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 == 0);
 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");
 556                 body = lustre_msg_buf(req->rq_repmsg, offset, 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");
 565         } else {
 566                 body = lustre_msg_buf(req->rq_repmsg, offset, 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 == 0);
 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                                 obdo_free(oa);
 756                                 GOTO(cleanup, rc);
 757                         }
 758
 759                         /* now, add new dir entry for it */
 760                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 761                         if (IS_ERR(handle)) {
 762                                 obdo_free(oa);
 763                                 GOTO(cleanup, rc = PTR_ERR(handle));
 764                         }
 765                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 766                                                   rec->ur_namelen - 1,
 767                                                   oa->o_id, oa->o_generation,
 768                                                   i);
 769                         LASSERT(rc == 0);
 770
 771                         /* fill reply */
 772                         body = lustre_msg_buf(req->rq_repmsg,
 773                                               offset, sizeof (*body));
 774                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 775                         body->fid1.id = oa->o_id;
 776                         body->fid1.mds = i;
 777                         body->fid1.generation = oa->o_generation;
 778                         obdo_free(oa);
 779                 } else {
 780                         /* requested name exists in the directory */
 781                         rc = -EEXIST;
 782                 }
 783                 EXIT;
 784                 break;
 785         }
 786         case S_IFLNK:{
 787                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 788                 if (IS_ERR(handle))
 789                         GOTO(cleanup, rc = PTR_ERR(handle));
 790                 if (rec->ur_tgt == NULL)        /* no target supplied */
 791                         rc = -EINVAL;           /* -EPROTO? */
 792                 else
 793                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 794                 EXIT;
 795                 break;
 796         }
 797         case S_IFCHR:
 798         case S_IFBLK:
 799         case S_IFIFO:
 800         case S_IFSOCK:{
 801                 int rdev = rec->ur_rdev;
 802                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 803                 if (IS_ERR(handle))
 804                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 805                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 806                 EXIT;
 807                 break;
 808         }
 809         default:
 810                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 811                 dchild->d_fsdata = NULL;
 812                 GOTO(cleanup, rc = -EINVAL);
 813         }
 814
 815         /* In case we stored the desired inum in here, we want to clean up. */
 816         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 817                 dchild->d_fsdata = NULL;
 818
 819         if (rc) {
 820                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 821                 GOTO(cleanup, rc);
 822         } else if (dchild->d_inode) {
 823                 struct iattr iattr;
 824                 struct inode *inode = dchild->d_inode;
 825                 struct mds_body *body;
 826
 827                 created = 1;
 828                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 829                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 830                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 831                 iattr.ia_uid = rec->_ur_fsuid;
 832                 if (dir->i_mode & S_ISGID)
 833                         iattr.ia_gid = dir->i_gid;
 834                 else
 835                         iattr.ia_gid = rec->_ur_fsgid;
 836                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 837                         ATTR_MTIME | ATTR_CTIME;
 838
 839                 if (rec->ur_fid2->id) {
 840                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 841                         inode->i_generation = rec->ur_fid2->generation;
 842                         /* Dirtied and committed by the upcoming setattr. */
 843                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 844                                inode->i_ino, inode->i_generation);
 845                 } else {
 846 #if 0
 847                         struct lustre_handle child_ino_lockh;
 848 #endif
 849
 850                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 851                                inode->i_ino, inode->i_generation);
 852
 853 #if 0
 854                         /* The inode we were allocated may have just been freed
 855                          * by an unlink operation.  We take this lock to
 856                          * synchronize against the matching reply-ack-lock taken
 857                          * in unlink, to avoid replay problems if this reply
 858                          * makes it out to the client but the unlink's does not.
 859                          * See bug 2029 for more detail.*/
 860                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 861                         if (rc != ELDLM_OK) {
 862                                 CERROR("error locking for unlink/create sync: "
 863                                        "%d\n", rc);
 864                         } else {
 865                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 866                         }
 867 #endif
 868                 }
 869
 870                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 871                 if (rc)
 872                         CERROR("error on child setattr: rc = %d\n", rc);
 873
 874                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 875                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 876                 if (rc)
 877                         CERROR("error on parent setattr: rc = %d\n", rc);
 878
 879                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 880                 mds_pack_inode2fid(obd, &body->fid1, inode);
 881                 mds_pack_inode2body(obd, body, inode);
 882                 if (rc == 0) {
 883                         MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
 884                 }
 885         }
 886         EXIT;
 887
 888 cleanup:
 889         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 890
 891         if (rc && created) {
 892                 /* Destroy the file we just created. This should not need extra
 893                  * journal credits, as we have already modified all of the
 894                  * blocks needed in order to create the file in the first
 895                  * place. */
 896                 switch (type) {
 897                 case S_IFDIR:
 898                         err = vfs_rmdir(dir, dchild);
 899                         if (err)
 900                                 CERROR("rmdir in error path: %d\n", err);
 901                         break;
 902                 default:
 903                         err = vfs_unlink(dir, dchild);
 904                         if (err)
 905                                 CERROR("unlink in error path: %d\n", err);
 906                         break;
 907                 }
 908         } else {
 909                 rc = err;
 910         }
 911         switch (cleanup_phase) {
 912         case 2: /* child dentry */
 913                 l_dput(dchild);
 914         case 1: /* locked parent dentry */
 915 #ifdef S_PDIROPS
 916                 if (lockh[1].cookie != 0)
 917                         ldlm_lock_decref(lockh + 1, parent_mode);
 918 #endif
 919                 if (rc) {
 920                         ldlm_lock_decref(lockh, LCK_PW);
 921                 } else {
 922                         ptlrpc_save_lock(req, lockh, LCK_PW);
 923                 }
 924                 l_dput(dparent);
 925         case 0:
 926                 break;
 927         default:
 928                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 929                 LBUG();
 930         }
 931         if (mea)
 932                 OBD_FREE(mea, mea_size);
 933         req->rq_status = rc;
 934         return 0;
 935 }
 936
 937 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 938            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 939 {
 940         int i;
 941
 942         for (i = 0; i < RES_NAME_SIZE; i++) {
 943                 /* return 1 here, because enqueue_ordered will skip resources
 944                  * of all zeroes if they're sorted to the end of the list. */
 945                 if (res1->name[i] == 0 && res2->name[i] != 0)
 946                         return 1;
 947                 if (res2->name[i] == 0 && res1->name[i] != 0)
 948                         return 0;
 949
 950                 if (res1->name[i] > res2->name[i])
 951                         return 1;
 952                 if (res1->name[i] < res2->name[i])
 953                         return 0;
 954         }
 955
 956         if (!p1 || !p2)
 957                 return 0;
 958
 959         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 960                 return 1;
 961
 962         return 0;
 963 }
 964
 965 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 966  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 967  * because they take the place of local semaphores.
