lustre/mds/mds_reint.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/mds_reint.c
   5  *  Lustre Metadata Server (mds) reintegration routines
   6  *
   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>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  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.
  17  *
  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.
  22  *
  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.
  26  */
  27
  28 #ifndef EXPORT_SYMTAB
  29 # define EXPORT_SYMTAB
  30 #endif
  31 #define DEBUG_SUBSYSTEM S_MDS
  32
  33 #include <linux/fs.h>
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/obd.h>
  37 #include <linux/lustre_lib.h>
  38 #include <linux/lustre_idl.h>
  39 #include <linux/lustre_mds.h>
  40 #include <linux/lustre_dlm.h>
  41 #include <linux/lustre_fsfilt.h>
  42
  43 #include "mds_internal.h"
  44
  45 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  46                    int error)
  47 {
  48         obd_transno_commit_cb(obd, transno, error);
  49 }
  50
  51 struct mds_logcancel_data {
  52         struct lov_mds_md      *mlcd_lmm;
  53         int                     mlcd_size;
  54         int                     mlcd_cookielen;
  55         int                     mlcd_eadatalen;
  56         struct llog_cookie      mlcd_cookies[0];
  57 };
  58
  59
  60 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  61                                   void *cb_data, int error)
  62 {
  63         struct mds_logcancel_data *mlcd = cb_data;
  64         struct lov_stripe_md *lsm = NULL;
  65         struct llog_ctxt *ctxt;
  66         int rc;
  67
  68         obd_transno_commit_cb(obd, transno, error);
  69
  70         CDEBUG(D_HA, "cancelling %d cookies\n",
  71                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  72
  73         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  74                           mlcd->mlcd_eadatalen);
  75         if (rc < 0) {
  76                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  77                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  78                        rc);
  79         } else {
  80                 ///* XXX 0 normally, SENDNOW for debug */);
  81                 ctxt = llog_get_context(obd, mlcd->mlcd_cookies[0].lgc_subsys + 1);
  82                 rc = llog_cancel(ctxt, lsm,
  83                                          mlcd->mlcd_cookielen /
  84                                          sizeof(*mlcd->mlcd_cookies),
  85                                          mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  86                 if (rc)
  87                         CERROR("error cancelling %d log cookies: rc %d\n",
  88                                (int)(mlcd->mlcd_cookielen /
  89                                      sizeof(*mlcd->mlcd_cookies)), rc);
  90         }
  91
  92         OBD_FREE(mlcd, mlcd->mlcd_size);
  93 }
  94
  95 /* Assumes caller has already pushed us into the kernel context. */
  96 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  97                        struct ptlrpc_request *req, int rc, __u32 op_data)
  98 {
  99         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 100         struct mds_client_data *mcd = med->med_mcd;
 101         struct obd_device *obd = req->rq_export->exp_obd;
 102         int err;
 103         __u64 transno;
 104         loff_t off;
 105         int log_pri = D_HA;
 106         ENTRY;
 107
 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");
 111                 if (handle)
 112                         GOTO(commit, rc);
 113                 RETURN(rc);
 114         }
 115
 116         if (IS_ERR(handle))
 117                 RETURN(rc);
 118
 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));
 126                 }
 127         }
 128
 129         off = med->med_off;
 130
 131         transno = req->rq_reqmsg->transno;
 132         if (rc != 0) {
 133                 LASSERT(transno == 0);
 134         } else if (transno == 0) {
 135                 spin_lock(&mds->mds_transno_lock);
 136                 transno = ++mds->mds_last_transno;
 137                 spin_unlock(&mds->mds_transno_lock);
 138         } else {
 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);
 143         }
 144         req->rq_repmsg->transno = req->rq_transno = transno;
 145         mcd->mcd_last_transno = cpu_to_le64(transno);
 146         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 147         mcd->mcd_last_result = cpu_to_le32(rc);
 148         mcd->mcd_last_data = cpu_to_le32(op_data);
 149
 150         fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 151                               mds_commit_cb, NULL);
 152         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 153                                   &off, 0);
 154
 155         if (err) {
 156                 log_pri = D_ERROR;
 157                 if (rc == 0)
 158                         rc = err;
 159         }
 160
 161         DEBUG_REQ(log_pri, req,
 162                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 163                   transno, mcd->mcd_uuid, med->med_idx, err);
 164
 165         err = mds_lov_write_objids(obd);
 166         if (err) {
 167                 log_pri = D_ERROR;
 168                 if (rc == 0)
 169                         rc = err;
 170         }
 171         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 172
 173 commit:
 174         err = fsfilt_commit(obd, inode, handle, 0);
 175         if (err) {
 176                 CERROR("error committing transaction: %d\n", err);
 177                 if (!rc)
 178                         rc = err;
 179         }
 180
 181         RETURN(rc);
 182 }
 183
 184 /* this gives the same functionality as the code between
 185  * sys_chmod and inode_setattr
 186  * chown_common and inode_setattr
 187  * utimes and inode_setattr
 188  */
 189 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 190 {
 191         time_t now = LTIME_S(CURRENT_TIME);
 192         struct iattr *attr = &rec->ur_iattr;
 193         unsigned int ia_valid = attr->ia_valid;
 194         int error;
 195         ENTRY;
 196
 197         /* only fix up attrs if the client VFS didn't already */
 198         if (!(ia_valid & ATTR_RAW))
 199                 RETURN(0);
 200
 201         if (!(ia_valid & ATTR_CTIME_SET))
 202                 LTIME_S(attr->ia_ctime) = now;
 203         if (!(ia_valid & ATTR_ATIME_SET))
 204                 LTIME_S(attr->ia_atime) = now;
 205         if (!(ia_valid & ATTR_MTIME_SET))
 206                 LTIME_S(attr->ia_mtime) = now;
 207
 208         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 209                 RETURN(-EPERM);
 210
 211         /* times */
 212         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 213                 if (rec->ur_fsuid != inode->i_uid &&
 214                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 215                         RETURN(error);
 216         }
 217
 218         if (ia_valid & ATTR_SIZE) {
 219                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 220                         RETURN(error);
 221         }
 222
 223         if (ia_valid & ATTR_UID) {
 224                 /* chown */
 225                 error = -EPERM;
 226                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 227                         RETURN(-EPERM);
 228                 if (attr->ia_uid == (uid_t) -1)
 229                         attr->ia_uid = inode->i_uid;
 230                 if (attr->ia_gid == (gid_t) -1)
 231                         attr->ia_gid = inode->i_gid;
 232                 attr->ia_mode = inode->i_mode;
 233                 /*
 234                  * If the user or group of a non-directory has been
 235                  * changed by a non-root user, remove the setuid bit.
 236                  * 19981026 David C Niemi <niemi@tux.org>
 237                  *
 238                  * Changed this to apply to all users, including root,
 239                  * to avoid some races. This is the behavior we had in
 240                  * 2.0. The check for non-root was definitely wrong
 241                  * for 2.2 anyway, as it should have been using
 242                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 243                  */
 244                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 245                     !S_ISDIR(inode->i_mode)) {
 246                         attr->ia_mode &= ~S_ISUID;
 247                         attr->ia_valid |= ATTR_MODE;
 248                 }
 249                 /*
 250                  * Likewise, if the user or group of a non-directory
 251                  * has been changed by a non-root user, remove the
 252                  * setgid bit UNLESS there is no group execute bit
 253                  * (this would be a file marked for mandatory
 254                  * locking).  19981026 David C Niemi <niemi@tux.org>
 255                  *
 256                  * Removed the fsuid check (see the comment above) --
 257                  * 19990830 SD.
