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/lustre_dlm.h>
20 #include <linux/obd_filter.h>
21 #include <linux/ext3_jbd.h>
22 #include <linux/quotaops.h>
23 #include <linux/init.h>
27 #define FILTER_ROOTINO 2
30 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
32 [S_IFREG >> S_SHIFT] "R",
33 [S_IFDIR >> S_SHIFT] "D",
34 [S_IFCHR >> S_SHIFT] "C",
35 [S_IFBLK >> S_SHIFT] "B",
36 [S_IFIFO >> S_SHIFT] "F",
37 [S_IFSOCK >> S_SHIFT] "S",
38 [S_IFLNK >> S_SHIFT] "L"
41 static inline const char *obd_mode_to_type(int mode)
43 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
46 /* write the pathname into the string */
47 static int filter_id(char *buf, obd_id id, obd_mode mode)
49 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
52 static inline void f_dput(struct dentry *dentry)
54 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
55 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
56 LASSERT(atomic_read(&dentry->d_count) > 0);
60 /* setup the object store with correct subdirectories */
61 static int filter_prep(struct obd_device *obddev)
63 struct obd_run_ctxt saved;
64 struct filter_obd *filter = &obddev->u.filter;
65 struct dentry *dentry;
73 push_ctxt(&saved, &filter->fo_ctxt);
74 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
75 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
78 CERROR("cannot open/create O: rc = %d\n", rc);
81 filter->fo_dentry_O = dentry;
82 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
83 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
86 CERROR("cannot open/create P: rc = %d\n", rc);
89 CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
90 atomic_read(&dentry->d_count) - 1);
92 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
93 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
96 CERROR("cannot open/create D: rc = %d\n", rc);
99 CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
100 atomic_read(&dentry->d_count) - 1);
104 * Create directories and/or get dentries for each object type.
105 * This saves us from having to do multiple lookups for each one.
107 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
108 char *type = obd_type_by_mode[mode];
111 filter->fo_dentry_O_mode[mode] = NULL;
114 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
115 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
116 if (IS_ERR(dentry)) {
117 rc = PTR_ERR(dentry);
118 CERROR("cannot create O/%s: rc = %d\n", type, rc);
119 GOTO(out_O_mode, rc);
121 filter->fo_dentry_O_mode[mode] = dentry;
124 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
125 file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
128 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
130 GOTO(out_O_mode, rc);
134 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
135 if ( !file || IS_ERR(file) ) {
137 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
139 GOTO(out_O_mode, rc);
142 /* steal operations */
143 inode = file->f_dentry->d_inode;
144 filter->fo_fop = file->f_op;
145 filter->fo_iop = inode->i_op;
146 filter->fo_aops = inode->i_mapping->a_ops;
148 if (inode->i_size == 0) {
149 __u64 disk_lastino = cpu_to_le64(lastino);
150 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
151 sizeof(disk_lastino),
153 if (retval != sizeof(disk_lastino)) {
154 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
156 GOTO(out_O_mode, rc = -EIO);
160 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
161 sizeof(disk_lastino),
163 if (retval != sizeof(disk_lastino)) {
164 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
166 GOTO(out_O_mode, rc = -EIO);
168 lastino = le64_to_cpu(disk_lastino);
170 filter->fo_lastino = lastino;
181 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
184 filter->fo_dentry_O_mode[mode] = NULL;
188 f_dput(filter->fo_dentry_O);
189 filter->fo_dentry_O = NULL;
193 /* cleanup the filter: write last used object id to status file */
194 static void filter_post(struct obd_device *obddev)
196 struct obd_run_ctxt saved;
197 struct filter_obd *filter = &obddev->u.filter;
203 push_ctxt(&saved, &filter->fo_ctxt);
204 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
