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/jbd.h>
35 #include <linux/namei.h>
36 #include <linux/ext3_fs.h>
37 #include <linux/obd_support.h>
38 #include <linux/obd_class.h>
39 #include <linux/obd.h>
40 #include <linux/lustre_lib.h>
41 #include <linux/lustre_idl.h>
42 #include <linux/lustre_mds.h>
43 #include <linux/lustre_dlm.h>
44 #include <linux/lustre_log.h>
45 #include <linux/lustre_fsfilt.h>
46 #include <linux/lustre_lite.h>
47 #include "mds_internal.h"
49 struct mds_logcancel_data {
50 struct lov_mds_md *mlcd_lmm;
54 struct llog_cookie mlcd_cookies[0];
57 static void mds_cancel_cookies_cb(struct obd_device *obd,
58 __u64 transno, void *cb_data,
61 struct mds_logcancel_data *mlcd = cb_data;
62 struct lov_stripe_md *lsm = NULL;
63 struct llog_ctxt *ctxt;
66 obd_transno_commit_cb(obd, transno, error);
68 CDEBUG(D_HA, "cancelling %d cookies\n",
69 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
71 rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, mlcd->mlcd_lmm,
72 mlcd->mlcd_eadatalen);
74 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
75 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
78 ///* XXX 0 normally, SENDNOW for debug */);
79 ctxt = llog_get_context(&obd->obd_llogs,
80 mlcd->mlcd_cookies[0].lgc_subsys + 1);
81 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
82 sizeof(*mlcd->mlcd_cookies),
83 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
85 CERROR("error cancelling %d log cookies: rc %d\n",
86 (int)(mlcd->mlcd_cookielen /
87 sizeof(*mlcd->mlcd_cookies)), rc);
88 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
91 OBD_FREE(mlcd, mlcd->mlcd_size);
94 /* Assumes caller has already pushed us into the kernel context. */
95 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
96 struct ptlrpc_request *req, int rc, __u32 op_data)
98 struct mds_export_data *med = &req->rq_export->exp_mds_data;
99 struct obd_device *obd = req->rq_export->exp_obd;
100 struct mds_client_data *mcd = med->med_mcd;
101 int err, log_pri = D_HA;
106 /* if the export has already been failed, we have no last_rcvd slot */
107 if (req->rq_export->exp_failed) {
108 CERROR("committing transaction for disconnected client\n");
110 GOTO(out_commit, rc);
117 if (handle == NULL) {
118 /* if we're starting our own xaction, use our own inode */
119 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
120 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
121 if (IS_ERR(handle)) {
122 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
123 RETURN(PTR_ERR(handle));
129 transno = req->rq_reqmsg->transno;
131 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
132 } else if (transno == 0) {
133 spin_lock(&mds->mds_transno_lock);
134 transno = ++mds->mds_last_transno;
135 spin_unlock(&mds->mds_transno_lock);
137 spin_lock(&mds->mds_transno_lock);
138 if (transno > mds->mds_last_transno)
139 mds->mds_last_transno = transno;
140 spin_unlock(&mds->mds_transno_lock);
142 req->rq_repmsg->transno = req->rq_transno = transno;
143 if (req->rq_reqmsg->opc == MDS_CLOSE) {
144 mcd->mcd_last_close_transno = cpu_to_le64(transno);
145 mcd->mcd_last_close_xid = cpu_to_le64(req->rq_xid);
146 mcd->mcd_last_close_result = cpu_to_le32(rc);
147 mcd->mcd_last_close_data = cpu_to_le32(op_data);
149 mcd->mcd_last_transno = cpu_to_le64(transno);
150 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
151 mcd->mcd_last_result = cpu_to_le32(rc);
152 mcd->mcd_last_data = cpu_to_le32(op_data);
155 fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
156 mds_commit_last_transno_cb, NULL);
158 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
159 sizeof(*mcd), &off, 0);
167 DEBUG_REQ(log_pri, req,
168 "wrote trans #"LPU64" client %s at idx %u: err = %d",
169 transno, mcd->mcd_uuid, med->med_idx, err);
171 err = mds_update_last_fid(obd, handle, 0);
178 err = mds_dt_write_objids(obd);
184 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
188 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
189 req->rq_export->exp_sync);
191 CERROR("error committing transaction: %d\n", err);
199 /* this gives the same functionality as the code between
200 * sys_chmod and inode_setattr
201 * chown_common and inode_setattr
202 * utimes and inode_setattr
205 /* Just for the case if we have some clients that know about ATTR_RAW */
206 #define ATTR_RAW 8192
208 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
210 time_t now = LTIME_S(CURRENT_TIME);
211 struct iattr *attr = &rec->ur_iattr;
212 unsigned int ia_valid = attr->ia_valid;
216 /* only fix up attrs if the client VFS didn't already */
218 if (!(ia_valid & ATTR_RAW))
221 if (!(ia_valid & ATTR_CTIME_SET))
222 LTIME_S(attr->ia_ctime) = now;
223 if (!(ia_valid & ATTR_ATIME_SET))
224 LTIME_S(attr->ia_atime) = now;
225 if (!(ia_valid & ATTR_MTIME_SET))
226 LTIME_S(attr->ia_mtime) = now;
228 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
232 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
233 if (rec->ur_fsuid != inode->i_uid &&
234 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
238 if (ia_valid & ATTR_SIZE) {
239 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
243 if (ia_valid & ATTR_UID) {
246 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
248 if (attr->ia_uid == (uid_t) -1)
249 attr->ia_uid = inode->i_uid;
250 if (attr->ia_gid == (gid_t) -1)
251 attr->ia_gid = inode->i_gid;
252 attr->ia_mode = inode->i_mode;
254 * If the user or group of a non-directory has been
255 * changed by a non-root user, remove the setuid bit.
256 * 19981026 David C Niemi <niemi@tux.org>
258 * Changed this to apply to all users, including root,
259 * to avoid some races. This is the behavior we had in
260 * 2.0. The check for non-root was definitely wrong
261 * for 2.2 anyway, as it should have been using
262 * CAP_FSETID rather than fsuid -- 19990830 SD.
264 if ((inode->i_mode & S_ISUID) == S_ISUID &&
265 !S_ISDIR(inode->i_mode)) {
266 attr->ia_mode &= ~S_ISUID;
267 attr->ia_valid |= ATTR_MODE;
270 * Likewise, if the user or group of a non-directory
271 * has been changed by a non-root user, remove the
272 * setgid bit UNLESS there is no group execute bit
273 * (this would be a file marked for mandatory
274 * locking). 19981026 David C Niemi <niemi@tux.org>
276 * Removed the fsuid check (see the comment above) --
279 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
280 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
281 attr->ia_mode &= ~S_ISGID;
282 attr->ia_valid |= ATTR_MODE;
284 } else if (ia_valid & ATTR_MODE) {
285 int mode = attr->ia_mode;
287 if (attr->ia_mode == (mode_t) -1)
288 attr->ia_mode = inode->i_mode;
290 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
295 void mds_steal_ack_locks(struct ptlrpc_request *req)
297 struct obd_export *exp = req->rq_export;
298 char str[PTL_NALFMT_SIZE];
299 struct list_head *tmp;
300 struct ptlrpc_reply_state *oldrep;
301 struct ptlrpc_service *svc;
302 struct llog_create_locks *lcl;
306 /* CAVEAT EMPTOR: spinlock order */
307 spin_lock_irqsave (&exp->exp_lock, flags);
308 list_for_each (tmp, &exp->exp_outstanding_replies) {
309 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
311 if (oldrep->rs_xid != req->rq_xid)
314 if (oldrep->rs_msg->opc != req->rq_reqmsg->opc)
315 CERROR ("Resent req xid "LPX64" has mismatched opc: "
316 "new %d old %d\n", req->rq_xid,
317 req->rq_reqmsg->opc, oldrep->rs_msg->opc);
319 svc = oldrep->rs_srv_ni->sni_service;
320 spin_lock (&svc->srv_lock);
322 list_del_init (&oldrep->rs_exp_list);
324 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
325 " o%d NID %s\n", oldrep->rs_nlocks, oldrep,
326 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg->opc,
327 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
329 for (i = 0; i < oldrep->rs_nlocks; i++)
330 ptlrpc_save_lock(req,
331 &oldrep->rs_locks[i],
332 oldrep->rs_modes[i]);
333 oldrep->rs_nlocks = 0;
335 lcl = oldrep->rs_llog_locks;
336 oldrep->rs_llog_locks = NULL;
338 ptlrpc_save_llog_lock(req, lcl);
340 DEBUG_REQ(D_HA, req, "stole locks for");
341 ptlrpc_schedule_difficult_reply (oldrep);
343 spin_unlock (&svc->srv_lock);
344 spin_unlock_irqrestore (&exp->exp_lock, flags);
347 spin_unlock_irqrestore (&exp->exp_lock, flags);
350 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
352 if (req->rq_reqmsg->opc == MDS_CLOSE) {
353 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
354 mcd->mcd_last_close_transno, mcd->mcd_last_close_result);
355 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_close_transno;
356 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_close_result;
358 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
359 mcd->mcd_last_transno, mcd->mcd_last_result);
360 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
361 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
364 mds_steal_ack_locks(req);
367 static void reconstruct_reint_setattr(struct mds_update_record *rec,
368 int offset, struct ptlrpc_request *req)
370 struct mds_export_data *med = &req->rq_export->exp_mds_data;
371 struct mds_body *body;
374 mds_req_from_mcd(req, med->med_mcd);
376 de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
378 LASSERT(PTR_ERR(de) == req->rq_status);
382 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
383 mds_pack_inode2body(req2obd(req), body, de->d_inode, 1);
385 /* Don't return OST-specific attributes if we didn't just set them */
386 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
387 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
388 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
389 body->valid |= OBD_MD_FLMTIME;
390 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
391 body->valid |= OBD_MD_FLATIME;
396 /* In the raw-setattr case, we lock the child inode.
