1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
8 * This code is issued under the GNU General Public License.
9 * See the file COPYING in this distribution
11 * by Peter Braam <braam@clusterfs.com>
12 * and Andreas Dilger <adilger@clusterfs.com>
16 #define DEBUG_SUBSYSTEM S_FILTER
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
21 #include <linux/dcache.h>
22 #include <linux/obd_class.h>
23 #include <linux/lustre_dlm.h>
24 #include <linux/obd_filter.h>
25 #include <linux/ext3_jbd.h>
26 #include <linux/extN_jbd.h>
27 #include <linux/quotaops.h>
28 #include <linux/init.h>
29 #include <linux/stringify.h>
33 #define FILTER_ROOTINO 2
34 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
37 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
39 [S_IFREG >> S_SHIFT] "R",
40 [S_IFDIR >> S_SHIFT] "D",
41 [S_IFCHR >> S_SHIFT] "C",
42 [S_IFBLK >> S_SHIFT] "B",
43 [S_IFIFO >> S_SHIFT] "F",
44 [S_IFSOCK >> S_SHIFT] "S",
45 [S_IFLNK >> S_SHIFT] "L"
48 static inline const char *obd_mode_to_type(int mode)
50 return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
53 /* write the pathname into the string */
54 static int filter_id(char *buf, obd_id id, obd_mode mode)
56 return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
59 static inline void f_dput(struct dentry *dentry)
61 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
62 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
63 LASSERT(atomic_read(&dentry->d_count) > 0);
67 /* setup the object store with correct subdirectories */
68 static int filter_prep(struct obd_device *obddev)
70 struct obd_run_ctxt saved;
71 struct filter_obd *filter = &obddev->u.filter;
72 struct dentry *dentry;
80 push_ctxt(&saved, &filter->fo_ctxt, NULL);
81 dentry = simple_mkdir(current->fs->pwd, "O", 0700);
82 CDEBUG(D_INODE, "got/created O: %p\n", dentry);
85 CERROR("cannot open/create O: rc = %d\n", rc);
88 filter->fo_dentry_O = dentry;
89 dentry = simple_mkdir(current->fs->pwd, "P", 0700);
90 CDEBUG(D_INODE, "got/created P: %p\n", dentry);
93 CERROR("cannot open/create P: rc = %d\n", rc);
97 dentry = simple_mkdir(current->fs->pwd, "D", 0700);
98 CDEBUG(D_INODE, "got/created D: %p\n", dentry);
100 rc = PTR_ERR(dentry);
101 CERROR("cannot open/create D: rc = %d\n", rc);
105 root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
109 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
116 * Create directories and/or get dentries for each object type.
117 * This saves us from having to do multiple lookups for each one.
119 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
120 char *type = obd_type_by_mode[mode];
123 filter->fo_dentry_O_mode[mode] = NULL;
126 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
127 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
128 if (IS_ERR(dentry)) {
129 rc = PTR_ERR(dentry);
130 CERROR("cannot create O/%s: rc = %d\n", type, rc);
131 GOTO(out_O_mode, rc);
133 filter->fo_dentry_O_mode[mode] = dentry;
136 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
137 if ( !file || IS_ERR(file) ) {
139 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
141 GOTO(out_O_mode, rc);
144 /* steal operations */
145 inode = file->f_dentry->d_inode;
146 filter->fo_fop = file->f_op;
147 filter->fo_iop = inode->i_op;
148 filter->fo_aops = inode->i_mapping->a_ops;
150 if (inode->i_size == 0) {
151 __u64 disk_lastino = cpu_to_le64(lastino);
152 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
153 sizeof(disk_lastino),
155 if (retval != sizeof(disk_lastino)) {
156 CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
158 GOTO(out_O_mode, rc = -EIO);
162 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
163 sizeof(disk_lastino),
165 if (retval != sizeof(disk_lastino)) {
166 CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
168 GOTO(out_O_mode, rc = -EIO);
170 lastino = le64_to_cpu(disk_lastino);
172 filter->fo_lastino = lastino;
183 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
186 filter->fo_dentry_O_mode[mode] = NULL;
190 f_dput(filter->fo_dentry_O);
191 filter->fo_dentry_O = NULL;
195 /* cleanup the filter: write last used object id to status file */
196 static void filter_post(struct obd_device *obddev)
198 struct obd_run_ctxt saved;
199 struct filter_obd *filter = &obddev->u.filter;
