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>
16 #include <linux/version.h>
17 #include <linux/module.h>
19 #include <linux/stat.h>
20 #include <linux/locks.h>
21 #include <linux/ext2_fs.h>
22 #include <linux/quotaops.h>
23 #include <asm/unistd.h>
25 #define DEBUG_SUBSYSTEM S_FILTER
27 #include <linux/obd_class.h>
28 #include <linux/obd_ext2.h>
29 #include <linux/obd_filter.h>
31 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
34 #define FILTER_ROOTINO 2
37 static char * obd_type_by_mode[S_IFMT >> S_SHIFT] = {
39 [S_IFREG >> S_SHIFT] "R",
40 [S_IFDIR >> S_SHIFT] "D",
41 [S_IFCHR >> S_SHIFT] "C",
42 [S_IFBLK >> S_SHIFT] "B",
43 [S_IFIFO >> S_SHIFT] "F",
44 [S_IFSOCK >> S_SHIFT] "S",
45 [S_IFLNK >> S_SHIFT] "L"
49 /* write the pathname into the string */
50 static void filter_id(char *buf, obd_id id, obd_mode mode)
52 sprintf(buf, "O/%s/%Ld", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT],
56 /* setup the object store with correct subdirectories */
57 static int filter_prep(struct obd_device *obddev)
59 struct obd_run_ctxt saved;
67 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
68 rc = simple_mkdir(current->fs->pwd, "O", 0700);
69 rc = simple_mkdir(current->fs->pwd, "P", 0700);
70 rc = simple_mkdir(current->fs->pwd, "D", 0700);
71 file = filp_open("O", O_RDONLY, 0);
73 CERROR("cannot open O\n");
74 GOTO(out, rc = PTR_ERR(file));
76 rc = simple_mkdir(file->f_dentry, "R", 0700); /* regular */
77 rc = simple_mkdir(file->f_dentry, "D", 0700); /* directory */
78 rc = simple_mkdir(file->f_dentry, "L", 0700); /* symbolic links */
79 rc = simple_mkdir(file->f_dentry, "C", 0700); /* character devices */
80 rc = simple_mkdir(file->f_dentry, "B", 0700); /* block devices */
81 rc = simple_mkdir(file->f_dentry, "F", 0700); /* fifo's */
82 rc = simple_mkdir(file->f_dentry, "S", 0700); /* sockets */
83 filp_close(file, NULL);
85 filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
86 file = filp_open(rootid, O_RDWR | O_CREAT, 00755);
88 CERROR("OBD filter: cannot make root directory");
89 GOTO(out, rc = PTR_ERR(file));
92 rc = simple_mkdir(current->fs->pwd, rootid, 0755);
94 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
95 if ( !file || IS_ERR(file) ) {
96 CERROR("OBD filter: cannot open/create status file\n");
97 GOTO(out, rc = PTR_ERR(file));
100 /* steal operations */
101 inode = file->f_dentry->d_inode;
102 obddev->u.filter.fo_fop = file->f_op;
103 obddev->u.filter.fo_iop = inode->i_op;
104 obddev->u.filter.fo_aops = inode->i_mapping->a_ops;
107 if (inode->i_size == 0) {
108 ssize_t retval = file->f_op->write(file, (char *)&lastino,
109 sizeof(lastino), &off);
110 if (retval != sizeof(lastino)) {
111 CERROR("OBD filter: error writing lastino\n");
112 GOTO(out, rc = -EIO);
115 ssize_t retval = file->f_op->read(file, (char *)&lastino,
116 sizeof(lastino), &off);
117 if (retval != sizeof(lastino)) {
118 CERROR("OBD filter: error reading lastino\n");
119 GOTO(out, rc = -EIO);
122 obddev->u.filter.fo_lastino = lastino;
131 /* cleanup the filter: write last used object id to status file */
132 static void filter_post(struct obd_device *obddev)
134 struct obd_run_ctxt saved;
139 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
140 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
141 if ( !file || IS_ERR(file)) {
142 CERROR("OBD filter: cannot create status file\n");
145 rc = file->f_op->write(file, (char *)&obddev->u.filter.fo_lastino,
146 sizeof(obddev->u.filter.fo_lastino), &off);
147 if (rc != sizeof(obddev->u.filter.fo_lastino) ) {
148 CERROR("OBD filter: error writing lastino\n");
151 rc = filp_close(file, NULL);
153 CERROR("OBD filter: cannot close status file\n");
160 static __u64 filter_next_id(struct obd_device *obddev)
163 spin_lock(&obddev->u.filter.fo_lock);
164 obddev->u.filter.fo_lastino++;
165 id = obddev->u.filter.fo_lastino;
166 spin_unlock(&obddev->u.filter.fo_lock);
170 /* how to get files, dentries, inodes from object id's */
171 static struct file *filter_obj_open(struct obd_device *obddev,
172 __u64 id, __u32 type)