 258                  */
 259                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 260                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 261                         attr->ia_mode &= ~S_ISGID;
 262                         attr->ia_valid |= ATTR_MODE;
 263                 }
 264         } else if (ia_valid & ATTR_MODE) {
 265                 int mode = attr->ia_mode;
 266                 /* chmod */
 267                 if (attr->ia_mode == (mode_t) -1)
 268                         attr->ia_mode = inode->i_mode;
 269                 attr->ia_mode =
 270                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 271         }
 272         RETURN(0);
 273 }
 274
 275 void mds_steal_ack_locks(struct ptlrpc_request *req)
 276 {
 277         struct obd_export         *exp = req->rq_export;
 278         struct list_head          *tmp;
 279         struct ptlrpc_reply_state *oldrep;
 280         struct ptlrpc_service     *svc;
 281         unsigned long              flags;
 282         int                        i;
 283
 284         /* CAVEAT EMPTOR: spinlock order */
 285         spin_lock_irqsave (&exp->exp_lock, flags);
 286         list_for_each (tmp, &exp->exp_outstanding_replies) {
 287                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 288
 289                 if (oldrep->rs_xid != req->rq_xid)
 290                         continue;
 291
 292                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 293                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 294                                 "new %d old %d\n", req->rq_xid,
 295                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 296
 297                 svc = oldrep->rs_srv_ni->sni_service;
 298                 spin_lock (&svc->srv_lock);
 299
 300                 list_del_init (&oldrep->rs_exp_list);
 301
 302                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 303                       " o%d NID"LPX64"\n",
 304                       oldrep->rs_nlocks, oldrep,
 305                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 306                       exp->exp_connection->c_peer.peer_nid);
 307
 308                 for (i = 0; i < oldrep->rs_nlocks; i++)
 309                         ptlrpc_save_lock(req,
 310                                          &oldrep->rs_locks[i],
 311                                          oldrep->rs_modes[i]);
 312                 oldrep->rs_nlocks = 0;
 313
 314                 DEBUG_REQ(D_HA, req, "stole locks for");
 315                 ptlrpc_schedule_difficult_reply (oldrep);
 316
 317                 spin_unlock (&svc->srv_lock);
 318                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 319                 return;
 320         }
 321         spin_unlock_irqrestore (&exp->exp_lock, flags);
 322 }
 323
 324 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 325 {
 326         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 327                   mcd->mcd_last_transno, mcd->mcd_last_result);
 328         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 329         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 330
 331         mds_steal_ack_locks(req);
 332 }
 333
 334 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 335                                       int offset, struct ptlrpc_request *req)
 336 {
 337         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 338         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 339         struct dentry *de;
 340         struct mds_body *body;
 341
 342         mds_req_from_mcd(req, med->med_mcd);
 343
 344         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 345         if (IS_ERR(de)) {
 346                 LASSERT(PTR_ERR(de) == req->rq_status);
 347                 return;
 348         }
 349
 350         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 351         mds_pack_inode2fid(&body->fid1, de->d_inode);
 352         mds_pack_inode2body(body, de->d_inode);
 353
 354         /* Don't return OST-specific attributes if we didn't just set them */
 355         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 356                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 357         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 358                 body->valid |= OBD_MD_FLMTIME;
 359         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 360                 body->valid |= OBD_MD_FLATIME;
 361
 362         l_dput(de);
 363 }
 364
 365 /* In the raw-setattr case, we lock the child inode.
 366  * In the write-back case or if being called from open, the client holds a lock
 367  * already.
 368  *
 369  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 370 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 371                              struct ptlrpc_request *req,
 372                              struct lustre_handle *lh)
 373 {
 374         struct mds_obd *mds = mds_req2mds(req);
 375         struct obd_device *obd = req->rq_export->exp_obd;
 376         struct mds_body *body;
 377         struct dentry *de;
 378         struct inode *inode = NULL;
 379         struct lustre_handle lockh;
 380         void *handle = NULL;
 381         struct mds_logcancel_data *mlcd = NULL;
 382         int rc = 0, cleanup_phase = 0, err, locked = 0;
 383         ENTRY;
 384
 385         LASSERT(offset == 0);
 386
 387         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 388                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 389
 390         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 391
 392         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 393                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 394                 if (IS_ERR(de))
 395                         GOTO(cleanup, rc = PTR_ERR(de));
 396         } else {
 397                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 398                                            &lockh, NULL, 0);
 399                 if (IS_ERR(de))
 400                         GOTO(cleanup, rc = PTR_ERR(de));
 401                 locked = 1;
 402         }
 403
 404         cleanup_phase = 1;
 405         inode = de->d_inode;
 406         LASSERT(inode);
 407         if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
 408                 down(&inode->i_sem);
 409
 410         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 411
 412         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 413         if (IS_ERR(handle))
 414                 GOTO(cleanup, rc = PTR_ERR(handle));
 415
 416         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 417                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 418                        LTIME_S(rec->ur_iattr.ia_mtime),
 419                        LTIME_S(rec->ur_iattr.ia_ctime));
 420         rc = mds_fix_attr(inode, rec);
 421         if (rc)
 422                 GOTO(cleanup, rc);
 423
 424         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 425                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 426                                       (long)&rec->ur_iattr.ia_attr_flags);
 427         else                                            /* setattr */
 428                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 429
 430         if (rc == 0 && S_ISREG(inode->i_mode) && rec->ur_eadata != NULL) {
 431                 rc = fsfilt_set_md(obd, inode, handle,
 432                                    rec->ur_eadata, rec->ur_eadatalen);
 433         }
 434
 435         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 436         mds_pack_inode2fid(&body->fid1, inode);
 437         mds_pack_inode2body(body, inode);
 438
 439         /* Don't return OST-specific attributes if we didn't just set them */
 440         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 441                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 442         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 443                 body->valid |= OBD_MD_FLMTIME;
 444         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 445                 body->valid |= OBD_MD_FLATIME;
 446
 447         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 448                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 449                           rec->ur_eadatalen);
 450                 if (mlcd) {
 451                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 452                                 rec->ur_eadatalen;
 453                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 454                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 455                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 456                                 mlcd->mlcd_cookielen;
 457                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 458                                mlcd->mlcd_cookielen);
 459                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 460                                mlcd->mlcd_eadatalen);
 461                 } else {
 462                         CERROR("unable to allocate log cancel data\n");
 463                 }
 464         }
 465         EXIT;
 466  cleanup:
 467         if (mlcd != NULL)
 468                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 469                                       mds_cancel_cookies_cb, mlcd);
 470         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 471         switch (cleanup_phase) {
 472         case 1:
 473                 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