 968  *
 969  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 970  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 971 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 972                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 973                           ldlm_policy_data_t *p1_policy,
 974                           struct ldlm_res_id *p2_res_id,
 975                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 976                           ldlm_policy_data_t *p2_policy)
 977 {
 978         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 979         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 980         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 981         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 982         int rc, flags;
 983         ENTRY;
 984
 985         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 986
 987         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 988                res_id[1]->name[0]);
 989
 990         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 991                 handles[1] = p1_lockh;
 992                 handles[0] = p2_lockh;
 993                 res_id[1] = p1_res_id;
 994                 res_id[0] = p2_res_id;
 995                 lock_modes[1] = p1_lock_mode;
 996                 lock_modes[0] = p2_lock_mode;
 997                 policies[1] = p1_policy;
 998                 policies[0] = p2_policy;
 999         }
1000
1001         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1002                res_id[0]->name[0], res_id[1]->name[0]);
1003
1004         flags = LDLM_FL_LOCAL_ONLY;
1005         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1006                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1007                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1008                               NULL, 0, NULL, handles[0]);
1009         if (rc != ELDLM_OK)
1010                 RETURN(-EIO);
1011         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1012
1013         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1014             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1015                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1016                 ldlm_lock_addref(handles[1], lock_modes[1]);
1017         } else if (res_id[1]->name[0] != 0) {
1018                 flags = LDLM_FL_LOCAL_ONLY;
1019                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1020                                       *res_id[1], LDLM_IBITS, policies[1],
1021                                       lock_modes[1], &flags, mds_blocking_ast,
1022                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1023                                       NULL, handles[1]);
1024                 if (rc != ELDLM_OK) {
1025                         ldlm_lock_decref(handles[0], lock_modes[0]);
1026                         RETURN(-EIO);
1027                 }
1028                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1029         }
1030
1031         RETURN(0);
1032 }
1033
1034 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1035                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1036                            ldlm_policy_data_t *p1_policy,
1037                            struct ldlm_res_id *p2_res_id,
1038                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1039                            ldlm_policy_data_t *p2_policy,
1040                            struct ldlm_res_id *c1_res_id,
1041                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1042                            ldlm_policy_data_t *c1_policy,
1043                            struct ldlm_res_id *c2_res_id,
1044                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1045                            ldlm_policy_data_t *c2_policy)
1046 {
1047         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1048                                           c1_res_id, c2_res_id };
1049         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1050                                                  c1_lockh, c2_lockh };
1051         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1052                               c1_lock_mode, c2_lock_mode };
1053         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1054                                             c1_policy, c2_policy};
1055         int rc, i, j, sorted, flags;
1056         ENTRY;
1057
1058         CDEBUG(D_DLMTRACE,
1059                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1060                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1061                res_id[3]->name[0]);
1062
1063         /* simple insertion sort - we have at most 4 elements */
1064         for (i = 1; i < 4; i++) {
1065                 j = i - 1;
1066                 dlm_handles[4] = dlm_handles[i];
1067                 res_id[4] = res_id[i];
1068                 lock_modes[4] = lock_modes[i];
1069                 policies[4] = policies[i];
1070
1071                 sorted = 0;
1072                 do {
1073                         if (res_gt(res_id[j], res_id[4], policies[j],
1074                                    policies[4])) {
1075                                 dlm_handles[j + 1] = dlm_handles[j];
1076                                 res_id[j + 1] = res_id[j];
1077                                 lock_modes[j + 1] = lock_modes[j];
1078                                 policies[j + 1] = policies[j];
1079                                 j--;
1080                         } else {
1081                                 sorted = 1;
1082                         }
1083                 } while (j >= 0 && !sorted);
1084
1085                 dlm_handles[j + 1] = dlm_handles[4];
1086                 res_id[j + 1] = res_id[4];
1087                 lock_modes[j + 1] = lock_modes[4];
1088                 policies[j + 1] = policies[4];
1089         }
1090
1091         CDEBUG(D_DLMTRACE,
1092                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1093                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1094                res_id[3]->name[0]);
1095
1096         /* XXX we could send ASTs on all these locks first before blocking? */
1097         for (i = 0; i < 4; i++) {
1098                 flags = 0;
1099                 if (res_id[i]->name[0] == 0)
1100                         break;
1101                 if (i != 0 &&
1102                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1103                     (policies[i]->l_inodebits.bits &
1104                      policies[i-1]->l_inodebits.bits) ) {
1105                         memcpy(dlm_handles[i], dlm_handles[i-1],
1106                                sizeof(*(dlm_handles[i])));
1107                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1108                 } else {
1109                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1110                                               *res_id[i], LDLM_IBITS,
1111                                               policies[i],
1112                                               lock_modes[i], &flags,
1113                                               mds_blocking_ast,
1114                                               ldlm_completion_ast, NULL, NULL,
1115                                               NULL, 0, NULL, dlm_handles[i]);
1116                         if (rc != ELDLM_OK)
1117                                 GOTO(out_err, rc = -EIO);
1118                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1119                 }
1120         }
1121
1122         RETURN(0);
1123 out_err:
1124         while (i-- > 0)
1125                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1126
1127         return rc;
1128 }
1129
1130 /* In the unlikely case that the child changed while we were waiting
1131  * on the lock, we need to drop the lock on the old child and either:
1132  * - if the child has a lower resource name, then we have to also
1133  *   drop the parent lock and regain the locks in the right order
1134  * - in the rename case, if the child has a lower resource name than one of
1135  *   the other parent/child resources (maxres) we also need to reget the locks
1136  * - if the child has a higher resource name (this is the common case)
1137  *   we can just get the lock on the new child (still in lock order)
1138  *
1139  * Returns 0 if the child did not change or if it changed but could be locked.
1140  * Returns 1 if the child changed and we need to re-lock (no locks held).
1141  * Returns -ve error with a valid dchild (no locks held). */
1142 static int mds_verify_child(struct obd_device *obd,
1143                             struct ldlm_res_id *parent_res_id,
1144                             struct lustre_handle *parent_lockh,
1145                             struct dentry *dparent, int parent_mode,
1146                             struct ldlm_res_id *child_res_id,
1147                             struct lustre_handle *child_lockh,
1148                             struct dentry **dchildp, int child_mode,
1149                             ldlm_policy_data_t *child_policy,
1150                             const char *name, int namelen,
1151                             struct ldlm_res_id *maxres)
1152 {
1153         struct dentry *vchild, *dchild = *dchildp;
1154         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1155         ENTRY;
1156
1157         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1158         if (IS_ERR(vchild))
1159                 GOTO(cleanup, rc = PTR_ERR(vchild));
1160
1161         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1162                 if  (child_res_id->name[0] == vchild->d_inum &&
1163                                 child_res_id->name[1] == vchild->d_generation) {
1164                         if (dchild != NULL)
1165                                 l_dput(dchild);
1166                         *dchildp = vchild;
1167                         RETURN(0);
1168                 }
1169                 goto changed;
1170         }
1171
1172         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1173                    (vchild->d_inode != NULL &&
1174                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1175                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1176                 if (dchild != NULL)
1177                         l_dput(dchild);
1178                 *dchildp = vchild;
1179
1180                 RETURN(0);
1181         }
1182
1183 changed:
1184         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1185                vchild->d_inode, dchild ? dchild->d_inode : 0,
1186                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1187                child_res_id->name[0]);
1188         if (child_res_id->name[0] != 0)
1189                 ldlm_lock_decref(child_lockh, child_mode);
1190         if (dchild)
1191                 l_dput(dchild);
1192
1193         cleanup_phase = 1; /* parent lock only */
1194         *dchildp = dchild = vchild;
1195
1196         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1197                 int flags = 0;
1198                 if (dchild->d_inode) {
1199                         child_res_id->name[0] = dchild->d_inode->i_ino;
1200                         child_res_id->name[1] = dchild->d_inode->i_generation;
1201                 } else {
1202                         child_res_id->name[0] = dchild->d_inum;
1203                         child_res_id->name[1] = dchild->d_generation;
1204                 }
1205
1206                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1207                     res_gt(maxres, child_res_id, NULL, NULL)) {
1208                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1209                                child_res_id->name[0], parent_res_id->name[0],
1210                                maxres->name[0]);
1211                         GOTO(cleanup, rc = 1);
1212                 }
1213
1214                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1215                                       *child_res_id, LDLM_IBITS, child_policy,
