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) 2002-2005 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 the Lustre file system, http://www.lustre.org
  13  *   Lustre is a trademark of Cluster File Systems, Inc.
  14  *
  15  *   You may have signed or agreed to another license before downloading
  16  *   this software.  If so, you are bound by the terms and conditions
  17  *   of that agreement, and the following does not apply to you.  See the
  18  *   LICENSE file included with this distribution for more information.
  19  *
  20  *   If you did not agree to a different license, then this copy of Lustre
  21  *   is open source software; you can redistribute it and/or modify it
  22  *   under the terms of version 2 of the GNU General Public License as
  23  *   published by the Free Software Foundation.
  24  *
  25  *   In either case, Lustre is distributed in the hope that it will be
  26  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  27  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  28  *   license text for more details.
  29  */
  30
  31 #ifndef EXPORT_SYMTAB
  32 # define EXPORT_SYMTAB
  33 #endif
  34 #define DEBUG_SUBSYSTEM S_MDS
  35
  36 #include <linux/fs.h>
  37 #include <obd_support.h>
  38 #include <obd_class.h>
  39 #include <obd.h>
  40 #include <lustre_lib.h>
  41 #include <lustre/lustre_idl.h>
  42 #include <lustre_mds.h>
  43 #include <lustre_dlm.h>
  44 #include <lustre_fsfilt.h>
  45 #include <lustre_ucache.h>
  46
  47 #include "mds_internal.h"
  48
  49 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  50                    int error)
  51 {
  52         obd_transno_commit_cb(obd, transno, error);
  53 }
  54
  55 struct mds_logcancel_data {
  56         struct lov_mds_md      *mlcd_lmm;
  57         int                     mlcd_size;
  58         int                     mlcd_cookielen;
  59         int                     mlcd_eadatalen;
  60         struct llog_cookie      mlcd_cookies[0];
  61 };
  62
  63
  64 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  65                                   void *cb_data, int error)
  66 {
  67         struct mds_logcancel_data *mlcd = cb_data;
  68         struct lov_stripe_md *lsm = NULL;
  69         struct llog_ctxt *ctxt;
  70         int rc;
  71
  72         obd_transno_commit_cb(obd, transno, error);
  73
  74         CDEBUG(D_HA, "cancelling %d cookies\n",
  75                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  76
  77         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  78                           mlcd->mlcd_eadatalen);
  79         if (rc < 0) {
  80                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  81                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  82                        rc);
  83         } else {
  84                 ///* XXX 0 normally, SENDNOW for debug */);
  85                 rc = obd_checkmd(obd->u.mds.mds_osc_exp, obd->obd_self_export,
  86                                  lsm);
  87                 if (rc)
  88                         CERROR("Can not revalidate lsm %p \n", lsm);
  89
  90                 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
  91                 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
  92                                                 sizeof(*mlcd->mlcd_cookies),
  93                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  94                 if (rc)
  95                         CERROR("error cancelling %d log cookies: rc %d\n",
  96                                (int)(mlcd->mlcd_cookielen /
  97                                      sizeof(*mlcd->mlcd_cookies)), rc);
  98         }
  99
 100         OBD_FREE(mlcd, mlcd->mlcd_size);
 101 }
 102
 103 /* Assumes caller has already pushed us into the kernel context. */
 104 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
 105                        struct ptlrpc_request *req, int rc, __u32 op_data)
 106 {
 107         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 108         struct mds_client_data *mcd = med->med_mcd;
 109         struct obd_device *obd = req->rq_export->exp_obd;
 110         int err;
 111         __u64 transno;
 112         loff_t off;
 113         int log_pri = D_HA;
 114         ENTRY;
 115
 116         if (IS_ERR(handle)) {
 117                 LASSERT(rc != 0);
 118                 RETURN(rc);
 119         }
 120
 121         /* if the export has already been failed, we have no last_rcvd slot */
 122         if (req->rq_export->exp_failed) {
 123                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 124                       req->rq_export->exp_client_uuid.uuid, rc);
 125                 if (rc == 0)
 126                         rc = -ENOTCONN;
 127                 if (handle)
 128                         GOTO(commit, rc);
 129                 RETURN(rc);
 130         }
 131
 132         if (handle == NULL) {
 133                 /* if we're starting our own xaction, use our own inode */
 134                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 135                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 136                 if (IS_ERR(handle)) {
 137                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 138                         RETURN(PTR_ERR(handle));
 139                 }
 140         }
 141
 142         off = med->med_lr_off;
 143
 144         transno = req->rq_reqmsg->transno;
 145         if (rc != 0) {
 146                 if (transno != 0) {
 147                         CERROR("%s: replay %s transno "LPU64" failed: rc %d\n",
 148                                obd->obd_name,
 149                                libcfs_nid2str(req->rq_export->exp_connection->c_peer.nid),
 150                                transno, rc);
 151                         transno = 0;
 152                 }
 153         } else if (transno == 0) {
 154                 spin_lock(&mds->mds_transno_lock);
 155                 transno = ++mds->mds_last_transno;
 156                 spin_unlock(&mds->mds_transno_lock);
 157         } else {
 158                 spin_lock(&mds->mds_transno_lock);
 159                 if (transno > mds->mds_last_transno)
 160                         mds->mds_last_transno = transno;
 161                 spin_unlock(&mds->mds_transno_lock);
 162         }
 163         req->rq_repmsg->transno = req->rq_transno = transno;
 164         mcd->mcd_last_transno = cpu_to_le64(transno);
 165         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 166         mcd->mcd_last_result = cpu_to_le32(rc);
 167         mcd->mcd_last_data = cpu_to_le32(op_data);
 168
 169         if (off <= 0) {
 170                 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
 171                 err = -EINVAL;
 172         } else {
 173                 fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 174                                       mds_commit_cb, NULL);
 175                 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
 176                                           sizeof(*mcd), &off, 0);
 177         }
 178
 179         if (err) {
 180                 log_pri = D_ERROR;
 181                 if (rc == 0)
 182                         rc = err;
 183         }
 184
 185         DEBUG_REQ_EX(log_pri, req,
 186                      "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
 187                      transno, rc, mcd->mcd_uuid, med->med_lr_idx, err);
 188
 189         err = mds_lov_write_objids(obd);
 190         if (err) {
 191                 log_pri = D_ERROR;
 192                 if (rc == 0)
 193                         rc = err;
 194         }
 195         CDEBUG_EX(log_pri, "wrote objids: err = %d\n", err);
 196
 197 commit:
 198         err = fsfilt_commit(obd, inode, handle, 0);
 199         if (err) {
 200                 CERROR("error committing transaction: %d\n", err);
 201                 if (!rc)
 202                         rc = err;
 203         }
 204
 205         RETURN(rc);
 206 }
 207
 208 /* this gives the same functionality as the code between
 209  * sys_chmod and inode_setattr
 210  * chown_common and inode_setattr
 211  * utimes and inode_setattr
 212  */
 213 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 214 {
 215         time_t now = CURRENT_SECONDS;
 216         struct iattr *attr = &rec->ur_iattr;
 217         unsigned int ia_valid = attr->ia_valid;
 218         int error;
 219         ENTRY;
 220
 221         if (ia_valid & ATTR_RAW)
 222                 attr->ia_valid &= ~ATTR_RAW;
 223
 224         if (!(ia_valid & ATTR_CTIME_SET))
 225                 LTIME_S(attr->ia_ctime) = now;
 226         else
 227                 attr->ia_valid &= ~ATTR_CTIME_SET;
 228         if (!(ia_valid & ATTR_ATIME_SET))
 229                 LTIME_S(attr->ia_atime) = now;
 230         if (!(ia_valid & ATTR_MTIME_SET))
 231                 LTIME_S(attr->ia_mtime) = now;
 232
 233         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 234                 RETURN((attr->ia_valid & ~ATTR_ATTR_FLAG) ? -EPERM : 0);
 235
 236         /* times */
 237         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 238                 if (current->fsuid != inode->i_uid &&
 239                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 240                         RETURN(error);
 241         }
 242
 243         if (ia_valid & ATTR_SIZE &&
 244             /* NFSD hack for open(O_CREAT|O_TRUNC)=mknod+truncate (bug 5781) */
 245             !(rec->ur_uc.luc_fsuid == inode->i_uid &&
 246               ia_valid & MDS_OPEN_OWNEROVERRIDE)) {
 247                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 248                         RETURN(error);
 249         }
 250
 251         if (ia_valid & (ATTR_UID | ATTR_GID)) {
 252                 /* chown */
 253                 error = -EPERM;
 254                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 255                         RETURN(-EPERM);
 256                 if (attr->ia_uid == (uid_t) -1)
 257                         attr->ia_uid = inode->i_uid;
 258                 if (attr->ia_gid == (gid_t) -1)
 259                         attr->ia_gid = inode->i_gid;
 260                 if (!(ia_valid & ATTR_MODE))
 261                         attr->ia_mode = inode->i_mode;
 262                 /*
 263                  * If the user or group of a non-directory has been
 264                  * changed by a non-root user, remove the setuid bit.
 265                  * 19981026 David C Niemi <niemi@tux.org>
 266                  *
 267                  * Changed this to apply to all users, including root,
 268                  * to avoid some races. This is the behavior we had in
 269                  * 2.0. The check for non-root was definitely wrong
 270                  * for 2.2 anyway, as it should have been using
 271                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 272                  */
 273                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 274                     !S_ISDIR(inode->i_mode)) {
 275                         attr->ia_mode &= ~S_ISUID;
 276                         attr->ia_valid |= ATTR_MODE;
 277                 }
 278                 /*
 279                  * Likewise, if the user or group of a non-directory
 280                  * has been changed by a non-root user, remove the
 281                  * setgid bit UNLESS there is no group execute bit
 282                  * (this would be a file marked for mandatory
 283                  * locking).  19981026 David C Niemi <niemi@tux.org>
 284                  *
 285                  * Removed the fsuid check (see the comment above) --
 286                  * 19990830 SD.