205 push_ctxt(&saved, &filter->fo_ctxt, NULL);
206 file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
208 CERROR("OBD filter: cannot create status file\n");
213 disk_lastino = cpu_to_le64(filter->fo_lastino);
214 rc = file->f_op->write(file, (char *)&disk_lastino,
215 sizeof(disk_lastino), &file->f_pos);
216 if (rc != sizeof(disk_lastino))
217 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
219 rc = filp_close(file, NULL);
221 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
223 for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
224 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
227 filter->fo_dentry_O_mode[mode] = NULL;
230 f_dput(filter->fo_dentry_O);
236 static __u64 filter_next_id(struct obd_device *obddev)
240 spin_lock(&obddev->u.filter.fo_lock);
241 id = ++obddev->u.filter.fo_lastino;
242 spin_unlock(&obddev->u.filter.fo_lock);
244 /* FIXME: write the lastino to disk here */
248 /* how to get files, dentries, inodes from object id's */
249 /* parent i_sem is already held if needed for exclusivity */
250 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
251 struct dentry *dparent,
252 __u64 id, __u32 type)
254 struct super_block *sb = obddev->u.filter.fo_sb;
255 struct dentry *dchild;
260 if (!sb || !sb->s_dev) {
261 CERROR("fatal: device not initialized.\n");
262 RETURN(ERR_PTR(-ENXIO));
266 CERROR("fatal: invalid object #0\n");
268 RETURN(ERR_PTR(-ESTALE));
271 if (!(type & S_IFMT)) {
272 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
273 __FUNCTION__, id, type);
274 RETURN(ERR_PTR(-EINVAL));
277 len = sprintf(name, LPU64, id);
278 CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
280 dchild = lookup_one_len(name, dparent, len);
281 if (IS_ERR(dchild)) {
282 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
286 CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
287 obd_mode_to_type(type), name, dchild,
288 atomic_read(&dchild->d_count));
290 LASSERT(atomic_read(&dchild->d_count) > 0);
295 static struct file *filter_obj_open(struct obd_device *obddev,
296 __u64 id, __u32 type)
298 struct super_block *sb = obddev->u.filter.fo_sb;
299 struct obd_run_ctxt saved;
304 if (!sb || !sb->s_dev) {
305 CERROR("fatal: device not initialized.\n");
306 RETURN(ERR_PTR(-ENXIO));
310 CERROR("fatal: invalid obdo "LPU64"\n", id);
311 RETURN(ERR_PTR(-ESTALE));
314 if (!(type & S_IFMT)) {
315 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
316 __FUNCTION__, id, type);
317 RETURN(ERR_PTR(-EINVAL));
320 filter_id(name, id, type);
321 push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
322 file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
325 CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
332 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
334 struct filter_obd *filter = &obddev->u.filter;
336 return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
341 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
342 obd_uuid_t cluuid, struct recovd_obd *recovd,
343 ptlrpc_recovery_cb_t recover)
348 rc = class_connect(conn, obd, cluuid);
354 static int filter_disconnect(struct lustre_handle *conn)
359 rc = class_disconnect(conn);
363 /* XXX cleanup preallocated inodes */
367 /* mount the file system (secretly) */
368 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
370 struct obd_ioctl_data* data = buf;
371 struct filter_obd *filter;
372 struct vfsmount *mnt;
376 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
380 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
385 filter = &obddev->u.filter;;
386 filter->fo_vfsmnt = mnt;
387 filter->fo_fstype = strdup(data->ioc_inlbuf2);
388 filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
389 /* XXX is this even possible if do_kern_mount succeeded? */
391 GOTO(err_kfree, err = -ENODEV);
393 OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
394 filter->fo_ctxt.pwdmnt = mnt;
395 filter->fo_ctxt.pwd = mnt->mnt_root;
396 filter->fo_ctxt.fs = get_ds();
398 err = filter_prep(obddev);
400 GOTO(err_kfree, err);
401 spin_lock_init(&filter->fo_lock);
403 obddev->obd_namespace =
404 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
405 if (obddev->obd_namespace == NULL)
408 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
409 "filter_ldlm_client", &obddev->obd_ldlm_client);
414 kfree(filter->fo_fstype);
