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_acl.h>
47 #include <linux/lustre_lite.h>
48 #include <linux/lustre_smfs.h>
49 #include "mds_internal.h"
51 struct mds_logcancel_data {
52 struct lov_mds_md *mlcd_lmm;
56 struct llog_cookie mlcd_cookies[0];
59 static void mds_cancel_cookies_cb(struct obd_device *obd,
60 __u64 transno, void *cb_data,
63 struct mds_logcancel_data *mlcd = cb_data;
64 struct lov_stripe_md *lsm = NULL;
65 struct llog_ctxt *ctxt;
68 obd_transno_commit_cb(obd, transno, error);
70 CDEBUG(D_HA, "cancelling %d cookies\n",
71 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
73 rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, mlcd->mlcd_lmm,
74 mlcd->mlcd_eadatalen);
76 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
77 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
80 ///* XXX 0 normally, SENDNOW for debug */);
81 ctxt = llog_get_context(&obd->obd_llogs,
82 mlcd->mlcd_cookies[0].lgc_subsys + 1);
83 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
84 sizeof(*mlcd->mlcd_cookies),
85 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
87 CERROR("error cancelling %d log cookies: rc %d\n",
88 (int)(mlcd->mlcd_cookielen /
89 sizeof(*mlcd->mlcd_cookies)), rc);
90 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
93 OBD_FREE(mlcd, mlcd->mlcd_size);
96 /* Assumes caller has already pushed us into the kernel context. */
97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
98 struct ptlrpc_request *req, int rc, __u32 op_data)
100 struct mds_export_data *med = &req->rq_export->exp_mds_data;
101 struct obd_device *obd = req->rq_export->exp_obd;
102 struct mds_client_data *mcd = med->med_mcd;
103 int err, log_pri = D_HA;
108 /* if the export has already been failed, we have no last_rcvd slot */
109 if (req->rq_export->exp_failed) {
110 CERROR("committing transaction for disconnected client\n");
112 GOTO(out_commit, rc);
119 if (handle == NULL) {
120 /* if we're starting our own xaction, use our own inode */
121 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
122 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
123 if (IS_ERR(handle)) {
124 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
125 RETURN(PTR_ERR(handle));
131 transno = req->rq_reqmsg->transno;
133 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
134 } else if (transno == 0) {
135 spin_lock(&mds->mds_transno_lock);
136 transno = ++mds->mds_last_transno;
137 spin_unlock(&mds->mds_transno_lock);
139 spin_lock(&mds->mds_transno_lock);
140 if (transno > mds->mds_last_transno)
141 mds->mds_last_transno = transno;
142 spin_unlock(&mds->mds_transno_lock);
144 req->rq_repmsg->transno = req->rq_transno = transno;
145 if (req->rq_reqmsg->opc == MDS_CLOSE) {
146 mcd->mcd_last_close_transno = cpu_to_le64(transno);
147 mcd->mcd_last_close_xid = cpu_to_le64(req->rq_xid);
148 mcd->mcd_last_close_result = cpu_to_le32(rc);
149 mcd->mcd_last_close_data = cpu_to_le32(op_data);
151 mcd->mcd_last_transno = cpu_to_le64(transno);
152 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
153 mcd->mcd_last_result = cpu_to_le32(rc);
154 mcd->mcd_last_data = cpu_to_le32(op_data);
157 fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
158 mds_commit_last_transno_cb, NULL);
160 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
161 sizeof(*mcd), &off, 0);
169 DEBUG_REQ(log_pri, req,
170 "wrote trans #"LPU64" client %s at idx %u: err = %d",
171 transno, mcd->mcd_uuid, med->med_idx, err);
173 err = mds_update_last_fid(obd, handle, 0);
180 err = mds_dt_write_objids(obd);
186 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
190 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
191 req->rq_export->exp_sync);
193 CERROR("error committing transaction: %d\n", err);
201 /* this gives the same functionality as the code between
202 * sys_chmod and inode_setattr
203 * chown_common and inode_setattr
204 * utimes and inode_setattr
207 /* Just for the case if we have some clients that know about ATTR_RAW */
208 #define ATTR_RAW 8192
210 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
212 time_t now = LTIME_S(CURRENT_TIME);
213 struct iattr *attr = &rec->ur_iattr;
214 unsigned int ia_valid = attr->ia_valid;
218 /* only fix up attrs if the client VFS didn't already */
220 if (!(ia_valid & ATTR_RAW))
223 if (!(ia_valid & ATTR_CTIME_SET))
224 LTIME_S(attr->ia_ctime) = now;
225 if (!(ia_valid & ATTR_ATIME_SET))
226 LTIME_S(attr->ia_atime) = now;
227 if (!(ia_valid & ATTR_MTIME_SET))
228 LTIME_S(attr->ia_mtime) = now;
230 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
234 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
235 if (rec->ur_fsuid != inode->i_uid &&
236 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
240 if (ia_valid & ATTR_SIZE) {
241 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
245 if (ia_valid & ATTR_UID) {
248 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
250 if (attr->ia_uid == (uid_t) -1)
251 attr->ia_uid = inode->i_uid;
252 if (attr->ia_gid == (gid_t) -1)
253 attr->ia_gid = inode->i_gid;
254 attr->ia_mode = inode->i_mode;
256 * If the user or group of a non-directory has been
257 * changed by a non-root user, remove the setuid bit.
258 * 19981026 David C Niemi <niemi@tux.org>
260 * Changed this to apply to all users, including root,
261 * to avoid some races. This is the behavior we had in
262 * 2.0. The check for non-root was definitely wrong
263 * for 2.2 anyway, as it should have been using
264 * CAP_FSETID rather than fsuid -- 19990830 SD.
266 if ((inode->i_mode & S_ISUID) == S_ISUID &&
267 !S_ISDIR(inode->i_mode)) {
268 attr->ia_mode &= ~S_ISUID;
269 attr->ia_valid |= ATTR_MODE;
272 * Likewise, if the user or group of a non-directory
273 * has been changed by a non-root user, remove the
274 * setgid bit UNLESS there is no group execute bit
275 * (this would be a file marked for mandatory
276 * locking). 19981026 David C Niemi <niemi@tux.org>
278 * Removed the fsuid check (see the comment above) --
281 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
282 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
283 attr->ia_mode &= ~S_ISGID;
284 attr->ia_valid |= ATTR_MODE;
286 } else if (ia_valid & ATTR_MODE) {
287 int mode = attr->ia_mode;
289 if (attr->ia_mode == (mode_t) -1)
290 attr->ia_mode = inode->i_mode;
292 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
297 void mds_steal_ack_locks(struct ptlrpc_request *req)
299 struct obd_export *exp = req->rq_export;
300 char str[PTL_NALFMT_SIZE];
301 struct list_head *tmp;
302 struct ptlrpc_reply_state *oldrep;
303 struct ptlrpc_service *svc;
304 struct llog_create_locks *lcl;
308 /* CAVEAT EMPTOR: spinlock order */
309 spin_lock_irqsave (&exp->exp_lock, flags);
310 list_for_each (tmp, &exp->exp_outstanding_replies) {
311 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
313 if (oldrep->rs_xid != req->rq_xid)
316 if (oldrep->rs_msg->opc != req->rq_reqmsg->opc)
317 CERROR ("Resent req xid "LPX64" has mismatched opc: "
318 "new %d old %d\n", req->rq_xid,
319 req->rq_reqmsg->opc, oldrep->rs_msg->opc);
321 svc = oldrep->rs_srv_ni->sni_service;
322 spin_lock (&svc->srv_lock);
324 list_del_init (&oldrep->rs_exp_list);
326 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
327 " o%d NID %s\n", oldrep->rs_nlocks, oldrep,
328 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg->opc,
329 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
331 for (i = 0; i < oldrep->rs_nlocks; i++)
332 ptlrpc_save_lock(req,
333 &oldrep->rs_locks[i],
334 oldrep->rs_modes[i]);
335 oldrep->rs_nlocks = 0;
337 lcl = oldrep->rs_llog_locks;
338 oldrep->rs_llog_locks = NULL;
340 ptlrpc_save_llog_lock(req, lcl);
342 DEBUG_REQ(D_HA, req, "stole locks for");
343 ptlrpc_schedule_difficult_reply (oldrep);
345 spin_unlock (&svc->srv_lock);
346 spin_unlock_irqrestore (&exp->exp_lock, flags);
349 spin_unlock_irqrestore (&exp->exp_lock, flags);
352 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
354 if (req->rq_reqmsg->opc == MDS_CLOSE) {
355 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
356 mcd->mcd_last_close_transno, mcd->mcd_last_close_result);
357 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_close_transno;
358 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_close_result;
360 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
361 mcd->mcd_last_transno, mcd->mcd_last_result);
362 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
363 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
366 mds_steal_ack_locks(req);
369 static void reconstruct_reint_setattr(struct mds_update_record *rec,
370 int offset, struct ptlrpc_request *req)
372 struct mds_export_data *med = &req->rq_export->exp_mds_data;
373 struct mds_body *body;
376 mds_req_from_mcd(req, med->med_mcd);
378 de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
380 LASSERT(PTR_ERR(de) == req->rq_status);
384 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
385 mds_pack_inode2body(req2obd(req), body, de->d_inode, 1);
387 /* Don't return OST-specific attributes if we didn't just set them */
388 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
389 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
390 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
391 body->valid |= OBD_MD_FLMTIME;
392 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
393 body->valid |= OBD_MD_FLATIME;
398 static int mds_reint_remote_setfacl(struct obd_device *obd,
399 struct mds_export_data *med,
400 struct mds_update_record *rec,
401 struct ptlrpc_request *req)
403 struct rmtacl_upcall_desc desc;
406 struct mds_body *body;
408 int repsize[2] = { sizeof(*body), LUSTRE_ACL_SIZE_MAX };
411 rc = lustre_pack_reply(req, 2, repsize, NULL);
415 de = mds_id2dentry(obd, rec->ur_id1, NULL);
417 GOTO(out, rc = PTR_ERR(de));
422 /* setxattr from remote client:
424 memset(&desc, 0, sizeof(desc));
425 desc.cmd = (char *) rec->ur_ea2data;
426 desc.cmdlen = rec->ur_ea2datalen;
427 desc.res = lustre_msg_buf(req->rq_repmsg, 1, LUSTRE_ACL_SIZE_MAX);
428 desc.reslen = LUSTRE_ACL_SIZE_MAX;
430 mds_do_remote_acl_upcall(&desc);
431 if (desc.upcall_status)
432 GOTO(out_put, rc = desc.upcall_status);
435 desc.status = -desc.status;
437 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
440 /* client (lmv) will do limited checking upon replied mds_body,
441 * we pack it as normal, but "steal" field "flags" field to store
442 * the acl execution status.
444 mds_pack_inode2body(obd, body, inode, 1);
445 body->flags = desc.status;
446 mds_body_do_reverse_map(med, body);
456 /*This is a tmp fix for cmobd setattr reint*/
458 #define XATTR_LUSTRE_MDS_LOV_EA "lov"
459 #define XATTR_LUSTRE_MDS_MEA_EA "mea"
460 #define XATTR_LUSTRE_MDS_MID_EA "mid"
461 #define XATTR_LUSTRE_MDS_SID_EA "sid"
463 static int mds_get_md_type(char *name)
465 if (!strcmp(name, XATTR_LUSTRE_MDS_LOV_EA))
467 if (!strcmp(name, XATTR_LUSTRE_MDS_MEA_EA))
469 if (!strcmp(name, XATTR_LUSTRE_MDS_MID_EA))
471 if (!strcmp(name, XATTR_LUSTRE_MDS_SID_EA))
475 /* In the raw-setattr case, we lock the child inode.