1216                                       child_mode, &flags, mds_blocking_ast,
1217                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1218                                       NULL, child_lockh);
1219                 if (rc != ELDLM_OK)
1220                         GOTO(cleanup, rc = -EIO);
1221
1222         } else {
1223                 memset(child_res_id, 0, sizeof(*child_res_id));
1224         }
1225
1226         EXIT;
1227 cleanup:
1228         if (rc) {
1229                 switch(cleanup_phase) {
1230                 case 2:
1231                         if (child_res_id->name[0] != 0)
1232                                 ldlm_lock_decref(child_lockh, child_mode);
1233                 case 1:
1234                         ldlm_lock_decref(parent_lockh, parent_mode);
1235                 }
1236         }
1237         return rc;
1238 }
1239
1240 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1241                                 struct ll_fid *fid,
1242                                 struct lustre_handle *parent_lockh,
1243                                 struct dentry **dparentp, int parent_mode,
1244                                 __u64 parent_lockpart, int *update_mode,
1245                                 char *name, int namelen,
1246                                 struct lustre_handle *child_lockh,
1247                                 struct dentry **dchildp, int child_mode,
1248                                 __u64 child_lockpart, void *clone_info)
1249 {
1250         struct ldlm_res_id child_res_id = { .name = {0} };
1251         struct ldlm_res_id parent_res_id = { .name = {0} };
1252         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1253         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1254         struct inode *inode;
1255         int rc = 0, cleanup_phase = 0;
1256         ENTRY;
1257
1258         /* Step 1: Lookup parent */
1259         *dparentp = mds_fid2dentry(mds, fid, NULL);
1260         if (IS_ERR(*dparentp)) {
1261                 rc = PTR_ERR(*dparentp);
1262                 *dparentp = NULL;
1263                 RETURN(rc);
1264         }
1265
1266         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1267                (*dparentp)->d_inode->i_ino, name);
1268
1269         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1270         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1271 #ifdef S_PDIROPS
1272         parent_lockh[1].cookie = 0;
1273         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1274                 struct ldlm_res_id res_id = { .name = {0} };
1275                 ldlm_policy_data_t policy;
1276                 int flags = 0;
1277                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1278                 if (*update_mode) {
1279                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1280                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1281                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1282                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1283                                               res_id, LDLM_IBITS, &policy,
1284                                               *update_mode, &flags,
1285                                               mds_blocking_ast,
1286                                               ldlm_completion_ast,
1287                                               NULL, NULL, NULL, 0, NULL,
1288                                               parent_lockh + 1);
1289                         if (rc != ELDLM_OK)
1290                                 RETURN(-ENOLCK);
1291                 }
1292
1293                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1294                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1295                        (*dparentp)->d_inode->i_ino,
1296                        (*dparentp)->d_inode->i_generation,
1297                        parent_res_id.name[2]);
1298         }
1299 #endif
1300
1301         cleanup_phase = 1; /* parent dentry */
1302 #ifdef CONFIG_SNAPFS
1303         if (clone_info) {
1304                 /*FIXME is there any other FUNC will use d_fsdata,
1305                  *excepet creating inode according inum*/
1306                 (*dparentp)->d_fsdata = clone_info;
1307         }
1308 #endif
1309
1310         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1311         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1312         if (IS_ERR(*dchildp)) {
1313                 rc = PTR_ERR(*dchildp);
1314                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1315                 GOTO(cleanup, rc);
1316         }
1317
1318         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1319                 /* inode lives on another MDS: return * mds/ino/gen
1320                  * and LOOKUP lock. drop possible UPDATE lock! */
1321                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1322                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1323                 child_res_id.name[0] = (*dchildp)->d_inum;
1324                 child_res_id.name[1] = (*dchildp)->d_generation;
1325                 goto retry_locks;
1326         }
1327
1328         inode = (*dchildp)->d_inode;
1329         if (inode != NULL)
1330                 inode = igrab(inode);
1331         if (inode == NULL)
1332                 goto retry_locks;
1333
1334         child_res_id.name[0] = inode->i_ino;
1335         child_res_id.name[1] = inode->i_generation;
1336
1337         iput(inode);
1338
1339 retry_locks:
1340         cleanup_phase = 2; /* child dentry */
1341
1342         /* Step 3: Lock parent and child in resource order.  If child doesn't
1343          *         exist, we still have to lock the parent and re-lookup. */
1344         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1345                                    &parent_policy, &child_res_id, child_lockh,
1346                                    child_mode, &child_policy);
1347         if (rc)
1348                 GOTO(cleanup, rc);
1349
1350         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1351                 cleanup_phase = 4; /* child lock */
1352         else
1353                 cleanup_phase = 3; /* parent lock */
1354
1355         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1356         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1357                               parent_mode, &child_res_id, child_lockh,
1358                               dchildp, child_mode, &child_policy,
1359                               name, namelen, &parent_res_id);
1360         if (rc > 0)
1361                 goto retry_locks;
1362         if (rc < 0) {
1363                 cleanup_phase = 3;
1364                 GOTO(cleanup, rc);
1365         }
1366
1367 cleanup:
1368         if (rc) {
1369                 switch (cleanup_phase) {
1370                 case 4:
1371                         ldlm_lock_decref(child_lockh, child_mode);
1372                 case 3:
1373                         ldlm_lock_decref(parent_lockh, parent_mode);
1374                 case 2:
1375                         l_dput(*dchildp);
1376                 case 1:
1377 #ifdef S_PDIROPS
1378                         if (parent_lockh[1].cookie)
1379                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1380 #endif
1381                         l_dput(*dparentp);
1382                 default: ;
1383                 }
1384         }
1385         return rc;
1386 }
1387
1388 void mds_reconstruct_generic(struct ptlrpc_request *req)
1389 {
1390         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1391
1392         mds_req_from_mcd(req, med->med_mcd);
1393 }
1394
1395 int mds_create_local_dentry(struct mds_update_record *rec,
1396                            struct obd_device *obd)
1397 {
1398         struct mds_obd *mds = &obd->u.mds;
1399         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1400         int fidlen = 0, rc, cleanup_phase = 0;
1401         struct dentry *new_child = NULL;
1402         char *fidname = rec->ur_name;
1403         struct dentry *child = NULL;
1404         struct lustre_handle lockh[2] = {{0}, {0}};
1405         void *handle;
1406         ENTRY;
1407
1408         down(&fids_dir->i_sem);
1409         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1410         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1411                         fidname, (unsigned) rec->ur_fid1->id,
1412                         (unsigned) rec->ur_fid1->generation);
1413
1414         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1415         up(&fids_dir->i_sem);
1416         if (IS_ERR(new_child)) {
1417                 CERROR("can't lookup %s: %d\n", fidname,
1418                                 (int) PTR_ERR(new_child));
1419                 GOTO(cleanup, rc = PTR_ERR(new_child));
1420         }
1421         cleanup_phase = 1;
1422
1423         if (new_child->d_inode != NULL) {
1424                 /* nice. we've already have local dentry! */
1425                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1426                        (unsigned) new_child->d_inode->i_ino,
1427                        (unsigned) new_child->d_inode->i_generation);
1428                 rec->ur_fid1->id = fids_dir->i_ino;
1429                 rec->ur_fid1->generation = fids_dir->i_generation;
1430                 rec->ur_namelen = fidlen + 1;
1431                 GOTO(cleanup, rc = 0);
1432         }
1433
1434         /* new, local dentry will be added soon. we need no aliases here */
1435         d_drop(new_child);
1436
1437         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1438                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1439         } else {
1440                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1441                                               LCK_EX, lockh, NULL, NULL, 0,
1442                                               MDS_INODELOCK_UPDATE);
1443         }
1444
1445         if (IS_ERR(child)) {
1446                 rc = PTR_ERR(child);
1447                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1448                         CERROR("can't get victim: %d\n", rc);
1449                 GOTO(cleanup, rc);
1450         }
1451         cleanup_phase = 2;
1452
1453         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1454         if (IS_ERR(handle))
1455                 GOTO(cleanup, rc = PTR_ERR(handle));
1456
1457         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1458                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1459                                   mds->mds_num);
1460         if (rc)
1461                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1462                        (unsigned long) child->d_inode->i_ino,
1463                        (unsigned long) child->d_inode->i_generation, rc);
1464         else {
1465                 if (S_ISDIR(child->d_inode->i_mode)) {
1466                         fids_dir->i_nlink++;
1467                         mark_inode_dirty(fids_dir);
1468                 }
1469                 mark_inode_dirty(child->d_inode);
1470         }
1471         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1472
1473         rec->ur_fid1->id = fids_dir->i_ino;
1474         rec->ur_fid1->generation = fids_dir->i_generation;
1475         rec->ur_namelen = fidlen + 1;
1476
1477 cleanup:
1478         switch(cleanup_phase) {
1479                 case 2:
1480                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1481                                 ldlm_lock_decref(lockh, LCK_EX);
1482                         dput(child);
1483                 case 1:
1484                         dput(new_child);
1485                 case 0:
1486                        break;
1487         }
1488         RETURN(rc);
1489 }
1490
1491 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1492                                         struct ptlrpc_request *slave)
1493 {
1494         void *cookie, *cookie2;
1495         struct mds_body *body2;