 287                  */
 288                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 289                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 290                         attr->ia_mode &= ~S_ISGID;
 291                         attr->ia_valid |= ATTR_MODE;
 292                 }
 293         } else if (ia_valid & ATTR_MODE) {
 294                 int mode = attr->ia_mode;
 295                 /* chmod */
 296                 if (attr->ia_mode == (umode_t)-1)
 297                         mode = inode->i_mode;
 298                 attr->ia_mode =
 299                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 300         }
 301         RETURN(0);
 302 }
 303
 304 void mds_steal_ack_locks(struct ptlrpc_request *req)
 305 {
 306         struct obd_export         *exp = req->rq_export;
 307         struct list_head          *tmp;
 308         struct ptlrpc_reply_state *oldrep;
 309         struct ptlrpc_service     *svc;
 310         unsigned long              flags;
 311         int                        i;
 312
 313         /* CAVEAT EMPTOR: spinlock order */
 314         spin_lock_irqsave (&exp->exp_lock, flags);
 315         list_for_each (tmp, &exp->exp_outstanding_replies) {
 316                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 317
 318                 if (oldrep->rs_xid != req->rq_xid)
 319                         continue;
 320
 321                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 322                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 323                                 "new %d old %d\n", req->rq_xid,
 324                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 325
 326                 svc = oldrep->rs_service;
 327                 spin_lock (&svc->srv_lock);
 328
 329                 list_del_init (&oldrep->rs_exp_list);
 330
 331                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 332                       " o%d NID %s\n",
 333                       oldrep->rs_nlocks, oldrep,
 334                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 335                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 336
 337                 for (i = 0; i < oldrep->rs_nlocks; i++)
 338                         ptlrpc_save_lock(req,
 339                                          &oldrep->rs_locks[i],
 340                                          oldrep->rs_modes[i]);
 341                 oldrep->rs_nlocks = 0;
 342
 343                 DEBUG_REQ(D_HA, req, "stole locks for");
 344                 ptlrpc_schedule_difficult_reply (oldrep);
 345
 346                 spin_unlock (&svc->srv_lock);
 347                 break;
 348         }
 349         spin_unlock_irqrestore (&exp->exp_lock, flags);
 350 }
 351
 352 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 353 {
 354         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 355                   mcd->mcd_last_transno, mcd->mcd_last_result);
 356         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 357         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 358
 359         mds_steal_ack_locks(req);
 360 }
 361
 362 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 363                                       int offset, struct ptlrpc_request *req)
 364 {
 365         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 366         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 367         struct dentry *de;
 368         struct mds_body *body;
 369
 370         mds_req_from_mcd(req, med->med_mcd);
 371
 372         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 373         if (IS_ERR(de)) {
 374                 LASSERT(PTR_ERR(de) == req->rq_status);
 375                 return;
 376         }
 377
 378         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 379         mds_pack_inode2fid(&body->fid1, de->d_inode);
 380         mds_pack_inode2body(body, de->d_inode);
 381
 382         /* Don't return OST-specific attributes if we didn't just set them */
 383         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 384                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 385         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 386                 body->valid |= OBD_MD_FLMTIME;
 387         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 388                 body->valid |= OBD_MD_FLATIME;
 389
 390         l_dput(de);
 391 }
 392
 393 int mds_osc_setattr_async(struct obd_device *obd, struct inode *inode,
 394                           struct lov_mds_md *lmm, int lmm_size,
 395                           struct llog_cookie *logcookies, struct ll_fid *fid)
 396 {
 397         struct mds_obd *mds = &obd->u.mds;
 398         struct lov_stripe_md *lsm = NULL;
 399         struct obd_trans_info oti = { 0 };
 400         struct obdo *oa = NULL;
 401         int rc;
 402         ENTRY;
 403
 404         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
 405                 RETURN(0);
 406
 407         /* first get memory EA */
 408         oa = obdo_alloc();
 409         if (!oa)
 410                 RETURN(-ENOMEM);
 411
 412         LASSERT(lmm);
 413
 414         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
 415         if (rc < 0) {
 416                 CERROR("Error unpack md %p for inode %lu\n", lmm, inode->i_ino);
 417                 GOTO(out, rc);
 418         }
 419
 420         rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, lsm);
 421         if (rc) {
 422                 CERROR("Error revalidate lsm %p \n", lsm);
 423                 GOTO(out, rc);
 424         }
 425
 426         /* then fill oa */
 427         oa->o_id = lsm->lsm_object_id;
 428         oa->o_uid = inode->i_uid;
 429         oa->o_gid = inode->i_gid;
 430         oa->o_valid = OBD_MD_FLID | OBD_MD_FLUID | OBD_MD_FLGID;
 431         if (logcookies) {
 432                 oa->o_valid |= OBD_MD_FLCOOKIE;
 433                 oti.oti_logcookies = logcookies;
 434         }
 435
 436         LASSERT(fid != NULL);
 437         oa->o_fid = fid->id;
 438         oa->o_generation = fid->generation;
 439         oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 440
 441         /* do setattr from mds to ost asynchronously */
 442         rc = obd_setattr_async(mds->mds_osc_exp, oa, lsm, &oti);
 443         if (rc)
 444                 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64
 445                        " on ost error %d\n", lsm->lsm_object_id, rc);
 446 out:
 447         if (lsm)
 448                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 449         obdo_free(oa);
 450         RETURN(rc);
 451 }
 452
 453 /* In the raw-setattr case, we lock the child inode.
 454  * In the write-back case or if being called from open, the client holds a lock
 455  * already.
 456  *
 457  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 458 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 459                              struct ptlrpc_request *req,
 460                              struct lustre_handle *lh)
 461 {
 462         unsigned int ia_valid = rec->ur_iattr.ia_valid;
 463         struct mds_obd *mds = mds_req2mds(req);
 464         struct obd_device *obd = req->rq_export->exp_obd;
 465         struct mds_body *body;
 466         struct dentry *de;
 467         struct inode *inode = NULL;
 468         struct lustre_handle lockh;
 469         void *handle = NULL;
 470         struct mds_logcancel_data *mlcd = NULL;
 471         struct lov_mds_md *lmm = NULL;
 472         struct llog_cookie *logcookies = NULL;
 473         int lmm_size = 0, need_lock = 1, cookie_size = 0;
 474         int rc = 0, cleanup_phase = 0, err, locked = 0;
 475         unsigned int qcids[MAXQUOTAS] = {0, 0};
 476         unsigned int qpids[MAXQUOTAS] = {rec->ur_iattr.ia_uid,
 477                                          rec->ur_iattr.ia_gid};
 478         ENTRY;
 479
 480         LASSERT(offset == MDS_REQ_REC_OFF);
 481
 482         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 483                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 484
 485         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 486
 487         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
 488             (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
 489                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 490                 if (IS_ERR(de))
 491                         GOTO(cleanup, rc = PTR_ERR(de));
 492                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
 493                         GOTO(cleanup, rc = -EROFS);
 494         } else {
 495                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 496                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
 497                         lockpart |= MDS_INODELOCK_LOOKUP;
 498
 499                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
 500                                            &lockh, lockpart);
 501                 if (IS_ERR(de))
 502                         GOTO(cleanup, rc = PTR_ERR(de));
 503                 locked = 1;
 504         }
 505
 506         cleanup_phase = 1;
 507         inode = de->d_inode;
 508         LASSERT(inode);
 509
 510         /* save uid/gid for quota acq/rel */
 511         qcids[USRQUOTA] = inode->i_uid;
 512         qcids[GRPQUOTA] = inode->i_gid;
 513
 514         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 515             rec->ur_eadata != NULL) {
 516                 LOCK_INODE_MUTEX(inode);
 517                 need_lock = 0;
 518         }
 519
 520         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 521
 522         /* start a log jounal handle if needed */
 523         if (S_ISREG(inode->i_mode) &&
 524             rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 525                 lmm_size = mds->mds_max_mdsize;
 526                 OBD_ALLOC(lmm, lmm_size);
 527                 if (lmm == NULL)
 528                         GOTO(cleanup, rc = -ENOMEM);
 529
 530                 cleanup_phase = 2;
 531                 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock);
 532                 if (rc < 0)
 533                         GOTO(cleanup, rc);
 534                 rc = 0;
 535
 536                 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
 537                                           le32_to_cpu(lmm->lmm_stripe_count));
 538         } else {
 539                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 540         }
 541         if (IS_ERR(handle))
 542                 GOTO(cleanup, rc = PTR_ERR(handle));
 543
 544         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 545                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 546                        LTIME_S(rec->ur_iattr.ia_mtime),
 547                        LTIME_S(rec->ur_iattr.ia_ctime));
 548         rc = mds_fix_attr(inode, rec);
 549         if (rc)
 550                 GOTO(cleanup, rc);
 551
 552         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) {  /* ioctl */
 553                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 554                                       (long)&rec->ur_iattr.ia_attr_flags);
 555         } else if (rec->ur_iattr.ia_valid) {            /* setattr */
 556                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 557                 /* journal chown/chgrp in llog, just like unlink */
 558                 if (rc == 0 && lmm_size){
 559                         cookie_size = mds_get_cookie_size(obd, lmm);
 560                         OBD_ALLOC(logcookies, cookie_size);
 561                         if (logcookies == NULL)
 562                                 GOTO(cleanup, rc = -ENOMEM);
 563
 564                         if (mds_log_op_setattr(obd, inode, lmm, lmm_size,
 565                                                logcookies, cookie_size) <= 0) {
 566                                 OBD_FREE(logcookies, cookie_size);
 567                                 logcookies = NULL;
 568                         }
 569                 }
 570         }
 571
 572         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 573             rec->ur_eadata != NULL) {
 574                 struct lov_stripe_md *lsm = NULL;
 575                 struct lov_user_md *lum = NULL;
 576
 577                 rc = ll_permission(inode, MAY_WRITE, NULL);
 578                 if (rc < 0)
 579                         GOTO(cleanup, rc);
 580
 581                 lum = rec->ur_eadata;
 582                 /* if { size, offset, count } = { 0, -1, 0 } (i.e. all default
 583                  * values specified) then delete default striping from dir. */
 584                 if (S_ISDIR(inode->i_mode) &&
 585                     ((lum->lmm_stripe_size == 0 &&
 586                       lum->lmm_stripe_offset == (typeof(lum->lmm_stripe_offset))(-1) &&
 587                       lum->lmm_stripe_count == 0) ||
 588                     /* lmm_stripe_size == -1 is deprecated in 1.4.6 */
 589                     lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1))){
 590                         rc = fsfilt_set_md(obd, inode, handle, NULL, 0, "lov");
