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                 mds_pack_dentry2body(body, child);
 551                 body->valid |= OBD_MD_MDS;
 552         } else if (child->d_inode == NULL) {
 553                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 554                           rec->ur_fid1->id, rec->ur_fid1->generation,
 555                           rec->ur_name, rec->ur_mode);
 556                 LASSERTF(child->d_inode != NULL, "BUG 3869");
 557         } else {
 558                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 559                 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 560                 mds_pack_inode2body(req2obd(req), body, child->d_inode);
 561         }
 562         l_dput(parent);
 563         l_dput(child);
 564         EXIT;
 565 }
 566
 567 static int mds_reint_create(struct mds_update_record *rec, int offset,
 568                             struct ptlrpc_request *req,
 569                             struct lustre_handle *lh)
 570 {
 571         struct dentry *dparent = NULL;
 572         struct mds_obd *mds = mds_req2mds(req);
 573         struct obd_device *obd = req->rq_export->exp_obd;
 574         struct dentry *dchild = NULL;
 575         struct inode *dir = NULL;
 576         void *handle = NULL;
 577         struct lustre_handle lockh[2] = {{0}, {0}};
 578         int parent_mode;
 579         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 580         int created = 0;
 581         struct dentry_params dp;
 582         struct mea *mea = NULL;
 583         int mea_size;
 584         ENTRY;
 585
 586         LASSERT(offset == 0);
 587         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 588
 589         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 590                   rec->ur_fid1->id, rec->ur_fid1->generation,
 591                   rec->ur_name, rec->ur_mode);
 592
 593         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 594
 595         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 596                 GOTO(cleanup, rc = -ESTALE);
 597
 598         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 599                                         &parent_mode, rec->ur_name,
 600                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 601         if (IS_ERR(dparent)) {
 602                 rc = PTR_ERR(dparent);
 603                 CERROR("parent lookup error %d\n", rc);
 604                 GOTO(cleanup, rc);
 605         }
 606         cleanup_phase = 1; /* locked parent dentry */
 607         dir = dparent->d_inode;
 608         LASSERT(dir);
 609
 610         ldlm_lock_dump_handle(D_OTHER, lockh);
 611
 612         /* try to retrieve MEA data for this dir */
 613         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 614         if (rc)
 615                 GOTO(cleanup, rc);
 616
 617         if (mea != NULL) {
 618                 /* dir is already splitted, check is requested filename
 619                  * should live at this MDS or at another one */
 620                 int i;
 621                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 622                 if (mea->mea_master != mea->mea_fids[i].mds) {
 623                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 624                                " should be %d(%d)\n",
 625                                mea->mea_master, dparent->d_inode->i_ino,
 626                                dparent->d_inode->i_generation, rec->ur_name,
 627                                mea->mea_fids[i].mds, i);
 628                         GOTO(cleanup, rc = -ERESTART);
 629                 }
 630         }
 631
 632         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 633         if (IS_ERR(dchild)) {
 634                 rc = PTR_ERR(dchild);
 635                 CERROR("child lookup error %d\n", rc);
 636                 GOTO(cleanup, rc);
 637         }
 638
 639         cleanup_phase = 2; /* child dentry */
 640
 641         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 642
 643         if (type == S_IFREG || type == S_IFDIR) {
 644                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 645                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 646                        obd->obd_name, dparent->d_inode->i_ino,
 647                        dparent->d_inode->i_generation, rc, parent_mode);
 648                 if (rc > 0) {
 649                         /* dir got splitted */
 650                         GOTO(cleanup, rc = -ERESTART);
 651                 } else if (rc < 0) {
 652                         /* error happened during spitting. */
 653                         GOTO(cleanup, rc);
 654                 }
 655         }
 656
 657         if (dir->i_mode & S_ISGID) {
 658                 if (S_ISDIR(rec->ur_mode))
 659                         rec->ur_mode |= S_ISGID;
 660         }
 661
 662         dchild->d_fsdata = (void *)&dp;
 663         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 664         dp.p_ptr = req;
 665
 666         switch (type) {
 667         case S_IFREG:{
 668                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 669                 if (IS_ERR(handle))
 670                         GOTO(cleanup, rc = PTR_ERR(handle));
 671                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 672                 EXIT;
 673                 break;
 674         }
 675         case S_IFDIR:{
 676                 int nstripes = 0;
 677                 int i;
 678
 679                 /* as Peter asked, mkdir() should distribute new directories
 680                  * over the whole cluster in order to distribute namespace
 681                  * processing load. first, we calculate which MDS to use to put
 682                  * new directory's inode in. */
 683                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 684                                       rec->ur_flags);
 685                 if (i == mds->mds_num) {
 686                         /* inode will be created locally */
 687                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 688                         if (IS_ERR(handle))
 689                                 GOTO(cleanup, rc = PTR_ERR(handle));
 690
 691                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 692
 693                         if (rec->ur_eadata)
 694                                 nstripes = *(u16 *)rec->ur_eadata;
 695
 696                         if (rc == 0 && nstripes) {
 697                                 /* we pass LCK_EX to split routine to
 698                                  * signalthat we have exclusive access
 699                                  * to the directory. simple because
 700                                  * nobody knows it already exists -bzzz */
 701                                 rc = mds_try_to_split_dir(obd, dchild,
 702                                                           NULL, nstripes,
 703                                                           LCK_EX);
 704                                 if (rc > 0) {
 705                                         /* dir got splitted */
 706                                         rc = 0;
 707                                 } else if (rc < 0) {
 708                                         /* an error occured during
 709                                          * splitting. */
 710                                         GOTO(cleanup, rc);
 711                                 }
 712                         }
 713                 } else if (!DENTRY_VALID(dchild)) {
 714                         /* inode will be created on another MDS */
 715                         struct obdo *oa = NULL;
 716                         struct mds_body *body;
 717
 718                         /* first, create that inode */
 719                         oa = obdo_alloc();
 720                         LASSERT(oa != NULL);
 721                         oa->o_mds = i;
 722                         oa->o_easize = 0;
 723                         if (rec->ur_eadata) {
 724                                 /* user asks for creating splitted dir */
 725                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 726                         }
 727
 728                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 729                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 730                                         OBD_MD_FLUID | OBD_MD_FLGID);
 731                         oa->o_mode = dir->i_mode;
 732                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 733                                         obd->obd_name, i);
 734                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 735                                 CWARN("%s: replay dir creation %*s -> %u/%u\n",
 736                                       obd->obd_name, rec->ur_namelen - 1,
 737                                       rec->ur_name, (unsigned) rec->ur_fid2->id,
 738                                       (unsigned) rec->ur_fid2->generation);
 739                                 oa->o_id = rec->ur_fid2->id;
 740                                 oa->o_generation = rec->ur_fid2->generation;
 741                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 742                         }
 743
 744                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 745                         if (rc) {
 746                                 obdo_free(oa);
 747                                 GOTO(cleanup, rc);
 748                         }
 749
 750                         /* now, add new dir entry for it */
 751                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 752                         if (IS_ERR(handle)) {
 753                                 obdo_free(oa);
 754                                 GOTO(cleanup, rc = PTR_ERR(handle));
 755                         }
 756                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 757                                                   rec->ur_namelen - 1,
 758                                                   oa->o_id, oa->o_generation,
 759                                                   i);
 760                         LASSERT(rc == 0);
 761
 762                         /* fill reply */
 763                         body = lustre_msg_buf(req->rq_repmsg,
 764                                               offset, sizeof (*body));
 765                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 766                         body->fid1.id = oa->o_id;
 767                         body->fid1.mds = i;
 768                         body->fid1.generation = oa->o_generation;
 769                         obdo_free(oa);
 770                 } else {
 771                         /* requested name exists in the directory */
 772                         rc = -EEXIST;
 773                 }
 774                 EXIT;
 775                 break;
 776         }
 777         case S_IFLNK:{
 778                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 779                 if (IS_ERR(handle))
 780                         GOTO(cleanup, rc = PTR_ERR(handle));
 781                 if (rec->ur_tgt == NULL)        /* no target supplied */
 782                         rc = -EINVAL;           /* -EPROTO? */
 783                 else
 784                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 785                 EXIT;
 786                 break;
 787         }
 788         case S_IFCHR:
 789         case S_IFBLK:
 790         case S_IFIFO:
 791         case S_IFSOCK:{
 792                 int rdev = rec->ur_rdev;
 793                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 794                 if (IS_ERR(handle))
 795                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 796                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 797                 EXIT;
 798                 break;
 799         }
 800         default:
 801                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 802                 dchild->d_fsdata = NULL;
 803                 GOTO(cleanup, rc = -EINVAL);
 804         }
 805
 806         /* In case we stored the desired inum in here, we want to clean up. */
 807         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 808                 dchild->d_fsdata = NULL;
 809
 810         if (rc) {
 811                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 812                 GOTO(cleanup, rc);
 813         } else if (dchild->d_inode) {
 814                 struct iattr iattr;
 815                 struct inode *inode = dchild->d_inode;
 816                 struct mds_body *body;
 817
 818                 created = 1;
 819                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 820                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 821                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 822                 iattr.ia_uid = rec->_ur_fsuid;
 823                 if (dir->i_mode & S_ISGID)
 824                         iattr.ia_gid = dir->i_gid;
 825                 else
 826                         iattr.ia_gid = rec->_ur_fsgid;
 827                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 828                         ATTR_MTIME | ATTR_CTIME;
 829
 830                 if (rec->ur_fid2->id) {
 831                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 832                         inode->i_generation = rec->ur_fid2->generation;
 833                         /* Dirtied and committed by the upcoming setattr. */
 834                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 835                                inode->i_ino, inode->i_generation);
 836                 } else {
 837 #if 0
 838                         struct lustre_handle child_ino_lockh;
 839 #endif
 840
 841                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 842                                inode->i_ino, inode->i_generation);
 843
 844 #if 0
 845                         /* The inode we were allocated may have just been freed
 846                          * by an unlink operation.  We take this lock to
 847                          * synchronize against the matching reply-ack-lock taken
 848                          * in unlink, to avoid replay problems if this reply
 849                          * makes it out to the client but the unlink's does not.
