1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
5 * Lustre Metadata Server (mds) reintegration routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/obd_support.h>
35 #include <linux/obd_class.h>
36 #include <linux/obd.h>
37 #include <linux/lustre_lib.h>
38 #include <linux/lustre_idl.h>
39 #include <linux/lustre_mds.h>
40 #include <linux/lustre_dlm.h>
41 #include <linux/lustre_fsfilt.h>
43 #include "mds_internal.h"
45 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
48 obd_transno_commit_cb(obd, transno, error);
51 struct mds_logcancel_data {
52 struct lov_mds_md *mlcd_lmm;
56 struct llog_cookie mlcd_cookies[0];
60 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
61 void *cb_data, int error)
63 struct mds_logcancel_data *mlcd = cb_data;
64 struct lov_stripe_md *lsm = NULL;
65 struct llog_ctxt *ctxt;
68 obd_transno_commit_cb(obd, transno, error);
70 CDEBUG(D_HA, "cancelling %d cookies\n",
71 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
73 rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
74 mlcd->mlcd_eadatalen);
76 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
77 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
80 ///* XXX 0 normally, SENDNOW for debug */);
81 ctxt = llog_get_context(obd, mlcd->mlcd_cookies[0].lgc_subsys + 1);
82 rc = llog_cancel(ctxt, lsm,
83 mlcd->mlcd_cookielen /
84 sizeof(*mlcd->mlcd_cookies),
85 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
87 CERROR("error cancelling %d log cookies: rc %d\n",
88 (int)(mlcd->mlcd_cookielen /
89 sizeof(*mlcd->mlcd_cookies)), rc);
92 OBD_FREE(mlcd, mlcd->mlcd_size);
95 /* Assumes caller has already pushed us into the kernel context. */
96 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
97 struct ptlrpc_request *req, int rc, __u32 op_data)
99 struct mds_export_data *med = &req->rq_export->exp_mds_data;
100 struct mds_client_data *mcd = med->med_mcd;
101 struct obd_device *obd = req->rq_export->exp_obd;
108 /* if the export has already been failed, we have no last_rcvd slot */
109 if (req->rq_export->exp_failed) {
110 CERROR("committing transaction for disconnected client\n");
119 if (handle == NULL) {
120 /* if we're starting our own xaction, use our own inode */
121 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
122 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
123 if (IS_ERR(handle)) {
124 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
125 RETURN(PTR_ERR(handle));
131 transno = req->rq_reqmsg->transno;
133 LASSERT(transno == 0);
134 } else if (transno == 0) {
135 spin_lock(&mds->mds_transno_lock);
136 transno = ++mds->mds_last_transno;
137 spin_unlock(&mds->mds_transno_lock);
139 spin_lock(&mds->mds_transno_lock);
140 if (transno > mds->mds_last_transno)
141 mds->mds_last_transno = transno;
142 spin_unlock(&mds->mds_transno_lock);
144 req->rq_repmsg->transno = req->rq_transno = transno;
145 mcd->mcd_last_transno = cpu_to_le64(transno);
146 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
147 mcd->mcd_last_result = cpu_to_le32(rc);
148 mcd->mcd_last_data = cpu_to_le32(op_data);
150 fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
151 mds_commit_cb, NULL);
152 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
161 DEBUG_REQ(log_pri, req,
162 "wrote trans #"LPU64" client %s at idx %u: err = %d",
163 transno, mcd->mcd_uuid, med->med_idx, err);
165 err = mds_lov_write_objids(obd);
171 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
174 err = fsfilt_commit(obd, inode, handle, 0);
176 CERROR("error committing transaction: %d\n", err);
184 /* this gives the same functionality as the code between
185 * sys_chmod and inode_setattr
186 * chown_common and inode_setattr
187 * utimes and inode_setattr
189 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
191 time_t now = LTIME_S(CURRENT_TIME);
192 struct iattr *attr = &rec->ur_iattr;
193 unsigned int ia_valid = attr->ia_valid;
197 /* only fix up attrs if the client VFS didn't already */
198 if (!(ia_valid & ATTR_RAW))
201 if (!(ia_valid & ATTR_CTIME_SET))
202 LTIME_S(attr->ia_ctime) = now;
203 if (!(ia_valid & ATTR_ATIME_SET))
204 LTIME_S(attr->ia_atime) = now;
205 if (!(ia_valid & ATTR_MTIME_SET))
206 LTIME_S(attr->ia_mtime) = now;
208 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
212 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
213 if (rec->ur_fsuid != inode->i_uid &&
214 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
218 if (ia_valid & ATTR_SIZE) {
219 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
223 if (ia_valid & ATTR_UID) {
226 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
228 if (attr->ia_uid == (uid_t) -1)
229 attr->ia_uid = inode->i_uid;
230 if (attr->ia_gid == (gid_t) -1)
231 attr->ia_gid = inode->i_gid;
232 attr->ia_mode = inode->i_mode;
234 * If the user or group of a non-directory has been
235 * changed by a non-root user, remove the setuid bit.
236 * 19981026 David C Niemi <niemi@tux.org>
238 * Changed this to apply to all users, including root,
239 * to avoid some races. This is the behavior we had in
240 * 2.0. The check for non-root was definitely wrong
241 * for 2.2 anyway, as it should have been using
242 * CAP_FSETID rather than fsuid -- 19990830 SD.
