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>
30 #define FILTER_ROOTINO 2
31 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
34 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
36 [S_IFREG >> S_SHIFT] "R",
37 [S_IFDIR >> S_SHIFT] "D",
38 [S_IFCHR >> S_SHIFT] "C",
39 [S_IFBLK >> S_SHIFT] "B",
40 [S_IFIFO >> S_SHIFT] "F",
41 [S_IFSOCK >> S_SHIFT] "S",
42 [S_IFLNK >> S_SHIFT] "L"
45 static inline const char *obd_mode_to_type(int mode)
47 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
50 /* write the pathname into the string */
51 static int filter_id(char *buf, obd_id id, obd_mode mode)
53 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
56 static inline void f_dput(struct dentry *dentry)
58 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
59 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
60 LASSERT(atomic_read(&dentry->d_count) > 0);
64 /* setup the object store with correct subdirectories */
65 static int filter_prep(struct obd_device *obddev)
67 struct obd_run_ctxt saved;
68 struct filter_obd *filter = &obddev->u.filter;
69 struct dentry *dentry;
77 push_ctxt(&saved, &filter->fo_ctxt, NULL);
78 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
79 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
82 CERROR("cannot open/create O: rc = %d\n", rc);
85 filter->fo_dentry_O = dentry;
86 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
87 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
90 CERROR("cannot open/create P: rc = %d\n", rc);
94 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
95 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
98 CERROR("cannot open/create D: rc = %d\n", rc);
102 root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
106 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
113 * Create directories and/or get dentries for each object type.
114 * This saves us from having to do multiple lookups for each one.
116 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
117 char *type = obd_type_by_mode[mode];
120 filter->fo_dentry_O_mode[mode] = NULL;
123 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
124 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
125 if (IS_ERR(dentry)) {
126 rc = PTR_ERR(dentry);
127 CERROR("cannot create O/%s: rc = %d\n", type, rc);
128 GOTO(out_O_mode, rc);
130 filter->fo_dentry_O_mode[mode] = dentry;
133 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
134 if ( !file || IS_ERR(file) ) {
136 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
138 GOTO(out_O_mode, rc);
141 /* steal operations */
142 inode = file->f_dentry->d_inode;
143 filter->fo_fop = file->f_op;
144 filter->fo_iop = inode->i_op;
145 filter->fo_aops = inode->i_mapping->a_ops;
147 if (inode->i_size == 0) {
148 __u64 disk_lastino = cpu_to_le64(lastino);
149 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
150 sizeof(disk_lastino),
152 if (retval != sizeof(disk_lastino)) {
153 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
155 GOTO(out_O_mode, rc = -EIO);
159 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
160 sizeof(disk_lastino),
162 if (retval != sizeof(disk_lastino)) {
163 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
165 GOTO(out_O_mode, rc = -EIO);
167 lastino = le64_to_cpu(disk_lastino);
169 filter->fo_lastino = lastino;
180 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
183 filter->fo_dentry_O_mode[mode] = NULL;
187 f_dput(filter->fo_dentry_O);
188 filter->fo_dentry_O = NULL;
192 /* cleanup the filter: write last used object id to status file */
193 static void filter_post(struct obd_device *obddev)
195 struct obd_run_ctxt saved;
196 struct filter_obd *filter = &obddev->u.filter;
202 push_ctxt(&saved, &filter->fo_ctxt, NULL);
203 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
205 CERROR("OBD filter: cannot create status file\n");
210 disk_lastino = cpu_to_le64(filter->fo_lastino);
211 rc = file->f_op->write(file, (char *)&disk_lastino,
212 sizeof(disk_lastino), &file->f_pos);
213 if (rc != sizeof(disk_lastino))
214 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
216 rc = filp_close(file, NULL);
218 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
220 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
221 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
224 filter->fo_dentry_O_mode[mode] = NULL;
227 f_dput(filter->fo_dentry_O);
233 static __u64 filter_next_id(struct obd_device *obddev)
