1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (C) 2001, 2002 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>
12 * and Andreas Dilger <adilger@clusterfs.com>
16 #define DEBUG_SUBSYSTEM S_FILTER
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
21 #include <linux/dcache.h>
22 #include <linux/obd_class.h>
23 #include <linux/lustre_dlm.h>
24 #include <linux/obd_filter.h>
25 #include <linux/ext3_jbd.h>
26 #include <linux/extN_jbd.h>
27 #include <linux/quotaops.h>
28 #include <linux/init.h>
29 #include <linux/stringify.h>
33 #define FILTER_ROOTINO 2
34 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
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"
48 static inline const char *obd_mode_to_type(int mode)
50 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
53 /* write the pathname into the string */
54 static int filter_id(char *buf, obd_id id, obd_mode mode)
56 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
59 static inline void f_dput(struct dentry *dentry)
61 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
62 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
63 LASSERT(atomic_read(&dentry->d_count) > 0);
67 /* setup the object store with correct subdirectories */
68 static int filter_prep(struct obd_device *obddev)
70 struct obd_run_ctxt saved;
71 struct filter_obd *filter = &obddev->u.filter;
72 struct dentry *dentry;
80 push_ctxt(&saved, &filter->fo_ctxt, NULL);
81 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
82 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
85 CERROR("cannot open/create O: rc = %d\n", rc);
88 filter->fo_dentry_O = dentry;
89 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
90 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
93 CERROR("cannot open/create P: rc = %d\n", rc);
97 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
98 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
100 rc = PTR_ERR(dentry);
101 CERROR("cannot open/create D: rc = %d\n", rc);
105 root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
109 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
116 * Create directories and/or get dentries for each object type.
117 * This saves us from having to do multiple lookups for each one.
119 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
120 char *type = obd_type_by_mode[mode];
123 filter->fo_dentry_O_mode[mode] = NULL;
126 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
127 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
128 if (IS_ERR(dentry)) {
129 rc = PTR_ERR(dentry);
130 CERROR("cannot create O/%s: rc = %d\n", type, rc);
131 GOTO(out_O_mode, rc);
133 filter->fo_dentry_O_mode[mode] = dentry;
136 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
137 if ( !file || IS_ERR(file) ) {
139 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
141 GOTO(out_O_mode, rc);
144 /* steal operations */
145 inode = file->f_dentry->d_inode;
146 filter->fo_fop = file->f_op;
147 filter->fo_iop = inode->i_op;
148 filter->fo_aops = inode->i_mapping->a_ops;
150 if (inode->i_size == 0) {
151 __u64 disk_lastino = cpu_to_le64(lastino);
152 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
153 sizeof(disk_lastino),
155 if (retval != sizeof(disk_lastino)) {
156 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
158 GOTO(out_O_mode, rc = -EIO);
162 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
163 sizeof(disk_lastino),
165 if (retval != sizeof(disk_lastino)) {
166 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
168 GOTO(out_O_mode, rc = -EIO);
170 lastino = le64_to_cpu(disk_lastino);
172 filter->fo_lastino = lastino;
183 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
186 filter->fo_dentry_O_mode[mode] = NULL;
190 f_dput(filter->fo_dentry_O);
191 filter->fo_dentry_O = NULL;
195 /* cleanup the filter: write last used object id to status file */
196 static void filter_post(struct obd_device *obddev)
198 struct obd_run_ctxt saved;
199 struct filter_obd *filter = &obddev->u.filter;
205 push_ctxt(&saved, &filter->fo_ctxt, NULL);
206 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
208 CERROR("OBD filter: cannot create status file\n");
213 disk_lastino = cpu_to_le64(filter->fo_lastino);
214 rc = file->f_op->write(file, (char *)&disk_lastino,
215 sizeof(disk_lastino), &file->f_pos);
216 if (rc != sizeof(disk_lastino))
217 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
219 rc = filp_close(file, NULL);
221 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
