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/jbd.h>
  35 #include <linux/ext3_fs.h>
  36 #include <linux/obd_support.h>
  37 #include <linux/obd_class.h>
  38 #include <linux/obd.h>
  39 #include <linux/lustre_lib.h>
  40 #include <linux/lustre_idl.h>
  41 #include <linux/lustre_mds.h>
  42 #include <linux/lustre_dlm.h>
  43 #include <linux/lustre_log.h>
  44 #include <linux/lustre_fsfilt.h>
  45 #include "mds_internal.h"
  46
  47 struct mds_logcancel_data {
  48         struct lov_mds_md      *mlcd_lmm;
  49         int                     mlcd_size;
  50         int                     mlcd_cookielen;
  51         int                     mlcd_eadatalen;
  52         struct llog_cookie      mlcd_cookies[0];
  53 };
  54
  55 static void mds_cancel_cookies_cb(struct obd_device *obd,
  56                                   __u64 transno, void *cb_data,
  57                                   int error)
  58 {
  59         struct mds_logcancel_data *mlcd = cb_data;
  60         struct lov_stripe_md *lsm = NULL;
  61         struct llog_ctxt *ctxt;
  62         int rc;
  63
  64         obd_transno_commit_cb(obd, transno, error);
  65
  66         CDEBUG(D_HA, "cancelling %d cookies\n",
  67                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  68
  69         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, mlcd->mlcd_lmm,
  70                           mlcd->mlcd_eadatalen);
  71         if (rc < 0) {
  72                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  73                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  74                        rc);
  75         } else {
  76                 ///* XXX 0 normally, SENDNOW for debug */);
  77                 ctxt = llog_get_context(&obd->obd_llogs,
  78                                         mlcd->mlcd_cookies[0].lgc_subsys + 1);
  79                 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
  80                                  sizeof(*mlcd->mlcd_cookies),
  81                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
  82                 if (rc)
  83                         CERROR("error cancelling %d log cookies: rc %d\n",
  84                                (int)(mlcd->mlcd_cookielen /
  85                                      sizeof(*mlcd->mlcd_cookies)), rc);
  86         }
  87
  88         OBD_FREE(mlcd, mlcd->mlcd_size);
  89 }
  90
  91 /* Assumes caller has already pushed us into the kernel context. */
  92 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  93                        struct ptlrpc_request *req, int rc, __u32 op_data)
  94 {
  95         struct mds_export_data *med = &req->rq_export->exp_mds_data;
  96         struct obd_device *obd = req->rq_export->exp_obd;
  97         struct mds_client_data *mcd = med->med_mcd;
  98         int err, log_pri = D_HA;
  99         __u64 transno;
 100         loff_t off;
 101         ENTRY;
 102
 103         /* if the export has already been failed, we have no last_rcvd slot */
 104         if (req->rq_export->exp_failed) {
 105                 CERROR("committing transaction for disconnected client\n");
 106                 if (handle)
 107                         GOTO(out_commit, rc);
 108                 RETURN(rc);
 109         }
 110
 111         if (IS_ERR(handle))
 112                 RETURN(rc);
 113
 114         if (handle == NULL) {
 115                 /* if we're starting our own xaction, use our own inode */
 116                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 117                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 118                 if (IS_ERR(handle)) {
 119                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 120                         RETURN(PTR_ERR(handle));
 121                 }
 122         }
 123
 124         off = med->med_off;
 125
 126         transno = req->rq_reqmsg->transno;
 127         if (rc != 0) {
 128                 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
 129         } else if (transno == 0) {
 130                 spin_lock(&mds->mds_transno_lock);
 131                 transno = ++mds->mds_last_transno;
 132                 spin_unlock(&mds->mds_transno_lock);
 133         } else {
 134                 spin_lock(&mds->mds_transno_lock);
 135                 if (transno > mds->mds_last_transno)
 136                         mds->mds_last_transno = transno;
 137                 spin_unlock(&mds->mds_transno_lock);
 138         }
 139         req->rq_repmsg->transno = req->rq_transno = transno;
 140         mcd->mcd_last_transno = cpu_to_le64(transno);
 141         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 142         mcd->mcd_last_result = cpu_to_le32(rc);
 143         mcd->mcd_last_data = cpu_to_le32(op_data);
 144
 145         fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
 146                               mds_commit_last_transno_cb, NULL);
 147
 148         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
 149                                   sizeof(*mcd), &off, 0);
 150
 151         if (err) {
 152                 log_pri = D_ERROR;
 153                 if (rc == 0)
 154                         rc = err;
 155         }
 156
 157         DEBUG_REQ(log_pri, req,
 158                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 159                   transno, mcd->mcd_uuid, med->med_idx, err);
 160
 161         err = mds_update_last_fid(obd, handle, 0);
 162         if (err) {
 163                 log_pri = D_ERROR;
 164                 if (rc == 0)
 165                         rc = err;
 166         }
 167
 168         err = mds_dt_write_objids(obd);
 169         if (err) {
 170                 log_pri = D_ERROR;
 171                 if (rc == 0)
 172                         rc = err;
 173         }
 174         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 175
 176         EXIT;
 177 out_commit:
 178         err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 179         if (err) {
 180                 CERROR("error committing transaction: %d\n", err);
 181                 if (!rc)
 182                         rc = err;
 183         }
 184
 185         return rc;
 186 }
 187
 188 /* this gives the same functionality as the code between
 189  * sys_chmod and inode_setattr
 190  * chown_common and inode_setattr
 191  * utimes and inode_setattr
 192  */
 193 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 194 {
 195         time_t now = LTIME_S(CURRENT_TIME);
 196         struct iattr *attr = &rec->ur_iattr;
 197         unsigned int ia_valid = attr->ia_valid;
 198         int error;
 199         ENTRY;
 200
 201         /* only fix up attrs if the client VFS didn't already */
 202         if (!(ia_valid & ATTR_RAW))
 203                 RETURN(0);
 204
 205         if (!(ia_valid & ATTR_CTIME_SET))
 206                 LTIME_S(attr->ia_ctime) = now;
 207         if (!(ia_valid & ATTR_ATIME_SET))
 208                 LTIME_S(attr->ia_atime) = now;
 209         if (!(ia_valid & ATTR_MTIME_SET))
 210                 LTIME_S(attr->ia_mtime) = now;
 211
 212         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 213                 RETURN(-EPERM);
 214
 215         /* times */
 216         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 217                 if (rec->ur_fsuid != inode->i_uid &&
 218                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 219                         RETURN(error);
 220         }
 221
 222         if (ia_valid & ATTR_SIZE) {
 223                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 224                         RETURN(error);
 225         }
 226
 227         if (ia_valid & ATTR_UID) {
 228                 /* chown */
 229                 error = -EPERM;
 230                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 231                         RETURN(-EPERM);
 232                 if (attr->ia_uid == (uid_t) -1)
 233                         attr->ia_uid = inode->i_uid;
 234                 if (attr->ia_gid == (gid_t) -1)
 235                         attr->ia_gid = inode->i_gid;
 236                 attr->ia_mode = inode->i_mode;
 237                 /*
 238                  * If the user or group of a non-directory has been
 239                  * changed by a non-root user, remove the setuid bit.
 240                  * 19981026 David C Niemi <niemi@tux.org>
 241                  *
 242                  * Changed this to apply to all users, including root,
 243                  * to avoid some races. This is the behavior we had in
 244                  * 2.0. The check for non-root was definitely wrong
 245                  * for 2.2 anyway, as it should have been using
 246                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 247                  */
 248                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 249                     !S_ISDIR(inode->i_mode)) {
 250                         attr->ia_mode &= ~S_ISUID;
 251                         attr->ia_valid |= ATTR_MODE;
 252                 }
 253                 /*
 254                  * Likewise, if the user or group of a non-directory
 255                  * has been changed by a non-root user, remove the
 256                  * setgid bit UNLESS there is no group execute bit
 257                  * (this would be a file marked for mandatory
 258                  * locking).  19981026 David C Niemi <niemi@tux.org>
 259                  *
 260                  * Removed the fsuid check (see the comment above) --
 261                  * 19990830 SD.
 262                  */
 263                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 264                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 265                         attr->ia_mode &= ~S_ISGID;
 266                         attr->ia_valid |= ATTR_MODE;
 267                 }
 268         } else if (ia_valid & ATTR_MODE) {
 269                 int mode = attr->ia_mode;
 270                 /* chmod */
 271                 if (attr->ia_mode == (mode_t) -1)
 272                         attr->ia_mode = inode->i_mode;
 273                 attr->ia_mode =
 274                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 275         }
 276         RETURN(0);
 277 }
 278
 279 void mds_steal_ack_locks(struct ptlrpc_request *req)
 280 {
 281         struct obd_export         *exp = req->rq_export;
 282         char                       str[PTL_NALFMT_SIZE];
 283         struct list_head          *tmp;
 284         struct ptlrpc_reply_state *oldrep;
 285         struct ptlrpc_service     *svc;
 286         struct llog_create_locks  *lcl;
 287         unsigned long              flags;
 288         int                        i;
 289
 290         /* CAVEAT EMPTOR: spinlock order */
 291         spin_lock_irqsave (&exp->exp_lock, flags);
 292         list_for_each (tmp, &exp->exp_outstanding_replies) {
 293                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 294
 295                 if (oldrep->rs_xid != req->rq_xid)
 296                         continue;
 297
 298                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 299                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 300                                 "new %d old %d\n", req->rq_xid,
 301                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 302
 303                 svc = oldrep->rs_srv_ni->sni_service;
 304                 spin_lock (&svc->srv_lock);
 305
 306                 list_del_init (&oldrep->rs_exp_list);
 307
 308                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 309                       " o%d NID %s\n", oldrep->rs_nlocks, oldrep,
 310                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 311                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 312
 313                 for (i = 0; i < oldrep->rs_nlocks; i++)
 314                         ptlrpc_save_lock(req,
 315                                          &oldrep->rs_locks[i],
 316                                          oldrep->rs_modes[i]);
 317                 oldrep->rs_nlocks = 0;
 318
 319                 lcl = oldrep->rs_llog_locks;
 320                 oldrep->rs_llog_locks = NULL;
 321                 if (lcl != NULL)
 322                         ptlrpc_save_llog_lock(req, lcl);
 323
 324                 DEBUG_REQ(D_HA, req, "stole locks for");
 325                 ptlrpc_schedule_difficult_reply (oldrep);
 326
 327                 spin_unlock (&svc->srv_lock);
 328                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 329                 return;
 330         }
 331         spin_unlock_irqrestore (&exp->exp_lock, flags);
 332 }
 333
 334 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 335 {
 336         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 337                   mcd->mcd_last_transno, mcd->mcd_last_result);
 338         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 339         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 340
 341         mds_steal_ack_locks(req);
 342 }
 343
 344 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 345                                       int offset, struct ptlrpc_request *req)
 346 {
 347         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 348         struct mds_body *body;
 349         struct dentry *de;
 350
 351         mds_req_from_mcd(req, med->med_mcd);
 352
 353         de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
 354         if (IS_ERR(de)) {
 355                 LASSERT(PTR_ERR(de) == req->rq_status);
 356                 return;
 357         }
 358
 359         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 360         mds_pack_inode2body(req2obd(req), body, de->d_inode, 1);
 361
 362         /* Don't return OST-specific attributes if we didn't just set them */
 363         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 364                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 365         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 366                 body->valid |= OBD_MD_FLMTIME;
 367         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 368                 body->valid |= OBD_MD_FLATIME;
 369
 370         l_dput(de);
 371 }
 372
 373 /* In the raw-setattr case, we lock the child inode.
 374  * In the write-back case or if being called from open, the client holds a lock
 375  * already.
