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>
20 #include <linux/obd_class.h>
21 #include <linux/lustre_dlm.h>
22 #include <linux/obd_filter.h>
23 #include <linux/ext3_jbd.h>
24 #include <linux/extN_jbd.h>
25 #include <linux/quotaops.h>
26 #include <linux/init.h>
27 #include <linux/stringify.h>
31 #define FILTER_ROOTINO 2
32 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
35 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
37 [S_IFREG >> S_SHIFT] "R",
38 [S_IFDIR >> S_SHIFT] "D",
39 [S_IFCHR >> S_SHIFT] "C",
40 [S_IFBLK >> S_SHIFT] "B",
41 [S_IFIFO >> S_SHIFT] "F",
42 [S_IFSOCK >> S_SHIFT] "S",
43 [S_IFLNK >> S_SHIFT] "L"
46 static inline const char *obd_mode_to_type(int mode)
48 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
51 /* write the pathname into the string */
52 static int filter_id(char *buf, obd_id id, obd_mode mode)
54 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
57 static inline void f_dput(struct dentry *dentry)
59 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
60 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
61 LASSERT(atomic_read(&dentry->d_count) > 0);
65 /* setup the object store with correct subdirectories */
66 static int filter_prep(struct obd_device *obddev)
68 struct obd_run_ctxt saved;
69 struct filter_obd *filter = &obddev->u.filter;
70 struct dentry *dentry;
78 push_ctxt(&saved, &filter->fo_ctxt, NULL);
79 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
80 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
83 CERROR("cannot open/create O: rc = %d\n", rc);
86 filter->fo_dentry_O = dentry;
87 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
88 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
91 CERROR("cannot open/create P: rc = %d\n", rc);
95 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
96 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
99 CERROR("cannot open/create D: rc = %d\n", rc);
103 root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
107 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
114 * Create directories and/or get dentries for each object type.
115 * This saves us from having to do multiple lookups for each one.
117 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
118 char *type = obd_type_by_mode[mode];
121 filter->fo_dentry_O_mode[mode] = NULL;
124 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
125 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
126 if (IS_ERR(dentry)) {
127 rc = PTR_ERR(dentry);
128 CERROR("cannot create O/%s: rc = %d\n", type, rc);
129 GOTO(out_O_mode, rc);
131 filter->fo_dentry_O_mode[mode] = dentry;
134 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
135 if ( !file || IS_ERR(file) ) {
137 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
139 GOTO(out_O_mode, rc);
142 /* steal operations */
143 inode = file->f_dentry->d_inode;
144 filter->fo_fop = file->f_op;
145 filter->fo_iop = inode->i_op;
146 filter->fo_aops = inode->i_mapping->a_ops;
148 if (inode->i_size == 0) {
149 __u64 disk_lastino = cpu_to_le64(lastino);
150 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
151 sizeof(disk_lastino),
153 if (retval != sizeof(disk_lastino)) {
154 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
156 GOTO(out_O_mode, rc = -EIO);
160 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
161 sizeof(disk_lastino),
163 if (retval != sizeof(disk_lastino)) {
164 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
166 GOTO(out_O_mode, rc = -EIO);
168 lastino = le64_to_cpu(disk_lastino);
170 filter->fo_lastino = lastino;
181 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
184 filter->fo_dentry_O_mode[mode] = NULL;
188 f_dput(filter->fo_dentry_O);
189 filter->fo_dentry_O = NULL;
193 /* cleanup the filter: write last used object id to status file */
194 static void filter_post(struct obd_device *obddev)
196 struct obd_run_ctxt saved;
197 struct filter_obd *filter = &obddev->u.filter;
203 push_ctxt(&saved, &filter->fo_ctxt, NULL);
204 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
206 CERROR("OBD filter: cannot create status file\n");
211 disk_lastino = cpu_to_le64(filter->fo_lastino);
212 rc = file->f_op->write(file, (char *)&disk_lastino,
213 sizeof(disk_lastino), &file->f_pos);
214 if (rc != sizeof(disk_lastino))
215 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
217 rc = filp_close(file, NULL);
219 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
221 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
