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_log.h>
42 #include <linux/lustre_fsfilt.h>
44 #include "mds_internal.h"
46 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
49 obd_transno_commit_cb(obd, transno, error);
52 struct mds_logcancel_data {
53 struct lov_mds_md *mlcd_lmm;
57 struct llog_cookie mlcd_cookies[0];
61 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
62 void *cb_data, int error)
64 struct mds_logcancel_data *mlcd = cb_data;
65 struct lov_stripe_md *lsm = NULL;
66 struct llog_ctxt *ctxt;
69 obd_transno_commit_cb(obd, transno, error);
71 CDEBUG(D_HA, "cancelling %d cookies\n",
72 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
74 rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
75 mlcd->mlcd_eadatalen);
77 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
78 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
81 ///* XXX 0 normally, SENDNOW for debug */);
82 ctxt = llog_get_context(&obd->obd_llogs,
83 mlcd->mlcd_cookies[0].lgc_subsys + 1);
84 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
85 sizeof(*mlcd->mlcd_cookies),
86 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
88 CERROR("error cancelling %d log cookies: rc %d\n",
89 (int)(mlcd->mlcd_cookielen /
90 sizeof(*mlcd->mlcd_cookies)), rc);
93 OBD_FREE(mlcd, mlcd->mlcd_size);
96 /* Assumes caller has already pushed us into the kernel context. */
97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
98 struct ptlrpc_request *req, int rc, __u32 op_data)
100 struct mds_export_data *med = &req->rq_export->exp_mds_data;
101 struct mds_client_data *mcd = med->med_mcd;
102 struct obd_device *obd = req->rq_export->exp_obd;
109 /* if the export has already been failed, we have no last_rcvd slot */
110 if (req->rq_export->exp_failed) {
111 CERROR("committing transaction for disconnected client\n");
120 if (handle == NULL) {
121 /* if we're starting our own xaction, use our own inode */
122 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
123 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
124 if (IS_ERR(handle)) {
125 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
126 RETURN(PTR_ERR(handle));
132 transno = req->rq_reqmsg->transno;
134 LASSERT(transno == 0);
135 } else if (transno == 0) {
136 spin_lock(&mds->mds_transno_lock);
137 transno = ++mds->mds_last_transno;
138 spin_unlock(&mds->mds_transno_lock);
140 spin_lock(&mds->mds_transno_lock);
141 if (transno > mds->mds_last_transno)
142 mds->mds_last_transno = transno;
143 spin_unlock(&mds->mds_transno_lock);
145 req->rq_repmsg->transno = req->rq_transno = transno;
146 mcd->mcd_last_transno = cpu_to_le64(transno);
147 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
148 mcd->mcd_last_result = cpu_to_le32(rc);
149 mcd->mcd_last_data = cpu_to_le32(op_data);
152 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb,
153 transno, handle, mds_commit_cb, NULL);
155 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
164 DEBUG_REQ(log_pri, req,
165 "wrote trans #"LPU64" client %s at idx %u: err = %d",
166 transno, mcd->mcd_uuid, med->med_idx, err);
168 err = mds_lov_write_objids(obd);
174 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
177 err = fsfilt_commit(obd, inode, handle, 0);
179 CERROR("error committing transaction: %d\n", err);
187 /* this gives the same functionality as the code between
188 * sys_chmod and inode_setattr
189 * chown_common and inode_setattr
190 * utimes and inode_setattr
192 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
194 time_t now = LTIME_S(CURRENT_TIME);
195 struct iattr *attr = &rec->ur_iattr;
196 unsigned int ia_valid = attr->ia_valid;
200 /* only fix up attrs if the client VFS didn't already */
201 if (!(ia_valid & ATTR_RAW))
204 if (!(ia_valid & ATTR_CTIME_SET))
205 LTIME_S(attr->ia_ctime) = now;
206 if (!(ia_valid & ATTR_ATIME_SET))
207 LTIME_S(attr->ia_atime) = now;
208 if (!(ia_valid & ATTR_MTIME_SET))
209 LTIME_S(attr->ia_mtime) = now;
211 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
215 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
216 if (rec->_ur_fsuid != inode->i_uid &&
217 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
221 if (ia_valid & ATTR_SIZE) {
222 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
226 if (ia_valid & ATTR_UID) {
229 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
231 if (attr->ia_uid == (uid_t) -1)
232 attr->ia_uid = inode->i_uid;
233 if (attr->ia_gid == (gid_t) -1)
234 attr->ia_gid = inode->i_gid;
235 attr->ia_mode = inode->i_mode;
237 * If the user or group of a non-directory has been
238 * changed by a non-root user, remove the setuid bit.
239 * 19981026 David C Niemi <niemi@tux.org>
241 * Changed this to apply to all users, including root,
242 * to avoid some races. This is the behavior we had in
243 * 2.0. The check for non-root was definitely wrong
244 * for 2.2 anyway, as it should have been using
245 * CAP_FSETID rather than fsuid -- 19990830 SD.
247 if ((inode->i_mode & S_ISUID) == S_ISUID &&
248 !S_ISDIR(inode->i_mode)) {
249 attr->ia_mode &= ~S_ISUID;
250 attr->ia_valid |= ATTR_MODE;
253 * Likewise, if the user or group of a non-directory
254 * has been changed by a non-root user, remove the
255 * setgid bit UNLESS there is no group execute bit
256 * (this would be a file marked for mandatory
257 * locking). 19981026 David C Niemi <niemi@tux.org>
259 * Removed the fsuid check (see the comment above) --
262 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
263 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
264 attr->ia_mode &= ~S_ISGID;
265 attr->ia_valid |= ATTR_MODE;
267 } else if (ia_valid & ATTR_MODE) {
268 int mode = attr->ia_mode;
270 if (attr->ia_mode == (mode_t) -1)
271 attr->ia_mode = inode->i_mode;
273 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
278 void mds_steal_ack_locks(struct ptlrpc_request *req)
280 struct obd_export *exp = req->rq_export;
281 struct list_head *tmp;
282 struct ptlrpc_reply_state *oldrep;
283 struct ptlrpc_service *svc;
285 char str[PTL_NALFMT_SIZE];
288 /* CAVEAT EMPTOR: spinlock order */
289 spin_lock_irqsave (&exp->exp_lock, flags);
290 list_for_each (tmp, &exp->exp_outstanding_replies) {
291 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
293 if (oldrep->rs_xid != req->rq_xid)
296 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
297 CERROR ("Resent req xid "LPX64" has mismatched opc: "
298 "new %d old %d\n", req->rq_xid,
299 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
301 svc = oldrep->rs_srv_ni->sni_service;
302 spin_lock (&svc->srv_lock);
304 list_del_init (&oldrep->rs_exp_list);
306 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
308 oldrep->rs_nlocks, oldrep,
309 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
310 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
312 for (i = 0; i < oldrep->rs_nlocks; i++)
313 ptlrpc_save_lock(req,
314 &oldrep->rs_locks[i],
315 oldrep->rs_modes[i]);
316 oldrep->rs_nlocks = 0;
318 DEBUG_REQ(D_HA, req, "stole locks for");
319 ptlrpc_schedule_difficult_reply (oldrep);
321 spin_unlock (&svc->srv_lock);
322 spin_unlock_irqrestore (&exp->exp_lock, flags);
325 spin_unlock_irqrestore (&exp->exp_lock, flags);
328 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
330 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
331 mcd->mcd_last_transno, mcd->mcd_last_result);
332 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
333 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
335 mds_steal_ack_locks(req);
338 static void reconstruct_reint_setattr(struct mds_update_record *rec,
339 int offset, struct ptlrpc_request *req)
341 struct mds_export_data *med = &req->rq_export->exp_mds_data;
342 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
344 struct mds_body *body;
346 mds_req_from_mcd(req, med->med_mcd);
348 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
350 LASSERT(PTR_ERR(de) == req->rq_status);
354 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
355 mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
356 mds_pack_inode2body(req2obd(req), body, de->d_inode);
358 /* Don't return OST-specific attributes if we didn't just set them */
359 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
360 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
361 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
362 body->valid |= OBD_MD_FLMTIME;
363 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
364 body->valid |= OBD_MD_FLATIME;
369 /* In the raw-setattr case, we lock the child inode.
