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/quotaops.h>
25 #include <linux/init.h>
29 #define FILTER_ROOTINO 2
30 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
33 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
35 [S_IFREG >> S_SHIFT] "R",
36 [S_IFDIR >> S_SHIFT] "D",
37 [S_IFCHR >> S_SHIFT] "C",
38 [S_IFBLK >> S_SHIFT] "B",
39 [S_IFIFO >> S_SHIFT] "F",
40 [S_IFSOCK >> S_SHIFT] "S",
41 [S_IFLNK >> S_SHIFT] "L"
44 static inline const char *obd_mode_to_type(int mode)
46 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
49 /* write the pathname into the string */
50 static int filter_id(char *buf, obd_id id, obd_mode mode)
52 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
55 static inline void f_dput(struct dentry *dentry)
57 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
58 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
59 LASSERT(atomic_read(&dentry->d_count) > 0);
63 /* setup the object store with correct subdirectories */
64 static int filter_prep(struct obd_device *obddev)
66 struct obd_run_ctxt saved;
67 struct filter_obd *filter = &obddev->u.filter;
68 struct dentry *dentry;
76 push_ctxt(&saved, &filter->fo_ctxt, NULL);
77 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
78 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
81 CERROR("cannot open/create O: rc = %d\n", rc);
84 filter->fo_dentry_O = dentry;
85 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
86 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
89 CERROR("cannot open/create P: rc = %d\n", rc);
93 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
94 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
97 CERROR("cannot open/create D: rc = %d\n", rc);
101 root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
105 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
112 * Create directories and/or get dentries for each object type.
113 * This saves us from having to do multiple lookups for each one.
115 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
116 char *type = obd_type_by_mode[mode];
119 filter->fo_dentry_O_mode[mode] = NULL;
122 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
123 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
124 if (IS_ERR(dentry)) {
125 rc = PTR_ERR(dentry);
126 CERROR("cannot create O/%s: rc = %d\n", type, rc);
127 GOTO(out_O_mode, rc);
129 filter->fo_dentry_O_mode[mode] = dentry;
132 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
133 if ( !file || IS_ERR(file) ) {
135 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
137 GOTO(out_O_mode, rc);
140 /* steal operations */
141 inode = file->f_dentry->d_inode;
142 filter->fo_fop = file->f_op;
143 filter->fo_iop = inode->i_op;
144 filter->fo_aops = inode->i_mapping->a_ops;
146 if (inode->i_size == 0) {
147 __u64 disk_lastino = cpu_to_le64(lastino);
148 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
149 sizeof(disk_lastino),
151 if (retval != sizeof(disk_lastino)) {
152 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
154 GOTO(out_O_mode, rc = -EIO);
158 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
159 sizeof(disk_lastino),
161 if (retval != sizeof(disk_lastino)) {
162 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
164 GOTO(out_O_mode, rc = -EIO);
166 lastino = le64_to_cpu(disk_lastino);
168 filter->fo_lastino = lastino;
179 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
182 filter->fo_dentry_O_mode[mode] = NULL;
186 f_dput(filter->fo_dentry_O);
187 filter->fo_dentry_O = NULL;
191 /* cleanup the filter: write last used object id to status file */
192 static void filter_post(struct obd_device *obddev)
194 struct obd_run_ctxt saved;
195 struct filter_obd *filter = &obddev->u.filter;
201 push_ctxt(&saved, &filter->fo_ctxt, NULL);
202 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
204 CERROR("OBD filter: cannot create status file\n");
209 disk_lastino = cpu_to_le64(filter->fo_lastino);
210 rc = file->f_op->write(file, (char *)&disk_lastino,
211 sizeof(disk_lastino), &file->f_pos);
212 if (rc != sizeof(disk_lastino))
213 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
215 rc = filp_close(file, NULL);
217 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
219 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
220 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
223 filter->fo_dentry_O_mode[mode] = NULL;
226 f_dput(filter->fo_dentry_O);