223 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
224 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
227 filter->fo_dentry_O_mode[mode] = NULL;
230 f_dput(filter->fo_dentry_O);
236 static __u64 filter_next_id(struct obd_device *obddev)
240 spin_lock(&obddev->u.filter.fo_lock);
241 id = ++obddev->u.filter.fo_lastino;
242 spin_unlock(&obddev->u.filter.fo_lock);
244 /* FIXME: write the lastino to disk here */
248 /* how to get files, dentries, inodes from object id's */
249 /* parent i_sem is already held if needed for exclusivity */
250 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
251 struct dentry *dparent,
252 __u64 id, __u32 type)
254 struct super_block *sb = obddev->u.filter.fo_sb;
255 struct dentry *dchild;
260 if (!sb || !sb->s_dev) {
261 CERROR("fatal: device not initialized.\n");
262 RETURN(ERR_PTR(-ENXIO));
266 CERROR("fatal: invalid object #0\n");
268 RETURN(ERR_PTR(-ESTALE));
271 if (!(type & S_IFMT)) {
272 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
273 __FUNCTION__, id, type);
274 RETURN(ERR_PTR(-EINVAL));
277 len = sprintf(name, LPU64, id);
278 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
280 dchild = lookup_one_len(name, dparent, len);
281 if (IS_ERR(dchild)) {
282 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
286 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
287 obd_mode_to_type(type), name, dchild,
288 atomic_read(&dchild->d_count));
290 LASSERT(atomic_read(&dchild->d_count) > 0);
295 static struct file *filter_obj_open(struct obd_device *obddev,
296 __u64 id, __u32 type)
298 struct super_block *sb = obddev->u.filter.fo_sb;
299 struct obd_run_ctxt saved;
304 if (!sb || !sb->s_dev) {
305 CERROR("fatal: device not initialized.\n");
306 RETURN(ERR_PTR(-ENXIO));
310 CERROR("fatal: invalid obdo "LPU64"\n", id);
311 RETURN(ERR_PTR(-ESTALE));
314 if (!(type & S_IFMT)) {
315 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
316 __FUNCTION__, id, type);
317 RETURN(ERR_PTR(-EINVAL));
320 filter_id(name, id, type);
321 push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
322 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
325 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
332 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
334 struct filter_obd *filter = &obddev->u.filter;
336 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
341 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
342 obd_uuid_t cluuid, struct recovd_obd *recovd,
343 ptlrpc_recovery_cb_t recover)
348 rc = class_connect(conn, obd, cluuid);
354 static int filter_disconnect(struct lustre_handle *conn)
356 struct obd_export *export = class_conn2export(conn);
360 ldlm_cancel_locks_for_export(export);
362 rc = class_disconnect(conn);
366 /* XXX cleanup preallocated inodes */
370 /* mount the file system (secretly) */
371 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
373 struct obd_ioctl_data* data = buf;
374 struct filter_obd *filter;
375 struct vfsmount *mnt;
379 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
383 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
388 filter = &obddev->u.filter;;
389 filter->fo_vfsmnt = mnt;
390 filter->fo_fstype = strdup(data->ioc_inlbuf2);
391 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
392 /* XXX is this even possible if do_kern_mount succeeded? */
394 GOTO(err_kfree, err = -ENODEV);
396 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
397 filter->fo_ctxt.pwdmnt = mnt;
398 filter->fo_ctxt.pwd = mnt->mnt_root;
399 filter->fo_ctxt.fs = get_ds();
401 err = filter_prep(obddev);
403 GOTO(err_kfree, err);
404 spin_lock_init(&filter->fo_lock);
406 obddev->obd_namespace =
407 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
408 if (obddev->obd_namespace == NULL)
411 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
412 "filter_ldlm_client", &obddev->obd_ldlm_client);
417 kfree(filter->fo_fstype);
419 mntput(filter->fo_vfsmnt);
429 static int filter_cleanup(struct obd_device * obddev)
431 struct super_block *sb;
434 if (!list_empty(&obddev->obd_exports)) {
435 CERROR("still has clients!\n");
436 class_disconnect_all(obddev);
437 if (!list_empty(&obddev->obd_exports)) {
438 CERROR("still has exports after forced cleanup?\n");
443 ldlm_namespace_free(obddev->obd_namespace);
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,
467 int type = oa->o_mode & S_IFMT;
470 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
471 inode->i_ino, inode, (long)oa->o_id, valid);