237 spin_lock(&obddev->u.filter.fo_lock);
238 id = ++obddev->u.filter.fo_lastino;
239 spin_unlock(&obddev->u.filter.fo_lock);
241 /* FIXME: write the lastino to disk here */
245 /* how to get files, dentries, inodes from object id's */
246 /* parent i_sem is already held if needed for exclusivity */
247 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
248 struct dentry *dparent,
249 __u64 id, __u32 type)
251 struct super_block *sb = obddev->u.filter.fo_sb;
252 struct dentry *dchild;
257 if (!sb || !sb->s_dev) {
258 CERROR("fatal: device not initialized.\n");
259 RETURN(ERR_PTR(-ENXIO));
263 CERROR("fatal: invalid object #0\n");
265 RETURN(ERR_PTR(-ESTALE));
268 if (!(type & S_IFMT)) {
269 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
270 __FUNCTION__, id, type);
271 RETURN(ERR_PTR(-EINVAL));
274 len = sprintf(name, LPU64, id);
275 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
277 dchild = lookup_one_len(name, dparent, len);
278 if (IS_ERR(dchild)) {
279 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
283 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
284 obd_mode_to_type(type), name, dchild,
285 atomic_read(&dchild->d_count));
287 LASSERT(atomic_read(&dchild->d_count) > 0);
292 static struct file *filter_obj_open(struct obd_device *obddev,
293 __u64 id, __u32 type)
295 struct super_block *sb = obddev->u.filter.fo_sb;
296 struct obd_run_ctxt saved;
301 if (!sb || !sb->s_dev) {
302 CERROR("fatal: device not initialized.\n");
303 RETURN(ERR_PTR(-ENXIO));
307 CERROR("fatal: invalid obdo "LPU64"\n", id);
308 RETURN(ERR_PTR(-ESTALE));
311 if (!(type & S_IFMT)) {
312 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
313 __FUNCTION__, id, type);
314 RETURN(ERR_PTR(-EINVAL));
317 filter_id(name, id, type);
318 push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
319 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
322 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
329 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
331 struct filter_obd *filter = &obddev->u.filter;
333 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
338 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
344 rc = class_connect(conn, obd, cluuid);
350 static int filter_disconnect(struct lustre_handle *conn)
355 rc = class_disconnect(conn);
359 /* XXX cleanup preallocated inodes */
363 /* mount the file system (secretly) */
364 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
366 struct obd_ioctl_data* data = buf;
367 struct filter_obd *filter;
368 struct vfsmount *mnt;
372 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
376 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
381 filter = &obddev->u.filter;;
382 filter->fo_vfsmnt = mnt;
383 filter->fo_fstype = strdup(data->ioc_inlbuf2);
384 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
385 /* XXX is this even possible if do_kern_mount succeeded? */
387 GOTO(err_kfree, err = -ENODEV);
389 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
390 filter->fo_ctxt.pwdmnt = mnt;
391 filter->fo_ctxt.pwd = mnt->mnt_root;
392 filter->fo_ctxt.fs = get_ds();
394 err = filter_prep(obddev);
396 GOTO(err_kfree, err);
397 spin_lock_init(&filter->fo_lock);
399 obddev->obd_namespace =
400 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
401 if (obddev->obd_namespace == NULL)
404 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
405 "filter_ldlm_client", &obddev->obd_ldlm_client);
410 kfree(filter->fo_fstype);
412 mntput(filter->fo_vfsmnt);
422 static int filter_cleanup(struct obd_device * obddev)
424 struct super_block *sb;
427 if (!list_empty(&obddev->obd_exports)) {
428 CERROR("still has clients!\n");
429 class_disconnect_all(obddev);
430 if (!list_empty(&obddev->obd_exports)) {
431 CERROR("still has exports after forced cleanup?\n");
436 ldlm_namespace_free(obddev->obd_namespace);
438 sb = obddev->u.filter.fo_sb;
439 if (!obddev->u.filter.fo_sb)
444 shrink_dcache_parent(sb->s_root);
446 mntput(obddev->u.filter.fo_vfsmnt);
447 obddev->u.filter.fo_sb = 0;
448 kfree(obddev->u.filter.fo_fstype);
457 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
460 int type = oa->o_mode & S_IFMT;
463 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
464 inode->i_ino, inode, (long)oa->o_id, valid);
465 /* Don't copy the inode number in place of the object ID */
466 obdo_from_inode(oa, inode, valid);
467 oa->o_mode &= ~S_IFMT;