476 * In the write-back case or if being called from open, the client holds a lock
479 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
480 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
481 struct ptlrpc_request *req, struct lustre_handle *lh)
483 struct mds_obd *mds = mds_req2mds(req);
484 struct obd_device *obd = req->rq_export->exp_obd;
485 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
486 struct mds_body *body;
487 struct dentry *de = NULL;
488 struct inode *inode = NULL;
489 struct lustre_handle lockh[2] = {{0}, {0}};
492 struct mds_logcancel_data *mlcd = NULL;
493 int rc = 0, cleanup_phase = 0, err;
494 int repsize = sizeof(*body), locked = 0;
497 LASSERT(offset == 1);
499 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
500 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
501 rec->ur_iattr.ia_valid);
503 /* remote setfacl need special handling */
504 if ((rec->ur_iattr.ia_valid & ATTR_EA) &&
505 !strcmp(rec->ur_eadata, XATTR_NAME_LUSTRE_ACL)) {
506 return mds_reint_remote_setfacl(obd, med, rec, req);
509 rc = lustre_pack_reply(req, 1, &repsize, NULL);
513 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
514 MD_COUNTER_INCREMENT(obd, setattr);
516 if (med->med_remote) {
517 if (rec->ur_iattr.ia_valid & ATTR_GID) {
518 CWARN("Deny chgrp from remote client\n");
519 GOTO(cleanup, rc = -EPERM);
521 if (rec->ur_iattr.ia_valid & ATTR_UID) {
524 uid = mds_idmap_lookup_uid(med->med_idmap, 0,
525 rec->ur_iattr.ia_uid);
526 if (uid == MDS_IDMAP_NOTFOUND) {
527 CWARN("Deny chown to uid %u\n",
528 rec->ur_iattr.ia_uid);
529 GOTO(cleanup, rc = -EPERM);
531 rec->ur_iattr.ia_uid = uid;
535 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
536 de = mds_id2dentry(obd, rec->ur_id1, NULL);
538 GOTO(cleanup, rc = PTR_ERR(de));
540 __u64 lockpart = MDS_INODELOCK_UPDATE;
541 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
542 lockpart |= MDS_INODELOCK_LOOKUP;
543 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
544 lockh, &parent_mode, NULL, 0, lockpart);
546 GOTO(cleanup, rc = PTR_ERR(de));
554 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
555 rec->ur_eadata != NULL)
558 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
560 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
562 GOTO(cleanup, rc = PTR_ERR(handle));
564 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
565 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
566 LTIME_S(rec->ur_iattr.ia_mtime),
567 LTIME_S(rec->ur_iattr.ia_ctime));
568 rc = mds_fix_attr(inode, rec);
572 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
573 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
574 (long)&rec->ur_iattr.ia_attr_flags);
576 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
579 if (rec->ur_iattr.ia_valid & ATTR_EA) {
580 int flags = (int) rec->ur_iattr.ia_attr_flags;
583 if (!med->med_remote && inode->i_op &&
584 inode->i_op->setxattr)
585 rc = inode->i_op->setxattr(
590 } else if (rec->ur_iattr.ia_valid & ATTR_EA_RM) {
592 if (!med->med_remote && inode->i_op &&
593 inode->i_op->removexattr)
594 rc = inode->i_op->removexattr(
596 } else if (rec->ur_iattr.ia_valid & ATTR_EA_CMOBD) {
599 /*tmp fix for cmobd set md reint*/
600 LASSERT(rec->ur_eadata != NULL);
601 LASSERT(rec->ur_ea2data != NULL);
602 name = rec->ur_eadata;
603 CDEBUG(D_INFO, "set %s EA for cmobd \n", name);
604 type = mds_get_md_type(name);
606 rc = fsfilt_set_md(obd, inode, handle,
608 rec->ur_ea2datalen, type);
611 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) {
612 struct lov_stripe_md *lsm = NULL;
613 struct lov_user_md *lum = NULL;
615 if (rec->ur_eadata != NULL) {
616 rc = ll_permission(inode, MAY_WRITE, NULL);
620 lum = rec->ur_eadata;
622 /* if lmm_stripe_size is -1 delete default
624 if (S_ISDIR(inode->i_mode) &&
625 lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1)){
626 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, EA_LOV);
630 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
632 &lsm, rec->ur_eadata);
636 obd_free_memmd(mds->mds_dt_exp, &lsm);
637 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
638 rec->ur_eadatalen, EA_LOV);
646 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
647 mds_pack_inode2body(obd, body, inode, 1);
649 /* Don't return OST-specific attributes if we didn't just set them */
650 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
651 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
652 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
653 body->valid |= OBD_MD_FLMTIME;
654 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
655 body->valid |= OBD_MD_FLATIME;
657 mds_body_do_reverse_map(med, body);
659 /* The logcookie should be no use anymore, why nobody remove
660 * following code block?
662 LASSERT(rec->ur_cookielen == 0);
663 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_dt_obd)) {
664 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
667 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
669 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
670 mlcd->mlcd_cookielen = rec->ur_cookielen;
671 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
672 mlcd->mlcd_cookielen;
673 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
674 mlcd->mlcd_cookielen);
675 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
676 mlcd->mlcd_eadatalen);
678 CERROR("unable to allocate log cancel data\n");
684 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
685 handle, mds_cancel_cookies_cb, mlcd);
686 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
687 switch (cleanup_phase) {
689 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
690 rec->ur_eadata != NULL)
695 if (lockh[1].cookie != 0)
696 ldlm_lock_decref(lockh + 1, parent_mode);
699 ldlm_lock_decref(lockh, LCK_PW);
701 ptlrpc_save_lock (req, lockh, LCK_PW);
716 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
717 struct ptlrpc_request *req)
719 struct mds_export_data *med = &req->rq_export->exp_mds_data;
720 struct dentry *parent, *child;
721 struct mds_body *body;
724 mds_req_from_mcd(req, med->med_mcd);
726 if (req->rq_status) {
731 parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
732 LASSERT(!IS_ERR(parent));
733 child = ll_lookup_one_len(rec->ur_name, parent,
734 rec->ur_namelen - 1);
735 LASSERT(!IS_ERR(child));
736 if ((child->d_flags & DCACHE_CROSS_REF)) {
737 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
738 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
739 mds_pack_dentry2body(req2obd(req), body, child, 1);
740 } else if (child->d_inode == NULL) {
741 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
742 OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
743 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
745 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
746 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
753 static int mds_get_default_acl(struct inode *dir, void **pacl)
755 struct dentry de = { .d_inode = dir };
758 LASSERT(S_ISDIR(dir->i_mode));
760 if (!dir->i_op->getxattr)
763 size = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, NULL, 0);
764 if (size == 0 || size == -ENODATA || size == -EOPNOTSUPP)
769 OBD_ALLOC(*pacl, size);
773 size2 = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, *pacl, size);
775 /* since we already locked the dir, it should not change
776 * between the 2 getxattr calls
778 CERROR("2'nd getxattr got %d, expect %d\n", size2, size);
779 OBD_FREE(*pacl, size);
786 static int mds_reint_create(struct mds_update_record *rec, int offset,
787 struct ptlrpc_request *req,
788 struct lustre_handle *lh)
790 struct dentry *dparent = NULL;
791 struct mds_obd *mds = mds_req2mds(req);
792 struct obd_device *obd = req->rq_export->exp_obd;
793 struct dentry *dchild = NULL;
794 struct inode *dir = NULL;
796 struct lustre_handle lockh[2] = {{0}, {0}};
798 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
800 struct dentry_params dp;
801 struct mea *mea = NULL;
806 LASSERT(offset == 1);
808 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
809 OBD_MDS_DEVICENAME));
811 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
812 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
813 rec->ur_name, rec->ur_mode);
815 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
817 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
818 GOTO(cleanup, rc = -ESTALE);
820 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
821 lockh, &parent_mode, rec->ur_name,
822 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
823 if (IS_ERR(dparent)) {
824 rc = PTR_ERR(dparent);
825 CERROR("parent lookup error %d\n", rc);
828 cleanup_phase = 1; /* locked parent dentry */
829 dir = dparent->d_inode;
832 ldlm_lock_dump_handle(D_OTHER, lockh);
834 /* try to retrieve MEA data for this dir */
835 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
841 * dir is already splitted, check is requested filename should
842 * live at this MDS or at another one.
844 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
845 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
846 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
847 " should be %lu(%d)\n",
848 mea->mea_master, dparent->d_inode->i_ino,
849 dparent->d_inode->i_generation, rec->ur_name,
850 (unsigned long)id_group(&mea->mea_ids[i]), i);
851 GOTO(cleanup, rc = -ERESTART);
855 dchild = ll_lookup_one_len(rec->ur_name, dparent,
856 rec->ur_namelen - 1);
857 if (IS_ERR(dchild)) {
858 rc = PTR_ERR(dchild);
859 CERROR("Can't find "DLID4"/%s, error %d\n",
860 OLID4(rec->ur_id1), rec->ur_name, rc);
864 cleanup_phase = 2; /* child dentry */
866 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
868 if (type == S_IFREG || type == S_IFDIR) {
869 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
870 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
871 obd->obd_name, dparent->d_inode->i_ino,
872 dparent->d_inode->i_generation, rc, parent_mode);
874 /* dir got splitted */
875 GOTO(cleanup, rc = -ERESTART);
877 /* error happened during spitting. */
882 if (dir->i_mode & S_ISGID) {
883 if (S_ISDIR(rec->ur_mode))
884 rec->ur_mode |= S_ISGID;
888 * here inode number should be used only in the case of replaying. It is
889 * needed to check if object already created in the case of creating
892 if (id_ino(rec->ur_id2))
893 fid = id_fid(rec->ur_id2);
895 fid = mds_alloc_fid(obd);
896 dchild->d_fsdata = (void *)&dp;
897 dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
900 dp.p_group = mds->mds_num;
904 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
906 GOTO(cleanup, rc = PTR_ERR(handle));
907 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
915 * as Peter asked, mkdir() should distribute new directories
916 * over the whole cluster in order to distribute namespace
917 * processing load. first, we calculate which MDS to use to put
918 * new directory's inode in.
920 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
921 rec->ur_flags, &req->rq_peer, dir);
922 if (i == mds->mds_num) {
923 /* inode will be created locally */
924 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
926 GOTO(cleanup, rc = PTR_ERR(handle));
928 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
931 "Can't create dir \"%s\", rc = %d\n",
932 dchild->d_name.name, rc);
936 down(&dchild->d_inode->i_sem);
938 rc = mds_update_inode_sid(obd, dchild->d_inode,
939 handle, rec->ur_id2);
941 CERROR("mds_update_inode_sid() failed, inode %lu, "
942 "rc %d\n", dchild->d_inode->i_ino, rc);
946 * make sure, that fid is up-to-date.