206 CERROR("OBD filter: cannot create status file\n");
211 disk_lastino = cpu_to_le64(filter->fo_lastino);
212 rc = file->f_op->write(file, (char *)&disk_lastino,
213 sizeof(disk_lastino), &file->f_pos);
214 if (rc != sizeof(disk_lastino))
215 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
217 rc = filp_close(file, NULL);
219 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
221 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
222 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
225 filter->fo_dentry_O_mode[mode] = NULL;
228 f_dput(filter->fo_dentry_O);
234 static __u64 filter_next_id(struct obd_device *obddev)
238 spin_lock(&obddev->u.filter.fo_lock);
239 id = ++obddev->u.filter.fo_lastino;
240 spin_unlock(&obddev->u.filter.fo_lock);
242 /* FIXME: write the lastino to disk here */
246 /* how to get files, dentries, inodes from object id's */
247 /* parent i_sem is already held if needed for exclusivity */
248 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
249 struct dentry *dparent,
250 __u64 id, __u32 type)
252 struct super_block *sb = obddev->u.filter.fo_sb;
253 struct dentry *dchild;
258 if (!sb || !sb->s_dev) {
259 CERROR("fatal: device not initialized.\n");
260 RETURN(ERR_PTR(-ENXIO));
264 CERROR("fatal: invalid object #0\n");
266 RETURN(ERR_PTR(-ESTALE));
269 if (!(type & S_IFMT)) {
270 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
271 __FUNCTION__, id, type);
272 RETURN(ERR_PTR(-EINVAL));
275 len = sprintf(name, LPU64, id);
276 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
278 dchild = lookup_one_len(name, dparent, len);
279 if (IS_ERR(dchild)) {
280 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
284 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
285 obd_mode_to_type(type), name, dchild,
286 atomic_read(&dchild->d_count));
288 LASSERT(atomic_read(&dchild->d_count) > 0);
293 static struct file *filter_obj_open(struct obd_device *obddev,
294 __u64 id, __u32 type)
296 struct super_block *sb = obddev->u.filter.fo_sb;
297 struct obd_run_ctxt saved;
302 if (!sb || !sb->s_dev) {
303 CERROR("fatal: device not initialized.\n");
304 RETURN(ERR_PTR(-ENXIO));
308 CERROR("fatal: invalid obdo "LPU64"\n", id);
309 RETURN(ERR_PTR(-ESTALE));
312 if (!(type & S_IFMT)) {
313 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
314 __FUNCTION__, id, type);
315 RETURN(ERR_PTR(-EINVAL));
318 filter_id(name, id, type);
319 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
320 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
323 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
330 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
332 struct filter_obd *filter = &obddev->u.filter;
334 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
339 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
345 rc = class_connect(conn, obd, cluuid);
351 static int filter_disconnect(struct lustre_handle *conn)
356 rc = class_disconnect(conn);
360 /* XXX cleanup preallocated inodes */
364 /* mount the file system (secretly) */
365 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
367 struct obd_ioctl_data* data = buf;
368 struct filter_obd *filter;
369 struct vfsmount *mnt;
373 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
377 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
382 filter = &obddev->u.filter;;
383 filter->fo_vfsmnt = mnt;
384 filter->fo_fstype = strdup(data->ioc_inlbuf2);
385 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
386 /* XXX is this even possible if do_kern_mount succeeded? */
388 GOTO(err_kfree, err = -ENODEV);
390 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
391 filter->fo_ctxt.pwdmnt = mnt;
392 filter->fo_ctxt.pwd = mnt->mnt_root;
393 filter->fo_ctxt.fs = get_ds();
395 err = filter_prep(obddev);
397 GOTO(err_kfree, err);
398 spin_lock_init(&filter->fo_lock);
400 obddev->obd_namespace =
401 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
402 if (obddev->obd_namespace == NULL)
405 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
406 "filter_ldlm_client", &obddev->obd_ldlm_client);