 850                          * See bug 2029 for more detail.*/
 851                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 852                         if (rc != ELDLM_OK) {
 853                                 CERROR("error locking for unlink/create sync: "
 854                                        "%d\n", rc);
 855                         } else {
 856                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 857                         }
 858 #endif
 859                 }
 860
 861                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 862                 if (rc)
 863                         CERROR("error on child setattr: rc = %d\n", rc);
 864
 865                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 866                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 867                 if (rc)
 868                         CERROR("error on parent setattr: rc = %d\n", rc);
 869
 870                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 871                 mds_pack_inode2fid(obd, &body->fid1, inode);
 872                 mds_pack_inode2body(obd, body, inode);
 873         }
 874         EXIT;
 875
 876 cleanup:
 877         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 878
 879         if (rc && created) {
 880                 /* Destroy the file we just created. This should not need extra
 881                  * journal credits, as we have already modified all of the
 882                  * blocks needed in order to create the file in the first
 883                  * place. */
 884                 switch (type) {
 885                 case S_IFDIR:
 886                         err = vfs_rmdir(dir, dchild);
 887                         if (err)
 888                                 CERROR("rmdir in error path: %d\n", err);
 889                         break;
 890                 default:
 891                         err = vfs_unlink(dir, dchild);
 892                         if (err)
 893                                 CERROR("unlink in error path: %d\n", err);
 894                         break;
 895                 }
 896         } else {
 897                 rc = err;
 898         }
 899         switch (cleanup_phase) {
 900         case 2: /* child dentry */
 901                 l_dput(dchild);
 902         case 1: /* locked parent dentry */
 903 #ifdef S_PDIROPS
 904                 if (lockh[1].cookie != 0)
 905                         ldlm_lock_decref(lockh + 1, parent_mode);
 906 #endif
 907                 if (rc) {
 908                         ldlm_lock_decref(lockh, LCK_PW);
 909                 } else {
 910                         ptlrpc_save_lock(req, lockh, LCK_PW);
 911                 }
 912                 l_dput(dparent);
 913         case 0:
 914                 break;
 915         default:
 916                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 917                 LBUG();
 918         }
 919         if (mea)
 920                 OBD_FREE(mea, mea_size);
 921         req->rq_status = rc;
 922         return 0;
 923 }
 924
 925 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 926            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 927 {
 928         int i;
 929
 930         for (i = 0; i < RES_NAME_SIZE; i++) {
 931                 /* return 1 here, because enqueue_ordered will skip resources
 932                  * of all zeroes if they're sorted to the end of the list. */
 933                 if (res1->name[i] == 0 && res2->name[i] != 0)
 934                         return 1;
 935                 if (res2->name[i] == 0 && res1->name[i] != 0)
 936                         return 0;
 937
 938                 if (res1->name[i] > res2->name[i])
 939                         return 1;
 940                 if (res1->name[i] < res2->name[i])
 941                         return 0;
 942         }
 943
 944         if (!p1 || !p2)
 945                 return 0;
 946
 947         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 948                 return 1;
 949
 950         return 0;
 951 }
 952
 953 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 954  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 955  * because they take the place of local semaphores.
 956  *
 957  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 958  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 959 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 960                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 961                           ldlm_policy_data_t *p1_policy,
 962                           struct ldlm_res_id *p2_res_id,
 963                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 964                           ldlm_policy_data_t *p2_policy)
 965 {
 966         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 967         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 968         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 969         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 970         int rc, flags;
 971         ENTRY;
 972
 973         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 974
 975         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 976                res_id[1]->name[0]);
 977
 978         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 979                 handles[1] = p1_lockh;
 980                 handles[0] = p2_lockh;
 981                 res_id[1] = p1_res_id;
 982                 res_id[0] = p2_res_id;
 983                 lock_modes[1] = p1_lock_mode;
 984                 lock_modes[0] = p2_lock_mode;
 985                 policies[1] = p1_policy;
 986                 policies[0] = p2_policy;
 987         }
 988
 989         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 990                res_id[0]->name[0], res_id[1]->name[0]);
 991
 992         flags = LDLM_FL_LOCAL_ONLY;
 993         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 994                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 995                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 996                               NULL, 0, NULL, handles[0]);
 997         if (rc != ELDLM_OK)
 998                 RETURN(-EIO);
 999         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1000
1001         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1002             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1003                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1004                 ldlm_lock_addref(handles[1], lock_modes[1]);
1005         } else if (res_id[1]->name[0] != 0) {
1006                 flags = LDLM_FL_LOCAL_ONLY;
1007                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1008                                       *res_id[1], LDLM_IBITS, policies[1],
1009                                       lock_modes[1], &flags, mds_blocking_ast,
1010                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1011                                       NULL, handles[1]);
1012                 if (rc != ELDLM_OK) {
1013                         ldlm_lock_decref(handles[0], lock_modes[0]);
1014                         RETURN(-EIO);
1015                 }
1016                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1017         }
1018
1019         RETURN(0);
1020 }
1021
1022 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1023                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1024                            ldlm_policy_data_t *p1_policy,
1025                            struct ldlm_res_id *p2_res_id,
1026                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1027                            ldlm_policy_data_t *p2_policy,
1028                            struct ldlm_res_id *c1_res_id,
1029                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1030                            ldlm_policy_data_t *c1_policy,
1031                            struct ldlm_res_id *c2_res_id,
1032                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1033                            ldlm_policy_data_t *c2_policy)
1034 {
1035         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1036                                           c1_res_id, c2_res_id };
1037         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1038                                                  c1_lockh, c2_lockh };
1039         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1040                               c1_lock_mode, c2_lock_mode };
1041         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1042                                             c1_policy, c2_policy};
1043         int rc, i, j, sorted, flags;
1044         ENTRY;
1045
1046         CDEBUG(D_DLMTRACE,
1047                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1048                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1049                res_id[3]->name[0]);
1050
1051         /* simple insertion sort - we have at most 4 elements */
1052         for (i = 1; i < 4; i++) {
1053                 j = i - 1;
1054                 dlm_handles[4] = dlm_handles[i];
1055                 res_id[4] = res_id[i];
1056                 lock_modes[4] = lock_modes[i];
1057                 policies[4] = policies[i];
1058
1059                 sorted = 0;
1060                 do {
1061                         if (res_gt(res_id[j], res_id[4], policies[j],
1062                                    policies[4])) {
1063                                 dlm_handles[j + 1] = dlm_handles[j];
1064                                 res_id[j + 1] = res_id[j];
1065                                 lock_modes[j + 1] = lock_modes[j];
1066                                 policies[j + 1] = policies[j];
1067                                 j--;
1068                         } else {
1069                                 sorted = 1;
1070                         }
1071                 } while (j >= 0 && !sorted);
1072
1073                 dlm_handles[j + 1] = dlm_handles[4];
1074                 res_id[j + 1] = res_id[4];
1075                 lock_modes[j + 1] = lock_modes[4];
1076                 policies[j + 1] = policies[4];
1077         }
1078
1079         CDEBUG(D_DLMTRACE,
1080                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1081                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1082                res_id[3]->name[0]);
1083
1084         /* XXX we could send ASTs on all these locks first before blocking? */
1085         for (i = 0; i < 4; i++) {
1086                 flags = 0;
1087                 if (res_id[i]->name[0] == 0)
1088                         break;
1089                 if (i != 0 &&
1090                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1091                     (policies[i]->l_inodebits.bits &
1092                      policies[i-1]->l_inodebits.bits) ) {
1093                         memcpy(dlm_handles[i], dlm_handles[i-1],
1094                                sizeof(*(dlm_handles[i])));
1095                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1096                 } else {
1097                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1098                                               *res_id[i], LDLM_IBITS,
1099                                               policies[i],
1100                                               lock_modes[i], &flags,
1101                                               mds_blocking_ast,
1102                                               ldlm_completion_ast, NULL, NULL,
1103                                               NULL, 0, NULL, dlm_handles[i]);
1104                         if (rc != ELDLM_OK)
1105                                 GOTO(out_err, rc = -EIO);
1106                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1107                 }
1108         }
1109
1110         RETURN(0);
1111 out_err:
1112         while (i-- > 0)
1113                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1114
1115         return rc;
1116 }
1117
1118 /* In the unlikely case that the child changed while we were waiting
1119  * on the lock, we need to drop the lock on the old child and either:
1120  * - if the child has a lower resource name, then we have to also
1121  *   drop the parent lock and regain the locks in the right order
1122  * - in the rename case, if the child has a lower resource name than one of
1123  *   the other parent/child resources (maxres) we also need to reget the locks
1124  * - if the child has a higher resource name (this is the common case)
1125  *   we can just get the lock on the new child (still in lock order)
1126  *
1127  * Returns 0 if the child did not change or if it changed but could be locked.