948 mds_set_last_fid(obd, id_fid(rec->ur_id2));
950 rc = mds_set_inode_sid(obd, dchild->d_inode,
953 CERROR("mds_set_inode_sid() failed, inode %lu, "
954 "rc %d\n", dchild->d_inode->i_ino, rc);
957 up(&dchild->d_inode->i_sem);
963 nstripes = *(u16 *)rec->ur_eadata;
965 if (rc == 0 && nstripes) {
967 * we pass LCK_EX to split routine to signal,
968 * that we have exclusive access to the
969 * directory. Simple because nobody knows it
970 * already exists -bzzz
972 rc = mds_try_to_split_dir(obd, dchild,
976 /* dir got splitted */
979 /* an error occured during
984 } else if (!DENTRY_VALID(dchild)) {
985 /* inode will be created on another MDS */
986 struct obdo *oa = NULL;
987 struct mds_body *body;
991 /* first, create that inode */
994 GOTO(cleanup, rc = -ENOMEM);
999 if (rec->ur_eadata) {
1000 /* user asks for creating splitted dir */
1001 oa->o_easize = *((u16 *) rec->ur_eadata);
1004 obdo_from_inode(oa, dir, OBD_MD_FLATIME |
1005 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1007 /* adjust the uid/gid/mode bits */
1008 oa->o_mode = rec->ur_mode;
1009 oa->o_uid = current->fsuid;
1010 oa->o_gid = (dir->i_mode & S_ISGID) ?
1011 dir->i_gid : current->fsgid;
1012 oa->o_valid |= OBD_MD_FLTYPE|OBD_MD_FLUID|OBD_MD_FLGID;
1014 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
1017 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1019 * here inode number and generation are
1020 * important, as this is replay request and we
1021 * need them to check if such an object is
1024 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
1025 obd->obd_name, rec->ur_namelen - 1,
1026 rec->ur_name, (unsigned)id_ino(rec->ur_id2),
1027 (unsigned)id_gen(rec->ur_id2));
1028 oa->o_id = id_ino(rec->ur_id2);
1029 oa->o_fid = id_fid(rec->ur_id2);
1030 oa->o_generation = id_gen(rec->ur_id2);
1031 oa->o_flags |= OBD_FL_RECREATE_OBJS;
1032 LASSERT(oa->o_fid != 0);
1035 /* obtain default ACL */
1036 acl_size = mds_get_default_acl(dir, &acl);
1039 GOTO(cleanup, rc = -ENOMEM);
1043 * before obd_create() is called, o_fid is not known if
1044 * this is not recovery of cause.
1046 rc = obd_create(mds->mds_md_exp, oa, acl, acl_size,
1050 OBD_FREE(acl, acl_size);
1053 CERROR("can't create remote inode: %d\n", rc);
1054 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
1055 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1056 rec->ur_name, rec->ur_mode);
1061 LASSERT(oa->o_fid != 0);
1063 /* now, add new dir entry for it */
1064 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1065 if (IS_ERR(handle)) {
1067 GOTO(cleanup, rc = PTR_ERR(handle));
1070 /* creating local dentry for remote inode. */
1071 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
1072 rec->ur_namelen - 1, oa->o_id,
1073 oa->o_generation, i, oa->o_fid);
1076 CERROR("Can't create local entry %*s for "
1077 "remote inode.\n", rec->ur_namelen - 1,
1083 body = lustre_msg_buf(req->rq_repmsg,
1085 body->valid |= OBD_MD_FLID | OBD_MD_MDS | OBD_MD_FID;
1087 obdo2id(&body->id1, oa);
1090 /* requested name exists in the directory */
1097 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
1099 GOTO(cleanup, rc = PTR_ERR(handle));
1100 if (rec->ur_tgt == NULL) /* no target supplied */
1101 rc = -EINVAL; /* -EPROTO? */
1103 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
1111 int rdev = rec->ur_rdev;
1112 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
1114 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
1115 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
1120 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
1121 dchild->d_fsdata = NULL;
1122 GOTO(cleanup, rc = -EINVAL);
1125 /* In case we stored the desired inum in here, we want to clean up. */
1126 if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
1127 dchild->d_fsdata = NULL;
1130 CDEBUG(D_INODE, "error during create: %d\n", rc);
1132 } else if (dchild->d_inode) {
1133 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1135 struct mds_body *body;
1136 struct inode *inode = dchild->d_inode;
1139 iattr.ia_uid = rec->ur_fsuid;
1140 LTIME_S(iattr.ia_atime) = rec->ur_time;
1141 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1142 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1144 if (dir->i_mode & S_ISGID)
1145 iattr.ia_gid = dir->i_gid;
1147 iattr.ia_gid = rec->ur_fsgid;
1149 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1150 ATTR_MTIME | ATTR_CTIME;
1152 if (id_ino(rec->ur_id2)) {
1153 LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
1154 inode->i_generation = id_gen(rec->ur_id2);
1156 if (type != S_IFDIR) {
1157 down(&inode->i_sem);
1158 rc = mds_update_inode_sid(obd, inode,
1159 handle, rec->ur_id2);
1162 CERROR("Can't update inode self id, "
1167 * make sure, that fid is up-to-date.
1169 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1172 /* dirtied and committed by the upcoming setattr. */
1173 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1174 inode->i_ino, inode->i_generation);
1176 struct lustre_handle child_ino_lockh;
1178 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
1179 inode->i_ino, inode->i_generation);
1181 if (type != S_IFDIR) {
1183 * allocate new id for @inode if it is not dir,
1184 * because for dir it was already done.
1186 down(&inode->i_sem);
1187 rc = mds_set_inode_sid(obd, inode,
1191 CERROR("mds_set_inode_sid() failed, "
1192 "inode %lu, rc %d\n", inode->i_ino,
1199 * the inode we were allocated may have just
1200 * been freed by an unlink operation. We take
1201 * this lock to synchronize against the matching
1202 * reply-ack-lock taken in unlink, to avoid
1203 * replay problems if this reply makes it out to
1204 * the client but the unlink's does not. See
1205 * bug 2029 for more detail.
1207 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
1208 if (rc != ELDLM_OK) {
1209 CERROR("error locking for unlink/create sync: "
1212 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
1217 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1219 CERROR("error on child setattr: rc = %d\n", rc);
1221 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1222 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1224 CERROR("error on parent setattr: rc = %d\n", rc);
1226 MD_COUNTER_INCREMENT(obd, create);
1228 /* take care of default stripe inheritance */
1229 if (type == S_IFDIR) {
1230 struct lov_mds_md lmm;
1231 int lmm_size = sizeof(lmm);
1233 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0);
1235 down(&inode->i_sem);
1236 rc = fsfilt_set_md(obd, inode, handle,
1237 &lmm, lmm_size, EA_LOV);
1241 CERROR("error on copy stripe info: rc = %d\n",
1247 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1248 mds_pack_inode2body(obd, body, inode, 1);
1249 mds_body_do_reverse_map(med, body);
1254 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
1256 if (rc && created) {
1257 /* Destroy the file we just created. This should not need extra
1258 * journal credits, as we have already modified all of the
1259 * blocks needed in order to create the file in the first
1263 err = vfs_rmdir(dir, dchild);
1265 CERROR("rmdir in error path: %d\n", err);
1268 err = vfs_unlink(dir, dchild);
1270 CERROR("unlink in error path: %d\n", err);
1276 switch (cleanup_phase) {
1277 case 2: /* child dentry */
1279 case 1: /* locked parent dentry */
1281 if (lockh[1].cookie != 0)
1282 ldlm_lock_decref(lockh + 1, parent_mode);
1285 ldlm_lock_decref(lockh, LCK_PW);
1287 ptlrpc_save_lock(req, lockh, LCK_PW);
1293 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1297 OBD_FREE(mea, mea_size);
1298 req->rq_status = rc;
1303 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1304 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1308 for (i = 0; i < RES_NAME_SIZE; i++) {
1310 * this is needed to make zeroed res_id entries to be put at the
1311 * end of list in *ordered_locks() .
1313 if (res1->name[i] == 0 && res2->name[i] != 0)
1315 if (res2->name[i] == 0 && res1->name[i] != 0)
1317 if (res1->name[i] > res2->name[i])
1319 if (res1->name[i] < res2->name[i])
1326 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1332 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1333 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1334 * because they take the place of local semaphores.
1336 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1337 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1338 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1339 struct lustre_handle *p1_lockh, int p1_lock_mode,
1340 ldlm_policy_data_t *p1_policy,
1341 struct ldlm_res_id *p2_res_id,
1342 struct lustre_handle *p2_lockh, int p2_lock_mode,
1343 ldlm_policy_data_t *p2_policy)
1345 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1346 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1347 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1348 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1352 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1354 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1355 res_id[0]->name[0], res_id[1]->name[0]);
1357 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1358 handles[1] = p1_lockh;
1359 handles[0] = p2_lockh;
1360 res_id[1] = p1_res_id;
1361 res_id[0] = p2_res_id;
1362 lock_modes[1] = p1_lock_mode;
1363 lock_modes[0] = p2_lock_mode;
1364 policies[1] = p1_policy;
1365 policies[0] = p2_policy;
1368 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1369 res_id[0]->name[0], res_id[1]->name[0]);
1371 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1372 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1373 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1374 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1375 NULL, 0, NULL, handles[0]);
1378 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1380 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1381 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1382 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1383 ldlm_lock_addref(handles[1], lock_modes[1]);
1384 } else if (res_id[1]->name[0] != 0) {
1385 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1386 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1387 *res_id[1], LDLM_IBITS, policies[1],
1388 lock_modes[1], &flags, mds_blocking_ast,
1389 ldlm_completion_ast, NULL, NULL, NULL, 0,
1391 if (rc != ELDLM_OK) {
1392 ldlm_lock_decref(handles[0], lock_modes[0]);
1395 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1401 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1402 struct lustre_handle *p1_lockh, int p1_lock_mode,
1403 ldlm_policy_data_t *p1_policy,
1404 struct ldlm_res_id *p2_res_id,
1405 struct lustre_handle *p2_lockh, int p2_lock_mode,
1406 ldlm_policy_data_t *p2_policy,
1407 struct ldlm_res_id *c1_res_id,
1408 struct lustre_handle *c1_lockh, int c1_lock_mode,
1409 ldlm_policy_data_t *c1_policy,
1410 struct ldlm_res_id *c2_res_id,
1411 struct lustre_handle *c2_lockh, int c2_lock_mode,
1412 ldlm_policy_data_t *c2_policy)
1414 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1415 c1_res_id, c2_res_id };
1416 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1417 c1_lockh, c2_lockh };
1418 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1419 c1_lock_mode, c2_lock_mode };
1420 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1421 c1_policy, c2_policy};
1422 int rc, i, j, sorted, flags;
1425 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1426 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1427 res_id[3]->name[0]);
1430 * simple insertion sort - we have at most 4 elements. Note, that zeroed
1431 * res_id should be at the end of list after sorting is finished.