222 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
225 filter->fo_dentry_O_mode[mode] = NULL;
228 f_dput(filter->fo_dentry_O);
234 static __u64 filter_next_id(struct obd_device *obddev)
238 spin_lock(&obddev->u.filter.fo_lock);
239 id = ++obddev->u.filter.fo_lastino;
240 spin_unlock(&obddev->u.filter.fo_lock);
242 /* FIXME: write the lastino to disk here */
246 /* how to get files, dentries, inodes from object id's */
247 /* parent i_sem is already held if needed for exclusivity */
248 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
249 struct dentry *dparent,
250 __u64 id, __u32 type)
252 struct super_block *sb = obddev->u.filter.fo_sb;
253 struct dentry *dchild;
258 if (!sb || !sb->s_dev) {
259 CERROR("fatal: device not initialized.\n");
260 RETURN(ERR_PTR(-ENXIO));
264 CERROR("fatal: invalid object #0\n");
266 RETURN(ERR_PTR(-ESTALE));
269 if (!(type & S_IFMT)) {
270 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
271 __FUNCTION__, id, type);
272 RETURN(ERR_PTR(-EINVAL));
275 len = sprintf(name, LPU64, id);
276 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
278 dchild = lookup_one_len(name, dparent, len);
279 if (IS_ERR(dchild)) {
280 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
284 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
285 obd_mode_to_type(type), name, dchild,
286 atomic_read(&dchild->d_count));
288 LASSERT(atomic_read(&dchild->d_count) > 0);
293 static struct file *filter_obj_open(struct obd_device *obddev,
294 __u64 id, __u32 type)
296 struct super_block *sb = obddev->u.filter.fo_sb;
297 struct obd_run_ctxt saved;
302 if (!sb || !sb->s_dev) {
303 CERROR("fatal: device not initialized.\n");
304 RETURN(ERR_PTR(-ENXIO));
308 CERROR("fatal: invalid obdo "LPU64"\n", id);
309 RETURN(ERR_PTR(-ESTALE));
312 if (!(type & S_IFMT)) {
313 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
314 __FUNCTION__, id, type);
315 RETURN(ERR_PTR(-EINVAL));
318 filter_id(name, id, type);
319 push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
320 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
323 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
330 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
332 struct filter_obd *filter = &obddev->u.filter;
334 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
339 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
345 rc = class_connect(conn, obd, cluuid);
351 static int filter_disconnect(struct lustre_handle *conn)
356 rc = class_disconnect(conn);
360 /* XXX cleanup preallocated inodes */
364 /* mount the file system (secretly) */
365 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
367 struct obd_ioctl_data* data = buf;
368 struct filter_obd *filter;
369 struct vfsmount *mnt;
373 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
377 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
382 filter = &obddev->u.filter;;
383 filter->fo_vfsmnt = mnt;
384 filter->fo_fstype = strdup(data->ioc_inlbuf2);
385 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
386 /* XXX is this even possible if do_kern_mount succeeded? */
388 GOTO(err_kfree, err = -ENODEV);
390 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
391 filter->fo_ctxt.pwdmnt = mnt;
392 filter->fo_ctxt.pwd = mnt->mnt_root;
393 filter->fo_ctxt.fs = get_ds();
395 err = filter_prep(obddev);
397 GOTO(err_kfree, err);
398 spin_lock_init(&filter->fo_lock);
400 obddev->obd_namespace =
401 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
402 if (obddev->obd_namespace == NULL)
405 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
406 "filter_ldlm_client", &obddev->obd_ldlm_client);
411 kfree(filter->fo_fstype);
413 mntput(filter->fo_vfsmnt);
423 static int filter_cleanup(struct obd_device * obddev)
425 struct super_block *sb;
428 if (!list_empty(&obddev->obd_exports)) {
429 CERROR("still has clients!\n");
430 class_disconnect_all(obddev);
431 if (!list_empty(&obddev->obd_exports)) {
432 CERROR("still has exports after forced cleanup?\n");
437 ldlm_namespace_free(obddev->obd_namespace);
439 sb = obddev->u.filter.fo_sb;
440 if (!obddev->u.filter.fo_sb)
445 shrink_dcache_parent(sb->s_root);
447 mntput(obddev->u.filter.fo_vfsmnt);
448 obddev->u.filter.fo_sb = 0;
449 kfree(obddev->u.filter.fo_fstype);
458 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
461 int type = oa->o_mode & S_IFMT;
464 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
465 inode->i_ino, inode, (long)oa->o_id, valid);
466 /* Don't copy the inode number in place of the object ID */