397 * In the write-back case or if being called from open, the client holds a lock
400 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
401 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
402 struct ptlrpc_request *req, struct lustre_handle *lh)
404 struct mds_obd *mds = mds_req2mds(req);
405 struct obd_device *obd = req->rq_export->exp_obd;
406 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
407 struct mds_body *body;
409 struct inode *inode = NULL;
410 struct lustre_handle lockh[2] = {{0}, {0}};
413 struct mds_logcancel_data *mlcd = NULL;
414 int rc = 0, cleanup_phase = 0, err;
418 LASSERT(offset == 1);
420 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
421 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
422 rec->ur_iattr.ia_valid);
424 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
425 MD_COUNTER_INCREMENT(obd, setattr);
427 if (med->med_remote) {
428 if (rec->ur_iattr.ia_valid & ATTR_GID) {
429 CWARN("Deny chgrp from remote client\n");
430 GOTO(cleanup, rc = -EPERM);
432 if (rec->ur_iattr.ia_valid & ATTR_UID) {
435 uid = mds_idmap_lookup_uid(med->med_idmap, 0,
436 rec->ur_iattr.ia_uid);
437 if (uid == MDS_IDMAP_NOTFOUND) {
438 CWARN("Deny chown to uid %u\n",
439 rec->ur_iattr.ia_uid);
440 GOTO(cleanup, rc = -EPERM);
442 rec->ur_iattr.ia_uid = uid;
446 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
447 de = mds_id2dentry(obd, rec->ur_id1, NULL);
449 GOTO(cleanup, rc = PTR_ERR(de));
451 __u64 lockpart = MDS_INODELOCK_UPDATE;
452 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
453 lockpart |= MDS_INODELOCK_LOOKUP;
454 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
455 lockh, &parent_mode, NULL, 0, lockpart);
457 GOTO(cleanup, rc = PTR_ERR(de));
465 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
466 rec->ur_eadata != NULL)
469 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
471 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
473 GOTO(cleanup, rc = PTR_ERR(handle));
475 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
476 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
477 LTIME_S(rec->ur_iattr.ia_mtime),
478 LTIME_S(rec->ur_iattr.ia_ctime));
479 rc = mds_fix_attr(inode, rec);
483 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
484 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
485 (long)&rec->ur_iattr.ia_attr_flags);
487 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
490 if (rec->ur_iattr.ia_valid & ATTR_EA) {
491 int flags = (int)rec->ur_iattr.ia_attr_flags;
494 if (inode->i_op && inode->i_op->setxattr)
495 rc = inode->i_op->setxattr(de, rec->ur_eadata,
496 rec->ur_ea2data, rec->ur_ea2datalen,
498 } else if (rec->ur_iattr.ia_valid & ATTR_EA_RM) {
500 if (inode->i_op && inode->i_op->removexattr)
501 rc = inode->i_op->removexattr(de,
503 } else if ((S_ISREG(inode->i_mode) ||
504 S_ISDIR(inode->i_mode)) && rec->ur_eadata != NULL) {
505 struct lov_stripe_md *lsm = NULL;
506 struct lov_user_md *lum = NULL;
508 rc = ll_permission(inode, MAY_WRITE, NULL);
512 lum = rec->ur_eadata;
513 /* if lmm_stripe_size is -1 delete default stripe from dir */
514 if (S_ISDIR(inode->i_mode) &&
515 lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1)){
516 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, EA_LOV);
520 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_dt_exp,
521 0, &lsm, rec->ur_eadata);
525 obd_free_memmd(mds->mds_dt_exp, &lsm);
526 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
527 rec->ur_eadatalen, EA_LOV);
534 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
535 mds_pack_inode2body(obd, body, inode, 1);
537 /* Don't return OST-specific attributes if we didn't just set them */
538 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
539 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
540 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
541 body->valid |= OBD_MD_FLMTIME;
542 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
543 body->valid |= OBD_MD_FLATIME;
545 mds_body_do_reverse_map(med, body);
547 /* The logcookie should be no use anymore, why nobody remove
548 * following code block?
550 LASSERT(rec->ur_cookielen == 0);
551 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_dt_obd)) {
552 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
555 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
557 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
558 mlcd->mlcd_cookielen = rec->ur_cookielen;
559 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
560 mlcd->mlcd_cookielen;
561 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
562 mlcd->mlcd_cookielen);
563 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
564 mlcd->mlcd_eadatalen);
566 CERROR("unable to allocate log cancel data\n");
572 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
573 handle, mds_cancel_cookies_cb, mlcd);
574 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
575 switch (cleanup_phase) {
577 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
578 rec->ur_eadata != NULL)
583 if (lockh[1].cookie != 0)
584 ldlm_lock_decref(lockh + 1, parent_mode);
587 ldlm_lock_decref(lockh, LCK_PW);
589 ptlrpc_save_lock (req, lockh, LCK_PW);
604 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
605 struct ptlrpc_request *req)
607 struct mds_export_data *med = &req->rq_export->exp_mds_data;
608 struct dentry *parent, *child;
609 struct mds_body *body;
612 mds_req_from_mcd(req, med->med_mcd);
614 if (req->rq_status) {
619 parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
620 LASSERT(!IS_ERR(parent));
621 child = ll_lookup_one_len(rec->ur_name, parent,
622 rec->ur_namelen - 1);
623 LASSERT(!IS_ERR(child));
624 if ((child->d_flags & DCACHE_CROSS_REF)) {
625 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
626 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
627 mds_pack_dentry2body(req2obd(req), body, child, 1);
628 } else if (child->d_inode == NULL) {
629 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
630 OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
631 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
633 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
634 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
641 static int mds_reint_create(struct mds_update_record *rec, int offset,
642 struct ptlrpc_request *req,
643 struct lustre_handle *lh)
645 struct dentry *dparent = NULL;
646 struct mds_obd *mds = mds_req2mds(req);
647 struct obd_device *obd = req->rq_export->exp_obd;
648 struct dentry *dchild = NULL;
649 struct inode *dir = NULL;
651 struct lustre_handle lockh[2] = {{0}, {0}};
653 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
655 struct dentry_params dp;
656 struct mea *mea = NULL;
660 LASSERT(offset == 1);
662 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
665 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
666 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
667 rec->ur_name, rec->ur_mode);
669 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
671 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
672 GOTO(cleanup, rc = -ESTALE);
674 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
675 lockh, &parent_mode, rec->ur_name,
676 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
677 if (IS_ERR(dparent)) {
678 rc = PTR_ERR(dparent);
679 CERROR("parent lookup error %d\n", rc);
682 cleanup_phase = 1; /* locked parent dentry */
683 dir = dparent->d_inode;
686 ldlm_lock_dump_handle(D_OTHER, lockh);
688 /* try to retrieve MEA data for this dir */
689 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
695 * dir is already splitted, check is requested filename should
696 * live at this MDS or at another one.
698 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
699 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
700 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
701 " should be %lu(%d)\n",
702 mea->mea_master, dparent->d_inode->i_ino,
703 dparent->d_inode->i_generation, rec->ur_name,
704 (unsigned long)id_group(&mea->mea_ids[i]), i);
705 GOTO(cleanup, rc = -ERESTART);
709 dchild = ll_lookup_one_len(rec->ur_name, dparent,
710 rec->ur_namelen - 1);
711 if (IS_ERR(dchild)) {
712 rc = PTR_ERR(dchild);
713 CERROR("Can't find "DLID4"/%s, error %d\n",
714 OLID4(rec->ur_id1), rec->ur_name, rc);
718 cleanup_phase = 2; /* child dentry */
720 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
722 if (type == S_IFREG || type == S_IFDIR) {
723 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
724 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
725 obd->obd_name, dparent->d_inode->i_ino,
726 dparent->d_inode->i_generation, rc, parent_mode);
728 /* dir got splitted */
729 GOTO(cleanup, rc = -ERESTART);
731 /* error happened during spitting. */
736 if (dir->i_mode & S_ISGID) {
737 if (S_ISDIR(rec->ur_mode))
738 rec->ur_mode |= S_ISGID;
742 * here inode number should be used only in the case of replaying. It is
743 * needed to check if object already created in the case of creating
746 dchild->d_fsdata = (void *)&dp;
747 dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
752 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
754 GOTO(cleanup, rc = PTR_ERR(handle));
755 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
763 * as Peter asked, mkdir() should distribute new directories
764 * over the whole cluster in order to distribute namespace
765 * processing load. first, we calculate which MDS to use to put
766 * new directory's inode in.
768 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
770 if (i == mds->mds_num) {
771 /* inode will be created locally */
772 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
774 GOTO(cleanup, rc = PTR_ERR(handle));
776 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
778 CERROR("Can't create dir \"%s\", rc = %d\n",
779 dchild->d_name.name, rc);
783 down(&dchild->d_inode->i_sem);
785 rc = mds_update_inode_sid(obd, dchild->d_inode,
786 handle, rec->ur_id2);
788 CERROR("mds_update_inode_sid() failed, inode %lu, "
789 "rc %d\n", dchild->d_inode->i_ino, rc);
793 * make sure, that fid is up-to-date.
795 mds_set_last_fid(obd, id_fid(rec->ur_id2));
797 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
800 CERROR("mds_alloc_inode_sid() failed, inode %lu, "
801 "rc %d\n", dchild->d_inode->i_ino, rc);
804 up(&dchild->d_inode->i_sem);
810 nstripes = *(u16 *)rec->ur_eadata;
812 if (rc == 0 && nstripes) {
814 * we pass LCK_EX to split routine to signal,
815 * that we have exclusive access to the
816 * directory. Simple because nobody knows it
817 * already exists -bzzz
819 rc = mds_try_to_split_dir(obd, dchild,
823 /* dir got splitted */
826 /* an error occured during
831 } else if (!DENTRY_VALID(dchild)) {
832 /* inode will be created on another MDS */
833 struct obdo *oa = NULL;
834 struct mds_body *body;
836 /* first, create that inode */
839 GOTO(cleanup, rc = -ENOMEM);
844 if (rec->ur_eadata) {
845 /* user asks for creating splitted dir */
846 oa->o_easize = *((u16 *) rec->ur_eadata);
849 obdo_from_inode(oa, dir, OBD_MD_FLATIME |
850 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
852 /* adjust the uid/gid/mode bits */
853 oa->o_mode = rec->ur_mode;
854 oa->o_uid = current->fsuid;
855 oa->o_gid = (dir->i_mode & S_ISGID) ?
856 dir->i_gid : current->fsgid;
857 oa->o_valid |= OBD_MD_FLTYPE|OBD_MD_FLUID|OBD_MD_FLGID;
859 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
862 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
864 * here inode number and generation are
865 * important, as this is replay request and we
866 * need them to check if such an object is
869 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
870 obd->obd_name, rec->ur_namelen - 1,
871 rec->ur_name, (unsigned)id_ino(rec->ur_id2),
872 (unsigned)id_gen(rec->ur_id2));
873 oa->o_id = id_ino(rec->ur_id2);
874 oa->o_fid = id_fid(rec->ur_id2);
875 oa->o_generation = id_gen(rec->ur_id2);
876 oa->o_flags |= OBD_FL_RECREATE_OBJS;
877 LASSERT(oa->o_fid != 0);
881 * before obd_create() is called, o_fid is not known if
882 * this is not recovery of cause.