244 if ((inode->i_mode & S_ISUID) == S_ISUID &&
245 !S_ISDIR(inode->i_mode)) {
246 attr->ia_mode &= ~S_ISUID;
247 attr->ia_valid |= ATTR_MODE;
250 * Likewise, if the user or group of a non-directory
251 * has been changed by a non-root user, remove the
252 * setgid bit UNLESS there is no group execute bit
253 * (this would be a file marked for mandatory
254 * locking). 19981026 David C Niemi <niemi@tux.org>
256 * Removed the fsuid check (see the comment above) --
259 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
260 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
261 attr->ia_mode &= ~S_ISGID;
262 attr->ia_valid |= ATTR_MODE;
264 } else if (ia_valid & ATTR_MODE) {
265 int mode = attr->ia_mode;
267 if (attr->ia_mode == (mode_t) -1)
268 attr->ia_mode = inode->i_mode;
270 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
275 void mds_steal_ack_locks(struct ptlrpc_request *req)
277 struct obd_export *exp = req->rq_export;
278 struct list_head *tmp;
279 struct ptlrpc_reply_state *oldrep;
280 struct ptlrpc_service *svc;
284 /* CAVEAT EMPTOR: spinlock order */
285 spin_lock_irqsave (&exp->exp_lock, flags);
286 list_for_each (tmp, &exp->exp_outstanding_replies) {
287 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
289 if (oldrep->rs_xid != req->rq_xid)
292 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
293 CERROR ("Resent req xid "LPX64" has mismatched opc: "
294 "new %d old %d\n", req->rq_xid,
295 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
297 svc = oldrep->rs_srv_ni->sni_service;
298 spin_lock (&svc->srv_lock);
300 list_del_init (&oldrep->rs_exp_list);
302 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
304 oldrep->rs_nlocks, oldrep,
305 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
306 exp->exp_connection->c_peer.peer_nid);
308 for (i = 0; i < oldrep->rs_nlocks; i++)
309 ptlrpc_save_lock(req,
310 &oldrep->rs_locks[i],
311 oldrep->rs_modes[i]);
312 oldrep->rs_nlocks = 0;
314 DEBUG_REQ(D_HA, req, "stole locks for");
315 ptlrpc_schedule_difficult_reply (oldrep);
317 spin_unlock (&svc->srv_lock);
318 spin_unlock_irqrestore (&exp->exp_lock, flags);
321 spin_unlock_irqrestore (&exp->exp_lock, flags);
324 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
326 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
327 mcd->mcd_last_transno, mcd->mcd_last_result);
328 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
329 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
331 mds_steal_ack_locks(req);
334 static void reconstruct_reint_setattr(struct mds_update_record *rec,
335 int offset, struct ptlrpc_request *req)
337 struct mds_export_data *med = &req->rq_export->exp_mds_data;
338 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
340 struct mds_body *body;
342 mds_req_from_mcd(req, med->med_mcd);
344 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
346 LASSERT(PTR_ERR(de) == req->rq_status);
350 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
351 mds_pack_inode2fid(&body->fid1, de->d_inode);
352 mds_pack_inode2body(body, de->d_inode);
354 /* Don't return OST-specific attributes if we didn't just set them */
355 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
356 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
357 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
358 body->valid |= OBD_MD_FLMTIME;
359 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
360 body->valid |= OBD_MD_FLATIME;
365 /* In the raw-setattr case, we lock the child inode.
366 * In the write-back case or if being called from open, the client holds a lock
369 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
370 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
371 struct ptlrpc_request *req,
372 struct lustre_handle *lh)
374 struct mds_obd *mds = mds_req2mds(req);
375 struct obd_device *obd = req->rq_export->exp_obd;
376 struct mds_body *body;
378 struct inode *inode = NULL;
379 struct lustre_handle lockh;
381 struct mds_logcancel_data *mlcd = NULL;
382 int rc = 0, cleanup_phase = 0, err, locked = 0;
385 LASSERT(offset == 0);
387 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
388 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
390 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
392 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
393 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
395 GOTO(cleanup, rc = PTR_ERR(de));
397 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
400 GOTO(cleanup, rc = PTR_ERR(de));
407 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
410 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
412 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
414 GOTO(cleanup, rc = PTR_ERR(handle));
416 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
417 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
418 LTIME_S(rec->ur_iattr.ia_mtime),
419 LTIME_S(rec->ur_iattr.ia_ctime));
420 rc = mds_fix_attr(inode, rec);
424 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
425 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
426 (long)&rec->ur_iattr.ia_attr_flags);
428 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
430 if (rc == 0 && S_ISREG(inode->i_mode) && rec->ur_eadata != NULL) {
431 rc = fsfilt_set_md(obd, inode, handle,
432 rec->ur_eadata, rec->ur_eadatalen);
435 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
436 mds_pack_inode2fid(&body->fid1, inode);
437 mds_pack_inode2body(body, inode);
439 /* Don't return OST-specific attributes if we didn't just set them */
440 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
441 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
442 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
443 body->valid |= OBD_MD_FLMTIME;
444 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
445 body->valid |= OBD_MD_FLATIME;
447 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
448 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
451 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
453 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
454 mlcd->mlcd_cookielen = rec->ur_cookielen;
455 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
456 mlcd->mlcd_cookielen;
457 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
458 mlcd->mlcd_cookielen);
459 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
460 mlcd->mlcd_eadatalen);
462 CERROR("unable to allocate log cancel data\n");
468 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
469 mds_cancel_cookies_cb, mlcd);
470 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
471 switch (cleanup_phase) {
473 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