467 obdo_from_inode(oa, inode, valid);
468 oa->o_mode &= ~S_IFMT;
471 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
472 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
474 oa->o_valid |= OBD_MD_FLRDEV;
480 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
481 struct lov_stripe_md *md)
483 struct obd_device *obddev = class_conn2obd(conn);
484 struct dentry *dentry;
488 if (!class_conn2export(conn)) {
489 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
493 obddev = class_conn2obd(conn);
494 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
495 oa->o_id, oa->o_mode);
497 RETURN(PTR_ERR(dentry));
499 if (!dentry->d_inode) {
500 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
501 GOTO(out_getattr, rc = -ENOENT);
504 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
511 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
512 struct lov_stripe_md *md)
514 struct obd_run_ctxt saved;
515 struct obd_device *obd = class_conn2obd(conn);
516 struct dentry *dentry;
522 iattr_from_obdo(&iattr, oa, oa->o_valid);
523 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
524 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
525 oa->o_id, iattr.ia_mode);
527 RETURN(PTR_ERR(dentry));
529 inode = dentry->d_inode;
531 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
532 GOTO(out_setattr, rc = -ENOENT);
536 if (iattr.ia_valid & ATTR_SIZE)
538 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
539 if (inode->i_op->setattr)
540 rc = inode->i_op->setattr(dentry, &iattr);
542 rc = inode_setattr(inode, &iattr);
544 if (iattr.ia_valid & ATTR_SIZE) {
546 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
547 obdo_from_inode(oa, inode, oa->o_valid);
556 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
557 struct lov_stripe_md *ea)
559 struct obd_export *export;
560 struct obd_device *obd;
561 struct dentry *dentry;
564 export = class_conn2export(conn);
566 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
570 obd = export->exp_obd;
571 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
572 oa->o_id, oa->o_mode);
574 RETURN(PTR_ERR(dentry));
576 if (!dentry->d_inode) {
577 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
581 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
582 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
587 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
588 struct lov_stripe_md *ea)
590 struct obd_device *obd;
591 struct dentry *dentry;
594 obd = class_conn2obd(conn);
596 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
600 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
601 oa->o_id, oa->o_mode);
603 RETURN(PTR_ERR(dentry));
604 LASSERT(atomic_read(&dentry->d_count) > 1);
606 f_dput(dentry); /* for the close */
607 f_dput(dentry); /* for this call */
611 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
612 struct lov_stripe_md **ea)
615 struct obd_run_ctxt saved;
618 struct obd_device *obd = class_conn2obd(conn);
619 struct filter_obd *filter = &obd->u.filter;
624 CERROR("invalid client "LPX64"\n", conn->addr);
628 if (!(oa->o_mode && S_IFMT)) {
629 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
630 __FUNCTION__, oa->o_id, oa->o_mode);
634 oa->o_id = filter_next_id(obd);
636 //filter_id(name, oa->o_id, oa->o_mode);
637 sprintf(name, LPU64, oa->o_id);
638 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
639 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
640 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
644 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
648 /* Set flags for fields we have set in the inode struct */
649 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
650 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
651 filter_from_inode(oa, new->d_inode, oa->o_valid);
657 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
658 struct lov_stripe_md *ea)
660 struct obd_device *obd;
661 struct filter_obd *filter;
662 struct obd_run_ctxt saved;
664 struct dentry *dir_dentry, *object_dentry;
668 obd = class_conn2obd(conn);
670 CERROR("invalid client "LPX64"\n", conn->addr);
674 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
676 dir_dentry = filter_parent(obd, oa->o_mode);
677 down(&dir_dentry->d_inode->i_sem);
679 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
681 if (IS_ERR(object_dentry))
682 GOTO(out, rc = -ENOENT);
684 inode = object_dentry->d_inode;
686 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