232 static __u64 filter_next_id(struct obd_device *obddev)
236 spin_lock(&obddev->u.filter.fo_lock);
237 id = ++obddev->u.filter.fo_lastino;
238 spin_unlock(&obddev->u.filter.fo_lock);
240 /* FIXME: write the lastino to disk here */
244 /* how to get files, dentries, inodes from object id's */
245 /* parent i_sem is already held if needed for exclusivity */
246 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
247 struct dentry *dparent,
248 __u64 id, __u32 type)
250 struct super_block *sb = obddev->u.filter.fo_sb;
251 struct dentry *dchild;
256 if (!sb || !sb->s_dev) {
257 CERROR("fatal: device not initialized.\n");
258 RETURN(ERR_PTR(-ENXIO));
262 CERROR("fatal: invalid object #0\n");
264 RETURN(ERR_PTR(-ESTALE));
267 if (!(type & S_IFMT)) {
268 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
269 __FUNCTION__, id, type);
270 RETURN(ERR_PTR(-EINVAL));
273 len = sprintf(name, LPU64, id);
274 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
276 dchild = lookup_one_len(name, dparent, len);
277 if (IS_ERR(dchild)) {
278 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
282 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
283 obd_mode_to_type(type), name, dchild,
284 atomic_read(&dchild->d_count));
286 LASSERT(atomic_read(&dchild->d_count) > 0);
291 static struct file *filter_obj_open(struct obd_device *obddev,
292 __u64 id, __u32 type)
294 struct super_block *sb = obddev->u.filter.fo_sb;
295 struct obd_run_ctxt saved;
300 if (!sb || !sb->s_dev) {
301 CERROR("fatal: device not initialized.\n");
302 RETURN(ERR_PTR(-ENXIO));
306 CERROR("fatal: invalid obdo "LPU64"\n", id);
307 RETURN(ERR_PTR(-ESTALE));
310 if (!(type & S_IFMT)) {
311 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
312 __FUNCTION__, id, type);
313 RETURN(ERR_PTR(-EINVAL));
316 filter_id(name, id, type);
317 push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
318 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
321 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
328 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
330 struct filter_obd *filter = &obddev->u.filter;
332 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
337 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
343 rc = class_connect(conn, obd, cluuid);
349 static int filter_disconnect(struct lustre_handle *conn)
354 rc = class_disconnect(conn);
358 /* XXX cleanup preallocated inodes */
362 /* mount the file system (secretly) */
363 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
365 struct obd_ioctl_data* data = buf;
366 struct filter_obd *filter;
367 struct vfsmount *mnt;
371 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
375 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
380 filter = &obddev->u.filter;;
381 filter->fo_vfsmnt = mnt;
382 filter->fo_fstype = strdup(data->ioc_inlbuf2);
383 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
384 /* XXX is this even possible if do_kern_mount succeeded? */
386 GOTO(err_kfree, err = -ENODEV);
388 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
389 filter->fo_ctxt.pwdmnt = mnt;
390 filter->fo_ctxt.pwd = mnt->mnt_root;
391 filter->fo_ctxt.fs = get_ds();
393 err = filter_prep(obddev);
395 GOTO(err_kfree, err);
396 spin_lock_init(&filter->fo_lock);
398 obddev->obd_namespace =
399 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
400 if (obddev->obd_namespace == NULL)
403 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
404 "filter_ldlm_client", &obddev->obd_ldlm_client);
409 kfree(filter->fo_fstype);
411 mntput(filter->fo_vfsmnt);
421 static int filter_cleanup(struct obd_device * obddev)
423 struct super_block *sb;
426 if (!list_empty(&obddev->obd_exports)) {
427 CERROR("still has clients!\n");
428 class_disconnect_all(obddev);
429 if (!list_empty(&obddev->obd_exports)) {
430 CERROR("still has exports after forced cleanup?\n");
435 ldlm_namespace_free(obddev->obd_namespace);
437 sb = obddev->u.filter.fo_sb;
438 if (!obddev->u.filter.fo_sb)
443 shrink_dcache_parent(sb->s_root);
445 mntput(obddev->u.filter.fo_vfsmnt);
446 obddev->u.filter.fo_sb = 0;
447 kfree(obddev->u.filter.fo_fstype);
456 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
459 int type = oa->o_mode & S_IFMT;
462 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
463 inode->i_ino, inode, (long)oa->o_id, valid);
464 /* Don't copy the inode number in place of the object ID */
465 obdo_from_inode(oa, inode, valid);
