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_fsfilt.h>
  42
  43 #include "mds_internal.h"
  44
  45 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  46                    int error)
  47 {
  48         obd_transno_commit_cb(obd, transno, error);
  49 }
  50
  51 struct mds_logcancel_data {
  52         struct lov_mds_md      *mlcd_lmm;
  53         int                     mlcd_size;
  54         int                     mlcd_cookielen;
  55         int                     mlcd_eadatalen;
  56         struct llog_cookie      mlcd_cookies[0];
  57 };
  58
  59
  60 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  61                                   void *cb_data, int error)
  62 {
  63         struct mds_logcancel_data *mlcd = cb_data;
  64         struct lov_stripe_md *lsm = NULL;
  65         struct llog_ctxt *ctxt;
  66         int rc;
  67
  68         obd_transno_commit_cb(obd, transno, error);
  69
  70         CDEBUG(D_HA, "cancelling %d cookies\n",
  71                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  72
  73         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  74                           mlcd->mlcd_eadatalen);
  75         if (rc < 0) {
  76                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  77                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  78                        rc);
  79         } else {
  80                 ///* XXX 0 normally, SENDNOW for debug */);
  81                 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
  82                 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
  83                                                 sizeof(*mlcd->mlcd_cookies),
  84                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  85                 if (rc)
  86                         CERROR("error cancelling %d log cookies: rc %d\n",
  87                                (int)(mlcd->mlcd_cookielen /
  88                                      sizeof(*mlcd->mlcd_cookies)), rc);
  89         }
  90
  91         OBD_FREE(mlcd, mlcd->mlcd_size);
  92 }
  93
  94 /* Assumes caller has already pushed us into the kernel context. */
  95 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  96                        struct ptlrpc_request *req, int rc, __u32 op_data)
  97 {
  98         struct mds_export_data *med = &req->rq_export->exp_mds_data;
  99         struct mds_client_data *mcd = med->med_mcd;
 100         struct obd_device *obd = req->rq_export->exp_obd;
 101         int err;
 102         __u64 transno;
 103         loff_t off;
 104         int log_pri = D_HA;
 105         ENTRY;
 106
 107         /* if the export has already been failed, we have no last_rcvd slot */
 108         if (req->rq_export->exp_failed) {
 109                 CERROR("committing transaction for disconnected client\n");
 110                 if (handle)
 111                         GOTO(commit, rc);
 112                 RETURN(rc);
 113         }
 114
 115         if (IS_ERR(handle))
 116                 RETURN(rc);
 117
 118         if (handle == NULL) {
 119                 /* if we're starting our own xaction, use our own inode */
 120                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 121                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 122                 if (IS_ERR(handle)) {
 123                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 124                         RETURN(PTR_ERR(handle));
 125                 }
 126         }
 127
 128         off = med->med_off;
 129
 130         transno = req->rq_reqmsg->transno;
 131         if (rc != 0) {
 132                 LASSERT(transno == 0);
 133         } else if (transno == 0) {
 134                 spin_lock(&mds->mds_transno_lock);
 135                 transno = ++mds->mds_last_transno;
 136                 spin_unlock(&mds->mds_transno_lock);
 137         } else {
 138                 spin_lock(&mds->mds_transno_lock);
 139                 if (transno > mds->mds_last_transno)
 140                         mds->mds_last_transno = transno;
 141                 spin_unlock(&mds->mds_transno_lock);
 142         }
 143         req->rq_repmsg->transno = req->rq_transno = transno;
 144         mcd->mcd_last_transno = cpu_to_le64(transno);
 145         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 146         mcd->mcd_last_result = cpu_to_le32(rc);
 147         mcd->mcd_last_data = cpu_to_le32(op_data);
 148
