lustre/mds/mds_reint.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/mds_reint.c
   5  *  Lustre Metadata Server (mds) reintegration routines
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #ifndef EXPORT_SYMTAB
  29 # define EXPORT_SYMTAB
  30 #endif
  31 #define DEBUG_SUBSYSTEM S_MDS
  32
  33 #include <linux/fs.h>
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/obd.h>
  37 #include <linux/lustre_lib.h>
  38 #include <linux/lustre_idl.h>
  39 #include <linux/lustre_mds.h>
  40 #include <linux/lustre_dlm.h>
  41 #include <linux/lustre_log.h>
  42 #include <linux/lustre_fsfilt.h>
  43
  44 #include "mds_internal.h"
  45
  46 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  47                    int error)
  48 {
  49         obd_transno_commit_cb(obd, transno, error);
  50 }
  51
  52 struct mds_logcancel_data {
  53         struct lov_mds_md      *mlcd_lmm;
  54         int                     mlcd_size;
  55         int                     mlcd_cookielen;
  56         int                     mlcd_eadatalen;
  57         struct llog_cookie      mlcd_cookies[0];
  58 };
  59
  60
  61 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  62                                   void *cb_data, int error)
  63 {
  64         struct mds_logcancel_data *mlcd = cb_data;
  65         struct lov_stripe_md *lsm = NULL;
  66         struct llog_ctxt *ctxt;
  67         int rc;
  68
  69         obd_transno_commit_cb(obd, transno, error);
  70
  71         CDEBUG(D_HA, "cancelling %d cookies\n",
  72                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  73
  74         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  75                           mlcd->mlcd_eadatalen);
  76         if (rc < 0) {
  77                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  78                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  79                        rc);
  80         } else {
  81                 ///* XXX 0 normally, SENDNOW for debug */);
  82                 ctxt = llog_get_context(&obd->obd_llogs,
  83                                         mlcd->mlcd_cookies[0].lgc_subsys + 1);
  84                 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
  85                                  sizeof(*mlcd->mlcd_cookies),
  86                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
  87                 if (rc)
  88                         CERROR("error cancelling %d log cookies: rc %d\n",
  89                                (int)(mlcd->mlcd_cookielen /
  90                                      sizeof(*mlcd->mlcd_cookies)), rc);
  91         }
  92
  93         OBD_FREE(mlcd, mlcd->mlcd_size);
  94 }
  95
  96 /* Assumes caller has already pushed us into the kernel context. */
  97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  98                        struct ptlrpc_request *req, int rc, __u32 op_data)
  99 {
 100         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 101         struct mds_client_data *mcd = med->med_mcd;
 102         struct obd_device *obd = req->rq_export->exp_obd;
 103         int err;
 104         __u64 transno;
 105         loff_t off;
 106         int log_pri = D_HA;
 107         ENTRY;
 108
 109         /* if the export has already been failed, we have no last_rcvd slot */
 110         if (req->rq_export->exp_failed) {
 111                 CERROR("committing transaction for disconnected client\n");
 112                 if (handle)
 113                         GOTO(commit, rc);
 114                 RETURN(rc);
 115         }
 116
 117         if (IS_ERR(handle))
 118                 RETURN(rc);
 119
 120         if (handle == NULL) {
 121                 /* if we're starting our own xaction, use our own inode */
 122                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 123                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 124                 if (IS_ERR(handle)) {
 125                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 126                         RETURN(PTR_ERR(handle));
 127                 }
 128         }
 129
 130         off = med->med_off;
 131
 132         transno = req->rq_reqmsg->transno;
 133         if (rc != 0) {
 134                 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
 135         } else if (transno == 0) {
 136                 spin_lock(&mds->mds_transno_lock);
 137                 transno = ++mds->mds_last_transno;
 138                 spin_unlock(&mds->mds_transno_lock);
 139         } else {
 140                 spin_lock(&mds->mds_transno_lock);
 141                 if (transno > mds->mds_last_transno)
 142                         mds->mds_last_transno = transno;
 143                 spin_unlock(&mds->mds_transno_lock);
 144         }
 145         req->rq_repmsg->transno = req->rq_transno = transno;
 146         mcd->mcd_last_transno = cpu_to_le64(transno);
 147         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 148         mcd->mcd_last_result = cpu_to_le32(rc);
 149         mcd->mcd_last_data = cpu_to_le32(op_data);
 150
 151
 152         fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb,
 153                               transno, handle, mds_commit_cb, NULL);
 154
 155         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 156                                   &off, 0);
 157
 158         if (err) {
 159                 log_pri = D_ERROR;
 160                 if (rc == 0)
 161                         rc = err;
 162         }
 163
 164         DEBUG_REQ(log_pri, req,
 165                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 166                   transno, mcd->mcd_uuid, med->med_idx, err);
 167
 168         err = mds_lov_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 commit:
 177         err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 178         if (err) {
 179                 CERROR("error committing transaction: %d\n", err);
 180                 if (!rc)
 181                         rc = err;
 182         }
 183
 184         RETURN(rc);
 185 }
 186
 187 /* this gives the same functionality as the code between
 188  * sys_chmod and inode_setattr
 189  * chown_common and inode_setattr
 190  * utimes and inode_setattr
 191  */
 192 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 193 {
 194         time_t now = LTIME_S(CURRENT_TIME);
 195         struct iattr *attr = &rec->ur_iattr;
 196         unsigned int ia_valid = attr->ia_valid;
 197         int error;
 198         ENTRY;
 199
 200         /* only fix up attrs if the client VFS didn't already */
 201         if (!(ia_valid & ATTR_RAW))
 202                 RETURN(0);
 203
 204         if (!(ia_valid & ATTR_CTIME_SET))
 205                 LTIME_S(attr->ia_ctime) = now;
 206         if (!(ia_valid & ATTR_ATIME_SET))
 207                 LTIME_S(attr->ia_atime) = now;
 208         if (!(ia_valid & ATTR_MTIME_SET))
 209                 LTIME_S(attr->ia_mtime) = now;
 210
 211         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 212                 RETURN(-EPERM);
 213
 214         /* times */
 215         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 216                 if (rec->_ur_fsuid != inode->i_uid &&
 217                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 218                         RETURN(error);
 219         }
 220
 221         if (ia_valid & ATTR_SIZE) {
 222                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 223                         RETURN(error);
 224         }
 225
 226         if (ia_valid & ATTR_UID) {
 227                 /* chown */
 228                 error = -EPERM;
 229                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 230                         RETURN(-EPERM);
 231                 if (attr->ia_uid == (uid_t) -1)
 232                         attr->ia_uid = inode->i_uid;
 233                 if (attr->ia_gid == (gid_t) -1)
 234                         attr->ia_gid = inode->i_gid;
 235                 attr->ia_mode = inode->i_mode;
 236                 /*
 237                  * If the user or group of a non-directory has been
 238                  * changed by a non-root user, remove the setuid bit.
 239                  * 19981026 David C Niemi <niemi@tux.org>
 240                  *
 241                  * Changed this to apply to all users, including root,
 242                  * to avoid some races. This is the behavior we had in
 243                  * 2.0. The check for non-root was definitely wrong
 244                  * for 2.2 anyway, as it should have been using
 245                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 246                  */
 247                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 248                     !S_ISDIR(inode->i_mode)) {
 249                         attr->ia_mode &= ~S_ISUID;
 250                         attr->ia_valid |= ATTR_MODE;
 251                 }
 252                 /*
 253                  * Likewise, if the user or group of a non-directory
 254                  * has been changed by a non-root user, remove the
 255                  * setgid bit UNLESS there is no group execute bit
 256                  * (this would be a file marked for mandatory
 257                  * locking).  19981026 David C Niemi <niemi@tux.org>
 258                  *
 259                  * Removed the fsuid check (see the comment above) --
 260                  * 19990830 SD.
 261                  */
 262                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 263                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 264                         attr->ia_mode &= ~S_ISGID;
 265                         attr->ia_valid |= ATTR_MODE;
 266                 }
 267         } else if (ia_valid & ATTR_MODE) {
 268                 int mode = attr->ia_mode;
 269                 /* chmod */
 270                 if (attr->ia_mode == (mode_t) -1)
 271                         attr->ia_mode = inode->i_mode;
 272                 attr->ia_mode =
 273                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 274         }
 275         RETURN(0);
 276 }
 277
 278 void mds_steal_ack_locks(struct ptlrpc_request *req)
 279 {
 280         struct obd_export         *exp = req->rq_export;
 281         struct list_head          *tmp;
 282         struct ptlrpc_reply_state *oldrep;
 283         struct ptlrpc_service     *svc;
 284         struct llog_create_locks  *lcl;
 285         unsigned long              flags;
 286         char                       str[PTL_NALFMT_SIZE];
 287         int                        i;
 288
 289         /* CAVEAT EMPTOR: spinlock order */
 290         spin_lock_irqsave (&exp->exp_lock, flags);
 291         list_for_each (tmp, &exp->exp_outstanding_replies) {
 292                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 293
 294                 if (oldrep->rs_xid != req->rq_xid)
 295                         continue;
 296
 297                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 298                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 299                                 "new %d old %d\n", req->rq_xid,
 300                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 301
 302                 svc = oldrep->rs_srv_ni->sni_service;
 303                 spin_lock (&svc->srv_lock);
 304
 305                 list_del_init (&oldrep->rs_exp_list);
 306
 307                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 308                       " o%d NID %s\n",
 309                       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_obd *obd = &req->rq_export->exp_obd->u.mds;
 349         struct dentry *de;
 350         struct mds_body *body;
 351
 352         mds_req_from_mcd(req, med->med_mcd);
 353
 354         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 355         if (IS_ERR(de)) {
 356                 LASSERT(PTR_ERR(de) == req->rq_status);
 357                 return;
 358         }
 359
 360         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 361         mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
 362         mds_pack_inode2body(req2obd(req), body, de->d_inode);
 363
 364         /* Don't return OST-specific attributes if we didn't just set them */
 365         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 366                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 367         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 368                 body->valid |= OBD_MD_FLMTIME;
 369         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 370                 body->valid |= OBD_MD_FLATIME;
 371
 372         l_dput(de);
 373 }
 374
 375 /* In the raw-setattr case, we lock the child inode.
 376  * In the write-back case or if being called from open, the client holds a lock
 377  * already.