370 * In the write-back case or if being called from open, the client holds a lock
373 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
374 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
375 struct ptlrpc_request *req,
376 struct lustre_handle *lh)
378 struct mds_obd *mds = mds_req2mds(req);
379 struct obd_device *obd = req->rq_export->exp_obd;
380 struct mds_body *body;
382 struct inode *inode = NULL;
383 struct lustre_handle lockh[2] = {{0}, {0}};
386 struct mds_logcancel_data *mlcd = NULL;
387 int rc = 0, cleanup_phase = 0, err, locked = 0;
390 LASSERT(offset == 0);
392 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
393 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
395 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
397 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
398 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
400 GOTO(cleanup, rc = PTR_ERR(de));
402 __u64 lockpart = MDS_INODELOCK_UPDATE;
403 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
404 lockpart |= MDS_INODELOCK_LOOKUP;
405 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
406 lockh, &parent_mode, NULL, 0, lockpart);
408 GOTO(cleanup, rc = PTR_ERR(de));
416 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
417 rec->ur_eadata != NULL)
420 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
422 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
424 GOTO(cleanup, rc = PTR_ERR(handle));
426 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
427 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
428 LTIME_S(rec->ur_iattr.ia_mtime),
429 LTIME_S(rec->ur_iattr.ia_ctime));
430 rc = mds_fix_attr(inode, rec);
434 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
435 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
436 (long)&rec->ur_iattr.ia_attr_flags);
438 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
440 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
441 rec->ur_eadata != NULL) {
442 struct lov_stripe_md *lsm = NULL;
444 rc = ll_permission(inode, MAY_WRITE, NULL);
448 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
449 mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
453 obd_free_memmd(mds->mds_osc_exp, &lsm);
455 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
461 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
462 mds_pack_inode2fid(obd, &body->fid1, inode);
463 mds_pack_inode2body(obd, body, inode);
465 /* Don't return OST-specific attributes if we didn't just set them */
466 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
467 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
468 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
469 body->valid |= OBD_MD_FLMTIME;
470 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
471 body->valid |= OBD_MD_FLATIME;
473 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
474 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
477 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
479 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
480 mlcd->mlcd_cookielen = rec->ur_cookielen;
481 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
482 mlcd->mlcd_cookielen;
483 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
484 mlcd->mlcd_cookielen);
485 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
486 mlcd->mlcd_eadatalen);
488 CERROR("unable to allocate log cancel data\n");
494 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
495 handle, mds_cancel_cookies_cb, mlcd);
496 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
497 switch (cleanup_phase) {
499 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
500 rec->ur_eadata != NULL)
505 if (lockh[1].cookie != 0)
506 ldlm_lock_decref(lockh + 1, parent_mode);
509 ldlm_lock_decref(lockh, LCK_PW);
511 ptlrpc_save_lock (req, lockh, LCK_PW);
526 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
527 struct ptlrpc_request *req)
529 struct mds_export_data *med = &req->rq_export->exp_mds_data;
530 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
531 struct dentry *parent, *child;
532 struct mds_body *body;
534 mds_req_from_mcd(req, med->med_mcd);
539 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
540 LASSERT(!IS_ERR(parent));
541 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
542 LASSERT(!IS_ERR(child));
543 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
544 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
545 mds_pack_inode2body(req2obd(req), body, child->d_inode);
550 static int mds_reint_create(struct mds_update_record *rec, int offset,
551 struct ptlrpc_request *req,
552 struct lustre_handle *lh)
554 struct dentry *dparent = NULL;
555 struct mds_obd *mds = mds_req2mds(req);
556 struct obd_device *obd = req->rq_export->exp_obd;
557 struct dentry *dchild = NULL;
558 struct inode *dir = NULL;
560 struct lustre_handle lockh[2] = {{0}, {0}};
562 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
564 struct dentry_params dp;
565 struct mea *mea = NULL;
569 LASSERT(offset == 0);
570 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
572 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
573 rec->ur_fid1->id, rec->ur_fid1->generation,
574 rec->ur_name, rec->ur_mode);
576 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
578 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
579 GOTO(cleanup, rc = -ESTALE);
581 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
582 &parent_mode, rec->ur_name,
583 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
584 if (IS_ERR(dparent)) {
585 rc = PTR_ERR(dparent);
586 CERROR("parent lookup error %d\n", rc);
589 cleanup_phase = 1; /* locked parent dentry */
590 dir = dparent->d_inode;
593 ldlm_lock_dump_handle(D_OTHER, lockh);
595 /* try to retrieve MEA data for this dir */
596 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
601 /* dir is already splitted, check is requested filename
602 * should live at this MDS or at another one */
604 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
605 if (mea->mea_master != i) {
606 CERROR("inapropriate MDS(%d) for %s. should be %d\n",
607 mea->mea_master, rec->ur_name, i);
608 GOTO(cleanup, rc = -ERESTART);
612 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
613 if (IS_ERR(dchild)) {
614 rc = PTR_ERR(dchild);
615 CERROR("child lookup error %d\n", rc);
619 cleanup_phase = 2; /* child dentry */
621 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
623 if (type == S_IFREG || type == S_IFDIR) {
624 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
626 /* dir got splitted */
627 GOTO(cleanup, rc = -ERESTART);
629 /* error happened during spitting */
635 if (dir->i_mode & S_ISGID) {
636 if (S_ISDIR(rec->ur_mode))
637 rec->ur_mode |= S_ISGID;
640 dchild->d_fsdata = (void *)&dp;
641 dp.p_inum = (unsigned long)rec->ur_fid2->id;
646 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
648 GOTO(cleanup, rc = PTR_ERR(handle));
649 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
657 /* as Peter asked, mkdir() should distribute new directories
658 * over the whole cluster in order to distribute namespace
659 * processing load. first, we calculate which MDS to use to
660 * put new directory's inode in */
661 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
663 if (i == mds->mds_num) {
664 /* inode will be created locally */
666 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
668 GOTO(cleanup, rc = PTR_ERR(handle));
670 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
673 nstripes = *(u16 *)rec->ur_eadata;
675 if (rc == 0 && nstripes) {
676 /* FIXME: error handling here */
677 mds_try_to_split_dir(obd, dchild,
680 } else if (!DENTRY_VALID(dchild)) {
681 /* inode will be created on another MDS */
682 struct obdo *oa = NULL;
683 struct mds_body *body;
685 /* first, create that inode */
689 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
690 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
691 OBD_MD_FLUID | OBD_MD_FLGID);
692 oa->o_mode = dir->i_mode;
693 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
695 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
696 CWARN("%s: replay dir creation %*s -> %u/%u\n",
697 obd->obd_name, rec->ur_namelen - 1,
698 rec->ur_name, (unsigned) rec->ur_fid2->id,
699 (unsigned) rec->ur_fid2->generation);
700 oa->o_id = rec->ur_fid2->id;
701 oa->o_generation = rec->ur_fid2->generation;
702 oa->o_flags |= OBD_FL_RECREATE_OBJS;
705 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
711 /* now, add new dir entry for it */
712 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
713 if (IS_ERR(handle)) {
715 GOTO(cleanup, rc = PTR_ERR(handle));
717 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
719 oa->o_id, oa->o_generation,
724 body = lustre_msg_buf(req->rq_repmsg,
725 offset, sizeof (*body));
726 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
727 body->fid1.id = oa->o_id;
729 body->fid1.generation = oa->o_generation;
732 /* requested name exists in the directory */
739 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
741 GOTO(cleanup, rc = PTR_ERR(handle));
742 if (rec->ur_tgt == NULL) /* no target supplied */
743 rc = -EINVAL; /* -EPROTO? */
745 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
753 int rdev = rec->ur_rdev;
754 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
756 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
757 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
762 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
763 dchild->d_fsdata = NULL;
764 GOTO(cleanup, rc = -EINVAL);
767 /* In case we stored the desired inum in here, we want to clean up. */
768 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
769 dchild->d_fsdata = NULL;
772 CDEBUG(D_INODE, "error during create: %d\n", rc);
774 } else if (dchild->d_inode) {
776 struct inode *inode = dchild->d_inode;
777 struct mds_body *body;
780 LTIME_S(iattr.ia_atime) = rec->ur_time;
781 LTIME_S(iattr.ia_ctime) = rec->ur_time;
782 LTIME_S(iattr.ia_mtime) = rec->ur_time;
783 iattr.ia_uid = rec->_ur_fsuid;
784 if (dir->i_mode & S_ISGID)
785 iattr.ia_gid = dir->i_gid;
787 iattr.ia_gid = rec->_ur_fsgid;
788 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
789 ATTR_MTIME | ATTR_CTIME;
791 if (rec->ur_fid2->id) {
792 LASSERT(rec->ur_fid2->id == inode->i_ino);
793 inode->i_generation = rec->ur_fid2->generation;
794 /* Dirtied and committed by the upcoming setattr. */
795 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
796 inode->i_ino, inode->i_generation);
798 struct lustre_handle child_ino_lockh;
800 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
801 inode->i_ino, inode->i_generation);
803 /* The inode we were allocated may have just been freed
804 * by an unlink operation. We take this lock to
805 * synchronize against the matching reply-ack-lock taken
806 * in unlink, to avoid replay problems if this reply
807 * makes it out to the client but the unlink's does not.