416 mntput(filter->fo_vfsmnt);
426 static int filter_cleanup(struct obd_device * obddev)
428 struct super_block *sb;
431 if (!list_empty(&obddev->obd_exports)) {
432 CERROR("still has clients!\n");
433 class_disconnect_all(obddev);
434 if (!list_empty(&obddev->obd_exports)) {
435 CERROR("still has exports after forced cleanup?\n");
440 ldlm_namespace_free(obddev->obd_namespace);
442 sb = obddev->u.filter.fo_sb;
443 if (!obddev->u.filter.fo_sb)
448 shrink_dcache_parent(sb->s_root);
450 mntput(obddev->u.filter.fo_vfsmnt);
451 obddev->u.filter.fo_sb = 0;
452 kfree(obddev->u.filter.fo_fstype);
461 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
464 int type = oa->o_mode & S_IFMT;
467 CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
468 inode->i_ino, inode, (long)oa->o_id, valid);
469 /* Don't copy the inode number in place of the object ID */
470 obdo_from_inode(oa, inode, valid);
471 oa->o_mode &= ~S_IFMT;
474 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
475 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
477 oa->o_valid |= OBD_MD_FLRDEV;
483 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
484 struct lov_stripe_md *md)
486 struct obd_device *obddev = class_conn2obd(conn);
487 struct dentry *dentry;
491 if (!class_conn2export(conn)) {
492 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
496 obddev = class_conn2obd(conn);
497 dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
498 oa->o_id, oa->o_mode);
500 RETURN(PTR_ERR(dentry));
502 if (!dentry->d_inode) {
503 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
504 GOTO(out_getattr, rc = -ENOENT);
507 filter_from_inode(oa, dentry->d_inode, oa->o_valid);
514 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
515 struct lov_stripe_md *md)
517 struct obd_run_ctxt saved;
518 struct obd_device *obd = class_conn2obd(conn);
519 struct dentry *dentry;
525 iattr_from_obdo(&iattr, oa, oa->o_valid);
526 iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
527 dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
528 oa->o_id, iattr.ia_mode);
530 RETURN(PTR_ERR(dentry));
532 inode = dentry->d_inode;
534 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
535 GOTO(out_setattr, rc = -ENOENT);
539 if (iattr.ia_valid & ATTR_SIZE)
541 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
542 if (inode->i_op->setattr)
543 rc = inode->i_op->setattr(dentry, &iattr);
545 rc = inode_setattr(inode, &iattr);
547 if (iattr.ia_valid & ATTR_SIZE) {
549 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
550 obdo_from_inode(oa, inode, oa->o_valid);
559 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
560 struct lov_stripe_md *ea)
562 struct obd_export *export;
563 struct obd_device *obd;
564 struct dentry *dentry;
567 export = class_conn2export(conn);
569 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
573 obd = export->exp_obd;
574 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
575 oa->o_id, oa->o_mode);
577 RETURN(PTR_ERR(dentry));
579 if (!dentry->d_inode) {
580 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
584 filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
585 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
590 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
591 struct lov_stripe_md *ea)
593 struct obd_device *obd;
594 struct dentry *dentry;
597 obd = class_conn2obd(conn);
599 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
603 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
604 oa->o_id, oa->o_mode);
606 RETURN(PTR_ERR(dentry));
607 LASSERT(atomic_read(&dentry->d_count) > 1);
609 f_dput(dentry); /* for the close */
610 f_dput(dentry); /* for this call */
614 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
615 struct lov_stripe_md **ea)
618 struct obd_run_ctxt saved;
621 struct obd_device *obd = class_conn2obd(conn);
622 struct filter_obd *filter = &obd->u.filter;
627 CERROR("invalid client "LPX64"\n", conn->addr);
631 if (!(oa->o_mode && S_IFMT)) {
632 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
633 __FUNCTION__, oa->o_id, oa->o_mode);
637 oa->o_id = filter_next_id(obd);
639 //filter_id(name, oa->o_id, oa->o_mode);
640 sprintf(name, LPU64, oa->o_id);
641 mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
642 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