1433 for (i = 1; i < 4; i++) {
1435 dlm_handles[4] = dlm_handles[i];
1436 res_id[4] = res_id[i];
1437 lock_modes[4] = lock_modes[i];
1438 policies[4] = policies[i];
1442 if (res_gt(res_id[j], res_id[4], policies[j],
1444 dlm_handles[j + 1] = dlm_handles[j];
1445 res_id[j + 1] = res_id[j];
1446 lock_modes[j + 1] = lock_modes[j];
1447 policies[j + 1] = policies[j];
1452 } while (j >= 0 && !sorted);
1454 dlm_handles[j + 1] = dlm_handles[4];
1455 res_id[j + 1] = res_id[4];
1456 lock_modes[j + 1] = lock_modes[4];
1457 policies[j + 1] = policies[4];
1460 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1461 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1462 res_id[3]->name[0]);
1464 /* XXX we could send ASTs on all these locks first before blocking? */
1465 for (i = 0; i < 4; i++) {
1469 * nevertheless zeroed res_ids should be at the end of list, and
1470 * could use break here, I think, that it is more correctly for
1471 * clear understanding of code to have continue here, as it
1472 * clearly means, that zeroed res_id should be skipped and does
1473 * not mean, that if we meet zeroed res_id we should stop
1476 if (res_id[i]->name[0] == 0)
1480 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1481 (policies[i]->l_inodebits.bits &
1482 policies[i-1]->l_inodebits.bits) ) {
1483 memcpy(dlm_handles[i], dlm_handles[i-1],
1484 sizeof(*(dlm_handles[i])));
1485 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1487 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1488 *res_id[i], LDLM_IBITS,
1490 lock_modes[i], &flags,
1492 ldlm_completion_ast, NULL, NULL,
1493 NULL, 0, NULL, dlm_handles[i]);
1495 GOTO(out_err, rc = -EIO);
1496 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1503 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1508 /* In the unlikely case that the child changed while we were waiting
1509 * on the lock, we need to drop the lock on the old child and either:
1510 * - if the child has a lower resource name, then we have to also
1511 * drop the parent lock and regain the locks in the right order
1512 * - in the rename case, if the child has a lower resource name than one of
1513 * the other parent/child resources (maxres) we also need to reget the locks
1514 * - if the child has a higher resource name (this is the common case)
1515 * we can just get the lock on the new child (still in lock order)
1517 * Returns 0 if the child did not change or if it changed but could be locked.
1518 * Returns 1 if the child changed and we need to re-lock (no locks held).
1519 * Returns -ve error with a valid dchild (no locks held). */
1520 static int mds_verify_child(struct obd_device *obd,
1521 struct ldlm_res_id *parent_res_id,
1522 struct lustre_handle *parent_lockh,
1523 struct dentry *dparent, int parent_mode,
1524 struct ldlm_res_id *child_res_id,
1525 struct lustre_handle *child_lockh,
1526 struct dentry **dchildp, int child_mode,
1527 ldlm_policy_data_t *child_policy,
1528 const char *name, int namelen,
1529 struct ldlm_res_id *maxres,
1530 unsigned long child_ino, __u32 child_gen)
1532 struct lustre_id sid;
1533 struct dentry *vchild, *dchild = *dchildp;
1534 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1537 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1539 GOTO(cleanup, rc = PTR_ERR(vchild));
1541 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1542 if (child_gen == vchild->d_generation &&
1543 child_ino == vchild->d_inum) {
1552 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1553 (vchild->d_inode != NULL &&
1554 child_gen == vchild->d_inode->i_generation &&
1555 child_ino == vchild->d_inode->i_ino))) {
1563 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1564 vchild->d_inode, dchild ? dchild->d_inode : 0,
1565 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1566 child_res_id->name[0]);
1568 if (child_res_id->name[0] != 0)
1569 ldlm_lock_decref(child_lockh, child_mode);
1573 cleanup_phase = 1; /* parent lock only */
1574 *dchildp = dchild = vchild;
1576 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1577 int flags = LDLM_FL_ATOMIC_CB;
1579 if (dchild->d_inode) {
1580 down(&dchild->d_inode->i_sem);
1581 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1582 up(&dchild->d_inode->i_sem);
1584 CERROR("Can't read inode self id, inode %lu,"
1585 " rc %d\n", dchild->d_inode->i_ino, rc);
1588 child_res_id->name[0] = id_fid(&sid);
1589 child_res_id->name[1] = id_group(&sid);
1591 child_res_id->name[0] = dchild->d_fid;
1592 child_res_id->name[1] = dchild->d_mdsnum;
1595 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1596 res_gt(maxres, child_res_id, NULL, NULL)) {
1597 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1598 child_res_id->name[0], parent_res_id->name[0],
1600 GOTO(cleanup, rc = 1);
1603 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1604 *child_res_id, LDLM_IBITS, child_policy,
1605 child_mode, &flags, mds_blocking_ast,
1606 ldlm_completion_ast, NULL, NULL, NULL, 0,
1609 GOTO(cleanup, rc = -EIO);
1612 memset(child_res_id, 0, sizeof(*child_res_id));
1618 switch(cleanup_phase) {
1620 if (child_res_id->name[0] != 0)
1621 ldlm_lock_decref(child_lockh, child_mode);
1623 ldlm_lock_decref(parent_lockh, parent_mode);
1629 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1630 struct lustre_id *id,
1631 struct lustre_handle *parent_lockh,
1632 struct dentry **dparentp, int parent_mode,
1633 __u64 parent_lockpart, int *update_mode,
1634 char *name, int namelen,
1635 struct lustre_handle *child_lockh,
1636 struct dentry **dchildp, int child_mode,
1637 __u64 child_lockpart)
1639 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1640 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1641 struct ldlm_res_id parent_res_id = { .name = {0} };
1642 struct ldlm_res_id child_res_id = { .name = {0} };
1643 unsigned long child_ino = 0; __u32 child_gen = 0;
1644 int rc = 0, cleanup_phase = 0;
1645 struct lustre_id sid;
1646 struct inode *inode;
1649 /* Step 1: Lookup parent */
1650 *dparentp = mds_id2dentry(obd, id, NULL);
1651 if (IS_ERR(*dparentp)) {
1652 rc = PTR_ERR(*dparentp);
1657 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1658 (*dparentp)->d_inode->i_ino, name);
1660 parent_res_id.name[0] = id_fid(id);
1661 parent_res_id.name[1] = id_group(id);
1664 parent_lockh[1].cookie = 0;
1665 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1666 struct ldlm_res_id res_id = { .name = {0} };
1667 ldlm_policy_data_t policy;
1668 int flags = LDLM_FL_ATOMIC_CB;
1670 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1672 res_id.name[0] = id_fid(id);
1673 res_id.name[1] = id_group(id);
1674 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1676 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1677 res_id, LDLM_IBITS, &policy,
1678 *update_mode, &flags,
1680 ldlm_completion_ast,
1681 NULL, NULL, NULL, 0, NULL,
1687 parent_res_id.name[2] = full_name_hash((unsigned char *)name,
1690 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1691 (unsigned long)id_fid(id), (unsigned long)id_group(id),
1692 parent_res_id.name[2]);
1696 cleanup_phase = 1; /* parent dentry */
1698 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1699 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1700 if (IS_ERR(*dchildp)) {
1701 rc = PTR_ERR(*dchildp);
1702 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1706 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1708 * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1709 * lock. Drop possible UPDATE lock!
1711 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1712 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1714 child_res_id.name[0] = (*dchildp)->d_fid;
1715 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1716 child_ino = (*dchildp)->d_inum;
1717 child_gen = (*dchildp)->d_generation;
1721 inode = (*dchildp)->d_inode;
1723 inode = igrab(inode);
1727 down(&inode->i_sem);
1728 rc = mds_read_inode_sid(obd, inode, &sid);
1731 CERROR("Can't read inode self id, inode %lu, "
1732 "rc %d\n", inode->i_ino, rc);
1737 child_ino = inode->i_ino;
1738 child_gen = inode->i_generation;
1739 child_res_id.name[0] = id_fid(&sid);
1740 child_res_id.name[1] = id_group(&sid);
1744 cleanup_phase = 2; /* child dentry */
1746 /* Step 3: Lock parent and child in resource order. If child doesn't
1747 * exist, we still have to lock the parent and re-lookup. */
1748 rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1749 &parent_policy, &child_res_id, child_lockh,
1750 child_mode, &child_policy);
1754 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1755 cleanup_phase = 4; /* child lock */
1757 cleanup_phase = 3; /* parent lock */
1759 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1760 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1761 parent_mode, &child_res_id, child_lockh,
1762 dchildp, child_mode, &child_policy,
1763 name, namelen, &parent_res_id, child_ino,
1775 switch (cleanup_phase) {
1777 ldlm_lock_decref(child_lockh, child_mode);
1779 ldlm_lock_decref(parent_lockh, parent_mode);
1784 if (parent_lockh[1].cookie)
1785 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1793 void mds_reconstruct_generic(struct ptlrpc_request *req)
1795 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1796 mds_req_from_mcd(req, med->med_mcd);
1799 /* If we are unlinking an open file/dir (i.e. creating an orphan) then we
1800 * instead link the inode into the PENDING directory until it is finally
1801 * released. We can't simply call mds_reint_rename() or some part thereof,
1802 * because we don't have the inode to check for link count/open status until
1803 * after it is locked.
1805 * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1806 * before starting journal transaction.