 591                         if (rc)
 592                                 GOTO(cleanup, rc);
 593                 } else {
 594                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 595                                            mds->mds_osc_exp, 0,
 596                                            &lsm, rec->ur_eadata);
 597                         if (rc)
 598                                 GOTO(cleanup, rc);
 599
 600                         obd_free_memmd(mds->mds_osc_exp, &lsm);
 601
 602                         rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 603                                            rec->ur_eadatalen, "lov");
 604                         if (rc)
 605                                 GOTO(cleanup, rc);
 606                 }
 607         }
 608
 609         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 610         mds_pack_inode2fid(&body->fid1, inode);
 611         mds_pack_inode2body(body, inode);
 612
 613         /* don't return OST-specific attributes if we didn't just set them. */
 614         if (ia_valid & ATTR_SIZE)
 615                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 616         if (ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 617                 body->valid |= OBD_MD_FLMTIME;
 618         if (ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 619                 body->valid |= OBD_MD_FLATIME;
 620
 621         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 622                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 623                           rec->ur_eadatalen);
 624                 if (mlcd) {
 625                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 626                                 rec->ur_eadatalen;
 627                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 628                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 629                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 630                                 mlcd->mlcd_cookielen;
 631                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 632                                mlcd->mlcd_cookielen);
 633                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 634                                mlcd->mlcd_eadatalen);
 635                 } else {
 636                         CERROR("unable to allocate log cancel data\n");
 637                 }
 638         }
 639         EXIT;
 640  cleanup:
 641         if (mlcd != NULL)
 642                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 643                                       mds_cancel_cookies_cb, mlcd);
 644         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 645         /* do mds to ost setattr if needed */
 646         if (!rc && !err && lmm_size)
 647                 mds_osc_setattr_async(obd, inode, lmm, lmm_size,
 648                                       logcookies, rec->ur_fid1);
 649
 650         switch (cleanup_phase) {
 651         case 2:
 652                 OBD_FREE(lmm, mds->mds_max_mdsize);
 653                 if (logcookies)
 654                         OBD_FREE(logcookies, cookie_size);
 655         case 1:
 656                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 657                     rec->ur_eadata != NULL)
 658                         UNLOCK_INODE_MUTEX(inode);
 659                 l_dput(de);
 660                 if (locked) {
 661                         if (rc) {
 662                                 ldlm_lock_decref(&lockh, LCK_EX);
 663                         } else {
 664                                 ptlrpc_save_lock (req, &lockh, LCK_EX);
 665                         }
 666                 }
 667         case 0:
 668                 break;
 669         default:
 670                 LBUG();
 671         }
 672         if (err && !rc)
 673                 rc = err;
 674
 675         req->rq_status = rc;
 676
 677         /* trigger dqrel/dqacq for original owner and new owner */
 678         if (ia_valid & (ATTR_UID | ATTR_GID))
 679                 lquota_adjust(quota_interface, obd, qcids, qpids, rc,
 680                               FSFILT_OP_SETATTR);
 681
 682         return 0;
 683 }
 684
 685 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 686                                      struct ptlrpc_request *req)
 687 {
 688         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 689         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 690         struct dentry *parent, *child;
 691         struct mds_body *body;
 692
 693         mds_req_from_mcd(req, med->med_mcd);
 694
 695         if (req->rq_status)
 696                 return;
 697
 698         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 699         LASSERT(!IS_ERR(parent));
 700         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 701         LASSERT(!IS_ERR(child));
 702         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 703         mds_pack_inode2fid(&body->fid1, child->d_inode);
 704         mds_pack_inode2body(body, child->d_inode);
 705         l_dput(parent);
 706         l_dput(child);
 707 }
 708
 709 static int mds_reint_create(struct mds_update_record *rec, int offset,
 710                             struct ptlrpc_request *req,
 711                             struct lustre_handle *lh)
 712 {
 713         struct dentry *dparent = NULL;
 714         struct mds_obd *mds = mds_req2mds(req);
 715         struct obd_device *obd = req->rq_export->exp_obd;
 716         struct dentry *dchild = NULL;
 717         struct inode *dir = NULL;
 718         void *handle = NULL;
 719         struct lustre_handle lockh;
 720         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 721         int created = 0;
 722         unsigned int qcids[MAXQUOTAS] = {current->fsuid, current->fsgid};
 723         unsigned int qpids[MAXQUOTAS] = {0, 0};
 724         struct dentry_params dp;
 725         ENTRY;
 726
 727         LASSERT(offset == MDS_REQ_REC_OFF);
 728         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
 729                         LUSTRE_MDS_NAME));
 730
 731         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 732                   rec->ur_fid1->id, rec->ur_fid1->generation,
 733                   rec->ur_name, rec->ur_mode);
 734
 735         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 736
 737         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 738                 GOTO(cleanup, rc = -ESTALE);
 739
 740         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX, &lockh,
 741                                         MDS_INODELOCK_UPDATE);
 742         if (IS_ERR(dparent)) {
 743                 rc = PTR_ERR(dparent);
 744                 if (rc != -ENOENT)
 745                         CERROR("parent "LPU64"/%u lookup error %d\n",
 746                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 747                 GOTO(cleanup, rc);
 748         }
 749         cleanup_phase = 1; /* locked parent dentry */
 750         dir = dparent->d_inode;
 751         LASSERT(dir);
 752
 753         ldlm_lock_dump_handle(D_OTHER, &lockh);
 754
 755         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 756         if (IS_ERR(dchild)) {
 757                 rc = PTR_ERR(dchild);
 758                 if (rc != -ENAMETOOLONG)
 759                 CERROR("child lookup error %d\n", rc);
 760                 GOTO(cleanup, rc);
 761         }
 762
 763         cleanup_phase = 2; /* child dentry */
 764
 765         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 766
 767         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
 768                 if (dchild->d_inode)
 769                         GOTO(cleanup, rc = -EEXIST);
 770                 GOTO(cleanup, rc = -EROFS);
 771         }
 772
 773         if (dir->i_mode & S_ISGID && S_ISDIR(rec->ur_mode))
 774                 rec->ur_mode |= S_ISGID;
 775
 776         dchild->d_fsdata = (void *)&dp;
 777         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 778         dp.p_ptr = req;
 779
 780         switch (type) {
 781         case S_IFREG:{
 782                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 783                 if (IS_ERR(handle))
 784                         GOTO(cleanup, rc = PTR_ERR(handle));
 785                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 786                 EXIT;
 787                 break;
 788         }
 789         case S_IFDIR:{
 790                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 791                 if (IS_ERR(handle))
 792                         GOTO(cleanup, rc = PTR_ERR(handle));
 793                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 794                 EXIT;
 795                 break;
 796         }
 797         case S_IFLNK:{
 798                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 799                 if (IS_ERR(handle))
 800                         GOTO(cleanup, rc = PTR_ERR(handle));
 801                 if (rec->ur_tgt == NULL)        /* no target supplied */
 802                         rc = -EINVAL;           /* -EPROTO? */
 803                 else
 804                         rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
 805                 EXIT;
 806                 break;
 807         }
 808         case S_IFCHR:
 809         case S_IFBLK:
 810         case S_IFIFO:
 811         case S_IFSOCK:{
 812                 int rdev = rec->ur_rdev;
 813                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 814                 if (IS_ERR(handle))
 815                         GOTO(cleanup, rc = PTR_ERR(handle));
 816                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 817                 EXIT;
 818                 break;
 819         }
 820         default:
 821                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 822                 dchild->d_fsdata = NULL;
 823                 GOTO(cleanup, rc = -EINVAL);
 824         }
 825
 826         /* In case we stored the desired inum in here, we want to clean up. */
 827         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 828                 dchild->d_fsdata = NULL;
 829
 830         if (rc) {
 831                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 832                 GOTO(cleanup, rc);
 833         } else {
 834                 struct iattr iattr;
 835                 struct inode *inode = dchild->d_inode;
 836                 struct mds_body *body;
 837
 838                 created = 1;
 839                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 840                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 841                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 842                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
 843                 if (dir->i_mode & S_ISGID)
 844                         iattr.ia_gid = dir->i_gid;
 845                 else
 846                         iattr.ia_gid = current->fsgid;
 847                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 848                         ATTR_MTIME | ATTR_CTIME;
 849
 850                 if (rec->ur_fid2->id) {
 851                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 852                         inode->i_generation = rec->ur_fid2->generation;
 853                         /* Dirtied and committed by the upcoming setattr. */
 854                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 855                                inode->i_ino, inode->i_generation);
 856                 } else {
 857                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 858                                inode->i_ino, inode->i_generation);
 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                 if (S_ISDIR(inode->i_mode)) {
 871                         struct lov_mds_md lmm;
 872                         int lmm_size = sizeof(lmm);
 873                         rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1);
 874                         if (rc > 0) {
 875                                 LOCK_INODE_MUTEX(inode);
 876                                 rc = fsfilt_set_md(obd, inode, handle,
 877                                                    &lmm, lmm_size, "lov");
 878                                 UNLOCK_INODE_MUTEX(inode);
 879                         }
 880                         if (rc)
 881                                 CERROR("error on copy stripe info: rc = %d\n",
 882                                         rc);
 883                 }
 884
 885                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 886                 mds_pack_inode2fid(&body->fid1, inode);
 887                 mds_pack_inode2body(body, inode);
 888         }
 889         EXIT;
 890
 891 cleanup:
 892         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 893
 894         if (rc && created) {
 895                 /* Destroy the file we just created.  This should not need
 896                  * extra journal credits, as we have already modified all of
 897                  * the blocks needed in order to create the file in the first
 898                  * place.
