1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
6 * Lustre Metadata Server (mds) reintegration routines
8 * Copyright (C) 2002 Cluster File Systems, Inc.
9 * author: Peter Braam <braam@clusterfs.com>
11 * This code is issued under the GNU General Public License.
12 * See the file COPYING in this distribution
16 // XXX - add transaction sequence numbers
19 #define DEBUG_SUBSYSTEM S_MDS
21 #include <linux/obd_support.h>
22 #include <linux/obd_class.h>
23 #include <linux/obd.h>
24 #include <linux/lustre_lib.h>
25 #include <linux/lustre_idl.h>
26 #include <linux/lustre_mds.h>
27 #include <linux/lustre_dlm.h>
28 #include <linux/obd_class.h>
30 extern inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req);
32 /* Assumes caller has already pushed us into the kernel context. */
33 int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
34 struct ptlrpc_request *req)
36 struct mds_export_data *med = &req->rq_export->exp_mds_data;
37 struct mds_client_data *mcd = med->med_mcd;
41 off = MDS_LR_CLIENT + med->med_off * MDS_LR_SIZE;
43 ++mds->mds_last_rcvd; /* lock this, or make it an LDLM function? */
44 req->rq_repmsg->transno = HTON__u64(mds->mds_last_rcvd);
45 mcd->mcd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
46 mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
47 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
49 mds_fs_set_last_rcvd(mds, handle);
50 rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mcd, sizeof(*mcd), &off);
51 CDEBUG(D_INODE, "wrote trans #%Ld for client '%s' at #%d: rc = %d\n",
52 mds->mds_last_rcvd, mcd->mcd_uuid, med->med_off, rc);
53 // store new value and last committed value in req struct
55 if (rc == sizeof(*mcd))
58 CERROR("error writing to last_rcvd file: rc = %d\n", rc);
66 /* In the write-back case, the client holds a lock on a subtree.
67 * In the intent case, the client holds a lock on the child inode.
68 * In the pathname case, the client (may) hold a lock on the child inode. */
69 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
70 struct ptlrpc_request *req)
72 struct mds_obd *mds = mds_req2mds(req);
73 struct obd_device *obd = req->rq_export->exp_obd;
76 struct lustre_handle child_lockh;
79 if (req->rq_reqmsg->bufcount > offset + 1) {
81 struct lustre_handle dir_lockh;
85 /* a name was supplied by the client; fid1 is the directory */
86 dir = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PR,
90 GOTO(out_setattr, rc = PTR_ERR(dir));
93 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
94 namelen = req->rq_reqmsg->buflens[offset + 1] - 1;
95 de = mds_name2locked_dentry(obd, dir, NULL, name, namelen,
96 0, &child_lockh, LCK_PR);
100 GOTO(out_setattr_de, rc = PTR_ERR(de));
103 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
104 if (!de || IS_ERR(de)) {
106 GOTO(out_setattr_de, rc = -ESTALE);
109 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
111 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE,
112 de->d_inode->i_sb->s_dev);
115 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
117 GOTO(out_unlock, rc = PTR_ERR(handle));
118 rc = mds_fs_setattr(mds, de, handle, &rec->ur_iattr);
121 rc = mds_update_last_rcvd(mds, handle, req);
123 err = mds_fs_commit(mds, de->d_inode, handle);
125 CERROR("error on commit: err = %d\n", err);
140 static int mds_reint_recreate(struct mds_update_record *rec, int offset,
141 struct ptlrpc_request *req)
143 struct dentry *de = NULL;
144 struct mds_obd *mds = mds_req2mds(req);
145 struct dentry *dchild = NULL;
150 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
151 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
153 GOTO(out_create_de, rc = -ESTALE);
156 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
159 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
160 if (IS_ERR(dchild)) {
161 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
164 GOTO(out_create_dchild, rc = -ESTALE);
167 if (dchild->d_inode) {
168 struct mds_body *body;
170 body = lustre_msg_buf(req->rq_repmsg, 0);
171 body->ino = dchild->d_inode->i_ino;
172 body->generation = dchild->d_inode->i_generation;
174 CERROR("child doesn't exist (dir %ld, name %s)\n",
175 dir->i_ino, rec->ur_name);
189 static int mds_reint_create(struct mds_update_record *rec, int offset,