 376  *
 377  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 378 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 379                              struct ptlrpc_request *req, struct lustre_handle *lh)
 380 {
 381         struct mds_obd *mds = mds_req2mds(req);
 382         struct obd_device *obd = req->rq_export->exp_obd;
 383         struct mds_body *body;
 384         struct dentry *de;
 385         struct inode *inode = NULL;
 386         struct lustre_handle lockh[2] = {{0}, {0}};
 387         int parent_mode;
 388         void *handle = NULL;
 389         struct mds_logcancel_data *mlcd = NULL;
 390         int rc = 0, cleanup_phase = 0, err;
 391         int locked = 0;
 392         ENTRY;
 393
 394         LASSERT(offset == 1);
 395
 396         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
 397                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
 398                   rec->ur_iattr.ia_valid);
 399
 400         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 401         MD_COUNTER_INCREMENT(obd, setattr);
 402
 403         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 404                 de = mds_id2dentry(obd, rec->ur_id1, NULL);
 405                 if (IS_ERR(de))
 406                         GOTO(cleanup, rc = PTR_ERR(de));
 407         } else {
 408                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 409                 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
 410                         lockpart |= MDS_INODELOCK_LOOKUP;
 411                 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
 412                                           lockh, &parent_mode, NULL, 0, lockpart);
 413                 if (IS_ERR(de))
 414                         GOTO(cleanup, rc = PTR_ERR(de));
 415                 locked = 1;
 416         }
 417
 418         cleanup_phase = 1;
 419
 420         inode = de->d_inode;
 421         LASSERT(inode);
 422         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 423             rec->ur_eadata != NULL)
 424                 down(&inode->i_sem);
 425
 426         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 427
 428         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 429         if (IS_ERR(handle))
 430                 GOTO(cleanup, rc = PTR_ERR(handle));
 431
 432         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 433                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 434                        LTIME_S(rec->ur_iattr.ia_mtime),
 435                        LTIME_S(rec->ur_iattr.ia_ctime));
 436         rc = mds_fix_attr(inode, rec);
 437         if (rc)
 438                 GOTO(cleanup, rc);
 439
 440         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 441                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 442                                       (long)&rec->ur_iattr.ia_attr_flags);
 443         else                                            /* setattr */
 444                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 445
 446         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 447             rec->ur_eadata != NULL) {
 448                 struct lov_stripe_md *lsm = NULL;
 449
 450                 rc = ll_permission(inode, MAY_WRITE, NULL);
 451                 if (rc < 0)
 452                         GOTO(cleanup, rc);
 453
 454                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_dt_exp,
 455                                    0, &lsm, rec->ur_eadata);
 456                 if (rc)
 457                         GOTO(cleanup, rc);
 458
 459                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 460
 461                 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 462                                    rec->ur_eadatalen);
 463                 if (rc)
 464                         GOTO(cleanup, rc);
 465         }
 466
 467         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 468         mds_pack_inode2body(obd, body, inode, 1);
 469
 470         /* Don't return OST-specific attributes if we didn't just set them */
 471         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 472                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 473         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 474                 body->valid |= OBD_MD_FLMTIME;
 475         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 476                 body->valid |= OBD_MD_FLATIME;
 477
 478         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_dt_obd)) {
 479                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 480                           rec->ur_eadatalen);
 481                 if (mlcd) {
 482                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 483                                 rec->ur_eadatalen;
 484                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 485                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 486                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 487                                 mlcd->mlcd_cookielen;
 488                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 489                                mlcd->mlcd_cookielen);
 490                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 491                                mlcd->mlcd_eadatalen);
 492                 } else {
 493                         CERROR("unable to allocate log cancel data\n");
 494                 }
 495         }
 496         EXIT;
 497  cleanup:
 498         if (mlcd != NULL)
 499                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 500                                       handle, mds_cancel_cookies_cb, mlcd);
 501         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 502         switch (cleanup_phase) {
 503         case 1:
 504                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 505                     rec->ur_eadata != NULL)
 506                         up(&inode->i_sem);
 507                 l_dput(de);
 508                 if (locked) {
 509 #ifdef S_PDIROPS
 510                         if (lockh[1].cookie != 0)
 511                                 ldlm_lock_decref(lockh + 1, parent_mode);
 512 #endif
 513                         if (rc) {
 514                                 ldlm_lock_decref(lockh, LCK_PW);
 515                         } else {
 516                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 517                         }
 518                 }
 519         case 0:
 520                 break;
 521         default:
 522                 LBUG();
 523         }
 524         if (err && !rc)
 525                 rc = err;
 526
 527         req->rq_status = rc;
 528         return 0;
 529 }
 530
 531 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 532                                      struct ptlrpc_request *req)
 533 {
 534         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 535         struct dentry *parent, *child;
 536         struct mds_body *body;
 537         ENTRY;
 538
 539         mds_req_from_mcd(req, med->med_mcd);
 540
 541         if (req->rq_status) {
 542                 EXIT;
 543                 return;
 544         }
 545
 546         parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
 547         LASSERT(!IS_ERR(parent));
 548         child = ll_lookup_one_len(rec->ur_name, parent,
 549                                   rec->ur_namelen - 1);
 550         LASSERT(!IS_ERR(child));
 551         if ((child->d_flags & DCACHE_CROSS_REF)) {
 552                 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
 553                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 554                 mds_pack_dentry2body(req2obd(req), body, child, 1);
 555         } else if (child->d_inode == NULL) {
 556                 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
 557                           OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
 558                 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
 559         } else {
 560                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 561                 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
 562         }
 563         l_dput(parent);
 564         l_dput(child);
 565         EXIT;
 566 }
 567
 568 static int mds_reint_create(struct mds_update_record *rec, int offset,
 569                             struct ptlrpc_request *req,
 570                             struct lustre_handle *lh)
 571 {
 572         struct dentry *dparent = NULL;
 573         struct mds_obd *mds = mds_req2mds(req);
 574         struct obd_device *obd = req->rq_export->exp_obd;
 575         struct dentry *dchild = NULL;
 576         struct inode *dir = NULL;
 577         void *handle = NULL;
 578         struct lustre_handle lockh[2] = {{0}, {0}};
 579         int parent_mode;
 580         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 581         int created = 0;
 582         struct dentry_params dp;
 583         struct mea *mea = NULL;
 584         int mea_size;
 585         ENTRY;
 586
 587         LASSERT(offset == 1);
 588
 589         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
 590                         LUSTRE_MDS_NAME));
 591
 592         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 593                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
 594                   rec->ur_name, rec->ur_mode);
 595
 596         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 597
 598         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 599                 GOTO(cleanup, rc = -ESTALE);
 600
 601         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
 602                                        lockh, &parent_mode, rec->ur_name,
 603                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 604         if (IS_ERR(dparent)) {
 605                 rc = PTR_ERR(dparent);
 606                 CERROR("parent lookup error %d\n", rc);
 607                 GOTO(cleanup, rc);
 608         }
 609         cleanup_phase = 1; /* locked parent dentry */
 610         dir = dparent->d_inode;
 611         LASSERT(dir);
 612
 613         ldlm_lock_dump_handle(D_OTHER, lockh);
 614
 615         /* try to retrieve MEA data for this dir */
 616         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
 617         if (rc)
 618                 GOTO(cleanup, rc);
 619
 620         if (mea != NULL) {
 621                 /*
 622                  * dir is already splitted, check is requested filename should
 623                  * live at this MDS or at another one.
 624                  */
 625                 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 626                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
 627                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 628                                " should be %lu(%d)\n",
 629                                mea->mea_master, dparent->d_inode->i_ino,
 630                                dparent->d_inode->i_generation, rec->ur_name,
 631                                (unsigned long)id_group(&mea->mea_ids[i]), i);
 632                         GOTO(cleanup, rc = -ERESTART);
 633                 }
 634         }
 635
 636         dchild = ll_lookup_one_len(rec->ur_name, dparent,
 637                                    rec->ur_namelen - 1);
 638         if (IS_ERR(dchild)) {
 639                 rc = PTR_ERR(dchild);
 640                 CERROR("Can't find "DLID4"/%s, error %d\n",
 641                        OLID4(rec->ur_id1), rec->ur_name, rc);
 642                 GOTO(cleanup, rc);
 643         }
 644
 645         cleanup_phase = 2; /* child dentry */
 646
 647         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 648
 649         if (type == S_IFREG || type == S_IFDIR) {
 650                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 651                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 652                        obd->obd_name, dparent->d_inode->i_ino,
 653                        dparent->d_inode->i_generation, rc, parent_mode);
 654                 if (rc > 0) {
 655                         /* dir got splitted */
 656                         GOTO(cleanup, rc = -ERESTART);
 657                 } else if (rc < 0) {
 658                         /* error happened during spitting. */
 659                         GOTO(cleanup, rc);
 660                 }
 661         }
 662
 663         if (dir->i_mode & S_ISGID) {
 664                 if (S_ISDIR(rec->ur_mode))
 665                         rec->ur_mode |= S_ISGID;
 666         }
 667
 668         /*
 669          * here inode number should be used only in the case of replaying. It is
 670          * needed to check if object already created in the case of creating
 671          * remote inode.
 672          */
 673         dchild->d_fsdata = (void *)&dp;
 674         dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
 675         dp.p_ptr = req;
 676
 677         switch (type) {
 678         case S_IFREG:{
 679                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 680                 if (IS_ERR(handle))
 681                         GOTO(cleanup, rc = PTR_ERR(handle));
 682                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 683                 EXIT;
 684                 break;
 685         }
 686         case S_IFDIR:{
 687                 int i, nstripes = 0;
 688
 689                 /*
 690                  * as Peter asked, mkdir() should distribute new directories
 691                  * over the whole cluster in order to distribute namespace
 692                  * processing load. first, we calculate which MDS to use to put
 693                  * new directory's inode in.
 694                  */
 695                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 696                                       rec->ur_flags);
 697                 if (i == mds->mds_num) {
 698                         /* inode will be created locally */
 699                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 700                         if (IS_ERR(handle))
 701                                 GOTO(cleanup, rc = PTR_ERR(handle));
 702
 703                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 704                         if (rc) {
 705                                 CERROR("Can't create dir \"%s\", rc = %d\n",
 706                                        dchild->d_name.name, rc);
 707                                 GOTO(cleanup, rc);
 708                         }
 709
 710                         down(&dchild->d_inode->i_sem);
 711                         if (dp.p_inum) {
 712                                 rc = mds_update_inode_sid(obd, dchild->d_inode,
 713                                                           handle, rec->ur_id2);
 714                                 if (rc) {
 715                                         CERROR("mds_update_inode_sid() failed, inode %lu, "
 716                                                "rc %d\n", dchild->d_inode->i_ino, rc);
 717                                 }
 718
 719                                 /*
 720                                  * make sure, that fid is up-to-date.
 721                                  */
 722                                 mds_set_last_fid(obd, id_fid(rec->ur_id2));
 723                         } else {
 724                                 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
 725                                                          handle, NULL);
 726                                 if (rc) {
 727                                         CERROR("mds_alloc_inode_sid() failed, inode %lu, "
 728                                                "rc %d\n", dchild->d_inode->i_ino, rc);
 729                                 }
 730                         }
 731                         up(&dchild->d_inode->i_sem);
 732
 733                         if (rc)
 734                                 GOTO(cleanup, rc);
 735
 736                         if (rec->ur_eadata)
 737                                 nstripes = *(u16 *)rec->ur_eadata;
 738
 739                         if (rc == 0 && nstripes) {
 740                                 /*
 741                                  * we pass LCK_EX to split routine to signal,
 742                                  * that we have exclusive access to the
 743                                  * directory. Simple because nobody knows it
 744                                  * already exists -bzzz
 745                                  */
 746                                 rc = mds_try_to_split_dir(obd, dchild,
 747                                                           NULL, nstripes,
 748                                                           LCK_EX);
 749                                 if (rc > 0) {
 750                                         /* dir got splitted */
 751                                         rc = 0;
 752                                 } else if (rc < 0) {
 753                                         /* an error occured during
 754                                          * splitting. */
 755                                         GOTO(cleanup, rc);
 756                                 }
 757                         }
 758                 } else if (!DENTRY_VALID(dchild)) {
 759                         /* inode will be created on another MDS */
 760                         struct obdo *oa = NULL;
 761                         struct mds_body *body;
 762
 763                         /* first, create that inode */
 764                         oa = obdo_alloc();
 765                         if (!oa)
 766                                 GOTO(cleanup, rc = -ENOMEM);
 767
 768                         oa->o_mds = i;
 769                         oa->o_easize = 0;
 770
 771                         if (rec->ur_eadata) {
 772                                 /* user asks for creating splitted dir */
 773                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 774                         }
 775
 776                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 777                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 778                                         OBD_MD_FLUID | OBD_MD_FLGID);
 779
 780                         oa->o_mode = dir->i_mode;
 781
 782                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 783                                obd->obd_name, i);
 784
 785                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 786                                 /*
 787                                  * here inode number and generation are
 788                                  * important, as this is replay request and we
 789                                  * need them to check if such an object is
 790                                  * already created.
 791                                  */
 792                                 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
 793                                        obd->obd_name, rec->ur_namelen - 1,
 794                                        rec->ur_name, (unsigned)id_ino(rec->ur_id2),
 795                                        (unsigned)id_gen(rec->ur_id2));
 796                                 oa->o_id = id_ino(rec->ur_id2);
 797                                 oa->o_fid = id_fid(rec->ur_id2);
 798                                 oa->o_generation = id_gen(rec->ur_id2);
 799                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 800
 801                                 /*
 802                                  * fid should be defined here. It should be
 803                                  * passedfrom client.
 804                                  */
 805                                 LASSERT(oa->o_fid != 0);
 806                         }
 807
 808                         /*
 809                          * before obd_create() is called, o_fid is not known if
 810                          * this is not recovery of cause.
 811                          */
 812                         rc = obd_create(mds->mds_md_exp, oa, NULL, NULL);
 813                         if (rc) {
 814                                 CERROR("can't create remote inode: %d\n", rc);
 815                                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 816                                           id_ino(rec->ur_id1), id_gen(rec->ur_id1),
 817                                           rec->ur_name, rec->ur_mode);
 818                                 obdo_free(oa);
 819                                 GOTO(cleanup, rc);
 820                         }
 821
 822                         LASSERT(oa->o_fid != 0);
 823
 824                         /* now, add new dir entry for it */
 825                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 826                         if (IS_ERR(handle)) {
 827                                 obdo_free(oa);
 828                                 GOTO(cleanup, rc = PTR_ERR(handle));
 829                         }
 830
 831                         /* creating local dentry for remote inode. */
 832                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 833                                                   rec->ur_namelen - 1, oa->o_id,
 834                                                   oa->o_generation, i, oa->o_fid);
 835
 836                         if (rc) {
 837                                 CERROR("Can't create local entry %*s for "
 838                                        "remote inode.\n", rec->ur_namelen - 1,
 839                                         rec->ur_name);
 840                                 GOTO(cleanup, rc);
 841                         }
 842
 843                         /* fill reply */
 844                         body = lustre_msg_buf(req->rq_repmsg,
 845                                               0, sizeof(*body));
 846                         body->valid |= OBD_MD_FLID | OBD_MD_MDS | OBD_MD_FID;
 847
 848                         obdo2id(&body->id1, oa);
 849                         obdo_free(oa);
 850                 } else {
 851                         /* requested name exists in the directory */
 852                         rc = -EEXIST;
 853                 }
 854                 EXIT;
 855                 break;
 856         }
 857         case S_IFLNK:{
 858                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 859                 if (IS_ERR(handle))
 860                         GOTO(cleanup, rc = PTR_ERR(handle));
 861                 if (rec->ur_tgt == NULL)        /* no target supplied */
 862                         rc = -EINVAL;           /* -EPROTO? */
 863                 else
 864                         rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
 865                 EXIT;
 866                 break;
 867         }
 868         case S_IFCHR:
 869         case S_IFBLK:
 870         case S_IFIFO:
 871         case S_IFSOCK:{
 872                 int rdev = rec->ur_rdev;
 873                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 874                 if (IS_ERR(handle))
 875                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 876                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 877                 EXIT;
 878                 break;
 879         }
 880         default:
 881                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 882                 dchild->d_fsdata = NULL;
 883                 GOTO(cleanup, rc = -EINVAL);
 884         }
 885
 886         /* In case we stored the desired inum in here, we want to clean up. */
 887         if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
 888                 dchild->d_fsdata = NULL;
 889
 890         if (rc) {
 891                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 892                 GOTO(cleanup, rc);
 893         } else if (dchild->d_inode) {
 894                 struct iattr iattr;
 895                 struct mds_body *body;
 896                 struct inode *inode = dchild->d_inode;
 897
 898                 created = 1;
 899                 iattr.ia_uid = rec->ur_fsuid;
 900                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 901                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 902                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 903
 904                 if (dir->i_mode & S_ISGID)
 905                         iattr.ia_gid = dir->i_gid;
 906                 else
 907                         iattr.ia_gid = rec->ur_fsgid;
 908
 909                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 910                         ATTR_MTIME | ATTR_CTIME;
 911
 912                 if (id_ino(rec->ur_id2)) {
 913                         LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
 914                         inode->i_generation = id_gen(rec->ur_id2);
 915
 916                         if (type != S_IFDIR) {
 917                                 /*
 918                                  * updating inode self id, as inode already
 919                                  * exists and we should make sure, its sid will
 920                                  * be the same as we reveived.