 378  *
 379  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 380 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 381                              struct ptlrpc_request *req,
 382                              struct lustre_handle *lh)
 383 {
 384         struct mds_obd *mds = mds_req2mds(req);
 385         struct obd_device *obd = req->rq_export->exp_obd;
 386         struct mds_body *body;
 387         struct dentry *de;
 388         struct inode *inode = NULL;
 389         struct lustre_handle lockh[2] = {{0}, {0}};
 390         int parent_mode;
 391         void *handle = NULL;
 392         struct mds_logcancel_data *mlcd = NULL;
 393         int rc = 0, cleanup_phase = 0, err, locked = 0;
 394         ENTRY;
 395
 396         LASSERT(offset == 0);
 397
 398         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 399                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 400
 401         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 402
 403         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 404                 de = mds_fid2dentry(mds, rec->ur_fid1, 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_fid2locked_dentry(obd, rec->ur_fid1, 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,
 455                                    mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
 456                 if (rc)
 457                         GOTO(cleanup, rc);
 458
 459                 obd_free_memmd(mds->mds_osc_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_inode2fid(obd, &body->fid1, inode);
 469         mds_pack_inode2body(obd, body, inode);
 470
 471         /* Don't return OST-specific attributes if we didn't just set them */
 472         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 473                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 474         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 475                 body->valid |= OBD_MD_FLMTIME;
 476         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 477                 body->valid |= OBD_MD_FLATIME;
 478
 479         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 480                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 481                           rec->ur_eadatalen);
 482                 if (mlcd) {
 483                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 484                                 rec->ur_eadatalen;
 485                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 486                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 487                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 488                                 mlcd->mlcd_cookielen;
 489                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 490                                mlcd->mlcd_cookielen);
 491                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 492                                mlcd->mlcd_eadatalen);
 493                 } else {
 494                         CERROR("unable to allocate log cancel data\n");
 495                 }
 496         }
 497         EXIT;
 498  cleanup:
 499         if (mlcd != NULL)
 500                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 501                                       handle, mds_cancel_cookies_cb, mlcd);
 502         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 503         switch (cleanup_phase) {
 504         case 1:
 505                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 506                     rec->ur_eadata != NULL)
 507                         up(&inode->i_sem);
 508                 l_dput(de);
 509                 if (locked) {
 510 #ifdef S_PDIROPS
 511                         if (lockh[1].cookie != 0)
 512                                 ldlm_lock_decref(lockh + 1, parent_mode);
 513 #endif
 514                         if (rc) {
 515                                 ldlm_lock_decref(lockh, LCK_PW);
 516                         } else {
 517                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 518                         }
 519                 }
 520         case 0:
 521                 break;
 522         default:
 523                 LBUG();
 524         }
 525         if (err && !rc)
 526                 rc = err;
 527
 528         req->rq_status = rc;
 529         return 0;
 530 }
 531
 532 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 533                                      struct ptlrpc_request *req)
 534 {
 535         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 536         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 537         struct dentry *parent, *child;
 538         struct mds_body *body;
 539         ENTRY;
 540
 541         mds_req_from_mcd(req, med->med_mcd);
 542
 543         if (req->rq_status) {
 544                 EXIT;
 545                 return;
 546         }
 547
 548         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 549         LASSERT(!IS_ERR(parent));
 550         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 551         LASSERT(!IS_ERR(child));
 552         if ((child->d_flags & DCACHE_CROSS_REF)) {
 553                 LASSERTF(child->d_inode == NULL, "BUG 3869");
 554                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 555                 mds_pack_dentry2fid(&body->fid1, child);
 556                 mds_pack_dentry2body(body, child);
 557                 body->valid |= OBD_MD_MDS;
 558         } else if (child->d_inode == NULL) {
 559                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 560                           rec->ur_fid1->id, rec->ur_fid1->generation,
 561                           rec->ur_name, rec->ur_mode);
 562                 LASSERTF(child->d_inode != NULL, "BUG 3869");
 563         } else {
 564                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 565                 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 566                 mds_pack_inode2body(req2obd(req), body, child->d_inode);
 567         }
 568         l_dput(parent);
 569         l_dput(child);
 570         EXIT;
 571 }
 572
 573 static int mds_reint_create(struct mds_update_record *rec, int offset,
 574                             struct ptlrpc_request *req,
 575                             struct lustre_handle *lh)
 576 {
 577         struct dentry *dparent = NULL;
 578         struct mds_obd *mds = mds_req2mds(req);
 579         struct obd_device *obd = req->rq_export->exp_obd;
 580         struct dentry *dchild = NULL;
 581         struct inode *dir = NULL;
 582         void *handle = NULL;
 583         struct lustre_handle lockh[2] = {{0}, {0}};
 584         int parent_mode;
 585         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 586         int created = 0;
 587         struct dentry_params dp;
 588         struct mea *mea = NULL;
 589         int mea_size;
 590         ENTRY;
 591
 592         LASSERT(offset == 0);
 593         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 594
 595         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 596                   rec->ur_fid1->id, rec->ur_fid1->generation,
 597                   rec->ur_name, rec->ur_mode);
 598
 599         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 600
 601         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 602                 GOTO(cleanup, rc = -ESTALE);
 603
 604         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 605                                         &parent_mode, rec->ur_name,
 606                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 607         if (IS_ERR(dparent)) {
 608                 rc = PTR_ERR(dparent);
 609                 CERROR("parent lookup error %d\n", rc);
 610                 GOTO(cleanup, rc);
 611         }
 612         cleanup_phase = 1; /* locked parent dentry */
 613         dir = dparent->d_inode;
 614         LASSERT(dir);
 615
 616         ldlm_lock_dump_handle(D_OTHER, lockh);
 617
 618         /* try to retrieve MEA data for this dir */
 619         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 620         if (rc)
 621                 GOTO(cleanup, rc);
 622
 623         if (mea != NULL) {
 624                 /* dir is already splitted, check is requested filename
 625                  * should live at this MDS or at another one */
 626                 int i;
 627                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 628                 if (mea->mea_master != mea->mea_fids[i].mds) {
 629                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 630                                " should be %d(%d)\n",
 631                                mea->mea_master, dparent->d_inode->i_ino,
 632                                dparent->d_inode->i_generation, rec->ur_name,
 633                                mea->mea_fids[i].mds, i);
 634                         GOTO(cleanup, rc = -ERESTART);
 635                 }
 636         }
 637
 638         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 639         if (IS_ERR(dchild)) {
 640                 rc = PTR_ERR(dchild);
 641                 CERROR("child lookup error %d\n", 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         dchild->d_fsdata = (void *)&dp;
 669         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 670         dp.p_ptr = req;
 671
 672         switch (type) {
 673         case S_IFREG:{
 674                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 675                 if (IS_ERR(handle))
 676                         GOTO(cleanup, rc = PTR_ERR(handle));
 677                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 678                 EXIT;
 679                 break;
 680         }
 681         case S_IFDIR:{
 682                 int nstripes = 0;
 683                 int i;
 684
 685                 /* as Peter asked, mkdir() should distribute new directories
 686                  * over the whole cluster in order to distribute namespace
 687                  * processing load. first, we calculate which MDS to use to put
 688                  * new directory's inode in. */
 689                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 690                                       rec->ur_flags);
 691                 if (i == mds->mds_num) {
 692                         /* inode will be created locally */
 693                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 694                         if (IS_ERR(handle))
 695                                 GOTO(cleanup, rc = PTR_ERR(handle));
 696
 697                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 698
 699                         if (rec->ur_eadata)
 700                                 nstripes = *(u16 *)rec->ur_eadata;
 701
 702                         if (rc == 0 && nstripes) {
 703                                 /* we pass LCK_EX to split routine to
 704                                  * signalthat we have exclusive access
 705                                  * to the directory. simple because
 706                                  * nobody knows it already exists -bzzz */
 707                                 rc = mds_try_to_split_dir(obd, dchild,
 708                                                           NULL, nstripes,
 709                                                           LCK_EX);
 710                                 if (rc > 0) {
 711                                         /* dir got splitted */
 712                                         rc = 0;
 713                                 } else if (rc < 0) {
 714                                         /* an error occured during
 715                                          * splitting. */
 716                                         GOTO(cleanup, rc);
 717                                 }
 718                         }
 719                 } else if (!DENTRY_VALID(dchild)) {
 720                         /* inode will be created on another MDS */
 721                         struct obdo *oa = NULL;
 722                         struct mds_body *body;
 723
 724                         /* first, create that inode */
 725                         oa = obdo_alloc();
 726                         LASSERT(oa != NULL);
 727                         oa->o_mds = i;
 728                         oa->o_easize = 0;
 729                         if (rec->ur_eadata) {
 730                                 /* user asks for creating splitted dir */
 731                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 732                         }
 733
 734                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 735                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 736                                         OBD_MD_FLUID | OBD_MD_FLGID);
 737                         oa->o_mode = dir->i_mode;
 738                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 739                                         obd->obd_name, i);
 740                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 741                                 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
 742                                        obd->obd_name, rec->ur_namelen - 1,
 743                                        rec->ur_name, (unsigned) rec->ur_fid2->id,
 744                                        (unsigned) rec->ur_fid2->generation);
 745                                 oa->o_id = rec->ur_fid2->id;
 746                                 oa->o_generation = rec->ur_fid2->generation;
 747                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 748                         }
 749
 750                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 751                         if (rc) {
 752                                 obdo_free(oa);
 753                                 GOTO(cleanup, rc);
 754                         }
 755
 756                         /* now, add new dir entry for it */
 757                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 758                         if (IS_ERR(handle)) {
 759                                 obdo_free(oa);
 760                                 GOTO(cleanup, rc = PTR_ERR(handle));
 761                         }
 762                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 763                                                   rec->ur_namelen - 1,
 764                                                   oa->o_id, oa->o_generation,
 765                                                   i);
 766                         LASSERT(rc == 0);
 767
 768                         /* fill reply */
 769                         body = lustre_msg_buf(req->rq_repmsg,
 770                                               offset, sizeof (*body));
 771                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 772                         body->fid1.id = oa->o_id;
 773                         body->fid1.mds = i;
 774                         body->fid1.generation = oa->o_generation;
 775                         obdo_free(oa);
 776                 } else {
 777                         /* requested name exists in the directory */
 778                         rc = -EEXIST;
 779                 }
 780                 EXIT;
 781                 break;
 782         }
 783         case S_IFLNK:{
 784                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 785                 if (IS_ERR(handle))
 786                         GOTO(cleanup, rc = PTR_ERR(handle));
 787                 if (rec->ur_tgt == NULL)        /* no target supplied */
 788                         rc = -EINVAL;           /* -EPROTO? */
 789                 else
 790                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 791                 EXIT;
 792                 break;
 793         }
 794         case S_IFCHR:
 795         case S_IFBLK:
 796         case S_IFIFO:
 797         case S_IFSOCK:{
 798                 int rdev = rec->ur_rdev;
 799                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 800                 if (IS_ERR(handle))
 801                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 802                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 803                 EXIT;
 804                 break;
 805         }
 806         default:
 807                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 808                 dchild->d_fsdata = NULL;
 809                 GOTO(cleanup, rc = -EINVAL);
 810         }
 811
 812         /* In case we stored the desired inum in here, we want to clean up. */
 813         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 814                 dchild->d_fsdata = NULL;
 815
 816         if (rc) {
 817                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 818                 GOTO(cleanup, rc);
 819         } else if (dchild->d_inode) {
 820                 struct iattr iattr;
 821                 struct inode *inode = dchild->d_inode;
 822                 struct mds_body *body;
 823
 824                 created = 1;
 825                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 826                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 827                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 828                 iattr.ia_uid = rec->_ur_fsuid;
 829                 if (dir->i_mode & S_ISGID)
 830                         iattr.ia_gid = dir->i_gid;
 831                 else
 832                         iattr.ia_gid = rec->_ur_fsgid;
 833                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 834                         ATTR_MTIME | ATTR_CTIME;
 835
 836                 if (rec->ur_fid2->id) {
 837                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 838                         inode->i_generation = rec->ur_fid2->generation;
 839                         /* Dirtied and committed by the upcoming setattr. */
 840                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 841                                inode->i_ino, inode->i_generation);
 842                 } else {
 843 #if 0
 844                         struct lustre_handle child_ino_lockh;
 845 #endif
 846
 847                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 848                                inode->i_ino, inode->i_generation);
 849
 850 #if 0
 851                         /* The inode we were allocated may have just been freed
 852                          * by an unlink operation.  We take this lock to
 853                          * synchronize against the matching reply-ack-lock taken
 854                          * in unlink, to avoid replay problems if this reply
 855                          * makes it out to the client but the unlink's does not.