411 kfree(filter->fo_fstype);
413 mntput(filter->fo_vfsmnt);
423 static int filter_cleanup(struct obd_device * obddev)
425 struct super_block *sb;
428 if (!list_empty(&obddev->obd_exports)) {
429 CERROR("still has clients!\n");
430 class_disconnect_all(obddev);
431 if (!list_empty(&obddev->obd_exports)) {
432 CERROR("still has exports after forced cleanup?\n");
437 ldlm_namespace_free(obddev->obd_namespace);
439 sb = obddev->u.filter.fo_sb;
440 if (!obddev->u.filter.fo_sb)
445 shrink_dcache_parent(sb->s_root);
447 mntput(obddev->u.filter.fo_vfsmnt);
448 obddev->u.filter.fo_sb = 0;
449 kfree(obddev->u.filter.fo_fstype);
458 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
461 int type = oa->o_mode & S_IFMT;
464 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
465 inode->i_ino, inode, (long)oa->o_id, valid);
466 /* Don't copy the inode number in place of the object ID */
467 obdo_from_inode(oa, inode, valid);
468 oa->o_mode &= ~S_IFMT;
471 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
472 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
474 oa->o_valid |= OBD_MD_FLRDEV;
480 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
481 struct lov_stripe_md *md)
483 struct obd_device *obddev = class_conn2obd(conn);
484 struct dentry *dentry;
487 if (!class_conn2export(conn)) {
488 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
492 obddev = class_conn2obd(conn);
493 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
494 oa->o_id, oa->o_mode);
496 RETURN(PTR_ERR(dentry));
498 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
504 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
505 struct lov_stripe_md *md)
507 struct obd_run_ctxt saved;
508 struct obd_device *obd = class_conn2obd(conn);
509 struct dentry *dentry;
515 iattr_from_obdo(&iattr, oa, oa->o_valid);
516 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
517 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
518 oa->o_id, iattr.ia_mode);
520 RETURN(PTR_ERR(dentry));
522 inode = dentry->d_inode;
524 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
525 GOTO(out_setattr, rc = -ENOENT);
529 if (iattr.ia_valid & ATTR_SIZE)
531 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
532 if (inode->i_op->setattr)
533 rc = inode->i_op->setattr(dentry, &iattr);
535 rc = inode_setattr(inode, &iattr);
537 if (iattr.ia_valid & ATTR_SIZE) {
539 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
540 obdo_from_inode(oa, inode, oa->o_valid);
549 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
550 struct lov_stripe_md *ea)
552 struct obd_export *export;
553 struct obd_device *obd;
554 struct dentry *dentry;
557 export = class_conn2export(conn);
559 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
563 obd = export->exp_obd;
564 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
565 oa->o_id, oa->o_mode);
567 RETURN(PTR_ERR(dentry));
569 if (!dentry->d_inode) {
570 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
574 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
575 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
580 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
581 struct lov_stripe_md *ea)
583 struct obd_device *obd;
584 struct dentry *dentry;
587 obd = class_conn2obd(conn);
589 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
593 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
594 oa->o_id, oa->o_mode);
596 RETURN(PTR_ERR(dentry));
597 LASSERT(atomic_read(&dentry->d_count) > 1);
599 f_dput(dentry); /* for the close */
600 f_dput(dentry); /* for this call */
604 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
605 struct lov_stripe_md **ea)
608 struct obd_run_ctxt saved;
611 struct obd_device *obd = class_conn2obd(conn);
612 struct filter_obd *filter = &obd->u.filter;
617 CERROR("invalid client "LPX64"\n", conn->addr);
621 if (!(oa->o_mode && S_IFMT)) {
622 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
623 __FUNCTION__, oa->o_id, oa->o_mode);
627 oa->o_id = filter_next_id(obd);
629 //filter_id(name, oa->o_id, oa->o_mode);
630 sprintf(name, LPU64, oa->o_id);
631 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
