1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/filter/filter.c
6 * Copyright (C) 2001 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>
15 #define DEBUG_SUBSYSTEM S_FILTER
17 #include <linux/module.h>
18 #include <linux/obd_filter.h>
19 #include <linux/ext3_jbd.h>
20 #include <linux/quotaops.h>
22 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
25 #define FILTER_ROOTINO 2
28 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
30 [S_IFREG >> S_SHIFT] "R",
31 [S_IFDIR >> S_SHIFT] "D",
32 [S_IFCHR >> S_SHIFT] "C",
33 [S_IFBLK >> S_SHIFT] "B",
34 [S_IFIFO >> S_SHIFT] "F",
35 [S_IFSOCK >> S_SHIFT] "S",
36 [S_IFLNK >> S_SHIFT] "L"
39 static inline const char *obd_mode_to_type(int mode)
41 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
44 /* write the pathname into the string */
45 static int filter_id(char *buf, obd_id id, obd_mode mode)
47 return sprintf(buf, "O/%s/%Ld", obd_mode_to_type(mode),
48 (unsigned long long)id);
51 /* setup the object store with correct subdirectories */
52 static int filter_prep(struct obd_device *obddev)
54 struct obd_run_ctxt saved;
55 struct filter_obd *filter = &obddev->u.filter;
56 struct dentry *dentry;
65 push_ctxt(&saved, &filter->fo_ctxt);
66 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
67 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
70 CERROR("cannot open/create O: rc = %d\n", rc);
73 filter->fo_dentry_O = dentry;
74 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
75 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
78 CERROR("cannot open/create P: rc = %d\n", rc);
81 CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
82 atomic_read(&dentry->d_count) - 1);
84 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
85 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
88 CERROR("cannot open/create D: rc = %d\n", rc);
91 CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
92 atomic_read(&dentry->d_count) - 1);
96 * Create directories and/or get dentries for each object type.
97 * This saves us from having to do multiple lookups for each one.
99 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
100 char *type = obd_type_by_mode[mode];
103 filter->fo_dentry_O_mode[mode] = NULL;
106 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
107 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
108 if (IS_ERR(dentry)) {
109 rc = PTR_ERR(dentry);
110 CERROR("cannot create O/%s: rc = %d\n", type, rc);
111 GOTO(out_O_mode, rc);
113 filter->fo_dentry_O_mode[mode] = dentry;
116 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
117 file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
120 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
122 GOTO(out_O_mode, rc);
126 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
127 if ( !file || IS_ERR(file) ) {
129 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
131 GOTO(out_O_mode, rc);
134 /* steal operations */
135 inode = file->f_dentry->d_inode;
136 filter->fo_fop = file->f_op;
137 filter->fo_iop = inode->i_op;
138 filter->fo_aops = inode->i_mapping->a_ops;
141 if (inode->i_size == 0) {
142 ssize_t retval = file->f_op->write(file, (char *)&lastino,
143 sizeof(lastino), &off);
144 if (retval != sizeof(lastino)) {
145 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
147 GOTO(out_O_mode, rc = -EIO);
150 ssize_t retval = file->f_op->read(file, (char *)&lastino,
151 sizeof(lastino), &off);
152 if (retval != sizeof(lastino)) {
153 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
155 GOTO(out_O_mode, rc = -EIO);
158 filter->fo_lastino = lastino;
168 while (--mode >= 0) {
169 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
171 CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
172 obd_type_by_mode[mode], dentry,
173 atomic_read(&dentry->d_count) - 1);
175 filter->fo_dentry_O_mode[mode] = NULL;
179 CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
180 atomic_read(&filter->fo_dentry_O->d_count) - 1);
181 dput(filter->fo_dentry_O);
182 filter->fo_dentry_O = NULL;
186 /* cleanup the filter: write last used object id to status file */
187 static void filter_post(struct obd_device *obddev)
189 struct obd_run_ctxt saved;
190 struct filter_obd *filter = &obddev->u.filter;
196 push_ctxt(&saved, &filter->fo_ctxt);
197 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
199 CERROR("OBD filter: cannot create status file\n");
202 rc = file->f_op->write(file, (char *)&filter->fo_lastino,
203 sizeof(filter->fo_lastino), &off);
204 if (rc != sizeof(filter->fo_lastino))
205 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
207 rc = filp_close(file, NULL);
209 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
211 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
