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         ENTRY;
 534
 535         mds_req_from_mcd(req, med->med_mcd);
 536
 537         if (req->rq_status) {
 538                 EXIT;
 539                 return;
 540         }
 541
 542         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 543         LASSERT(!IS_ERR(parent));
 544         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 545         LASSERT(!IS_ERR(child));
 546         if ((child->d_flags & DCACHE_CROSS_REF)) {
 547                 LASSERTF(child->d_inode == NULL, "BUG 3869");
 548                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 549                 mds_pack_dentry2fid(&body->fid1, child);
 550         } else if (child->d_inode == NULL) {
 551                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 552                           rec->ur_fid1->id, rec->ur_fid1->generation,
 553                           rec->ur_name, rec->ur_mode);
 554                 LASSERTF(child->d_inode != NULL, "BUG 3869");
 555         } else {
 556                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 557                 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 558                 mds_pack_inode2body(req2obd(req), body, child->d_inode);
 559         }
 560         l_dput(parent);
 561         l_dput(child);
 562         EXIT;
 563 }
 564
 565 static int mds_reint_create(struct mds_update_record *rec, int offset,
 566                             struct ptlrpc_request *req,
 567                             struct lustre_handle *lh)
 568 {
 569         struct dentry *dparent = NULL;
 570         struct mds_obd *mds = mds_req2mds(req);
 571         struct obd_device *obd = req->rq_export->exp_obd;
 572         struct dentry *dchild = NULL;
 573         struct inode *dir = NULL;
 574         void *handle = NULL;
 575         struct lustre_handle lockh[2] = {{0}, {0}};
 576         int parent_mode;
 577         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 578         int created = 0;
 579         struct dentry_params dp;
 580         struct mea *mea = NULL;
 581         int mea_size;
 582         ENTRY;
 583
 584         LASSERT(offset == 0);
 585         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 586
 587         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 588                   rec->ur_fid1->id, rec->ur_fid1->generation,
 589                   rec->ur_name, rec->ur_mode);
 590
 591         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 592
 593         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 594                 GOTO(cleanup, rc = -ESTALE);
 595
 596         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 597                                         &parent_mode, rec->ur_name,
 598                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 599         if (IS_ERR(dparent)) {
 600                 rc = PTR_ERR(dparent);
 601                 CERROR("parent lookup error %d\n", rc);
 602                 GOTO(cleanup, rc);
 603         }
 604         cleanup_phase = 1; /* locked parent dentry */
 605         dir = dparent->d_inode;
 606         LASSERT(dir);
 607
 608         ldlm_lock_dump_handle(D_OTHER, lockh);
 609
 610         /* try to retrieve MEA data for this dir */
 611         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 612         if (rc)
 613                 GOTO(cleanup, rc);
 614
 615         if (mea != NULL) {
 616                 /* dir is already splitted, check is requested filename
 617                  * should live at this MDS or at another one */
 618                 int i;
 619                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 620                 if (mea->mea_master != mea->mea_fids[i].mds) {
 621                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 622                                " should be %d(%d)\n",
 623                                mea->mea_master, dparent->d_inode->i_ino,
 624                                dparent->d_inode->i_generation, rec->ur_name,
 625                                mea->mea_fids[i].mds, i);
 626                         GOTO(cleanup, rc = -ERESTART);
 627                 }
 628         }
 629
 630         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 631         if (IS_ERR(dchild)) {
 632                 rc = PTR_ERR(dchild);
 633                 CERROR("child lookup error %d\n", rc);
 634                 GOTO(cleanup, rc);
 635         }
 636
 637         cleanup_phase = 2; /* child dentry */
 638
 639         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 640
 641         if (type == S_IFREG || type == S_IFDIR) {
 642                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 643                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 644                        obd->obd_name, dparent->d_inode->i_ino,
 645                        dparent->d_inode->i_generation, rc, parent_mode);
 646                 if (rc > 0) {
 647                         /* dir got splitted */
 648                         GOTO(cleanup, rc = -ERESTART);
 649                 } else if (rc < 0) {
 650                         /* error happened during spitting. */
 651                         GOTO(cleanup, rc);
 652                 }
 653         }
 654
 655         if (dir->i_mode & S_ISGID) {
 656                 if (S_ISDIR(rec->ur_mode))
 657                         rec->ur_mode |= S_ISGID;
 658         }
 659
 660         dchild->d_fsdata = (void *)&dp;
 661         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 662         dp.p_ptr = req;
 663
 664         switch (type) {
 665         case S_IFREG:{
 666                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 667                 if (IS_ERR(handle))
 668                         GOTO(cleanup, rc = PTR_ERR(handle));
 669                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 670                 EXIT;
 671                 break;
 672         }
 673         case S_IFDIR:{
 674                 int nstripes = 0;
 675                 int i;
 676
 677                 /* as Peter asked, mkdir() should distribute new directories
 678                  * over the whole cluster in order to distribute namespace
 679                  * processing load. first, we calculate which MDS to use to put
 680                  * new directory's inode in. */
 681                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 682                                       rec->ur_flags);
 683                 if (i == mds->mds_num) {
 684                         /* inode will be created locally */
 685                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 686                         if (IS_ERR(handle))
 687                                 GOTO(cleanup, rc = PTR_ERR(handle));
 688
 689                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 690
 691                         if (rec->ur_eadata)
 692                                 nstripes = *(u16 *)rec->ur_eadata;
 693
 694                         if (rc == 0 && nstripes) {
 695                                 /* we pass LCK_EX to split routine to
 696                                  * signalthat we have exclusive access
 697                                  * to the directory. simple because
 698                                  * nobody knows it already exists -bzzz */
 699                                 rc = mds_try_to_split_dir(obd, dchild,
 700                                                           NULL, nstripes,
 701                                                           LCK_EX);
 702                                 if (rc > 0) {
 703                                         /* dir got splitted */
 704                                         rc = 0;
 705                                 } else if (rc < 0) {
 706                                         /* an error occured during
 707                                          * splitting. */
 708                                         GOTO(cleanup, rc);
 709                                 }
 710                         }
 711                 } else if (!DENTRY_VALID(dchild)) {
 712                         /* inode will be created on another MDS */
 713                         struct obdo *oa = NULL;
 714                         struct mds_body *body;
 715
 716                         /* first, create that inode */
 717                         oa = obdo_alloc();
 718                         LASSERT(oa != NULL);
 719                         oa->o_mds = i;
 720                         oa->o_easize = 0;
 721                         if (rec->ur_eadata) {
 722                                 /* user asks for creating splitted dir */
 723                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 724                         }
 725
 726                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 727                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 728                                         OBD_MD_FLUID | OBD_MD_FLGID);
 729                         oa->o_mode = dir->i_mode;
 730                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 731                                         obd->obd_name, i);
 732                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 733                                 CWARN("%s: replay dir creation %*s -> %u/%u\n",
 734                                       obd->obd_name, rec->ur_namelen - 1,
 735                                       rec->ur_name, (unsigned) rec->ur_fid2->id,
 736                                       (unsigned) rec->ur_fid2->generation);
 737                                 oa->o_id = rec->ur_fid2->id;
 738                                 oa->o_generation = rec->ur_fid2->generation;
 739                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 740                         }
 741
 742                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 743                         if (rc) {
 744                                 obdo_free(oa);
 745                                 GOTO(cleanup, rc);
 746                         }
 747
 748                         /* now, add new dir entry for it */
 749                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 750                         if (IS_ERR(handle)) {
 751                                 obdo_free(oa);
 752                                 GOTO(cleanup, rc = PTR_ERR(handle));
 753                         }
 754                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 755                                                   rec->ur_namelen - 1,
 756                                                   oa->o_id, oa->o_generation,
 757                                                   i);
 758                         LASSERT(rc == 0);
 759
 760                         /* fill reply */
 761                         body = lustre_msg_buf(req->rq_repmsg,
 762                                               offset, sizeof (*body));
 763                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 764                         body->fid1.id = oa->o_id;
 765                         body->fid1.mds = i;
 766                         body->fid1.generation = oa->o_generation;
 767                         obdo_free(oa);
 768                 } else {
 769                         /* requested name exists in the directory */
 770                         rc = -EEXIST;
 771                 }
 772                 EXIT;
 773                 break;
 774         }
 775         case S_IFLNK:{
 776                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 777                 if (IS_ERR(handle))
 778                         GOTO(cleanup, rc = PTR_ERR(handle));
 779                 if (rec->ur_tgt == NULL)        /* no target supplied */
 780                         rc = -EINVAL;           /* -EPROTO? */
 781                 else
 782                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 783                 EXIT;
 784                 break;
 785         }
 786         case S_IFCHR:
 787         case S_IFBLK:
 788         case S_IFIFO:
 789         case S_IFSOCK:{
 790                 int rdev = rec->ur_rdev;
 791                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 792                 if (IS_ERR(handle))
 793                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 794                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 795                 EXIT;
 796                 break;
 797         }
 798         default:
 799                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 800                 dchild->d_fsdata = NULL;
 801                 GOTO(cleanup, rc = -EINVAL);
 802         }
 803
 804         /* In case we stored the desired inum in here, we want to clean up. */
 805         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 806                 dchild->d_fsdata = NULL;
 807
 808         if (rc) {
 809                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 810                 GOTO(cleanup, rc);
 811         } else if (dchild->d_inode) {
 812                 struct iattr iattr;
 813                 struct inode *inode = dchild->d_inode;
 814                 struct mds_body *body;
 815
 816                 created = 1;
 817                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 818                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 819                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 820                 iattr.ia_uid = rec->_ur_fsuid;
 821                 if (dir->i_mode & S_ISGID)
 822                         iattr.ia_gid = dir->i_gid;
 823                 else
 824                         iattr.ia_gid = rec->_ur_fsgid;
 825                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 826                         ATTR_MTIME | ATTR_CTIME;
 827
 828                 if (rec->ur_fid2->id) {
 829                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 830                         inode->i_generation = rec->ur_fid2->generation;
 831                         /* Dirtied and committed by the upcoming setattr. */
 832                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 833                                inode->i_ino, inode->i_generation);
 834                 } else {
 835                         struct lustre_handle child_ino_lockh;
 836
 837                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 838                                inode->i_ino, inode->i_generation);
 839
 840                         /* The inode we were allocated may have just been freed
 841                          * by an unlink operation.  We take this lock to
 842                          * synchronize against the matching reply-ack-lock taken
 843                          * in unlink, to avoid replay problems if this reply
 844                          * makes it out to the client but the unlink's does not.