190 struct ptlrpc_request *req)
192 struct dentry *de = NULL;
193 struct mds_obd *mds = mds_req2mds(req);
194 struct obd_device *obd = req->rq_export->exp_obd;
195 struct dentry *dchild = NULL;
198 struct lustre_handle lockh;
199 int rc = 0, err, lock_mode, type = rec->ur_mode & S_IFMT;
202 /* requests were at offset 2, replies go back at 1 */
206 if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
209 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
211 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
212 GOTO(out_create, rc = -ESTALE);
214 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, lock_mode, &lockh);
217 CERROR("parent lookup error %d\n", rc);
219 GOTO(out_create, rc);
222 CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
223 dir->i_ino, rec->ur_name, rec->ur_mode);
225 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
228 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
229 if (IS_ERR(dchild)) {
230 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
232 GOTO(out_create_de, rc = -ESTALE);
235 if (dchild->d_inode) {
236 struct mds_body *body;
237 struct inode *inode = dchild->d_inode;
238 struct lov_stripe_md *md;
239 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
240 dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
242 body = lustre_msg_buf(req->rq_repmsg, offset);
243 mds_pack_inode2fid(&body->fid1, inode);
244 mds_pack_inode2body(body, inode);
245 if (S_ISREG(inode->i_mode)) {
246 md = lustre_msg_buf(req->rq_repmsg, offset + 1);
247 md->lmd_easize = mds->mds_max_mdsize;
249 if (mds_fs_get_md(mds, inode, md) < 0)
250 memset(md, 0, md->lmd_easize);
252 /* now a normal case for intent locking */
253 GOTO(out_create_dchild, rc = -EEXIST);
256 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
258 if (dir->i_mode & S_ISGID) {
259 rec->ur_gid = dir->i_gid;
260 if (S_ISDIR(rec->ur_mode))
261 rec->ur_mode |= S_ISGID;
266 handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
268 GOTO(out_create_dchild, PTR_ERR(handle));
269 rc = vfs_create(dir, dchild, rec->ur_mode);
274 handle = mds_fs_start(mds, dir, MDS_FSOP_MKDIR);
276 GOTO(out_create_dchild, PTR_ERR(handle));
277 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
282 handle = mds_fs_start(mds, dir, MDS_FSOP_SYMLINK);
284 GOTO(out_create_dchild, PTR_ERR(handle));
285 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
293 int rdev = rec->ur_rdev;
294 handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
296 GOTO(out_create_dchild, PTR_ERR(handle));
297 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
302 CERROR("bad file type %o for create of %s\n", type,
304 GOTO(out_create_dchild, rc = -EINVAL);
308 CERROR("error during create: %d\n", rc);
311 GOTO(out_create_commit, rc);
314 struct inode *inode = dchild->d_inode;
315 struct mds_body *body;
317 CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
318 if (!offset && type == S_IFREG) {
319 struct lov_stripe_md *md;
320 md = lustre_msg_buf(req->rq_reqmsg, 2);
321 rc = mds_fs_set_md(mds, inode, handle, md);
323 CERROR("error %d setting LOV MD for %ld\n",
325 GOTO(out_create_unlink, rc);
329 iattr.ia_atime = rec->ur_time;
330 iattr.ia_ctime = rec->ur_time;
331 iattr.ia_mtime = rec->ur_time;
332 iattr.ia_uid = rec->ur_uid;
333 iattr.ia_gid = rec->ur_gid;
334 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
335 ATTR_MTIME | ATTR_CTIME;
337 rc = mds_fs_setattr(mds, dchild, handle, &iattr);
339 CERROR("error on setattr: rc = %d\n", rc);
340 /* XXX should we abort here in case of error? */
343 rc = mds_update_last_rcvd(mds, handle, req);
345 CERROR("error on update_last_rcvd: rc = %d\n", rc);
346 /* XXX should we abort here in case of error? */
349 body = lustre_msg_buf(req->rq_repmsg, offset);
350 body->ino = inode->i_ino;
351 body->generation = inode->i_generation;
352 body->valid = OBD_MD_FLID | OBD_MD_FLGENER;
356 err = mds_fs_commit(mds, dir, handle);
358 CERROR("error on commit: err = %d\n", err);
364 ldlm_lock_decref(&lockh, lock_mode);
373 /* Destroy the file we just created. This should not need extra
374 * journal credits, as we have already modified all of the blocks
375 * needed in order to create the file in the first place.