472 /* Don't copy the inode number in place of the object ID */
473 obdo_from_inode(oa, inode, valid);
474 oa->o_mode &= ~S_IFMT;
477 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
478 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
480 oa->o_valid |= OBD_MD_FLRDEV;
486 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
487 struct lov_stripe_md *md)
489 struct obd_device *obddev = class_conn2obd(conn);
490 struct dentry *dentry;
494 if (!class_conn2export(conn)) {
495 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
499 obddev = class_conn2obd(conn);
500 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
501 oa->o_id, oa->o_mode);
503 RETURN(PTR_ERR(dentry));
505 if (!dentry->d_inode) {
506 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
507 GOTO(out_getattr, rc = -ENOENT);
510 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
517 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
518 struct lov_stripe_md *md)
520 struct obd_run_ctxt saved;
521 struct obd_device *obd = class_conn2obd(conn);
522 struct dentry *dentry;
528 iattr_from_obdo(&iattr, oa, oa->o_valid);
529 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
530 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
531 oa->o_id, iattr.ia_mode);
533 RETURN(PTR_ERR(dentry));
535 inode = dentry->d_inode;
537 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
538 GOTO(out_setattr, rc = -ENOENT);
542 if (iattr.ia_valid & ATTR_SIZE)
544 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
545 if (inode->i_op->setattr)
546 rc = inode->i_op->setattr(dentry, &iattr);
548 rc = inode_setattr(inode, &iattr);
550 if (iattr.ia_valid & ATTR_SIZE) {
552 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
553 obdo_from_inode(oa, inode, oa->o_valid);
562 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
563 struct lov_stripe_md *ea)
565 struct obd_export *export;
566 struct obd_device *obd;
567 struct dentry *dentry;
570 export = class_conn2export(conn);
572 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
576 obd = export->exp_obd;
577 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
578 oa->o_id, oa->o_mode);
580 RETURN(PTR_ERR(dentry));
582 if (!dentry->d_inode) {
583 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
587 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
588 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
593 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
594 struct lov_stripe_md *ea)
596 struct obd_device *obd;
597 struct dentry *dentry;
600 obd = class_conn2obd(conn);
602 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
606 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
607 oa->o_id, oa->o_mode);
609 RETURN(PTR_ERR(dentry));
610 LASSERT(atomic_read(&dentry->d_count) > 1);
612 f_dput(dentry); /* for the close */
613 f_dput(dentry); /* for this call */
617 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
618 struct lov_stripe_md **ea)
621 struct obd_run_ctxt saved;
624 struct obd_device *obd = class_conn2obd(conn);
625 struct filter_obd *filter = &obd->u.filter;
630 CERROR("invalid client "LPX64"\n", conn->addr);
634 if (!(oa->o_mode && S_IFMT)) {
635 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
636 __FUNCTION__, oa->o_id, oa->o_mode);
640 oa->o_id = filter_next_id(obd);
642 //filter_id(name, oa->o_id, oa->o_mode);
643 sprintf(name, LPU64, oa->o_id);
644 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
645 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
646 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
650 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
654 /* Set flags for fields we have set in the inode struct */
655 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
656 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
657 filter_from_inode(oa, new->d_inode, oa->o_valid);
663 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
664 struct lov_stripe_md *ea)
666 struct obd_device *obd;
667 struct filter_obd *filter;
668 struct obd_run_ctxt saved;
670 struct dentry *dir_dentry, *object_dentry;
674 obd = class_conn2obd(conn);
676 CERROR("invalid client "LPX64"\n", conn->addr);
680 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
682 dir_dentry = filter_parent(obd, oa->o_mode);
683 down(&dir_dentry->d_inode->i_sem);
685 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
687 if (IS_ERR(object_dentry))
688 GOTO(out, rc = -ENOENT);
690 inode = object_dentry->d_inode;
692 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
