2 * linux/fs/ext2_obd/ext2_obd.c
4 * Copyright (C) 2001 Cluster File Systems, Inc.
6 * This code is issued under the GNU General Public License.
7 * See the file COPYING in this distribution
9 * by Peter Braam <braam@clusterfs.com>
14 #include <linux/version.h>
15 #include <linux/module.h>
17 #include <linux/stat.h>
18 #include <linux/locks.h>
19 #include <linux/ext2_fs.h>
20 #include <linux/quotaops.h>
21 #include <asm/unistd.h>
23 #define DEBUG_SUBSYSTEM S_FILTER
25 #include <linux/obd_support.h>
26 #include <linux/obd_class.h>
27 #include <linux/obd_ext2.h>
29 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
32 #define FILTER_ROOTINO 2
36 static char * obd_type_by_mode[S_IFMT >> S_SHIFT] = {
38 [S_IFREG >> S_SHIFT] "R",
39 [S_IFDIR >> S_SHIFT] "D",
40 [S_IFCHR >> S_SHIFT] "C",
41 [S_IFBLK >> S_SHIFT] "B",
42 [S_IFIFO >> S_SHIFT] "F",
43 [S_IFSOCK >> S_SHIFT] "S",
44 [S_IFLNK >> S_SHIFT] "L"
47 static void filter_id(char *buf, obd_id id, obd_mode mode)
49 sprintf(buf, "O/%s/%Ld", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT],
53 static struct file *filter_parent(obd_id id, obd_mode mode)
57 sprintf(path, "O/%s", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT]);
59 file = filp_open(path, O_RDONLY, 0);
63 void push_ctxt(struct run_ctxt *save, struct run_ctxt *new)
66 save->pwd = dget(current->fs->pwd);
67 save->pwdmnt = mntget(current->fs->pwdmnt);
70 set_fs_pwd(current->fs, new->pwdmnt, new->pwd);
73 void pop_ctxt(struct run_ctxt *saved)
76 set_fs_pwd(current->fs, saved->pwdmnt, saved->pwd);
79 mntput(saved->pwdmnt);
82 static int simple_mkdir(struct dentry *dir, char *name, int mode)
84 struct dentry *dchild;
88 dchild = lookup_one_len(name, dir, strlen(name));
91 return PTR_ERR(dchild);
94 if (dchild->d_inode) {
100 err = vfs_mkdir(dir->d_inode, dchild, mode);
107 static int simple_unlink(struct dentry *dir, char *name)
109 struct dentry *dchild;
113 dchild = lookup_one_len(name, dir, strlen(name));
114 if (IS_ERR(dchild)) {
116 return PTR_ERR(dchild);
119 if (!dchild->d_inode) {
125 err = vfs_unlink(dir->d_inode, dchild);
132 static void filter_prep(struct obd_device *obddev)
134 struct run_ctxt saved;
142 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
143 rc = simple_mkdir(current->fs->pwd, "O", 0700);
144 rc = simple_mkdir(current->fs->pwd, "P", 0700);
145 rc = simple_mkdir(current->fs->pwd, "D", 0700);
146 file = filp_open("O", O_RDONLY, 0);
147 if (!file || IS_ERR(file)) {
148 CERROR("cannot open O\n");
151 rc = simple_mkdir(file->f_dentry, "R", 0700); /* regular */
152 rc = simple_mkdir(file->f_dentry, "D", 0700); /* directory */
153 rc = simple_mkdir(file->f_dentry, "L", 0700); /* symbolic links */
154 rc = simple_mkdir(file->f_dentry, "C", 0700); /* character devices */
155 rc = simple_mkdir(file->f_dentry, "B", 0700); /* block devices */
156 rc = simple_mkdir(file->f_dentry, "F", 0700); /* fifo's */
157 rc = simple_mkdir(file->f_dentry, "S", 0700); /* sockets */
158 filp_close(file, NULL);
160 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
161 file = filp_open(rootid, O_RDWR | O_CREAT, 00755);
163 CERROR("OBD filter: cannot make root directory");
167 rc = simple_mkdir(current->fs->pwd, rootid, 0755);
169 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
170 if ( !file || IS_ERR(file) ) {
171 CERROR("OBD filter: cannot open/create status file\n");
175 /* steal operations */
176 inode = file->f_dentry->d_inode;
177 obddev->u.filter.fo_fop = file->f_op;
178 obddev->u.filter.fo_iop = inode->i_op;
179 obddev->u.filter.fo_aops = inode->i_mapping->a_ops;
182 if (inode->i_size == 0) {
183 rc = file->f_op->write(file, (char *)&lastino,
184 sizeof(lastino), &off);
185 if (rc != sizeof(lastino)) {
186 CERROR("OBD filter: error writing lastino\n");