 921                                  */
 922                                 down(&inode->i_sem);
 923                                 rc = mds_update_inode_sid(obd, inode,
 924                                                           handle, rec->ur_id2);
 925                                 up(&inode->i_sem);
 926                                 if (rc) {
 927                                         CERROR("Can't update inode self id, "
 928                                                "rc = %d.\n", rc);
 929                                 }
 930
 931                                 /*
 932                                  * make sure, that fid is up-to-date.
 933                                  */
 934                                 mds_set_last_fid(obd, id_fid(rec->ur_id2));
 935                         }
 936
 937                         /* dirtied and committed by the upcoming setattr. */
 938                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 939                                inode->i_ino, inode->i_generation);
 940                 } else {
 941                         struct lustre_handle child_ino_lockh;
 942
 943                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 944                                inode->i_ino, inode->i_generation);
 945
 946                         if (type != S_IFDIR) {
 947                                 /*
 948                                  * allocate new id for @inode if it is not dir,
 949                                  * because for dir it was already done.
 950                                  */
 951                                 down(&inode->i_sem);
 952                                 rc = mds_alloc_inode_sid(obd, inode,
 953                                                          handle, NULL);
 954                                 up(&inode->i_sem);
 955                                 if (rc) {
 956                                         CERROR("mds_alloc_inode_sid() failed, "
 957                                                "inode %lu, rc %d\n", inode->i_ino,
 958                                                rc);
 959                                 }
 960                         }
 961
 962                         if (rc == 0) {
 963                                 /*
 964                                  * the inode we were allocated may have just
 965                                  * been freed by an unlink operation.  We take
 966                                  * this lock to synchronize against the matching
 967                                  * reply-ack-lock taken in unlink, to avoid
 968                                  * replay problems if this reply makes it out to
 969                                  * the client but the unlink's does not.  See
 970                                  * bug 2029 for more detail.
 971                                  */
 972                                 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 973                                 if (rc != ELDLM_OK) {
 974                                         CERROR("error locking for unlink/create sync: "
 975                                                "%d\n", rc);
 976                                 } else {
 977                                         ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 978                                 }
 979                         }
 980                 }
 981
 982                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 983                 if (rc)
 984                         CERROR("error on child setattr: rc = %d\n", rc);
 985
 986                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 987                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 988                 if (rc)
 989                         CERROR("error on parent setattr: rc = %d\n", rc);
 990                 else
 991                         MD_COUNTER_INCREMENT(obd, create);
 992
 993                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 994                 mds_pack_inode2body(obd, body, inode, 1);
 995         }
 996
 997         EXIT;
 998 cleanup:
 999         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
1000
1001         if (rc && created) {
1002                 /* Destroy the file we just created. This should not need extra
1003                  * journal credits, as we have already modified all of the
1004                  * blocks needed in order to create the file in the first
1005                  * place. */
1006                 switch (type) {
1007                 case S_IFDIR:
1008                         err = vfs_rmdir(dir, dchild);
1009                         if (err)
1010                                 CERROR("rmdir in error path: %d\n", err);
1011                         break;
1012                 default:
1013                         err = vfs_unlink(dir, dchild);
1014                         if (err)
1015                                 CERROR("unlink in error path: %d\n", err);
1016                         break;
1017                 }
1018         } else {
1019                 rc = err;
1020         }
1021         switch (cleanup_phase) {
1022         case 2: /* child dentry */
1023                 l_dput(dchild);
1024         case 1: /* locked parent dentry */
1025 #ifdef S_PDIROPS
1026                 if (lockh[1].cookie != 0)
1027                         ldlm_lock_decref(lockh + 1, parent_mode);
1028 #endif
1029                 if (rc) {
1030                         ldlm_lock_decref(lockh, LCK_PW);
1031                 } else {
1032                         ptlrpc_save_lock(req, lockh, LCK_PW);
1033                 }
1034                 l_dput(dparent);
1035         case 0:
1036                 break;
1037         default:
1038                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1039                 LBUG();
1040         }
1041         if (mea)
1042                 OBD_FREE(mea, mea_size);
1043         req->rq_status = rc;
1044         return 0;
1045 }
1046
1047 static inline int
1048 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1049        ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1050 {
1051         int i;
1052
1053         for (i = 0; i < RES_NAME_SIZE; i++) {
1054                 /*
1055                  * this is needed to make zeroed res_id entries to be put at the
1056                  * end of list in *ordered_locks() .
1057                  */
1058                 if (res1->name[i] == 0 && res2->name[i] != 0)
1059                         return 1;
1060                 if (res2->name[i] == 0 && res1->name[i] != 0)
1061                         return 0;
1062                 if (res1->name[i] > res2->name[i])
1063                         return 1;
1064                 if (res1->name[i] < res2->name[i])
1065                         return 0;
1066         }
1067
1068         if (!p1 || !p2)
1069                 return 0;
1070
1071         if (memcmp(p1, p2, sizeof(*p1)) < 0)
1072                 return 1;
1073
1074         return 0;
1075 }
1076
1077 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1078  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1079  * because they take the place of local semaphores.
1080  *
1081  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
1082  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
1083 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1084                           struct lustre_handle *p1_lockh, int p1_lock_mode,
1085                           ldlm_policy_data_t *p1_policy,
1086                           struct ldlm_res_id *p2_res_id,
1087                           struct lustre_handle *p2_lockh, int p2_lock_mode,
1088                           ldlm_policy_data_t *p2_policy)
1089 {
1090         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1091         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1092         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1093         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1094         int rc, flags;
1095         ENTRY;
1096
1097         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1098
1099         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1100                res_id[0]->name[0], res_id[1]->name[0]);
1101
1102         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1103                 handles[1] = p1_lockh;
1104                 handles[0] = p2_lockh;
1105                 res_id[1] = p1_res_id;
1106                 res_id[0] = p2_res_id;
1107                 lock_modes[1] = p1_lock_mode;
1108                 lock_modes[0] = p2_lock_mode;
1109                 policies[1] = p1_policy;
1110                 policies[0] = p2_policy;
1111         }
1112
1113         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1114                res_id[0]->name[0], res_id[1]->name[0]);
1115
1116         flags = LDLM_FL_LOCAL_ONLY;
1117         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1118                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1119                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1120                               NULL, 0, NULL, handles[0]);
1121         if (rc != ELDLM_OK)
1122                 RETURN(-EIO);
1123         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1124
1125         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1126             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1127                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1128                 ldlm_lock_addref(handles[1], lock_modes[1]);
1129         } else if (res_id[1]->name[0] != 0) {
1130                 flags = LDLM_FL_LOCAL_ONLY;
1131                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1132                                       *res_id[1], LDLM_IBITS, policies[1],
1133                                       lock_modes[1], &flags, mds_blocking_ast,
1134                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1135                                       NULL, handles[1]);
1136                 if (rc != ELDLM_OK) {
1137                         ldlm_lock_decref(handles[0], lock_modes[0]);
1138                         RETURN(-EIO);
1139                 }
1140                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1141         }
1142
1143         RETURN(0);
1144 }
1145
1146 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1147                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1148                            ldlm_policy_data_t *p1_policy,
1149                            struct ldlm_res_id *p2_res_id,
1150                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1151                            ldlm_policy_data_t *p2_policy,
1152                            struct ldlm_res_id *c1_res_id,
1153                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1154                            ldlm_policy_data_t *c1_policy,
1155                            struct ldlm_res_id *c2_res_id,
1156                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1157                            ldlm_policy_data_t *c2_policy)
1158 {
1159         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1160                                           c1_res_id, c2_res_id };
1161         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1162                                                  c1_lockh, c2_lockh };
1163         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1164                               c1_lock_mode, c2_lock_mode };
1165         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1166                                             c1_policy, c2_policy};
1167         int rc, i, j, sorted, flags;
1168         ENTRY;
1169
1170         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1171                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1172                res_id[3]->name[0]);
1173
1174         /*
1175          * simple insertion sort - we have at most 4 elements. Note, that zeroed
1176          * res_id should be at the end of list after sorting is finished.
1177          */
1178         for (i = 1; i < 4; i++) {
1179                 j = i - 1;
1180                 dlm_handles[4] = dlm_handles[i];
1181                 res_id[4] = res_id[i];
1182                 lock_modes[4] = lock_modes[i];
1183                 policies[4] = policies[i];
1184
1185                 sorted = 0;
1186                 do {
1187                         if (res_gt(res_id[j], res_id[4], policies[j],
1188                                    policies[4])) {
1189                                 dlm_handles[j + 1] = dlm_handles[j];
1190                                 res_id[j + 1] = res_id[j];
1191                                 lock_modes[j + 1] = lock_modes[j];
1192                                 policies[j + 1] = policies[j];
1193                                 j--;
1194                         } else {
1195                                 sorted = 1;
1196                         }
1197                 } while (j >= 0 && !sorted);
1198
1199                 dlm_handles[j + 1] = dlm_handles[4];
1200                 res_id[j + 1] = res_id[4];
1201                 lock_modes[j + 1] = lock_modes[4];
1202                 policies[j + 1] = policies[4];
1203         }
1204
1205         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1206                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1207                res_id[3]->name[0]);
1208
1209         /* XXX we could send ASTs on all these locks first before blocking? */
1210         for (i = 0; i < 4; i++) {
1211                 flags = 0;
1212
1213                 /*
1214                  * nevertheless zeroed res_ids should be at the end of list, and
1215                  * could use break here, I think, that it is more correctly for
1216                  * clear understanding of code to have continue here, as it
1217                  * clearly means, that zeroed res_id should be skipped and does
1218                  * not mean, that if we meet zeroed res_id we should stop
1219                  * locking loop.
1220                  */
1221                 if (res_id[i]->name[0] == 0)
1222                         continue;
1223
1224                 if (i != 0 &&
1225                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1226                     (policies[i]->l_inodebits.bits &
1227                      policies[i-1]->l_inodebits.bits) ) {
1228                         memcpy(dlm_handles[i], dlm_handles[i-1],
1229                                sizeof(*(dlm_handles[i])));
1230                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1231                 } else {
1232                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1233                                               *res_id[i], LDLM_IBITS,
1234                                               policies[i],
1235                                               lock_modes[i], &flags,
1236                                               mds_blocking_ast,
1237                                               ldlm_completion_ast, NULL, NULL,
1238                                               NULL, 0, NULL, dlm_handles[i]);
1239                         if (rc != ELDLM_OK)
1240                                 GOTO(out_err, rc = -EIO);
1241                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1242                 }
1243         }
1244
1245         RETURN(0);
1246 out_err:
1247         while (i-- > 0)
1248                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1249
1250         return rc;
1251 }
1252
1253 /* In the unlikely case that the child changed while we were waiting
1254  * on the lock, we need to drop the lock on the old child and either:
1255  * - if the child has a lower resource name, then we have to also
1256  *   drop the parent lock and regain the locks in the right order
1257  * - in the rename case, if the child has a lower resource name than one of
1258  *   the other parent/child resources (maxres) we also need to reget the locks
1259  * - if the child has a higher resource name (this is the common case)
1260  *   we can just get the lock on the new child (still in lock order)
1261  *
1262  * Returns 0 if the child did not change or if it changed but could be locked.
1263  * Returns 1 if the child changed and we need to re-lock (no locks held).
1264  * Returns -ve error with a valid dchild (no locks held). */
1265 static int mds_verify_child(struct obd_device *obd,
1266                             struct ldlm_res_id *parent_res_id,
1267                             struct lustre_handle *parent_lockh,
1268                             struct dentry *dparent, int parent_mode,
1269                             struct ldlm_res_id *child_res_id,
1270                             struct lustre_handle *child_lockh,
1271                             struct dentry **dchildp, int child_mode,
1272                             ldlm_policy_data_t *child_policy,
1273                             const char *name, int namelen,
1274                             struct ldlm_res_id *maxres)
1275 {
1276         struct dentry *vchild, *dchild = *dchildp;
1277         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1278         struct lustre_id sid;
1279         ENTRY;
1280
1281         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1282         if (IS_ERR(vchild))
1283                 GOTO(cleanup, rc = PTR_ERR(vchild));
1284
1285         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1286                 /*
1287                  * cross-ref case, comparing child res_id with values stored in
1288                  * dentry.
1289                  */
1290                 id_fid(&sid) = vchild->d_fid;
1291                 id_group(&sid) = vchild->d_mdsnum;
1292
1293                 if (child_res_id->name[0] != id_fid(&sid) ||
1294                     child_res_id->name[1] != id_group(&sid))
1295                 {
1296                         goto changed;
1297                 }
1298         } else {
1299                 if (vchild->d_inode != NULL) {
1300                         /*
1301                          * name found and child res_id is not zero, getting
1302                          * inode fid and compare it with child res_id.
