lustre/mds/mds_reint.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/mds_reint.c
   5  *  Lustre Metadata Server (mds) reintegration routines
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #ifndef EXPORT_SYMTAB
  29 # define EXPORT_SYMTAB
  30 #endif
  31 #define DEBUG_SUBSYSTEM S_MDS
  32
  33 #include <linux/fs.h>
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/obd.h>
  37 #include <linux/lustre_lib.h>
  38 #include <linux/lustre_idl.h>
  39 #include <linux/lustre_mds.h>
  40 #include <linux/lustre_dlm.h>
  41 #include <linux/lustre_log.h>
  42 #include <linux/lustre_fsfilt.h>
  43
  44 #include "mds_internal.h"
  45
  46 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  47                    int error)
  48 {
  49         obd_transno_commit_cb(obd, transno, error);
  50 }
  51
  52 struct mds_logcancel_data {
  53         struct lov_mds_md      *mlcd_lmm;
  54         int                     mlcd_size;
  55         int                     mlcd_cookielen;
  56         int                     mlcd_eadatalen;
  57         struct llog_cookie      mlcd_cookies[0];
  58 };
  59
  60
  61 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  62                                   void *cb_data, int error)
  63 {
  64         struct mds_logcancel_data *mlcd = cb_data;
  65         struct lov_stripe_md *lsm = NULL;
  66         struct llog_ctxt *ctxt;
  67         int rc;
  68
  69         obd_transno_commit_cb(obd, transno, error);
  70
  71         CDEBUG(D_HA, "cancelling %d cookies\n",
  72                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  73
  74         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  75                           mlcd->mlcd_eadatalen);
  76         if (rc < 0) {
  77                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  78                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  79                        rc);
  80         } else {
  81                 ///* XXX 0 normally, SENDNOW for debug */);
  82                 ctxt = llog_get_context(&obd->obd_llogs,
  83                                         mlcd->mlcd_cookies[0].lgc_subsys + 1);
  84                 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
  85                                  sizeof(*mlcd->mlcd_cookies),
  86                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
  87                 if (rc)
  88                         CERROR("error cancelling %d log cookies: rc %d\n",
  89                                (int)(mlcd->mlcd_cookielen /
  90                                      sizeof(*mlcd->mlcd_cookies)), rc);
  91         }
  92
  93         OBD_FREE(mlcd, mlcd->mlcd_size);
  94 }
  95
  96 /* Assumes caller has already pushed us into the kernel context. */
  97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  98                        struct ptlrpc_request *req, int rc, __u32 op_data)
  99 {
 100         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 101         struct mds_client_data *mcd = med->med_mcd;
 102         struct obd_device *obd = req->rq_export->exp_obd;
 103         int err;
 104         __u64 transno;
 105         loff_t off;
 106         int log_pri = D_HA;
 107         ENTRY;
 108
 109         /* if the export has already been failed, we have no last_rcvd slot */
 110         if (req->rq_export->exp_failed) {
 111                 CERROR("committing transaction for disconnected client\n");
 112                 if (handle)
 113                         GOTO(commit, rc);
 114                 RETURN(rc);
 115         }
 116
 117         if (IS_ERR(handle))
 118                 RETURN(rc);
 119
 120         if (handle == NULL) {
 121                 /* if we're starting our own xaction, use our own inode */
 122                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 123                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 124                 if (IS_ERR(handle)) {
 125                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 126                         RETURN(PTR_ERR(handle));
 127                 }
 128         }
 129
 130         off = med->med_off;
 131
 132         transno = req->rq_reqmsg->transno;
 133         if (rc != 0) {
 134                 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
 135         } else if (transno == 0) {
 136                 spin_lock(&mds->mds_transno_lock);
 137                 transno = ++mds->mds_last_transno;
 138                 spin_unlock(&mds->mds_transno_lock);
 139         } else {
 140                 spin_lock(&mds->mds_transno_lock);
 141                 if (transno > mds->mds_last_transno)
 142                         mds->mds_last_transno = transno;
 143                 spin_unlock(&mds->mds_transno_lock);
 144         }
 145         req->rq_repmsg->transno = req->rq_transno = transno;
 146         mcd->mcd_last_transno = cpu_to_le64(transno);
 147         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 148         mcd->mcd_last_result = cpu_to_le32(rc);
 149         mcd->mcd_last_data = cpu_to_le32(op_data);
 150
 151
 152         fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb,
 153                               transno, handle, mds_commit_cb, NULL);
 154
 155         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 156                                   &off, 0);
 157
 158         if (err) {
 159                 log_pri = D_ERROR;
 160                 if (rc == 0)
 161                         rc = err;
 162         }
 163
 164         DEBUG_REQ(log_pri, req,
 165                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 166                   transno, mcd->mcd_uuid, med->med_idx, err);
 167
 168         err = mds_lov_write_objids(obd);
 169         if (err) {
 170                 log_pri = D_ERROR;
 171                 if (rc == 0)
 172                         rc = err;
 173         }
 174         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 175
 176 commit:
 177         err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 178         if (err) {
 179                 CERROR("error committing transaction: %d\n", err);
 180                 if (!rc)
 181                         rc = err;
 182         }
 183
 184         RETURN(rc);
 185 }
 186
 187 /* this gives the same functionality as the code between
 188  * sys_chmod and inode_setattr
 189  * chown_common and inode_setattr
 190  * utimes and inode_setattr
 191  */
 192 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 193 {
 194         time_t now = LTIME_S(CURRENT_TIME);
 195         struct iattr *attr = &rec->ur_iattr;
 196         unsigned int ia_valid = attr->ia_valid;
 197         int error;
 198         ENTRY;
 199
 200         /* only fix up attrs if the client VFS didn't already */
 201         if (!(ia_valid & ATTR_RAW))
 202                 RETURN(0);
 203
 204         if (!(ia_valid & ATTR_CTIME_SET))
 205                 LTIME_S(attr->ia_ctime) = now;
 206         if (!(ia_valid & ATTR_ATIME_SET))
 207                 LTIME_S(attr->ia_atime) = now;
 208         if (!(ia_valid & ATTR_MTIME_SET))
 209                 LTIME_S(attr->ia_mtime) = now;
 210
 211         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 212                 RETURN(-EPERM);
 213
 214         /* times */
 215         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 216                 if (rec->_ur_fsuid != inode->i_uid &&
 217                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 218                         RETURN(error);
 219         }
 220
 221         if (ia_valid & ATTR_SIZE) {
 222                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 223                         RETURN(error);
 224         }
 225
 226         if (ia_valid & ATTR_UID) {
 227                 /* chown */
 228                 error = -EPERM;
 229                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 230                         RETURN(-EPERM);
 231                 if (attr->ia_uid == (uid_t) -1)
 232                         attr->ia_uid = inode->i_uid;
 233                 if (attr->ia_gid == (gid_t) -1)
 234                         attr->ia_gid = inode->i_gid;
 235                 attr->ia_mode = inode->i_mode;
 236                 /*
 237                  * If the user or group of a non-directory has been
 238                  * changed by a non-root user, remove the setuid bit.
 239                  * 19981026 David C Niemi <niemi@tux.org>
 240                  *
 241                  * Changed this to apply to all users, including root,
 242                  * to avoid some races. This is the behavior we had in
 243                  * 2.0. The check for non-root was definitely wrong
 244                  * for 2.2 anyway, as it should have been using
 245                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 246                  */
 247                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 248                     !S_ISDIR(inode->i_mode)) {
 249                         attr->ia_mode &= ~S_ISUID;
 250                         attr->ia_valid |= ATTR_MODE;
 251                 }
 252                 /*
 253                  * Likewise, if the user or group of a non-directory
 254                  * has been changed by a non-root user, remove the
 255                  * setgid bit UNLESS there is no group execute bit
 256                  * (this would be a file marked for mandatory
 257                  * locking).  19981026 David C Niemi <niemi@tux.org>
 258                  *
 259                  * Removed the fsuid check (see the comment above) --
 260                  * 19990830 SD.
 261                  */
 262                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 263                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 264                         attr->ia_mode &= ~S_ISGID;
 265                         attr->ia_valid |= ATTR_MODE;
 266                 }
 267         } else if (ia_valid & ATTR_MODE) {
 268                 int mode = attr->ia_mode;
 269                 /* chmod */
 270                 if (attr->ia_mode == (mode_t) -1)
 271                         attr->ia_mode = inode->i_mode;
 272                 attr->ia_mode =
 273                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 274         }
 275         RETURN(0);
 276 }
 277
 278 void mds_steal_ack_locks(struct ptlrpc_request *req)
 279 {
 280         struct obd_export         *exp = req->rq_export;
 281         struct list_head          *tmp;
 282         struct ptlrpc_reply_state *oldrep;
 283         struct ptlrpc_service     *svc;
 284         struct llog_create_locks  *lcl;
 285         unsigned long              flags;
 286         char                       str[PTL_NALFMT_SIZE];
 287         int                        i;
 288
 289         /* CAVEAT EMPTOR: spinlock order */
 290         spin_lock_irqsave (&exp->exp_lock, flags);
 291         list_for_each (tmp, &exp->exp_outstanding_replies) {
 292                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 293
 294                 if (oldrep->rs_xid != req->rq_xid)
 295                         continue;
 296
 297                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 298                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 299                                 "new %d old %d\n", req->rq_xid,
 300                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 301
 302                 svc = oldrep->rs_srv_ni->sni_service;
 303                 spin_lock (&svc->srv_lock);
 304
 305                 list_del_init (&oldrep->rs_exp_list);
 306
 307                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 308                       " o%d NID %s\n",
 309                       oldrep->rs_nlocks, oldrep,
 310                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 311                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 312
 313                 for (i = 0; i < oldrep->rs_nlocks; i++)
 314                         ptlrpc_save_lock(req,
 315                                          &oldrep->rs_locks[i],
 316                                          oldrep->rs_modes[i]);
 317                 oldrep->rs_nlocks = 0;
 318
 319                 lcl = oldrep->rs_llog_locks;
 320                 oldrep->rs_llog_locks = NULL;
 321                 if (lcl != NULL)
 322                         ptlrpc_save_llog_lock(req, lcl);
 323
 324                 DEBUG_REQ(D_HA, req, "stole locks for");
 325                 ptlrpc_schedule_difficult_reply (oldrep);
 326
 327                 spin_unlock (&svc->srv_lock);
 328                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 329                 return;
 330         }
 331         spin_unlock_irqrestore (&exp->exp_lock, flags);
 332 }
 333
 334 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 335 {
 336         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 337                   mcd->mcd_last_transno, mcd->mcd_last_result);
 338         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 339         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 340
 341         mds_steal_ack_locks(req);
 342 }
 343
 344 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 345                                       int offset, struct ptlrpc_request *req)
 346 {
 347         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 348         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 349         struct dentry *de;
 350         struct mds_body *body;
 351
 352         mds_req_from_mcd(req, med->med_mcd);
 353
 354         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 355         if (IS_ERR(de)) {
 356                 LASSERT(PTR_ERR(de) == req->rq_status);
 357                 return;
 358         }
 359
 360         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 361         mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
 362         mds_pack_inode2body(req2obd(req), body, de->d_inode);
 363
 364         /* Don't return OST-specific attributes if we didn't just set them */
 365         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 366                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 367         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 368                 body->valid |= OBD_MD_FLMTIME;
 369         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 370                 body->valid |= OBD_MD_FLATIME;
 371
 372         l_dput(de);
 373 }
 374
 375 /* In the raw-setattr case, we lock the child inode.
 376  * In the write-back case or if being called from open, the client holds a lock
 377  * already.