470 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
471 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
473 oa->o_valid |= OBD_MD_FLRDEV;
479 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
480 struct lov_stripe_md *md)
482 struct obd_device *obddev = class_conn2obd(conn);
483 struct dentry *dentry;
487 if (!class_conn2export(conn)) {
488 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
492 obddev = class_conn2obd(conn);
493 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
494 oa->o_id, oa->o_mode);
496 RETURN(PTR_ERR(dentry));
498 if (!dentry->d_inode) {
499 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
500 GOTO(out_getattr, rc = -ENOENT);
503 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
510 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
511 struct lov_stripe_md *md)
513 struct obd_run_ctxt saved;
514 struct obd_device *obd = class_conn2obd(conn);
515 struct dentry *dentry;
521 iattr_from_obdo(&iattr, oa, oa->o_valid);
522 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
523 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
524 oa->o_id, iattr.ia_mode);
526 RETURN(PTR_ERR(dentry));
528 inode = dentry->d_inode;
530 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
531 GOTO(out_setattr, rc = -ENOENT);
535 if (iattr.ia_valid & ATTR_SIZE)
537 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
538 if (inode->i_op->setattr)
539 rc = inode->i_op->setattr(dentry, &iattr);
541 rc = inode_setattr(inode, &iattr);
543 if (iattr.ia_valid & ATTR_SIZE) {
545 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
546 obdo_from_inode(oa, inode, oa->o_valid);
555 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
556 struct lov_stripe_md *ea)
558 struct obd_export *export;
559 struct obd_device *obd;
560 struct dentry *dentry;
563 export = class_conn2export(conn);
565 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
569 obd = export->exp_obd;
570 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
571 oa->o_id, oa->o_mode);
573 RETURN(PTR_ERR(dentry));
575 if (!dentry->d_inode) {
576 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
580 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
581 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
586 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
587 struct lov_stripe_md *ea)
589 struct obd_device *obd;
590 struct dentry *dentry;
593 obd = class_conn2obd(conn);
595 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
599 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
600 oa->o_id, oa->o_mode);
602 RETURN(PTR_ERR(dentry));
603 LASSERT(atomic_read(&dentry->d_count) > 1);
605 f_dput(dentry); /* for the close */
606 f_dput(dentry); /* for this call */
610 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
611 struct lov_stripe_md **ea)
614 struct obd_run_ctxt saved;
617 struct obd_device *obd = class_conn2obd(conn);
618 struct filter_obd *filter = &obd->u.filter;
623 CERROR("invalid client "LPX64"\n", conn->addr);
627 if (!(oa->o_mode && S_IFMT)) {
628 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
629 __FUNCTION__, oa->o_id, oa->o_mode);
633 oa->o_id = filter_next_id(obd);
635 //filter_id(name, oa->o_id, oa->o_mode);
636 sprintf(name, LPU64, oa->o_id);
637 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
638 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
639 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
643 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
647 /* Set flags for fields we have set in the inode struct */
648 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
649 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
650 filter_from_inode(oa, new->d_inode, oa->o_valid);
656 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
657 struct lov_stripe_md *ea)
659 struct obd_device *obd;
660 struct filter_obd *filter;
661 struct obd_run_ctxt saved;
663 struct dentry *dir_dentry, *object_dentry;
667 obd = class_conn2obd(conn);
669 CERROR("invalid client "LPX64"\n", conn->addr);
673 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
675 dir_dentry = filter_parent(obd, oa->o_mode);
676 down(&dir_dentry->d_inode->i_sem);
678 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
680 if (IS_ERR(object_dentry))
681 GOTO(out, rc = -ENOENT);
683 inode = object_dentry->d_inode;
685 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
686 GOTO(out, rc = -ENOENT);
689 if (inode->i_nlink != 1) {