643 new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
647 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
651 /* Set flags for fields we have set in the inode struct */
652 oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
653 OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
654 filter_from_inode(oa, new->d_inode, oa->o_valid);
660 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
661 struct lov_stripe_md *ea)
663 struct obd_device *obd;
664 struct filter_obd *filter;
665 struct obd_run_ctxt saved;
667 struct dentry *dir_dentry, *object_dentry;
671 obd = class_conn2obd(conn);
673 CERROR("invalid client "LPX64"\n", conn->addr);
677 CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
679 dir_dentry = filter_parent(obd, oa->o_mode);
680 down(&dir_dentry->d_inode->i_sem);
682 object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
684 if (IS_ERR(object_dentry))
685 GOTO(out, rc = -ENOENT);
687 inode = object_dentry->d_inode;
689 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
690 GOTO(out, rc = -ENOENT);
693 if (inode->i_nlink != 1) {
694 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
698 inode->i_mode = S_IFREG;
700 if (atomic_read(&inode->i_count) > 1) {
701 #warning FIXME: need to handle open-unlinked case and move to hold dir
702 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
705 filter = &obd->u.filter;
706 push_ctxt(&saved, &filter->fo_ctxt, NULL);
708 rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
710 f_dput(object_dentry);
714 up(&dir_dentry->d_inode->i_sem);
718 /* NB count and offset are used for punch, but not truncate */
719 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
720 struct lov_stripe_md *md,
721 obd_off start, obd_off end)
726 if (end != OBD_OBJECT_EOF)
727 CERROR("PUNCH not supported, only truncate works\n");
729 CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
730 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
732 error = filter_setattr(conn, oa, NULL);
736 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
737 struct lov_stripe_md *lsm, obd_count oa_bufs,
738 struct brw_page *pga, brw_callback_t callback,
739 struct io_cb_data *data)
741 struct obd_run_ctxt saved;
742 struct super_block *sb;
743 int pnum; /* index to pages (bufs) */
744 unsigned long retval;
747 struct obd_device *obd = class_conn2obd(conn);
752 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
756 sb = obd->u.filter.fo_sb;
757 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
758 pnum = 0; /* pnum indexes buf 0..num_pages */
760 file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
762 GOTO(out, retval = PTR_ERR(file));
764 /* count doubles as retval */
765 for (pg = 0; pg < oa_bufs; pg++) {
766 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
767 ") off count ("LPU64",%d)\n",
768 cmd, pnum, file->f_dentry->d_inode->i_ino,
769 pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
770 (int)pga[pnum].count);
771 if (cmd & OBD_BRW_WRITE) {
775 buffer = kmap(pga[pnum].pg);
776 retval = file->f_op->write(file, buffer,
779 kunmap(pga[pnum].pg);
780 CDEBUG(D_INODE, "retval %ld\n", retval);
782 loff_t off = pga[pnum].off;
783 char *buffer = kmap(pga[pnum].pg);
785 if (off >= file->f_dentry->d_inode->i_size) {
786 memset(buffer, 0, pga[pnum].count);
787 retval = pga[pnum].count;
789 retval = file->f_op->read(file, buffer,
790 pga[pnum].count, &off);
792 kunmap(pga[pnum].pg);
794 if (retval != pga[pnum].count) {
796 GOTO(out, retval = -EIO);
798 CDEBUG(D_INODE, "retval %ld\n", retval);
802 /* sizes and blocks are set by generic_file_write */
803 /* ctimes/mtimes will follow with a setattr call */
806 /* XXX: do something with callback if it is set? */
811 error = (retval >= 0) ? 0 : retval;
816 * Calculate the number of buffer credits needed to write multiple pages in
817 * a single ext3/extN transaction. No, this shouldn't be here, but as yet
818 * ext3 doesn't have a nice API for calculating this sort of thing in advance.
820 * See comment above ext3_writepage_trans_blocks for details. We assume
821 * no data journaling is being done, but it does allow for all of the pages
822 * being non-contiguous. If we are guaranteed contiguous pages we could
823 * reduce the number of (d)indirect blocks a lot.