687 GOTO(out, rc = -ENOENT);
690 if (inode->i_nlink != 1) {
691 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
695 inode->i_mode = S_IFREG;
697 if (atomic_read(&inode->i_count) > 1) {
698 #warning FIXME: need to handle open-unlinked case and move to hold dir
699 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
702 filter = &obd->u.filter;
703 push_ctxt(&saved, &filter->fo_ctxt, NULL);
705 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
707 f_dput(object_dentry);
711 up(&dir_dentry->d_inode->i_sem);
715 /* NB count and offset are used for punch, but not truncate */
716 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
717 struct lov_stripe_md *md,
718 obd_off start, obd_off end)
723 if (end != OBD_OBJECT_EOF)
724 CERROR("PUNCH not supported, only truncate works\n");
726 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
727 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
729 error = filter_setattr(conn, oa, NULL);
733 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
734 struct lov_stripe_md *lsm, obd_count oa_bufs,
735 struct brw_page *pga, brw_callback_t callback,
736 struct io_cb_data *data)
738 struct obd_run_ctxt saved;
739 struct super_block *sb;
740 int pnum; /* index to pages (bufs) */
741 unsigned long retval;
744 struct obd_device *obd = class_conn2obd(conn);
749 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
753 sb = obd->u.filter.fo_sb;
754 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
755 pnum = 0; /* pnum indexes buf 0..num_pages */
757 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
759 GOTO(out, retval = PTR_ERR(file));
761 /* count doubles as retval */
762 for (pg = 0; pg < oa_bufs; pg++) {
763 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
764 ") off count ("LPU64",%d)\n",
765 cmd, pnum, file->f_dentry->d_inode->i_ino,
766 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
767 (int)pga[pnum].count);
768 if (cmd & OBD_BRW_WRITE) {
772 buffer = kmap(pga[pnum].pg);
773 retval = file->f_op->write(file, buffer,
776 kunmap(pga[pnum].pg);
777 CDEBUG(D_INODE, "retval %ld\n", retval);
779 loff_t off = pga[pnum].off;
780 char *buffer = kmap(pga[pnum].pg);
782 if (off >= file->f_dentry->d_inode->i_size) {
783 memset(buffer, 0, pga[pnum].count);
784 retval = pga[pnum].count;
786 retval = file->f_op->read(file, buffer,
787 pga[pnum].count, &off);
789 kunmap(pga[pnum].pg);
791 if (retval != pga[pnum].count) {
793 GOTO(out, retval = -EIO);
795 CDEBUG(D_INODE, "retval %ld\n", retval);
799 /* sizes and blocks are set by generic_file_write */
800 /* ctimes/mtimes will follow with a setattr call */
803 /* XXX: do something with callback if it is set? */
808 error = (retval >= 0) ? 0 : retval;
813 * Calculate the number of buffer credits needed to write multiple pages in
814 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
815 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
817 * See comment above ext3_writepage_trans_blocks for details. We assume
818 * no data journaling is being done, but it does allow for all of the pages
819 * being non-contiguous. If we are guaranteed contiguous pages we could
820 * reduce the number of (d)indirect blocks a lot.
822 * With N blocks per page and P pages, for each inode we have at most:
824 * min(N*P, blocksize/4 + 1) dindirect blocks
827 * For the entire filesystem, we have at most:
828 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
829 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
832 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
834 static int ext3_credits_needed(struct super_block *sb, int objcount,
835 struct obd_ioobj *obj)
837 struct obd_ioobj *o = obj;
838 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
839 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
842 int needed = objcount + 1;
845 for (i = 0; i < objcount; i++, o++) {
846 int nblocks = o->ioo_bufcnt * blockpp;
847 int ndindirect = min(nblocks, addrpp + 1);
848 int nindir = nblocks + ndindirect + 1;
850 nbitmaps += nindir + nblocks;
851 ngdblocks += nindir + nblocks;
856 /* Assumes ext3 and extN have same sb_info layout at the start. */
857 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
858 nbitmaps = EXT3_SB(sb)->s_groups_count;
859 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
860 ngdblocks = EXT3_SB(sb)->s_gdb_count;
862 needed += nbitmaps + ngdblocks;
865 /* We assume that there will be 1 bit set in s_dquot.flags for each
866 * quota file that is active. This is at least true for now.