 474                         up(&inode->i_sem);
 475                 l_dput(de);
 476                 if (locked) {
 477                         if (rc) {
 478                                 ldlm_lock_decref(&lockh, LCK_PW);
 479                         } else {
 480                                 ptlrpc_save_lock (req, &lockh, LCK_PW);
 481                         }
 482                 }
 483         case 0:
 484                 break;
 485         default:
 486                 LBUG();
 487         }
 488         if (err && !rc)
 489                 rc = err;
 490
 491         req->rq_status = rc;
 492         return 0;
 493 }
 494
 495 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 496                                      struct ptlrpc_request *req)
 497 {
 498         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 499         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 500         struct dentry *parent, *child;
 501         struct mds_body *body;
 502
 503         mds_req_from_mcd(req, med->med_mcd);
 504
 505         if (req->rq_status)
 506                 return;
 507
 508         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 509         LASSERT(!IS_ERR(parent));
 510         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 511         LASSERT(!IS_ERR(child));
 512         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 513         mds_pack_inode2fid(&body->fid1, child->d_inode);
 514         mds_pack_inode2body(body, child->d_inode);
 515         l_dput(parent);
 516         l_dput(child);
 517 }
 518
 519 static int mds_reint_create(struct mds_update_record *rec, int offset,
 520                             struct ptlrpc_request *req,
 521                             struct lustre_handle *lh)
 522 {
 523         struct dentry *dparent = NULL;
 524         struct mds_obd *mds = mds_req2mds(req);
 525         struct obd_device *obd = req->rq_export->exp_obd;
 526         struct dentry *dchild = NULL;
 527         struct inode *dir = NULL;
 528         void *handle = NULL;
 529         struct lustre_handle lockh;
 530         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 531         int created = 0;
 532         struct dentry_params dp;
 533         ENTRY;
 534
 535         LASSERT(offset == 0);
 536         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 537
 538         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 539                   rec->ur_fid1->id, rec->ur_fid1->generation,
 540                   rec->ur_name, rec->ur_mode);
 541
 542         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 543
 544         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 545                 GOTO(cleanup, rc = -ESTALE);
 546
 547         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
 548                                         rec->ur_name, rec->ur_namelen - 1);
 549         if (IS_ERR(dparent)) {
 550                 rc = PTR_ERR(dparent);
 551                 CERROR("parent lookup error %d\n", rc);
 552                 GOTO(cleanup, rc);
 553         }
 554         cleanup_phase = 1; /* locked parent dentry */
 555         dir = dparent->d_inode;
 556         LASSERT(dir);
 557
 558         ldlm_lock_dump_handle(D_OTHER, &lockh);
 559
 560         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 561         if (IS_ERR(dchild)) {
 562                 rc = PTR_ERR(dchild);
 563                 CERROR("child lookup error %d\n", rc);
 564                 GOTO(cleanup, rc);
 565         }
 566
 567         cleanup_phase = 2; /* child dentry */
 568
 569         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 570
 571         if (dir->i_mode & S_ISGID) {
 572                 if (S_ISDIR(rec->ur_mode))
 573                         rec->ur_mode |= S_ISGID;
 574         }
 575
 576         dchild->d_fsdata = (void *)&dp;
 577         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 578         dp.p_ptr = req;
 579
 580         switch (type) {
 581         case S_IFREG:{
 582                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 583                 if (IS_ERR(handle))
 584                         GOTO(cleanup, rc = PTR_ERR(handle));
 585                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 586                 EXIT;
 587                 break;
 588         }
 589         case S_IFDIR:{
 590                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 591                 if (IS_ERR(handle))
 592                         GOTO(cleanup, rc = PTR_ERR(handle));
 593                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 594                 EXIT;
 595                 break;
 596         }
 597         case S_IFLNK:{
 598                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 599                 if (IS_ERR(handle))
 600                         GOTO(cleanup, rc = PTR_ERR(handle));
 601                 if (rec->ur_tgt == NULL)        /* no target supplied */
 602                         rc = -EINVAL;           /* -EPROTO? */
 603                 else
 604                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 605                 EXIT;
 606                 break;
 607         }
 608         case S_IFCHR:
 609         case S_IFBLK:
 610         case S_IFIFO:
 611         case S_IFSOCK:{
 612                 int rdev = rec->ur_rdev;
 613                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 614                 if (IS_ERR(handle))
 615                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 616                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 617                 EXIT;
 618                 break;
 619         }
 620         default:
 621                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 622                 dchild->d_fsdata = NULL;
 623                 GOTO(cleanup, rc = -EINVAL);
 624         }
 625
 626         /* In case we stored the desired inum in here, we want to clean up. */
 627         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 628                 dchild->d_fsdata = NULL;
 629
 630         if (rc) {
 631                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 632                 GOTO(cleanup, rc);
 633         } else {
 634                 struct iattr iattr;
 635                 struct inode *inode = dchild->d_inode;
 636                 struct mds_body *body;
 637
 638                 created = 1;
 639                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 640                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 641                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 642                 iattr.ia_uid = rec->ur_fsuid;
 643                 if (dir->i_mode & S_ISGID)
 644                         iattr.ia_gid = dir->i_gid;
 645                 else
 646                         iattr.ia_gid = rec->ur_fsgid;
 647                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 648                         ATTR_MTIME | ATTR_CTIME;
 649
 650                 if (rec->ur_fid2->id) {
 651                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 652                         inode->i_generation = rec->ur_fid2->generation;
 653                         /* Dirtied and committed by the upcoming setattr. */
 654                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 655                                inode->i_ino, inode->i_generation);
 656                 } else {
 657                         struct lustre_handle child_ino_lockh;
 658
 659                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 660                                inode->i_ino, inode->i_generation);
 661
 662                         /* The inode we were allocated may have just been freed
 663                          * by an unlink operation.  We take this lock to
 664                          * synchronize against the matching reply-ack-lock taken
 665                          * in unlink, to avoid replay problems if this reply
 666                          * makes it out to the client but the unlink's does not.
 667                          * See bug 2029 for more detail.*/
 668                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 669                         if (rc != ELDLM_OK) {
 670                                 CERROR("error locking for unlink/create sync: "
 671                                        "%d\n", rc);
 672                         } else {
 673                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 674                         }
 675                 }
 676
 677                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 678                 if (rc)
 679                         CERROR("error on child setattr: rc = %d\n", rc);
 680
 681                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 682                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 683                 if (rc)
 684                         CERROR("error on parent setattr: rc = %d\n", rc);
 685
 686                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 687                 mds_pack_inode2fid(&body->fid1, inode);
 688                 mds_pack_inode2body(body, inode);
 689         }
 690         EXIT;
 691
 692 cleanup:
 693         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 694
 695         if (rc && created) {
 696                 /* Destroy the file we just created.  This should not need
 697                  * extra journal credits, as we have already modified all of
 698                  * the blocks needed in order to create the file in the first
 699                  * place.