884 rc = obd_create(mds->mds_md_exp, oa, NULL, NULL);
886 CERROR("can't create remote inode: %d\n", rc);
887 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
888 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
889 rec->ur_name, rec->ur_mode);
894 LASSERT(oa->o_fid != 0);
896 /* now, add new dir entry for it */
897 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
898 if (IS_ERR(handle)) {
900 GOTO(cleanup, rc = PTR_ERR(handle));
903 /* creating local dentry for remote inode. */
904 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
905 rec->ur_namelen - 1, oa->o_id,
906 oa->o_generation, i, oa->o_fid);
909 CERROR("Can't create local entry %*s for "
910 "remote inode.\n", rec->ur_namelen - 1,
916 body = lustre_msg_buf(req->rq_repmsg,
918 body->valid |= OBD_MD_FLID | OBD_MD_MDS | OBD_MD_FID;
920 obdo2id(&body->id1, oa);
923 /* requested name exists in the directory */
930 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
932 GOTO(cleanup, rc = PTR_ERR(handle));
933 if (rec->ur_tgt == NULL) /* no target supplied */
934 rc = -EINVAL; /* -EPROTO? */
936 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
944 int rdev = rec->ur_rdev;
945 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
947 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
948 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
953 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
954 dchild->d_fsdata = NULL;
955 GOTO(cleanup, rc = -EINVAL);
958 /* In case we stored the desired inum in here, we want to clean up. */
959 if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
960 dchild->d_fsdata = NULL;
963 CDEBUG(D_INODE, "error during create: %d\n", rc);
965 } else if (dchild->d_inode) {
966 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
968 struct mds_body *body;
969 struct inode *inode = dchild->d_inode;
972 iattr.ia_uid = rec->ur_fsuid;
973 LTIME_S(iattr.ia_atime) = rec->ur_time;
974 LTIME_S(iattr.ia_ctime) = rec->ur_time;
975 LTIME_S(iattr.ia_mtime) = rec->ur_time;
977 if (dir->i_mode & S_ISGID)
978 iattr.ia_gid = dir->i_gid;
980 iattr.ia_gid = rec->ur_fsgid;
982 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
983 ATTR_MTIME | ATTR_CTIME;
985 if (id_ino(rec->ur_id2)) {
986 LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
987 inode->i_generation = id_gen(rec->ur_id2);
989 if (type != S_IFDIR) {
991 rc = mds_update_inode_sid(obd, inode,
992 handle, rec->ur_id2);
995 CERROR("Can't update inode self id, "
1000 * make sure, that fid is up-to-date.
1002 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1005 /* dirtied and committed by the upcoming setattr. */
1006 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1007 inode->i_ino, inode->i_generation);
1009 struct lustre_handle child_ino_lockh;
1011 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
1012 inode->i_ino, inode->i_generation);
1014 if (type != S_IFDIR) {
1016 * allocate new id for @inode if it is not dir,
1017 * because for dir it was already done.
1019 down(&inode->i_sem);
1020 rc = mds_alloc_inode_sid(obd, inode,
1024 CERROR("mds_alloc_inode_sid() failed, "
1025 "inode %lu, rc %d\n", inode->i_ino,
1032 * the inode we were allocated may have just
1033 * been freed by an unlink operation. We take
1034 * this lock to synchronize against the matching
1035 * reply-ack-lock taken in unlink, to avoid
1036 * replay problems if this reply makes it out to
1037 * the client but the unlink's does not. See
1038 * bug 2029 for more detail.
1040 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
1041 if (rc != ELDLM_OK) {
1042 CERROR("error locking for unlink/create sync: "
1045 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
1050 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1052 CERROR("error on child setattr: rc = %d\n", rc);
1054 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1055 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1057 CERROR("error on parent setattr: rc = %d\n", rc);
1059 MD_COUNTER_INCREMENT(obd, create);
1061 /* take care of default stripe inheritance */
1062 if (type == S_IFDIR) {
1063 struct lov_mds_md lmm;
1064 int lmm_size = sizeof(lmm);
1066 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0);
1068 down(&inode->i_sem);
1069 rc = fsfilt_set_md(obd, inode, handle,
1070 &lmm, lmm_size, EA_LOV);
1074 CERROR("error on copy stripe info: rc = %d\n",
1080 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1081 mds_pack_inode2body(obd, body, inode, 1);
1082 mds_body_do_reverse_map(med, body);
1087 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
1089 if (rc && created) {
1090 /* Destroy the file we just created. This should not need extra
1091 * journal credits, as we have already modified all of the
1092 * blocks needed in order to create the file in the first
1096 err = vfs_rmdir(dir, dchild);
1098 CERROR("rmdir in error path: %d\n", err);
1101 err = vfs_unlink(dir, dchild);
1103 CERROR("unlink in error path: %d\n", err);
1109 switch (cleanup_phase) {
1110 case 2: /* child dentry */
1112 case 1: /* locked parent dentry */
1114 if (lockh[1].cookie != 0)
1115 ldlm_lock_decref(lockh + 1, parent_mode);
1118 ldlm_lock_decref(lockh, LCK_PW);
1120 ptlrpc_save_lock(req, lockh, LCK_PW);
1126 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1130 OBD_FREE(mea, mea_size);
1131 req->rq_status = rc;
1136 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1137 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1141 for (i = 0; i < RES_NAME_SIZE; i++) {
1143 * this is needed to make zeroed res_id entries to be put at the
1144 * end of list in *ordered_locks() .
1146 if (res1->name[i] == 0 && res2->name[i] != 0)
1148 if (res2->name[i] == 0 && res1->name[i] != 0)
1150 if (res1->name[i] > res2->name[i])
1152 if (res1->name[i] < res2->name[i])
1159 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1165 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1166 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1167 * because they take the place of local semaphores.
1169 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1170 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1171 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1172 struct lustre_handle *p1_lockh, int p1_lock_mode,
1173 ldlm_policy_data_t *p1_policy,
1174 struct ldlm_res_id *p2_res_id,
1175 struct lustre_handle *p2_lockh, int p2_lock_mode,
1176 ldlm_policy_data_t *p2_policy)
1178 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1179 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1180 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1181 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1185 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1187 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1188 res_id[0]->name[0], res_id[1]->name[0]);
1190 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1191 handles[1] = p1_lockh;
1192 handles[0] = p2_lockh;
1193 res_id[1] = p1_res_id;
1194 res_id[0] = p2_res_id;
1195 lock_modes[1] = p1_lock_mode;
1196 lock_modes[0] = p2_lock_mode;
1197 policies[1] = p1_policy;
1198 policies[0] = p2_policy;
1201 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1202 res_id[0]->name[0], res_id[1]->name[0]);
1204 flags = LDLM_FL_LOCAL_ONLY;
1205 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1206 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1207 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1208 NULL, 0, NULL, handles[0]);
1211 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1213 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1214 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1215 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1216 ldlm_lock_addref(handles[1], lock_modes[1]);
1217 } else if (res_id[1]->name[0] != 0) {
1218 flags = LDLM_FL_LOCAL_ONLY;
1219 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1220 *res_id[1], LDLM_IBITS, policies[1],
1221 lock_modes[1], &flags, mds_blocking_ast,
1222 ldlm_completion_ast, NULL, NULL, NULL, 0,
1224 if (rc != ELDLM_OK) {
1225 ldlm_lock_decref(handles[0], lock_modes[0]);
1228 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1234 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1235 struct lustre_handle *p1_lockh, int p1_lock_mode,
1236 ldlm_policy_data_t *p1_policy,
1237 struct ldlm_res_id *p2_res_id,
1238 struct lustre_handle *p2_lockh, int p2_lock_mode,
1239 ldlm_policy_data_t *p2_policy,
1240 struct ldlm_res_id *c1_res_id,
1241 struct lustre_handle *c1_lockh, int c1_lock_mode,
1242 ldlm_policy_data_t *c1_policy,
1243 struct ldlm_res_id *c2_res_id,
1244 struct lustre_handle *c2_lockh, int c2_lock_mode,
1245 ldlm_policy_data_t *c2_policy)
1247 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1248 c1_res_id, c2_res_id };
1249 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1250 c1_lockh, c2_lockh };
1251 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1252 c1_lock_mode, c2_lock_mode };
1253 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1254 c1_policy, c2_policy};
1255 int rc, i, j, sorted, flags;
1258 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1259 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1260 res_id[3]->name[0]);
1263 * simple insertion sort - we have at most 4 elements. Note, that zeroed
1264 * res_id should be at the end of list after sorting is finished.
1266 for (i = 1; i < 4; i++) {
1268 dlm_handles[4] = dlm_handles[i];
1269 res_id[4] = res_id[i];
1270 lock_modes[4] = lock_modes[i];
1271 policies[4] = policies[i];
1275 if (res_gt(res_id[j], res_id[4], policies[j],
1277 dlm_handles[j + 1] = dlm_handles[j];
1278 res_id[j + 1] = res_id[j];
1279 lock_modes[j + 1] = lock_modes[j];
1280 policies[j + 1] = policies[j];
1285 } while (j >= 0 && !sorted);
1287 dlm_handles[j + 1] = dlm_handles[4];
1288 res_id[j + 1] = res_id[4];
1289 lock_modes[j + 1] = lock_modes[4];
1290 policies[j + 1] = policies[4];
1293 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1294 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1295 res_id[3]->name[0]);
1297 /* XXX we could send ASTs on all these locks first before blocking? */
1298 for (i = 0; i < 4; i++) {
1302 * nevertheless zeroed res_ids should be at the end of list, and
1303 * could use break here, I think, that it is more correctly for
1304 * clear understanding of code to have continue here, as it
1305 * clearly means, that zeroed res_id should be skipped and does
1306 * not mean, that if we meet zeroed res_id we should stop
1309 if (res_id[i]->name[0] == 0)
1313 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1314 (policies[i]->l_inodebits.bits &
1315 policies[i-1]->l_inodebits.bits) ) {
1316 memcpy(dlm_handles[i], dlm_handles[i-1],
1317 sizeof(*(dlm_handles[i])));
1318 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1320 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1321 *res_id[i], LDLM_IBITS,
1323 lock_modes[i], &flags,
1325 ldlm_completion_ast, NULL, NULL,
1326 NULL, 0, NULL, dlm_handles[i]);
1328 GOTO(out_err, rc = -EIO);
1329 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1336 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1341 /* In the unlikely case that the child changed while we were waiting
1342 * on the lock, we need to drop the lock on the old child and either:
1343 * - if the child has a lower resource name, then we have to also
1344 * drop the parent lock and regain the locks in the right order
1345 * - in the rename case, if the child has a lower resource name than one of
1346 * the other parent/child resources (maxres) we also need to reget the locks
1347 * - if the child has a higher resource name (this is the common case)
1348 * we can just get the lock on the new child (still in lock order)
1350 * Returns 0 if the child did not change or if it changed but could be locked.
1351 * Returns 1 if the child changed and we need to re-lock (no locks held).