808 * See bug 2029 for more detail.*/
809 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
810 if (rc != ELDLM_OK) {
811 CERROR("error locking for unlink/create sync: "
814 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
818 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
820 CERROR("error on child setattr: rc = %d\n", rc);
822 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
823 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
825 CERROR("error on parent setattr: rc = %d\n", rc);
827 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
828 mds_pack_inode2fid(obd, &body->fid1, inode);
829 mds_pack_inode2body(obd, body, inode);
834 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
837 /* Destroy the file we just created. This should not need
838 * extra journal credits, as we have already modified all of
839 * the blocks needed in order to create the file in the first
844 err = vfs_rmdir(dir, dchild);
846 CERROR("rmdir in error path: %d\n", err);
849 err = vfs_unlink(dir, dchild);
851 CERROR("unlink in error path: %d\n", err);
857 switch (cleanup_phase) {
858 case 2: /* child dentry */
860 case 1: /* locked parent dentry */
862 if (lockh[1].cookie != 0)
863 ldlm_lock_decref(lockh + 1, parent_mode);
866 ldlm_lock_decref(lockh, LCK_PW);
868 ptlrpc_save_lock (req, lockh, LCK_PW);
874 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
878 OBD_FREE(mea, mea_size);
883 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
884 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
888 for (i = 0; i < RES_NAME_SIZE; i++) {
889 /* return 1 here, because enqueue_ordered will skip resources
890 * of all zeroes if they're sorted to the end of the list. */
891 if (res1->name[i] == 0 && res2->name[i] != 0)
893 if (res2->name[i] == 0 && res1->name[i] != 0)
896 if (res1->name[i] > res2->name[i])
898 if (res1->name[i] < res2->name[i])
905 if (memcmp(p1, p2, sizeof(*p1)) < 0)
911 /* This function doesn't use ldlm_match_or_enqueue because we're always called
912 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
913 * because they take the place of local semaphores.
915 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
916 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
917 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
918 struct lustre_handle *p1_lockh, int p1_lock_mode,
919 ldlm_policy_data_t *p1_policy,
920 struct ldlm_res_id *p2_res_id,
921 struct lustre_handle *p2_lockh, int p2_lock_mode,
922 ldlm_policy_data_t *p2_policy)
924 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
925 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
926 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
927 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
931 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
933 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
936 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
937 handles[1] = p1_lockh;
938 handles[0] = p2_lockh;
939 res_id[1] = p1_res_id;
940 res_id[0] = p2_res_id;
941 lock_modes[1] = p1_lock_mode;
942 lock_modes[0] = p2_lock_mode;
943 policies[1] = p1_policy;
944 policies[0] = p2_policy;
947 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
948 res_id[0]->name[0], res_id[1]->name[0]);
950 flags = LDLM_FL_LOCAL_ONLY;
951 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
952 LDLM_IBITS, policies[0], lock_modes[0], &flags,
953 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
954 NULL, 0, NULL, handles[0]);
957 ldlm_lock_dump_handle(D_OTHER, handles[0]);
959 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
960 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
961 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
962 ldlm_lock_addref(handles[1], lock_modes[1]);
963 } else if (res_id[1]->name[0] != 0) {
964 flags = LDLM_FL_LOCAL_ONLY;
965 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
966 *res_id[1], LDLM_IBITS, policies[1],
967 lock_modes[1], &flags, mds_blocking_ast,
968 ldlm_completion_ast, NULL, NULL, NULL, 0,
970 if (rc != ELDLM_OK) {
971 ldlm_lock_decref(handles[0], lock_modes[0]);
974 ldlm_lock_dump_handle(D_OTHER, handles[1]);
980 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
981 struct lustre_handle *p1_lockh, int p1_lock_mode,
982 ldlm_policy_data_t *p1_policy,
983 struct ldlm_res_id *p2_res_id,
984 struct lustre_handle *p2_lockh, int p2_lock_mode,
985 ldlm_policy_data_t *p2_policy,
986 struct ldlm_res_id *c1_res_id,
987 struct lustre_handle *c1_lockh, int c1_lock_mode,
988 ldlm_policy_data_t *c1_policy,
989 struct ldlm_res_id *c2_res_id,
990 struct lustre_handle *c2_lockh, int c2_lock_mode,
991 ldlm_policy_data_t *c2_policy)
993 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
994 c1_res_id, c2_res_id };
995 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
996 c1_lockh, c2_lockh };
997 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
998 c1_lock_mode, c2_lock_mode };
999 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1000 c1_policy, c2_policy};
1001 int rc, i, j, sorted, flags;
1005 "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1006 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1007 res_id[3]->name[0]);
1009 /* simple insertion sort - we have at most 4 elements */
1010 for (i = 1; i < 4; i++) {
1012 dlm_handles[4] = dlm_handles[i];
1013 res_id[4] = res_id[i];
1014 lock_modes[4] = lock_modes[i];
1015 policies[4] = policies[i];
1019 if (res_gt(res_id[j], res_id[4], policies[j],
1021 dlm_handles[j + 1] = dlm_handles[j];
1022 res_id[j + 1] = res_id[j];
1023 lock_modes[j + 1] = lock_modes[j];
1024 policies[j + 1] = policies[j];
1029 } while (j >= 0 && !sorted);
1031 dlm_handles[j + 1] = dlm_handles[4];
1032 res_id[j + 1] = res_id[4];
1033 lock_modes[j + 1] = lock_modes[4];
1034 policies[j + 1] = policies[4];
1038 "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1039 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1040 res_id[3]->name[0]);
1042 /* XXX we could send ASTs on all these locks first before blocking? */
1043 for (i = 0; i < 4; i++) {
1045 if (res_id[i]->name[0] == 0)
1048 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1049 (policies[i]->l_inodebits.bits &
1050 policies[i-1]->l_inodebits.bits) ) {
1051 memcpy(dlm_handles[i], dlm_handles[i-1],
1052 sizeof(*(dlm_handles[i])));
1053 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1055 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1056 *res_id[i], LDLM_IBITS,
1058 lock_modes[i], &flags,
1060 ldlm_completion_ast, NULL, NULL,
1061 NULL, 0, NULL, dlm_handles[i]);
1063 GOTO(out_err, rc = -EIO);
1064 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1071 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1076 /* In the unlikely case that the child changed while we were waiting
1077 * on the lock, we need to drop the lock on the old child and either:
1078 * - if the child has a lower resource name, then we have to also
1079 * drop the parent lock and regain the locks in the right order
1080 * - in the rename case, if the child has a lower resource name than one of
1081 * the other parent/child resources (maxres) we also need to reget the locks
1082 * - if the child has a higher resource name (this is the common case)
1083 * we can just get the lock on the new child (still in lock order)
1085 * Returns 0 if the child did not change or if it changed but could be locked.
1086 * Returns 1 if the child changed and we need to re-lock (no locks held).