 845                          * See bug 2029 for more detail.*/
 846                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 847                         if (rc != ELDLM_OK) {
 848                                 CERROR("error locking for unlink/create sync: "
 849                                        "%d\n", rc);
 850                         } else {
 851                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 852                         }
 853                 }
 854
 855                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 856                 if (rc)
 857                         CERROR("error on child setattr: rc = %d\n", rc);
 858
 859                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 860                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 861                 if (rc)
 862                         CERROR("error on parent setattr: rc = %d\n", rc);
 863
 864                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 865                 mds_pack_inode2fid(obd, &body->fid1, inode);
 866                 mds_pack_inode2body(obd, body, inode);
 867         }
 868         EXIT;
 869
 870 cleanup:
 871         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 872
 873         if (rc && created) {
 874                 /* Destroy the file we just created. This should not need extra
 875                  * journal credits, as we have already modified all of the
 876                  * blocks needed in order to create the file in the first
 877                  * place. */
 878                 switch (type) {
 879                 case S_IFDIR:
 880                         err = vfs_rmdir(dir, dchild);
 881                         if (err)
 882                                 CERROR("rmdir in error path: %d\n", err);
 883                         break;
 884                 default:
 885                         err = vfs_unlink(dir, dchild);
 886                         if (err)
 887                                 CERROR("unlink in error path: %d\n", err);
 888                         break;
 889                 }
 890         } else {
 891                 rc = err;
 892         }
 893         switch (cleanup_phase) {
 894         case 2: /* child dentry */
 895                 l_dput(dchild);
 896         case 1: /* locked parent dentry */
 897 #ifdef S_PDIROPS
 898                 if (lockh[1].cookie != 0)
 899                         ldlm_lock_decref(lockh + 1, parent_mode);
 900 #endif
 901                 if (rc) {
 902                         ldlm_lock_decref(lockh, LCK_PW);
 903                 } else {
 904                         ptlrpc_save_lock(req, lockh, LCK_PW);
 905                 }
 906                 l_dput(dparent);
 907         case 0:
 908                 break;
 909         default:
 910                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 911                 LBUG();
 912         }
 913         if (mea)
 914                 OBD_FREE(mea, mea_size);
 915         req->rq_status = rc;
 916         return 0;
 917 }
 918
 919 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 920            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 921 {
 922         int i;
 923
 924         for (i = 0; i < RES_NAME_SIZE; i++) {
 925                 /* return 1 here, because enqueue_ordered will skip resources
 926                  * of all zeroes if they're sorted to the end of the list. */
 927                 if (res1->name[i] == 0 && res2->name[i] != 0)
 928                         return 1;
 929                 if (res2->name[i] == 0 && res1->name[i] != 0)
 930                         return 0;
 931
 932                 if (res1->name[i] > res2->name[i])
 933                         return 1;
 934                 if (res1->name[i] < res2->name[i])
 935                         return 0;
 936         }
 937
 938         if (!p1 || !p2)
 939                 return 0;
 940
 941         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 942                 return 1;
 943
 944         return 0;
 945 }
 946
 947 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 948  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 949  * because they take the place of local semaphores.
 950  *
 951  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 952  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 953 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 954                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 955                           ldlm_policy_data_t *p1_policy,
 956                           struct ldlm_res_id *p2_res_id,
 957                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 958                           ldlm_policy_data_t *p2_policy)
 959 {
 960         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 961         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 962         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 963         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 964         int rc, flags;
 965         ENTRY;
 966
 967         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 968
 969         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 970                res_id[1]->name[0]);
 971
 972         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 973                 handles[1] = p1_lockh;
 974                 handles[0] = p2_lockh;
 975                 res_id[1] = p1_res_id;
 976                 res_id[0] = p2_res_id;
 977                 lock_modes[1] = p1_lock_mode;
 978                 lock_modes[0] = p2_lock_mode;
 979                 policies[1] = p1_policy;
 980                 policies[0] = p2_policy;
 981         }
 982
 983         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 984                res_id[0]->name[0], res_id[1]->name[0]);
 985
 986         flags = LDLM_FL_LOCAL_ONLY;
 987         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 988                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 989                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 990                               NULL, 0, NULL, handles[0]);
 991         if (rc != ELDLM_OK)
 992                 RETURN(-EIO);
 993         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 994
 995         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
 996             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
 997                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 998                 ldlm_lock_addref(handles[1], lock_modes[1]);
 999         } else if (res_id[1]->name[0] != 0) {
1000                 flags = LDLM_FL_LOCAL_ONLY;
1001                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1002                                       *res_id[1], LDLM_IBITS, policies[1],
1003                                       lock_modes[1], &flags, mds_blocking_ast,
1004                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1005                                       NULL, handles[1]);
1006                 if (rc != ELDLM_OK) {
1007                         ldlm_lock_decref(handles[0], lock_modes[0]);
1008                         RETURN(-EIO);
1009                 }
1010                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1011         }
1012
1013         RETURN(0);
1014 }
1015
1016 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1017                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1018                            ldlm_policy_data_t *p1_policy,
1019                            struct ldlm_res_id *p2_res_id,
1020                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1021                            ldlm_policy_data_t *p2_policy,
1022                            struct ldlm_res_id *c1_res_id,
1023                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1024                            ldlm_policy_data_t *c1_policy,
1025                            struct ldlm_res_id *c2_res_id,
1026                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1027                            ldlm_policy_data_t *c2_policy)
1028 {
1029         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1030                                           c1_res_id, c2_res_id };
1031         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1032                                                  c1_lockh, c2_lockh };
1033         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1034                               c1_lock_mode, c2_lock_mode };
1035         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1036                                             c1_policy, c2_policy};
1037         int rc, i, j, sorted, flags;
1038         ENTRY;
1039
1040         CDEBUG(D_DLMTRACE,
1041                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1042                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1043                res_id[3]->name[0]);
1044
1045         /* simple insertion sort - we have at most 4 elements */
1046         for (i = 1; i < 4; i++) {
1047                 j = i - 1;
1048                 dlm_handles[4] = dlm_handles[i];
1049                 res_id[4] = res_id[i];
1050                 lock_modes[4] = lock_modes[i];
1051                 policies[4] = policies[i];
1052
1053                 sorted = 0;
1054                 do {
1055                         if (res_gt(res_id[j], res_id[4], policies[j],
1056                                    policies[4])) {
1057                                 dlm_handles[j + 1] = dlm_handles[j];
1058                                 res_id[j + 1] = res_id[j];
1059                                 lock_modes[j + 1] = lock_modes[j];
1060                                 policies[j + 1] = policies[j];
1061                                 j--;
1062                         } else {
1063                                 sorted = 1;
1064                         }
1065                 } while (j >= 0 && !sorted);
1066
1067                 dlm_handles[j + 1] = dlm_handles[4];
1068                 res_id[j + 1] = res_id[4];
1069                 lock_modes[j + 1] = lock_modes[4];
1070                 policies[j + 1] = policies[4];
1071         }
1072
1073         CDEBUG(D_DLMTRACE,
1074                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1075                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1076                res_id[3]->name[0]);
1077
1078         /* XXX we could send ASTs on all these locks first before blocking? */
1079         for (i = 0; i < 4; i++) {
1080                 flags = 0;
1081                 if (res_id[i]->name[0] == 0)
1082                         break;
1083                 if (i != 0 &&
1084                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1085                     (policies[i]->l_inodebits.bits &
1086                      policies[i-1]->l_inodebits.bits) ) {
1087                         memcpy(dlm_handles[i], dlm_handles[i-1],
1088                                sizeof(*(dlm_handles[i])));
1089                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1090                 } else {
1091                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1092                                               *res_id[i], LDLM_IBITS,
1093                                               policies[i],
1094                                               lock_modes[i], &flags,
1095                                               mds_blocking_ast,
1096                                               ldlm_completion_ast, NULL, NULL,
1097                                               NULL, 0, NULL, dlm_handles[i]);
1098                         if (rc != ELDLM_OK)
1099                                 GOTO(out_err, rc = -EIO);
1100                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1101                 }
1102         }
1103
1104         RETURN(0);
1105 out_err:
1106         while (i-- > 0)
1107                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1108
1109         return rc;
1110 }
1111
1112 /* In the unlikely case that the child changed while we were waiting
1113  * on the lock, we need to drop the lock on the old child and either:
1114  * - if the child has a lower resource name, then we have to also
1115  *   drop the parent lock and regain the locks in the right order
1116  * - in the rename case, if the child has a lower resource name than one of
1117  *   the other parent/child resources (maxres) we also need to reget the locks
1118  * - if the child has a higher resource name (this is the common case)
1119  *   we can just get the lock on the new child (still in lock order)
1120  *
1121  * Returns 0 if the child did not change or if it changed but could be locked.