174 struct obd_run_ctxt saved;
176 struct super_block *sb;
179 sb = obddev->u.filter.fo_sb;
180 if (!sb || !sb->s_dev) {
181 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
187 CDEBUG(D_INODE, "fatal: invalid obdo %Lu\n", id);
192 if ( ! (type & S_IFMT) ) {
193 CERROR("OBD filter_obj_open, no type (%Ld), mode %o!\n",
197 filter_id(name, id, type);
198 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
199 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0);
202 CDEBUG(D_INODE, "opening obdo %s\n", name);
207 static struct file *filter_parent(obd_id id, obd_mode mode)
212 sprintf(path, "O/%s", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT]);
214 file = filp_open(path, O_RDONLY, 0);
219 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
220 __u64 id, __u32 type)
225 file = filter_obj_open(obddev, id, type);
227 CERROR("filter_inode_from_obdo failed\n");
231 inode = iget(file->f_dentry->d_inode->i_sb,
232 file->f_dentry->d_inode->i_ino);
238 static int filter_connect(struct obd_conn *conn)
243 rc = gen_connect(conn);
251 static int filter_disconnect(struct obd_conn *conn)
255 rc = gen_disconnect(conn);
259 /* XXX cleanup preallocated inodes */
263 /* mount the file system (secretly) */
264 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
266 struct obd_ioctl_data* data = buf;
267 struct vfsmount *mnt;
272 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
277 /* XXX is this even possible if do_kern_mount succeeded? */
278 obddev->u.filter.fo_sb = mnt->mnt_root->d_inode->i_sb;
279 if (!obddev->u.filter.fo_sb)
280 GOTO(err_put, err = -ENODEV);
282 obddev->u.filter.fo_vfsmnt = mnt;
283 obddev->u.filter.fo_fstype = strdup(data->ioc_inlbuf2);
285 obddev->u.filter.fo_ctxt.pwdmnt = mnt;
286 obddev->u.filter.fo_ctxt.pwd = mnt->mnt_root;
287 obddev->u.filter.fo_ctxt.fs = KERNEL_DS;
289 err = filter_prep(obddev);
291 GOTO(err_kfree, err);
292 spin_lock_init(&obddev->u.filter.fo_lock);
297 kfree(obddev->u.filter.fo_fstype);
300 mntput(obddev->u.filter.fo_vfsmnt);
301 obddev->u.filter.fo_sb = 0;
310 static int filter_cleanup(struct obd_device * obddev)
312 struct super_block *sb;
316 if ( !(obddev->obd_flags & OBD_SET_UP) )
319 if ( !list_empty(&obddev->obd_gen_clients) ) {
320 CERROR("still has clients!\n");
324 sb = obddev->u.filter.fo_sb;
325 if (!obddev->u.filter.fo_sb)
331 mntput(obddev->u.filter.fo_vfsmnt);
332 obddev->u.filter.fo_sb = 0;
333 kfree(obddev->u.filter.fo_fstype);
342 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
344 int type = oa->o_mode & S_IFMT;
347 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
348 inode->i_ino, (long)oa->o_id, oa->o_valid);
349 obdo_from_inode(oa, inode);
350 oa->o_mode &= ~S_IFMT;
353 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
354 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
355 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
357 *((obd_rdev *)oa->o_inline) = rdev;
358 oa->o_obdflags |= OBD_FL_INLINEDATA;
359 oa->o_valid |= OBD_MD_FLINLINE;
363 else if (ext2obd_has_inline(inode)) {
364 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
365 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
366 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
367 oa->o_obdflags |= OBD_FL_INLINEDATA;
368 oa->o_valid |= OBD_MD_FLINLINE;
371 if (ext2obd_has_obdmd(inode)) {
372 /* XXX this will change when we don't store the obdmd in data */
373 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
374 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
375 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
376 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
377 oa->o_valid |= OBD_MD_FLOBDMD;
383 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
388 if ( !gen_client(conn) ) {
389 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
394 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
395 oa->o_id, oa->o_mode)) ) {
400 oa->o_valid &= ~OBD_MD_FLID;
401 filter_from_inode(oa, inode);
408 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
415 if (!gen_client(conn)) {