 149         fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 150                               mds_commit_cb, NULL);
 151         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 152                                   &off, 0);
 153
 154         if (err) {
 155                 log_pri = D_ERROR;
 156                 if (rc == 0)
 157                         rc = err;
 158         }
 159
 160         DEBUG_REQ(log_pri, req,
 161                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 162                   transno, mcd->mcd_uuid, med->med_idx, err);
 163
 164         err = mds_lov_write_objids(obd);
 165         if (err) {
 166                 log_pri = D_ERROR;
 167                 if (rc == 0)
 168                         rc = err;
 169         }
 170         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 171
 172 commit:
 173         err = fsfilt_commit(obd, inode, handle, 0);
 174         if (err) {
 175                 CERROR("error committing transaction: %d\n", err);
 176                 if (!rc)
 177                         rc = err;
 178         }
 179
 180         RETURN(rc);
 181 }
 182
 183 /* this gives the same functionality as the code between
 184  * sys_chmod and inode_setattr
 185  * chown_common and inode_setattr
 186  * utimes and inode_setattr
 187  */
 188 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 189 {
 190         time_t now = LTIME_S(CURRENT_TIME);
 191         struct iattr *attr = &rec->ur_iattr;
 192         unsigned int ia_valid = attr->ia_valid;
 193         int error;
 194         ENTRY;
 195
 196         /* only fix up attrs if the client VFS didn't already */
 197         if (!(ia_valid & ATTR_RAW))
 198                 RETURN(0);
 199
 200         if (!(ia_valid & ATTR_CTIME_SET))
 201                 LTIME_S(attr->ia_ctime) = now;
 202         if (!(ia_valid & ATTR_ATIME_SET))
 203                 LTIME_S(attr->ia_atime) = now;
 204         if (!(ia_valid & ATTR_MTIME_SET))
 205                 LTIME_S(attr->ia_mtime) = now;
 206
 207         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 208                 RETURN(-EPERM);
 209
 210         /* times */
 211         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 212                 if (rec->ur_fsuid != inode->i_uid &&
 213                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 214                         RETURN(error);
 215         }
 216
 217         if (ia_valid & ATTR_SIZE) {
 218                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 219                         RETURN(error);
 220         }
 221
 222         if (ia_valid & ATTR_UID) {
 223                 /* chown */
 224                 error = -EPERM;
 225                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 226                         RETURN(-EPERM);
 227                 if (attr->ia_uid == (uid_t) -1)
 228                         attr->ia_uid = inode->i_uid;
 229                 if (attr->ia_gid == (gid_t) -1)
 230                         attr->ia_gid = inode->i_gid;
 231                 attr->ia_mode = inode->i_mode;
 232                 /*
 233                  * If the user or group of a non-directory has been
 234                  * changed by a non-root user, remove the setuid bit.
 235                  * 19981026 David C Niemi <niemi@tux.org>
 236                  *
 237                  * Changed this to apply to all users, including root,
 238                  * to avoid some races. This is the behavior we had in
 239                  * 2.0. The check for non-root was definitely wrong
 240                  * for 2.2 anyway, as it should have been using
 241                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 242                  */
 243                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 244                     !S_ISDIR(inode->i_mode)) {
 245                         attr->ia_mode &= ~S_ISUID;
 246                         attr->ia_valid |= ATTR_MODE;
 247                 }
 248                 /*
 249                  * Likewise, if the user or group of a non-directory
 250                  * has been changed by a non-root user, remove the
 251                  * setgid bit UNLESS there is no group execute bit
 252                  * (this would be a file marked for mandatory
 253                  * locking).  19981026 David C Niemi <niemi@tux.org>
 254                  *
 255                  * Removed the fsuid check (see the comment above) --
 256                  * 19990830 SD.