 700                  */
 701                 switch (type) {
 702                 case S_IFDIR:
 703                         err = vfs_rmdir(dir, dchild);
 704                         if (err)
 705                                 CERROR("rmdir in error path: %d\n", err);
 706                         break;
 707                 default:
 708                         err = vfs_unlink(dir, dchild);
 709                         if (err)
 710                                 CERROR("unlink in error path: %d\n", err);
 711                         break;
 712                 }
 713         } else {
 714                 rc = err;
 715         }
 716         switch (cleanup_phase) {
 717         case 2: /* child dentry */
 718                 l_dput(dchild);
 719         case 1: /* locked parent dentry */
 720                 if (rc) {
 721                         ldlm_lock_decref(&lockh, LCK_PW);
 722                 } else {
 723                         ptlrpc_save_lock (req, &lockh, LCK_PW);
 724                 }
 725                 l_dput(dparent);
 726         case 0:
 727                 break;
 728         default:
 729                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 730                 LBUG();
 731         }
 732         req->rq_status = rc;
 733         return 0;
 734 }
 735
 736 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
 737 {
 738         int i;
 739
 740         for (i = 0; i < RES_NAME_SIZE; i++) {
 741                 /* return 1 here, because enqueue_ordered will skip resources
 742                  * of all zeroes if they're sorted to the end of the list. */
 743                 if (res1->name[i] == 0 && res2->name[i] != 0)
 744                         return 1;
 745                 if (res2->name[i] == 0 && res1->name[i] != 0)
 746                         return 0;
 747
 748                 if (res1->name[i] > res2->name[i])
 749                         return 1;
 750                 if (res1->name[i] < res2->name[i])
 751                         return 0;
 752         }
 753         return 0;
 754 }
 755
 756 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 757  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 758  * because they take the place of local semaphores.
 759  *
 760  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 761  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 762 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 763                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 764                           struct ldlm_res_id *p2_res_id,
 765                           struct lustre_handle *p2_lockh, int p2_lock_mode)
 766 {
 767         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 768         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 769         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 770         int rc, flags;
 771         ENTRY;
 772
 773         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 774
 775         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
 776                res_id[0]->name[0], res_id[1]->name[0]);
 777
 778         if (res_gt(p1_res_id, p2_res_id)) {
 779                 handles[1] = p1_lockh;
 780                 handles[0] = p2_lockh;
 781                 res_id[1] = p1_res_id;
 782                 res_id[0] = p2_res_id;
 783                 lock_modes[1] = p1_lock_mode;
 784                 lock_modes[0] = p2_lock_mode;
 785         }
 786
 787         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 788                res_id[0]->name[0], res_id[1]->name[0]);
 789
 790         flags = LDLM_FL_LOCAL_ONLY;
 791         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 792                               LDLM_PLAIN, NULL, lock_modes[0], &flags,
 793                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 794                               NULL, 0, NULL, handles[0]);
 795         if (rc != ELDLM_OK)
 796                 RETURN(-EIO);
 797         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 798
 799         if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0) {
 800                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 801                 ldlm_lock_addref(handles[1], lock_modes[1]);
 802         } else if (res_id[1]->name[0] != 0) {
 803                 flags = LDLM_FL_LOCAL_ONLY;
 804                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 805                                       *res_id[1], LDLM_PLAIN, NULL,
 806                                       lock_modes[1], &flags, mds_blocking_ast,
 807                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 808                                       NULL, handles[1]);
 809                 if (rc != ELDLM_OK) {
 810                         ldlm_lock_decref(handles[0], lock_modes[0]);
 811                         RETURN(-EIO);
 812                 }
 813                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 814         }
 815
 816         RETURN(0);
 817 }
 818
 819 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
 820                            struct lustre_handle *p1_lockh, int p1_lock_mode,
 821                            struct ldlm_res_id *p2_res_id,
 822                            struct lustre_handle *p2_lockh, int p2_lock_mode,
 823                            struct ldlm_res_id *c1_res_id,
 824                            struct lustre_handle *c1_lockh, int c1_lock_mode,
 825                            struct ldlm_res_id *c2_res_id,
 826                            struct lustre_handle *c2_lockh, int c2_lock_mode)
 827 {
 828         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
 829                                           c1_res_id, c2_res_id };
 830         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
 831                                                  c1_lockh, c2_lockh };
 832         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
 833                               c1_lock_mode, c2_lock_mode };
 834         int rc, i, j, sorted, flags;
 835         ENTRY;
 836
 837         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 838                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 839                res_id[3]->name[0]);
 840
 841         /* simple insertion sort - we have at most 4 elements */
 842         for (i = 1; i < 4; i++) {
 843                 j = i - 1;
 844                 dlm_handles[4] = dlm_handles[i];
 845                 res_id[4] = res_id[i];
 846                 lock_modes[4] = lock_modes[i];
 847
 848                 sorted = 0;
 849                 do {
 850                         if (res_gt(res_id[j], res_id[4])) {
 851                                 dlm_handles[j + 1] = dlm_handles[j];
 852                                 res_id[j + 1] = res_id[j];
 853                                 lock_modes[j + 1] = lock_modes[j];
 854                                 j--;
 855                         } else {
 856                                 sorted = 1;
 857                         }
 858                 } while (j >= 0 && !sorted);
 859
 860                 dlm_handles[j + 1] = dlm_handles[4];
 861                 res_id[j + 1] = res_id[4];
 862                 lock_modes[j + 1] = lock_modes[4];
 863         }
 864
 865         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 866                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 867                res_id[3]->name[0]);
 868
 869         /* XXX we could send ASTs on all these locks first before blocking? */
 870         for (i = 0; i < 4; i++) {
 871                 flags = 0;
 872                 if (res_id[i]->name[0] == 0)
 873                         break;
 874                 if (i != 0 &&
 875                     memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) == 0) {
 876                         memcpy(dlm_handles[i], dlm_handles[i-1],
 877                                sizeof(*(dlm_handles[i])));
 878                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
 879                 } else {
 880                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 881                                               *res_id[i], LDLM_PLAIN, NULL,
 882                                               lock_modes[i], &flags,
 883                                               mds_blocking_ast,
 884                                               ldlm_completion_ast, NULL, NULL,
 885                                               NULL, 0, NULL, dlm_handles[i]);
 886                         if (rc != ELDLM_OK)
 887                                 GOTO(out_err, rc = -EIO);
 888                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
 889                 }
 890         }
 891
 892         RETURN(0);
 893 out_err:
 894         while (i-- > 0)
 895                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
 896
 897         return rc;
 898 }
 899
 900 /* In the unlikely case that the child changed while we were waiting
 901  * on the lock, we need to drop the lock on the old child and either:
 902  * - if the child has a lower resource name, then we have to also
 903  *   drop the parent lock and regain the locks in the right order
 904  * - in the rename case, if the child has a lower resource name than one of
 905  *   the other parent/child resources (maxres) we also need to reget the locks
 906  * - if the child has a higher resource name (this is the common case)
 907  *   we can just get the lock on the new child (still in lock order)
 908  *
 909  * Returns 0 if the child did not change or if it changed but could be locked.
 910  * Returns 1 if the child changed and we need to re-lock (no locks held).