868 needed += hweight32(sb_any_quota_enabled(sb)) *
869 EXT3_SINGLEDATA_TRANS_BLOCKS;
875 /* We have to start a huge journal transaction here to hold all of the
876 * metadata for the pages being written here. This is necessitated by
877 * the fact that we do lots of prepare_write operations before we do
878 * any of the matching commit_write operations, so even if we split
879 * up to use "smaller" transactions none of them could complete until
880 * all of them were opened. By having a single journal transaction,
881 * we eliminate duplicate reservations for common blocks like the
882 * superblock and group descriptors or bitmaps.
884 * We will start the transaction here, but each prepare_write will
885 * add a refcount to the transaction, and each commit_write will
886 * remove a refcount. The transaction will be closed when all of
887 * the pages have been written.
889 static void *ext3_filter_journal_start(struct filter_obd *filter,
890 int objcount, struct obd_ioobj *obj,
891 int niocount, struct niobuf_remote *nb)
893 journal_t *journal = NULL;
894 handle_t *handle = NULL;
897 /* It appears that some kernels have different values for
898 * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
899 * assume anything after s_inode_bitmap_number is the same.
901 if (!strcmp(filter->fo_fstype, "ext3"))
902 journal = EXT3_SB(filter->fo_sb)->s_journal;
903 else if (!strcmp(filter->fo_fstype, "extN"))
904 journal = EXTN_SB(filter->fo_sb)->s_journal;
905 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
907 /* The number of blocks we could _possibly_ dirty can very large.
908 * We reduce our request if it is absurd (and we couldn't get that
909 * many credits for a single handle anyways).
911 * At some point we have to limit the size of I/Os sent at one time,
912 * increase the size of the journal, or we have to calculate the
913 * actual journal requirements more carefully by checking all of
914 * the blocks instead of being maximally pessimistic. It remains to
915 * be seen if this is a real problem or not.
917 if (needed > journal->j_max_transaction_buffers) {
918 CERROR("want too many journal credits (%d) using %d instead\n",
919 needed, journal->j_max_transaction_buffers);
920 needed = journal->j_max_transaction_buffers;
924 handle = journal_start(journal, needed);
927 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
933 static void *filter_journal_start(void **journal_save,
934 struct filter_obd *filter,
935 int objcount, struct obd_ioobj *obj,
936 int niocount, struct niobuf_remote *nb)
940 /* This may not be necessary - we probably never have a
941 * transaction started when we enter here, so we can
942 * remove the saving of the journal state entirely.
943 * For now leave it in just to see if it ever happens.
945 *journal_save = current->journal_info;
947 CERROR("Already have handle %p???\n", *journal_save);
949 current->journal_info = NULL;
952 if (!strcmp(filter->fo_fstype, "ext3") ||
953 !strcmp(filter->fo_fstype, "extN"))
954 handle = ext3_filter_journal_start(filter, objcount, obj,
959 static int ext3_filter_journal_stop(void *handle)
963 /* We got a refcount on the handle for each call to prepare_write,
964 * so we can drop the "parent" handle here to avoid the need for
965 * osc to call back into filterobd to close the handle. The
966 * remaining references will be dropped in commit_write.