 257                  */
 258                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 259                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 260                         attr->ia_mode &= ~S_ISGID;
 261                         attr->ia_valid |= ATTR_MODE;
 262                 }
 263         } else if (ia_valid & ATTR_MODE) {
 264                 int mode = attr->ia_mode;
 265                 /* chmod */
 266                 if (attr->ia_mode == (mode_t) -1)
 267                         attr->ia_mode = inode->i_mode;
 268                 attr->ia_mode =
 269                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 270         }
 271         RETURN(0);
 272 }
 273
 274 void mds_steal_ack_locks(struct ptlrpc_request *req)
 275 {
 276         struct obd_export         *exp = req->rq_export;
 277         struct list_head          *tmp;
 278         struct ptlrpc_reply_state *oldrep;
 279         struct ptlrpc_service     *svc;
 280         unsigned long              flags;
 281         char                       str[PTL_NALFMT_SIZE];
 282         int                        i;
 283
 284         /* CAVEAT EMPTOR: spinlock order */
 285         spin_lock_irqsave (&exp->exp_lock, flags);
 286         list_for_each (tmp, &exp->exp_outstanding_replies) {
 287                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 288
 289                 if (oldrep->rs_xid != req->rq_xid)
 290                         continue;
 291
 292                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 293                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 294                                 "new %d old %d\n", req->rq_xid,
 295                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 296
 297                 svc = oldrep->rs_srv_ni->sni_service;
 298                 spin_lock (&svc->srv_lock);
 299
 300                 list_del_init (&oldrep->rs_exp_list);
 301
 302                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 303                       " o%d NID %s\n",
 304                       oldrep->rs_nlocks, oldrep,
 305                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 306                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 307
 308                 for (i = 0; i < oldrep->rs_nlocks; i++)
 309                         ptlrpc_save_lock(req,
 310                                          &oldrep->rs_locks[i],
 311                                          oldrep->rs_modes[i]);
 312                 oldrep->rs_nlocks = 0;
 313
 314                 DEBUG_REQ(D_HA, req, "stole locks for");
 315                 ptlrpc_schedule_difficult_reply (oldrep);
 316
 317                 spin_unlock (&svc->srv_lock);
 318                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 319                 return;
 320         }
 321         spin_unlock_irqrestore (&exp->exp_lock, flags);
 322 }
 323
 324 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 325 {
 326         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 327                   mcd->mcd_last_transno, mcd->mcd_last_result);
 328         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 329         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 330
 331         mds_steal_ack_locks(req);
 332 }
 333
 334 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 335                                       int offset, struct ptlrpc_request *req)
 336 {
 337         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 338         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 339         struct dentry *de;
 340         struct mds_body *body;
 341
 342         mds_req_from_mcd(req, med->med_mcd);
 343
 344         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 345         if (IS_ERR(de)) {
 346                 LASSERT(PTR_ERR(de) == req->rq_status);
 347                 return;
 348         }
 349
 350         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 351         mds_pack_inode2fid(&body->fid1, de->d_inode);
 352         mds_pack_inode2body(body, de->d_inode);
 353
 354         /* Don't return OST-specific attributes if we didn't just set them */
 355         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 356                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 357         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 358                 body->valid |= OBD_MD_FLMTIME;
 359         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 360                 body->valid |= OBD_MD_FLATIME;
 361
 362         l_dput(de);
 363 }
 364
 365 /* In the raw-setattr case, we lock the child inode.
 366  * In the write-back case or if being called from open, the client holds a lock
 367  * already.
 368  *
 369  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 370 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 371                              struct ptlrpc_request *req,
 372                              struct lustre_handle *lh)
 373 {
 374         struct mds_obd *mds = mds_req2mds(req);
 375         struct obd_device *obd = req->rq_export->exp_obd;
 376         struct mds_body *body;
 377         struct dentry *de;
 378         struct inode *inode = NULL;
 379         struct lustre_handle lockh;
 380         void *handle = NULL;
 381         struct mds_logcancel_data *mlcd = NULL;
 382         int rc = 0, cleanup_phase = 0, err, locked = 0;
 383         ENTRY;
 384
 385         LASSERT(offset == 0);
 386
 387         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 388                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 389
 390         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 391
 392         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 393                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 394                 if (IS_ERR(de))
 395                         GOTO(cleanup, rc = PTR_ERR(de));
 396         } else {
 397                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 398                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
 399                         lockpart |= MDS_INODELOCK_LOOKUP;
 400                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 401                                            &lockh, NULL, 0, lockpart);
 402                 if (IS_ERR(de))
 403                         GOTO(cleanup, rc = PTR_ERR(de));
 404                 locked = 1;
 405         }
 406
 407         cleanup_phase = 1;
 408
 409         inode = de->d_inode;
 410         LASSERT(inode);
 411         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 412             rec->ur_eadata != NULL)
 413                 down(&inode->i_sem);
 414
 415         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 416
 417         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 418         if (IS_ERR(handle))