 899                  */
 900                 switch (type) {
 901                 case S_IFDIR:
 902                         err = vfs_rmdir(dir, dchild);
 903                         if (err)
 904                                 CERROR("rmdir in error path: %d\n", err);
 905                         break;
 906                 default:
 907                         err = vfs_unlink(dir, dchild);
 908                         if (err)
 909                                 CERROR("unlink in error path: %d\n", err);
 910                         break;
 911                 }
 912         } else if (created) {
 913                 /* The inode we were allocated may have just been freed
 914                  * by an unlink operation.  We take this lock to
 915                  * synchronize against the matching reply-ack-lock taken
 916                  * in unlink, to avoid replay problems if this reply
 917                  * makes it out to the client but the unlink's does not.
 918                  * See bug 2029 for more detail.*/
 919                 mds_lock_new_child(obd, dchild->d_inode, NULL);
 920                 /* save uid/gid of create inode and parent */
 921                 qpids[USRQUOTA] = dir->i_uid;
 922                 qpids[GRPQUOTA] = dir->i_gid;
 923         } else {
 924                 rc = err;
 925         }
 926
 927         switch (cleanup_phase) {
 928         case 2: /* child dentry */
 929                 l_dput(dchild);
 930         case 1: /* locked parent dentry */
 931                 if (rc) {
 932                         ldlm_lock_decref(&lockh, LCK_EX);
 933                 } else {
 934                         ptlrpc_save_lock (req, &lockh, LCK_EX);
 935                 }
 936                 l_dput(dparent);
 937         case 0:
 938                 break;
 939         default:
 940                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 941                 LBUG();
 942         }
 943         req->rq_status = rc;
 944
 945         /* trigger dqacq on the owner of child and parent */
 946         lquota_adjust(quota_interface, obd, qcids, qpids, rc, FSFILT_OP_CREATE);
 947         return 0;
 948 }
 949
 950 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 951            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 952 {
 953         int i;
 954
 955         for (i = 0; i < RES_NAME_SIZE; i++) {
 956                 /* return 1 here, because enqueue_ordered will skip resources
 957                  * of all zeroes if they're sorted to the end of the list. */
 958                 if (res1->name[i] == 0 && res2->name[i] != 0)
 959                         return 1;
 960                 if (res2->name[i] == 0 && res1->name[i] != 0)
 961                         return 0;
 962
 963                 if (res1->name[i] > res2->name[i])
 964                         return 1;
 965                 if (res1->name[i] < res2->name[i])
 966                         return 0;
 967         }
 968         if (!p1 || !p2)
 969                 return 0;
 970         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 971                 return 1;
 972         return 0;
 973 }
 974
 975 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 976  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 977  * because they take the place of local semaphores.
 978  *
 979  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 980  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 981 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 982                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 983                           ldlm_policy_data_t *p1_policy,
 984                           struct ldlm_res_id *p2_res_id,
 985                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 986                           ldlm_policy_data_t *p2_policy)
 987 {
 988         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 989         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 990         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 991         ldlm_policy_data_t *policies[2] = {p1_policy, p2_policy};
 992         int rc, flags;
 993         ENTRY;
 994
 995         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 996
 997         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
 998                res_id[0]->name[0], res_id[1]->name[0]);
 999
1000         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1001                 handles[1] = p1_lockh;
1002                 handles[0] = p2_lockh;
1003                 res_id[1] = p1_res_id;
1004                 res_id[0] = p2_res_id;
1005                 lock_modes[1] = p1_lock_mode;
1006                 lock_modes[0] = p2_lock_mode;
1007                 policies[1] = p1_policy;
1008                 policies[0] = p2_policy;
1009         }
1010
1011         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1012                res_id[0]->name[0], res_id[1]->name[0]);
1013
1014         flags = LDLM_FL_LOCAL_ONLY;
1015         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1016                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1017                               ldlm_blocking_ast, ldlm_completion_ast,
1018                               NULL, NULL, NULL, 0, NULL, handles[0]);
1019         if (rc != ELDLM_OK)
1020                 RETURN(-EIO);
1021         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1022
1023         if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0 &&
1024             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1025                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1026                 ldlm_lock_addref(handles[1], lock_modes[1]);
1027         } else if (res_id[1]->name[0] != 0) {
1028                 flags = LDLM_FL_LOCAL_ONLY;
1029                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1030                                       *res_id[1], LDLM_IBITS, policies[1],
1031                                       lock_modes[1], &flags,
1032                                       ldlm_blocking_ast, ldlm_completion_ast,
1033                                       NULL, NULL, NULL, 0, NULL, handles[1]);
1034                 if (rc != ELDLM_OK) {
1035                         ldlm_lock_decref(handles[0], lock_modes[0]);
1036                         RETURN(-EIO);
1037                 }
1038                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1039         }
1040
1041         RETURN(0);
1042 }
1043
1044 static inline int res_eq(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
1045 {
1046         return !memcmp(res1, res2, sizeof(*res1));
1047 }
1048
1049 static inline void
1050 try_to_aggregate_locks(struct ldlm_res_id *res1, ldlm_policy_data_t *p1,
1051                         struct ldlm_res_id *res2, ldlm_policy_data_t *p2)
1052 {
1053         if (!res_eq(res1, res2))
1054                 return;
1055         /* XXX: any additional inodebits (to current LOOKUP and UPDATE)
1056          * should be taken with great care here */
1057         p1->l_inodebits.bits |= p2->l_inodebits.bits;
1058 }
1059
1060 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1061                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1062                            ldlm_policy_data_t *p1_policy,
1063                            struct ldlm_res_id *p2_res_id,
1064                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1065                            ldlm_policy_data_t *p2_policy,
1066                            struct ldlm_res_id *c1_res_id,
1067                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1068                            ldlm_policy_data_t *c1_policy,
1069                            struct ldlm_res_id *c2_res_id,
1070                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1071                            ldlm_policy_data_t *c2_policy)
1072 {
1073         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1074                                           c1_res_id, c2_res_id };
1075         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1076                                                  c1_lockh, c2_lockh };
1077         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1078                               c1_lock_mode, c2_lock_mode };
1079         ldlm_policy_data_t *policies[5] = {p1_policy, p2_policy,
1080                                            c1_policy, c2_policy};
1081         int rc, i, j, sorted, flags;
1082         ENTRY;
1083
1084         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1085                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1086                res_id[3]->name[0]);
1087
1088         /* simple insertion sort - we have at most 4 elements */
1089         for (i = 1; i < 4; i++) {
1090                 j = i - 1;
1091                 dlm_handles[4] = dlm_handles[i];
1092                 res_id[4] = res_id[i];
1093                 lock_modes[4] = lock_modes[i];
1094                 policies[4] = policies[i];
1095
1096                 sorted = 0;
1097                 do {
1098                         if (res_gt(res_id[j], res_id[4], policies[j],
1099                                    policies[4])) {
1100                                 dlm_handles[j + 1] = dlm_handles[j];
1101                                 res_id[j + 1] = res_id[j];
1102                                 lock_modes[j + 1] = lock_modes[j];
1103                                 policies[j + 1] = policies[j];
1104                                 j--;
1105                         } else {
1106                                 sorted = 1;
1107                         }
1108                 } while (j >= 0 && !sorted);
1109
1110                 dlm_handles[j + 1] = dlm_handles[4];
1111                 res_id[j + 1] = res_id[4];
1112                 lock_modes[j + 1] = lock_modes[4];
1113                 policies[j + 1] = policies[4];
1114         }
1115
1116         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1117                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1118                res_id[3]->name[0]);
1119
1120         /* XXX we could send ASTs on all these locks first before blocking? */
1121         for (i = 0; i < 4; i++) {
1122                 flags = 0;
1123                 if (res_id[i]->name[0] == 0)
1124                         break;
1125                 if (i && res_eq(res_id[i], res_id[i-1])) {
1126                         memcpy(dlm_handles[i], dlm_handles[i-1],
1127                                sizeof(*(dlm_handles[i])));
1128                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1129                 } else {
1130                         /* we need to enqueue locks with different inodebits
1131                          * at once, because otherwise concurrent thread can
1132                          * hit the windown between these two locks and we'll
1133                          * get to deadlock. see bug 10360. note also, that it
1134                          * is impossible to have >2 equal res. */
1135                         if (i < 3)
1136                                 try_to_aggregate_locks(res_id[i], policies[i],
1137                                                        res_id[i+1], policies[i+1]);
1138                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1139                                               *res_id[i], LDLM_IBITS,
1140                                               policies[i],
1141                                               lock_modes[i], &flags,
1142                                               ldlm_blocking_ast,
1143                                               ldlm_completion_ast, NULL, NULL,
1144                                               NULL, 0, NULL, dlm_handles[i]);
1145                         if (rc != ELDLM_OK)
1146                                 GOTO(out_err, rc = -EIO);
1147                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1148                 }
1149         }
1150
1151         RETURN(0);
1152 out_err:
1153         while (i-- > 0)
1154                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1155
1156         return rc;
1157 }
1158
1159 /* In the unlikely case that the child changed while we were waiting
1160  * on the lock, we need to drop the lock on the old child and either:
1161  * - if the child has a lower resource name, then we have to also
1162  *   drop the parent lock and regain the locks in the right order
1163  * - in the rename case, if the child has a lower resource name than one of
1164  *   the other parent/child resources (maxres) we also need to reget the locks
1165  * - if the child has a higher resource name (this is the common case)
1166  *   we can just get the lock on the new child (still in lock order)
1167  *
1168  * Returns 0 if the child did not change or if it changed but could be locked.