 856                          * See bug 2029 for more detail.*/
 857                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 858                         if (rc != ELDLM_OK) {
 859                                 CERROR("error locking for unlink/create sync: "
 860                                        "%d\n", rc);
 861                         } else {
 862                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 863                         }
 864 #endif
 865                 }
 866
 867                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 868                 if (rc)
 869                         CERROR("error on child setattr: rc = %d\n", rc);
 870
 871                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 872                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 873                 if (rc)
 874                         CERROR("error on parent setattr: rc = %d\n", rc);
 875
 876                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 877                 mds_pack_inode2fid(obd, &body->fid1, inode);
 878                 mds_pack_inode2body(obd, body, inode);
 879         }
 880         EXIT;
 881
 882 cleanup:
 883         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 884
 885         if (rc && created) {
 886                 /* Destroy the file we just created. This should not need extra
 887                  * journal credits, as we have already modified all of the
 888                  * blocks needed in order to create the file in the first
 889                  * place. */
 890                 switch (type) {
 891                 case S_IFDIR:
 892                         err = vfs_rmdir(dir, dchild);
 893                         if (err)
 894                                 CERROR("rmdir in error path: %d\n", err);
 895                         break;
 896                 default:
 897                         err = vfs_unlink(dir, dchild);
 898                         if (err)
 899                                 CERROR("unlink in error path: %d\n", err);
 900                         break;
 901                 }
 902         } else {
 903                 rc = err;
 904         }
 905         switch (cleanup_phase) {
 906         case 2: /* child dentry */
 907                 l_dput(dchild);
 908         case 1: /* locked parent dentry */
 909 #ifdef S_PDIROPS
 910                 if (lockh[1].cookie != 0)
 911                         ldlm_lock_decref(lockh + 1, parent_mode);
 912 #endif
 913                 if (rc) {
 914                         ldlm_lock_decref(lockh, LCK_PW);
 915                 } else {
 916                         ptlrpc_save_lock(req, lockh, LCK_PW);
 917                 }
 918                 l_dput(dparent);
 919         case 0:
 920                 break;
 921         default:
 922                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 923                 LBUG();
 924         }
 925         if (mea)
 926                 OBD_FREE(mea, mea_size);
 927         req->rq_status = rc;
 928         return 0;
 929 }
 930
 931 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 932            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 933 {
 934         int i;
 935
 936         for (i = 0; i < RES_NAME_SIZE; i++) {
 937                 /* return 1 here, because enqueue_ordered will skip resources
 938                  * of all zeroes if they're sorted to the end of the list. */
 939                 if (res1->name[i] == 0 && res2->name[i] != 0)
 940                         return 1;
 941                 if (res2->name[i] == 0 && res1->name[i] != 0)
 942                         return 0;
 943
 944                 if (res1->name[i] > res2->name[i])
 945                         return 1;
 946                 if (res1->name[i] < res2->name[i])
 947                         return 0;
 948         }
 949
 950         if (!p1 || !p2)
 951                 return 0;
 952
 953         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 954                 return 1;
 955
 956         return 0;
 957 }
 958
 959 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 960  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 961  * because they take the place of local semaphores.
 962  *
 963  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 964  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 965 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 966                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 967                           ldlm_policy_data_t *p1_policy,
 968                           struct ldlm_res_id *p2_res_id,
 969                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 970                           ldlm_policy_data_t *p2_policy)
 971 {
 972         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 973         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 974         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 975         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 976         int rc, flags;
 977         ENTRY;
 978
 979         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 980
 981         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 982                res_id[1]->name[0]);
 983
 984         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 985                 handles[1] = p1_lockh;
 986                 handles[0] = p2_lockh;
 987                 res_id[1] = p1_res_id;
 988                 res_id[0] = p2_res_id;
 989                 lock_modes[1] = p1_lock_mode;
 990                 lock_modes[0] = p2_lock_mode;
 991                 policies[1] = p1_policy;
 992                 policies[0] = p2_policy;
 993         }
 994
 995         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 996                res_id[0]->name[0], res_id[1]->name[0]);
 997
 998         flags = LDLM_FL_LOCAL_ONLY;
 999         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1000                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1001                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1002                               NULL, 0, NULL, handles[0]);
1003         if (rc != ELDLM_OK)
1004                 RETURN(-EIO);
1005         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1006
1007         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1008             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1009                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1010                 ldlm_lock_addref(handles[1], lock_modes[1]);
1011         } else if (res_id[1]->name[0] != 0) {
1012                 flags = LDLM_FL_LOCAL_ONLY;
1013                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1014                                       *res_id[1], LDLM_IBITS, policies[1],
1015                                       lock_modes[1], &flags, mds_blocking_ast,
1016                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1017                                       NULL, handles[1]);
1018                 if (rc != ELDLM_OK) {
1019                         ldlm_lock_decref(handles[0], lock_modes[0]);
1020                         RETURN(-EIO);
1021                 }
1022                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1023         }
1024
1025         RETURN(0);
1026 }
1027
1028 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1029                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1030                            ldlm_policy_data_t *p1_policy,
1031                            struct ldlm_res_id *p2_res_id,
1032                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1033                            ldlm_policy_data_t *p2_policy,
1034                            struct ldlm_res_id *c1_res_id,
1035                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1036                            ldlm_policy_data_t *c1_policy,
1037                            struct ldlm_res_id *c2_res_id,
1038                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1039                            ldlm_policy_data_t *c2_policy)
1040 {
1041         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1042                                           c1_res_id, c2_res_id };
1043         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1044                                                  c1_lockh, c2_lockh };
1045         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1046                               c1_lock_mode, c2_lock_mode };
1047         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1048                                             c1_policy, c2_policy};
1049         int rc, i, j, sorted, flags;
1050         ENTRY;
1051
1052         CDEBUG(D_DLMTRACE,
1053                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1054                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1055                res_id[3]->name[0]);
1056
1057         /* simple insertion sort - we have at most 4 elements */
1058         for (i = 1; i < 4; i++) {
1059                 j = i - 1;
1060                 dlm_handles[4] = dlm_handles[i];
1061                 res_id[4] = res_id[i];
1062                 lock_modes[4] = lock_modes[i];
1063                 policies[4] = policies[i];
1064
1065                 sorted = 0;
1066                 do {
1067                         if (res_gt(res_id[j], res_id[4], policies[j],
1068                                    policies[4])) {
1069                                 dlm_handles[j + 1] = dlm_handles[j];
1070                                 res_id[j + 1] = res_id[j];
1071                                 lock_modes[j + 1] = lock_modes[j];
1072                                 policies[j + 1] = policies[j];
1073                                 j--;
1074                         } else {
1075                                 sorted = 1;
1076                         }
1077                 } while (j >= 0 && !sorted);
1078
1079                 dlm_handles[j + 1] = dlm_handles[4];
1080                 res_id[j + 1] = res_id[4];
1081                 lock_modes[j + 1] = lock_modes[4];
1082                 policies[j + 1] = policies[4];
1083         }
1084
1085         CDEBUG(D_DLMTRACE,
1086                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1087                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1088                res_id[3]->name[0]);
1089
1090         /* XXX we could send ASTs on all these locks first before blocking? */
1091         for (i = 0; i < 4; i++) {
1092                 flags = 0;
1093                 if (res_id[i]->name[0] == 0)
1094                         break;
1095                 if (i != 0 &&
1096                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1097                     (policies[i]->l_inodebits.bits &
1098                      policies[i-1]->l_inodebits.bits) ) {
1099                         memcpy(dlm_handles[i], dlm_handles[i-1],
1100                                sizeof(*(dlm_handles[i])));
1101                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1102                 } else {
1103                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1104                                               *res_id[i], LDLM_IBITS,
1105                                               policies[i],
1106                                               lock_modes[i], &flags,
1107                                               mds_blocking_ast,
1108                                               ldlm_completion_ast, NULL, NULL,
1109                                               NULL, 0, NULL, dlm_handles[i]);
1110                         if (rc != ELDLM_OK)
1111                                 GOTO(out_err, rc = -EIO);
1112                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1113                 }
1114         }
1115
1116         RETURN(0);
1117 out_err:
1118         while (i-- > 0)
1119                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1120
1121         return rc;
1122 }
1123
1124 /* In the unlikely case that the child changed while we were waiting
1125  * on the lock, we need to drop the lock on the old child and either:
1126  * - if the child has a lower resource name, then we have to also
1127  *   drop the parent lock and regain the locks in the right order
1128  * - in the rename case, if the child has a lower resource name than one of
1129  *   the other parent/child resources (maxres) we also need to reget the locks
1130  * - if the child has a higher resource name (this is the common case)
1131  *   we can just get the lock on the new child (still in lock order)
1132  *
1133  * Returns 0 if the child did not change or if it changed but could be locked.
1134  * Returns 1 if the child changed and we need to re-lock (no locks held).