212 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
214 CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
215 obd_type_by_mode[mode], dentry,
216 atomic_read(&dentry->d_count) - 1);
218 filter->fo_dentry_O_mode[mode] = NULL;
221 CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
222 atomic_read(&filter->fo_dentry_O->d_count) - 1);
223 dput(filter->fo_dentry_O);
229 static __u64 filter_next_id(struct obd_device *obddev)
232 spin_lock(&obddev->u.filter.fo_lock);
233 obddev->u.filter.fo_lastino++;
234 id = obddev->u.filter.fo_lastino;
235 spin_unlock(&obddev->u.filter.fo_lock);
239 /* how to get files, dentries, inodes from object id's */
240 /* parent i_sem is already held if needed for exclusivity */
241 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
242 struct dentry *dparent,
243 __u64 id, __u32 type)
245 struct super_block *sb = obddev->u.filter.fo_sb;
246 struct dentry *dchild;
251 if (!sb || !sb->s_dev) {
252 CERROR("fatal: device not initialized.\n");
253 RETURN(ERR_PTR(-ENXIO));
257 CERROR("fatal: invalid object #0\n");
259 RETURN(ERR_PTR(-ESTALE));
262 if (!(type & S_IFMT)) {
263 CERROR("OBD %s, object %Lu has bad type: %o\n", __FUNCTION__,
264 (unsigned long long)id, type);
265 RETURN(ERR_PTR(-EINVAL));
268 len = sprintf(name, "%Ld", id);
269 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
271 dchild = lookup_one_len(name, dparent, len);
272 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
273 obd_mode_to_type(type), name, dchild,
274 atomic_read(&dchild->d_count));
276 if (IS_ERR(dchild)) {
277 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
284 static struct file *filter_obj_open(struct obd_device *obddev,
285 __u64 id, __u32 type)
287 struct super_block *sb = obddev->u.filter.fo_sb;
288 struct obd_run_ctxt saved;
293 if (!sb || !sb->s_dev) {
294 CERROR("fatal: device not initialized.\n");
295 RETURN(ERR_PTR(-ENXIO));
299 CERROR("fatal: invalid obdo %Lu\n", (unsigned long long)id);
300 RETURN(ERR_PTR(-ESTALE));
303 if (!(type & S_IFMT)) {
304 CERROR("OBD %s, no type (%Ld), mode %o!\n", __FUNCTION__,
305 (unsigned long long)id, type);
306 RETURN(ERR_PTR(-EINVAL));
309 filter_id(name, id, type);
310 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
311 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
314 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
321 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
323 struct filter_obd *filter = &obddev->u.filter;
325 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
329 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
330 __u64 id, __u32 type)
332 struct dentry *dentry;
335 dentry = filter_fid2dentry(obddev, filter_parent(obddev, type),
337 if (IS_ERR(dentry)) {
338 CERROR("%s: lookup failed: rc = %ld\n", __FUNCTION__,
344 inode = iget(dentry->d_inode->i_sb, dentry->d_inode->i_ino);
346 CDEBUG(D_INODE, "put child %p, count = %d\n", dentry,
347 atomic_read(&dentry->d_count) - 1);
349 CDEBUG(D_INODE, "got inode %p (%ld), count = %d\n", inode, inode->i_ino,
350 atomic_read(&inode->i_count));
355 static int filter_connect(struct obd_conn *conn)
360 rc = gen_connect(conn);
368 static int filter_disconnect(struct obd_conn *conn)
372 rc = gen_disconnect(conn);
376 /* XXX cleanup preallocated inodes */
380 /* mount the file system (secretly) */
381 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
383 struct obd_ioctl_data* data = buf;
384 struct filter_obd *filter;
385 struct vfsmount *mnt;
389 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
393 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
398 filter = &obddev->u.filter;;
399 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
400 /* XXX is this even possible if do_kern_mount succeeded? */
402 GOTO(err_put, err = -ENODEV);
404 filter->fo_vfsmnt = mnt;
405 filter->fo_fstype = strdup(data->ioc_inlbuf2);
407 filter->fo_ctxt.pwdmnt = mnt;
408 filter->fo_ctxt.pwd = mnt->mnt_root;
409 filter->fo_ctxt.fs = KERNEL_DS;
411 err = filter_prep(obddev);
413 GOTO(err_kfree, err);
414 spin_lock_init(&filter->fo_lock);
419 kfree(filter->fo_fstype);
422 mntput(filter->fo_vfsmnt);
432 static int filter_cleanup(struct obd_device * obddev)
434 struct super_block *sb;
437 if (!(obddev->obd_flags & OBD_SET_UP))
440 if (!list_empty(&obddev->obd_gen_clients)) {
441 CERROR("still has clients!\n");
445 sb = obddev->u.filter.fo_sb;
446 if (!obddev->u.filter.fo_sb)
451 shrink_dcache_parent(sb->s_root);
453 mntput(obddev->u.filter.fo_vfsmnt);
454 obddev->u.filter.fo_sb = 0;
455 kfree(obddev->u.filter.fo_fstype);
464 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
466 int type = oa->o_mode & S_IFMT;