478 ldlm_lock_decref(&lockh, LCK_PW);
480 ptlrpc_save_lock (req, &lockh, LCK_PW);
495 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
496 struct ptlrpc_request *req)
498 struct mds_export_data *med = &req->rq_export->exp_mds_data;
499 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
500 struct dentry *parent, *child;
501 struct mds_body *body;
503 mds_req_from_mcd(req, med->med_mcd);
508 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
509 LASSERT(!IS_ERR(parent));
510 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
511 LASSERT(!IS_ERR(child));
512 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
513 mds_pack_inode2fid(&body->fid1, child->d_inode);
514 mds_pack_inode2body(body, child->d_inode);
519 static int mds_reint_create(struct mds_update_record *rec, int offset,
520 struct ptlrpc_request *req,
521 struct lustre_handle *lh)
523 struct dentry *dparent = NULL;
524 struct mds_obd *mds = mds_req2mds(req);
525 struct obd_device *obd = req->rq_export->exp_obd;
526 struct dentry *dchild = NULL;
527 struct inode *dir = NULL;
529 struct lustre_handle lockh;
530 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
532 struct dentry_params dp;
535 LASSERT(offset == 0);
536 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
538 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
539 rec->ur_fid1->id, rec->ur_fid1->generation,
540 rec->ur_name, rec->ur_mode);
542 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
544 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
545 GOTO(cleanup, rc = -ESTALE);
547 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
548 rec->ur_name, rec->ur_namelen - 1);
549 if (IS_ERR(dparent)) {
550 rc = PTR_ERR(dparent);
551 CERROR("parent lookup error %d\n", rc);
554 cleanup_phase = 1; /* locked parent dentry */
555 dir = dparent->d_inode;
558 ldlm_lock_dump_handle(D_OTHER, &lockh);
560 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
561 if (IS_ERR(dchild)) {
562 rc = PTR_ERR(dchild);
563 CERROR("child lookup error %d\n", rc);
567 cleanup_phase = 2; /* child dentry */
569 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
571 if (dir->i_mode & S_ISGID) {
572 if (S_ISDIR(rec->ur_mode))
573 rec->ur_mode |= S_ISGID;
576 dchild->d_fsdata = (void *)&dp;
577 dp.p_inum = (unsigned long)rec->ur_fid2->id;
582 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
584 GOTO(cleanup, rc = PTR_ERR(handle));
585 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
590 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
592 GOTO(cleanup, rc = PTR_ERR(handle));
593 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
598 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
600 GOTO(cleanup, rc = PTR_ERR(handle));
601 if (rec->ur_tgt == NULL) /* no target supplied */
602 rc = -EINVAL; /* -EPROTO? */
604 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
612 int rdev = rec->ur_rdev;
613 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
615 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
616 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
621 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
622 dchild->d_fsdata = NULL;
623 GOTO(cleanup, rc = -EINVAL);
626 /* In case we stored the desired inum in here, we want to clean up. */
627 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
628 dchild->d_fsdata = NULL;
631 CDEBUG(D_INODE, "error during create: %d\n", rc);
635 struct inode *inode = dchild->d_inode;
636 struct mds_body *body;
639 LTIME_S(iattr.ia_atime) = rec->ur_time;
640 LTIME_S(iattr.ia_ctime) = rec->ur_time;
641 LTIME_S(iattr.ia_mtime) = rec->ur_time;
642 iattr.ia_uid = rec->ur_fsuid;
643 if (dir->i_mode & S_ISGID)
644 iattr.ia_gid = dir->i_gid;
646 iattr.ia_gid = rec->ur_fsgid;
647 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
648 ATTR_MTIME | ATTR_CTIME;
650 if (rec->ur_fid2->id) {
651 LASSERT(rec->ur_fid2->id == inode->i_ino);
652 inode->i_generation = rec->ur_fid2->generation;
653 /* Dirtied and committed by the upcoming setattr. */
654 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
655 inode->i_ino, inode->i_generation);
657 struct lustre_handle child_ino_lockh;
658 struct ldlm_res_id child_res_id =
659 { .name = { inode->i_ino, 0 } };
662 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
663 inode->i_ino, inode->i_generation);
665 /* The inode we were allocated may have just been freed
666 * by an unlink operation. We take this lock to
667 * synchronize against the matching reply-ack-lock taken
668 * in unlink, to avoid replay problems if this reply
669 * makes it out to the client but the unlink's does not.
670 * See bug 2029 for more detail.*/
671 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
672 child_res_id, LDLM_PLAIN, NULL,
675 ldlm_completion_ast, NULL, NULL,
676 NULL, 0, NULL, &child_ino_lockh);
677 if (rc != ELDLM_OK) {
678 CERROR("error locking for unlink/create sync: "
681 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
685 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
687 CERROR("error on child setattr: rc = %d\n", rc);
689 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
690 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
692 CERROR("error on parent setattr: rc = %d\n", rc);
694 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
695 mds_pack_inode2fid(&body->fid1, inode);
696 mds_pack_inode2body(body, inode);
701 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
704 /* Destroy the file we just created. This should not need
705 * extra journal credits, as we have already modified all of
706 * the blocks needed in order to create the file in the first
711 err = vfs_rmdir(dir, dchild);
713 CERROR("rmdir in error path: %d\n", err);
716 err = vfs_unlink(dir, dchild);
718 CERROR("unlink in error path: %d\n", err);
724 switch (cleanup_phase) {
725 case 2: /* child dentry */
727 case 1: /* locked parent dentry */
729 ldlm_lock_decref(&lockh, LCK_PW);
731 ptlrpc_save_lock (req, &lockh, LCK_PW);
737 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
744 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
748 for (i = 0; i < RES_NAME_SIZE; i++) {
749 /* return 1 here, because enqueue_ordered will skip resources
750 * of all zeroes if they're sorted to the end of the list. */
751 if (res1->name[i] == 0 && res2->name[i] != 0)
753 if (res2->name[i] == 0 && res1->name[i] != 0)
756 if (res1->name[i] > res2->name[i])
758 if (res1->name[i] < res2->name[i])
764 /* This function doesn't use ldlm_match_or_enqueue because we're always called
765 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
766 * because they take the place of local semaphores.