1135  * Returns -ve error with a valid dchild (no locks held). */
1136 static int mds_verify_child(struct obd_device *obd,
1137                             struct ldlm_res_id *parent_res_id,
1138                             struct lustre_handle *parent_lockh,
1139                             struct dentry *dparent, int parent_mode,
1140                             struct ldlm_res_id *child_res_id,
1141                             struct lustre_handle *child_lockh,
1142                             struct dentry **dchildp, int child_mode,
1143                             ldlm_policy_data_t *child_policy,
1144                             const char *name, int namelen,
1145                             struct ldlm_res_id *maxres)
1146 {
1147         struct dentry *vchild, *dchild = *dchildp;
1148         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1149         ENTRY;
1150
1151         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1152         if (IS_ERR(vchild))
1153                 GOTO(cleanup, rc = PTR_ERR(vchild));
1154
1155         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1156                 if  (child_res_id->name[0] == vchild->d_inum &&
1157                                 child_res_id->name[1] == vchild->d_generation) {
1158                         if (dchild != NULL)
1159                                 l_dput(dchild);
1160                         *dchildp = vchild;
1161                         RETURN(0);
1162                 }
1163                 goto changed;
1164         }
1165
1166         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1167                    (vchild->d_inode != NULL &&
1168                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1169                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1170                 if (dchild != NULL)
1171                         l_dput(dchild);
1172                 *dchildp = vchild;
1173
1174                 RETURN(0);
1175         }
1176
1177 changed:
1178         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1179                vchild->d_inode, dchild ? dchild->d_inode : 0,
1180                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1181                child_res_id->name[0]);
1182         if (child_res_id->name[0] != 0)
1183                 ldlm_lock_decref(child_lockh, child_mode);
1184         if (dchild)
1185                 l_dput(dchild);
1186
1187         cleanup_phase = 1; /* parent lock only */
1188         *dchildp = dchild = vchild;
1189
1190         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1191                 int flags = 0;
1192                 if (dchild->d_inode) {
1193                         child_res_id->name[0] = dchild->d_inode->i_ino;
1194                         child_res_id->name[1] = dchild->d_inode->i_generation;
1195                 } else {
1196                         child_res_id->name[0] = dchild->d_inum;
1197                         child_res_id->name[1] = dchild->d_generation;
1198                 }
1199
1200                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1201                     res_gt(maxres, child_res_id, NULL, NULL)) {
1202                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1203                                child_res_id->name[0], parent_res_id->name[0],
1204                                maxres->name[0]);
1205                         GOTO(cleanup, rc = 1);
1206                 }
1207
1208                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1209                                       *child_res_id, LDLM_IBITS, child_policy,
1210                                       child_mode, &flags, mds_blocking_ast,
1211                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1212                                       NULL, child_lockh);
1213                 if (rc != ELDLM_OK)
1214                         GOTO(cleanup, rc = -EIO);
1215
1216         } else {
1217                 memset(child_res_id, 0, sizeof(*child_res_id));
1218         }
1219
1220         EXIT;
1221 cleanup:
1222         if (rc) {
1223                 switch(cleanup_phase) {
1224                 case 2:
1225                         if (child_res_id->name[0] != 0)
1226                                 ldlm_lock_decref(child_lockh, child_mode);
1227                 case 1:
1228                         ldlm_lock_decref(parent_lockh, parent_mode);
1229                 }
1230         }
1231         return rc;
1232 }
1233
1234 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1235                                 struct ll_fid *fid,
1236                                 struct lustre_handle *parent_lockh,
1237                                 struct dentry **dparentp, int parent_mode,
1238                                 __u64 parent_lockpart, int *update_mode,
1239                                 char *name, int namelen,
1240                                 struct lustre_handle *child_lockh,
1241                                 struct dentry **dchildp, int child_mode,
1242                                 __u64 child_lockpart, void *clone_info)
1243 {
1244         struct ldlm_res_id child_res_id = { .name = {0} };
1245         struct ldlm_res_id parent_res_id = { .name = {0} };
1246         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1247         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1248         struct inode *inode;
1249         int rc = 0, cleanup_phase = 0;
1250         ENTRY;
1251
1252         /* Step 1: Lookup parent */
1253         *dparentp = mds_fid2dentry(mds, fid, NULL);
1254         if (IS_ERR(*dparentp)) {
1255                 rc = PTR_ERR(*dparentp);
1256                 *dparentp = NULL;
1257                 RETURN(rc);
1258         }
1259
1260         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1261                (*dparentp)->d_inode->i_ino, name);
1262
1263         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1264         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1265 #ifdef S_PDIROPS
1266         parent_lockh[1].cookie = 0;
1267         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1268                 struct ldlm_res_id res_id = { .name = {0} };
1269                 ldlm_policy_data_t policy;
1270                 int flags = 0;
1271                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1272                 if (*update_mode) {
1273                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1274                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1275                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1276                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1277                                               res_id, LDLM_IBITS, &policy,
1278                                               *update_mode, &flags,
1279                                               mds_blocking_ast,
1280                                               ldlm_completion_ast,
1281                                               NULL, NULL, NULL, 0, NULL,
1282                                               parent_lockh + 1);
1283                         if (rc != ELDLM_OK)
1284                                 RETURN(-ENOLCK);
1285                 }
1286
1287                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1288                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1289                        (*dparentp)->d_inode->i_ino,
1290                        (*dparentp)->d_inode->i_generation,
1291                        parent_res_id.name[2]);
1292         }
1293 #endif
1294
1295         cleanup_phase = 1; /* parent dentry */
1296 #ifdef CONFIG_SNAPFS
1297         if (clone_info) {
1298                 /*FIXME is there any other FUNC will use d_fsdata,
1299                  *excepet creating inode according inum*/
1300                 (*dparentp)->d_fsdata = clone_info;
1301         }
1302 #endif
1303
1304         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1305         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1306         if (IS_ERR(*dchildp)) {
1307                 rc = PTR_ERR(*dchildp);
1308                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1309                 GOTO(cleanup, rc);
1310         }
1311
1312         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1313                 /* inode lives on another MDS: return * mds/ino/gen
1314                  * and LOOKUP lock. drop possible UPDATE lock! */
1315                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1316                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1317                 child_res_id.name[0] = (*dchildp)->d_inum;
1318                 child_res_id.name[1] = (*dchildp)->d_generation;
1319                 goto retry_locks;
1320         }
1321
1322         inode = (*dchildp)->d_inode;
1323         if (inode != NULL)
1324                 inode = igrab(inode);
1325         if (inode == NULL)
1326                 goto retry_locks;
1327
1328         child_res_id.name[0] = inode->i_ino;
1329         child_res_id.name[1] = inode->i_generation;
1330
1331         iput(inode);
1332
1333 retry_locks:
1334         cleanup_phase = 2; /* child dentry */
1335
1336         /* Step 3: Lock parent and child in resource order.  If child doesn't
1337          *         exist, we still have to lock the parent and re-lookup. */
1338         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1339                                    &parent_policy, &child_res_id, child_lockh,
1340                                    child_mode, &child_policy);
1341         if (rc)
1342                 GOTO(cleanup, rc);
1343
1344         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1345                 cleanup_phase = 4; /* child lock */
1346         else
1347                 cleanup_phase = 3; /* parent lock */
1348
1349         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1350         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1351                               parent_mode, &child_res_id, child_lockh,
1352                               dchildp, child_mode, &child_policy,
1353                               name, namelen, &parent_res_id);
1354         if (rc > 0)
1355                 goto retry_locks;
1356         if (rc < 0) {
1357                 cleanup_phase = 3;
1358                 GOTO(cleanup, rc);
1359         }
1360
1361 cleanup:
1362         if (rc) {
1363                 switch (cleanup_phase) {
1364                 case 4:
1365                         ldlm_lock_decref(child_lockh, child_mode);
1366                 case 3:
1367                         ldlm_lock_decref(parent_lockh, parent_mode);
1368                 case 2:
1369                         l_dput(*dchildp);
1370                 case 1:
1371 #ifdef S_PDIROPS
1372                         if (parent_lockh[1].cookie)
1373                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1374 #endif
1375                         l_dput(*dparentp);
1376                 default: ;
1377                 }
1378         }
1379         return rc;
1380 }
1381
1382 void mds_reconstruct_generic(struct ptlrpc_request *req)
1383 {
1384         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1385
1386         mds_req_from_mcd(req, med->med_mcd);
1387 }
1388
1389 int mds_create_local_dentry(struct mds_update_record *rec,
1390                            struct obd_device *obd)
1391 {
1392         struct mds_obd *mds = &obd->u.mds;
1393         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1394         int fidlen = 0, rc, cleanup_phase = 0;
1395         struct dentry *new_child = NULL;
1396         char *fidname = rec->ur_name;
1397         struct dentry *child = NULL;
1398         struct lustre_handle lockh[2] = {{0}, {0}};
1399         void *handle;
1400         ENTRY;
1401
1402         down(&fids_dir->i_sem);
1403         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1404         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1405                         fidname, (unsigned) rec->ur_fid1->id,
1406                         (unsigned) rec->ur_fid1->generation);
1407
1408         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1409         up(&fids_dir->i_sem);
1410         if (IS_ERR(new_child)) {
1411                 CERROR("can't lookup %s: %d\n", fidname,
1412                                 (int) PTR_ERR(new_child));
1413                 GOTO(cleanup, rc = PTR_ERR(new_child));
1414         }
1415         cleanup_phase = 1;
1416
1417         if (new_child->d_inode != NULL) {
1418                 /* nice. we've already have local dentry! */
1419                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1420                        (unsigned) new_child->d_inode->i_ino,
1421                        (unsigned) new_child->d_inode->i_generation);
1422                 rec->ur_fid1->id = fids_dir->i_ino;
1423                 rec->ur_fid1->generation = fids_dir->i_generation;
1424                 rec->ur_namelen = fidlen + 1;
1425                 GOTO(cleanup, rc = 0);
1426         }
1427
1428         /* new, local dentry will be added soon. we need no aliases here */
1429         d_drop(new_child);
1430
1431         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1432                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1433         } else {
1434                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1435                                               LCK_EX, lockh, NULL, NULL, 0,
1436                                               MDS_INODELOCK_UPDATE);
1437         }
1438
1439         if (IS_ERR(child)) {
1440                 rc = PTR_ERR(child);
1441                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1442                         CERROR("can't get victim: %d\n", rc);
1443                 GOTO(cleanup, rc);
1444         }
1445         cleanup_phase = 2;
1446
1447         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1448         if (IS_ERR(handle))
1449                 GOTO(cleanup, rc = PTR_ERR(handle));
1450
1451         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1452                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1453                                   mds->mds_num);
1454         if (rc)
1455                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1456                        (unsigned long) child->d_inode->i_ino,
1457                        (unsigned long) child->d_inode->i_generation, rc);
1458         else {
1459                 if (S_ISDIR(child->d_inode->i_mode)) {
1460                         fids_dir->i_nlink++;
1461                         mark_inode_dirty(fids_dir);
1462                 }
1463                 mark_inode_dirty(child->d_inode);
1464         }
1465         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1466
1467         rec->ur_fid1->id = fids_dir->i_ino;
1468         rec->ur_fid1->generation = fids_dir->i_generation;
1469         rec->ur_namelen = fidlen + 1;
1470
1471 cleanup:
1472         switch(cleanup_phase) {
1473                 case 2:
1474                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1475                                 ldlm_lock_decref(lockh, LCK_EX);
1476                         dput(child);
1477                 case 1:
1478                         dput(new_child);
1479                 case 0:
1480                        break;
1481         }
1482         RETURN(rc);
1483 }
1484
1485 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1486                                         struct ptlrpc_request *slave)
1487 {
1488         void *cookie, *cookie2;
1489         struct mds_body *body2;
1490         struct mds_body *body;
1491         void *ea, *ea2;
1492         ENTRY;
1493
1494         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1495         LASSERT(body != NULL);
1496
1497         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1498         LASSERT(body2 != NULL);
1499
1500         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1501                 RETURN(0);
1502         }
1503
1504         memcpy(body2, body, sizeof(*body));
1505         body2->valid &= ~OBD_MD_FLCOOKIE;
1506
1507         if (!(body->valid & OBD_MD_FLEASIZE) &&
1508             !(body->valid & OBD_MD_FLDIREA))
1509                 RETURN(0);
1510
1511         if (body->eadatasize == 0) {
1512                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1513                 RETURN(0);
1514         }
1515
1516         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1517
1518         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1519         LASSERT(ea != NULL);
1520
1521         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1522         LASSERT(ea2 != NULL);
1523
1524         memcpy(ea2, ea, body->eadatasize);
1525
1526         if (body->valid & OBD_MD_FLCOOKIE) {
1527                 LASSERT(master->rq_repmsg->buflens[2] >=