1169  * Returns 1 if the child changed and we need to re-lock (no locks held).
1170  * Returns -ve error with a valid dchild (no locks held). */
1171 static int mds_verify_child(struct obd_device *obd,
1172                             struct ldlm_res_id *parent_res_id,
1173                             struct lustre_handle *parent_lockh,
1174                             struct dentry *dparent, int parent_mode,
1175                             struct ldlm_res_id *child_res_id,
1176                             struct lustre_handle *child_lockh,
1177                             struct dentry **dchildp, int child_mode,
1178                             ldlm_policy_data_t *child_policy,
1179                             const char *name, int namelen,
1180                             struct ldlm_res_id *maxres)
1181 {
1182         struct dentry *vchild, *dchild = *dchildp;
1183         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1184         ENTRY;
1185
1186         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1187         if (IS_ERR(vchild))
1188                 GOTO(cleanup, rc = PTR_ERR(vchild));
1189
1190         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1191                    (vchild->d_inode != NULL &&
1192                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1193                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1194                 if (dchild != NULL)
1195                         l_dput(dchild);
1196                 *dchildp = vchild;
1197
1198                 RETURN(0);
1199         }
1200
1201         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1202                vchild->d_inode, dchild ? dchild->d_inode : 0,
1203                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1204                child_res_id->name[0]);
1205         if (child_res_id->name[0] != 0)
1206                 ldlm_lock_decref(child_lockh, child_mode);
1207         if (dchild)
1208                 l_dput(dchild);
1209
1210         cleanup_phase = 1; /* parent lock only */
1211         *dchildp = dchild = vchild;
1212
1213         if (dchild->d_inode) {
1214                 int flags = 0;
1215                 child_res_id->name[0] = dchild->d_inode->i_ino;
1216                 child_res_id->name[1] = dchild->d_inode->i_generation;
1217
1218                 /* Make sure that we don't try to re-enqueue a lock on the
1219                  * same resource if it happens that the source is renamed to
1220                  * the target by another thread (bug 9974, thanks racer :-) */
1221                 if (!res_gt(child_res_id, parent_res_id, NULL, NULL) ||
1222                     !res_gt(child_res_id, maxres, NULL, NULL)) {
1223                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1224                                child_res_id->name[0], parent_res_id->name[0],
1225                                maxres->name[0]);
1226                         GOTO(cleanup, rc = 1);
1227                 }
1228
1229                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1230                                       *child_res_id, LDLM_IBITS, child_policy,
1231                                       child_mode, &flags, ldlm_blocking_ast,
1232                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1233                                       NULL, child_lockh);
1234                 if (rc != ELDLM_OK)
1235                         GOTO(cleanup, rc = -EIO);
1236         } else {
1237                 memset(child_res_id, 0, sizeof(*child_res_id));
1238         }
1239
1240         EXIT;
1241 cleanup:
1242         if (rc) {
1243                 switch(cleanup_phase) {
1244                 case 2:
1245                         if (child_res_id->name[0] != 0)
1246                                 ldlm_lock_decref(child_lockh, child_mode);
1247                 case 1:
1248                         ldlm_lock_decref(parent_lockh, parent_mode);
1249                 }
1250         }
1251         return rc;
1252 }
1253
1254 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1255                                 struct ll_fid *fid,
1256                                 struct lustre_handle *parent_lockh,
1257                                 struct dentry **dparentp, int parent_mode,
1258                                 __u64 parent_lockpart,
1259                                 char *name, int namelen,
1260                                 struct lustre_handle *child_lockh,
1261                                 struct dentry **dchildp, int child_mode,
1262                                 __u64 child_lockpart)
1263 {
1264         struct ldlm_res_id child_res_id = { .name = {0} };
1265         struct ldlm_res_id parent_res_id = { .name = {0} };
1266         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1267         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1268         struct inode *inode;
1269         int rc = 0, cleanup_phase = 0;
1270         ENTRY;
1271
1272         /* Step 1: Lookup parent */
1273         *dparentp = mds_fid2dentry(mds, fid, NULL);
1274         if (IS_ERR(*dparentp)) {
1275                 rc = PTR_ERR(*dparentp);
1276                 *dparentp = NULL;
1277                 RETURN(rc);
1278         }
1279
1280         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1281                (*dparentp)->d_inode->i_ino, name);
1282
1283         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1284         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1285
1286         cleanup_phase = 1; /* parent dentry */
1287
1288         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1289         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1290         if (IS_ERR(*dchildp)) {
1291                 rc = PTR_ERR(*dchildp);
1292                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1293                 GOTO(cleanup, rc);
1294         }
1295
1296         cleanup_phase = 2; /* child dentry */
1297         inode = (*dchildp)->d_inode;
1298         if (inode != NULL) {
1299                 if (is_bad_inode(inode)) {
1300                         CERROR("bad inode returned %lu/%u\n",
1301                                inode->i_ino, inode->i_generation);
1302                         GOTO(cleanup, rc = -ENOENT);
1303                 }
1304                 inode = igrab(inode);
1305         }
1306         if (inode == NULL)
1307                 goto retry_locks;
1308
1309         child_res_id.name[0] = inode->i_ino;
1310         child_res_id.name[1] = inode->i_generation;
1311         iput(inode);
1312
1313 retry_locks:
1314         cleanup_phase = 2; /* child dentry */
1315
1316         /* Step 3: Lock parent and child in resource order.  If child doesn't
1317          *         exist, we still have to lock the parent and re-lookup. */
1318         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1319                                    &parent_policy,
1320                                    &child_res_id, child_lockh, child_mode,
1321                                    &child_policy);
1322         if (rc)
1323                 GOTO(cleanup, rc);
1324
1325         if (!(*dchildp)->d_inode)
1326                 cleanup_phase = 3; /* parent lock */
1327         else
1328                 cleanup_phase = 4; /* child lock */
1329
1330         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1331         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1332                               parent_mode, &child_res_id, child_lockh, dchildp,
1333                               child_mode,&child_policy, name, namelen, &parent_res_id);
1334         if (rc > 0)
1335                 goto retry_locks;
1336         if (rc < 0) {
1337                 cleanup_phase = 2;
1338                 GOTO(cleanup, rc);
1339         }
1340
1341 cleanup:
1342         if (rc) {
1343                 switch (cleanup_phase) {
1344                 case 4:
1345                         ldlm_lock_decref(child_lockh, child_mode);
1346                 case 3:
1347                         ldlm_lock_decref(parent_lockh, parent_mode);
1348                 case 2:
1349                         l_dput(*dchildp);
1350                 case 1:
1351                         l_dput(*dparentp);
1352                 default: ;
1353                 }
1354         }
1355         return rc;
1356 }
1357
1358 void mds_reconstruct_generic(struct ptlrpc_request *req)
1359 {
1360         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1361
1362         mds_req_from_mcd(req, med->med_mcd);
1363 }
1364
1365 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1366  * we instead link the inode into the PENDING directory until it is
1367  * finally released.  We can't simply call mds_reint_rename() or some
1368  * part thereof, because we don't have the inode to check for link
1369  * count/open status until after it is locked.
1370  *
1371  * For lock ordering, caller must get child->i_mutex first, then
1372  * pending->i_mutex before starting journal transaction.