1496         struct mds_body *body;
1497         void *ea, *ea2;
1498         ENTRY;
1499
1500         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1501         LASSERT(body != NULL);
1502
1503         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1504         LASSERT(body2 != NULL);
1505
1506         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1507                 RETURN(0);
1508         }
1509
1510         memcpy(body2, body, sizeof(*body));
1511         body2->valid &= ~OBD_MD_FLCOOKIE;
1512
1513         if (!(body->valid & OBD_MD_FLEASIZE) &&
1514             !(body->valid & OBD_MD_FLDIREA))
1515                 RETURN(0);
1516
1517         if (body->eadatasize == 0) {
1518                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1519                 RETURN(0);
1520         }
1521
1522         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1523
1524         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1525         LASSERT(ea != NULL);
1526
1527         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1528         LASSERT(ea2 != NULL);
1529
1530         memcpy(ea2, ea, body->eadatasize);
1531
1532         if (body->valid & OBD_MD_FLCOOKIE) {
1533                 LASSERT(master->rq_repmsg->buflens[2] >=
1534                                 slave->rq_repmsg->buflens[2]);
1535                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1536                                 slave->rq_repmsg->buflens[2]);
1537                 LASSERT(cookie != NULL);
1538
1539                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1540                                 master->rq_repmsg->buflens[2]);
1541                 LASSERT(cookie2 != NULL);
1542                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1543                 body2->valid |= OBD_MD_FLCOOKIE;
1544         }
1545         RETURN(0);
1546 }
1547
1548 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1549                                    struct ptlrpc_request *req,
1550                                    struct lustre_handle *parent_lockh,
1551                                    int update_mode,
1552                                    struct dentry *dparent,
1553                                    struct lustre_handle *child_lockh,
1554                                    struct dentry *dchild)
1555 {
1556         struct obd_device *obd = req->rq_export->exp_obd;
1557         struct mds_obd *mds = mds_req2mds(req);
1558         struct mdc_op_data op_data;
1559         int rc = 0, cleanup_phase = 0;
1560         struct ptlrpc_request *request = NULL;
1561         void *handle;
1562         ENTRY;
1563
1564         LASSERT(offset == 0 || offset == 2);
1565
1566         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1567                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1568                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1569         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1570                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1571                           rec->ur_namelen - 1, rec->ur_name,
1572                           (unsigned)dchild->d_mdsnum,
1573                           (unsigned) dchild->d_inum,
1574                           (unsigned) dchild->d_generation);
1575
1576         /* time to drop i_nlink on remote MDS */
1577         memset(&op_data, 0, sizeof(op_data));
1578         op_data.fid1.mds = dchild->d_mdsnum;
1579         op_data.fid1.id = dchild->d_inum;
1580         op_data.fid1.generation = dchild->d_generation;
1581         op_data.create_mode = rec->ur_mode;
1582         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1583                 op_data.create_mode |= MDS_MODE_REPLAY;
1584         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1585         cleanup_phase = 2;
1586         if (request) {
1587                 if (rc == 0)
1588                         mds_copy_unlink_reply(req, request);
1589                 ptlrpc_req_finished(request);
1590         }
1591         if (rc == 0) {
1592                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1593                                       NULL);
1594                 if (IS_ERR(handle))
1595                         GOTO(cleanup, rc = PTR_ERR(handle));
1596                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1597                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1598                                         rc, 0);
1599         }
1600 cleanup:
1601         req->rq_status = rc;
1602
1603 #ifdef S_PDIROPS
1604         if (parent_lockh[1].cookie != 0)
1605                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1606 #endif
1607         ldlm_lock_decref(child_lockh, LCK_EX);
1608         if (rc)
1609                 ldlm_lock_decref(parent_lockh, LCK_PW);
1610         else
1611                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1612         l_dput(dchild);
1613         l_dput(dparent);
1614
1615         return 0;
1616 }
1617
1618 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1619                             struct ptlrpc_request *req,
1620                             struct lustre_handle *lh)
1621 {
1622         struct dentry *dparent, *dchild;
1623         struct mds_obd *mds = mds_req2mds(req);
1624         struct obd_device *obd = req->rq_export->exp_obd;
1625         struct mds_body *body = NULL;
1626         struct inode *child_inode;
1627         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1628         struct lustre_handle child_lockh = {0};
1629 #if 0
1630         struct lustre_handle child_reuse_lockh = {0};
1631 #endif
1632         struct lustre_handle * slave_lockh = NULL;
1633         struct llog_create_locks *lcl = NULL;
1634         char fidname[LL_FID_NAMELEN];
1635         void *handle = NULL;
1636         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1637         int unlink_by_fid = 0;
1638         int update_mode;
1639         ENTRY;
1640
1641         LASSERT(offset == 0 || offset == 2);
1642
1643         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1644                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1645
1646         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1647
1648         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1649                 DEBUG_REQ(D_HA, req, "unlink replay\n");
1650                 memcpy(lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1651                        lustre_msg_buf(req->rq_reqmsg, 2, 0),
1652                        req->rq_repmsg->buflens[offset + 2]);
1653         }
1654
1655         MDS_UPDATE_COUNTER(mds, MDS_UNLINK_COUNT);
1656
1657
1658         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1659                 GOTO(cleanup, rc = -ENOENT);
1660
1661         if (rec->ur_namelen == 1) {
1662                 /* this is request to drop i_nlink on local inode */
1663                 unlink_by_fid = 1;
1664                 rec->ur_name = fidname;
1665                 rc = mds_create_local_dentry(rec, obd);
1666                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1667                         DEBUG_REQ(D_HA, req,
1668                                   "drop nlink on inode %u/%u/%u (replay)",
1669                                   (unsigned) rec->ur_fid1->mds,
1670                                   (unsigned) rec->ur_fid1->id,
1671                                   (unsigned) rec->ur_fid1->generation);
1672                         req->rq_status = 0;
1673                         RETURN(0);
1674                 }
1675         }
1676
1677         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1678                 /* master mds for directory asks slave removing
1679                  * inode is already locked */
1680                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1681                                                LCK_PW, parent_lockh,
1682                                                &update_mode, rec->ur_name,
1683                                                rec->ur_namelen,
1684                                                MDS_INODELOCK_UPDATE);
1685                 if (IS_ERR(dparent))
1686                         GOTO(cleanup, rc = PTR_ERR(dparent));
1687                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1688                                            rec->ur_namelen - 1);
1689                 if (IS_ERR(dchild))
1690                         GOTO(cleanup, rc = PTR_ERR(dchild));
1691                 child_lockh.cookie = 0;
1692                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1693                 LASSERT(dchild->d_inode != NULL);
1694                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1695         } else {
1696                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1697                                                  parent_lockh, &dparent,
1698                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1699                                                  &update_mode, rec->ur_name,
1700                                                  rec->ur_namelen, &child_lockh,
1701                                                  &dchild, LCK_EX,
1702                                                  MDS_INODELOCK_LOOKUP |
1703                                                  MDS_INODELOCK_UPDATE, NULL);
1704         }
1705         if (rc)
1706                 GOTO(cleanup, rc);
1707
1708         if (dchild->d_flags & DCACHE_CROSS_REF) {
1709                 /* we should have parent lock only here */
1710                 LASSERT(unlink_by_fid == 0);
1711                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1712                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1713                                         update_mode, dparent, &child_lockh, dchild);
1714                 RETURN(0);
1715         }
1716
1717         cleanup_phase = 1; /* dchild, dparent, locks */
1718
1719         dget(dchild);
1720         child_inode = dchild->d_inode;
1721         if (child_inode == NULL) {
1722                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1723                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1724                 GOTO(cleanup, rc = -ENOENT);
1725         }
1726
1727         cleanup_phase = 2; /* dchild has a lock */
1728
1729         /* We have to do these checks ourselves, in case we are making an
1730          * orphan.  The client tells us whether rmdir() or unlink() was called,
1731          * so we need to return appropriate errors (bug 72).
1732          *
1733          * We don't have to check permissions, because vfs_rename (called from
1734          * mds_open_unlink_rename) also calls may_delete. */
1735         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1736                 if (!S_ISDIR(child_inode->i_mode))
1737                         GOTO(cleanup, rc = -ENOTDIR);
1738         } else {
1739                 if (S_ISDIR(child_inode->i_mode))
1740                         GOTO(cleanup, rc = -EISDIR);
1741         }
1742
1743         /* handle splitted dir */
1744         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1745         if (rc)
1746                 GOTO(cleanup, rc);
1747
1748 #if 0
1749         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1750          * reuse (see bug 2029). */
1751         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1752         if (rc != ELDLM_OK)
1753                 GOTO(cleanup, rc);
1754 #endif
1755         cleanup_phase = 3; /* child inum lock */
1756
1757         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1758
1759         /* ldlm_reply in buf[0] if called via intent */
1760         if (offset)
1761                 offset = 1;
1762
1763         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1764         LASSERT(body != NULL);
1765
1766         /* If this is the last reference to this inode, get the OBD EA
1767          * data first so the client can destroy OST objects.
1768          * we only do the object removal if no open files remain.