1128  * Returns 1 if the child changed and we need to re-lock (no locks held).
1129  * Returns -ve error with a valid dchild (no locks held). */
1130 static int mds_verify_child(struct obd_device *obd,
1131                             struct ldlm_res_id *parent_res_id,
1132                             struct lustre_handle *parent_lockh,
1133                             struct dentry *dparent, int parent_mode,
1134                             struct ldlm_res_id *child_res_id,
1135                             struct lustre_handle *child_lockh,
1136                             struct dentry **dchildp, int child_mode,
1137                             ldlm_policy_data_t *child_policy,
1138                             const char *name, int namelen,
1139                             struct ldlm_res_id *maxres)
1140 {
1141         struct dentry *vchild, *dchild = *dchildp;
1142         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1143         ENTRY;
1144
1145         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1146         if (IS_ERR(vchild))
1147                 GOTO(cleanup, rc = PTR_ERR(vchild));
1148
1149         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1150                 if  (child_res_id->name[0] == vchild->d_inum &&
1151                                 child_res_id->name[1] == vchild->d_generation) {
1152                         if (dchild != NULL)
1153                                 l_dput(dchild);
1154                         *dchildp = vchild;
1155                         RETURN(0);
1156                 }
1157                 goto changed;
1158         }
1159
1160         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1161                    (vchild->d_inode != NULL &&
1162                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1163                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1164                 if (dchild != NULL)
1165                         l_dput(dchild);
1166                 *dchildp = vchild;
1167
1168                 RETURN(0);
1169         }
1170
1171 changed:
1172         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1173                vchild->d_inode, dchild ? dchild->d_inode : 0,
1174                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1175                child_res_id->name[0]);
1176         if (child_res_id->name[0] != 0)
1177                 ldlm_lock_decref(child_lockh, child_mode);
1178         if (dchild)
1179                 l_dput(dchild);
1180
1181         cleanup_phase = 1; /* parent lock only */
1182         *dchildp = dchild = vchild;
1183
1184         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1185                 int flags = 0;
1186                 if (dchild->d_inode) {
1187                         child_res_id->name[0] = dchild->d_inode->i_ino;
1188                         child_res_id->name[1] = dchild->d_inode->i_generation;
1189                 } else {
1190                         child_res_id->name[0] = dchild->d_inum;
1191                         child_res_id->name[1] = dchild->d_generation;
1192                 }
1193
1194                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1195                     res_gt(maxres, child_res_id, NULL, NULL)) {
1196                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1197                                child_res_id->name[0], parent_res_id->name[0],
1198                                maxres->name[0]);
1199                         GOTO(cleanup, rc = 1);
1200                 }
1201
1202                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1203                                       *child_res_id, LDLM_IBITS, child_policy,
1204                                       child_mode, &flags, mds_blocking_ast,
1205                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1206                                       NULL, child_lockh);
1207                 if (rc != ELDLM_OK)
1208                         GOTO(cleanup, rc = -EIO);
1209
1210         } else {
1211                 memset(child_res_id, 0, sizeof(*child_res_id));
1212         }
1213
1214         EXIT;
1215 cleanup:
1216         if (rc) {
1217                 switch(cleanup_phase) {
1218                 case 2:
1219                         if (child_res_id->name[0] != 0)
1220                                 ldlm_lock_decref(child_lockh, child_mode);
1221                 case 1:
1222                         ldlm_lock_decref(parent_lockh, parent_mode);
1223                 }
1224         }
1225         return rc;
1226 }
1227
1228 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1229                                 struct ll_fid *fid,
1230                                 struct lustre_handle *parent_lockh,
1231                                 struct dentry **dparentp, int parent_mode,
1232                                 __u64 parent_lockpart, int *update_mode,
1233                                 char *name, int namelen,
1234                                 struct lustre_handle *child_lockh,
1235                                 struct dentry **dchildp, int child_mode,
1236                                 __u64 child_lockpart, void *clone_info)
1237 {
1238         struct ldlm_res_id child_res_id = { .name = {0} };
1239         struct ldlm_res_id parent_res_id = { .name = {0} };
1240         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1241         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1242         struct inode *inode;
1243         int rc = 0, cleanup_phase = 0;
1244         ENTRY;
1245
1246         /* Step 1: Lookup parent */
1247         *dparentp = mds_fid2dentry(mds, fid, NULL);
1248         if (IS_ERR(*dparentp)) {
1249                 rc = PTR_ERR(*dparentp);
1250                 *dparentp = NULL;
1251                 RETURN(rc);
1252         }
1253
1254         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1255                (*dparentp)->d_inode->i_ino, name);
1256
1257         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1258         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1259 #ifdef S_PDIROPS
1260         parent_lockh[1].cookie = 0;
1261         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1262                 struct ldlm_res_id res_id = { .name = {0} };
1263                 ldlm_policy_data_t policy;
1264                 int flags = 0;
1265                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1266                 if (*update_mode) {
1267                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1268                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1269                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1270                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1271                                               res_id, LDLM_IBITS, &policy,
1272                                               *update_mode, &flags,
1273                                               mds_blocking_ast,
1274                                               ldlm_completion_ast,
1275                                               NULL, NULL, NULL, 0, NULL,
1276                                               parent_lockh + 1);
1277                         if (rc != ELDLM_OK)
1278                                 RETURN(-ENOLCK);
1279                 }
1280
1281                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1282                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1283                        (*dparentp)->d_inode->i_ino,
1284                        (*dparentp)->d_inode->i_generation,
1285                        parent_res_id.name[2]);
1286         }
1287 #endif
1288
1289         cleanup_phase = 1; /* parent dentry */
1290 #ifdef CONFIG_SNAPFS
1291         if (clone_info) {
1292                 /*FIXME is there any other FUNC will use d_fsdata,
1293                  *excepet creating inode according inum*/
1294                 (*dparentp)->d_fsdata = clone_info;
1295         }
1296 #endif
1297
1298         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1299         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1300         if (IS_ERR(*dchildp)) {
1301                 rc = PTR_ERR(*dchildp);
1302                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1303                 GOTO(cleanup, rc);
1304         }
1305
1306         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1307                 /* inode lives on another MDS: return * mds/ino/gen
1308                  * and LOOKUP lock. drop possible UPDATE lock! */
1309                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1310                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1311                 child_res_id.name[0] = (*dchildp)->d_inum;
1312                 child_res_id.name[1] = (*dchildp)->d_generation;
1313                 goto retry_locks;
1314         }
1315
1316         inode = (*dchildp)->d_inode;
1317         if (inode != NULL)
1318                 inode = igrab(inode);
1319         if (inode == NULL)
1320                 goto retry_locks;
1321
1322         child_res_id.name[0] = inode->i_ino;
1323         child_res_id.name[1] = inode->i_generation;
1324
1325         iput(inode);
1326
1327 retry_locks:
1328         cleanup_phase = 2; /* child dentry */
1329
1330         /* Step 3: Lock parent and child in resource order.  If child doesn't
1331          *         exist, we still have to lock the parent and re-lookup. */
1332         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1333                                    &parent_policy, &child_res_id, child_lockh,
1334                                    child_mode, &child_policy);
1335         if (rc)
1336                 GOTO(cleanup, rc);
1337
1338         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1339                 cleanup_phase = 4; /* child lock */
1340         else
1341                 cleanup_phase = 3; /* parent lock */
1342
1343         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1344         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1345                               parent_mode, &child_res_id, child_lockh,
1346                               dchildp, child_mode, &child_policy,
1347                               name, namelen, &parent_res_id);
1348         if (rc > 0)
1349                 goto retry_locks;
1350         if (rc < 0) {
1351                 cleanup_phase = 3;
1352                 GOTO(cleanup, rc);
1353         }
1354
1355 cleanup:
1356         if (rc) {
1357                 switch (cleanup_phase) {
1358                 case 4:
1359                         ldlm_lock_decref(child_lockh, child_mode);
1360                 case 3:
1361                         ldlm_lock_decref(parent_lockh, parent_mode);
1362                 case 2:
1363                         l_dput(*dchildp);
1364                 case 1:
1365 #ifdef S_PDIROPS
1366                         if (parent_lockh[1].cookie)
1367                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1368 #endif
1369                         l_dput(*dparentp);
1370                 default: ;
1371                 }
1372         }
1373         return rc;
1374 }
1375
1376 void mds_reconstruct_generic(struct ptlrpc_request *req)
1377 {
1378         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1379
1380         mds_req_from_mcd(req, med->med_mcd);
1381 }
1382
1383 int mds_create_local_dentry(struct mds_update_record *rec,
1384                            struct obd_device *obd)
1385 {
1386         struct mds_obd *mds = &obd->u.mds;
1387         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1388         int fidlen = 0, rc, cleanup_phase = 0;
1389         struct dentry *new_child = NULL;
1390         char *fidname = rec->ur_name;
1391         struct dentry *child = NULL;
1392         struct lustre_handle lockh[2] = {{0}, {0}};
1393         void *handle;
1394         ENTRY;
1395
1396         down(&fids_dir->i_sem);
1397         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1398         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1399                         fidname, (unsigned) rec->ur_fid1->id,
1400                         (unsigned) rec->ur_fid1->generation);
1401
1402         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1403         up(&fids_dir->i_sem);
1404         if (IS_ERR(new_child)) {
1405                 CERROR("can't lookup %s: %d\n", fidname,
1406                                 (int) PTR_ERR(new_child));
1407                 GOTO(cleanup, rc = PTR_ERR(new_child));
1408         }
1409         cleanup_phase = 1;
1410
1411         if (new_child->d_inode != NULL) {
1412                 /* nice. we've already have local dentry! */
1413                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1414                        (unsigned) new_child->d_inode->i_ino,
1415                        (unsigned) new_child->d_inode->i_generation);