825 * With N blocks per page and P pages, for each inode we have at most:
827 * min(N*P, blocksize/4 + 1) dindirect blocks
830 * For the entire filesystem, we have at most:
831 * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
832 * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
835 * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
837 static int ext3_credits_needed(struct super_block *sb, int objcount,
838 struct obd_ioobj *obj)
840 struct obd_ioobj *o = obj;
841 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
842 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
845 int needed = objcount + 1;
848 for (i = 0; i < objcount; i++, o++) {
849 int nblocks = o->ioo_bufcnt * blockpp;
850 int ndindirect = min(nblocks, addrpp + 1);
851 int nindir = nblocks + ndindirect + 1;
853 nbitmaps += nindir + nblocks;
854 ngdblocks += nindir + nblocks;
859 /* Assumes ext3 and extN have same sb_info layout at the start. */
860 if (nbitmaps > EXT3_SB(sb)->s_groups_count)
861 nbitmaps = EXT3_SB(sb)->s_groups_count;
862 if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
863 ngdblocks = EXT3_SB(sb)->s_gdb_count;
865 needed += nbitmaps + ngdblocks;
868 /* We assume that there will be 1 bit set in s_dquot.flags for each
869 * quota file that is active. This is at least true for now.
871 needed += hweight32(sb_any_quota_enabled(sb)) *
872 EXT3_SINGLEDATA_TRANS_BLOCKS;
878 /* We have to start a huge journal transaction here to hold all of the
879 * metadata for the pages being written here. This is necessitated by
880 * the fact that we do lots of prepare_write operations before we do
881 * any of the matching commit_write operations, so even if we split
882 * up to use "smaller" transactions none of them could complete until
883 * all of them were opened. By having a single journal transaction,
884 * we eliminate duplicate reservations for common blocks like the
885 * superblock and group descriptors or bitmaps.
887 * We will start the transaction here, but each prepare_write will
888 * add a refcount to the transaction, and each commit_write will
889 * remove a refcount. The transaction will be closed when all of
890 * the pages have been written.
892 static void *ext3_filter_journal_start(struct filter_obd *filter,
893 int objcount, struct obd_ioobj *obj,
894 int niocount, struct niobuf_remote *nb)
896 journal_t *journal = NULL;
897 handle_t *handle = NULL;
900 /* It appears that some kernels have different values for
901 * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
902 * assume anything after s_inode_bitmap_number is the same.
904 if (!strcmp(filter->fo_fstype, "ext3"))
905 journal = EXT3_SB(filter->fo_sb)->s_journal;
906 else if (!strcmp(filter->fo_fstype, "extN"))
907 journal = EXTN_SB(filter->fo_sb)->s_journal;
908 needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
910 /* The number of blocks we could _possibly_ dirty can very large.
911 * We reduce our request if it is absurd (and we couldn't get that
912 * many credits for a single handle anyways).
914 * At some point we have to limit the size of I/Os sent at one time,
915 * increase the size of the journal, or we have to calculate the
916 * actual journal requirements more carefully by checking all of
917 * the blocks instead of being maximally pessimistic. It remains to
918 * be seen if this is a real problem or not.
920 if (needed > journal->j_max_transaction_buffers) {
921 CERROR("want too many journal credits (%d) using %d instead\n",
922 needed, journal->j_max_transaction_buffers);
923 needed = journal->j_max_transaction_buffers;
927 handle = journal_start(journal, needed);
930 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
936 static void *filter_journal_start(void **journal_save,
937 struct filter_obd *filter,
938 int objcount, struct obd_ioobj *obj,
939 int niocount, struct niobuf_remote *nb)
943 /* This may not be necessary - we probably never have a
944 * transaction started when we enter here, so we can
945 * remove the saving of the journal state entirely.
946 * For now leave it in just to see if it ever happens.
948 *journal_save = current->journal_info;
950 CERROR("Already have handle %p???\n", *journal_save);
952 current->journal_info = NULL;
955 if (!strcmp(filter->fo_fstype, "ext3") ||
956 !strcmp(filter->fo_fstype, "extN"))
957 handle = ext3_filter_journal_start(filter, objcount, obj,
962 static int ext3_filter_journal_stop(void *handle)
966 /* We got a refcount on the handle for each call to prepare_write,
967 * so we can drop the "parent" handle here to avoid the need for
968 * osc to call back into filterobd to close the handle. The
969 * remaining references will be dropped in commit_write.
