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 void 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);
72 if (!file || IS_ERR(file)) {
73 CERROR("cannot open O\n");
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");
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");
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 rc = file->f_op->write(file, (char *)&lastino,
109 sizeof(lastino), &off);
110 if (rc != sizeof(lastino)) {
111 CERROR("OBD filter: error writing lastino\n");
115 rc = file->f_op->read(file, (char *)&lastino, sizeof(lastino),
117 if (rc != sizeof(lastino)) {
118 CERROR("OBD filter: error reading lastino\n");
122 obddev->u.filter.fo_lastino = lastino;
129 /* cleanup the filter: write last used object id to status file */
130 static void filter_post(struct obd_device *obddev)
132 struct obd_run_ctxt saved;
137 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
138 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
139 if ( !file || IS_ERR(file)) {
140 CERROR("OBD filter: cannot create status file\n");
143 rc = file->f_op->write(file, (char *)&obddev->u.filter.fo_lastino,
144 sizeof(obddev->u.filter.fo_lastino), &off);
145 if (rc != sizeof(obddev->u.filter.fo_lastino) ) {
146 CERROR("OBD filter: error writing lastino\n");
149 rc = filp_close(file, NULL);
151 CERROR("OBD filter: cannot close status file\n");
158 static __u64 filter_next_id(struct obd_device *obddev)
161 spin_lock(&obddev->u.filter.fo_lock);
162 obddev->u.filter.fo_lastino++;
163 id = obddev->u.filter.fo_lastino;
164 spin_unlock(&obddev->u.filter.fo_lock);
168 /* how to get files, dentries, inodes from object id's */
169 static struct file *filter_obj_open(struct obd_device *obddev,
170 __u64 id, __u32 type)
172 struct obd_run_ctxt saved;
174 struct super_block *sb;
177 sb = obddev->u.filter.fo_sb;
178 if (!sb || !sb->s_dev) {
179 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
185 CDEBUG(D_INODE, "fatal: invalid obdo %Lu\n", id);
190 if ( ! (type & S_IFMT) ) {
191 CERROR("OBD filter_obj_open, no type (%Ld), mode %o!\n",
195 filter_id(name, id, type);
196 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
197 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0);
200 CDEBUG(D_INODE, "opening obdo %s\n", name);
205 static struct file *filter_parent(obd_id id, obd_mode mode)
210 sprintf(path, "O/%s", obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT]);
212 file = filp_open(path, O_RDONLY, 0);
217 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
218 __u64 id, __u32 type)
223 file = filter_obj_open(obddev, id, type);
225 CERROR("filter_inode_from_obdo failed\n");
229 inode = iget(file->f_dentry->d_inode->i_sb,
230 file->f_dentry->d_inode->i_ino);
236 static int filter_connect(struct obd_conn *conn)
241 rc = gen_connect(conn);
249 static int filter_disconnect(struct obd_conn *conn)
253 rc = gen_disconnect(conn);
257 /* XXX cleanup preallocated inodes */
261 /* mount the file system (secretly) */
262 static int filter_setup(struct obd_device *obddev, obd_count len,
266 struct obd_ioctl_data* data = buf;
267 struct vfsmount *mnt;
272 mnt = do_kern_mount(data->ioc_inlbuf2, 0,
273 data->ioc_inlbuf1, NULL);
280 obddev->u.filter.fo_sb = mnt->mnt_root->d_inode->i_sb;
281 if (!obddev->u.filter.fo_sb) {
286 obddev->u.filter.fo_vfsmnt = mnt;
287 obddev->u.filter.fo_fstype = strdup(data->ioc_inlbuf2);
289 obddev->u.filter.fo_ctxt.pwdmnt = mnt;
290 obddev->u.filter.fo_ctxt.pwd = mnt->mnt_root;
291 obddev->u.filter.fo_ctxt.fs = KERNEL_DS;
294 spin_lock_init(&obddev->u.filter.fo_lock);
302 static int filter_cleanup(struct obd_device * obddev)
304 struct super_block *sb;
308 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
313 if ( !list_empty(&obddev->obd_gen_clients) ) {
314 CERROR("still has clients!\n");
319 sb = obddev->u.filter.fo_sb;
320 if (!obddev->u.filter.fo_sb){
327 mntput(obddev->u.filter.fo_vfsmnt);
328 obddev->u.filter.fo_sb = 0;
329 kfree(obddev->u.filter.fo_fstype);
339 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
341 int type = oa->o_mode & S_IFMT;
344 CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
345 inode->i_ino, (long)oa->o_id, oa->o_valid);
346 obdo_from_inode(oa, inode);
347 oa->o_mode &= ~S_IFMT;
350 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
351 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
352 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
354 *((obd_rdev *)oa->o_inline) = rdev;
355 oa->o_obdflags |= OBD_FL_INLINEDATA;
356 oa->o_valid |= OBD_MD_FLINLINE;
360 else if (ext2obd_has_inline(inode)) {
361 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