1087 * Returns -ve error with a valid dchild (no locks held). */
1088 static int mds_verify_child(struct obd_device *obd,
1089 struct ldlm_res_id *parent_res_id,
1090 struct lustre_handle *parent_lockh,
1091 struct dentry *dparent, int parent_mode,
1092 struct ldlm_res_id *child_res_id,
1093 struct lustre_handle *child_lockh,
1094 struct dentry **dchildp, int child_mode,
1095 ldlm_policy_data_t *child_policy,
1096 const char *name, int namelen,
1097 struct ldlm_res_id *maxres)
1099 struct dentry *vchild, *dchild = *dchildp;
1100 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1103 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1105 GOTO(cleanup, rc = PTR_ERR(vchild));
1107 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1108 if (child_res_id->name[0] == vchild->d_inum &&
1109 child_res_id->name[1] == vchild->d_generation) {
1118 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1119 (vchild->d_inode != NULL &&
1120 child_res_id->name[0] == vchild->d_inode->i_ino &&
1121 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1130 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1131 vchild->d_inode, dchild ? dchild->d_inode : 0,
1132 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1133 child_res_id->name[0]);
1134 if (child_res_id->name[0] != 0)
1135 ldlm_lock_decref(child_lockh, child_mode);
1139 cleanup_phase = 1; /* parent lock only */
1140 *dchildp = dchild = vchild;
1142 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1144 if (dchild->d_inode) {
1145 child_res_id->name[0] = dchild->d_inode->i_ino;
1146 child_res_id->name[1] = dchild->d_inode->i_generation;
1148 child_res_id->name[0] = dchild->d_inum;
1149 child_res_id->name[1] = dchild->d_generation;
1152 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1153 res_gt(maxres, child_res_id, NULL, NULL)) {
1154 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1155 child_res_id->name[0], parent_res_id->name[0],
1157 GOTO(cleanup, rc = 1);
1160 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1161 *child_res_id, LDLM_IBITS, child_policy,
1162 child_mode, &flags, mds_blocking_ast,
1163 ldlm_completion_ast, NULL, NULL, NULL, 0,
1166 GOTO(cleanup, rc = -EIO);
1169 memset(child_res_id, 0, sizeof(*child_res_id));
1175 switch(cleanup_phase) {
1177 if (child_res_id->name[0] != 0)
1178 ldlm_lock_decref(child_lockh, child_mode);
1180 ldlm_lock_decref(parent_lockh, parent_mode);
1186 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1188 struct lustre_handle *parent_lockh,
1189 struct dentry **dparentp, int parent_mode,
1190 __u64 parent_lockpart, int *update_mode,
1191 char *name, int namelen,
1192 struct lustre_handle *child_lockh,
1193 struct dentry **dchildp, int child_mode,
1194 __u64 child_lockpart)
1196 struct ldlm_res_id child_res_id = { .name = {0} };
1197 struct ldlm_res_id parent_res_id = { .name = {0} };
1198 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1199 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1200 struct inode *inode;
1201 int rc = 0, cleanup_phase = 0;
1204 /* Step 1: Lookup parent */
1205 *dparentp = mds_fid2dentry(mds, fid, NULL);
1206 if (IS_ERR(*dparentp)) {
1207 rc = PTR_ERR(*dparentp);
1212 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1213 (*dparentp)->d_inode->i_ino, name);
1215 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1216 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1218 parent_lockh[1].cookie = 0;
1219 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1220 struct ldlm_res_id res_id = { .name = {0} };
1221 ldlm_policy_data_t policy;
1223 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1224 res_id.name[0] = (*dparentp)->d_inode->i_ino;
1225 res_id.name[1] = (*dparentp)->d_inode->i_generation;
1226 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1227 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
1228 LDLM_IBITS, &policy, *update_mode, &flags,
1229 mds_blocking_ast, ldlm_completion_ast,
1230 NULL, NULL, NULL, 0, NULL, parent_lockh+1);
1234 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1235 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1236 (*dparentp)->d_inode->i_ino,
1237 (*dparentp)->d_inode->i_generation,
1238 parent_res_id.name[2]);
1242 cleanup_phase = 1; /* parent dentry */
1244 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1245 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1246 if (IS_ERR(*dchildp)) {
1247 rc = PTR_ERR(*dchildp);
1248 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1252 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1253 /* inode lives on another MDS: return * mds/ino/gen
1254 * and LOOKUP lock. drop possible UPDATE lock! */
1255 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1256 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1257 child_res_id.name[0] = (*dchildp)->d_inum;
1258 child_res_id.name[1] = (*dchildp)->d_generation;
1262 inode = (*dchildp)->d_inode;
1264 inode = igrab(inode);
1268 child_res_id.name[0] = inode->i_ino;
1269 child_res_id.name[1] = inode->i_generation;
1274 cleanup_phase = 2; /* child dentry */
1276 /* Step 3: Lock parent and child in resource order. If child doesn't
1277 * exist, we still have to lock the parent and re-lookup. */
1278 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1279 &parent_policy, &child_res_id, child_lockh,
1280 child_mode, &child_policy);
1284 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1285 cleanup_phase = 4; /* child lock */
1287 cleanup_phase = 3; /* parent lock */
1289 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1290 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1291 parent_mode, &child_res_id, child_lockh,
1292 dchildp, child_mode, &child_policy,
1293 name, namelen, &parent_res_id);
1303 switch (cleanup_phase) {
1305 ldlm_lock_decref(child_lockh, child_mode);
1307 ldlm_lock_decref(parent_lockh, parent_mode);
1312 if (parent_lockh[1].cookie)
1313 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1322 void mds_reconstruct_generic(struct ptlrpc_request *req)
1324 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1326 mds_req_from_mcd(req, med->med_mcd);
1329 int mds_create_local_dentry(struct mds_update_record *rec,
1330 struct obd_device *obd)
1332 struct mds_obd *mds = &obd->u.mds;
1333 struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1334 int fidlen = 0, rc, cleanup_phase = 0;
1335 struct dentry *new_child = NULL;
1336 char *fidname = rec->ur_name;
1337 struct dentry *child = NULL;
1338 struct lustre_handle lockh[2] = {{0}, {0}};
1342 down(&fids_dir->i_sem);
1343 fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1344 CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1345 fidname, (unsigned) rec->ur_fid1->id,
1346 (unsigned) rec->ur_fid1->generation);
1348 new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1349 up(&fids_dir->i_sem);
1350 if (IS_ERR(new_child)) {
1351 CERROR("can't lookup %s: %d\n", fidname,
1352 (int) PTR_ERR(new_child));
1353 GOTO(cleanup, rc = PTR_ERR(new_child));
1357 if (new_child->d_inode != NULL) {
1358 /* nice. we've already have local dentry! */
1359 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1360 (unsigned) new_child->d_inode->i_ino,
1361 (unsigned) new_child->d_inode->i_generation);
1362 rec->ur_fid1->id = fids_dir->i_ino;
1363 rec->ur_fid1->generation = fids_dir->i_generation;
1364 rec->ur_namelen = fidlen + 1;
1365 GOTO(cleanup, rc = 0);
1368 /* new, local dentry will be added soon. we need no aliases here */
1371 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
1372 lockh, NULL, NULL, 0,
1373 MDS_INODELOCK_UPDATE);
1374 if (IS_ERR(child)) {
1375 CERROR("can't get victim\n");
1376 GOTO(cleanup, rc = PTR_ERR(child));
1380 handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1382 GOTO(cleanup, rc = PTR_ERR(handle));
1384 rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1385 rec->ur_fid1->id, rec->ur_fid1->generation,
1388 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1389 (unsigned long) child->d_inode->i_ino,
1390 (unsigned long) child->d_inode->i_generation, rc);
1392 if (S_ISDIR(child->d_inode->i_mode)) {
1393 fids_dir->i_nlink++;
1394 mark_inode_dirty(fids_dir);
1396 mark_inode_dirty(child->d_inode);
1398 fsfilt_commit(obd, fids_dir, handle, 0);
1400 rec->ur_fid1->id = fids_dir->i_ino;
1401 rec->ur_fid1->generation = fids_dir->i_generation;
1402 rec->ur_namelen = fidlen + 1;
1405 switch(cleanup_phase) {
1407 ldlm_lock_decref(lockh, LCK_EX);
1417 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1418 struct ptlrpc_request *slave)
1420 void *cookie, *cookie2;
1421 struct mds_body *body2;
1422 struct mds_body *body;
1426 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1427 LASSERT(body != NULL);
1429 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1430 LASSERT(body2 != NULL);
1432 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1436 memcpy(body2, body, sizeof(*body));
1437 body2->valid &= ~OBD_MD_FLCOOKIE;
1439 if (!(body->valid & OBD_MD_FLEASIZE) &&
1440 !(body->valid & OBD_MD_FLDIREA))
1443 if (body->eadatasize == 0) {
1444 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1448 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1450 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1451 LASSERT(ea != NULL);
1453 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1454 LASSERT(ea2 != NULL);