 378  *
 379  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 380 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 381                              struct ptlrpc_request *req,
 382                              struct lustre_handle *lh)
 383 {
 384         struct mds_obd *mds = mds_req2mds(req);
 385         struct obd_device *obd = req->rq_export->exp_obd;
 386         struct mds_body *body;
 387         struct dentry *de;
 388         struct inode *inode = NULL;
 389         struct lustre_handle lockh[2] = {{0}, {0}};
 390         int parent_mode;
 391         void *handle = NULL;
 392         struct mds_logcancel_data *mlcd = NULL;
 393         int rc = 0, cleanup_phase = 0, err, locked = 0;
 394         ENTRY;
 395
 396         LASSERT(offset == 1);
 397
 398         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 399                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 400
 401         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 402
 403         MDS_UPDATE_COUNTER(mds, MDS_SETATTR_COUNT);
 404
 405         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 406                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 407                 if (IS_ERR(de))
 408                         GOTO(cleanup, rc = PTR_ERR(de));
 409         } else {
 410                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 411                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
 412                         lockpart |= MDS_INODELOCK_LOOKUP;
 413                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 414                                            lockh, &parent_mode, NULL, 0, lockpart);
 415                 if (IS_ERR(de))
 416                         GOTO(cleanup, rc = PTR_ERR(de));
 417                 locked = 1;
 418         }
 419
 420         cleanup_phase = 1;
 421
 422         inode = de->d_inode;
 423         LASSERT(inode);
 424         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 425             rec->ur_eadata != NULL)
 426                 down(&inode->i_sem);
 427
 428         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 429
 430         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 431         if (IS_ERR(handle))
 432                 GOTO(cleanup, rc = PTR_ERR(handle));
 433
 434         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 435                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 436                        LTIME_S(rec->ur_iattr.ia_mtime),
 437                        LTIME_S(rec->ur_iattr.ia_ctime));
 438         rc = mds_fix_attr(inode, rec);
 439         if (rc)
 440                 GOTO(cleanup, rc);
 441
 442         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 443                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 444                                       (long)&rec->ur_iattr.ia_attr_flags);
 445         else                                            /* setattr */
 446                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 447
 448         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 449             rec->ur_eadata != NULL) {
 450                 struct lov_stripe_md *lsm = NULL;
 451
 452                 rc = ll_permission(inode, MAY_WRITE, NULL);
 453                 if (rc < 0)
 454                         GOTO(cleanup, rc);
 455
 456                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 457                                    mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
 458                 if (rc)
 459                         GOTO(cleanup, rc);
 460
 461                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 462
 463                 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 464                                    rec->ur_eadatalen);
 465                 if (rc)
 466                         GOTO(cleanup, rc);
 467         }
 468
 469         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 470         mds_pack_inode2fid(obd, &body->fid1, inode);
 471         mds_pack_inode2body(obd, body, inode);
 472
 473         /* Don't return OST-specific attributes if we didn't just set them */
 474         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 475                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 476         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 477                 body->valid |= OBD_MD_FLMTIME;
 478         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 479                 body->valid |= OBD_MD_FLATIME;
 480
 481         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 482                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 483                           rec->ur_eadatalen);
 484                 if (mlcd) {
 485                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 486                                 rec->ur_eadatalen;
 487                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 488                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 489                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 490                                 mlcd->mlcd_cookielen;
 491                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 492                                mlcd->mlcd_cookielen);
 493                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 494                                mlcd->mlcd_eadatalen);
 495                 } else {
 496                         CERROR("unable to allocate log cancel data\n");
 497                 }
 498         }
 499         EXIT;
 500  cleanup:
 501         if (mlcd != NULL)
 502                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 503                                       handle, mds_cancel_cookies_cb, mlcd);
 504         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 505         switch (cleanup_phase) {
 506         case 1:
 507                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 508                     rec->ur_eadata != NULL)
 509                         up(&inode->i_sem);
 510                 l_dput(de);
 511                 if (locked) {
 512 #ifdef S_PDIROPS
 513                         if (lockh[1].cookie != 0)
 514                                 ldlm_lock_decref(lockh + 1, parent_mode);
 515 #endif
 516                         if (rc) {
 517                                 ldlm_lock_decref(lockh, LCK_PW);
 518                         } else {
 519                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 520                         }
 521                 }
 522         case 0:
 523                 break;
 524         default:
 525                 LBUG();
 526         }
 527         if (err && !rc)
 528                 rc = err;
 529
 530         req->rq_status = rc;
 531         return 0;
 532 }
 533
 534 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 535                                      struct ptlrpc_request *req)
 536 {
 537         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 538         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 539         struct dentry *parent, *child;
 540         struct mds_body *body;
 541         ENTRY;
 542
 543         mds_req_from_mcd(req, med->med_mcd);
 544
 545         if (req->rq_status) {
 546                 EXIT;
 547                 return;
 548         }
 549
 550         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 551         LASSERT(!IS_ERR(parent));
 552         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 553         LASSERT(!IS_ERR(child));
 554         if ((child->d_flags & DCACHE_CROSS_REF)) {
 555                 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
 556                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 557                 mds_pack_dentry2fid(&body->fid1, child);
 558                 mds_pack_dentry2body(body, child);
 559                 body->valid |= OBD_MD_MDS;
 560         } else if (child->d_inode == NULL) {
 561                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 562                           rec->ur_fid1->id, rec->ur_fid1->generation,
 563                           rec->ur_name, rec->ur_mode);
 564                 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
 565         } else {
 566                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 567                 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 568                 mds_pack_inode2body(req2obd(req), body, child->d_inode);
 569         }
 570         l_dput(parent);
 571         l_dput(child);
 572         EXIT;
 573 }
 574
 575 static int mds_reint_create(struct mds_update_record *rec, int offset,
 576                             struct ptlrpc_request *req,
 577                             struct lustre_handle *lh)
 578 {
 579         struct dentry *dparent = NULL;
 580         struct mds_obd *mds = mds_req2mds(req);
 581         struct obd_device *obd = req->rq_export->exp_obd;
 582         struct dentry *dchild = NULL;
 583         struct inode *dir = NULL;
 584         void *handle = NULL;
 585         struct lustre_handle lockh[2] = {{0}, {0}};
 586         int parent_mode;
 587         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 588         int created = 0;
 589         struct dentry_params dp;
 590         struct mea *mea = NULL;
 591         int mea_size;
 592         ENTRY;
 593
 594         LASSERT(offset == 1);
 595         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 596
 597         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 598                   rec->ur_fid1->id, rec->ur_fid1->generation,
 599                   rec->ur_name, rec->ur_mode);
 600
 601         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 602
 603
 604         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 605                 GOTO(cleanup, rc = -ESTALE);
 606
 607         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 608                                         &parent_mode, rec->ur_name,
 609                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 610         if (IS_ERR(dparent)) {
 611                 rc = PTR_ERR(dparent);
 612                 CERROR("parent lookup error %d\n", rc);
 613                 GOTO(cleanup, rc);
 614         }
 615         cleanup_phase = 1; /* locked parent dentry */
 616         dir = dparent->d_inode;
 617         LASSERT(dir);
 618
 619         ldlm_lock_dump_handle(D_OTHER, lockh);
 620
 621         /* try to retrieve MEA data for this dir */
 622         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 623         if (rc)
 624                 GOTO(cleanup, rc);
 625
 626         if (mea != NULL) {
 627                 /* dir is already splitted, check is requested filename
 628                  * should live at this MDS or at another one */
 629                 int i;
 630                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 631                 if (mea->mea_master != mea->mea_fids[i].mds) {
 632                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 633                                " should be %d(%d)\n",
 634                                mea->mea_master, dparent->d_inode->i_ino,
 635                                dparent->d_inode->i_generation, rec->ur_name,
 636                                mea->mea_fids[i].mds, i);
 637                         GOTO(cleanup, rc = -ERESTART);
 638                 }
 639         }
 640
 641         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 642         if (IS_ERR(dchild)) {
 643                 rc = PTR_ERR(dchild);
 644                 CERROR("child lookup error %d\n", rc);
 645                 GOTO(cleanup, rc);
 646         }
 647
 648         cleanup_phase = 2; /* child dentry */
 649
 650         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 651
 652         if (type == S_IFREG || type == S_IFDIR) {
 653                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 654                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 655                        obd->obd_name, dparent->d_inode->i_ino,
 656                        dparent->d_inode->i_generation, rc, parent_mode);
 657                 if (rc > 0) {
 658                         /* dir got splitted */
 659                         GOTO(cleanup, rc = -ERESTART);
 660                 } else if (rc < 0) {
 661                         /* error happened during spitting. */
 662                         GOTO(cleanup, rc);
 663                 }
 664         }
 665
 666         if (dir->i_mode & S_ISGID) {
 667                 if (S_ISDIR(rec->ur_mode))
 668                         rec->ur_mode |= S_ISGID;
 669         }
 670
 671         dchild->d_fsdata = (void *)&dp;
 672         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 673         dp.p_ptr = req;
 674
 675         switch (type) {
 676         case S_IFREG:{
 677                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 678                 if (IS_ERR(handle))
 679                         GOTO(cleanup, rc = PTR_ERR(handle));
 680                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 681                 EXIT;
 682                 break;
 683         }
 684         case S_IFDIR:{
 685                 int nstripes = 0;
 686                 int i;
 687
 688                 /* as Peter asked, mkdir() should distribute new directories
 689                  * over the whole cluster in order to distribute namespace
 690                  * processing load. first, we calculate which MDS to use to put
 691                  * new directory's inode in. */
 692                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 693                                       rec->ur_flags);
 694                 if (i == mds->mds_num) {
 695                         /* inode will be created locally */
 696                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 697                         if (IS_ERR(handle))
 698                                 GOTO(cleanup, rc = PTR_ERR(handle));
 699
 700                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 701
 702                         if (rec->ur_eadata)
 703                                 nstripes = *(u16 *)rec->ur_eadata;
 704
 705                         if (rc == 0 && nstripes) {
 706                                 /* we pass LCK_EX to split routine to
 707                                  * signalthat we have exclusive access
 708                                  * to the directory. simple because
 709                                  * nobody knows it already exists -bzzz */
 710                                 rc = mds_try_to_split_dir(obd, dchild,
 711                                                           NULL, nstripes,
 712                                                           LCK_EX);
 713                                 if (rc > 0) {
 714                                         /* dir got splitted */
 715                                         rc = 0;
 716                                 } else if (rc < 0) {
 717                                         /* an error occured during
 718                                          * splitting. */
 719                                         GOTO(cleanup, rc);
 720                                 }
 721                         }
 722                 } else if (!DENTRY_VALID(dchild)) {
 723                         /* inode will be created on another MDS */
 724                         struct obdo *oa = NULL;
 725                         struct mds_body *body;
 726
 727                         /* first, create that inode */
 728                         oa = obdo_alloc();
 729                         LASSERT(oa != NULL);
 730                         oa->o_mds = i;
 731                         oa->o_easize = 0;
 732                         if (rec->ur_eadata) {
 733                                 /* user asks for creating splitted dir */
 734                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 735                         }
 736
 737                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 738                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 739                                         OBD_MD_FLUID | OBD_MD_FLGID);
 740                         oa->o_mode = dir->i_mode;
 741                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 742                                         obd->obd_name, i);
 743                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 744                                 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
 745                                        obd->obd_name, rec->ur_namelen - 1,
 746                                        rec->ur_name, (unsigned) rec->ur_fid2->id,
 747                                        (unsigned) rec->ur_fid2->generation);
 748                                 oa->o_id = rec->ur_fid2->id;
 749                                 oa->o_generation = rec->ur_fid2->generation;
 750                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 751                         }
 752
 753                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 754                         if (rc) {
 755                                 CERROR("can't create remote inode: %d\n", rc);
 756                                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 757                                           rec->ur_fid1->id, rec->ur_fid1->generation,
 758                                           rec->ur_name, rec->ur_mode);
 759                                 obdo_free(oa);
 760                                 GOTO(cleanup, rc);
 761                         }
 762
 763                         /* now, add new dir entry for it */
 764                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 765                         if (IS_ERR(handle)) {
 766                                 obdo_free(oa);
 767                                 GOTO(cleanup, rc = PTR_ERR(handle));
 768                         }
 769                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 770                                                   rec->ur_namelen - 1,
 771                                                   oa->o_id, oa->o_generation,
 772                                                   i);
 773                         LASSERT(rc == 0);
 774
 775                         /* fill reply */
 776                         body = lustre_msg_buf(req->rq_repmsg, 0,
 777                                               sizeof (*body));
 778                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 779                         body->fid1.id = oa->o_id;
 780                         body->fid1.mds = i;
 781                         body->fid1.generation = oa->o_generation;
 782                         obdo_free(oa);
 783                 } else {
 784                         /* requested name exists in the directory */
 785                         rc = -EEXIST;
 786                 }
 787                 EXIT;
 788                 break;
 789         }
 790         case S_IFLNK:{
 791                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 792                 if (IS_ERR(handle))
 793                         GOTO(cleanup, rc = PTR_ERR(handle));
 794                 if (rec->ur_tgt == NULL)        /* no target supplied */
 795                         rc = -EINVAL;           /* -EPROTO? */
 796                 else
 797                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 798                 EXIT;
 799                 break;
 800         }
 801         case S_IFCHR:
 802         case S_IFBLK:
 803         case S_IFIFO:
 804         case S_IFSOCK:{
 805                 int rdev = rec->ur_rdev;
 806                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 807                 if (IS_ERR(handle))
 808                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 809                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 810                 EXIT;
 811                 break;
 812         }
 813         default:
 814                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 815                 dchild->d_fsdata = NULL;
 816                 GOTO(cleanup, rc = -EINVAL);
 817         }
 818
 819         /* In case we stored the desired inum in here, we want to clean up. */
 820         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 821                 dchild->d_fsdata = NULL;
 822
 823         if (rc) {
 824                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 825                 GOTO(cleanup, rc);
 826         } else if (dchild->d_inode) {
 827                 struct iattr iattr;
 828                 struct inode *inode = dchild->d_inode;
 829                 struct mds_body *body;
 830
 831                 created = 1;
 832                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 833                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 834                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 835                 iattr.ia_uid = rec->_ur_fsuid;
 836                 if (dir->i_mode & S_ISGID)
 837                         iattr.ia_gid = dir->i_gid;
 838                 else
 839                         iattr.ia_gid = rec->_ur_fsgid;
 840                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 841                         ATTR_MTIME | ATTR_CTIME;
 842
 843                 if (rec->ur_fid2->id) {
 844                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 845                         inode->i_generation = rec->ur_fid2->generation;
 846                         /* Dirtied and committed by the upcoming setattr. */
 847                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 848                                inode->i_ino, inode->i_generation);
 849                 } else {
 850 #if 0
 851                         struct lustre_handle child_ino_lockh;
 852 #endif
 853
 854                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 855                                inode->i_ino, inode->i_generation);
 856
 857 #if 0
 858                         /* The inode we were allocated may have just been freed
 859                          * by an unlink operation.  We take this lock to
 860                          * synchronize against the matching reply-ack-lock taken
 861                          * in unlink, to avoid replay problems if this reply
 862                          * makes it out to the client but the unlink's does not.