362 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
363 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
364 oa->o_obdflags |= OBD_FL_INLINEDATA;
365 oa->o_valid |= OBD_MD_FLINLINE;
368 if (ext2obd_has_obdmd(inode)) {
369 /* XXX this will change when we don't store the obdmd in data */
370 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
371 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
372 MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
373 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
374 oa->o_valid |= OBD_MD_FLOBDMD;
380 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
385 if ( !gen_client(conn) ) {
386 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
391 if ( !(inode = filter_inode_from_obj(conn->oc_dev,
392 oa->o_id, oa->o_mode)) ) {
397 oa->o_valid &= ~OBD_MD_FLID;
398 filter_from_inode(oa, inode);
405 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
412 if (!gen_client(conn)) {
413 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
417 inode = filter_inode_from_obj(conn->oc_dev, oa->o_id, oa->o_mode);
423 iattr_from_obdo(&iattr, oa);
424 iattr.ia_mode &= ~S_IFMT;
425 iattr.ia_mode |= S_IFREG;
427 if ( inode->i_op->setattr ) {
428 rc = inode->i_op->setattr(&de, &iattr);
430 rc = inode_setattr(inode, &iattr);
438 static int filter_create (struct obd_conn* conn, struct obdo *oa)
441 struct obd_run_ctxt saved;
444 struct obd_device *obddev = conn->oc_dev;
448 if (!gen_client(conn)) {
449 CERROR("invalid client %u\n", conn->oc_id);
453 oa->o_id = filter_next_id(conn->oc_dev);
454 if ( !(oa->o_mode && S_IFMT) ) {
455 CERROR("filter obd: no type!\n");
459 filter_id(name, oa->o_id, oa->o_mode);
460 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
464 file = filp_open(name, O_RDONLY | O_CREAT, mode);
467 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
472 /* Set flags for fields we have set in ext2_new_inode */
473 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
474 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
475 OBD_MD_FLUID | OBD_MD_FLGID;
479 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
481 struct obd_device * obddev;
482 struct obd_client * cli;
483 struct inode * inode;
487 struct obd_run_ctxt saved;
489 if (!(cli = gen_client(conn))) {
490 CERROR("invalid client %u\n", conn->oc_id);
495 obddev = conn->oc_dev;
496 object = filter_obj_open(obddev, oa->o_id, oa->o_mode);
497 if (!object || IS_ERR(object)) {
502 inode = object->f_dentry->d_inode;
504 inode->i_mode = 010000;
506 push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
507 dir = filter_parent(oa->o_id, oa->o_mode);
514 dget(object->f_dentry);
515 rc = vfs_unlink(dir->f_dentry->d_inode, object->f_dentry);
518 filp_close(object, 0);
525 static int filter_truncate(struct obd_conn *conn, struct obdo *oa, obd_size count,
530 error = filter_setattr(conn, oa);
531 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
537 /* buffer must lie in user memory here */
538 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
539 obd_size *count, obd_off offset)
542 unsigned long retval;
545 if (!gen_client(conn)) {
546 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
551 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
552 if (!file || IS_ERR(file)) {
554 return -PTR_ERR(file);
557 /* count doubles as retval */
558 retval = file->f_op->read(file, buf, *count, &offset);
573 /* buffer must lie in user memory here */
574 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
575 obd_size *count, obd_off offset)
579 unsigned long retval;
582 if (!gen_client(conn)) {
583 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
588 file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
589 if (!file || IS_ERR(file)) {
591 return -PTR_ERR(file);
594 /* count doubles as retval */
595 retval = file->f_op->write(file, buf, *count, &offset);
609 } /* ext2obd_write */
611 static int filter_pgcache_brw(int rw, struct obd_conn *conn,
620 struct super_block *sb;
622 int onum; /* index to oas */
623 int pnum; /* index to pages (bufs) */
624 unsigned long retval;
630 if (!gen_client(conn)) {
631 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
636 sb = conn->oc_dev->u.filter.fo_sb;
640 pnum = 0; /* pnum indexes buf 0..num_pages */
641 for (onum = 0; onum < num_oa; onum++) {
644 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
646 if (!file || IS_ERR(file)) {
652 /* count doubles as retval */
653 for (pg = 0; pg < oa_bufs[onum]; pg++) {
654 CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) off count (%Ld,%Ld)\n",
655 rw, onum, pnum, file->f_dentry->d_inode->i_ino,