969 rc = journal_stop((handle_t *)handle);
975 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
980 if (!strcmp(filter->fo_fstype, "ext3") ||
981 !strcmp(filter->fo_fstype, "extN"))
982 rc = ext3_filter_journal_stop(handle);
985 CERROR("error on journal stop: rc = %d\n", rc);
987 current->journal_info = journal_save;
992 static inline void lustre_put_page(struct page *page)
995 page_cache_release(page);
999 lustre_get_page_read(struct inode *inode, unsigned long index)
1001 struct address_space *mapping = inode->i_mapping;
1005 page = read_cache_page(mapping, index,
1006 (filler_t*)mapping->a_ops->readpage, NULL);
1007 if (!IS_ERR(page)) {
1010 if (!Page_Uptodate(page)) {
1011 CERROR("page index %lu not uptodate\n", index);
1012 GOTO(err_page, rc = -EIO);
1014 if (PageError(page)) {
1015 CERROR("page index %lu has error\n", index);
1016 GOTO(err_page, rc = -EIO);
1022 lustre_put_page(page);
1026 static struct page *
1027 lustre_get_page_write(struct inode *inode, unsigned long index)
1029 struct address_space *mapping = inode->i_mapping;
1033 page = grab_cache_page(mapping, index); /* locked page */
1035 if (!IS_ERR(page)) {
1037 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1038 * a no-op for most filesystems, because we write the whole
1039 * page. For partial-page I/O this will read in the page.
1041 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1043 CERROR("page index %lu, rc = %d\n", index, rc);
1046 GOTO(err_unlock, rc);
1048 /* XXX not sure if we need this if we are overwriting page */
1049 if (PageError(page)) {
1050 CERROR("error on page index %lu, rc = %d\n", index, rc);
1052 GOTO(err_unlock, rc = -EIO);
1059 lustre_put_page(page);
1063 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1065 struct inode *inode = page->mapping->host;
1068 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1069 if (!err && IS_SYNC(inode))
1070 err = waitfor_one_page(page);
1072 //SetPageUptodate(page); // the client commit_write will do this
1074 SetPageReferenced(page);
1076 lustre_put_page(page);
1080 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1081 struct niobuf_local *lnb, int *pglocked)
1083 struct address_space *mapping = inode->i_mapping;
1087 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1089 page = grab_cache_page_nowait(mapping, index); /* locked page */
1091 page = grab_cache_page(mapping, index); /* locked page */
1094 /* This page is currently locked, so get a temporary page instead. */
1095 /* XXX I believe this is a very dangerous thing to do - consider if
1096 * we had multiple writers for the same file (definitely the case
1097 * if we are using this codepath). If writer A locks the page,
1098 * writer B writes to a copy (as here), writer A drops the page
1099 * lock, and writer C grabs the lock before B does, then B will
1100 * later overwrite the data from C, even if C had LDLM locked
1101 * and initiated the write after B did.
1105 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1106 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1108 CERROR("no memory for a temp page\n");
1110 GOTO(err, rc = -ENOMEM);
1112 page = virt_to_page(addr);
1114 page->index = index;
1115 lnb->flags |= N_LOCAL_TEMP_PAGE;
1116 } else if (!IS_ERR(page)) {
1120 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1121 * a no-op for most filesystems, because we write the whole
1122 * page. For partial-page I/O this will read in the page.
1124 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1126 CERROR("page index %lu, rc = %d\n", index, rc);
1129 GOTO(err_unlock, rc);
1131 /* XXX not sure if we need this if we are overwriting page */
1132 if (PageError(page)) {
1133 CERROR("error on page index %lu, rc = %d\n", index, rc);
1135 GOTO(err_unlock, rc = -EIO);
1142 lustre_put_page(page);
1148 * We need to balance prepare_write() calls with commit_write() calls.
1149 * If the page has been prepared, but we have no data for it, we don't
1150 * want to overwrite valid data on disk, but we still need to zero out
1151 * data for space which was newly allocated. Like part of what happens
1152 * in __block_prepare_write() for newly allocated blocks.
1154 * XXX currently __block_prepare_write() creates buffers for all the
1155 * pages, and the filesystems mark these buffers as BH_New if they
1156 * were newly allocated from disk. We use the BH_New flag similarly.