1352 * Returns -ve error with a valid dchild (no locks held). */
1353 static int mds_verify_child(struct obd_device *obd,
1354 struct ldlm_res_id *parent_res_id,
1355 struct lustre_handle *parent_lockh,
1356 struct dentry *dparent, int parent_mode,
1357 struct ldlm_res_id *child_res_id,
1358 struct lustre_handle *child_lockh,
1359 struct dentry **dchildp, int child_mode,
1360 ldlm_policy_data_t *child_policy,
1361 const char *name, int namelen,
1362 struct ldlm_res_id *maxres,
1363 unsigned long child_ino, __u32 child_gen)
1365 struct lustre_id sid;
1366 struct dentry *vchild, *dchild = *dchildp;
1367 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1370 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1372 GOTO(cleanup, rc = PTR_ERR(vchild));
1374 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1375 if (child_gen == vchild->d_generation &&
1376 child_ino == vchild->d_inum) {
1385 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1386 (vchild->d_inode != NULL &&
1387 child_gen == vchild->d_inode->i_generation &&
1388 child_ino == vchild->d_inode->i_ino))) {
1396 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1397 vchild->d_inode, dchild ? dchild->d_inode : 0,
1398 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1399 child_res_id->name[0]);
1401 if (child_res_id->name[0] != 0)
1402 ldlm_lock_decref(child_lockh, child_mode);
1406 cleanup_phase = 1; /* parent lock only */
1407 *dchildp = dchild = vchild;
1409 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1412 if (dchild->d_inode) {
1413 down(&dchild->d_inode->i_sem);
1414 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1415 up(&dchild->d_inode->i_sem);
1417 CERROR("Can't read inode self id, inode %lu,"
1418 " rc %d\n", dchild->d_inode->i_ino, rc);
1421 child_res_id->name[0] = id_fid(&sid);
1422 child_res_id->name[1] = id_group(&sid);
1424 child_res_id->name[0] = dchild->d_fid;
1425 child_res_id->name[1] = dchild->d_mdsnum;
1428 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1429 res_gt(maxres, child_res_id, NULL, NULL)) {
1430 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1431 child_res_id->name[0], parent_res_id->name[0],
1433 GOTO(cleanup, rc = 1);
1436 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1437 *child_res_id, LDLM_IBITS, child_policy,
1438 child_mode, &flags, mds_blocking_ast,
1439 ldlm_completion_ast, NULL, NULL, NULL, 0,
1442 GOTO(cleanup, rc = -EIO);
1445 memset(child_res_id, 0, sizeof(*child_res_id));
1451 switch(cleanup_phase) {
1453 if (child_res_id->name[0] != 0)
1454 ldlm_lock_decref(child_lockh, child_mode);
1456 ldlm_lock_decref(parent_lockh, parent_mode);
1462 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1463 struct lustre_id *id,
1464 struct lustre_handle *parent_lockh,
1465 struct dentry **dparentp, int parent_mode,
1466 __u64 parent_lockpart, int *update_mode,
1467 char *name, int namelen,
1468 struct lustre_handle *child_lockh,
1469 struct dentry **dchildp, int child_mode,
1470 __u64 child_lockpart)
1472 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1473 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1474 struct ldlm_res_id parent_res_id = { .name = {0} };
1475 struct ldlm_res_id child_res_id = { .name = {0} };
1476 unsigned long child_ino = 0; __u32 child_gen = 0;
1477 int rc = 0, cleanup_phase = 0;
1478 struct lustre_id sid;
1479 struct inode *inode;
1482 /* Step 1: Lookup parent */
1483 *dparentp = mds_id2dentry(obd, id, NULL);
1484 if (IS_ERR(*dparentp)) {
1485 rc = PTR_ERR(*dparentp);
1490 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1491 (*dparentp)->d_inode->i_ino, name);
1493 parent_res_id.name[0] = id_fid(id);
1494 parent_res_id.name[1] = id_group(id);
1497 parent_lockh[1].cookie = 0;
1498 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1499 struct ldlm_res_id res_id = { .name = {0} };
1500 ldlm_policy_data_t policy;
1503 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1505 res_id.name[0] = id_fid(id);
1506 res_id.name[1] = id_group(id);
1507 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1509 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1510 res_id, LDLM_IBITS, &policy,
1511 *update_mode, &flags,
1513 ldlm_completion_ast,
1514 NULL, NULL, NULL, 0, NULL,
1520 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1522 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1523 (unsigned long)id_fid(id), (unsigned long)id_group(id),
1524 parent_res_id.name[2]);
1528 cleanup_phase = 1; /* parent dentry */
1530 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1531 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1532 if (IS_ERR(*dchildp)) {
1533 rc = PTR_ERR(*dchildp);
1534 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1538 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1540 * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1541 * lock. Drop possible UPDATE lock!
1543 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1544 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1546 child_res_id.name[0] = (*dchildp)->d_fid;
1547 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1548 child_ino = (*dchildp)->d_inum;
1549 child_gen = (*dchildp)->d_generation;
1553 inode = (*dchildp)->d_inode;
1555 inode = igrab(inode);
1559 down(&inode->i_sem);
1560 rc = mds_read_inode_sid(obd, inode, &sid);
1563 CERROR("Can't read inode self id, inode %lu, "
1564 "rc %d\n", inode->i_ino, rc);
1569 child_ino = inode->i_ino;
1570 child_gen = inode->i_generation;
1571 child_res_id.name[0] = id_fid(&sid);
1572 child_res_id.name[1] = id_group(&sid);
1576 cleanup_phase = 2; /* child dentry */
1578 /* Step 3: Lock parent and child in resource order. If child doesn't
1579 * exist, we still have to lock the parent and re-lookup. */
1580 rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1581 &parent_policy, &child_res_id, child_lockh,
1582 child_mode, &child_policy);
1586 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1587 cleanup_phase = 4; /* child lock */
1589 cleanup_phase = 3; /* parent lock */
1591 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1592 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1593 parent_mode, &child_res_id, child_lockh,
1594 dchildp, child_mode, &child_policy,
1595 name, namelen, &parent_res_id, child_ino,
1607 switch (cleanup_phase) {
1609 ldlm_lock_decref(child_lockh, child_mode);
1611 ldlm_lock_decref(parent_lockh, parent_mode);
1616 if (parent_lockh[1].cookie)
1617 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1625 void mds_reconstruct_generic(struct ptlrpc_request *req)
1627 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1628 mds_req_from_mcd(req, med->med_mcd);
1631 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1632 * we instead link the inode into the PENDING directory until it is
1633 * finally released. We can't simply call mds_reint_rename() or some
1634 * part thereof, because we don't have the inode to check for link
1635 * count/open status until after it is locked.
1637 * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1638 * before starting journal transaction.
1640 * returns 1 on success
1641 * returns 0 if we lost a race and didn't make a new link
1642 * returns negative on error
1644 static int mds_orphan_add_link(struct mds_update_record *rec,
1645 struct obd_device *obd, struct dentry *dentry)
1647 struct mds_obd *mds = &obd->u.mds;
1648 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1649 struct inode *inode = dentry->d_inode;
1650 struct dentry *pending_child;
1651 char idname[LL_ID_NAMELEN];
1652 int idlen = 0, rc, mode;
1655 LASSERT(inode != NULL);
1656 LASSERT(!mds_inode_is_orphan(inode));
1657 #ifndef HAVE_I_ALLOC_SEM
1658 LASSERT(down_trylock(&inode->i_sem) != 0);
1660 LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1662 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1664 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1665 mds_orphan_open_count(inode), inode->i_nlink,
1666 S_ISDIR(inode->i_mode) ? "dir" :
1667 S_ISREG(inode->i_mode) ? "file" : "other",
1668 rec->ur_name, idname);
1670 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1673 pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1674 if (IS_ERR(pending_child))
1675 RETURN(PTR_ERR(pending_child));
1677 if (pending_child->d_inode != NULL) {
1678 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1679 LASSERT(pending_child->d_inode == inode);
1680 GOTO(out_dput, rc = 0);
1683 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1684 * for linking and return real mode back then -bzzz */
1685 mode = inode->i_mode;
1686 inode->i_mode = S_IFREG;
1687 rc = vfs_link(dentry, pending_dir, pending_child);
1689 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1692 mds_inode_set_orphan(inode);
1694 /* return mode and correct i_nlink if inode is directory */
1695 inode->i_mode = mode;
1696 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1697 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1699 if (S_ISDIR(mode)) {
1701 pending_dir->i_nlink++;
1702 mark_inode_dirty(inode);
1703 mark_inode_dirty(pending_dir);
1708 l_dput(pending_child);
1712 int mds_create_local_dentry(struct mds_update_record *rec,
1713 struct obd_device *obd)
1715 struct mds_obd *mds = &obd->u.mds;
1716 struct inode *id_dir = mds->mds_id_dir->d_inode;
1717 int idlen = 0, rc, cleanup_phase = 0;
1718 struct dentry *new_child = NULL;
1719 char *idname = rec->ur_name;
1720 struct dentry *child = NULL;
1721 struct lustre_handle lockh[2] = {{0}, {0}};
1722 struct lustre_id sid;
1726 down(&id_dir->i_sem);
1727 idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1728 id_gen(rec->ur_id1));
1730 CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1731 idname, OLID4(rec->ur_id1));
1733 new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1736 if (IS_ERR(new_child)) {
1737 CERROR("can't lookup %s: %d\n", idname,
1738 (int) PTR_ERR(new_child));
1739 GOTO(cleanup, rc = PTR_ERR(new_child));
1743 down(&id_dir->i_sem);
1744 rc = mds_read_inode_sid(obd, id_dir, &sid);
1747 CERROR("Can't read inode self id, inode %lu, "
1748 "rc %d\n", id_dir->i_ino, rc);
1752 if (new_child->d_inode != NULL) {
1753 /* nice. we've already have local dentry! */
1754 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1755 (unsigned)new_child->d_inode->i_ino,
1756 (unsigned)new_child->d_inode->i_generation);
1758 id_ino(rec->ur_id1) = id_dir->i_ino;
1759 id_gen(rec->ur_id1) = id_dir->i_generation;
1760 rec->ur_namelen = idlen + 1;
1762 id_fid(rec->ur_id1) = id_fid(&sid);
1763 id_group(rec->ur_id1) = id_group(&sid);
1765 GOTO(cleanup, rc = 0);
1768 /* new, local dentry will be added soon. we need no aliases here */
1771 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1772 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1774 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1775 LCK_EX, lockh, NULL, NULL, 0,
1776 MDS_INODELOCK_UPDATE);
1779 if (IS_ERR(child)) {
1780 rc = PTR_ERR(child);
1781 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1782 CERROR("can't get victim: %d\n", rc);
1787 handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1789 GOTO(cleanup, rc = PTR_ERR(handle));
1791 rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
1792 idlen, id_ino(rec->ur_id1),
1793 id_gen(rec->ur_id1), mds->mds_num,
1794 id_fid(rec->ur_id1));
1796 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1797 (unsigned long)child->d_inode->i_ino,
1798 (unsigned long)child->d_inode->i_generation, rc);
1800 if (S_ISDIR(child->d_inode->i_mode)) {
1802 mark_inode_dirty(id_dir);
1804 mark_inode_dirty(child->d_inode);
1806 fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
1808 id_ino(rec->ur_id1) = id_dir->i_ino;
1809 id_gen(rec->ur_id1) = id_dir->i_generation;
1810 rec->ur_namelen = idlen + 1;
1812 id_fid(rec->ur_id1) = id_fid(&sid);
1813 id_group(rec->ur_id1) = id_group(&sid);
1817 switch(cleanup_phase) {