1122  * Returns 1 if the child changed and we need to re-lock (no locks held).
1123  * Returns -ve error with a valid dchild (no locks held). */
1124 static int mds_verify_child(struct obd_device *obd,
1125                             struct ldlm_res_id *parent_res_id,
1126                             struct lustre_handle *parent_lockh,
1127                             struct dentry *dparent, int parent_mode,
1128                             struct ldlm_res_id *child_res_id,
1129                             struct lustre_handle *child_lockh,
1130                             struct dentry **dchildp, int child_mode,
1131                             ldlm_policy_data_t *child_policy,
1132                             const char *name, int namelen,
1133                             struct ldlm_res_id *maxres)
1134 {
1135         struct dentry *vchild, *dchild = *dchildp;
1136         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1137         ENTRY;
1138
1139         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1140         if (IS_ERR(vchild))
1141                 GOTO(cleanup, rc = PTR_ERR(vchild));
1142
1143         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1144                 if  (child_res_id->name[0] == vchild->d_inum &&
1145                                 child_res_id->name[1] == vchild->d_generation) {
1146                         if (dchild != NULL)
1147                                 l_dput(dchild);
1148                         *dchildp = vchild;
1149                         RETURN(0);
1150                 }
1151                 goto changed;
1152         }
1153
1154         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1155                    (vchild->d_inode != NULL &&
1156                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1157                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1158                 if (dchild != NULL)
1159                         l_dput(dchild);
1160                 *dchildp = vchild;
1161
1162                 RETURN(0);
1163         }
1164
1165 changed:
1166         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1167                vchild->d_inode, dchild ? dchild->d_inode : 0,
1168                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1169                child_res_id->name[0]);
1170         if (child_res_id->name[0] != 0)
1171                 ldlm_lock_decref(child_lockh, child_mode);
1172         if (dchild)
1173                 l_dput(dchild);
1174
1175         cleanup_phase = 1; /* parent lock only */
1176         *dchildp = dchild = vchild;
1177
1178         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1179                 int flags = 0;
1180                 if (dchild->d_inode) {
1181                         child_res_id->name[0] = dchild->d_inode->i_ino;
1182                         child_res_id->name[1] = dchild->d_inode->i_generation;
1183                 } else {
1184                         child_res_id->name[0] = dchild->d_inum;
1185                         child_res_id->name[1] = dchild->d_generation;
1186                 }
1187
1188                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1189                     res_gt(maxres, child_res_id, NULL, NULL)) {
1190                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1191                                child_res_id->name[0], parent_res_id->name[0],
1192                                maxres->name[0]);
1193                         GOTO(cleanup, rc = 1);
1194                 }
1195
1196                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1197                                       *child_res_id, LDLM_IBITS, child_policy,
1198                                       child_mode, &flags, mds_blocking_ast,
1199                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1200                                       NULL, child_lockh);
1201                 if (rc != ELDLM_OK)
1202                         GOTO(cleanup, rc = -EIO);
1203
1204         } else {
1205                 memset(child_res_id, 0, sizeof(*child_res_id));
1206         }
1207
1208         EXIT;
1209 cleanup:
1210         if (rc) {
1211                 switch(cleanup_phase) {
1212                 case 2:
1213                         if (child_res_id->name[0] != 0)
1214                                 ldlm_lock_decref(child_lockh, child_mode);
1215                 case 1:
1216                         ldlm_lock_decref(parent_lockh, parent_mode);
1217                 }
1218         }
1219         return rc;
1220 }
1221
1222 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1223                                 struct ll_fid *fid,
1224                                 struct lustre_handle *parent_lockh,
1225                                 struct dentry **dparentp, int parent_mode,
1226                                 __u64 parent_lockpart, int *update_mode,
1227                                 char *name, int namelen,
1228                                 struct lustre_handle *child_lockh,
1229                                 struct dentry **dchildp, int child_mode,
1230                                 __u64 child_lockpart, void *clone_info)
1231 {
1232         struct ldlm_res_id child_res_id = { .name = {0} };
1233         struct ldlm_res_id parent_res_id = { .name = {0} };
1234         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1235         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1236         struct inode *inode;
1237         int rc = 0, cleanup_phase = 0;
1238         ENTRY;
1239
1240         /* Step 1: Lookup parent */
1241         *dparentp = mds_fid2dentry(mds, fid, NULL);
1242         if (IS_ERR(*dparentp)) {
1243                 rc = PTR_ERR(*dparentp);
1244                 *dparentp = NULL;
1245                 RETURN(rc);
1246         }
1247
1248         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1249                (*dparentp)->d_inode->i_ino, name);
1250
1251         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1252         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1253 #ifdef S_PDIROPS
1254         parent_lockh[1].cookie = 0;
1255         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1256                 struct ldlm_res_id res_id = { .name = {0} };
1257                 ldlm_policy_data_t policy;
1258                 int flags = 0;
1259                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1260                 if (*update_mode) {
1261                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1262                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1263                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1264                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1265                                               res_id, LDLM_IBITS, &policy,
1266                                               *update_mode, &flags,
1267                                               mds_blocking_ast,
1268                                               ldlm_completion_ast,
1269                                               NULL, NULL, NULL, 0, NULL,
1270                                               parent_lockh + 1);
1271                         if (rc != ELDLM_OK)
1272                                 RETURN(-ENOLCK);
1273                 }
1274
1275                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1276                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1277                        (*dparentp)->d_inode->i_ino,
1278                        (*dparentp)->d_inode->i_generation,
1279                        parent_res_id.name[2]);
1280         }
1281 #endif
1282
1283         cleanup_phase = 1; /* parent dentry */
1284 #ifdef CONFIG_SNAPFS
1285         if (clone_info) {
1286                 /*FIXME is there any other FUNC will use d_fsdata,
1287                  *excepet creating inode according inum*/
1288                 (*dparentp)->d_fsdata = clone_info;
1289         }
1290 #endif
1291
1292         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1293         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1294         if (IS_ERR(*dchildp)) {
1295                 rc = PTR_ERR(*dchildp);
1296                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1297                 GOTO(cleanup, rc);
1298         }
1299
1300         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1301                 /* inode lives on another MDS: return * mds/ino/gen
1302                  * and LOOKUP lock. drop possible UPDATE lock! */
1303                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1304                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1305                 child_res_id.name[0] = (*dchildp)->d_inum;
1306                 child_res_id.name[1] = (*dchildp)->d_generation;
1307                 goto retry_locks;
1308         }
1309
1310         inode = (*dchildp)->d_inode;
1311         if (inode != NULL)
1312                 inode = igrab(inode);
1313         if (inode == NULL)
1314                 goto retry_locks;
1315
1316         child_res_id.name[0] = inode->i_ino;
1317         child_res_id.name[1] = inode->i_generation;
1318
1319         iput(inode);
1320
1321 retry_locks:
1322         cleanup_phase = 2; /* child dentry */
1323
1324         /* Step 3: Lock parent and child in resource order.  If child doesn't
1325          *         exist, we still have to lock the parent and re-lookup. */
1326         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1327                                    &parent_policy, &child_res_id, child_lockh,
1328                                    child_mode, &child_policy);
1329         if (rc)
1330                 GOTO(cleanup, rc);
1331
1332         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1333                 cleanup_phase = 4; /* child lock */
1334         else
1335                 cleanup_phase = 3; /* parent lock */
1336
1337         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1338         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1339                               parent_mode, &child_res_id, child_lockh,
1340                               dchildp, child_mode, &child_policy,
1341                               name, namelen, &parent_res_id);
1342         if (rc > 0)
1343                 goto retry_locks;
1344         if (rc < 0) {
1345                 cleanup_phase = 3;
1346                 GOTO(cleanup, rc);
1347         }
1348
1349 cleanup:
1350         if (rc) {
1351                 switch (cleanup_phase) {
1352                 case 4:
1353                         ldlm_lock_decref(child_lockh, child_mode);
1354                 case 3:
1355                         ldlm_lock_decref(parent_lockh, parent_mode);
1356                 case 2:
1357                         l_dput(*dchildp);
1358                 case 1:
1359 #ifdef S_PDIROPS
1360                         if (parent_lockh[1].cookie)
1361                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1362 #endif
1363                         l_dput(*dparentp);
1364                 default: ;
1365                 }
1366         }
1367         return rc;
1368 }
1369
1370 void mds_reconstruct_generic(struct ptlrpc_request *req)
1371 {
1372         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1373
1374         mds_req_from_mcd(req, med->med_mcd);
1375 }
1376
1377 int mds_create_local_dentry(struct mds_update_record *rec,
1378                            struct obd_device *obd)
1379 {
1380         struct mds_obd *mds = &obd->u.mds;
1381         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1382         int fidlen = 0, rc, cleanup_phase = 0;
1383         struct dentry *new_child = NULL;
1384         char *fidname = rec->ur_name;
1385         struct dentry *child = NULL;
1386         struct lustre_handle lockh[2] = {{0}, {0}};
1387         void *handle;
1388         ENTRY;
1389
1390         down(&fids_dir->i_sem);
1391         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1392         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1393                         fidname, (unsigned) rec->ur_fid1->id,
1394                         (unsigned) rec->ur_fid1->generation);
1395
1396         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1397         up(&fids_dir->i_sem);
1398         if (IS_ERR(new_child)) {
1399                 CERROR("can't lookup %s: %d\n", fidname,
1400                                 (int) PTR_ERR(new_child));
1401                 GOTO(cleanup, rc = PTR_ERR(new_child));
1402         }
1403         cleanup_phase = 1;
1404
1405         if (new_child->d_inode != NULL) {
1406                 /* nice. we've already have local dentry! */
1407                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1408                        (unsigned) new_child->d_inode->i_ino,