466 oa->o_mode &= ~S_IFMT;
469 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
470 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
472 oa->o_valid |= OBD_MD_FLRDEV;
478 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
479 struct lov_stripe_md *md)
481 struct obd_device *obddev = class_conn2obd(conn);
482 struct dentry *dentry;
486 if (!class_conn2export(conn)) {
487 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
491 obddev = class_conn2obd(conn);
492 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
493 oa->o_id, oa->o_mode);
495 RETURN(PTR_ERR(dentry));
497 if (!dentry->d_inode) {
498 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
499 GOTO(out_getattr, rc = -ENOENT);
502 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
509 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
510 struct lov_stripe_md *md)
512 struct obd_run_ctxt saved;
513 struct obd_device *obd = class_conn2obd(conn);
514 struct dentry *dentry;
520 iattr_from_obdo(&iattr, oa, oa->o_valid);
521 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
522 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
523 oa->o_id, iattr.ia_mode);
525 RETURN(PTR_ERR(dentry));
527 inode = dentry->d_inode;
529 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
530 GOTO(out_setattr, rc = -ENOENT);
534 if (iattr.ia_valid & ATTR_SIZE)
536 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
537 if (inode->i_op->setattr)
538 rc = inode->i_op->setattr(dentry, &iattr);
540 rc = inode_setattr(inode, &iattr);
542 if (iattr.ia_valid & ATTR_SIZE) {
544 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
545 obdo_from_inode(oa, inode, oa->o_valid);
554 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
555 struct lov_stripe_md *ea)
557 struct obd_export *export;
558 struct obd_device *obd;
559 struct dentry *dentry;
562 export = class_conn2export(conn);
564 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
568 obd = export->exp_obd;
569 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
570 oa->o_id, oa->o_mode);
572 RETURN(PTR_ERR(dentry));
574 if (!dentry->d_inode) {
575 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
579 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
580 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
585 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
586 struct lov_stripe_md *ea)
588 struct obd_device *obd;
589 struct dentry *dentry;
592 obd = class_conn2obd(conn);
594 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
598 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
599 oa->o_id, oa->o_mode);
601 RETURN(PTR_ERR(dentry));
602 LASSERT(atomic_read(&dentry->d_count) > 1);
604 f_dput(dentry); /* for the close */
605 f_dput(dentry); /* for this call */
609 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
610 struct lov_stripe_md **ea)
613 struct obd_run_ctxt saved;
616 struct obd_device *obd = class_conn2obd(conn);
617 struct filter_obd *filter = &obd->u.filter;
622 CERROR("invalid client "LPX64"\n", conn->addr);
626 if (!(oa->o_mode && S_IFMT)) {
627 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
628 __FUNCTION__, oa->o_id, oa->o_mode);
632 oa->o_id = filter_next_id(obd);
634 //filter_id(name, oa->o_id, oa->o_mode);
635 sprintf(name, LPU64, oa->o_id);
636 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
637 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
638 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
642 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
646 /* Set flags for fields we have set in the inode struct */
647 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
648 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
649 filter_from_inode(oa, new->d_inode, oa->o_valid);
655 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
656 struct lov_stripe_md *ea)
658 struct obd_device *obd;
659 struct filter_obd *filter;
660 struct obd_run_ctxt saved;
662 struct dentry *dir_dentry, *object_dentry;
666 obd = class_conn2obd(conn);
668 CERROR("invalid client "LPX64"\n", conn->addr);
672 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
674 dir_dentry = filter_parent(obd, oa->o_mode);
675 down(&dir_dentry->d_inode->i_sem);
677 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
679 if (IS_ERR(object_dentry))
680 GOTO(out, rc = -ENOENT);
682 inode = object_dentry->d_inode;
684 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