1528                                 slave->rq_repmsg->buflens[2]);
1529                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1530                                 slave->rq_repmsg->buflens[2]);
1531                 LASSERT(cookie != NULL);
1532
1533                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1534                                 master->rq_repmsg->buflens[2]);
1535                 LASSERT(cookie2 != NULL);
1536                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1537                 body2->valid |= OBD_MD_FLCOOKIE;
1538         }
1539         RETURN(0);
1540 }
1541
1542 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1543                                    struct ptlrpc_request *req,
1544                                    struct lustre_handle *parent_lockh,
1545                                    int update_mode,
1546                                    struct dentry *dparent,
1547                                    struct lustre_handle *child_lockh,
1548                                    struct dentry *dchild)
1549 {
1550         struct obd_device *obd = req->rq_export->exp_obd;
1551         struct mds_obd *mds = mds_req2mds(req);
1552         struct mdc_op_data op_data;
1553         int rc = 0, cleanup_phase = 0;
1554         struct ptlrpc_request *request = NULL;
1555         void *handle;
1556         ENTRY;
1557
1558         LASSERT(offset == 0 || offset == 2);
1559
1560         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1561                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1562                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1563         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1564                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1565                           rec->ur_namelen - 1, rec->ur_name,
1566                           (unsigned)dchild->d_mdsnum,
1567                           (unsigned) dchild->d_inum,
1568                           (unsigned) dchild->d_generation);
1569
1570         /* time to drop i_nlink on remote MDS */
1571         memset(&op_data, 0, sizeof(op_data));
1572         op_data.fid1.mds = dchild->d_mdsnum;
1573         op_data.fid1.id = dchild->d_inum;
1574         op_data.fid1.generation = dchild->d_generation;
1575         op_data.create_mode = rec->ur_mode;
1576         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1577                 op_data.create_mode |= MDS_MODE_REPLAY;
1578         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1579         cleanup_phase = 2;
1580         if (request) {
1581                 if (rc == 0)
1582                         mds_copy_unlink_reply(req, request);
1583                 ptlrpc_req_finished(request);
1584         }
1585         if (rc == 0) {
1586                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1587                                       NULL);
1588                 if (IS_ERR(handle))
1589                         GOTO(cleanup, rc = PTR_ERR(handle));
1590                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1591                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1592                                         rc, 0);
1593         }
1594 cleanup:
1595         req->rq_status = rc;
1596
1597 #ifdef S_PDIROPS
1598         if (parent_lockh[1].cookie != 0)
1599                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1600 #endif
1601         ldlm_lock_decref(child_lockh, LCK_EX);
1602         if (rc)
1603                 ldlm_lock_decref(parent_lockh, LCK_PW);
1604         else
1605                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1606         l_dput(dchild);
1607         l_dput(dparent);
1608
1609         return 0;
1610 }
1611
1612 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1613                             struct ptlrpc_request *req,
1614                             struct lustre_handle *lh)
1615 {
1616         struct dentry *dparent, *dchild;
1617         struct mds_obd *mds = mds_req2mds(req);
1618         struct obd_device *obd = req->rq_export->exp_obd;
1619         struct mds_body *body = NULL;
1620         struct inode *child_inode;
1621         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1622         struct lustre_handle child_lockh = {0};
1623 #if 0
1624         struct lustre_handle child_reuse_lockh = {0};
1625 #endif
1626         struct lustre_handle * slave_lockh = NULL;
1627         struct llog_create_locks *lcl = NULL;
1628         char fidname[LL_FID_NAMELEN];
1629         void *handle = NULL;
1630         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1631         int unlink_by_fid = 0;
1632         int update_mode;
1633         ENTRY;
1634
1635         LASSERT(offset == 0 || offset == 2);
1636
1637         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1638                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1639
1640         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1641
1642         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1643                 DEBUG_REQ(D_HA, req, "unlink replay\n");
1644                 memcpy(lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1645                        lustre_msg_buf(req->rq_reqmsg, 2, 0),
1646                        req->rq_repmsg->buflens[offset + 2]);
1647         }
1648
1649         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1650                 GOTO(cleanup, rc = -ENOENT);
1651
1652         if (rec->ur_namelen == 1) {
1653                 /* this is request to drop i_nlink on local inode */
1654                 unlink_by_fid = 1;
1655                 rec->ur_name = fidname;
1656                 rc = mds_create_local_dentry(rec, obd);
1657                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1658                         DEBUG_REQ(D_HA, req,
1659                                   "drop nlink on inode %u/%u/%u (replay)",
1660                                   (unsigned) rec->ur_fid1->mds,
1661                                   (unsigned) rec->ur_fid1->id,
1662                                   (unsigned) rec->ur_fid1->generation);
1663                         req->rq_status = 0;
1664                         RETURN(0);
1665                 }
1666         }
1667
1668         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1669                 /* master mds for directory asks slave removing
1670                  * inode is already locked */
1671                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1672                                                LCK_PW, parent_lockh,
1673                                                &update_mode, rec->ur_name,
1674                                                rec->ur_namelen,
1675                                                MDS_INODELOCK_UPDATE);
1676                 if (IS_ERR(dparent))
1677                         GOTO(cleanup, rc = PTR_ERR(dparent));
1678                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1679                                            rec->ur_namelen - 1);
1680                 if (IS_ERR(dchild))
1681                         GOTO(cleanup, rc = PTR_ERR(dchild));
1682                 child_lockh.cookie = 0;
1683                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1684                 LASSERT(dchild->d_inode != NULL);
1685                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1686         } else {
1687                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1688                                                  parent_lockh, &dparent,
1689                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1690                                                  &update_mode, rec->ur_name,
1691                                                  rec->ur_namelen, &child_lockh,
1692                                                  &dchild, LCK_EX,
1693                                                  MDS_INODELOCK_LOOKUP |
1694                                                  MDS_INODELOCK_UPDATE, NULL);
1695         }
1696         if (rc)
1697                 GOTO(cleanup, rc);
1698
1699         if (dchild->d_flags & DCACHE_CROSS_REF) {
1700                 /* we should have parent lock only here */
1701                 LASSERT(unlink_by_fid == 0);
1702                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1703                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1704                                         update_mode, dparent, &child_lockh, dchild);
1705                 RETURN(0);
1706         }
1707
1708         cleanup_phase = 1; /* dchild, dparent, locks */
1709
1710         dget(dchild);
1711         child_inode = dchild->d_inode;
1712         if (child_inode == NULL) {
1713                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1714                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1715                 GOTO(cleanup, rc = -ENOENT);
1716         }
1717
1718         cleanup_phase = 2; /* dchild has a lock */
1719
1720         /* We have to do these checks ourselves, in case we are making an
1721          * orphan.  The client tells us whether rmdir() or unlink() was called,
1722          * so we need to return appropriate errors (bug 72).
1723          *
1724          * We don't have to check permissions, because vfs_rename (called from
1725          * mds_open_unlink_rename) also calls may_delete. */
1726         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1727                 if (!S_ISDIR(child_inode->i_mode))
1728                         GOTO(cleanup, rc = -ENOTDIR);
1729         } else {
1730                 if (S_ISDIR(child_inode->i_mode))
1731                         GOTO(cleanup, rc = -EISDIR);
1732         }
1733
1734         /* handle splitted dir */
1735         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1736         if (rc)
1737                 GOTO(cleanup, rc);
1738
1739 #if 0
1740         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1741          * reuse (see bug 2029). */
1742         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1743         if (rc != ELDLM_OK)
1744                 GOTO(cleanup, rc);
1745 #endif
1746         cleanup_phase = 3; /* child inum lock */
1747
1748         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1749
1750         /* ldlm_reply in buf[0] if called via intent */
1751         if (offset)
1752                 offset = 1;
1753
1754         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1755         LASSERT(body != NULL);
1756
1757         /* If this is the last reference to this inode, get the OBD EA
1758          * data first so the client can destroy OST objects.
1759          * we only do the object removal if no open files remain.
1760          * Nobody can get at this name anymore because of the locks so
1761          * we make decisions here as to whether to remove the inode */
1762         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1763             mds_open_orphan_count(child_inode) == 0) {
1764                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1765                 mds_pack_inode2body(obd, body, child_inode);
1766                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1767                             child_inode, 1);
1768                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1769                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1770                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1771                 } else {
1772                         log_unlink = 1;
1773                 }
1774         }
1775
1776         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1777         switch (child_inode->i_mode & S_IFMT) {
1778         case S_IFDIR:
1779                 /* Drop any lingering child directories before we start our
1780                  * transaction, to avoid doing multiple inode dirty/delete
1781                  * in our compound transaction (bug 1321). */
1782                 shrink_dcache_parent(dchild);
1783                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1784                                       NULL);
1785                 if (IS_ERR(handle))
1786                         GOTO(cleanup, rc = PTR_ERR(handle));
1787                 cleanup_phase = 4; /* transaction */
1788                 rc = vfs_rmdir(dparent->d_inode, dchild);
1789                 break;
1790         case S_IFREG: {
1791 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1792                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1793                                                         offset + 1, 0);
1794                 handle = fsfilt_start_log(obd, dparent->d_inode,
1795                                           FSFILT_OP_UNLINK, NULL,
1796                                           le32_to_cpu(lmm->lmm_stripe_count));
1797                 if (IS_ERR(handle))
1798                         GOTO(cleanup, rc = PTR_ERR(handle));
1799
1800                 cleanup_phase = 4; /* transaction */
1801                 rc = vfs_unlink(dparent->d_inode, dchild);
1802
1803                 if (!rc && log_unlink)
1804                         if (mds_log_op_unlink(obd, child_inode,
1805                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1806                                 req->rq_repmsg->buflens[offset + 1],
1807                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1808                                 req->rq_repmsg->buflens[offset + 2], &lcl) > 0)
1809                                 body->valid |= OBD_MD_FLCOOKIE;
1810                 break;
1811         }
1812         case S_IFLNK:
1813         case S_IFCHR:
1814         case S_IFBLK:
1815         case S_IFIFO:
1816         case S_IFSOCK:
1817                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1818                                       NULL);
1819                 if (IS_ERR(handle))
1820                         GOTO(cleanup, rc = PTR_ERR(handle));
1821                 cleanup_phase = 4; /* transaction */
1822                 rc = vfs_unlink(dparent->d_inode, dchild);
1823                 break;
1824         default:
1825                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1826                        rec->ur_name);
1827                 LBUG();
1828                 GOTO(cleanup, rc = -EINVAL);
1829         }
1830
1831  cleanup:
1832         if (rc == 0) {
1833                 struct iattr iattr;
1834                 int err;
1835
1836                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1837                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1838                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1839
1840                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1841                 if (err)
1842                         CERROR("error on parent setattr: rc = %d\n", err);
1843         }
1844
1845         switch(cleanup_phase) {
1846         case 4:
1847                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1848                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1849                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1850                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1851
1852                         /* filesystem is really going to destroy an inode
1853                          * we have to delay this till inode is opened -bzzz */
1854                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1855                 }
1856                 /* handle splitted dir */
1857                 if (rc == 0) {
1858                         /* master directory can be non-empty or something else ... */
1859                         mds_unlink_slave_objs(obd, dchild);
1860                 }
1861                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1862                                         rc, 0);
1863                 if (!rc)
1864                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1865                                            "unlinked", 0, NULL);
1866                 if (lcl != NULL)
1867                         ptlrpc_save_llog_lock(req, lcl);
1868         case 3: /* child ino-reuse lock */
1869 #if 0
1870                 if (rc && body != NULL) {
1871                         // Don't unlink the OST objects if the MDS unlink failed
1872                         body->valid = 0;
1873                 }
1874                 if (rc)
1875                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1876                 else
1877                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1878 #endif
1879         case 2: /* child lock */
1880                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1881                 if (child_lockh.cookie)
1882                         ldlm_lock_decref(&child_lockh, LCK_EX);
1883         case 1: /* child and parent dentry, parent lock */