 911  * Returns -ve error with a valid dchild (no locks held). */
 912 static int mds_verify_child(struct obd_device *obd,
 913                             struct ldlm_res_id *parent_res_id,
 914                             struct lustre_handle *parent_lockh,
 915                             struct dentry *dparent, int parent_mode,
 916                             struct ldlm_res_id *child_res_id,
 917                             struct lustre_handle *child_lockh,
 918                             struct dentry **dchildp, int child_mode,
 919                             const char *name, int namelen,
 920                             struct ldlm_res_id *maxres)
 921 {
 922         struct dentry *vchild, *dchild = *dchildp;
 923         int rc = 0, cleanup_phase = 2; /* parent, child locks */
 924         ENTRY;
 925
 926         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
 927         if (IS_ERR(vchild))
 928                 GOTO(cleanup, rc = PTR_ERR(vchild));
 929
 930         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
 931                    (vchild->d_inode != NULL &&
 932                     child_res_id->name[0] == vchild->d_inode->i_ino &&
 933                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
 934                 if (dchild != NULL)
 935                         l_dput(dchild);
 936                 *dchildp = vchild;
 937
 938                 RETURN(0);
 939         }
 940
 941         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
 942                vchild->d_inode, dchild ? dchild->d_inode : 0,
 943                vchild->d_inode ? vchild->d_inode->i_ino : 0,
 944                child_res_id->name[0]);
 945         if (child_res_id->name[0] != 0)
 946                 ldlm_lock_decref(child_lockh, child_mode);
 947         if (dchild)
 948                 l_dput(dchild);
 949
 950         cleanup_phase = 1; /* parent lock only */
 951         *dchildp = dchild = vchild;
 952
 953         if (dchild->d_inode) {
 954                 int flags = 0;
 955                 child_res_id->name[0] = dchild->d_inode->i_ino;
 956                 child_res_id->name[1] = dchild->d_inode->i_generation;
 957
 958                 if (res_gt(parent_res_id, child_res_id) ||
 959                     res_gt(maxres, child_res_id)) {
 960                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
 961                                child_res_id->name[0], parent_res_id->name[0],
 962                                maxres->name[0]);
 963                         GOTO(cleanup, rc = 1);
 964                 }
 965
 966                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 967                                       *child_res_id, LDLM_PLAIN, NULL,
 968                                       child_mode, &flags, mds_blocking_ast,
 969                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 970                                       NULL, child_lockh);
 971                 if (rc != ELDLM_OK)
 972                         GOTO(cleanup, rc = -EIO);
 973         } else {
 974                 memset(child_res_id, 0, sizeof(*child_res_id));
 975         }
 976
 977         EXIT;
 978 cleanup:
 979         if (rc) {
 980                 switch(cleanup_phase) {
 981                 case 2:
 982                         if (child_res_id->name[0] != 0)
 983                                 ldlm_lock_decref(child_lockh, child_mode);
 984                 case 1:
 985                         ldlm_lock_decref(parent_lockh, parent_mode);
 986                 }
 987         }
 988         return rc;
 989 }
 990
 991 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
 992                                 struct ll_fid *fid,
 993                                 struct lustre_handle *parent_lockh,
 994                                 struct dentry **dparentp, int parent_mode,
 995                                 char *name, int namelen,
 996                                 struct lustre_handle *child_lockh,
 997                                 struct dentry **dchildp, int child_mode)
 998 {
 999         struct ldlm_res_id child_res_id = { .name = {0} };
1000         struct ldlm_res_id parent_res_id = { .name = {0} };
1001         struct inode *inode;
1002         int rc = 0, cleanup_phase = 0;
1003         ENTRY;
1004
1005         /* Step 1: Lookup parent */
1006         *dparentp = mds_fid2dentry(mds, fid, NULL);
1007         if (IS_ERR(*dparentp))
1008                 RETURN(rc = PTR_ERR(*dparentp));
1009         LASSERT((*dparentp)->d_inode);
1010
1011         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1012                (*dparentp)->d_inode->i_ino, name);
1013
1014         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1015         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1016
1017         cleanup_phase = 1; /* parent dentry */
1018
1019         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1020         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1021         if (IS_ERR(*dchildp)) {
1022                 rc = PTR_ERR(*dchildp);
1023                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1024                 GOTO(cleanup, rc);
1025         }
1026
1027         inode = (*dchildp)->d_inode;
1028         if (inode != NULL)
1029                 inode = igrab(inode);
1030         if (inode == NULL)
1031                 goto retry_locks;
1032
1033         child_res_id.name[0] = inode->i_ino;
1034         child_res_id.name[1] = inode->i_generation;
1035         iput(inode);
1036
1037 retry_locks:
1038         cleanup_phase = 2; /* child dentry */
1039
1040         /* Step 3: Lock parent and child in resource order.  If child doesn't
1041          *         exist, we still have to lock the parent and re-lookup. */
1042         rc = enqueue_ordered_locks(obd, &parent_res_id,parent_lockh,parent_mode,
1043                                    &child_res_id, child_lockh, child_mode);
1044         if (rc)
1045                 GOTO(cleanup, rc);
1046
1047         if (!(*dchildp)->d_inode)
1048                 cleanup_phase = 3; /* parent lock */
1049         else
1050                 cleanup_phase = 4; /* child lock */
1051
1052         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1053         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1054                               parent_mode, &child_res_id, child_lockh, dchildp,
1055                               child_mode, name, namelen, &parent_res_id);
1056         if (rc > 0)
1057                 goto retry_locks;
1058         if (rc < 0) {
1059                 cleanup_phase = 3;
1060                 GOTO(cleanup, rc);
1061         }
1062
1063 cleanup:
1064         if (rc) {
1065                 switch (cleanup_phase) {
1066                 case 4:
1067                         ldlm_lock_decref(child_lockh, child_mode);
1068                 case 3:
1069                         ldlm_lock_decref(parent_lockh, parent_mode);
1070                 case 2:
1071                         l_dput(*dchildp);
1072                 case 1:
1073                         l_dput(*dparentp);
1074                 default: ;
1075                 }
1076         }
1077         return rc;
1078 }
1079
1080 void mds_reconstruct_generic(struct ptlrpc_request *req)
1081 {
1082         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1083
1084         mds_req_from_mcd(req, med->med_mcd);
1085 }
1086
1087 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1088                             struct ptlrpc_request *req,
1089                             struct lustre_handle *lh)
1090 {
1091         struct dentry *dparent, *dchild;
1092         struct mds_obd *mds = mds_req2mds(req);
1093         struct obd_device *obd = req->rq_export->exp_obd;
1094         struct mds_body *body = NULL;
1095         struct inode *child_inode;
1096         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1097         void *handle = NULL;
1098         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1099         ENTRY;
1100
1101         LASSERT(offset == 0 || offset == 2);
1102
1103         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1104                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1105
1106         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1107
1108         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1109                 GOTO(cleanup, rc = -ENOENT);
1110
1111         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1112                                          &parent_lockh, &dparent, LCK_PW,
1113                                          rec->ur_name, rec->ur_namelen,
1114                                          &child_lockh, &dchild, LCK_EX);
1115         if (rc)
1116                 GOTO(cleanup, rc);
1117
1118         cleanup_phase = 1; /* dchild, dparent, locks */
1119
1120         child_inode = dchild->d_inode;
1121         if (child_inode == NULL) {
1122                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1123                        dparent->d_inode->i_ino, rec->ur_name);
1124                 GOTO(cleanup, rc = -ENOENT);
1125         }
1126
1127         cleanup_phase = 2; /* dchild has a lock */
1128
1129         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1130          * reuse (see bug 2029). */
1131         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1132         if (rc != ELDLM_OK)
1133                 GOTO(cleanup, rc);
1134
1135         cleanup_phase = 3; /* child inum lock */
1136
1137         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1138
1139         /* ldlm_reply in buf[0] if called via intent */
1140         if (offset)
1141                 offset = 1;
1142
1143         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1144         LASSERT(body != NULL);
1145
1146         /* If this is the last reference to this inode, get the OBD EA
1147          * data first so the client can destroy OST objects.