768 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
769 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
770 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
771 struct lustre_handle *p1_lockh, int p1_lock_mode,
772 struct ldlm_res_id *p2_res_id,
773 struct lustre_handle *p2_lockh, int p2_lock_mode)
775 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
776 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
777 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
781 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
783 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
784 res_id[0]->name[0], res_id[1]->name[0]);
786 if (res_gt(p1_res_id, p2_res_id)) {
787 handles[1] = p1_lockh;
788 handles[0] = p2_lockh;
789 res_id[1] = p1_res_id;
790 res_id[0] = p2_res_id;
791 lock_modes[1] = p1_lock_mode;
792 lock_modes[0] = p2_lock_mode;
795 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
796 res_id[0]->name[0], res_id[1]->name[0]);
798 flags = LDLM_FL_LOCAL_ONLY;
799 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
800 LDLM_PLAIN, NULL, lock_modes[0], &flags,
801 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
802 NULL, 0, NULL, handles[0]);
805 ldlm_lock_dump_handle(D_OTHER, handles[0]);
807 if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0) {
808 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
809 ldlm_lock_addref(handles[1], lock_modes[1]);
810 } else if (res_id[1]->name[0] != 0) {
811 flags = LDLM_FL_LOCAL_ONLY;
812 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
813 *res_id[1], LDLM_PLAIN, NULL,
814 lock_modes[1], &flags, mds_blocking_ast,
815 ldlm_completion_ast, NULL, NULL, NULL, 0,
817 if (rc != ELDLM_OK) {
818 ldlm_lock_decref(handles[0], lock_modes[0]);
821 ldlm_lock_dump_handle(D_OTHER, handles[1]);
827 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
828 struct lustre_handle *p1_lockh, int p1_lock_mode,
829 struct ldlm_res_id *p2_res_id,
830 struct lustre_handle *p2_lockh, int p2_lock_mode,
831 struct ldlm_res_id *c1_res_id,
832 struct lustre_handle *c1_lockh, int c1_lock_mode,
833 struct ldlm_res_id *c2_res_id,
834 struct lustre_handle *c2_lockh, int c2_lock_mode)
836 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
837 c1_res_id, c2_res_id };
838 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
839 c1_lockh, c2_lockh };
840 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
841 c1_lock_mode, c2_lock_mode };
842 int rc, i, j, sorted, flags;
845 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
846 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
849 /* simple insertion sort - we have at most 4 elements */
850 for (i = 1; i < 4; i++) {
852 dlm_handles[4] = dlm_handles[i];
853 res_id[4] = res_id[i];
854 lock_modes[4] = lock_modes[i];
858 if (res_gt(res_id[j], res_id[4])) {
859 dlm_handles[j + 1] = dlm_handles[j];
860 res_id[j + 1] = res_id[j];
861 lock_modes[j + 1] = lock_modes[j];
866 } while (j >= 0 && !sorted);
868 dlm_handles[j + 1] = dlm_handles[4];
869 res_id[j + 1] = res_id[4];
870 lock_modes[j + 1] = lock_modes[4];
873 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
874 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
877 /* XXX we could send ASTs on all these locks first before blocking? */
878 for (i = 0; i < 4; i++) {
880 if (res_id[i]->name[0] == 0)
883 memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) == 0) {
884 memcpy(dlm_handles[i], dlm_handles[i-1],
885 sizeof(*(dlm_handles[i])));
886 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
888 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
889 *res_id[i], LDLM_PLAIN, NULL,
890 lock_modes[i], &flags,
892 ldlm_completion_ast, NULL, NULL,
893 NULL, 0, NULL, dlm_handles[i]);
895 GOTO(out_err, rc = -EIO);
896 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
903 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
908 /* In the unlikely case that the child changed while we were waiting
909 * on the lock, we need to drop the lock on the old child and either:
910 * - if the child has a lower resource name, then we have to also
911 * drop the parent lock and regain the locks in the right order
912 * - in the rename case, if the child has a lower resource name than one of
913 * the other parent/child resources (maxres) we also need to reget the locks
914 * - if the child has a higher resource name (this is the common case)
915 * we can just get the lock on the new child (still in lock order)
917 * Returns 0 if the child did not change or if it changed but could be locked.
918 * Returns 1 if the child changed and we need to re-lock (no locks held).