190 rc = file->f_op->read(file, (char *)&lastino, sizeof(lastino),
192 if (rc != sizeof(lastino)) {
193 CERROR("OBD filter: error reading lastino\n");
197 obddev->u.filter.fo_lastino = lastino;
203 static void filter_post(struct obd_device *obddev)
205 struct run_ctxt saved;
210 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
211 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
212 if ( !file || IS_ERR(file)) {
213 CERROR("OBD filter: cannot create status file\n");
216 rc = file->f_op->write(file, (char *)&obddev->u.filter.fo_lastino,
217 sizeof(obddev->u.filter.fo_lastino), &off);
218 if (rc != sizeof(sizeof(obddev->u.filter.fo_lastino)) ) {
219 CERROR("OBD filter: error writing lastino\n");
222 rc = filp_close(file, NULL);
224 CERROR("OBD filter: cannot close status file\n");
231 /* release per client resources */
232 static int filter_disconnect(struct obd_conn *conn)
234 /* XXX cleanup preallocated inodes */
235 return gen_disconnect(conn);
236 } /* ext2obd_disconnect */
238 /* mount the file system (secretly) */
239 static int filter_setup(struct obd_device *obddev, obd_count len,
243 struct obd_ioctl_data* data = buf;
244 struct vfsmount *mnt;
249 mnt = do_kern_mount(data->ioc_inlbuf2, 0,
250 data->ioc_inlbuf1, NULL);
257 obddev->u.filter.fo_sb = mnt->mnt_root->d_inode->i_sb;
258 if (!obddev->u.filter.fo_sb) {
263 obddev->u.filter.fo_vfsmnt = mnt;
264 obddev->u.filter.fo_fstype = strdup(data->ioc_inlbuf2);
266 obddev->u.filter.fo_ctxt.pwdmnt = mnt;
267 obddev->u.filter.fo_ctxt.pwd = mnt->mnt_root;
268 obddev->u.filter.fo_ctxt.fs = KERNEL_DS;
271 spin_lock_init(&obddev->u.filter.fo_lock);
278 static __u64 filter_next_id(struct obd_device *obddev)
281 spin_lock(&obddev->u.filter.fo_lock);
282 obddev->u.filter.fo_lastino++;
283 id = obddev->u.filter.fo_lastino;
284 spin_unlock(&obddev->u.filter.fo_lock);
288 static int filter_cleanup(struct obd_device * obddev)
290 struct super_block *sb;
294 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
299 if ( !list_empty(&obddev->obd_gen_clients) ) {
300 CERROR("still has clients!\n");
305 sb = obddev->u.filter.fo_sb;
306 if (!obddev->u.filter.fo_sb){
313 mntput(obddev->u.filter.fo_vfsmnt);
314 obddev->u.filter.fo_sb = 0;
315 kfree(obddev->u.filter.fo_fstype);
325 static struct file *filter_obj_open(struct obd_device *obddev,
331 struct run_ctxt saved;
332 struct super_block *sb;
334 sb = obddev->u.filter.fo_sb;
335 if (!sb || !sb->s_dev) {
336 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
342 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)oa->o_id);
347 if ( ! (oa->o_mode & S_IFMT) ) {
348 CERROR("OBD filter_obj_open, no type (%Ld), mode %o!\n",
349 oa->o_id, oa->o_mode);
351 filter_id(id, oa->o_id, oa->o_mode);
352 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
353 file = filp_open(id , O_RDONLY | O_LARGEFILE, 0);
357 error = PTR_ERR(file);
360 CDEBUG(D_INODE, "opening obdo %s\n", id);
365 static struct inode *filter_inode_from_obdo(struct obd_device *obddev,
371 file = filter_obj_open(obddev, oa);
373 CERROR("filter_inode_from_obdo failed\n");
377 inode = iget(file->f_dentry->d_inode->i_sb, file->f_dentry->d_inode->i_ino);
382 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
384 int type = oa->o_mode & S_IFMT;
387 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
388 inode->i_ino, (long)oa->o_id, oa->o_valid);
389 obdo_from_inode(oa, inode);
390 oa->o_mode &= ~S_IFMT;
393 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
394 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
395 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
397 *((obd_rdev *)oa->o_inline) = rdev;
398 oa->o_obdflags |= OBD_FL_INLINEDATA;
399 oa->o_valid |= OBD_MD_FLINLINE;