1884 #ifdef S_PDIROPS
1885                 if (parent_lockh[1].cookie != 0)
1886                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1887 #endif
1888                 if (rc)
1889                         ldlm_lock_decref(parent_lockh, LCK_PW);
1890                 else
1891                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1892                 l_dput(dchild);
1893                 l_dput(dchild);
1894                 l_dput(dparent);
1895         case 0:
1896                 break;
1897         default:
1898                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1899                 LBUG();
1900         }
1901         req->rq_status = rc;
1902         return 0;
1903 }
1904
1905 /*
1906  * to service requests from remote MDS to increment i_nlink
1907  */
1908 static int mds_reint_link_acquire(struct mds_update_record *rec,
1909                                  int offset, struct ptlrpc_request *req,
1910                                  struct lustre_handle *lh)
1911 {
1912         struct obd_device *obd = req->rq_export->exp_obd;
1913         struct ldlm_res_id src_res_id = { .name = {0} };
1914         struct lustre_handle *handle = NULL, src_lockh = {0};
1915         struct mds_obd *mds = mds_req2mds(req);
1916         int rc = 0, cleanup_phase = 0;
1917         struct dentry *de_src = NULL;
1918         ldlm_policy_data_t policy;
1919         int flags = 0;
1920         ENTRY;
1921
1922         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1923                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1924                   (unsigned) rec->ur_fid1->id,
1925                   (unsigned) rec->ur_fid1->generation);
1926
1927         /* Step 1: Lookup the source inode and target directory by FID */
1928         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1929         if (IS_ERR(de_src))
1930                 GOTO(cleanup, rc = PTR_ERR(de_src));
1931         cleanup_phase = 1; /* source dentry */
1932
1933         src_res_id.name[0] = de_src->d_inode->i_ino;
1934         src_res_id.name[1] = de_src->d_inode->i_generation;
1935         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1936
1937         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1938                         src_res_id, LDLM_IBITS, &policy,
1939                         LCK_EX, &flags, mds_blocking_ast,
1940                         ldlm_completion_ast, NULL, NULL,
1941                         NULL, 0, NULL, &src_lockh);
1942         if (rc != ELDLM_OK)
1943                 GOTO(cleanup, rc = -ENOLCK);
1944         cleanup_phase = 2; /* lock */
1945
1946         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1947
1948         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1949         if (IS_ERR(handle)) {
1950                 rc = PTR_ERR(handle);
1951                 GOTO(cleanup, rc);
1952         }
1953         de_src->d_inode->i_nlink++;
1954         mark_inode_dirty(de_src->d_inode);
1955
1956 cleanup:
1957         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1958                                         handle, req, rc, 0);
1959         EXIT;
1960         switch (cleanup_phase) {
1961                 case 2:
1962                         if (rc)
1963                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1964                         else
1965                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1966                 case 1:
1967                         l_dput(de_src);
1968                 case 0:
1969                         break;
1970                 default:
1971                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1972                         LBUG();
1973         }
1974         req->rq_status = rc;
1975         return 0;
1976 }
1977
1978 /*
1979  * request to link to foreign inode:
1980  *  - acquire i_nlinks on this inode
1981  *  - add dentry
1982  */
1983 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1984                                     int offset, struct ptlrpc_request *req,
1985                                     struct lustre_handle *lh)
1986 {
1987         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1988         struct obd_device *obd = req->rq_export->exp_obd;
1989         struct dentry *de_tgt_dir = NULL;
1990         struct mds_obd *mds = mds_req2mds(req);
1991         int rc = 0, cleanup_phase = 0;
1992         struct mdc_op_data op_data;
1993         struct ptlrpc_request *request = NULL;
1994         int update_mode;
1995         ENTRY;
1996
1997 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1998         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1999                   (unsigned) rec->ur_fid2->mds,
2000                   (unsigned) rec->ur_fid2->id,
2001                   (unsigned) rec->ur_fid2->generation,
2002                   rec->ur_namelen - 1, rec->ur_name,
2003                   (unsigned) rec->ur_fid1->mds,
2004                   (unsigned) rec->ur_fid1->id,
2005                   (unsigned)rec->ur_fid1->generation);
2006
2007         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
2008                                            tgt_dir_lockh, &update_mode,
2009                                            rec->ur_name, rec->ur_namelen - 1,
2010                                            MDS_INODELOCK_UPDATE);
2011         if (IS_ERR(de_tgt_dir))
2012                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2013         cleanup_phase = 1;
2014
2015         op_data.fid1 = *(rec->ur_fid1);
2016         op_data.namelen = 0;
2017         op_data.name = NULL;
2018         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2019         if (rc)
2020                 GOTO(cleanup, rc);
2021
2022         cleanup_phase = 2;
2023         if (request)
2024                 ptlrpc_req_finished(request);
2025
2026         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2027
2028         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2029         if (IS_ERR(handle)) {
2030                 rc = PTR_ERR(handle);
2031                 GOTO(cleanup, rc);
2032         }
2033
2034         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2035                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2036                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2037         cleanup_phase = 3;
2038
2039 cleanup:
2040         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2041                                 handle, req, rc, 0);
2042         EXIT;
2043
2044         switch (cleanup_phase) {
2045                 case 3:
2046                         if (rc) {
2047                                 /* FIXME: drop i_nlink on remote inode here */
2048                                 CERROR("MUST drop drop i_nlink here\n");
2049                         }
2050                 case 2:
2051                 case 1:
2052                         if (rc) {
2053                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2054 #ifdef S_PDIROPS
2055                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2056 #endif
2057                         } else {
2058                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2059 #ifdef S_PDIROPS
2060                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2061 #endif
2062                         }
2063                         l_dput(de_tgt_dir);
2064                         break;
2065                 default:
2066                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2067                         LBUG();
2068         }
2069         req->rq_status = rc;
2070         return 0;
2071 }
2072
2073 static int mds_reint_link(struct mds_update_record *rec, int offset,
2074                           struct ptlrpc_request *req,
2075                           struct lustre_handle *lh)
2076 {
2077         struct obd_device *obd = req->rq_export->exp_obd;
2078         struct dentry *de_src = NULL;
2079         struct dentry *de_tgt_dir = NULL;
2080         struct dentry *dchild = NULL;
2081         struct mds_obd *mds = mds_req2mds(req);
2082         struct lustre_handle *handle = NULL;
2083         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2084         struct ldlm_res_id src_res_id = { .name = {0} };
2085         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2086         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2087         ldlm_policy_data_t tgt_dir_policy =
2088                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2089         int rc = 0, cleanup_phase = 0;
2090         int update_mode = 0;
2091         ENTRY;
2092
2093         LASSERT(offset == 0);
2094
2095         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2096                   rec->ur_fid1->id, rec->ur_fid1->generation,
2097                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2098
2099         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2100
2101 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2102         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2103                 GOTO(cleanup, rc = -ENOENT);
2104
2105         if (rec->ur_fid1->mds != mds->mds_num) {
2106                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2107                 RETURN(rc);
2108         }
2109
2110         if (rec->ur_namelen == 1) {
2111                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2112                 RETURN(rc);
2113         }
2114
2115         /* Step 1: Lookup the source inode and target directory by FID */
2116         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2117         if (IS_ERR(de_src))
2118                 GOTO(cleanup, rc = PTR_ERR(de_src));
2119
2120         cleanup_phase = 1; /* source dentry */
2121
2122         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2123         if (IS_ERR(de_tgt_dir)) {
2124                 rc = PTR_ERR(de_tgt_dir);
2125                 de_tgt_dir = NULL;
2126                 GOTO(cleanup, rc);
2127         }
2128
2129         cleanup_phase = 2; /* target directory dentry */
2130
2131         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2132                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2133                de_src->d_inode->i_ino);
2134
2135         /* Step 2: Take the two locks */
2136         src_res_id.name[0] = de_src->d_inode->i_ino;
2137         src_res_id.name[1] = de_src->d_inode->i_generation;
2138         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2139         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2140 #ifdef S_PDIROPS
2141         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2142                 int flags = 0;
2143                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2144                 if (update_mode) {
2145                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2146                                               tgt_dir_res_id, LDLM_IBITS,
2147                                               &src_policy, update_mode, &flags,
2148                                               mds_blocking_ast,
2149                                               ldlm_completion_ast, NULL, NULL,
2150                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2151                         if (rc != ELDLM_OK)
2152                                 GOTO(cleanup, rc = -ENOLCK);
2153                 }
2154
2155                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2156                                                         rec->ur_namelen - 1);
2157                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2158                        de_tgt_dir->d_inode->i_ino,
2159                        de_tgt_dir->d_inode->i_generation,
2160                        tgt_dir_res_id.name[2]);
2161         }
2162 #endif
2163         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2164                                    &src_policy,
2165                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2166                                    &tgt_dir_policy);
2167         if (rc)
2168                 GOTO(cleanup, rc);
2169
2170         cleanup_phase = 3; /* locks */
2171
2172         /* Step 3: Lookup the child */
2173         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2174         if (IS_ERR(dchild)) {
2175                 rc = PTR_ERR(dchild);
2176                 if (rc != -EPERM && rc != -EACCES)
2177                         CERROR("child lookup error %d\n", rc);
2178                 GOTO(cleanup, rc);
2179         }
2180
2181         cleanup_phase = 4; /* child dentry */
2182
2183         if (dchild->d_inode) {
2184                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2185                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2186                 rc = -EEXIST;
2187                 GOTO(cleanup, rc);
2188         }
2189
2190         /* Step 4: Do it. */
2191         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2192
2193         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2194         if (IS_ERR(handle)) {
2195                 rc = PTR_ERR(handle);
2196                 GOTO(cleanup, rc);
2197         }
2198
2199         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2200         if (rc && rc != -EPERM && rc != -EACCES)
2201                 CERROR("vfs_link error %d\n", rc);
2202 cleanup:
2203         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2204                                 handle, req, rc, 0);
2205         EXIT;
2206
2207         switch (cleanup_phase) {
2208         case 4: /* child dentry */
2209                 l_dput(dchild);
2210         case 3: /* locks */
2211                 if (rc) {
2212                         ldlm_lock_decref(&src_lockh, LCK_EX);
2213                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2214                 } else {
2215                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2216                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2217                 }
2218         case 2: /* target dentry */
2219 #ifdef S_PDIROPS
2220                 if (tgt_dir_lockh[1].cookie && update_mode)
2221                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2222 #endif
2223                 if (de_tgt_dir)
2224                         l_dput(de_tgt_dir);
2225         case 1: /* source dentry */
2226                 l_dput(de_src);
2227         case 0:
2228                 break;
2229         default:
2230                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2231                 LBUG();
2232         }
2233         req->rq_status = rc;
2234         return 0;
2235 }
2236
2237 /*
2238  * add a hard link in the PENDING directory, only used by rename()
2239  */
2240 static int mds_add_link_orphan(struct mds_update_record *rec,
2241                                struct obd_device *obd,
2242                                struct dentry *dentry)
2243 {
2244         struct mds_obd *mds = &obd->u.mds;
2245         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2246         struct dentry *pending_child;
2247         char fidname[LL_FID_NAMELEN];
2248         int fidlen = 0, rc;
2249         ENTRY;
2250
2251         LASSERT(dentry->d_inode);
2252         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2253
2254         down(&pending_dir->i_sem);
2255         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2256                             dentry->d_inode->i_generation);
2257
2258         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2259                mds_open_orphan_count(dentry->d_inode),
2260                rec->ur_name, fidname);
2261
2262         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2263         if (IS_ERR(pending_child))
2264                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2265
2266         if (pending_child->d_inode != NULL) {
2267                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2268                 LASSERT(pending_child->d_inode == dentry->d_inode);
2269                 GOTO(out_dput, rc = 0);
2270         }
2271
2272         lock_kernel();
2273         rc = vfs_link(dentry, pending_dir, pending_child);
2274         unlock_kernel();
2275         if (rc)
2276                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2277                        rec->ur_name, rc);
2278         else
2279                 mds_inode_set_orphan(dentry->d_inode);
2280 out_dput:
2281         l_dput(pending_child);
2282 out_lock:
2283         up(&pending_dir->i_sem);
2284         RETURN(rc);
2285 }
2286
2287 /* The idea here is that we need to get four locks in the end:
2288  * one on each parent directory, one on each child.  We need to take
2289  * these locks in some kind of order (to avoid deadlocks), and the order
2290  * I selected is "increasing resource number" order.  We need to look up
2291  * the children, however, before we know what the resource number(s) are.