1373  *
1374  * returns 1 on success
1375  * returns 0 if we lost a race and didn't make a new link
1376  * returns negative on error
1377  */
1378 static int mds_orphan_add_link(struct mds_update_record *rec,
1379                                struct obd_device *obd, struct dentry *dentry)
1380 {
1381         struct mds_obd *mds = &obd->u.mds;
1382         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1383         struct inode *inode = dentry->d_inode;
1384         struct dentry *pending_child;
1385         char fidname[LL_FID_NAMELEN];
1386         int fidlen = 0, rc, mode;
1387         ENTRY;
1388
1389         LASSERT(inode != NULL);
1390         LASSERT(!mds_inode_is_orphan(inode));
1391 #ifndef HAVE_I_ALLOC_SEM
1392         LASSERT(TRYLOCK_INODE_MUTEX(inode) == 0);
1393 #endif
1394         LASSERT(TRYLOCK_INODE_MUTEX(pending_dir) == 0);
1395
1396         fidlen = mds_fid2str(fidname, inode->i_ino, inode->i_generation);
1397
1398         CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1399                mds_orphan_open_count(inode), inode->i_nlink,
1400                S_ISDIR(inode->i_mode) ? "dir" :
1401                 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1402
1403         if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1404                 RETURN(0);
1405
1406         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1407         if (IS_ERR(pending_child))
1408                 RETURN(PTR_ERR(pending_child));
1409
1410         if (pending_child->d_inode != NULL) {
1411                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1412                 LASSERT(pending_child->d_inode == inode);
1413                 GOTO(out_dput, rc = 0);
1414         }
1415
1416         /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1417          * for linking and return real mode back then -bzzz */
1418         mode = inode->i_mode;
1419         inode->i_mode = S_IFREG;
1420         rc = vfs_link(dentry, pending_dir, pending_child);
1421         if (rc)
1422                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1423                        rec->ur_name, rc);
1424         else
1425                 mds_inode_set_orphan(inode);
1426
1427         /* return mode and correct i_nlink if inode is directory */
1428         inode->i_mode = mode;
1429         LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1430                  S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1431                  inode->i_nlink);
1432         if (S_ISDIR(mode)) {
1433                 inode->i_nlink++;
1434                 pending_dir->i_nlink++;
1435                 mark_inode_dirty(inode);
1436                 mark_inode_dirty(pending_dir);
1437         }
1438
1439         GOTO(out_dput, rc = 1);
1440 out_dput:
1441         l_dput(pending_child);
1442         RETURN(rc);
1443 }
1444
1445 int mds_get_cookie_size(struct obd_device *obd, struct lov_mds_md *lmm)
1446 {
1447         int count = le32_to_cpu(lmm->lmm_stripe_count);
1448         int real_csize = count * sizeof(struct llog_cookie);
1449         return real_csize;
1450 }
1451
1452 void mds_shrink_reply(struct obd_device *obd, struct ptlrpc_request *req,
1453                       struct mds_body *body)
1454 {
1455         int cookie_size = 0, md_size = 0;
1456
1457         if (body && body->valid & OBD_MD_FLEASIZE) {
1458                 md_size = body->eadatasize;
1459         }
1460         if (body && body->valid & OBD_MD_FLCOOKIE) {
1461                 LASSERT(body->valid & OBD_MD_FLEASIZE);
1462                 cookie_size = mds_get_cookie_size(obd, lustre_msg_buf(
1463                                                   req->rq_repmsg, 1, 0));
1464         }
1465
1466         CDEBUG(D_INFO, "Shrink to md_size %d cookie_size %d \n", md_size,
1467                cookie_size);
1468
1469         lustre_shrink_reply(req, 1, md_size, 1);
1470
1471         lustre_shrink_reply(req, md_size? 2:1, cookie_size, 0);
1472 }
1473
1474 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1475                             struct ptlrpc_request *req,
1476                             struct lustre_handle *lh)
1477 {
1478         struct dentry *dparent = NULL, *dchild;
1479         struct mds_obd *mds = mds_req2mds(req);
1480         struct obd_device *obd = req->rq_export->exp_obd;
1481         struct mds_body *body = NULL;
1482         struct inode *child_inode = NULL;
1483         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1484         void *handle = NULL;
1485         int rc = 0, cleanup_phase = 0;
1486         unsigned int qcids [MAXQUOTAS] = {0, 0};
1487         unsigned int qpids [MAXQUOTAS] = {0, 0};
1488         ENTRY;
1489
1490         LASSERT(offset == MDS_REQ_REC_OFF || offset == 2);
1491
1492         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1493                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1494
1495         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1496
1497         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1498                 GOTO(cleanup, rc = -ENOENT);
1499
1500         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1501                                          &parent_lockh, &dparent, LCK_EX,
1502                                          MDS_INODELOCK_UPDATE,
1503                                          rec->ur_name, rec->ur_namelen,
1504                                          &child_lockh, &dchild, LCK_EX,
1505                                          MDS_INODELOCK_FULL);
1506         if (rc)
1507                 GOTO(cleanup, rc);
1508
1509         cleanup_phase = 1; /* dchild, dparent, locks */
1510
1511         dget(dchild);
1512         child_inode = dchild->d_inode;
1513         if (child_inode == NULL) {
1514                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1515                        dparent->d_inode->i_ino, rec->ur_name);
1516                 GOTO(cleanup, rc = -ENOENT);
1517         }
1518
1519         /* save uid/gid for quota acquire/release */
1520         qcids[USRQUOTA] = child_inode->i_uid;
1521         qcids[GRPQUOTA] = child_inode->i_gid;
1522         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1523         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1524
1525         cleanup_phase = 2; /* dchild has a lock */
1526
1527         /* We have to do these checks ourselves, in case we are making an
1528          * orphan.  The client tells us whether rmdir() or unlink() was called,
1529          * so we need to return appropriate errors (bug 72). */
1530         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1531                 if (!S_ISDIR(child_inode->i_mode))
1532                         GOTO(cleanup, rc = -ENOTDIR);
1533         } else {
1534                 if (S_ISDIR(child_inode->i_mode))
1535                         GOTO(cleanup, rc = -EISDIR);
1536         }
1537
1538         /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1539         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1540                 GOTO(cleanup, rc = -EROFS);
1541
1542         /* Step 3: Get a lock on the ino to sync with creation WRT inode
1543          * reuse (see bug 2029). */
1544         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1545         if (rc != ELDLM_OK)
1546                 GOTO(cleanup, rc);
1547
1548         cleanup_phase = 3; /* child inum lock */
1549
1550         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1551
1552         /* ldlm_reply in buf[0] if called via intent */
1553         if (offset)
1554                 offset = 1;
1555
1556         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1557         LASSERT(body != NULL);
1558
1559         /* child orphan sem protects orphan_dec_test && is_orphan race */
1560         MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1561         cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1562
1563         /* If this is potentially the last reference to this inode, get the
1564          * OBD EA data first so the client can destroy OST objects.  We
1565          * only do the object removal later if no open files/links remain. */
1566         if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1567             child_inode->i_nlink == 1) {
1568                 if (mds_orphan_open_count(child_inode) > 0) {
1569                         /* need to lock pending_dir before transaction */
1570                         LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1571                         cleanup_phase = 5; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
1572                 } else if (S_ISREG(child_inode->i_mode)) {
1573                         mds_pack_inode2fid(&body->fid1, child_inode);
1574                         mds_pack_inode2body(body, child_inode);
1575                         mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1576                                     child_inode, MDS_PACK_MD_LOCK);
1577                 }
1578         }
1579
1580         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1581         switch (child_inode->i_mode & S_IFMT) {
1582         case S_IFDIR:
1583                 /* Drop any lingering child directories before we start our
1584                  * transaction, to avoid doing multiple inode dirty/delete
1585                  * in our compound transaction (bug 1321). */
1586                 shrink_dcache_parent(dchild);
1587                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1588                                       NULL);
1589                 if (IS_ERR(handle))
1590                         GOTO(cleanup, rc = PTR_ERR(handle));
1591                 rc = vfs_rmdir(dparent->d_inode, dchild);
1592                 break;
1593         case S_IFREG: {
1594                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1595                                                         offset + 1, 0);
1596                 handle = fsfilt_start_log(obd, dparent->d_inode,
1597                                           FSFILT_OP_UNLINK, NULL,
1598                                           le32_to_cpu(lmm->lmm_stripe_count));
1599                 if (IS_ERR(handle))
1600                         GOTO(cleanup, rc = PTR_ERR(handle));
1601                 rc = vfs_unlink(dparent->d_inode, dchild);
1602                 break;
1603         }
1604         case S_IFLNK:
1605         case S_IFCHR:
1606         case S_IFBLK:
1607         case S_IFIFO:
1608         case S_IFSOCK:
1609                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1610                                       NULL);
1611                 if (IS_ERR(handle))
1612                         GOTO(cleanup, rc = PTR_ERR(handle));
1613                 rc = vfs_unlink(dparent->d_inode, dchild);
1614                 break;
1615         default:
1616                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1617                        rec->ur_name);
1618                 LBUG();
1619                 GOTO(cleanup, rc = -EINVAL);
1620         }
1621
1622         if (rc == 0 && child_inode->i_nlink == 0) {
1623                 if (mds_orphan_open_count(child_inode) > 0)
1624                         rc = mds_orphan_add_link(rec, obd, dchild);
1625
1626                 if (rc == 1)
1627                         GOTO(cleanup, rc = 0);
1628
1629                 if (!S_ISREG(child_inode->i_mode))
1630                         GOTO(cleanup, rc);
1631
1632                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1633                         body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1634                                        OBD_MD_FLATIME | OBD_MD_FLMTIME);
1635                 } else if (mds_log_op_unlink(obd, child_inode,
1636                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1637                                         req->rq_repmsg->buflens[offset + 1],
1638                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1639                                         req->rq_repmsg->buflens[offset+2]) > 0){
1640                         body->valid |= OBD_MD_FLCOOKIE;
1641                 }
1642         }
1643
1644         GOTO(cleanup, rc);
1645 cleanup:
1646         if (rc == 0) {
1647                 struct iattr iattr;
1648                 int err;
1649
1650                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1651                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1652                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1653
1654                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1655                 if (err)
1656                         CERROR("error on parent setattr: rc = %d\n", err);
1657         }
1658
1659         rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
1660                                 handle, req, rc, 0);
1661         if (!rc)
1662                 (void)obd_set_info_async(mds->mds_osc_exp, strlen("unlinked"),
1663                                          "unlinked", 0, NULL, NULL);
1664         switch(cleanup_phase) {
1665         case 5: /* pending_dir semaphore */
1666                 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1667         case 4: /* child inode semaphore */
1668                 MDS_UP_READ_ORPHAN_SEM(child_inode);
1669         case 3: /* child ino-reuse lock */
1670                 if (rc && body != NULL) {
1671                         // Don't unlink the OST objects if the MDS unlink failed
1672                         body->valid = 0;
