1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/ext2_obd/ext2_obd.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_ext2.h>
19 #include <linux/obd_filter.h>
21 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
24 #define FILTER_ROOTINO 2
27 static char * obd_type_by_mode[S_IFMT >> S_SHIFT] = {
29 [S_IFREG >> S_SHIFT] "R",
30 [S_IFDIR >> S_SHIFT] "D",
31 [S_IFCHR >> S_SHIFT] "C",
32 [S_IFBLK >> S_SHIFT] "B",
33 [S_IFIFO >> S_SHIFT] "F",
34 [S_IFSOCK >> S_SHIFT] "S",
35 [S_IFLNK >> S_SHIFT] "L"
39 /* write the pathname into the string */
40 static void filter_id(char *buf, obd_id id, obd_mode mode)
42 sprintf(buf, "O/%s/%Ld", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT],
43 (unsigned long long)id);
46 /* setup the object store with correct subdirectories */
47 static int filter_prep(struct obd_device *obddev)
49 struct obd_run_ctxt saved;
57 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
58 rc = simple_mkdir(current->fs->pwd, "O", 0700);
59 rc = simple_mkdir(current->fs->pwd, "P", 0700);
60 rc = simple_mkdir(current->fs->pwd, "D", 0700);
61 file = filp_open("O", O_RDONLY, 0);
63 CERROR("cannot open O\n");
64 GOTO(out, rc = PTR_ERR(file));
66 rc = simple_mkdir(file->f_dentry, "R", 0700); /* regular */
67 rc = simple_mkdir(file->f_dentry, "D", 0700); /* directory */
68 rc = simple_mkdir(file->f_dentry, "L", 0700); /* symbolic links */
69 rc = simple_mkdir(file->f_dentry, "C", 0700); /* character devices */
70 rc = simple_mkdir(file->f_dentry, "B", 0700); /* block devices */
71 rc = simple_mkdir(file->f_dentry, "F", 0700); /* fifo's */
72 rc = simple_mkdir(file->f_dentry, "S", 0700); /* sockets */
73 filp_close(file, NULL);
75 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
76 file = filp_open(rootid, O_RDWR | O_CREAT, 00755);
78 CERROR("OBD filter: cannot make root directory");
79 GOTO(out, rc = PTR_ERR(file));
82 /* FIXME: this is the same as the _file_ we just created? */
83 rc = simple_mkdir(current->fs->pwd, rootid, 0755);
85 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
86 if ( !file || IS_ERR(file) ) {
87 CERROR("OBD filter: cannot open/create status file\n");
88 GOTO(out, rc = PTR_ERR(file));
91 /* steal operations */
92 inode = file->f_dentry->d_inode;
93 obddev->u.filter.fo_fop = file->f_op;
94 obddev->u.filter.fo_iop = inode->i_op;
95 obddev->u.filter.fo_aops = inode->i_mapping->a_ops;
98 if (inode->i_size == 0) {
99 ssize_t retval = file->f_op->write(file, (char *)&lastino,
100 sizeof(lastino), &off);
101 if (retval != sizeof(lastino)) {
102 CERROR("OBD filter: error writing lastino\n");
103 GOTO(out, rc = -EIO);
106 ssize_t retval = file->f_op->read(file, (char *)&lastino,
107 sizeof(lastino), &off);
108 if (retval != sizeof(lastino)) {
109 CERROR("OBD filter: error reading lastino\n");
110 GOTO(out, rc = -EIO);
113 obddev->u.filter.fo_lastino = lastino;
123 /* cleanup the filter: write last used object id to status file */
124 static void filter_post(struct obd_device *obddev)
126 struct obd_run_ctxt saved;
131 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
132 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
133 if ( !file || IS_ERR(file)) {
134 CERROR("OBD filter: cannot create status file\n");
137 rc = file->f_op->write(file, (char *)&obddev->u.filter.fo_lastino,
138 sizeof(obddev->u.filter.fo_lastino), &off);
139 if (rc != sizeof(obddev->u.filter.fo_lastino) ) {
140 CERROR("OBD filter: error writing lastino\n");
143 rc = filp_close(file, NULL);
145 CERROR("OBD filter: cannot close status file\n");
152 static __u64 filter_next_id(struct obd_device *obddev)
155 spin_lock(&obddev->u.filter.fo_lock);
156 obddev->u.filter.fo_lastino++;
157 id = obddev->u.filter.fo_lastino;
158 spin_unlock(&obddev->u.filter.fo_lock);
162 /* how to get files, dentries, inodes from object id's */
163 static struct file *filter_obj_open(struct obd_device *obddev,
164 __u64 id, __u32 type)
166 struct obd_run_ctxt saved;
168 struct super_block *sb;
172 sb = obddev->u.filter.fo_sb;
173 if (!sb || !sb->s_dev) {
174 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
179 CDEBUG(D_INODE, "fatal: invalid obdo %Lu\n",
180 (unsigned long long)id);
184 if ( ! (type & S_IFMT) ) {
185 CERROR("OBD filter_obj_open, no type (%Ld), mode %o!\n",
186 (unsigned long long)id, type);