416 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
420 inode = filter_inode_from_obj(conn->oc_dev, oa->o_id, oa->o_mode);
426 iattr_from_obdo(&iattr, oa);
427 iattr.ia_mode &= ~S_IFMT;
428 iattr.ia_mode |= S_IFREG;
430 if ( inode->i_op->setattr ) {
431 rc = inode->i_op->setattr(&de, &iattr);
433 rc = inode_setattr(inode, &iattr);
441 static int filter_create (struct obd_conn* conn, struct obdo *oa)
444 struct obd_run_ctxt saved;
447 struct obd_device *obddev = conn->oc_dev;
451 if (!gen_client(conn)) {
452 CERROR("invalid client %u\n", conn->oc_id);
456 oa->o_id = filter_next_id(conn->oc_dev);
457 if ( !(oa->o_mode && S_IFMT) ) {
458 CERROR("filter obd: no type!\n");
462 filter_id(name, oa->o_id, oa->o_mode);
463 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
467 file = filp_open(name, O_RDONLY | O_CREAT, mode);
470 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
475 /* Set flags for fields we have set in ext2_new_inode */
476 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
477 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
478 OBD_MD_FLUID | OBD_MD_FLGID;
482 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
484 struct obd_device * obddev;
485 struct obd_client * cli;
486 struct inode * inode;
490 struct obd_run_ctxt saved;
492 if (!(cli = gen_client(conn))) {
493 CERROR("invalid client %u\n", conn->oc_id);
498 obddev = conn->oc_dev;
499 object = filter_obj_open(obddev, oa->o_id, oa->o_mode);
500 if (!object || IS_ERR(object)) {
505 inode = object->f_dentry->d_inode;
507 inode->i_mode = 010000;
509 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
510 dir = filter_parent(oa->o_id, oa->o_mode);
517 dget(object->f_dentry);
518 rc = vfs_unlink(dir->f_dentry->d_inode, object->f_dentry);
521 filp_close(object, 0);
528 static int filter_truncate(struct obd_conn *conn, struct obdo *oa, obd_size count,
533 error = filter_setattr(conn, oa);
534 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
540 /* buffer must lie in user memory here */
541 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
542 obd_size *count, obd_off offset)
545 unsigned long retval;
548 if (!gen_client(conn)) {
549 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
554 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
555 if (!file || IS_ERR(file)) {
557 return -PTR_ERR(file);
560 /* count doubles as retval */
561 retval = file->f_op->read(file, buf, *count, &offset);
576 /* buffer must lie in user memory here */
577 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
578 obd_size *count, obd_off offset)
582 unsigned long retval;
585 if (!gen_client(conn)) {
586 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
591 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
592 if (!file || IS_ERR(file)) {
594 return -PTR_ERR(file);
597 /* count doubles as retval */
598 retval = file->f_op->write(file, buf, *count, &offset);
612 } /* ext2obd_write */
614 static int filter_pgcache_brw(int rw, struct obd_conn *conn,
623 struct super_block *sb;
625 int onum; /* index to oas */
626 int pnum; /* index to pages (bufs) */
627 unsigned long retval;
633 if (!gen_client(conn)) {
634 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
639 sb = conn->oc_dev->u.filter.fo_sb;
643 pnum = 0; /* pnum indexes buf 0..num_pages */
644 for (onum = 0; onum < num_oa; onum++) {
647 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
649 if (!file || IS_ERR(file)) {
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) off count (%Ld,%Ld)\n",
658 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
659 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
660 offset[pnum], count[pnum]);
665 buffer = kmap(pages[pnum]);
666 retval = file->f_op->write(file, buffer, count[pnum], &off);
668 CDEBUG(D_INODE, "retval %ld\n", retval);
670 loff_t off = offset[pnum];
671 char *buffer = kmap(pages[pnum]);
673 if (off >= file->f_dentry->d_inode->i_size) {
674 memset(buffer, 0, count[pnum]);
675 retval = count[pnum];
677 retval = file->f_op->read(file, buffer, count[pnum], &off);