693 GOTO(out, rc = -ENOENT);
696 if (inode->i_nlink != 1) {
697 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
701 inode->i_mode = S_IFREG;
703 if (atomic_read(&inode->i_count) > 1) {
704 #warning FIXME: need to handle open-unlinked case and move to hold dir
705 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
708 filter = &obd->u.filter;
709 push_ctxt(&saved, &filter->fo_ctxt, NULL);
711 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
713 f_dput(object_dentry);
717 up(&dir_dentry->d_inode->i_sem);
721 /* NB count and offset are used for punch, but not truncate */
722 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
723 struct lov_stripe_md *md,
724 obd_off start, obd_off end)
729 if (end != OBD_OBJECT_EOF)
730 CERROR("PUNCH not supported, only truncate works\n");
732 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
733 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
735 error = filter_setattr(conn, oa, NULL);
739 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
740 struct lov_stripe_md *lsm, obd_count oa_bufs,
741 struct brw_page *pga, brw_callback_t callback,
742 struct io_cb_data *data)
744 struct obd_run_ctxt saved;
745 struct super_block *sb;
746 int pnum; /* index to pages (bufs) */
747 unsigned long retval;
750 struct obd_device *obd = class_conn2obd(conn);
755 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
759 sb = obd->u.filter.fo_sb;
760 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
761 pnum = 0; /* pnum indexes buf 0..num_pages */
763 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
765 GOTO(out, retval = PTR_ERR(file));
767 /* count doubles as retval */
768 for (pg = 0; pg < oa_bufs; pg++) {
769 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
770 ") off count ("LPU64",%d)\n",
771 cmd, pnum, file->f_dentry->d_inode->i_ino,
772 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
773 (int)pga[pnum].count);
774 if (cmd & OBD_BRW_WRITE) {
778 buffer = kmap(pga[pnum].pg);
779 retval = file->f_op->write(file, buffer,
782 kunmap(pga[pnum].pg);
783 CDEBUG(D_INODE, "retval %ld\n", retval);
785 loff_t off = pga[pnum].off;
786 char *buffer = kmap(pga[pnum].pg);
788 if (off >= file->f_dentry->d_inode->i_size) {
789 memset(buffer, 0, pga[pnum].count);
790 retval = pga[pnum].count;
792 retval = file->f_op->read(file, buffer,
793 pga[pnum].count, &off);
795 kunmap(pga[pnum].pg);
797 if (retval != pga[pnum].count) {
799 GOTO(out, retval = -EIO);
801 CDEBUG(D_INODE, "retval %ld\n", retval);
805 /* sizes and blocks are set by generic_file_write */
806 /* ctimes/mtimes will follow with a setattr call */
809 /* XXX: do something with callback if it is set? */
814 error = (retval >= 0) ? 0 : retval;
819 * Calculate the number of buffer credits needed to write multiple pages in
820 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
821 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
823 * See comment above ext3_writepage_trans_blocks for details. We assume
824 * no data journaling is being done, but it does allow for all of the pages
825 * being non-contiguous. If we are guaranteed contiguous pages we could
826 * reduce the number of (d)indirect blocks a lot.
828 * With N blocks per page and P pages, for each inode we have at most:
830 * min(N*P, blocksize/4 + 1) dindirect blocks
833 * For the entire filesystem, we have at most:
834 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
835 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
838 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
840 static int ext3_credits_needed(struct super_block *sb, int objcount,
841 struct obd_ioobj *obj)
843 struct obd_ioobj *o = obj;
844 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
845 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
848 int needed = objcount + 1;
851 for (i = 0; i < objcount; i++, o++) {
852 int nblocks = o->ioo_bufcnt * blockpp;
853 int ndindirect = min(nblocks, addrpp + 1);
854 int nindir = nblocks + ndindirect + 1;
856 nbitmaps += nindir + nblocks;
857 ngdblocks += nindir + nblocks;
862 /* Assumes ext3 and extN have same sb_info layout at the start. */
863 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
864 nbitmaps = EXT3_SB(sb)->s_groups_count;
865 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
866 ngdblocks = EXT3_SB(sb)->s_gdb_count;
868 needed += nbitmaps + ngdblocks;
871 /* We assume that there will be 1 bit set in s_dquot.flags for each
872 * quota file that is active. This is at least true for now.
