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                 LASSERT(transno == 0);
 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         unsigned long              flags;
 285         char                       str[PTL_NALFMT_SIZE];
 286         int                        i;
 287
 288         /* CAVEAT EMPTOR: spinlock order */
 289         spin_lock_irqsave (&exp->exp_lock, flags);
 290         list_for_each (tmp, &exp->exp_outstanding_replies) {
 291                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 292
 293                 if (oldrep->rs_xid != req->rq_xid)
 294                         continue;
 295
 296                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 297                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 298                                 "new %d old %d\n", req->rq_xid,
 299                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 300
 301                 svc = oldrep->rs_srv_ni->sni_service;
 302                 spin_lock (&svc->srv_lock);
 303
 304                 list_del_init (&oldrep->rs_exp_list);
 305
 306                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 307                       " o%d NID %s\n",
 308                       oldrep->rs_nlocks, oldrep,
 309                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 310                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 311
 312                 for (i = 0; i < oldrep->rs_nlocks; i++)
 313                         ptlrpc_save_lock(req,
 314                                          &oldrep->rs_locks[i],
 315                                          oldrep->rs_modes[i]);
 316                 oldrep->rs_nlocks = 0;
 317
 318                 DEBUG_REQ(D_HA, req, "stole locks for");
 319                 ptlrpc_schedule_difficult_reply (oldrep);
 320
 321                 spin_unlock (&svc->srv_lock);
 322                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 323                 return;
 324         }
 325         spin_unlock_irqrestore (&exp->exp_lock, flags);
 326 }
 327
 328 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 329 {
 330         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 331                   mcd->mcd_last_transno, mcd->mcd_last_result);
 332         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 333         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 334
 335         mds_steal_ack_locks(req);
 336 }
 337
 338 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 339                                       int offset, struct ptlrpc_request *req)
 340 {
 341         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 342         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 343         struct dentry *de;
 344         struct mds_body *body;
 345
 346         mds_req_from_mcd(req, med->med_mcd);
 347
 348         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 349         if (IS_ERR(de)) {
 350                 LASSERT(PTR_ERR(de) == req->rq_status);
 351                 return;
 352         }
 353
 354         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 355         mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
 356         mds_pack_inode2body(req2obd(req), body, de->d_inode);
 357
 358         /* Don't return OST-specific attributes if we didn't just set them */
 359         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 360                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 361         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 362                 body->valid |= OBD_MD_FLMTIME;
 363         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 364                 body->valid |= OBD_MD_FLATIME;
 365
 366         l_dput(de);
 367 }
 368
 369 /* In the raw-setattr case, we lock the child inode.
 370  * In the write-back case or if being called from open, the client holds a lock
 371  * already.
 372  *
 373  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 374 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 375                              struct ptlrpc_request *req,
 376                              struct lustre_handle *lh)
 377 {
 378         struct mds_obd *mds = mds_req2mds(req);
 379         struct obd_device *obd = req->rq_export->exp_obd;
 380         struct mds_body *body;
 381         struct dentry *de;
 382         struct inode *inode = NULL;
 383         struct lustre_handle lockh[2] = {{0}, {0}};
 384         int parent_mode;
 385         void *handle = NULL;
 386         struct mds_logcancel_data *mlcd = NULL;
 387         int rc = 0, cleanup_phase = 0, err, locked = 0;
 388         ENTRY;
 389
 390         LASSERT(offset == 0);
 391
 392         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 393                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 394
 395         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 396
 397         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 398                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 399                 if (IS_ERR(de))
 400                         GOTO(cleanup, rc = PTR_ERR(de));
 401         } else {
 402                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 403                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
 404                         lockpart |= MDS_INODELOCK_LOOKUP;
 405                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 406                                            lockh, &parent_mode, NULL, 0, lockpart);
 407                 if (IS_ERR(de))
 408                         GOTO(cleanup, rc = PTR_ERR(de));
 409                 locked = 1;
 410         }
 411
 412         cleanup_phase = 1;
 413
 414         inode = de->d_inode;
 415         LASSERT(inode);
 416         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 417             rec->ur_eadata != NULL)
 418                 down(&inode->i_sem);
 419
 420         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 421
 422         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 423         if (IS_ERR(handle))
 424                 GOTO(cleanup, rc = PTR_ERR(handle));
 425
 426         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 427                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 428                        LTIME_S(rec->ur_iattr.ia_mtime),
 429                        LTIME_S(rec->ur_iattr.ia_ctime));
 430         rc = mds_fix_attr(inode, rec);
 431         if (rc)
 432                 GOTO(cleanup, rc);
 433
 434         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 435                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 436                                       (long)&rec->ur_iattr.ia_attr_flags);
 437         else                                            /* setattr */
 438                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 439
 440         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 441             rec->ur_eadata != NULL) {
 442                 struct lov_stripe_md *lsm = NULL;
 443
 444                 rc = ll_permission(inode, MAY_WRITE, NULL);
 445                 if (rc < 0)
 446                         GOTO(cleanup, rc);
 447
 448                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 449                                    mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
 450                 if (rc)
 451                         GOTO(cleanup, rc);
 452
 453                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 454
 455                 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 456                                    rec->ur_eadatalen);
 457                 if (rc)
 458                         GOTO(cleanup, rc);
 459         }
 460
 461         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 462         mds_pack_inode2fid(obd, &body->fid1, inode);
 463         mds_pack_inode2body(obd, body, inode);
 464
 465         /* Don't return OST-specific attributes if we didn't just set them */
 466         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 467                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 468         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 469                 body->valid |= OBD_MD_FLMTIME;
 470         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 471                 body->valid |= OBD_MD_FLATIME;
 472
 473         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 474                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 475                           rec->ur_eadatalen);
 476                 if (mlcd) {
 477                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 478                                 rec->ur_eadatalen;
 479                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 480                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 481                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 482                                 mlcd->mlcd_cookielen;
 483                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 484                                mlcd->mlcd_cookielen);
 485                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 486                                mlcd->mlcd_eadatalen);
 487                 } else {
 488                         CERROR("unable to allocate log cancel data\n");
 489                 }
 490         }
 491         EXIT;
 492  cleanup:
 493         if (mlcd != NULL)
 494                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 495                                       handle, mds_cancel_cookies_cb, mlcd);
 496         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 497         switch (cleanup_phase) {
 498         case 1:
 499                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 500                     rec->ur_eadata != NULL)
 501                         up(&inode->i_sem);
 502                 l_dput(de);
 503                 if (locked) {
 504 #ifdef S_PDIROPS
 505                         if (lockh[1].cookie != 0)
 506                                 ldlm_lock_decref(lockh + 1, parent_mode);
 507 #endif
 508                         if (rc) {
 509                                 ldlm_lock_decref(lockh, LCK_PW);
 510                         } else {
 511                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 512                         }
 513                 }
 514         case 0:
 515                 break;
 516         default:
 517                 LBUG();
 518         }
 519         if (err && !rc)
 520                 rc = err;
 521
 522         req->rq_status = rc;
 523         return 0;
 524 }
 525
 526 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 527                                      struct ptlrpc_request *req)
 528 {
 529         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 530         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 531         struct dentry *parent, *child;
 532         struct mds_body *body;
 533
 534         mds_req_from_mcd(req, med->med_mcd);
 535
 536         if (req->rq_status)
 537                 return;
 538
 539         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 540         LASSERT(!IS_ERR(parent));
 541         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 542         LASSERT(!IS_ERR(child));
 543         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 544         mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 545         mds_pack_inode2body(req2obd(req), body, child->d_inode);
 546         l_dput(parent);
 547         l_dput(child);
 548 }
 549
 550 static int mds_reint_create(struct mds_update_record *rec, int offset,
 551                             struct ptlrpc_request *req,
 552                             struct lustre_handle *lh)
 553 {
 554         struct dentry *dparent = NULL;
 555         struct mds_obd *mds = mds_req2mds(req);
 556         struct obd_device *obd = req->rq_export->exp_obd;
 557         struct dentry *dchild = NULL;
 558         struct inode *dir = NULL;
 559         void *handle = NULL;
 560         struct lustre_handle lockh[2] = {{0}, {0}};
 561         int parent_mode;
 562         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 563         int created = 0;
 564         struct dentry_params dp;
 565         struct mea *mea = NULL;
 566         int mea_size;
 567         ENTRY;
 568
 569         LASSERT(offset == 0);
 570         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 571
 572         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 573                   rec->ur_fid1->id, rec->ur_fid1->generation,
 574                   rec->ur_name, rec->ur_mode);
 575
 576         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 577
 578         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 579                 GOTO(cleanup, rc = -ESTALE);
 580
 581         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 582                                         &parent_mode, rec->ur_name,
 583                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 584         if (IS_ERR(dparent)) {
 585                 rc = PTR_ERR(dparent);
 586                 CERROR("parent lookup error %d\n", rc);
 587                 GOTO(cleanup, rc);
 588         }
 589         cleanup_phase = 1; /* locked parent dentry */
 590         dir = dparent->d_inode;
 591         LASSERT(dir);
 592
 593         ldlm_lock_dump_handle(D_OTHER, lockh);
 594
 595         /* try to retrieve MEA data for this dir */
 596         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 597         if (rc)
 598                 GOTO(cleanup, rc);
 599
 600         if (mea != NULL) {
 601                 /* dir is already splitted, check is requested filename
 602                  * should live at this MDS or at another one */
 603                 int i;
 604                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 605                 if (mea->mea_master != mea->mea_fids[i].mds) {
 606                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 607                                " should be %d(%d)\n",
 608                                mea->mea_master, dparent->d_inode->i_ino,
 609                                dparent->d_inode->i_generation, rec->ur_name,
 610                                mea->mea_fids[i].mds, i);
 611                         GOTO(cleanup, rc = -ERESTART);
 612                 }
 613         }
 614
 615         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 616         if (IS_ERR(dchild)) {
 617                 rc = PTR_ERR(dchild);
 618                 CERROR("child lookup error %d\n", rc);
 619                 GOTO(cleanup, rc);
