;
1142 struct buffer_head *bh, *head = page->buffers;
1143 unsigned blocksize = head->b_size;
1144 void *addr = page_address(page);
1146 /* debugging: just seeing if this ever happens */
1147 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1148 page->index, page->mapping->host->i_ino, err);
1150 /* Currently one buffer per page, but in the future... */
1151 for (bh = head, block_start = 0; bh != head || !block_start;
1152 block_start = block_end, bh = bh->b_this_page) {
1153 block_end = block_start + blocksize;
1155 memset(addr + block_start, 0, blocksize);
1159 return lustre_commit_write(page, from, to);
1162 static int filter_preprw(int cmd, struct lustre_handle *conn,
1163 int objcount, struct obd_ioobj *obj,
1164 int niocount, struct niobuf_remote *nb,
1165 struct niobuf_local *res, void **desc_private)
1167 struct obd_run_ctxt saved;
1168 struct obd_device *obd;
1169 struct obd_ioobj *o = obj;
1170 struct niobuf_remote *b = nb;
1171 struct niobuf_local *r = res;
1172 void *journal_save = NULL;
1178 obd = class_conn2obd(conn);
1180 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1183 memset(res, 0, sizeof(*res) * niocount);
1185 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1187 if (cmd & OBD_BRW_WRITE) {
1188 *desc_private = filter_journal_start(&journal_save,
1190 objcount, obj, niocount,
1192 if (IS_ERR(*desc_private))
1193 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1196 for (i = 0; i < objcount; i++, o++) {
1197 struct dentry *dentry;
1198 struct inode *inode;
1201 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1202 o->ioo_id, S_IFREG);
1204 GOTO(out_clean, rc = PTR_ERR(dentry));
1205 inode = dentry->d_inode;
1207 CERROR("trying to BRW to non-existent file "LPU64"\n",
1210 GOTO(out_clean, rc = -ENOENT);
1213 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1214 unsigned long index = b->offset >> PAGE_SHIFT;
1220 r->dentry = dget(dentry);
1222 if (cmd & OBD_BRW_WRITE)
1223 page = filter_get_page_write(inode, index, r,
1226 page = lustre_get_page_read(inode, index);
1230 GOTO(out_clean, rc = PTR_ERR(page));
1233 r->addr = page_address(page);
1234 r->offset = b->offset;
1241 if (cmd & OBD_BRW_WRITE) {
1242 int err = filter_journal_stop(journal_save, &obd->u.filter,
1252 CERROR("error cleanup on brw\n");
1254 if (cmd & OBD_BRW_WRITE)
1255 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1257 lustre_put_page(r->page);
1262 static int filter_write_locked_page(struct niobuf_local *lnb)
1267 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1268 if (IS_ERR(lpage)) {
1269 /* It is highly unlikely that we would ever get an error here.
1270 * The page we want to get was previously locked, so it had to
1271 * have already allocated the space, and we were just writing
1272 * over the same data, so there would be no hole in the file.
1274 * XXX: possibility of a race with truncate could exist, need
1275 * to check that. There are no guarantees w.r.t.
1276 * write order even on a local filesystem, although the
1277 * normal response would be to return the number of bytes
1278 * successfully written and leave the rest to the app.