1409                        (unsigned) new_child->d_inode->i_generation);
1410                 rec->ur_fid1->id = fids_dir->i_ino;
1411                 rec->ur_fid1->generation = fids_dir->i_generation;
1412                 rec->ur_namelen = fidlen + 1;
1413                 GOTO(cleanup, rc = 0);
1414         }
1415
1416         /* new, local dentry will be added soon. we need no aliases here */
1417         d_drop(new_child);
1418
1419         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1420                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1421         } else {
1422                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1423                                               LCK_EX, lockh, NULL, NULL, 0,
1424                                               MDS_INODELOCK_UPDATE);
1425         }
1426
1427         if (IS_ERR(child)) {
1428                 CERROR("can't get victim: %d\n", (int) PTR_ERR(child));
1429                 GOTO(cleanup, rc = PTR_ERR(child));
1430         }
1431         cleanup_phase = 2;
1432
1433         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1434         if (IS_ERR(handle))
1435                 GOTO(cleanup, rc = PTR_ERR(handle));
1436
1437         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1438                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1439                                   mds->mds_num);
1440         if (rc)
1441                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1442                        (unsigned long) child->d_inode->i_ino,
1443                        (unsigned long) child->d_inode->i_generation, rc);
1444         else {
1445                 if (S_ISDIR(child->d_inode->i_mode)) {
1446                         fids_dir->i_nlink++;
1447                         mark_inode_dirty(fids_dir);
1448                 }
1449                 mark_inode_dirty(child->d_inode);
1450         }
1451         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1452
1453         rec->ur_fid1->id = fids_dir->i_ino;
1454         rec->ur_fid1->generation = fids_dir->i_generation;
1455         rec->ur_namelen = fidlen + 1;
1456
1457 cleanup:
1458         switch(cleanup_phase) {
1459                 case 2:
1460                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1461                                 ldlm_lock_decref(lockh, LCK_EX);
1462                         dput(child);
1463                 case 1:
1464                         dput(new_child);
1465                 case 0:
1466                        break;
1467         }
1468         RETURN(rc);
1469 }
1470
1471 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1472                                         struct ptlrpc_request *slave)
1473 {
1474         void *cookie, *cookie2;
1475         struct mds_body *body2;
1476         struct mds_body *body;
1477         void *ea, *ea2;
1478         ENTRY;
1479
1480         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1481         LASSERT(body != NULL);
1482
1483         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1484         LASSERT(body2 != NULL);
1485
1486         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1487                 RETURN(0);
1488         }
1489
1490         memcpy(body2, body, sizeof(*body));
1491         body2->valid &= ~OBD_MD_FLCOOKIE;
1492
1493         if (!(body->valid & OBD_MD_FLEASIZE) &&
1494             !(body->valid & OBD_MD_FLDIREA))
1495                 RETURN(0);
1496
1497         if (body->eadatasize == 0) {
1498                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1499                 RETURN(0);
1500         }
1501
1502         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1503
1504         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1505         LASSERT(ea != NULL);
1506
1507         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1508         LASSERT(ea2 != NULL);
1509
1510         memcpy(ea2, ea, body->eadatasize);
1511
1512         if (body->valid & OBD_MD_FLCOOKIE) {
1513                 LASSERT(master->rq_repmsg->buflens[2] >=
1514                                 slave->rq_repmsg->buflens[2]);
1515                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1516                                 slave->rq_repmsg->buflens[2]);
1517                 LASSERT(cookie != NULL);
1518
1519                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1520                                 master->rq_repmsg->buflens[2]);
1521                 LASSERT(cookie2 != NULL);
1522                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1523                 body2->valid |= OBD_MD_FLCOOKIE;
1524         }
1525         RETURN(0);
1526 }
1527
1528 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1529                                    struct ptlrpc_request *req,
1530                                    struct lustre_handle *parent_lockh,
1531                                    int update_mode,
1532                                    struct dentry *dparent,
1533                                    struct lustre_handle *child_lockh,
1534                                    struct dentry *dchild)
1535 {
1536         struct obd_device *obd = req->rq_export->exp_obd;
1537         struct mds_obd *mds = mds_req2mds(req);
1538         struct mdc_op_data op_data;
1539         int rc = 0, cleanup_phase = 0;
1540         struct ptlrpc_request *request = NULL;
1541         void *handle;
1542         ENTRY;
1543
1544         LASSERT(offset == 0 || offset == 2);
1545
1546         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1547                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1548                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1549         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1550                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1551                           rec->ur_namelen - 1, rec->ur_name,
1552                           (unsigned)dchild->d_mdsnum,
1553                           (unsigned) dchild->d_inum,
1554                           (unsigned) dchild->d_generation);
1555
1556         /* time to drop i_nlink on remote MDS */
1557         memset(&op_data, 0, sizeof(op_data));
1558         op_data.fid1.mds = dchild->d_mdsnum;
1559         op_data.fid1.id = dchild->d_inum;
1560         op_data.fid1.generation = dchild->d_generation;
1561         op_data.create_mode = rec->ur_mode;
1562         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1563                 op_data.create_mode |= MDS_MODE_REPLAY;
1564         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1565         cleanup_phase = 2;
1566         if (request) {
1567                 if (rc == 0)
1568                         mds_copy_unlink_reply(req, request);
1569                 ptlrpc_req_finished(request);
1570         }
1571         if (rc == 0) {
1572                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1573                                       NULL);
1574                 if (IS_ERR(handle))
1575                         GOTO(cleanup, rc = PTR_ERR(handle));
1576                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1577                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1578                                         rc, 0);
1579         }
1580 cleanup:
1581         req->rq_status = rc;
1582
1583 #ifdef S_PDIROPS
1584         if (parent_lockh[1].cookie != 0)
1585                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1586 #endif
1587         ldlm_lock_decref(child_lockh, LCK_EX);
1588         if (rc)
1589                 ldlm_lock_decref(parent_lockh, LCK_PW);
1590         else
1591                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1592         l_dput(dchild);
1593         l_dput(dparent);
1594
1595         return 0;
1596 }
1597
1598 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1599                             struct ptlrpc_request *req,
1600                             struct lustre_handle *lh)
1601 {
1602         struct dentry *dparent, *dchild;
1603         struct mds_obd *mds = mds_req2mds(req);
1604         struct obd_device *obd = req->rq_export->exp_obd;
1605         struct mds_body *body = NULL;
1606         struct inode *child_inode;
1607         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1608         struct lustre_handle child_lockh = {0}, child_reuse_lockh = {0};
1609         struct lustre_handle * slave_lockh = NULL;
1610         char fidname[LL_FID_NAMELEN];
1611         void *handle = NULL;
1612         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1613         int unlink_by_fid = 0;
1614         int update_mode;
1615         ENTRY;
1616
1617         LASSERT(offset == 0 || offset == 2);
1618
1619         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1620                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1621
1622         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1623
1624         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1625                 GOTO(cleanup, rc = -ENOENT);
1626
1627         if (rec->ur_namelen == 1) {
1628                 /* this is request to drop i_nlink on local inode */
1629                 unlink_by_fid = 1;
1630                 rec->ur_name = fidname;
1631                 rc = mds_create_local_dentry(rec, obd);
1632                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1633                         DEBUG_REQ(D_HA, req,
1634                                   "drop nlink on inode %u/%u/%u (replay)",
1635                                   (unsigned) rec->ur_fid1->mds,
1636                                   (unsigned) rec->ur_fid1->id,
1637                                   (unsigned) rec->ur_fid1->generation);
1638                         req->rq_status = 0;
1639                         RETURN(0);
1640                 }
1641         }
1642
1643         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1644                 /* master mds for directory asks slave removing
1645                  * inode is already locked */
1646                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1647                                                LCK_PW, parent_lockh,
1648                                                &update_mode, rec->ur_name,
1649                                                rec->ur_namelen,
1650                                                MDS_INODELOCK_UPDATE);
1651                 if (IS_ERR(dparent))
1652                         GOTO(cleanup, rc = PTR_ERR(dparent));
1653                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1654                                            rec->ur_namelen - 1);
1655                 if (IS_ERR(dchild))
1656                         GOTO(cleanup, rc = PTR_ERR(dchild));
1657                 child_lockh.cookie = 0;
1658                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1659                 LASSERT(dchild->d_inode != NULL);
1660                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1661         } else {
1662                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1663                                                  parent_lockh, &dparent,
1664                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1665                                                  &update_mode, rec->ur_name,
1666                                                  rec->ur_namelen, &child_lockh,
1667                                                  &dchild, LCK_EX,
1668                                                  MDS_INODELOCK_LOOKUP |
1669                                                    MDS_INODELOCK_UPDATE, NULL);
1670         }
1671         if (rc)
1672                 GOTO(cleanup, rc);
1673
1674         if (dchild->d_flags & DCACHE_CROSS_REF) {
1675                 /* we should have parent lock only here */
1676                 LASSERT(unlink_by_fid == 0);
1677                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1678                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1679                                         update_mode, dparent, &child_lockh, dchild);
1680                 RETURN(0);
1681         }
1682
1683         cleanup_phase = 1; /* dchild, dparent, locks */
1684
1685         dget(dchild);
1686         child_inode = dchild->d_inode;
1687         if (child_inode == NULL) {
1688                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1689                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1690                 GOTO(cleanup, rc = -ENOENT);
1691         }
1692
1693         cleanup_phase = 2; /* dchild has a lock */
1694
1695         /* We have to do these checks ourselves, in case we are making an
1696          * orphan.  The client tells us whether rmdir() or unlink() was called,
1697          * so we need to return appropriate errors (bug 72).