469 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
470 inode->i_ino, (long)oa->o_id, oa->o_valid);
471 obdo_from_inode(oa, inode);
472 oa->o_mode &= ~S_IFMT;
475 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
476 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
477 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
479 *((obd_rdev *)oa->o_inline) = rdev;
480 oa->o_obdflags |= OBD_FL_INLINEDATA;
481 oa->o_valid |= OBD_MD_FLINLINE;
485 else if (filter_has_inline(inode)) {
486 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
487 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
488 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
489 oa->o_obdflags |= OBD_FL_INLINEDATA;
490 oa->o_valid |= OBD_MD_FLINLINE;
493 if (filter_has_obdmd(inode)) {
494 /* XXX this will change when we don't store the obdmd in data */
495 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
496 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
497 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
498 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
499 oa->o_valid |= OBD_MD_FLOBDMD;
505 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
507 struct obd_device *obddev;
508 struct dentry *dentry;
511 if (!gen_client(conn)) {
512 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
516 obddev = conn->oc_dev;
517 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
518 oa->o_id, oa->o_mode);
520 RETURN(PTR_ERR(dentry));
522 oa->o_valid &= ~OBD_MD_FLID;
523 filter_from_inode(oa, dentry->d_inode);
529 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
531 struct obd_run_ctxt saved;
532 struct obd_device *obddev;
533 struct dentry *dentry;
539 if (!gen_client(conn)) {
540 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
544 obddev = conn->oc_dev;
545 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
546 oa->o_id, oa->o_mode);
548 RETURN(PTR_ERR(dentry));
550 inode = dentry->d_inode;
551 iattr_from_obdo(&iattr, oa);
552 iattr.ia_mode &= ~S_IFMT;
553 iattr.ia_mode |= S_IFREG;
555 if (iattr.ia_mode & ATTR_SIZE)
557 push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
558 if (inode->i_op->setattr)
559 rc = inode->i_op->setattr(dentry, &iattr);
561 rc = inode_setattr(inode, &iattr);
563 if (iattr.ia_mode & ATTR_SIZE)
567 CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
568 atomic_read(&dentry->d_count) - 1);
573 static int filter_open(struct obd_conn *conn, struct obdo *oa)
575 struct obd_device *obddev;
576 struct dentry *dentry;
579 if (!gen_client(conn)) {
580 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
584 obddev = conn->oc_dev;
585 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
586 oa->o_id, oa->o_mode);
588 RETURN(PTR_ERR(dentry));
593 static int filter_close(struct obd_conn *conn, struct obdo *oa)
595 struct obd_device *obddev;
596 struct dentry *dentry;
599 if (!gen_client(conn)) {
600 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
604 obddev = conn->oc_dev;
605 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
606 oa->o_id, oa->o_mode);
608 RETURN(PTR_ERR(dentry));
610 CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
611 atomic_read(&dentry->d_count) - 1);
612 dput(dentry); /* for the close */
613 CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
614 atomic_read(&dentry->d_count) - 1);
615 dput(dentry); /* for this call */
619 static int filter_create(struct obd_conn* conn, struct obdo *oa)
622 struct obd_run_ctxt saved;
625 struct obd_device *obddev = conn->oc_dev;
629 if (!gen_client(conn)) {
630 CERROR("invalid client %u\n", conn->oc_id);
634 oa->o_id = filter_next_id(conn->oc_dev);
635 if (!(oa->o_mode && S_IFMT)) {
636 CERROR("filter obd: no type!\n");
640 filter_id(name, oa->o_id, oa->o_mode);
641 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
642 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
643 file = filp_open(name, O_RDONLY | O_CREAT, mode);
646 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
651 /* Set flags for fields we have set in the inode struct */
652 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
653 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
654 OBD_MD_FLUID | OBD_MD_FLGID;
656 /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
660 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
662 struct obd_run_ctxt saved;
663 struct obd_device *obddev;
664 struct obd_client *cli;
666 struct dentry *dir_dentry, *object_dentry;
670 if (!(cli = gen_client(conn))) {
671 CERROR("invalid client %u\n", conn->oc_id);
675 CDEBUG(D_INODE, "destroying object %Ld\n",oa->o_id);
676 obddev = conn->oc_dev;
678 dir_dentry = filter_parent(obddev, oa->o_mode);
679 down(&dir_dentry->d_inode->i_sem);