1808 * returns 1 on success
1809 * returns 0 if we lost a race and didn't make a new link
1810 * returns negative on error
1812 static int mds_orphan_add_link(struct mds_update_record *rec,
1813 struct obd_device *obd, struct dentry *dentry)
1815 struct mds_obd *mds = &obd->u.mds;
1816 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1817 struct inode *inode = dentry->d_inode;
1818 struct dentry *pending_child;
1819 char idname[LL_ID_NAMELEN];
1820 int idlen = 0, rc, mode;
1823 LASSERT(inode != NULL);
1824 LASSERT(!mds_inode_is_orphan(inode));
1825 #ifndef HAVE_I_ALLOC_SEM
1826 LASSERT(down_trylock(&inode->i_sem) != 0);
1828 LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1830 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1832 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1833 mds_orphan_open_count(inode), inode->i_nlink,
1834 S_ISDIR(inode->i_mode) ? "dir" :
1835 S_ISREG(inode->i_mode) ? "file" : "other",
1836 rec->ur_name, idname);
1838 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1841 pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1842 if (IS_ERR(pending_child))
1843 RETURN(PTR_ERR(pending_child));
1845 if (pending_child->d_inode != NULL) {
1846 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1847 LASSERT(pending_child->d_inode == inode);
1848 GOTO(out_dput, rc = 0);
1852 * link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG for
1853 * linking and return real mode back then -bzzz
1855 mode = inode->i_mode;
1856 inode->i_mode = S_IFREG;
1857 rc = vfs_link(dentry, pending_dir, pending_child);
1859 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1862 mds_inode_set_orphan(inode);
1864 /* return mode and correct i_nlink if inode is directory */
1865 inode->i_mode = mode;
1866 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1867 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1869 if (S_ISDIR(mode)) {
1871 i_nlink_inc(pending_dir);
1872 mark_inode_dirty(inode);
1873 mark_inode_dirty(pending_dir);
1876 GOTO(out_dput, rc = 1);
1878 l_dput(pending_child);
1882 int mds_create_local_dentry(struct mds_update_record *rec,
1883 struct obd_device *obd)
1885 struct mds_obd *mds = &obd->u.mds;
1886 struct inode *id_dir = mds->mds_id_dir->d_inode;
1887 int idlen = 0, rc, cleanup_phase = 0;
1888 struct dentry *new_child = NULL;
1889 char *idname = rec->ur_name;
1890 struct dentry *child = NULL;
1891 struct lustre_handle lockh[2] = {{0}, {0}};
1892 struct lustre_id sid;
1896 down(&id_dir->i_sem);
1897 idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1898 id_gen(rec->ur_id1));
1900 CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1901 idname, OLID4(rec->ur_id1));
1903 new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1906 if (IS_ERR(new_child)) {
1907 CERROR("can't lookup %s: %d\n", idname,
1908 (int) PTR_ERR(new_child));
1909 GOTO(cleanup, rc = PTR_ERR(new_child));
1913 down(&id_dir->i_sem);
1914 rc = mds_read_inode_sid(obd, id_dir, &sid);
1917 CERROR("Can't read inode self id, inode %lu, "
1918 "rc %d\n", id_dir->i_ino, rc);
1922 if (new_child->d_inode != NULL) {
1923 /* nice. we've already have local dentry! */
1924 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1925 (unsigned)new_child->d_inode->i_ino,
1926 (unsigned)new_child->d_inode->i_generation);
1928 id_ino(rec->ur_id1) = id_dir->i_ino;
1929 id_gen(rec->ur_id1) = id_dir->i_generation;
1930 rec->ur_namelen = idlen + 1;
1932 id_fid(rec->ur_id1) = id_fid(&sid);
1933 id_group(rec->ur_id1) = id_group(&sid);
1935 GOTO(cleanup, rc = 0);
1938 /* new, local dentry will be added soon. we need no aliases here */
1941 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1942 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1944 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1945 LCK_EX, lockh, NULL, NULL, 0,
1946 MDS_INODELOCK_UPDATE);
1949 if (IS_ERR(child)) {
1950 rc = PTR_ERR(child);
1951 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1952 CERROR("can't get victim: %d\n", rc);
1957 handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1959 GOTO(cleanup, rc = PTR_ERR(handle));
1961 rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
1962 idlen, id_ino(rec->ur_id1),
1963 id_gen(rec->ur_id1), mds->mds_num,
1964 id_fid(rec->ur_id1));
1966 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1967 (unsigned long)child->d_inode->i_ino,
1968 (unsigned long)child->d_inode->i_generation, rc);
1970 if (S_ISDIR(child->d_inode->i_mode)) {
1971 i_nlink_inc(id_dir);
1972 mark_inode_dirty(id_dir);
1974 mark_inode_dirty(child->d_inode);
1976 fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
1978 id_ino(rec->ur_id1) = id_dir->i_ino;
1979 id_gen(rec->ur_id1) = id_dir->i_generation;
1980 rec->ur_namelen = idlen + 1;
1982 id_fid(rec->ur_id1) = id_fid(&sid);
1983 id_group(rec->ur_id1) = id_group(&sid);
1987 switch(cleanup_phase) {
1989 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1990 ldlm_lock_decref(lockh, LCK_EX);
2000 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
2001 struct ptlrpc_request *slave)
2003 void *cookie, *cookie2;
2004 struct mds_body *body2;
2005 struct mds_body *body;
2009 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
2010 LASSERT(body != NULL);
2012 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
2013 LASSERT(body2 != NULL);
2015 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
2018 memcpy(body2, body, sizeof(*body));
2019 body2->valid &= ~OBD_MD_FLCOOKIE;
2021 if (!(body->valid & OBD_MD_FLEASIZE) &&
2022 !(body->valid & OBD_MD_FLDIREA))
2025 if (body->eadatasize == 0) {
2026 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
2030 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
2032 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
2033 LASSERT(ea != NULL);
2035 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
2036 LASSERT(ea2 != NULL);
2038 memcpy(ea2, ea, body->eadatasize);
2040 if (body->valid & OBD_MD_FLCOOKIE) {
2041 LASSERT(master->rq_repmsg->buflens[2] >=
2042 slave->rq_repmsg->buflens[2]);
2043 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
2044 slave->rq_repmsg->buflens[2]);
2045 LASSERT(cookie != NULL);
2047 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
2048 master->rq_repmsg->buflens[2]);
2049 LASSERT(cookie2 != NULL);
2050 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
2051 body2->valid |= OBD_MD_FLCOOKIE;
2056 static int mds_reint_unlink_remote(struct mds_update_record *rec,
2057 int offset, struct ptlrpc_request *req,
2058 struct lustre_handle *parent_lockh,
2059 int update_mode, struct dentry *dparent,
2060 struct lustre_handle *child_lockh,
2061 struct dentry *dchild)
2063 struct obd_device *obd = req->rq_export->exp_obd;
2064 struct mds_obd *mds = mds_req2mds(req);
2065 struct ptlrpc_request *request = NULL;
2066 int rc = 0, cleanup_phase = 0;
2067 struct mdc_op_data *op_data;
2071 LASSERT(offset == 1 || offset == 3);
2073 /* time to drop i_nlink on remote MDS */
2074 OBD_ALLOC(op_data, sizeof(*op_data));
2075 if (op_data == NULL)
2078 memset(op_data, 0, sizeof(*op_data));
2079 mds_pack_dentry2id(obd, &op_data->id1, dchild, 1);
2080 op_data->create_mode = rec->ur_mode;
2082 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
2083 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2085 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2086 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
2087 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2090 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2091 op_data->create_mode |= MDS_MODE_REPLAY;
2093 rc = md_unlink(mds->mds_md_exp, op_data, &request);
2094 OBD_FREE(op_data, sizeof(*op_data));
2099 mds_copy_unlink_reply(req, request);
2100 ptlrpc_req_finished(request);
2104 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2107 GOTO(cleanup, rc = PTR_ERR(handle));
2108 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
2109 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
2114 req->rq_status = rc;
2117 if (parent_lockh[1].cookie != 0)
2118 ldlm_lock_decref(parent_lockh + 1, update_mode);
2120 ldlm_lock_decref(child_lockh, LCK_EX);
2122 ldlm_lock_decref(parent_lockh, LCK_PW);
2124 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2131 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
2132 struct ptlrpc_request *req,
2133 struct lustre_handle *lh)
2135 struct dentry *dparent = NULL, *dchild;
2136 struct mds_obd *mds = mds_req2mds(req);
2137 struct obd_device *obd = req->rq_export->exp_obd;
2138 struct mds_body *body = NULL;
2139 struct inode *child_inode = NULL;
2140 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2141 struct lustre_handle child_lockh = {0};
2142 struct lustre_handle child_reuse_lockh = {0};
2143 struct lustre_handle *slave_lockh = NULL;
2144 char idname[LL_ID_NAMELEN];
2145 struct llog_create_locks *lcl = NULL;
2146 void *handle = NULL;
2147 int rc = 0, cleanup_phase = 0;
2148 int unlink_by_id = 0;
2152 LASSERT(offset == 1 || offset == 3);
2154 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
2155 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2158 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2160 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2161 DEBUG_REQ(D_HA, req, "unlink replay");
2162 LASSERT(offset == 1); /* should not come from intent */
2163 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2164 lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
2165 req->rq_repmsg->buflens[2]);
2168 MD_COUNTER_INCREMENT(obd, unlink);
2170 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
2171 GOTO(cleanup, rc = -ENOENT);
2173 if (rec->ur_namelen == 1) {
2174 /* this is request to drop i_nlink on local inode */
2176 rec->ur_name = idname;
2177 rc = mds_create_local_dentry(rec, obd);
2178 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
2179 DEBUG_REQ(D_HA, req,
2180 "drop nlink on inode "DLID4" (replay)",
2181 OLID4(rec->ur_id1));
2187 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
2188 /* master mds for directory asks slave removing inode is already
2190 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
2191 LCK_PW, parent_lockh,
2192 &update_mode, rec->ur_name,
2194 MDS_INODELOCK_UPDATE);
2195 if (IS_ERR(dparent))
2196 GOTO(cleanup, rc = PTR_ERR(dparent));
2197 dchild = ll_lookup_one_len(rec->ur_name, dparent,
2198 rec->ur_namelen - 1);
2200 GOTO(cleanup, rc = PTR_ERR(dchild));
2201 child_lockh.cookie = 0;
2202 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
2203 LASSERT(dchild->d_inode != NULL);
2204 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
2206 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
2207 parent_lockh, &dparent,
2208 LCK_PW, MDS_INODELOCK_UPDATE,
2209 &update_mode, rec->ur_name,
2210 rec->ur_namelen, &child_lockh,
2212 MDS_INODELOCK_LOOKUP |
2213 MDS_INODELOCK_UPDATE);
2218 if (dchild->d_flags & DCACHE_CROSS_REF) {
2219 /* we should have parent lock only here */
2220 LASSERT(unlink_by_id == 0);
2221 LASSERT(dchild->d_mdsnum != mds->mds_num);
2222 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
2223 update_mode, dparent, &child_lockh, dchild);
2227 cleanup_phase = 1; /* dchild, dparent, locks */
2230 child_inode = dchild->d_inode;
2231 if (child_inode == NULL) {
2232 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
2233 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
2234 GOTO(cleanup, rc = -ENOENT);
2237 cleanup_phase = 2; /* dchild has a lock */
2239 /* We have to do these checks ourselves, in case we are making an
2240 * orphan. The client tells us whether rmdir() or unlink() was called,
2241 * so we need to return appropriate errors (bug 72).