 863                          * See bug 2029 for more detail.*/
 864                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 865                         if (rc != ELDLM_OK) {
 866                                 CERROR("error locking for unlink/create sync: "
 867                                        "%d\n", rc);
 868                         } else {
 869                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 870                         }
 871 #endif
 872                 }
 873
 874                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 875                 if (rc)
 876                         CERROR("error on child setattr: rc = %d\n", rc);
 877
 878                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 879                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 880                 if (rc)
 881                         CERROR("error on parent setattr: rc = %d\n", rc);
 882                 else
 883                         MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
 884
 885                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 886                 mds_pack_inode2fid(obd, &body->fid1, inode);
 887                 mds_pack_inode2body(obd, 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 extra
 896                  * journal credits, as we have already modified all of the
 897                  * blocks needed in order to create the file in the first
 898                  * place. */
 899                 switch (type) {
 900                 case S_IFDIR:
 901                         err = vfs_rmdir(dir, dchild);
 902                         if (err)
 903                                 CERROR("rmdir in error path: %d\n", err);
 904                         break;
 905                 default:
 906                         err = vfs_unlink(dir, dchild);
 907                         if (err)
 908                                 CERROR("unlink in error path: %d\n", err);
 909                         break;
 910                 }
 911         } else {
 912                 rc = err;
 913         }
 914         switch (cleanup_phase) {
 915         case 2: /* child dentry */
 916                 l_dput(dchild);
 917         case 1: /* locked parent dentry */
 918 #ifdef S_PDIROPS
 919                 if (lockh[1].cookie != 0)
 920                         ldlm_lock_decref(lockh + 1, parent_mode);
 921 #endif
 922                 if (rc) {
 923                         ldlm_lock_decref(lockh, LCK_PW);
 924                 } else {
 925                         ptlrpc_save_lock(req, lockh, LCK_PW);
 926                 }
 927                 l_dput(dparent);
 928         case 0:
 929                 break;
 930         default:
 931                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 932                 LBUG();
 933         }
 934         if (mea)
 935                 OBD_FREE(mea, mea_size);
 936         req->rq_status = rc;
 937         return 0;
 938 }
 939
 940 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 941            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 942 {
 943         int i;
 944
 945         for (i = 0; i < RES_NAME_SIZE; i++) {
 946                 /* return 1 here, because enqueue_ordered will skip resources
 947                  * of all zeroes if they're sorted to the end of the list. */
 948                 if (res1->name[i] == 0 && res2->name[i] != 0)
 949                         return 1;
 950                 if (res2->name[i] == 0 && res1->name[i] != 0)
 951                         return 0;
 952
 953                 if (res1->name[i] > res2->name[i])
 954                         return 1;
 955                 if (res1->name[i] < res2->name[i])
 956                         return 0;
 957         }
 958
 959         if (!p1 || !p2)
 960                 return 0;
 961
 962         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 963                 return 1;
 964
 965         return 0;
 966 }
 967
 968 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 969  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 970  * because they take the place of local semaphores.
 971  *
 972  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 973  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 974 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 975                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 976                           ldlm_policy_data_t *p1_policy,
 977                           struct ldlm_res_id *p2_res_id,
 978                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 979                           ldlm_policy_data_t *p2_policy)
 980 {
 981         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 982         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 983         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 984         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 985         int rc, flags;
 986         ENTRY;
 987
 988         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 989
 990         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 991                res_id[1]->name[0]);
 992
 993         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 994                 handles[1] = p1_lockh;
 995                 handles[0] = p2_lockh;
 996                 res_id[1] = p1_res_id;
 997                 res_id[0] = p2_res_id;
 998                 lock_modes[1] = p1_lock_mode;
 999                 lock_modes[0] = p2_lock_mode;
1000                 policies[1] = p1_policy;
1001                 policies[0] = p2_policy;
1002         }
1003
1004         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1005                res_id[0]->name[0], res_id[1]->name[0]);
1006
1007         flags = LDLM_FL_LOCAL_ONLY;
1008         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1009                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1010                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1011                               NULL, 0, NULL, handles[0]);
1012         if (rc != ELDLM_OK)
1013                 RETURN(-EIO);
1014         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1015
1016         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1017             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1018                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1019                 ldlm_lock_addref(handles[1], lock_modes[1]);
1020         } else if (res_id[1]->name[0] != 0) {
1021                 flags = LDLM_FL_LOCAL_ONLY;
1022                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1023                                       *res_id[1], LDLM_IBITS, policies[1],
1024                                       lock_modes[1], &flags, mds_blocking_ast,
1025                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1026                                       NULL, handles[1]);
1027                 if (rc != ELDLM_OK) {
1028                         ldlm_lock_decref(handles[0], lock_modes[0]);
1029                         RETURN(-EIO);
1030                 }
1031                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1032         }
1033
1034         RETURN(0);
1035 }
1036
1037 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1038                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1039                            ldlm_policy_data_t *p1_policy,
1040                            struct ldlm_res_id *p2_res_id,
1041                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1042                            ldlm_policy_data_t *p2_policy,
1043                            struct ldlm_res_id *c1_res_id,
1044                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1045                            ldlm_policy_data_t *c1_policy,
1046                            struct ldlm_res_id *c2_res_id,
1047                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1048                            ldlm_policy_data_t *c2_policy)
1049 {
1050         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1051                                           c1_res_id, c2_res_id };
1052         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1053                                                  c1_lockh, c2_lockh };
1054         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1055                               c1_lock_mode, c2_lock_mode };
1056         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1057                                             c1_policy, c2_policy};
1058         int rc, i, j, sorted, flags;
1059         ENTRY;
1060
1061         CDEBUG(D_DLMTRACE,
1062                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1063                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1064                res_id[3]->name[0]);
1065
1066         /* simple insertion sort - we have at most 4 elements */
1067         for (i = 1; i < 4; i++) {
1068                 j = i - 1;
1069                 dlm_handles[4] = dlm_handles[i];
1070                 res_id[4] = res_id[i];
1071                 lock_modes[4] = lock_modes[i];
1072                 policies[4] = policies[i];
1073
1074                 sorted = 0;
1075                 do {
1076                         if (res_gt(res_id[j], res_id[4], policies[j],
1077                                    policies[4])) {
1078                                 dlm_handles[j + 1] = dlm_handles[j];
1079                                 res_id[j + 1] = res_id[j];
1080                                 lock_modes[j + 1] = lock_modes[j];
1081                                 policies[j + 1] = policies[j];
1082                                 j--;
1083                         } else {
1084                                 sorted = 1;
1085                         }
1086                 } while (j >= 0 && !sorted);
1087
1088                 dlm_handles[j + 1] = dlm_handles[4];
1089                 res_id[j + 1] = res_id[4];
1090                 lock_modes[j + 1] = lock_modes[4];
1091                 policies[j + 1] = policies[4];
1092         }
1093
1094         CDEBUG(D_DLMTRACE,
1095                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1096                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1097                res_id[3]->name[0]);
1098
1099         /* XXX we could send ASTs on all these locks first before blocking? */
1100         for (i = 0; i < 4; i++) {
1101                 flags = 0;
1102                 if (res_id[i]->name[0] == 0)
1103                         break;
1104                 if (i != 0 &&
1105                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1106                     (policies[i]->l_inodebits.bits &
1107                      policies[i-1]->l_inodebits.bits) ) {
1108                         memcpy(dlm_handles[i], dlm_handles[i-1],
1109                                sizeof(*(dlm_handles[i])));
1110                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1111                 } else {
1112                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1113                                               *res_id[i], LDLM_IBITS,
1114                                               policies[i],
1115                                               lock_modes[i], &flags,
1116                                               mds_blocking_ast,
1117                                               ldlm_completion_ast, NULL, NULL,
1118                                               NULL, 0, NULL, dlm_handles[i]);
1119                         if (rc != ELDLM_OK)
1120                                 GOTO(out_err, rc = -EIO);
1121                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1122                 }
1123         }
1124
1125         RETURN(0);
1126 out_err:
1127         while (i-- > 0)
1128                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1129
1130         return rc;
1131 }
1132
1133 /* In the unlikely case that the child changed while we were waiting
1134  * on the lock, we need to drop the lock on the old child and either:
1135  * - if the child has a lower resource name, then we have to also
1136  *   drop the parent lock and regain the locks in the right order
1137  * - in the rename case, if the child has a lower resource name than one of
1138  *   the other parent/child resources (maxres) we also need to reget the locks
1139  * - if the child has a higher resource name (this is the common case)
1140  *   we can just get the lock on the new child (still in lock order)
1141  *
1142  * Returns 0 if the child did not change or if it changed but could be locked.