681 object_dentry = filter_fid2dentry(obddev, dir_dentry, oa->o_id,
683 if (IS_ERR(object_dentry))
684 GOTO(out, rc = -ENOENT);
686 inode = object_dentry->d_inode;
687 if (inode->i_nlink != 1) {
688 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
691 inode->i_mode = S_IFREG;
693 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
695 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
697 CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
698 atomic_read(&object_dentry->d_count) - 1);
703 up(&dir_dentry->d_inode->i_sem);
707 /* NB count and offset are used for punch, but not truncate */
708 static int filter_truncate(struct obd_conn *conn, struct obdo *oa,
709 obd_size count, obd_off offset)
714 CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
715 "o_size = %Ld\n", oa->o_id, oa->o_valid, oa->o_size);
716 error = filter_setattr(conn, oa);
717 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
722 /* buffer must lie in user memory here */
723 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
724 obd_size *count, obd_off offset)
727 unsigned long retval;
731 if (!gen_client(conn)) {
732 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
736 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
738 RETURN(PTR_ERR(file));
740 /* count doubles as retval */
741 retval = file->f_op->read(file, buf, *count, (loff_t *)&offset);
756 /* buffer must lie in user memory here */
757 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
758 obd_size *count, obd_off offset)
760 struct obd_run_ctxt saved;
763 unsigned long retval;
766 if (!gen_client(conn)) {
767 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
771 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
773 RETURN(PTR_ERR(file));
775 /* count doubles as retval */
776 push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
777 retval = file->f_op->write(file, buf, *count, (loff_t *)&offset);
794 static int filter_pgcache_brw(int rw, struct obd_conn *conn, obd_count num_oa,
795 struct obdo **oa, obd_count *oa_bufs,
796 struct page **pages, obd_size *count,
797 obd_off *offset, obd_flag *flags, void *callback)
799 struct obd_run_ctxt saved;
800 struct super_block *sb;
801 int onum; /* index to oas */
802 int pnum; /* index to pages (bufs) */
803 unsigned long retval;
808 if (!gen_client(conn)) {
809 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
813 sb = conn->oc_dev->u.filter.fo_sb;
815 push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
816 pnum = 0; /* pnum indexes buf 0..num_pages */
817 for (onum = 0; onum < num_oa; onum++) {
820 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
823 GOTO(out, retval = PTR_ERR(file));
825 /* count doubles as retval */
826 for (pg = 0; pg < oa_bufs[onum]; pg++) {
827 CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) "
828 "off count (%Ld,%Ld)\n",
829 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
830 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
831 (unsigned long long)offset[pnum],
832 (unsigned long long)count[pnum]);
837 buffer = kmap(pages[pnum]);
838 retval = file->f_op->write(file, buffer, count[pnum], &off);
840 CDEBUG(D_INODE, "retval %ld\n", retval);
842 loff_t off = offset[pnum];
843 char *buffer = kmap(pages[pnum]);
845 if (off >= file->f_dentry->d_inode->i_size) {
846 memset(buffer, 0, count[pnum]);
847 retval = count[pnum];
849 retval = file->f_op->read(file, buffer, count[pnum], &off);
853 if (retval != count[pnum]) {
855 GOTO(out, retval = -EIO);
857 CDEBUG(D_INODE, "retval %ld\n", retval);
861 /* sizes and blocks are set by generic_file_write */
862 /* ctimes/mtimes will follow with a setattr call */
870 error = (retval >= 0) ? 0 : retval;
875 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
877 struct super_block *sb = conn->oc_dev->u.filter.fo_sb;
878 struct inode *inode = NULL;
881 if (!sb || !sb->s_dev) {
882 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
887 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
891 inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
892 if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
893 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
894 (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
895 "no links" : "NULL");
904 * Calculate the number of buffer credits needed to write multiple pages in
905 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
906 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
908 * See comment above ext3_writepage_trans_blocks for details. We assume
909 * no data journaling is being done, but it does allow for all of the pages
910 * being non-contiguous. If we are guaranteed contiguous pages we could
911 * reduce the number of (d)indirect blocks a lot.