874 needed += hweight32(sb_any_quota_enabled(sb)) *
875 EXT3_SINGLEDATA_TRANS_BLOCKS;
881 /* We have to start a huge journal transaction here to hold all of the
882 * metadata for the pages being written here. This is necessitated by
883 * the fact that we do lots of prepare_write operations before we do
884 * any of the matching commit_write operations, so even if we split
885 * up to use "smaller" transactions none of them could complete until
886 * all of them were opened. By having a single journal transaction,
887 * we eliminate duplicate reservations for common blocks like the
888 * superblock and group descriptors or bitmaps.
890 * We will start the transaction here, but each prepare_write will
891 * add a refcount to the transaction, and each commit_write will
892 * remove a refcount. The transaction will be closed when all of
893 * the pages have been written.
895 static void *ext3_filter_journal_start(struct filter_obd *filter,
896 int objcount, struct obd_ioobj *obj,
897 int niocount, struct niobuf_remote *nb)
899 journal_t *journal = NULL;
900 handle_t *handle = NULL;
903 /* It appears that some kernels have different values for
904 * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
905 * assume anything after s_inode_bitmap_number is the same.
907 if (!strcmp(filter->fo_fstype, "ext3"))
908 journal = EXT3_SB(filter->fo_sb)->s_journal;
909 else if (!strcmp(filter->fo_fstype, "extN"))
910 journal = EXTN_SB(filter->fo_sb)->s_journal;
911 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
913 /* The number of blocks we could _possibly_ dirty can very large.
914 * We reduce our request if it is absurd (and we couldn't get that
915 * many credits for a single handle anyways).
917 * At some point we have to limit the size of I/Os sent at one time,
918 * increase the size of the journal, or we have to calculate the
919 * actual journal requirements more carefully by checking all of
920 * the blocks instead of being maximally pessimistic. It remains to
921 * be seen if this is a real problem or not.
923 if (needed > journal->j_max_transaction_buffers) {
924 CERROR("want too many journal credits (%d) using %d instead\n",
925 needed, journal->j_max_transaction_buffers);
926 needed = journal->j_max_transaction_buffers;
930 handle = journal_start(journal, needed);
933 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
939 static void *filter_journal_start(void **journal_save,
940 struct filter_obd *filter,
941 int objcount, struct obd_ioobj *obj,
942 int niocount, struct niobuf_remote *nb)
946 /* This may not be necessary - we probably never have a
947 * transaction started when we enter here, so we can
948 * remove the saving of the journal state entirely.
949 * For now leave it in just to see if it ever happens.
951 *journal_save = current->journal_info;
953 CERROR("Already have handle %p???\n", *journal_save);
955 current->journal_info = NULL;
958 if (!strcmp(filter->fo_fstype, "ext3") ||
959 !strcmp(filter->fo_fstype, "extN"))
960 handle = ext3_filter_journal_start(filter, objcount, obj,
965 static int ext3_filter_journal_stop(void *handle)
969 /* We got a refcount on the handle for each call to prepare_write,
970 * so we can drop the "parent" handle here to avoid the need for
971 * osc to call back into filterobd to close the handle. The
972 * remaining references will be dropped in commit_write.