1148          * we only do the object removal if no open files remain.
1149          * Nobody can get at this name anymore because of the locks so
1150          * we make decisions here as to whether to remove the inode */
1151         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1152             mds_open_orphan_count(child_inode) == 0) {
1153                 mds_pack_inode2fid(&body->fid1, child_inode);
1154                 mds_pack_inode2body(body, child_inode);
1155                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1156                             child_inode, 1);
1157                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1158                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1159                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1160                 } else {
1161                         log_unlink = 1;
1162                 }
1163         }
1164
1165         /* We have to do these checks ourselves, in case we are making an
1166          * orphan.  The client tells us whether rmdir() or unlink() was called,
1167          * so we need to return appropriate errors (bug 72).
1168          *
1169          * We don't have to check permissions, because vfs_rename (called from
1170          * mds_open_unlink_rename) also calls may_delete. */
1171         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1172                 if (!S_ISDIR(child_inode->i_mode))
1173                         GOTO(cleanup, rc = -ENOTDIR);
1174         } else {
1175                 if (S_ISDIR(child_inode->i_mode))
1176                         GOTO(cleanup, rc = -EISDIR);
1177         }
1178
1179         if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1180             mds_open_orphan_count(child_inode) > 0) {
1181                 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1182                 cleanup_phase = 4; /* transaction */
1183                 GOTO(cleanup, rc);
1184         }
1185
1186         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1187         switch (child_inode->i_mode & S_IFMT) {
1188         case S_IFDIR:
1189                 /* Drop any lingering child directories before we start our
1190                  * transaction, to avoid doing multiple inode dirty/delete
1191                  * in our compound transaction (bug 1321). */
1192                 shrink_dcache_parent(dchild);
1193                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1194                                       NULL);
1195                 if (IS_ERR(handle))
1196                         GOTO(cleanup, rc = PTR_ERR(handle));
1197                 cleanup_phase = 4; /* transaction */
1198                 rc = vfs_rmdir(dparent->d_inode, dchild);
1199                 break;
1200         case S_IFREG: {
1201                 handle = fsfilt_start(obd, dparent->d_inode,
1202                                       FSFILT_OP_UNLINK_LOG, NULL);
1203                 if (IS_ERR(handle))
1204                         GOTO(cleanup, rc = PTR_ERR(handle));
1205
1206                 cleanup_phase = 4; /* transaction */
1207                 rc = vfs_unlink(dparent->d_inode, dchild);
1208
1209                 if (!rc && log_unlink)
1210                         if (mds_log_op_unlink(obd, child_inode,
1211                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1212                                 req->rq_repmsg->buflens[offset + 1],
1213                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1214                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1215                                 body->valid |= OBD_MD_FLCOOKIE;
1216                 break;
1217         }
1218         case S_IFLNK:
1219         case S_IFCHR:
1220         case S_IFBLK:
1221         case S_IFIFO:
1222         case S_IFSOCK:
1223                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1224                                       NULL);
1225                 if (IS_ERR(handle))
1226                         GOTO(cleanup, rc = PTR_ERR(handle));
1227                 cleanup_phase = 4; /* transaction */
1228                 rc = vfs_unlink(dparent->d_inode, dchild);
1229                 break;
1230         default:
1231                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1232                        rec->ur_name);
1233                 LBUG();
1234                 GOTO(cleanup, rc = -EINVAL);
1235         }
1236
1237  cleanup:
1238         if (rc == 0) {
1239                 struct iattr iattr;
1240                 int err;
1241
1242                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1243                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1244                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1245
1246                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1247                 if (err)
1248                         CERROR("error on parent setattr: rc = %d\n", err);
1249         }
1250
1251         switch(cleanup_phase) {
1252         case 4:
1253                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1254                                         rc, 0);
1255                 if (!rc)
1256                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1257                                            "unlinked", 0, NULL);
1258         case 3: /* child ino-reuse lock */
1259                 if (rc && body != NULL) {
1260                         // Don't unlink the OST objects if the MDS unlink failed
1261                         body->valid = 0;
1262                 }
1263                 if (rc)
1264                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1265                 else
1266                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1267         case 2: /* child lock */
1268                 ldlm_lock_decref(&child_lockh, LCK_EX);
1269         case 1: /* child and parent dentry, parent lock */
1270                 if (rc)
1271                         ldlm_lock_decref(&parent_lockh, LCK_PW);
1272                 else
1273                         ptlrpc_save_lock(req, &parent_lockh, LCK_PW);
1274                 l_dput(dchild);
1275                 l_dput(dparent);
1276         case 0:
1277                 break;
1278         default:
1279                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1280                 LBUG();
1281         }
1282         req->rq_status = rc;
1283         return 0;
1284 }
1285
1286 static int mds_reint_link(struct mds_update_record *rec, int offset,
1287                           struct ptlrpc_request *req,
1288                           struct lustre_handle *lh)
1289 {
1290         struct obd_device *obd = req->rq_export->exp_obd;
1291         struct dentry *de_src = NULL;
1292         struct dentry *de_tgt_dir = NULL;
1293         struct dentry *dchild = NULL;
1294         struct mds_obd *mds = mds_req2mds(req);
1295         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1296         struct ldlm_res_id src_res_id = { .name = {0} };
1297         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1298         int rc = 0, cleanup_phase = 0;
1299         ENTRY;
1300
1301         LASSERT(offset == 0);
1302
1303         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1304                   rec->ur_fid1->id, rec->ur_fid1->generation,
1305                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1306
1307         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1308
1309         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1310                 GOTO(cleanup, rc = -ENOENT);
1311
1312         /* Step 1: Lookup the source inode and target directory by FID */
1313         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1314         if (IS_ERR(de_src))
1315                 GOTO(cleanup, rc = PTR_ERR(de_src));
1316
1317         cleanup_phase = 1; /* source dentry */
1318
1319         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1320         if (IS_ERR(de_tgt_dir))
1321                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1322
1323         cleanup_phase = 2; /* target directory dentry */
1324
1325         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1326                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1327                de_src->d_inode->i_ino);
1328
1329         /* Step 2: Take the two locks */
1330         src_res_id.name[0] = de_src->d_inode->i_ino;
1331         src_res_id.name[1] = de_src->d_inode->i_generation;
1332         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1333         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1334
1335         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1336                                    &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX);
1337         if (rc)
1338                 GOTO(cleanup, rc);
1339
1340         cleanup_phase = 3; /* locks */
1341
1342         /* Step 3: Lookup the child */
1343         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1344         if (IS_ERR(dchild)) {
1345                 rc = PTR_ERR(dchild);
1346                 if (rc != -EPERM && rc != -EACCES)
1347                         CERROR("child lookup error %d\n", rc);
1348                 GOTO(cleanup, rc);
1349         }
1350
1351         cleanup_phase = 4; /* child dentry */
1352
1353         if (dchild->d_inode) {
1354                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1355                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1356                 rc = -EEXIST;
1357                 GOTO(cleanup, rc);