685 GOTO(out, rc = -ENOENT);
688 if (inode->i_nlink != 1) {
689 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
693 inode->i_mode = S_IFREG;
695 if (atomic_read(&inode->i_count) > 1) {
696 #warning FIXME: need to handle open-unlinked case and move to hold dir
697 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
700 filter = &obd->u.filter;
701 push_ctxt(&saved, &filter->fo_ctxt, NULL);
703 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
705 f_dput(object_dentry);
709 up(&dir_dentry->d_inode->i_sem);
713 /* NB count and offset are used for punch, but not truncate */
714 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
715 struct lov_stripe_md *md,
716 obd_off start, obd_off end)
721 if (end != OBD_PUNCH_EOF)
722 CERROR("PUNCH not supported, only truncate works\n");
724 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
725 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
727 error = filter_setattr(conn, oa, NULL);
731 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
732 struct lov_stripe_md *lsm, obd_count oa_bufs,
733 struct brw_page *pga, brw_callback_t callback,
734 struct io_cb_data *data)
736 struct obd_run_ctxt saved;
737 struct super_block *sb;
738 int pnum; /* index to pages (bufs) */
739 unsigned long retval;
742 struct obd_device *obd = class_conn2obd(conn);
747 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
751 sb = obd->u.filter.fo_sb;
752 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
753 pnum = 0; /* pnum indexes buf 0..num_pages */
755 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
757 GOTO(out, retval = PTR_ERR(file));
759 /* count doubles as retval */
760 for (pg = 0; pg < oa_bufs; pg++) {
761 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
762 ") off count ("LPU64",%d)\n",
763 cmd, pnum, file->f_dentry->d_inode->i_ino,
764 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
765 (int)pga[pnum].count);
766 if (cmd & OBD_BRW_WRITE) {
770 buffer = kmap(pga[pnum].pg);
771 retval = file->f_op->write(file, buffer,
774 kunmap(pga[pnum].pg);
775 CDEBUG(D_INODE, "retval %ld\n", retval);
777 loff_t off = pga[pnum].off;
778 char *buffer = kmap(pga[pnum].pg);
780 if (off >= file->f_dentry->d_inode->i_size) {
781 memset(buffer, 0, pga[pnum].count);
782 retval = pga[pnum].count;
784 retval = file->f_op->read(file, buffer,
785 pga[pnum].count, &off);
787 kunmap(pga[pnum].pg);
789 if (retval != pga[pnum].count) {
791 GOTO(out, retval = -EIO);
793 CDEBUG(D_INODE, "retval %ld\n", retval);
797 /* sizes and blocks are set by generic_file_write */
798 /* ctimes/mtimes will follow with a setattr call */
801 /* XXX: do something with callback if it is set? */
806 error = (retval >= 0) ? 0 : retval;
811 * Calculate the number of buffer credits needed to write multiple pages in
812 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
813 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
815 * See comment above ext3_writepage_trans_blocks for details. We assume
816 * no data journaling is being done, but it does allow for all of the pages
817 * being non-contiguous. If we are guaranteed contiguous pages we could
818 * reduce the number of (d)indirect blocks a lot.
820 * With N blocks per page and P pages, for each inode we have at most:
822 * min(N*P, blocksize/4 + 1) dindirect blocks
825 * For the entire filesystem, we have at most:
826 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
827 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
830 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
832 static int ext3_credits_needed(struct super_block *sb, int objcount,
833 struct obd_ioobj *obj)
835 struct obd_ioobj *o = obj;
836 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
837 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
840 int needed = objcount + 1;
843 for (i = 0; i < objcount; i++, o++) {
844 int nblocks = o->ioo_bufcnt * blockpp;
845 int ndindirect = min(nblocks, addrpp + 1);
846 int nindir = nblocks + ndindirect + 1;
848 nbitmaps += nindir + nblocks;
849 ngdblocks += nindir + nblocks;
854 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
855 nbitmaps = EXT3_SB(sb)->s_groups_count;
856 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
857 ngdblocks = EXT3_SB(sb)->s_gdb_count;
859 needed += nbitmaps + ngdblocks;
862 /* We assume that there will be 1 bit set in s_dquot.flags for each
863 * quota file that is active. This is at least true for now.