681 if ( retval != count[pnum] ) {
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;
709 if (!sb || !sb->s_dev) {
710 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
716 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
721 inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
722 if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
723 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
724 (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
725 "no links" : "NULL");
735 static int filter_preprw(int cmd, struct obd_conn *conn,
736 int objcount, struct obd_ioobj *obj,
737 int niocount, struct niobuf *nb,
740 struct obd_ioobj *o = obj;
741 struct niobuf *b = nb;
742 struct niobuf *r = res;
746 memset(res, 0, sizeof(*res) * niocount);
748 for (i = 0; i < objcount; i++, o++) {
750 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
751 struct inode *inode = ioobj_to_inode(conn, o);
754 /* FIXME: we need to iput all inodes on error */
758 page = lustre_get_page(inode, b->offset >> PAGE_SHIFT);
760 RETURN(PTR_ERR(page));
762 if (cmd == OBD_BRW_WRITE) {
763 int rc = lustre_prepare_page(0, PAGE_SIZE,page);
765 CERROR("i %d j %d objcount %d bufcnt %d , rc %d, offset %Ld\n", i, j, objcount, o->ioo_bufcnt, rc, b->offset);
768 r->addr = (__u64)(unsigned long)page_address(page);
769 r->offset = b->offset;
777 static int filter_commitrw(int cmd, struct obd_conn *conn,
778 int objcount, struct obd_ioobj *obj,
779 int niocount, struct niobuf *res)
781 struct obd_ioobj *o = obj;
782 struct niobuf *r = res;
786 for (i = 0; i < objcount; i++, obj++) {
788 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
789 struct page *page = r->page;
794 if (cmd == OBD_BRW_WRITE) {
795 int rc = lustre_commit_page(page, 0, PAGE_SIZE);
799 lustre_put_page(page);
801 iput(page->mapping->host);
807 static int filter_statfs (struct obd_conn *conn, struct statfs * statfs)
809 struct super_block *sb;
814 if (!gen_client(conn)) {
815 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
820 sb = conn->oc_dev->u.filter.fo_sb;
822 err = sb->s_op->statfs(sb, statfs);
825 } /* ext2obd_statfs */
828 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
829 void *key, obd_count *vallen, void **val)
831 struct obd_device *obddev;
832 struct obd_client * cli;
835 if (!(cli = gen_client(conn))) {
836 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
840 obddev = conn->oc_dev;
842 if ( keylen == strlen("blocksize") &&
843 memcmp(key, "blocksize", keylen) == 0 ) {
844 *vallen = sizeof(int);
845 *val = (void *)obddev->u.filter.fo_sb->s_blocksize;
850 if ( keylen == strlen("blocksize_bits") &&
851 memcmp(key, "blocksize_bits", keylen) == 0 ){
852 *vallen = sizeof(int);
853 *val = (void *)(int)obddev->u.filter.fo_sb->s_blocksize_bits;
858 if ( keylen == strlen("root_ino") &&
859 memcmp(key, "root_ino", keylen) == 0 ){
860 *vallen = sizeof(int);
861 *val = (void *)(int) FILTER_ROOTINO;
866 CDEBUG(D_IOCTL, "invalid key\n");
871 struct obd_ops filter_obd_ops = {
873 o_get_info: filter_get_info,
874 o_setup: filter_setup,
875 o_cleanup: filter_cleanup,
876 o_connect: filter_connect,
877 o_disconnect: filter_disconnect,
878 o_statfs: filter_statfs,
879 o_getattr: filter_getattr,
880 o_create: filter_create,
881 o_setattr: filter_setattr,
882 o_destroy: filter_destroy,
884 o_write: filter_write,
885 o_brw: filter_pgcache_brw,
886 o_punch: filter_truncate,
887 o_preprw: filter_preprw,
888 o_commitrw: filter_commitrw
890 o_preallocate: ext2obd_preallocate_inodes,
891 o_migrate: ext2obd_migrate,
892 o_copy: gen_copy_data,
893 o_iterate: ext2obd_iterate
898 static int __init obdfilter_init(void)
900 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
901 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
904 static void __exit obdfilter_exit(void)
906 obd_unregister_type(OBD_FILTER_DEVICENAME);
909 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
910 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
911 MODULE_LICENSE("GPL");
913 module_init(obdfilter_init);
914 module_exit(obdfilter_exit);