2292  * Thus the following plan:
2293  *
2294  * 1,2. Look up the parents
2295  * 3,4. Look up the children
2296  * 5. Take locks on the parents and children, in order
2297  * 6. Verify that the children haven't changed since they were looked up
2298  *
2299  * If there was a race and the children changed since they were first looked
2300  * up, it is possible that mds_verify_child() will be able to just grab the
2301  * lock on the new child resource (if it has a higher resource than any other)
2302  * but we need to compare against not only its parent, but also against the
2303  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2304  *
2305  * We need the fancy igrab() on the child inodes because we aren't holding a
2306  * lock on the parent after the lookup is done, so dentry->d_inode may change
2307  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2308  */
2309 static int mds_get_parents_children_locked(struct obd_device *obd,
2310                                            struct mds_obd *mds,
2311                                            struct ll_fid *p1_fid,
2312                                            struct dentry **de_srcdirp,
2313                                            struct ll_fid *p2_fid,
2314                                            struct dentry **de_tgtdirp,
2315                                            int parent_mode,
2316                                            const char *old_name, int old_len,
2317                                            struct dentry **de_oldp,
2318                                            const char *new_name, int new_len,
2319                                            struct dentry **de_newp,
2320                                            struct lustre_handle *dlm_handles,
2321                                            int child_mode)
2322 {
2323         struct ldlm_res_id p1_res_id = { .name = {0} };
2324         struct ldlm_res_id p2_res_id = { .name = {0} };
2325         struct ldlm_res_id c1_res_id = { .name = {0} };
2326         struct ldlm_res_id c2_res_id = { .name = {0} };
2327         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2328         /* Only dentry should change, but the inode itself would be
2329            intact otherwise */
2330         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2331         /* If something is going to be replaced, both dentry and inode locks are
2332            needed */
2333         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2334                                                         MDS_INODELOCK_UPDATE}};
2335         struct ldlm_res_id *maxres_src, *maxres_tgt;
2336         struct inode *inode;
2337         int rc = 0, cleanup_phase = 0;
2338         ENTRY;
2339
2340         /* Step 1: Lookup the source directory */
2341         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2342         if (IS_ERR(*de_srcdirp))
2343                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2344
2345         cleanup_phase = 1; /* source directory dentry */
2346
2347         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2348         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2349
2350         /* Step 2: Lookup the target directory */
2351         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2352                 *de_tgtdirp = dget(*de_srcdirp);
2353         } else {
2354                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2355                 if (IS_ERR(*de_tgtdirp)) {
2356                         rc = PTR_ERR(*de_tgtdirp);
2357                         *de_tgtdirp = NULL;
2358                         GOTO(cleanup, rc);
2359                 }
2360         }
2361
2362         cleanup_phase = 2; /* target directory dentry */
2363
2364         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2365         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2366
2367 #ifdef S_PDIROPS
2368         dlm_handles[5].cookie = 0;
2369         dlm_handles[6].cookie = 0;
2370         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2371                 /* Get a temp lock on just ino, gen to flush client cache and
2372                  * to protect dirs from concurrent splitting */
2373                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2374                                            LCK_PW, &p_policy, &p2_res_id,
2375                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2376                 if (rc != ELDLM_OK)
2377                         GOTO(cleanup, rc);
2378                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2379                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2380                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2381                        (*de_srcdirp)->d_inode->i_ino,
2382                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2383                        (*de_tgtdirp)->d_inode->i_ino,
2384                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2385         }
2386         cleanup_phase = 3;
2387 #endif
2388
2389         /* Step 3: Lookup the source child entry */
2390         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2391         if (IS_ERR(*de_oldp)) {
2392                 rc = PTR_ERR(*de_oldp);
2393                 CERROR("old child lookup error (%*s): %d\n",
2394                        old_len - 1, old_name, rc);
2395                 GOTO(cleanup, rc);
2396         }
2397
2398         cleanup_phase = 4; /* original name dentry */
2399
2400         inode = (*de_oldp)->d_inode;
2401         if (inode != NULL) {
2402                 inode = igrab(inode);
2403                 if (inode == NULL)
2404                         GOTO(cleanup, rc = -ENOENT);
2405
2406                 c1_res_id.name[0] = inode->i_ino;
2407                 c1_res_id.name[1] = inode->i_generation;
2408                 iput(inode);
2409         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2410                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2411                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2412         }
2413
2414         /* Step 4: Lookup the target child entry */
2415         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2416         if (IS_ERR(*de_newp)) {
2417                 rc = PTR_ERR(*de_newp);
2418                 CERROR("new child lookup error (%*s): %d\n",
2419                        old_len - 1, old_name, rc);
2420                 GOTO(cleanup, rc);
2421         }
2422
2423         cleanup_phase = 5; /* target dentry */
2424
2425         inode = (*de_newp)->d_inode;
2426         if (inode != NULL) {
2427                 inode = igrab(inode);
2428                 if (inode == NULL)
2429                         goto retry_locks;
2430
2431                 c2_res_id.name[0] = inode->i_ino;
2432                 c2_res_id.name[1] = inode->i_generation;
2433                 iput(inode);
2434         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2435                 c2_res_id.name[0] = (*de_newp)->d_inum;
2436                 c2_res_id.name[1] = (*de_newp)->d_generation;
2437         }
2438
2439 retry_locks:
2440         /* Step 5: Take locks on the parents and child(ren) */
2441         maxres_src = &p1_res_id;
2442         maxres_tgt = &p2_res_id;
2443         cleanup_phase = 5; /* target dentry */
2444
2445         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2446                 maxres_src = &c1_res_id;
2447         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2448                 maxres_tgt = &c2_res_id;
2449
2450         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2451                                     &p_policy,
2452                                     &p2_res_id, &dlm_handles[1], parent_mode,
2453                                     &p_policy,
2454                                     &c1_res_id, &dlm_handles[2], child_mode,
2455                                     &c1_policy,
2456                                     &c2_res_id, &dlm_handles[3], child_mode,
2457                                     &c2_policy);
2458         if (rc)
2459                 GOTO(cleanup, rc);
2460
2461         cleanup_phase = 6; /* parent and child(ren) locks */
2462
2463         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2464         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2465                               parent_mode, &c1_res_id, &dlm_handles[2],
2466                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2467                               maxres_tgt);
2468         if (rc) {
2469                 if (c2_res_id.name[0] != 0)
2470                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2471                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2472                 cleanup_phase = 5;
2473                 if (rc > 0)
2474                         goto retry_locks;
2475                 GOTO(cleanup, rc);
2476         }
2477
2478         if (!DENTRY_VALID(*de_oldp))
2479                 GOTO(cleanup, rc = -ENOENT);
2480
2481         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2482         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2483                               parent_mode, &c2_res_id, &dlm_handles[3],
2484                               de_newp, child_mode, &c2_policy, new_name,
2485                               new_len, maxres_src);
2486         if (rc) {
2487                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2488                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2489                 cleanup_phase = 5;
2490                 if (rc > 0)
2491                         goto retry_locks;
2492                 GOTO(cleanup, rc);
2493         }
2494
2495         EXIT;
2496 cleanup:
2497         if (rc) {
2498                 switch (cleanup_phase) {
2499                 case 6: /* child lock(s) */
2500                         if (c2_res_id.name[0] != 0)
2501                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2502                         if (c1_res_id.name[0] != 0)
2503                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2504                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2505                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2506                 case 5: /* target dentry */
2507                         l_dput(*de_newp);
2508                 case 4: /* source dentry */
2509                         l_dput(*de_oldp);
2510                 case 3:
2511 #ifdef S_PDIROPS
2512                         if (dlm_handles[5].cookie != 0)
2513                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2514                         if (dlm_handles[6].cookie != 0)
2515                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2516 #endif
2517                 case 2: /* target directory dentry */
2518                         l_dput(*de_tgtdirp);
2519                 case 1: /* source directry dentry */
2520                         l_dput(*de_srcdirp);
2521                 }
2522         }
2523
2524         return rc;
2525 }
2526
2527 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2528                                         int offset, struct ptlrpc_request *req)
2529 {
2530         struct obd_device *obd = req->rq_export->exp_obd;
2531         struct dentry *de_srcdir = NULL;
2532         struct dentry *de_new = NULL;
2533         struct mds_obd *mds = mds_req2mds(req);
2534         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2535         struct lustre_handle child_lockh = {0};
2536         int cleanup_phase = 0;
2537         void *handle = NULL;
2538         int update_mode, rc = 0;
2539         ENTRY;
2540
2541         /* another MDS executing rename operation has asked us
2542          * to create target name. such a creation should destroy
2543          * existing target name */
2544
2545         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2546                         obd->obd_name, rec->ur_tgt,
2547                         (unsigned long) rec->ur_fid1->mds,
2548                         (unsigned long) rec->ur_fid1->id,
2549                         (unsigned long) rec->ur_fid1->generation);
2550
2551         /* first, lookup the target */
2552         child_lockh.cookie = 0;
2553         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2554                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2555                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2556                                          &child_lockh, &de_new, LCK_EX,
2557                                          MDS_INODELOCK_LOOKUP, NULL);
2558         if (rc)
2559                 GOTO(cleanup, rc);
2560
2561         cleanup_phase = 1;
2562
2563         LASSERT(de_srcdir);
2564         LASSERT(de_srcdir->d_inode);
2565         LASSERT(de_new);
2566
2567         if (de_new->d_inode) {
2568                 /* name exists and points to local inode
2569                  * try to unlink this name and create new one */
2570                 CERROR("%s: %s points to local inode %lu/%lu\n",
2571                        obd->obd_name, rec->ur_tgt,
2572                        (unsigned long) de_new->d_inode->i_ino,
2573                        (unsigned long) de_new->d_inode->i_generation);
2574                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2575                                       FSFILT_OP_RENAME, NULL);
2576                 if (IS_ERR(handle))
2577                         GOTO(cleanup, rc = PTR_ERR(handle));
2578                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2579                 if (rc)
2580                         GOTO(cleanup, rc);
2581         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2582                 /* name exists adn points to remove inode */
2583                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2584                        obd->obd_name, rec->ur_tgt,