1819 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1820 ldlm_lock_decref(lockh, LCK_EX);
1830 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1831 struct ptlrpc_request *slave)
1833 void *cookie, *cookie2;
1834 struct mds_body *body2;
1835 struct mds_body *body;
1839 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1840 LASSERT(body != NULL);
1842 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1843 LASSERT(body2 != NULL);
1845 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
1848 memcpy(body2, body, sizeof(*body));
1849 body2->valid &= ~OBD_MD_FLCOOKIE;
1851 if (!(body->valid & OBD_MD_FLEASIZE) &&
1852 !(body->valid & OBD_MD_FLDIREA))
1855 if (body->eadatasize == 0) {
1856 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1860 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1862 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1863 LASSERT(ea != NULL);
1865 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1866 LASSERT(ea2 != NULL);
1868 memcpy(ea2, ea, body->eadatasize);
1870 if (body->valid & OBD_MD_FLCOOKIE) {
1871 LASSERT(master->rq_repmsg->buflens[2] >=
1872 slave->rq_repmsg->buflens[2]);
1873 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1874 slave->rq_repmsg->buflens[2]);
1875 LASSERT(cookie != NULL);
1877 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1878 master->rq_repmsg->buflens[2]);
1879 LASSERT(cookie2 != NULL);
1880 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1881 body2->valid |= OBD_MD_FLCOOKIE;
1886 static int mds_reint_unlink_remote(struct mds_update_record *rec,
1887 int offset, struct ptlrpc_request *req,
1888 struct lustre_handle *parent_lockh,
1889 int update_mode, struct dentry *dparent,
1890 struct lustre_handle *child_lockh,
1891 struct dentry *dchild)
1893 struct obd_device *obd = req->rq_export->exp_obd;
1894 struct mds_obd *mds = mds_req2mds(req);
1895 struct ptlrpc_request *request = NULL;
1896 int rc = 0, cleanup_phase = 0;
1897 struct mdc_op_data *op_data;
1901 LASSERT(offset == 1 || offset == 3);
1903 /* time to drop i_nlink on remote MDS */
1904 OBD_ALLOC(op_data, sizeof(*op_data));
1905 if (op_data == NULL)
1908 memset(op_data, 0, sizeof(*op_data));
1909 mds_pack_dentry2id(obd, &op_data->id1, dchild, 1);
1910 op_data->create_mode = rec->ur_mode;
1912 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
1913 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
1915 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1916 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
1917 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
1920 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1921 op_data->create_mode |= MDS_MODE_REPLAY;
1923 rc = md_unlink(mds->mds_md_exp, op_data, &request);
1924 OBD_FREE(op_data, sizeof(*op_data));
1929 mds_copy_unlink_reply(req, request);
1930 ptlrpc_req_finished(request);
1934 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1937 GOTO(cleanup, rc = PTR_ERR(handle));
1938 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1939 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1944 req->rq_status = rc;
1947 if (parent_lockh[1].cookie != 0)
1948 ldlm_lock_decref(parent_lockh + 1, update_mode);
1950 ldlm_lock_decref(child_lockh, LCK_EX);
1952 ldlm_lock_decref(parent_lockh, LCK_PW);
1954 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1961 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1962 struct ptlrpc_request *req,
1963 struct lustre_handle *lh)
1965 struct dentry *dparent = NULL, *dchild;
1966 struct mds_obd *mds = mds_req2mds(req);
1967 struct obd_device *obd = req->rq_export->exp_obd;
1968 struct mds_body *body = NULL;
1969 struct inode *child_inode = NULL;
1970 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1971 struct lustre_handle child_lockh = {0};
1972 struct lustre_handle child_reuse_lockh = {0};
1973 struct lustre_handle *slave_lockh = NULL;
1974 char idname[LL_ID_NAMELEN];
1975 struct llog_create_locks *lcl = NULL;
1976 void *handle = NULL;
1977 int rc = 0, cleanup_phase = 0;
1978 int unlink_by_id = 0;
1982 LASSERT(offset == 1 || offset == 3);
1984 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1985 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1988 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1990 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1991 DEBUG_REQ(D_HA, req, "unlink replay");
1992 LASSERT(offset == 1); /* should not come from intent */
1993 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1994 lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
1995 req->rq_repmsg->buflens[2]);
1998 MD_COUNTER_INCREMENT(obd, unlink);
2000 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
2001 GOTO(cleanup, rc = -ENOENT);
2003 if (rec->ur_namelen == 1) {
2004 /* this is request to drop i_nlink on local inode */
2006 rec->ur_name = idname;
2007 rc = mds_create_local_dentry(rec, obd);
2008 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
2009 DEBUG_REQ(D_HA, req,
2010 "drop nlink on inode "DLID4" (replay)",
2011 OLID4(rec->ur_id1));
2017 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
2018 /* master mds for directory asks slave removing inode is already
2020 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
2021 LCK_PW, parent_lockh,
2022 &update_mode, rec->ur_name,
2024 MDS_INODELOCK_UPDATE);
2025 if (IS_ERR(dparent))
2026 GOTO(cleanup, rc = PTR_ERR(dparent));
2027 dchild = ll_lookup_one_len(rec->ur_name, dparent,
2028 rec->ur_namelen - 1);
2030 GOTO(cleanup, rc = PTR_ERR(dchild));
2031 child_lockh.cookie = 0;
2032 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
2033 LASSERT(dchild->d_inode != NULL);
2034 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
2036 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
2037 parent_lockh, &dparent,
2038 LCK_PW, MDS_INODELOCK_UPDATE,
2039 &update_mode, rec->ur_name,
2040 rec->ur_namelen, &child_lockh,
2042 MDS_INODELOCK_LOOKUP |
2043 MDS_INODELOCK_UPDATE);
2048 if (dchild->d_flags & DCACHE_CROSS_REF) {
2049 /* we should have parent lock only here */
2050 LASSERT(unlink_by_id == 0);
2051 LASSERT(dchild->d_mdsnum != mds->mds_num);
2052 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
2053 update_mode, dparent, &child_lockh, dchild);
2057 cleanup_phase = 1; /* dchild, dparent, locks */
2060 child_inode = dchild->d_inode;
2061 if (child_inode == NULL) {
2062 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
2063 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
2064 GOTO(cleanup, rc = -ENOENT);
2067 cleanup_phase = 2; /* dchild has a lock */
2069 /* We have to do these checks ourselves, in case we are making an
2070 * orphan. The client tells us whether rmdir() or unlink() was called,
2071 * so we need to return appropriate errors (bug 72).
2073 * We don't have to check permissions, because vfs_rename (called from
2074 * mds_open_unlink_rename) also calls may_delete. */
2075 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
2076 if (!S_ISDIR(child_inode->i_mode))
2077 GOTO(cleanup, rc = -ENOTDIR);
2079 if (S_ISDIR(child_inode->i_mode))
2080 GOTO(cleanup, rc = -EISDIR);
2083 /* handle splitted dir */
2084 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
2088 /* Step 4: Get a lock on the ino to sync with creation WRT inode
2089 * reuse (see bug 2029). */
2090 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
2093 cleanup_phase = 3; /* child inum lock */
2095 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2097 /* ldlm_reply in buf[0] if called via intent */
2103 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2104 LASSERT(body != NULL);
2106 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2107 DOWN_READ_I_ALLOC_SEM(child_inode);
2108 cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2110 /* If this is potentially the last reference to this inode, get the
2111 * OBD EA data first so the client can destroy OST objects. We
2112 * only do the object removal later if no open files/links remain. */
2113 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2114 child_inode->i_nlink == 1) {
2115 if (mds_orphan_open_count(child_inode) > 0) {
2116 /* need to lock pending_dir before transaction */
2117 down(&mds->mds_pending_dir->d_inode->i_sem);
2118 cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2119 } else if (S_ISREG(child_inode->i_mode)) {
2120 mds_pack_inode2body(obd, body, child_inode, 0);
2121 mds_pack_md(obd, req->rq_repmsg, offset + 1,
2122 body, child_inode, MDS_PACK_MD_LOCK, 0);
2126 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2127 switch (child_inode->i_mode & S_IFMT) {
2129 /* Drop any lingering child directories before we start our
2130 * transaction, to avoid doing multiple inode dirty/delete
2131 * in our compound transaction (bug 1321). */
2132 shrink_dcache_parent(dchild);
2133 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2136 GOTO(cleanup, rc = PTR_ERR(handle));
2137 rc = vfs_rmdir(dparent->d_inode, dchild);
2140 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2141 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2143 handle = fsfilt_start_log(obd, dparent->d_inode,
2144 FSFILT_OP_UNLINK, NULL,
2145 le32_to_cpu(lmm->lmm_stripe_count));
2147 GOTO(cleanup, rc = PTR_ERR(handle));
2148 rc = vfs_unlink(dparent->d_inode, dchild);
2156 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2159 GOTO(cleanup, rc = PTR_ERR(handle));
2160 rc = vfs_unlink(dparent->d_inode, dchild);
2163 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2166 GOTO(cleanup, rc = -EINVAL);
2169 if (rc == 0 && child_inode->i_nlink == 0) {
2170 if (mds_orphan_open_count(child_inode) > 0)
2171 rc = mds_orphan_add_link(rec, obd, dchild);
2174 GOTO(cleanup, rc = 0);
2176 if (!S_ISREG(child_inode->i_mode))
2179 if (!(body->valid & OBD_MD_FLEASIZE)) {
2180 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2181 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2182 } else if (mds_log_op_unlink(obd, child_inode,
2183 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2184 req->rq_repmsg->buflens[offset + 1],
2185 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2186 req->rq_repmsg->buflens[offset+2],
2188 body->valid |= OBD_MD_FLCOOKIE;
2199 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2200 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2201 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2203 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2205 CERROR("error on parent setattr: rc = %d\n", err);
2207 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2208 handle, req, rc, 0);
2210 (void)obd_set_info(mds->mds_dt_exp, strlen("unlinked"),
2211 "unlinked", 0, NULL);
2212 switch(cleanup_phase) {
2213 case 5: /* pending_dir semaphore */
2214 up(&mds->mds_pending_dir->d_inode->i_sem);
2215 case 4: /* child inode semaphore */
2216 UP_READ_I_ALLOC_SEM(child_inode);
2217 /* handle splitted dir */
2219 /* master directory can be non-empty or something else ... */
2220 mds_unlink_slave_objs(obd, dchild);
2223 ptlrpc_save_llog_lock(req, lcl);
2224 case 3: /* child ino-reuse lock */
2225 if (rc && body != NULL) {
2226 // Don't unlink the OST objects if the MDS unlink failed
2230 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2232 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2233 case 2: /* child lock */
2234 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2235 if (child_lockh.cookie)
2236 ldlm_lock_decref(&child_lockh, LCK_EX);
2237 case 1: /* child and parent dentry, parent lock */
2239 if (parent_lockh[1].cookie != 0)
2240 ldlm_lock_decref(parent_lockh + 1, update_mode);
2243 ldlm_lock_decref(parent_lockh, LCK_PW);
2245 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2252 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2255 req->rq_status = rc;
2260 * to service requests from remote MDS to increment i_nlink
2262 static int mds_reint_link_acquire(struct mds_update_record *rec,
2263 int offset, struct ptlrpc_request *req,
2264 struct lustre_handle *lh)