2243 * We don't have to check permissions, because vfs_rename (called from
2244 * mds_open_unlink_rename) also calls may_delete. */
2245 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
2246 if (!S_ISDIR(child_inode->i_mode))
2247 GOTO(cleanup, rc = -ENOTDIR);
2249 if (S_ISDIR(child_inode->i_mode))
2250 GOTO(cleanup, rc = -EISDIR);
2253 /* handle splitted dir */
2254 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
2258 /* Step 4: Get a lock on the ino to sync with creation WRT inode
2259 * reuse (see bug 2029). */
2260 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
2263 cleanup_phase = 3; /* child inum lock */
2265 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2267 /* ldlm_reply in buf[0] if called via intent */
2273 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2274 LASSERT(body != NULL);
2276 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2277 DOWN_READ_I_ALLOC_SEM(child_inode);
2278 cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2280 /* If this is potentially the last reference to this inode, get the
2281 * OBD EA data first so the client can destroy OST objects. We
2282 * only do the object removal later if no open files/links remain. */
2283 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2284 child_inode->i_nlink == 1) {
2285 if (mds_orphan_open_count(child_inode) > 0) {
2286 /* need to lock pending_dir before transaction */
2287 down(&mds->mds_pending_dir->d_inode->i_sem);
2288 cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2289 } else if (S_ISREG(child_inode->i_mode)) {
2290 mds_pack_inode2body(obd, body, child_inode, 0);
2291 mds_pack_md(obd, req->rq_repmsg, offset + 1,
2292 body, child_inode, MDS_PACK_MD_LOCK, 0);
2296 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2297 switch (child_inode->i_mode & S_IFMT) {
2299 /* Drop any lingering child directories before we start our
2300 * transaction, to avoid doing multiple inode dirty/delete
2301 * in our compound transaction (bug 1321). */
2302 shrink_dcache_parent(dchild);
2303 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2306 GOTO(cleanup, rc = PTR_ERR(handle));
2307 rc = vfs_rmdir(dparent->d_inode, dchild);
2310 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2311 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2313 handle = fsfilt_start_log(obd, dparent->d_inode,
2314 FSFILT_OP_UNLINK, NULL,
2315 le32_to_cpu(lmm->lmm_stripe_count));
2317 GOTO(cleanup, rc = PTR_ERR(handle));
2318 rc = vfs_unlink(dparent->d_inode, dchild);
2326 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2329 GOTO(cleanup, rc = PTR_ERR(handle));
2330 rc = vfs_unlink(dparent->d_inode, dchild);
2333 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2336 GOTO(cleanup, rc = -EINVAL);
2339 if (rc == 0 && child_inode->i_nlink == 0) {
2340 if (mds_orphan_open_count(child_inode) > 0)
2341 rc = mds_orphan_add_link(rec, obd, dchild);
2344 GOTO(cleanup, rc = 0);
2346 if (!S_ISREG(child_inode->i_mode))
2349 if (!(body->valid & OBD_MD_FLEASIZE)) {
2350 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2351 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2352 } else if (mds_log_op_unlink(obd, child_inode,
2353 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2354 req->rq_repmsg->buflens[offset + 1],
2355 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2356 req->rq_repmsg->buflens[offset + 2],
2358 body->valid |= OBD_MD_FLCOOKIE;
2361 rc = mds_destroy_object(obd, child_inode, 1);
2363 CERROR("can't remove OST object, err %d\n",
2375 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2376 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2377 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2379 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2381 CERROR("error on parent setattr: rc = %d\n", err);
2383 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2384 handle, req, rc, 0);
2386 (void)obd_set_info(mds->mds_dt_exp, strlen("unlinked"),
2387 "unlinked", 0, NULL);
2388 switch(cleanup_phase) {
2389 case 5: /* pending_dir semaphore */
2390 up(&mds->mds_pending_dir->d_inode->i_sem);
2391 case 4: /* child inode semaphore */
2392 UP_READ_I_ALLOC_SEM(child_inode);
2393 /* handle splitted dir */
2395 /* master directory can be non-empty or something else ... */
2396 mds_unlink_slave_objs(obd, dchild);
2399 ptlrpc_save_llog_lock(req, lcl);
2400 case 3: /* child ino-reuse lock */
2401 if (rc && body != NULL) {
2402 // Don't unlink the OST objects if the MDS unlink failed
2406 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2408 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2409 case 2: /* child lock */
2410 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2411 if (child_lockh.cookie)
2412 ldlm_lock_decref(&child_lockh, LCK_EX);
2413 case 1: /* child and parent dentry, parent lock */
2415 if (parent_lockh[1].cookie != 0)
2416 ldlm_lock_decref(parent_lockh + 1, update_mode);
2419 ldlm_lock_decref(parent_lockh, LCK_PW);
2421 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2428 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2431 req->rq_status = rc;
2436 * to service requests from remote MDS to increment i_nlink
2438 static int mds_reint_link_acquire(struct mds_update_record *rec,
2439 int offset, struct ptlrpc_request *req,
2440 struct lustre_handle *lh)
2442 struct obd_device *obd = req->rq_export->exp_obd;
2443 struct ldlm_res_id src_res_id = { .name = {0} };
2444 struct lustre_handle *handle = NULL, src_lockh = {0};
2445 struct mds_obd *mds = mds_req2mds(req);
2446 int rc = 0, cleanup_phase = 0;
2447 struct dentry *de_src = NULL;
2448 ldlm_policy_data_t policy;
2449 int flags = LDLM_FL_ATOMIC_CB;
2452 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2453 obd->obd_name, OLID4(rec->ur_id1));
2455 /* Step 1: Lookup the source inode and target directory by ID */
2456 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2458 GOTO(cleanup, rc = PTR_ERR(de_src));
2459 cleanup_phase = 1; /* source dentry */
2461 src_res_id.name[0] = id_fid(rec->ur_id1);
2462 src_res_id.name[1] = id_group(rec->ur_id1);
2463 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2465 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2466 src_res_id, LDLM_IBITS, &policy,
2467 LCK_EX, &flags, mds_blocking_ast,
2468 ldlm_completion_ast, NULL, NULL,
2469 NULL, 0, NULL, &src_lockh);
2471 GOTO(cleanup, rc = -ENOLCK);
2472 cleanup_phase = 2; /* lock */
2474 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2476 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2477 if (IS_ERR(handle)) {
2478 rc = PTR_ERR(handle);
2481 i_nlink_inc(de_src->d_inode);
2482 mark_inode_dirty(de_src->d_inode);
2486 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2487 handle, req, rc, 0);
2488 switch (cleanup_phase) {
2491 ldlm_lock_decref(&src_lockh, LCK_EX);
2493 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2499 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2502 req->rq_status = rc;
2507 * request to link to foreign inode:
2508 * - acquire i_nlinks on this inode
2511 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2512 int offset, struct ptlrpc_request *req,
2513 struct lustre_handle *lh)
2515 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2516 struct obd_device *obd = req->rq_export->exp_obd;
2517 struct dentry *de_tgt_dir = NULL;
2518 struct mds_obd *mds = mds_req2mds(req);
2519 int rc = 0, cleanup_phase = 0;
2520 struct mdc_op_data *op_data;
2521 struct ptlrpc_request *request = NULL;
2525 DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2526 ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2527 OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2528 OLID4(rec->ur_id1));
2530 de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2531 tgt_dir_lockh, &update_mode,
2532 rec->ur_name, rec->ur_namelen - 1,
2533 MDS_INODELOCK_UPDATE);
2534 if (IS_ERR(de_tgt_dir))
2535 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2538 OBD_ALLOC(op_data, sizeof(*op_data));
2539 if (op_data == NULL)
2540 GOTO(cleanup, rc = -ENOMEM);
2542 memset(op_data, 0, sizeof(*op_data));
2543 op_data->id1 = *(rec->ur_id1);
2544 rc = md_link(mds->mds_md_exp, op_data, &request);
2545 OBD_FREE(op_data, sizeof(*op_data));
2548 ptlrpc_req_finished(request);
2554 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2556 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2557 if (IS_ERR(handle)) {
2558 rc = PTR_ERR(handle);
2564 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2565 rec->ur_namelen - 1, id_ino(rec->ur_id1),
2566 id_gen(rec->ur_id1), id_group(rec->ur_id1),
2567 id_fid(rec->ur_id1));
2570 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2571 handle, req, rc, 0);
2573 switch (cleanup_phase) {
2576 OBD_ALLOC(op_data, sizeof(*op_data));
2577 if (op_data != NULL) {
2579 memset(op_data, 0, sizeof(*op_data));
2581 op_data->id1 = *(rec->ur_id1);
2582 op_data->create_mode = rec->ur_mode;
2584 rc = md_unlink(mds->mds_md_exp, op_data, &request);
2585 OBD_FREE(op_data, sizeof(*op_data));
2587 ptlrpc_req_finished(request);
2589 CERROR("error %d while dropping i_nlink on "
2590 "remote inode\n", rc);
2593 CERROR("rc %d prevented dropping i_nlink on "
2594 "remote inode\n", -ENOMEM);
2600 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2602 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2605 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2607 ptlrpc_save_lock(req, tgt_dir_lockh + 1, update_mode);
2613 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2616 req->rq_status = rc;
2620 static int mds_reint_link(struct mds_update_record *rec, int offset,
2621 struct ptlrpc_request *req,
2622 struct lustre_handle *lh)
2624 struct obd_device *obd = req->rq_export->exp_obd;
2625 struct dentry *de_src = NULL;
2626 struct dentry *de_tgt_dir = NULL;
2627 struct dentry *dchild = NULL;
2628 struct mds_obd *mds = mds_req2mds(req);
2629 struct lustre_handle *handle = NULL;
2630 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2631 struct ldlm_res_id src_res_id = { .name = {0} };
2632 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2633 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2634 ldlm_policy_data_t tgt_dir_policy =
2635 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2636 int rc = 0, cleanup_phase = 0;
2638 int update_mode = 0;
2642 LASSERT(offset == 1);
2644 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2645 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2646 id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2649 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2650 MD_COUNTER_INCREMENT(obd, link);
2652 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2653 GOTO(cleanup, rc = -ENOENT);
2655 if (id_group(rec->ur_id1) != mds->mds_num) {
2656 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2660 if (rec->ur_namelen == 1) {
2661 rc = mds_reint_link_acquire(rec, offset, req, lh);
2665 /* Step 1: Lookup the source inode and target directory by ID */
2666 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2668 GOTO(cleanup, rc = PTR_ERR(de_src));
2670 cleanup_phase = 1; /* source dentry */
2672 de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2673 if (IS_ERR(de_tgt_dir)) {
2674 rc = PTR_ERR(de_tgt_dir);
2679 cleanup_phase = 2; /* target directory dentry */
2681 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2682 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2683 rec->ur_name, de_src->d_inode->i_ino);
2685 /* Step 2: Take the two locks */
2686 src_res_id.name[0] = id_fid(rec->ur_id1);
2687 src_res_id.name[1] = id_group(rec->ur_id1);
2688 tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2689 tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2692 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2693 int flags = LDLM_FL_ATOMIC_CB;
2694 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2696 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2697 tgt_dir_res_id, LDLM_IBITS,
2698 &src_policy, update_mode, &flags,
2700 ldlm_completion_ast, NULL, NULL,
2701 NULL, 0, NULL, tgt_dir_lockh + 1);
2703 GOTO(cleanup, rc = -ENOLCK);
2706 tgt_dir_res_id.name[2] = full_name_hash((unsigned char *)rec->ur_name,
2707 rec->ur_namelen - 1);
2708 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2709 (unsigned long)id_fid(rec->ur_id2),
2710 (unsigned long)id_group(rec->ur_id2),
2711 tgt_dir_res_id.name[2]);
2714 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2715 &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2716 LCK_EX, &tgt_dir_policy);
2720 cleanup_phase = 3; /* locks */
2722 /* Step 3: Lookup the child */
2723 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2724 rec->ur_namelen - 1);
2725 if (IS_ERR(dchild)) {
2726 rc = PTR_ERR(dchild);
2727 if (rc != -EPERM && rc != -EACCES)
2728 CERROR("child lookup error %d\n", rc);
2732 cleanup_phase = 4; /* child dentry */
2734 if (dchild->d_inode) {
2735 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2736 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2741 /* Step 4: Do it. */
2742 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2744 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2745 if (IS_ERR(handle)) {
2746 rc = PTR_ERR(handle);
2750 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2751 if (rc && rc != -EPERM && rc != -EACCES)
2752 CERROR("vfs_link error %d\n", rc);
2754 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2755 handle, req, rc, 0);
2758 switch (cleanup_phase) {
2759 case 4: /* child dentry */
2763 ldlm_lock_decref(&src_lockh, LCK_EX);
2764 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2766 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2767 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2769 case 2: /* target dentry */
2771 if (tgt_dir_lockh[1].cookie && update_mode)
2772 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2776 case 1: /* source dentry */
2781 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2784 req->rq_status = rc;
2788 /* The idea here is that we need to get four locks in the end:
2789 * one on each parent directory, one on each child. We need to take
2790 * these locks in some kind of order (to avoid deadlocks), and the order
2791 * I selected is "increasing resource number" order. We need to look up
2792 * the children, however, before we know what the resource number(s) are.