 419                 GOTO(cleanup, rc = PTR_ERR(handle));
 420
 421         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 422                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 423                        LTIME_S(rec->ur_iattr.ia_mtime),
 424                        LTIME_S(rec->ur_iattr.ia_ctime));
 425         rc = mds_fix_attr(inode, rec);
 426         if (rc)
 427                 GOTO(cleanup, rc);
 428
 429         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 430                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 431                                       (long)&rec->ur_iattr.ia_attr_flags);
 432         else                                            /* setattr */
 433                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 434
 435         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 436             rec->ur_eadata != NULL) {
 437                 struct lov_stripe_md *lsm = NULL;
 438
 439                 rc = ll_permission(inode, MAY_WRITE, NULL);
 440                 if (rc < 0)
 441                         GOTO(cleanup, rc);
 442
 443                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 444                                    mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
 445                 if (rc)
 446                         GOTO(cleanup, rc);
 447
 448                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 449
 450                 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 451                                    rec->ur_eadatalen);
 452                 if (rc)
 453                         GOTO(cleanup, rc);
 454         }
 455
 456         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 457         mds_pack_inode2fid(&body->fid1, inode);
 458         mds_pack_inode2body(body, inode);
 459
 460         /* Don't return OST-specific attributes if we didn't just set them */
 461         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 462                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 463         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 464                 body->valid |= OBD_MD_FLMTIME;
 465         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 466                 body->valid |= OBD_MD_FLATIME;
 467
 468         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 469                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 470                           rec->ur_eadatalen);
 471                 if (mlcd) {
 472                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 473                                 rec->ur_eadatalen;
 474                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 475                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 476                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 477                                 mlcd->mlcd_cookielen;
 478                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 479                                mlcd->mlcd_cookielen);
 480                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 481                                mlcd->mlcd_eadatalen);
 482                 } else {
 483                         CERROR("unable to allocate log cancel data\n");
 484                 }
 485         }
 486         EXIT;
 487  cleanup:
 488         if (mlcd != NULL)
 489                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 490                                       mds_cancel_cookies_cb, mlcd);
 491         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 492         switch (cleanup_phase) {
 493         case 1:
 494                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 495                     rec->ur_eadata != NULL)
 496                         up(&inode->i_sem);
 497                 l_dput(de);
 498                 if (locked) {
 499                         if (rc) {
 500                                 ldlm_lock_decref(&lockh, LCK_PW);
 501                         } else {
 502                                 ptlrpc_save_lock (req, &lockh, LCK_PW);
 503                         }
 504                 }
 505         case 0:
 506                 break;
 507         default:
 508                 LBUG();
 509         }
 510         if (err && !rc)
 511                 rc = err;
 512
 513         req->rq_status = rc;
 514         return 0;
 515 }
 516
 517 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 518                                      struct ptlrpc_request *req)
 519 {
 520         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 521         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 522         struct dentry *parent, *child;
 523         struct mds_body *body;
 524
 525         mds_req_from_mcd(req, med->med_mcd);
 526
 527         if (req->rq_status)
 528                 return;
 529
 530         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 531         LASSERT(!IS_ERR(parent));
 532         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 533         LASSERT(!IS_ERR(child));
 534         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 535         mds_pack_inode2fid(&body->fid1, child->d_inode);
 536         mds_pack_inode2body(body, child->d_inode);
 537         l_dput(parent);
 538         l_dput(child);
 539 }
 540
 541 static int mds_reint_create(struct mds_update_record *rec, int offset,
 542                             struct ptlrpc_request *req,
 543                             struct lustre_handle *lh)
 544 {
 545         struct dentry *dparent = NULL;
 546         struct mds_obd *mds = mds_req2mds(req);
 547         struct obd_device *obd = req->rq_export->exp_obd;
 548         struct dentry *dchild = NULL;
 549         struct inode *dir = NULL;
 550         void *handle = NULL;
 551         struct lustre_handle lockh;
 552         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 553         int created = 0;
 554         struct dentry_params dp;
 555         ENTRY;
 556
 557         LASSERT(offset == 0);
 558         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 559
 560         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 561                   rec->ur_fid1->id, rec->ur_fid1->generation,
 562                   rec->ur_name, rec->ur_mode);
 563
 564         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 565
 566         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 567                 GOTO(cleanup, rc = -ESTALE);
 568
 569         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
 570                                         rec->ur_name, rec->ur_namelen - 1,
 571                                         MDS_INODELOCK_UPDATE);
 572         if (IS_ERR(dparent)) {
 573                 rc = PTR_ERR(dparent);