1416                 rec->ur_fid1->id = fids_dir->i_ino;
1417                 rec->ur_fid1->generation = fids_dir->i_generation;
1418                 rec->ur_namelen = fidlen + 1;
1419                 GOTO(cleanup, rc = 0);
1420         }
1421
1422         /* new, local dentry will be added soon. we need no aliases here */
1423         d_drop(new_child);
1424
1425         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1426                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1427         } else {
1428                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1429                                               LCK_EX, lockh, NULL, NULL, 0,
1430                                               MDS_INODELOCK_UPDATE);
1431         }
1432
1433         if (IS_ERR(child)) {
1434                 CERROR("can't get victim: %d\n", (int) PTR_ERR(child));
1435                 GOTO(cleanup, rc = PTR_ERR(child));
1436         }
1437         cleanup_phase = 2;
1438
1439         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1440         if (IS_ERR(handle))
1441                 GOTO(cleanup, rc = PTR_ERR(handle));
1442
1443         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1444                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1445                                   mds->mds_num);
1446         if (rc)
1447                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1448                        (unsigned long) child->d_inode->i_ino,
1449                        (unsigned long) child->d_inode->i_generation, rc);
1450         else {
1451                 if (S_ISDIR(child->d_inode->i_mode)) {
1452                         fids_dir->i_nlink++;
1453                         mark_inode_dirty(fids_dir);
1454                 }
1455                 mark_inode_dirty(child->d_inode);
1456         }
1457         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1458
1459         rec->ur_fid1->id = fids_dir->i_ino;
1460         rec->ur_fid1->generation = fids_dir->i_generation;
1461         rec->ur_namelen = fidlen + 1;
1462
1463 cleanup:
1464         switch(cleanup_phase) {
1465                 case 2:
1466                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1467                                 ldlm_lock_decref(lockh, LCK_EX);
1468                         dput(child);
1469                 case 1:
1470                         dput(new_child);
1471                 case 0:
1472                        break;
1473         }
1474         RETURN(rc);
1475 }
1476
1477 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1478                                         struct ptlrpc_request *slave)
1479 {
1480         void *cookie, *cookie2;
1481         struct mds_body *body2;
1482         struct mds_body *body;
1483         void *ea, *ea2;
1484         ENTRY;
1485
1486         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1487         LASSERT(body != NULL);
1488
1489         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1490         LASSERT(body2 != NULL);
1491
1492         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1493                 RETURN(0);
1494         }
1495
1496         memcpy(body2, body, sizeof(*body));
1497         body2->valid &= ~OBD_MD_FLCOOKIE;
1498
1499         if (!(body->valid & OBD_MD_FLEASIZE) &&
1500             !(body->valid & OBD_MD_FLDIREA))
1501                 RETURN(0);
1502
1503         if (body->eadatasize == 0) {
1504                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1505                 RETURN(0);
1506         }
1507
1508         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1509
1510         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1511         LASSERT(ea != NULL);
1512
1513         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1514         LASSERT(ea2 != NULL);
1515
1516         memcpy(ea2, ea, body->eadatasize);
1517
1518         if (body->valid & OBD_MD_FLCOOKIE) {
1519                 LASSERT(master->rq_repmsg->buflens[2] >=
1520                                 slave->rq_repmsg->buflens[2]);
1521                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1522                                 slave->rq_repmsg->buflens[2]);
1523                 LASSERT(cookie != NULL);
1524
1525                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1526                                 master->rq_repmsg->buflens[2]);
1527                 LASSERT(cookie2 != NULL);
1528                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1529                 body2->valid |= OBD_MD_FLCOOKIE;
1530         }
1531         RETURN(0);
1532 }
1533
1534 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1535                                    struct ptlrpc_request *req,
1536                                    struct lustre_handle *parent_lockh,
1537                                    int update_mode,
1538                                    struct dentry *dparent,
1539                                    struct lustre_handle *child_lockh,
1540                                    struct dentry *dchild)
1541 {
1542         struct obd_device *obd = req->rq_export->exp_obd;
1543         struct mds_obd *mds = mds_req2mds(req);
1544         struct mdc_op_data op_data;
1545         int rc = 0, cleanup_phase = 0;
1546         struct ptlrpc_request *request = NULL;
1547         void *handle;
1548         ENTRY;
1549
1550         LASSERT(offset == 0 || offset == 2);
1551
1552         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1553                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1554                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1555         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1556                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1557                           rec->ur_namelen - 1, rec->ur_name,
1558                           (unsigned)dchild->d_mdsnum,
1559                           (unsigned) dchild->d_inum,
1560                           (unsigned) dchild->d_generation);
1561
1562         /* time to drop i_nlink on remote MDS */
1563         memset(&op_data, 0, sizeof(op_data));
1564         op_data.fid1.mds = dchild->d_mdsnum;
1565         op_data.fid1.id = dchild->d_inum;
1566         op_data.fid1.generation = dchild->d_generation;
1567         op_data.create_mode = rec->ur_mode;
1568         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1569                 op_data.create_mode |= MDS_MODE_REPLAY;
1570         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1571         cleanup_phase = 2;
1572         if (request) {
1573                 if (rc == 0)
1574                         mds_copy_unlink_reply(req, request);
1575                 ptlrpc_req_finished(request);
1576         }
1577         if (rc == 0) {
1578                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1579                                       NULL);
1580                 if (IS_ERR(handle))
1581                         GOTO(cleanup, rc = PTR_ERR(handle));
1582                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1583                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1584                                         rc, 0);
1585         }
1586 cleanup:
1587         req->rq_status = rc;
1588
1589 #ifdef S_PDIROPS
1590         if (parent_lockh[1].cookie != 0)
1591                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1592 #endif
1593         ldlm_lock_decref(child_lockh, LCK_EX);
1594         if (rc)
1595                 ldlm_lock_decref(parent_lockh, LCK_PW);
1596         else
1597                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1598         l_dput(dchild);
1599         l_dput(dparent);
1600
1601         return 0;
1602 }
1603
1604 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1605                             struct ptlrpc_request *req,
1606                             struct lustre_handle *lh)
1607 {
1608         struct dentry *dparent, *dchild;
1609         struct mds_obd *mds = mds_req2mds(req);
1610         struct obd_device *obd = req->rq_export->exp_obd;
1611         struct mds_body *body = NULL;
1612         struct inode *child_inode;
1613         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1614         struct lustre_handle child_lockh = {0};
1615 #if 0
1616         struct lustre_handle child_reuse_lockh = {0};
1617 #endif
1618         struct lustre_handle * slave_lockh = NULL;
1619         char fidname[LL_FID_NAMELEN];
1620         void *handle = NULL;
1621         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1622         int unlink_by_fid = 0;
1623         int update_mode;
1624         ENTRY;
1625
1626         LASSERT(offset == 0 || offset == 2);
1627
1628         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1629                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1630
1631         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1632
1633         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1634                 GOTO(cleanup, rc = -ENOENT);
1635
1636         if (rec->ur_namelen == 1) {
1637                 /* this is request to drop i_nlink on local inode */
1638                 unlink_by_fid = 1;
1639                 rec->ur_name = fidname;
1640                 rc = mds_create_local_dentry(rec, obd);
1641                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1642                         DEBUG_REQ(D_HA, req,
1643                                   "drop nlink on inode %u/%u/%u (replay)",
1644                                   (unsigned) rec->ur_fid1->mds,
1645                                   (unsigned) rec->ur_fid1->id,
1646                                   (unsigned) rec->ur_fid1->generation);
1647                         req->rq_status = 0;
1648                         RETURN(0);
1649                 }
1650         }
1651
1652         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1653                 /* master mds for directory asks slave removing
1654                  * inode is already locked */
1655                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1656                                                LCK_PW, parent_lockh,
1657                                                &update_mode, rec->ur_name,
1658                                                rec->ur_namelen,
1659                                                MDS_INODELOCK_UPDATE);
1660                 if (IS_ERR(dparent))
1661                         GOTO(cleanup, rc = PTR_ERR(dparent));
1662                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1663                                            rec->ur_namelen - 1);
1664                 if (IS_ERR(dchild))
1665                         GOTO(cleanup, rc = PTR_ERR(dchild));
1666                 child_lockh.cookie = 0;
1667                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1668                 LASSERT(dchild->d_inode != NULL);
1669                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1670         } else {
1671                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1672                                                  parent_lockh, &dparent,
1673                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1674                                                  &update_mode, rec->ur_name,
1675                                                  rec->ur_namelen, &child_lockh,
1676                                                  &dchild, LCK_EX,
1677                                                  MDS_INODELOCK_LOOKUP |
1678                                                    MDS_INODELOCK_UPDATE, NULL);
1679         }
1680         if (rc)
1681                 GOTO(cleanup, rc);
1682
1683         if (dchild->d_flags & DCACHE_CROSS_REF) {
1684                 /* we should have parent lock only here */
1685                 LASSERT(unlink_by_fid == 0);
1686                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1687                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1688                                         update_mode, dparent, &child_lockh, dchild);
1689                 RETURN(0);
1690         }
1691
1692         cleanup_phase = 1; /* dchild, dparent, locks */
1693
1694         dget(dchild);
1695         child_inode = dchild->d_inode;
1696         if (child_inode == NULL) {
1697                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1698                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1699                 GOTO(cleanup, rc = -ENOENT);
1700         }
1701
1702         cleanup_phase = 2; /* dchild has a lock */
1703
1704         /* We have to do these checks ourselves, in case we are making an
1705          * orphan.  The client tells us whether rmdir() or unlink() was called,
1706          * so we need to return appropriate errors (bug 72).