189 filter_id(name, id, type);
190 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
191 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0);
194 CDEBUG(D_INODE, "opening obdo %s\n", name);
199 static struct file *filter_parent(obd_id id, obd_mode mode)
204 sprintf(path, "O/%s", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT]);
206 file = filp_open(path, O_RDONLY, 0);
211 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
212 __u64 id, __u32 type)
217 file = filter_obj_open(obddev, id, type);
219 CERROR("filter_inode_from_obdo failed\n");
223 inode = iget(file->f_dentry->d_inode->i_sb,
224 file->f_dentry->d_inode->i_ino);
230 static int filter_connect(struct obd_conn *conn)
235 rc = gen_connect(conn);
243 static int filter_disconnect(struct obd_conn *conn)
247 rc = gen_disconnect(conn);
251 /* XXX cleanup preallocated inodes */
255 /* mount the file system (secretly) */
256 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
258 struct obd_ioctl_data* data = buf;
259 struct vfsmount *mnt;
263 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
267 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
272 /* XXX is this even possible if do_kern_mount succeeded? */
273 obddev->u.filter.fo_sb = mnt->mnt_root->d_inode->i_sb;
274 if (!obddev->u.filter.fo_sb)
275 GOTO(err_put, err = -ENODEV);
277 obddev->u.filter.fo_vfsmnt = mnt;
278 obddev->u.filter.fo_fstype = strdup(data->ioc_inlbuf2);
280 obddev->u.filter.fo_ctxt.pwdmnt = mnt;
281 obddev->u.filter.fo_ctxt.pwd = mnt->mnt_root;
282 obddev->u.filter.fo_ctxt.fs = KERNEL_DS;
284 err = filter_prep(obddev);
286 GOTO(err_kfree, err);
287 spin_lock_init(&obddev->u.filter.fo_lock);
292 kfree(obddev->u.filter.fo_fstype);
295 mntput(obddev->u.filter.fo_vfsmnt);
296 obddev->u.filter.fo_sb = 0;
305 static int filter_cleanup(struct obd_device * obddev)
307 struct super_block *sb;
311 if ( !list_empty(&obddev->obd_gen_clients) ) {
312 CERROR("still has clients!\n");
316 sb = obddev->u.filter.fo_sb;
317 if (!obddev->u.filter.fo_sb)
323 mntput(obddev->u.filter.fo_vfsmnt);
324 obddev->u.filter.fo_sb = 0;
325 kfree(obddev->u.filter.fo_fstype);
334 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
336 int type = oa->o_mode & S_IFMT;
339 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
340 inode->i_ino, (long)oa->o_id, oa->o_valid);
341 obdo_from_inode(oa, inode);
342 oa->o_mode &= ~S_IFMT;
345 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
346 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
347 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
349 *((obd_rdev *)oa->o_inline) = rdev;
350 oa->o_obdflags |= OBD_FL_INLINEDATA;
351 oa->o_valid |= OBD_MD_FLINLINE;
355 else if (filter_has_inline(inode)) {
356 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
357 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
358 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
359 oa->o_obdflags |= OBD_FL_INLINEDATA;
360 oa->o_valid |= OBD_MD_FLINLINE;
363 if (filter_has_obdmd(inode)) {
364 /* XXX this will change when we don't store the obdmd in data */
365 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
366 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
367 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
368 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
369 oa->o_valid |= OBD_MD_FLOBDMD;
375 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
380 if ( !gen_client(conn) ) {
381 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
385 inode = filter_inode_from_obj(conn->oc_dev, oa->o_id, oa->o_mode);
389 oa->o_valid &= ~OBD_MD_FLID;
390 filter_from_inode(oa, inode);
396 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
404 if (!gen_client(conn)) {
405 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
409 inode = filter_inode_from_obj(conn->oc_dev, oa->o_id, oa->o_mode);
413 iattr_from_obdo(&iattr, oa);
414 iattr.ia_mode &= ~S_IFMT;
415 iattr.ia_mode |= S_IFREG;
417 if ( inode->i_op->setattr )
418 rc = inode->i_op->setattr(&de, &iattr);
420 rc = inode_setattr(inode, &iattr);
426 static int filter_open(struct obd_conn *conn, struct obdo *oa)
431 if (!gen_client(conn))
434 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
435 oa->o_id, oa->o_mode)) )
441 static int filter_close(struct obd_conn *conn, struct obdo *oa)