975 rc = journal_stop((handle_t *)handle);
981 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
986 if (!strcmp(filter->fo_fstype, "ext3") ||
987 !strcmp(filter->fo_fstype, "extN"))
988 rc = ext3_filter_journal_stop(handle);
991 CERROR("error on journal stop: rc = %d\n", rc);
993 current->journal_info = journal_save;
998 static inline void lustre_put_page(struct page *page)
1001 page_cache_release(page);
1005 #ifndef PageUptodate
1006 #define PageUptodate(page) Page_Uptodate(page)
1008 static struct page *
1009 lustre_get_page_read(struct inode *inode, unsigned long index)
1011 struct address_space *mapping = inode->i_mapping;
1015 page = read_cache_page(mapping, index,
1016 (filler_t*)mapping->a_ops->readpage, NULL);
1017 if (!IS_ERR(page)) {
1020 if (!PageUptodate(page)) {
1021 CERROR("page index %lu not uptodate\n", index);
1022 GOTO(err_page, rc = -EIO);
1024 if (PageError(page)) {
1025 CERROR("page index %lu has error\n", index);
1026 GOTO(err_page, rc = -EIO);
1032 lustre_put_page(page);
1036 static struct page *
1037 lustre_get_page_write(struct inode *inode, unsigned long index)
1039 struct address_space *mapping = inode->i_mapping;
1043 page = grab_cache_page(mapping, index); /* locked page */
1045 if (!IS_ERR(page)) {
1047 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1048 * a no-op for most filesystems, because we write the whole
1049 * page. For partial-page I/O this will read in the page.
1051 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1053 CERROR("page index %lu, rc = %d\n", index, rc);
1056 GOTO(err_unlock, rc);
1058 /* XXX not sure if we need this if we are overwriting page */
1059 if (PageError(page)) {
1060 CERROR("error on page index %lu, rc = %d\n", index, rc);
1062 GOTO(err_unlock, rc = -EIO);
1069 lustre_put_page(page);
1073 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1075 struct inode *inode = page->mapping->host;
1078 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1079 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1080 if (!err && IS_SYNC(inode))
1081 err = waitfor_one_page(page);
1083 #warn ADD 2.5 waiting code here?
1085 //SetPageUptodate(page); // the client commit_write will do this
1087 SetPageReferenced(page);
1089 lustre_put_page(page);
1093 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1094 struct niobuf_local *lnb, int *pglocked)
1096 struct address_space *mapping = inode->i_mapping;
1100 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1102 page = grab_cache_page_nowait(mapping, index); /* locked page */
1104 page = grab_cache_page(mapping, index); /* locked page */
1107 /* This page is currently locked, so get a temporary page instead. */
1108 /* XXX I believe this is a very dangerous thing to do - consider if
1109 * we had multiple writers for the same file (definitely the case
1110 * if we are using this codepath). If writer A locks the page,
1111 * writer B writes to a copy (as here), writer A drops the page
1112 * lock, and writer C grabs the lock before B does, then B will
1113 * later overwrite the data from C, even if C had LDLM locked
1114 * and initiated the write after B did.
1118 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1119 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1121 CERROR("no memory for a temp page\n");
1123 GOTO(err, rc = -ENOMEM);
1125 page = virt_to_page(addr);
1127 page->index = index;
1128 lnb->flags |= N_LOCAL_TEMP_PAGE;
1129 } else if (!IS_ERR(page)) {
1133 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1134 * a no-op for most filesystems, because we write the whole
1135 * page. For partial-page I/O this will read in the page.
1137 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1139 CERROR("page index %lu, rc = %d\n", index, rc);
1142 GOTO(err_unlock, rc);
1144 /* XXX not sure if we need this if we are overwriting page */
1145 if (PageError(page)) {
1146 CERROR("error on page index %lu, rc = %d\n", index, rc);
1148 GOTO(err_unlock, rc = -EIO);
1155 lustre_put_page(page);
1161 * We need to balance prepare_write() calls with commit_write() calls.
1162 * If the page has been prepared, but we have no data for it, we don't
1163 * want to overwrite valid data on disk, but we still need to zero out
1164 * data for space which was newly allocated. Like part of what happens
1165 * in __block_prepare_write() for newly allocated blocks.
1167 * XXX currently __block_prepare_write() creates buffers for all the
1168 * pages, and the filesystems mark these buffers as BH_New if they
1169 * were newly allocated from disk. We use the BH_New flag similarly.