1143  * Returns 1 if the child changed and we need to re-lock (no locks held).
1144  * Returns -ve error with a valid dchild (no locks held). */
1145 static int mds_verify_child(struct obd_device *obd,
1146                             struct ldlm_res_id *parent_res_id,
1147                             struct lustre_handle *parent_lockh,
1148                             struct dentry *dparent, int parent_mode,
1149                             struct ldlm_res_id *child_res_id,
1150                             struct lustre_handle *child_lockh,
1151                             struct dentry **dchildp, int child_mode,
1152                             ldlm_policy_data_t *child_policy,
1153                             const char *name, int namelen,
1154                             struct ldlm_res_id *maxres)
1155 {
1156         struct dentry *vchild, *dchild = *dchildp;
1157         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1158         ENTRY;
1159
1160         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1161         if (IS_ERR(vchild))
1162                 GOTO(cleanup, rc = PTR_ERR(vchild));
1163
1164         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1165                 if  (child_res_id->name[0] == vchild->d_inum &&
1166                                 child_res_id->name[1] == vchild->d_generation) {
1167                         if (dchild != NULL)
1168                                 l_dput(dchild);
1169                         *dchildp = vchild;
1170                         RETURN(0);
1171                 }
1172                 goto changed;
1173         }
1174
1175         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1176                    (vchild->d_inode != NULL &&
1177                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1178                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1179                 if (dchild != NULL)
1180                         l_dput(dchild);
1181                 *dchildp = vchild;
1182
1183                 RETURN(0);
1184         }
1185
1186 changed:
1187         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1188                vchild->d_inode, dchild ? dchild->d_inode : 0,
1189                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1190                child_res_id->name[0]);
1191         if (child_res_id->name[0] != 0)
1192                 ldlm_lock_decref(child_lockh, child_mode);
1193         if (dchild)
1194                 l_dput(dchild);
1195
1196         cleanup_phase = 1; /* parent lock only */
1197         *dchildp = dchild = vchild;
1198
1199         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1200                 int flags = 0;
1201                 if (dchild->d_inode) {
1202                         child_res_id->name[0] = dchild->d_inode->i_ino;
1203                         child_res_id->name[1] = dchild->d_inode->i_generation;
1204                 } else {
1205                         child_res_id->name[0] = dchild->d_inum;
1206                         child_res_id->name[1] = dchild->d_generation;
1207                 }
1208
1209                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1210                     res_gt(maxres, child_res_id, NULL, NULL)) {
1211                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1212                                child_res_id->name[0], parent_res_id->name[0],
1213                                maxres->name[0]);
1214                         GOTO(cleanup, rc = 1);
1215                 }
1216
1217                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1218                                       *child_res_id, LDLM_IBITS, child_policy,
1219                                       child_mode, &flags, mds_blocking_ast,
1220                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1221                                       NULL, child_lockh);
1222                 if (rc != ELDLM_OK)
1223                         GOTO(cleanup, rc = -EIO);
1224
1225         } else {
1226                 memset(child_res_id, 0, sizeof(*child_res_id));
1227         }
1228
1229         EXIT;
1230 cleanup:
1231         if (rc) {
1232                 switch(cleanup_phase) {
1233                 case 2:
1234                         if (child_res_id->name[0] != 0)
1235                                 ldlm_lock_decref(child_lockh, child_mode);
1236                 case 1:
1237                         ldlm_lock_decref(parent_lockh, parent_mode);
1238                 }
1239         }
1240         return rc;
1241 }
1242
1243 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1244                                 struct ll_fid *fid,
1245                                 struct lustre_handle *parent_lockh,
1246                                 struct dentry **dparentp, int parent_mode,
1247                                 __u64 parent_lockpart, int *update_mode,
1248                                 char *name, int namelen,
1249                                 struct lustre_handle *child_lockh,
1250                                 struct dentry **dchildp, int child_mode,
1251                                 __u64 child_lockpart, void *clone_info)
1252 {
1253         struct ldlm_res_id child_res_id = { .name = {0} };
1254         struct ldlm_res_id parent_res_id = { .name = {0} };
1255         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1256         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1257         struct inode *inode;
1258         int rc = 0, cleanup_phase = 0;
1259         ENTRY;
1260
1261         /* Step 1: Lookup parent */
1262         *dparentp = mds_fid2dentry(mds, fid, NULL);
1263         if (IS_ERR(*dparentp)) {
1264                 rc = PTR_ERR(*dparentp);
1265                 *dparentp = NULL;
1266                 RETURN(rc);
1267         }
1268
1269         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1270                (*dparentp)->d_inode->i_ino, name);
1271
1272         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1273         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1274 #ifdef S_PDIROPS
1275         parent_lockh[1].cookie = 0;
1276         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1277                 struct ldlm_res_id res_id = { .name = {0} };
1278                 ldlm_policy_data_t policy;
1279                 int flags = 0;
1280                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1281                 if (*update_mode) {
1282                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1283                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1284                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1285                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1286                                               res_id, LDLM_IBITS, &policy,
1287                                               *update_mode, &flags,
1288                                               mds_blocking_ast,
1289                                               ldlm_completion_ast,
1290                                               NULL, NULL, NULL, 0, NULL,
1291                                               parent_lockh + 1);
1292                         if (rc != ELDLM_OK)
1293                                 RETURN(-ENOLCK);
1294                 }
1295
1296                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1297                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1298                        (*dparentp)->d_inode->i_ino,
1299                        (*dparentp)->d_inode->i_generation,
1300                        parent_res_id.name[2]);
1301         }
1302 #endif
1303
1304         cleanup_phase = 1; /* parent dentry */
1305 #ifdef CONFIG_SNAPFS
1306         if (clone_info) {
1307                 /*FIXME is there any other FUNC will use d_fsdata,
1308                  *excepet creating inode according inum*/
1309                 (*dparentp)->d_fsdata = clone_info;
1310         }
1311 #endif
1312
1313         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1314         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1315         if (IS_ERR(*dchildp)) {
1316                 rc = PTR_ERR(*dchildp);
1317                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1318                 GOTO(cleanup, rc);
1319         }
1320
1321         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1322                 /* inode lives on another MDS: return * mds/ino/gen
1323                  * and LOOKUP lock. drop possible UPDATE lock! */
1324                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1325                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1326                 child_res_id.name[0] = (*dchildp)->d_inum;
1327                 child_res_id.name[1] = (*dchildp)->d_generation;
1328                 goto retry_locks;
1329         }
1330
1331         inode = (*dchildp)->d_inode;
1332         if (inode != NULL)
1333                 inode = igrab(inode);
1334         if (inode == NULL)
1335                 goto retry_locks;
1336
1337         child_res_id.name[0] = inode->i_ino;
1338         child_res_id.name[1] = inode->i_generation;
1339
1340         iput(inode);
1341
1342 retry_locks:
1343         cleanup_phase = 2; /* child dentry */
1344
1345         /* Step 3: Lock parent and child in resource order.  If child doesn't
1346          *         exist, we still have to lock the parent and re-lookup. */
1347         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1348                                    &parent_policy, &child_res_id, child_lockh,
1349                                    child_mode, &child_policy);
1350         if (rc)
1351                 GOTO(cleanup, rc);
1352
1353         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1354                 cleanup_phase = 4; /* child lock */
1355         else
1356                 cleanup_phase = 3; /* parent lock */
1357
1358         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1359         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1360                               parent_mode, &child_res_id, child_lockh,
1361                               dchildp, child_mode, &child_policy,
1362                               name, namelen, &parent_res_id);
1363         if (rc > 0)
1364                 goto retry_locks;
1365         if (rc < 0) {
1366                 cleanup_phase = 3;
1367                 GOTO(cleanup, rc);
1368         }
1369
1370 cleanup:
1371         if (rc) {
1372                 switch (cleanup_phase) {
1373                 case 4:
1374                         ldlm_lock_decref(child_lockh, child_mode);
1375                 case 3:
1376                         ldlm_lock_decref(parent_lockh, parent_mode);
1377                 case 2:
1378                         l_dput(*dchildp);
1379                 case 1:
1380 #ifdef S_PDIROPS
1381                         if (parent_lockh[1].cookie)
1382                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1383 #endif
1384                         l_dput(*dparentp);
1385                 default: ;
1386                 }
1387         }
1388         return rc;
1389 }
1390
1391 void mds_reconstruct_generic(struct ptlrpc_request *req)
1392 {
1393         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1394
1395         mds_req_from_mcd(req, med->med_mcd);
1396 }
1397
1398 int mds_create_local_dentry(struct mds_update_record *rec,
1399                            struct obd_device *obd)
1400 {
1401         struct mds_obd *mds = &obd->u.mds;
1402         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1403         int fidlen = 0, rc, cleanup_phase = 0;
1404         struct dentry *new_child = NULL;
1405         char *fidname = rec->ur_name;
1406         struct dentry *child = NULL;
1407         struct lustre_handle lockh[2] = {{0}, {0}};
1408         void *handle;
1409         ENTRY;
1410
1411         down(&fids_dir->i_sem);
1412         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1413         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1414                         fidname, (unsigned) rec->ur_fid1->id,
1415                         (unsigned) rec->ur_fid1->generation);
1416
1417         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1418         up(&fids_dir->i_sem);
1419         if (IS_ERR(new_child)) {
1420                 CERROR("can't lookup %s: %d\n", fidname,
1421                                 (int) PTR_ERR(new_child));
1422                 GOTO(cleanup, rc = PTR_ERR(new_child));
1423         }
1424         cleanup_phase = 1;
1425
1426         if (new_child->d_inode != NULL) {
1427                 /* nice. we've already have local dentry! */
1428                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1429                        (unsigned) new_child->d_inode->i_ino,
1430                        (unsigned) new_child->d_inode->i_generation);
1431                 rec->ur_fid1->id = fids_dir->i_ino;
1432                 rec->ur_fid1->generation = fids_dir->i_generation;
1433                 rec->ur_namelen = fidlen + 1;
1434                 GOTO(cleanup, rc = 0);
1435         }
1436
1437         /* new, local dentry will be added soon. we need no aliases here */
1438         d_drop(new_child);
1439
1440         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1441                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1442         } else {
1443                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1444                                               LCK_EX, lockh, NULL, NULL, 0,
1445                                               MDS_INODELOCK_UPDATE);
1446         }
1447
1448         if (IS_ERR(child)) {
1449                 rc = PTR_ERR(child);
1450                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1451                         CERROR("can't get victim: %d\n", rc);
1452                 GOTO(cleanup, rc);
1453         }
1454         cleanup_phase = 2;
1455