1769          * Nobody can get at this name anymore because of the locks so
1770          * we make decisions here as to whether to remove the inode */
1771         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1772             mds_open_orphan_count(child_inode) == 0) {
1773                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1774                 mds_pack_inode2body(obd, body, child_inode);
1775                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1776                             child_inode, 1);
1777                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1778                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1779                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1780                 } else {
1781                         log_unlink = 1;
1782                 }
1783         }
1784
1785         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1786         switch (child_inode->i_mode & S_IFMT) {
1787         case S_IFDIR:
1788                 /* Drop any lingering child directories before we start our
1789                  * transaction, to avoid doing multiple inode dirty/delete
1790                  * in our compound transaction (bug 1321). */
1791                 shrink_dcache_parent(dchild);
1792                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1793                                       NULL);
1794                 if (IS_ERR(handle))
1795                         GOTO(cleanup, rc = PTR_ERR(handle));
1796                 cleanup_phase = 4; /* transaction */
1797                 rc = vfs_rmdir(dparent->d_inode, dchild);
1798                 break;
1799         case S_IFREG: {
1800 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1801                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1802                                                         offset + 1, 0);
1803                 handle = fsfilt_start_log(obd, dparent->d_inode,
1804                                           FSFILT_OP_UNLINK, NULL,
1805                                           le32_to_cpu(lmm->lmm_stripe_count));
1806                 if (IS_ERR(handle))
1807                         GOTO(cleanup, rc = PTR_ERR(handle));
1808
1809                 cleanup_phase = 4; /* transaction */
1810                 rc = vfs_unlink(dparent->d_inode, dchild);
1811
1812                 if (!rc && log_unlink)
1813                         if (mds_log_op_unlink(obd, child_inode,
1814                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1815                                 req->rq_repmsg->buflens[offset + 1],
1816                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1817                                 req->rq_repmsg->buflens[offset + 2], &lcl) > 0)
1818                                 body->valid |= OBD_MD_FLCOOKIE;
1819                 break;
1820         }
1821         case S_IFLNK:
1822         case S_IFCHR:
1823         case S_IFBLK:
1824         case S_IFIFO:
1825         case S_IFSOCK:
1826                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1827                                       NULL);
1828                 if (IS_ERR(handle))
1829                         GOTO(cleanup, rc = PTR_ERR(handle));
1830                 cleanup_phase = 4; /* transaction */
1831                 rc = vfs_unlink(dparent->d_inode, dchild);
1832                 break;
1833         default:
1834                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1835                        rec->ur_name);
1836                 LBUG();
1837                 GOTO(cleanup, rc = -EINVAL);
1838         }
1839
1840  cleanup:
1841         if (rc == 0) {
1842                 struct iattr iattr;
1843                 int err;
1844
1845                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1846                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1847                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1848
1849                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1850                 if (err)
1851                         CERROR("error on parent setattr: rc = %d\n", err);
1852         }
1853
1854         switch(cleanup_phase) {
1855         case 4:
1856                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1857                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1858                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1859                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1860
1861                         /* filesystem is really going to destroy an inode
1862                          * we have to delay this till inode is opened -bzzz */
1863                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1864                 }
1865                 /* handle splitted dir */
1866                 if (rc == 0) {
1867                         /* master directory can be non-empty or something else ... */
1868                         mds_unlink_slave_objs(obd, dchild);
1869                 }
1870                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1871                                         rc, 0);
1872                 if (!rc)
1873                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1874                                            "unlinked", 0, NULL);
1875                 if (lcl != NULL)
1876                         ptlrpc_save_llog_lock(req, lcl);
1877         case 3: /* child ino-reuse lock */
1878 #if 0
1879                 if (rc && body != NULL) {
1880                         // Don't unlink the OST objects if the MDS unlink failed
1881                         body->valid = 0;
1882                 }
1883                 if (rc)
1884                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1885                 else
1886                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1887 #endif
1888         case 2: /* child lock */
1889                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1890                 if (child_lockh.cookie)
1891                         ldlm_lock_decref(&child_lockh, LCK_EX);
1892         case 1: /* child and parent dentry, parent lock */
1893 #ifdef S_PDIROPS
1894                 if (parent_lockh[1].cookie != 0)
1895                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1896 #endif
1897                 if (rc)
1898                         ldlm_lock_decref(parent_lockh, LCK_PW);
1899                 else
1900                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1901                 l_dput(dchild);
1902                 l_dput(dchild);
1903                 l_dput(dparent);
1904         case 0:
1905                 break;
1906         default:
1907                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1908                 LBUG();
1909         }
1910         req->rq_status = rc;
1911         return 0;
1912 }
1913
1914 /*
1915  * to service requests from remote MDS to increment i_nlink
1916  */
1917 static int mds_reint_link_acquire(struct mds_update_record *rec,
1918                                  int offset, struct ptlrpc_request *req,
1919                                  struct lustre_handle *lh)
1920 {
1921         struct obd_device *obd = req->rq_export->exp_obd;
1922         struct ldlm_res_id src_res_id = { .name = {0} };
1923         struct lustre_handle *handle = NULL, src_lockh = {0};
1924         struct mds_obd *mds = mds_req2mds(req);
1925         int rc = 0, cleanup_phase = 0;
1926         struct dentry *de_src = NULL;
1927         ldlm_policy_data_t policy;
1928         int flags = 0;
1929         ENTRY;
1930
1931         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1932                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1933                   (unsigned) rec->ur_fid1->id,
1934                   (unsigned) rec->ur_fid1->generation);
1935
1936         /* Step 1: Lookup the source inode and target directory by FID */
1937         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1938         if (IS_ERR(de_src))
1939                 GOTO(cleanup, rc = PTR_ERR(de_src));
1940         cleanup_phase = 1; /* source dentry */
1941
1942         src_res_id.name[0] = de_src->d_inode->i_ino;
1943         src_res_id.name[1] = de_src->d_inode->i_generation;
1944         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1945
1946         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1947                         src_res_id, LDLM_IBITS, &policy,
1948                         LCK_EX, &flags, mds_blocking_ast,
1949                         ldlm_completion_ast, NULL, NULL,
1950                         NULL, 0, NULL, &src_lockh);
1951         if (rc != ELDLM_OK)
1952                 GOTO(cleanup, rc = -ENOLCK);
1953         cleanup_phase = 2; /* lock */
1954
1955         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1956
1957         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1958         if (IS_ERR(handle)) {
1959                 rc = PTR_ERR(handle);
1960                 GOTO(cleanup, rc);
1961         }
1962         de_src->d_inode->i_nlink++;
1963         mark_inode_dirty(de_src->d_inode);
1964
1965 cleanup:
1966         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1967                                         handle, req, rc, 0);
1968         EXIT;
1969         switch (cleanup_phase) {
1970                 case 2:
1971                         if (rc)
1972                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1973                         else
1974                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1975                 case 1:
1976                         l_dput(de_src);
1977                 case 0:
1978                         break;
1979                 default:
1980                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1981                         LBUG();
1982         }
1983         req->rq_status = rc;
1984         return 0;
1985 }
1986
1987 /*
1988  * request to link to foreign inode:
1989  *  - acquire i_nlinks on this inode
1990  *  - add dentry
1991  */
1992 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1993                                     int offset, struct ptlrpc_request *req,
1994                                     struct lustre_handle *lh)
1995 {
1996         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1997         struct obd_device *obd = req->rq_export->exp_obd;
1998         struct dentry *de_tgt_dir = NULL;
1999         struct mds_obd *mds = mds_req2mds(req);
2000         int rc = 0, cleanup_phase = 0;
2001         struct mdc_op_data op_data;
2002         struct ptlrpc_request *request = NULL;
2003         int update_mode;
2004         ENTRY;
2005
2006 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
2007         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
2008                   (unsigned) rec->ur_fid2->mds,
2009                   (unsigned) rec->ur_fid2->id,
2010                   (unsigned) rec->ur_fid2->generation,
2011                   rec->ur_namelen - 1, rec->ur_name,
2012                   (unsigned) rec->ur_fid1->mds,
2013                   (unsigned) rec->ur_fid1->id,
2014                   (unsigned)rec->ur_fid1->generation);
2015
2016         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
2017                                            tgt_dir_lockh, &update_mode,
2018                                            rec->ur_name, rec->ur_namelen - 1,
2019                                            MDS_INODELOCK_UPDATE);
2020         if (IS_ERR(de_tgt_dir))
2021                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2022         cleanup_phase = 1;
2023
2024         op_data.fid1 = *(rec->ur_fid1);
2025         op_data.namelen = 0;
2026         op_data.name = NULL;
2027         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2028         if (rc)
2029                 GOTO(cleanup, rc);
2030
2031         cleanup_phase = 2;
2032         if (request)
2033                 ptlrpc_req_finished(request);
2034
2035         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2036
2037         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2038         if (IS_ERR(handle)) {
2039                 rc = PTR_ERR(handle);
2040                 GOTO(cleanup, rc);
2041         }
2042
2043         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2044                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2045                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2046         cleanup_phase = 3;
2047
2048 cleanup:
2049         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2050                                 handle, req, rc, 0);
2051         EXIT;
2052
2053         switch (cleanup_phase) {