632 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
633 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name, oa->o_mode);
636 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
640 /* Set flags for fields we have set in the inode struct */
641 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
642 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
643 filter_from_inode(oa, new->d_inode, oa->o_valid);
649 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
650 struct lov_stripe_md *ea)
652 struct obd_device *obd;
653 struct filter_obd *filter;
654 struct obd_run_ctxt saved;
656 struct dentry *dir_dentry, *object_dentry;
660 obd = class_conn2obd(conn);
662 CERROR("invalid client "LPX64"\n", conn->addr);
666 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
668 dir_dentry = filter_parent(obd, oa->o_mode);
669 down(&dir_dentry->d_inode->i_sem);
671 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
673 if (IS_ERR(object_dentry))
674 GOTO(out, rc = -ENOENT);
676 inode = object_dentry->d_inode;
678 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
679 GOTO(out, rc = -ENOENT);
682 if (inode->i_nlink != 1) {
683 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
687 inode->i_mode = S_IFREG;
689 filter = &obd->u.filter;
690 push_ctxt(&saved, &filter->fo_ctxt);
692 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
694 f_dput(object_dentry);
698 up(&dir_dentry->d_inode->i_sem);
702 /* NB count and offset are used for punch, but not truncate */
703 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
704 struct lov_stripe_md *md,
705 obd_off start, obd_off end)
710 if (end != OBD_PUNCH_EOF)
711 CERROR("PUNCH not supported, only truncate works\n");
713 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
714 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
716 error = filter_setattr(conn, oa, NULL);
720 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
721 struct lov_stripe_md *lsm, obd_count oa_bufs,
722 struct brw_page *pga, brw_callback_t callback,
723 struct io_cb_data *data)
725 struct obd_run_ctxt saved;
726 struct super_block *sb;
727 int pnum; /* index to pages (bufs) */
728 unsigned long retval;
731 struct obd_device *obd = class_conn2obd(conn);
736 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
740 sb = obd->u.filter.fo_sb;
741 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
742 pnum = 0; /* pnum indexes buf 0..num_pages */
744 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
746 GOTO(out, retval = PTR_ERR(file));
748 /* count doubles as retval */
749 for (pg = 0; pg < oa_bufs; pg++) {
750 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
751 ") off count ("LPU64",%d)\n",
752 cmd, pnum, file->f_dentry->d_inode->i_ino,
753 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
754 (int)pga[pnum].count);
755 if (cmd & OBD_BRW_WRITE) {
759 buffer = kmap(pga[pnum].pg);
760 retval = file->f_op->write(file, buffer,
763 kunmap(pga[pnum].pg);
764 CDEBUG(D_INODE, "retval %ld\n", retval);
766 loff_t off = pga[pnum].off;
767 char *buffer = kmap(pga[pnum].pg);
769 if (off >= file->f_dentry->d_inode->i_size) {
770 memset(buffer, 0, pga[pnum].count);
771 retval = pga[pnum].count;
773 retval = file->f_op->read(file, buffer,
774 pga[pnum].count, &off);
776 kunmap(pga[pnum].pg);
778 if (retval != pga[pnum].count) {
780 GOTO(out, retval = -EIO);
782 CDEBUG(D_INODE, "retval %ld\n", retval);
786 /* sizes and blocks are set by generic_file_write */
787 /* ctimes/mtimes will follow with a setattr call */
790 /* XXX: do something with callback if it is set? */
795 error = (retval >= 0) ? 0 : retval;
800 * Calculate the number of buffer credits needed to write multiple pages in
801 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
802 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
804 * See comment above ext3_writepage_trans_blocks for details. We assume
805 * no data journaling is being done, but it does allow for all of the pages
806 * being non-contiguous. If we are guaranteed contiguous pages we could
807 * reduce the number of (d)indirect blocks a lot.