690 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
694 inode->i_mode = S_IFREG;
696 if (atomic_read(&inode->i_count) > 1) {
697 #warning FIXME: need to handle open-unlinked case and move to hold dir
698 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
701 filter = &obd->u.filter;
702 push_ctxt(&saved, &filter->fo_ctxt, NULL);
704 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
706 f_dput(object_dentry);
710 up(&dir_dentry->d_inode->i_sem);
714 /* NB count and offset are used for punch, but not truncate */
715 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
716 struct lov_stripe_md *md,
717 obd_off start, obd_off end)
722 if (end != OBD_PUNCH_EOF)
723 CERROR("PUNCH not supported, only truncate works\n");
725 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
726 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
728 error = filter_setattr(conn, oa, NULL);
732 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
733 struct lov_stripe_md *lsm, obd_count oa_bufs,
734 struct brw_page *pga, brw_callback_t callback,
735 struct io_cb_data *data)
737 struct obd_run_ctxt saved;
738 struct super_block *sb;
739 int pnum; /* index to pages (bufs) */
740 unsigned long retval;
743 struct obd_device *obd = class_conn2obd(conn);
748 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
752 sb = obd->u.filter.fo_sb;
753 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
754 pnum = 0; /* pnum indexes buf 0..num_pages */
756 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
758 GOTO(out, retval = PTR_ERR(file));
760 /* count doubles as retval */
761 for (pg = 0; pg < oa_bufs; pg++) {
762 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
763 ") off count ("LPU64",%d)\n",
764 cmd, pnum, file->f_dentry->d_inode->i_ino,
765 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
766 (int)pga[pnum].count);
767 if (cmd & OBD_BRW_WRITE) {
771 buffer = kmap(pga[pnum].pg);
772 retval = file->f_op->write(file, buffer,
775 kunmap(pga[pnum].pg);
776 CDEBUG(D_INODE, "retval %ld\n", retval);
778 loff_t off = pga[pnum].off;
779 char *buffer = kmap(pga[pnum].pg);
781 if (off >= file->f_dentry->d_inode->i_size) {
782 memset(buffer, 0, pga[pnum].count);
783 retval = pga[pnum].count;
785 retval = file->f_op->read(file, buffer,
786 pga[pnum].count, &off);
788 kunmap(pga[pnum].pg);
790 if (retval != pga[pnum].count) {
792 GOTO(out, retval = -EIO);
794 CDEBUG(D_INODE, "retval %ld\n", retval);
798 /* sizes and blocks are set by generic_file_write */
799 /* ctimes/mtimes will follow with a setattr call */
802 /* XXX: do something with callback if it is set? */
807 error = (retval >= 0) ? 0 : retval;
812 * Calculate the number of buffer credits needed to write multiple pages in
813 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
814 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
816 * See comment above ext3_writepage_trans_blocks for details. We assume
817 * no data journaling is being done, but it does allow for all of the pages
818 * being non-contiguous. If we are guaranteed contiguous pages we could
819 * reduce the number of (d)indirect blocks a lot.
821 * With N blocks per page and P pages, for each inode we have at most:
823 * min(N*P, blocksize/4 + 1) dindirect blocks
826 * For the entire filesystem, we have at most:
827 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
828 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
831 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
833 static int ext3_credits_needed(struct super_block *sb, int objcount,
834 struct obd_ioobj *obj)
836 struct obd_ioobj *o = obj;
837 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
838 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
841 int needed = objcount + 1;
844 for (i = 0; i < objcount; i++, o++) {
845 int nblocks = o->ioo_bufcnt * blockpp;
846 int ndindirect = min(nblocks, addrpp + 1);
847 int nindir = nblocks + ndindirect + 1;
849 nbitmaps += nindir + nblocks;
850 ngdblocks += nindir + nblocks;
855 /* Assumes ext3 and extN have same sb_info layout at the start. */
856 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
857 nbitmaps = EXT3_SB(sb)->s_groups_count;
858 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
859 ngdblocks = EXT3_SB(sb)->s_gdb_count;
861 needed += nbitmaps + ngdblocks;
864 /* We assume that there will be 1 bit set in s_dquot.flags for each
865 * quota file that is active. This is at least true for now.
