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 static int mds_get_md_type(char *name)
458 if (!strcmp(name, XATTR_LUSTRE_MDS_LOV_EA))
460 if (!strcmp(name, XATTR_LUSTRE_MDS_MEA_EA))
462 if (!strcmp(name, XATTR_LUSTRE_MDS_SID_EA))
464 if (!strcmp(name, XATTR_LUSTRE_MDS_PID_EA))
466 if (!strcmp(name, XATTR_LUSTRE_MDS_KEY_EA))
472 /* In the raw-setattr case, we lock the child inode.
473 * In the write-back case or if being called from open, the client holds a lock
476 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
477 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
478 struct ptlrpc_request *req, struct lustre_handle *lh)
480 struct mds_obd *mds = mds_req2mds(req);
481 struct obd_device *obd = req->rq_export->exp_obd;
482 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
483 struct mds_body *body;
484 struct dentry *de = NULL;
485 struct inode *inode = NULL;
486 struct lustre_handle lockh[2] = {{0}, {0}};
489 struct mds_logcancel_data *mlcd = NULL;
490 int rc = 0, cleanup_phase = 0, err;
491 int repcnt = 1, repsize[2] = { sizeof(*body) };
492 int locked = 0, do_trunc = 0;
495 LASSERT(offset == 1);
497 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
498 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
499 rec->ur_iattr.ia_valid);
501 /* remote setfacl need special handling */
502 if ((rec->ur_iattr.ia_valid & ATTR_EA) &&
503 !strcmp(rec->ur_eadata, XATTR_NAME_LUSTRE_ACL)) {
504 return mds_reint_remote_setfacl(obd, med, rec, req);
507 if (rec->ur_iattr.ia_valid & ATTR_SIZE) {
508 repsize[repcnt++] = sizeof(struct lustre_capa);
509 do_trunc = 1; /* XXX: ATTR_SIZE will be cleared from ia_valid */
512 rc = lustre_pack_reply(req, repcnt, repsize, NULL);
516 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
517 MD_COUNTER_INCREMENT(obd, setattr);
519 if (med->med_remote) {
520 if (rec->ur_iattr.ia_valid & ATTR_GID) {
521 CWARN("Deny chgrp from remote client\n");
522 GOTO(cleanup, rc = -EPERM);
524 if (rec->ur_iattr.ia_valid & ATTR_UID) {
527 uid = mds_idmap_lookup_uid(med->med_idmap, 0,
528 rec->ur_iattr.ia_uid);
529 if (uid == MDS_IDMAP_NOTFOUND) {
530 CWARN("Deny chown to uid %u\n",
531 rec->ur_iattr.ia_uid);
532 GOTO(cleanup, rc = -EPERM);
534 rec->ur_iattr.ia_uid = uid;
538 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
539 de = mds_id2dentry(obd, rec->ur_id1, NULL);
541 GOTO(cleanup, rc = PTR_ERR(de));
543 __u64 lockpart = MDS_INODELOCK_UPDATE;
544 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
545 lockpart |= MDS_INODELOCK_LOOKUP;
546 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
547 lockh, &parent_mode, NULL, 0, lockpart);
549 GOTO(cleanup, rc = PTR_ERR(de));
557 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
558 rec->ur_eadata != NULL)
561 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
563 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
565 GOTO(cleanup, rc = PTR_ERR(handle));
567 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
568 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
569 LTIME_S(rec->ur_iattr.ia_mtime),
570 LTIME_S(rec->ur_iattr.ia_ctime));
571 rc = mds_fix_attr(inode, rec);
575 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
576 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
577 (long)&rec->ur_iattr.ia_attr_flags);
579 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
582 if (rec->ur_iattr.ia_valid & ATTR_EA) {
583 int flags = (int) rec->ur_iattr.ia_attr_flags;
586 if (!med->med_remote && inode->i_op &&
587 inode->i_op->setxattr)
588 rc = inode->i_op->setxattr(
593 } else if (rec->ur_iattr.ia_valid & ATTR_EA_RM) {
595 if (inode->i_op && inode->i_op->removexattr)
596 rc = inode->i_op->removexattr(de,
598 } else if (rec->ur_iattr.ia_valid & ATTR_EA_CMOBD) {
602 LASSERT(rec->ur_eadata != NULL);
603 LASSERT(rec->ur_ea2data != NULL);
604 name = rec->ur_eadata;
606 /* XXX: tmp fix for setting LOV EA from CMOBD */
607 type = mds_get_md_type(name);
609 if (type == EA_LOV) {
610 CDEBUG(D_INFO, "set %s EA for cmobd\n", name);
612 rc = fsfilt_set_md(obd, inode, handle,
617 CERROR("fsfilt_set_md() failed, err %d\n",
622 } else if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
623 !((rec->ur_iattr.ia_valid & ATTR_KEY) ||
624 (rec->ur_iattr.ia_valid & ATTR_MAC))) {
626 struct lov_stripe_md *lsm = NULL;
627 struct lov_user_md *lum = NULL;
629 if (rec->ur_eadata != NULL) {
630 rc = ll_permission(inode, MAY_WRITE, NULL);
634 lum = rec->ur_eadata;
636 /* if lmm_stripe_size is -1 delete default
638 if (S_ISDIR(inode->i_mode) &&
639 lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1)){
640 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, EA_LOV);
644 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
646 &lsm, rec->ur_eadata);
650 obd_free_memmd(mds->mds_dt_exp, &lsm);
651 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
652 rec->ur_eadatalen, EA_LOV);
659 if ((rec->ur_iattr.ia_valid & ATTR_KEY) || (rec->ur_iattr.ia_valid & ATTR_MAC)) {
663 LASSERT(rec->ur_eadatalen || rec->ur_ea3datalen);
664 LASSERT(rec->ur_eadata || rec->ur_ea3data);
665 key = rec->ur_ea3data ? rec->ur_ea3data : rec->ur_eadata;
666 keylen = rec->ur_ea3datalen ? rec->ur_ea3datalen :
668 mds_set_gskey(obd, handle, inode, key, keylen,
669 rec->ur_iattr.ia_valid);
673 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
674 mds_pack_inode2body(obd, body, inode, 1);
676 /* Don't return OST-specific attributes if we didn't just set them */
677 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
678 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
679 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
680 body->valid |= OBD_MD_FLMTIME;
681 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
682 body->valid |= OBD_MD_FLATIME;
685 struct lustre_capa capa = {
686 .lc_uid = rec->ur_uc.luc_uid,
688 .lc_ino = inode->i_ino,
689 .lc_igen = inode->i_generation,
690 .lc_mdsid = mds->mds_num,
694 LASSERT(capa.lc_mdsid == mds->mds_num);
695 rc = mds_pack_capa(obd, med, NULL, &capa, req,
698 CERROR("mds_pack_capa: rc = %d\n", rc);
703 mds_body_do_reverse_map(med, body);
705 /* The logcookie should be no use anymore, why nobody remove
706 * following code block?
708 LASSERT(rec->ur_cookielen == 0);
709 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_dt_obd)) {
710 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
713 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
715 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
716 mlcd->mlcd_cookielen = rec->ur_cookielen;
717 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
718 mlcd->mlcd_cookielen;
719 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
720 mlcd->mlcd_cookielen);
721 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
722 mlcd->mlcd_eadatalen);
724 CERROR("unable to allocate log cancel data\n");
730 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
731 handle, mds_cancel_cookies_cb, mlcd);
732 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
733 switch (cleanup_phase) {
735 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
736 rec->ur_eadata != NULL)
741 if (lockh[1].cookie != 0)
742 ldlm_lock_decref(lockh + 1, parent_mode);
745 ldlm_lock_decref(lockh, LCK_PW);
747 ptlrpc_save_lock (req, lockh, LCK_PW);
762 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
763 struct ptlrpc_request *req)
765 struct mds_export_data *med = &req->rq_export->exp_mds_data;
766 struct dentry *parent, *child;
767 struct mds_body *body;
770 mds_req_from_mcd(req, med->med_mcd);
772 if (req->rq_status) {
777 parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
778 LASSERT(!IS_ERR(parent));
779 child = ll_lookup_one_len(rec->ur_name, parent,
780 rec->ur_namelen - 1);
781 LASSERT(!IS_ERR(child));
782 if ((child->d_flags & DCACHE_CROSS_REF)) {
783 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
784 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
785 mds_pack_dentry2body(req2obd(req), body, child, 1);
786 } else if (child->d_inode == NULL) {
787 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
788 OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
789 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
791 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
792 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
799 static int mds_get_default_acl(struct inode *dir, void **pacl)
801 struct dentry de = { .d_inode = dir };
804 LASSERT(S_ISDIR(dir->i_mode));
806 if (!dir->i_op->getxattr)
809 size = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, NULL, 0);
810 if (size == 0 || size == -ENODATA || size == -EOPNOTSUPP)
815 OBD_ALLOC(*pacl, size);
819 size2 = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, *pacl, size);
821 /* since we already locked the dir, it should not change
822 * between the 2 getxattr calls
824 CERROR("2'nd getxattr got %d, expect %d\n", size2, size);
825 OBD_FREE(*pacl, size);
832 static int mds_reint_create(struct mds_update_record *rec, int offset,
833 struct ptlrpc_request *req,
834 struct lustre_handle *lh)
836 struct dentry *dparent = NULL;
837 struct mds_obd *mds = mds_req2mds(req);
838 struct obd_device *obd = req->rq_export->exp_obd;
839 struct mds_body *body = NULL;
840 struct dentry *dchild = NULL;
841 struct inode *dir = NULL;
843 struct lustre_handle lockh[2] = {{0}, {0}};
845 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
847 struct dentry_params dp;
848 struct mea *mea = NULL;
850 struct lustre_id sid;
854 LASSERT(offset == 1);
856 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
857 OBD_MDS_DEVICENAME));
859 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
860 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
861 rec->ur_name, rec->ur_mode);
863 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
865 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
866 GOTO(cleanup, rc = -ESTALE);
868 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
869 lockh, &parent_mode, rec->ur_name,
870 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
871 if (IS_ERR(dparent)) {
872 rc = PTR_ERR(dparent);
873 CERROR("parent lookup error %d, id "DLID4"\n",
874 rc, OLID4(rec->ur_id1));
877 cleanup_phase = 1; /* locked parent dentry */
878 dir = dparent->d_inode;
881 ldlm_lock_dump_handle(D_OTHER, lockh);
883 /* get parent id: ldlm lock on the parent protects ea */
884 rc = mds_read_inode_sid(obd, dir, &sid);
886 CERROR("can't read parent id. ino(%lu) rc(%d)\n",
891 /* try to retrieve MEA data for this dir */
892 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
896 if (mea != NULL && mea->mea_count) {
898 * dir is already splitted, check is requested filename should
899 * live at this MDS or at another one.