913 * With N blocks per page and P pages, for each inode we have at most:
915 * min(N*P, blocksize/4 + 1) dindirect blocks
918 * For the entire filesystem, we have at most:
919 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
920 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
923 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
925 static int ext3_credits_needed(struct super_block *sb, int objcount,
926 struct obd_ioobj *obj)
928 struct obd_ioobj *o = obj;
929 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
930 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
933 int needed = objcount + 1;
936 for (i = 0; i < objcount; i++, o++) {
937 int nblocks = o->ioo_bufcnt * blockpp;
938 int ndindirect = min(nblocks, addrpp + 1);
939 int nindir = nblocks + ndindirect + 1;
941 nbitmaps += nindir + nblocks;
942 ngdblocks += nindir + nblocks;
947 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
948 nbitmaps = EXT3_SB(sb)->s_groups_count;
949 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
950 ngdblocks = EXT3_SB(sb)->s_gdb_count;
952 needed += nbitmaps + ngdblocks;
955 /* We assume that there will be 1 bit set in s_dquot.flags for each
956 * quota file that is active. This is at least true for now.
958 needed += hweight32(sb_any_quota_enabled(sb)) *
959 EXT3_SINGLEDATA_TRANS_BLOCKS;
965 /* We have to start a huge journal transaction here to hold all of the
966 * metadata for the pages being written here. This is necessitated by
967 * the fact that we do lots of prepare_write operations before we do
968 * any of the matching commit_write operations, so even if we split
969 * up to use "smaller" transactions none of them could complete until
970 * all of them were opened. By having a single journal transaction,
971 * we eliminate duplicate reservations for common blocks like the
972 * superblock and group descriptors or bitmaps.
974 * We will start the transaction here, but each prepare_write will
975 * add a refcount to the transaction, and each commit_write will
976 * remove a refcount. The transaction will be closed when all of
977 * the pages have been written.
979 static void *ext3_filter_journal_start(struct filter_obd *filter,
980 int objcount, struct obd_ioobj *obj,
981 int niocount, struct niobuf_remote *nb)
983 journal_t *journal = NULL;
984 handle_t *handle = NULL;
987 /* Assumes ext3 and extN have same sb_info layout, but avoids issues
988 * with having extN built properly before filterobd for now.
990 journal = EXT3_SB(filter->fo_sb)->s_journal;
991 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
993 /* The number of blocks we could _possibly_ dirty can very large.
994 * We reduce our request if it is absurd (and we couldn't get that
995 * many credits for a single handle anyways).
997 * At some point we have to limit the size of I/Os sent at one time,
998 * increase the size of the journal, or we have to calculate the
999 * actual journal requirements more carefully by checking all of
1000 * the blocks instead of being maximally pessimistic. It remains to
1001 * be seen if this is a real problem or not.
1003 if (needed > journal->j_max_transaction_buffers) {
1004 CERROR("want too many journal credits (%d) using %d instead\n",
1005 needed, journal->j_max_transaction_buffers);
1006 needed = journal->j_max_transaction_buffers;
1009 handle = journal_start(journal, needed);
1011 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1017 static void *filter_journal_start(void **journal_save,
1018 struct filter_obd *filter,
1019 int objcount, struct obd_ioobj *obj,
1020 int niocount, struct niobuf_remote *nb)
1022 void *handle = NULL;
1024 /* This may not be necessary - we probably never have a
1025 * transaction started when we enter here, so we can
1026 * remove the saving of the journal state entirely.
1027 * For now leave it in just to see if it ever happens.