809 * With N blocks per page and P pages, for each inode we have at most:
811 * min(N*P, blocksize/4 + 1) dindirect blocks
814 * For the entire filesystem, we have at most:
815 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
816 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
819 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
821 static int ext3_credits_needed(struct super_block *sb, int objcount,
822 struct obd_ioobj *obj)
824 struct obd_ioobj *o = obj;
825 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
826 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
829 int needed = objcount + 1;
832 for (i = 0; i < objcount; i++, o++) {
833 int nblocks = o->ioo_bufcnt * blockpp;
834 int ndindirect = min(nblocks, addrpp + 1);
835 int nindir = nblocks + ndindirect + 1;
837 nbitmaps += nindir + nblocks;
838 ngdblocks += nindir + nblocks;
843 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
844 nbitmaps = EXT3_SB(sb)->s_groups_count;
845 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
846 ngdblocks = EXT3_SB(sb)->s_gdb_count;
848 needed += nbitmaps + ngdblocks;
851 /* We assume that there will be 1 bit set in s_dquot.flags for each
852 * quota file that is active. This is at least true for now.
854 needed += hweight32(sb_any_quota_enabled(sb)) *
855 EXT3_SINGLEDATA_TRANS_BLOCKS;
861 /* We have to start a huge journal transaction here to hold all of the
862 * metadata for the pages being written here. This is necessitated by
863 * the fact that we do lots of prepare_write operations before we do
864 * any of the matching commit_write operations, so even if we split
865 * up to use "smaller" transactions none of them could complete until
866 * all of them were opened. By having a single journal transaction,
867 * we eliminate duplicate reservations for common blocks like the
868 * superblock and group descriptors or bitmaps.
870 * We will start the transaction here, but each prepare_write will
871 * add a refcount to the transaction, and each commit_write will
872 * remove a refcount. The transaction will be closed when all of
873 * the pages have been written.
875 static void *ext3_filter_journal_start(struct filter_obd *filter,
876 int objcount, struct obd_ioobj *obj,
877 int niocount, struct niobuf_remote *nb)
879 journal_t *journal = NULL;
880 handle_t *handle = NULL;
883 /* Assumes ext3 and extN have same sb_info layout, but avoids issues
884 * with having extN built properly before filterobd for now.
886 journal = EXT3_SB(filter->fo_sb)->s_journal;
887 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
889 /* The number of blocks we could _possibly_ dirty can very large.
890 * We reduce our request if it is absurd (and we couldn't get that
891 * many credits for a single handle anyways).
893 * At some point we have to limit the size of I/Os sent at one time,
894 * increase the size of the journal, or we have to calculate the
895 * actual journal requirements more carefully by checking all of
896 * the blocks instead of being maximally pessimistic. It remains to
897 * be seen if this is a real problem or not.
899 if (needed > journal->j_max_transaction_buffers) {
900 CERROR("want too many journal credits (%d) using %d instead\n",
901 needed, journal->j_max_transaction_buffers);
902 needed = journal->j_max_transaction_buffers;
905 handle = journal_start(journal, needed);
907 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
913 static void *filter_journal_start(void **journal_save,
914 struct filter_obd *filter,
915 int objcount, struct obd_ioobj *obj,
916 int niocount, struct niobuf_remote *nb)
920 /* This may not be necessary - we probably never have a
921 * transaction started when we enter here, so we can
922 * remove the saving of the journal state entirely.
923 * For now leave it in just to see if it ever happens.
925 *journal_save = current->journal_info;
927 CERROR("Already have handle %p???\n", *journal_save);
929 current->journal_info = NULL;
932 if (!strcmp(filter->fo_fstype, "ext3") ||
933 !strcmp(filter->fo_fstype, "extN"))
934 handle = ext3_filter_journal_start(filter, objcount, obj,
939 static int ext3_filter_journal_stop(void *handle)
943 /* We got a refcount on the handle for each call to prepare_write,
944 * so we can drop the "parent" handle here to avoid the need for
945 * osc to call back into filterobd to close the handle. The
946 * remaining references will be dropped in commit_write.