919 * Returns -ve error with a valid dchild (no locks held). */
920 static int mds_verify_child(struct obd_device *obd,
921 struct ldlm_res_id *parent_res_id,
922 struct lustre_handle *parent_lockh,
923 struct dentry *dparent, int parent_mode,
924 struct ldlm_res_id *child_res_id,
925 struct lustre_handle *child_lockh,
926 struct dentry **dchildp, int child_mode,
927 const char *name, int namelen,
928 struct ldlm_res_id *maxres)
930 struct dentry *vchild, *dchild = *dchildp;
931 int rc = 0, cleanup_phase = 2; /* parent, child locks */
934 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
936 GOTO(cleanup, rc = PTR_ERR(vchild));
938 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
939 (vchild->d_inode != NULL &&
940 child_res_id->name[0] == vchild->d_inode->i_ino &&
941 child_res_id->name[1] == vchild->d_inode->i_generation))) {
949 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
950 vchild->d_inode, dchild ? dchild->d_inode : 0,
951 vchild->d_inode ? vchild->d_inode->i_ino : 0,
952 child_res_id->name[0]);
953 if (child_res_id->name[0] != 0)
954 ldlm_lock_decref(child_lockh, child_mode);
958 cleanup_phase = 1; /* parent lock only */
959 *dchildp = dchild = vchild;
961 if (dchild->d_inode) {
963 child_res_id->name[0] = dchild->d_inode->i_ino;
964 child_res_id->name[1] = dchild->d_inode->i_generation;
966 if (res_gt(parent_res_id, child_res_id) ||
967 res_gt(maxres, child_res_id)) {
968 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
969 child_res_id->name[0], parent_res_id->name[0],
971 GOTO(cleanup, rc = 1);
974 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
975 *child_res_id, LDLM_PLAIN, NULL,
976 child_mode, &flags, mds_blocking_ast,
977 ldlm_completion_ast, NULL, NULL, NULL, 0,
980 GOTO(cleanup, rc = -EIO);
982 memset(child_res_id, 0, sizeof(*child_res_id));
988 switch(cleanup_phase) {
990 if (child_res_id->name[0] != 0)
991 ldlm_lock_decref(child_lockh, child_mode);
993 ldlm_lock_decref(parent_lockh, parent_mode);
999 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1001 struct lustre_handle *parent_lockh,
1002 struct dentry **dparentp, int parent_mode,
1003 char *name, int namelen,
1004 struct lustre_handle *child_lockh,
1005 struct dentry **dchildp, int child_mode)
1007 struct ldlm_res_id child_res_id = { .name = {0} };
1008 struct ldlm_res_id parent_res_id = { .name = {0} };
1009 struct inode *inode;
1010 int rc = 0, cleanup_phase = 0;
1013 /* Step 1: Lookup parent */
1014 *dparentp = mds_fid2dentry(mds, fid, NULL);
1015 if (IS_ERR(*dparentp))
1016 RETURN(rc = PTR_ERR(*dparentp));
1017 LASSERT((*dparentp)->d_inode);
1019 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1020 (*dparentp)->d_inode->i_ino, name);
1022 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1023 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1025 cleanup_phase = 1; /* parent dentry */
1027 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1028 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1029 if (IS_ERR(*dchildp)) {
1030 rc = PTR_ERR(*dchildp);
1031 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1035 inode = (*dchildp)->d_inode;
1037 inode = igrab(inode);
1041 child_res_id.name[0] = inode->i_ino;
1042 child_res_id.name[1] = inode->i_generation;
1046 cleanup_phase = 2; /* child dentry */
1048 /* Step 3: Lock parent and child in resource order. If child doesn't
1049 * exist, we still have to lock the parent and re-lookup. */
1050 rc = enqueue_ordered_locks(obd, &parent_res_id,parent_lockh,parent_mode,
1051 &child_res_id, child_lockh, child_mode);
1055 if (!(*dchildp)->d_inode)
1056 cleanup_phase = 3; /* parent lock */
1058 cleanup_phase = 4; /* child lock */
1060 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1061 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1062 parent_mode, &child_res_id, child_lockh, dchildp,
1063 child_mode, name, namelen, &parent_res_id);
1073 switch (cleanup_phase) {
1075 ldlm_lock_decref(child_lockh, child_mode);
1077 ldlm_lock_decref(parent_lockh, parent_mode);
1088 void mds_reconstruct_generic(struct ptlrpc_request *req)
1090 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1092 mds_req_from_mcd(req, med->med_mcd);
1095 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1096 struct ptlrpc_request *req,
1097 struct lustre_handle *lh)
1099 struct dentry *dparent, *dchild;
1100 struct mds_obd *mds = mds_req2mds(req);
1101 struct obd_device *obd = req->rq_export->exp_obd;
1102 struct mds_body *body = NULL;
1103 struct inode *child_inode;
1104 struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1105 void *handle = NULL;
1106 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1109 LASSERT(offset == 0 || offset == 2);
1111 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1112 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1114 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1116 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1117 GOTO(cleanup, rc = -ENOENT);
1119 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1120 &parent_lockh, &dparent, LCK_PW,
1121 rec->ur_name, rec->ur_namelen,
1122 &child_lockh, &dchild, LCK_EX);
1126 cleanup_phase = 1; /* dchild, dparent, locks */
1128 child_inode = dchild->d_inode;
1129 if (child_inode == NULL) {
1130 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1131 dparent->d_inode->i_ino, rec->ur_name);
1132 GOTO(cleanup, rc = -ENOENT);
1135 cleanup_phase = 2; /* dchild has a lock */
1137 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1138 * reuse (see bug 2029). */
1139 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1143 cleanup_phase = 3; /* child inum lock */
1145 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1147 /* ldlm_reply in buf[0] if called via intent */
1151 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1152 LASSERT(body != NULL);
1154 /* If this is the last reference to this inode, get the OBD EA
1155 * data first so the client can destroy OST objects.
1156 * we only do the object removal if no open files remain.