 620         }
 621
 622         cleanup_phase = 2; /* child dentry */
 623
 624         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 625
 626         if (type == S_IFREG || type == S_IFDIR) {
 627                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 628                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 629                        obd->obd_name, dparent->d_inode->i_ino,
 630                        dparent->d_inode->i_generation, rc, parent_mode);
 631                 if (rc > 0) {
 632                         /* dir got splitted */
 633                         GOTO(cleanup, rc = -ERESTART);
 634                 } else if (rc < 0) {
 635                         /* error happened during spitting. */
 636                         GOTO(cleanup, rc);
 637                 }
 638         }
 639
 640         if (dir->i_mode & S_ISGID) {
 641                 if (S_ISDIR(rec->ur_mode))
 642                         rec->ur_mode |= S_ISGID;
 643         }
 644
 645         dchild->d_fsdata = (void *)&dp;
 646         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 647         dp.p_ptr = req;
 648
 649         switch (type) {
 650         case S_IFREG:{
 651                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 652                 if (IS_ERR(handle))
 653                         GOTO(cleanup, rc = PTR_ERR(handle));
 654                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 655                 EXIT;
 656                 break;
 657         }
 658         case S_IFDIR:{
 659                 int nstripes = 0;
 660                 int i;
 661
 662                 /* as Peter asked, mkdir() should distribute new directories
 663                  * over the whole cluster in order to distribute namespace
 664                  * processing load. first, we calculate which MDS to use to put
 665                  * new directory's inode in. */
 666                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 667                                       rec->ur_flags);
 668                 if (i == mds->mds_num) {
 669                         /* inode will be created locally */
 670                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 671                         if (IS_ERR(handle))
 672                                 GOTO(cleanup, rc = PTR_ERR(handle));
 673
 674                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 675
 676                         if (rec->ur_eadata)
 677                                 nstripes = *(u16 *)rec->ur_eadata;
 678
 679                         if (rc == 0 && nstripes) {
 680                                 /* we pass LCK_EX to split routine to
 681                                  * signalthat we have exclusive access
 682                                  * to the directory. simple because
 683                                  * nobody knows it already exists -bzzz */
 684                                 rc = mds_try_to_split_dir(obd, dchild,
 685                                                           NULL, nstripes,
 686                                                           LCK_EX);
 687                                 if (rc > 0) {
 688                                         /* dir got splitted */
 689                                         rc = 0;
 690                                 } else if (rc < 0) {
 691                                         /* an error occured during
 692                                          * splitting. */
 693                                         GOTO(cleanup, rc);
 694                                 }
 695                         }
 696                 } else if (!DENTRY_VALID(dchild)) {
 697                         /* inode will be created on another MDS */
 698                         struct obdo *oa = NULL;
 699                         struct mds_body *body;
 700
 701                         /* first, create that inode */
 702                         oa = obdo_alloc();
 703                         LASSERT(oa != NULL);
 704                         oa->o_mds = i;
 705                         oa->o_easize = 0;
 706                         if (rec->ur_eadata) {
 707                                 /* user asks for creating splitted dir */
 708                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 709                         }
 710
 711                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 712                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 713                                         OBD_MD_FLUID | OBD_MD_FLGID);
 714                         oa->o_mode = dir->i_mode;
 715                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 716                                         obd->obd_name, i);
 717                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 718                                 CWARN("%s: replay dir creation %*s -> %u/%u\n",
 719                                       obd->obd_name, rec->ur_namelen - 1,
 720                                       rec->ur_name, (unsigned) rec->ur_fid2->id,
 721                                       (unsigned) rec->ur_fid2->generation);
 722                                 oa->o_id = rec->ur_fid2->id;
 723                                 oa->o_generation = rec->ur_fid2->generation;
 724                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 725                         }
 726
 727                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 728                         if (rc) {
 729                                 obdo_free(oa);
 730                                 GOTO(cleanup, rc);
 731                         }
 732
 733                         /* now, add new dir entry for it */
 734                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 735                         if (IS_ERR(handle)) {
 736                                 obdo_free(oa);
 737                                 GOTO(cleanup, rc = PTR_ERR(handle));
 738                         }
 739                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 740                                                   rec->ur_namelen - 1,
 741                                                   oa->o_id, oa->o_generation,
 742                                                   i);
 743                         LASSERT(rc == 0);
 744
 745                         /* fill reply */
 746                         body = lustre_msg_buf(req->rq_repmsg,
 747                                               offset, sizeof (*body));
 748                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 749                         body->fid1.id = oa->o_id;
 750                         body->fid1.mds = i;
 751                         body->fid1.generation = oa->o_generation;
 752                         obdo_free(oa);
 753                 } else {
 754                         /* requested name exists in the directory */
 755                         rc = -EEXIST;
 756                 }
 757                 EXIT;
 758                 break;
 759         }
 760         case S_IFLNK:{
 761                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 762                 if (IS_ERR(handle))
 763                         GOTO(cleanup, rc = PTR_ERR(handle));
 764                 if (rec->ur_tgt == NULL)        /* no target supplied */
 765                         rc = -EINVAL;           /* -EPROTO? */
 766                 else
 767                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 768                 EXIT;
 769                 break;
 770         }
 771         case S_IFCHR:
 772         case S_IFBLK:
 773         case S_IFIFO:
 774         case S_IFSOCK:{
 775                 int rdev = rec->ur_rdev;
 776                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 777                 if (IS_ERR(handle))
 778                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 779                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 780                 EXIT;
 781                 break;
 782         }
 783         default:
 784                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 785                 dchild->d_fsdata = NULL;
 786                 GOTO(cleanup, rc = -EINVAL);
 787         }
 788
 789         /* In case we stored the desired inum in here, we want to clean up. */
 790         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 791                 dchild->d_fsdata = NULL;
 792
 793         if (rc) {
 794                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 795                 GOTO(cleanup, rc);
 796         } else if (dchild->d_inode) {
 797                 struct iattr iattr;
 798                 struct inode *inode = dchild->d_inode;
 799                 struct mds_body *body;
 800
 801                 created = 1;
 802                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 803                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 804                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 805                 iattr.ia_uid = rec->_ur_fsuid;
 806                 if (dir->i_mode & S_ISGID)
 807                         iattr.ia_gid = dir->i_gid;
 808                 else
 809                         iattr.ia_gid = rec->_ur_fsgid;
 810                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 811                         ATTR_MTIME | ATTR_CTIME;
 812
 813                 if (rec->ur_fid2->id) {
 814                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 815                         inode->i_generation = rec->ur_fid2->generation;
 816                         /* Dirtied and committed by the upcoming setattr. */
 817                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 818                                inode->i_ino, inode->i_generation);
 819                 } else {
 820                         struct lustre_handle child_ino_lockh;
 821
 822                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 823                                inode->i_ino, inode->i_generation);
 824
 825                         /* The inode we were allocated may have just been freed
 826                          * by an unlink operation.  We take this lock to
 827                          * synchronize against the matching reply-ack-lock taken
 828                          * in unlink, to avoid replay problems if this reply
 829                          * makes it out to the client but the unlink's does not.
 830                          * See bug 2029 for more detail.*/
 831                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 832                         if (rc != ELDLM_OK) {
 833                                 CERROR("error locking for unlink/create sync: "
 834                                        "%d\n", rc);
 835                         } else {
 836                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 837                         }
 838                 }
 839
 840                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 841                 if (rc)
 842                         CERROR("error on child setattr: rc = %d\n", rc);
 843
 844                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 845                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 846                 if (rc)
 847                         CERROR("error on parent setattr: rc = %d\n", rc);
 848
 849                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 850                 mds_pack_inode2fid(obd, &body->fid1, inode);
 851                 mds_pack_inode2body(obd, body, inode);
 852         }
 853         EXIT;
 854
 855 cleanup:
 856         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 857
 858         if (rc && created) {
 859                 /* Destroy the file we just created. This should not need extra
 860                  * journal credits, as we have already modified all of the
 861                  * blocks needed in order to create the file in the first
 862                  * place. */
 863                 switch (type) {
 864                 case S_IFDIR:
 865                         err = vfs_rmdir(dir, dchild);
 866                         if (err)
 867                                 CERROR("rmdir in error path: %d\n", err);
 868                         break;
 869                 default:
 870                         err = vfs_unlink(dir, dchild);
 871                         if (err)
 872                                 CERROR("unlink in error path: %d\n", err);
 873                         break;
 874                 }
 875         } else {
 876                 rc = err;
 877         }
 878         switch (cleanup_phase) {
 879         case 2: /* child dentry */
 880                 l_dput(dchild);
 881         case 1: /* locked parent dentry */
 882 #ifdef S_PDIROPS
 883                 if (lockh[1].cookie != 0)
 884                         ldlm_lock_decref(lockh + 1, parent_mode);
 885 #endif
 886                 if (rc) {
 887                         ldlm_lock_decref(lockh, LCK_PW);
 888                 } else {
 889                         ptlrpc_save_lock(req, lockh, LCK_PW);
 890                 }
 891                 l_dput(dparent);
 892         case 0:
 893                 break;
 894         default:
 895                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 896                 LBUG();
 897         }
 898         if (mea)
 899                 OBD_FREE(mea, mea_size);
 900         req->rq_status = rc;
 901         return 0;
 902 }
 903
 904 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 905            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 906 {
 907         int i;
 908
 909         for (i = 0; i < RES_NAME_SIZE; i++) {
 910                 /* return 1 here, because enqueue_ordered will skip resources
 911                  * of all zeroes if they're sorted to the end of the list. */
 912                 if (res1->name[i] == 0 && res2->name[i] != 0)
 913                         return 1;
 914                 if (res2->name[i] == 0 && res1->name[i] != 0)
 915                         return 0;
 916
 917                 if (res1->name[i] > res2->name[i])
 918                         return 1;
 919                 if (res1->name[i] < res2->name[i])
 920                         return 0;
 921         }
 922
 923         if (!p1 || !p2)
 924                 return 0;
 925
 926         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 927                 return 1;
 928
 929         return 0;
 930 }
 931
 932 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 933  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 934  * because they take the place of local semaphores.