1303                          */
1304                         down(&vchild->d_inode->i_sem);
1305                         rc = mds_read_inode_sid(obd, vchild->d_inode, &sid);
1306                         up(&vchild->d_inode->i_sem);
1307                         if (rc) {
1308                                 CERROR("Can't read inode self id, inode %lu,"
1309                                        " rc %d\n",  vchild->d_inode->i_ino, rc);
1310                                 GOTO(cleanup, rc);
1311                         }
1312
1313                         if (child_res_id->name[0] != id_fid(&sid) ||
1314                             child_res_id->name[1] != id_group(&sid))
1315                         {
1316                                 goto changed;
1317                         }
1318                 }
1319         }
1320
1321         if (dchild)
1322                 l_dput(dchild);
1323         *dchildp = vchild;
1324         RETURN(0);
1325
1326 changed:
1327         /* child inode is chnaged, analizing it. */
1328         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1329                vchild->d_inode, dchild ? dchild->d_inode : 0,
1330                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1331                child_res_id->name[0]);
1332
1333         if (child_res_id->name[0] != 0)
1334                 ldlm_lock_decref(child_lockh, child_mode);
1335         if (dchild)
1336                 l_dput(dchild);
1337
1338         cleanup_phase = 1; /* parent lock only */
1339         *dchildp = dchild = vchild;
1340
1341         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1342                 int flags = 0;
1343
1344                 child_res_id->name[0] = id_fid(&sid);
1345                 child_res_id->name[1] = id_group(&sid);
1346
1347                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1348                     res_gt(maxres, child_res_id, NULL, NULL)) {
1349                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1350                                child_res_id->name[0], parent_res_id->name[0],
1351                                maxres->name[0]);
1352                         GOTO(cleanup, rc = 1);
1353                 }
1354
1355                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1356                                       *child_res_id, LDLM_IBITS, child_policy,
1357                                       child_mode, &flags, mds_blocking_ast,
1358                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1359                                       NULL, child_lockh);
1360                 if (rc != ELDLM_OK)
1361                         GOTO(cleanup, rc = -EIO);
1362
1363         } else {
1364                 memset(child_res_id, 0, sizeof(*child_res_id));
1365         }
1366
1367         EXIT;
1368 cleanup:
1369         if (rc) {
1370                 switch(cleanup_phase) {
1371                 case 2:
1372                         if (child_res_id->name[0] != 0)
1373                                 ldlm_lock_decref(child_lockh, child_mode);
1374                 case 1:
1375                         ldlm_lock_decref(parent_lockh, parent_mode);
1376                 }
1377         }
1378         return rc;
1379 }
1380
1381 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1382                                 struct lustre_id *id,
1383                                 struct lustre_handle *parent_lockh,
1384                                 struct dentry **dparentp, int parent_mode,
1385                                 __u64 parent_lockpart, int *update_mode,
1386                                 char *name, int namelen,
1387                                 struct lustre_handle *child_lockh,
1388                                 struct dentry **dchildp, int child_mode,
1389                                 __u64 child_lockpart)
1390 {
1391         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1392         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1393         struct ldlm_res_id parent_res_id = { .name = {0} };
1394         struct ldlm_res_id child_res_id = { .name = {0} };
1395         int rc = 0, cleanup_phase = 0;
1396         struct lustre_id sid;
1397         struct inode *inode;
1398         ENTRY;
1399
1400         /* Step 1: Lookup parent */
1401         *dparentp = mds_id2dentry(obd, id, NULL);
1402         if (IS_ERR(*dparentp)) {
1403                 rc = PTR_ERR(*dparentp);
1404                 *dparentp = NULL;
1405                 RETURN(rc);
1406         }
1407
1408         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1409                (*dparentp)->d_inode->i_ino, name);
1410
1411         parent_res_id.name[0] = id_fid(id);
1412         parent_res_id.name[1] = id_group(id);
1413
1414 #ifdef S_PDIROPS
1415         parent_lockh[1].cookie = 0;
1416         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1417                 struct ldlm_res_id res_id = { .name = {0} };
1418                 ldlm_policy_data_t policy;
1419                 int flags = 0;
1420
1421                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1422                 if (*update_mode) {
1423                         res_id.name[0] = id_fid(id);
1424                         res_id.name[1] = id_group(id);
1425                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1426
1427                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1428                                               res_id, LDLM_IBITS, &policy,
1429                                               *update_mode, &flags,
1430                                               mds_blocking_ast,
1431                                               ldlm_completion_ast,
1432                                               NULL, NULL, NULL, 0, NULL,
1433                                               parent_lockh + 1);
1434                         if (rc != ELDLM_OK)
1435                                 RETURN(-ENOLCK);
1436                 }
1437
1438                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1439
1440                 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1441                        (unsigned long)id_fid(id), (unsigned long)id_group(id),
1442                        parent_res_id.name[2]);
1443         }
1444 #endif
1445
1446         cleanup_phase = 1; /* parent dentry */
1447
1448         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1449         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1450         if (IS_ERR(*dchildp)) {
1451                 rc = PTR_ERR(*dchildp);
1452                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1453                 GOTO(cleanup, rc);
1454         }
1455
1456         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1457                 /*
1458                  * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1459                  * lock. Drop possible UPDATE lock!
1460                  */
1461                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1462                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1463
1464                 child_res_id.name[0] = (*dchildp)->d_fid;
1465                 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1466                 goto retry_locks;
1467         }
1468
1469         inode = (*dchildp)->d_inode;
1470         if (inode != NULL)
1471                 inode = igrab(inode);
1472         if (inode == NULL)
1473                 goto retry_locks;
1474
1475         down(&inode->i_sem);
1476         rc = mds_read_inode_sid(obd, inode, &sid);
1477         up(&inode->i_sem);
1478         if (rc) {
1479                 CERROR("Can't read inode self id, inode %lu, "
1480                        "rc %d\n", inode->i_ino, rc);
1481                 iput(inode);
1482                 GOTO(cleanup, rc);
1483         }
1484
1485         child_res_id.name[0] = id_fid(&sid);
1486         child_res_id.name[1] = id_group(&sid);
1487         iput(inode);
1488
1489 retry_locks:
1490         cleanup_phase = 2; /* child dentry */
1491
1492         /* Step 3: Lock parent and child in resource order.  If child doesn't
1493          * exist, we still have to lock the parent and re-lookup. */
1494         rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1495                                    &parent_policy, &child_res_id, child_lockh,
1496                                    child_mode, &child_policy);
1497         if (rc)
1498                 GOTO(cleanup, rc);
1499
1500         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1501                 cleanup_phase = 4; /* child lock */
1502         else
1503                 cleanup_phase = 3; /* parent lock */
1504
1505         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1506         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1507                               parent_mode, &child_res_id, child_lockh,
1508                               dchildp, child_mode, &child_policy,
1509                               name, namelen, &parent_res_id);
1510         if (rc > 0)
1511                 goto retry_locks;
1512         if (rc < 0) {
1513                 cleanup_phase = 3;
1514                 GOTO(cleanup, rc);
1515         }
1516
1517         EXIT;
1518 cleanup:
1519         if (rc) {
1520                 switch (cleanup_phase) {
1521                 case 4:
1522                         ldlm_lock_decref(child_lockh, child_mode);
1523                 case 3:
1524                         ldlm_lock_decref(parent_lockh, parent_mode);
1525                 case 2:
1526                         l_dput(*dchildp);
1527                 case 1:
1528 #ifdef S_PDIROPS
1529                         if (parent_lockh[1].cookie)
1530                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1531 #endif
1532                         l_dput(*dparentp);
1533                 }
1534         }
1535         return rc;
1536 }
1537
1538 void mds_reconstruct_generic(struct ptlrpc_request *req)
1539 {
1540         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1541         mds_req_from_mcd(req, med->med_mcd);
1542 }
1543
1544 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1545  * we instead link the inode into the PENDING directory until it is
1546  * finally released.  We can't simply call mds_reint_rename() or some
1547  * part thereof, because we don't have the inode to check for link
1548  * count/open status until after it is locked.
1549  *
1550  * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1551  * before starting journal transaction.
1552  *
1553  * returns 1 on success
1554  * returns 0 if we lost a race and didn't make a new link
1555  * returns negative on error
1556  */
1557 static int mds_orphan_add_link(struct mds_update_record *rec,
1558                                struct obd_device *obd, struct dentry *dentry)
1559 {
1560         struct mds_obd *mds = &obd->u.mds;
1561         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1562         struct inode *inode = dentry->d_inode;
1563         struct dentry *pending_child;
1564         char idname[LL_ID_NAMELEN];
1565         int idlen = 0, rc, mode;
1566         ENTRY;
1567
1568         LASSERT(inode != NULL);
1569         LASSERT(!mds_inode_is_orphan(inode));
1570 #ifndef HAVE_I_ALLOC_SEM
1571         LASSERT(down_trylock(&inode->i_sem) != 0);
1572 #endif
1573         LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1574
1575         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1576
1577         CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1578                mds_orphan_open_count(inode), inode->i_nlink,
1579                S_ISDIR(inode->i_mode) ? "dir" :
1580                S_ISREG(inode->i_mode) ? "file" : "other",
1581                rec->ur_name, idname);
1582
1583         if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1584                 RETURN(0);
1585
1586         pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1587         if (IS_ERR(pending_child))
1588                 RETURN(PTR_ERR(pending_child));
1589
1590         if (pending_child->d_inode != NULL) {
1591                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1592                 LASSERT(pending_child->d_inode == inode);
1593                 GOTO(out_dput, rc = 0);
1594         }
1595
1596         /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1597          * for linking and return real mode back then -bzzz */
1598         mode = inode->i_mode;
1599         inode->i_mode = S_IFREG;
1600         rc = vfs_link(dentry, pending_dir, pending_child);
1601         if (rc)
1602                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1603                        rec->ur_name, rc);
1604         else
1605                 mds_inode_set_orphan(inode);
1606
1607         /* return mode and correct i_nlink if inode is directory */
1608         inode->i_mode = mode;
1609         LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1610                  S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1611                  inode->i_nlink);
1612         if (S_ISDIR(mode)) {
1613                 inode->i_nlink++;
1614                 pending_dir->i_nlink++;
1615                 mark_inode_dirty(inode);
1616                 mark_inode_dirty(pending_dir);
1617         }
1618
1619         EXIT;
1620 out_dput:
1621         l_dput(pending_child);
1622         return rc;
1623 }
1624
1625 int mds_create_local_dentry(struct mds_update_record *rec,
1626                             struct obd_device *obd)
1627 {
1628         struct mds_obd *mds = &obd->u.mds;
1629         struct inode *id_dir = mds->mds_id_dir->d_inode;
1630         int idlen = 0, rc, cleanup_phase = 0;
1631         struct dentry *new_child = NULL;
1632         char *idname = rec->ur_name;
1633         struct dentry *child = NULL;
1634         struct lustre_handle lockh[2] = {{0}, {0}};
1635         struct lustre_id sid;
1636         void *handle;
1637         ENTRY;
1638
1639         down(&id_dir->i_sem);
1640         idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1641                           id_gen(rec->ur_id1));
1642
1643         CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1644                idname, OLID4(rec->ur_id1));
1645
1646         new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1647                                       idlen);
1648         up(&id_dir->i_sem);
1649         if (IS_ERR(new_child)) {
1650                 CERROR("can't lookup %s: %d\n", idname,
1651                        (int) PTR_ERR(new_child));
1652                 GOTO(cleanup, rc = PTR_ERR(new_child));
1653         }
1654         cleanup_phase = 1;
1655
1656         down(&id_dir->i_sem);
1657         rc = mds_read_inode_sid(obd, id_dir, &sid);
1658         up(&id_dir->i_sem);
1659         if (rc) {
1660                 CERROR("Can't read inode self id, inode %lu, "
1661                        "rc %d\n", id_dir->i_ino, rc);
1662                 GOTO(cleanup, rc);
1663         }
1664
1665         if (new_child->d_inode != NULL) {
1666                 /* nice. we've already have local dentry! */
1667                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1668                        (unsigned)new_child->d_inode->i_ino,
1669                        (unsigned)new_child->d_inode->i_generation);
1670
1671                 id_ino(rec->ur_id1) = id_dir->i_ino;
1672                 id_gen(rec->ur_id1) = id_dir->i_generation;
1673                 rec->ur_namelen = idlen + 1;
1674
1675                 id_fid(rec->ur_id1) = id_fid(&sid);
1676                 id_group(rec->ur_id1) = id_group(&sid);
1677
1678                 GOTO(cleanup, rc = 0);
1679         }
1680
1681         /* new, local dentry will be added soon. we need no aliases here */
1682         d_drop(new_child);
1683
1684         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1685                 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1686         } else {
1687                 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1688                                              LCK_EX, lockh, NULL, NULL, 0,
1689                                              MDS_INODELOCK_UPDATE);
1690         }
1691
1692         if (IS_ERR(child)) {
1693                 rc = PTR_ERR(child);
1694                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1695                         CERROR("can't get victim: %d\n", rc);
1696                 GOTO(cleanup, rc);
1697         }
1698         cleanup_phase = 2;
1699
1700         handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1701         if (IS_ERR(handle))
1702                 GOTO(cleanup, rc = PTR_ERR(handle));
1703
1704         rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
1705                                   idlen, id_ino(rec->ur_id1),
1706                                   id_gen(rec->ur_id1), mds->mds_num,
1707                                   id_fid(rec->ur_id1));
1708         if (rc)
1709                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1710                        (unsigned long)child->d_inode->i_ino,
1711                        (unsigned long)child->d_inode->i_generation, rc);
1712         else {
1713                 if (S_ISDIR(child->d_inode->i_mode)) {
1714                         id_dir->i_nlink++;
1715                         mark_inode_dirty(id_dir);
1716                 }
1717                 mark_inode_dirty(child->d_inode);
1718         }
1719         fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
1720
1721         id_ino(rec->ur_id1) = id_dir->i_ino;
1722         id_gen(rec->ur_id1) = id_dir->i_generation;
1723         rec->ur_namelen = idlen + 1;
1724
1725         id_fid(rec->ur_id1) = id_fid(&sid);
1726         id_group(rec->ur_id1) = id_group(&sid);
1727
1728         EXIT;
1729 cleanup:
1730         switch(cleanup_phase) {
1731                 case 2:
1732                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1733                                 ldlm_lock_decref(lockh, LCK_EX);
1734                         dput(child);
1735                 case 1:
1736                         dput(new_child);
1737                 case 0:
1738                        break;
1739         }
1740         return rc;
1741 }
1742
1743 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1744                                  struct ptlrpc_request *slave)
1745 {
1746         void *cookie, *cookie2;
1747         struct mds_body *body2;
1748         struct mds_body *body;
1749         void *ea, *ea2;
1750         ENTRY;
1751
1752         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1753         LASSERT(body != NULL);
1754
1755         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1756         LASSERT(body2 != NULL);
1757
1758         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
1759                 RETURN(0);
1760
1761         memcpy(body2, body, sizeof(*body));
1762         body2->valid &= ~OBD_MD_FLCOOKIE;
1763
1764         if (!(body->valid & OBD_MD_FLEASIZE) &&
1765             !(body->valid & OBD_MD_FLDIREA))