1456         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1457         if (IS_ERR(handle))
1458                 GOTO(cleanup, rc = PTR_ERR(handle));
1459
1460         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1461                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1462                                   mds->mds_num);
1463         if (rc)
1464                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1465                        (unsigned long) child->d_inode->i_ino,
1466                        (unsigned long) child->d_inode->i_generation, rc);
1467         else {
1468                 if (S_ISDIR(child->d_inode->i_mode)) {
1469                         fids_dir->i_nlink++;
1470                         mark_inode_dirty(fids_dir);
1471                 }
1472                 mark_inode_dirty(child->d_inode);
1473         }
1474         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1475
1476         rec->ur_fid1->id = fids_dir->i_ino;
1477         rec->ur_fid1->generation = fids_dir->i_generation;
1478         rec->ur_namelen = fidlen + 1;
1479
1480 cleanup:
1481         switch(cleanup_phase) {
1482                 case 2:
1483                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1484                                 ldlm_lock_decref(lockh, LCK_EX);
1485                         dput(child);
1486                 case 1:
1487                         dput(new_child);
1488                 case 0:
1489                        break;
1490         }
1491         RETURN(rc);
1492 }
1493
1494 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1495                                         struct ptlrpc_request *slave)
1496 {
1497         void *cookie, *cookie2;
1498         struct mds_body *body2;
1499         struct mds_body *body;
1500         void *ea, *ea2;
1501         ENTRY;
1502
1503         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1504         LASSERT(body != NULL);
1505
1506         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1507         LASSERT(body2 != NULL);
1508
1509         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1510                 RETURN(0);
1511         }
1512
1513         memcpy(body2, body, sizeof(*body));
1514         body2->valid &= ~OBD_MD_FLCOOKIE;
1515
1516         if (!(body->valid & OBD_MD_FLEASIZE) &&
1517             !(body->valid & OBD_MD_FLDIREA))
1518                 RETURN(0);
1519
1520         if (body->eadatasize == 0) {
1521                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1522                 RETURN(0);
1523         }
1524
1525         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1526
1527         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1528         LASSERT(ea != NULL);
1529
1530         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1531         LASSERT(ea2 != NULL);
1532
1533         memcpy(ea2, ea, body->eadatasize);
1534
1535         if (body->valid & OBD_MD_FLCOOKIE) {
1536                 LASSERT(master->rq_repmsg->buflens[2] >=
1537                                 slave->rq_repmsg->buflens[2]);
1538                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1539                                 slave->rq_repmsg->buflens[2]);
1540                 LASSERT(cookie != NULL);
1541
1542                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1543                                 master->rq_repmsg->buflens[2]);
1544                 LASSERT(cookie2 != NULL);
1545                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1546                 body2->valid |= OBD_MD_FLCOOKIE;
1547         }
1548         RETURN(0);
1549 }
1550
1551 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1552                                    struct ptlrpc_request *req,
1553                                    struct lustre_handle *parent_lockh,
1554                                    int update_mode,
1555                                    struct dentry *dparent,
1556                                    struct lustre_handle *child_lockh,
1557                                    struct dentry *dchild)
1558 {
1559         struct obd_device *obd = req->rq_export->exp_obd;
1560         struct mds_obd *mds = mds_req2mds(req);
1561         struct mdc_op_data op_data;
1562         int rc = 0, cleanup_phase = 0;
1563         struct ptlrpc_request *request = NULL;
1564         void *handle;
1565         ENTRY;
1566
1567         LASSERT(offset == 1 || offset == 3);
1568
1569         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1570                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1571                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1572         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1573                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1574                           rec->ur_namelen - 1, rec->ur_name,
1575                           (unsigned)dchild->d_mdsnum,
1576                           (unsigned) dchild->d_inum,
1577                           (unsigned) dchild->d_generation);
1578
1579         /* time to drop i_nlink on remote MDS */
1580         memset(&op_data, 0, sizeof(op_data));
1581         op_data.fid1.mds = dchild->d_mdsnum;
1582         op_data.fid1.id = dchild->d_inum;
1583         op_data.fid1.generation = dchild->d_generation;
1584         op_data.create_mode = rec->ur_mode;
1585         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1586                 op_data.create_mode |= MDS_MODE_REPLAY;
1587         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1588         cleanup_phase = 2;
1589         if (request) {
1590                 if (rc == 0)
1591                         mds_copy_unlink_reply(req, request);
1592                 ptlrpc_req_finished(request);
1593         }
1594         if (rc == 0) {
1595                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1596                                       NULL);
1597                 if (IS_ERR(handle))
1598                         GOTO(cleanup, rc = PTR_ERR(handle));
1599                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1600                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1601                                         rc, 0);
1602         }
1603 cleanup:
1604         req->rq_status = rc;
1605
1606 #ifdef S_PDIROPS
1607         if (parent_lockh[1].cookie != 0)
1608                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1609 #endif
1610         ldlm_lock_decref(child_lockh, LCK_EX);
1611         if (rc)
1612                 ldlm_lock_decref(parent_lockh, LCK_PW);
1613         else
1614                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1615         l_dput(dchild);
1616         l_dput(dparent);
1617
1618         return 0;
1619 }
1620
1621 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1622                             struct ptlrpc_request *req,
1623                             struct lustre_handle *lh)
1624 {
1625         struct dentry *dparent, *dchild;
1626         struct mds_obd *mds = mds_req2mds(req);
1627         struct obd_device *obd = req->rq_export->exp_obd;
1628         struct mds_body *body = NULL;
1629         struct inode *child_inode;
1630         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1631         struct lustre_handle child_lockh = {0};
1632 #if 0
1633         struct lustre_handle child_reuse_lockh = {0};
1634 #endif
1635         struct lustre_handle * slave_lockh = NULL;
1636         struct llog_create_locks *lcl = NULL;
1637         char fidname[LL_FID_NAMELEN];
1638         void *handle = NULL;
1639         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1640         int unlink_by_fid = 0;
1641         int update_mode;
1642         ENTRY;
1643
1644         LASSERT(offset == 1 || offset == 3);
1645
1646         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1647                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1648
1649         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1650
1651         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1652                 DEBUG_REQ(D_HA, req, "unlink replay\n");
1653                 LASSERT(offset == 1); /* should not come from intent */
1654                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1655                        lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
1656                        req->rq_repmsg->buflens[2]);
1657         }
1658
1659         MDS_UPDATE_COUNTER(mds, MDS_UNLINK_COUNT);
1660
1661         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1662                 GOTO(cleanup, rc = -ENOENT);
1663
1664         if (rec->ur_namelen == 1) {
1665                 /* this is request to drop i_nlink on local inode */
1666                 unlink_by_fid = 1;
1667                 rec->ur_name = fidname;
1668                 rc = mds_create_local_dentry(rec, obd);
1669                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1670                         DEBUG_REQ(D_HA, req,
1671                                   "drop nlink on inode %u/%u/%u (replay)",
1672                                   (unsigned) rec->ur_fid1->mds,
1673                                   (unsigned) rec->ur_fid1->id,
1674                                   (unsigned) rec->ur_fid1->generation);
1675                         req->rq_status = 0;
1676                         RETURN(0);
1677                 }
1678         }
1679
1680         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1681                 /* master mds for directory asks slave removing
1682                  * inode is already locked */
1683                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1684                                                LCK_PW, parent_lockh,
1685                                                &update_mode, rec->ur_name,
1686                                                rec->ur_namelen,
1687                                                MDS_INODELOCK_UPDATE);
1688                 if (IS_ERR(dparent))
1689                         GOTO(cleanup, rc = PTR_ERR(dparent));
1690                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1691                                            rec->ur_namelen - 1);
1692                 if (IS_ERR(dchild))
1693                         GOTO(cleanup, rc = PTR_ERR(dchild));
1694                 child_lockh.cookie = 0;
1695                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1696                 LASSERT(dchild->d_inode != NULL);
1697                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1698         } else {
1699                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1700                                                  parent_lockh, &dparent,
1701                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1702                                                  &update_mode, rec->ur_name,
1703                                                  rec->ur_namelen, &child_lockh,
1704                                                  &dchild, LCK_EX,
1705                                                  MDS_INODELOCK_LOOKUP |
1706                                                  MDS_INODELOCK_UPDATE, NULL);
1707         }
1708         if (rc)
1709                 GOTO(cleanup, rc);
1710
1711         if (dchild->d_flags & DCACHE_CROSS_REF) {
1712                 /* we should have parent lock only here */
1713                 LASSERT(unlink_by_fid == 0);
1714                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1715                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1716                                         update_mode, dparent, &child_lockh, dchild);
1717                 RETURN(0);
1718         }
1719
1720         cleanup_phase = 1; /* dchild, dparent, locks */
1721
1722         dget(dchild);
1723         child_inode = dchild->d_inode;
1724         if (child_inode == NULL) {
1725                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1726                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1727                 GOTO(cleanup, rc = -ENOENT);
1728         }
1729
1730         cleanup_phase = 2; /* dchild has a lock */
1731
1732         /* We have to do these checks ourselves, in case we are making an
1733          * orphan.  The client tells us whether rmdir() or unlink() was called,
1734          * so we need to return appropriate errors (bug 72).
1735          *
1736          * We don't have to check permissions, because vfs_rename (called from
1737          * mds_open_unlink_rename) also calls may_delete. */
1738         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1739                 if (!S_ISDIR(child_inode->i_mode))
1740                         GOTO(cleanup, rc = -ENOTDIR);
1741         } else {
1742                 if (S_ISDIR(child_inode->i_mode))
1743                         GOTO(cleanup, rc = -EISDIR);
1744         }
1745
1746         /* handle splitted dir */
1747         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1748         if (rc)
1749                 GOTO(cleanup, rc);
1750
1751 #if 0
1752         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1753          * reuse (see bug 2029). */
1754         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1755         if (rc != ELDLM_OK)
1756                 GOTO(cleanup, rc);
1757 #endif
1758         cleanup_phase = 3; /* child inum lock */
1759
1760         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1761
1762         /* ldlm_reply in buf[0] if called via intent */
1763         if (offset == 3)
1764                 offset = 1;
1765         else
1766                 offset = 0;
1767
1768         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1769         LASSERT(body != NULL);
1770
1771         /* If this is the last reference to this inode, get the OBD EA
1772          * data first so the client can destroy OST objects.
1773          * we only do the object removal if no open files remain.