1698          *
1699          * We don't have to check permissions, because vfs_rename (called from
1700          * mds_open_unlink_rename) also calls may_delete. */
1701         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1702                 if (!S_ISDIR(child_inode->i_mode))
1703                         GOTO(cleanup, rc = -ENOTDIR);
1704         } else {
1705                 if (S_ISDIR(child_inode->i_mode))
1706                         GOTO(cleanup, rc = -EISDIR);
1707         }
1708
1709         /* handle splitted dir */
1710         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1711         if (rc)
1712                 GOTO(cleanup, rc);
1713
1714         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1715          * reuse (see bug 2029). */
1716         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1717         if (rc != ELDLM_OK)
1718                 GOTO(cleanup, rc);
1719
1720         cleanup_phase = 3; /* child inum lock */
1721
1722         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1723
1724         /* ldlm_reply in buf[0] if called via intent */
1725         if (offset)
1726                 offset = 1;
1727
1728         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1729         LASSERT(body != NULL);
1730
1731         /* If this is the last reference to this inode, get the OBD EA
1732          * data first so the client can destroy OST objects.
1733          * we only do the object removal if no open files remain.
1734          * Nobody can get at this name anymore because of the locks so
1735          * we make decisions here as to whether to remove the inode */
1736         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1737             mds_open_orphan_count(child_inode) == 0) {
1738                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1739                 mds_pack_inode2body(obd, body, child_inode);
1740                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1741                             child_inode, 1);
1742                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1743                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1744                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1745                 } else {
1746                         log_unlink = 1;
1747                 }
1748         }
1749
1750         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1751         switch (child_inode->i_mode & S_IFMT) {
1752         case S_IFDIR:
1753                 /* Drop any lingering child directories before we start our
1754                  * transaction, to avoid doing multiple inode dirty/delete
1755                  * in our compound transaction (bug 1321). */
1756                 shrink_dcache_parent(dchild);
1757                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1758                                       NULL);
1759                 if (IS_ERR(handle))
1760                         GOTO(cleanup, rc = PTR_ERR(handle));
1761                 cleanup_phase = 4; /* transaction */
1762                 rc = vfs_rmdir(dparent->d_inode, dchild);
1763                 break;
1764         case S_IFREG: {
1765 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1766                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1767                                                         offset + 1, 0);
1768                 handle = fsfilt_start_log(obd, dparent->d_inode,
1769                                           FSFILT_OP_UNLINK, NULL,
1770                                           le32_to_cpu(lmm->lmm_stripe_count));
1771                 if (IS_ERR(handle))
1772                         GOTO(cleanup, rc = PTR_ERR(handle));
1773
1774                 cleanup_phase = 4; /* transaction */
1775                 rc = vfs_unlink(dparent->d_inode, dchild);
1776
1777                 if (!rc && log_unlink)
1778                         if (mds_log_op_unlink(obd, child_inode,
1779                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1780                                 req->rq_repmsg->buflens[offset + 1],
1781                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1782                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1783                                 body->valid |= OBD_MD_FLCOOKIE;
1784                 break;
1785         }
1786         case S_IFLNK:
1787         case S_IFCHR:
1788         case S_IFBLK:
1789         case S_IFIFO:
1790         case S_IFSOCK:
1791                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1792                                       NULL);
1793                 if (IS_ERR(handle))
1794                         GOTO(cleanup, rc = PTR_ERR(handle));
1795                 cleanup_phase = 4; /* transaction */
1796                 rc = vfs_unlink(dparent->d_inode, dchild);
1797                 break;
1798         default:
1799                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1800                        rec->ur_name);
1801                 LBUG();
1802                 GOTO(cleanup, rc = -EINVAL);
1803         }
1804
1805  cleanup:
1806         if (rc == 0) {
1807                 struct iattr iattr;
1808                 int err;
1809
1810                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1811                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1812                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1813
1814                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1815                 if (err)
1816                         CERROR("error on parent setattr: rc = %d\n", err);
1817         }
1818
1819         switch(cleanup_phase) {
1820         case 4:
1821                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1822                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1823                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1824                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1825
1826                         /* filesystem is really going to destroy an inode
1827                          * we have to delay this till inode is opened -bzzz */
1828                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1829                 }
1830                 /* handle splitted dir */
1831                 mds_unlink_slave_objs(obd, dchild);
1832                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1833                                         rc, 0);
1834                 if (!rc)
1835                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1836                                            "unlinked", 0, NULL);
1837         case 3: /* child ino-reuse lock */
1838                 if (rc && body != NULL) {
1839                         // Don't unlink the OST objects if the MDS unlink failed
1840                         body->valid = 0;
1841                 }
1842                 if (rc)
1843                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1844                 else
1845                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1846         case 2: /* child lock */
1847                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1848                 if (child_lockh.cookie)
1849                         ldlm_lock_decref(&child_lockh, LCK_EX);
1850         case 1: /* child and parent dentry, parent lock */
1851 #ifdef S_PDIROPS
1852                 if (parent_lockh[1].cookie != 0)
1853                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1854 #endif
1855                 if (rc)
1856                         ldlm_lock_decref(parent_lockh, LCK_PW);
1857                 else
1858                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1859                 l_dput(dchild);
1860                 l_dput(dchild);
1861                 l_dput(dparent);
1862         case 0:
1863                 break;
1864         default:
1865                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1866                 LBUG();
1867         }
1868         req->rq_status = rc;
1869         return 0;
1870 }
1871
1872 /*
1873  * to service requests from remote MDS to increment i_nlink
1874  */
1875 static int mds_reint_link_acquire(struct mds_update_record *rec,
1876                                  int offset, struct ptlrpc_request *req,
1877                                  struct lustre_handle *lh)
1878 {
1879         struct obd_device *obd = req->rq_export->exp_obd;
1880         struct ldlm_res_id src_res_id = { .name = {0} };
1881         struct lustre_handle *handle = NULL, src_lockh = {0};
1882         struct mds_obd *mds = mds_req2mds(req);
1883         int rc = 0, cleanup_phase = 0;
1884         struct dentry *de_src = NULL;
1885         ldlm_policy_data_t policy;
1886         int flags = 0;
1887         ENTRY;
1888
1889         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1890                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1891                   (unsigned) rec->ur_fid1->id,
1892                   (unsigned) rec->ur_fid1->generation);
1893
1894         /* Step 1: Lookup the source inode and target directory by FID */
1895         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1896         if (IS_ERR(de_src))
1897                 GOTO(cleanup, rc = PTR_ERR(de_src));
1898         cleanup_phase = 1; /* source dentry */
1899
1900         src_res_id.name[0] = de_src->d_inode->i_ino;
1901         src_res_id.name[1] = de_src->d_inode->i_generation;
1902         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1903
1904         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1905                         src_res_id, LDLM_IBITS, &policy,
1906                         LCK_EX, &flags, mds_blocking_ast,
1907                         ldlm_completion_ast, NULL, NULL,
1908                         NULL, 0, NULL, &src_lockh);
1909         if (rc != ELDLM_OK)
1910                 GOTO(cleanup, rc = -ENOLCK);
1911         cleanup_phase = 2; /* lock */
1912
1913         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1914
1915         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1916         if (IS_ERR(handle)) {
1917                 rc = PTR_ERR(handle);
1918                 GOTO(cleanup, rc);
1919         }
1920         de_src->d_inode->i_nlink++;
1921         mark_inode_dirty(de_src->d_inode);
1922
1923 cleanup:
1924         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1925                                         handle, req, rc, 0);
1926         EXIT;
1927         switch (cleanup_phase) {
1928                 case 2:
1929                         if (rc)
1930                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1931                         else
1932                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1933                 case 1:
1934                         l_dput(de_src);
1935                 case 0:
1936                         break;
1937                 default:
1938                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1939                         LBUG();
1940         }
1941         req->rq_status = rc;
1942         return 0;
1943 }
1944
1945 /*
1946  * request to link to foreign inode:
1947  *  - acquire i_nlinks on this inode
1948  *  - add dentry
1949  */
1950 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1951                                     int offset, struct ptlrpc_request *req,
1952                                     struct lustre_handle *lh)
1953 {
1954         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1955         struct obd_device *obd = req->rq_export->exp_obd;
1956         struct dentry *de_tgt_dir = NULL;
1957         struct mds_obd *mds = mds_req2mds(req);
1958         int rc = 0, cleanup_phase = 0;
1959         struct mdc_op_data op_data;
1960         struct ptlrpc_request *request = NULL;
1961         int update_mode;
1962         ENTRY;
1963
1964 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1965         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1966                   (unsigned) rec->ur_fid2->mds,
1967                   (unsigned) rec->ur_fid2->id,