 935  *
 936  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 937  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 938 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 939                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 940                           ldlm_policy_data_t *p1_policy,
 941                           struct ldlm_res_id *p2_res_id,
 942                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 943                           ldlm_policy_data_t *p2_policy)
 944 {
 945         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 946         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 947         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 948         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 949         int rc, flags;
 950         ENTRY;
 951
 952         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 953
 954         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 955                res_id[1]->name[0]);
 956
 957         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 958                 handles[1] = p1_lockh;
 959                 handles[0] = p2_lockh;
 960                 res_id[1] = p1_res_id;
 961                 res_id[0] = p2_res_id;
 962                 lock_modes[1] = p1_lock_mode;
 963                 lock_modes[0] = p2_lock_mode;
 964                 policies[1] = p1_policy;
 965                 policies[0] = p2_policy;
 966         }
 967
 968         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 969                res_id[0]->name[0], res_id[1]->name[0]);
 970
 971         flags = LDLM_FL_LOCAL_ONLY;
 972         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 973                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 974                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 975                               NULL, 0, NULL, handles[0]);
 976         if (rc != ELDLM_OK)
 977                 RETURN(-EIO);
 978         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 979
 980         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
 981             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
 982                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 983                 ldlm_lock_addref(handles[1], lock_modes[1]);
 984         } else if (res_id[1]->name[0] != 0) {
 985                 flags = LDLM_FL_LOCAL_ONLY;
 986                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 987                                       *res_id[1], LDLM_IBITS, policies[1],
 988                                       lock_modes[1], &flags, mds_blocking_ast,
 989                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 990                                       NULL, handles[1]);
 991                 if (rc != ELDLM_OK) {
 992                         ldlm_lock_decref(handles[0], lock_modes[0]);
 993                         RETURN(-EIO);
 994                 }
 995                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 996         }
 997
 998         RETURN(0);
 999 }
1000
1001 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1002                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1003                            ldlm_policy_data_t *p1_policy,
1004                            struct ldlm_res_id *p2_res_id,
1005                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1006                            ldlm_policy_data_t *p2_policy,
1007                            struct ldlm_res_id *c1_res_id,
1008                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1009                            ldlm_policy_data_t *c1_policy,
1010                            struct ldlm_res_id *c2_res_id,
1011                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1012                            ldlm_policy_data_t *c2_policy)
1013 {
1014         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1015                                           c1_res_id, c2_res_id };
1016         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1017                                                  c1_lockh, c2_lockh };
1018         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1019                               c1_lock_mode, c2_lock_mode };
1020         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1021                                             c1_policy, c2_policy};
1022         int rc, i, j, sorted, flags;
1023         ENTRY;
1024
1025         CDEBUG(D_DLMTRACE,
1026                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1027                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1028                res_id[3]->name[0]);
1029
1030         /* simple insertion sort - we have at most 4 elements */
1031         for (i = 1; i < 4; i++) {
1032                 j = i - 1;
1033                 dlm_handles[4] = dlm_handles[i];
1034                 res_id[4] = res_id[i];
1035                 lock_modes[4] = lock_modes[i];
1036                 policies[4] = policies[i];
1037
1038                 sorted = 0;
1039                 do {
1040                         if (res_gt(res_id[j], res_id[4], policies[j],
1041                                    policies[4])) {
1042                                 dlm_handles[j + 1] = dlm_handles[j];
1043                                 res_id[j + 1] = res_id[j];
1044                                 lock_modes[j + 1] = lock_modes[j];
1045                                 policies[j + 1] = policies[j];
1046                                 j--;
1047                         } else {
1048                                 sorted = 1;
1049                         }
1050                 } while (j >= 0 && !sorted);
1051
1052                 dlm_handles[j + 1] = dlm_handles[4];
1053                 res_id[j + 1] = res_id[4];
1054                 lock_modes[j + 1] = lock_modes[4];
1055                 policies[j + 1] = policies[4];
1056         }
1057
1058         CDEBUG(D_DLMTRACE,
1059                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1060                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1061                res_id[3]->name[0]);
1062
1063         /* XXX we could send ASTs on all these locks first before blocking? */
1064         for (i = 0; i < 4; i++) {
1065                 flags = 0;
1066                 if (res_id[i]->name[0] == 0)
1067                         break;
1068                 if (i != 0 &&
1069                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1070                     (policies[i]->l_inodebits.bits &
1071                      policies[i-1]->l_inodebits.bits) ) {
1072                         memcpy(dlm_handles[i], dlm_handles[i-1],
1073                                sizeof(*(dlm_handles[i])));
1074                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1075                 } else {
1076                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1077                                               *res_id[i], LDLM_IBITS,
1078                                               policies[i],
1079                                               lock_modes[i], &flags,
1080                                               mds_blocking_ast,
1081                                               ldlm_completion_ast, NULL, NULL,
1082                                               NULL, 0, NULL, dlm_handles[i]);
1083                         if (rc != ELDLM_OK)
1084                                 GOTO(out_err, rc = -EIO);
1085                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1086                 }
1087         }
1088
1089         RETURN(0);
1090 out_err:
1091         while (i-- > 0)
1092                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1093
1094         return rc;
1095 }
1096
1097 /* In the unlikely case that the child changed while we were waiting
1098  * on the lock, we need to drop the lock on the old child and either:
1099  * - if the child has a lower resource name, then we have to also
1100  *   drop the parent lock and regain the locks in the right order
1101  * - in the rename case, if the child has a lower resource name than one of
1102  *   the other parent/child resources (maxres) we also need to reget the locks
1103  * - if the child has a higher resource name (this is the common case)
1104  *   we can just get the lock on the new child (still in lock order)
1105  *
1106  * Returns 0 if the child did not change or if it changed but could be locked.
1107  * Returns 1 if the child changed and we need to re-lock (no locks held).
1108  * Returns -ve error with a valid dchild (no locks held). */
1109 static int mds_verify_child(struct obd_device *obd,
1110                             struct ldlm_res_id *parent_res_id,
1111                             struct lustre_handle *parent_lockh,
1112                             struct dentry *dparent, int parent_mode,
1113                             struct ldlm_res_id *child_res_id,
1114                             struct lustre_handle *child_lockh,
1115                             struct dentry **dchildp, int child_mode,
1116                             ldlm_policy_data_t *child_policy,
1117                             const char *name, int namelen,
1118                             struct ldlm_res_id *maxres)
1119 {
1120         struct dentry *vchild, *dchild = *dchildp;
1121         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1122         ENTRY;
1123
1124         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1125         if (IS_ERR(vchild))
1126                 GOTO(cleanup, rc = PTR_ERR(vchild));
1127
1128         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1129                 if  (child_res_id->name[0] == vchild->d_inum &&
1130                                 child_res_id->name[1] == vchild->d_generation) {
1131                         if (dchild != NULL)
1132                                 l_dput(dchild);
1133                         *dchildp = vchild;
1134                         RETURN(0);
1135                 }
1136                 goto changed;
1137         }
1138
1139         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1140                    (vchild->d_inode != NULL &&
1141                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1142                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1143                 if (dchild != NULL)
1144                         l_dput(dchild);
1145                 *dchildp = vchild;
1146
1147                 RETURN(0);
1148         }
1149
1150 changed:
1151         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1152                vchild->d_inode, dchild ? dchild->d_inode : 0,
1153                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1154                child_res_id->name[0]);
1155         if (child_res_id->name[0] != 0)
1156                 ldlm_lock_decref(child_lockh, child_mode);
1157         if (dchild)
1158                 l_dput(dchild);
1159
1160         cleanup_phase = 1; /* parent lock only */
1161         *dchildp = dchild = vchild;
1162
1163         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1164                 int flags = 0;
1165                 if (dchild->d_inode) {
1166                         child_res_id->name[0] = dchild->d_inode->i_ino;
1167                         child_res_id->name[1] = dchild->d_inode->i_generation;
1168                 } else {
1169                         child_res_id->name[0] = dchild->d_inum;
1170                         child_res_id->name[1] = dchild->d_generation;
1171                 }
1172
1173                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1174                     res_gt(maxres, child_res_id, NULL, NULL)) {
1175                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1176                                child_res_id->name[0], parent_res_id->name[0],
1177                                maxres->name[0]);
1178                         GOTO(cleanup, rc = 1);
1179                 }
1180
1181                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1182                                       *child_res_id, LDLM_IBITS, child_policy,
1183                                       child_mode, &flags, mds_blocking_ast,
1184                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1185                                       NULL, child_lockh);
1186                 if (rc != ELDLM_OK)
1187                         GOTO(cleanup, rc = -EIO);
1188
1189         } else {
1190                 memset(child_res_id, 0, sizeof(*child_res_id));
1191         }
1192
1193         EXIT;
1194 cleanup:
1195         if (rc) {
1196                 switch(cleanup_phase) {
1197                 case 2:
1198                         if (child_res_id->name[0] != 0)
1199                                 ldlm_lock_decref(child_lockh, child_mode);
1200                 case 1:
1201                         ldlm_lock_decref(parent_lockh, parent_mode);
1202                 }
1203         }
1204         return rc;
1205 }
1206
1207 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1208                                 struct ll_fid *fid,
1209                                 struct lustre_handle *parent_lockh,
1210                                 struct dentry **dparentp, int parent_mode,
1211                                 __u64 parent_lockpart, int *update_mode,
1212                                 char *name, int namelen,
1213                                 struct lustre_handle *child_lockh,
1214                                 struct dentry **dchildp, int child_mode,
1215                                 __u64 child_lockpart)
1216 {
1217         struct ldlm_res_id child_res_id = { .name = {0} };
1218         struct ldlm_res_id parent_res_id = { .name = {0} };
1219         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1220         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1221         struct inode *inode;
1222         int rc = 0, cleanup_phase = 0;
1223         ENTRY;
1224
1225         /* Step 1: Lookup parent */
1226         *dparentp = mds_fid2dentry(mds, fid, NULL);
1227         if (IS_ERR(*dparentp)) {
1228                 rc = PTR_ERR(*dparentp);
1229                 *dparentp = NULL;
1230                 RETURN(rc);
1231         }
1232
1233         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1234                (*dparentp)->d_inode->i_ino, name);
1235
1236         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1237         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1238 #ifdef S_PDIROPS