2793 * Thus the following plan:
2795 * 1,2. Look up the parents
2796 * 3,4. Look up the children
2797 * 5. Take locks on the parents and children, in order
2798 * 6. Verify that the children haven't changed since they were looked up
2800 * If there was a race and the children changed since they were first looked
2801 * up, it is possible that mds_verify_child() will be able to just grab the
2802 * lock on the new child resource (if it has a higher resource than any other)
2803 * but we need to compare against not only its parent, but also against the
2804 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2806 * We need the fancy igrab() on the child inodes because we aren't holding a
2807 * lock on the parent after the lookup is done, so dentry->d_inode may change
2808 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2810 static int mds_get_parents_children_locked(struct obd_device *obd,
2811 struct mds_obd *mds,
2812 struct lustre_id *p1_id,
2813 struct dentry **de_srcdirp,
2814 struct lustre_id *p2_id,
2815 struct dentry **de_tgtdirp,
2817 const char *old_name, int old_len,
2818 struct dentry **de_oldp,
2819 const char *new_name, int new_len,
2820 struct dentry **de_newp,
2821 struct lustre_handle *dlm_handles,
2824 struct ldlm_res_id p1_res_id = { .name = {0} };
2825 struct ldlm_res_id p2_res_id = { .name = {0} };
2826 struct ldlm_res_id c1_res_id = { .name = {0} };
2827 struct ldlm_res_id c2_res_id = { .name = {0} };
2828 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2829 /* Only dentry should disappear, but the inode itself would be
2830 intact otherwise. */
2831 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2832 /* If something is going to be replaced, both dentry and inode locks are
2834 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2835 struct ldlm_res_id *maxres_src, *maxres_tgt;
2836 struct inode *inode;
2837 int rc = 0, cleanup_phase = 0;
2838 __u32 child_gen1 = 0;
2839 __u32 child_gen2 = 0;
2840 unsigned long child_ino1 = 0;
2841 unsigned long child_ino2 = 0;
2844 /* Step 1: Lookup the source directory */
2845 *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2846 if (IS_ERR(*de_srcdirp))
2847 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2849 cleanup_phase = 1; /* source directory dentry */
2851 p1_res_id.name[0] = id_fid(p1_id);
2852 p1_res_id.name[1] = id_group(p1_id);
2854 /* Step 2: Lookup the target directory */
2855 if (id_equal_stc(p1_id, p2_id)) {
2856 *de_tgtdirp = dget(*de_srcdirp);
2858 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2859 if (IS_ERR(*de_tgtdirp)) {
2860 rc = PTR_ERR(*de_tgtdirp);
2866 cleanup_phase = 2; /* target directory dentry */
2868 p2_res_id.name[0] = id_fid(p2_id);
2869 p2_res_id.name[1] = id_group(p2_id);
2872 dlm_handles[5].cookie = 0;
2873 dlm_handles[6].cookie = 0;
2875 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2877 * get a temp lock on just fid, group to flush client cache and
2878 * to protect dirs from concurrent splitting.
2880 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2881 LCK_PW, &p_policy, &p2_res_id,
2882 &dlm_handles[6], LCK_PW, &p_policy);
2886 p1_res_id.name[2] = full_name_hash((unsigned char *)old_name,
2888 p2_res_id.name[2] = full_name_hash((unsigned char *)new_name,
2891 CDEBUG(D_INFO, "take locks on "
2892 LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2893 p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2894 p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2899 /* Step 3: Lookup the source child entry */
2900 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2902 if (IS_ERR(*de_oldp)) {
2903 rc = PTR_ERR(*de_oldp);
2904 CERROR("old child lookup error (%.*s): %d\n",
2905 old_len - 1, old_name, rc);
2909 cleanup_phase = 4; /* original name dentry */
2911 inode = (*de_oldp)->d_inode;
2912 if (inode != NULL) {
2913 struct lustre_id sid;
2915 inode = igrab(inode);
2917 GOTO(cleanup, rc = -ENOENT);
2919 down(&inode->i_sem);
2920 rc = mds_read_inode_sid(obd, inode, &sid);
2923 CERROR("Can't read inode self id, inode %lu, "
2924 "rc %d\n", inode->i_ino, rc);
2929 child_ino1 = inode->i_ino;
2930 child_gen1 = inode->i_generation;
2931 c1_res_id.name[0] = id_fid(&sid);
2932 c1_res_id.name[1] = id_group(&sid);
2934 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2935 child_ino1 = (*de_oldp)->d_inum;
2936 child_gen1 = (*de_oldp)->d_generation;
2937 c1_res_id.name[0] = (*de_oldp)->d_fid;
2938 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2940 GOTO(cleanup, rc = -ENOENT);
2943 /* Step 4: Lookup the target child entry */
2944 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2946 if (IS_ERR(*de_newp)) {
2947 rc = PTR_ERR(*de_newp);
2948 CERROR("new child lookup error (%.*s): %d\n",
2949 old_len - 1, old_name, rc);
2953 cleanup_phase = 5; /* target dentry */
2955 inode = (*de_newp)->d_inode;
2956 if (inode != NULL) {
2957 struct lustre_id sid;
2959 inode = igrab(inode);
2963 down(&inode->i_sem);
2964 rc = mds_read_inode_sid(obd, inode, &sid);
2967 CERROR("Can't read inode self id, inode %lu, "
2968 "rc %d\n", inode->i_ino, rc);
2972 child_ino2 = inode->i_ino;
2973 child_gen2 = inode->i_generation;
2974 c2_res_id.name[0] = id_fid(&sid);
2975 c2_res_id.name[1] = id_group(&sid);
2977 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2978 child_ino2 = (*de_newp)->d_inum;
2979 child_gen2 = (*de_newp)->d_generation;
2980 c2_res_id.name[0] = (*de_newp)->d_fid;
2981 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
2985 /* Step 5: Take locks on the parents and child(ren) */
2986 maxres_src = &p1_res_id;
2987 maxres_tgt = &p2_res_id;
2988 cleanup_phase = 5; /* target dentry */
2990 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
2991 maxres_src = &c1_res_id;
2992 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
2993 maxres_tgt = &c2_res_id;
2995 rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
2997 &p2_res_id, &dlm_handles[1], parent_mode,
2999 &c1_res_id, &dlm_handles[2], child_mode,
3001 &c2_res_id, &dlm_handles[3], child_mode,
3006 cleanup_phase = 6; /* parent and child(ren) locks */
3008 /* Step 6a: Re-lookup source child to verify it hasn't changed */
3009 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
3010 parent_mode, &c1_res_id, &dlm_handles[2],
3011 de_oldp, child_mode, &c1_policy, old_name, old_len,
3012 maxres_tgt, child_ino1, child_gen1);
3014 if (c2_res_id.name[0] != 0)
3015 ldlm_lock_decref(&dlm_handles[3], child_mode);
3016 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3023 if (!DENTRY_VALID(*de_oldp))
3024 GOTO(cleanup, rc = -ENOENT);
3026 /* Step 6b: Re-lookup target child to verify it hasn't changed */
3027 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
3028 parent_mode, &c2_res_id, &dlm_handles[3],
3029 de_newp, child_mode, &c2_policy, new_name,
3030 new_len, maxres_src, child_ino2, child_gen2);
3032 ldlm_lock_decref(&dlm_handles[2], child_mode);
3033 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3043 switch (cleanup_phase) {
3044 case 6: /* child lock(s) */
3045 if (c2_res_id.name[0] != 0)
3046 ldlm_lock_decref(&dlm_handles[3], child_mode);
3047 if (c1_res_id.name[0] != 0)
3048 ldlm_lock_decref(&dlm_handles[2], child_mode);
3049 if (dlm_handles[1].cookie != 0)
3050 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3051 if (dlm_handles[0].cookie != 0)
3052 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3053 case 5: /* target dentry */
3055 case 4: /* source dentry */
3059 if (dlm_handles[5].cookie != 0)
3060 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3061 if (dlm_handles[6].cookie != 0)
3062 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3064 case 2: /* target directory dentry */
3065 l_dput(*de_tgtdirp);
3066 case 1: /* source directry dentry */
3067 l_dput(*de_srcdirp);
3075 * checks if dentry can be removed. This function also handles cross-ref
3078 static int mds_check_for_rename(struct obd_device *obd,
3079 struct dentry *dentry)
3081 struct mds_obd *mds = &obd->u.mds;
3082 struct lustre_handle *rlockh;
3083 struct ptlrpc_request *req;
3084 struct mdc_op_data *op_data;
3085 struct lookup_intent it;
3086 int handle_size, rc = 0;
3089 LASSERT(dentry != NULL);
3091 if (dentry->d_inode) {
3092 if (S_ISDIR(dentry->d_inode->i_mode) &&
3093 !mds_is_dir_empty(obd, dentry))
3096 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
3097 handle_size = sizeof(struct lustre_handle);
3099 OBD_ALLOC(rlockh, handle_size);
3103 memset(rlockh, 0, handle_size);
3104 OBD_ALLOC(op_data, sizeof(*op_data));
3105 if (op_data == NULL) {
3106 OBD_FREE(rlockh, handle_size);
3109 memset(op_data, 0, sizeof(*op_data));
3110 mds_pack_dentry2id(obd, &op_data->id1, dentry, 1);
3112 it.it_op = IT_UNLINK;
3113 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
3116 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
3117 op_data, rlockh, NULL, 0, ldlm_completion_ast,
3118 mds_blocking_ast, NULL);
3119 OBD_FREE(op_data, sizeof(*op_data));
3123 OBD_FREE(it.d.fs_data,
3124 sizeof(struct lustre_intent_data));
3127 if (rlockh->cookie != 0)
3128 ldlm_lock_decref(rlockh, LCK_EX);
3130 if (LUSTRE_IT(&it)->it_data) {
3131 req = (struct ptlrpc_request *)LUSTRE_IT(&it)->it_data;
3132 ptlrpc_req_finished(req);
3135 if (LUSTRE_IT(&it)->it_status)
3136 rc = LUSTRE_IT(&it)->it_status;
3137 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
3138 OBD_FREE(rlockh, handle_size);
3143 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
3144 struct ptlrpc_request *req, struct lustre_id *id,
3145 struct dentry *de_dir, struct dentry *de)
3147 struct obd_device *obd = req->rq_export->exp_obd;
3148 struct mds_obd *mds = mds_req2mds(req);
3149 void *handle = NULL;
3155 * name exists and points to local inode try to unlink this name
3156 * and create new one.
3158 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
3159 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
3160 (unsigned long)de->d_inode->i_generation);
3162 /* checking if we can remove local dentry. */
3163 rc = mds_check_for_rename(obd, de);
3167 handle = fsfilt_start(obd, de_dir->d_inode,
3168 FSFILT_OP_RENAME, NULL);
3170 GOTO(cleanup, rc = PTR_ERR(handle));
3171 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3174 } else if (de->d_flags & DCACHE_CROSS_REF) {
3175 CDEBUG(D_OTHER, "%s: %s points to remote inode %lu/%lu\n",
3176 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_mdsnum,
3177 (unsigned long)de->d_fid);
3179 /* checking if we can remove local dentry. */
3180 rc = mds_check_for_rename(obd, de);
3185 * to be fully POSIX compatible, we should add one more check:
3187 * if de_new is subdir of dir rec->ur_id1. If so - return
3190 * I do not know how to implement it right now, because
3191 * inodes/dentries for new and old names lie on different MDS,
3192 * so add this notice here just to make it visible for the rest
3193 * of developers and do not forget about. And when this check
3194 * will be added, del_cross_ref should gone, that is local
3195 * dentry is able to be removed if all checks passed. --umka
3198 handle = fsfilt_start(obd, de_dir->d_inode,
3199 FSFILT_OP_RENAME, NULL);
3201 GOTO(cleanup, rc = PTR_ERR(handle));
3202 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3206 /* name doesn't exist. the simplest case. */
3207 handle = fsfilt_start(obd, de_dir->d_inode,
3208 FSFILT_OP_LINK, NULL);
3210 GOTO(cleanup, rc = PTR_ERR(handle));
3213 rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
3214 rec->ur_tgtlen - 1, id_ino(id),
3215 id_gen(id), id_group(id), id_fid(id));
3217 CERROR("add_dir_entry() returned error %d\n", rc);
3223 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3224 handle, req, rc, 0);
3229 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
3230 struct ptlrpc_request *req, struct dentry *de_dir,
3233 struct obd_device *obd = req->rq_export->exp_obd;
3234 struct mds_obd *mds = mds_req2mds(req);
3235 void *handle = NULL;
3239 handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
3241 GOTO(cleanup, rc = PTR_ERR(handle));
3242 rc = fsfilt_del_dir_entry(obd, de);
3247 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3248 handle, req, rc, 0);
3252 static int mds_reint_rename_create_name(struct mds_update_record *rec,
3253 int offset, struct ptlrpc_request *req)
3255 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3256 struct obd_device *obd = req->rq_export->exp_obd;
3257 struct mds_obd *mds = mds_req2mds(req);
3258 struct lustre_handle child_lockh = {0};
3259 struct dentry *de_tgtdir = NULL;
3260 struct dentry *de_new = NULL;
3261 int cleanup_phase = 0;
3262 int update_mode, rc = 0;
3266 * another MDS executing rename operation has asked us to create target
3267 * name. such a creation should destroy existing target name.