403 else if (ext2obd_has_inline(inode)) {
404 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
405 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
406 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
407 oa->o_obdflags |= OBD_FL_INLINEDATA;
408 oa->o_valid |= OBD_MD_FLINLINE;
411 if (ext2obd_has_obdmd(inode)) {
412 /* XXX this will change when we don't store the obdmd in data */
413 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
414 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
415 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
416 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
417 oa->o_valid |= OBD_MD_FLOBDMD;
423 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
428 if ( !gen_client(conn) ) {
429 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
434 if ( !(inode = filter_inode_from_obdo(conn->oc_dev, oa)) ) {
439 oa->o_valid &= ~OBD_MD_FLID;
440 filter_from_inode(oa, inode);
447 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
454 if (!gen_client(conn)) {
455 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
459 inode = filter_inode_from_obdo(conn->oc_dev, oa);
465 iattr_from_obdo(&iattr, oa);
466 iattr.ia_mode &= ~S_IFMT;
467 iattr.ia_mode |= S_IFREG;
469 if ( inode->i_op->setattr ) {
470 rc = inode->i_op->setattr(&de, &iattr);
472 rc = inode_setattr(inode, &iattr);
480 static int filter_create (struct obd_conn* conn, struct obdo *oa)
483 struct run_ctxt saved;
486 struct obd_device *obddev = conn->oc_dev;
490 if (!gen_client(conn)) {
491 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
494 CDEBUG(D_IOCTL, "\n");
496 oa->o_id = filter_next_id(conn->oc_dev);
497 if ( !(oa->o_mode && S_IFMT) ) {
498 CERROR("filter obd: no type!\n");
502 filter_id(name, oa->o_id, oa->o_mode);
503 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
507 file = filp_open(name, O_RDONLY | O_CREAT, mode);
510 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
514 CDEBUG(D_IOCTL, "\n");
516 /* Set flags for fields we have set in ext2_new_inode */
517 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
518 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
519 OBD_MD_FLUID | OBD_MD_FLGID;
523 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
525 struct obd_device * obddev;
526 struct obd_client * cli;
527 struct inode * inode;
530 struct run_ctxt saved;
533 if (!(cli = gen_client(conn))) {
534 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
539 obddev = conn->oc_dev;
540 inode = filter_inode_from_obdo(obddev, oa);
548 inode->i_mode = 010000;
551 filter_id(id, oa->o_id, oa->o_mode);
552 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
553 dir = filter_parent(oa->o_id, oa->o_mode);
554 if (!dir || IS_ERR(dir)) {
558 rc = simple_unlink(dir->f_dentry, id);
567 static int filter_truncate(struct obd_conn *conn, struct obdo *oa, obd_size count,
572 error = filter_setattr(conn, oa);
573 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
579 /* buffer must lie in user memory here */
580 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
581 obd_size *count, obd_off offset)
584 unsigned long retval;
587 if (!gen_client(conn)) {
588 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
593 file = filter_obj_open(conn->oc_dev, oa);
594 if (!file || IS_ERR(file)) {
596 return -PTR_ERR(file);
599 /* count doubles as retval */
600 retval = file->f_op->read(file, buf, *count, &offset);
615 /* buffer must lie in user memory here */
616 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
617 obd_size *count, obd_off offset)
621 unsigned long retval;
624 if (!gen_client(conn)) {
625 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
630 file = filter_obj_open(conn->oc_dev, oa);
631 if (!file || IS_ERR(file)) {