867 needed += hweight32(sb_any_quota_enabled(sb)) *
868 EXT3_SINGLEDATA_TRANS_BLOCKS;
874 /* We have to start a huge journal transaction here to hold all of the
875 * metadata for the pages being written here. This is necessitated by
876 * the fact that we do lots of prepare_write operations before we do
877 * any of the matching commit_write operations, so even if we split
878 * up to use "smaller" transactions none of them could complete until
879 * all of them were opened. By having a single journal transaction,
880 * we eliminate duplicate reservations for common blocks like the
881 * superblock and group descriptors or bitmaps.
883 * We will start the transaction here, but each prepare_write will
884 * add a refcount to the transaction, and each commit_write will
885 * remove a refcount. The transaction will be closed when all of
886 * the pages have been written.
888 static void *ext3_filter_journal_start(struct filter_obd *filter,
889 int objcount, struct obd_ioobj *obj,
890 int niocount, struct niobuf_remote *nb)
892 journal_t *journal = NULL;
893 handle_t *handle = NULL;
896 /* It appears that some kernels have different values for
897 * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
898 * assume anything after s_inode_bitmap_number is the same.
900 if (!strcmp(filter->fo_fstype, "ext3"))
901 journal = EXT3_SB(filter->fo_sb)->s_journal;
902 else if (!strcmp(filter->fo_fstype, "extN"))
903 journal = EXTN_SB(filter->fo_sb)->s_journal;
904 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
906 /* The number of blocks we could _possibly_ dirty can very large.
907 * We reduce our request if it is absurd (and we couldn't get that
908 * many credits for a single handle anyways).
910 * At some point we have to limit the size of I/Os sent at one time,
911 * increase the size of the journal, or we have to calculate the
912 * actual journal requirements more carefully by checking all of
913 * the blocks instead of being maximally pessimistic. It remains to
914 * be seen if this is a real problem or not.
916 if (needed > journal->j_max_transaction_buffers) {
917 CERROR("want too many journal credits (%d) using %d instead\n",
918 needed, journal->j_max_transaction_buffers);
919 needed = journal->j_max_transaction_buffers;
923 handle = journal_start(journal, needed);
926 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
932 static void *filter_journal_start(void **journal_save,
933 struct filter_obd *filter,
934 int objcount, struct obd_ioobj *obj,
935 int niocount, struct niobuf_remote *nb)
939 /* This may not be necessary - we probably never have a
940 * transaction started when we enter here, so we can
941 * remove the saving of the journal state entirely.
942 * For now leave it in just to see if it ever happens.
944 *journal_save = current->journal_info;
946 CERROR("Already have handle %p???\n", *journal_save);
948 current->journal_info = NULL;
951 if (!strcmp(filter->fo_fstype, "ext3") ||
952 !strcmp(filter->fo_fstype, "extN"))
953 handle = ext3_filter_journal_start(filter, objcount, obj,
958 static int ext3_filter_journal_stop(void *handle)
962 /* We got a refcount on the handle for each call to prepare_write,
963 * so we can drop the "parent" handle here to avoid the need for
964 * osc to call back into filterobd to close the handle. The
965 * remaining references will be dropped in commit_write.