1707          *
1708          * We don't have to check permissions, because vfs_rename (called from
1709          * mds_open_unlink_rename) also calls may_delete. */
1710         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1711                 if (!S_ISDIR(child_inode->i_mode))
1712                         GOTO(cleanup, rc = -ENOTDIR);
1713         } else {
1714                 if (S_ISDIR(child_inode->i_mode))
1715                         GOTO(cleanup, rc = -EISDIR);
1716         }
1717
1718         /* handle splitted dir */
1719         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1720         if (rc)
1721                 GOTO(cleanup, rc);
1722
1723 #if 0
1724         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1725          * reuse (see bug 2029). */
1726         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1727         if (rc != ELDLM_OK)
1728                 GOTO(cleanup, rc);
1729 #endif
1730         cleanup_phase = 3; /* child inum lock */
1731
1732         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1733
1734         /* ldlm_reply in buf[0] if called via intent */
1735         if (offset)
1736                 offset = 1;
1737
1738         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1739         LASSERT(body != NULL);
1740
1741         /* If this is the last reference to this inode, get the OBD EA
1742          * data first so the client can destroy OST objects.
1743          * we only do the object removal if no open files remain.
1744          * Nobody can get at this name anymore because of the locks so
1745          * we make decisions here as to whether to remove the inode */
1746         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1747             mds_open_orphan_count(child_inode) == 0) {
1748                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1749                 mds_pack_inode2body(obd, body, child_inode);
1750                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1751                             child_inode, 1);
1752                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1753                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1754                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1755                 } else {
1756                         log_unlink = 1;
1757                 }
1758         }
1759
1760         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1761         switch (child_inode->i_mode & S_IFMT) {
1762         case S_IFDIR:
1763                 /* Drop any lingering child directories before we start our
1764                  * transaction, to avoid doing multiple inode dirty/delete
1765                  * in our compound transaction (bug 1321). */
1766                 shrink_dcache_parent(dchild);
1767                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1768                                       NULL);
1769                 if (IS_ERR(handle))
1770                         GOTO(cleanup, rc = PTR_ERR(handle));
1771                 cleanup_phase = 4; /* transaction */
1772                 rc = vfs_rmdir(dparent->d_inode, dchild);
1773                 break;
1774         case S_IFREG: {
1775 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1776                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1777                                                         offset + 1, 0);
1778                 handle = fsfilt_start_log(obd, dparent->d_inode,
1779                                           FSFILT_OP_UNLINK, NULL,
1780                                           le32_to_cpu(lmm->lmm_stripe_count));
1781                 if (IS_ERR(handle))
1782                         GOTO(cleanup, rc = PTR_ERR(handle));
1783
1784                 cleanup_phase = 4; /* transaction */
1785                 rc = vfs_unlink(dparent->d_inode, dchild);
1786
1787                 if (!rc && log_unlink)
1788                         if (mds_log_op_unlink(obd, child_inode,
1789                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1790                                 req->rq_repmsg->buflens[offset + 1],
1791                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1792                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1793                                 body->valid |= OBD_MD_FLCOOKIE;
1794                 break;
1795         }
1796         case S_IFLNK:
1797         case S_IFCHR:
1798         case S_IFBLK:
1799         case S_IFIFO:
1800         case S_IFSOCK:
1801                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1802                                       NULL);
1803                 if (IS_ERR(handle))
1804                         GOTO(cleanup, rc = PTR_ERR(handle));
1805                 cleanup_phase = 4; /* transaction */
1806                 rc = vfs_unlink(dparent->d_inode, dchild);
1807                 break;
1808         default:
1809                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1810                        rec->ur_name);
1811                 LBUG();
1812                 GOTO(cleanup, rc = -EINVAL);
1813         }
1814
1815  cleanup:
1816         if (rc == 0) {
1817                 struct iattr iattr;
1818                 int err;
1819
1820                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1821                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1822                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1823
1824                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1825                 if (err)
1826                         CERROR("error on parent setattr: rc = %d\n", err);
1827         }
1828
1829         switch(cleanup_phase) {
1830         case 4:
1831                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1832                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1833                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1834                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1835
1836                         /* filesystem is really going to destroy an inode
1837                          * we have to delay this till inode is opened -bzzz */
1838                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1839                 }
1840                 /* handle splitted dir */
1841                 mds_unlink_slave_objs(obd, dchild);
1842                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1843                                         rc, 0);
1844                 if (!rc)
1845                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1846                                            "unlinked", 0, NULL);
1847         case 3: /* child ino-reuse lock */
1848 #if 0
1849                 if (rc && body != NULL) {
1850                         // Don't unlink the OST objects if the MDS unlink failed
1851                         body->valid = 0;
1852                 }
1853                 if (rc)
1854                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1855                 else
1856                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1857 #endif
1858         case 2: /* child lock */
1859                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1860                 if (child_lockh.cookie)
1861                         ldlm_lock_decref(&child_lockh, LCK_EX);
1862         case 1: /* child and parent dentry, parent lock */
1863 #ifdef S_PDIROPS
1864                 if (parent_lockh[1].cookie != 0)
1865                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1866 #endif
1867                 if (rc)
1868                         ldlm_lock_decref(parent_lockh, LCK_PW);
1869                 else
1870                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1871                 l_dput(dchild);
1872                 l_dput(dchild);
1873                 l_dput(dparent);
1874         case 0:
1875                 break;
1876         default:
1877                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1878                 LBUG();
1879         }
1880         req->rq_status = rc;
1881         return 0;
1882 }
1883
1884 /*
1885  * to service requests from remote MDS to increment i_nlink
1886  */
1887 static int mds_reint_link_acquire(struct mds_update_record *rec,
1888                                  int offset, struct ptlrpc_request *req,
1889                                  struct lustre_handle *lh)
1890 {
1891         struct obd_device *obd = req->rq_export->exp_obd;
1892         struct ldlm_res_id src_res_id = { .name = {0} };
1893         struct lustre_handle *handle = NULL, src_lockh = {0};
1894         struct mds_obd *mds = mds_req2mds(req);
1895         int rc = 0, cleanup_phase = 0;
1896         struct dentry *de_src = NULL;
1897         ldlm_policy_data_t policy;
1898         int flags = 0;
1899         ENTRY;
1900
1901         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1902                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1903                   (unsigned) rec->ur_fid1->id,
1904                   (unsigned) rec->ur_fid1->generation);
1905
1906         /* Step 1: Lookup the source inode and target directory by FID */
1907         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1908         if (IS_ERR(de_src))
1909                 GOTO(cleanup, rc = PTR_ERR(de_src));
1910         cleanup_phase = 1; /* source dentry */
1911
1912         src_res_id.name[0] = de_src->d_inode->i_ino;
1913         src_res_id.name[1] = de_src->d_inode->i_generation;
1914         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1915
1916         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1917                         src_res_id, LDLM_IBITS, &policy,
1918                         LCK_EX, &flags, mds_blocking_ast,
1919                         ldlm_completion_ast, NULL, NULL,
1920                         NULL, 0, NULL, &src_lockh);
1921         if (rc != ELDLM_OK)
1922                 GOTO(cleanup, rc = -ENOLCK);
1923         cleanup_phase = 2; /* lock */
1924
1925         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1926
1927         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1928         if (IS_ERR(handle)) {
1929                 rc = PTR_ERR(handle);
1930                 GOTO(cleanup, rc);
1931         }
1932         de_src->d_inode->i_nlink++;
1933         mark_inode_dirty(de_src->d_inode);
1934
1935 cleanup:
1936         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1937                                         handle, req, rc, 0);
1938         EXIT;
1939         switch (cleanup_phase) {
1940                 case 2:
1941                         if (rc)
1942                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1943                         else
1944                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1945                 case 1:
1946                         l_dput(de_src);
1947                 case 0:
1948                         break;
1949                 default:
1950                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1951                         LBUG();
1952         }
1953         req->rq_status = rc;
1954         return 0;
1955 }
1956
1957 /*
1958  * request to link to foreign inode:
1959  *  - acquire i_nlinks on this inode
1960  *  - add dentry
1961  */
1962 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1963                                     int offset, struct ptlrpc_request *req,
1964                                     struct lustre_handle *lh)
1965 {
1966         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1967         struct obd_device *obd = req->rq_export->exp_obd;
1968         struct dentry *de_tgt_dir = NULL;
1969         struct mds_obd *mds = mds_req2mds(req);
1970         int rc = 0, cleanup_phase = 0;
1971         struct mdc_op_data op_data;
1972         struct ptlrpc_request *request = NULL;
1973         int update_mode;
1974         ENTRY;
1975
1976 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1977         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1978                   (unsigned) rec->ur_fid2->mds,
1979                   (unsigned) rec->ur_fid2->id,