1029 *journal_save = current->journal_info;
1030 if (*journal_save) {
1031 CERROR("Already have handle %p???\n", *journal_save);
1033 current->journal_info = NULL;
1036 if (!strcmp(filter->fo_fstype, "ext3") ||
1037 !strcmp(filter->fo_fstype, "extN"))
1038 handle = ext3_filter_journal_start(filter, objcount, obj,
1043 static int ext3_filter_journal_stop(void *handle)
1047 /* We got a refcount on the handle for each call to prepare_write,
1048 * so we can drop the "parent" handle here to avoid the need for
1049 * osc to call back into filterobd to close the handle. The
1050 * remaining references will be dropped in commit_write.
1052 rc = journal_stop((handle_t *)handle);
1057 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1062 if (!strcmp(filter->fo_fstype, "ext3") ||
1063 !strcmp(filter->fo_fstype, "extN"))
1064 rc = ext3_filter_journal_stop(handle);
1066 current->journal_info = journal_save;
1071 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1072 struct niobuf_local *lnb)
1074 struct address_space *mapping = inode->i_mapping;
1078 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1079 page = grab_cache_page_nowait(mapping, index); /* locked page */
1081 /* This page is currently locked, so we grab a new one temporarily */
1084 addr = __get_free_pages(GFP_KERNEL, 0);
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)) {
1095 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1096 * a no-op for most filesystems, because we write the whole
1097 * page. For partial-page I/O this will read in the page.
1099 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1101 CERROR("page index %lu, rc = %d\n", index, rc);
1104 GOTO(err_unlock, rc);
1106 /* XXX not sure if we need this if we are overwriting page */
1107 if (PageError(page)) {
1108 CERROR("error on page index %lu, rc = %d\n", index, rc);
1110 GOTO(err_unlock, rc = -EIO);
1119 lustre_put_page(page);
1124 static int filter_preprw(int cmd, struct obd_conn *conn,
1125 int objcount, struct obd_ioobj *obj,
1126 int niocount, struct niobuf_remote *nb,
1127 struct niobuf_local *res, void **desc_private)
1129 struct obd_run_ctxt saved;
1130 struct obd_device *obddev;
1131 struct obd_ioobj *o = obj;
1132 struct niobuf_remote *b = nb;
1133 struct niobuf_local *r = res;
1134 void *journal_save = NULL;
1139 memset(res, 0, sizeof(*res) * niocount);
1140 obddev = conn->oc_dev;
1142 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
1144 if (cmd == OBD_BRW_WRITE) {
1145 *desc_private = filter_journal_start(&journal_save,
1147 objcount, obj, niocount,
1149 if (IS_ERR(*desc_private))
1150 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1153 for (i = 0; i < objcount; i++, o++) {
1154 struct dentry *dentry;
1155 struct inode *inode;
1158 dentry = filter_fid2dentry(obddev,
1159 filter_parent(obddev, S_IFREG),
1160 o->ioo_id, S_IFREG);
1162 GOTO(out_ctxt, rc = PTR_ERR(dentry));
1163 inode = dentry->d_inode;
1165 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1166 unsigned long index = b->offset >> PAGE_SHIFT;
1169 /* XXX We _might_ change this to a dcount if we
1170 * wanted to pass a dentry pointer in the niobuf
1171 * to avoid doing so many igets on an inode we
1172 * already have. It appears to be solely for the
1173 * purpose of having a refcount that we can drop
1174 * in commitrw where we get one call per page.
1177 r->dentry = dget(dentry);
1181 /* FIXME: we need to iput all inodes on error */
1183 GOTO(out_ctxt, rc = -EINVAL);
1185 if (cmd == OBD_BRW_WRITE) {
1186 page = filter_get_page_write(inode, index, r);
1188 /* We unlock the page to avoid deadlocks with
1189 * the page I/O because we are preparing
1190 * multiple pages at one time and we have lock
1191 * ordering problems. Lustre I/O and disk I/O
1192 * on this page can happen concurrently.