901 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
902 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
903 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
904 " should be %lu(%d)\n",
905 mea->mea_master, dparent->d_inode->i_ino,
906 dparent->d_inode->i_generation, rec->ur_name,
907 (unsigned long)id_group(&mea->mea_ids[i]), i);
908 GOTO(cleanup, rc = -ERESTART);
912 dchild = ll_lookup_one_len(rec->ur_name, dparent,
913 rec->ur_namelen - 1);
914 if (IS_ERR(dchild)) {
915 rc = PTR_ERR(dchild);
916 CERROR("Can't find "DLID4"/%s, error %d\n",
917 OLID4(rec->ur_id1), rec->ur_name, rc);
921 cleanup_phase = 2; /* child dentry */
923 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
925 if (type == S_IFREG || type == S_IFDIR) {
926 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
927 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
928 obd->obd_name, dparent->d_inode->i_ino,
929 dparent->d_inode->i_generation, rc, parent_mode);
931 /* dir got splitted */
932 GOTO(cleanup, rc = -ERESTART);
934 /* error happened during spitting. */
939 if (dir->i_mode & S_ISGID) {
940 if (S_ISDIR(rec->ur_mode))
941 rec->ur_mode |= S_ISGID;
944 /* for reint case stor ecookie should be zero */
945 if (rec->ur_flags & MDS_REINT_REQ) {
946 LASSERT(id_ino(rec->ur_id1) == 0);
947 LASSERT(id_ino(rec->ur_id2) == 0);
950 if (id_fid(rec->ur_id2))
951 fid = id_fid(rec->ur_id2);
953 fid = mds_alloc_fid(obd);
955 dchild->d_fsdata = (void *)&dp;
956 dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
960 dp.p_group = mds->mds_num;
962 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
966 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
968 GOTO(cleanup, rc = PTR_ERR(handle));
969 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
971 /* XXX: here we should check what type of EA is in ur_eadata
972 * and do appropriate actions. --umka */
973 if (rec->ur_eadata && rec->ur_eadatalen &&
974 rc == 0 && dchild->d_inode != NULL) {
975 if (rec->ur_flags & MDS_REINT_REQ) {
976 /* for CMOBD to set lov md info when cmobd reint
978 CDEBUG(D_INFO, "set lsm %p, len %d to inode %lu \n",
979 rec->ur_eadata, rec->ur_eadatalen,
980 dchild->d_inode->i_ino);
981 rc = fsfilt_set_md(obd, dchild->d_inode, handle,
982 rec->ur_eadata, rec->ur_eadatalen,
985 CERROR("fsfilt_set_md() failed, err %d\n",
990 /* assumption: when ur_eadata is not NULL,
991 * ur_eadata is crypto key, should fix it later,
993 rc = mds_set_gskey(obd, handle, dchild->d_inode,
994 rec->ur_eadata, rec->ur_eadatalen,
995 ATTR_MAC | ATTR_KEY);
997 CWARN("mds_set_gskey() failed, err %d\n",
1008 * as Peter asked, mkdir() should distribute new directories
1009 * over the whole cluster in order to distribute namespace
1010 * processing load. first, we calculate which MDS to use to put
1011 * new directory's inode in.
1014 /* XXX: here we order mds_choose_mdsnum() to use local mdsnum
1015 * for reint requests. This should be gone when real flushing on
1016 * LMV is fixed. --umka */
1017 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
1018 rec->ur_flags, &req->rq_peer, dir,
1019 (rec->ur_flags & MDS_REINT_REQ) ? 1 : 0);
1021 if (i == mds->mds_num) {
1022 /* inode will be created locally */
1023 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1025 GOTO(cleanup, rc = PTR_ERR(handle));
1027 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
1030 "Can't create dir \"%s\", rc = %d\n",
1031 dchild->d_name.name, rc);
1035 } else if (!DENTRY_VALID(dchild)) {
1036 /* inode will be created on another MDS */
1037 struct obdo *oa = NULL;
1041 /* first, create that inode */
1044 GOTO(cleanup, rc = -ENOMEM);
1049 if (rec->ur_eadata) {
1050 /* user asks for creating splitted dir */
1051 oa->o_easize = *((u16 *) rec->ur_eadata);
1054 obdo_from_inode(oa, dir, OBD_MD_FLATIME |
1055 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1057 /* adjust the uid/gid/mode bits */
1058 oa->o_mode = rec->ur_mode;
1059 oa->o_uid = current->fsuid;
1060 oa->o_gid = (dir->i_mode & S_ISGID) ?
1061 dir->i_gid : current->fsgid;
1063 /* letting remote MDS know that this is reint
1065 if (rec->ur_flags & MDS_REINT_REQ)
1066 oa->o_flags |= OBD_FL_REINT;
1068 /* transfer parent id to remote inode */
1069 memcpy(obdo_id(oa), &sid, sizeof(sid));
1070 oa->o_valid |= OBD_MD_FLTYPE | OBD_MD_FLUID |
1071 OBD_MD_FLGID | OBD_MD_FLIFID;
1073 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
1076 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1078 * here inode number and generation are
1079 * important, as this is replay request and we
1080 * need them to check if such an object is
1083 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
1084 obd->obd_name, rec->ur_namelen - 1,
1085 rec->ur_name, (unsigned)id_ino(rec->ur_id2),
1086 (unsigned)id_gen(rec->ur_id2));
1087 oa->o_id = id_ino(rec->ur_id2);
1088 oa->o_fid = id_fid(rec->ur_id2);
1089 oa->o_generation = id_gen(rec->ur_id2);
1090 oa->o_flags |= OBD_FL_RECREATE_OBJS;
1091 LASSERT(oa->o_fid != 0);
1094 /* obtain default ACL */
1095 acl_size = mds_get_default_acl(dir, &acl);
1098 GOTO(cleanup, rc = -ENOMEM);
1102 * before obd_create() is called, o_fid is not known if
1103 * this is not recovery of cause.
1105 rc = obd_create(mds->mds_md_exp, oa, acl, acl_size,
1109 OBD_FREE(acl, acl_size);
1112 CERROR("can't create remote inode: %d\n", rc);
1113 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
1114 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1115 rec->ur_name, rec->ur_mode);
1120 LASSERT(oa->o_fid != 0);
1122 /* now, add new dir entry for it */
1123 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1124 if (IS_ERR(handle)) {
1126 GOTO(cleanup, rc = PTR_ERR(handle));
1129 /* creating local dentry for remote inode. */
1130 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
1131 rec->ur_namelen - 1, oa->o_id,
1132 oa->o_generation, i, oa->o_fid);
1135 CERROR("Can't create local entry %*s for "
1136 "remote inode.\n", rec->ur_namelen - 1,
1142 body->valid |= OBD_MD_FLID | OBD_MD_MDS | OBD_MD_FID;
1144 obdo2id(&body->id1, oa);
1147 /* requested name exists in the directory */
1153 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
1155 GOTO(cleanup, rc = PTR_ERR(handle));
1156 if (rec->ur_tgt == NULL) /* no target supplied */
1157 rc = -EINVAL; /* -EPROTO? */
1159 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
1166 int rdev = rec->ur_rdev;
1167 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
1169 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
1170 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
1174 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
1175 dchild->d_fsdata = NULL;
1176 GOTO(cleanup, rc = -EINVAL);
1179 /* In case we stored the desired inum in here, we want to clean up. */
1180 if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
1181 dchild->d_fsdata = NULL;
1184 CDEBUG(D_INODE, "error during create: %d\n", rc);
1186 } else if (dchild->d_inode) {
1187 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1188 struct inode *inode = dchild->d_inode;
1192 iattr.ia_uid = rec->ur_fsuid;
1193 LTIME_S(iattr.ia_atime) = rec->ur_time;
1194 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1195 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1197 if (dir->i_mode & S_ISGID)
1198 iattr.ia_gid = dir->i_gid;
1200 iattr.ia_gid = rec->ur_fsgid;
1202 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1203 ATTR_MTIME | ATTR_CTIME;
1205 if (id_ino(rec->ur_id2)) {
1206 LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
1207 inode->i_generation = id_gen(rec->ur_id2);
1208 /* dirtied and committed by the upcoming setattr. */
1209 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1210 inode->i_ino, inode->i_generation);
1212 mds_inode2id(obd, &body->id1, dchild->d_inode, fid);
1213 mds_update_inode_ids(obd, inode, handle, &body->id1, &sid);
1215 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1217 CERROR("error on child setattr: rc = %d\n", rc);
1219 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1220 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1222 CERROR("error on parent setattr: rc = %d\n", rc);
1224 MD_COUNTER_INCREMENT(obd, create);
1226 /* take care of default stripe inheritance */
1227 if (type == S_IFDIR) {
1228 struct lov_mds_md lmm;
1229 int lmm_size = sizeof(lmm), nstripes = 0;
1231 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0);
1233 down(&inode->i_sem);
1234 rc = fsfilt_set_md(obd, inode, handle,
1235 &lmm, lmm_size, EA_LOV);
1239 CERROR("error on copy stripe info: rc = %d\n",
1245 nstripes = *(u16 *)rec->ur_eadata;
1249 * we pass LCK_EX to split routine to signal,
1250 * that we have exclusive access to the
1251 * directory. Simple because nobody knows it
1252 * already exists -bzzz
1254 rc = mds_try_to_split_dir(obd, dchild,
1258 /* dir got splitted */
1260 } else if (rc < 0) {
1261 /* an error occured during
1269 mds_pack_inode2body(obd, body, inode, 1);
1270 mds_body_do_reverse_map(med, body);
1272 if (rec->ur_flags & MDS_REINT_REQ) {
1273 LASSERT(body != NULL);
1274 rc = mds_fidmap_add(obd, &body->id1);
1276 CERROR("can't create fid->ino mapping, "
1286 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
1288 if (rc && created) {
1289 /* Destroy the file we just created. This should not need extra
1290 * journal credits, as we have already modified all of the
1291 * blocks needed in order to create the file in the first
1295 err = vfs_rmdir(dir, dchild);
1297 CERROR("rmdir in error path: %d\n", err);
1300 err = vfs_unlink(dir, dchild);
1302 CERROR("unlink in error path: %d\n", err);
1305 } else if (created) {
1306 /* The inode we were allocated may have just been freed
1307 * by an unlink operation. We take this lock to
1308 * synchronize against the matching reply-ack-lock taken
1309 * in unlink, to avoid replay problems if this reply
1310 * makes it out to the client but the unlink's does not.
1311 * See bug 2029 for more detail.*/
1312 mds_lock_new_child(obd, dchild->d_inode, NULL);
1316 switch (cleanup_phase) {
1317 case 2: /* child dentry */
1319 mds_audit(req, dchild, rec->ur_name,
1320 rec->ur_namelen - 1, AUDIT_CREATE, rc);
1322 case 1: /* locked parent dentry */
1324 if (lockh[1].cookie != 0)
1325 ldlm_lock_decref(lockh + 1, parent_mode);
1328 ldlm_lock_decref(lockh, LCK_PW);
1330 ptlrpc_save_lock(req, lockh, LCK_PW);
1336 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1340 OBD_FREE(mea, mea_size);
1341 req->rq_status = rc;
1346 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1347 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1351 for (i = 0; i < RES_NAME_SIZE; i++) {
1353 * this is needed to make zeroed res_id entries to be put at the
1354 * end of list in *ordered_locks() .