1358         }
1359
1360         /* Step 4: Do it. */
1361         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1362
1363         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1364         if (IS_ERR(handle)) {
1365                 rc = PTR_ERR(handle);
1366                 GOTO(cleanup, rc);
1367         }
1368
1369         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1370         if (rc && rc != -EPERM && rc != -EACCES)
1371                 CERROR("vfs_link error %d\n", rc);
1372 cleanup:
1373         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1374                                 handle, req, rc, 0);
1375         EXIT;
1376
1377         switch (cleanup_phase) {
1378         case 4: /* child dentry */
1379                 l_dput(dchild);
1380         case 3: /* locks */
1381                 if (rc) {
1382                         ldlm_lock_decref(&src_lockh, LCK_EX);
1383                         ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1384                 } else {
1385                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1386                         ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1387                 }
1388         case 2: /* target dentry */
1389                 l_dput(de_tgt_dir);
1390         case 1: /* source dentry */
1391                 l_dput(de_src);
1392         case 0:
1393                 break;
1394         default:
1395                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1396                 LBUG();
1397         }
1398         req->rq_status = rc;
1399         return 0;
1400 }
1401
1402 /*
1403  * add a hard link in the PENDING directory, only used by rename()
1404  */
1405 static int mds_add_link_orphan(struct mds_update_record *rec,
1406                                struct obd_device *obd,
1407                                struct dentry *dentry)
1408 {
1409         struct mds_obd *mds = &obd->u.mds;
1410         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1411         struct dentry *pending_child;
1412         char fidname[LL_FID_NAMELEN];
1413         int fidlen = 0, rc;
1414         ENTRY;
1415
1416         LASSERT(dentry->d_inode);
1417         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
1418
1419         down(&pending_dir->i_sem);
1420         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
1421                             dentry->d_inode->i_generation);
1422
1423         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
1424                mds_open_orphan_count(dentry->d_inode),
1425                rec->ur_name, fidname);
1426
1427         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1428         if (IS_ERR(pending_child))
1429                 GOTO(out_lock, rc = PTR_ERR(pending_child));
1430
1431         if (pending_child->d_inode != NULL) {
1432                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1433                 LASSERT(pending_child->d_inode == dentry->d_inode);
1434                 GOTO(out_dput, rc = 0);
1435         }
1436
1437         lock_kernel();
1438         rc = vfs_link(dentry, pending_dir, pending_child);
1439         unlock_kernel();
1440         if (rc)
1441                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
1442                        rec->ur_name, rc);
1443         else
1444                 mds_inode_set_orphan(dentry->d_inode);
1445 out_dput:
1446         l_dput(pending_child);
1447 out_lock:
1448         up(&pending_dir->i_sem);
1449         RETURN(rc);
1450 }
1451
1452 /* The idea here is that we need to get four locks in the end:
1453  * one on each parent directory, one on each child.  We need to take
1454  * these locks in some kind of order (to avoid deadlocks), and the order
1455  * I selected is "increasing resource number" order.  We need to look up
1456  * the children, however, before we know what the resource number(s) are.
1457  * Thus the following plan:
1458  *
1459  * 1,2. Look up the parents
1460  * 3,4. Look up the children
1461  * 5. Take locks on the parents and children, in order
1462  * 6. Verify that the children haven't changed since they were looked up
1463  *
1464  * If there was a race and the children changed since they were first looked
1465  * up, it is possible that mds_verify_child() will be able to just grab the
1466  * lock on the new child resource (if it has a higher resource than any other)
1467  * but we need to compare against not only its parent, but also against the
1468  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1469  *
1470  * We need the fancy igrab() on the child inodes because we aren't holding a
1471  * lock on the parent after the lookup is done, so dentry->d_inode may change
1472  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1473  */
1474 static int mds_get_parents_children_locked(struct obd_device *obd,
1475                                            struct mds_obd *mds,
1476                                            struct ll_fid *p1_fid,
1477                                            struct dentry **de_srcdirp,
1478                                            struct ll_fid *p2_fid,
1479                                            struct dentry **de_tgtdirp,
1480                                            int parent_mode,
1481                                            const char *old_name, int old_len,
1482                                            struct dentry **de_oldp,
1483                                            const char *new_name, int new_len,
1484                                            struct dentry **de_newp,
1485                                            struct lustre_handle *dlm_handles,
1486                                            int child_mode)
1487 {
1488         struct ldlm_res_id p1_res_id = { .name = {0} };
1489         struct ldlm_res_id p2_res_id = { .name = {0} };
1490         struct ldlm_res_id c1_res_id = { .name = {0} };
1491         struct ldlm_res_id c2_res_id = { .name = {0} };
1492         struct ldlm_res_id *maxres_src, *maxres_tgt;
1493         struct inode *inode;
1494         int rc = 0, cleanup_phase = 0;
1495         ENTRY;
1496
1497         /* Step 1: Lookup the source directory */
1498         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1499         if (IS_ERR(*de_srcdirp))
1500                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1501
1502         cleanup_phase = 1; /* source directory dentry */
1503
1504         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1505         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1506
1507         /* Step 2: Lookup the target directory */
1508         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1509                 *de_tgtdirp = dget(*de_srcdirp);
1510         } else {
1511                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1512                 if (IS_ERR(*de_tgtdirp))
1513                         GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
1514         }
1515
1516         cleanup_phase = 2; /* target directory dentry */
1517
1518         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1519         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1520
1521         /* Step 3: Lookup the source child entry */
1522         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1523         if (IS_ERR(*de_oldp)) {
1524                 rc = PTR_ERR(*de_oldp);
1525                 CERROR("old child lookup error (%*s): %d\n",
1526                        old_len - 1, old_name, rc);
1527                 GOTO(cleanup, rc);
1528         }
1529
1530         cleanup_phase = 3; /* original name dentry */
1531
1532         inode = (*de_oldp)->d_inode;
1533         if (inode != NULL)
1534                 inode = igrab(inode);
1535         if (inode == NULL)
1536                 GOTO(cleanup, rc = -ENOENT);
1537
1538         c1_res_id.name[0] = inode->i_ino;
1539         c1_res_id.name[1] = inode->i_generation;
1540         iput(inode);
1541
1542         /* Step 4: Lookup the target child entry */
1543         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1544         if (IS_ERR(*de_newp)) {
1545                 rc = PTR_ERR(*de_newp);
1546                 CERROR("new child lookup error (%*s): %d\n",
1547                        old_len - 1, old_name, rc);
1548                 GOTO(cleanup, rc);
1549         }
1550
1551         cleanup_phase = 4; /* target dentry */
1552
1553         inode = (*de_newp)->d_inode;
1554         if (inode != NULL)
1555                 inode = igrab(inode);
1556         if (inode == NULL)
1557                 goto retry_locks;
1558
1559         c2_res_id.name[0] = inode->i_ino;
1560         c2_res_id.name[1] = inode->i_generation;
1561         iput(inode);
1562
1563 retry_locks:
1564         /* Step 5: Take locks on the parents and child(ren) */
1565         maxres_src = &p1_res_id;
1566         maxres_tgt = &p2_res_id;
1567         cleanup_phase = 4; /* target dentry */
1568
1569         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id))
1570                 maxres_src = &c1_res_id;
1571         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id))
1572                 maxres_tgt = &c2_res_id;
1573
1574         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1575                                     &p2_res_id, &dlm_handles[1], parent_mode,
1576                                     &c1_res_id, &dlm_handles[2], child_mode,
1577                                     &c2_res_id, &dlm_handles[3], child_mode);
1578         if (rc)
1579                 GOTO(cleanup, rc);
1580
1581         cleanup_phase = 6; /* parent and child(ren) locks */