2054                 case 3:
2055                         if (rc) {
2056                                 /* FIXME: drop i_nlink on remote inode here */
2057                                 CERROR("MUST drop drop i_nlink here\n");
2058                         }
2059                 case 2:
2060                 case 1:
2061                         if (rc) {
2062                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2063 #ifdef S_PDIROPS
2064                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2065 #endif
2066                         } else {
2067                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2068 #ifdef S_PDIROPS
2069                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2070 #endif
2071                         }
2072                         l_dput(de_tgt_dir);
2073                         break;
2074                 default:
2075                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2076                         LBUG();
2077         }
2078         req->rq_status = rc;
2079         return 0;
2080 }
2081
2082 static int mds_reint_link(struct mds_update_record *rec, int offset,
2083                           struct ptlrpc_request *req,
2084                           struct lustre_handle *lh)
2085 {
2086         struct obd_device *obd = req->rq_export->exp_obd;
2087         struct dentry *de_src = NULL;
2088         struct dentry *de_tgt_dir = NULL;
2089         struct dentry *dchild = NULL;
2090         struct mds_obd *mds = mds_req2mds(req);
2091         struct lustre_handle *handle = NULL;
2092         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2093         struct ldlm_res_id src_res_id = { .name = {0} };
2094         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2095         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2096         ldlm_policy_data_t tgt_dir_policy =
2097                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2098         int rc = 0, cleanup_phase = 0;
2099         int update_mode = 0;
2100         ENTRY;
2101
2102         LASSERT(offset == 0);
2103
2104         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2105                   rec->ur_fid1->id, rec->ur_fid1->generation,
2106                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2107
2108         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2109
2110         MDS_UPDATE_COUNTER(mds, MDS_LINK_COUNT);
2111
2112
2113 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2114         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2115                 GOTO(cleanup, rc = -ENOENT);
2116
2117         if (rec->ur_fid1->mds != mds->mds_num) {
2118                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2119                 RETURN(rc);
2120         }
2121
2122         if (rec->ur_namelen == 1) {
2123                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2124                 RETURN(rc);
2125         }
2126
2127         /* Step 1: Lookup the source inode and target directory by FID */
2128         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2129         if (IS_ERR(de_src))
2130                 GOTO(cleanup, rc = PTR_ERR(de_src));
2131
2132         cleanup_phase = 1; /* source dentry */
2133
2134         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2135         if (IS_ERR(de_tgt_dir)) {
2136                 rc = PTR_ERR(de_tgt_dir);
2137                 de_tgt_dir = NULL;
2138                 GOTO(cleanup, rc);
2139         }
2140
2141         cleanup_phase = 2; /* target directory dentry */
2142
2143         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2144                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2145                de_src->d_inode->i_ino);
2146
2147         /* Step 2: Take the two locks */
2148         src_res_id.name[0] = de_src->d_inode->i_ino;
2149         src_res_id.name[1] = de_src->d_inode->i_generation;
2150         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2151         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2152 #ifdef S_PDIROPS
2153         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2154                 int flags = 0;
2155                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2156                 if (update_mode) {
2157                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2158                                               tgt_dir_res_id, LDLM_IBITS,
2159                                               &src_policy, update_mode, &flags,
2160                                               mds_blocking_ast,
2161                                               ldlm_completion_ast, NULL, NULL,
2162                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2163                         if (rc != ELDLM_OK)
2164                                 GOTO(cleanup, rc = -ENOLCK);
2165                 }
2166
2167                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2168                                                         rec->ur_namelen - 1);
2169                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2170                        de_tgt_dir->d_inode->i_ino,
2171                        de_tgt_dir->d_inode->i_generation,
2172                        tgt_dir_res_id.name[2]);
2173         }
2174 #endif
2175         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2176                                    &src_policy,
2177                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2178                                    &tgt_dir_policy);
2179         if (rc)
2180                 GOTO(cleanup, rc);
2181
2182         cleanup_phase = 3; /* locks */
2183
2184         /* Step 3: Lookup the child */
2185         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2186         if (IS_ERR(dchild)) {
2187                 rc = PTR_ERR(dchild);
2188                 if (rc != -EPERM && rc != -EACCES)
2189                         CERROR("child lookup error %d\n", rc);
2190                 GOTO(cleanup, rc);
2191         }
2192
2193         cleanup_phase = 4; /* child dentry */
2194
2195         if (dchild->d_inode) {
2196                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2197                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2198                 rc = -EEXIST;
2199                 GOTO(cleanup, rc);
2200         }
2201
2202         /* Step 4: Do it. */
2203         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2204
2205         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2206         if (IS_ERR(handle)) {
2207                 rc = PTR_ERR(handle);
2208                 GOTO(cleanup, rc);
2209         }
2210
2211         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2212         if (rc && rc != -EPERM && rc != -EACCES)
2213                 CERROR("vfs_link error %d\n", rc);
2214 cleanup:
2215         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2216                                 handle, req, rc, 0);
2217         EXIT;
2218
2219         switch (cleanup_phase) {
2220         case 4: /* child dentry */
2221                 l_dput(dchild);
2222         case 3: /* locks */
2223                 if (rc) {
2224                         ldlm_lock_decref(&src_lockh, LCK_EX);
2225                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2226                 } else {
2227                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2228                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2229                 }
2230         case 2: /* target dentry */
2231 #ifdef S_PDIROPS
2232                 if (tgt_dir_lockh[1].cookie && update_mode)
2233                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2234 #endif
2235                 if (de_tgt_dir)
2236                         l_dput(de_tgt_dir);
2237         case 1: /* source dentry */
2238                 l_dput(de_src);
2239         case 0:
2240                 break;
2241         default:
2242                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2243                 LBUG();
2244         }
2245         req->rq_status = rc;
2246         return 0;
2247 }
2248
2249 /*
2250  * add a hard link in the PENDING directory, only used by rename()
2251  */
2252 static int mds_add_link_orphan(struct mds_update_record *rec,
2253                                struct obd_device *obd,
2254                                struct dentry *dentry)
2255 {
2256         struct mds_obd *mds = &obd->u.mds;
2257         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2258         struct dentry *pending_child;
2259         char fidname[LL_FID_NAMELEN];
2260         int fidlen = 0, rc;
2261         ENTRY;
2262
2263         LASSERT(dentry->d_inode);
2264         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2265
2266         down(&pending_dir->i_sem);
2267         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2268                             dentry->d_inode->i_generation);
2269
2270         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2271                mds_open_orphan_count(dentry->d_inode),
2272                rec->ur_name, fidname);
2273
2274         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2275         if (IS_ERR(pending_child))
2276                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2277
2278         if (pending_child->d_inode != NULL) {
2279                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2280                 LASSERT(pending_child->d_inode == dentry->d_inode);
2281                 GOTO(out_dput, rc = 0);
2282         }
2283
2284         lock_kernel();
2285         rc = vfs_link(dentry, pending_dir, pending_child);
2286         unlock_kernel();
2287         if (rc)
2288                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2289                        rec->ur_name, rc);
2290         else
2291                 mds_inode_set_orphan(dentry->d_inode);
2292 out_dput:
2293         l_dput(pending_child);
2294 out_lock:
2295         up(&pending_dir->i_sem);
2296         RETURN(rc);
2297 }
2298
2299 /* The idea here is that we need to get four locks in the end:
2300  * one on each parent directory, one on each child.  We need to take
2301  * these locks in some kind of order (to avoid deadlocks), and the order
2302  * I selected is "increasing resource number" order.  We need to look up
2303  * the children, however, before we know what the resource number(s) are.
2304  * Thus the following plan:
2305  *
2306  * 1,2. Look up the parents
2307  * 3,4. Look up the children
2308  * 5. Take locks on the parents and children, in order
2309  * 6. Verify that the children haven't changed since they were looked up
2310  *
2311  * If there was a race and the children changed since they were first looked
2312  * up, it is possible that mds_verify_child() will be able to just grab the
2313  * lock on the new child resource (if it has a higher resource than any other)
2314  * but we need to compare against not only its parent, but also against the
2315  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2316  *
2317  * We need the fancy igrab() on the child inodes because we aren't holding a
2318  * lock on the parent after the lookup is done, so dentry->d_inode may change
2319  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2320  */
2321 static int mds_get_parents_children_locked(struct obd_device *obd,
2322                                            struct mds_obd *mds,
2323                                            struct ll_fid *p1_fid,
2324                                            struct dentry **de_srcdirp,
2325                                            struct ll_fid *p2_fid,
2326                                            struct dentry **de_tgtdirp,
2327                                            int parent_mode,
2328                                            const char *old_name, int old_len,
2329                                            struct dentry **de_oldp,
2330                                            const char *new_name, int new_len,
2331                                            struct dentry **de_newp,
2332                                            struct lustre_handle *dlm_handles,
2333                                            int child_mode)
2334 {
2335         struct ldlm_res_id p1_res_id = { .name = {0} };
2336         struct ldlm_res_id p2_res_id = { .name = {0} };
2337         struct ldlm_res_id c1_res_id = { .name = {0} };
2338         struct ldlm_res_id c2_res_id = { .name = {0} };
2339         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2340         /* Only dentry should change, but the inode itself would be
2341            intact otherwise */