865 needed += hweight32(sb_any_quota_enabled(sb)) *
866 EXT3_SINGLEDATA_TRANS_BLOCKS;
872 /* We have to start a huge journal transaction here to hold all of the
873 * metadata for the pages being written here. This is necessitated by
874 * the fact that we do lots of prepare_write operations before we do
875 * any of the matching commit_write operations, so even if we split
876 * up to use "smaller" transactions none of them could complete until
877 * all of them were opened. By having a single journal transaction,
878 * we eliminate duplicate reservations for common blocks like the
879 * superblock and group descriptors or bitmaps.
881 * We will start the transaction here, but each prepare_write will
882 * add a refcount to the transaction, and each commit_write will
883 * remove a refcount. The transaction will be closed when all of
884 * the pages have been written.
886 static void *ext3_filter_journal_start(struct filter_obd *filter,
887 int objcount, struct obd_ioobj *obj,
888 int niocount, struct niobuf_remote *nb)
890 journal_t *journal = NULL;
891 handle_t *handle = NULL;
894 /* Assumes ext3 and extN have same sb_info layout, but avoids issues
895 * with having extN built properly before filterobd for now.
897 journal = EXT3_SB(filter->fo_sb)->s_journal;
898 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
900 /* The number of blocks we could _possibly_ dirty can very large.
901 * We reduce our request if it is absurd (and we couldn't get that
902 * many credits for a single handle anyways).
904 * At some point we have to limit the size of I/Os sent at one time,
905 * increase the size of the journal, or we have to calculate the
906 * actual journal requirements more carefully by checking all of
907 * the blocks instead of being maximally pessimistic. It remains to
908 * be seen if this is a real problem or not.
910 if (needed > journal->j_max_transaction_buffers) {
911 CERROR("want too many journal credits (%d) using %d instead\n",
912 needed, journal->j_max_transaction_buffers);
913 needed = journal->j_max_transaction_buffers;
916 handle = journal_start(journal, needed);
918 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
924 static void *filter_journal_start(void **journal_save,
925 struct filter_obd *filter,
926 int objcount, struct obd_ioobj *obj,
927 int niocount, struct niobuf_remote *nb)
931 /* This may not be necessary - we probably never have a
932 * transaction started when we enter here, so we can
933 * remove the saving of the journal state entirely.
934 * For now leave it in just to see if it ever happens.
936 *journal_save = current->journal_info;
938 CERROR("Already have handle %p???\n", *journal_save);
940 current->journal_info = NULL;
943 if (!strcmp(filter->fo_fstype, "ext3") ||
944 !strcmp(filter->fo_fstype, "extN"))
945 handle = ext3_filter_journal_start(filter, objcount, obj,
950 static int ext3_filter_journal_stop(void *handle)
954 /* We got a refcount on the handle for each call to prepare_write,
955 * so we can drop the "parent" handle here to avoid the need for
956 * osc to call back into filterobd to close the handle. The
957 * remaining references will be dropped in commit_write.