1239         parent_lockh[1].cookie = 0;
1240         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1241                 struct ldlm_res_id res_id = { .name = {0} };
1242                 ldlm_policy_data_t policy;
1243                 int flags = 0;
1244                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1245                 if (*update_mode) {
1246                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1247                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1248                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1249                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1250                                               res_id, LDLM_IBITS, &policy,
1251                                               *update_mode, &flags,
1252                                               mds_blocking_ast,
1253                                               ldlm_completion_ast,
1254                                               NULL, NULL, NULL, 0, NULL,
1255                                               parent_lockh + 1);
1256                         if (rc != ELDLM_OK)
1257                                 RETURN(-ENOLCK);
1258                 }
1259
1260                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1261                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1262                        (*dparentp)->d_inode->i_ino,
1263                        (*dparentp)->d_inode->i_generation,
1264                        parent_res_id.name[2]);
1265         }
1266 #endif
1267
1268         cleanup_phase = 1; /* parent dentry */
1269
1270         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1271         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1272         if (IS_ERR(*dchildp)) {
1273                 rc = PTR_ERR(*dchildp);
1274                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1275                 GOTO(cleanup, rc);
1276         }
1277
1278         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1279                 /* inode lives on another MDS: return * mds/ino/gen
1280                  * and LOOKUP lock. drop possible UPDATE lock! */
1281                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1282                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1283                 child_res_id.name[0] = (*dchildp)->d_inum;
1284                 child_res_id.name[1] = (*dchildp)->d_generation;
1285                 goto retry_locks;
1286         }
1287
1288         inode = (*dchildp)->d_inode;
1289         if (inode != NULL)
1290                 inode = igrab(inode);
1291         if (inode == NULL)
1292                 goto retry_locks;
1293
1294         child_res_id.name[0] = inode->i_ino;
1295         child_res_id.name[1] = inode->i_generation;
1296
1297         iput(inode);
1298
1299 retry_locks:
1300         cleanup_phase = 2; /* child dentry */
1301
1302         /* Step 3: Lock parent and child in resource order.  If child doesn't
1303          *         exist, we still have to lock the parent and re-lookup. */
1304         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1305                                    &parent_policy, &child_res_id, child_lockh,
1306                                    child_mode, &child_policy);
1307         if (rc)
1308                 GOTO(cleanup, rc);
1309
1310         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1311                 cleanup_phase = 4; /* child lock */
1312         else
1313                 cleanup_phase = 3; /* parent lock */
1314
1315         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1316         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1317                               parent_mode, &child_res_id, child_lockh,
1318                               dchildp, child_mode, &child_policy,
1319                               name, namelen, &parent_res_id);
1320         if (rc > 0)
1321                 goto retry_locks;
1322         if (rc < 0) {
1323                 cleanup_phase = 3;
1324                 GOTO(cleanup, rc);
1325         }
1326
1327 cleanup:
1328         if (rc) {
1329                 switch (cleanup_phase) {
1330                 case 4:
1331                         ldlm_lock_decref(child_lockh, child_mode);
1332                 case 3:
1333                         ldlm_lock_decref(parent_lockh, parent_mode);
1334                 case 2:
1335                         l_dput(*dchildp);
1336                 case 1:
1337 #ifdef S_PDIROPS
1338                         if (parent_lockh[1].cookie)
1339                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1340 #endif
1341                         l_dput(*dparentp);
1342                 default: ;
1343                 }
1344         }
1345         return rc;
1346 }
1347
1348 void mds_reconstruct_generic(struct ptlrpc_request *req)
1349 {
1350         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1351
1352         mds_req_from_mcd(req, med->med_mcd);
1353 }
1354
1355 int mds_create_local_dentry(struct mds_update_record *rec,
1356                            struct obd_device *obd)
1357 {
1358         struct mds_obd *mds = &obd->u.mds;
1359         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1360         int fidlen = 0, rc, cleanup_phase = 0;
1361         struct dentry *new_child = NULL;
1362         char *fidname = rec->ur_name;
1363         struct dentry *child = NULL;
1364         struct lustre_handle lockh[2] = {{0}, {0}};
1365         void *handle;
1366         ENTRY;
1367
1368         down(&fids_dir->i_sem);
1369         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1370         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1371                         fidname, (unsigned) rec->ur_fid1->id,
1372                         (unsigned) rec->ur_fid1->generation);
1373
1374         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1375         up(&fids_dir->i_sem);
1376         if (IS_ERR(new_child)) {
1377                 CERROR("can't lookup %s: %d\n", fidname,
1378                                 (int) PTR_ERR(new_child));
1379                 GOTO(cleanup, rc = PTR_ERR(new_child));
1380         }
1381         cleanup_phase = 1;
1382
1383         if (new_child->d_inode != NULL) {
1384                 /* nice. we've already have local dentry! */
1385                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1386                        (unsigned) new_child->d_inode->i_ino,
1387                        (unsigned) new_child->d_inode->i_generation);
1388                 rec->ur_fid1->id = fids_dir->i_ino;
1389                 rec->ur_fid1->generation = fids_dir->i_generation;
1390                 rec->ur_namelen = fidlen + 1;
1391                 GOTO(cleanup, rc = 0);
1392         }
1393
1394         /* new, local dentry will be added soon. we need no aliases here */
1395         d_drop(new_child);
1396
1397         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1398                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1399         } else {
1400                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1401                                               LCK_EX, lockh, NULL, NULL, 0,
1402                                               MDS_INODELOCK_UPDATE);
1403         }
1404
1405         if (IS_ERR(child)) {
1406                 CERROR("can't get victim: %d\n", (int) PTR_ERR(child));
1407                 GOTO(cleanup, rc = PTR_ERR(child));
1408         }
1409         cleanup_phase = 2;
1410
1411         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1412         if (IS_ERR(handle))
1413                 GOTO(cleanup, rc = PTR_ERR(handle));
1414
1415         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1416                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1417                                   mds->mds_num);
1418         if (rc)
1419                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1420                        (unsigned long) child->d_inode->i_ino,
1421                        (unsigned long) child->d_inode->i_generation, rc);
1422         else {
1423                 if (S_ISDIR(child->d_inode->i_mode)) {
1424                         fids_dir->i_nlink++;
1425                         mark_inode_dirty(fids_dir);
1426                 }
1427                 mark_inode_dirty(child->d_inode);
1428         }
1429         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1430
1431         rec->ur_fid1->id = fids_dir->i_ino;
1432         rec->ur_fid1->generation = fids_dir->i_generation;
1433         rec->ur_namelen = fidlen + 1;
1434
1435 cleanup:
1436         switch(cleanup_phase) {
1437                 case 2:
1438                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1439                                 ldlm_lock_decref(lockh, LCK_EX);
1440                         dput(child);
1441                 case 1:
1442                         dput(new_child);
1443                 case 0:
1444                        break;
1445         }
1446         RETURN(rc);
1447 }
1448
1449 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1450                                         struct ptlrpc_request *slave)
1451 {
1452         void *cookie, *cookie2;
1453         struct mds_body *body2;
1454         struct mds_body *body;
1455         void *ea, *ea2;
1456         ENTRY;
1457
1458         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1459         LASSERT(body != NULL);
1460
1461         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1462         LASSERT(body2 != NULL);
1463
1464         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1465                 RETURN(0);
1466         }
1467
1468         memcpy(body2, body, sizeof(*body));
1469         body2->valid &= ~OBD_MD_FLCOOKIE;
1470
1471         if (!(body->valid & OBD_MD_FLEASIZE) &&
1472             !(body->valid & OBD_MD_FLDIREA))
1473                 RETURN(0);
1474
1475         if (body->eadatasize == 0) {
1476                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1477                 RETURN(0);
1478         }
1479
1480         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1481
1482         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1483         LASSERT(ea != NULL);
1484
1485         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1486         LASSERT(ea2 != NULL);
1487
1488         memcpy(ea2, ea, body->eadatasize);
1489
1490         if (body->valid & OBD_MD_FLCOOKIE) {
1491                 LASSERT(master->rq_repmsg->buflens[2] >=
1492                                 slave->rq_repmsg->buflens[2]);
1493                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1494                                 slave->rq_repmsg->buflens[2]);
1495                 LASSERT(cookie != NULL);
1496
1497                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1498                                 master->rq_repmsg->buflens[2]);
1499                 LASSERT(cookie2 != NULL);
1500                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1501                 body2->valid |= OBD_MD_FLCOOKIE;
1502         }
1503         RETURN(0);
1504 }
1505
1506 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1507                                    struct ptlrpc_request *req,
1508                                    struct lustre_handle *parent_lockh,
1509                                    int update_mode,
1510                                    struct dentry *dparent,
1511                                    struct lustre_handle *child_lockh,
1512                                    struct dentry *dchild)
1513 {
1514         struct obd_device *obd = req->rq_export->exp_obd;
1515         struct mds_obd *mds = mds_req2mds(req);
1516         struct mdc_op_data op_data;
1517         int rc = 0, cleanup_phase = 0;
1518         struct ptlrpc_request *request = NULL;
1519         void *handle;
1520         ENTRY;
1521
1522         LASSERT(offset == 0 || offset == 2);
1523
1524         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1525                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1526                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1527         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1528                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1529                           rec->ur_namelen - 1, rec->ur_name,
1530                           (unsigned)dchild->d_mdsnum,
1531                           (unsigned) dchild->d_inum,
1532                           (unsigned) dchild->d_generation);
1533
1534         /* time to drop i_nlink on remote MDS */
1535         memset(&op_data, 0, sizeof(op_data));
1536         op_data.fid1.mds = dchild->d_mdsnum;
1537         op_data.fid1.id = dchild->d_inum;
1538         op_data.fid1.generation = dchild->d_generation;
1539         op_data.create_mode = rec->ur_mode;
1540         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1541                 op_data.create_mode |= MDS_MODE_REPLAY;
1542         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1543         cleanup_phase = 2;
1544         if (request) {
1545                 mds_copy_unlink_reply(req, request);
1546                 ptlrpc_req_finished(request);
1547         }
1548         if (rc == 0) {
1549                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1550                                       NULL);
1551                 if (IS_ERR(handle))
1552                         GOTO(cleanup, rc = PTR_ERR(handle));
1553                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1554                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1555                                         rc, 0);
1556         }
1557 cleanup:
1558         req->rq_status = rc;
1559
1560 #ifdef S_PDIROPS
1561         if (parent_lockh[1].cookie != 0)
1562                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1563 #endif
1564         ldlm_lock_decref(child_lockh, LCK_EX);
1565         if (rc)
1566                 ldlm_lock_decref(parent_lockh, LCK_PW);
1567         else
1568                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1569         l_dput(dchild);
1570         l_dput(dparent);
1571
1572         return 0;
1573 }
1574
1575 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1576                             struct ptlrpc_request *req,
1577                             struct lustre_handle *lh)
1578 {
1579         struct dentry *dparent, *dchild;
1580         struct mds_obd *mds = mds_req2mds(req);