1356 if (res1->name[i] == 0 && res2->name[i] != 0)
1358 if (res2->name[i] == 0 && res1->name[i] != 0)
1360 if (res1->name[i] > res2->name[i])
1362 if (res1->name[i] < res2->name[i])
1369 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1375 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1376 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1377 * because they take the place of local semaphores.
1379 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1380 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1381 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1382 struct lustre_handle *p1_lockh, int p1_lock_mode,
1383 ldlm_policy_data_t *p1_policy,
1384 struct ldlm_res_id *p2_res_id,
1385 struct lustre_handle *p2_lockh, int p2_lock_mode,
1386 ldlm_policy_data_t *p2_policy)
1388 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1389 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1390 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1391 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1395 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1397 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1398 res_id[0]->name[0], res_id[1]->name[0]);
1400 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1401 handles[1] = p1_lockh;
1402 handles[0] = p2_lockh;
1403 res_id[1] = p1_res_id;
1404 res_id[0] = p2_res_id;
1405 lock_modes[1] = p1_lock_mode;
1406 lock_modes[0] = p2_lock_mode;
1407 policies[1] = p1_policy;
1408 policies[0] = p2_policy;
1411 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1412 res_id[0]->name[0], res_id[1]->name[0]);
1414 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1415 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1416 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1417 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1418 NULL, 0, NULL, handles[0]);
1421 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1423 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1424 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1425 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1426 ldlm_lock_addref(handles[1], lock_modes[1]);
1427 } else if (res_id[1]->name[0] != 0) {
1428 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1429 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1430 *res_id[1], LDLM_IBITS, policies[1],
1431 lock_modes[1], &flags, mds_blocking_ast,
1432 ldlm_completion_ast, NULL, NULL, NULL, 0,
1434 if (rc != ELDLM_OK) {
1435 ldlm_lock_decref(handles[0], lock_modes[0]);
1438 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1444 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1445 struct lustre_handle *p1_lockh, int p1_lock_mode,
1446 ldlm_policy_data_t *p1_policy,
1447 struct ldlm_res_id *p2_res_id,
1448 struct lustre_handle *p2_lockh, int p2_lock_mode,
1449 ldlm_policy_data_t *p2_policy,
1450 struct ldlm_res_id *c1_res_id,
1451 struct lustre_handle *c1_lockh, int c1_lock_mode,
1452 ldlm_policy_data_t *c1_policy,
1453 struct ldlm_res_id *c2_res_id,
1454 struct lustre_handle *c2_lockh, int c2_lock_mode,
1455 ldlm_policy_data_t *c2_policy)
1457 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1458 c1_res_id, c2_res_id };
1459 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1460 c1_lockh, c2_lockh };
1461 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1462 c1_lock_mode, c2_lock_mode };
1463 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1464 c1_policy, c2_policy};
1465 int rc, i, j, sorted, flags;
1468 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1469 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1470 res_id[3]->name[0]);
1473 * simple insertion sort - we have at most 4 elements. Note, that zeroed
1474 * res_id should be at the end of list after sorting is finished.
1476 for (i = 1; i < 4; i++) {
1478 dlm_handles[4] = dlm_handles[i];
1479 res_id[4] = res_id[i];
1480 lock_modes[4] = lock_modes[i];
1481 policies[4] = policies[i];
1485 if (res_gt(res_id[j], res_id[4], policies[j],
1487 dlm_handles[j + 1] = dlm_handles[j];
1488 res_id[j + 1] = res_id[j];
1489 lock_modes[j + 1] = lock_modes[j];
1490 policies[j + 1] = policies[j];
1495 } while (j >= 0 && !sorted);
1497 dlm_handles[j + 1] = dlm_handles[4];
1498 res_id[j + 1] = res_id[4];
1499 lock_modes[j + 1] = lock_modes[4];
1500 policies[j + 1] = policies[4];
1503 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1504 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1505 res_id[3]->name[0]);
1507 /* XXX we could send ASTs on all these locks first before blocking? */
1508 for (i = 0; i < 4; i++) {
1509 flags = LDLM_FL_ATOMIC_CB;
1512 * nevertheless zeroed res_ids should be at the end of list, and
1513 * could use break here, I think, that it is more correctly for
1514 * clear understanding of code to have continue here, as it
1515 * clearly means, that zeroed res_id should be skipped and does
1516 * not mean, that if we meet zeroed res_id we should stop
1519 if (res_id[i]->name[0] == 0)
1523 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1524 (policies[i]->l_inodebits.bits &
1525 policies[i-1]->l_inodebits.bits) ) {
1526 memcpy(dlm_handles[i], dlm_handles[i-1],
1527 sizeof(*(dlm_handles[i])));
1528 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1530 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1531 *res_id[i], LDLM_IBITS,
1533 lock_modes[i], &flags,
1535 ldlm_completion_ast, NULL, NULL,
1536 NULL, 0, NULL, dlm_handles[i]);
1538 GOTO(out_err, rc = -EIO);
1539 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1546 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1551 /* In the unlikely case that the child changed while we were waiting
1552 * on the lock, we need to drop the lock on the old child and either:
1553 * - if the child has a lower resource name, then we have to also
1554 * drop the parent lock and regain the locks in the right order
1555 * - in the rename case, if the child has a lower resource name than one of
1556 * the other parent/child resources (maxres) we also need to reget the locks
1557 * - if the child has a higher resource name (this is the common case)
1558 * we can just get the lock on the new child (still in lock order)
1560 * Returns 0 if the child did not change or if it changed but could be locked.
1561 * Returns 1 if the child changed and we need to re-lock (no locks held).
1562 * Returns -ve error with a valid dchild (no locks held). */
1563 static int mds_verify_child(struct obd_device *obd,
1564 struct ldlm_res_id *parent_res_id,
1565 struct lustre_handle *parent_lockh,
1566 struct dentry *dparent, int parent_mode,
1567 struct ldlm_res_id *child_res_id,
1568 struct lustre_handle *child_lockh,
1569 struct dentry **dchildp, int child_mode,
1570 ldlm_policy_data_t *child_policy,
1571 const char *name, int namelen,
1572 struct ldlm_res_id *maxres,
1573 unsigned long child_ino, __u32 child_gen)
1575 struct lustre_id sid;
1576 struct dentry *vchild, *dchild = *dchildp;
1577 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1580 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1582 GOTO(cleanup, rc = PTR_ERR(vchild));
1584 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1585 if (child_gen == vchild->d_generation &&
1586 child_ino == vchild->d_inum) {
1595 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1596 (vchild->d_inode != NULL &&
1597 child_gen == vchild->d_inode->i_generation &&
1598 child_ino == vchild->d_inode->i_ino))) {
1606 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1607 vchild->d_inode, dchild ? dchild->d_inode : 0,
1608 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1609 child_res_id->name[0]);
1611 if (child_res_id->name[0] != 0)
1612 ldlm_lock_decref(child_lockh, child_mode);
1616 cleanup_phase = 1; /* parent lock only */
1617 *dchildp = dchild = vchild;
1619 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1620 int flags = LDLM_FL_ATOMIC_CB;
1622 if (dchild->d_inode) {
1623 down(&dchild->d_inode->i_sem);
1624 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1625 up(&dchild->d_inode->i_sem);
1627 CERROR("Can't read inode self id, inode %lu,"
1628 " rc %d\n", dchild->d_inode->i_ino, rc);
1631 child_res_id->name[0] = id_fid(&sid);
1632 child_res_id->name[1] = id_group(&sid);
1634 child_res_id->name[0] = dchild->d_fid;
1635 child_res_id->name[1] = dchild->d_mdsnum;
1638 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1639 res_gt(maxres, child_res_id, NULL, NULL)) {
1640 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1641 child_res_id->name[0], parent_res_id->name[0],
1643 GOTO(cleanup, rc = 1);
1646 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1647 *child_res_id, LDLM_IBITS, child_policy,
1648 child_mode, &flags, mds_blocking_ast,
1649 ldlm_completion_ast, NULL, NULL, NULL, 0,
1652 GOTO(cleanup, rc = -EIO);
1655 memset(child_res_id, 0, sizeof(*child_res_id));
1661 switch(cleanup_phase) {
1663 if (child_res_id->name[0] != 0)
1664 ldlm_lock_decref(child_lockh, child_mode);
1666 ldlm_lock_decref(parent_lockh, parent_mode);
1672 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1673 struct lustre_id *id,
1674 struct lustre_handle *parent_lockh,
1675 struct dentry **dparentp, int parent_mode,
1676 __u64 parent_lockpart, int *update_mode,
1677 char *name, int namelen,
1678 struct lustre_handle *child_lockh,
1679 struct dentry **dchildp, int child_mode,
1680 __u64 child_lockpart)
1682 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1683 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1684 struct ldlm_res_id parent_res_id = { .name = {0} };
1685 struct ldlm_res_id child_res_id = { .name = {0} };
1686 unsigned long child_ino = 0; __u32 child_gen = 0;
1687 int rc = 0, cleanup_phase = 0;
1688 struct lustre_id sid;
1689 struct inode *inode;
1692 /* Step 1: Lookup parent */
1693 *dparentp = mds_id2dentry(obd, id, NULL);
1694 if (IS_ERR(*dparentp)) {
1695 rc = PTR_ERR(*dparentp);
1700 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1701 (*dparentp)->d_inode->i_ino, name);
1703 parent_res_id.name[0] = id_fid(id);
1704 parent_res_id.name[1] = id_group(id);
1707 parent_lockh[1].cookie = 0;
1708 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1709 struct ldlm_res_id res_id = { .name = {0} };
1710 ldlm_policy_data_t policy;
1711 int flags = LDLM_FL_ATOMIC_CB;
1713 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1715 res_id.name[0] = id_fid(id);
1716 res_id.name[1] = id_group(id);
1717 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1719 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1720 res_id, LDLM_IBITS, &policy,
1721 *update_mode, &flags,
1723 ldlm_completion_ast,
1724 NULL, NULL, NULL, 0, NULL,
1730 parent_res_id.name[2] = full_name_hash((unsigned char *)name,
1733 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1734 (unsigned long)id_fid(id), (unsigned long)id_group(id),
1735 parent_res_id.name[2]);
1739 cleanup_phase = 1; /* parent dentry */
1741 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1742 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1743 if (IS_ERR(*dchildp)) {
1744 rc = PTR_ERR(*dchildp);
1745 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1749 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1751 * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1752 * lock. Drop possible UPDATE lock!