2266 struct obd_device *obd = req->rq_export->exp_obd;
2267 struct ldlm_res_id src_res_id = { .name = {0} };
2268 struct lustre_handle *handle = NULL, src_lockh = {0};
2269 struct mds_obd *mds = mds_req2mds(req);
2270 int rc = 0, cleanup_phase = 0;
2271 struct dentry *de_src = NULL;
2272 ldlm_policy_data_t policy;
2276 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2277 obd->obd_name, OLID4(rec->ur_id1));
2279 /* Step 1: Lookup the source inode and target directory by ID */
2280 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2282 GOTO(cleanup, rc = PTR_ERR(de_src));
2283 cleanup_phase = 1; /* source dentry */
2285 src_res_id.name[0] = id_fid(rec->ur_id1);
2286 src_res_id.name[1] = id_group(rec->ur_id1);
2287 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2289 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2290 src_res_id, LDLM_IBITS, &policy,
2291 LCK_EX, &flags, mds_blocking_ast,
2292 ldlm_completion_ast, NULL, NULL,
2293 NULL, 0, NULL, &src_lockh);
2295 GOTO(cleanup, rc = -ENOLCK);
2296 cleanup_phase = 2; /* lock */
2298 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2300 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2301 if (IS_ERR(handle)) {
2302 rc = PTR_ERR(handle);
2305 de_src->d_inode->i_nlink++;
2306 mark_inode_dirty(de_src->d_inode);
2310 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2311 handle, req, rc, 0);
2312 switch (cleanup_phase) {
2315 ldlm_lock_decref(&src_lockh, LCK_EX);
2317 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2323 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2326 req->rq_status = rc;
2331 * request to link to foreign inode:
2332 * - acquire i_nlinks on this inode
2335 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2336 int offset, struct ptlrpc_request *req,
2337 struct lustre_handle *lh)
2339 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2340 struct obd_device *obd = req->rq_export->exp_obd;
2341 struct dentry *de_tgt_dir = NULL;
2342 struct mds_obd *mds = mds_req2mds(req);
2343 int rc = 0, cleanup_phase = 0;
2344 struct mdc_op_data *op_data;
2345 struct ptlrpc_request *request = NULL;
2349 DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2350 ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2351 OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2352 OLID4(rec->ur_id1));
2354 de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2355 tgt_dir_lockh, &update_mode,
2356 rec->ur_name, rec->ur_namelen - 1,
2357 MDS_INODELOCK_UPDATE);
2358 if (IS_ERR(de_tgt_dir))
2359 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2362 OBD_ALLOC(op_data, sizeof(*op_data));
2363 if (op_data == NULL)
2364 GOTO(cleanup, rc = -ENOMEM);
2366 memset(op_data, 0, sizeof(*op_data));
2367 op_data->id1 = *(rec->ur_id1);
2368 rc = md_link(mds->mds_md_exp, op_data, &request);
2369 OBD_FREE(op_data, sizeof(*op_data));
2372 ptlrpc_req_finished(request);
2378 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2380 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2381 if (IS_ERR(handle)) {
2382 rc = PTR_ERR(handle);
2388 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2389 rec->ur_namelen - 1, id_ino(rec->ur_id1),
2390 id_gen(rec->ur_id1), id_group(rec->ur_id1),
2391 id_fid(rec->ur_id1));
2394 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2395 handle, req, rc, 0);
2397 switch (cleanup_phase) {
2400 OBD_ALLOC(op_data, sizeof(*op_data));
2401 if (op_data != NULL) {
2403 memset(op_data, 0, sizeof(*op_data));
2405 op_data->id1 = *(rec->ur_id1);
2406 op_data->create_mode = rec->ur_mode;
2408 rc = md_unlink(mds->mds_md_exp, op_data, &request);
2409 OBD_FREE(op_data, sizeof(*op_data));
2411 ptlrpc_req_finished(request);
2413 CERROR("error %d while dropping i_nlink on "
2414 "remote inode\n", rc);
2417 CERROR("rc %d prevented dropping i_nlink on "
2418 "remote inode\n", -ENOMEM);
2424 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2426 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2429 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2431 ptlrpc_save_lock(req, tgt_dir_lockh + 1, update_mode);
2437 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2440 req->rq_status = rc;
2444 static int mds_reint_link(struct mds_update_record *rec, int offset,
2445 struct ptlrpc_request *req,
2446 struct lustre_handle *lh)
2448 struct obd_device *obd = req->rq_export->exp_obd;
2449 struct dentry *de_src = NULL;
2450 struct dentry *de_tgt_dir = NULL;
2451 struct dentry *dchild = NULL;
2452 struct mds_obd *mds = mds_req2mds(req);
2453 struct lustre_handle *handle = NULL;
2454 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2455 struct ldlm_res_id src_res_id = { .name = {0} };
2456 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2457 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2458 ldlm_policy_data_t tgt_dir_policy =
2459 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2460 int rc = 0, cleanup_phase = 0;
2462 int update_mode = 0;
2466 LASSERT(offset == 1);
2468 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2469 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2470 id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2473 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2474 MD_COUNTER_INCREMENT(obd, link);
2476 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2477 GOTO(cleanup, rc = -ENOENT);
2479 if (id_group(rec->ur_id1) != mds->mds_num) {
2480 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2484 if (rec->ur_namelen == 1) {
2485 rc = mds_reint_link_acquire(rec, offset, req, lh);
2489 /* Step 1: Lookup the source inode and target directory by ID */
2490 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2492 GOTO(cleanup, rc = PTR_ERR(de_src));
2494 cleanup_phase = 1; /* source dentry */
2496 de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2497 if (IS_ERR(de_tgt_dir)) {
2498 rc = PTR_ERR(de_tgt_dir);
2503 cleanup_phase = 2; /* target directory dentry */
2505 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2506 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2507 rec->ur_name, de_src->d_inode->i_ino);
2509 /* Step 2: Take the two locks */
2510 src_res_id.name[0] = id_fid(rec->ur_id1);
2511 src_res_id.name[1] = id_group(rec->ur_id1);
2512 tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2513 tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2516 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2518 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2520 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2521 tgt_dir_res_id, LDLM_IBITS,
2522 &src_policy, update_mode, &flags,
2524 ldlm_completion_ast, NULL, NULL,
2525 NULL, 0, NULL, tgt_dir_lockh + 1);
2527 GOTO(cleanup, rc = -ENOLCK);
2530 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2531 rec->ur_namelen - 1);
2532 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2533 (unsigned long)id_fid(rec->ur_id2),
2534 (unsigned long)id_group(rec->ur_id2),
2535 tgt_dir_res_id.name[2]);
2538 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2539 &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2540 LCK_EX, &tgt_dir_policy);
2544 cleanup_phase = 3; /* locks */
2546 /* Step 3: Lookup the child */
2547 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2548 rec->ur_namelen - 1);
2549 if (IS_ERR(dchild)) {
2550 rc = PTR_ERR(dchild);
2551 if (rc != -EPERM && rc != -EACCES)
2552 CERROR("child lookup error %d\n", rc);
2556 cleanup_phase = 4; /* child dentry */
2558 if (dchild->d_inode) {
2559 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2560 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2565 /* Step 4: Do it. */
2566 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2568 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2569 if (IS_ERR(handle)) {
2570 rc = PTR_ERR(handle);
2574 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2575 if (rc && rc != -EPERM && rc != -EACCES)
2576 CERROR("vfs_link error %d\n", rc);
2578 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2579 handle, req, rc, 0);
2582 switch (cleanup_phase) {
2583 case 4: /* child dentry */
2587 ldlm_lock_decref(&src_lockh, LCK_EX);
2588 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2590 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2591 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2593 case 2: /* target dentry */
2595 if (tgt_dir_lockh[1].cookie && update_mode)
2596 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2600 case 1: /* source dentry */
2605 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2608 req->rq_status = rc;
2612 /* The idea here is that we need to get four locks in the end:
2613 * one on each parent directory, one on each child. We need to take
2614 * these locks in some kind of order (to avoid deadlocks), and the order
2615 * I selected is "increasing resource number" order. We need to look up
2616 * the children, however, before we know what the resource number(s) are.
2617 * Thus the following plan:
2619 * 1,2. Look up the parents
2620 * 3,4. Look up the children
2621 * 5. Take locks on the parents and children, in order
2622 * 6. Verify that the children haven't changed since they were looked up
2624 * If there was a race and the children changed since they were first looked
2625 * up, it is possible that mds_verify_child() will be able to just grab the
2626 * lock on the new child resource (if it has a higher resource than any other)
2627 * but we need to compare against not only its parent, but also against the
2628 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2630 * We need the fancy igrab() on the child inodes because we aren't holding a
2631 * lock on the parent after the lookup is done, so dentry->d_inode may change
2632 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2634 static int mds_get_parents_children_locked(struct obd_device *obd,
2635 struct mds_obd *mds,
2636 struct lustre_id *p1_id,
2637 struct dentry **de_srcdirp,
2638 struct lustre_id *p2_id,
2639 struct dentry **de_tgtdirp,
2641 const char *old_name, int old_len,
2642 struct dentry **de_oldp,
2643 const char *new_name, int new_len,
2644 struct dentry **de_newp,
2645 struct lustre_handle *dlm_handles,
2648 struct ldlm_res_id p1_res_id = { .name = {0} };
2649 struct ldlm_res_id p2_res_id = { .name = {0} };
2650 struct ldlm_res_id c1_res_id = { .name = {0} };
2651 struct ldlm_res_id c2_res_id = { .name = {0} };
2652 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2653 /* Only dentry should disappear, but the inode itself would be
2654 intact otherwise. */
2655 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2656 /* If something is going to be replaced, both dentry and inode locks are
2658 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2659 struct ldlm_res_id *maxres_src, *maxres_tgt;
2660 struct inode *inode;
2661 int rc = 0, cleanup_phase = 0;
2662 __u32 child_gen1 = 0;
2663 __u32 child_gen2 = 0;
2664 unsigned long child_ino1 = 0;
2665 unsigned long child_ino2 = 0;
2668 /* Step 1: Lookup the source directory */
2669 *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2670 if (IS_ERR(*de_srcdirp))
2671 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2673 cleanup_phase = 1; /* source directory dentry */
2675 p1_res_id.name[0] = id_fid(p1_id);
2676 p1_res_id.name[1] = id_group(p1_id);
2678 /* Step 2: Lookup the target directory */
2679 if (id_equal_stc(p1_id, p2_id)) {
2680 *de_tgtdirp = dget(*de_srcdirp);
2682 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2683 if (IS_ERR(*de_tgtdirp)) {
2684 rc = PTR_ERR(*de_tgtdirp);
2690 cleanup_phase = 2; /* target directory dentry */
2692 p2_res_id.name[0] = id_fid(p2_id);
2693 p2_res_id.name[1] = id_group(p2_id);
2696 dlm_handles[5].cookie = 0;
2697 dlm_handles[6].cookie = 0;
2699 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2701 * get a temp lock on just fid, group to flush client cache and
2702 * to protect dirs from concurrent splitting.