1673                 }
1674                 if (rc)
1675                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1676                 else
1677                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1678         case 2: /* child lock */
1679                 ldlm_lock_decref(&child_lockh, LCK_EX);
1680         case 1: /* child and parent dentry, parent lock */
1681                 if (rc)
1682                         ldlm_lock_decref(&parent_lockh, LCK_EX);
1683                 else
1684                         ptlrpc_save_lock(req, &parent_lockh, LCK_EX);
1685                 l_dput(dchild);
1686                 l_dput(dchild);
1687                 l_dput(dparent);
1688         case 0:
1689                 break;
1690         default:
1691                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1692                 LBUG();
1693         }
1694         req->rq_status = rc;
1695
1696         mds_shrink_reply(obd, req, body);
1697
1698         /* trigger dqrel on the owner of child and parent */
1699         lquota_adjust(quota_interface, obd, qcids, qpids, rc, FSFILT_OP_UNLINK);
1700         return 0;
1701 }
1702
1703 static int mds_reint_link(struct mds_update_record *rec, int offset,
1704                           struct ptlrpc_request *req,
1705                           struct lustre_handle *lh)
1706 {
1707         struct obd_device *obd = req->rq_export->exp_obd;
1708         struct dentry *de_src = NULL;
1709         struct dentry *de_tgt_dir = NULL;
1710         struct dentry *dchild = NULL;
1711         struct mds_obd *mds = mds_req2mds(req);
1712         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1713         struct ldlm_res_id src_res_id = { .name = {0} };
1714         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1715         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1716         ldlm_policy_data_t tgt_dir_policy =
1717                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1718         int rc = 0, cleanup_phase = 0;
1719         ENTRY;
1720
1721         LASSERT(offset == MDS_REQ_REC_OFF);
1722
1723         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1724                   rec->ur_fid1->id, rec->ur_fid1->generation,
1725                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1726
1727         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1728
1729         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1730                 GOTO(cleanup, rc = -ENOENT);
1731
1732         /* Step 1: Lookup the source inode and target directory by FID */
1733         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1734         if (IS_ERR(de_src))
1735                 GOTO(cleanup, rc = PTR_ERR(de_src));
1736
1737         cleanup_phase = 1; /* source dentry */
1738
1739         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1740         if (IS_ERR(de_tgt_dir)) {
1741                 rc = PTR_ERR(de_tgt_dir);
1742                 de_tgt_dir = NULL;
1743                 GOTO(cleanup, rc);
1744         }
1745
1746         cleanup_phase = 2; /* target directory dentry */
1747
1748         CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
1749                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1750                de_src->d_inode->i_ino);
1751
1752         /* Step 2: Take the two locks */
1753         src_res_id.name[0] = de_src->d_inode->i_ino;
1754         src_res_id.name[1] = de_src->d_inode->i_generation;
1755         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1756         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1757
1758         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1759                                    &src_policy,
1760                                    &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
1761                                    &tgt_dir_policy);
1762         if (rc)
1763                 GOTO(cleanup, rc);
1764
1765         cleanup_phase = 3; /* locks */
1766
1767         if (mds_inode_is_orphan(de_src->d_inode)) {
1768                 CDEBUG(D_INODE, "an attempt to link an orphan inode %lu/%u\n",
1769                        de_src->d_inode->i_ino,
1770                        de_src->d_inode->i_generation);
1771                 GOTO(cleanup, rc = -ENOENT);
1772         }
1773
1774         /* Step 3: Lookup the child */
1775         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1776         if (IS_ERR(dchild)) {
1777                 rc = PTR_ERR(dchild);
1778                 if (rc != -EPERM && rc != -EACCES && rc != -ENAMETOOLONG)
1779                         CERROR("child lookup error %d\n", rc);
1780                 GOTO(cleanup, rc);
1781         }
1782
1783         cleanup_phase = 4; /* child dentry */
1784
1785         if (dchild->d_inode) {
1786                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1787                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1788                 rc = -EEXIST;
1789                 GOTO(cleanup, rc);
1790         }
1791
1792         /* Step 4: Do it. */
1793         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1794
1795         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1796                 GOTO(cleanup, rc = -EROFS);
1797
1798         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1799         if (IS_ERR(handle))
1800                 GOTO(cleanup, rc = PTR_ERR(handle));
1801
1802         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1803         if (rc && rc != -EPERM && rc != -EACCES)
1804                 CERROR("vfs_link error %d\n", rc);
1805 cleanup:
1806         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1807                                 handle, req, rc, 0);
1808         EXIT;
1809
1810         switch (cleanup_phase) {
1811         case 4: /* child dentry */
1812                 l_dput(dchild);
1813         case 3: /* locks */
1814                 if (rc) {
1815                         ldlm_lock_decref(&src_lockh, LCK_EX);
1816                         ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1817                 } else {
1818                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1819                         ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1820                 }
1821         case 2: /* target dentry */
1822                 l_dput(de_tgt_dir);
1823         case 1: /* source dentry */
1824                 l_dput(de_src);
1825         case 0:
1826                 break;
1827         default:
1828                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1829                 LBUG();
1830         }
1831         req->rq_status = rc;
1832         return 0;
1833 }
1834
1835 /* The idea here is that we need to get four locks in the end:
1836  * one on each parent directory, one on each child.  We need to take
1837  * these locks in some kind of order (to avoid deadlocks), and the order
1838  * I selected is "increasing resource number" order.  We need to look up
1839  * the children, however, before we know what the resource number(s) are.
1840  * Thus the following plan:
1841  *
1842  * 1,2. Look up the parents
1843  * 3,4. Look up the children
1844  * 5. Take locks on the parents and children, in order
1845  * 6. Verify that the children haven't changed since they were looked up
1846  *
1847  * If there was a race and the children changed since they were first looked
1848  * up, it is possible that mds_verify_child() will be able to just grab the
1849  * lock on the new child resource (if it has a higher resource than any other)
1850  * but we need to compare against not only its parent, but also against the
1851  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1852  *
1853  * We need the fancy igrab() on the child inodes because we aren't holding a
1854  * lock on the parent after the lookup is done, so dentry->d_inode may change
1855  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1856  */
1857 int mds_get_parents_children_locked(struct obd_device *obd,
1858                                     struct mds_obd *mds,
1859                                     struct ll_fid *p1_fid,
1860                                     struct dentry **de_srcdirp,
1861                                     struct ll_fid *p2_fid,
1862                                     struct dentry **de_tgtdirp,
1863                                     int parent_mode,
1864                                     const char *old_name, int old_len,
1865                                     struct dentry **de_oldp,
1866                                     const char *new_name, int new_len,
1867                                     struct dentry **de_newp,
1868                                     struct lustre_handle *dlm_handles,
1869                                     int child_mode)
1870 {
1871         struct ldlm_res_id p1_res_id = { .name = {0} };
1872         struct ldlm_res_id p2_res_id = { .name = {0} };
1873         struct ldlm_res_id c1_res_id = { .name = {0} };
1874         struct ldlm_res_id c2_res_id = { .name = {0} };
1875         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1876         /* Only dentry should disappear, but the inode itself would be
1877            intact otherwise. */
1878         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
1879         /* If something is going to be replaced, both dentry and inode locks are           needed */
1880         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
1881         struct ldlm_res_id *maxres_src, *maxres_tgt;
1882         struct inode *inode;
1883         int rc = 0, cleanup_phase = 0;
1884         ENTRY;
1885
1886         /* Step 1: Lookup the source directory */
1887         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1888         if (IS_ERR(*de_srcdirp))
1889                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1890
1891         cleanup_phase = 1; /* source directory dentry */
1892
1893         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1894         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1895
1896         /* Step 2: Lookup the target directory */
1897         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1898                 *de_tgtdirp = dget(*de_srcdirp);
1899         } else {
1900                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1901                 if (IS_ERR(*de_tgtdirp)) {
1902                         rc = PTR_ERR(*de_tgtdirp);
1903                         *de_tgtdirp = NULL;
1904                         GOTO(cleanup, rc);
1905                 }
1906         }
1907
1908         cleanup_phase = 2; /* target directory dentry */
1909
1910         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1911         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1912
1913         /* Step 3: Lookup the source child entry */
1914         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1915         if (IS_ERR(*de_oldp)) {
1916                 rc = PTR_ERR(*de_oldp);
1917                 CERROR("old child lookup error (%.*s): %d\n",
1918                        old_len - 1, old_name, rc);
1919                 GOTO(cleanup, rc);
1920         }
1921
1922         cleanup_phase = 3; /* original name dentry */
1923
1924         inode = (*de_oldp)->d_inode;
1925         if (inode != NULL)
1926                 inode = igrab(inode);
1927         if (inode == NULL)
1928                 GOTO(cleanup, rc = -ENOENT);
1929
1930         c1_res_id.name[0] = inode->i_ino;
1931         c1_res_id.name[1] = inode->i_generation;
1932
1933         iput(inode);
1934
1935         /* Step 4: Lookup the target child entry */
1936         if (!new_name)
1937                 GOTO(retry_locks, rc);
1938         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1939         if (IS_ERR(*de_newp)) {
1940                 rc = PTR_ERR(*de_newp);
1941                 if (rc != -ENAMETOOLONG)
1942                 CERROR("new child lookup error (%.*s): %d\n",
1943                        old_len - 1, old_name, rc);
1944                 GOTO(cleanup, rc);
1945         }
1946
1947         cleanup_phase = 4; /* target dentry */
1948
1949         inode = (*de_newp)->d_inode;
1950         if (inode != NULL)
1951                 inode = igrab(inode);
1952         if (inode == NULL)
1953                 goto retry_locks;
1954
1955         c2_res_id.name[0] = inode->i_ino;
1956         c2_res_id.name[1] = inode->i_generation;
1957         iput(inode);
1958
1959 retry_locks:
1960         /* Step 5: Take locks on the parents and child(ren) */
1961         maxres_src = &p1_res_id;
1962         maxres_tgt = &p2_res_id;
1963         cleanup_phase = 4; /* target dentry */
1964
1965         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id,NULL,NULL))
1966                 maxres_src = &c1_res_id;
1967         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id,NULL,NULL))
1968                 maxres_tgt = &c2_res_id;
1969
1970         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1971                                     &p_policy,
1972                                     &p2_res_id, &dlm_handles[1], parent_mode,
1973                                     &p_policy,
1974                                     &c1_res_id, &dlm_handles[2], child_mode,
1975                                     &c1_policy,
1976                                     &c2_res_id, &dlm_handles[3], child_mode,
1977                                     &c2_policy);
1978         if (rc)