1754 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1755 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1757 child_res_id.name[0] = (*dchildp)->d_fid;
1758 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1759 child_ino = (*dchildp)->d_inum;
1760 child_gen = (*dchildp)->d_generation;
1764 inode = (*dchildp)->d_inode;
1766 inode = igrab(inode);
1770 down(&inode->i_sem);
1771 rc = mds_read_inode_sid(obd, inode, &sid);
1774 CERROR("Can't read inode self id, inode %lu, "
1775 "rc %d\n", inode->i_ino, rc);
1780 child_ino = inode->i_ino;
1781 child_gen = inode->i_generation;
1782 child_res_id.name[0] = id_fid(&sid);
1783 child_res_id.name[1] = id_group(&sid);
1787 cleanup_phase = 2; /* child dentry */
1789 /* Step 3: Lock parent and child in resource order. If child doesn't
1790 * exist, we still have to lock the parent and re-lookup. */
1791 rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1792 &parent_policy, &child_res_id, child_lockh,
1793 child_mode, &child_policy);
1797 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1798 cleanup_phase = 4; /* child lock */
1800 cleanup_phase = 3; /* parent lock */
1802 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1803 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1804 parent_mode, &child_res_id, child_lockh,
1805 dchildp, child_mode, &child_policy,
1806 name, namelen, &parent_res_id, child_ino,
1818 switch (cleanup_phase) {
1820 ldlm_lock_decref(child_lockh, child_mode);
1822 ldlm_lock_decref(parent_lockh, parent_mode);
1827 if (parent_lockh[1].cookie)
1828 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1836 void mds_reconstruct_generic(struct ptlrpc_request *req)
1838 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1839 mds_req_from_mcd(req, med->med_mcd);
1842 /* If we are unlinking an open file/dir (i.e. creating an orphan) then we
1843 * instead link the inode into the PENDING directory until it is finally
1844 * released. We can't simply call mds_reint_rename() or some part thereof,
1845 * because we don't have the inode to check for link count/open status until
1846 * after it is locked.
1848 * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1849 * before starting journal transaction.
1851 * returns 1 on success
1852 * returns 0 if we lost a race and didn't make a new link
1853 * returns negative on error
1855 static int mds_orphan_add_link(struct mds_update_record *rec,
1856 struct obd_device *obd, struct dentry *dentry)
1858 struct mds_obd *mds = &obd->u.mds;
1859 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1860 struct inode *inode = dentry->d_inode;
1861 struct dentry *pending_child;
1862 char idname[LL_ID_NAMELEN];
1863 int idlen = 0, rc, mode;
1866 LASSERT(inode != NULL);
1867 LASSERT(!mds_inode_is_orphan(inode));
1868 #ifndef HAVE_I_ALLOC_SEM
1869 LASSERT(down_trylock(&inode->i_sem) != 0);
1871 LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1873 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1875 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1876 mds_orphan_open_count(inode), inode->i_nlink,
1877 S_ISDIR(inode->i_mode) ? "dir" :
1878 S_ISREG(inode->i_mode) ? "file" : "other",
1879 rec->ur_name, idname);
1881 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1884 pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1885 if (IS_ERR(pending_child))
1886 RETURN(PTR_ERR(pending_child));
1888 if (pending_child->d_inode != NULL) {
1889 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1890 LASSERT(pending_child->d_inode == inode);
1891 GOTO(out_dput, rc = 0);
1895 * link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG for
1896 * linking and return real mode back then -bzzz
1898 mode = inode->i_mode;
1899 inode->i_mode = S_IFREG;
1900 rc = vfs_link(dentry, pending_dir, pending_child);
1902 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1905 mds_inode_set_orphan(inode);
1907 /* return mode and correct i_nlink if inode is directory */
1908 inode->i_mode = mode;
1909 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1910 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1912 if (S_ISDIR(mode)) {
1914 i_nlink_inc(pending_dir);
1915 mark_inode_dirty(inode);
1916 mark_inode_dirty(pending_dir);
1919 GOTO(out_dput, rc = 1);
1921 l_dput(pending_child);
1925 int mds_create_local_dentry(struct mds_update_record *rec,
1926 struct obd_device *obd)
1928 struct mds_obd *mds = &obd->u.mds;
1929 struct inode *id_dir = mds->mds_id_dir->d_inode;
1930 int idlen = 0, rc, cleanup_phase = 0;
1931 struct dentry *new_child = NULL;
1932 char *idname = rec->ur_name;
1933 struct dentry *child = NULL;
1934 struct lustre_handle lockh[2] = {{0}, {0}};
1935 struct lustre_id sid;
1939 down(&id_dir->i_sem);
1940 idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1941 id_gen(rec->ur_id1));
1943 CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1944 idname, OLID4(rec->ur_id1));
1946 new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1949 if (IS_ERR(new_child)) {
1950 CERROR("can't lookup %s: %d\n", idname,
1951 (int) PTR_ERR(new_child));
1952 GOTO(cleanup, rc = PTR_ERR(new_child));
1956 down(&id_dir->i_sem);
1957 rc = mds_read_inode_sid(obd, id_dir, &sid);
1960 CERROR("Can't read inode self id, inode %lu, "
1961 "rc %d\n", id_dir->i_ino, rc);
1965 if (new_child->d_inode != NULL) {
1966 /* nice. we've already have local dentry! */
1967 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1968 (unsigned)new_child->d_inode->i_ino,
1969 (unsigned)new_child->d_inode->i_generation);
1971 id_ino(rec->ur_id1) = id_dir->i_ino;
1972 id_gen(rec->ur_id1) = id_dir->i_generation;
1973 rec->ur_namelen = idlen + 1;
1975 id_fid(rec->ur_id1) = id_fid(&sid);
1976 id_group(rec->ur_id1) = id_group(&sid);
1978 GOTO(cleanup, rc = 0);
1981 /* new, local dentry will be added soon. we need no aliases here */
1984 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1985 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1987 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1988 LCK_EX, lockh, NULL, NULL, 0,
1989 MDS_INODELOCK_UPDATE);
1992 if (IS_ERR(child)) {
1993 rc = PTR_ERR(child);
1994 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1995 CERROR("can't get victim: %d\n", rc);
2000 handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
2002 GOTO(cleanup, rc = PTR_ERR(handle));
2004 rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
2005 idlen, id_ino(rec->ur_id1),
2006 id_gen(rec->ur_id1), mds->mds_num,
2007 id_fid(rec->ur_id1));
2009 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
2010 (unsigned long)child->d_inode->i_ino,
2011 (unsigned long)child->d_inode->i_generation, rc);
2013 if (S_ISDIR(child->d_inode->i_mode)) {
2014 i_nlink_inc(id_dir);
2015 mark_inode_dirty(id_dir);
2017 mark_inode_dirty(child->d_inode);
2019 fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
2021 id_ino(rec->ur_id1) = id_dir->i_ino;
2022 id_gen(rec->ur_id1) = id_dir->i_generation;
2023 rec->ur_namelen = idlen + 1;
2025 id_fid(rec->ur_id1) = id_fid(&sid);
2026 id_group(rec->ur_id1) = id_group(&sid);
2030 switch(cleanup_phase) {
2032 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
2033 ldlm_lock_decref(lockh, LCK_EX);
2043 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
2044 struct ptlrpc_request *slave)
2046 void *cookie, *cookie2;
2047 struct mds_body *body2;
2048 struct mds_body *body;
2052 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
2053 LASSERT(body != NULL);
2055 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
2056 LASSERT(body2 != NULL);
2058 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
2061 memcpy(body2, body, sizeof(*body));
2062 body2->valid &= ~OBD_MD_FLCOOKIE;
2064 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)))
2067 if (body->eadatasize == 0) {
2068 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
2072 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
2074 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
2075 LASSERT(ea != NULL);
2077 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
2078 LASSERT(ea2 != NULL);
2080 memcpy(ea2, ea, body->eadatasize);
2082 if (body->valid & OBD_MD_FLCOOKIE) {
2083 LASSERT(master->rq_repmsg->buflens[2] >=
2084 slave->rq_repmsg->buflens[2]);
2085 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
2086 slave->rq_repmsg->buflens[2]);
2087 LASSERT(cookie != NULL);
2089 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
2090 master->rq_repmsg->buflens[2]);
2091 LASSERT(cookie2 != NULL);
2092 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
2093 body2->valid |= OBD_MD_FLCOOKIE;
2098 static int mds_reint_unlink_remote(struct mds_update_record *rec,
2099 int offset, struct ptlrpc_request *req,
2100 struct lustre_handle *parent_lockh,
2101 int update_mode, struct dentry *dparent,
2102 struct lustre_handle *child_lockh,
2103 struct dentry *dchild)
2105 struct obd_device *obd = req->rq_export->exp_obd;
2106 struct mds_obd *mds = mds_req2mds(req);
2107 struct ptlrpc_request *request = NULL;
2108 int rc = 0, cleanup_phase = 0;
2109 struct mdc_op_data *op_data;
2113 LASSERT(offset == 1 || offset == 3);
2115 /* time to drop i_nlink on remote MDS */
2116 OBD_ALLOC(op_data, sizeof(*op_data));
2117 if (op_data == NULL)
2120 memset(op_data, 0, sizeof(*op_data));
2121 mds_pack_dentry2id(obd, &op_data->id1, dchild, 1);
2122 op_data->create_mode = rec->ur_mode;
2124 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
2125 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2127 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2128 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
2129 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2132 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2133 op_data->create_mode |= MDS_MODE_REPLAY;
2135 rc = md_unlink(mds->mds_md_exp, op_data, &request);
2140 /* with MOD this is not really needed, as cookie and EA
2141 * is not needed on client as it does not call
2142 * OST_DESTROY anymore, MDS does it. */
2143 mds_copy_unlink_reply(req, request);
2144 ptlrpc_req_finished(request);
2148 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2150 if (IS_ERR(handle)) {
2151 CERROR("can't start transaction, err %d\n",
2152 (int)PTR_ERR(handle));
2153 GOTO(cleanup, rc = PTR_ERR(handle));
2155 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
2157 CERROR("can't remove direntry: %d\n", rc);
2158 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
2161 CERROR("can't finish transno: %d\n", rc);
2166 struct dentry *chkdentry;
2167 chkdentry = ll_lookup_one_len(rec->ur_name, dparent,
2168 rec->ur_namelen - 1);
2169 if (!IS_ERR(chkdentry)) {
2170 if (chkdentry->d_inode) {
2171 /* it was cross-ref before */
2172 CDEBUG(D_ERROR, "%lu/%u:%*s(%d,%x,%p) -> %lu/%u, was %p\n",
2173 dparent->d_inode->i_ino,
2174 dparent->d_inode->i_generation,
2175 chkdentry->d_name.len,
2176 chkdentry->d_name.name,
2177 atomic_read(&chkdentry->d_count),
2178 chkdentry->d_flags, chkdentry,
2179 chkdentry->d_inode->i_ino,
2180 chkdentry->d_inode->i_generation, dchild);
2181 } else if (chkdentry->d_flags & DCACHE_CROSS_REF) {