948 rc = journal_stop((handle_t *)handle);
953 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
958 if (!strcmp(filter->fo_fstype, "ext3") ||
959 !strcmp(filter->fo_fstype, "extN"))
960 rc = ext3_filter_journal_stop(handle);
963 CERROR("error on journal stop: rc = %d\n", rc);
965 current->journal_info = journal_save;
970 static inline void lustre_put_page(struct page *page)
973 page_cache_release(page);
977 lustre_get_page_read(struct inode *inode, unsigned long index)
979 struct address_space *mapping = inode->i_mapping;
983 page = read_cache_page(mapping, index,
984 (filler_t*)mapping->a_ops->readpage, NULL);
988 if (!Page_Uptodate(page)) {
989 CERROR("page index %lu not uptodate\n", index);
990 GOTO(err_page, rc = -EIO);
992 if (PageError(page)) {
993 CERROR("page index %lu has error\n", index);
994 GOTO(err_page, rc = -EIO);
1000 lustre_put_page(page);
1004 static struct page *
1005 lustre_get_page_write(struct inode *inode, unsigned long index)
1007 struct address_space *mapping = inode->i_mapping;
1011 page = grab_cache_page(mapping, index); /* locked page */
1013 if (!IS_ERR(page)) {
1015 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1016 * a no-op for most filesystems, because we write the whole
1017 * page. For partial-page I/O this will read in the page.
1019 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1021 CERROR("page index %lu, rc = %d\n", index, rc);
1024 GOTO(err_unlock, rc);
1026 /* XXX not sure if we need this if we are overwriting page */
1027 if (PageError(page)) {
1028 CERROR("error on page index %lu, rc = %d\n", index, rc);
1030 GOTO(err_unlock, rc = -EIO);
1037 lustre_put_page(page);
1041 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1043 struct inode *inode = page->mapping->host;
1046 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1047 if (!err && IS_SYNC(inode))
1048 err = waitfor_one_page(page);
1050 //SetPageUptodate(page); // the client commit_write will do this
1052 SetPageReferenced(page);
1054 lustre_put_page(page);
1058 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1059 struct niobuf_local *lnb, int *pglocked)
1061 struct address_space *mapping = inode->i_mapping;
1065 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1067 page = grab_cache_page_nowait(mapping, index); /* locked page */
1069 page = grab_cache_page(mapping, index); /* locked page */
1072 /* This page is currently locked, so get a temporary page instead. */
1073 /* XXX I believe this is a very dangerous thing to do - consider if
1074 * we had multiple writers for the same file (definitely the case
1075 * if we are using this codepath). If writer A locks the page,
1076 * writer B writes to a copy (as here), writer A drops the page
1077 * lock, and writer C grabs the lock before B does, then B will
1078 * later overwrite the data from C, even if C had LDLM locked
1079 * and initiated the write after B did.
1083 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1084 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1086 CERROR("no memory for a temp page\n");
1088 GOTO(err, rc = -ENOMEM);
1090 page = virt_to_page(addr);
1092 page->index = index;
1093 lnb->flags |= N_LOCAL_TEMP_PAGE;
1094 } else if (!IS_ERR(page)) {
1098 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1099 * a no-op for most filesystems, because we write the whole
1100 * page. For partial-page I/O this will read in the page.
1102 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1104 CERROR("page index %lu, rc = %d\n", index, rc);
1107 GOTO(err_unlock, rc);
1109 /* XXX not sure if we need this if we are overwriting page */
1110 if (PageError(page)) {
1111 CERROR("error on page index %lu, rc = %d\n", index, rc);
1113 GOTO(err_unlock, rc = -EIO);
1120 lustre_put_page(page);
1126 * We need to balance prepare_write() calls with commit_write() calls.
1127 * If the page has been prepared, but we have no data for it, we don't
1128 * want to overwrite valid data on disk, but we still need to zero out
1129 * data for space which was newly allocated. Like part of what happens
1130 * in __block_prepare_write() for newly allocated blocks.
1132 * XXX currently __block_prepare_write() creates buffers for all the
1133 * pages, and the filesystems mark these buffers as BH_New if they
1134 * were newly allocated from disk. We use the BH_New flag similarly.