1766                 RETURN(0);
1767
1768         if (body->eadatasize == 0) {
1769                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1770                 RETURN(0);
1771         }
1772
1773         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1774
1775         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1776         LASSERT(ea != NULL);
1777
1778         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1779         LASSERT(ea2 != NULL);
1780
1781         memcpy(ea2, ea, body->eadatasize);
1782
1783         if (body->valid & OBD_MD_FLCOOKIE) {
1784                 LASSERT(master->rq_repmsg->buflens[2] >=
1785                                 slave->rq_repmsg->buflens[2]);
1786                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1787                                 slave->rq_repmsg->buflens[2]);
1788                 LASSERT(cookie != NULL);
1789
1790                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1791                                 master->rq_repmsg->buflens[2]);
1792                 LASSERT(cookie2 != NULL);
1793                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1794                 body2->valid |= OBD_MD_FLCOOKIE;
1795         }
1796         RETURN(0);
1797 }
1798
1799 static int mds_reint_unlink_remote(struct mds_update_record *rec,
1800                                    int offset, struct ptlrpc_request *req,
1801                                    struct lustre_handle *parent_lockh,
1802                                    int update_mode, struct dentry *dparent,
1803                                    struct lustre_handle *child_lockh,
1804                                    struct dentry *dchild)
1805 {
1806         struct obd_device *obd = req->rq_export->exp_obd;
1807         struct mds_obd *mds = mds_req2mds(req);
1808         struct ptlrpc_request *request = NULL;
1809         int rc = 0, cleanup_phase = 0;
1810         struct mdc_op_data *op_data;
1811         void *handle;
1812         ENTRY;
1813
1814         LASSERT(offset == 1 || offset == 3);
1815
1816         /* time to drop i_nlink on remote MDS */
1817         OBD_ALLOC(op_data, sizeof(*op_data));
1818         if (op_data == NULL)
1819                 RETURN(-ENOMEM);
1820
1821         memset(op_data, 0, sizeof(*op_data));
1822         mds_pack_dentry2id(obd, &op_data->id1, dchild, 1);
1823         op_data->create_mode = rec->ur_mode;
1824
1825         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
1826                   rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
1827
1828         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1829                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
1830                           rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
1831         }
1832
1833         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1834                 op_data->create_mode |= MDS_MODE_REPLAY;
1835
1836         rc = md_unlink(mds->mds_md_exp, op_data, &request);
1837         OBD_FREE(op_data, sizeof(*op_data));
1838         cleanup_phase = 2;
1839
1840         if (request) {
1841                 if (rc == 0)
1842                         mds_copy_unlink_reply(req, request);
1843                 ptlrpc_req_finished(request);
1844         }
1845
1846         if (rc == 0) {
1847                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1848                                       NULL);
1849                 if (IS_ERR(handle))
1850                         GOTO(cleanup, rc = PTR_ERR(handle));
1851                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1852                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1853                                         rc, 0);
1854         }
1855         EXIT;
1856 cleanup:
1857         req->rq_status = rc;
1858
1859 #ifdef S_PDIROPS
1860         if (parent_lockh[1].cookie != 0)
1861                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1862 #endif
1863         ldlm_lock_decref(child_lockh, LCK_EX);
1864         if (rc)
1865                 ldlm_lock_decref(parent_lockh, LCK_PW);
1866         else
1867                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1868         l_dput(dchild);
1869         l_dput(dparent);
1870
1871         return 0;
1872 }
1873
1874 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1875                             struct ptlrpc_request *req,
1876                             struct lustre_handle *lh)
1877 {
1878         struct dentry *dparent = NULL, *dchild;
1879         struct mds_obd *mds = mds_req2mds(req);
1880         struct obd_device *obd = req->rq_export->exp_obd;
1881         struct mds_body *body = NULL;
1882         struct inode *child_inode = NULL;
1883         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1884         struct lustre_handle child_lockh = {0};
1885         struct lustre_handle child_reuse_lockh = {0};
1886         struct lustre_handle *slave_lockh = NULL;
1887         char idname[LL_ID_NAMELEN];
1888         struct llog_create_locks *lcl = NULL;
1889         void *handle = NULL;
1890         int rc = 0, cleanup_phase = 0;
1891         int unlink_by_id = 0;
1892         int update_mode;
1893         ENTRY;
1894
1895         LASSERT(offset == 1 || offset == 3);
1896
1897         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1898                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1899                   rec->ur_name);
1900
1901         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1902
1903         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1904                 DEBUG_REQ(D_HA, req, "unlink replay");
1905                 LASSERT(offset == 1); /* should not come from intent */
1906                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1907                        lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
1908                        req->rq_repmsg->buflens[2]);
1909         }
1910
1911         MD_COUNTER_INCREMENT(obd, unlink);
1912
1913         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1914                 GOTO(cleanup, rc = -ENOENT);
1915
1916         if (rec->ur_namelen == 1) {
1917                 /* this is request to drop i_nlink on local inode */
1918                 unlink_by_id = 1;
1919                 rec->ur_name = idname;
1920                 rc = mds_create_local_dentry(rec, obd);
1921                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1922                         DEBUG_REQ(D_HA, req,
1923                                   "drop nlink on inode "DLID4" (replay)",
1924                                   OLID4(rec->ur_id1));
1925                         req->rq_status = 0;
1926                         RETURN(0);
1927                 }
1928         }
1929
1930         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1931                 /* master mds for directory asks slave removing inode is already
1932                  * locked */
1933                 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1934                                                LCK_PW, parent_lockh,
1935                                                &update_mode, rec->ur_name,
1936                                                rec->ur_namelen,
1937                                                MDS_INODELOCK_UPDATE);
1938                 if (IS_ERR(dparent))
1939                         GOTO(cleanup, rc = PTR_ERR(dparent));
1940                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1941                                            rec->ur_namelen - 1);
1942                 if (IS_ERR(dchild))
1943                         GOTO(cleanup, rc = PTR_ERR(dchild));
1944                 child_lockh.cookie = 0;
1945                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1946                 LASSERT(dchild->d_inode != NULL);
1947                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1948         } else {
1949                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1950                                                  parent_lockh, &dparent,
1951                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1952                                                  &update_mode, rec->ur_name,
1953                                                  rec->ur_namelen, &child_lockh,
1954                                                  &dchild, LCK_EX,
1955                                                  MDS_INODELOCK_LOOKUP |
1956                                                  MDS_INODELOCK_UPDATE);
1957         }
1958         if (rc)
1959                 GOTO(cleanup, rc);
1960
1961         if (dchild->d_flags & DCACHE_CROSS_REF) {
1962                 /* we should have parent lock only here */
1963                 LASSERT(unlink_by_id == 0);
1964                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1965                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1966                                         update_mode, dparent, &child_lockh, dchild);
1967                 RETURN(0);
1968         }
1969
1970         cleanup_phase = 1; /* dchild, dparent, locks */
1971
1972         dget(dchild);
1973         child_inode = dchild->d_inode;
1974         if (child_inode == NULL) {
1975                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1976                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1977                 GOTO(cleanup, rc = -ENOENT);
1978         }
1979
1980         cleanup_phase = 2; /* dchild has a lock */
1981
1982         /* We have to do these checks ourselves, in case we are making an
1983          * orphan.  The client tells us whether rmdir() or unlink() was called,
1984          * so we need to return appropriate errors (bug 72).
1985          *
1986          * We don't have to check permissions, because vfs_rename (called from
1987          * mds_open_unlink_rename) also calls may_delete. */
1988         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1989                 if (!S_ISDIR(child_inode->i_mode))
1990                         GOTO(cleanup, rc = -ENOTDIR);
1991         } else {
1992                 if (S_ISDIR(child_inode->i_mode))
1993                         GOTO(cleanup, rc = -EISDIR);
1994         }
1995
1996         /* handle splitted dir */
1997         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1998         if (rc)
1999                 GOTO(cleanup, rc);
2000
2001         /* Step 4: Get a lock on the ino to sync with creation WRT inode
2002          * reuse (see bug 2029). */
2003         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
2004         if (rc != ELDLM_OK)
2005                 GOTO(cleanup, rc);
2006         cleanup_phase = 3; /* child inum lock */
2007
2008         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2009
2010         /* ldlm_reply in buf[0] if called via intent */
2011         if (offset == 3)
2012                 offset = 1;
2013         else
2014                 offset = 0;
2015
2016         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2017         LASSERT(body != NULL);
2018
2019         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2020         DOWN_READ_I_ALLOC_SEM(child_inode);
2021         cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2022
2023         /* If this is potentially the last reference to this inode, get the
2024          * OBD EA data first so the client can destroy OST objects.  We
2025          * only do the object removal later if no open files/links remain. */
2026         if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2027             child_inode->i_nlink == 1) {
2028                 if (mds_orphan_open_count(child_inode) > 0) {
2029                         /* need to lock pending_dir before transaction */
2030                         down(&mds->mds_pending_dir->d_inode->i_sem);
2031                         cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2032                 } else if (S_ISREG(child_inode->i_mode)) {
2033                         mds_pack_inode2body(obd, body, child_inode, 0);
2034                         mds_pack_md(obd, req->rq_repmsg, offset + 1,
2035                                     body, child_inode, MDS_PACK_MD_LOCK);
2036                 }
2037         }
2038
2039         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2040         switch (child_inode->i_mode & S_IFMT) {
2041         case S_IFDIR:
2042                 /* Drop any lingering child directories before we start our
2043                  * transaction, to avoid doing multiple inode dirty/delete
2044                  * in our compound transaction (bug 1321). */
2045                 shrink_dcache_parent(dchild);
2046                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2047                                       NULL);
2048                 if (IS_ERR(handle))
2049                         GOTO(cleanup, rc = PTR_ERR(handle));
2050                 rc = vfs_rmdir(dparent->d_inode, dchild);
2051                 break;
2052         case S_IFREG: {
2053 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2054                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2055                                                         offset + 1, 0);
2056                 handle = fsfilt_start_log(obd, dparent->d_inode,
2057                                           FSFILT_OP_UNLINK, NULL,
2058                                           le32_to_cpu(lmm->lmm_stripe_count));
2059                 if (IS_ERR(handle))
2060                         GOTO(cleanup, rc = PTR_ERR(handle));
2061                 rc = vfs_unlink(dparent->d_inode, dchild);
2062                 break;
2063         }
2064         case S_IFLNK:
2065         case S_IFCHR:
2066         case S_IFBLK:
2067         case S_IFIFO:
2068         case S_IFSOCK:
2069                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2070                                       NULL);
2071                 if (IS_ERR(handle))
2072                         GOTO(cleanup, rc = PTR_ERR(handle));
2073                 rc = vfs_unlink(dparent->d_inode, dchild);
2074                 break;
2075         default:
2076                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2077                        rec->ur_name);
2078                 LBUG();
2079                 GOTO(cleanup, rc = -EINVAL);
2080         }
2081
2082         if (rc == 0 && child_inode->i_nlink == 0) {
2083                 if (mds_orphan_open_count(child_inode) > 0)
2084                         rc = mds_orphan_add_link(rec, obd, dchild);
2085
2086                 if (rc == 1)
2087                         GOTO(cleanup, rc = 0);
2088
2089                 if (!S_ISREG(child_inode->i_mode))
2090                         GOTO(cleanup, rc);
2091
2092                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2093                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2094                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2095                 } else if (mds_log_op_unlink(obd, child_inode,
2096                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2097                                         req->rq_repmsg->buflens[offset + 1],
2098                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2099                                         req->rq_repmsg->buflens[offset+2],
2100                                 &lcl) > 0){
2101                         body->valid |= OBD_MD_FLCOOKIE;
2102                 }
2103         }
2104
2105         GOTO(cleanup, rc);
2106
2107 cleanup:
2108         if (rc == 0) {
2109                 struct iattr iattr;
2110                 int err;
2111
2112                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2113                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2114                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2115
2116                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2117                 if (err)
2118                         CERROR("error on parent setattr: rc = %d\n", err);
2119         }
2120         rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2121                                 handle, req, rc, 0);
2122         if (!rc)
2123                 (void)obd_set_info(mds->mds_dt_exp, strlen("unlinked"),
2124                                    "unlinked", 0, NULL);
2125         switch(cleanup_phase) {
2126         case 5: /* pending_dir semaphore */
2127                 up(&mds->mds_pending_dir->d_inode->i_sem);
2128         case 4: /* child inode semaphore */
2129                 UP_READ_I_ALLOC_SEM(child_inode);
2130                  /* handle splitted dir */
2131                 if (rc == 0) {
2132                         /* master directory can be non-empty or something else ... */
2133                         mds_unlink_slave_objs(obd, dchild);
2134                 }
2135                 if (lcl != NULL)
2136                         ptlrpc_save_llog_lock(req, lcl);
2137         case 3: /* child ino-reuse lock */
2138                 if (rc && body != NULL) {
2139                         // Don't unlink the OST objects if the MDS unlink failed
2140                         body->valid = 0;
2141                 }
2142                 if (rc)
2143                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2144                 else
2145                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2146         case 2: /* child lock */
2147                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2148                 if (child_lockh.cookie)
2149                         ldlm_lock_decref(&child_lockh, LCK_EX);
2150         case 1: /* child and parent dentry, parent lock */
2151 #ifdef S_PDIROPS
2152                 if (parent_lockh[1].cookie != 0)
2153                         ldlm_lock_decref(parent_lockh + 1, update_mode);
2154 #endif
2155                 if (rc)
2156                         ldlm_lock_decref(parent_lockh, LCK_PW);
2157                 else
2158                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2159                 l_dput(dchild);
2160                 l_dput(dchild);
2161                 l_dput(dparent);
2162         case 0:
2163                 break;
2164         default:
2165                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2166                 LBUG();
2167         }
2168         req->rq_status = rc;
2169         return 0;
2170 }
2171
2172 /*
2173  * to service requests from remote MDS to increment i_nlink
2174  */
2175 static int mds_reint_link_acquire(struct mds_update_record *rec,
2176                                   int offset, struct ptlrpc_request *req,
2177                                   struct lustre_handle *lh)
2178 {
2179         struct obd_device *obd = req->rq_export->exp_obd;
2180         struct ldlm_res_id src_res_id = { .name = {0} };
2181         struct lustre_handle *handle = NULL, src_lockh = {0};
2182         struct mds_obd *mds = mds_req2mds(req);
2183         int rc = 0, cleanup_phase = 0;