1195 page = lustre_get_page_read(inode, index);
1197 /* FIXME: we need to clean up here... */
1199 GOTO(out_ctxt, rc = PTR_ERR(page));
1201 r->addr = page_address(page);
1202 r->offset = b->offset;
1208 if (cmd == OBD_BRW_WRITE) {
1209 /* FIXME: need to clean up here */
1210 rc = filter_journal_stop(journal_save, &obddev->u.filter,
1218 static int filter_write_locked_page(struct niobuf_local *lnb)
1223 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1226 memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1228 __free_pages(lnb->page, 0);
1230 rc = lustre_commit_page(lpage, 0, PAGE_SIZE);
1236 static int filter_commitrw(int cmd, struct obd_conn *conn,
1237 int objcount, struct obd_ioobj *obj,
1238 int niocount, struct niobuf_local *res,
1241 struct obd_run_ctxt saved;
1242 struct obd_ioobj *o = obj;
1243 struct niobuf_local *r = res;
1245 int found_locked = 0;
1249 // if (cmd == OBD_BRW_WRITE)
1250 push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
1251 journal_save = current->journal_info;
1253 CERROR("Existing handle %p???\n", journal_save);
1254 current->journal_info = private;
1255 for (i = 0; i < objcount; i++, obj++) {
1257 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1258 struct page *page = r->page;
1260 /* If there was an error setting up a particular page
1261 * for I/O we still need to continue with the rest of
1262 * the pages in order to balance prepate/commit_write
1263 * calls, and to complete as much I/O as possible.
1268 if (r->flags & N_LOCAL_TEMP_PAGE) {
1273 if (cmd == OBD_BRW_WRITE) {
1275 rc = lustre_commit_page(page, 0, PAGE_SIZE);
1277 /* FIXME: still need to iput the other inodes */
1281 lustre_put_page(page);
1284 "put inode %p (%ld), count = %d, nlink = %d\n",
1285 page->mapping->host,
1286 page->mapping->host->i_ino,
1287 atomic_read(&page->mapping->host->i_count) - 1,
1288 page->mapping->host->i_nlink);
1295 for (i = 0; i < objcount; i++, obj++) {
1297 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1299 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1302 rc = filter_write_locked_page(r);
1308 current->journal_info = journal_save;
1313 static int filter_statfs(struct obd_conn *conn, struct statfs * statfs)
1315 struct super_block *sb;
1319 if (!gen_client(conn)) {
1320 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
1324 sb = conn->oc_dev->u.filter.fo_sb;
1326 err = sb->s_op->statfs(sb, statfs);
1328 } /* filter_statfs */
1331 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
1332 void *key, obd_count *vallen, void **val)
1334 struct obd_device *obddev;
1335 struct obd_client * cli;
1338 if (!(cli = gen_client(conn))) {
1339 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
1343 obddev = conn->oc_dev;
1345 if ( keylen == strlen("blocksize") &&
1346 memcmp(key, "blocksize", keylen) == 0 ) {
1347 *vallen = sizeof(long);
1348 *val = (void *)(long)obddev->u.filter.fo_sb->s_blocksize;
1352 if ( keylen == strlen("blocksize_bits") &&
1353 memcmp(key, "blocksize_bits", keylen) == 0 ){
1354 *vallen = sizeof(long);
1355 *val = (void *)(long)obddev->u.filter.fo_sb->s_blocksize_bits;
1359 if ( keylen == strlen("root_ino") &&
1360 memcmp(key, "root_ino", keylen) == 0 ){
1361 *vallen = sizeof(long);
1362 *val = (void *)(long)FILTER_ROOTINO;
1366 CDEBUG(D_IOCTL, "invalid key\n");
1371 static struct obd_ops filter_obd_ops = {
1372 o_get_info: filter_get_info,
1373 o_setup: filter_setup,
1374 o_cleanup: filter_cleanup,
1375 o_connect: filter_connect,
1376 o_disconnect: filter_disconnect,
1377 o_statfs: filter_statfs,
1378 o_getattr: filter_getattr,
1379 o_create: filter_create,
1380 o_setattr: filter_setattr,
1381 o_destroy: filter_destroy,
1382 o_open: filter_open,
1383 o_close: filter_close,
1384 o_read: filter_read,
1385 o_write: filter_write,
1386 o_brw: filter_pgcache_brw,
1387 o_punch: filter_truncate,
1388 o_preprw: filter_preprw,
1389 o_commitrw: filter_commitrw
1391 o_preallocate: filter_preallocate_inodes,
1392 o_migrate: filter_migrate,
1393 o_copy: gen_copy_data,
1394 o_iterate: filter_iterate
1399 static int __init obdfilter_init(void)
1401 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
1402 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1405 static void __exit obdfilter_exit(void)
1407 obd_unregister_type(OBD_FILTER_DEVICENAME);
1410 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
1411 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1412 MODULE_LICENSE("GPL");
1414 module_init(obdfilter_init);
1415 module_exit(obdfilter_exit);