972 rc = journal_stop((handle_t *)handle);
978 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
983 if (!strcmp(filter->fo_fstype, "ext3") ||
984 !strcmp(filter->fo_fstype, "extN"))
985 rc = ext3_filter_journal_stop(handle);
988 CERROR("error on journal stop: rc = %d\n", rc);
990 current->journal_info = journal_save;
995 static inline void lustre_put_page(struct page *page)
998 page_cache_release(page);
1001 #ifndef PageUptodate
1002 #define PageUptodate(page) Page_Uptodate(page)
1005 static struct page *
1006 lustre_get_page_read(struct inode *inode, unsigned long index)
1008 struct address_space *mapping = inode->i_mapping;
1012 page = read_cache_page(mapping, index,
1013 (filler_t*)mapping->a_ops->readpage, NULL);
1014 if (!IS_ERR(page)) {
1017 if (!PageUptodate(page)) {
1018 CERROR("page index %lu not uptodate\n", index);
1019 GOTO(err_page, rc = -EIO);
1021 if (PageError(page)) {
1022 CERROR("page index %lu has error\n", index);
1023 GOTO(err_page, rc = -EIO);
1029 lustre_put_page(page);
1033 static struct page *
1034 lustre_get_page_write(struct inode *inode, unsigned long index)
1036 struct address_space *mapping = inode->i_mapping;
1040 page = grab_cache_page(mapping, index); /* locked page */
1042 if (!IS_ERR(page)) {
1044 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1045 * a no-op for most filesystems, because we write the whole
1046 * page. For partial-page I/O this will read in the page.
1048 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1050 CERROR("page index %lu, rc = %d\n", index, rc);
1053 GOTO(err_unlock, rc);
1055 /* XXX not sure if we need this if we are overwriting page */
1056 if (PageError(page)) {
1057 CERROR("error on page index %lu, rc = %d\n", index, rc);
1059 GOTO(err_unlock, rc = -EIO);
1066 lustre_put_page(page);
1070 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1072 struct inode *inode = page->mapping->host;
1075 err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1076 if (!err && IS_SYNC(inode))
1077 err = waitfor_one_page(page);
1079 //SetPageUptodate(page); // the client commit_write will do this
1081 SetPageReferenced(page);
1083 lustre_put_page(page);
1087 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1088 struct niobuf_local *lnb, int *pglocked)
1090 struct address_space *mapping = inode->i_mapping;
1094 //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1096 page = grab_cache_page_nowait(mapping, index); /* locked page */
1098 page = grab_cache_page(mapping, index); /* locked page */
1101 /* This page is currently locked, so get a temporary page instead. */
1102 /* XXX I believe this is a very dangerous thing to do - consider if
1103 * we had multiple writers for the same file (definitely the case
1104 * if we are using this codepath). If writer A locks the page,
1105 * writer B writes to a copy (as here), writer A drops the page
1106 * lock, and writer C grabs the lock before B does, then B will
1107 * later overwrite the data from C, even if C had LDLM locked
1108 * and initiated the write after B did.
1112 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1113 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1115 CERROR("no memory for a temp page\n");
1117 GOTO(err, rc = -ENOMEM);
1119 page = virt_to_page(addr);
1121 page->index = index;
1122 lnb->flags |= N_LOCAL_TEMP_PAGE;
1123 } else if (!IS_ERR(page)) {
1127 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1128 * a no-op for most filesystems, because we write the whole
1129 * page. For partial-page I/O this will read in the page.
1131 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1133 CERROR("page index %lu, rc = %d\n", index, rc);
1136 GOTO(err_unlock, rc);
1138 /* XXX not sure if we need this if we are overwriting page */
1139 if (PageError(page)) {
1140 CERROR("error on page index %lu, rc = %d\n", index, rc);
1142 GOTO(err_unlock, rc = -EIO);
1149 lustre_put_page(page);
1155 * We need to balance prepare_write() calls with commit_write() calls.
1156 * If the page has been prepared, but we have no data for it, we don't
1157 * want to overwrite valid data on disk, but we still need to zero out
1158 * data for space which was newly allocated. Like part of what happens
1159 * in __block_prepare_write() for newly allocated blocks.
1161 * XXX currently __block_prepare_write() creates buffers for all the
1162 * pages, and the filesystems mark these buffers as BH_New if they
1163 * were newly allocated from disk. We use the BH_New flag similarly.