1582
1583         /* Step 6a: Re-lookup source child to verify it hasn't changed */
1584         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1585                               parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
1586                               child_mode, old_name, old_len, maxres_tgt);
1587         if (rc) {
1588                 if (c2_res_id.name[0] != 0)
1589                         ldlm_lock_decref(&dlm_handles[3], child_mode);
1590                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1591                 cleanup_phase = 4;
1592                 if (rc > 0)
1593                         goto retry_locks;
1594                 GOTO(cleanup, rc);
1595         }
1596
1597         if ((*de_oldp)->d_inode == NULL)
1598                 GOTO(cleanup, rc = -ENOENT);
1599
1600         /* Step 6b: Re-lookup target child to verify it hasn't changed */
1601         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
1602                               parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
1603                               child_mode, new_name, new_len, maxres_src);
1604         if (rc) {
1605                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1606                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1607                 cleanup_phase = 4;
1608                 if (rc > 0)
1609                         goto retry_locks;
1610                 GOTO(cleanup, rc);
1611         }
1612
1613         EXIT;
1614 cleanup:
1615         if (rc) {
1616                 switch (cleanup_phase) {
1617                 case 6: /* child lock(s) */
1618                         if (c2_res_id.name[0] != 0)
1619                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
1620                         if (c1_res_id.name[0] != 0)
1621                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1622                 case 5: /* parent locks */
1623                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
1624                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
1625                 case 4: /* target dentry */
1626                         l_dput(*de_newp);
1627                 case 3: /* source dentry */
1628                         l_dput(*de_oldp);
1629                 case 2: /* target directory dentry */
1630                         l_dput(*de_tgtdirp);
1631                 case 1: /* source directry dentry */
1632                         l_dput(*de_srcdirp);
1633                 }
1634         }
1635
1636         return rc;
1637 }
1638
1639 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1640                             struct ptlrpc_request *req,
1641                             struct lustre_handle *lockh)
1642 {
1643         struct obd_device *obd = req->rq_export->exp_obd;
1644         struct dentry *de_srcdir = NULL;
1645         struct dentry *de_tgtdir = NULL;
1646         struct dentry *de_old = NULL;
1647         struct dentry *de_new = NULL;
1648         struct mds_obd *mds = mds_req2mds(req);
1649         struct lustre_handle dlm_handles[4];
1650         struct mds_body *body = NULL;
1651         int rc = 0, lock_count = 3;
1652         int cleanup_phase = 0;
1653         void *handle = NULL;
1654         ENTRY;
1655
1656         LASSERT(offset == 0);
1657
1658         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
1659                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
1660                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
1661
1662         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1663
1664         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
1665                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
1666                                              rec->ur_name, rec->ur_namelen,
1667                                              &de_old, rec->ur_tgt,
1668                                              rec->ur_tgtlen, &de_new,
1669                                              dlm_handles, LCK_EX);
1670         if (rc)
1671                 GOTO(cleanup, rc);
1672
1673         cleanup_phase = 1; /* parent(s), children, locks */
1674
1675         if (de_new->d_inode)
1676                 lock_count = 4;
1677
1678         /* sanity check for src inode */
1679         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1680             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
1681                 GOTO(cleanup, rc = -EINVAL);
1682
1683         /* sanity check for dest inode */
1684         if (de_new->d_inode &&
1685             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1686              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
1687                 GOTO(cleanup, rc = -EINVAL);
1688
1689         if (de_old->d_inode == de_new->d_inode) {
1690                 GOTO(cleanup, rc = 0);
1691         }
1692
1693         /* if we are about to remove the target at first, pass the EA of
1694          * that inode to client to perform and cleanup on OST */
1695         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1696         LASSERT(body != NULL);
1697
1698         if (de_new->d_inode &&
1699             S_ISREG(de_new->d_inode->i_mode) &&
1700             de_new->d_inode->i_nlink == 1 &&
1701             mds_open_orphan_count(de_new->d_inode) == 0) {
1702                 mds_pack_inode2fid(&body->fid1, de_new->d_inode);
1703                 mds_pack_inode2body(body, de_new->d_inode);
1704                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
1705                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1706                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1707                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1708                 } else {
1709                         /* XXX need log unlink? */
1710                 }
1711         }
1712
1713         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
1714                        de_srcdir->d_inode->i_sb);
1715
1716         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
1717         if (IS_ERR(handle))
1718                 GOTO(cleanup, rc = PTR_ERR(handle));
1719
1720         /* FIXME need adjust the journal block count? */
1721         /* if the target should be moved to PENDING, we at first increase the
1722          * link and later vfs_rename() will decrease the link count again */
1723         if (de_new->d_inode &&
1724             S_ISREG(de_new->d_inode->i_mode) &&
1725             de_new->d_inode->i_nlink == 1 &&
1726             mds_open_orphan_count(de_new->d_inode) > 0) {
1727                 rc = mds_add_link_orphan(rec, obd, de_new);
1728                 if (rc)
1729                         GOTO(cleanup, rc);
1730         }
1731
1732         lock_kernel();
1733         de_old->d_fsdata = req;
1734         de_new->d_fsdata = req;
1735         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
1736         unlock_kernel();
1737
1738         GOTO(cleanup, rc);
1739 cleanup:
1740         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
1741                                 handle, req, rc, 0);
1742         switch (cleanup_phase) {
1743         case 1:
1744                 if (rc) {
1745                         if (lock_count == 4)
1746                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
1747                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
1748                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
1749                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
1750                 } else {
1751                         if (lock_count == 4)
1752                                 ptlrpc_save_lock(req,
1753                                               &(dlm_handles[3]), LCK_EX);
1754                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
1755                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
1756                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
1757                 }
1758                 l_dput(de_new);
1759                 l_dput(de_old);
1760                 l_dput(de_tgtdir);
1761                 l_dput(de_srcdir);
1762         case 0:
1763                 break;
1764         default:
1765                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1766                 LBUG();
1767         }
1768         req->rq_status = rc;
1769         return 0;
1770 }
1771
1772 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
1773                            struct ptlrpc_request *, struct lustre_handle *);
1774
1775 static mds_reinter reinters[REINT_MAX + 1] = {
1776         [REINT_SETATTR] mds_reint_setattr,
1777         [REINT_CREATE] mds_reint_create,
1778         [REINT_LINK] mds_reint_link,
1779         [REINT_UNLINK] mds_reint_unlink,
1780         [REINT_RENAME] mds_reint_rename,
1781         [REINT_OPEN] mds_open
1782 };
1783
1784 int mds_reint_rec(struct mds_update_record *rec, int offset,
1785                   struct ptlrpc_request *req, struct lustre_handle *lockh)
1786 {
1787         struct obd_device *obd = req->rq_export->exp_obd;
1788         struct obd_run_ctxt saved;
1789         int rc;
1790         ENTRY;
1791
1792         /* checked by unpacker */
1793         LASSERT(rec->ur_opcode <= REINT_MAX &&
1794                 reinters[rec->ur_opcode] != NULL);
1795
1796         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1797         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
1798         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1799
1800         RETURN(rc);
1801 }