1158 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1162 unsigned block_start, block_end;
1163 struct buffer_head *bh, *head = page->buffers;
1164 unsigned blocksize = head->b_size;
1165 void *addr = page_address(page);
1167 /* debugging: just seeing if this ever happens */
1168 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1169 page->index, page->mapping->host->i_ino, err);
1171 /* Currently one buffer per page, but in the future... */
1172 for (bh = head, block_start = 0; bh != head || !block_start;
1173 block_start = block_end, bh = bh->b_this_page) {
1174 block_end = block_start + blocksize;
1176 memset(addr + block_start, 0, blocksize);
1180 return lustre_commit_write(page, from, to);
1183 static int filter_preprw(int cmd, struct lustre_handle *conn,
1184 int objcount, struct obd_ioobj *obj,
1185 int niocount, struct niobuf_remote *nb,
1186 struct niobuf_local *res, void **desc_private)
1188 struct obd_run_ctxt saved;
1189 struct obd_device *obd;
1190 struct obd_ioobj *o = obj;
1191 struct niobuf_remote *b = nb;
1192 struct niobuf_local *r = res;
1193 void *journal_save = NULL;
1199 obd = class_conn2obd(conn);
1201 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1204 memset(res, 0, sizeof(*res) * niocount);
1206 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1208 if (cmd & OBD_BRW_WRITE) {
1209 *desc_private = filter_journal_start(&journal_save,
1211 objcount, obj, niocount,
1213 if (IS_ERR(*desc_private))
1214 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1217 for (i = 0; i < objcount; i++, o++) {
1218 struct dentry *dentry;
1219 struct inode *inode;
1222 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1223 o->ioo_id, S_IFREG);
1225 GOTO(out_clean, rc = PTR_ERR(dentry));
1226 inode = dentry->d_inode;
1228 CERROR("trying to BRW to non-existent file "LPU64"\n",
1231 GOTO(out_clean, rc = -ENOENT);
1234 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1235 unsigned long index = b->offset >> PAGE_SHIFT;
1241 r->dentry = dget(dentry);
1243 if (cmd & OBD_BRW_WRITE)
1244 page = filter_get_page_write(inode, index, r,
1247 page = lustre_get_page_read(inode, index);
1251 GOTO(out_clean, rc = PTR_ERR(page));
1254 r->addr = page_address(page);
1255 r->offset = b->offset;
1262 if (cmd & OBD_BRW_WRITE) {
1263 int err = filter_journal_stop(journal_save, &obd->u.filter,
1273 CERROR("error cleanup on brw\n");
1275 if (cmd & OBD_BRW_WRITE)
1276 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1278 lustre_put_page(r->page);
1283 static int filter_write_locked_page(struct niobuf_local *lnb)
1288 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1289 if (IS_ERR(lpage)) {
1290 /* It is highly unlikely that we would ever get an error here.
1291 * The page we want to get was previously locked, so it had to
1292 * have already allocated the space, and we were just writing
1293 * over the same data, so there would be no hole in the file.
1295 * XXX: possibility of a race with truncate could exist, need
1296 * to check that. There are no guarantees w.r.t.
1297 * write order even on a local filesystem, although the
1298 * normal response would be to return the number of bytes
1299 * successfully written and leave the rest to the app.