379 err = vfs_rmdir(dir, dchild);
381 CERROR("failed rmdir in error path: rc = %d\n", err);
384 err = vfs_unlink(dir, dchild);
386 CERROR("failed unlink in error path: rc = %d\n", err);
389 goto out_create_commit;
392 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
393 struct ptlrpc_request *req)
395 struct dentry *de = NULL;
396 struct dentry *dchild = NULL;
397 struct mds_obd *mds = mds_req2mds(req);
398 struct obd_device *obd = req->rq_export->exp_obd;
401 struct inode *dir, *inode;
403 struct lustre_handle lockh, child_lockh;
409 /* a name was supplied by the client; fid1 is the directory */
410 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
411 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, lock_mode, &lockh);
417 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
418 namelen = req->rq_reqmsg->buflens[offset + 1] - 1;
419 dchild = mds_name2locked_dentry(obd, de, NULL, name, namelen,
420 LCK_EX, &child_lockh, lock_mode);
422 if (IS_ERR(dchild) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) {
424 GOTO(out_unlink, rc = PTR_ERR(dchild));
428 inode = dchild->d_inode;
429 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
432 CDEBUG(D_INODE, "child doesn't exist (dir %ld, name %s\n",
433 dir->i_ino, rec->ur_name);
434 /* XXX should be out_unlink_cancel, or do we keep child_lockh? */
435 GOTO(out_unlink_dchild, rc = -ENOENT);
437 struct mds_body *body = lustre_msg_buf(req->rq_repmsg, 1);
438 mds_pack_inode2fid(&body->fid1, inode);
439 mds_pack_inode2body(body, inode);
442 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
444 switch (rec->ur_mode) {
446 handle = mds_fs_start(mds, dir, MDS_FSOP_RMDIR);
448 GOTO(out_unlink_cancel, rc = PTR_ERR(handle));
449 rc = vfs_rmdir(dir, dchild);
452 if ((inode->i_mode & S_IFMT) == S_IFREG && offset) {
453 struct lov_stripe_md *md;
455 md = lustre_msg_buf(req->rq_repmsg, 2);
456 md->lmd_easize = mds->mds_max_mdsize;
457 if ((rc = mds_fs_get_md(mds, inode, md)) < 0) {
458 CDEBUG(D_INFO, "No md for ino %ld: rc = %d\n",
460 memset(md, 0, md->lmd_easize);
463 handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
465 GOTO(out_unlink_cancel, rc = PTR_ERR(handle));
466 rc = vfs_unlink(dir, dchild);
473 rc = mds_update_last_rcvd(mds, handle, req);
474 err = mds_fs_commit(mds, dir, handle);
476 CERROR("error on commit: err = %d\n", err);
484 ldlm_lock_decref(&child_lockh, LCK_EX);
485 err = ldlm_cli_cancel(&child_lockh);
487 CERROR("failed to cancel child inode lock: err = %d\n", err);
489 rc = -ENOLCK; /*XXX translate LDLM lock error */
495 ldlm_lock_decref(&lockh, lock_mode);
501 static int mds_reint_link(struct mds_update_record *rec, int offset,
502 struct ptlrpc_request *req)
504 struct dentry *de_src = NULL;
505 struct dentry *de_tgt_dir = NULL;
506 struct dentry *dchild = NULL;
507 struct mds_obd *mds = mds_req2mds(req);
513 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
514 if (IS_ERR(de_src) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) {
515 GOTO(out_link, rc = -ESTALE);
518 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
519 if (IS_ERR(de_tgt_dir)) {
520 GOTO(out_link_de_src, rc = -ESTALE);
523 down(&de_tgt_dir->d_inode->i_sem);
524 dchild = lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen - 1);
525 if (IS_ERR(dchild)) {
526 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
527 GOTO(out_link_de_tgt_dir, rc = -ESTALE);
530 if (dchild->d_inode) {
531 CERROR("child exists (dir %ld, name %s\n",
532 de_tgt_dir->d_inode->i_ino, rec->ur_name);
533 GOTO(out_link_dchild, rc = -EEXIST);
536 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
537 de_src->d_inode->i_sb->s_dev);
539 handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
541 GOTO(out_link_dchild, rc = PTR_ERR(handle));
543 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
546 rc = mds_update_last_rcvd(mds, handle, req);
548 err = mds_fs_commit(mds, de_tgt_dir->d_inode, handle);
550 CERROR("error on commit: err = %d\n", err);
559 up(&de_tgt_dir->d_inode->i_sem);
568 static int mds_reint_rename(struct mds_update_record *rec, int offset,
569 struct ptlrpc_request *req)
571 struct obd_device *obd = req->rq_export->exp_obd;
572 struct dentry *de_srcdir = NULL;
573 struct dentry *de_tgtdir = NULL;