1581         struct obd_device *obd = req->rq_export->exp_obd;
1582         struct mds_body *body = NULL;
1583         struct inode *child_inode;
1584         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1585         struct lustre_handle child_lockh = {0}, child_reuse_lockh = {0};
1586         struct lustre_handle * slave_lockh = NULL;
1587         char fidname[LL_FID_NAMELEN];
1588         void *handle = NULL;
1589         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1590         int unlink_by_fid = 0;
1591         int update_mode;
1592         ENTRY;
1593
1594         LASSERT(offset == 0 || offset == 2);
1595
1596         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1597                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1598
1599         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1600
1601         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1602                 GOTO(cleanup, rc = -ENOENT);
1603
1604         if (rec->ur_namelen == 1) {
1605                 /* this is request to drop i_nlink on local inode */
1606                 unlink_by_fid = 1;
1607                 rec->ur_name = fidname;
1608                 rc = mds_create_local_dentry(rec, obd);
1609                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1610                         DEBUG_REQ(D_HA, req,
1611                                   "drop nlink on inode %u/%u/%u (replay)",
1612                                   (unsigned) rec->ur_fid1->mds,
1613                                   (unsigned) rec->ur_fid1->id,
1614                                   (unsigned) rec->ur_fid1->generation);
1615                         req->rq_status = 0;
1616                         RETURN(0);
1617                 }
1618         }
1619
1620         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1621                 /* master mds for directory asks slave removing
1622                  * inode is already locked */
1623                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1624                                                LCK_PW, parent_lockh,
1625                                                &update_mode, rec->ur_name,
1626                                                rec->ur_namelen,
1627                                                MDS_INODELOCK_UPDATE);
1628                 if (IS_ERR(dparent))
1629                         GOTO(cleanup, rc = PTR_ERR(dparent));
1630                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1631                                            rec->ur_namelen - 1);
1632                 if (IS_ERR(dchild))
1633                         GOTO(cleanup, rc = PTR_ERR(dchild));
1634                 child_lockh.cookie = 0;
1635                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1636                 LASSERT(dchild->d_inode != NULL);
1637                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1638         } else {
1639                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1640                                                  parent_lockh, &dparent,
1641                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1642                                                  &update_mode, rec->ur_name,
1643                                                  rec->ur_namelen, &child_lockh,
1644                                                  &dchild, LCK_EX,
1645                                                  MDS_INODELOCK_LOOKUP |
1646                                                         MDS_INODELOCK_UPDATE);
1647         }
1648         if (rc)
1649                 GOTO(cleanup, rc);
1650
1651         if (dchild->d_flags & DCACHE_CROSS_REF) {
1652                 /* we should have parent lock only here */
1653                 LASSERT(unlink_by_fid == 0);
1654                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1655                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1656                                         update_mode, dparent, &child_lockh, dchild);
1657                 RETURN(0);
1658         }
1659
1660         cleanup_phase = 1; /* dchild, dparent, locks */
1661
1662         dget(dchild);
1663         child_inode = dchild->d_inode;
1664         if (child_inode == NULL) {
1665                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1666                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1667                 GOTO(cleanup, rc = -ENOENT);
1668         }
1669
1670         cleanup_phase = 2; /* dchild has a lock */
1671
1672         /* We have to do these checks ourselves, in case we are making an
1673          * orphan.  The client tells us whether rmdir() or unlink() was called,
1674          * so we need to return appropriate errors (bug 72).
1675          *
1676          * We don't have to check permissions, because vfs_rename (called from
1677          * mds_open_unlink_rename) also calls may_delete. */
1678         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1679                 if (!S_ISDIR(child_inode->i_mode))
1680                         GOTO(cleanup, rc = -ENOTDIR);
1681         } else {
1682                 if (S_ISDIR(child_inode->i_mode))
1683                         GOTO(cleanup, rc = -EISDIR);
1684         }
1685
1686         /* handle splitted dir */
1687         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1688         if (rc)
1689                 GOTO(cleanup, rc);
1690
1691         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1692          * reuse (see bug 2029). */
1693         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1694         if (rc != ELDLM_OK)
1695                 GOTO(cleanup, rc);
1696
1697         cleanup_phase = 3; /* child inum lock */
1698
1699         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1700
1701         /* ldlm_reply in buf[0] if called via intent */
1702         if (offset)
1703                 offset = 1;
1704
1705         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1706         LASSERT(body != NULL);
1707
1708         /* If this is the last reference to this inode, get the OBD EA
1709          * data first so the client can destroy OST objects.
1710          * we only do the object removal if no open files remain.
1711          * Nobody can get at this name anymore because of the locks so
1712          * we make decisions here as to whether to remove the inode */
1713         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1714             mds_open_orphan_count(child_inode) == 0) {
1715                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1716                 mds_pack_inode2body(obd, body, child_inode);
1717                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1718                             child_inode, 1);
1719                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1720                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1721                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1722                 } else {
1723                         log_unlink = 1;
1724                 }
1725         }
1726
1727         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1728         switch (child_inode->i_mode & S_IFMT) {
1729         case S_IFDIR:
1730                 /* Drop any lingering child directories before we start our
1731                  * transaction, to avoid doing multiple inode dirty/delete
1732                  * in our compound transaction (bug 1321). */
1733                 shrink_dcache_parent(dchild);
1734                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1735                                       NULL);
1736                 if (IS_ERR(handle))
1737                         GOTO(cleanup, rc = PTR_ERR(handle));
1738                 cleanup_phase = 4; /* transaction */
1739                 rc = vfs_rmdir(dparent->d_inode, dchild);
1740                 break;
1741         case S_IFREG: {
1742 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1743                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1744                                                         offset + 1, 0);
1745                 handle = fsfilt_start_log(obd, dparent->d_inode,
1746                                           FSFILT_OP_UNLINK, NULL,
1747                                           le32_to_cpu(lmm->lmm_stripe_count));
1748                 if (IS_ERR(handle))
1749                         GOTO(cleanup, rc = PTR_ERR(handle));
1750
1751                 cleanup_phase = 4; /* transaction */
1752                 rc = vfs_unlink(dparent->d_inode, dchild);
1753
1754                 if (!rc && log_unlink)
1755                         if (mds_log_op_unlink(obd, child_inode,
1756                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1757                                 req->rq_repmsg->buflens[offset + 1],
1758                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1759                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1760                                 body->valid |= OBD_MD_FLCOOKIE;
1761                 break;
1762         }
1763         case S_IFLNK:
1764         case S_IFCHR:
1765         case S_IFBLK:
1766         case S_IFIFO:
1767         case S_IFSOCK:
1768                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1769                                       NULL);
1770                 if (IS_ERR(handle))
1771                         GOTO(cleanup, rc = PTR_ERR(handle));
1772                 cleanup_phase = 4; /* transaction */
1773                 rc = vfs_unlink(dparent->d_inode, dchild);
1774                 break;
1775         default:
1776                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1777                        rec->ur_name);
1778                 LBUG();
1779                 GOTO(cleanup, rc = -EINVAL);
1780         }
1781
1782  cleanup:
1783         if (rc == 0) {
1784                 struct iattr iattr;
1785                 int err;
1786
1787                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1788                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1789                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1790
1791                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1792                 if (err)
1793                         CERROR("error on parent setattr: rc = %d\n", err);
1794         }
1795
1796         switch(cleanup_phase) {
1797         case 4:
1798                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1799                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1800                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1801                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1802
1803                         /* filesystem is really going to destroy an inode
1804                          * we have to delay this till inode is opened -bzzz */
1805                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1806                 }
1807                 /* handle splitted dir */
1808                 mds_unlink_slave_objs(obd, dchild);
1809                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1810                                         rc, 0);
1811                 if (!rc)
1812                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1813                                            "unlinked", 0, NULL);
1814         case 3: /* child ino-reuse lock */
1815                 if (rc && body != NULL) {
1816                         // Don't unlink the OST objects if the MDS unlink failed
1817                         body->valid = 0;
1818                 }
1819                 if (rc)
1820                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1821                 else
1822                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1823         case 2: /* child lock */
1824                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1825                 if (child_lockh.cookie)
1826                         ldlm_lock_decref(&child_lockh, LCK_EX);
1827         case 1: /* child and parent dentry, parent lock */
1828 #ifdef S_PDIROPS
1829                 if (parent_lockh[1].cookie != 0)
1830                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1831 #endif
1832                 if (rc)
1833                         ldlm_lock_decref(parent_lockh, LCK_PW);
1834                 else
1835                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1836                 l_dput(dchild);
1837                 l_dput(dchild);
1838                 l_dput(dparent);
1839         case 0:
1840                 break;
1841         default:
1842                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1843                 LBUG();
1844         }
1845         req->rq_status = rc;
1846         return 0;
1847 }
1848
1849 /*
1850  * to service requests from remote MDS to increment i_nlink
1851  */
1852 static int mds_reint_link_acquire(struct mds_update_record *rec,
1853                                  int offset, struct ptlrpc_request *req,
1854                                  struct lustre_handle *lh)
1855 {
1856         struct obd_device *obd = req->rq_export->exp_obd;
1857         struct ldlm_res_id src_res_id = { .name = {0} };
1858         struct lustre_handle *handle = NULL, src_lockh = {0};
1859         struct mds_obd *mds = mds_req2mds(req);
1860         int rc = 0, cleanup_phase = 0;
1861         struct dentry *de_src = NULL;
1862         ldlm_policy_data_t policy;
1863         int flags = 0;
1864         ENTRY;
1865
1866         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1867                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1868                   (unsigned) rec->ur_fid1->id,
1869                   (unsigned) rec->ur_fid1->generation);
1870
1871         /* Step 1: Lookup the source inode and target directory by FID */
1872         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1873         if (IS_ERR(de_src))
1874                 GOTO(cleanup, rc = PTR_ERR(de_src));
1875         cleanup_phase = 1; /* source dentry */
1876
1877         src_res_id.name[0] = de_src->d_inode->i_ino;
1878         src_res_id.name[1] = de_src->d_inode->i_generation;
1879         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1880
1881         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1882                         src_res_id, LDLM_IBITS, &policy,
1883                         LCK_EX, &flags, mds_blocking_ast,