446 if (!gen_client(conn))
449 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
450 oa->o_id, oa->o_mode)) )
453 iput(inode); /* for the close */
454 iput(inode); /* for this call */
458 static int filter_create (struct obd_conn* conn, struct obdo *oa)
461 struct obd_run_ctxt saved;
464 struct obd_device *obddev = conn->oc_dev;
468 if (!gen_client(conn)) {
469 CERROR("invalid client %u\n", conn->oc_id);
473 oa->o_id = filter_next_id(conn->oc_dev);
474 if ( !(oa->o_mode && S_IFMT) ) {
475 CERROR("filter obd: no type!\n");
479 filter_id(name, oa->o_id, oa->o_mode);
480 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
484 file = filp_open(name, O_RDONLY | O_CREAT, mode);
487 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
492 /* Set flags for fields we have set in ext2_new_inode */
493 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
494 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
495 OBD_MD_FLUID | OBD_MD_FLGID;
499 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
501 struct obd_device * obddev;
502 struct obd_client * cli;
505 struct dentry *dir_dentry, *object_dentry;
507 struct obd_run_ctxt saved;
510 if (!(cli = gen_client(conn))) {
511 CERROR("invalid client %u\n", conn->oc_id);
515 obddev = conn->oc_dev;
516 object = filter_obj_open(obddev, oa->o_id, oa->o_mode);
517 if (!object || IS_ERR(object))
520 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
521 dir = filter_parent(oa->o_id, oa->o_mode);
523 GOTO(out, rc = PTR_ERR(dir));
524 dir_dentry = dget(dir->f_dentry);
525 object_dentry = dget(object->f_dentry);
527 filp_close(object, 0);
528 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
538 static int filter_truncate(struct obd_conn *conn, struct obdo *oa,
539 obd_size count, obd_off offset)
544 error = filter_setattr(conn, oa);
545 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
550 /* buffer must lie in user memory here */
551 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
552 obd_size *count, obd_off offset)
555 unsigned long retval;
559 if (!gen_client(conn)) {
560 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
564 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
565 if (!file || IS_ERR(file))
566 RETURN(-PTR_ERR(file));
568 /* count doubles as retval */
569 retval = file->f_op->read(file, buf, *count, (loff_t *)&offset);
584 /* buffer must lie in user memory here */
585 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
586 obd_size *count, obd_off offset)
590 unsigned long retval;
593 if (!gen_client(conn)) {
594 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
598 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
599 if (!file || IS_ERR(file))
600 RETURN(-PTR_ERR(file));
602 /* count doubles as retval */
603 retval = file->f_op->write(file, buf, *count, (loff_t *)&offset);
619 static int filter_pgcache_brw(int rw, struct obd_conn *conn,
628 struct super_block *sb;
630 int onum; /* index to oas */
631 int pnum; /* index to pages (bufs) */
632 unsigned long retval;
637 if (!gen_client(conn)) {
638 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
642 sb = conn->oc_dev->u.filter.fo_sb;
646 pnum = 0; /* pnum indexes buf 0..num_pages */
647 for (onum = 0; onum < num_oa; onum++) {
650 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
652 if (!file || IS_ERR(file))
653 GOTO(ERROR, error = -ENOENT);
655 /* count doubles as retval */
656 for (pg = 0; pg < oa_bufs[onum]; pg++) {
657 CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) "
658 "off count (%Ld,%Ld)\n",
659 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
660 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
661 (unsigned long long)offset[pnum],
662 (unsigned long long)count[pnum]);
667 buffer = kmap(pages[pnum]);
668 retval = file->f_op->write(file, buffer, count[pnum], &off);
670 CDEBUG(D_INODE, "retval %ld\n", retval);
672 loff_t off = offset[pnum];
673 char *buffer = kmap(pages[pnum]);
675 if (off >= file->f_dentry->d_inode->i_size) {
676 memset(buffer, 0, count[pnum]);
677 retval = count[pnum];
679 retval = file->f_op->read(file, buffer, count[pnum], &off);
683 if ( retval != count[pnum] ) {
685 GOTO(ERROR, retval = -EIO);
687 CDEBUG(D_INODE, "retval %ld\n", retval);
691 /* sizes and blocks are set by generic_file_write */