2182 /* it's still in dcache? on a platter?! */
2183 CDEBUG(D_ERROR, "%lu/%u:%*s(%d,%x,%p) -> %u/%u, was %p\n",
2184 dparent->d_inode->i_ino,
2185 dparent->d_inode->i_generation,
2186 chkdentry->d_name.len,
2187 chkdentry->d_name.name,
2188 atomic_read(&chkdentry->d_count),
2189 chkdentry->d_flags, chkdentry,
2191 chkdentry->d_generation, dchild);
2193 /* didn't find removed dir entry - fine! */
2197 CDEBUG(D_ERROR, "child lookup error %ld\n",
2198 PTR_ERR(chkdentry));
2199 } else if (rc != -EISDIR)
2200 CERROR("error while unlinking inode "DLID4": %d\n",
2201 OLID4(&op_data->id1), rc);
2202 req->rq_status = rc;
2205 if (parent_lockh[1].cookie != 0)
2206 ldlm_lock_decref(parent_lockh + 1, update_mode);
2208 ldlm_lock_decref(child_lockh, LCK_EX);
2210 ldlm_lock_decref(parent_lockh, LCK_PW);
2212 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2216 OBD_FREE(op_data, sizeof(*op_data));
2220 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
2221 struct ptlrpc_request *req, struct lustre_handle *lh)
2223 struct dentry *dparent = NULL, *dchild;
2224 struct mds_obd *mds = mds_req2mds(req);
2225 struct obd_device *obd = req->rq_export->exp_obd;
2226 struct mds_body *body = NULL;
2227 struct inode *child_inode = NULL;
2228 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2229 struct lustre_handle child_lockh = {0};
2230 struct lustre_handle child_reuse_lockh = {0};
2231 struct lustre_handle *slave_lockh = NULL;
2232 char idname[LL_ID_NAMELEN];
2233 struct llog_create_locks *lcl = NULL;
2234 void *handle = NULL;
2235 int rc = 0, cleanup_phase = 0;
2236 int unlink_by_id = 0;
2240 LASSERT(offset == 1 || offset == 3);
2242 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
2243 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2246 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2248 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2249 DEBUG_REQ(D_HA, req, "unlink replay");
2250 LASSERT(offset == 1); /* should not come from intent */
2251 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2252 lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
2253 req->rq_repmsg->buflens[2]);
2256 MD_COUNTER_INCREMENT(obd, unlink);
2258 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
2259 GOTO(cleanup, rc = -ENOENT);
2261 if (rec->ur_namelen == 1) {
2262 /* this is request to drop i_nlink on local inode */
2264 rec->ur_name = idname;
2265 rc = mds_create_local_dentry(rec, obd);
2266 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
2267 DEBUG_REQ(D_HA, req,
2268 "drop nlink on inode "DLID4" (replay)",
2269 OLID4(rec->ur_id1));
2275 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
2276 /* master mds for directory asks slave removing inode is already
2278 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
2279 LCK_PW, parent_lockh,
2280 &update_mode, rec->ur_name,
2282 MDS_INODELOCK_UPDATE);
2283 if (IS_ERR(dparent))
2284 GOTO(cleanup, rc = PTR_ERR(dparent));
2285 dchild = ll_lookup_one_len(rec->ur_name, dparent,
2286 rec->ur_namelen - 1);
2288 GOTO(cleanup, rc = PTR_ERR(dchild));
2289 child_lockh.cookie = 0;
2290 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
2291 LASSERT(dchild->d_inode != NULL);
2292 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
2294 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
2295 parent_lockh, &dparent,
2296 LCK_PW, MDS_INODELOCK_UPDATE,
2297 &update_mode, rec->ur_name,
2298 rec->ur_namelen, &child_lockh,
2300 (MDS_INODELOCK_LOOKUP |
2301 MDS_INODELOCK_UPDATE));
2306 if (dchild->d_flags & DCACHE_CROSS_REF) {
2307 /* we should have parent lock only here */
2308 LASSERT(unlink_by_id == 0);
2309 LASSERT(dchild->d_mdsnum != mds->mds_num);
2310 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
2311 update_mode, dparent, &child_lockh, dchild);
2315 cleanup_phase = 1; /* dchild, dparent, locks */
2318 child_inode = dchild->d_inode;
2319 if (child_inode == NULL) {
2320 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
2321 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
2322 GOTO(cleanup, rc = -ENOENT);
2325 cleanup_phase = 2; /* dchild has a lock */
2327 /* We have to do these checks ourselves, in case we are making an
2328 * orphan. The client tells us whether rmdir() or unlink() was called,
2329 * so we need to return appropriate errors (bug 72).
2331 * We don't have to check permissions, because vfs_rename (called from
2332 * mds_open_unlink_rename) also calls may_delete. */
2333 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
2334 if (!S_ISDIR(child_inode->i_mode))
2335 GOTO(cleanup, rc = -ENOTDIR);
2337 if (S_ISDIR(child_inode->i_mode))
2338 GOTO(cleanup, rc = -EISDIR);
2341 /* handle splitted dir */
2342 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
2346 /* Step 4: Get a lock on the ino to sync with creation WRT inode
2347 * reuse (see bug 2029). */
2348 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
2351 cleanup_phase = 3; /* child inum lock */
2353 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2355 /* ldlm_reply in buf[0] if called via intent */
2361 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2362 LASSERT(body != NULL);
2364 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2365 DOWN_READ_I_ALLOC_SEM(child_inode);
2366 cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2368 /* If this is potentially the last reference to this inode, get the
2369 * OBD EA data first so the client can destroy OST objects. We
2370 * only do the object removal later if no open files/links remain. */
2371 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2372 child_inode->i_nlink == 1) {
2373 if (mds_orphan_open_count(child_inode) > 0) {
2374 /* need to lock pending_dir before transaction */
2375 down(&mds->mds_pending_dir->d_inode->i_sem);
2376 cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2377 } else if (S_ISREG(child_inode->i_mode)) {
2378 mds_pack_inode2body(obd, body, child_inode, 0);
2379 mds_pack_md(obd, req->rq_repmsg, offset + 1,
2380 body, child_inode, MDS_PACK_MD_LOCK, 0);
2384 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2385 switch (child_inode->i_mode & S_IFMT) {
2387 /* Drop any lingering child directories before we start our
2388 * transaction, to avoid doing multiple inode dirty/delete
2389 * in our compound transaction (bug 1321). */
2390 shrink_dcache_parent(dchild);
2391 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2393 if (IS_ERR(handle)) {
2394 CERROR("can't start transaction, err %d\n",
2395 (int)PTR_ERR(handle));
2396 GOTO(cleanup, rc = PTR_ERR(handle));
2398 rc = vfs_rmdir(dparent->d_inode, dchild);
2401 /* optimization is possible here: we could drop nlink
2402 * w/o removing local dentry in FIDS */
2403 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2405 handle = fsfilt_start_log(obd, dparent->d_inode,
2406 FSFILT_OP_UNLINK, NULL,
2407 le32_to_cpu(lmm->lmm_stripe_count));
2409 GOTO(cleanup, rc = PTR_ERR(handle));
2410 rc = vfs_unlink(dparent->d_inode, dchild);
2418 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2421 GOTO(cleanup, rc = PTR_ERR(handle));
2422 rc = vfs_unlink(dparent->d_inode, dchild);
2425 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2428 GOTO(cleanup, rc = -EINVAL);
2431 if (rc == 0 && child_inode->i_nlink == 0) {
2432 if (mds_orphan_open_count(child_inode) > 0)
2433 rc = mds_orphan_add_link(rec, obd, dchild);
2436 GOTO(cleanup, rc = 0);
2438 if (!S_ISREG(child_inode->i_mode))
2441 if (!(body->valid & OBD_MD_FLEASIZE)) {
2442 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2443 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2444 } else if (mds_log_op_unlink(obd, child_inode,
2445 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2446 req->rq_repmsg->buflens[offset + 1],
2447 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2448 req->rq_repmsg->buflens[offset + 2],
2450 body->valid |= OBD_MD_FLCOOKIE;
2453 rc = mds_destroy_object(obd, child_inode, 1);
2455 CERROR("can't remove OST object, err %d\n",
2459 if (child_inode->i_nlink == 0)
2460 mds_fidmap_del(obd, &body->id1);
2470 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2471 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2472 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2474 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2476 CERROR("error on parent setattr: rc = %d\n", err);
2479 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2480 handle, req, rc, 0);
2482 (void)obd_set_info(mds->mds_dt_exp, strlen("unlinked"),
2483 "unlinked", 0, NULL);
2485 switch(cleanup_phase) {
2486 case 5: /* pending_dir semaphore */
2487 up(&mds->mds_pending_dir->d_inode->i_sem);
2488 case 4: /* child inode semaphore */
2489 UP_READ_I_ALLOC_SEM(child_inode);
2490 /* handle splitted dir */
2492 /* master directory can be non-empty or something else ... */
2493 mds_unlink_slave_objs(obd, dchild);
2496 ptlrpc_save_llog_lock(req, lcl);
2497 case 3: /* child ino-reuse lock */
2498 if (rc && body != NULL) {
2499 // Don't unlink the OST objects if the MDS unlink failed
2503 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2505 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2506 case 2: /* child lock */
2507 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2508 if (child_lockh.cookie)
2509 ldlm_lock_decref(&child_lockh, LCK_EX);
2510 case 1: /* child and parent dentry, parent lock */
2512 if (parent_lockh[1].cookie != 0)
2513 ldlm_lock_decref(parent_lockh + 1, update_mode);
2516 ldlm_lock_decref(parent_lockh, LCK_PW);
2518 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2519 if (dchild->d_inode && rc && (dchild->d_inode->i_nlink == 0 ||
2520 mds_inode_is_orphan(dchild->d_inode)))
2521 CDEBUG(D_ERROR, "unlink, but return %d\n", rc);
2523 /* catching failed permissions check for audit */
2525 mds_audit(req, dchild, NULL, 0, AUDIT_UNLINK, rc);
2533 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2536 req->rq_status = rc;
2541 * to service requests from remote MDS to increment i_nlink
2543 static int mds_reint_link_acquire(struct mds_update_record *rec,
2544 int offset, struct ptlrpc_request *req,
2545 struct lustre_handle *lh)
2547 struct obd_device *obd = req->rq_export->exp_obd;
2548 struct ldlm_res_id src_res_id = { .name = {0} };
2549 struct lustre_handle *handle = NULL, src_lockh = {0};
2550 struct mds_obd *mds = mds_req2mds(req);
2551 int rc = 0, cleanup_phase = 0;
2552 struct dentry *de_src = NULL;
2553 ldlm_policy_data_t policy;
2554 int flags = LDLM_FL_ATOMIC_CB;
2557 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2558 obd->obd_name, OLID4(rec->ur_id1));
2560 /* Step 1: Lookup the source inode and target directory by ID */
2561 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2563 GOTO(cleanup, rc = PTR_ERR(de_src));
2564 cleanup_phase = 1; /* source dentry */
2566 src_res_id.name[0] = id_fid(rec->ur_id1);
2567 src_res_id.name[1] = id_group(rec->ur_id1);
2568 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2570 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2571 src_res_id, LDLM_IBITS, &policy,