2184         struct dentry *de_src = NULL;
2185         ldlm_policy_data_t policy;
2186         int flags = 0;
2187         ENTRY;
2188
2189         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2190                   obd->obd_name, OLID4(rec->ur_id1));
2191
2192         /* Step 1: Lookup the source inode and target directory by ID */
2193         de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2194         if (IS_ERR(de_src))
2195                 GOTO(cleanup, rc = PTR_ERR(de_src));
2196         cleanup_phase = 1; /* source dentry */
2197
2198         src_res_id.name[0] = id_fid(rec->ur_id1);
2199         src_res_id.name[1] = id_group(rec->ur_id1);
2200         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2201
2202         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2203                               src_res_id, LDLM_IBITS, &policy,
2204                               LCK_EX, &flags, mds_blocking_ast,
2205                               ldlm_completion_ast, NULL, NULL,
2206                               NULL, 0, NULL, &src_lockh);
2207         if (rc != ELDLM_OK)
2208                 GOTO(cleanup, rc = -ENOLCK);
2209         cleanup_phase = 2; /* lock */
2210
2211         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2212
2213         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2214         if (IS_ERR(handle)) {
2215                 rc = PTR_ERR(handle);
2216                 GOTO(cleanup, rc);
2217         }
2218         de_src->d_inode->i_nlink++;
2219         mark_inode_dirty(de_src->d_inode);
2220
2221         EXIT;
2222 cleanup:
2223         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2224                                         handle, req, rc, 0);
2225         switch (cleanup_phase) {
2226                 case 2:
2227                         if (rc)
2228                                 ldlm_lock_decref(&src_lockh, LCK_EX);
2229                         else
2230                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2231                 case 1:
2232                         l_dput(de_src);
2233                 case 0:
2234                         break;
2235                 default:
2236                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2237                         LBUG();
2238         }
2239         req->rq_status = rc;
2240         return 0;
2241 }
2242
2243 /*
2244  * request to link to foreign inode:
2245  *  - acquire i_nlinks on this inode
2246  *  - add dentry
2247  */
2248 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2249                                     int offset, struct ptlrpc_request *req,
2250                                     struct lustre_handle *lh)
2251 {
2252         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2253         struct obd_device *obd = req->rq_export->exp_obd;
2254         struct dentry *de_tgt_dir = NULL;
2255         struct mds_obd *mds = mds_req2mds(req);
2256         int rc = 0, cleanup_phase = 0;
2257         struct mdc_op_data *op_data;
2258         struct ptlrpc_request *request = NULL;
2259         int update_mode;
2260         ENTRY;
2261
2262         DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2263                   ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2264                   OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2265                   OLID4(rec->ur_id1));
2266
2267         de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2268                                           tgt_dir_lockh, &update_mode,
2269                                           rec->ur_name, rec->ur_namelen - 1,
2270                                           MDS_INODELOCK_UPDATE);
2271         if (IS_ERR(de_tgt_dir))
2272                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2273         cleanup_phase = 1;
2274
2275         OBD_ALLOC(op_data, sizeof(*op_data));
2276         if (op_data == NULL)
2277                 GOTO(cleanup, rc = -ENOMEM);
2278
2279         memset(op_data, 0, sizeof(*op_data));
2280         op_data->id1 = *(rec->ur_id1);
2281         rc = md_link(mds->mds_md_exp, op_data, &request);
2282         OBD_FREE(op_data, sizeof(*op_data));
2283         if (rc)
2284                 GOTO(cleanup, rc);
2285
2286         cleanup_phase = 2;
2287         if (request)
2288                 ptlrpc_req_finished(request);
2289
2290         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2291
2292         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2293         if (IS_ERR(handle)) {
2294                 rc = PTR_ERR(handle);
2295                 GOTO(cleanup, rc);
2296         }
2297
2298         cleanup_phase = 3;
2299
2300         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2301                                   rec->ur_namelen - 1, id_ino(rec->ur_id1),
2302                                   id_gen(rec->ur_id1), id_group(rec->ur_id1),
2303                                   id_fid(rec->ur_id1));
2304         EXIT;
2305 cleanup:
2306         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2307                                 handle, req, rc, 0);
2308
2309         switch (cleanup_phase) {
2310                 case 3:
2311                         if (rc) {
2312                                 OBD_ALLOC(op_data, sizeof(*op_data));
2313                                 if (op_data != NULL) {
2314                                         request = NULL;
2315                                         memset(op_data, 0, sizeof(*op_data));
2316
2317                                         op_data->id1 = *(rec->ur_id1);
2318                                         op_data->create_mode = rec->ur_mode;
2319
2320                                         rc = md_unlink(mds->mds_md_exp, op_data, &request);
2321                                         OBD_FREE(op_data, sizeof(*op_data));
2322                                         if (request)
2323                                                 ptlrpc_req_finished(request);
2324                                         if (rc) {
2325                                                 CERROR("error %d while dropping i_nlink on "
2326                                                        "remote inode\n", rc);
2327                                         }
2328                                 } else {
2329                                         CERROR("rc %d prevented dropping i_nlink on "
2330                                                "remote inode\n", -ENOMEM);
2331                                 }
2332                         }
2333                 case 2:
2334                 case 1:
2335                         if (rc) {
2336                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2337 #ifdef S_PDIROPS
2338                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2339 #endif
2340                         } else {
2341                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2342 #ifdef S_PDIROPS
2343                                 ptlrpc_save_lock(req, tgt_dir_lockh + 1, update_mode);
2344 #endif
2345                         }
2346                         l_dput(de_tgt_dir);
2347                         break;
2348                 default:
2349                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2350                         LBUG();
2351         }
2352         req->rq_status = rc;
2353         return 0;
2354 }
2355
2356 static int mds_reint_link(struct mds_update_record *rec, int offset,
2357                           struct ptlrpc_request *req,
2358                           struct lustre_handle *lh)
2359 {
2360         struct obd_device *obd = req->rq_export->exp_obd;
2361         struct dentry *de_src = NULL;
2362         struct dentry *de_tgt_dir = NULL;
2363         struct dentry *dchild = NULL;
2364         struct mds_obd *mds = mds_req2mds(req);
2365         struct lustre_handle *handle = NULL;
2366         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2367         struct ldlm_res_id src_res_id = { .name = {0} };
2368         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2369         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2370         ldlm_policy_data_t tgt_dir_policy =
2371                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2372         int rc = 0, cleanup_phase = 0;
2373 #ifdef S_PDIROPS
2374         int update_mode = 0;
2375 #endif
2376         ENTRY;
2377
2378         LASSERT(offset == 1);
2379
2380         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2381                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2382                   id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2383                   rec->ur_name);
2384
2385         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2386         MD_COUNTER_INCREMENT(obd, link);
2387
2388         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2389                 GOTO(cleanup, rc = -ENOENT);
2390
2391         if (id_group(rec->ur_id1) != mds->mds_num) {
2392                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2393                 RETURN(rc);
2394         }
2395
2396         if (rec->ur_namelen == 1) {
2397                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2398                 RETURN(rc);
2399         }
2400
2401         /* Step 1: Lookup the source inode and target directory by ID */
2402         de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2403         if (IS_ERR(de_src))
2404                 GOTO(cleanup, rc = PTR_ERR(de_src));
2405
2406         cleanup_phase = 1; /* source dentry */
2407
2408         de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2409         if (IS_ERR(de_tgt_dir)) {
2410                 rc = PTR_ERR(de_tgt_dir);
2411                 de_tgt_dir = NULL;
2412                 GOTO(cleanup, rc);
2413         }
2414
2415         cleanup_phase = 2; /* target directory dentry */
2416
2417         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2418                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2419                rec->ur_name, de_src->d_inode->i_ino);
2420
2421         /* Step 2: Take the two locks */
2422         src_res_id.name[0] = id_fid(rec->ur_id1);
2423         src_res_id.name[1] = id_group(rec->ur_id1);
2424         tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2425         tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2426
2427 #ifdef S_PDIROPS
2428         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2429                 int flags = 0;
2430                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2431                 if (update_mode) {
2432                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2433                                               tgt_dir_res_id, LDLM_IBITS,
2434                                               &src_policy, update_mode, &flags,
2435                                               mds_blocking_ast,
2436                                               ldlm_completion_ast, NULL, NULL,
2437                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2438                         if (rc != ELDLM_OK)
2439                                 GOTO(cleanup, rc = -ENOLCK);
2440                 }
2441
2442                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2443                                                         rec->ur_namelen - 1);
2444                 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2445                        (unsigned long)id_fid(rec->ur_id2),
2446                        (unsigned long)id_group(rec->ur_id2),
2447                        tgt_dir_res_id.name[2]);
2448         }
2449 #endif
2450         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2451                                    &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2452                                    LCK_EX, &tgt_dir_policy);
2453         if (rc)
2454                 GOTO(cleanup, rc);
2455
2456         cleanup_phase = 3; /* locks */
2457
2458         /* Step 3: Lookup the child */
2459         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2460                                    rec->ur_namelen - 1);
2461         if (IS_ERR(dchild)) {
2462                 rc = PTR_ERR(dchild);
2463                 if (rc != -EPERM && rc != -EACCES)
2464                         CERROR("child lookup error %d\n", rc);
2465                 GOTO(cleanup, rc);
2466         }
2467
2468         cleanup_phase = 4; /* child dentry */
2469
2470         if (dchild->d_inode) {
2471                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2472                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2473                 rc = -EEXIST;
2474                 GOTO(cleanup, rc);
2475         }
2476
2477         /* Step 4: Do it. */
2478         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2479
2480         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2481         if (IS_ERR(handle)) {
2482                 rc = PTR_ERR(handle);
2483                 GOTO(cleanup, rc);
2484         }
2485
2486         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2487         if (rc && rc != -EPERM && rc != -EACCES)
2488                 CERROR("vfs_link error %d\n", rc);
2489 cleanup:
2490         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2491                                 handle, req, rc, 0);
2492         EXIT;
2493
2494         switch (cleanup_phase) {
2495         case 4: /* child dentry */
2496                 l_dput(dchild);
2497         case 3: /* locks */
2498                 if (rc) {
2499                         ldlm_lock_decref(&src_lockh, LCK_EX);
2500                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2501                 } else {
2502                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2503                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2504                 }
2505         case 2: /* target dentry */
2506 #ifdef S_PDIROPS
2507                 if (tgt_dir_lockh[1].cookie && update_mode)
2508                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2509 #endif
2510                 if (de_tgt_dir)
2511                         l_dput(de_tgt_dir);
2512         case 1: /* source dentry */
2513                 l_dput(de_src);
2514         case 0:
2515                 break;
2516         default:
2517                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2518                 LBUG();
2519         }
2520         req->rq_status = rc;
2521         return 0;
2522 }
2523
2524 /* The idea here is that we need to get four locks in the end:
2525  * one on each parent directory, one on each child.  We need to take
2526  * these locks in some kind of order (to avoid deadlocks), and the order
2527  * I selected is "increasing resource number" order.  We need to look up
2528  * the children, however, before we know what the resource number(s) are.
2529  * Thus the following plan:
2530  *
2531  * 1,2. Look up the parents
2532  * 3,4. Look up the children
2533  * 5. Take locks on the parents and children, in order
2534  * 6. Verify that the children haven't changed since they were looked up
2535  *
2536  * If there was a race and the children changed since they were first looked
2537  * up, it is possible that mds_verify_child() will be able to just grab the
2538  * lock on the new child resource (if it has a higher resource than any other)
2539  * but we need to compare against not only its parent, but also against the
2540  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2541  *
2542  * We need the fancy igrab() on the child inodes because we aren't holding a
2543  * lock on the parent after the lookup is done, so dentry->d_inode may change
2544  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2545  */
2546 static int mds_get_parents_children_locked(struct obd_device *obd,
2547                                            struct mds_obd *mds,
2548                                            struct lustre_id *p1_id,
2549                                            struct dentry **de_srcdirp,
2550                                            struct lustre_id *p2_id,
2551                                            struct dentry **de_tgtdirp,
2552                                            int parent_mode,
2553                                            const char *old_name, int old_len,
2554                                            struct dentry **de_oldp,
2555                                            const char *new_name, int new_len,
2556                                            struct dentry **de_newp,
2557                                            struct lustre_handle *dlm_handles,
2558                                            int child_mode)
2559 {
2560         struct ldlm_res_id p1_res_id = { .name = {0} };
2561         struct ldlm_res_id p2_res_id = { .name = {0} };
2562         struct ldlm_res_id c1_res_id = { .name = {0} };
2563         struct ldlm_res_id c2_res_id = { .name = {0} };
2564         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2565         /* Only dentry should change, but the inode itself would be
2566            intact otherwise */
2567         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2568         /* If something is going to be replaced, both dentry and inode locks are
2569            needed */
2570         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2571                                                         MDS_INODELOCK_UPDATE}};
2572         struct ldlm_res_id *maxres_src, *maxres_tgt;
2573         struct inode *inode;
2574         int rc = 0, cleanup_phase = 0;
2575         ENTRY;
2576
2577         /* Step 1: Lookup the source directory */
2578         *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2579         if (IS_ERR(*de_srcdirp))
2580                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2581
2582         cleanup_phase = 1; /* source directory dentry */
2583
2584         p1_res_id.name[0] = id_fid(p1_id);
2585         p1_res_id.name[1] = id_group(p1_id);
2586
2587         /* Step 2: Lookup the target directory */
2588         if (id_equal_stc(p1_id, p2_id)) {
2589                 *de_tgtdirp = dget(*de_srcdirp);
2590         } else {
2591                 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2592                 if (IS_ERR(*de_tgtdirp)) {
2593                         rc = PTR_ERR(*de_tgtdirp);
2594                         *de_tgtdirp = NULL;
2595                         GOTO(cleanup, rc);
2596                 }
2597         }
2598
2599         cleanup_phase = 2; /* target directory dentry */
2600
2601         p2_res_id.name[0] = id_fid(p2_id);
2602         p2_res_id.name[1] = id_group(p2_id);
2603
2604 #ifdef S_PDIROPS
2605         dlm_handles[5].cookie = 0;
2606         dlm_handles[6].cookie = 0;
2607
2608         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2609                 /*
2610                  * get a temp lock on just fid, group to flush client cache and
2611                  * to protect dirs from concurrent splitting.