2704 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2705 LCK_PW, &p_policy, &p2_res_id,
2706 &dlm_handles[6], LCK_PW, &p_policy);
2710 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2711 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2713 CDEBUG(D_INFO, "take locks on "
2714 LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2715 p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2716 p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2721 /* Step 3: Lookup the source child entry */
2722 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2724 if (IS_ERR(*de_oldp)) {
2725 rc = PTR_ERR(*de_oldp);
2726 CERROR("old child lookup error (%.*s): %d\n",
2727 old_len - 1, old_name, rc);
2731 cleanup_phase = 4; /* original name dentry */
2733 inode = (*de_oldp)->d_inode;
2734 if (inode != NULL) {
2735 struct lustre_id sid;
2737 inode = igrab(inode);
2739 GOTO(cleanup, rc = -ENOENT);
2741 down(&inode->i_sem);
2742 rc = mds_read_inode_sid(obd, inode, &sid);
2745 CERROR("Can't read inode self id, inode %lu, "
2746 "rc %d\n", inode->i_ino, rc);
2751 child_ino1 = inode->i_ino;
2752 child_gen1 = inode->i_generation;
2753 c1_res_id.name[0] = id_fid(&sid);
2754 c1_res_id.name[1] = id_group(&sid);
2756 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2757 child_ino1 = (*de_oldp)->d_inum;
2758 child_gen1 = (*de_oldp)->d_generation;
2759 c1_res_id.name[0] = (*de_oldp)->d_fid;
2760 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2762 GOTO(cleanup, rc = -ENOENT);
2765 /* Step 4: Lookup the target child entry */
2766 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2768 if (IS_ERR(*de_newp)) {
2769 rc = PTR_ERR(*de_newp);
2770 CERROR("new child lookup error (%.*s): %d\n",
2771 old_len - 1, old_name, rc);
2775 cleanup_phase = 5; /* target dentry */
2777 inode = (*de_newp)->d_inode;
2778 if (inode != NULL) {
2779 struct lustre_id sid;
2781 inode = igrab(inode);
2785 down(&inode->i_sem);
2786 rc = mds_read_inode_sid(obd, inode, &sid);
2789 CERROR("Can't read inode self id, inode %lu, "
2790 "rc %d\n", inode->i_ino, rc);
2794 child_ino2 = inode->i_ino;
2795 child_gen2 = inode->i_generation;
2796 c2_res_id.name[0] = id_fid(&sid);
2797 c2_res_id.name[1] = id_group(&sid);
2799 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2800 child_ino2 = (*de_newp)->d_inum;
2801 child_gen2 = (*de_newp)->d_generation;
2802 c2_res_id.name[0] = (*de_newp)->d_fid;
2803 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
2807 /* Step 5: Take locks on the parents and child(ren) */
2808 maxres_src = &p1_res_id;
2809 maxres_tgt = &p2_res_id;
2810 cleanup_phase = 5; /* target dentry */
2812 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
2813 maxres_src = &c1_res_id;
2814 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
2815 maxres_tgt = &c2_res_id;
2817 rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
2819 &p2_res_id, &dlm_handles[1], parent_mode,
2821 &c1_res_id, &dlm_handles[2], child_mode,
2823 &c2_res_id, &dlm_handles[3], child_mode,
2828 cleanup_phase = 6; /* parent and child(ren) locks */
2830 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2831 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2832 parent_mode, &c1_res_id, &dlm_handles[2],
2833 de_oldp, child_mode, &c1_policy, old_name, old_len,
2834 maxres_tgt, child_ino1, child_gen1);
2836 if (c2_res_id.name[0] != 0)
2837 ldlm_lock_decref(&dlm_handles[3], child_mode);
2838 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2845 if (!DENTRY_VALID(*de_oldp))
2846 GOTO(cleanup, rc = -ENOENT);
2848 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2849 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2850 parent_mode, &c2_res_id, &dlm_handles[3],
2851 de_newp, child_mode, &c2_policy, new_name,
2852 new_len, maxres_src, child_ino2, child_gen2);
2854 ldlm_lock_decref(&dlm_handles[2], child_mode);
2855 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2865 switch (cleanup_phase) {
2866 case 6: /* child lock(s) */
2867 if (c2_res_id.name[0] != 0)
2868 ldlm_lock_decref(&dlm_handles[3], child_mode);
2869 if (c1_res_id.name[0] != 0)
2870 ldlm_lock_decref(&dlm_handles[2], child_mode);
2871 if (dlm_handles[1].cookie != 0)
2872 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2873 if (dlm_handles[0].cookie != 0)
2874 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2875 case 5: /* target dentry */
2877 case 4: /* source dentry */
2881 if (dlm_handles[5].cookie != 0)
2882 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2883 if (dlm_handles[6].cookie != 0)
2884 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2886 case 2: /* target directory dentry */
2887 l_dput(*de_tgtdirp);
2888 case 1: /* source directry dentry */
2889 l_dput(*de_srcdirp);
2897 * checks if dentry can be removed. This function also handles cross-ref
2900 static int mds_check_for_rename(struct obd_device *obd,
2901 struct dentry *dentry)
2903 struct mds_obd *mds = &obd->u.mds;
2904 struct lustre_handle *rlockh;
2905 struct ptlrpc_request *req;
2906 struct mdc_op_data *op_data;
2907 struct lookup_intent it;
2908 int handle_size, rc = 0;
2911 LASSERT(dentry != NULL);
2913 if (dentry->d_inode) {
2914 if (S_ISDIR(dentry->d_inode->i_mode) &&
2915 !mds_is_dir_empty(obd, dentry))
2918 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
2919 handle_size = sizeof(struct lustre_handle);
2921 OBD_ALLOC(rlockh, handle_size);
2925 memset(rlockh, 0, handle_size);
2926 OBD_ALLOC(op_data, sizeof(*op_data));
2927 if (op_data == NULL) {
2928 OBD_FREE(rlockh, handle_size);
2931 memset(op_data, 0, sizeof(*op_data));
2932 mds_pack_dentry2id(obd, &op_data->id1, dentry, 1);
2934 it.it_op = IT_UNLINK;
2935 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
2938 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
2939 op_data, rlockh, NULL, 0, ldlm_completion_ast,
2940 mds_blocking_ast, NULL);
2941 OBD_FREE(op_data, sizeof(*op_data));
2945 OBD_FREE(it.d.fs_data,
2946 sizeof(struct lustre_intent_data));
2949 if (rlockh->cookie != 0)
2950 ldlm_lock_decref(rlockh, LCK_EX);
2952 if (LUSTRE_IT(&it)->it_data) {
2953 req = (struct ptlrpc_request *)LUSTRE_IT(&it)->it_data;
2954 ptlrpc_req_finished(req);
2957 if (LUSTRE_IT(&it)->it_status)
2958 rc = LUSTRE_IT(&it)->it_status;
2959 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
2960 OBD_FREE(rlockh, handle_size);
2965 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
2966 struct ptlrpc_request *req, struct lustre_id *id,
2967 struct dentry *de_dir, struct dentry *de)
2969 struct obd_device *obd = req->rq_export->exp_obd;
2970 struct mds_obd *mds = mds_req2mds(req);
2971 void *handle = NULL;
2977 * name exists and points to local inode try to unlink this name
2978 * and create new one.
2980 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
2981 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
2982 (unsigned long)de->d_inode->i_generation);
2984 /* checking if we can remove local dentry. */
2985 rc = mds_check_for_rename(obd, de);
2989 handle = fsfilt_start(obd, de_dir->d_inode,
2990 FSFILT_OP_RENAME, NULL);
2992 GOTO(cleanup, rc = PTR_ERR(handle));
2993 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2996 } else if (de->d_flags & DCACHE_CROSS_REF) {
2997 CDEBUG(D_OTHER, "%s: %s points to remote inode %lu/%lu\n",
2998 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_mdsnum,
2999 (unsigned long)de->d_fid);
3001 /* checking if we can remove local dentry. */
3002 rc = mds_check_for_rename(obd, de);
3007 * to be fully POSIX compatible, we should add one more check:
3009 * if de_new is subdir of dir rec->ur_id1. If so - return
3012 * I do not know how to implement it right now, because
3013 * inodes/dentries for new and old names lie on different MDS,
3014 * so add this notice here just to make it visible for the rest
3015 * of developers and do not forget about. And when this check
3016 * will be added, del_cross_ref should gone, that is local
3017 * dentry is able to be removed if all checks passed. --umka
3020 handle = fsfilt_start(obd, de_dir->d_inode,
3021 FSFILT_OP_RENAME, NULL);
3023 GOTO(cleanup, rc = PTR_ERR(handle));
3024 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3028 /* name doesn't exist. the simplest case. */
3029 handle = fsfilt_start(obd, de_dir->d_inode,
3030 FSFILT_OP_LINK, NULL);
3032 GOTO(cleanup, rc = PTR_ERR(handle));
3035 rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
3036 rec->ur_tgtlen - 1, id_ino(id),
3037 id_gen(id), id_group(id), id_fid(id));
3039 CERROR("add_dir_entry() returned error %d\n", rc);
3045 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3046 handle, req, rc, 0);
3051 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
3052 struct ptlrpc_request *req, struct dentry *de_dir,
3055 struct obd_device *obd = req->rq_export->exp_obd;
3056 struct mds_obd *mds = mds_req2mds(req);
3057 void *handle = NULL;
3061 handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
3063 GOTO(cleanup, rc = PTR_ERR(handle));
3064 rc = fsfilt_del_dir_entry(obd, de);
3069 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3070 handle, req, rc, 0);
3074 static int mds_reint_rename_create_name(struct mds_update_record *rec,
3075 int offset, struct ptlrpc_request *req)
3077 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3078 struct obd_device *obd = req->rq_export->exp_obd;
3079 struct mds_obd *mds = mds_req2mds(req);
3080 struct lustre_handle child_lockh = {0};
3081 struct dentry *de_tgtdir = NULL;
3082 struct dentry *de_new = NULL;
3083 int cleanup_phase = 0;
3084 int update_mode, rc = 0;
3088 * another MDS executing rename operation has asked us to create target
3089 * name. such a creation should destroy existing target name.