959 rc = journal_stop((handle_t *)handle);
964 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
969 if (!strcmp(filter->fo_fstype, "ext3") ||
970 !strcmp(filter->fo_fstype, "extN"))
971 rc = ext3_filter_journal_stop(handle);
974 CERROR("error on journal stop: rc = %d\n", rc);
976 current->journal_info = journal_save;
981 static inline void lustre_put_page(struct page *page)
984 page_cache_release(page);
988 lustre_get_page_read(struct inode *inode, unsigned long index)
990 struct address_space *mapping = inode->i_mapping;
994 page = read_cache_page(mapping, index,
995 (filler_t*)mapping->a_ops->readpage, NULL);
999 if (!Page_Uptodate(page)) {
1000 CERROR("page index %lu not uptodate\n", index);
1001 GOTO(err_page, rc = -EIO);
1003 if (PageError(page)) {
1004 CERROR("page index %lu has error\n", index);
1005 GOTO(err_page, rc = -EIO);
1011 lustre_put_page(page);
1015 static struct page *
1016 lustre_get_page_write(struct inode *inode, unsigned long index)
1018 struct address_space *mapping = inode->i_mapping;
1022 page = grab_cache_page(mapping, index); /* locked page */
1024 if (!IS_ERR(page)) {
1026 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1027 * a no-op for most filesystems, because we write the whole
1028 * page. For partial-page I/O this will read in the page.
1030 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1032 CERROR("page index %lu, rc = %d\n", index, rc);
1035 GOTO(err_unlock, rc);
1037 /* XXX not sure if we need this if we are overwriting page */
1038 if (PageError(page)) {
1039 CERROR("error on page index %lu, rc = %d\n", index, rc);
1041 GOTO(err_unlock, rc = -EIO);
1048 lustre_put_page(page);
1052 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1054 struct inode *inode = page->mapping->host;
1057 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1058 if (!err && IS_SYNC(inode))
1059 err = waitfor_one_page(page);
1061 //SetPageUptodate(page); // the client commit_write will do this
1063 SetPageReferenced(page);
1065 lustre_put_page(page);
1069 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1070 struct niobuf_local *lnb, int *pglocked)
1072 struct address_space *mapping = inode->i_mapping;
1076 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1078 page = grab_cache_page_nowait(mapping, index); /* locked page */
1080 page = grab_cache_page(mapping, index); /* locked page */
1083 /* This page is currently locked, so get a temporary page instead. */
1084 /* XXX I believe this is a very dangerous thing to do - consider if
1085 * we had multiple writers for the same file (definitely the case
1086 * if we are using this codepath). If writer A locks the page,
1087 * writer B writes to a copy (as here), writer A drops the page
1088 * lock, and writer C grabs the lock before B does, then B will
1089 * later overwrite the data from C, even if C had LDLM locked
1090 * and initiated the write after B did.
1094 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1095 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1097 CERROR("no memory for a temp page\n");
1099 GOTO(err, rc = -ENOMEM);
1101 page = virt_to_page(addr);
1103 page->index = index;
1104 lnb->flags |= N_LOCAL_TEMP_PAGE;
1105 } else if (!IS_ERR(page)) {
1109 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1110 * a no-op for most filesystems, because we write the whole
1111 * page. For partial-page I/O this will read in the page.
1113 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1115 CERROR("page index %lu, rc = %d\n", index, rc);
1118 GOTO(err_unlock, rc);
1120 /* XXX not sure if we need this if we are overwriting page */
1121 if (PageError(page)) {
1122 CERROR("error on page index %lu, rc = %d\n", index, rc);
1124 GOTO(err_unlock, rc = -EIO);
1131 lustre_put_page(page);
1137 * We need to balance prepare_write() calls with commit_write() calls.
1138 * If the page has been prepared, but we have no data for it, we don't
1139 * want to overwrite valid data on disk, but we still need to zero out
1140 * data for space which was newly allocated. Like part of what happens
1141 * in __block_prepare_write() for newly allocated blocks.
1143 * XXX currently __block_prepare_write() creates buffers for all the
1144 * pages, and the filesystems mark these buffers as BH_New if they
1145 * were newly allocated from disk. We use the BH_New flag similarly.