1157 * Nobody can get at this name anymore because of the locks so
1158 * we make decisions here as to whether to remove the inode */
1159 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1160 mds_open_orphan_count(child_inode) == 0) {
1161 mds_pack_inode2fid(&body->fid1, child_inode);
1162 mds_pack_inode2body(body, child_inode);
1163 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1165 if (!(body->valid & OBD_MD_FLEASIZE)) {
1166 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1167 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1173 /* We have to do these checks ourselves, in case we are making an
1174 * orphan. The client tells us whether rmdir() or unlink() was called,
1175 * so we need to return appropriate errors (bug 72).
1177 * We don't have to check permissions, because vfs_rename (called from
1178 * mds_open_unlink_rename) also calls may_delete. */
1179 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1180 if (!S_ISDIR(child_inode->i_mode))
1181 GOTO(cleanup, rc = -ENOTDIR);
1183 if (S_ISDIR(child_inode->i_mode))
1184 GOTO(cleanup, rc = -EISDIR);
1187 if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1188 mds_open_orphan_count(child_inode) > 0) {
1189 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1190 cleanup_phase = 4; /* transaction */
1194 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1195 switch (child_inode->i_mode & S_IFMT) {
1197 /* Drop any lingering child directories before we start our
1198 * transaction, to avoid doing multiple inode dirty/delete
1199 * in our compound transaction (bug 1321). */
1200 shrink_dcache_parent(dchild);
1201 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1204 GOTO(cleanup, rc = PTR_ERR(handle));
1205 cleanup_phase = 4; /* transaction */
1206 rc = vfs_rmdir(dparent->d_inode, dchild);
1209 handle = fsfilt_start(obd, dparent->d_inode,
1210 FSFILT_OP_UNLINK_LOG, NULL);
1212 GOTO(cleanup, rc = PTR_ERR(handle));
1214 cleanup_phase = 4; /* transaction */
1215 rc = vfs_unlink(dparent->d_inode, dchild);
1217 if (!rc && log_unlink)
1218 if (mds_log_op_unlink(obd, child_inode,
1219 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1220 req->rq_repmsg->buflens[offset + 1],
1221 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1222 req->rq_repmsg->buflens[offset + 2]) > 0)
1223 body->valid |= OBD_MD_FLCOOKIE;
1231 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1234 GOTO(cleanup, rc = PTR_ERR(handle));
1235 cleanup_phase = 4; /* transaction */
1236 rc = vfs_unlink(dparent->d_inode, dchild);
1239 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1242 GOTO(cleanup, rc = -EINVAL);
1250 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1251 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1252 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1254 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1256 CERROR("error on parent setattr: rc = %d\n", err);
1259 switch(cleanup_phase) {
1261 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1264 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1265 "unlinked", 0, NULL);
1266 case 3: /* child ino-reuse lock */
1267 if (rc && body != NULL) {
1268 // Don't unlink the OST objects if the MDS unlink failed
1272 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1274 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1275 case 2: /* child lock */
1276 ldlm_lock_decref(&child_lockh, LCK_EX);
1277 case 1: /* child and parent dentry, parent lock */
1279 ldlm_lock_decref(&parent_lockh, LCK_PW);
1281 ptlrpc_save_lock(req, &parent_lockh, LCK_PW);
1287 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1290 req->rq_status = rc;
1294 static int mds_reint_link(struct mds_update_record *rec, int offset,
1295 struct ptlrpc_request *req,
1296 struct lustre_handle *lh)
1298 struct obd_device *obd = req->rq_export->exp_obd;
1299 struct dentry *de_src = NULL;
1300 struct dentry *de_tgt_dir = NULL;
1301 struct dentry *dchild = NULL;
1302 struct mds_obd *mds = mds_req2mds(req);
1303 struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1304 struct ldlm_res_id src_res_id = { .name = {0} };
1305 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1306 int rc = 0, cleanup_phase = 0;
1309 LASSERT(offset == 0);
1311 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1312 rec->ur_fid1->id, rec->ur_fid1->generation,
1313 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1315 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1317 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1318 GOTO(cleanup, rc = -ENOENT);
1320 /* Step 1: Lookup the source inode and target directory by FID */
1321 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1323 GOTO(cleanup, rc = PTR_ERR(de_src));
1325 cleanup_phase = 1; /* source dentry */
1327 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1328 if (IS_ERR(de_tgt_dir))
1329 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1331 cleanup_phase = 2; /* target directory dentry */
1333 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1334 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1335 de_src->d_inode->i_ino);
1337 /* Step 2: Take the two locks */
1338 src_res_id.name[0] = de_src->d_inode->i_ino;
1339 src_res_id.name[1] = de_src->d_inode->i_generation;
1340 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1341 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1343 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1344 &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX);
1348 cleanup_phase = 3; /* locks */
1350 /* Step 3: Lookup the child */
1351 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1352 if (IS_ERR(dchild)) {
1353 rc = PTR_ERR(dchild);
1354 if (rc != -EPERM && rc != -EACCES)
1355 CERROR("child lookup error %d\n", rc);
1359 cleanup_phase = 4; /* child dentry */
1361 if (dchild->d_inode) {
1362 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1363 de_tgt_dir->d_inode->i_ino, rec->ur_name);
1368 /* Step 4: Do it. */
1369 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1371 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1372 if (IS_ERR(handle)) {
1373 rc = PTR_ERR(handle);
1377 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1378 if (rc && rc != -EPERM && rc != -EACCES)
1379 CERROR("vfs_link error %d\n", rc);
1381 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1382 handle, req, rc, 0);
1385 switch (cleanup_phase) {
1386 case 4: /* child dentry */
1390 ldlm_lock_decref(&src_lockh, LCK_EX);
1391 ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1393 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1394 ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1396 case 2: /* target dentry */
1398 case 1: /* source dentry */
1403 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1406 req->rq_status = rc;
1411 * add a hard link in the PENDING directory, only used by rename()
1413 static int mds_add_link_orphan(struct mds_update_record *rec,
1414 struct obd_device *obd,
1415 struct dentry *dentry)
1417 struct mds_obd *mds = &obd->u.mds;
1418 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1419 struct dentry *pending_child;
1420 char fidname[LL_FID_NAMELEN];
1424 LASSERT(dentry->d_inode);
1425 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
1427 down(&pending_dir->i_sem);
1428 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
1429 dentry->d_inode->i_generation);
1431 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
1432 mds_open_orphan_count(dentry->d_inode),
1433 rec->ur_name, fidname);
1435 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1436 if (IS_ERR(pending_child))
1437 GOTO(out_lock, rc = PTR_ERR(pending_child));
1439 if (pending_child->d_inode != NULL) {
1440 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1441 LASSERT(pending_child->d_inode == dentry->d_inode);
1442 GOTO(out_dput, rc = 0);
1446 rc = vfs_link(dentry, pending_dir, pending_child);
1449 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
1452 mds_inode_set_orphan(dentry->d_inode);
1454 l_dput(pending_child);
1456 up(&pending_dir->i_sem);
1460 /* The idea here is that we need to get four locks in the end:
1461 * one on each parent directory, one on each child. We need to take
1462 * these locks in some kind of order (to avoid deadlocks), and the order
1463 * I selected is "increasing resource number" order. We need to look up
1464 * the children, however, before we know what the resource number(s) are.