1136 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1140 unsigned block_start, block_end;
1141 struct buffer_head *bh, *head = page->buffers;
1142 unsigned blocksize = head->b_size;
1143 void *addr = page_address(page);
1145 /* debugging: just seeing if this ever happens */
1146 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1147 page->index, page->mapping->host->i_ino, err);
1149 /* Currently one buffer per page, but in the future... */
1150 for (bh = head, block_start = 0; bh != head || !block_start;
1151 block_start = block_end, bh = bh->b_this_page) {
1152 block_end = block_start + blocksize;
1154 memset(addr + block_start, 0, blocksize);
1158 return lustre_commit_write(page, from, to);
1161 static int filter_preprw(int cmd, struct lustre_handle *conn,
1162 int objcount, struct obd_ioobj *obj,
1163 int niocount, struct niobuf_remote *nb,
1164 struct niobuf_local *res, void **desc_private)
1166 struct obd_run_ctxt saved;
1167 struct obd_device *obd;
1168 struct obd_ioobj *o = obj;
1169 struct niobuf_remote *b = nb;
1170 struct niobuf_local *r = res;
1171 void *journal_save = NULL;
1177 obd = class_conn2obd(conn);
1179 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1182 memset(res, 0, sizeof(*res) * niocount);
1184 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1186 if (cmd & OBD_BRW_WRITE) {
1187 *desc_private = filter_journal_start(&journal_save,
1189 objcount, obj, niocount,
1191 if (IS_ERR(*desc_private))
1192 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1195 for (i = 0; i < objcount; i++, o++) {
1196 struct dentry *dentry;
1197 struct inode *inode;
1200 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1201 o->ioo_id, S_IFREG);
1203 GOTO(out_clean, rc = PTR_ERR(dentry));
1204 inode = dentry->d_inode;
1206 CERROR("trying to BRW to non-existent file "LPU64"\n",
1209 GOTO(out_clean, rc = -ENOENT);
1212 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1213 unsigned long index = b->offset >> PAGE_SHIFT;
1219 r->dentry = dget(dentry);
1221 if (cmd & OBD_BRW_WRITE)
1222 page = filter_get_page_write(inode, index, r,
1225 page = lustre_get_page_read(inode, index);
1229 GOTO(out_clean, rc = PTR_ERR(page));
1232 r->addr = page_address(page);
1233 r->offset = b->offset;
1240 if (cmd & OBD_BRW_WRITE) {
1241 int err = filter_journal_stop(journal_save, &obd->u.filter,
1251 CERROR("error cleanup on brw\n");
1253 if (cmd & OBD_BRW_WRITE)
1254 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1256 lustre_put_page(r->page);
1261 static int filter_write_locked_page(struct niobuf_local *lnb)
1266 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1267 if (IS_ERR(lpage)) {
1268 /* It is highly unlikely that we would ever get an error here.
1269 * The page we want to get was previously locked, so it had to
1270 * have already allocated the space, and we were just writing
1271 * over the same data, so there would be no hole in the file.
1273 * XXX: possibility of a race with truncate could exist, need
1274 * to check that. There are no guarantees w.r.t.
1275 * write order even on a local filesystem, although the
1276 * normal response would be to return the number of bytes
1277 * successfully written and leave the rest to the app.