2572 LCK_EX, &flags, mds_blocking_ast,
2573 ldlm_completion_ast, NULL, NULL,
2574 NULL, 0, NULL, &src_lockh);
2576 GOTO(cleanup, rc = -ENOLCK);
2577 cleanup_phase = 2; /* lock */
2579 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2581 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2582 if (IS_ERR(handle)) {
2583 rc = PTR_ERR(handle);
2586 i_nlink_inc(de_src->d_inode);
2587 mark_inode_dirty(de_src->d_inode);
2591 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2592 handle, req, rc, 0);
2593 switch (cleanup_phase) {
2596 ldlm_lock_decref(&src_lockh, LCK_EX);
2598 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2604 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2607 req->rq_status = rc;
2612 * request to link to foreign inode:
2613 * - acquire i_nlinks on this inode
2616 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2617 int offset, struct ptlrpc_request *req,
2618 struct lustre_handle *lh)
2620 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2621 struct obd_device *obd = req->rq_export->exp_obd;
2622 struct dentry *de_tgt_dir = NULL;
2623 struct mds_obd *mds = mds_req2mds(req);
2624 int rc = 0, cleanup_phase = 0;
2625 struct mdc_op_data *op_data;
2626 struct ptlrpc_request *request = NULL;
2630 DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2631 ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2632 OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2633 OLID4(rec->ur_id1));
2635 de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2636 tgt_dir_lockh, &update_mode,
2637 rec->ur_name, rec->ur_namelen - 1,
2638 MDS_INODELOCK_UPDATE);
2639 if (IS_ERR(de_tgt_dir))
2640 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2643 OBD_ALLOC(op_data, sizeof(*op_data));
2644 if (op_data == NULL)
2645 GOTO(cleanup, rc = -ENOMEM);
2647 memset(op_data, 0, sizeof(*op_data));
2648 op_data->id1 = *(rec->ur_id1);
2649 rc = md_link(mds->mds_md_exp, op_data, &request);
2650 OBD_FREE(op_data, sizeof(*op_data));
2653 ptlrpc_req_finished(request);
2659 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2661 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2662 if (IS_ERR(handle)) {
2663 rc = PTR_ERR(handle);
2669 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2670 rec->ur_namelen - 1, id_ino(rec->ur_id1),
2671 id_gen(rec->ur_id1), id_group(rec->ur_id1),
2672 id_fid(rec->ur_id1));
2675 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2676 handle, req, rc, 0);
2678 switch (cleanup_phase) {
2681 OBD_ALLOC(op_data, sizeof(*op_data));
2682 if (op_data != NULL) {
2684 memset(op_data, 0, sizeof(*op_data));
2686 op_data->id1 = *(rec->ur_id1);
2687 op_data->create_mode = rec->ur_mode;
2689 rc = md_unlink(mds->mds_md_exp, op_data, &request);
2690 OBD_FREE(op_data, sizeof(*op_data));
2692 ptlrpc_req_finished(request);
2694 CERROR("error %d while dropping i_nlink on "
2695 "remote inode\n", rc);
2698 CERROR("rc %d prevented dropping i_nlink on "
2699 "remote inode\n", -ENOMEM);
2705 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2707 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2710 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2712 ptlrpc_save_lock(req, tgt_dir_lockh + 1, update_mode);
2718 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2721 req->rq_status = rc;
2725 static int mds_reint_link(struct mds_update_record *rec, int offset,
2726 struct ptlrpc_request *req, struct lustre_handle *lh)
2728 struct obd_device *obd = req->rq_export->exp_obd;
2729 struct dentry *de_src = NULL;
2730 struct dentry *de_tgt_dir = NULL;
2731 struct dentry *dchild = NULL;
2732 struct mds_obd *mds = mds_req2mds(req);
2733 struct lustre_handle *handle = NULL;
2734 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2735 struct ldlm_res_id src_res_id = { .name = {0} };
2736 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2737 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2738 ldlm_policy_data_t tgt_dir_policy =
2739 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2740 int rc = 0, cleanup_phase = 0;
2742 int update_mode = 0;
2746 LASSERT(offset == 1);
2748 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2749 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2750 id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2753 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2754 MD_COUNTER_INCREMENT(obd, link);
2756 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2757 GOTO(cleanup, rc = -ENOENT);
2759 if (id_group(rec->ur_id1) != mds->mds_num) {
2760 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2764 if (rec->ur_namelen == 1) {
2765 rc = mds_reint_link_acquire(rec, offset, req, lh);
2769 /* Step 1: Lookup the source inode and target directory by ID */
2770 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2772 GOTO(cleanup, rc = PTR_ERR(de_src));
2774 cleanup_phase = 1; /* source dentry */
2776 de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2777 if (IS_ERR(de_tgt_dir)) {
2778 rc = PTR_ERR(de_tgt_dir);
2783 cleanup_phase = 2; /* target directory dentry */
2785 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2786 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2787 rec->ur_name, de_src->d_inode->i_ino);
2789 /* Step 2: Take the two locks */
2790 src_res_id.name[0] = id_fid(rec->ur_id1);
2791 src_res_id.name[1] = id_group(rec->ur_id1);
2792 tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2793 tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2796 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2797 int flags = LDLM_FL_ATOMIC_CB;
2798 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2800 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2801 tgt_dir_res_id, LDLM_IBITS,
2802 &src_policy, update_mode, &flags,
2804 ldlm_completion_ast, NULL, NULL,
2805 NULL, 0, NULL, tgt_dir_lockh + 1);
2807 GOTO(cleanup, rc = -ENOLCK);
2810 tgt_dir_res_id.name[2] = full_name_hash((unsigned char *)rec->ur_name,
2811 rec->ur_namelen - 1);
2812 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2813 (unsigned long)id_fid(rec->ur_id2),
2814 (unsigned long)id_group(rec->ur_id2),
2815 tgt_dir_res_id.name[2]);
2818 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2819 &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2820 LCK_EX, &tgt_dir_policy);
2824 cleanup_phase = 3; /* locks */
2826 /* Step 3: Lookup the child */
2827 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2828 rec->ur_namelen - 1);
2829 if (IS_ERR(dchild)) {
2830 rc = PTR_ERR(dchild);
2831 if (rc != -EPERM && rc != -EACCES)
2832 CERROR("child lookup error %d\n", rc);
2836 cleanup_phase = 4; /* child dentry */
2838 if (dchild->d_inode) {
2839 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2840 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2845 /* Step 4: Do it. */
2846 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2848 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2849 if (IS_ERR(handle)) {
2850 rc = PTR_ERR(handle);
2854 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2855 if (rc && rc != -EPERM && rc != -EACCES)
2856 CERROR("vfs_link error %d\n", rc);
2858 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2859 handle, req, rc, 0);
2862 switch (cleanup_phase) {
2863 case 4: /* child dentry */
2867 ldlm_lock_decref(&src_lockh, LCK_EX);
2868 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2870 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2871 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2873 case 2: /* target dentry */
2875 if (tgt_dir_lockh[1].cookie && update_mode)
2876 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2880 case 1: /* source dentry */
2885 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2888 req->rq_status = rc;
2892 /* The idea here is that we need to get four locks in the end:
2893 * one on each parent directory, one on each child. We need to take
2894 * these locks in some kind of order (to avoid deadlocks), and the order
2895 * I selected is "increasing resource number" order. We need to look up
2896 * the children, however, before we know what the resource number(s) are.
2897 * Thus the following plan:
2899 * 1,2. Look up the parents
2900 * 3,4. Look up the children
2901 * 5. Take locks on the parents and children, in order
2902 * 6. Verify that the children haven't changed since they were looked up
2904 * If there was a race and the children changed since they were first looked
2905 * up, it is possible that mds_verify_child() will be able to just grab the
2906 * lock on the new child resource (if it has a higher resource than any other)
2907 * but we need to compare against not only its parent, but also against the
2908 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2910 * We need the fancy igrab() on the child inodes because we aren't holding a
2911 * lock on the parent after the lookup is done, so dentry->d_inode may change
2912 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2914 static int mds_get_parents_children_locked(struct obd_device *obd,
2915 struct mds_obd *mds,
2916 struct lustre_id *p1_id,
2917 struct dentry **de_srcdirp,
2918 struct lustre_id *p2_id,
2919 struct dentry **de_tgtdirp,
2921 const char *old_name, int old_len,
2922 struct dentry **de_oldp,
2923 const char *new_name, int new_len,
2924 struct dentry **de_newp,
2925 struct lustre_handle *dlm_handles,
2928 struct ldlm_res_id p1_res_id = { .name = {0} };
2929 struct ldlm_res_id p2_res_id = { .name = {0} };
2930 struct ldlm_res_id c1_res_id = { .name = {0} };
2931 struct ldlm_res_id c2_res_id = { .name = {0} };
2932 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2933 /* Only dentry should disappear, but the inode itself would be
2934 intact otherwise. */
2935 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2936 /* If something is going to be replaced, both dentry and inode locks are
2938 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2939 struct ldlm_res_id *maxres_src, *maxres_tgt;
2940 struct inode *inode;
2941 int rc = 0, cleanup_phase = 0;
2942 __u32 child_gen1 = 0;
2943 __u32 child_gen2 = 0;
2944 unsigned long child_ino1 = 0;
2945 unsigned long child_ino2 = 0;
2948 /* Step 1: Lookup the source directory */
2949 *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2950 if (IS_ERR(*de_srcdirp))
2951 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2953 cleanup_phase = 1; /* source directory dentry */
2955 p1_res_id.name[0] = id_fid(p1_id);
2956 p1_res_id.name[1] = id_group(p1_id);
2958 /* Step 2: Lookup the target directory */
2959 if (id_equal_stc(p1_id, p2_id)) {
2960 *de_tgtdirp = dget(*de_srcdirp);
2962 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2963 if (IS_ERR(*de_tgtdirp)) {
2964 rc = PTR_ERR(*de_tgtdirp);
2970 cleanup_phase = 2; /* target directory dentry */
2972 p2_res_id.name[0] = id_fid(p2_id);
2973 p2_res_id.name[1] = id_group(p2_id);
2976 dlm_handles[5].cookie = 0;
2977 dlm_handles[6].cookie = 0;
2979 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2981 * get a temp lock on just fid, group to flush client cache and
2982 * to protect dirs from concurrent splitting.