3269 CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
3270 obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
3272 /* first, lookup the target */
3273 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
3274 &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
3275 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
3276 &child_lockh, &de_new, LCK_EX,
3277 MDS_INODELOCK_LOOKUP);
3284 LASSERT(de_tgtdir->d_inode);
3287 rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
3293 if (cleanup_phase == 1) {
3295 if (parent_lockh[1].cookie != 0)
3296 ldlm_lock_decref(parent_lockh + 1, update_mode);
3298 ldlm_lock_decref(parent_lockh, LCK_PW);
3299 if (child_lockh.cookie != 0)
3300 ldlm_lock_decref(&child_lockh, LCK_EX);
3305 req->rq_status = rc;
3309 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3310 struct ptlrpc_request *req)
3312 struct obd_device *obd = req->rq_export->exp_obd;
3313 struct ptlrpc_request *req2 = NULL;
3314 struct dentry *de_srcdir = NULL;
3315 struct dentry *de_old = NULL;
3316 struct mds_obd *mds = mds_req2mds(req);
3317 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3318 struct lustre_handle child_lockh = {0};
3319 struct mdc_op_data *op_data;
3320 int update_mode, rc = 0;
3323 CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3324 obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3326 OBD_ALLOC(op_data, sizeof(*op_data));
3327 if (op_data == NULL)
3329 memset(op_data, 0, sizeof(*op_data));
3331 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3332 &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3333 &update_mode, rec->ur_name,
3334 rec->ur_namelen, &child_lockh, &de_old,
3335 LCK_EX, MDS_INODELOCK_LOOKUP);
3337 OBD_FREE(op_data, sizeof(*op_data));
3342 LASSERT(de_srcdir->d_inode);
3346 * we already know the target should be created on another MDS so, we
3347 * have to request that MDS to do it.
3350 /* prepare source id */
3351 if (de_old->d_flags & DCACHE_CROSS_REF) {
3352 LASSERT(de_old->d_inode == NULL);
3353 CDEBUG(D_OTHER, "request to move remote name\n");
3354 mds_pack_dentry2id(obd, &op_data->id1, de_old, 1);
3355 } else if (de_old->d_inode == NULL) {
3356 /* oh, source doesn't exist */
3357 OBD_FREE(op_data, sizeof(*op_data));
3358 GOTO(cleanup, rc = -ENOENT);
3360 struct lustre_id sid;
3361 struct inode *inode = de_old->d_inode;
3363 LASSERT(inode != NULL);
3364 CDEBUG(D_OTHER, "request to move local name\n");
3365 id_ino(&op_data->id1) = inode->i_ino;
3366 id_group(&op_data->id1) = mds->mds_num;
3367 id_gen(&op_data->id1) = inode->i_generation;
3369 down(&inode->i_sem);
3370 rc = mds_read_inode_sid(obd, inode, &sid);
3373 CERROR("Can't read inode self id, "
3374 "inode %lu, rc = %d\n",
3379 id_fid(&op_data->id1) = id_fid(&sid);
3382 op_data->id2 = *rec->ur_id2;
3383 rc = md_rename(mds->mds_md_exp, op_data, NULL, 0,
3384 rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3385 OBD_FREE(op_data, sizeof(*op_data));
3390 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3396 ptlrpc_req_finished(req2);
3399 if (parent_lockh[1].cookie != 0)
3400 ldlm_lock_decref(parent_lockh + 1, update_mode);
3402 ldlm_lock_decref(parent_lockh, LCK_PW);
3403 if (child_lockh.cookie != 0)
3404 ldlm_lock_decref(&child_lockh, LCK_EX);
3409 req->rq_status = rc;
3413 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3414 struct ptlrpc_request *req, struct lustre_handle *lockh)
3416 struct obd_device *obd = req->rq_export->exp_obd;
3417 struct dentry *de_srcdir = NULL;
3418 struct dentry *de_tgtdir = NULL;
3419 struct dentry *de_old = NULL;
3420 struct dentry *de_new = NULL;
3421 struct inode *old_inode = NULL, *new_inode = NULL;
3422 struct mds_obd *mds = mds_req2mds(req);
3423 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3424 struct mds_body *body = NULL;
3425 struct llog_create_locks *lcl = NULL;
3426 struct lov_mds_md *lmm = NULL;
3427 int rc = 0, cleanup_phase = 0;
3428 void *handle = NULL;
3431 LASSERT(offset == 1);
3433 DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3434 OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3437 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3439 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3440 DEBUG_REQ(D_HA, req, "rename replay");
3441 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3442 lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3443 req->rq_repmsg->buflens[2]);
3446 MD_COUNTER_INCREMENT(obd, rename);
3448 if (rec->ur_namelen == 1) {
3449 rc = mds_reint_rename_create_name(rec, offset, req);
3453 /* check if new name should be located on remote target. */
3454 if (id_group(rec->ur_id2) != mds->mds_num) {
3455 rc = mds_reint_rename_to_remote(rec, offset, req);
3459 rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3460 rec->ur_id2, &de_tgtdir, LCK_PW,
3461 rec->ur_name, rec->ur_namelen,
3462 &de_old, rec->ur_tgt,
3463 rec->ur_tgtlen, &de_new,
3464 dlm_handles, LCK_EX);
3468 cleanup_phase = 1; /* parent(s), children, locks */
3469 old_inode = de_old->d_inode;
3470 new_inode = de_new->d_inode;
3472 /* sanity check for src inode */
3473 if (de_old->d_flags & DCACHE_CROSS_REF) {
3474 LASSERT(de_old->d_inode == NULL);
3477 * in the case of cross-ref dir, we can perform this check only
3478 * if child and parent lie on the same mds. This is because
3479 * otherwise they can have the same inode numbers.
3481 if (de_old->d_mdsnum == mds->mds_num) {
3482 if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3483 de_old->d_inum == de_tgtdir->d_inode->i_ino)
3484 GOTO(cleanup, rc = -EINVAL);
3487 LASSERT(de_old->d_inode != NULL);
3488 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3489 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3490 GOTO(cleanup, rc = -EINVAL);
3493 /* sanity check for dest inode */
3494 if (de_new->d_flags & DCACHE_CROSS_REF) {
3495 LASSERT(new_inode == NULL);
3497 /* the same check about target dentry. */
3498 if (de_new->d_mdsnum == mds->mds_num) {
3499 if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3500 de_new->d_inum == de_tgtdir->d_inode->i_ino)
3501 GOTO(cleanup, rc = -EINVAL);
3505 * regular files usualy do not have ->rename() implemented. But
3506 * we handle only this case when @de_new is cross-ref entry,
3507 * because in other cases it will be handled by vfs_rename().
3509 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3510 !de_old->d_inode->i_op->rename))
3511 GOTO(cleanup, rc = -EPERM);
3514 (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3515 new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3516 GOTO(cleanup, rc = -EINVAL);
3521 * check if inodes point to each other. This should be checked before
3522 * is_subdir() check, as for the same entries it will think that they
3525 if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3526 !(de_new->d_flags & DCACHE_CROSS_REF) &&
3527 old_inode == new_inode)
3528 GOTO(cleanup, rc = 0);
3530 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3532 * check if we are moving old entry into its child. 2.6 does not check
3533 * for this in vfs_rename() anymore.
3535 if (is_subdir(de_new, de_old))
3536 GOTO(cleanup, rc = -EINVAL);
3540 * if we are about to remove the target at first, pass the EA of that
3541 * inode to client to perform and cleanup on OST.
3543 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3544 LASSERT(body != NULL);
3546 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3548 DOWN_READ_I_ALLOC_SEM(new_inode);
3550 cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3552 if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3553 new_inode->i_nlink == 2) ||
3554 new_inode->i_nlink == 1)) {
3555 if (mds_orphan_open_count(new_inode) > 0) {
3556 /* need to lock pending_dir before transaction */
3557 down(&mds->mds_pending_dir->d_inode->i_sem);
3558 cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3559 } else if (S_ISREG(new_inode->i_mode)) {
3560 mds_pack_inode2body(obd, body, new_inode, 0);
3561 mds_pack_md(obd, req->rq_repmsg, 1, body,
3562 new_inode, MDS_PACK_MD_LOCK, 0);
3566 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3567 de_srcdir->d_inode->i_sb);
3569 if (de_old->d_flags & DCACHE_CROSS_REF) {
3570 struct lustre_id old_id;
3572 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3574 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3579 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3584 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3585 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3586 NULL, le32_to_cpu(lmm->lmm_stripe_count));
3589 GOTO(cleanup, rc = PTR_ERR(handle));
3592 de_old->d_fsdata = req;
3593 de_new->d_fsdata = req;
3594 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3597 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3598 if (mds_orphan_open_count(new_inode) > 0)
3599 rc = mds_orphan_add_link(rec, obd, de_new);
3602 GOTO(cleanup, rc = 0);
3604 if (!S_ISREG(new_inode->i_mode))
3607 if (!(body->valid & OBD_MD_FLEASIZE)) {
3608 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3609 OBD_MD_FLATIME | OBD_MD_FLMTIME);
3610 } else if (mds_log_op_unlink(obd, new_inode,
3611 lustre_msg_buf(req->rq_repmsg,1,0),
3612 req->rq_repmsg->buflens[1],
3613 lustre_msg_buf(req->rq_repmsg,2,0),
3614 req->rq_repmsg->buflens[2],
3616 body->valid |= OBD_MD_FLCOOKIE;
3619 rc = mds_destroy_object(obd, old_inode, 1);
3621 CERROR("can't remove OST object, err %d\n",
3628 rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3629 handle, req, rc, 0);
3631 switch (cleanup_phase) {
3633 up(&mds->mds_pending_dir->d_inode->i_sem);
3636 UP_READ_I_ALLOC_SEM(new_inode);
3639 if (dlm_handles[5].cookie != 0)
3640 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3641 if (dlm_handles[6].cookie != 0)
3642 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3645 ptlrpc_save_llog_lock(req, lcl);
3648 if (dlm_handles[3].cookie != 0)
3649 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3650 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3651 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3652 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3654 if (dlm_handles[3].cookie != 0)
3655 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3656 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3657 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3658 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3667 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3670 req->rq_status = rc;
3674 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3675 struct ptlrpc_request *, struct lustre_handle *);
3677 static mds_reinter reinters[REINT_MAX + 1] = {
3678 [REINT_SETATTR] mds_reint_setattr,
3679 [REINT_CREATE] mds_reint_create,
3680 [REINT_LINK] mds_reint_link,
3681 [REINT_UNLINK] mds_reint_unlink,
3682 [REINT_RENAME] mds_reint_rename,
3683 [REINT_OPEN] mds_open
3686 int mds_reint_rec(struct mds_update_record *rec, int offset,
3687 struct ptlrpc_request *req, struct lustre_handle *lockh)
3689 struct obd_device *obd = req->rq_export->exp_obd;
3690 struct lvfs_run_ctxt saved;
3693 /* checked by unpacker */
3694 LASSERT(rec->ur_opcode <= REINT_MAX &&
3695 reinters[rec->ur_opcode] != NULL);
3697 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3698 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3699 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);