1774          * Nobody can get at this name anymore because of the locks so
1775          * we make decisions here as to whether to remove the inode */
1776         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1777             mds_open_orphan_count(child_inode) == 0) {
1778                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1779                 mds_pack_inode2body(obd, body, child_inode);
1780                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1781                             child_inode, 1);
1782                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1783                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1784                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1785                 } else {
1786                         log_unlink = 1;
1787                 }
1788         }
1789
1790         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1791         switch (child_inode->i_mode & S_IFMT) {
1792         case S_IFDIR:
1793                 /* Drop any lingering child directories before we start our
1794                  * transaction, to avoid doing multiple inode dirty/delete
1795                  * in our compound transaction (bug 1321). */
1796                 shrink_dcache_parent(dchild);
1797                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1798                                       NULL);
1799                 if (IS_ERR(handle))
1800                         GOTO(cleanup, rc = PTR_ERR(handle));
1801                 cleanup_phase = 4; /* transaction */
1802                 rc = vfs_rmdir(dparent->d_inode, dchild);
1803                 break;
1804         case S_IFREG: {
1805 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1806                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1807                                                         offset + 1, 0);
1808                 handle = fsfilt_start_log(obd, dparent->d_inode,
1809                                           FSFILT_OP_UNLINK, NULL,
1810                                           le32_to_cpu(lmm->lmm_stripe_count));
1811                 if (IS_ERR(handle))
1812                         GOTO(cleanup, rc = PTR_ERR(handle));
1813
1814                 cleanup_phase = 4; /* transaction */
1815                 rc = vfs_unlink(dparent->d_inode, dchild);
1816
1817                 if (!rc && log_unlink)
1818                         if (mds_log_op_unlink(obd, child_inode,
1819                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1820                                 req->rq_repmsg->buflens[offset + 1],
1821                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1822                                 req->rq_repmsg->buflens[offset + 2], &lcl) > 0)
1823                                 body->valid |= OBD_MD_FLCOOKIE;
1824                 break;
1825         }
1826         case S_IFLNK:
1827         case S_IFCHR:
1828         case S_IFBLK:
1829         case S_IFIFO:
1830         case S_IFSOCK:
1831                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1832                                       NULL);
1833                 if (IS_ERR(handle))
1834                         GOTO(cleanup, rc = PTR_ERR(handle));
1835                 cleanup_phase = 4; /* transaction */
1836                 rc = vfs_unlink(dparent->d_inode, dchild);
1837                 break;
1838         default:
1839                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1840                        rec->ur_name);
1841                 LBUG();
1842                 GOTO(cleanup, rc = -EINVAL);
1843         }
1844
1845  cleanup:
1846         if (rc == 0) {
1847                 struct iattr iattr;
1848                 int err;
1849
1850                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1851                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1852                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1853
1854                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1855                 if (err)
1856                         CERROR("error on parent setattr: rc = %d\n", err);
1857         }
1858
1859         switch(cleanup_phase) {
1860         case 4:
1861                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1862                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1863                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1864                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1865
1866                         /* filesystem is really going to destroy an inode
1867                          * we have to delay this till inode is opened -bzzz */
1868                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1869                 }
1870                 /* handle splitted dir */
1871                 if (rc == 0) {
1872                         /* master directory can be non-empty or something else ... */
1873                         mds_unlink_slave_objs(obd, dchild);
1874                 }
1875                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1876                                         rc, 0);
1877                 if (!rc)
1878                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1879                                            "unlinked", 0, NULL);
1880                 if (lcl != NULL)
1881                         ptlrpc_save_llog_lock(req, lcl);
1882         case 3: /* child ino-reuse lock */
1883 #if 0
1884                 if (rc && body != NULL) {
1885                         // Don't unlink the OST objects if the MDS unlink failed
1886                         body->valid = 0;
1887                 }
1888                 if (rc)
1889                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1890                 else
1891                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1892 #endif
1893         case 2: /* child lock */
1894                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1895                 if (child_lockh.cookie)
1896                         ldlm_lock_decref(&child_lockh, LCK_EX);
1897         case 1: /* child and parent dentry, parent lock */
1898 #ifdef S_PDIROPS
1899                 if (parent_lockh[1].cookie != 0)
1900                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1901 #endif
1902                 if (rc)
1903                         ldlm_lock_decref(parent_lockh, LCK_PW);
1904                 else
1905                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1906                 l_dput(dchild);
1907                 l_dput(dchild);
1908                 l_dput(dparent);
1909         case 0:
1910                 break;
1911         default:
1912                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1913                 LBUG();
1914         }
1915         req->rq_status = rc;
1916         return 0;
1917 }
1918
1919 /*
1920  * to service requests from remote MDS to increment i_nlink
1921  */
1922 static int mds_reint_link_acquire(struct mds_update_record *rec,
1923                                  int offset, struct ptlrpc_request *req,
1924                                  struct lustre_handle *lh)
1925 {
1926         struct obd_device *obd = req->rq_export->exp_obd;
1927         struct ldlm_res_id src_res_id = { .name = {0} };
1928         struct lustre_handle *handle = NULL, src_lockh = {0};
1929         struct mds_obd *mds = mds_req2mds(req);
1930         int rc = 0, cleanup_phase = 0;
1931         struct dentry *de_src = NULL;
1932         ldlm_policy_data_t policy;
1933         int flags = 0;
1934         ENTRY;
1935
1936         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1937                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1938                   (unsigned) rec->ur_fid1->id,
1939                   (unsigned) rec->ur_fid1->generation);
1940
1941         /* Step 1: Lookup the source inode and target directory by FID */
1942         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1943         if (IS_ERR(de_src))
1944                 GOTO(cleanup, rc = PTR_ERR(de_src));
1945         cleanup_phase = 1; /* source dentry */
1946
1947         src_res_id.name[0] = de_src->d_inode->i_ino;
1948         src_res_id.name[1] = de_src->d_inode->i_generation;
1949         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1950
1951         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1952                         src_res_id, LDLM_IBITS, &policy,
1953                         LCK_EX, &flags, mds_blocking_ast,
1954                         ldlm_completion_ast, NULL, NULL,
1955                         NULL, 0, NULL, &src_lockh);
1956         if (rc != ELDLM_OK)
1957                 GOTO(cleanup, rc = -ENOLCK);
1958         cleanup_phase = 2; /* lock */
1959
1960         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1961
1962         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1963         if (IS_ERR(handle)) {
1964                 rc = PTR_ERR(handle);
1965                 GOTO(cleanup, rc);
1966         }
1967         de_src->d_inode->i_nlink++;
1968         mark_inode_dirty(de_src->d_inode);
1969
1970 cleanup:
1971         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1972                                         handle, req, rc, 0);
1973         EXIT;
1974         switch (cleanup_phase) {
1975                 case 2:
1976                         if (rc)
1977                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1978                         else
1979                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1980                 case 1:
1981                         l_dput(de_src);
1982                 case 0:
1983                         break;
1984                 default:
1985                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1986                         LBUG();
1987         }
1988         req->rq_status = rc;
1989         return 0;
1990 }
1991
1992 /*
1993  * request to link to foreign inode:
1994  *  - acquire i_nlinks on this inode
1995  *  - add dentry
1996  */
1997 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1998                                     int offset, struct ptlrpc_request *req,
1999                                     struct lustre_handle *lh)
2000 {
2001         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2002         struct obd_device *obd = req->rq_export->exp_obd;
2003         struct dentry *de_tgt_dir = NULL;
2004         struct mds_obd *mds = mds_req2mds(req);
2005         int rc = 0, cleanup_phase = 0;
2006         struct mdc_op_data op_data;
2007         struct ptlrpc_request *request = NULL;
2008         int update_mode;
2009         ENTRY;
2010
2011 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
2012         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
2013                   (unsigned) rec->ur_fid2->mds,
2014                   (unsigned) rec->ur_fid2->id,
2015                   (unsigned) rec->ur_fid2->generation,
2016                   rec->ur_namelen - 1, rec->ur_name,
2017                   (unsigned) rec->ur_fid1->mds,
2018                   (unsigned) rec->ur_fid1->id,
2019                   (unsigned)rec->ur_fid1->generation);
2020
2021         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
2022                                            tgt_dir_lockh, &update_mode,
2023                                            rec->ur_name, rec->ur_namelen - 1,
2024                                            MDS_INODELOCK_UPDATE);
2025         if (IS_ERR(de_tgt_dir))
2026                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2027         cleanup_phase = 1;
2028
2029         op_data.fid1 = *(rec->ur_fid1);
2030         op_data.namelen = 0;
2031         op_data.name = NULL;
2032         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2033         if (rc)
2034                 GOTO(cleanup, rc);
2035
2036         cleanup_phase = 2;
2037         if (request)
2038                 ptlrpc_req_finished(request);
2039
2040         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2041
2042         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2043         if (IS_ERR(handle)) {
2044                 rc = PTR_ERR(handle);
2045                 GOTO(cleanup, rc);
2046         }
2047
2048         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2049                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2050                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2051         cleanup_phase = 3;
2052
2053 cleanup:
2054         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2055                                 handle, req, rc, 0);
2056         EXIT;
2057
2058         switch (cleanup_phase) {
2059                 case 3:
2060                         if (rc) {
2061                                 /* FIXME: drop i_nlink on remote inode here */
2062                                 CERROR("MUST drop drop i_nlink here\n");
2063                         }
2064                 case 2:
2065                 case 1:
2066                         if (rc) {
2067                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2068 #ifdef S_PDIROPS
2069                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2070 #endif
2071                         } else {
2072                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2073 #ifdef S_PDIROPS
2074                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2075 #endif
2076                         }
2077                         l_dput(de_tgt_dir);
2078                         break;
2079                 default:
2080                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2081                         LBUG();
2082         }
2083         req->rq_status = rc;
2084         return 0;
2085 }
2086
2087 static int mds_reint_link(struct mds_update_record *rec, int offset,
2088                           struct ptlrpc_request *req,
2089                           struct lustre_handle *lh)
2090 {
2091         struct obd_device *obd = req->rq_export->exp_obd;
2092         struct dentry *de_src = NULL;
2093         struct dentry *de_tgt_dir = NULL;
2094         struct dentry *dchild = NULL;
2095         struct mds_obd *mds = mds_req2mds(req);
2096         struct lustre_handle *handle = NULL;
2097         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2098         struct ldlm_res_id src_res_id = { .name = {0} };
2099         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2100         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2101         ldlm_policy_data_t tgt_dir_policy =
2102                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2103         int rc = 0, cleanup_phase = 0;
2104         int update_mode = 0;
2105         ENTRY;
2106
2107         LASSERT(offset == 1);
2108
2109         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2110                   rec->ur_fid1->id, rec->ur_fid1->generation,
2111                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2112
2113         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2114
2115         MDS_UPDATE_COUNTER(mds, MDS_LINK_COUNT);
2116
2117 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2118         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2119                 GOTO(cleanup, rc = -ENOENT);
2120
2121         if (rec->ur_fid1->mds != mds->mds_num) {
2122                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2123                 RETURN(rc);
2124         }
2125
2126         if (rec->ur_namelen == 1) {
2127                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2128                 RETURN(rc);
2129         }
2130
2131         /* Step 1: Lookup the source inode and target directory by FID */
2132         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2133         if (IS_ERR(de_src))
2134                 GOTO(cleanup, rc = PTR_ERR(de_src));
2135
2136         cleanup_phase = 1; /* source dentry */
2137
2138         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2139         if (IS_ERR(de_tgt_dir)) {
2140                 rc = PTR_ERR(de_tgt_dir);
2141                 de_tgt_dir = NULL;
2142                 GOTO(cleanup, rc);
2143         }
2144
2145         cleanup_phase = 2; /* target directory dentry */
2146
2147         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2148                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2149                de_src->d_inode->i_ino);
2150
2151         /* Step 2: Take the two locks */
2152         src_res_id.name[0] = de_src->d_inode->i_ino;
2153         src_res_id.name[1] = de_src->d_inode->i_generation;
2154         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2155         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2156 #ifdef S_PDIROPS
2157         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2158                 int flags = 0;
2159                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2160                 if (update_mode) {
2161                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2162                                               tgt_dir_res_id, LDLM_IBITS,
2163                                               &src_policy, update_mode, &flags,
2164                                               mds_blocking_ast,
2165                                               ldlm_completion_ast, NULL, NULL,
2166                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2167                         if (rc != ELDLM_OK)
2168                                 GOTO(cleanup, rc = -ENOLCK);
2169                 }
2170
2171                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2172                                                         rec->ur_namelen - 1);
2173                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2174                        de_tgt_dir->d_inode->i_ino,
2175                        de_tgt_dir->d_inode->i_generation,
2176                        tgt_dir_res_id.name[2]);
2177         }
2178 #endif
2179         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2180                                    &src_policy,
2181                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2182                                    &tgt_dir_policy);
2183         if (rc)
2184                 GOTO(cleanup, rc);
2185
2186         cleanup_phase = 3; /* locks */
2187
2188         /* Step 3: Lookup the child */
2189         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2190         if (IS_ERR(dchild)) {
2191                 rc = PTR_ERR(dchild);
2192                 if (rc != -EPERM && rc != -EACCES)
2193                         CERROR("child lookup error %d\n", rc);
2194                 GOTO(cleanup, rc);
2195         }
2196
2197         cleanup_phase = 4; /* child dentry */
2198
2199         if (dchild->d_inode) {
2200                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2201                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2202                 rc = -EEXIST;
2203                 GOTO(cleanup, rc);
2204         }
2205
2206         /* Step 4: Do it. */
2207         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2208
2209         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2210         if (IS_ERR(handle)) {
2211                 rc = PTR_ERR(handle);
2212                 GOTO(cleanup, rc);
2213         }
2214
2215         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2216         if (rc && rc != -EPERM && rc != -EACCES)
2217                 CERROR("vfs_link error %d\n", rc);
2218 cleanup:
2219         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2220                                 handle, req, rc, 0);
2221         EXIT;
2222
2223         switch (cleanup_phase) {
2224         case 4: /* child dentry */
2225                 l_dput(dchild);
2226         case 3: /* locks */
2227                 if (rc) {
2228                         ldlm_lock_decref(&src_lockh, LCK_EX);
2229                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2230                 } else {
2231                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2232                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2233                 }
2234         case 2: /* target dentry */
2235 #ifdef S_PDIROPS
2236                 if (tgt_dir_lockh[1].cookie && update_mode)
2237                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2238 #endif
2239                 if (de_tgt_dir)
2240                         l_dput(de_tgt_dir);
2241         case 1: /* source dentry */
2242                 l_dput(de_src);
2243         case 0:
2244                 break;
2245         default:
2246                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2247                 LBUG();
2248         }
2249         req->rq_status = rc;
2250         return 0;
2251 }
2252
2253 /*
2254  * add a hard link in the PENDING directory, only used by rename()
2255  */
2256 static int mds_add_link_orphan(struct mds_update_record *rec,
2257                                struct obd_device *obd,
2258                                struct dentry *dentry)
2259 {
2260         struct mds_obd *mds = &obd->u.mds;
2261         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2262         struct dentry *pending_child;
2263         char fidname[LL_FID_NAMELEN];
2264         int fidlen = 0, rc;
2265         ENTRY;
2266
2267         LASSERT(dentry->d_inode);
2268         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2269
2270         down(&pending_dir->i_sem);
2271         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2272                             dentry->d_inode->i_generation);
2273
2274         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2275                mds_open_orphan_count(dentry->d_inode),
2276                rec->ur_name, fidname);
2277
2278         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2279         if (IS_ERR(pending_child))
2280                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2281
2282         if (pending_child->d_inode != NULL) {
2283                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2284                 LASSERT(pending_child->d_inode == dentry->d_inode);
2285                 GOTO(out_dput, rc = 0);
2286         }
2287
2288         lock_kernel();
2289         rc = vfs_link(dentry, pending_dir, pending_child);
2290         unlock_kernel();
2291         if (rc)
2292                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2293                        rec->ur_name, rc);
2294         else
2295                 mds_inode_set_orphan(dentry->d_inode);
2296 out_dput:
2297         l_dput(pending_child);
2298 out_lock:
2299         up(&pending_dir->i_sem);
2300         RETURN(rc);
2301 }
2302
2303 /* The idea here is that we need to get four locks in the end:
2304  * one on each parent directory, one on each child.  We need to take
2305  * these locks in some kind of order (to avoid deadlocks), and the order
2306  * I selected is "increasing resource number" order.  We need to look up
2307  * the children, however, before we know what the resource number(s) are.