2984 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2985 LCK_PW, &p_policy, &p2_res_id,
2986 &dlm_handles[6], LCK_PW, &p_policy);
2990 p1_res_id.name[2] = full_name_hash((unsigned char *)old_name,
2992 p2_res_id.name[2] = full_name_hash((unsigned char *)new_name,
2995 CDEBUG(D_INFO, "take locks on "
2996 LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2997 p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2998 p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
3003 /* Step 3: Lookup the source child entry */
3004 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
3006 if (IS_ERR(*de_oldp)) {
3007 rc = PTR_ERR(*de_oldp);
3008 CERROR("old child lookup error (%.*s): %d\n",
3009 old_len - 1, old_name, rc);
3013 cleanup_phase = 4; /* original name dentry */
3015 inode = (*de_oldp)->d_inode;
3016 if (inode != NULL) {
3017 struct lustre_id sid;
3019 inode = igrab(inode);
3021 GOTO(cleanup, rc = -ENOENT);
3023 down(&inode->i_sem);
3024 rc = mds_read_inode_sid(obd, inode, &sid);
3027 CERROR("Can't read inode self id, inode %lu, "
3028 "rc %d\n", inode->i_ino, rc);
3033 child_ino1 = inode->i_ino;
3034 child_gen1 = inode->i_generation;
3035 c1_res_id.name[0] = id_fid(&sid);
3036 c1_res_id.name[1] = id_group(&sid);
3038 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
3039 child_ino1 = (*de_oldp)->d_inum;
3040 child_gen1 = (*de_oldp)->d_generation;
3041 c1_res_id.name[0] = (*de_oldp)->d_fid;
3042 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
3044 GOTO(cleanup, rc = -ENOENT);
3047 /* Step 4: Lookup the target child entry */
3048 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
3050 if (IS_ERR(*de_newp)) {
3051 rc = PTR_ERR(*de_newp);
3052 CERROR("new child lookup error (%.*s): %d\n",
3053 old_len - 1, old_name, rc);
3057 cleanup_phase = 5; /* target dentry */
3059 inode = (*de_newp)->d_inode;
3060 if (inode != NULL) {
3061 struct lustre_id sid;
3063 inode = igrab(inode);
3067 down(&inode->i_sem);
3068 rc = mds_read_inode_sid(obd, inode, &sid);
3071 CERROR("Can't read inode self id, inode %lu, "
3072 "rc %d\n", inode->i_ino, rc);
3076 child_ino2 = inode->i_ino;
3077 child_gen2 = inode->i_generation;
3078 c2_res_id.name[0] = id_fid(&sid);
3079 c2_res_id.name[1] = id_group(&sid);
3081 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
3082 child_ino2 = (*de_newp)->d_inum;
3083 child_gen2 = (*de_newp)->d_generation;
3084 c2_res_id.name[0] = (*de_newp)->d_fid;
3085 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
3089 /* Step 5: Take locks on the parents and child(ren) */
3090 maxres_src = &p1_res_id;
3091 maxres_tgt = &p2_res_id;
3092 cleanup_phase = 5; /* target dentry */
3094 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
3095 maxres_src = &c1_res_id;
3096 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
3097 maxres_tgt = &c2_res_id;
3099 rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
3101 &p2_res_id, &dlm_handles[1], parent_mode,
3103 &c1_res_id, &dlm_handles[2], child_mode,
3105 &c2_res_id, &dlm_handles[3], child_mode,
3110 cleanup_phase = 6; /* parent and child(ren) locks */
3112 /* Step 6a: Re-lookup source child to verify it hasn't changed */
3113 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
3114 parent_mode, &c1_res_id, &dlm_handles[2],
3115 de_oldp, child_mode, &c1_policy, old_name, old_len,
3116 maxres_tgt, child_ino1, child_gen1);
3118 if (c2_res_id.name[0] != 0)
3119 ldlm_lock_decref(&dlm_handles[3], child_mode);
3120 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3127 if (!DENTRY_VALID(*de_oldp))
3128 GOTO(cleanup, rc = -ENOENT);
3130 /* Step 6b: Re-lookup target child to verify it hasn't changed */
3131 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
3132 parent_mode, &c2_res_id, &dlm_handles[3],
3133 de_newp, child_mode, &c2_policy, new_name,
3134 new_len, maxres_src, child_ino2, child_gen2);
3136 ldlm_lock_decref(&dlm_handles[2], child_mode);
3137 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3147 switch (cleanup_phase) {
3148 case 6: /* child lock(s) */
3149 if (c2_res_id.name[0] != 0)
3150 ldlm_lock_decref(&dlm_handles[3], child_mode);
3151 if (c1_res_id.name[0] != 0)
3152 ldlm_lock_decref(&dlm_handles[2], child_mode);
3153 if (dlm_handles[1].cookie != 0)
3154 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3155 if (dlm_handles[0].cookie != 0)
3156 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3157 case 5: /* target dentry */
3159 case 4: /* source dentry */
3163 if (dlm_handles[5].cookie != 0)
3164 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3165 if (dlm_handles[6].cookie != 0)
3166 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3168 case 2: /* target directory dentry */
3169 l_dput(*de_tgtdirp);
3170 case 1: /* source directry dentry */
3171 l_dput(*de_srcdirp);
3179 * checks if dentry can be removed. This function also handles cross-ref
3182 static int mds_check_for_rename(struct obd_device *obd,
3183 struct dentry *dentry)
3185 struct mds_obd *mds = &obd->u.mds;
3186 struct lustre_handle *rlockh;
3187 struct ptlrpc_request *req;
3188 struct mdc_op_data *op_data;
3189 struct lookup_intent it;
3190 int handle_size, rc = 0;
3193 LASSERT(dentry != NULL);
3195 if (dentry->d_inode) {
3196 if (S_ISDIR(dentry->d_inode->i_mode) &&
3197 !mds_is_dir_empty(obd, dentry))
3200 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
3201 handle_size = sizeof(struct lustre_handle);
3203 OBD_ALLOC(rlockh, handle_size);
3207 memset(rlockh, 0, handle_size);
3208 OBD_ALLOC(op_data, sizeof(*op_data));
3209 if (op_data == NULL) {
3210 OBD_FREE(rlockh, handle_size);
3213 memset(op_data, 0, sizeof(*op_data));
3214 mds_pack_dentry2id(obd, &op_data->id1, dentry, 1);
3216 it.it_op = IT_UNLINK;
3217 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
3220 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
3221 op_data, rlockh, NULL, 0, ldlm_completion_ast,
3222 mds_blocking_ast, NULL);
3223 OBD_FREE(op_data, sizeof(*op_data));
3227 OBD_FREE(it.d.fs_data,
3228 sizeof(struct lustre_intent_data));
3231 if (rlockh->cookie != 0)
3232 ldlm_lock_decref(rlockh, LCK_EX);
3234 if (LUSTRE_IT(&it)->it_data) {
3235 req = (struct ptlrpc_request *)LUSTRE_IT(&it)->it_data;
3236 ptlrpc_req_finished(req);
3239 if (LUSTRE_IT(&it)->it_status)
3240 rc = LUSTRE_IT(&it)->it_status;
3241 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
3242 OBD_FREE(rlockh, handle_size);
3247 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
3248 struct ptlrpc_request *req, struct lustre_id *id,
3249 struct dentry *de_dir, struct dentry *de)
3251 struct obd_device *obd = req->rq_export->exp_obd;
3252 struct mds_obd *mds = mds_req2mds(req);
3253 void *handle = NULL;
3259 * name exists and points to local inode try to unlink this name
3260 * and create new one.
3262 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
3263 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
3264 (unsigned long)de->d_inode->i_generation);
3266 /* checking if we can remove local dentry. */
3267 rc = mds_check_for_rename(obd, de);
3271 handle = fsfilt_start(obd, de_dir->d_inode,
3272 FSFILT_OP_RENAME, NULL);
3274 GOTO(cleanup, rc = PTR_ERR(handle));
3275 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3278 } else if (de->d_flags & DCACHE_CROSS_REF) {
3279 CDEBUG(D_OTHER, "%s: %s points to remote inode %lu/%lu\n",
3280 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_mdsnum,
3281 (unsigned long)de->d_fid);
3283 /* checking if we can remove local dentry. */
3284 rc = mds_check_for_rename(obd, de);
3289 * to be fully POSIX compatible, we should add one more check:
3291 * if de_new is subdir of dir rec->ur_id1. If so - return
3294 * I do not know how to implement it right now, because
3295 * inodes/dentries for new and old names lie on different MDS,
3296 * so add this notice here just to make it visible for the rest
3297 * of developers and do not forget about. And when this check
3298 * will be added, del_cross_ref should gone, that is local
3299 * dentry is able to be removed if all checks passed. --umka
3302 handle = fsfilt_start(obd, de_dir->d_inode,
3303 FSFILT_OP_RENAME, NULL);
3305 GOTO(cleanup, rc = PTR_ERR(handle));
3306 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3310 /* name doesn't exist. the simplest case. */
3311 handle = fsfilt_start(obd, de_dir->d_inode,
3312 FSFILT_OP_LINK, NULL);
3314 GOTO(cleanup, rc = PTR_ERR(handle));
3317 rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
3318 rec->ur_tgtlen - 1, id_ino(id),
3319 id_gen(id), id_group(id), id_fid(id));
3321 CERROR("add_dir_entry() returned error %d\n", rc);
3327 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3328 handle, req, rc, 0);
3333 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
3334 struct ptlrpc_request *req, struct dentry *de_dir,
3337 struct obd_device *obd = req->rq_export->exp_obd;
3338 struct mds_obd *mds = mds_req2mds(req);
3339 void *handle = NULL;
3343 handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
3345 GOTO(cleanup, rc = PTR_ERR(handle));
3346 rc = fsfilt_del_dir_entry(obd, de);
3351 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3352 handle, req, rc, 0);
3356 static int mds_reint_rename_create_name(struct mds_update_record *rec,
3357 int offset, struct ptlrpc_request *req)
3359 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3360 struct obd_device *obd = req->rq_export->exp_obd;
3361 struct mds_obd *mds = mds_req2mds(req);
3362 struct lustre_handle child_lockh = {0};
3363 struct dentry *de_tgtdir = NULL;
3364 struct dentry *de_new = NULL;
3365 int cleanup_phase = 0;
3366 int update_mode, rc = 0;
3367 struct lustre_id ids[2]; /* sid, pid */
3368 struct obd_export *set_exp;
3372 * another MDS executing rename operation has asked us to create target
3373 * name. such a creation should destroy existing target name.
3375 CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
3376 obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
3378 /* first, lookup the target */
3379 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
3380 &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
3381 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
3382 &child_lockh, &de_new, LCK_EX,
3383 MDS_INODELOCK_LOOKUP);
3387 /* get parent id: ldlm lock on the parent protects ea */
3388 rc = mds_read_inode_sid(obd, de_tgtdir->d_inode, &ids[1]);
3395 LASSERT(de_tgtdir->d_inode);
3398 rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
3403 ids[0] = *(rec->ur_id1);
3404 if (id_group(ids) == mds->mds_num)
3405 set_exp = req->rq_export;
3407 set_exp = mds->mds_md_exp;
3408 rc = obd_set_info(set_exp, strlen("ids"), "ids",
3409 sizeof(struct lustre_id) * 2, ids);
3414 if (cleanup_phase == 1) {
3416 if (parent_lockh[1].cookie != 0)
3417 ldlm_lock_decref(parent_lockh + 1, update_mode);
3419 ldlm_lock_decref(parent_lockh, LCK_PW);
3420 if (child_lockh.cookie != 0)
3421 ldlm_lock_decref(&child_lockh, LCK_EX);
3426 req->rq_status = rc;
3430 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3431 struct ptlrpc_request *req)
3433 struct obd_device *obd = req->rq_export->exp_obd;
3434 struct ptlrpc_request *req2 = NULL;
3435 struct dentry *de_srcdir = NULL;
3436 struct dentry *de_old = NULL;
3437 struct mds_obd *mds = mds_req2mds(req);
3438 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3439 struct lustre_handle child_lockh = {0};
3440 struct mdc_op_data *op_data;
3441 int update_mode, rc = 0;
3444 CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3445 obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3447 OBD_ALLOC(op_data, sizeof(*op_data));
3448 if (op_data == NULL)
3450 memset(op_data, 0, sizeof(*op_data));
3452 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3453 &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3454 &update_mode, rec->ur_name,
3455 rec->ur_namelen, &child_lockh, &de_old,
3456 LCK_EX, MDS_INODELOCK_LOOKUP);
3458 OBD_FREE(op_data, sizeof(*op_data));
3463 LASSERT(de_srcdir->d_inode);
3467 * we already know the target should be created on another MDS so, we
3468 * have to request that MDS to do it.