1171 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1175 unsigned block_start, block_end;
1176 struct buffer_head *bh, *head = page->buffers;
1177 unsigned blocksize = head->b_size;
1178 void *addr = page_address(page);
1180 /* debugging: just seeing if this ever happens */
1181 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1182 page->index, page->mapping->host->i_ino, err);
1184 /* Currently one buffer per page, but in the future... */
1185 for (bh = head, block_start = 0; bh != head || !block_start;
1186 block_start = block_end, bh = bh->b_this_page) {
1187 block_end = block_start + blocksize;
1189 memset(addr + block_start, 0, blocksize);
1193 return lustre_commit_write(page, from, to);
1196 static int filter_preprw(int cmd, struct lustre_handle *conn,
1197 int objcount, struct obd_ioobj *obj,
1198 int niocount, struct niobuf_remote *nb,
1199 struct niobuf_local *res, void **desc_private)
1201 struct obd_run_ctxt saved;
1202 struct obd_device *obd;
1203 struct obd_ioobj *o = obj;
1204 struct niobuf_remote *b = nb;
1205 struct niobuf_local *r = res;
1206 void *journal_save = NULL;
1212 obd = class_conn2obd(conn);
1214 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1217 memset(res, 0, sizeof(*res) * niocount);
1219 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1221 if (cmd & OBD_BRW_WRITE) {
1222 *desc_private = filter_journal_start(&journal_save,
1224 objcount, obj, niocount,
1226 if (IS_ERR(*desc_private))
1227 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1230 for (i = 0; i < objcount; i++, o++) {
1231 struct dentry *dentry;
1232 struct inode *inode;
1235 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1236 o->ioo_id, S_IFREG);
1238 GOTO(out_clean, rc = PTR_ERR(dentry));
1239 inode = dentry->d_inode;
1241 CERROR("trying to BRW to non-existent file "LPU64"\n",
1244 GOTO(out_clean, rc = -ENOENT);
1247 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1248 unsigned long index = b->offset >> PAGE_SHIFT;
1254 r->dentry = dget(dentry);
1256 if (cmd & OBD_BRW_WRITE)
1257 page = filter_get_page_write(inode, index, r,
1260 page = lustre_get_page_read(inode, index);
1264 GOTO(out_clean, rc = PTR_ERR(page));
1267 r->addr = page_address(page);
1268 r->offset = b->offset;
1275 if (cmd & OBD_BRW_WRITE) {
1276 int err = filter_journal_stop(journal_save, &obd->u.filter,
1286 CERROR("error cleanup on brw\n");
1288 if (cmd & OBD_BRW_WRITE)
1289 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1291 lustre_put_page(r->page);
1296 static int filter_write_locked_page(struct niobuf_local *lnb)
1301 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1302 if (IS_ERR(lpage)) {
1303 /* It is highly unlikely that we would ever get an error here.
1304 * The page we want to get was previously locked, so it had to
1305 * have already allocated the space, and we were just writing
1306 * over the same data, so there would be no hole in the file.
1308 * XXX: possibility of a race with truncate could exist, need
1309 * to check that. There are no guarantees w.r.t.
1310 * write order even on a local filesystem, although the
1311 * normal response would be to return the number of bytes
1312 * successfully written and leave the rest to the app.