3091 CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
3092 obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
3094 /* first, lookup the target */
3095 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
3096 &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
3097 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
3098 &child_lockh, &de_new, LCK_EX,
3099 MDS_INODELOCK_LOOKUP);
3106 LASSERT(de_tgtdir->d_inode);
3109 rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
3115 if (cleanup_phase == 1) {
3117 if (parent_lockh[1].cookie != 0)
3118 ldlm_lock_decref(parent_lockh + 1, update_mode);
3120 ldlm_lock_decref(parent_lockh, LCK_PW);
3121 if (child_lockh.cookie != 0)
3122 ldlm_lock_decref(&child_lockh, LCK_EX);
3127 req->rq_status = rc;
3131 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3132 struct ptlrpc_request *req)
3134 struct obd_device *obd = req->rq_export->exp_obd;
3135 struct ptlrpc_request *req2 = NULL;
3136 struct dentry *de_srcdir = NULL;
3137 struct dentry *de_old = NULL;
3138 struct mds_obd *mds = mds_req2mds(req);
3139 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3140 struct lustre_handle child_lockh = {0};
3141 struct mdc_op_data *op_data;
3142 int update_mode, rc = 0;
3145 CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3146 obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3148 OBD_ALLOC(op_data, sizeof(*op_data));
3149 if (op_data == NULL)
3151 memset(op_data, 0, sizeof(*op_data));
3153 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3154 &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3155 &update_mode, rec->ur_name,
3156 rec->ur_namelen, &child_lockh, &de_old,
3157 LCK_EX, MDS_INODELOCK_LOOKUP);
3159 OBD_FREE(op_data, sizeof(*op_data));
3164 LASSERT(de_srcdir->d_inode);
3168 * we already know the target should be created on another MDS so, we
3169 * have to request that MDS to do it.
3172 /* prepare source id */
3173 if (de_old->d_flags & DCACHE_CROSS_REF) {
3174 LASSERT(de_old->d_inode == NULL);
3175 CDEBUG(D_OTHER, "request to move remote name\n");
3176 mds_pack_dentry2id(obd, &op_data->id1, de_old, 1);
3177 } else if (de_old->d_inode == NULL) {
3178 /* oh, source doesn't exist */
3179 OBD_FREE(op_data, sizeof(*op_data));
3180 GOTO(cleanup, rc = -ENOENT);
3182 struct lustre_id sid;
3183 struct inode *inode = de_old->d_inode;
3185 LASSERT(inode != NULL);
3186 CDEBUG(D_OTHER, "request to move local name\n");
3187 id_ino(&op_data->id1) = inode->i_ino;
3188 id_group(&op_data->id1) = mds->mds_num;
3189 id_gen(&op_data->id1) = inode->i_generation;
3191 down(&inode->i_sem);
3192 rc = mds_read_inode_sid(obd, inode, &sid);
3195 CERROR("Can't read inode self id, "
3196 "inode %lu, rc = %d\n",
3201 id_fid(&op_data->id1) = id_fid(&sid);
3204 op_data->id2 = *rec->ur_id2;
3205 rc = md_rename(mds->mds_md_exp, op_data, NULL, 0,
3206 rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3207 OBD_FREE(op_data, sizeof(*op_data));
3212 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3218 ptlrpc_req_finished(req2);
3221 if (parent_lockh[1].cookie != 0)
3222 ldlm_lock_decref(parent_lockh + 1, update_mode);
3224 ldlm_lock_decref(parent_lockh, LCK_PW);
3225 if (child_lockh.cookie != 0)
3226 ldlm_lock_decref(&child_lockh, LCK_EX);
3231 req->rq_status = rc;
3235 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3236 struct ptlrpc_request *req, struct lustre_handle *lockh)
3238 struct obd_device *obd = req->rq_export->exp_obd;
3239 struct dentry *de_srcdir = NULL;
3240 struct dentry *de_tgtdir = NULL;
3241 struct dentry *de_old = NULL;
3242 struct dentry *de_new = NULL;
3243 struct inode *old_inode = NULL, *new_inode = NULL;
3244 struct mds_obd *mds = mds_req2mds(req);
3245 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3246 struct mds_body *body = NULL;
3247 struct llog_create_locks *lcl = NULL;
3248 struct lov_mds_md *lmm = NULL;
3249 int rc = 0, cleanup_phase = 0;
3250 void *handle = NULL;
3253 LASSERT(offset == 1);
3255 DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3256 OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3259 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3261 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3262 DEBUG_REQ(D_HA, req, "rename replay\n");
3263 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3264 lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3265 req->rq_repmsg->buflens[2]);
3268 MD_COUNTER_INCREMENT(obd, rename);
3270 if (rec->ur_namelen == 1) {
3271 rc = mds_reint_rename_create_name(rec, offset, req);
3275 /* check if new name should be located on remote target. */
3276 if (id_group(rec->ur_id2) != mds->mds_num) {
3277 rc = mds_reint_rename_to_remote(rec, offset, req);
3281 rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3282 rec->ur_id2, &de_tgtdir, LCK_PW,
3283 rec->ur_name, rec->ur_namelen,
3284 &de_old, rec->ur_tgt,
3285 rec->ur_tgtlen, &de_new,
3286 dlm_handles, LCK_EX);
3290 cleanup_phase = 1; /* parent(s), children, locks */
3291 old_inode = de_old->d_inode;
3292 new_inode = de_new->d_inode;
3294 /* sanity check for src inode */
3295 if (de_old->d_flags & DCACHE_CROSS_REF) {
3296 LASSERT(de_old->d_inode == NULL);
3299 * in the case of cross-ref dir, we can perform this check only
3300 * if child and parent lie on the same mds. This is because
3301 * otherwise they can have the same inode numbers.
3303 if (de_old->d_mdsnum == mds->mds_num) {
3304 if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3305 de_old->d_inum == de_tgtdir->d_inode->i_ino)
3306 GOTO(cleanup, rc = -EINVAL);
3309 LASSERT(de_old->d_inode != NULL);
3310 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3311 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3312 GOTO(cleanup, rc = -EINVAL);
3315 /* sanity check for dest inode */
3316 if (de_new->d_flags & DCACHE_CROSS_REF) {
3317 LASSERT(new_inode == NULL);
3319 /* the same check about target dentry. */
3320 if (de_new->d_mdsnum == mds->mds_num) {
3321 if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3322 de_new->d_inum == de_tgtdir->d_inode->i_ino)
3323 GOTO(cleanup, rc = -EINVAL);
3327 * regular files usualy do not have ->rename() implemented. But
3328 * we handle only this case when @de_new is cross-ref entry,
3329 * because in other cases it will be handled by vfs_rename().
3331 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3332 !de_old->d_inode->i_op->rename))
3333 GOTO(cleanup, rc = -EPERM);
3336 (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3337 new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3338 GOTO(cleanup, rc = -EINVAL);
3343 * check if inodes point to each other. This should be checked before
3344 * is_subdir() check, as for the same entries it will think that they
3347 if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3348 !(de_new->d_flags & DCACHE_CROSS_REF) &&
3349 old_inode == new_inode)
3350 GOTO(cleanup, rc = 0);
3352 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3354 * check if we are moving old entry into its child. 2.6 does not check
3355 * for this in vfs_rename() anymore.
3357 if (is_subdir(de_new, de_old))
3358 GOTO(cleanup, rc = -EINVAL);
3362 * if we are about to remove the target at first, pass the EA of that
3363 * inode to client to perform and cleanup on OST.
3365 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3366 LASSERT(body != NULL);
3368 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3370 DOWN_READ_I_ALLOC_SEM(new_inode);
3372 cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3374 if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3375 new_inode->i_nlink == 2) ||
3376 new_inode->i_nlink == 1)) {
3377 if (mds_orphan_open_count(new_inode) > 0) {
3378 /* need to lock pending_dir before transaction */
3379 down(&mds->mds_pending_dir->d_inode->i_sem);
3380 cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3381 } else if (S_ISREG(new_inode->i_mode)) {
3382 mds_pack_inode2body(obd, body, new_inode, 0);
3383 mds_pack_md(obd, req->rq_repmsg, 1, body,
3384 new_inode, MDS_PACK_MD_LOCK, 0);
3388 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3389 de_srcdir->d_inode->i_sb);
3391 if (de_old->d_flags & DCACHE_CROSS_REF) {
3392 struct lustre_id old_id;
3394 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3396 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3401 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3406 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3407 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3408 NULL, le32_to_cpu(lmm->lmm_stripe_count));
3411 GOTO(cleanup, rc = PTR_ERR(handle));
3414 de_old->d_fsdata = req;
3415 de_new->d_fsdata = req;
3416 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3419 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3420 if (mds_orphan_open_count(new_inode) > 0)
3421 rc = mds_orphan_add_link(rec, obd, de_new);
3424 GOTO(cleanup, rc = 0);
3426 if (!S_ISREG(new_inode->i_mode))
3429 if (!(body->valid & OBD_MD_FLEASIZE)) {
3430 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3431 OBD_MD_FLATIME | OBD_MD_FLMTIME);
3432 } else if (mds_log_op_unlink(obd, new_inode,
3433 lustre_msg_buf(req->rq_repmsg,1,0),
3434 req->rq_repmsg->buflens[1],
3435 lustre_msg_buf(req->rq_repmsg,2,0),
3436 req->rq_repmsg->buflens[2],
3438 body->valid |= OBD_MD_FLCOOKIE;
3444 rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3445 handle, req, rc, 0);
3447 switch (cleanup_phase) {
3449 up(&mds->mds_pending_dir->d_inode->i_sem);
3452 UP_READ_I_ALLOC_SEM(new_inode);
3455 if (dlm_handles[5].cookie != 0)
3456 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3457 if (dlm_handles[6].cookie != 0)
3458 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3461 ptlrpc_save_llog_lock(req, lcl);
3464 if (dlm_handles[3].cookie != 0)
3465 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3466 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3467 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3468 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3470 if (dlm_handles[3].cookie != 0)
3471 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3472 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3473 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3474 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3483 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3486 req->rq_status = rc;
3490 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3491 struct ptlrpc_request *, struct lustre_handle *);
3493 static mds_reinter reinters[REINT_MAX + 1] = {
3494 [REINT_SETATTR] mds_reint_setattr,
3495 [REINT_CREATE] mds_reint_create,
3496 [REINT_LINK] mds_reint_link,
3497 [REINT_UNLINK] mds_reint_unlink,
3498 [REINT_RENAME] mds_reint_rename,
3499 [REINT_OPEN] mds_open
3502 int mds_reint_rec(struct mds_update_record *rec, int offset,
3503 struct ptlrpc_request *req, struct lustre_handle *lockh)
3505 struct obd_device *obd = req->rq_export->exp_obd;
3506 struct lvfs_run_ctxt saved;
3509 /* checked by unpacker */
3510 LASSERT(rec->ur_opcode <= REINT_MAX &&
3511 reinters[rec->ur_opcode] != NULL);
3513 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3514 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3515 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);