2585                        (unsigned long) de_new->d_mdsnum,
2586                        (unsigned long) de_new->d_inum,
2587                        (unsigned long) de_new->d_generation);
2588         } else {
2589                 /* name doesn't exist. the simplest case */
2590                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2591                                       FSFILT_OP_LINK, NULL);
2592                 if (IS_ERR(handle))
2593                         GOTO(cleanup, rc = PTR_ERR(handle));
2594         }
2595
2596         cleanup_phase = 2;
2597         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2598                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2599                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2600         if (rc)
2601                 CERROR("add_dir_entry() returned error %d\n", rc);
2602 cleanup:
2603         EXIT;
2604         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2605                                 handle, req, rc, 0);
2606         switch(cleanup_phase) {
2607                 case 2:
2608                 case 1:
2609 #ifdef S_PDIROPS
2610                         if (parent_lockh[1].cookie != 0)
2611                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2612 #endif
2613                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2614                         if (child_lockh.cookie != 0)
2615                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2616                         l_dput(de_new);
2617                         l_dput(de_srcdir);
2618                         break;
2619                 default:
2620                         LBUG();
2621         }
2622
2623         req->rq_status = rc;
2624
2625         RETURN(0);
2626 }
2627
2628 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2629                                       struct ptlrpc_request *req)
2630 {
2631         struct obd_device *obd = req->rq_export->exp_obd;
2632         struct ptlrpc_request *req2 = NULL;
2633         struct dentry *de_srcdir = NULL;
2634         struct dentry *de_old = NULL;
2635         struct mds_obd *mds = mds_req2mds(req);
2636         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2637         struct lustre_handle child_lockh = {0};
2638         struct mdc_op_data opdata;
2639         void *handle = NULL;
2640         int update_mode, rc = 0;
2641         ENTRY;
2642
2643         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2644                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2645         memset(&opdata, 0, sizeof(opdata));
2646
2647         child_lockh.cookie = 0;
2648         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2649                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2650                                          &update_mode, rec->ur_name,
2651                                          rec->ur_namelen, &child_lockh, &de_old,
2652                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2653         LASSERT(rc == 0);
2654         LASSERT(de_srcdir);
2655         LASSERT(de_srcdir->d_inode);
2656         LASSERT(de_old);
2657
2658         /* we already know the target should be created on another MDS
2659          * so, we have to request that MDS to do it */
2660
2661         /* prepare source fid */
2662         if (de_old->d_flags & DCACHE_CROSS_REF) {
2663                 LASSERT(de_old->d_inode == NULL);
2664                 CDEBUG(D_OTHER, "request to move remote name\n");
2665                 opdata.fid1.mds = de_old->d_mdsnum;
2666                 opdata.fid1.id = de_old->d_inum;
2667                 opdata.fid1.generation = de_old->d_generation;
2668         } else if (de_old->d_inode == NULL) {
2669                 /* oh, source doesn't exist */
2670                 GOTO(cleanup, rc = -ENOENT);
2671         } else {
2672                 LASSERT(de_old->d_inode != NULL);
2673                 CDEBUG(D_OTHER, "request to move local name\n");
2674                 opdata.fid1.mds = mds->mds_num;
2675                 opdata.fid1.id = de_old->d_inode->i_ino;
2676                 opdata.fid1.generation = de_old->d_inode->i_generation;
2677         }
2678
2679         opdata.fid2 = *(rec->ur_fid2);
2680         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2681                        rec->ur_tgtlen - 1, &req2);
2682
2683         if (rc)
2684                 GOTO(cleanup, rc);
2685
2686         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2687         if (IS_ERR(handle))
2688                 GOTO(cleanup, rc = PTR_ERR(handle));
2689         rc = fsfilt_del_dir_entry(obd, de_old);
2690         d_drop(de_old);
2691
2692 cleanup:
2693         EXIT;
2694         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2695                                 handle, req, rc, 0);
2696         if (req2)
2697                 ptlrpc_req_finished(req2);
2698
2699 #ifdef S_PDIROPS
2700         if (parent_lockh[1].cookie != 0)
2701                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2702 #endif
2703         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2704         if (child_lockh.cookie != 0)
2705                 ldlm_lock_decref(&child_lockh, LCK_EX);
2706
2707         l_dput(de_old);
2708         l_dput(de_srcdir);
2709
2710         req->rq_status = rc;
2711         RETURN(0);
2712
2713 }
2714
2715 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2716                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2717 {
2718         struct obd_device *obd = req->rq_export->exp_obd;
2719         struct dentry *de_srcdir = NULL;
2720         struct dentry *de_tgtdir = NULL;
2721         struct dentry *de_old = NULL;
2722         struct dentry *de_new = NULL;
2723         struct mds_obd *mds = mds_req2mds(req);
2724         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2725         struct mds_body *body = NULL;
2726         struct llog_create_locks *lcl = NULL;
2727         int rc = 0, log_unlink = 0, lock_count = 3;
2728         int cleanup_phase = 0;
2729         void *handle = NULL;
2730         ENTRY;
2731
2732         LASSERT(offset == 0);
2733
2734         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2735                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2736                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2737
2738         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2739
2740         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2741                 DEBUG_REQ(D_HA, req, "rename replay\n");
2742                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2743                        lustre_msg_buf(req->rq_reqmsg, 3, 0),
2744                        req->rq_repmsg->buflens[2]);
2745         }
2746
2747         if (rec->ur_namelen == 1) {
2748                 rc = mds_reint_rename_create_name(rec, offset, req);
2749                 RETURN(rc);
2750         }
2751
2752         if (rec->ur_fid2->mds != mds->mds_num) {
2753                 rc = mds_reint_rename_to_remote(rec, offset, req);
2754                 RETURN(rc);
2755         }
2756
2757         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2758                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2759                                              rec->ur_name, rec->ur_namelen,
2760                                              &de_old, rec->ur_tgt,
2761                                              rec->ur_tgtlen, &de_new,
2762                                              dlm_handles, LCK_EX);
2763         if (rc)
2764                 GOTO(cleanup, rc);
2765
2766         cleanup_phase = 1; /* parent(s), children, locks */
2767
2768         if (de_new->d_inode)
2769                 lock_count = 4;
2770
2771         /* sanity check for src inode */
2772         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2773             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2774                 GOTO(cleanup, rc = -EINVAL);
2775
2776         /* sanity check for dest inode */
2777         if (de_new->d_inode &&
2778             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2779              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2780                 GOTO(cleanup, rc = -EINVAL);
2781
2782         if (de_old->d_inode == de_new->d_inode)
2783                 GOTO(cleanup, rc = 0);
2784
2785         /* if we are about to remove the target at first, pass the EA of
2786          * that inode to client to perform and cleanup on OST */
2787         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2788         LASSERT(body != NULL);
2789
2790         if (de_new->d_inode &&
2791             S_ISREG(de_new->d_inode->i_mode) &&
2792             de_new->d_inode->i_nlink == 1 &&
2793             mds_open_orphan_count(de_new->d_inode) == 0) {
2794                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2795                 mds_pack_inode2body(obd, body, de_new->d_inode);
2796                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2797                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2798                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2799                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2800                 } else {
2801                         log_unlink = 1;
2802                 }
2803         }
2804
2805         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2806                        de_srcdir->d_inode->i_sb);
2807
2808         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2809         if (IS_ERR(handle))
2810                 GOTO(cleanup, rc = PTR_ERR(handle));
2811
2812         /* FIXME need adjust the journal block count? */
2813         /* if the target should be moved to PENDING, we at first increase the
2814          * link and later vfs_rename() will decrease the link count again */
2815         if (de_new->d_inode &&
2816             S_ISREG(de_new->d_inode->i_mode) &&
2817             de_new->d_inode->i_nlink == 1 &&
2818             mds_open_orphan_count(de_new->d_inode) > 0) {
2819                 rc = mds_add_link_orphan(rec, obd, de_new);
2820                 if (rc)
2821                         GOTO(cleanup, rc);
2822         }
2823
2824         lock_kernel();
2825         de_old->d_fsdata = req;
2826         de_new->d_fsdata = req;
2827         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2828         unlock_kernel();
2829
2830         if (!rc && log_unlink) {
2831                 if (mds_log_op_unlink(obd, de_tgtdir->d_inode,
2832                                 lustre_msg_buf(req->rq_repmsg, 1, 0),
2833                                 req->rq_repmsg->buflens[1],
2834                                 lustre_msg_buf(req->rq_repmsg, 2, 0),
2835                                 req->rq_repmsg->buflens[2], &lcl) > 0) {
2836                         body->valid |= OBD_MD_FLCOOKIE;
2837                 }
2838         }
2839
2840         GOTO(cleanup, rc);
2841 cleanup:
2842         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2843                                 handle, req, rc, 0);
2844         switch (cleanup_phase) {
2845         case 1:
2846 #ifdef S_PDIROPS
2847                 if (dlm_handles[5].cookie != 0)
2848                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2849                 if (dlm_handles[6].cookie != 0)
2850                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2851 #endif
2852                 if (lcl != NULL)
2853                         ptlrpc_save_llog_lock(req, lcl);
2854
2855                 if (rc) {
2856                         if (lock_count == 4)
2857                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2858                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2859                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2860                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2861                 } else {
2862                         if (lock_count == 4)
2863                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2864                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2865                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2866                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2867                 }
2868                 l_dput(de_new);
2869                 l_dput(de_old);
2870                 l_dput(de_tgtdir);
2871                 l_dput(de_srcdir);
2872         case 0:
2873                 break;
2874         default:
2875                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2876                 LBUG();
2877         }
2878         req->rq_status = rc;
2879         return 0;
2880 }
2881
2882 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2883                            struct ptlrpc_request *, struct lustre_handle *);
2884
2885 static mds_reinter reinters[REINT_MAX + 1] = {
2886         [REINT_SETATTR] mds_reint_setattr,
2887         [REINT_CREATE] mds_reint_create,
2888         [REINT_LINK] mds_reint_link,
2889         [REINT_UNLINK] mds_reint_unlink,
2890         [REINT_RENAME] mds_reint_rename,
2891         [REINT_OPEN] mds_open
2892 };
2893
2894 int mds_reint_rec(struct mds_update_record *rec, int offset,
2895                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2896 {
2897         struct obd_device *obd = req->rq_export->exp_obd;
2898         struct lvfs_run_ctxt saved;
2899         int rc;
2900
2901         /* checked by unpacker */
2902         LASSERT(rec->ur_opcode <= REINT_MAX &&
2903                 reinters[rec->ur_opcode] != NULL);
2904
2905         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2906         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2907         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2908
2909         return rc;
2910 }