968 rc = journal_stop((handle_t *)handle);
974 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
979 if (!strcmp(filter->fo_fstype, "ext3") ||
980 !strcmp(filter->fo_fstype, "extN"))
981 rc = ext3_filter_journal_stop(handle);
984 CERROR("error on journal stop: rc = %d\n", rc);
986 current->journal_info = journal_save;
991 static inline void lustre_put_page(struct page *page)
994 page_cache_release(page);
998 lustre_get_page_read(struct inode *inode, unsigned long index)
1000 struct address_space *mapping = inode->i_mapping;
1004 page = read_cache_page(mapping, index,
1005 (filler_t*)mapping->a_ops->readpage, NULL);
1006 if (!IS_ERR(page)) {
1009 if (!Page_Uptodate(page)) {
1010 CERROR("page index %lu not uptodate\n", index);
1011 GOTO(err_page, rc = -EIO);
1013 if (PageError(page)) {
1014 CERROR("page index %lu has error\n", index);
1015 GOTO(err_page, rc = -EIO);
1021 lustre_put_page(page);
1025 static struct page *
1026 lustre_get_page_write(struct inode *inode, unsigned long index)
1028 struct address_space *mapping = inode->i_mapping;
1032 page = grab_cache_page(mapping, index); /* locked page */
1034 if (!IS_ERR(page)) {
1036 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1037 * a no-op for most filesystems, because we write the whole
1038 * page. For partial-page I/O this will read in the page.
1040 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1042 CERROR("page index %lu, rc = %d\n", index, rc);
1045 GOTO(err_unlock, rc);
1047 /* XXX not sure if we need this if we are overwriting page */
1048 if (PageError(page)) {
1049 CERROR("error on page index %lu, rc = %d\n", index, rc);
1051 GOTO(err_unlock, rc = -EIO);
1058 lustre_put_page(page);
1062 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1064 struct inode *inode = page->mapping->host;
1067 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1068 if (!err && IS_SYNC(inode))
1069 err = waitfor_one_page(page);
1071 //SetPageUptodate(page); // the client commit_write will do this
1073 SetPageReferenced(page);
1075 lustre_put_page(page);
1079 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1080 struct niobuf_local *lnb, int *pglocked)
1082 struct address_space *mapping = inode->i_mapping;
1086 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1088 page = grab_cache_page_nowait(mapping, index); /* locked page */
1090 page = grab_cache_page(mapping, index); /* locked page */
1093 /* This page is currently locked, so get a temporary page instead. */
1094 /* XXX I believe this is a very dangerous thing to do - consider if
1095 * we had multiple writers for the same file (definitely the case
1096 * if we are using this codepath). If writer A locks the page,
1097 * writer B writes to a copy (as here), writer A drops the page
1098 * lock, and writer C grabs the lock before B does, then B will
1099 * later overwrite the data from C, even if C had LDLM locked
1100 * and initiated the write after B did.
1104 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1105 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1107 CERROR("no memory for a temp page\n");
1109 GOTO(err, rc = -ENOMEM);
1111 page = virt_to_page(addr);
1113 page->index = index;
1114 lnb->flags |= N_LOCAL_TEMP_PAGE;
1115 } else if (!IS_ERR(page)) {
1119 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1120 * a no-op for most filesystems, because we write the whole
1121 * page. For partial-page I/O this will read in the page.
1123 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1125 CERROR("page index %lu, rc = %d\n", index, rc);
1128 GOTO(err_unlock, rc);
1130 /* XXX not sure if we need this if we are overwriting page */
1131 if (PageError(page)) {
1132 CERROR("error on page index %lu, rc = %d\n", index, rc);
1134 GOTO(err_unlock, rc = -EIO);
1141 lustre_put_page(page);
1147 * We need to balance prepare_write() calls with commit_write() calls.
1148 * If the page has been prepared, but we have no data for it, we don't
1149 * want to overwrite valid data on disk, but we still need to zero out
1150 * data for space which was newly allocated. Like part of what happens
1151 * in __block_prepare_write() for newly allocated blocks.
1153 * XXX currently __block_prepare_write() creates buffers for all the
1154 * pages, and the filesystems mark these buffers as BH_New if they
1155 * were newly allocated from disk. We use the BH_New flag similarly.