1465 * Thus the following plan:
1467 * 1,2. Look up the parents
1468 * 3,4. Look up the children
1469 * 5. Take locks on the parents and children, in order
1470 * 6. Verify that the children haven't changed since they were looked up
1472 * If there was a race and the children changed since they were first looked
1473 * up, it is possible that mds_verify_child() will be able to just grab the
1474 * lock on the new child resource (if it has a higher resource than any other)
1475 * but we need to compare against not only its parent, but also against the
1476 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1478 * We need the fancy igrab() on the child inodes because we aren't holding a
1479 * lock on the parent after the lookup is done, so dentry->d_inode may change
1480 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1482 static int mds_get_parents_children_locked(struct obd_device *obd,
1483 struct mds_obd *mds,
1484 struct ll_fid *p1_fid,
1485 struct dentry **de_srcdirp,
1486 struct ll_fid *p2_fid,
1487 struct dentry **de_tgtdirp,
1489 const char *old_name, int old_len,
1490 struct dentry **de_oldp,
1491 const char *new_name, int new_len,
1492 struct dentry **de_newp,
1493 struct lustre_handle *dlm_handles,
1496 struct ldlm_res_id p1_res_id = { .name = {0} };
1497 struct ldlm_res_id p2_res_id = { .name = {0} };
1498 struct ldlm_res_id c1_res_id = { .name = {0} };
1499 struct ldlm_res_id c2_res_id = { .name = {0} };
1500 struct ldlm_res_id *maxres_src, *maxres_tgt;
1501 struct inode *inode;
1502 int rc = 0, cleanup_phase = 0;
1505 /* Step 1: Lookup the source directory */
1506 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1507 if (IS_ERR(*de_srcdirp))
1508 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1510 cleanup_phase = 1; /* source directory dentry */
1512 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1513 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1515 /* Step 2: Lookup the target directory */
1516 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1517 *de_tgtdirp = dget(*de_srcdirp);
1519 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1520 if (IS_ERR(*de_tgtdirp))
1521 GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
1524 cleanup_phase = 2; /* target directory dentry */
1526 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1527 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1529 /* Step 3: Lookup the source child entry */
1530 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1531 if (IS_ERR(*de_oldp)) {
1532 rc = PTR_ERR(*de_oldp);
1533 CERROR("old child lookup error (%*s): %d\n",
1534 old_len - 1, old_name, rc);
1538 cleanup_phase = 3; /* original name dentry */
1540 inode = (*de_oldp)->d_inode;
1542 inode = igrab(inode);
1544 GOTO(cleanup, rc = -ENOENT);
1546 c1_res_id.name[0] = inode->i_ino;
1547 c1_res_id.name[1] = inode->i_generation;
1550 /* Step 4: Lookup the target child entry */
1551 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1552 if (IS_ERR(*de_newp)) {
1553 rc = PTR_ERR(*de_newp);
1554 CERROR("new child lookup error (%*s): %d\n",
1555 old_len - 1, old_name, rc);
1559 cleanup_phase = 4; /* target dentry */
1561 inode = (*de_newp)->d_inode;
1563 inode = igrab(inode);
1567 c2_res_id.name[0] = inode->i_ino;
1568 c2_res_id.name[1] = inode->i_generation;
1572 /* Step 5: Take locks on the parents and child(ren) */
1573 maxres_src = &p1_res_id;
1574 maxres_tgt = &p2_res_id;
1575 cleanup_phase = 4; /* target dentry */
1577 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id))
1578 maxres_src = &c1_res_id;
1579 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id))
1580 maxres_tgt = &c2_res_id;
1582 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1583 &p2_res_id, &dlm_handles[1], parent_mode,
1584 &c1_res_id, &dlm_handles[2], child_mode,
1585 &c2_res_id, &dlm_handles[3], child_mode);
1589 cleanup_phase = 6; /* parent and child(ren) locks */
1591 /* Step 6a: Re-lookup source child to verify it hasn't changed */
1592 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1593 parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
1594 child_mode, old_name, old_len, maxres_tgt);
1596 if (c2_res_id.name[0] != 0)
1597 ldlm_lock_decref(&dlm_handles[3], child_mode);
1598 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1605 if ((*de_oldp)->d_inode == NULL)
1606 GOTO(cleanup, rc = -ENOENT);
1608 /* Step 6b: Re-lookup target child to verify it hasn't changed */
1609 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
1610 parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
1611 child_mode, new_name, new_len, maxres_src);
1613 ldlm_lock_decref(&dlm_handles[2], child_mode);
1614 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1624 switch (cleanup_phase) {
1625 case 6: /* child lock(s) */
1626 if (c2_res_id.name[0] != 0)
1627 ldlm_lock_decref(&dlm_handles[3], child_mode);
1628 if (c1_res_id.name[0] != 0)
1629 ldlm_lock_decref(&dlm_handles[2], child_mode);
1630 case 5: /* parent locks */