574 struct dentry *de_old = NULL;
575 struct dentry *de_new = NULL;
576 struct mds_obd *mds = mds_req2mds(req);
577 struct lustre_handle tgtlockh, srclockh, oldhandle;
578 int flags, lock_mode, rc = 0, err;
580 __u64 res_id[3] = { 0 };
583 de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
584 if (IS_ERR(de_srcdir))
585 GOTO(out_rename, rc = -ESTALE);
587 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
588 res_id[0] = de_srcdir->d_inode->i_ino;
590 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
591 NULL, 0, lock_mode, &srclockh);
593 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
594 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
595 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
596 &flags, ldlm_completion_ast,
597 mds_blocking_ast, NULL, 0, &srclockh);
598 if (rc != ELDLM_OK) {
599 CERROR("lock enqueue: err: %d\n", rc);
600 GOTO(out_rename_srcput, rc = -EIO);
603 ldlm_lock_dump((void *)(unsigned long)srclockh.addr);
605 de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
606 if (IS_ERR(de_tgtdir))
607 GOTO(out_rename_srcdir, rc = -ESTALE);
609 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
610 res_id[0] = de_tgtdir->d_inode->i_ino;
612 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
613 NULL, 0, lock_mode, &tgtlockh);
615 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
616 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
617 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
618 &flags, ldlm_completion_ast,
619 mds_blocking_ast, NULL, 0, &tgtlockh);
620 if (rc != ELDLM_OK) {
621 CERROR("lock enqueue: err: %d\n", rc);
622 GOTO(out_rename_tgtput, rc = -EIO);
625 ldlm_lock_dump((void *)(unsigned long)tgtlockh.addr);
627 double_lock(de_tgtdir, de_srcdir);
629 de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
630 if (IS_ERR(de_old)) {
631 CERROR("old child lookup error (%*s): %ld\n",
632 rec->ur_namelen - 1, rec->ur_name, PTR_ERR(de_old));
633 GOTO(out_rename_tgtdir, rc = -ENOENT);
636 de_new = lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
637 if (IS_ERR(de_new)) {
638 CERROR("new child lookup error (%*s): %ld\n",
639 rec->ur_tgtlen - 1, rec->ur_tgt, PTR_ERR(de_new));
640 GOTO(out_rename_deold, rc = -ENOENT);
643 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
644 de_srcdir->d_inode->i_sb->s_dev);
646 handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
648 GOTO(out_rename_denew, rc = PTR_ERR(handle));
650 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new,
655 rc = mds_update_last_rcvd(mds, handle, req);
657 err = mds_fs_commit(mds, de_tgtdir->d_inode, handle);
659 CERROR("error on commit: err = %d\n", err);
669 res_id[0] = de_old->d_inode->i_ino;
670 /* Take an exclusive lock on the resource that we're
671 * about to free, to force everyone to drop their
673 LDLM_DEBUG_NOLOCK("getting EX lock res %Lu", res_id[0]);
674 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
675 res_id, LDLM_PLAIN, NULL, 0, LCK_EX,
676 &flags, ldlm_completion_ast,
677 mds_blocking_ast, NULL, 0, &oldhandle);
679 CERROR("failed to get child inode lock (child ino %Ld, "
681 res_id[0], de_old->d_inode->i_ino);
687 ldlm_lock_decref(&oldhandle, LCK_EX);
688 rc = ldlm_cli_cancel(&oldhandle);
690 CERROR("failed to cancel child inode lock ino "
691 "%Ld: %d\n", res_id[0], rc);
694 double_up(&de_srcdir->d_inode->i_sem, &de_tgtdir->d_inode->i_sem);
695 ldlm_lock_decref(&tgtlockh, lock_mode);
699 ldlm_lock_decref(&srclockh, lock_mode);
707 typedef int (*mds_reinter) (struct mds_update_record *, int offset,
708 struct ptlrpc_request *);
710 static mds_reinter reinters[REINT_MAX + 1] = {
711 [REINT_SETATTR] mds_reint_setattr,
712 [REINT_CREATE] mds_reint_create,
713 [REINT_UNLINK] mds_reint_unlink,
714 [REINT_LINK] mds_reint_link,
715 [REINT_RENAME] mds_reint_rename,
716 [REINT_RECREATE] mds_reint_recreate,
719 int mds_reint_rec(struct mds_update_record *rec, int offset,
720 struct ptlrpc_request *req)
722 struct mds_obd *mds = mds_req2mds(req);
723 struct obd_run_ctxt saved;
727 if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
728 CERROR("opcode %d not valid\n", rec->ur_opcode);
729 rc = req->rq_status = -EINVAL;
733 push_ctxt(&saved, &mds->mds_ctxt);
734 rc = reinters[rec->ur_opcode] (rec, offset, req);