2612                  */
2613                 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2614                                            LCK_PW, &p_policy, &p2_res_id,
2615                                            &dlm_handles[6], LCK_PW, &p_policy);
2616                 if (rc != ELDLM_OK)
2617                         GOTO(cleanup, rc);
2618
2619                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2620                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2621
2622                 CDEBUG(D_INFO, "take locks on "
2623                        LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2624                        p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2625                        p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2626         }
2627         cleanup_phase = 3;
2628 #endif
2629
2630         /* Step 3: Lookup the source child entry */
2631         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2632                                      old_len - 1);
2633         if (IS_ERR(*de_oldp)) {
2634                 rc = PTR_ERR(*de_oldp);
2635                 CERROR("old child lookup error (%*s): %d\n",
2636                        old_len - 1, old_name, rc);
2637                 GOTO(cleanup, rc);
2638         }
2639
2640         cleanup_phase = 4; /* original name dentry */
2641
2642         inode = (*de_oldp)->d_inode;
2643         if (inode != NULL) {
2644                 struct lustre_id sid;
2645
2646                 inode = igrab(inode);
2647                 if (inode == NULL)
2648                         GOTO(cleanup, rc = -ENOENT);
2649
2650                 down(&inode->i_sem);
2651                 rc = mds_read_inode_sid(obd, inode, &sid);
2652                 up(&inode->i_sem);
2653                 if (rc) {
2654                         CERROR("Can't read inode self id, inode %lu, "
2655                                "rc %d\n", inode->i_ino, rc);
2656                         iput(inode);
2657                         GOTO(cleanup, rc);
2658                 }
2659
2660                 c1_res_id.name[0] = id_fid(&sid);
2661                 c1_res_id.name[1] = id_group(&sid);
2662                 iput(inode);
2663         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2664                 c1_res_id.name[0] = (*de_oldp)->d_fid;
2665                 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2666         } else {
2667                 GOTO(cleanup, rc = -ENOENT);
2668         }
2669
2670         /* Step 4: Lookup the target child entry */
2671         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2672                                      new_len - 1);
2673         if (IS_ERR(*de_newp)) {
2674                 rc = PTR_ERR(*de_newp);
2675                 CERROR("new child lookup error (%*s): %d\n",
2676                        old_len - 1, old_name, rc);
2677                 GOTO(cleanup, rc);
2678         }
2679
2680         cleanup_phase = 5; /* target dentry */
2681
2682         inode = (*de_newp)->d_inode;
2683         if (inode != NULL) {
2684                 struct lustre_id sid;
2685
2686                 inode = igrab(inode);
2687                 if (inode == NULL)
2688                         goto retry_locks;
2689
2690                 down(&inode->i_sem);
2691                 rc = mds_read_inode_sid(obd, inode, &sid);
2692                 up(&inode->i_sem);
2693                 if (rc) {
2694                         CERROR("Can't read inode self id, inode %lu, "
2695                                "rc %d\n", inode->i_ino, rc);
2696                         GOTO(cleanup, rc);
2697                 }
2698
2699                 c2_res_id.name[0] = id_fid(&sid);
2700                 c2_res_id.name[1] = id_group(&sid);
2701                 iput(inode);
2702         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2703                 c2_res_id.name[0] = (*de_newp)->d_fid;
2704                 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
2705         }
2706
2707 retry_locks:
2708         /* Step 5: Take locks on the parents and child(ren) */
2709         maxres_src = &p1_res_id;
2710         maxres_tgt = &p2_res_id;
2711         cleanup_phase = 5; /* target dentry */
2712
2713         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
2714                 maxres_src = &c1_res_id;
2715         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
2716                 maxres_tgt = &c2_res_id;
2717
2718         rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
2719                                     &p_policy,
2720                                     &p2_res_id, &dlm_handles[1], parent_mode,
2721                                     &p_policy,
2722                                     &c1_res_id, &dlm_handles[2], child_mode,
2723                                     &c1_policy,
2724                                     &c2_res_id, &dlm_handles[3], child_mode,
2725                                     &c2_policy);
2726         if (rc)
2727                 GOTO(cleanup, rc);
2728
2729         cleanup_phase = 6; /* parent and child(ren) locks */
2730
2731         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2732         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2733                               parent_mode, &c1_res_id, &dlm_handles[2],
2734                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2735                               maxres_tgt);
2736         if (rc) {
2737                 if (c2_res_id.name[0] != 0)
2738                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2739                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2740                 cleanup_phase = 5;
2741                 if (rc > 0)
2742                         goto retry_locks;
2743                 GOTO(cleanup, rc);
2744         }
2745
2746         if (!DENTRY_VALID(*de_oldp))
2747                 GOTO(cleanup, rc = -ENOENT);
2748
2749         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2750         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2751                               parent_mode, &c2_res_id, &dlm_handles[3],
2752                               de_newp, child_mode, &c2_policy, new_name,
2753                               new_len, maxres_src);
2754         if (rc) {
2755                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2756                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2757                 cleanup_phase = 5;
2758                 if (rc > 0)
2759                         goto retry_locks;
2760                 GOTO(cleanup, rc);
2761         }
2762
2763         EXIT;
2764 cleanup:
2765         if (rc) {
2766                 switch (cleanup_phase) {
2767                 case 6: /* child lock(s) */
2768                         if (c2_res_id.name[0] != 0)
2769                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2770                         if (c1_res_id.name[0] != 0)
2771                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2772                         if (dlm_handles[1].cookie != 0)
2773                                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2774                         if (dlm_handles[0].cookie != 0)
2775                                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2776                 case 5: /* target dentry */
2777                         l_dput(*de_newp);
2778                 case 4: /* source dentry */
2779                         l_dput(*de_oldp);
2780                 case 3:
2781 #ifdef S_PDIROPS
2782                         if (dlm_handles[5].cookie != 0)
2783                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2784                         if (dlm_handles[6].cookie != 0)
2785                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2786 #endif
2787                 case 2: /* target directory dentry */
2788                         l_dput(*de_tgtdirp);
2789                 case 1: /* source directry dentry */
2790                         l_dput(*de_srcdirp);
2791                 }
2792         }
2793
2794         return rc;
2795 }
2796
2797 /*
2798  * checks if dentry can be removed. This function also handles cross-ref
2799  * dentries.
2800  */
2801 static int mds_check_for_rename(struct obd_device *obd,
2802                                 struct dentry *dentry)
2803 {
2804         struct mds_obd *mds = &obd->u.mds;
2805         struct lustre_handle *rlockh;
2806         struct ptlrpc_request *req;
2807         struct mdc_op_data *op_data;
2808         struct lookup_intent it;
2809         int handle_size, rc = 0;
2810         ENTRY;
2811
2812         LASSERT(dentry != NULL);
2813
2814         if (dentry->d_inode) {
2815                 if (S_ISDIR(dentry->d_inode->i_mode) &&
2816                     !mds_is_dir_empty(obd, dentry))
2817                         rc = -ENOTEMPTY;
2818         } else {
2819                 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
2820                 handle_size = sizeof(struct lustre_handle);
2821
2822                 OBD_ALLOC(rlockh, handle_size);
2823                 if (rlockh == NULL)
2824                         RETURN(-ENOMEM);
2825
2826                 memset(rlockh, 0, handle_size);
2827                 OBD_ALLOC(op_data, sizeof(*op_data));
2828                 if (op_data == NULL) {
2829                         OBD_FREE(rlockh, handle_size);
2830                         RETURN(-ENOMEM);
2831                 }
2832                 memset(op_data, 0, sizeof(*op_data));
2833                 mds_pack_dentry2id(obd, &op_data->id1, dentry, 1);
2834
2835                 it.it_op = IT_UNLINK;
2836                 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
2837                                 op_data, rlockh, NULL, 0, ldlm_completion_ast,
2838                                 mds_blocking_ast, NULL);
2839                 OBD_FREE(op_data, sizeof(*op_data));
2840
2841                 if (rc)
2842                         RETURN(rc);
2843
2844                 if (rlockh->cookie != 0)
2845                         ldlm_lock_decref(rlockh, LCK_EX);
2846
2847                 if (it.d.lustre.it_data) {
2848                         req = (struct ptlrpc_request *)it.d.lustre.it_data;
2849                         ptlrpc_req_finished(req);
2850                 }
2851
2852                 if (it.d.lustre.it_status)
2853                         rc = it.d.lustre.it_status;
2854                 OBD_FREE(rlockh, handle_size);
2855         }
2856         RETURN(rc);
2857 }
2858
2859 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
2860                                 struct ptlrpc_request *req, struct lustre_id *id,
2861                                 struct dentry *de_dir, struct dentry *de)
2862 {
2863         struct obd_device *obd = req->rq_export->exp_obd;
2864         struct mds_obd *mds = mds_req2mds(req);
2865         void *handle = NULL;
2866         int rc = 0;
2867         ENTRY;
2868
2869         if (de->d_inode) {
2870                 /*
2871                  * name exists and points to local inode try to unlink this name
2872                  * and create new one.
2873                  */
2874                 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
2875                        obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
2876                        (unsigned long)de->d_inode->i_generation);
2877
2878                 /* checking if we can remove local dentry. */
2879                 rc = mds_check_for_rename(obd, de);
2880                 if (rc)
2881                         GOTO(cleanup, rc);
2882
2883                 handle = fsfilt_start(obd, de_dir->d_inode,
2884                                       FSFILT_OP_RENAME, NULL);
2885                 if (IS_ERR(handle))
2886                         GOTO(cleanup, rc = PTR_ERR(handle));
2887                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2888                 if (rc)
2889                         GOTO(cleanup, rc);
2890         } else if (de->d_flags & DCACHE_CROSS_REF) {
2891                 CDEBUG(D_OTHER, "%s: %s points to remote inode %lu/%lu\n",
2892                        obd->obd_name, rec->ur_tgt, (unsigned long)de->d_mdsnum,
2893                         (unsigned long)de->d_fid);
2894
2895                 /* checking if we can remove local dentry. */
2896                 rc = mds_check_for_rename(obd, de);
2897                 if (rc)
2898                         GOTO(cleanup, rc);
2899
2900                 /*
2901                  * to be fully POSIX compatible, we should add one more check:
2902                  *
2903                  * if de_new is subdir of dir rec->ur_id1. If so - return
2904                  * -EINVAL.
2905                  *
2906                  * I do not know how to implement it right now, because
2907                  * inodes/dentries for new and old names lie on different MDS,
2908                  * so add this notice here just to make it visible for the rest
2909                  * of developers and do not forget about. And when this check
2910                  * will be added, del_cross_ref should gone, that is local
2911                  * dentry is able to be removed if all checks passed. --umka
2912                  */
2913
2914                 handle = fsfilt_start(obd, de_dir->d_inode,
2915                                       FSFILT_OP_RENAME, NULL);
2916                 if (IS_ERR(handle))
2917                         GOTO(cleanup, rc = PTR_ERR(handle));
2918                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2919                 if (rc)
2920                         GOTO(cleanup, rc);
2921         } else {
2922                 /* name doesn't exist. the simplest case. */
2923                 handle = fsfilt_start(obd, de_dir->d_inode,
2924                                       FSFILT_OP_LINK, NULL);
2925                 if (IS_ERR(handle))
2926                         GOTO(cleanup, rc = PTR_ERR(handle));
2927         }
2928
2929         rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
2930                                   rec->ur_tgtlen - 1, id_ino(id),
2931                                   id_gen(id), id_group(id), id_fid(id));
2932         if (rc) {
2933                 CERROR("add_dir_entry() returned error %d\n", rc);
2934                 GOTO(cleanup, rc);
2935         }
2936
2937         EXIT;
2938 cleanup:
2939         rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2940                                 handle, req, rc, 0);
2941
2942         return rc;
2943 }
2944
2945 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
2946                                 struct ptlrpc_request *req, struct dentry *de_dir,
2947                                 struct dentry *de)
2948 {
2949         struct obd_device *obd = req->rq_export->exp_obd;
2950         struct mds_obd *mds = mds_req2mds(req);
2951         void *handle = NULL;
2952         int rc = 0;
2953         ENTRY;
2954
2955         handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
2956         if (IS_ERR(handle))
2957                 GOTO(cleanup, rc = PTR_ERR(handle));
2958         rc = fsfilt_del_dir_entry(obd, de);
2959         d_drop(de);
2960
2961         EXIT;
2962 cleanup:
2963         rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2964                                 handle, req, rc, 0);
2965         return rc;
2966 }
2967
2968 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2969                                         int offset, struct ptlrpc_request *req)
2970 {
2971         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2972         struct obd_device *obd = req->rq_export->exp_obd;
2973         struct mds_obd *mds = mds_req2mds(req);
2974         struct lustre_handle child_lockh = {0};
2975         struct dentry *de_tgtdir = NULL;
2976         struct dentry *de_new = NULL;
2977         int cleanup_phase = 0;
2978         int update_mode, rc = 0;
2979         ENTRY;
2980
2981         /*
2982          * another MDS executing rename operation has asked us to create target
2983          * name. such a creation should destroy existing target name.