1631 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1632 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1633 case 4: /* target dentry */
1635 case 3: /* source dentry */
1637 case 2: /* target directory dentry */
1638 l_dput(*de_tgtdirp);
1639 case 1: /* source directry dentry */
1640 l_dput(*de_srcdirp);
1647 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1648 struct ptlrpc_request *req,
1649 struct lustre_handle *lockh)
1651 struct obd_device *obd = req->rq_export->exp_obd;
1652 struct dentry *de_srcdir = NULL;
1653 struct dentry *de_tgtdir = NULL;
1654 struct dentry *de_old = NULL;
1655 struct dentry *de_new = NULL;
1656 struct mds_obd *mds = mds_req2mds(req);
1657 struct lustre_handle dlm_handles[4];
1658 struct mds_body *body = NULL;
1659 int rc = 0, lock_count = 3;
1660 int cleanup_phase = 0;
1661 void *handle = NULL;
1664 LASSERT(offset == 0);
1666 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
1667 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
1668 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
1670 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1672 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
1673 rec->ur_fid2, &de_tgtdir, LCK_PW,
1674 rec->ur_name, rec->ur_namelen,
1675 &de_old, rec->ur_tgt,
1676 rec->ur_tgtlen, &de_new,
1677 dlm_handles, LCK_EX);
1681 cleanup_phase = 1; /* parent(s), children, locks */
1683 if (de_new->d_inode)
1686 /* sanity check for src inode */
1687 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1688 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
1689 GOTO(cleanup, rc = -EINVAL);
1691 /* sanity check for dest inode */
1692 if (de_new->d_inode &&
1693 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1694 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
1695 GOTO(cleanup, rc = -EINVAL);
1697 if (de_old->d_inode == de_new->d_inode) {
1698 GOTO(cleanup, rc = 0);
1701 /* if we are about to remove the target at first, pass the EA of
1702 * that inode to client to perform and cleanup on OST */
1703 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1704 LASSERT(body != NULL);
1706 if (de_new->d_inode &&
1707 S_ISREG(de_new->d_inode->i_mode) &&
1708 de_new->d_inode->i_nlink == 1 &&
1709 mds_open_orphan_count(de_new->d_inode) == 0) {
1710 mds_pack_inode2fid(&body->fid1, de_new->d_inode);
1711 mds_pack_inode2body(body, de_new->d_inode);
1712 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
1713 if (!(body->valid & OBD_MD_FLEASIZE)) {
1714 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1715 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1717 /* XXX need log unlink? */
1721 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
1722 de_srcdir->d_inode->i_sb);
1724 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
1726 GOTO(cleanup, rc = PTR_ERR(handle));
1728 /* FIXME need adjust the journal block count? */
1729 /* if the target should be moved to PENDING, we at first increase the
1730 * link and later vfs_rename() will decrease the link count again */
1731 if (de_new->d_inode &&
1732 S_ISREG(de_new->d_inode->i_mode) &&
1733 de_new->d_inode->i_nlink == 1 &&
1734 mds_open_orphan_count(de_new->d_inode) > 0) {
1735 rc = mds_add_link_orphan(rec, obd, de_new);
1741 de_old->d_fsdata = req;
1742 de_new->d_fsdata = req;
1743 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
1748 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
1749 handle, req, rc, 0);
1750 switch (cleanup_phase) {
1753 if (lock_count == 4)
1754 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
1755 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
1756 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
1757 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
1759 if (lock_count == 4)
1760 ptlrpc_save_lock(req,
1761 &(dlm_handles[3]), LCK_EX);
1762 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
1763 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
1764 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
1773 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1776 req->rq_status = rc;
1780 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
1781 struct ptlrpc_request *, struct lustre_handle *);
1783 static mds_reinter reinters[REINT_MAX + 1] = {
1784 [REINT_SETATTR] mds_reint_setattr,
1785 [REINT_CREATE] mds_reint_create,
1786 [REINT_LINK] mds_reint_link,
1787 [REINT_UNLINK] mds_reint_unlink,
1788 [REINT_RENAME] mds_reint_rename,
1789 [REINT_OPEN] mds_open
1792 int mds_reint_rec(struct mds_update_record *rec, int offset,
1793 struct ptlrpc_request *req, struct lustre_handle *lockh)
1795 struct obd_device *obd = req->rq_export->exp_obd;
1796 struct obd_run_ctxt saved;
1800 /* checked by unpacker */
1801 LASSERT(rec->ur_opcode <= REINT_MAX &&
1802 reinters[rec->ur_opcode] != NULL);
1804 push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1805 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
1806 pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);