1456 memcpy(ea2, ea, body->eadatasize);
1458 if (body->valid & OBD_MD_FLCOOKIE) {
1459 LASSERT(master->rq_repmsg->buflens[2] >=
1460 slave->rq_repmsg->buflens[2]);
1461 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1462 slave->rq_repmsg->buflens[2]);
1463 LASSERT(cookie != NULL);
1465 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1466 master->rq_repmsg->buflens[2]);
1467 LASSERT(cookie2 != NULL);
1468 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1469 body2->valid |= OBD_MD_FLCOOKIE;
1474 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1475 struct ptlrpc_request *req,
1476 struct lustre_handle *parent_lockh,
1478 struct dentry *dparent,
1479 struct lustre_handle *child_lockh,
1480 struct dentry *dchild)
1482 struct mds_obd *mds = mds_req2mds(req);
1483 struct mdc_op_data op_data;
1484 int rc = 0, cleanup_phase = 0;
1485 struct ptlrpc_request *request = NULL;
1488 LASSERT(offset == 0 || offset == 2);
1490 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1491 rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1492 (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1494 /* time to drop i_nlink on remote MDS */
1495 op_data.fid1.mds = dchild->d_mdsnum;
1496 op_data.fid1.id = dchild->d_inum;
1497 op_data.fid1.generation = dchild->d_generation;
1498 op_data.create_mode = rec->ur_mode;
1499 op_data.namelen = 0;
1500 op_data.name = NULL;
1501 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1504 mds_copy_unlink_reply(req, request);
1505 ptlrpc_req_finished(request);
1508 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1509 req->rq_status = rc;
1512 if (parent_lockh[1].cookie != 0)
1513 ldlm_lock_decref(parent_lockh + 1, update_mode);
1515 ldlm_lock_decref(child_lockh, LCK_EX);
1517 ldlm_lock_decref(parent_lockh, LCK_PW);
1519 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1526 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1527 struct ptlrpc_request *req,
1528 struct lustre_handle *lh)
1530 struct dentry *dparent, *dchild;
1531 struct mds_obd *mds = mds_req2mds(req);
1532 struct obd_device *obd = req->rq_export->exp_obd;
1533 struct mds_body *body = NULL;
1534 struct inode *child_inode;
1535 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1536 struct lustre_handle child_lockh = {0}, child_reuse_lockh = {0};
1537 char fidname[LL_FID_NAMELEN];
1538 void *handle = NULL;
1539 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1540 int unlink_by_fid = 0;
1544 LASSERT(offset == 0 || offset == 2);
1546 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1547 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1549 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1551 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1552 GOTO(cleanup, rc = -ENOENT);
1554 if (rec->ur_namelen == 1) {
1555 /* this is request to drop i_nlink on local inode */
1557 rec->ur_name = fidname;
1558 rc = mds_create_local_dentry(rec, obd);
1561 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1562 parent_lockh, &dparent, LCK_PW,
1563 MDS_INODELOCK_UPDATE, &update_mode,
1564 rec->ur_name, rec->ur_namelen,
1565 &child_lockh, &dchild, LCK_EX,
1566 MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1570 if (dchild->d_flags & DCACHE_CROSS_REF) {
1571 /* we should have parent lock only here */
1572 LASSERT(unlink_by_fid == 0);
1573 LASSERT(dchild->d_mdsnum != mds->mds_num);
1574 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1575 update_mode, dparent, &child_lockh, dchild);
1579 cleanup_phase = 1; /* dchild, dparent, locks */
1582 child_inode = dchild->d_inode;
1583 if (child_inode == NULL) {
1584 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1585 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1586 GOTO(cleanup, rc = -ENOENT);
1589 cleanup_phase = 2; /* dchild has a lock */
1591 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1592 * reuse (see bug 2029). */
1593 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1597 cleanup_phase = 3; /* child inum lock */
1599 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1601 /* ldlm_reply in buf[0] if called via intent */
1605 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1606 LASSERT(body != NULL);
1608 /* If this is the last reference to this inode, get the OBD EA
1609 * data first so the client can destroy OST objects.
1610 * we only do the object removal if no open files remain.
1611 * Nobody can get at this name anymore because of the locks so
1612 * we make decisions here as to whether to remove the inode */
1613 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1614 mds_open_orphan_count(child_inode) == 0) {
1615 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1616 mds_pack_inode2body(obd, body, child_inode);
1617 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1619 if (!(body->valid & OBD_MD_FLEASIZE)) {
1620 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1621 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1627 /* We have to do these checks ourselves, in case we are making an
1628 * orphan. The client tells us whether rmdir() or unlink() was called,
1629 * so we need to return appropriate errors (bug 72).
1631 * We don't have to check permissions, because vfs_rename (called from
1632 * mds_open_unlink_rename) also calls may_delete. */
1633 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1634 if (!S_ISDIR(child_inode->i_mode))
1635 GOTO(cleanup, rc = -ENOTDIR);
1637 if (S_ISDIR(child_inode->i_mode))
1638 GOTO(cleanup, rc = -EISDIR);
1641 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1642 switch (child_inode->i_mode & S_IFMT) {
1644 /* Drop any lingering child directories before we start our
1645 * transaction, to avoid doing multiple inode dirty/delete
1646 * in our compound transaction (bug 1321). */
1647 shrink_dcache_parent(dchild);
1648 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1651 GOTO(cleanup, rc = PTR_ERR(handle));
1652 cleanup_phase = 4; /* transaction */
1653 rc = vfs_rmdir(dparent->d_inode, dchild);
1656 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1657 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1659 handle = fsfilt_start_log(obd, dparent->d_inode,
1660 FSFILT_OP_UNLINK, NULL,
1661 le32_to_cpu(lmm->lmm_stripe_count));
1663 GOTO(cleanup, rc = PTR_ERR(handle));
1665 cleanup_phase = 4; /* transaction */
1666 rc = vfs_unlink(dparent->d_inode, dchild);
1668 if (!rc && log_unlink)
1669 if (mds_log_op_unlink(obd, child_inode,
1670 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1671 req->rq_repmsg->buflens[offset + 1],
1672 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1673 req->rq_repmsg->buflens[offset + 2]) > 0)
1674 body->valid |= OBD_MD_FLCOOKIE;
1682 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1685 GOTO(cleanup, rc = PTR_ERR(handle));
1686 cleanup_phase = 4; /* transaction */
1687 rc = vfs_unlink(dparent->d_inode, dchild);
1690 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1693 GOTO(cleanup, rc = -EINVAL);
1701 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1702 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1703 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1705 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1707 CERROR("error on parent setattr: rc = %d\n", err);
1710 switch(cleanup_phase) {
1712 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1713 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1714 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1715 mds_open_orphan_count(dchild->d_inode) > 0) {
1716 /* filesystem is really going to destroy an inode
1717 * we have to delay this till inode is opened -bzzz */
1718 mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1720 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1723 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1724 "unlinked", 0, NULL);
1725 case 3: /* child ino-reuse lock */
1726 if (rc && body != NULL) {
1727 // Don't unlink the OST objects if the MDS unlink failed
1731 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1733 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1734 case 2: /* child lock */
1735 ldlm_lock_decref(&child_lockh, LCK_EX);
1736 case 1: /* child and parent dentry, parent lock */
1738 if (parent_lockh[1].cookie != 0)
1739 ldlm_lock_decref(parent_lockh + 1, update_mode);
1742 ldlm_lock_decref(parent_lockh, LCK_PW);
1744 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1751 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1754 req->rq_status = rc;
1759 * to service requests from remote MDS to increment i_nlink
1761 static int mds_reint_link_acquire(struct mds_update_record *rec,
1762 int offset, struct ptlrpc_request *req,
1763 struct lustre_handle *lh)
1765 struct obd_device *obd = req->rq_export->exp_obd;
1766 struct ldlm_res_id src_res_id = { .name = {0} };
1767 struct lustre_handle *handle = NULL, src_lockh = {0};
1768 struct mds_obd *mds = mds_req2mds(req);
1769 int rc = 0, cleanup_phase = 0;
1770 struct dentry *de_src = NULL;
1771 ldlm_policy_data_t policy;
1775 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1776 obd->obd_name, (unsigned) rec->ur_fid1->mds,
1777 (unsigned) rec->ur_fid1->id,
1778 (unsigned) rec->ur_fid1->generation);
1780 /* Step 1: Lookup the source inode and target directory by FID */