1147 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1151 unsigned block_start, block_end;
1152 struct buffer_head *bh, *head = page->buffers;
1153 unsigned blocksize = head->b_size;
1154 void *addr = page_address(page);
1156 /* debugging: just seeing if this ever happens */
1157 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1158 page->index, page->mapping->host->i_ino, err);
1160 /* Currently one buffer per page, but in the future... */
1161 for (bh = head, block_start = 0; bh != head || !block_start;
1162 block_start = block_end, bh = bh->b_this_page) {
1163 block_end = block_start + blocksize;
1165 memset(addr + block_start, 0, blocksize);
1169 return lustre_commit_write(page, from, to);
1172 static int filter_preprw(int cmd, struct lustre_handle *conn,
1173 int objcount, struct obd_ioobj *obj,
1174 int niocount, struct niobuf_remote *nb,
1175 struct niobuf_local *res, void **desc_private)
1177 struct obd_run_ctxt saved;
1178 struct obd_device *obd;
1179 struct obd_ioobj *o = obj;
1180 struct niobuf_remote *b = nb;
1181 struct niobuf_local *r = res;
1182 void *journal_save = NULL;
1188 obd = class_conn2obd(conn);
1190 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1193 memset(res, 0, sizeof(*res) * niocount);
1195 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1197 if (cmd & OBD_BRW_WRITE) {
1198 *desc_private = filter_journal_start(&journal_save,
1200 objcount, obj, niocount,
1202 if (IS_ERR(*desc_private))
1203 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1206 for (i = 0; i < objcount; i++, o++) {
1207 struct dentry *dentry;
1208 struct inode *inode;
1211 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1212 o->ioo_id, S_IFREG);
1214 GOTO(out_clean, rc = PTR_ERR(dentry));
1215 inode = dentry->d_inode;
1217 CERROR("trying to BRW to non-existent file "LPU64"\n",
1220 GOTO(out_clean, rc = -ENOENT);
1223 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1224 unsigned long index = b->offset >> PAGE_SHIFT;
1230 r->dentry = dget(dentry);
1232 if (cmd & OBD_BRW_WRITE)
1233 page = filter_get_page_write(inode, index, r,
1236 page = lustre_get_page_read(inode, index);
1240 GOTO(out_clean, rc = PTR_ERR(page));
1243 r->addr = page_address(page);
1244 r->offset = b->offset;
1251 if (cmd & OBD_BRW_WRITE) {
1252 int err = filter_journal_stop(journal_save, &obd->u.filter,
1262 CERROR("error cleanup on brw\n");
1264 if (cmd & OBD_BRW_WRITE)
1265 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1267 lustre_put_page(r->page);
1272 static int filter_write_locked_page(struct niobuf_local *lnb)
1277 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1278 if (IS_ERR(lpage)) {
1279 /* It is highly unlikely that we would ever get an error here.
1280 * The page we want to get was previously locked, so it had to
1281 * have already allocated the space, and we were just writing
1282 * over the same data, so there would be no hole in the file.
1284 * XXX: possibility of a race with truncate could exist, need
1285 * to check that. There are no guarantees w.r.t.
1286 * write order even on a local filesystem, although the
1287 * normal response would be to return the number of bytes
1288 * successfully written and leave the rest to the app.