1980                   (unsigned) rec->ur_fid2->generation,
1981                   rec->ur_namelen - 1, rec->ur_name,
1982                   (unsigned) rec->ur_fid1->mds,
1983                   (unsigned) rec->ur_fid1->id,
1984                   (unsigned)rec->ur_fid1->generation);
1985
1986         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1987                                            tgt_dir_lockh, &update_mode,
1988                                            rec->ur_name, rec->ur_namelen - 1,
1989                                            MDS_INODELOCK_UPDATE);
1990         if (IS_ERR(de_tgt_dir))
1991                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1992         cleanup_phase = 1;
1993
1994         op_data.fid1 = *(rec->ur_fid1);
1995         op_data.namelen = 0;
1996         op_data.name = NULL;
1997         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1998         if (rc)
1999                 GOTO(cleanup, rc);
2000
2001         cleanup_phase = 2;
2002         if (request)
2003                 ptlrpc_req_finished(request);
2004
2005         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2006
2007         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2008         if (IS_ERR(handle)) {
2009                 rc = PTR_ERR(handle);
2010                 GOTO(cleanup, rc);
2011         }
2012
2013         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2014                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2015                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2016         cleanup_phase = 3;
2017
2018 cleanup:
2019         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2020                                 handle, req, rc, 0);
2021         EXIT;
2022
2023         switch (cleanup_phase) {
2024                 case 3:
2025                         if (rc) {
2026                                 /* FIXME: drop i_nlink on remote inode here */
2027                                 CERROR("MUST drop drop i_nlink here\n");
2028                         }
2029                 case 2:
2030                 case 1:
2031                         if (rc) {
2032                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2033 #ifdef S_PDIROPS
2034                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2035 #endif
2036                         } else {
2037                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2038 #ifdef S_PDIROPS
2039                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2040 #endif
2041                         }
2042                         l_dput(de_tgt_dir);
2043                         break;
2044                 default:
2045                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2046                         LBUG();
2047         }
2048         req->rq_status = rc;
2049         return 0;
2050 }
2051
2052 static int mds_reint_link(struct mds_update_record *rec, int offset,
2053                           struct ptlrpc_request *req,
2054                           struct lustre_handle *lh)
2055 {
2056         struct obd_device *obd = req->rq_export->exp_obd;
2057         struct dentry *de_src = NULL;
2058         struct dentry *de_tgt_dir = NULL;
2059         struct dentry *dchild = NULL;
2060         struct mds_obd *mds = mds_req2mds(req);
2061         struct lustre_handle *handle = NULL;
2062         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2063         struct ldlm_res_id src_res_id = { .name = {0} };
2064         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2065         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2066         ldlm_policy_data_t tgt_dir_policy =
2067                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2068         int rc = 0, cleanup_phase = 0;
2069         int update_mode = 0;
2070         ENTRY;
2071
2072         LASSERT(offset == 0);
2073
2074         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2075                   rec->ur_fid1->id, rec->ur_fid1->generation,
2076                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2077
2078         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2079
2080 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2081         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2082                 GOTO(cleanup, rc = -ENOENT);
2083
2084         if (rec->ur_fid1->mds != mds->mds_num) {
2085                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2086                 RETURN(rc);
2087         }
2088
2089         if (rec->ur_namelen == 1) {
2090                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2091                 RETURN(rc);
2092         }
2093
2094         /* Step 1: Lookup the source inode and target directory by FID */
2095         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2096         if (IS_ERR(de_src))
2097                 GOTO(cleanup, rc = PTR_ERR(de_src));
2098
2099         cleanup_phase = 1; /* source dentry */
2100
2101         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2102         if (IS_ERR(de_tgt_dir)) {
2103                 rc = PTR_ERR(de_tgt_dir);
2104                 de_tgt_dir = NULL;
2105                 GOTO(cleanup, rc);
2106         }
2107
2108         cleanup_phase = 2; /* target directory dentry */
2109
2110         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2111                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2112                de_src->d_inode->i_ino);
2113
2114         /* Step 2: Take the two locks */
2115         src_res_id.name[0] = de_src->d_inode->i_ino;
2116         src_res_id.name[1] = de_src->d_inode->i_generation;
2117         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2118         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2119 #ifdef S_PDIROPS
2120         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2121                 int flags = 0;
2122                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2123                 if (update_mode) {
2124                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2125                                               tgt_dir_res_id, LDLM_IBITS,
2126                                               &src_policy, update_mode, &flags,
2127                                               mds_blocking_ast,
2128                                               ldlm_completion_ast, NULL, NULL,
2129                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2130                         if (rc != ELDLM_OK)
2131                                 GOTO(cleanup, rc = -ENOLCK);
2132                 }
2133
2134                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2135                                                         rec->ur_namelen - 1);
2136                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2137                        de_tgt_dir->d_inode->i_ino,
2138                        de_tgt_dir->d_inode->i_generation,
2139                        tgt_dir_res_id.name[2]);
2140         }
2141 #endif
2142         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2143                                    &src_policy,
2144                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2145                                    &tgt_dir_policy);
2146         if (rc)
2147                 GOTO(cleanup, rc);
2148
2149         cleanup_phase = 3; /* locks */
2150
2151         /* Step 3: Lookup the child */
2152         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2153         if (IS_ERR(dchild)) {
2154                 rc = PTR_ERR(dchild);
2155                 if (rc != -EPERM && rc != -EACCES)
2156                         CERROR("child lookup error %d\n", rc);
2157                 GOTO(cleanup, rc);
2158         }
2159
2160         cleanup_phase = 4; /* child dentry */
2161
2162         if (dchild->d_inode) {
2163                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2164                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2165                 rc = -EEXIST;
2166                 GOTO(cleanup, rc);
2167         }
2168
2169         /* Step 4: Do it. */
2170         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2171
2172         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2173         if (IS_ERR(handle)) {
2174                 rc = PTR_ERR(handle);
2175                 GOTO(cleanup, rc);
2176         }
2177
2178         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2179         if (rc && rc != -EPERM && rc != -EACCES)
2180                 CERROR("vfs_link error %d\n", rc);
2181 cleanup:
2182         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2183                                 handle, req, rc, 0);
2184         EXIT;
2185
2186         switch (cleanup_phase) {
2187         case 4: /* child dentry */
2188                 l_dput(dchild);
2189         case 3: /* locks */
2190                 if (rc) {
2191                         ldlm_lock_decref(&src_lockh, LCK_EX);
2192                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2193                 } else {
2194                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2195                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2196                 }
2197         case 2: /* target dentry */
2198 #ifdef S_PDIROPS
2199                 if (tgt_dir_lockh[1].cookie && update_mode)
2200                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2201 #endif
2202                 if (de_tgt_dir)
2203                         l_dput(de_tgt_dir);
2204         case 1: /* source dentry */
2205                 l_dput(de_src);
2206         case 0:
2207                 break;
2208         default:
2209                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2210                 LBUG();
2211         }
2212         req->rq_status = rc;
2213         return 0;
2214 }
2215
2216 /*
2217  * add a hard link in the PENDING directory, only used by rename()
2218  */
2219 static int mds_add_link_orphan(struct mds_update_record *rec,
2220                                struct obd_device *obd,
2221                                struct dentry *dentry)
2222 {
2223         struct mds_obd *mds = &obd->u.mds;
2224         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2225         struct dentry *pending_child;
2226         char fidname[LL_FID_NAMELEN];
2227         int fidlen = 0, rc;
2228         ENTRY;
2229
2230         LASSERT(dentry->d_inode);
2231         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2232
2233         down(&pending_dir->i_sem);
2234         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2235                             dentry->d_inode->i_generation);
2236
2237         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2238                mds_open_orphan_count(dentry->d_inode),
2239                rec->ur_name, fidname);
2240
2241         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2242         if (IS_ERR(pending_child))
2243                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2244
2245         if (pending_child->d_inode != NULL) {
2246                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2247                 LASSERT(pending_child->d_inode == dentry->d_inode);
2248                 GOTO(out_dput, rc = 0);
2249         }
2250
2251         lock_kernel();
2252         rc = vfs_link(dentry, pending_dir, pending_child);
2253         unlock_kernel();
2254         if (rc)
2255                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2256                        rec->ur_name, rc);
2257         else
2258                 mds_inode_set_orphan(dentry->d_inode);
2259 out_dput:
2260         l_dput(pending_child);
2261 out_lock:
2262         up(&pending_dir->i_sem);
2263         RETURN(rc);
2264 }
2265
2266 /* The idea here is that we need to get four locks in the end:
2267  * one on each parent directory, one on each child.  We need to take
2268  * these locks in some kind of order (to avoid deadlocks), and the order
2269  * I selected is "increasing resource number" order.  We need to look up
2270  * the children, however, before we know what the resource number(s) are.