1301 rc = PTR_ERR(lpage);
1302 CERROR("error getting locked page index %ld: rc = %d\n",
1303 lnb->page->index, rc);
1307 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1308 * lustre_commit_write() below, lnb->page was kmapped previously in
1309 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1311 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1312 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1314 CERROR("error committing locked page %ld: rc = %d\n",
1315 lnb->page->index, rc);
1317 lustre_put_page(lnb->page);
1322 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1323 int objcount, struct obd_ioobj *obj,
1324 int niocount, struct niobuf_local *res,
1327 struct obd_run_ctxt saved;
1328 struct obd_ioobj *o;
1329 struct niobuf_local *r;
1330 struct obd_device *obd = class_conn2obd(conn);
1332 int found_locked = 0;
1337 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1338 journal_save = current->journal_info;
1339 LASSERT(!journal_save);
1341 current->journal_info = private;
1342 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1344 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1345 struct page *page = r->page;
1350 if (r->flags & N_LOCAL_TEMP_PAGE) {
1355 if (cmd & OBD_BRW_WRITE) {
1356 int err = filter_commit_write(page, 0,
1362 lustre_put_page(page);
1367 current->journal_info = journal_save;
1372 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1374 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1376 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1379 err = filter_write_locked_page(r);
1391 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1393 struct obd_device *obd = class_conn2obd(conn);
1398 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1400 statfs_pack(osfs, &sfs);
1405 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1406 void *key, obd_count *vallen, void **val)
1408 struct obd_device *obd;
1411 obd = class_conn2obd(conn);
1413 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1417 if ( keylen == strlen("blocksize") &&
1418 memcmp(key, "blocksize", keylen) == 0 ) {
1419 *vallen = sizeof(long);
1420 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1424 if ( keylen == strlen("blocksize_bits") &&
1425 memcmp(key, "blocksize_bits", keylen) == 0 ){
1426 *vallen = sizeof(long);
1427 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1431 if ( keylen == strlen("root_ino") &&
1432 memcmp(key, "root_ino", keylen) == 0 ){
1433 *vallen = sizeof(obd_id);
1434 *val = (void *)(obd_id)FILTER_ROOTINO;
1438 CDEBUG(D_IOCTL, "invalid key\n");
1442 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1443 struct lustre_handle *src_conn, struct obdo *src,
1444 obd_size count, obd_off offset)
1447 struct lov_stripe_md srcmd, dstmd;
1448 unsigned long index = 0;
1451 memset(&srcmd, 0, sizeof(srcmd));
1452 memset(&dstmd, 0, sizeof(dstmd));
1453 srcmd.lsm_object_id = src->o_id;
1454 dstmd.lsm_object_id = dst->o_id;
1457 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1458 ", dst: ino "LPU64"\n",
1459 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1460 page = alloc_page(GFP_USER);
1464 while (TryLockPage(page))
1465 ___wait_on_page(page);
1467 /* XXX with brw vector I/O, we could batch up reads and writes here,
1468 * all we need to do is allocate multiple pages to handle the I/Os
1469 * and arrays to handle the request parameters.
1471 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1473 struct io_cb_data *cbd = ll_init_cb();
1482 pg.count = PAGE_SIZE;
1483 pg.off = (page->index) << PAGE_SHIFT;
1486 page->index = index;
1487 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1488 ll_sync_io_cb, cbd);
1501 pg.flag = OBD_BRW_CREATE;
1502 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1504 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1505 ll_sync_io_cb, cbd);
1507 /* XXX should handle dst->o_size, dst->o_blocks here */
1513 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1517 dst->o_size = src->o_size;
1518 dst->o_blocks = src->o_blocks;
1519 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1527 static struct obd_ops filter_obd_ops = {
1528 o_get_info: filter_get_info,
1529 o_setup: filter_setup,
1530 o_cleanup: filter_cleanup,
1531 o_connect: filter_connect,
1532 o_disconnect: filter_disconnect,
1533 o_statfs: filter_statfs,
1534 o_getattr: filter_getattr,
1535 o_create: filter_create,
1536 o_setattr: filter_setattr,
1537 o_destroy: filter_destroy,
1538 o_open: filter_open,
1539 o_close: filter_close,
1540 o_brw: filter_pgcache_brw,
1541 o_punch: filter_truncate,
1542 o_preprw: filter_preprw,
1543 o_commitrw: filter_commitrw
1545 o_preallocate: filter_preallocate_inodes,
1546 o_migrate: filter_migrate,
1547 o_copy: filter_copy_data,
1548 o_iterate: filter_iterate
1553 static int __init obdfilter_init(void)
1555 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1556 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1559 static void __exit obdfilter_exit(void)
1561 class_unregister_type(OBD_FILTER_DEVICENAME);
1564 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1565 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1566 MODULE_LICENSE("GPL");
1568 module_init(obdfilter_init);
1569 module_exit(obdfilter_exit);