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