633 return -PTR_ERR(file);
636 /* count doubles as retval */
637 retval = file->f_op->write(file, buf, *count, &offset);
651 } /* ext2obd_write */
653 static int filter_pgcache_brw(int rw, struct obd_conn *conn,
662 struct super_block *sb;
664 int onum; /* index to oas */
665 int pnum; /* index to pages (bufs) */
666 unsigned long retval;
672 if (!gen_client(conn)) {
673 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
678 sb = conn->oc_dev->u.filter.fo_sb;
682 pnum = 0; /* pnum indexes buf 0..num_pages */
683 for (onum = 0; onum < num_oa; onum++) {
686 file = filter_obj_open(conn->oc_dev, oa[onum]);
687 if (!file || IS_ERR(file)) {
693 /* count doubles as retval */
694 for (pg = 0; pg < oa_bufs[onum]; pg++) {
695 CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) off count (%Ld,%Ld)\n",
696 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
697 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
698 offset[pnum], count[pnum]);
703 buffer = kmap(pages[pnum]);
704 retval = file->f_op->write(file, buffer, count[pnum], &off);
706 CDEBUG(D_INODE, "retval %ld\n", retval);
708 loff_t off = offset[pnum];
709 char *buffer = kmap(pages[pnum]);
711 if (off >= file->f_dentry->d_inode->i_size) {
712 memset(buffer, 0, count[pnum]);
713 retval = count[pnum];
715 retval = file->f_op->read(file, buffer, count[pnum], &off);
719 if ( retval != count[pnum] ) {
725 CDEBUG(D_INODE, "retval %ld\n", retval);
729 /* sizes and blocks are set by generic_file_write */
730 /* ctimes/mtimes will follow with a setattr call */
737 error = (retval >= 0) ? 0 : retval;
741 static int filter_statfs (struct obd_conn *conn, struct statfs * statfs)
743 struct super_block *sb;
748 if (!gen_client(conn)) {
749 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
754 sb = conn->oc_dev->u.filter.fo_sb;
756 err = sb->s_op->statfs(sb, statfs);
759 } /* ext2obd_statfs */
762 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
763 void *key, obd_count *vallen, void **val)
765 struct obd_device *obddev;
766 struct obd_client * cli;
769 if (!(cli = gen_client(conn))) {
770 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
774 obddev = conn->oc_dev;
776 if ( keylen == strlen("blocksize") &&
777 memcmp(key, "blocksize", keylen) == 0 ) {
778 *vallen = sizeof(int);
779 *val = (void *)obddev->u.filter.fo_sb->s_blocksize;
784 if ( keylen == strlen("blocksize_bits") &&
785 memcmp(key, "blocksize_bits", keylen) == 0 ){
786 *vallen = sizeof(int);
787 *val = (void *)(int)obddev->u.filter.fo_sb->s_blocksize_bits;
792 if ( keylen == strlen("root_ino") &&
793 memcmp(key, "root_ino", keylen) == 0 ){
794 *vallen = sizeof(int);
795 *val = (void *)(int) FILTER_ROOTINO;
800 CDEBUG(D_IOCTL, "invalid key\n");
805 struct obd_ops filter_obd_ops = {
807 o_get_info: filter_get_info,
808 o_setup: filter_setup,
809 o_cleanup: filter_cleanup,
810 o_connect: gen_connect,
811 o_disconnect: filter_disconnect,
812 o_statfs: filter_statfs,
813 o_getattr: filter_getattr,
814 o_create: filter_create,
815 o_setattr: filter_setattr,
816 o_destroy: filter_destroy,
818 o_write: filter_write,
819 o_brw: filter_pgcache_brw,
820 o_punch: filter_truncate,
822 o_preallocate: ext2obd_preallocate_inodes,
823 o_migrate: ext2obd_migrate,
824 o_copy: gen_copy_data,
825 o_iterate: ext2obd_iterate
830 static int __init obdfilter_init(void)
832 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
833 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
836 static void __exit obdfilter_exit(void)
838 obd_unregister_type(OBD_FILTER_DEVICENAME);
841 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
842 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
843 MODULE_LICENSE("GPL");
845 module_init(obdfilter_init);
846 module_exit(obdfilter_exit);