1781 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1783 GOTO(cleanup, rc = PTR_ERR(de_src));
1784 cleanup_phase = 1; /* source dentry */
1786 src_res_id.name[0] = de_src->d_inode->i_ino;
1787 src_res_id.name[1] = de_src->d_inode->i_generation;
1788 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1790 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1791 src_res_id, LDLM_IBITS, &policy,
1792 LCK_EX, &flags, mds_blocking_ast,
1793 ldlm_completion_ast, NULL, NULL,
1794 NULL, 0, NULL, &src_lockh);
1796 GOTO(cleanup, rc = -ENOLCK);
1797 cleanup_phase = 2; /* lock */
1799 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1801 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1802 if (IS_ERR(handle)) {
1803 rc = PTR_ERR(handle);
1806 de_src->d_inode->i_nlink++;
1807 mark_inode_dirty(de_src->d_inode);
1810 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1811 handle, req, rc, 0);
1813 switch (cleanup_phase) {
1816 ldlm_lock_decref(&src_lockh, LCK_EX);
1818 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1824 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1827 req->rq_status = rc;
1832 * request to link to foreign inode:
1833 * - acquire i_nlinks on this inode
1836 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1837 int offset, struct ptlrpc_request *req,
1838 struct lustre_handle *lh)
1840 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1841 struct obd_device *obd = req->rq_export->exp_obd;
1842 struct dentry *de_tgt_dir = NULL;
1843 struct mds_obd *mds = mds_req2mds(req);
1844 int rc = 0, cleanup_phase = 0;
1845 struct mdc_op_data op_data;
1846 struct ptlrpc_request *request = NULL;
1850 #define fmt "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1851 DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1852 (unsigned) rec->ur_fid2->mds,
1853 (unsigned) rec->ur_fid2->id,
1854 (unsigned) rec->ur_fid2->generation,
1855 rec->ur_namelen - 1, rec->ur_name,
1856 (unsigned) rec->ur_fid1->mds,
1857 (unsigned) rec->ur_fid1->id,
1858 (unsigned)rec->ur_fid1->generation);
1860 de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1861 tgt_dir_lockh, &update_mode,
1862 rec->ur_name, rec->ur_namelen - 1,
1863 MDS_INODELOCK_UPDATE);
1864 if (IS_ERR(de_tgt_dir))
1865 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1868 op_data.fid1 = *(rec->ur_fid1);
1869 op_data.namelen = 0;
1870 op_data.name = NULL;
1871 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1877 ptlrpc_req_finished(request);
1879 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1881 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1882 if (IS_ERR(handle)) {
1883 rc = PTR_ERR(handle);
1887 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1888 rec->ur_namelen - 1, rec->ur_fid1->id,
1889 rec->ur_fid1->generation, rec->ur_fid1->mds);
1893 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1894 handle, req, rc, 0);
1897 switch (cleanup_phase) {
1900 /* FIXME: drop i_nlink on remote inode here */
1901 CERROR("MUST drop drop i_nlink here\n");
1906 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1908 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
1911 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1913 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
1919 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1922 req->rq_status = rc;
1926 static int mds_reint_link(struct mds_update_record *rec, int offset,
1927 struct ptlrpc_request *req,
1928 struct lustre_handle *lh)
1930 struct obd_device *obd = req->rq_export->exp_obd;
1931 struct dentry *de_src = NULL;
1932 struct dentry *de_tgt_dir = NULL;
1933 struct dentry *dchild = NULL;
1934 struct mds_obd *mds = mds_req2mds(req);
1935 struct lustre_handle *handle = NULL;
1936 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
1937 struct ldlm_res_id src_res_id = { .name = {0} };
1938 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1939 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1940 ldlm_policy_data_t tgt_dir_policy =
1941 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1942 int rc = 0, cleanup_phase = 0;
1943 int update_mode = 0;
1946 LASSERT(offset == 0);
1948 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1949 rec->ur_fid1->id, rec->ur_fid1->generation,
1950 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1952 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1954 // memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
1955 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1956 GOTO(cleanup, rc = -ENOENT);
1958 if (rec->ur_fid1->mds != mds->mds_num) {
1959 rc = mds_reint_link_to_remote(rec, offset, req, lh);
1963 if (rec->ur_namelen == 1) {
1964 rc = mds_reint_link_acquire(rec, offset, req, lh);
1968 /* Step 1: Lookup the source inode and target directory by FID */
1969 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1971 GOTO(cleanup, rc = PTR_ERR(de_src));
1973 cleanup_phase = 1; /* source dentry */
1975 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1976 if (IS_ERR(de_tgt_dir)) {
1977 rc = PTR_ERR(de_tgt_dir);
1982 cleanup_phase = 2; /* target directory dentry */
1984 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1985 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1986 de_src->d_inode->i_ino);
1988 /* Step 2: Take the two locks */
1989 src_res_id.name[0] = de_src->d_inode->i_ino;
1990 src_res_id.name[1] = de_src->d_inode->i_generation;
1991 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1992 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1994 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
1996 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
1997 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1998 tgt_dir_res_id, LDLM_IBITS, &src_policy,
1999 update_mode, &flags, mds_blocking_ast,
2000 ldlm_completion_ast, NULL, NULL,
2001 NULL, 0, NULL, tgt_dir_lockh + 1);
2003 GOTO(cleanup, rc = -ENOLCK);
2005 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2006 rec->ur_namelen - 1);
2007 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2008 de_tgt_dir->d_inode->i_ino,
2009 de_tgt_dir->d_inode->i_generation,
2010 tgt_dir_res_id.name[2]);
2013 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2015 &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2020 cleanup_phase = 3; /* locks */
2022 /* Step 3: Lookup the child */
2023 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2024 if (IS_ERR(dchild)) {
2025 rc = PTR_ERR(dchild);
2026 if (rc != -EPERM && rc != -EACCES)
2027 CERROR("child lookup error %d\n", rc);
2031 cleanup_phase = 4; /* child dentry */
2033 if (dchild->d_inode) {
2034 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2035 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2040 /* Step 4: Do it. */
2041 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2043 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2044 if (IS_ERR(handle)) {
2045 rc = PTR_ERR(handle);
2049 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2050 if (rc && rc != -EPERM && rc != -EACCES)
2051 CERROR("vfs_link error %d\n", rc);
2053 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2054 handle, req, rc, 0);
2057 switch (cleanup_phase) {
2058 case 4: /* child dentry */
2062 ldlm_lock_decref(&src_lockh, LCK_EX);
2063 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2065 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2066 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2068 case 2: /* target dentry */
2070 if (tgt_dir_lockh[1].cookie && update_mode)
2071 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2075 case 1: /* source dentry */
2080 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2083 req->rq_status = rc;
2088 * add a hard link in the PENDING directory, only used by rename()
2090 static int mds_add_link_orphan(struct mds_update_record *rec,
2091 struct obd_device *obd,
2092 struct dentry *dentry)
2094 struct mds_obd *mds = &obd->u.mds;
2095 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2096 struct dentry *pending_child;
2097 char fidname[LL_FID_NAMELEN];
2101 LASSERT(dentry->d_inode);
2102 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2104 down(&pending_dir->i_sem);
2105 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2106 dentry->d_inode->i_generation);
2108 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2109 mds_open_orphan_count(dentry->d_inode),
2110 rec->ur_name, fidname);
2112 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2113 if (IS_ERR(pending_child))
2114 GOTO(out_lock, rc = PTR_ERR(pending_child));
2116 if (pending_child->d_inode != NULL) {
2117 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2118 LASSERT(pending_child->d_inode == dentry->d_inode);
2119 GOTO(out_dput, rc = 0);
2123 rc = vfs_link(dentry, pending_dir, pending_child);
2126 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2129 mds_inode_set_orphan(dentry->d_inode);
2131 l_dput(pending_child);
2133 up(&pending_dir->i_sem);
2137 /* The idea here is that we need to get four locks in the end:
2138 * one on each parent directory, one on each child. We need to take
2139 * these locks in some kind of order (to avoid deadlocks), and the order
2140 * I selected is "increasing resource number" order. We need to look up
2141 * the children, however, before we know what the resource number(s) are.