1314 rc = PTR_ERR(lpage);
1315 CERROR("error getting locked page index %ld: rc = %d\n",
1316 lnb->page->index, rc);
1320 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1321 * lustre_commit_write() below, lnb->page was kmapped previously in
1322 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1324 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1325 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1327 CERROR("error committing locked page %ld: rc = %d\n",
1328 lnb->page->index, rc);
1330 lustre_put_page(lnb->page);
1335 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1336 int objcount, struct obd_ioobj *obj,
1337 int niocount, struct niobuf_local *res,
1340 struct obd_run_ctxt saved;
1341 struct obd_ioobj *o;
1342 struct niobuf_local *r;
1343 struct obd_device *obd = class_conn2obd(conn);
1345 int found_locked = 0;
1350 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1351 journal_save = current->journal_info;
1352 LASSERT(!journal_save);
1354 current->journal_info = private;
1355 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1357 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1358 struct page *page = r->page;
1363 if (r->flags & N_LOCAL_TEMP_PAGE) {
1368 if (cmd & OBD_BRW_WRITE) {
1369 int err = filter_commit_write(page, 0,
1375 lustre_put_page(page);
1380 current->journal_info = journal_save;
1385 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1387 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1389 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1392 err = filter_write_locked_page(r);
1404 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1406 struct obd_device *obd = class_conn2obd(conn);
1411 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1413 statfs_pack(osfs, &sfs);
1418 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1419 void *key, obd_count *vallen, void **val)
1421 struct obd_device *obd;
1424 obd = class_conn2obd(conn);
1426 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1430 if ( keylen == strlen("blocksize") &&
1431 memcmp(key, "blocksize", keylen) == 0 ) {
1432 *vallen = sizeof(long);
1433 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1437 if ( keylen == strlen("blocksize_bits") &&
1438 memcmp(key, "blocksize_bits", keylen) == 0 ){
1439 *vallen = sizeof(long);
1440 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1444 if ( keylen == strlen("root_ino") &&
1445 memcmp(key, "root_ino", keylen) == 0 ){
1446 *vallen = sizeof(obd_id);
1447 *val = (void *)(obd_id)FILTER_ROOTINO;
1451 CDEBUG(D_IOCTL, "invalid key\n");
1455 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1456 struct lustre_handle *src_conn, struct obdo *src,
1457 obd_size count, obd_off offset)
1460 struct lov_stripe_md srcmd, dstmd;
1461 unsigned long index = 0;
1464 memset(&srcmd, 0, sizeof(srcmd));
1465 memset(&dstmd, 0, sizeof(dstmd));
1466 srcmd.lsm_object_id = src->o_id;
1467 dstmd.lsm_object_id = dst->o_id;
1470 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1471 ", dst: ino "LPU64"\n",
1472 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1473 page = alloc_page(GFP_USER);
1477 while (TryLockPage(page))
1478 ___wait_on_page(page);
1480 /* XXX with brw vector I/O, we could batch up reads and writes here,
1481 * all we need to do is allocate multiple pages to handle the I/Os
1482 * and arrays to handle the request parameters.
1484 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1486 struct io_cb_data *cbd = ll_init_cb();
1495 pg.count = PAGE_SIZE;
1496 pg.off = (page->index) << PAGE_SHIFT;
1499 page->index = index;
1500 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1501 ll_sync_io_cb, cbd);
1514 pg.flag = OBD_BRW_CREATE;
1515 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1517 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1518 ll_sync_io_cb, cbd);
1520 /* XXX should handle dst->o_size, dst->o_blocks here */
1526 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1530 dst->o_size = src->o_size;
1531 dst->o_blocks = src->o_blocks;
1532 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1540 static struct obd_ops filter_obd_ops = {
1541 o_get_info: filter_get_info,
1542 o_setup: filter_setup,
1543 o_cleanup: filter_cleanup,
1544 o_connect: filter_connect,
1545 o_disconnect: filter_disconnect,
1546 o_statfs: filter_statfs,
1547 o_getattr: filter_getattr,
1548 o_create: filter_create,
1549 o_setattr: filter_setattr,
1550 o_destroy: filter_destroy,
1551 o_open: filter_open,
1552 o_close: filter_close,
1553 o_brw: filter_pgcache_brw,
1554 o_punch: filter_truncate,
1555 o_preprw: filter_preprw,
1556 o_commitrw: filter_commitrw
1558 o_preallocate: filter_preallocate_inodes,
1559 o_migrate: filter_migrate,
1560 o_copy: filter_copy_data,
1561 o_iterate: filter_iterate
1566 static int __init obdfilter_init(void)
1568 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1569 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1572 static void __exit obdfilter_exit(void)
1574 class_unregister_type(OBD_FILTER_DEVICENAME);
1577 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1578 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1579 MODULE_LICENSE("GPL");
1581 module_init(obdfilter_init);
1582 module_exit(obdfilter_exit);