1968                   (unsigned) rec->ur_fid2->generation,
1969                   rec->ur_namelen - 1, rec->ur_name,
1970                   (unsigned) rec->ur_fid1->mds,
1971                   (unsigned) rec->ur_fid1->id,
1972                   (unsigned)rec->ur_fid1->generation);
1973
1974         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1975                                            tgt_dir_lockh, &update_mode,
1976                                            rec->ur_name, rec->ur_namelen - 1,
1977                                            MDS_INODELOCK_UPDATE);
1978         if (IS_ERR(de_tgt_dir))
1979                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1980         cleanup_phase = 1;
1981
1982         op_data.fid1 = *(rec->ur_fid1);
1983         op_data.namelen = 0;
1984         op_data.name = NULL;
1985         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1986         if (rc)
1987                 GOTO(cleanup, rc);
1988
1989         cleanup_phase = 2;
1990         if (request)
1991                 ptlrpc_req_finished(request);
1992
1993         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1994
1995         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1996         if (IS_ERR(handle)) {
1997                 rc = PTR_ERR(handle);
1998                 GOTO(cleanup, rc);
1999         }
2000
2001         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2002                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2003                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2004         cleanup_phase = 3;
2005
2006 cleanup:
2007         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2008                                 handle, req, rc, 0);
2009         EXIT;
2010
2011         switch (cleanup_phase) {
2012                 case 3:
2013                         if (rc) {
2014                                 /* FIXME: drop i_nlink on remote inode here */
2015                                 CERROR("MUST drop drop i_nlink here\n");
2016                         }
2017                 case 2:
2018                 case 1:
2019                         if (rc) {
2020                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2021 #ifdef S_PDIROPS
2022                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2023 #endif
2024                         } else {
2025                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2026 #ifdef S_PDIROPS
2027                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2028 #endif
2029                         }
2030                         l_dput(de_tgt_dir);
2031                         break;
2032                 default:
2033                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2034                         LBUG();
2035         }
2036         req->rq_status = rc;
2037         return 0;
2038 }
2039
2040 static int mds_reint_link(struct mds_update_record *rec, int offset,
2041                           struct ptlrpc_request *req,
2042                           struct lustre_handle *lh)
2043 {
2044         struct obd_device *obd = req->rq_export->exp_obd;
2045         struct dentry *de_src = NULL;
2046         struct dentry *de_tgt_dir = NULL;
2047         struct dentry *dchild = NULL;
2048         struct mds_obd *mds = mds_req2mds(req);
2049         struct lustre_handle *handle = NULL;
2050         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2051         struct ldlm_res_id src_res_id = { .name = {0} };
2052         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2053         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2054         ldlm_policy_data_t tgt_dir_policy =
2055                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2056         int rc = 0, cleanup_phase = 0;
2057         int update_mode = 0;
2058         ENTRY;
2059
2060         LASSERT(offset == 0);
2061
2062         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2063                   rec->ur_fid1->id, rec->ur_fid1->generation,
2064                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2065
2066         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2067
2068 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2069         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2070                 GOTO(cleanup, rc = -ENOENT);
2071
2072         if (rec->ur_fid1->mds != mds->mds_num) {
2073                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2074                 RETURN(rc);
2075         }
2076
2077         if (rec->ur_namelen == 1) {
2078                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2079                 RETURN(rc);
2080         }
2081
2082         /* Step 1: Lookup the source inode and target directory by FID */
2083         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2084         if (IS_ERR(de_src))
2085                 GOTO(cleanup, rc = PTR_ERR(de_src));
2086
2087         cleanup_phase = 1; /* source dentry */
2088
2089         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2090         if (IS_ERR(de_tgt_dir)) {
2091                 rc = PTR_ERR(de_tgt_dir);
2092                 de_tgt_dir = NULL;
2093                 GOTO(cleanup, rc);
2094         }
2095
2096         cleanup_phase = 2; /* target directory dentry */
2097
2098         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2099                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2100                de_src->d_inode->i_ino);
2101
2102         /* Step 2: Take the two locks */
2103         src_res_id.name[0] = de_src->d_inode->i_ino;
2104         src_res_id.name[1] = de_src->d_inode->i_generation;
2105         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2106         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2107 #ifdef S_PDIROPS
2108         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2109                 int flags = 0;
2110                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2111                 if (update_mode) {
2112                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2113                                               tgt_dir_res_id, LDLM_IBITS,
2114                                               &src_policy, update_mode, &flags,
2115                                               mds_blocking_ast,
2116                                               ldlm_completion_ast, NULL, NULL,
2117                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2118                         if (rc != ELDLM_OK)
2119                                 GOTO(cleanup, rc = -ENOLCK);
2120                 }
2121
2122                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2123                                                         rec->ur_namelen - 1);
2124                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2125                        de_tgt_dir->d_inode->i_ino,
2126                        de_tgt_dir->d_inode->i_generation,
2127                        tgt_dir_res_id.name[2]);
2128         }
2129 #endif
2130         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2131                                    &src_policy,
2132                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2133                                    &tgt_dir_policy);
2134         if (rc)
2135                 GOTO(cleanup, rc);
2136
2137         cleanup_phase = 3; /* locks */
2138
2139         /* Step 3: Lookup the child */
2140         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2141         if (IS_ERR(dchild)) {
2142                 rc = PTR_ERR(dchild);
2143                 if (rc != -EPERM && rc != -EACCES)
2144                         CERROR("child lookup error %d\n", rc);
2145                 GOTO(cleanup, rc);
2146         }
2147
2148         cleanup_phase = 4; /* child dentry */
2149
2150         if (dchild->d_inode) {
2151                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2152                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2153                 rc = -EEXIST;
2154                 GOTO(cleanup, rc);
2155         }
2156
2157         /* Step 4: Do it. */
2158         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2159
2160         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2161         if (IS_ERR(handle)) {
2162                 rc = PTR_ERR(handle);
2163                 GOTO(cleanup, rc);
2164         }
2165
2166         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2167         if (rc && rc != -EPERM && rc != -EACCES)
2168                 CERROR("vfs_link error %d\n", rc);
2169 cleanup:
2170         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2171                                 handle, req, rc, 0);
2172         EXIT;
2173
2174         switch (cleanup_phase) {
2175         case 4: /* child dentry */
2176                 l_dput(dchild);
2177         case 3: /* locks */
2178                 if (rc) {
2179                         ldlm_lock_decref(&src_lockh, LCK_EX);
2180                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2181                 } else {
2182                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2183                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2184                 }
2185         case 2: /* target dentry */
2186 #ifdef S_PDIROPS
2187                 if (tgt_dir_lockh[1].cookie && update_mode)
2188                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2189 #endif
2190                 if (de_tgt_dir)
2191                         l_dput(de_tgt_dir);
2192         case 1: /* source dentry */
2193                 l_dput(de_src);
2194         case 0:
2195                 break;
2196         default:
2197                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2198                 LBUG();
2199         }
2200         req->rq_status = rc;
2201         return 0;
2202 }
2203
2204 /*
2205  * add a hard link in the PENDING directory, only used by rename()
2206  */
2207 static int mds_add_link_orphan(struct mds_update_record *rec,
2208                                struct obd_device *obd,
2209                                struct dentry *dentry)
2210 {
2211         struct mds_obd *mds = &obd->u.mds;
2212         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2213         struct dentry *pending_child;
2214         char fidname[LL_FID_NAMELEN];
2215         int fidlen = 0, rc;
2216         ENTRY;
2217
2218         LASSERT(dentry->d_inode);
2219         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2220
2221         down(&pending_dir->i_sem);
2222         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2223                             dentry->d_inode->i_generation);
2224
2225         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2226                mds_open_orphan_count(dentry->d_inode),
2227                rec->ur_name, fidname);
2228
2229         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2230         if (IS_ERR(pending_child))
2231                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2232
2233         if (pending_child->d_inode != NULL) {
2234                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2235                 LASSERT(pending_child->d_inode == dentry->d_inode);
2236                 GOTO(out_dput, rc = 0);
2237         }
2238
2239         lock_kernel();
2240         rc = vfs_link(dentry, pending_dir, pending_child);
2241         unlock_kernel();
2242         if (rc)
2243                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2244                        rec->ur_name, rc);
2245         else
2246                 mds_inode_set_orphan(dentry->d_inode);
2247 out_dput:
2248         l_dput(pending_child);
2249 out_lock:
2250         up(&pending_dir->i_sem);
2251         RETURN(rc);
2252 }
2253
2254 /* The idea here is that we need to get four locks in the end:
2255  * one on each parent directory, one on each child.  We need to take
2256  * these locks in some kind of order (to avoid deadlocks), and the order
2257  * I selected is "increasing resource number" order.  We need to look up
2258  * the children, however, before we know what the resource number(s) are.