2142 * Thus the following plan:
2144 * 1,2. Look up the parents
2145 * 3,4. Look up the children
2146 * 5. Take locks on the parents and children, in order
2147 * 6. Verify that the children haven't changed since they were looked up
2149 * If there was a race and the children changed since they were first looked
2150 * up, it is possible that mds_verify_child() will be able to just grab the
2151 * lock on the new child resource (if it has a higher resource than any other)
2152 * but we need to compare against not only its parent, but also against the
2153 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2155 * We need the fancy igrab() on the child inodes because we aren't holding a
2156 * lock on the parent after the lookup is done, so dentry->d_inode may change
2157 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2159 static int mds_get_parents_children_locked(struct obd_device *obd,
2160 struct mds_obd *mds,
2161 struct ll_fid *p1_fid,
2162 struct dentry **de_srcdirp,
2163 struct ll_fid *p2_fid,
2164 struct dentry **de_tgtdirp,
2166 const char *old_name, int old_len,
2167 struct dentry **de_oldp,
2168 const char *new_name, int new_len,
2169 struct dentry **de_newp,
2170 struct lustre_handle *dlm_handles,
2173 struct ldlm_res_id p1_res_id = { .name = {0} };
2174 struct ldlm_res_id p2_res_id = { .name = {0} };
2175 struct ldlm_res_id c1_res_id = { .name = {0} };
2176 struct ldlm_res_id c2_res_id = { .name = {0} };
2177 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2178 /* Only dentry should change, but the inode itself would be
2180 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2181 /* If something is going to be replaced, both dentry and inode locks are
2183 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2184 MDS_INODELOCK_UPDATE}};
2185 struct ldlm_res_id *maxres_src, *maxres_tgt;
2186 struct inode *inode;
2187 int rc = 0, cleanup_phase = 0;
2190 /* Step 1: Lookup the source directory */
2191 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2192 if (IS_ERR(*de_srcdirp))
2193 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2195 cleanup_phase = 1; /* source directory dentry */
2197 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2198 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2200 /* Step 2: Lookup the target directory */
2201 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2202 *de_tgtdirp = dget(*de_srcdirp);
2204 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2205 if (IS_ERR(*de_tgtdirp)) {
2206 rc = PTR_ERR(*de_tgtdirp);
2212 cleanup_phase = 2; /* target directory dentry */
2214 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2215 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2218 dlm_handles[5].cookie = 0;
2219 dlm_handles[6].cookie = 0;
2220 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2221 /* Get a temp lock on just ino, gen to flush client cache and
2222 * to protect dirs from concurrent splitting */
2223 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2224 LCK_PW, &p_policy, &p2_res_id,
2225 &(dlm_handles[6]),LCK_PW,&p_policy);
2229 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2230 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2231 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2232 (*de_srcdirp)->d_inode->i_ino,
2233 (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2234 (*de_tgtdirp)->d_inode->i_ino,
2235 (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2239 /* Step 3: Lookup the source child entry */
2240 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2241 if (IS_ERR(*de_oldp)) {
2242 rc = PTR_ERR(*de_oldp);
2243 CERROR("old child lookup error (%*s): %d\n",
2244 old_len - 1, old_name, rc);
2248 cleanup_phase = 3; /* original name dentry */
2250 inode = (*de_oldp)->d_inode;
2251 if (inode != NULL) {
2252 inode = igrab(inode);
2254 GOTO(cleanup, rc = -ENOENT);
2256 c1_res_id.name[0] = inode->i_ino;
2257 c1_res_id.name[1] = inode->i_generation;
2259 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2260 c1_res_id.name[0] = (*de_oldp)->d_inum;
2261 c1_res_id.name[1] = (*de_oldp)->d_generation;
2264 /* Step 4: Lookup the target child entry */
2265 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2266 if (IS_ERR(*de_newp)) {
2267 rc = PTR_ERR(*de_newp);
2268 CERROR("new child lookup error (%*s): %d\n",
2269 old_len - 1, old_name, rc);
2273 cleanup_phase = 4; /* target dentry */
2275 inode = (*de_newp)->d_inode;
2276 if (inode != NULL) {
2277 inode = igrab(inode);
2281 c2_res_id.name[0] = inode->i_ino;
2282 c2_res_id.name[1] = inode->i_generation;
2284 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2285 c2_res_id.name[0] = (*de_newp)->d_inum;
2286 c2_res_id.name[1] = (*de_newp)->d_generation;
2290 /* Step 5: Take locks on the parents and child(ren) */
2291 maxres_src = &p1_res_id;
2292 maxres_tgt = &p2_res_id;
2293 cleanup_phase = 4; /* target dentry */
2295 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2296 maxres_src = &c1_res_id;
2297 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2298 maxres_tgt = &c2_res_id;
2300 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2302 &p2_res_id, &dlm_handles[1], parent_mode,
2304 &c1_res_id, &dlm_handles[2], child_mode,
2306 &c2_res_id, &dlm_handles[3], child_mode,
2311 cleanup_phase = 6; /* parent and child(ren) locks */
2313 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2314 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2315 parent_mode, &c1_res_id, &dlm_handles[2],
2316 de_oldp, child_mode, &c1_policy, old_name,old_len,
2319 if (c2_res_id.name[0] != 0)
2320 ldlm_lock_decref(&dlm_handles[3], child_mode);
2321 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2328 if (!DENTRY_VALID(*de_oldp))
2329 GOTO(cleanup, rc = -ENOENT);
2331 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2332 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2333 parent_mode, &c2_res_id, &dlm_handles[3],
2334 de_newp, child_mode, &c2_policy, new_name,
2335 new_len, maxres_src);
2337 ldlm_lock_decref(&dlm_handles[2], child_mode);
2338 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2348 switch (cleanup_phase) {
2349 case 6: /* child lock(s) */
2350 if (c2_res_id.name[0] != 0)
2351 ldlm_lock_decref(&dlm_handles[3], child_mode);
2352 if (c1_res_id.name[0] != 0)
2353 ldlm_lock_decref(&dlm_handles[2], child_mode);
2354 case 5: /* parent locks */
2355 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2356 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2357 case 4: /* target dentry */
2359 case 3: /* source dentry */
2361 case 2: /* target directory dentry */
2362 l_dput(*de_tgtdirp);
2363 case 1: /* source directry dentry */
2364 l_dput(*de_srcdirp);
2371 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2372 int offset, struct ptlrpc_request *req)
2374 struct obd_device *obd = req->rq_export->exp_obd;
2375 struct dentry *de_srcdir = NULL;
2376 struct dentry *de_new = NULL;
2377 struct mds_obd *mds = mds_req2mds(req);
2378 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2379 struct lustre_handle child_lockh = {0};
2380 int cleanup_phase = 0;
2381 void *handle = NULL;
2382 int update_mode, rc = 0;
2385 /* another MDS executing rename operation has asked us
2386 * to create target name. such a creation should destroy
2387 * existing target name */
2389 CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2390 obd->obd_name, rec->ur_tgt,
2391 (unsigned long) rec->ur_fid1->mds,
2392 (unsigned long) rec->ur_fid1->id,
2393 (unsigned long) rec->ur_fid1->generation);
2395 /* first, lookup the target */
2396 child_lockh.cookie = 0;
2397 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2398 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2399 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2400 &child_lockh, &de_new, LCK_EX,
2401 MDS_INODELOCK_LOOKUP);
2408 LASSERT(de_srcdir->d_inode);
2411 if (de_new->d_inode) {
2412 /* name exists and points to local inode
2413 * try to unlink this name and create new one */
2414 CERROR("%s: %s points to local inode %lu/%lu\n",
2415 obd->obd_name, rec->ur_tgt,
2416 (unsigned long) de_new->d_inode->i_ino,
2417 (unsigned long) de_new->d_inode->i_generation);
2418 handle = fsfilt_start(obd, de_srcdir->d_inode,
2419 FSFILT_OP_RENAME, NULL);
2421 GOTO(cleanup, rc = PTR_ERR(handle));
2422 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2425 } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2426 /* name exists adn points to remove inode */