2342         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2343         /* If something is going to be replaced, both dentry and inode locks are
2344            needed */
2345         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2346                                                         MDS_INODELOCK_UPDATE}};
2347         struct ldlm_res_id *maxres_src, *maxres_tgt;
2348         struct inode *inode;
2349         int rc = 0, cleanup_phase = 0;
2350         ENTRY;
2351
2352         /* Step 1: Lookup the source directory */
2353         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2354         if (IS_ERR(*de_srcdirp))
2355                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2356
2357         cleanup_phase = 1; /* source directory dentry */
2358
2359         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2360         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2361
2362         /* Step 2: Lookup the target directory */
2363         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2364                 *de_tgtdirp = dget(*de_srcdirp);
2365         } else {
2366                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2367                 if (IS_ERR(*de_tgtdirp)) {
2368                         rc = PTR_ERR(*de_tgtdirp);
2369                         *de_tgtdirp = NULL;
2370                         GOTO(cleanup, rc);
2371                 }
2372         }
2373
2374         cleanup_phase = 2; /* target directory dentry */
2375
2376         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2377         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2378
2379 #ifdef S_PDIROPS
2380         dlm_handles[5].cookie = 0;
2381         dlm_handles[6].cookie = 0;
2382         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2383                 /* Get a temp lock on just ino, gen to flush client cache and
2384                  * to protect dirs from concurrent splitting */
2385                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2386                                            LCK_PW, &p_policy, &p2_res_id,
2387                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2388                 if (rc != ELDLM_OK)
2389                         GOTO(cleanup, rc);
2390                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2391                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2392                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2393                        (*de_srcdirp)->d_inode->i_ino,
2394                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2395                        (*de_tgtdirp)->d_inode->i_ino,
2396                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2397         }
2398         cleanup_phase = 3;
2399 #endif
2400
2401         /* Step 3: Lookup the source child entry */
2402         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2403         if (IS_ERR(*de_oldp)) {
2404                 rc = PTR_ERR(*de_oldp);
2405                 CERROR("old child lookup error (%*s): %d\n",
2406                        old_len - 1, old_name, rc);
2407                 GOTO(cleanup, rc);
2408         }
2409
2410         cleanup_phase = 4; /* original name dentry */
2411
2412         inode = (*de_oldp)->d_inode;
2413         if (inode != NULL) {
2414                 inode = igrab(inode);
2415                 if (inode == NULL)
2416                         GOTO(cleanup, rc = -ENOENT);
2417
2418                 c1_res_id.name[0] = inode->i_ino;
2419                 c1_res_id.name[1] = inode->i_generation;
2420                 iput(inode);
2421         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2422                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2423                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2424         }
2425
2426         /* Step 4: Lookup the target child entry */
2427         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2428         if (IS_ERR(*de_newp)) {
2429                 rc = PTR_ERR(*de_newp);
2430                 CERROR("new child lookup error (%*s): %d\n",
2431                        old_len - 1, old_name, rc);
2432                 GOTO(cleanup, rc);
2433         }
2434
2435         cleanup_phase = 5; /* target dentry */
2436
2437         inode = (*de_newp)->d_inode;
2438         if (inode != NULL) {
2439                 inode = igrab(inode);
2440                 if (inode == NULL)
2441                         goto retry_locks;
2442
2443                 c2_res_id.name[0] = inode->i_ino;
2444                 c2_res_id.name[1] = inode->i_generation;
2445                 iput(inode);
2446         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2447                 c2_res_id.name[0] = (*de_newp)->d_inum;
2448                 c2_res_id.name[1] = (*de_newp)->d_generation;
2449         }
2450
2451 retry_locks:
2452         /* Step 5: Take locks on the parents and child(ren) */
2453         maxres_src = &p1_res_id;
2454         maxres_tgt = &p2_res_id;
2455         cleanup_phase = 5; /* target dentry */
2456
2457         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2458                 maxres_src = &c1_res_id;
2459         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2460                 maxres_tgt = &c2_res_id;
2461
2462         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2463                                     &p_policy,
2464                                     &p2_res_id, &dlm_handles[1], parent_mode,
2465                                     &p_policy,
2466                                     &c1_res_id, &dlm_handles[2], child_mode,
2467                                     &c1_policy,
2468                                     &c2_res_id, &dlm_handles[3], child_mode,
2469                                     &c2_policy);
2470         if (rc)
2471                 GOTO(cleanup, rc);
2472
2473         cleanup_phase = 6; /* parent and child(ren) locks */
2474
2475         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2476         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2477                               parent_mode, &c1_res_id, &dlm_handles[2],
2478                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2479                               maxres_tgt);
2480         if (rc) {
2481                 if (c2_res_id.name[0] != 0)
2482                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2483                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2484                 cleanup_phase = 5;
2485                 if (rc > 0)
2486                         goto retry_locks;
2487                 GOTO(cleanup, rc);
2488         }
2489
2490         if (!DENTRY_VALID(*de_oldp))
2491                 GOTO(cleanup, rc = -ENOENT);
2492
2493         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2494         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2495                               parent_mode, &c2_res_id, &dlm_handles[3],
2496                               de_newp, child_mode, &c2_policy, new_name,
2497                               new_len, maxres_src);
2498         if (rc) {
2499                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2500                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2501                 cleanup_phase = 5;
2502                 if (rc > 0)
2503                         goto retry_locks;
2504                 GOTO(cleanup, rc);
2505         }
2506
2507         EXIT;
2508 cleanup:
2509         if (rc) {
2510                 switch (cleanup_phase) {
2511                 case 6: /* child lock(s) */
2512                         if (c2_res_id.name[0] != 0)
2513                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2514                         if (c1_res_id.name[0] != 0)
2515                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2516                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2517                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2518                 case 5: /* target dentry */
2519                         l_dput(*de_newp);
2520                 case 4: /* source dentry */
2521                         l_dput(*de_oldp);
2522                 case 3:
2523 #ifdef S_PDIROPS
2524                         if (dlm_handles[5].cookie != 0)
2525                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2526                         if (dlm_handles[6].cookie != 0)
2527                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2528 #endif
2529                 case 2: /* target directory dentry */
2530                         l_dput(*de_tgtdirp);
2531                 case 1: /* source directry dentry */
2532                         l_dput(*de_srcdirp);
2533                 }
2534         }
2535
2536         return rc;
2537 }
2538
2539 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2540                                         int offset, struct ptlrpc_request *req)
2541 {
2542         struct obd_device *obd = req->rq_export->exp_obd;
2543         struct dentry *de_srcdir = NULL;
2544         struct dentry *de_new = NULL;
2545         struct mds_obd *mds = mds_req2mds(req);
2546         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2547         struct lustre_handle child_lockh = {0};
2548         int cleanup_phase = 0;
2549         void *handle = NULL;
2550         int update_mode, rc = 0;
2551         ENTRY;
2552
2553         /* another MDS executing rename operation has asked us
2554          * to create target name. such a creation should destroy
2555          * existing target name */
2556
2557         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2558                         obd->obd_name, rec->ur_tgt,
2559                         (unsigned long) rec->ur_fid1->mds,
2560                         (unsigned long) rec->ur_fid1->id,
2561                         (unsigned long) rec->ur_fid1->generation);
2562
2563         /* first, lookup the target */
2564         child_lockh.cookie = 0;
2565         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2566                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2567                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2568                                          &child_lockh, &de_new, LCK_EX,
2569                                          MDS_INODELOCK_LOOKUP, NULL);
2570         if (rc)
2571                 GOTO(cleanup, rc);
2572
2573         cleanup_phase = 1;
2574
2575         LASSERT(de_srcdir);
2576         LASSERT(de_srcdir->d_inode);
2577         LASSERT(de_new);
2578
2579         if (de_new->d_inode) {
2580                 /* name exists and points to local inode
2581                  * try to unlink this name and create new one */
2582                 CERROR("%s: %s points to local inode %lu/%lu\n",
2583                        obd->obd_name, rec->ur_tgt,
2584                        (unsigned long) de_new->d_inode->i_ino,
2585                        (unsigned long) de_new->d_inode->i_generation);
2586                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2587                                       FSFILT_OP_RENAME, NULL);
2588                 if (IS_ERR(handle))
2589                         GOTO(cleanup, rc = PTR_ERR(handle));
2590                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2591                 if (rc)
2592                         GOTO(cleanup, rc);
2593         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2594                 /* name exists adn points to remove inode */
2595                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2596                        obd->obd_name, rec->ur_tgt,
2597                        (unsigned long) de_new->d_mdsnum,
2598                        (unsigned long) de_new->d_inum,
2599                        (unsigned long) de_new->d_generation);
2600         } else {
2601                 /* name doesn't exist. the simplest case */
2602                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2603                                       FSFILT_OP_LINK, NULL);
2604                 if (IS_ERR(handle))
2605                         GOTO(cleanup, rc = PTR_ERR(handle));
2606         }
2607
2608         cleanup_phase = 2;
2609         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2610                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2611                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2612         if (rc)
2613                 CERROR("add_dir_entry() returned error %d\n", rc);
2614 cleanup:
2615         EXIT;
2616         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2617                                 handle, req, rc, 0);
2618         switch(cleanup_phase) {
2619                 case 2:
2620                 case 1:
2621 #ifdef S_PDIROPS