1884                         ldlm_completion_ast, NULL, NULL,
1885                         NULL, 0, NULL, &src_lockh);
1886         if (rc != ELDLM_OK)
1887                 GOTO(cleanup, rc = -ENOLCK);
1888         cleanup_phase = 2; /* lock */
1889
1890         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1891
1892         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1893         if (IS_ERR(handle)) {
1894                 rc = PTR_ERR(handle);
1895                 GOTO(cleanup, rc);
1896         }
1897         de_src->d_inode->i_nlink++;
1898         mark_inode_dirty(de_src->d_inode);
1899
1900 cleanup:
1901         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1902                                         handle, req, rc, 0);
1903         EXIT;
1904         switch (cleanup_phase) {
1905                 case 2:
1906                         if (rc)
1907                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1908                         else
1909                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1910                 case 1:
1911                         l_dput(de_src);
1912                 case 0:
1913                         break;
1914                 default:
1915                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1916                         LBUG();
1917         }
1918         req->rq_status = rc;
1919         return 0;
1920 }
1921
1922 /*
1923  * request to link to foreign inode:
1924  *  - acquire i_nlinks on this inode
1925  *  - add dentry
1926  */
1927 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1928                                     int offset, struct ptlrpc_request *req,
1929                                     struct lustre_handle *lh)
1930 {
1931         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1932         struct obd_device *obd = req->rq_export->exp_obd;
1933         struct dentry *de_tgt_dir = NULL;
1934         struct mds_obd *mds = mds_req2mds(req);
1935         int rc = 0, cleanup_phase = 0;
1936         struct mdc_op_data op_data;
1937         struct ptlrpc_request *request = NULL;
1938         int update_mode;
1939         ENTRY;
1940
1941 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1942         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1943                   (unsigned) rec->ur_fid2->mds,
1944                   (unsigned) rec->ur_fid2->id,
1945                   (unsigned) rec->ur_fid2->generation,
1946                   rec->ur_namelen - 1, rec->ur_name,
1947                   (unsigned) rec->ur_fid1->mds,
1948                   (unsigned) rec->ur_fid1->id,
1949                   (unsigned)rec->ur_fid1->generation);
1950
1951         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1952                                            tgt_dir_lockh, &update_mode,
1953                                            rec->ur_name, rec->ur_namelen - 1,
1954                                            MDS_INODELOCK_UPDATE);
1955         if (IS_ERR(de_tgt_dir))
1956                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1957         cleanup_phase = 1;
1958
1959         op_data.fid1 = *(rec->ur_fid1);
1960         op_data.namelen = 0;
1961         op_data.name = NULL;
1962         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1963         if (rc)
1964                 GOTO(cleanup, rc);
1965
1966         cleanup_phase = 2;
1967         if (request)
1968                 ptlrpc_req_finished(request);
1969
1970         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1971
1972         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1973         if (IS_ERR(handle)) {
1974                 rc = PTR_ERR(handle);
1975                 GOTO(cleanup, rc);
1976         }
1977
1978         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1979                                   rec->ur_namelen - 1, rec->ur_fid1->id,
1980                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
1981         cleanup_phase = 3;
1982
1983 cleanup:
1984         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1985                                 handle, req, rc, 0);
1986         EXIT;
1987
1988         switch (cleanup_phase) {
1989                 case 3:
1990                         if (rc) {
1991                                 /* FIXME: drop i_nlink on remote inode here */
1992                                 CERROR("MUST drop drop i_nlink here\n");
1993                         }
1994                 case 2:
1995                 case 1:
1996                         if (rc) {
1997                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1998 #ifdef S_PDIROPS
1999                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2000 #endif
2001                         } else {
2002                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2003 #ifdef S_PDIROPS
2004                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2005 #endif
2006                         }
2007                         l_dput(de_tgt_dir);
2008                         break;
2009                 default:
2010                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2011                         LBUG();
2012         }
2013         req->rq_status = rc;
2014         return 0;
2015 }
2016
2017 static int mds_reint_link(struct mds_update_record *rec, int offset,
2018                           struct ptlrpc_request *req,
2019                           struct lustre_handle *lh)
2020 {
2021         struct obd_device *obd = req->rq_export->exp_obd;
2022         struct dentry *de_src = NULL;
2023         struct dentry *de_tgt_dir = NULL;
2024         struct dentry *dchild = NULL;
2025         struct mds_obd *mds = mds_req2mds(req);
2026         struct lustre_handle *handle = NULL;
2027         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2028         struct ldlm_res_id src_res_id = { .name = {0} };
2029         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2030         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2031         ldlm_policy_data_t tgt_dir_policy =
2032                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2033         int rc = 0, cleanup_phase = 0;
2034         int update_mode = 0;
2035         ENTRY;
2036
2037         LASSERT(offset == 0);
2038
2039         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2040                   rec->ur_fid1->id, rec->ur_fid1->generation,
2041                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2042
2043         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2044
2045 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2046         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2047                 GOTO(cleanup, rc = -ENOENT);
2048
2049         if (rec->ur_fid1->mds != mds->mds_num) {
2050                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2051                 RETURN(rc);
2052         }
2053
2054         if (rec->ur_namelen == 1) {
2055                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2056                 RETURN(rc);
2057         }
2058
2059         /* Step 1: Lookup the source inode and target directory by FID */
2060         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2061         if (IS_ERR(de_src))
2062                 GOTO(cleanup, rc = PTR_ERR(de_src));
2063
2064         cleanup_phase = 1; /* source dentry */
2065
2066         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2067         if (IS_ERR(de_tgt_dir)) {
2068                 rc = PTR_ERR(de_tgt_dir);
2069                 de_tgt_dir = NULL;
2070                 GOTO(cleanup, rc);
2071         }
2072
2073         cleanup_phase = 2; /* target directory dentry */
2074
2075         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2076                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2077                de_src->d_inode->i_ino);
2078
2079         /* Step 2: Take the two locks */
2080         src_res_id.name[0] = de_src->d_inode->i_ino;
2081         src_res_id.name[1] = de_src->d_inode->i_generation;
2082         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2083         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2084 #ifdef S_PDIROPS
2085         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2086                 int flags = 0;
2087                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2088                 if (update_mode) {
2089                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2090                                               tgt_dir_res_id, LDLM_IBITS,
2091                                               &src_policy, update_mode, &flags,
2092                                               mds_blocking_ast,
2093                                               ldlm_completion_ast, NULL, NULL,
2094                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2095                         if (rc != ELDLM_OK)
2096                                 GOTO(cleanup, rc = -ENOLCK);
2097                 }
2098
2099                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2100                                                         rec->ur_namelen - 1);
2101                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2102                        de_tgt_dir->d_inode->i_ino,
2103                        de_tgt_dir->d_inode->i_generation,
2104                        tgt_dir_res_id.name[2]);
2105         }
2106 #endif
2107         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2108                                    &src_policy,
2109                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2110                                    &tgt_dir_policy);
2111         if (rc)
2112                 GOTO(cleanup, rc);
2113
2114         cleanup_phase = 3; /* locks */
2115
2116         /* Step 3: Lookup the child */
2117         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2118         if (IS_ERR(dchild)) {
2119                 rc = PTR_ERR(dchild);
2120                 if (rc != -EPERM && rc != -EACCES)
2121                         CERROR("child lookup error %d\n", rc);
2122                 GOTO(cleanup, rc);
2123         }
2124
2125         cleanup_phase = 4; /* child dentry */
2126
2127         if (dchild->d_inode) {
2128                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2129                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2130                 rc = -EEXIST;
2131                 GOTO(cleanup, rc);
2132         }
2133
2134         /* Step 4: Do it. */
2135         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2136
2137         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2138         if (IS_ERR(handle)) {
2139                 rc = PTR_ERR(handle);
2140                 GOTO(cleanup, rc);
2141         }
2142
2143         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2144         if (rc && rc != -EPERM && rc != -EACCES)
2145                 CERROR("vfs_link error %d\n", rc);
2146 cleanup:
2147         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2148                                 handle, req, rc, 0);
2149         EXIT;
2150
2151         switch (cleanup_phase) {
2152         case 4: /* child dentry */
2153                 l_dput(dchild);
2154         case 3: /* locks */
2155                 if (rc) {
2156                         ldlm_lock_decref(&src_lockh, LCK_EX);
2157                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2158                 } else {
2159                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2160                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2161                 }
2162         case 2: /* target dentry */
2163 #ifdef S_PDIROPS
2164                 if (tgt_dir_lockh[1].cookie && update_mode)
2165                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2166 #endif
2167                 if (de_tgt_dir)
2168                         l_dput(de_tgt_dir);
2169         case 1: /* source dentry */
2170                 l_dput(de_src);
2171         case 0:
2172                 break;
2173         default:
2174                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2175                 LBUG();
2176         }
2177         req->rq_status = rc;
2178         return 0;
2179 }
2180
2181 /*
2182  * add a hard link in the PENDING directory, only used by rename()
2183  */
2184 static int mds_add_link_orphan(struct mds_update_record *rec,
2185                                struct obd_device *obd,
2186                                struct dentry *dentry)
2187 {
2188         struct mds_obd *mds = &obd->u.mds;
2189         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2190         struct dentry *pending_child;
2191         char fidname[LL_FID_NAMELEN];
2192         int fidlen = 0, rc;
2193         ENTRY;
2194
2195         LASSERT(dentry->d_inode);
2196         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2197
2198         down(&pending_dir->i_sem);
2199         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2200                             dentry->d_inode->i_generation);
2201
2202         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2203                mds_open_orphan_count(dentry->d_inode),
2204                rec->ur_name, fidname);
2205
2206         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2207         if (IS_ERR(pending_child))
2208                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2209
2210         if (pending_child->d_inode != NULL) {
2211                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2212                 LASSERT(pending_child->d_inode == dentry->d_inode);
2213                 GOTO(out_dput, rc = 0);
2214         }
2215
2216         lock_kernel();
2217         rc = vfs_link(dentry, pending_dir, pending_child);
2218         unlock_kernel();
2219         if (rc)
2220                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2221                        rec->ur_name, rc);
2222         else
2223                 mds_inode_set_orphan(dentry->d_inode);
2224 out_dput:
2225         l_dput(pending_child);
2226 out_lock:
2227         up(&pending_dir->i_sem);
2228         RETURN(rc);
2229 }
2230
2231 /* The idea here is that we need to get four locks in the end:
2232  * one on each parent directory, one on each child.  We need to take
2233  * these locks in some kind of order (to avoid deadlocks), and the order
2234  * I selected is "increasing resource number" order.  We need to look up
2235  * the children, however, before we know what the resource number(s) are.