1290 rc = PTR_ERR(lpage);
1291 CERROR("error getting locked page index %ld: rc = %d\n",
1292 lnb->page->index, rc);
1296 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1297 * lustre_commit_write() below, lnb->page was kmapped previously in
1298 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1300 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1301 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1303 CERROR("error committing locked page %ld: rc = %d\n",
1304 lnb->page->index, rc);
1306 lustre_put_page(lnb->page);
1311 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1312 int objcount, struct obd_ioobj *obj,
1313 int niocount, struct niobuf_local *res,
1316 struct obd_run_ctxt saved;
1317 struct obd_ioobj *o;
1318 struct niobuf_local *r;
1319 struct obd_device *obd = class_conn2obd(conn);
1321 int found_locked = 0;
1326 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1327 journal_save = current->journal_info;
1328 LASSERT(!journal_save);
1330 current->journal_info = private;
1331 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1333 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1334 struct page *page = r->page;
1339 if (r->flags & N_LOCAL_TEMP_PAGE) {
1344 if (cmd & OBD_BRW_WRITE) {
1345 int err = filter_commit_write(page, 0,
1351 lustre_put_page(page);
1356 current->journal_info = journal_save;
1361 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1363 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1365 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1368 err = filter_write_locked_page(r);
1380 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1382 struct obd_device *obd = class_conn2obd(conn);
1387 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1389 statfs_pack(osfs, &sfs);
1394 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1395 void *key, obd_count *vallen, void **val)
1397 struct obd_device *obd;
1400 obd = class_conn2obd(conn);
1402 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1406 if ( keylen == strlen("blocksize") &&
1407 memcmp(key, "blocksize", keylen) == 0 ) {
1408 *vallen = sizeof(long);
1409 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1413 if ( keylen == strlen("blocksize_bits") &&
1414 memcmp(key, "blocksize_bits", keylen) == 0 ){
1415 *vallen = sizeof(long);
1416 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1420 if ( keylen == strlen("root_ino") &&
1421 memcmp(key, "root_ino", keylen) == 0 ){
1422 *vallen = sizeof(obd_id);
1423 *val = (void *)(obd_id)FILTER_ROOTINO;
1427 CDEBUG(D_IOCTL, "invalid key\n");
1431 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1432 struct lustre_handle *src_conn, struct obdo *src,
1433 obd_size count, obd_off offset)
1436 struct lov_stripe_md srcmd, dstmd;
1437 unsigned long index = 0;
1440 memset(&srcmd, 0, sizeof(srcmd));
1441 memset(&dstmd, 0, sizeof(dstmd));
1442 srcmd.lsm_object_id = src->o_id;
1443 dstmd.lsm_object_id = dst->o_id;
1446 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1447 ", dst: ino "LPU64"\n",
1448 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1449 page = alloc_page(GFP_USER);
1453 while (TryLockPage(page))
1454 ___wait_on_page(page);
1456 /* XXX with brw vector I/O, we could batch up reads and writes here,
1457 * all we need to do is allocate multiple pages to handle the I/Os
1458 * and arrays to handle the request parameters.
1460 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1462 struct io_cb_data *cbd = ll_init_cb();
1471 pg.count = PAGE_SIZE;
1472 pg.off = (page->index) << PAGE_SHIFT;
1475 page->index = index;
1476 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1477 ll_sync_io_cb, cbd);
1490 pg.flag = OBD_BRW_CREATE;
1491 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1493 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1494 ll_sync_io_cb, cbd);
1496 /* XXX should handle dst->o_size, dst->o_blocks here */
1502 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1506 dst->o_size = src->o_size;
1507 dst->o_blocks = src->o_blocks;
1508 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1516 static struct obd_ops filter_obd_ops = {
1517 o_get_info: filter_get_info,
1518 o_setup: filter_setup,
1519 o_cleanup: filter_cleanup,
1520 o_connect: filter_connect,
1521 o_disconnect: filter_disconnect,
1522 o_statfs: filter_statfs,
1523 o_getattr: filter_getattr,
1524 o_create: filter_create,
1525 o_setattr: filter_setattr,
1526 o_destroy: filter_destroy,
1527 o_open: filter_open,
1528 o_close: filter_close,
1529 o_brw: filter_pgcache_brw,
1530 o_punch: filter_truncate,
1531 o_preprw: filter_preprw,
1532 o_commitrw: filter_commitrw
1534 o_preallocate: filter_preallocate_inodes,
1535 o_migrate: filter_migrate,
1536 o_copy: filter_copy_data,
1537 o_iterate: filter_iterate
1542 static int __init obdfilter_init(void)
1544 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1545 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1548 static void __exit obdfilter_exit(void)
1550 class_unregister_type(OBD_FILTER_DEVICENAME);
1553 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1554 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1555 MODULE_LICENSE("GPL");
1557 module_init(obdfilter_init);
1558 module_exit(obdfilter_exit);