2622                         if (parent_lockh[1].cookie != 0)
2623                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2624 #endif
2625                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2626                         if (child_lockh.cookie != 0)
2627                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2628                         l_dput(de_new);
2629                         l_dput(de_srcdir);
2630                         break;
2631                 default:
2632                         LBUG();
2633         }
2634
2635         req->rq_status = rc;
2636
2637         RETURN(0);
2638 }
2639
2640 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2641                                       struct ptlrpc_request *req)
2642 {
2643         struct obd_device *obd = req->rq_export->exp_obd;
2644         struct ptlrpc_request *req2 = NULL;
2645         struct dentry *de_srcdir = NULL;
2646         struct dentry *de_old = NULL;
2647         struct mds_obd *mds = mds_req2mds(req);
2648         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2649         struct lustre_handle child_lockh = {0};
2650         struct mdc_op_data opdata;
2651         void *handle = NULL;
2652         int update_mode, rc = 0;
2653         ENTRY;
2654
2655         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2656                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2657         memset(&opdata, 0, sizeof(opdata));
2658
2659         child_lockh.cookie = 0;
2660         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2661                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2662                                          &update_mode, rec->ur_name,
2663                                          rec->ur_namelen, &child_lockh, &de_old,
2664                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2665         LASSERT(rc == 0);
2666         LASSERT(de_srcdir);
2667         LASSERT(de_srcdir->d_inode);
2668         LASSERT(de_old);
2669
2670         /* we already know the target should be created on another MDS
2671          * so, we have to request that MDS to do it */
2672
2673         /* prepare source fid */
2674         if (de_old->d_flags & DCACHE_CROSS_REF) {
2675                 LASSERT(de_old->d_inode == NULL);
2676                 CDEBUG(D_OTHER, "request to move remote name\n");
2677                 opdata.fid1.mds = de_old->d_mdsnum;
2678                 opdata.fid1.id = de_old->d_inum;
2679                 opdata.fid1.generation = de_old->d_generation;
2680         } else if (de_old->d_inode == NULL) {
2681                 /* oh, source doesn't exist */
2682                 GOTO(cleanup, rc = -ENOENT);
2683         } else {
2684                 LASSERT(de_old->d_inode != NULL);
2685                 CDEBUG(D_OTHER, "request to move local name\n");
2686                 opdata.fid1.mds = mds->mds_num;
2687                 opdata.fid1.id = de_old->d_inode->i_ino;
2688                 opdata.fid1.generation = de_old->d_inode->i_generation;
2689         }
2690
2691         opdata.fid2 = *(rec->ur_fid2);
2692         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2693                        rec->ur_tgtlen - 1, &req2);
2694
2695         if (rc)
2696                 GOTO(cleanup, rc);
2697
2698         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2699         if (IS_ERR(handle))
2700                 GOTO(cleanup, rc = PTR_ERR(handle));
2701         rc = fsfilt_del_dir_entry(obd, de_old);
2702         d_drop(de_old);
2703
2704 cleanup:
2705         EXIT;
2706         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2707                                 handle, req, rc, 0);
2708         if (req2)
2709                 ptlrpc_req_finished(req2);
2710
2711 #ifdef S_PDIROPS
2712         if (parent_lockh[1].cookie != 0)
2713                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2714 #endif
2715         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2716         if (child_lockh.cookie != 0)
2717                 ldlm_lock_decref(&child_lockh, LCK_EX);
2718
2719         l_dput(de_old);
2720         l_dput(de_srcdir);
2721
2722         req->rq_status = rc;
2723         RETURN(0);
2724
2725 }
2726
2727 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2728                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2729 {
2730         struct obd_device *obd = req->rq_export->exp_obd;
2731         struct dentry *de_srcdir = NULL;
2732         struct dentry *de_tgtdir = NULL;
2733         struct dentry *de_old = NULL;
2734         struct dentry *de_new = NULL;
2735         struct mds_obd *mds = mds_req2mds(req);
2736         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2737         struct mds_body *body = NULL;
2738         struct llog_create_locks *lcl = NULL;
2739         int rc = 0, log_unlink = 0, lock_count = 3;
2740         int cleanup_phase = 0;
2741         void *handle = NULL;
2742         ENTRY;
2743
2744         LASSERT(offset == 0);
2745
2746         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2747                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2748                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2749
2750         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2751
2752         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2753                 DEBUG_REQ(D_HA, req, "rename replay\n");
2754                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2755                        lustre_msg_buf(req->rq_reqmsg, 3, 0),
2756                        req->rq_repmsg->buflens[2]);
2757         }
2758
2759         MDS_UPDATE_COUNTER(mds, MDS_RENAME_COUNT);
2760
2761         if (rec->ur_namelen == 1) {
2762                 rc = mds_reint_rename_create_name(rec, offset, req);
2763                 RETURN(rc);
2764         }
2765
2766         if (rec->ur_fid2->mds != mds->mds_num) {
2767                 rc = mds_reint_rename_to_remote(rec, offset, req);
2768                 RETURN(rc);
2769         }
2770
2771         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2772                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2773                                              rec->ur_name, rec->ur_namelen,
2774                                              &de_old, rec->ur_tgt,
2775                                              rec->ur_tgtlen, &de_new,
2776                                              dlm_handles, LCK_EX);
2777         if (rc)
2778                 GOTO(cleanup, rc);
2779
2780         cleanup_phase = 1; /* parent(s), children, locks */
2781
2782         if (de_new->d_inode)
2783                 lock_count = 4;
2784
2785         /* sanity check for src inode */
2786         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2787             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2788                 GOTO(cleanup, rc = -EINVAL);
2789
2790         /* sanity check for dest inode */
2791         if (de_new->d_inode &&
2792             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2793              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2794                 GOTO(cleanup, rc = -EINVAL);
2795
2796         if (de_old->d_inode == de_new->d_inode)
2797                 GOTO(cleanup, rc = 0);
2798
2799         /* if we are about to remove the target at first, pass the EA of
2800          * that inode to client to perform and cleanup on OST */
2801         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2802         LASSERT(body != NULL);
2803
2804         if (de_new->d_inode &&
2805             S_ISREG(de_new->d_inode->i_mode) &&
2806             de_new->d_inode->i_nlink == 1 &&
2807             mds_open_orphan_count(de_new->d_inode) == 0) {
2808                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2809                 mds_pack_inode2body(obd, body, de_new->d_inode);
2810                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2811                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2812                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2813                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2814                 } else {
2815                         log_unlink = 1;
2816                 }
2817         }
2818
2819         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2820                        de_srcdir->d_inode->i_sb);
2821
2822         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2823         if (IS_ERR(handle))
2824                 GOTO(cleanup, rc = PTR_ERR(handle));
2825
2826         /* FIXME need adjust the journal block count? */
2827         /* if the target should be moved to PENDING, we at first increase the
2828          * link and later vfs_rename() will decrease the link count again */
2829         if (de_new->d_inode &&
2830             S_ISREG(de_new->d_inode->i_mode) &&
2831             de_new->d_inode->i_nlink == 1 &&
2832             mds_open_orphan_count(de_new->d_inode) > 0) {
2833                 rc = mds_add_link_orphan(rec, obd, de_new);
2834                 if (rc)
2835                         GOTO(cleanup, rc);
2836         }
2837
2838         lock_kernel();
2839         de_old->d_fsdata = req;
2840         de_new->d_fsdata = req;
2841         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2842         unlock_kernel();
2843
2844         if (!rc && log_unlink) {
2845                 if (mds_log_op_unlink(obd, de_tgtdir->d_inode,
2846                                 lustre_msg_buf(req->rq_repmsg, 1, 0),
2847                                 req->rq_repmsg->buflens[1],
2848                                 lustre_msg_buf(req->rq_repmsg, 2, 0),
2849                                 req->rq_repmsg->buflens[2], &lcl) > 0) {
2850                         body->valid |= OBD_MD_FLCOOKIE;
2851                 }
2852         }
2853
2854         GOTO(cleanup, rc);
2855 cleanup:
2856         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2857                                 handle, req, rc, 0);
2858         switch (cleanup_phase) {
2859         case 1:
2860 #ifdef S_PDIROPS
2861                 if (dlm_handles[5].cookie != 0)
2862                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2863                 if (dlm_handles[6].cookie != 0)
2864                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2865 #endif
2866                 if (lcl != NULL)
2867                         ptlrpc_save_llog_lock(req, lcl);
2868
2869                 if (rc) {
2870                         if (lock_count == 4)
2871                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2872                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2873                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2874                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2875                 } else {
2876                         if (lock_count == 4)
2877                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2878                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2879                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2880                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2881                 }
2882                 l_dput(de_new);
2883                 l_dput(de_old);
2884                 l_dput(de_tgtdir);
2885                 l_dput(de_srcdir);
2886         case 0:
2887                 break;
2888         default:
2889                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2890                 LBUG();
2891         }
2892         req->rq_status = rc;
2893         return 0;
2894 }
2895
2896 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2897                            struct ptlrpc_request *, struct lustre_handle *);
2898
2899 static mds_reinter reinters[REINT_MAX + 1] = {
2900         [REINT_SETATTR] mds_reint_setattr,
2901         [REINT_CREATE] mds_reint_create,
2902         [REINT_LINK] mds_reint_link,
2903         [REINT_UNLINK] mds_reint_unlink,
2904         [REINT_RENAME] mds_reint_rename,
2905         [REINT_OPEN] mds_open
2906 };
2907
2908 int mds_reint_rec(struct mds_update_record *rec, int offset,
2909                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2910 {
2911         struct obd_device *obd = req->rq_export->exp_obd;
2912         struct lvfs_run_ctxt saved;
2913         int rc;
2914
2915         /* checked by unpacker */
2916         LASSERT(rec->ur_opcode <= REINT_MAX &&
2917                 reinters[rec->ur_opcode] != NULL);
2918
2919         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2920         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2921         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2922
2923         return rc;
2924 }