1279 rc = PTR_ERR(lpage);
1280 CERROR("error getting locked page index %ld: rc = %d\n",
1281 lnb->page->index, rc);
1285 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1286 * lustre_commit_write() below, lnb->page was kmapped previously in
1287 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1289 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1290 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1292 CERROR("error committing locked page %ld: rc = %d\n",
1293 lnb->page->index, rc);
1295 lustre_put_page(lnb->page);
1300 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1301 int objcount, struct obd_ioobj *obj,
1302 int niocount, struct niobuf_local *res,
1305 struct obd_run_ctxt saved;
1306 struct obd_ioobj *o;
1307 struct niobuf_local *r;
1308 struct obd_device *obd = class_conn2obd(conn);
1310 int found_locked = 0;
1315 push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1316 journal_save = current->journal_info;
1317 LASSERT(!journal_save);
1319 current->journal_info = private;
1320 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1322 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1323 struct page *page = r->page;
1328 if (r->flags & N_LOCAL_TEMP_PAGE) {
1333 if (cmd & OBD_BRW_WRITE) {
1334 int err = filter_commit_write(page, 0,
1340 lustre_put_page(page);
1345 current->journal_info = journal_save;
1350 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1352 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1354 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1357 err = filter_write_locked_page(r);
1369 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1371 struct obd_device *obd = class_conn2obd(conn);
1376 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1378 statfs_pack(osfs, &sfs);
1383 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1384 void *key, obd_count *vallen, void **val)
1386 struct obd_device *obd;
1389 obd = class_conn2obd(conn);
1391 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1395 if ( keylen == strlen("blocksize") &&
1396 memcmp(key, "blocksize", keylen) == 0 ) {
1397 *vallen = sizeof(long);
1398 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1402 if ( keylen == strlen("blocksize_bits") &&
1403 memcmp(key, "blocksize_bits", keylen) == 0 ){
1404 *vallen = sizeof(long);
1405 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1409 if ( keylen == strlen("root_ino") &&
1410 memcmp(key, "root_ino", keylen) == 0 ){
1411 *vallen = sizeof(obd_id);
1412 *val = (void *)(obd_id)FILTER_ROOTINO;
1416 CDEBUG(D_IOCTL, "invalid key\n");
1420 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1421 struct lustre_handle *src_conn, struct obdo *src,
1422 obd_size count, obd_off offset)
1425 struct lov_stripe_md srcmd, dstmd;
1426 unsigned long index = 0;
1429 memset(&srcmd, 0, sizeof(srcmd));
1430 memset(&dstmd, 0, sizeof(dstmd));
1431 srcmd.lsm_object_id = src->o_id;
1432 dstmd.lsm_object_id = dst->o_id;
1435 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1436 ", dst: ino "LPU64"\n",
1437 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1438 page = alloc_page(GFP_USER);
1442 while (TryLockPage(page))
1443 ___wait_on_page(page);
1445 /* XXX with brw vector I/O, we could batch up reads and writes here,
1446 * all we need to do is allocate multiple pages to handle the I/Os
1447 * and arrays to handle the request parameters.
1449 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1451 struct io_cb_data *cbd = ll_init_cb();
1460 pg.count = PAGE_SIZE;
1461 pg.off = (page->index) << PAGE_SHIFT;
1464 page->index = index;
1465 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1466 ll_sync_io_cb, cbd);
1479 pg.flag = OBD_BRW_CREATE;
1480 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1482 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1483 ll_sync_io_cb, cbd);
1485 /* XXX should handle dst->o_size, dst->o_blocks here */
1491 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1495 dst->o_size = src->o_size;
1496 dst->o_blocks = src->o_blocks;
1497 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1505 static struct obd_ops filter_obd_ops = {
1506 o_get_info: filter_get_info,
1507 o_setup: filter_setup,
1508 o_cleanup: filter_cleanup,
1509 o_connect: filter_connect,
1510 o_disconnect: filter_disconnect,
1511 o_statfs: filter_statfs,
1512 o_getattr: filter_getattr,
1513 o_create: filter_create,
1514 o_setattr: filter_setattr,
1515 o_destroy: filter_destroy,
1516 o_open: filter_open,
1517 o_close: filter_close,
1518 o_brw: filter_pgcache_brw,
1519 o_punch: filter_truncate,
1520 o_preprw: filter_preprw,
1521 o_commitrw: filter_commitrw
1523 o_preallocate: filter_preallocate_inodes,
1524 o_migrate: filter_migrate,
1525 o_copy: filter_copy_data,
1526 o_iterate: filter_iterate
1531 static int __init obdfilter_init(void)
1533 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1534 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1537 static void __exit obdfilter_exit(void)
1539 class_unregister_type(OBD_FILTER_DEVICENAME);
1542 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1543 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1544 MODULE_LICENSE("GPL");
1546 module_init(obdfilter_init);
1547 module_exit(obdfilter_exit);