2271  * Thus the following plan:
2272  *
2273  * 1,2. Look up the parents
2274  * 3,4. Look up the children
2275  * 5. Take locks on the parents and children, in order
2276  * 6. Verify that the children haven't changed since they were looked up
2277  *
2278  * If there was a race and the children changed since they were first looked
2279  * up, it is possible that mds_verify_child() will be able to just grab the
2280  * lock on the new child resource (if it has a higher resource than any other)
2281  * but we need to compare against not only its parent, but also against the
2282  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2283  *
2284  * We need the fancy igrab() on the child inodes because we aren't holding a
2285  * lock on the parent after the lookup is done, so dentry->d_inode may change
2286  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2287  */
2288 static int mds_get_parents_children_locked(struct obd_device *obd,
2289                                            struct mds_obd *mds,
2290                                            struct ll_fid *p1_fid,
2291                                            struct dentry **de_srcdirp,
2292                                            struct ll_fid *p2_fid,
2293                                            struct dentry **de_tgtdirp,
2294                                            int parent_mode,
2295                                            const char *old_name, int old_len,
2296                                            struct dentry **de_oldp,
2297                                            const char *new_name, int new_len,
2298                                            struct dentry **de_newp,
2299                                            struct lustre_handle *dlm_handles,
2300                                            int child_mode)
2301 {
2302         struct ldlm_res_id p1_res_id = { .name = {0} };
2303         struct ldlm_res_id p2_res_id = { .name = {0} };
2304         struct ldlm_res_id c1_res_id = { .name = {0} };
2305         struct ldlm_res_id c2_res_id = { .name = {0} };
2306         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2307         /* Only dentry should change, but the inode itself would be
2308            intact otherwise */
2309         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2310         /* If something is going to be replaced, both dentry and inode locks are
2311            needed */
2312         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2313                                                         MDS_INODELOCK_UPDATE}};
2314         struct ldlm_res_id *maxres_src, *maxres_tgt;
2315         struct inode *inode;
2316         int rc = 0, cleanup_phase = 0;
2317         ENTRY;
2318
2319         /* Step 1: Lookup the source directory */
2320         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2321         if (IS_ERR(*de_srcdirp))
2322                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2323
2324         cleanup_phase = 1; /* source directory dentry */
2325
2326         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2327         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2328
2329         /* Step 2: Lookup the target directory */
2330         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2331                 *de_tgtdirp = dget(*de_srcdirp);
2332         } else {
2333                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2334                 if (IS_ERR(*de_tgtdirp)) {
2335                         rc = PTR_ERR(*de_tgtdirp);
2336                         *de_tgtdirp = NULL;
2337                         GOTO(cleanup, rc);
2338                 }
2339         }
2340
2341         cleanup_phase = 2; /* target directory dentry */
2342
2343         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2344         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2345
2346 #ifdef S_PDIROPS
2347         dlm_handles[5].cookie = 0;
2348         dlm_handles[6].cookie = 0;
2349         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2350                 /* Get a temp lock on just ino, gen to flush client cache and
2351                  * to protect dirs from concurrent splitting */
2352                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2353                                            LCK_PW, &p_policy, &p2_res_id,
2354                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2355                 if (rc != ELDLM_OK)
2356                         GOTO(cleanup, rc);
2357                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2358                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2359                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2360                        (*de_srcdirp)->d_inode->i_ino,
2361                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2362                        (*de_tgtdirp)->d_inode->i_ino,
2363                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2364         }
2365         cleanup_phase = 3;
2366 #endif
2367
2368         /* Step 3: Lookup the source child entry */
2369         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2370         if (IS_ERR(*de_oldp)) {
2371                 rc = PTR_ERR(*de_oldp);
2372                 CERROR("old child lookup error (%*s): %d\n",
2373                        old_len - 1, old_name, rc);
2374                 GOTO(cleanup, rc);
2375         }
2376
2377         cleanup_phase = 4; /* original name dentry */
2378
2379         inode = (*de_oldp)->d_inode;
2380         if (inode != NULL) {
2381                 inode = igrab(inode);
2382                 if (inode == NULL)
2383                         GOTO(cleanup, rc = -ENOENT);
2384
2385                 c1_res_id.name[0] = inode->i_ino;
2386                 c1_res_id.name[1] = inode->i_generation;
2387                 iput(inode);
2388         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2389                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2390                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2391         }
2392
2393         /* Step 4: Lookup the target child entry */
2394         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2395         if (IS_ERR(*de_newp)) {
2396                 rc = PTR_ERR(*de_newp);
2397                 CERROR("new child lookup error (%*s): %d\n",
2398                        old_len - 1, old_name, rc);
2399                 GOTO(cleanup, rc);
2400         }
2401
2402         cleanup_phase = 5; /* target dentry */
2403
2404         inode = (*de_newp)->d_inode;
2405         if (inode != NULL) {
2406                 inode = igrab(inode);
2407                 if (inode == NULL)
2408                         goto retry_locks;
2409
2410                 c2_res_id.name[0] = inode->i_ino;
2411                 c2_res_id.name[1] = inode->i_generation;
2412                 iput(inode);
2413         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2414                 c2_res_id.name[0] = (*de_newp)->d_inum;
2415                 c2_res_id.name[1] = (*de_newp)->d_generation;
2416         }
2417
2418 retry_locks:
2419         /* Step 5: Take locks on the parents and child(ren) */
2420         maxres_src = &p1_res_id;
2421         maxres_tgt = &p2_res_id;
2422         cleanup_phase = 5; /* target dentry */
2423
2424         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2425                 maxres_src = &c1_res_id;
2426         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2427                 maxres_tgt = &c2_res_id;
2428
2429         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2430                                     &p_policy,
2431                                     &p2_res_id, &dlm_handles[1], parent_mode,
2432                                     &p_policy,
2433                                     &c1_res_id, &dlm_handles[2], child_mode,
2434                                     &c1_policy,
2435                                     &c2_res_id, &dlm_handles[3], child_mode,
2436                                     &c2_policy);
2437         if (rc)
2438                 GOTO(cleanup, rc);
2439
2440         cleanup_phase = 6; /* parent and child(ren) locks */
2441
2442         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2443         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2444                               parent_mode, &c1_res_id, &dlm_handles[2],
2445                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2446                               maxres_tgt);
2447         if (rc) {
2448                 if (c2_res_id.name[0] != 0)
2449                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2450                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2451                 cleanup_phase = 5;
2452                 if (rc > 0)
2453                         goto retry_locks;
2454                 GOTO(cleanup, rc);
2455         }
2456
2457         if (!DENTRY_VALID(*de_oldp))
2458                 GOTO(cleanup, rc = -ENOENT);
2459
2460         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2461         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2462                               parent_mode, &c2_res_id, &dlm_handles[3],
2463                               de_newp, child_mode, &c2_policy, new_name,
2464                               new_len, maxres_src);
2465         if (rc) {
2466                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2467                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2468                 cleanup_phase = 5;
2469                 if (rc > 0)
2470                         goto retry_locks;
2471                 GOTO(cleanup, rc);
2472         }
2473
2474         EXIT;
2475 cleanup:
2476         if (rc) {
2477                 switch (cleanup_phase) {
2478                 case 6: /* child lock(s) */
2479                         if (c2_res_id.name[0] != 0)
2480                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2481                         if (c1_res_id.name[0] != 0)
2482                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2483                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2484                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2485                 case 5: /* target dentry */
2486                         l_dput(*de_newp);
2487                 case 4: /* source dentry */
2488                         l_dput(*de_oldp);
2489                 case 3:
2490 #ifdef S_PDIROPS
2491                         if (dlm_handles[5].cookie != 0)
2492                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2493                         if (dlm_handles[6].cookie != 0)
2494                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2495 #endif
2496                 case 2: /* target directory dentry */
2497                         l_dput(*de_tgtdirp);
2498                 case 1: /* source directry dentry */
2499                         l_dput(*de_srcdirp);
2500                 }
2501         }
2502
2503         return rc;
2504 }
2505
2506 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2507                                         int offset, struct ptlrpc_request *req)
2508 {
2509         struct obd_device *obd = req->rq_export->exp_obd;
2510         struct dentry *de_srcdir = NULL;
2511         struct dentry *de_new = NULL;
2512         struct mds_obd *mds = mds_req2mds(req);
2513         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2514         struct lustre_handle child_lockh = {0};
2515         int cleanup_phase = 0;
2516         void *handle = NULL;
2517         int update_mode, rc = 0;
2518         ENTRY;
2519
2520         /* another MDS executing rename operation has asked us
2521          * to create target name. such a creation should destroy
2522          * existing target name */
2523
2524         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2525                         obd->obd_name, rec->ur_tgt,
2526                         (unsigned long) rec->ur_fid1->mds,
2527                         (unsigned long) rec->ur_fid1->id,
2528                         (unsigned long) rec->ur_fid1->generation);
2529
2530         /* first, lookup the target */
2531         child_lockh.cookie = 0;
2532         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2533                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2534                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2535                                          &child_lockh, &de_new, LCK_EX,
2536                                          MDS_INODELOCK_LOOKUP, NULL);
2537         if (rc)
2538                 GOTO(cleanup, rc);
2539
2540         cleanup_phase = 1;
2541
2542         LASSERT(de_srcdir);
2543         LASSERT(de_srcdir->d_inode);
2544         LASSERT(de_new);
2545
2546         if (de_new->d_inode) {
2547                 /* name exists and points to local inode
2548                  * try to unlink this name and create new one */
2549                 CERROR("%s: %s points to local inode %lu/%lu\n",
2550                        obd->obd_name, rec->ur_tgt,
2551                        (unsigned long) de_new->d_inode->i_ino,
2552                        (unsigned long) de_new->d_inode->i_generation);
2553                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2554                                       FSFILT_OP_RENAME, NULL);