656 (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
657 offset[pnum], count[pnum]);
662 buffer = kmap(pages[pnum]);
663 retval = file->f_op->write(file, buffer, count[pnum], &off);
665 CDEBUG(D_INODE, "retval %ld\n", retval);
667 loff_t off = offset[pnum];
668 char *buffer = kmap(pages[pnum]);
670 if (off >= file->f_dentry->d_inode->i_size) {
671 memset(buffer, 0, count[pnum]);
672 retval = count[pnum];
674 retval = file->f_op->read(file, buffer, count[pnum], &off);
678 if ( retval != count[pnum] ) {
684 CDEBUG(D_INODE, "retval %ld\n", retval);
688 /* sizes and blocks are set by generic_file_write */
689 /* ctimes/mtimes will follow with a setattr call */
696 error = (retval >= 0) ? 0 : retval;
701 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
703 struct inode *inode = NULL;
704 struct super_block *sb = conn->oc_dev->u.ext2.e2_sb;
706 if (!sb || !sb->s_dev) {
707 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
713 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
718 inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
719 if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
720 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
721 (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
722 "no links" : "NULL");
732 int filter_preprw(int cmd, struct obd_conn *conn,
733 int objcount, struct obd_ioobj *obj,
734 int niocount, struct niobuf *nb,
746 memset(res, 0, sizeof(*res) * niocount);
748 for (i=0; i < objcount; i++) {
751 for (j = 0 ; j < o->ioo_bufcnt ; j++) {
757 inode = ioobj_to_inode(conn, o);
760 /* FIXME: we need to iput all of previous inodes */
764 page = lustre_get_page(inode, b->offset >> PAGE_SHIFT);
767 return PTR_ERR(page);
769 if (cmd == OBD_BRW_WRITE) {
770 rc = lustre_prepare_page(0, PAGE_SIZE, page);
773 CERROR("i %d j %d objcount %d bufcnt %d , rc %d, offset %Ld\n", i, j, objcount, o->ioo_bufcnt, rc, b->offset);
776 r->addr = (__u64)(unsigned long)page_address(page);
777 r->offset = b->offset;
786 int filter_commitrw(int cmd, struct obd_conn *conn,
787 int objcount, struct obd_ioobj *obj,
788 int niocount, struct niobuf *res)
794 struct niobuf *r = NULL;
797 for (i=0; i < objcount; i++) {
801 for (j = 0 ; j < o->ioo_bufcnt ; j++) {
810 if (cmd == OBD_BRW_WRITE)
811 rc = lustre_commit_page(r->page, 0, PAGE_SIZE);
813 lustre_put_page(r->page);
821 inode = ((struct page *)r->page)->mapping->host;
828 static int filter_statfs (struct obd_conn *conn, struct statfs * statfs)
830 struct super_block *sb;
835 if (!gen_client(conn)) {
836 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
841 sb = conn->oc_dev->u.filter.fo_sb;
843 err = sb->s_op->statfs(sb, statfs);
846 } /* ext2obd_statfs */
849 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
850 void *key, obd_count *vallen, void **val)
852 struct obd_device *obddev;
853 struct obd_client * cli;
856 if (!(cli = gen_client(conn))) {
857 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
861 obddev = conn->oc_dev;
863 if ( keylen == strlen("blocksize") &&
864 memcmp(key, "blocksize", keylen) == 0 ) {
865 *vallen = sizeof(int);
866 *val = (void *)obddev->u.filter.fo_sb->s_blocksize;
871 if ( keylen == strlen("blocksize_bits") &&
872 memcmp(key, "blocksize_bits", keylen) == 0 ){
873 *vallen = sizeof(int);
874 *val = (void *)(int)obddev->u.filter.fo_sb->s_blocksize_bits;
879 if ( keylen == strlen("root_ino") &&
880 memcmp(key, "root_ino", keylen) == 0 ){
881 *vallen = sizeof(int);
882 *val = (void *)(int) FILTER_ROOTINO;
887 CDEBUG(D_IOCTL, "invalid key\n");
892 struct obd_ops filter_obd_ops = {
894 o_get_info: filter_get_info,
895 o_setup: filter_setup,
896 o_cleanup: filter_cleanup,
897 o_connect: filter_connect,
898 o_disconnect: filter_disconnect,
899 o_statfs: filter_statfs,
900 o_getattr: filter_getattr,
901 o_create: filter_create,
902 o_setattr: filter_setattr,
903 o_destroy: filter_destroy,
905 o_write: filter_write,
906 o_brw: filter_pgcache_brw,
907 o_punch: filter_truncate,
908 o_preprw: filter_preprw,
909 o_commitrw: filter_commitrw
911 o_preallocate: ext2obd_preallocate_inodes,
912 o_migrate: ext2obd_migrate,
913 o_copy: gen_copy_data,
914 o_iterate: ext2obd_iterate
919 static int __init obdfilter_init(void)
921 printk(KERN_INFO "Filtering OBD driver v0.001, braam@clusterfs.com\n");
922 return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
925 static void __exit obdfilter_exit(void)
927 obd_unregister_type(OBD_FILTER_DEVICENAME);
930 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
931 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
932 MODULE_LICENSE("GPL");
934 module_init(obdfilter_init);
935 module_exit(obdfilter_exit);