692 /* ctimes/mtimes will follow with a setattr call */
699 error = (retval >= 0) ? 0 : retval;
704 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
706 struct inode *inode = NULL;
707 struct super_block *sb = conn->oc_dev->u.ext2.e2_sb;
710 if (!sb || !sb->s_dev) {
711 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
716 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
720 inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
721 if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
722 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
723 (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
724 "no links" : "NULL");
733 static int filter_preprw(int cmd, struct obd_conn *conn,
734 int objcount, struct obd_ioobj *obj,
735 int niocount, struct niobuf_remote *nb,
736 struct niobuf_local *res)
738 struct obd_ioobj *o = obj;
739 struct niobuf_remote *b = nb;
740 struct niobuf_local *r = res;
744 memset(res, 0, sizeof(*res) * niocount);
746 for (i = 0; i < objcount; i++, o++) {
748 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
749 unsigned long index = b->offset >> PAGE_SHIFT;
750 struct inode *inode = ioobj_to_inode(conn, o);
753 /* FIXME: we need to iput all inodes on error */
757 if (cmd == OBD_BRW_WRITE)
758 page = lustre_get_page_write(inode, index);
760 page = lustre_get_page_read(inode, index);
762 RETURN(PTR_ERR(page));
764 r->addr = (__u64)(unsigned long)page_address(page);
765 r->offset = b->offset;
773 static int filter_commitrw(int cmd, struct obd_conn *conn,
774 int objcount, struct obd_ioobj *obj,
775 int niocount, struct niobuf_local *res)
777 struct obd_ioobj *o = obj;
778 struct niobuf_local *r = res;
782 for (i = 0; i < objcount; i++, obj++) {
784 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
785 struct page *page = r->page;
790 if (cmd == OBD_BRW_WRITE) {
791 int rc = lustre_commit_page(page, 0, PAGE_SIZE);
793 /* FIXME: still need to iput the other inodes */
797 lustre_put_page(page);
799 iput(page->mapping->host);
805 static int filter_statfs (struct obd_conn *conn, struct statfs * statfs)
807 struct super_block *sb;
811 if (!gen_client(conn)) {
812 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
816 sb = conn->oc_dev->u.filter.fo_sb;
818 err = sb->s_op->statfs(sb, statfs);
820 } /* filter_statfs */
823 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
824 void *key, obd_count *vallen, void **val)
826 struct obd_device *obddev;
827 struct obd_client * cli;
830 if (!(cli = gen_client(conn))) {
831 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
835 obddev = conn->oc_dev;
837 if ( keylen == strlen("blocksize") &&
838 memcmp(key, "blocksize", keylen) == 0 ) {
839 *vallen = sizeof(int);
840 *val = (void *)obddev->u.filter.fo_sb->s_blocksize;
844 if ( keylen == strlen("blocksize_bits") &&
845 memcmp(key, "blocksize_bits", keylen) == 0 ){
846 *vallen = sizeof(int);
847 *val = (void *)(int)obddev->u.filter.fo_sb->s_blocksize_bits;
851 if ( keylen == strlen("root_ino") &&
852 memcmp(key, "root_ino", keylen) == 0 ){
853 *vallen = sizeof(int);
854 *val = (void *)(int) FILTER_ROOTINO;
858 CDEBUG(D_IOCTL, "invalid key\n");
863 struct obd_ops filter_obd_ops = {
865 o_get_info: filter_get_info,
866 o_setup: filter_setup,
867 o_cleanup: filter_cleanup,
868 o_connect: filter_connect,
869 o_disconnect: filter_disconnect,
870 o_statfs: filter_statfs,
871 o_getattr: filter_getattr,
872 o_create: filter_create,
873 o_setattr: filter_setattr,
874 o_destroy: filter_destroy,
876 o_close: filter_close,
878 o_write: filter_write,
879 o_brw: filter_pgcache_brw,
880 o_punch: filter_truncate,
881 o_preprw: filter_preprw,
882 o_commitrw: filter_commitrw
884 o_preallocate: filter_preallocate_inodes,
885 o_migrate: filter_migrate,
886 o_copy: gen_copy_data,
887 o_iterate: filter_iterate
892 static int __init obdfilter_init(void)
894 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
895 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
898 static void __exit obdfilter_exit(void)
900 obd_unregister_type(OBD_FILTER_DEVICENAME);
903 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
904 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
905 MODULE_LICENSE("GPL");
907 module_init(obdfilter_init);
908 module_exit(obdfilter_exit);