1157 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1161 unsigned block_start, block_end;
1162 struct buffer_head *bh, *head = page->buffers;
1163 unsigned blocksize = head->b_size;
1164 void *addr = page_address(page);
1166 /* debugging: just seeing if this ever happens */
1167 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1168 page->index, page->mapping->host->i_ino, err);
1170 /* Currently one buffer per page, but in the future... */
1171 for (bh = head, block_start = 0; bh != head || !block_start;
1172 block_start = block_end, bh = bh->b_this_page) {
1173 block_end = block_start + blocksize;
1175 memset(addr + block_start, 0, blocksize);
1179 return lustre_commit_write(page, from, to);
1182 static int filter_preprw(int cmd, struct lustre_handle *conn,
1183 int objcount, struct obd_ioobj *obj,
1184 int niocount, struct niobuf_remote *nb,
1185 struct niobuf_local *res, void **desc_private)
1187 struct obd_run_ctxt saved;
1188 struct obd_device *obd;
1189 struct obd_ioobj *o = obj;
1190 struct niobuf_remote *b = nb;
1191 struct niobuf_local *r = res;
1192 void *journal_save = NULL;
1198 obd = class_conn2obd(conn);
1200 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1203 memset(res, 0, sizeof(*res) * niocount);
1205 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1207 if (cmd & OBD_BRW_WRITE) {
1208 *desc_private = filter_journal_start(&journal_save,
1210 objcount, obj, niocount,
1212 if (IS_ERR(*desc_private))
1213 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1216 for (i = 0; i < objcount; i++, o++) {
1217 struct dentry *dentry;
1218 struct inode *inode;
1221 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1222 o->ioo_id, S_IFREG);
1224 GOTO(out_clean, rc = PTR_ERR(dentry));
1225 inode = dentry->d_inode;
1227 CERROR("trying to BRW to non-existent file "LPU64"\n",
1230 GOTO(out_clean, rc = -ENOENT);
1233 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1234 unsigned long index = b->offset >> PAGE_SHIFT;
1240 r->dentry = dget(dentry);
1242 if (cmd & OBD_BRW_WRITE)
1243 page = filter_get_page_write(inode, index, r,
1246 page = lustre_get_page_read(inode, index);
1250 GOTO(out_clean, rc = PTR_ERR(page));
1253 r->addr = page_address(page);
1254 r->offset = b->offset;
1261 if (cmd & OBD_BRW_WRITE) {
1262 int err = filter_journal_stop(journal_save, &obd->u.filter,
1272 CERROR("error cleanup on brw\n");
1274 if (cmd & OBD_BRW_WRITE)
1275 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1277 lustre_put_page(r->page);
1282 static int filter_write_locked_page(struct niobuf_local *lnb)
1287 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1288 if (IS_ERR(lpage)) {
1289 /* It is highly unlikely that we would ever get an error here.
1290 * The page we want to get was previously locked, so it had to
1291 * have already allocated the space, and we were just writing
1292 * over the same data, so there would be no hole in the file.
1294 * XXX: possibility of a race with truncate could exist, need
1295 * to check that. There are no guarantees w.r.t.
1296 * write order even on a local filesystem, although the
1297 * normal response would be to return the number of bytes
1298 * successfully written and leave the rest to the app.