1280 rc = PTR_ERR(lpage);
1281 CERROR("error getting locked page index %ld: rc = %d\n",
1282 lnb->page->index, rc);
1286 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1287 * lustre_commit_write() below, lnb->page was kmapped previously in
1288 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1290 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1291 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1293 CERROR("error committing locked page %ld: rc = %d\n",
1294 lnb->page->index, rc);
1296 lustre_put_page(lnb->page);
1301 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1302 int objcount, struct obd_ioobj *obj,
1303 int niocount, struct niobuf_local *res,
1306 struct obd_run_ctxt saved;
1307 struct obd_ioobj *o;
1308 struct niobuf_local *r;
1309 struct obd_device *obd = class_conn2obd(conn);
1311 int found_locked = 0;
1316 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1317 journal_save = current->journal_info;
1318 LASSERT(!journal_save);
1320 current->journal_info = private;
1321 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1323 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1324 struct page *page = r->page;
1329 if (r->flags & N_LOCAL_TEMP_PAGE) {
1334 if (cmd & OBD_BRW_WRITE) {
1335 int err = filter_commit_write(page, 0,
1341 lustre_put_page(page);
1346 current->journal_info = journal_save;
1351 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1353 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1355 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1358 err = filter_write_locked_page(r);
1370 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1372 struct obd_device *obd = class_conn2obd(conn);
1377 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1379 statfs_pack(osfs, &sfs);
1384 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1385 void *key, obd_count *vallen, void **val)
1387 struct obd_device *obd;
1390 obd = class_conn2obd(conn);
1392 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1396 if ( keylen == strlen("blocksize") &&
1397 memcmp(key, "blocksize", keylen) == 0 ) {
1398 *vallen = sizeof(long);
1399 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1403 if ( keylen == strlen("blocksize_bits") &&
1404 memcmp(key, "blocksize_bits", keylen) == 0 ){
1405 *vallen = sizeof(long);
1406 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1410 if ( keylen == strlen("root_ino") &&
1411 memcmp(key, "root_ino", keylen) == 0 ){
1412 *vallen = sizeof(obd_id);
1413 *val = (void *)(obd_id)FILTER_ROOTINO;
1417 CDEBUG(D_IOCTL, "invalid key\n");
1421 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1422 struct lustre_handle *src_conn, struct obdo *src,
1423 obd_size count, obd_off offset)
1426 struct lov_stripe_md srcmd, dstmd;
1427 unsigned long index = 0;
1430 memset(&srcmd, 0, sizeof(srcmd));
1431 memset(&dstmd, 0, sizeof(dstmd));
1432 srcmd.lsm_object_id = src->o_id;
1433 dstmd.lsm_object_id = dst->o_id;
1436 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1437 ", dst: ino "LPU64"\n",
1438 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1439 page = alloc_page(GFP_USER);
1443 while (TryLockPage(page))
1444 ___wait_on_page(page);
1446 /* XXX with brw vector I/O, we could batch up reads and writes here,
1447 * all we need to do is allocate multiple pages to handle the I/Os
1448 * and arrays to handle the request parameters.
1450 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1452 struct io_cb_data *cbd = ll_init_cb();
1461 pg.count = PAGE_SIZE;
1462 pg.off = (page->index) << PAGE_SHIFT;
1465 page->index = index;
1466 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1467 ll_sync_io_cb, cbd);
1480 pg.flag = OBD_BRW_CREATE;
1481 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1483 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1484 ll_sync_io_cb, cbd);
1486 /* XXX should handle dst->o_size, dst->o_blocks here */
1492 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1496 dst->o_size = src->o_size;
1497 dst->o_blocks = src->o_blocks;
1498 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1506 static struct obd_ops filter_obd_ops = {
1507 o_get_info: filter_get_info,
1508 o_setup: filter_setup,
1509 o_cleanup: filter_cleanup,
1510 o_connect: filter_connect,
1511 o_disconnect: filter_disconnect,
1512 o_statfs: filter_statfs,
1513 o_getattr: filter_getattr,
1514 o_create: filter_create,
1515 o_setattr: filter_setattr,
1516 o_destroy: filter_destroy,
1517 o_open: filter_open,
1518 o_close: filter_close,
1519 o_brw: filter_pgcache_brw,
1520 o_punch: filter_truncate,
1521 o_preprw: filter_preprw,
1522 o_commitrw: filter_commitrw
1524 o_preallocate: filter_preallocate_inodes,
1525 o_migrate: filter_migrate,
1526 o_copy: filter_copy_data,
1527 o_iterate: filter_iterate
1532 static int __init obdfilter_init(void)
1534 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1535 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1538 static void __exit obdfilter_exit(void)
1540 class_unregister_type(OBD_FILTER_DEVICENAME);
1543 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1544 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1545 MODULE_LICENSE("GPL");
1547 module_init(obdfilter_init);
1548 module_exit(obdfilter_exit);