2259  * Thus the following plan:
2260  *
2261  * 1,2. Look up the parents
2262  * 3,4. Look up the children
2263  * 5. Take locks on the parents and children, in order
2264  * 6. Verify that the children haven't changed since they were looked up
2265  *
2266  * If there was a race and the children changed since they were first looked
2267  * up, it is possible that mds_verify_child() will be able to just grab the
2268  * lock on the new child resource (if it has a higher resource than any other)
2269  * but we need to compare against not only its parent, but also against the
2270  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2271  *
2272  * We need the fancy igrab() on the child inodes because we aren't holding a
2273  * lock on the parent after the lookup is done, so dentry->d_inode may change
2274  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2275  */
2276 static int mds_get_parents_children_locked(struct obd_device *obd,
2277                                            struct mds_obd *mds,
2278                                            struct ll_fid *p1_fid,
2279                                            struct dentry **de_srcdirp,
2280                                            struct ll_fid *p2_fid,
2281                                            struct dentry **de_tgtdirp,
2282                                            int parent_mode,
2283                                            const char *old_name, int old_len,
2284                                            struct dentry **de_oldp,
2285                                            const char *new_name, int new_len,
2286                                            struct dentry **de_newp,
2287                                            struct lustre_handle *dlm_handles,
2288                                            int child_mode)
2289 {
2290         struct ldlm_res_id p1_res_id = { .name = {0} };
2291         struct ldlm_res_id p2_res_id = { .name = {0} };
2292         struct ldlm_res_id c1_res_id = { .name = {0} };
2293         struct ldlm_res_id c2_res_id = { .name = {0} };
2294         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2295         /* Only dentry should change, but the inode itself would be
2296            intact otherwise */
2297         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2298         /* If something is going to be replaced, both dentry and inode locks are
2299            needed */
2300         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2301                                                         MDS_INODELOCK_UPDATE}};
2302         struct ldlm_res_id *maxres_src, *maxres_tgt;
2303         struct inode *inode;
2304         int rc = 0, cleanup_phase = 0;
2305         ENTRY;
2306
2307         /* Step 1: Lookup the source directory */
2308         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2309         if (IS_ERR(*de_srcdirp))
2310                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2311
2312         cleanup_phase = 1; /* source directory dentry */
2313
2314         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2315         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2316
2317         /* Step 2: Lookup the target directory */
2318         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2319                 *de_tgtdirp = dget(*de_srcdirp);
2320         } else {
2321                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2322                 if (IS_ERR(*de_tgtdirp)) {
2323                         rc = PTR_ERR(*de_tgtdirp);
2324                         *de_tgtdirp = NULL;
2325                         GOTO(cleanup, rc);
2326                 }
2327         }
2328
2329         cleanup_phase = 2; /* target directory dentry */
2330
2331         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2332         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2333
2334 #ifdef S_PDIROPS
2335         dlm_handles[5].cookie = 0;
2336         dlm_handles[6].cookie = 0;
2337         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2338                 /* Get a temp lock on just ino, gen to flush client cache and
2339                  * to protect dirs from concurrent splitting */
2340                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2341                                            LCK_PW, &p_policy, &p2_res_id,
2342                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2343                 if (rc != ELDLM_OK)
2344                         GOTO(cleanup, rc);
2345                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2346                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2347                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2348                        (*de_srcdirp)->d_inode->i_ino,
2349                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2350                        (*de_tgtdirp)->d_inode->i_ino,
2351                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2352         }
2353         cleanup_phase = 3;
2354 #endif
2355
2356         /* Step 3: Lookup the source child entry */
2357         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2358         if (IS_ERR(*de_oldp)) {
2359                 rc = PTR_ERR(*de_oldp);
2360                 CERROR("old child lookup error (%*s): %d\n",
2361                        old_len - 1, old_name, rc);
2362                 GOTO(cleanup, rc);
2363         }
2364
2365         cleanup_phase = 4; /* original name dentry */
2366
2367         inode = (*de_oldp)->d_inode;
2368         if (inode != NULL) {
2369                 inode = igrab(inode);
2370                 if (inode == NULL)
2371                         GOTO(cleanup, rc = -ENOENT);
2372
2373                 c1_res_id.name[0] = inode->i_ino;
2374                 c1_res_id.name[1] = inode->i_generation;
2375                 iput(inode);
2376         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2377                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2378                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2379         }
2380
2381         /* Step 4: Lookup the target child entry */
2382         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2383         if (IS_ERR(*de_newp)) {
2384                 rc = PTR_ERR(*de_newp);
2385                 CERROR("new child lookup error (%*s): %d\n",
2386                        old_len - 1, old_name, rc);
2387                 GOTO(cleanup, rc);
2388         }
2389
2390         cleanup_phase = 5; /* target dentry */
2391
2392         inode = (*de_newp)->d_inode;
2393         if (inode != NULL) {
2394                 inode = igrab(inode);
2395                 if (inode == NULL)
2396                         goto retry_locks;
2397
2398                 c2_res_id.name[0] = inode->i_ino;
2399                 c2_res_id.name[1] = inode->i_generation;
2400                 iput(inode);
2401         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2402                 c2_res_id.name[0] = (*de_newp)->d_inum;
2403                 c2_res_id.name[1] = (*de_newp)->d_generation;
2404         }
2405
2406 retry_locks:
2407         /* Step 5: Take locks on the parents and child(ren) */
2408         maxres_src = &p1_res_id;
2409         maxres_tgt = &p2_res_id;
2410         cleanup_phase = 5; /* target dentry */
2411
2412         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2413                 maxres_src = &c1_res_id;
2414         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2415                 maxres_tgt = &c2_res_id;
2416
2417         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2418                                     &p_policy,
2419                                     &p2_res_id, &dlm_handles[1], parent_mode,
2420                                     &p_policy,
2421                                     &c1_res_id, &dlm_handles[2], child_mode,
2422                                     &c1_policy,
2423                                     &c2_res_id, &dlm_handles[3], child_mode,
2424                                     &c2_policy);
2425         if (rc)
2426                 GOTO(cleanup, rc);
2427
2428         cleanup_phase = 6; /* parent and child(ren) locks */
2429
2430         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2431         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2432                               parent_mode, &c1_res_id, &dlm_handles[2],
2433                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2434                               maxres_tgt);
2435         if (rc) {
2436                 if (c2_res_id.name[0] != 0)
2437                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2438                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2439                 cleanup_phase = 5;
2440                 if (rc > 0)
2441                         goto retry_locks;
2442                 GOTO(cleanup, rc);
2443         }
2444
2445         if (!DENTRY_VALID(*de_oldp))
2446                 GOTO(cleanup, rc = -ENOENT);
2447
2448         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2449         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2450                               parent_mode, &c2_res_id, &dlm_handles[3],
2451                               de_newp, child_mode, &c2_policy, new_name,
2452                               new_len, maxres_src);
2453         if (rc) {
2454                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2455                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2456                 cleanup_phase = 5;
2457                 if (rc > 0)
2458                         goto retry_locks;
2459                 GOTO(cleanup, rc);
2460         }
2461
2462         EXIT;
2463 cleanup:
2464         if (rc) {
2465                 switch (cleanup_phase) {
2466                 case 6: /* child lock(s) */
2467                         if (c2_res_id.name[0] != 0)
2468                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2469                         if (c1_res_id.name[0] != 0)
2470                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2471                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2472                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2473                 case 5: /* target dentry */
2474                         l_dput(*de_newp);
2475                 case 4: /* source dentry */
2476                         l_dput(*de_oldp);
2477                 case 3:
2478 #ifdef S_PDIROPS
2479                         if (dlm_handles[5].cookie != 0)
2480                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2481                         if (dlm_handles[6].cookie != 0)
2482                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2483 #endif
2484                 case 2: /* target directory dentry */
2485                         l_dput(*de_tgtdirp);
2486                 case 1: /* source directry dentry */
2487                         l_dput(*de_srcdirp);
2488                 }
2489         }
2490
2491         return rc;
2492 }
2493
2494 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2495                                         int offset, struct ptlrpc_request *req)
2496 {
2497         struct obd_device *obd = req->rq_export->exp_obd;
2498         struct dentry *de_srcdir = NULL;
2499         struct dentry *de_new = NULL;
2500         struct mds_obd *mds = mds_req2mds(req);
2501         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2502         struct lustre_handle child_lockh = {0};
2503         int cleanup_phase = 0;
2504         void *handle = NULL;
2505         int update_mode, rc = 0;
2506         ENTRY;
2507
2508         /* another MDS executing rename operation has asked us
2509          * to create target name. such a creation should destroy
2510          * existing target name */
2511
2512         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2513                         obd->obd_name, rec->ur_tgt,
2514                         (unsigned long) rec->ur_fid1->mds,
2515                         (unsigned long) rec->ur_fid1->id,
2516                         (unsigned long) rec->ur_fid1->generation);
2517
2518         /* first, lookup the target */
2519         child_lockh.cookie = 0;
2520         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2521                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2522                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2523                                          &child_lockh, &de_new, LCK_EX,
2524                                          MDS_INODELOCK_LOOKUP, NULL);
2525         if (rc)
2526                 GOTO(cleanup, rc);
2527
2528         cleanup_phase = 1;
2529
2530         LASSERT(de_srcdir);
2531         LASSERT(de_srcdir->d_inode);
2532         LASSERT(de_new);
2533
2534         if (de_new->d_inode) {
2535                 /* name exists and points to local inode
2536                  * try to unlink this name and create new one */
2537                 CERROR("%s: %s points to local inode %lu/%lu\n",
2538                        obd->obd_name, rec->ur_tgt,
2539                        (unsigned long) de_new->d_inode->i_ino,
2540                        (unsigned long) de_new->d_inode->i_generation);
2541                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2542                                       FSFILT_OP_RENAME, NULL);