2308  * Thus the following plan:
2309  *
2310  * 1,2. Look up the parents
2311  * 3,4. Look up the children
2312  * 5. Take locks on the parents and children, in order
2313  * 6. Verify that the children haven't changed since they were looked up
2314  *
2315  * If there was a race and the children changed since they were first looked
2316  * up, it is possible that mds_verify_child() will be able to just grab the
2317  * lock on the new child resource (if it has a higher resource than any other)
2318  * but we need to compare against not only its parent, but also against the
2319  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2320  *
2321  * We need the fancy igrab() on the child inodes because we aren't holding a
2322  * lock on the parent after the lookup is done, so dentry->d_inode may change
2323  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2324  */
2325 static int mds_get_parents_children_locked(struct obd_device *obd,
2326                                            struct mds_obd *mds,
2327                                            struct ll_fid *p1_fid,
2328                                            struct dentry **de_srcdirp,
2329                                            struct ll_fid *p2_fid,
2330                                            struct dentry **de_tgtdirp,
2331                                            int parent_mode,
2332                                            const char *old_name, int old_len,
2333                                            struct dentry **de_oldp,
2334                                            const char *new_name, int new_len,
2335                                            struct dentry **de_newp,
2336                                            struct lustre_handle *dlm_handles,
2337                                            int child_mode)
2338 {
2339         struct ldlm_res_id p1_res_id = { .name = {0} };
2340         struct ldlm_res_id p2_res_id = { .name = {0} };
2341         struct ldlm_res_id c1_res_id = { .name = {0} };
2342         struct ldlm_res_id c2_res_id = { .name = {0} };
2343         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2344         /* Only dentry should change, but the inode itself would be
2345            intact otherwise */
2346         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2347         /* If something is going to be replaced, both dentry and inode locks are
2348            needed */
2349         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2350                                                         MDS_INODELOCK_UPDATE}};
2351         struct ldlm_res_id *maxres_src, *maxres_tgt;
2352         struct inode *inode;
2353         int rc = 0, cleanup_phase = 0;
2354         ENTRY;
2355
2356         /* Step 1: Lookup the source directory */
2357         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2358         if (IS_ERR(*de_srcdirp))
2359                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2360
2361         cleanup_phase = 1; /* source directory dentry */
2362
2363         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2364         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2365
2366         /* Step 2: Lookup the target directory */
2367         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2368                 *de_tgtdirp = dget(*de_srcdirp);
2369         } else {
2370                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2371                 if (IS_ERR(*de_tgtdirp)) {
2372                         rc = PTR_ERR(*de_tgtdirp);
2373                         *de_tgtdirp = NULL;
2374                         GOTO(cleanup, rc);
2375                 }
2376         }
2377
2378         cleanup_phase = 2; /* target directory dentry */
2379
2380         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2381         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2382
2383 #ifdef S_PDIROPS
2384         dlm_handles[5].cookie = 0;
2385         dlm_handles[6].cookie = 0;
2386         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2387                 /* Get a temp lock on just ino, gen to flush client cache and
2388                  * to protect dirs from concurrent splitting */
2389                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2390                                            LCK_PW, &p_policy, &p2_res_id,
2391                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2392                 if (rc != ELDLM_OK)
2393                         GOTO(cleanup, rc);
2394                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2395                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2396                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2397                        (*de_srcdirp)->d_inode->i_ino,
2398                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2399                        (*de_tgtdirp)->d_inode->i_ino,
2400                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2401         }
2402         cleanup_phase = 3;
2403 #endif
2404
2405         /* Step 3: Lookup the source child entry */
2406         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2407         if (IS_ERR(*de_oldp)) {
2408                 rc = PTR_ERR(*de_oldp);
2409                 CERROR("old child lookup error (%*s): %d\n",
2410                        old_len - 1, old_name, rc);
2411                 GOTO(cleanup, rc);
2412         }
2413
2414         cleanup_phase = 4; /* original name dentry */
2415
2416         inode = (*de_oldp)->d_inode;
2417         if (inode != NULL) {
2418                 inode = igrab(inode);
2419                 if (inode == NULL)
2420                         GOTO(cleanup, rc = -ENOENT);
2421
2422                 c1_res_id.name[0] = inode->i_ino;
2423                 c1_res_id.name[1] = inode->i_generation;
2424                 iput(inode);
2425         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2426                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2427                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2428         }
2429
2430         /* Step 4: Lookup the target child entry */
2431         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2432         if (IS_ERR(*de_newp)) {
2433                 rc = PTR_ERR(*de_newp);
2434                 CERROR("new child lookup error (%*s): %d\n",
2435                        old_len - 1, old_name, rc);
2436                 GOTO(cleanup, rc);
2437         }
2438
2439         cleanup_phase = 5; /* target dentry */
2440
2441         inode = (*de_newp)->d_inode;
2442         if (inode != NULL) {
2443                 inode = igrab(inode);
2444                 if (inode == NULL)
2445                         goto retry_locks;
2446
2447                 c2_res_id.name[0] = inode->i_ino;
2448                 c2_res_id.name[1] = inode->i_generation;
2449                 iput(inode);
2450         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2451                 c2_res_id.name[0] = (*de_newp)->d_inum;
2452                 c2_res_id.name[1] = (*de_newp)->d_generation;
2453         }
2454
2455 retry_locks:
2456         /* Step 5: Take locks on the parents and child(ren) */
2457         maxres_src = &p1_res_id;
2458         maxres_tgt = &p2_res_id;
2459         cleanup_phase = 5; /* target dentry */
2460
2461         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2462                 maxres_src = &c1_res_id;
2463         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2464                 maxres_tgt = &c2_res_id;
2465
2466         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2467                                     &p_policy,
2468                                     &p2_res_id, &dlm_handles[1], parent_mode,
2469                                     &p_policy,
2470                                     &c1_res_id, &dlm_handles[2], child_mode,
2471                                     &c1_policy,
2472                                     &c2_res_id, &dlm_handles[3], child_mode,
2473                                     &c2_policy);
2474         if (rc)
2475                 GOTO(cleanup, rc);
2476
2477         cleanup_phase = 6; /* parent and child(ren) locks */
2478
2479         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2480         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2481                               parent_mode, &c1_res_id, &dlm_handles[2],
2482                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2483                               maxres_tgt);
2484         if (rc) {
2485                 if (c2_res_id.name[0] != 0)
2486                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2487                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2488                 cleanup_phase = 5;
2489                 if (rc > 0)
2490                         goto retry_locks;
2491                 GOTO(cleanup, rc);
2492         }
2493
2494         if (!DENTRY_VALID(*de_oldp))
2495                 GOTO(cleanup, rc = -ENOENT);
2496
2497         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2498         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2499                               parent_mode, &c2_res_id, &dlm_handles[3],
2500                               de_newp, child_mode, &c2_policy, new_name,
2501                               new_len, maxres_src);
2502         if (rc) {
2503                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2504                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2505                 cleanup_phase = 5;
2506                 if (rc > 0)
2507                         goto retry_locks;
2508                 GOTO(cleanup, rc);
2509         }
2510
2511         EXIT;
2512 cleanup:
2513         if (rc) {
2514                 switch (cleanup_phase) {
2515                 case 6: /* child lock(s) */
2516                         if (c2_res_id.name[0] != 0)
2517                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2518                         if (c1_res_id.name[0] != 0)
2519                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2520                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2521                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2522                 case 5: /* target dentry */
2523                         l_dput(*de_newp);
2524                 case 4: /* source dentry */
2525                         l_dput(*de_oldp);
2526                 case 3:
2527 #ifdef S_PDIROPS
2528                         if (dlm_handles[5].cookie != 0)
2529                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2530                         if (dlm_handles[6].cookie != 0)
2531                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2532 #endif
2533                 case 2: /* target directory dentry */
2534                         l_dput(*de_tgtdirp);
2535                 case 1: /* source directry dentry */
2536                         l_dput(*de_srcdirp);
2537                 }
2538         }
2539
2540         return rc;
2541 }
2542
2543 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2544                                         int offset, struct ptlrpc_request *req)
2545 {
2546         struct obd_device *obd = req->rq_export->exp_obd;
2547         struct dentry *de_srcdir = NULL;
2548         struct dentry *de_new = NULL;
2549         struct mds_obd *mds = mds_req2mds(req);
2550         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2551         struct lustre_handle child_lockh = {0};
2552         int cleanup_phase = 0;
2553         void *handle = NULL;
2554         int update_mode, rc = 0;
2555         ENTRY;
2556
2557         /* another MDS executing rename operation has asked us
2558          * to create target name. such a creation should destroy
2559          * existing target name */
2560
2561         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2562                         obd->obd_name, rec->ur_tgt,
2563                         (unsigned long) rec->ur_fid1->mds,
2564                         (unsigned long) rec->ur_fid1->id,
2565                         (unsigned long) rec->ur_fid1->generation);
2566
2567         /* first, lookup the target */
2568         child_lockh.cookie = 0;
2569         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2570                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2571                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2572                                          &child_lockh, &de_new, LCK_EX,
2573                                          MDS_INODELOCK_LOOKUP, NULL);
2574         if (rc)
2575                 GOTO(cleanup, rc);
2576
2577         cleanup_phase = 1;
2578
2579         LASSERT(de_srcdir);
2580         LASSERT(de_srcdir->d_inode);
2581         LASSERT(de_new);
2582
2583         if (de_new->d_inode) {
2584                 /* name exists and points to local inode
2585                  * try to unlink this name and create new one */
2586                 CERROR("%s: %s points to local inode %lu/%lu\n",
2587                        obd->obd_name, rec->ur_tgt,
2588                        (unsigned long) de_new->d_inode->i_ino,
2589                        (unsigned long) de_new->d_inode->i_generation);
2590                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2591                                       FSFILT_OP_RENAME, NULL);
2592                 if (IS_ERR(handle))
2593                         GOTO(cleanup, rc = PTR_ERR(handle));
2594                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2595                 if (rc)
2596                         GOTO(cleanup, rc);
2597         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2598                 /* name exists adn points to remove inode */
2599                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2600                        obd->obd_name, rec->ur_tgt,