1300 rc = PTR_ERR(lpage);
1301 CERROR("error getting locked page index %ld: rc = %d\n",
1302 lnb->page->index, rc);
1306 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1307 * lustre_commit_write() below, lnb->page was kmapped previously in
1308 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1310 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1311 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1313 CERROR("error committing locked page %ld: rc = %d\n",
1314 lnb->page->index, rc);
1316 lustre_put_page(lnb->page);
1321 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1322 int objcount, struct obd_ioobj *obj,
1323 int niocount, struct niobuf_local *res,
1326 struct obd_run_ctxt saved;
1327 struct obd_ioobj *o;
1328 struct niobuf_local *r;
1329 struct obd_device *obd = class_conn2obd(conn);
1331 int found_locked = 0;
1336 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1337 journal_save = current->journal_info;
1338 LASSERT(!journal_save);
1340 current->journal_info = private;
1341 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1343 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1344 struct page *page = r->page;
1349 if (r->flags & N_LOCAL_TEMP_PAGE) {
1354 if (cmd & OBD_BRW_WRITE) {
1355 int err = filter_commit_write(page, 0,
1361 lustre_put_page(page);
1366 current->journal_info = journal_save;
1371 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1373 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1375 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1378 err = filter_write_locked_page(r);
1390 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1392 struct obd_device *obd = class_conn2obd(conn);
1397 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1399 statfs_pack(osfs, &sfs);
1404 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1405 void *key, obd_count *vallen, void **val)
1407 struct obd_device *obd;
1410 obd = class_conn2obd(conn);
1412 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1416 if ( keylen == strlen("blocksize") &&
1417 memcmp(key, "blocksize", keylen) == 0 ) {
1418 *vallen = sizeof(long);
1419 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1423 if ( keylen == strlen("blocksize_bits") &&
1424 memcmp(key, "blocksize_bits", keylen) == 0 ){
1425 *vallen = sizeof(long);
1426 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1430 if ( keylen == strlen("root_ino") &&
1431 memcmp(key, "root_ino", keylen) == 0 ){
1432 *vallen = sizeof(obd_id);
1433 *val = (void *)(obd_id)FILTER_ROOTINO;
1437 CDEBUG(D_IOCTL, "invalid key\n");
1441 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1442 struct lustre_handle *src_conn, struct obdo *src,
1443 obd_size count, obd_off offset)
1446 struct lov_stripe_md srcmd, dstmd;
1447 unsigned long index = 0;
1450 memset(&srcmd, 0, sizeof(srcmd));
1451 memset(&dstmd, 0, sizeof(dstmd));
1452 srcmd.lsm_object_id = src->o_id;
1453 dstmd.lsm_object_id = dst->o_id;
1456 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1457 ", dst: ino "LPU64"\n",
1458 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1459 page = alloc_page(GFP_USER);
1463 while (TryLockPage(page))
1464 ___wait_on_page(page);
1466 /* XXX with brw vector I/O, we could batch up reads and writes here,
1467 * all we need to do is allocate multiple pages to handle the I/Os
1468 * and arrays to handle the request parameters.
1470 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1472 struct io_cb_data *cbd = ll_init_cb();
1481 pg.count = PAGE_SIZE;
1482 pg.off = (page->index) << PAGE_SHIFT;
1485 page->index = index;
1486 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1487 ll_sync_io_cb, cbd);
1500 pg.flag = OBD_BRW_CREATE;
1501 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1503 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1504 ll_sync_io_cb, cbd);
1506 /* XXX should handle dst->o_size, dst->o_blocks here */
1512 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1516 dst->o_size = src->o_size;
1517 dst->o_blocks = src->o_blocks;
1518 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1526 static struct obd_ops filter_obd_ops = {
1527 o_get_info: filter_get_info,
1528 o_setup: filter_setup,
1529 o_cleanup: filter_cleanup,
1530 o_connect: filter_connect,
1531 o_disconnect: filter_disconnect,
1532 o_statfs: filter_statfs,
1533 o_getattr: filter_getattr,
1534 o_create: filter_create,
1535 o_setattr: filter_setattr,
1536 o_destroy: filter_destroy,
1537 o_open: filter_open,
1538 o_close: filter_close,
1539 o_brw: filter_pgcache_brw,
1540 o_punch: filter_truncate,
1541 o_preprw: filter_preprw,
1542 o_commitrw: filter_commitrw
1544 o_preallocate: filter_preallocate_inodes,
1545 o_migrate: filter_migrate,
1546 o_copy: filter_copy_data,
1547 o_iterate: filter_iterate
1552 static int __init obdfilter_init(void)
1554 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1555 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1558 static void __exit obdfilter_exit(void)
1560 class_unregister_type(OBD_FILTER_DEVICENAME);
1563 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1564 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1565 MODULE_LICENSE("GPL");
1567 module_init(obdfilter_init);
1568 module_exit(obdfilter_exit);