1165 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1169 unsigned block_start, block_end;
1170 struct buffer_head *bh, *head = page->buffers;
1171 unsigned blocksize = head->b_size;
1172 void *addr = page_address(page);
1174 /* debugging: just seeing if this ever happens */
1175 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1176 page->index, page->mapping->host->i_ino, err);
1178 /* Currently one buffer per page, but in the future... */
1179 for (bh = head, block_start = 0; bh != head || !block_start;
1180 block_start = block_end, bh = bh->b_this_page) {
1181 block_end = block_start + blocksize;
1183 memset(addr + block_start, 0, blocksize);
1187 return lustre_commit_write(page, from, to);
1190 static int filter_preprw(int cmd, struct lustre_handle *conn,
1191 int objcount, struct obd_ioobj *obj,
1192 int niocount, struct niobuf_remote *nb,
1193 struct niobuf_local *res, void **desc_private)
1195 struct obd_run_ctxt saved;
1196 struct obd_device *obd;
1197 struct obd_ioobj *o = obj;
1198 struct niobuf_remote *b = nb;
1199 struct niobuf_local *r = res;
1200 void *journal_save = NULL;
1206 obd = class_conn2obd(conn);
1208 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1211 memset(res, 0, sizeof(*res) * niocount);
1213 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1215 if (cmd & OBD_BRW_WRITE) {
1216 *desc_private = filter_journal_start(&journal_save,
1218 objcount, obj, niocount,
1220 if (IS_ERR(*desc_private))
1221 GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1224 for (i = 0; i < objcount; i++, o++) {
1225 struct dentry *dentry;
1226 struct inode *inode;
1229 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1230 o->ioo_id, S_IFREG);
1232 GOTO(out_clean, rc = PTR_ERR(dentry));
1233 inode = dentry->d_inode;
1235 CERROR("trying to BRW to non-existent file "LPU64"\n",
1238 GOTO(out_clean, rc = -ENOENT);
1241 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1242 unsigned long index = b->offset >> PAGE_SHIFT;
1248 r->dentry = dget(dentry);
1250 if (cmd & OBD_BRW_WRITE)
1251 page = filter_get_page_write(inode, index, r,
1254 page = lustre_get_page_read(inode, index);
1258 GOTO(out_clean, rc = PTR_ERR(page));
1261 r->addr = page_address(page);
1262 r->offset = b->offset;
1269 if (cmd & OBD_BRW_WRITE) {
1270 int err = filter_journal_stop(journal_save, &obd->u.filter,
1280 CERROR("error cleanup on brw\n");
1282 if (cmd & OBD_BRW_WRITE)
1283 filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1285 lustre_put_page(r->page);
1290 static int filter_write_locked_page(struct niobuf_local *lnb)
1295 lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1296 if (IS_ERR(lpage)) {
1297 /* It is highly unlikely that we would ever get an error here.
1298 * The page we want to get was previously locked, so it had to
1299 * have already allocated the space, and we were just writing
1300 * over the same data, so there would be no hole in the file.
1302 * XXX: possibility of a race with truncate could exist, need
1303 * to check that. There are no guarantees w.r.t.
1304 * write order even on a local filesystem, although the
1305 * normal response would be to return the number of bytes
1306 * successfully written and leave the rest to the app.
1308 rc = PTR_ERR(lpage);
1309 CERROR("error getting locked page index %ld: rc = %d\n",
1310 lnb->page->index, rc);
1314 /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1315 * lustre_commit_write() below, lnb->page was kmapped previously in
1316 * filter_get_page_write() and kunmapped in lustre_put_page() below.