2555                 if (IS_ERR(handle))
2556                         GOTO(cleanup, rc = PTR_ERR(handle));
2557                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2558                 if (rc)
2559                         GOTO(cleanup, rc);
2560         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2561                 /* name exists adn points to remove inode */
2562                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2563                        obd->obd_name, rec->ur_tgt,
2564                        (unsigned long) de_new->d_mdsnum,
2565                        (unsigned long) de_new->d_inum,
2566                        (unsigned long) de_new->d_generation);
2567         } else {
2568                 /* name doesn't exist. the simplest case */
2569                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2570                                       FSFILT_OP_LINK, NULL);
2571                 if (IS_ERR(handle))
2572                         GOTO(cleanup, rc = PTR_ERR(handle));
2573         }
2574
2575         cleanup_phase = 2;
2576         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2577                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2578                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2579         if (rc)
2580                 CERROR("add_dir_entry() returned error %d\n", rc);
2581 cleanup:
2582         EXIT;
2583         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2584                                 handle, req, rc, 0);
2585         switch(cleanup_phase) {
2586                 case 2:
2587                 case 1:
2588 #ifdef S_PDIROPS
2589                         if (parent_lockh[1].cookie != 0)
2590                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2591 #endif
2592                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2593                         if (child_lockh.cookie != 0)
2594                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2595                         l_dput(de_new);
2596                         l_dput(de_srcdir);
2597                         break;
2598                 default:
2599                         LBUG();
2600         }
2601
2602         req->rq_status = rc;
2603
2604         RETURN(0);
2605 }
2606
2607 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2608                                       struct ptlrpc_request *req)
2609 {
2610         struct obd_device *obd = req->rq_export->exp_obd;
2611         struct ptlrpc_request *req2 = NULL;
2612         struct dentry *de_srcdir = NULL;
2613         struct dentry *de_old = NULL;
2614         struct mds_obd *mds = mds_req2mds(req);
2615         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2616         struct lustre_handle child_lockh = {0};
2617         struct mdc_op_data opdata;
2618         void *handle = NULL;
2619         int update_mode, rc = 0;
2620         ENTRY;
2621
2622         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2623                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2624         memset(&opdata, 0, sizeof(opdata));
2625
2626         child_lockh.cookie = 0;
2627         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2628                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2629                                          &update_mode, rec->ur_name,
2630                                          rec->ur_namelen, &child_lockh, &de_old,
2631                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2632         LASSERT(rc == 0);
2633         LASSERT(de_srcdir);
2634         LASSERT(de_srcdir->d_inode);
2635         LASSERT(de_old);
2636
2637         /* we already know the target should be created on another MDS
2638          * so, we have to request that MDS to do it */
2639
2640         /* prepare source fid */
2641         if (de_old->d_flags & DCACHE_CROSS_REF) {
2642                 LASSERT(de_old->d_inode == NULL);
2643                 CDEBUG(D_OTHER, "request to move remote name\n");
2644                 opdata.fid1.mds = de_old->d_mdsnum;
2645                 opdata.fid1.id = de_old->d_inum;
2646                 opdata.fid1.generation = de_old->d_generation;
2647         } else if (de_old->d_inode == NULL) {
2648                 /* oh, source doesn't exist */
2649                 GOTO(cleanup, rc = -ENOENT);
2650         } else {
2651                 LASSERT(de_old->d_inode != NULL);
2652                 CDEBUG(D_OTHER, "request to move local name\n");
2653                 opdata.fid1.mds = mds->mds_num;
2654                 opdata.fid1.id = de_old->d_inode->i_ino;
2655                 opdata.fid1.generation = de_old->d_inode->i_generation;
2656         }
2657
2658         opdata.fid2 = *(rec->ur_fid2);
2659         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2660                        rec->ur_tgtlen - 1, &req2);
2661
2662         if (rc)
2663                 GOTO(cleanup, rc);
2664
2665         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2666         if (IS_ERR(handle))
2667                 GOTO(cleanup, rc = PTR_ERR(handle));
2668         rc = fsfilt_del_dir_entry(obd, de_old);
2669         d_drop(de_old);
2670
2671 cleanup:
2672         EXIT;
2673         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2674                                 handle, req, rc, 0);
2675         if (req2)
2676                 ptlrpc_req_finished(req2);
2677
2678 #ifdef S_PDIROPS
2679         if (parent_lockh[1].cookie != 0)
2680                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2681 #endif
2682         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2683         if (child_lockh.cookie != 0)
2684                 ldlm_lock_decref(&child_lockh, LCK_EX);
2685
2686         l_dput(de_old);
2687         l_dput(de_srcdir);
2688
2689         req->rq_status = rc;
2690         RETURN(0);
2691
2692 }
2693
2694 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2695                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2696 {
2697         struct obd_device *obd = req->rq_export->exp_obd;
2698         struct dentry *de_srcdir = NULL;
2699         struct dentry *de_tgtdir = NULL;
2700         struct dentry *de_old = NULL;
2701         struct dentry *de_new = NULL;
2702         struct mds_obd *mds = mds_req2mds(req);
2703         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2704         struct mds_body *body = NULL;
2705         int rc = 0, lock_count = 3;
2706         int cleanup_phase = 0;
2707         void *handle = NULL;
2708         ENTRY;
2709
2710         LASSERT(offset == 0);
2711
2712         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2713                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2714                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2715
2716         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2717
2718         if (rec->ur_namelen == 1) {
2719                 rc = mds_reint_rename_create_name(rec, offset, req);
2720                 RETURN(rc);
2721         }
2722
2723         if (rec->ur_fid2->mds != mds->mds_num) {
2724                 rc = mds_reint_rename_to_remote(rec, offset, req);
2725                 RETURN(rc);
2726         }
2727
2728         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2729                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2730                                              rec->ur_name, rec->ur_namelen,
2731                                              &de_old, rec->ur_tgt,
2732                                              rec->ur_tgtlen, &de_new,
2733                                              dlm_handles, LCK_EX);
2734         if (rc)
2735                 GOTO(cleanup, rc);
2736
2737         cleanup_phase = 1; /* parent(s), children, locks */
2738
2739         if (de_new->d_inode)
2740                 lock_count = 4;
2741
2742         /* sanity check for src inode */
2743         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2744             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2745                 GOTO(cleanup, rc = -EINVAL);
2746
2747         /* sanity check for dest inode */
2748         if (de_new->d_inode &&
2749             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2750              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2751                 GOTO(cleanup, rc = -EINVAL);
2752
2753         if (de_old->d_inode == de_new->d_inode)
2754                 GOTO(cleanup, rc = 0);
2755
2756         /* if we are about to remove the target at first, pass the EA of
2757          * that inode to client to perform and cleanup on OST */
2758         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2759         LASSERT(body != NULL);
2760
2761         if (de_new->d_inode &&
2762             S_ISREG(de_new->d_inode->i_mode) &&
2763             de_new->d_inode->i_nlink == 1 &&
2764             mds_open_orphan_count(de_new->d_inode) == 0) {
2765                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2766                 mds_pack_inode2body(obd, body, de_new->d_inode);
2767                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2768                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2769                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2770                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2771                 } else {
2772                         /* XXX need log unlink? */
2773                 }
2774         }
2775
2776         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2777                        de_srcdir->d_inode->i_sb);
2778
2779         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2780         if (IS_ERR(handle))
2781                 GOTO(cleanup, rc = PTR_ERR(handle));
2782
2783         /* FIXME need adjust the journal block count? */
2784         /* if the target should be moved to PENDING, we at first increase the
2785          * link and later vfs_rename() will decrease the link count again */
2786         if (de_new->d_inode &&
2787             S_ISREG(de_new->d_inode->i_mode) &&
2788             de_new->d_inode->i_nlink == 1 &&
2789             mds_open_orphan_count(de_new->d_inode) > 0) {
2790                 rc = mds_add_link_orphan(rec, obd, de_new);
2791                 if (rc)
2792                         GOTO(cleanup, rc);
2793         }
2794
2795         lock_kernel();
2796         de_old->d_fsdata = req;
2797         de_new->d_fsdata = req;
2798         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2799         unlock_kernel();
2800
2801         GOTO(cleanup, rc);
2802 cleanup:
2803         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2804                                 handle, req, rc, 0);
2805         switch (cleanup_phase) {
2806         case 1:
2807 #ifdef S_PDIROPS
2808                 if (dlm_handles[5].cookie != 0)
2809                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2810                 if (dlm_handles[6].cookie != 0)
2811                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2812 #endif
2813                 if (rc) {
2814                         if (lock_count == 4)
2815                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2816                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2817                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2818                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2819                 } else {
2820                         if (lock_count == 4)
2821                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2822                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2823                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2824                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2825                 }
2826                 l_dput(de_new);
2827                 l_dput(de_old);
2828                 l_dput(de_tgtdir);
2829                 l_dput(de_srcdir);
2830         case 0:
2831                 break;
2832         default:
2833                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2834                 LBUG();
2835         }
2836         req->rq_status = rc;
2837         return 0;
2838 }
2839
2840 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2841                            struct ptlrpc_request *, struct lustre_handle *);
2842
2843 static mds_reinter reinters[REINT_MAX + 1] = {
2844         [REINT_SETATTR] mds_reint_setattr,
2845         [REINT_CREATE] mds_reint_create,
2846         [REINT_LINK] mds_reint_link,
2847         [REINT_UNLINK] mds_reint_unlink,
2848         [REINT_RENAME] mds_reint_rename,
2849         [REINT_OPEN] mds_open
2850 };
2851
2852 int mds_reint_rec(struct mds_update_record *rec, int offset,
2853                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2854 {
2855         struct obd_device *obd = req->rq_export->exp_obd;
2856         struct lvfs_run_ctxt saved;
2857         int rc;
2858
2859         /* checked by unpacker */
2860         LASSERT(rec->ur_opcode <= REINT_MAX &&
2861                 reinters[rec->ur_opcode] != NULL);
2862
2863         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2864         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2865         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2866
2867         return rc;
2868 }