2236  * Thus the following plan:
2237  *
2238  * 1,2. Look up the parents
2239  * 3,4. Look up the children
2240  * 5. Take locks on the parents and children, in order
2241  * 6. Verify that the children haven't changed since they were looked up
2242  *
2243  * If there was a race and the children changed since they were first looked
2244  * up, it is possible that mds_verify_child() will be able to just grab the
2245  * lock on the new child resource (if it has a higher resource than any other)
2246  * but we need to compare against not only its parent, but also against the
2247  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2248  *
2249  * We need the fancy igrab() on the child inodes because we aren't holding a
2250  * lock on the parent after the lookup is done, so dentry->d_inode may change
2251  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2252  */
2253 static int mds_get_parents_children_locked(struct obd_device *obd,
2254                                            struct mds_obd *mds,
2255                                            struct ll_fid *p1_fid,
2256                                            struct dentry **de_srcdirp,
2257                                            struct ll_fid *p2_fid,
2258                                            struct dentry **de_tgtdirp,
2259                                            int parent_mode,
2260                                            const char *old_name, int old_len,
2261                                            struct dentry **de_oldp,
2262                                            const char *new_name, int new_len,
2263                                            struct dentry **de_newp,
2264                                            struct lustre_handle *dlm_handles,
2265                                            int child_mode)
2266 {
2267         struct ldlm_res_id p1_res_id = { .name = {0} };
2268         struct ldlm_res_id p2_res_id = { .name = {0} };
2269         struct ldlm_res_id c1_res_id = { .name = {0} };
2270         struct ldlm_res_id c2_res_id = { .name = {0} };
2271         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2272         /* Only dentry should change, but the inode itself would be
2273            intact otherwise */
2274         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2275         /* If something is going to be replaced, both dentry and inode locks are
2276            needed */
2277         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2278                                                         MDS_INODELOCK_UPDATE}};
2279         struct ldlm_res_id *maxres_src, *maxres_tgt;
2280         struct inode *inode;
2281         int rc = 0, cleanup_phase = 0;
2282         ENTRY;
2283
2284         /* Step 1: Lookup the source directory */
2285         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2286         if (IS_ERR(*de_srcdirp))
2287                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2288
2289         cleanup_phase = 1; /* source directory dentry */
2290
2291         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2292         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2293
2294         /* Step 2: Lookup the target directory */
2295         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2296                 *de_tgtdirp = dget(*de_srcdirp);
2297         } else {
2298                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2299                 if (IS_ERR(*de_tgtdirp)) {
2300                         rc = PTR_ERR(*de_tgtdirp);
2301                         *de_tgtdirp = NULL;
2302                         GOTO(cleanup, rc);
2303                 }
2304         }
2305
2306         cleanup_phase = 2; /* target directory dentry */
2307
2308         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2309         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2310
2311 #ifdef S_PDIROPS
2312         dlm_handles[5].cookie = 0;
2313         dlm_handles[6].cookie = 0;
2314         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2315                 /* Get a temp lock on just ino, gen to flush client cache and
2316                  * to protect dirs from concurrent splitting */
2317                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2318                                            LCK_PW, &p_policy, &p2_res_id,
2319                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2320                 if (rc != ELDLM_OK)
2321                         GOTO(cleanup, rc);
2322                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2323                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2324                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2325                        (*de_srcdirp)->d_inode->i_ino,
2326                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2327                        (*de_tgtdirp)->d_inode->i_ino,
2328                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2329         }
2330         cleanup_phase = 3;
2331 #endif
2332
2333         /* Step 3: Lookup the source child entry */
2334         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2335         if (IS_ERR(*de_oldp)) {
2336                 rc = PTR_ERR(*de_oldp);
2337                 CERROR("old child lookup error (%*s): %d\n",
2338                        old_len - 1, old_name, rc);
2339                 GOTO(cleanup, rc);
2340         }
2341
2342         cleanup_phase = 4; /* original name dentry */
2343
2344         inode = (*de_oldp)->d_inode;
2345         if (inode != NULL) {
2346                 inode = igrab(inode);
2347                 if (inode == NULL)
2348                         GOTO(cleanup, rc = -ENOENT);
2349
2350                 c1_res_id.name[0] = inode->i_ino;
2351                 c1_res_id.name[1] = inode->i_generation;
2352                 iput(inode);
2353         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2354                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2355                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2356         }
2357
2358         /* Step 4: Lookup the target child entry */
2359         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2360         if (IS_ERR(*de_newp)) {
2361                 rc = PTR_ERR(*de_newp);
2362                 CERROR("new child lookup error (%*s): %d\n",
2363                        old_len - 1, old_name, rc);
2364                 GOTO(cleanup, rc);
2365         }
2366
2367         cleanup_phase = 5; /* target dentry */
2368
2369         inode = (*de_newp)->d_inode;
2370         if (inode != NULL) {
2371                 inode = igrab(inode);
2372                 if (inode == NULL)
2373                         goto retry_locks;
2374
2375                 c2_res_id.name[0] = inode->i_ino;
2376                 c2_res_id.name[1] = inode->i_generation;
2377                 iput(inode);
2378         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2379                 c2_res_id.name[0] = (*de_newp)->d_inum;
2380                 c2_res_id.name[1] = (*de_newp)->d_generation;
2381         }
2382
2383 retry_locks:
2384         /* Step 5: Take locks on the parents and child(ren) */
2385         maxres_src = &p1_res_id;
2386         maxres_tgt = &p2_res_id;
2387         cleanup_phase = 5; /* target dentry */
2388
2389         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2390                 maxres_src = &c1_res_id;
2391         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2392                 maxres_tgt = &c2_res_id;
2393
2394         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2395                                     &p_policy,
2396                                     &p2_res_id, &dlm_handles[1], parent_mode,
2397                                     &p_policy,
2398                                     &c1_res_id, &dlm_handles[2], child_mode,
2399                                     &c1_policy,
2400                                     &c2_res_id, &dlm_handles[3], child_mode,
2401                                     &c2_policy);
2402         if (rc)
2403                 GOTO(cleanup, rc);
2404
2405         cleanup_phase = 6; /* parent and child(ren) locks */
2406
2407         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2408         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2409                               parent_mode, &c1_res_id, &dlm_handles[2],
2410                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2411                               maxres_tgt);
2412         if (rc) {
2413                 if (c2_res_id.name[0] != 0)
2414                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2415                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2416                 cleanup_phase = 5;
2417                 if (rc > 0)
2418                         goto retry_locks;
2419                 GOTO(cleanup, rc);
2420         }
2421
2422         if (!DENTRY_VALID(*de_oldp))
2423                 GOTO(cleanup, rc = -ENOENT);
2424
2425         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2426         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2427                               parent_mode, &c2_res_id, &dlm_handles[3],
2428                               de_newp, child_mode, &c2_policy, new_name,
2429                               new_len, maxres_src);
2430         if (rc) {
2431                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2432                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2433                 cleanup_phase = 5;
2434                 if (rc > 0)
2435                         goto retry_locks;
2436                 GOTO(cleanup, rc);
2437         }
2438
2439         EXIT;
2440 cleanup:
2441         if (rc) {
2442                 switch (cleanup_phase) {
2443                 case 6: /* child lock(s) */
2444                         if (c2_res_id.name[0] != 0)
2445                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2446                         if (c1_res_id.name[0] != 0)
2447                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2448                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2449                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2450                 case 5: /* target dentry */
2451                         l_dput(*de_newp);
2452                 case 4: /* source dentry */
2453                         l_dput(*de_oldp);
2454                 case 3:
2455 #ifdef S_PDIROPS
2456                         if (dlm_handles[5].cookie != 0)
2457                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2458                         if (dlm_handles[6].cookie != 0)
2459                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2460 #endif
2461                 case 2: /* target directory dentry */
2462                         l_dput(*de_tgtdirp);
2463                 case 1: /* source directry dentry */
2464                         l_dput(*de_srcdirp);
2465                 }
2466         }
2467
2468         return rc;
2469 }
2470
2471 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2472                                         int offset, struct ptlrpc_request *req)
2473 {
2474         struct obd_device *obd = req->rq_export->exp_obd;
2475         struct dentry *de_srcdir = NULL;
2476         struct dentry *de_new = NULL;
2477         struct mds_obd *mds = mds_req2mds(req);
2478         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2479         struct lustre_handle child_lockh = {0};
2480         int cleanup_phase = 0;
2481         void *handle = NULL;
2482         int update_mode, rc = 0;
2483         ENTRY;
2484
2485         /* another MDS executing rename operation has asked us
2486          * to create target name. such a creation should destroy
2487          * existing target name */
2488
2489         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2490                         obd->obd_name, rec->ur_tgt,
2491                         (unsigned long) rec->ur_fid1->mds,
2492                         (unsigned long) rec->ur_fid1->id,
2493                         (unsigned long) rec->ur_fid1->generation);
2494
2495         /* first, lookup the target */
2496         child_lockh.cookie = 0;
2497         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2498                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2499                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2500                                          &child_lockh, &de_new, LCK_EX,
2501                                          MDS_INODELOCK_LOOKUP);
2502         if (rc)
2503                 GOTO(cleanup, rc);
2504
2505         cleanup_phase = 1;
2506
2507         LASSERT(de_srcdir);
2508         LASSERT(de_srcdir->d_inode);
2509         LASSERT(de_new);
2510
2511         if (de_new->d_inode) {
2512                 /* name exists and points to local inode
2513                  * try to unlink this name and create new one */
2514                 CERROR("%s: %s points to local inode %lu/%lu\n",
2515                        obd->obd_name, rec->ur_tgt,
2516                        (unsigned long) de_new->d_inode->i_ino,
2517                        (unsigned long) de_new->d_inode->i_generation);
2518                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2519                                       FSFILT_OP_RENAME, NULL);