2984          */
2985         CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
2986                obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
2987
2988         /* first, lookup the target */
2989         rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
2990                                          &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
2991                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2992                                          &child_lockh, &de_new, LCK_EX,
2993                                          MDS_INODELOCK_LOOKUP);
2994         if (rc)
2995                 GOTO(cleanup, rc);
2996
2997         cleanup_phase = 1;
2998
2999         LASSERT(de_tgtdir);
3000         LASSERT(de_tgtdir->d_inode);
3001         LASSERT(de_new);
3002
3003         rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
3004                                   de_tgtdir, de_new);
3005
3006         EXIT;
3007 cleanup:
3008
3009         if (cleanup_phase == 1) {
3010 #ifdef S_PDIROPS
3011                 if (parent_lockh[1].cookie != 0)
3012                         ldlm_lock_decref(parent_lockh + 1, update_mode);
3013 #endif
3014                 ldlm_lock_decref(parent_lockh, LCK_PW);
3015                 if (child_lockh.cookie != 0)
3016                         ldlm_lock_decref(&child_lockh, LCK_EX);
3017                 l_dput(de_new);
3018                 l_dput(de_tgtdir);
3019         }
3020
3021         req->rq_status = rc;
3022         return 0;
3023 }
3024
3025 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3026                                       struct ptlrpc_request *req)
3027 {
3028         struct obd_device *obd = req->rq_export->exp_obd;
3029         struct ptlrpc_request *req2 = NULL;
3030         struct dentry *de_srcdir = NULL;
3031         struct dentry *de_old = NULL;
3032         struct mds_obd *mds = mds_req2mds(req);
3033         struct lustre_handle parent_lockh[2] = {{0}, {0}};
3034         struct lustre_handle child_lockh = {0};
3035         struct mdc_op_data *op_data;
3036         int update_mode, rc = 0;
3037         ENTRY;
3038
3039         CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3040                obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3041
3042         OBD_ALLOC(op_data, sizeof(*op_data));
3043         if (op_data == NULL)
3044                 RETURN(-ENOMEM);
3045         memset(op_data, 0, sizeof(*op_data));
3046
3047         rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3048                                          &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3049                                          &update_mode, rec->ur_name,
3050                                          rec->ur_namelen, &child_lockh, &de_old,
3051                                          LCK_EX, MDS_INODELOCK_LOOKUP);
3052         LASSERT(rc == 0);
3053         LASSERT(de_srcdir);
3054         LASSERT(de_srcdir->d_inode);
3055         LASSERT(de_old);
3056
3057         /*
3058          * we already know the target should be created on another MDS so, we
3059          * have to request that MDS to do it.
3060          */
3061
3062         /* prepare source id */
3063         if (de_old->d_flags & DCACHE_CROSS_REF) {
3064                 LASSERT(de_old->d_inode == NULL);
3065                 CDEBUG(D_OTHER, "request to move remote name\n");
3066                 mds_pack_dentry2id(obd, &op_data->id1, de_old, 1);
3067         } else if (de_old->d_inode == NULL) {
3068                 /* oh, source doesn't exist */
3069                 OBD_FREE(op_data, sizeof(*op_data));
3070                 GOTO(cleanup, rc = -ENOENT);
3071         } else {
3072                 struct lustre_id sid;
3073                 struct inode *inode = de_old->d_inode;
3074
3075                 LASSERT(inode != NULL);
3076                 CDEBUG(D_OTHER, "request to move local name\n");
3077                 id_ino(&op_data->id1) = inode->i_ino;
3078                 id_group(&op_data->id1) = mds->mds_num;
3079                 id_gen(&op_data->id1) = inode->i_generation;
3080
3081                 down(&inode->i_sem);
3082                 rc = mds_read_inode_sid(obd, inode, &sid);
3083                 up(&inode->i_sem);
3084                 if (rc) {
3085                         CERROR("Can't read inode self id, "
3086                                "inode %lu, rc = %d\n",
3087                                inode->i_ino, rc);
3088                         GOTO(cleanup, rc);
3089                 }
3090
3091                 id_fid(&op_data->id1) = id_fid(&sid);
3092         }
3093
3094         op_data->id2 = *rec->ur_id2;
3095         rc = md_rename(mds->mds_md_exp, op_data, NULL, 0,
3096                        rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3097         OBD_FREE(op_data, sizeof(*op_data));
3098
3099         if (rc)
3100                 GOTO(cleanup, rc);
3101
3102         rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3103                                   de_old);
3104
3105         EXIT;
3106 cleanup:
3107         if (req2)
3108                 ptlrpc_req_finished(req2);
3109
3110 #ifdef S_PDIROPS
3111         if (parent_lockh[1].cookie != 0)
3112                 ldlm_lock_decref(parent_lockh + 1, update_mode);
3113 #endif
3114         ldlm_lock_decref(parent_lockh, LCK_PW);
3115         if (child_lockh.cookie != 0)
3116                 ldlm_lock_decref(&child_lockh, LCK_EX);
3117
3118         l_dput(de_old);
3119         l_dput(de_srcdir);
3120
3121         req->rq_status = rc;
3122         return 0;
3123 }
3124
3125 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3126                             struct ptlrpc_request *req, struct lustre_handle *lockh)
3127 {
3128         struct obd_device *obd = req->rq_export->exp_obd;
3129         struct dentry *de_srcdir = NULL;
3130         struct dentry *de_tgtdir = NULL;
3131         struct dentry *de_old = NULL;
3132         struct dentry *de_new = NULL;
3133         struct inode *old_inode = NULL, *new_inode = NULL;
3134         struct mds_obd *mds = mds_req2mds(req);
3135         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3136         struct mds_body *body = NULL;
3137         struct llog_create_locks *lcl = NULL;
3138         struct lov_mds_md *lmm = NULL;
3139         int rc = 0, cleanup_phase = 0;
3140         void *handle = NULL;
3141         ENTRY;
3142
3143         LASSERT(offset == 1);
3144
3145         DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3146                   OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3147                   rec->ur_tgt);
3148
3149         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3150
3151         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3152                 DEBUG_REQ(D_HA, req, "rename replay\n");
3153                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3154                        lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3155                        req->rq_repmsg->buflens[2]);
3156         }
3157
3158         MD_COUNTER_INCREMENT(obd, rename);
3159
3160         if (rec->ur_namelen == 1) {
3161                 rc = mds_reint_rename_create_name(rec, offset, req);
3162                 RETURN(rc);
3163         }
3164
3165         /* check if new name should be located on remote target. */
3166         if (id_group(rec->ur_id2) != mds->mds_num) {
3167                 rc = mds_reint_rename_to_remote(rec, offset, req);
3168                 RETURN(rc);
3169         }
3170
3171         rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3172                                              rec->ur_id2, &de_tgtdir, LCK_PW,
3173                                              rec->ur_name, rec->ur_namelen,
3174                                              &de_old, rec->ur_tgt,
3175                                              rec->ur_tgtlen, &de_new,
3176                                              dlm_handles, LCK_EX);
3177         if (rc)
3178                 GOTO(cleanup, rc);
3179
3180         cleanup_phase = 1; /* parent(s), children, locks */
3181         old_inode = de_old->d_inode;
3182         new_inode = de_new->d_inode;
3183
3184         /* sanity check for src inode */
3185         if (de_old->d_flags & DCACHE_CROSS_REF) {
3186                 LASSERT(de_old->d_inode == NULL);
3187
3188                 /*
3189                  * in the case of cross-ref dir, we can perform this check only
3190                  * if child and parent lie on the same mds. This is because
3191                  * otherwise they can have the same inode numbers.
3192                  */
3193                 if (de_old->d_mdsnum == mds->mds_num) {
3194                         if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3195                             de_old->d_inum == de_tgtdir->d_inode->i_ino)
3196                                 GOTO(cleanup, rc = -EINVAL);
3197                 }
3198         } else {
3199                 LASSERT(de_old->d_inode != NULL);
3200                 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3201                     de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3202                         GOTO(cleanup, rc = -EINVAL);
3203         }
3204
3205         /* sanity check for dest inode */
3206         if (de_new->d_flags & DCACHE_CROSS_REF) {
3207                 LASSERT(new_inode == NULL);
3208
3209                 /* the same check about target dentry. */
3210                 if (de_new->d_mdsnum == mds->mds_num) {
3211                         if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3212                             de_new->d_inum == de_tgtdir->d_inode->i_ino)
3213                                 GOTO(cleanup, rc = -EINVAL);
3214                 }
3215
3216                 /*
3217                  * regular files usualy do not have ->rename() implemented. But
3218                  * we handle only this case when @de_new is cross-ref entry,
3219                  * because in other cases it will be handled by vfs_rename().
3220                  */
3221                 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3222                     !de_old->d_inode->i_op->rename))
3223                         GOTO(cleanup, rc = -EPERM);
3224         } else {
3225                 if (new_inode &&
3226                     (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3227                      new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3228                         GOTO(cleanup, rc = -EINVAL);
3229
3230         }
3231
3232         /* check if inodes point to each other. */
3233         if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3234             !(de_new->d_flags & DCACHE_CROSS_REF) &&
3235             old_inode == new_inode)
3236                 GOTO(cleanup, rc = 0);
3237
3238 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3239         /*
3240          * check if we are moving old entry into its child. 2.6 does not check
3241          * for this in vfs_rename() anymore.
3242          */
3243         if (is_subdir(de_new, de_old))
3244                 GOTO(cleanup, rc = -EINVAL);
3245 #endif
3246
3247         /*
3248          * if we are about to remove the target at first, pass the EA of that
3249          * inode to client to perform and cleanup on OST.
3250          */
3251         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3252         LASSERT(body != NULL);
3253
3254         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3255         if (new_inode)
3256                 DOWN_READ_I_ALLOC_SEM(new_inode);
3257
3258         cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3259
3260         if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3261             new_inode->i_nlink == 2) ||
3262             new_inode->i_nlink == 1)) {
3263                 if (mds_orphan_open_count(new_inode) > 0) {
3264                         /* need to lock pending_dir before transaction */
3265                         down(&mds->mds_pending_dir->d_inode->i_sem);
3266                         cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3267                 } else if (S_ISREG(new_inode->i_mode)) {
3268                         mds_pack_inode2body(obd, body, new_inode, 0);
3269                         mds_pack_md(obd, req->rq_repmsg, 1, body,
3270                                     new_inode, MDS_PACK_MD_LOCK);
3271                  }
3272         }
3273
3274         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3275                        de_srcdir->d_inode->i_sb);
3276
3277         if (de_old->d_flags & DCACHE_CROSS_REF) {
3278                 struct lustre_id old_id;
3279
3280                 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3281
3282                 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3283                                           de_tgtdir, de_new);
3284                 if (rc)
3285                         GOTO(cleanup, rc);
3286
3287                 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3288                                           de_old);
3289                 GOTO(cleanup, rc);
3290         }
3291
3292         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3293         handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3294                                   NULL, le32_to_cpu(lmm->lmm_stripe_count));
3295
3296         if (IS_ERR(handle))
3297                 GOTO(cleanup, rc = PTR_ERR(handle));
3298
3299         lock_kernel();
3300         de_old->d_fsdata = req;
3301         de_new->d_fsdata = req;
3302         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3303         unlock_kernel();
3304
3305         if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3306                 if (mds_orphan_open_count(new_inode) > 0)
3307                         rc = mds_orphan_add_link(rec, obd, de_new);
3308
3309                 if (rc == 1)
3310                         GOTO(cleanup, rc = 0);
3311
3312                 if (!S_ISREG(new_inode->i_mode))
3313                         GOTO(cleanup, rc);
3314
3315                 if (!(body->valid & OBD_MD_FLEASIZE)) {
3316                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3317                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
3318                 } else if (mds_log_op_unlink(obd, new_inode,
3319                                              lustre_msg_buf(req->rq_repmsg,1,0),
3320                                              req->rq_repmsg->buflens[1],
3321                                              lustre_msg_buf(req->rq_repmsg,2,0),
3322                                              req->rq_repmsg->buflens[2],
3323                                              &lcl) > 0) {
3324                         body->valid |= OBD_MD_FLCOOKIE;
3325                 }
3326         }
3327
3328         EXIT;
3329 cleanup:
3330         rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3331                                 handle, req, rc, 0);
3332
3333         switch (cleanup_phase) {
3334         case 3:
3335                 up(&mds->mds_pending_dir->d_inode->i_sem);
3336         case 2:
3337                 if (new_inode)
3338                         UP_READ_I_ALLOC_SEM(new_inode);
3339         case 1:
3340 #ifdef S_PDIROPS
3341                 if (dlm_handles[5].cookie != 0)
3342                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3343                 if (dlm_handles[6].cookie != 0)
3344                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3345 #endif
3346                 if (lcl != NULL)
3347                         ptlrpc_save_llog_lock(req, lcl);
3348
3349                 if (rc) {
3350                         if (dlm_handles[3].cookie != 0)
3351                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3352                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3353                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3354                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3355                 } else {
3356                         if (dlm_handles[3].cookie != 0)
3357                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3358                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3359                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3360                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3361                 }
3362                 l_dput(de_new);
3363                 l_dput(de_old);
3364                 l_dput(de_tgtdir);
3365                 l_dput(de_srcdir);
3366         case 0:
3367                 break;
3368         default:
3369                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3370                 LBUG();
3371         }
3372         req->rq_status = rc;
3373         return 0;
3374 }
3375
3376 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3377                            struct ptlrpc_request *, struct lustre_handle *);
3378
3379 static mds_reinter reinters[REINT_MAX + 1] = {
3380         [REINT_SETATTR] mds_reint_setattr,
3381         [REINT_CREATE] mds_reint_create,
3382         [REINT_LINK] mds_reint_link,
3383         [REINT_UNLINK] mds_reint_unlink,
3384         [REINT_RENAME] mds_reint_rename,
3385         [REINT_OPEN] mds_open
3386 };
3387
3388 int mds_reint_rec(struct mds_update_record *rec, int offset,
3389                   struct ptlrpc_request *req, struct lustre_handle *lockh)
3390 {
3391         struct obd_device *obd = req->rq_export->exp_obd;
3392         struct lvfs_run_ctxt saved;
3393         int rc;
3394
3395         /* checked by unpacker */
3396         LASSERT(rec->ur_opcode <= REINT_MAX &&
3397                 reinters[rec->ur_opcode] != NULL);
3398
3399         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3400         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3401         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3402
3403         return rc;
3404 }