2543                 if (IS_ERR(handle))
2544                         GOTO(cleanup, rc = PTR_ERR(handle));
2545                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2546                 if (rc)
2547                         GOTO(cleanup, rc);
2548         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2549                 /* name exists adn points to remove inode */
2550                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2551                        obd->obd_name, rec->ur_tgt,
2552                        (unsigned long) de_new->d_mdsnum,
2553                        (unsigned long) de_new->d_inum,
2554                        (unsigned long) de_new->d_generation);
2555         } else {
2556                 /* name doesn't exist. the simplest case */
2557                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2558                                       FSFILT_OP_LINK, NULL);
2559                 if (IS_ERR(handle))
2560                         GOTO(cleanup, rc = PTR_ERR(handle));
2561         }
2562
2563         cleanup_phase = 2;
2564         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2565                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2566                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2567         if (rc)
2568                 CERROR("add_dir_entry() returned error %d\n", rc);
2569 cleanup:
2570         EXIT;
2571         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2572                                 handle, req, rc, 0);
2573         switch(cleanup_phase) {
2574                 case 2:
2575                 case 1:
2576 #ifdef S_PDIROPS
2577                         if (parent_lockh[1].cookie != 0)
2578                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2579 #endif
2580                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2581                         if (child_lockh.cookie != 0)
2582                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2583                         l_dput(de_new);
2584                         l_dput(de_srcdir);
2585                         break;
2586                 default:
2587                         LBUG();
2588         }
2589
2590         req->rq_status = rc;
2591
2592         RETURN(0);
2593 }
2594
2595 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2596                                       struct ptlrpc_request *req)
2597 {
2598         struct obd_device *obd = req->rq_export->exp_obd;
2599         struct ptlrpc_request *req2 = NULL;
2600         struct dentry *de_srcdir = NULL;
2601         struct dentry *de_old = NULL;
2602         struct mds_obd *mds = mds_req2mds(req);
2603         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2604         struct lustre_handle child_lockh = {0};
2605         struct mdc_op_data opdata;
2606         void *handle = NULL;
2607         int update_mode, rc = 0;
2608         ENTRY;
2609
2610         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2611                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2612         memset(&opdata, 0, sizeof(opdata));
2613
2614         child_lockh.cookie = 0;
2615         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2616                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2617                                          &update_mode, rec->ur_name,
2618                                          rec->ur_namelen, &child_lockh, &de_old,
2619                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2620         LASSERT(rc == 0);
2621         LASSERT(de_srcdir);
2622         LASSERT(de_srcdir->d_inode);
2623         LASSERT(de_old);
2624
2625         /* we already know the target should be created on another MDS
2626          * so, we have to request that MDS to do it */
2627
2628         /* prepare source fid */
2629         if (de_old->d_flags & DCACHE_CROSS_REF) {
2630                 LASSERT(de_old->d_inode == NULL);
2631                 CDEBUG(D_OTHER, "request to move remote name\n");
2632                 opdata.fid1.mds = de_old->d_mdsnum;
2633                 opdata.fid1.id = de_old->d_inum;
2634                 opdata.fid1.generation = de_old->d_generation;
2635         } else if (de_old->d_inode == NULL) {
2636                 /* oh, source doesn't exist */
2637                 GOTO(cleanup, rc = -ENOENT);
2638         } else {
2639                 LASSERT(de_old->d_inode != NULL);
2640                 CDEBUG(D_OTHER, "request to move local name\n");
2641                 opdata.fid1.mds = mds->mds_num;
2642                 opdata.fid1.id = de_old->d_inode->i_ino;
2643                 opdata.fid1.generation = de_old->d_inode->i_generation;
2644         }
2645
2646         opdata.fid2 = *(rec->ur_fid2);
2647         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2648                        rec->ur_tgtlen - 1, &req2);
2649
2650         if (rc)
2651                 GOTO(cleanup, rc);
2652
2653         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2654         if (IS_ERR(handle))
2655                 GOTO(cleanup, rc = PTR_ERR(handle));
2656         rc = fsfilt_del_dir_entry(obd, de_old);
2657         d_drop(de_old);
2658
2659 cleanup:
2660         EXIT;
2661         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2662                                 handle, req, rc, 0);
2663         if (req2)
2664                 ptlrpc_req_finished(req2);
2665
2666 #ifdef S_PDIROPS
2667         if (parent_lockh[1].cookie != 0)
2668                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2669 #endif
2670         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2671         if (child_lockh.cookie != 0)
2672                 ldlm_lock_decref(&child_lockh, LCK_EX);
2673
2674         l_dput(de_old);
2675         l_dput(de_srcdir);
2676
2677         req->rq_status = rc;
2678         RETURN(0);
2679
2680 }
2681
2682 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2683                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2684 {
2685         struct obd_device *obd = req->rq_export->exp_obd;
2686         struct dentry *de_srcdir = NULL;
2687         struct dentry *de_tgtdir = NULL;
2688         struct dentry *de_old = NULL;
2689         struct dentry *de_new = NULL;
2690         struct mds_obd *mds = mds_req2mds(req);
2691         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2692         struct mds_body *body = NULL;
2693         int rc = 0, lock_count = 3;
2694         int cleanup_phase = 0;
2695         void *handle = NULL;
2696         ENTRY;
2697
2698         LASSERT(offset == 0);
2699
2700         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2701                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2702                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2703
2704         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2705
2706         if (rec->ur_namelen == 1) {
2707                 rc = mds_reint_rename_create_name(rec, offset, req);
2708                 RETURN(rc);
2709         }
2710
2711         if (rec->ur_fid2->mds != mds->mds_num) {
2712                 rc = mds_reint_rename_to_remote(rec, offset, req);
2713                 RETURN(rc);
2714         }
2715
2716         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2717                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2718                                              rec->ur_name, rec->ur_namelen,
2719                                              &de_old, rec->ur_tgt,
2720                                              rec->ur_tgtlen, &de_new,
2721                                              dlm_handles, LCK_EX);
2722         if (rc)
2723                 GOTO(cleanup, rc);
2724
2725         cleanup_phase = 1; /* parent(s), children, locks */
2726
2727         if (de_new->d_inode)
2728                 lock_count = 4;
2729
2730         /* sanity check for src inode */
2731         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2732             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2733                 GOTO(cleanup, rc = -EINVAL);
2734
2735         /* sanity check for dest inode */
2736         if (de_new->d_inode &&
2737             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2738              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2739                 GOTO(cleanup, rc = -EINVAL);
2740
2741         if (de_old->d_inode == de_new->d_inode)
2742                 GOTO(cleanup, rc = 0);
2743
2744         /* if we are about to remove the target at first, pass the EA of
2745          * that inode to client to perform and cleanup on OST */
2746         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2747         LASSERT(body != NULL);
2748
2749         if (de_new->d_inode &&
2750             S_ISREG(de_new->d_inode->i_mode) &&
2751             de_new->d_inode->i_nlink == 1 &&
2752             mds_open_orphan_count(de_new->d_inode) == 0) {
2753                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2754                 mds_pack_inode2body(obd, body, de_new->d_inode);
2755                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2756                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2757                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2758                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2759                 } else {
2760                         /* XXX need log unlink? */
2761                 }
2762         }
2763
2764         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2765                        de_srcdir->d_inode->i_sb);
2766
2767         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2768         if (IS_ERR(handle))
2769                 GOTO(cleanup, rc = PTR_ERR(handle));
2770
2771         /* FIXME need adjust the journal block count? */
2772         /* if the target should be moved to PENDING, we at first increase the
2773          * link and later vfs_rename() will decrease the link count again */
2774         if (de_new->d_inode &&
2775             S_ISREG(de_new->d_inode->i_mode) &&
2776             de_new->d_inode->i_nlink == 1 &&
2777             mds_open_orphan_count(de_new->d_inode) > 0) {
2778                 rc = mds_add_link_orphan(rec, obd, de_new);
2779                 if (rc)
2780                         GOTO(cleanup, rc);
2781         }
2782
2783         lock_kernel();
2784         de_old->d_fsdata = req;
2785         de_new->d_fsdata = req;
2786         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2787         unlock_kernel();
2788
2789         GOTO(cleanup, rc);
2790 cleanup:
2791         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2792                                 handle, req, rc, 0);
2793         switch (cleanup_phase) {
2794         case 1:
2795 #ifdef S_PDIROPS
2796                 if (dlm_handles[5].cookie != 0)
2797                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2798                 if (dlm_handles[6].cookie != 0)
2799                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2800 #endif
2801                 if (rc) {
2802                         if (lock_count == 4)
2803                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2804                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2805                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2806                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2807                 } else {
2808                         if (lock_count == 4)
2809                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2810                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2811                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2812                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2813                 }
2814                 l_dput(de_new);
2815                 l_dput(de_old);
2816                 l_dput(de_tgtdir);
2817                 l_dput(de_srcdir);
2818         case 0:
2819                 break;
2820         default:
2821                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2822                 LBUG();
2823         }
2824         req->rq_status = rc;
2825         return 0;
2826 }
2827
2828 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2829                            struct ptlrpc_request *, struct lustre_handle *);
2830
2831 static mds_reinter reinters[REINT_MAX + 1] = {
2832         [REINT_SETATTR] mds_reint_setattr,
2833         [REINT_CREATE] mds_reint_create,
2834         [REINT_LINK] mds_reint_link,
2835         [REINT_UNLINK] mds_reint_unlink,
2836         [REINT_RENAME] mds_reint_rename,
2837         [REINT_OPEN] mds_open
2838 };
2839
2840 int mds_reint_rec(struct mds_update_record *rec, int offset,
2841                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2842 {
2843         struct obd_device *obd = req->rq_export->exp_obd;
2844         struct lvfs_run_ctxt saved;
2845         int rc;
2846
2847         /* checked by unpacker */
2848         LASSERT(rec->ur_opcode <= REINT_MAX &&
2849                 reinters[rec->ur_opcode] != NULL);
2850
2851         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2852         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2853         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2854
2855         return rc;
2856 }