2520                 if (IS_ERR(handle))
2521                         GOTO(cleanup, rc = PTR_ERR(handle));
2522                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2523                 if (rc)
2524                         GOTO(cleanup, rc);
2525         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2526                 /* name exists adn points to remove inode */
2527                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2528                        obd->obd_name, rec->ur_tgt,
2529                        (unsigned long) de_new->d_mdsnum,
2530                        (unsigned long) de_new->d_inum,
2531                        (unsigned long) de_new->d_generation);
2532         } else {
2533                 /* name doesn't exist. the simplest case */
2534                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2535                                       FSFILT_OP_LINK, NULL);
2536                 if (IS_ERR(handle))
2537                         GOTO(cleanup, rc = PTR_ERR(handle));
2538         }
2539
2540         cleanup_phase = 2;
2541         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2542                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2543                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2544         if (rc)
2545                 CERROR("add_dir_entry() returned error %d\n", rc);
2546 cleanup:
2547         EXIT;
2548         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2549                                 handle, req, rc, 0);
2550         switch(cleanup_phase) {
2551                 case 2:
2552                 case 1:
2553 #ifdef S_PDIROPS
2554                         if (parent_lockh[1].cookie != 0)
2555                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2556 #endif
2557                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2558                         if (child_lockh.cookie != 0)
2559                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2560                         l_dput(de_new);
2561                         l_dput(de_srcdir);
2562                         break;
2563                 default:
2564                         LBUG();
2565         }
2566
2567         req->rq_status = rc;
2568
2569         RETURN(0);
2570 }
2571
2572 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2573                                       struct ptlrpc_request *req)
2574 {
2575         struct obd_device *obd = req->rq_export->exp_obd;
2576         struct ptlrpc_request *req2 = NULL;
2577         struct dentry *de_srcdir = NULL;
2578         struct dentry *de_old = NULL;
2579         struct mds_obd *mds = mds_req2mds(req);
2580         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2581         struct lustre_handle child_lockh = {0};
2582         struct mdc_op_data opdata;
2583         void *handle = NULL;
2584         int update_mode, rc = 0;
2585         ENTRY;
2586
2587         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2588                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2589         memset(&opdata, 0, sizeof(opdata));
2590
2591         child_lockh.cookie = 0;
2592         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2593                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2594                                          &update_mode, rec->ur_name,
2595                                          rec->ur_namelen, &child_lockh, &de_old,
2596                                          LCK_EX, MDS_INODELOCK_LOOKUP);
2597         LASSERT(rc == 0);
2598         LASSERT(de_srcdir);
2599         LASSERT(de_srcdir->d_inode);
2600         LASSERT(de_old);
2601
2602         /* we already know the target should be created on another MDS
2603          * so, we have to request that MDS to do it */
2604
2605         /* prepare source fid */
2606         if (de_old->d_flags & DCACHE_CROSS_REF) {
2607                 LASSERT(de_old->d_inode == NULL);
2608                 CDEBUG(D_OTHER, "request to move remote name\n");
2609                 opdata.fid1.mds = de_old->d_mdsnum;
2610                 opdata.fid1.id = de_old->d_inum;
2611                 opdata.fid1.generation = de_old->d_generation;
2612         } else if (de_old->d_inode == NULL) {
2613                 /* oh, source doesn't exist */
2614                 GOTO(cleanup, rc = -ENOENT);
2615         } else {
2616                 LASSERT(de_old->d_inode != NULL);
2617                 CDEBUG(D_OTHER, "request to move local name\n");
2618                 opdata.fid1.mds = mds->mds_num;
2619                 opdata.fid1.id = de_old->d_inode->i_ino;
2620                 opdata.fid1.generation = de_old->d_inode->i_generation;
2621         }
2622
2623         opdata.fid2 = *(rec->ur_fid2);
2624         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2625                        rec->ur_tgtlen - 1, &req2);
2626
2627         if (rc)
2628                 GOTO(cleanup, rc);
2629
2630         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2631         if (IS_ERR(handle))
2632                 GOTO(cleanup, rc = PTR_ERR(handle));
2633         rc = fsfilt_del_dir_entry(obd, de_old);
2634         d_drop(de_old);
2635
2636 cleanup:
2637         EXIT;
2638         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2639                                 handle, req, rc, 0);
2640         if (req2)
2641                 ptlrpc_req_finished(req2);
2642
2643 #ifdef S_PDIROPS
2644         if (parent_lockh[1].cookie != 0)
2645                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2646 #endif
2647         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2648         if (child_lockh.cookie != 0)
2649                 ldlm_lock_decref(&child_lockh, LCK_EX);
2650
2651         l_dput(de_old);
2652         l_dput(de_srcdir);
2653
2654         req->rq_status = rc;
2655         RETURN(0);
2656
2657 }
2658
2659 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2660                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2661 {
2662         struct obd_device *obd = req->rq_export->exp_obd;
2663         struct dentry *de_srcdir = NULL;
2664         struct dentry *de_tgtdir = NULL;
2665         struct dentry *de_old = NULL;
2666         struct dentry *de_new = NULL;
2667         struct mds_obd *mds = mds_req2mds(req);
2668         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2669         struct mds_body *body = NULL;
2670         int rc = 0, lock_count = 3;
2671         int cleanup_phase = 0;
2672         void *handle = NULL;
2673         ENTRY;
2674
2675         LASSERT(offset == 0);
2676
2677         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2678                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2679                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2680
2681         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2682
2683         if (rec->ur_namelen == 1) {
2684                 rc = mds_reint_rename_create_name(rec, offset, req);
2685                 RETURN(rc);
2686         }
2687
2688         if (rec->ur_fid2->mds != mds->mds_num) {
2689                 rc = mds_reint_rename_to_remote(rec, offset, req);
2690                 RETURN(rc);
2691         }
2692
2693         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2694                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2695                                              rec->ur_name, rec->ur_namelen,
2696                                              &de_old, rec->ur_tgt,
2697                                              rec->ur_tgtlen, &de_new,
2698                                              dlm_handles, LCK_EX);
2699         if (rc)
2700                 GOTO(cleanup, rc);
2701
2702         cleanup_phase = 1; /* parent(s), children, locks */
2703
2704         if (de_new->d_inode)
2705                 lock_count = 4;
2706
2707         /* sanity check for src inode */
2708         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2709             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2710                 GOTO(cleanup, rc = -EINVAL);
2711
2712         /* sanity check for dest inode */
2713         if (de_new->d_inode &&
2714             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2715              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2716                 GOTO(cleanup, rc = -EINVAL);
2717
2718         if (de_old->d_inode == de_new->d_inode)
2719                 GOTO(cleanup, rc = 0);
2720
2721         /* if we are about to remove the target at first, pass the EA of
2722          * that inode to client to perform and cleanup on OST */
2723         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2724         LASSERT(body != NULL);
2725
2726         if (de_new->d_inode &&
2727             S_ISREG(de_new->d_inode->i_mode) &&
2728             de_new->d_inode->i_nlink == 1 &&
2729             mds_open_orphan_count(de_new->d_inode) == 0) {
2730                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2731                 mds_pack_inode2body(obd, body, de_new->d_inode);
2732                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2733                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2734                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2735                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2736                 } else {
2737                         /* XXX need log unlink? */
2738                 }
2739         }
2740
2741         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2742                        de_srcdir->d_inode->i_sb);
2743
2744         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2745         if (IS_ERR(handle))
2746                 GOTO(cleanup, rc = PTR_ERR(handle));
2747
2748         /* FIXME need adjust the journal block count? */
2749         /* if the target should be moved to PENDING, we at first increase the
2750          * link and later vfs_rename() will decrease the link count again */
2751         if (de_new->d_inode &&
2752             S_ISREG(de_new->d_inode->i_mode) &&
2753             de_new->d_inode->i_nlink == 1 &&
2754             mds_open_orphan_count(de_new->d_inode) > 0) {
2755                 rc = mds_add_link_orphan(rec, obd, de_new);
2756                 if (rc)
2757                         GOTO(cleanup, rc);
2758         }
2759
2760         lock_kernel();
2761         de_old->d_fsdata = req;
2762         de_new->d_fsdata = req;
2763         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2764         unlock_kernel();
2765
2766         GOTO(cleanup, rc);
2767 cleanup:
2768         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2769                                 handle, req, rc, 0);
2770         switch (cleanup_phase) {
2771         case 1:
2772 #ifdef S_PDIROPS
2773                 if (dlm_handles[5].cookie != 0)
2774                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2775                 if (dlm_handles[6].cookie != 0)
2776                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2777 #endif
2778                 if (rc) {
2779                         if (lock_count == 4)
2780                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2781                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2782                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2783                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2784                 } else {
2785                         if (lock_count == 4)
2786                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2787                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2788                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2789                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2790                 }
2791                 l_dput(de_new);
2792                 l_dput(de_old);
2793                 l_dput(de_tgtdir);
2794                 l_dput(de_srcdir);
2795         case 0:
2796                 break;
2797         default:
2798                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2799                 LBUG();
2800         }
2801         req->rq_status = rc;
2802         return 0;
2803 }
2804
2805 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2806                            struct ptlrpc_request *, struct lustre_handle *);
2807
2808 static mds_reinter reinters[REINT_MAX + 1] = {
2809         [REINT_SETATTR] mds_reint_setattr,
2810         [REINT_CREATE] mds_reint_create,
2811         [REINT_LINK] mds_reint_link,
2812         [REINT_UNLINK] mds_reint_unlink,
2813         [REINT_RENAME] mds_reint_rename,
2814         [REINT_OPEN] mds_open
2815 };
2816
2817 int mds_reint_rec(struct mds_update_record *rec, int offset,
2818                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2819 {
2820         struct obd_device *obd = req->rq_export->exp_obd;
2821         struct lvfs_run_ctxt saved;
2822         int rc;
2823
2824         /* checked by unpacker */
2825         LASSERT(rec->ur_opcode <= REINT_MAX &&
2826                 reinters[rec->ur_opcode] != NULL);
2827
2828         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2829         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2830         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2831
2832         return rc;
2833 }