3471 /* prepare source id */
3472 if (de_old->d_flags & DCACHE_CROSS_REF) {
3473 LASSERT(de_old->d_inode == NULL);
3474 CDEBUG(D_OTHER, "request to move remote name\n");
3475 mds_pack_dentry2id(obd, &op_data->id1, de_old, 1);
3476 } else if (de_old->d_inode == NULL) {
3477 /* oh, source doesn't exist */
3478 OBD_FREE(op_data, sizeof(*op_data));
3479 GOTO(cleanup, rc = -ENOENT);
3481 struct lustre_id sid;
3482 struct inode *inode = de_old->d_inode;
3484 LASSERT(inode != NULL);
3485 CDEBUG(D_OTHER, "request to move local name\n");
3486 id_ino(&op_data->id1) = inode->i_ino;
3487 id_group(&op_data->id1) = mds->mds_num;
3488 id_gen(&op_data->id1) = inode->i_generation;
3490 down(&inode->i_sem);
3491 rc = mds_read_inode_sid(obd, inode, &sid);
3494 CERROR("Can't read inode self id, "
3495 "inode %lu, rc = %d\n",
3500 id_fid(&op_data->id1) = id_fid(&sid);
3503 op_data->id2 = *rec->ur_id2;
3504 rc = md_rename(mds->mds_md_exp, op_data, NULL, 0,
3505 rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3506 OBD_FREE(op_data, sizeof(*op_data));
3511 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3517 ptlrpc_req_finished(req2);
3520 if (parent_lockh[1].cookie != 0)
3521 ldlm_lock_decref(parent_lockh + 1, update_mode);
3523 ldlm_lock_decref(parent_lockh, LCK_PW);
3524 if (child_lockh.cookie != 0)
3525 ldlm_lock_decref(&child_lockh, LCK_EX);
3530 req->rq_status = rc;
3534 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3535 struct ptlrpc_request *req, struct lustre_handle *lockh)
3537 struct obd_device *obd = req->rq_export->exp_obd;
3538 struct dentry *de_srcdir = NULL;
3539 struct dentry *de_tgtdir = NULL;
3540 struct dentry *de_old = NULL;
3541 struct dentry *de_new = NULL;
3542 struct inode *old_inode = NULL, *new_inode = NULL;
3543 struct mds_obd *mds = mds_req2mds(req);
3544 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3545 struct mds_body *body = NULL;
3546 struct llog_create_locks *lcl = NULL;
3547 struct lov_mds_md *lmm = NULL;
3548 int rc = 0, cleanup_phase = 0;
3549 struct lustre_id ids[2]; /* sid, pid */
3550 void *handle = NULL;
3553 LASSERT(offset == 1);
3555 DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3556 OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3559 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3561 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3562 DEBUG_REQ(D_HA, req, "rename replay");
3563 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3564 lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3565 req->rq_repmsg->buflens[2]);
3568 MD_COUNTER_INCREMENT(obd, rename);
3570 if (rec->ur_namelen == 1) {
3571 rc = mds_reint_rename_create_name(rec, offset, req);
3575 /* check if new name should be located on remote target. */
3576 if (id_group(rec->ur_id2) != mds->mds_num) {
3577 rc = mds_reint_rename_to_remote(rec, offset, req);
3581 rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3582 rec->ur_id2, &de_tgtdir, LCK_PW,
3583 rec->ur_name, rec->ur_namelen,
3584 &de_old, rec->ur_tgt,
3585 rec->ur_tgtlen, &de_new,
3586 dlm_handles, LCK_EX);
3590 cleanup_phase = 1; /* parent(s), children, locks */
3591 old_inode = de_old->d_inode;
3592 new_inode = de_new->d_inode;
3594 /* sanity check for src inode */
3595 if (de_old->d_flags & DCACHE_CROSS_REF) {
3596 LASSERT(de_old->d_inode == NULL);
3599 * in the case of cross-ref dir, we can perform this check only
3600 * if child and parent lie on the same mds. This is because
3601 * otherwise they can have the same inode numbers.
3603 if (de_old->d_mdsnum == mds->mds_num) {
3604 if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3605 de_old->d_inum == de_tgtdir->d_inode->i_ino)
3606 GOTO(cleanup, rc = -EINVAL);
3609 LASSERT(de_old->d_inode != NULL);
3610 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3611 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3612 GOTO(cleanup, rc = -EINVAL);
3615 /* sanity check for dest inode */
3616 if (de_new->d_flags & DCACHE_CROSS_REF) {
3617 LASSERT(new_inode == NULL);
3619 /* the same check about target dentry. */
3620 if (de_new->d_mdsnum == mds->mds_num) {
3621 if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3622 de_new->d_inum == de_tgtdir->d_inode->i_ino)
3623 GOTO(cleanup, rc = -EINVAL);
3627 * regular files usualy do not have ->rename() implemented. But
3628 * we handle only this case when @de_new is cross-ref entry,
3629 * because in other cases it will be handled by vfs_rename().
3631 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3632 !de_old->d_inode->i_op->rename))
3633 GOTO(cleanup, rc = -EPERM);
3636 (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3637 new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3638 GOTO(cleanup, rc = -EINVAL);
3643 * check if inodes point to each other. This should be checked before
3644 * is_subdir() check, as for the same entries it will think that they
3647 if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3648 !(de_new->d_flags & DCACHE_CROSS_REF) &&
3649 old_inode == new_inode)
3650 GOTO(cleanup, rc = 0);
3652 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3654 * check if we are moving old entry into its child. 2.6 does not check
3655 * for this in vfs_rename() anymore.
3657 if (is_subdir(de_new, de_old))
3658 GOTO(cleanup, rc = -EINVAL);
3662 * if we are about to remove the target at first, pass the EA of that
3663 * inode to client to perform and cleanup on OST.
3665 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3666 LASSERT(body != NULL);
3668 /* get new parent id: ldlm lock on the parent protects ea */
3669 rc = mds_read_inode_sid(obd, de_tgtdir->d_inode, &ids[1]);
3673 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3675 DOWN_READ_I_ALLOC_SEM(new_inode);
3677 cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3679 if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3680 new_inode->i_nlink == 2) ||
3681 new_inode->i_nlink == 1)) {
3682 if (mds_orphan_open_count(new_inode) > 0) {
3683 /* need to lock pending_dir before transaction */
3684 down(&mds->mds_pending_dir->d_inode->i_sem);
3685 cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3686 } else if (S_ISREG(new_inode->i_mode)) {
3687 mds_pack_inode2body(obd, body, new_inode, 0);
3688 mds_pack_md(obd, req->rq_repmsg, 1, body,
3689 new_inode, MDS_PACK_MD_LOCK, 0);
3693 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3694 de_srcdir->d_inode->i_sb);
3696 if (de_old->d_flags & DCACHE_CROSS_REF) {
3697 struct lustre_id old_id;
3698 struct obd_export *set_exp;
3701 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3703 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3708 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3714 if (id_group(ids) == mds->mds_num)
3715 set_exp = req->rq_export;
3717 set_exp = mds->mds_md_exp;
3718 rc = obd_set_info(set_exp, strlen("ids"), "ids",
3719 sizeof(struct lustre_id) * 2, ids);
3724 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3725 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3726 NULL, le32_to_cpu(lmm->lmm_stripe_count));
3729 GOTO(cleanup, rc = PTR_ERR(handle));
3732 de_old->d_fsdata = req;
3733 de_new->d_fsdata = req;
3734 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3737 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3738 if (mds_orphan_open_count(new_inode) > 0)
3739 rc = mds_orphan_add_link(rec, obd, de_new);
3742 GOTO(cleanup, rc = 0);
3744 if (!S_ISREG(new_inode->i_mode))
3747 if (!(body->valid & OBD_MD_FLEASIZE)) {
3748 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3749 OBD_MD_FLATIME | OBD_MD_FLMTIME);
3750 } else if (mds_log_op_unlink(obd, new_inode,
3751 lustre_msg_buf(req->rq_repmsg,1,0),
3752 req->rq_repmsg->buflens[1],
3753 lustre_msg_buf(req->rq_repmsg,2,0),
3754 req->rq_repmsg->buflens[2],
3756 body->valid |= OBD_MD_FLCOOKIE;
3759 rc = mds_destroy_object(obd, old_inode, 1);
3761 CERROR("can't remove OST object, err %d\n",
3767 rc = mds_update_inode_ids(obd, de_old->d_inode,
3768 handle, NULL, &ids[1]);
3772 rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3773 handle, req, rc, 0);
3775 switch (cleanup_phase) {
3777 up(&mds->mds_pending_dir->d_inode->i_sem);
3780 UP_READ_I_ALLOC_SEM(new_inode);
3783 if (dlm_handles[5].cookie != 0)
3784 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3785 if (dlm_handles[6].cookie != 0)
3786 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3789 ptlrpc_save_llog_lock(req, lcl);
3792 if (dlm_handles[3].cookie != 0)
3793 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3794 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3795 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3796 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3798 if (dlm_handles[3].cookie != 0)
3799 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3800 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3801 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3802 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3811 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3814 req->rq_status = rc;
3818 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3819 struct ptlrpc_request *, struct lustre_handle *);
3821 static mds_reinter reinters[REINT_MAX + 1] = {
3822 [REINT_SETATTR] mds_reint_setattr,
3823 [REINT_CREATE] mds_reint_create,
3824 [REINT_LINK] mds_reint_link,
3825 [REINT_UNLINK] mds_reint_unlink,
3826 [REINT_RENAME] mds_reint_rename,
3827 [REINT_OPEN] mds_open
3830 int mds_reint_rec(struct mds_update_record *rec, int offset,
3831 struct ptlrpc_request *req, struct lustre_handle *lockh)
3833 struct obd_device *obd = req->rq_export->exp_obd;
3834 struct lvfs_run_ctxt saved;
3837 /* checked by unpacker */
3838 LASSERT(rec->ur_opcode <= REINT_MAX &&
3839 reinters[rec->ur_opcode] != NULL);
3841 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3842 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3843 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);