2601                        (unsigned long) de_new->d_mdsnum,
2602                        (unsigned long) de_new->d_inum,
2603                        (unsigned long) de_new->d_generation);
2604         } else {
2605                 /* name doesn't exist. the simplest case */
2606                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2607                                       FSFILT_OP_LINK, NULL);
2608                 if (IS_ERR(handle))
2609                         GOTO(cleanup, rc = PTR_ERR(handle));
2610         }
2611
2612         cleanup_phase = 2;
2613         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2614                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2615                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2616         if (rc)
2617                 CERROR("add_dir_entry() returned error %d\n", rc);
2618 cleanup:
2619         EXIT;
2620         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2621                                 handle, req, rc, 0);
2622         switch(cleanup_phase) {
2623                 case 2:
2624                 case 1:
2625 #ifdef S_PDIROPS
2626                         if (parent_lockh[1].cookie != 0)
2627                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2628 #endif
2629                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2630                         if (child_lockh.cookie != 0)
2631                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2632                         l_dput(de_new);
2633                         l_dput(de_srcdir);
2634                         break;
2635                 default:
2636                         LBUG();
2637         }
2638
2639         req->rq_status = rc;
2640
2641         RETURN(0);
2642 }
2643
2644 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2645                                       struct ptlrpc_request *req)
2646 {
2647         struct obd_device *obd = req->rq_export->exp_obd;
2648         struct ptlrpc_request *req2 = NULL;
2649         struct dentry *de_srcdir = NULL;
2650         struct dentry *de_old = NULL;
2651         struct mds_obd *mds = mds_req2mds(req);
2652         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2653         struct lustre_handle child_lockh = {0};
2654         struct mdc_op_data opdata;
2655         void *handle = NULL;
2656         int update_mode, rc = 0;
2657         ENTRY;
2658
2659         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2660                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2661         memset(&opdata, 0, sizeof(opdata));
2662
2663         child_lockh.cookie = 0;
2664         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2665                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2666                                          &update_mode, rec->ur_name,
2667                                          rec->ur_namelen, &child_lockh, &de_old,
2668                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2669         LASSERT(rc == 0);
2670         LASSERT(de_srcdir);
2671         LASSERT(de_srcdir->d_inode);
2672         LASSERT(de_old);
2673
2674         /* we already know the target should be created on another MDS
2675          * so, we have to request that MDS to do it */
2676
2677         /* prepare source fid */
2678         if (de_old->d_flags & DCACHE_CROSS_REF) {
2679                 LASSERT(de_old->d_inode == NULL);
2680                 CDEBUG(D_OTHER, "request to move remote name\n");
2681                 opdata.fid1.mds = de_old->d_mdsnum;
2682                 opdata.fid1.id = de_old->d_inum;
2683                 opdata.fid1.generation = de_old->d_generation;
2684         } else if (de_old->d_inode == NULL) {
2685                 /* oh, source doesn't exist */
2686                 GOTO(cleanup, rc = -ENOENT);
2687         } else {
2688                 LASSERT(de_old->d_inode != NULL);
2689                 CDEBUG(D_OTHER, "request to move local name\n");
2690                 opdata.fid1.mds = mds->mds_num;
2691                 opdata.fid1.id = de_old->d_inode->i_ino;
2692                 opdata.fid1.generation = de_old->d_inode->i_generation;
2693         }
2694
2695         opdata.fid2 = *(rec->ur_fid2);
2696         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2697                        rec->ur_tgtlen - 1, &req2);
2698
2699         if (rc)
2700                 GOTO(cleanup, rc);
2701
2702         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2703         if (IS_ERR(handle))
2704                 GOTO(cleanup, rc = PTR_ERR(handle));
2705         rc = fsfilt_del_dir_entry(obd, de_old);
2706         d_drop(de_old);
2707
2708 cleanup:
2709         EXIT;
2710         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2711                                 handle, req, rc, 0);
2712         if (req2)
2713                 ptlrpc_req_finished(req2);
2714
2715 #ifdef S_PDIROPS
2716         if (parent_lockh[1].cookie != 0)
2717                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2718 #endif
2719         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2720         if (child_lockh.cookie != 0)
2721                 ldlm_lock_decref(&child_lockh, LCK_EX);
2722
2723         l_dput(de_old);
2724         l_dput(de_srcdir);
2725
2726         req->rq_status = rc;
2727         RETURN(0);
2728
2729 }
2730
2731 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2732                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2733 {
2734         struct obd_device *obd = req->rq_export->exp_obd;
2735         struct dentry *de_srcdir = NULL;
2736         struct dentry *de_tgtdir = NULL;
2737         struct dentry *de_old = NULL;
2738         struct dentry *de_new = NULL;
2739         struct mds_obd *mds = mds_req2mds(req);
2740         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2741         struct mds_body *body = NULL;
2742         struct llog_create_locks *lcl = NULL;
2743         int rc = 0, log_unlink = 0, lock_count = 3;
2744         int cleanup_phase = 0;
2745         void *handle = NULL;
2746         ENTRY;
2747
2748         LASSERT(offset == 1);
2749
2750         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2751                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2752                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2753
2754         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2755
2756         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2757                 DEBUG_REQ(D_HA, req, "rename replay\n");
2758                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2759                        lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
2760                        req->rq_repmsg->buflens[2]);
2761         }
2762
2763         MDS_UPDATE_COUNTER(mds, MDS_RENAME_COUNT);
2764
2765         if (rec->ur_namelen == 1) {
2766                 rc = mds_reint_rename_create_name(rec, offset, req);
2767                 RETURN(rc);
2768         }
2769
2770         if (rec->ur_fid2->mds != mds->mds_num) {
2771                 rc = mds_reint_rename_to_remote(rec, offset, req);
2772                 RETURN(rc);
2773         }
2774
2775         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2776                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2777                                              rec->ur_name, rec->ur_namelen,
2778                                              &de_old, rec->ur_tgt,
2779                                              rec->ur_tgtlen, &de_new,
2780                                              dlm_handles, LCK_EX);
2781         if (rc)
2782                 GOTO(cleanup, rc);
2783
2784         cleanup_phase = 1; /* parent(s), children, locks */
2785
2786         if (de_new->d_inode)
2787                 lock_count = 4;
2788
2789         /* sanity check for src inode */
2790         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2791             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2792                 GOTO(cleanup, rc = -EINVAL);
2793
2794         /* sanity check for dest inode */
2795         if (de_new->d_inode &&
2796             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2797              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2798                 GOTO(cleanup, rc = -EINVAL);
2799
2800         if (de_old->d_inode == de_new->d_inode)
2801                 GOTO(cleanup, rc = 0);
2802
2803         /* if we are about to remove the target at first, pass the EA of
2804          * that inode to client to perform and cleanup on OST */
2805         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2806         LASSERT(body != NULL);
2807
2808         if (de_new->d_inode &&
2809             S_ISREG(de_new->d_inode->i_mode) &&
2810             de_new->d_inode->i_nlink == 1 &&
2811             mds_open_orphan_count(de_new->d_inode) == 0) {
2812                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2813                 mds_pack_inode2body(obd, body, de_new->d_inode);
2814                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2815                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2816                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2817                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2818                 } else {
2819                         log_unlink = 1;
2820                 }
2821         }
2822
2823         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2824                        de_srcdir->d_inode->i_sb);
2825
2826         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2827         if (IS_ERR(handle))
2828                 GOTO(cleanup, rc = PTR_ERR(handle));
2829
2830         /* FIXME need adjust the journal block count? */
2831         /* if the target should be moved to PENDING, we at first increase the
2832          * link and later vfs_rename() will decrease the link count again */
2833         if (de_new->d_inode &&
2834             S_ISREG(de_new->d_inode->i_mode) &&
2835             de_new->d_inode->i_nlink == 1 &&
2836             mds_open_orphan_count(de_new->d_inode) > 0) {
2837                 rc = mds_add_link_orphan(rec, obd, de_new);
2838                 if (rc)
2839                         GOTO(cleanup, rc);
2840         }
2841
2842         lock_kernel();
2843         de_old->d_fsdata = req;
2844         de_new->d_fsdata = req;
2845         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2846         unlock_kernel();
2847
2848         if (!rc && log_unlink) {
2849                 if (mds_log_op_unlink(obd, de_tgtdir->d_inode,
2850                                 lustre_msg_buf(req->rq_repmsg, 1, 0),
2851                                 req->rq_repmsg->buflens[1],
2852                                 lustre_msg_buf(req->rq_repmsg, 2, 0),
2853                                 req->rq_repmsg->buflens[2], &lcl) > 0) {
2854                         body->valid |= OBD_MD_FLCOOKIE;
2855                 }
2856         }
2857
2858         GOTO(cleanup, rc);
2859 cleanup:
2860         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2861                                 handle, req, rc, 0);
2862         switch (cleanup_phase) {
2863         case 1:
2864 #ifdef S_PDIROPS
2865                 if (dlm_handles[5].cookie != 0)
2866                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2867                 if (dlm_handles[6].cookie != 0)
2868                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2869 #endif
2870                 if (lcl != NULL)
2871                         ptlrpc_save_llog_lock(req, lcl);
2872
2873                 if (rc) {
2874                         if (lock_count == 4)
2875                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2876                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2877                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2878                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2879                 } else {
2880                         if (lock_count == 4)
2881                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2882                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2883                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2884                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2885                 }
2886                 l_dput(de_new);
2887                 l_dput(de_old);
2888                 l_dput(de_tgtdir);
2889                 l_dput(de_srcdir);
2890         case 0:
2891                 break;
2892         default:
2893                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2894                 LBUG();
2895         }
2896         req->rq_status = rc;
2897         return 0;
2898 }
2899
2900 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2901                            struct ptlrpc_request *, struct lustre_handle *);
2902
2903 static mds_reinter reinters[REINT_MAX + 1] = {
2904         [REINT_SETATTR] mds_reint_setattr,
2905         [REINT_CREATE] mds_reint_create,
2906         [REINT_LINK] mds_reint_link,
2907         [REINT_UNLINK] mds_reint_unlink,
2908         [REINT_RENAME] mds_reint_rename,
2909         [REINT_OPEN] mds_open
2910 };
2911
2912 int mds_reint_rec(struct mds_update_record *rec, int offset,
2913                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2914 {
2915         struct obd_device *obd = req->rq_export->exp_obd;
2916         struct lvfs_run_ctxt saved;
2917         int rc;
2918
2919         /* checked by unpacker */
2920         LASSERT(rec->ur_opcode <= REINT_MAX &&
2921                 reinters[rec->ur_opcode] != NULL);
2922
2923         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2924         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2925         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2926
2927         return rc;
2928 }