1979                 GOTO(cleanup, rc);
1980
1981         cleanup_phase = 6; /* parent and child(ren) locks */
1982
1983         /* Step 6a: Re-lookup source child to verify it hasn't changed */
1984         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1985                               parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
1986                               child_mode, &c1_policy, old_name, old_len,
1987                               maxres_tgt);
1988         if (rc) {
1989                 if (c2_res_id.name[0] != 0)
1990                         ldlm_lock_decref(&dlm_handles[3], child_mode);
1991                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1992                 cleanup_phase = 4;
1993                 if (rc > 0)
1994                         goto retry_locks;
1995                 GOTO(cleanup, rc);
1996         }
1997
1998         if ((*de_oldp)->d_inode == NULL)
1999                 GOTO(cleanup, rc = -ENOENT);
2000
2001         if (!new_name)
2002                 GOTO(cleanup, rc);
2003         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2004         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2005                               parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
2006                               child_mode, &c2_policy, new_name, new_len,
2007                               maxres_src);
2008         if (rc) {
2009                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2010                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2011                 cleanup_phase = 4;
2012                 if (rc > 0)
2013                         goto retry_locks;
2014                 GOTO(cleanup, rc);
2015         }
2016
2017         EXIT;
2018 cleanup:
2019         if (rc) {
2020                 switch (cleanup_phase) {
2021                 case 6: /* child lock(s) */
2022                         if (c2_res_id.name[0] != 0)
2023                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2024                         if (c1_res_id.name[0] != 0)
2025                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2026                 case 5: /* parent locks */
2027                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2028                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2029                 case 4: /* target dentry */
2030                         l_dput(*de_newp);
2031                 case 3: /* source dentry */
2032                         l_dput(*de_oldp);
2033                 case 2: /* target directory dentry */
2034                         l_dput(*de_tgtdirp);
2035                 case 1: /* source directry dentry */
2036                         l_dput(*de_srcdirp);
2037                 }
2038         }
2039
2040         return rc;
2041 }
2042
2043 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2044                             struct ptlrpc_request *req,
2045                             struct lustre_handle *lockh)
2046 {
2047         struct obd_device *obd = req->rq_export->exp_obd;
2048         struct dentry *de_srcdir = NULL;
2049         struct dentry *de_tgtdir = NULL;
2050         struct dentry *de_old = NULL;
2051         struct dentry *de_new = NULL;
2052         struct inode *old_inode = NULL, *new_inode = NULL;
2053         struct mds_obd *mds = mds_req2mds(req);
2054         struct lustre_handle dlm_handles[4];
2055         struct mds_body *body = NULL;
2056         struct lov_mds_md *lmm = NULL;
2057         int rc = 0, lock_count = 3, cleanup_phase = 0;
2058         void *handle = NULL;
2059         unsigned int qcids[MAXQUOTAS] = {0, 0};
2060         unsigned int qpids[4] = {0, 0, 0, 0};
2061         ENTRY;
2062
2063         LASSERT(offset == MDS_REQ_REC_OFF);
2064
2065         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2066                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2067                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2068
2069         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2070
2071         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2072                                              rec->ur_fid2, &de_tgtdir, LCK_EX,
2073                                              rec->ur_name, rec->ur_namelen,
2074                                              &de_old, rec->ur_tgt,
2075                                              rec->ur_tgtlen, &de_new,
2076                                              dlm_handles, LCK_EX);
2077         if (rc)
2078                 GOTO(cleanup, rc);
2079
2080         cleanup_phase = 1; /* parent(s), children, locks */
2081
2082         old_inode = de_old->d_inode;
2083         new_inode = de_new->d_inode;
2084
2085         if (new_inode != NULL)
2086                 lock_count = 4;
2087
2088         /* sanity check for src inode */
2089         if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
2090             old_inode->i_ino == de_tgtdir->d_inode->i_ino)
2091                 GOTO(cleanup, rc = -EINVAL);
2092
2093         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2094                 GOTO(cleanup, rc = -EROFS);
2095
2096         if (new_inode == NULL)
2097                 goto no_unlink;
2098
2099         igrab(new_inode);
2100         cleanup_phase = 2; /* iput(new_inode) when finished */
2101
2102         /* sanity check for dest inode */
2103         if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
2104             new_inode->i_ino == de_tgtdir->d_inode->i_ino)
2105                 GOTO(cleanup, rc = -EINVAL);
2106
2107         if (old_inode == new_inode)
2108                 GOTO(cleanup, rc = 0);
2109
2110         /* save uids/gids for qunit acquire/release */
2111         qcids[USRQUOTA] = old_inode->i_uid;
2112         qcids[GRPQUOTA] = old_inode->i_gid;
2113         qpids[USRQUOTA] = de_tgtdir->d_inode->i_uid;
2114         qpids[GRPQUOTA] = de_tgtdir->d_inode->i_gid;
2115         qpids[2] = de_srcdir->d_inode->i_uid;
2116         qpids[3] = de_srcdir->d_inode->i_gid;
2117
2118         /* if we are about to remove the target at first, pass the EA of
2119          * that inode to client to perform and cleanup on OST */
2120         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2121         LASSERT(body != NULL);
2122
2123         /* child orphan sem protects orphan_dec_test && is_orphan race */
2124         MDS_DOWN_READ_ORPHAN_SEM(new_inode);
2125         cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
2126
2127         if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
2128             new_inode->i_nlink == 1) {
2129                 if (mds_orphan_open_count(new_inode) > 0) {
2130                         /* need to lock pending_dir before transaction */
2131                         LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2132                         cleanup_phase = 4; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
2133                 } else if (S_ISREG(new_inode->i_mode)) {
2134                         mds_pack_inode2fid(&body->fid1, new_inode);
2135                         mds_pack_inode2body(body, new_inode);
2136                         mds_pack_md(obd, req->rq_repmsg, 1, body, new_inode,
2137                                     MDS_PACK_MD_LOCK);
2138                 }
2139         }
2140
2141 no_unlink:
2142         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2143                        de_srcdir->d_inode->i_sb);
2144
2145 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
2146         /* Check if we are moving old entry into its child. 2.6 does not
2147            check for this in vfs_rename() anymore */
2148         if (is_subdir(de_new, de_old))
2149                 GOTO(cleanup, rc = -EINVAL);
2150 #endif
2151
2152         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
2153         handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
2154                                   NULL, le32_to_cpu(lmm->lmm_stripe_count));
2155
2156         if (IS_ERR(handle))
2157                 GOTO(cleanup, rc = PTR_ERR(handle));
2158
2159         lock_kernel();
2160         de_old->d_fsdata = req;
2161         de_new->d_fsdata = req;
2162
2163         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2164         unlock_kernel();
2165
2166         if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2167                 if (mds_orphan_open_count(new_inode) > 0)
2168                         rc = mds_orphan_add_link(rec, obd, de_new);
2169
2170                 if (rc == 1)
2171                         GOTO(cleanup, rc = 0);
2172
2173                 if (!S_ISREG(new_inode->i_mode))
2174                         GOTO(cleanup, rc);
2175
2176                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2177                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2178                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2179                 } else if (mds_log_op_unlink(obd, new_inode,
2180                                              lustre_msg_buf(req->rq_repmsg,1,0),
2181                                              req->rq_repmsg->buflens[1],
2182                                              lustre_msg_buf(req->rq_repmsg,2,0),
2183                                              req->rq_repmsg->buflens[2]) > 0) {
2184                         body->valid |= OBD_MD_FLCOOKIE;
2185                 }
2186         }
2187
2188         GOTO(cleanup, rc);
2189 cleanup:
2190         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2191                                 handle, req, rc, 0);
2192
2193         switch (cleanup_phase) {
2194         case 4:
2195                 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2196         case 3:
2197                 MDS_UP_READ_ORPHAN_SEM(new_inode);
2198         case 2:
2199                 iput(new_inode);
2200         case 1:
2201                 if (rc) {
2202                         if (lock_count == 4)
2203                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2204                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2205                         ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
2206                         ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
2207                 } else {
2208                         if (lock_count == 4)
2209                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2210                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2211                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_EX);
2212                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_EX);
2213                 }
2214                 l_dput(de_new);
2215                 l_dput(de_old);
2216                 l_dput(de_tgtdir);
2217                 l_dput(de_srcdir);
2218         case 0:
2219                 break;
2220         default:
2221                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2222                 LBUG();
2223         }
2224         req->rq_status = rc;
2225
2226         /* acquire/release qunit */
2227         lquota_adjust(quota_interface, obd, qcids, qpids, rc, FSFILT_OP_RENAME);
2228         return 0;
2229 }
2230
2231 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2232                            struct ptlrpc_request *, struct lustre_handle *);
2233
2234 static mds_reinter reinters[REINT_MAX] = {
2235         [REINT_SETATTR] mds_reint_setattr,
2236         [REINT_CREATE] mds_reint_create,
2237         [REINT_LINK] mds_reint_link,
2238         [REINT_UNLINK] mds_reint_unlink,
2239         [REINT_RENAME] mds_reint_rename,
2240         [REINT_OPEN] mds_open
2241 };
2242
2243 int mds_reint_rec(struct mds_update_record *rec, int offset,
2244                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2245 {
2246         struct obd_device *obd = req->rq_export->exp_obd;
2247         struct mds_obd *mds = &obd->u.mds;
2248         struct lvfs_run_ctxt saved;
2249         int rc;
2250         ENTRY;
2251
2252 #if CRAY_XT3
2253         if (req->rq_uid != LNET_UID_ANY) {
2254                 /* non-root local cluster client
2255                  * NB root's creds are believed... */
2256                 LASSERT (req->rq_uid != 0);
2257                 rec->ur_uc.luc_fsuid = req->rq_uid;
2258                 rec->ur_uc.luc_cap = 0;
2259         }
2260 #endif
2261
2262         /* get group info of this user */
2263         rec->ur_uc.luc_uce = upcall_cache_get_entry(mds->mds_group_hash,
2264                                                     rec->ur_uc.luc_fsuid,
2265                                                     rec->ur_uc.luc_fsgid, 2,
2266                                                     &rec->ur_uc.luc_suppgid1);
2267
2268         if (IS_ERR(rec->ur_uc.luc_uce)) {
2269                 rc = PTR_ERR(rec->ur_uc.luc_uce);
2270                 rec->ur_uc.luc_uce = NULL;
2271                 RETURN(rc);
2272         }
2273
2274         /* checked by unpacker */
2275         LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2276
2277 #if CRAY_XT3
2278         if (rec->ur_uc.luc_uce)
2279                 rec->ur_uc.luc_fsgid = rec->ur_uc.luc_uce->ue_primary;
2280 #endif
2281
2282         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2283         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2284         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2285
2286         upcall_cache_put_entry(mds->mds_group_hash, rec->ur_uc.luc_uce);
2287         RETURN(rc);
2288 }