1318 memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1319 rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1321 CERROR("error committing locked page %ld: rc = %d\n",
1322 lnb->page->index, rc);
1324 lustre_put_page(lnb->page);
1329 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1330 int objcount, struct obd_ioobj *obj,
1331 int niocount, struct niobuf_local *res,
1334 struct obd_run_ctxt saved;
1335 struct obd_ioobj *o;
1336 struct niobuf_local *r;
1337 struct obd_device *obd = class_conn2obd(conn);
1339 int found_locked = 0;
1344 push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1345 journal_save = current->journal_info;
1346 LASSERT(!journal_save);
1348 current->journal_info = private;
1349 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1351 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1352 struct page *page = r->page;
1357 if (r->flags & N_LOCAL_TEMP_PAGE) {
1362 if (cmd & OBD_BRW_WRITE) {
1363 int err = filter_commit_write(page, 0,
1369 lustre_put_page(page);
1374 current->journal_info = journal_save;
1379 for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1381 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1383 if (!(r->flags & N_LOCAL_TEMP_PAGE))
1386 err = filter_write_locked_page(r);
1398 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1400 struct obd_device *obd = class_conn2obd(conn);
1405 rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1407 statfs_pack(osfs, &sfs);
1412 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1413 void *key, obd_count *vallen, void **val)
1415 struct obd_device *obd;
1418 obd = class_conn2obd(conn);
1420 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1424 if ( keylen == strlen("blocksize") &&
1425 memcmp(key, "blocksize", keylen) == 0 ) {
1426 *vallen = sizeof(long);
1427 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1431 if ( keylen == strlen("blocksize_bits") &&
1432 memcmp(key, "blocksize_bits", keylen) == 0 ){
1433 *vallen = sizeof(long);
1434 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1438 if ( keylen == strlen("root_ino") &&
1439 memcmp(key, "root_ino", keylen) == 0 ){
1440 *vallen = sizeof(obd_id);
1441 *val = (void *)(obd_id)FILTER_ROOTINO;
1445 CDEBUG(D_IOCTL, "invalid key\n");
1449 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1450 struct lustre_handle *src_conn, struct obdo *src,
1451 obd_size count, obd_off offset)
1454 struct lov_stripe_md srcmd, dstmd;
1455 unsigned long index = 0;
1458 memset(&srcmd, 0, sizeof(srcmd));
1459 memset(&dstmd, 0, sizeof(dstmd));
1460 srcmd.lsm_object_id = src->o_id;
1461 dstmd.lsm_object_id = dst->o_id;
1464 CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1465 ", dst: ino "LPU64"\n",
1466 src->o_id, src->o_blocks, src->o_size, dst->o_id);
1467 page = alloc_page(GFP_USER);
1471 while (TryLockPage(page))
1472 ___wait_on_page(page);
1474 /* XXX with brw vector I/O, we could batch up reads and writes here,
1475 * all we need to do is allocate multiple pages to handle the I/Os
1476 * and arrays to handle the request parameters.
1478 while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1480 struct io_cb_data *cbd = ll_init_cb();
1489 pg.count = PAGE_SIZE;
1490 pg.off = (page->index) << PAGE_SHIFT;
1493 page->index = index;
1494 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1495 ll_sync_io_cb, cbd);
1508 pg.flag = OBD_BRW_CREATE;
1509 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1511 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1512 ll_sync_io_cb, cbd);
1514 /* XXX should handle dst->o_size, dst->o_blocks here */
1520 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1524 dst->o_size = src->o_size;
1525 dst->o_blocks = src->o_blocks;
1526 dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1534 static struct obd_ops filter_obd_ops = {
1535 o_get_info: filter_get_info,
1536 o_setup: filter_setup,
1537 o_cleanup: filter_cleanup,
1538 o_connect: filter_connect,
1539 o_disconnect: filter_disconnect,
1540 o_statfs: filter_statfs,
1541 o_getattr: filter_getattr,
1542 o_create: filter_create,
1543 o_setattr: filter_setattr,
1544 o_destroy: filter_destroy,
1545 o_open: filter_open,
1546 o_close: filter_close,
1547 o_brw: filter_pgcache_brw,
1548 o_punch: filter_truncate,
1549 o_preprw: filter_preprw,
1550 o_commitrw: filter_commitrw
1552 o_preallocate: filter_preallocate_inodes,
1553 o_migrate: filter_migrate,
1554 o_copy: filter_copy_data,
1555 o_iterate: filter_iterate
1560 static int __init obdfilter_init(void)
1562 printk(KERN_INFO "Filtering OBD driver v0.001, info@clusterfs.com\n");
1563 return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1566 static void __exit obdfilter_exit(void)
1568 class_unregister_type(OBD_FILTER_DEVICENAME);
1571 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1572 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1573 MODULE_LICENSE("GPL");
1575 module_init(obdfilter_init);
1576 module_exit(obdfilter_exit);