2427 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2428 obd->obd_name, rec->ur_tgt,
2429 (unsigned long) de_new->d_mdsnum,
2430 (unsigned long) de_new->d_inum,
2431 (unsigned long) de_new->d_generation);
2433 /* name doesn't exist. the simplest case */
2434 handle = fsfilt_start(obd, de_srcdir->d_inode,
2435 FSFILT_OP_LINK, NULL);
2437 GOTO(cleanup, rc = PTR_ERR(handle));
2441 rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2442 rec->ur_tgtlen - 1, rec->ur_fid1->id,
2443 rec->ur_fid1->generation, rec->ur_fid1->mds);
2445 CERROR("add_dir_entry() returned error %d\n", rc);
2448 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2449 handle, req, rc, 0);
2450 switch(cleanup_phase) {
2454 if (parent_lockh[1].cookie != 0)
2455 ldlm_lock_decref(&parent_lockh[1], update_mode);
2457 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2458 if (child_lockh.cookie != 0)
2459 ldlm_lock_decref(&child_lockh, LCK_EX);
2467 req->rq_status = rc;
2472 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2473 struct ptlrpc_request *req)
2475 struct obd_device *obd = req->rq_export->exp_obd;
2476 struct ptlrpc_request *req2 = NULL;
2477 struct dentry *de_srcdir = NULL;
2478 struct dentry *de_old = NULL;
2479 struct mds_obd *mds = mds_req2mds(req);
2480 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2481 struct lustre_handle child_lockh = {0};
2482 struct mdc_op_data opdata;
2483 void *handle = NULL;
2484 int update_mode, rc = 0;
2487 CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2488 obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2489 memset(&opdata, 0, sizeof(opdata));
2491 child_lockh.cookie = 0;
2492 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2493 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2494 &update_mode, rec->ur_name,
2495 rec->ur_namelen, &child_lockh, &de_old,
2496 LCK_EX, MDS_INODELOCK_LOOKUP);
2499 LASSERT(de_srcdir->d_inode);
2502 /* we already know the target should be created on another MDS
2503 * so, we have to request that MDS to do it */
2505 /* prepare source fid */
2506 if (de_old->d_flags & DCACHE_CROSS_REF) {
2507 LASSERT(de_old->d_inode == NULL);
2508 CDEBUG(D_OTHER, "request to move remote name\n");
2509 opdata.fid1.mds = de_old->d_mdsnum;
2510 opdata.fid1.id = de_old->d_inum;
2511 opdata.fid1.generation = de_old->d_generation;
2512 } else if (de_old->d_inode == NULL) {
2513 /* oh, source doesn't exist */
2514 GOTO(cleanup, rc = -ENOENT);
2516 LASSERT(de_old->d_inode != NULL);
2517 CDEBUG(D_OTHER, "request to move local name\n");
2518 opdata.fid1.mds = mds->mds_num;
2519 opdata.fid1.id = de_old->d_inode->i_ino;
2520 opdata.fid1.generation = de_old->d_inode->i_generation;
2523 opdata.fid2 = *(rec->ur_fid2);
2524 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2525 rec->ur_tgtlen - 1, &req2);
2530 handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2532 GOTO(cleanup, rc = PTR_ERR(handle));
2533 rc = fsfilt_del_dir_entry(obd, de_old);
2538 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2539 handle, req, rc, 0);
2541 ptlrpc_req_finished(req2);
2544 if (parent_lockh[1].cookie != 0)
2545 ldlm_lock_decref(&parent_lockh[1], update_mode);
2547 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2548 if (child_lockh.cookie != 0)
2549 ldlm_lock_decref(&child_lockh, LCK_EX);
2554 req->rq_status = rc;
2559 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2560 struct ptlrpc_request *req,
2561 struct lustre_handle *lockh)
2563 struct obd_device *obd = req->rq_export->exp_obd;
2564 struct dentry *de_srcdir = NULL;
2565 struct dentry *de_tgtdir = NULL;
2566 struct dentry *de_old = NULL;
2567 struct dentry *de_new = NULL;
2568 struct mds_obd *mds = mds_req2mds(req);
2569 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2570 struct mds_body *body = NULL;
2571 int rc = 0, lock_count = 3;
2572 int cleanup_phase = 0;
2573 void *handle = NULL;
2576 LASSERT(offset == 0);
2578 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2579 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2580 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2582 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2584 if (rec->ur_namelen == 1) {
2585 rc = mds_reint_rename_create_name(rec, offset, req);
2589 if (rec->ur_fid2->mds != mds->mds_num) {
2590 rc = mds_reint_rename_to_remote(rec, offset, req);
2594 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2595 rec->ur_fid2, &de_tgtdir, LCK_PW,
2596 rec->ur_name, rec->ur_namelen,
2597 &de_old, rec->ur_tgt,
2598 rec->ur_tgtlen, &de_new,
2599 dlm_handles, LCK_EX);
2603 cleanup_phase = 1; /* parent(s), children, locks */
2605 if (de_new->d_inode)
2608 /* sanity check for src inode */
2609 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2610 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2611 GOTO(cleanup, rc = -EINVAL);
2613 /* sanity check for dest inode */
2614 if (de_new->d_inode &&
2615 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2616 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2617 GOTO(cleanup, rc = -EINVAL);
2619 if (de_old->d_inode == de_new->d_inode)
2620 GOTO(cleanup, rc = 0);
2622 /* if we are about to remove the target at first, pass the EA of
2623 * that inode to client to perform and cleanup on OST */
2624 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2625 LASSERT(body != NULL);
2627 if (de_new->d_inode &&
2628 S_ISREG(de_new->d_inode->i_mode) &&
2629 de_new->d_inode->i_nlink == 1 &&
2630 mds_open_orphan_count(de_new->d_inode) == 0) {
2631 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2632 mds_pack_inode2body(obd, body, de_new->d_inode);
2633 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2634 if (!(body->valid & OBD_MD_FLEASIZE)) {
2635 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2636 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2638 /* XXX need log unlink? */
2642 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2643 de_srcdir->d_inode->i_sb);
2645 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2647 GOTO(cleanup, rc = PTR_ERR(handle));
2649 /* FIXME need adjust the journal block count? */
2650 /* if the target should be moved to PENDING, we at first increase the
2651 * link and later vfs_rename() will decrease the link count again */
2652 if (de_new->d_inode &&
2653 S_ISREG(de_new->d_inode->i_mode) &&
2654 de_new->d_inode->i_nlink == 1 &&
2655 mds_open_orphan_count(de_new->d_inode) > 0) {
2656 rc = mds_add_link_orphan(rec, obd, de_new);
2662 de_old->d_fsdata = req;
2663 de_new->d_fsdata = req;
2664 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2669 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2670 handle, req, rc, 0);
2671 switch (cleanup_phase) {
2674 if (dlm_handles[5].cookie != 0)
2675 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2676 if (dlm_handles[6].cookie != 0)
2677 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2680 if (lock_count == 4)
2681 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2682 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2683 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2684 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2686 if (lock_count == 4)
2687 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2688 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2689 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2690 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2699 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2702 req->rq_status = rc;
2706 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2707 struct ptlrpc_request *, struct lustre_handle *);
2709 static mds_reinter reinters[REINT_MAX + 1] = {
2710 [REINT_SETATTR] mds_reint_setattr,
2711 [REINT_CREATE] mds_reint_create,
2712 [REINT_LINK] mds_reint_link,
2713 [REINT_UNLINK] mds_reint_unlink,
2714 [REINT_RENAME] mds_reint_rename,
2715 [REINT_OPEN] mds_open
2718 int mds_reint_rec(struct mds_update_record *rec, int offset,
2719 struct ptlrpc_request *req, struct lustre_handle *lockh)
2721 struct obd_device *obd = req->rq_export->exp_obd;
2722 struct lvfs_run_ctxt saved;
2725 /* checked by unpacker */
2726 LASSERT(rec->ur_opcode <= REINT_MAX &&
2727 reinters[rec->ur_opcode] != NULL);
2729 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2730 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2731 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);