 574                 CERROR("parent lookup error %d\n", rc);
 575                 GOTO(cleanup, rc);
 576         }
 577         cleanup_phase = 1; /* locked parent dentry */
 578         dir = dparent->d_inode;
 579         LASSERT(dir);
 580
 581         ldlm_lock_dump_handle(D_OTHER, &lockh);
 582
 583         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 584         if (IS_ERR(dchild)) {
 585                 rc = PTR_ERR(dchild);
 586                 CERROR("child lookup error %d\n", rc);
 587                 GOTO(cleanup, rc);
 588         }
 589
 590         cleanup_phase = 2; /* child dentry */
 591
 592         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 593
 594         if (dir->i_mode & S_ISGID) {
 595                 if (S_ISDIR(rec->ur_mode))
 596                         rec->ur_mode |= S_ISGID;
 597         }
 598
 599         dchild->d_fsdata = (void *)&dp;
 600         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 601         dp.p_ptr = req;
 602
 603         switch (type) {
 604         case S_IFREG:{
 605                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 606                 if (IS_ERR(handle))
 607                         GOTO(cleanup, rc = PTR_ERR(handle));
 608                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 609                 EXIT;
 610                 break;
 611         }
 612         case S_IFDIR:{
 613                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 614                 if (IS_ERR(handle))
 615                         GOTO(cleanup, rc = PTR_ERR(handle));
 616                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 617                 EXIT;
 618                 break;
 619         }
 620         case S_IFLNK:{
 621                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 622                 if (IS_ERR(handle))
 623                         GOTO(cleanup, rc = PTR_ERR(handle));
 624                 if (rec->ur_tgt == NULL)        /* no target supplied */
 625                         rc = -EINVAL;           /* -EPROTO? */
 626                 else
 627                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 628                 EXIT;
 629                 break;
 630         }
 631         case S_IFCHR:
 632         case S_IFBLK:
 633         case S_IFIFO:
 634         case S_IFSOCK:{
 635                 int rdev = rec->ur_rdev;
 636                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 637                 if (IS_ERR(handle))
 638                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 639                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 640                 EXIT;
 641                 break;
 642         }
 643         default:
 644                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 645                 dchild->d_fsdata = NULL;
 646                 GOTO(cleanup, rc = -EINVAL);
 647         }
 648
 649         /* In case we stored the desired inum in here, we want to clean up. */
 650         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 651                 dchild->d_fsdata = NULL;
 652
 653         if (rc) {
 654                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 655                 GOTO(cleanup, rc);
 656         } else {
 657                 struct iattr iattr;
 658                 struct inode *inode = dchild->d_inode;
 659                 struct mds_body *body;
 660
 661                 created = 1;
 662                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 663                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 664                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 665                 iattr.ia_uid = rec->ur_fsuid;
 666                 if (dir->i_mode & S_ISGID)
 667                         iattr.ia_gid = dir->i_gid;
 668                 else
 669                         iattr.ia_gid = rec->ur_fsgid;
 670                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 671                         ATTR_MTIME | ATTR_CTIME;
 672
 673                 if (rec->ur_fid2->id) {
 674                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 675                         inode->i_generation = rec->ur_fid2->generation;
 676                         /* Dirtied and committed by the upcoming setattr. */
 677                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 678                                inode->i_ino, inode->i_generation);
 679                 } else {
 680                         struct lustre_handle child_ino_lockh;
 681
 682                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 683                                inode->i_ino, inode->i_generation);
 684
 685                         /* The inode we were allocated may have just been freed
 686                          * by an unlink operation.  We take this lock to
 687                          * synchronize against the matching reply-ack-lock taken
 688                          * in unlink, to avoid replay problems if this reply
 689                          * makes it out to the client but the unlink's does not.
 690                          * See bug 2029 for more detail.*/
 691                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 692                         if (rc != ELDLM_OK) {
 693                                 CERROR("error locking for unlink/create sync: "
 694                                        "%d\n", rc);
 695                         } else {
 696                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 697                         }
 698                 }
 699
 700                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 701                 if (rc)
 702                         CERROR("error on child setattr: rc = %d\n", rc);
 703
 704                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 705                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 706                 if (rc)
 707                         CERROR("error on parent setattr: rc = %d\n", rc);
 708
 709                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 710                 mds_pack_inode2fid(&body->fid1, inode);
 711                 mds_pack_inode2body(body, inode);
 712         }
 713         EXIT;
 714
 715 cleanup:
 716         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 717
 718         if (rc && created) {
 719                 /* Destroy the file we just created.  This should not need
 720                  * extra journal credits, as we have already modified all of
 721                  * the blocks needed in order to create the file in the first
 722                  * place.
 723                  */
 724                 switch (type) {
 725                 case S_IFDIR:
 726                         err = vfs_rmdir(dir, dchild);
 727                         if (err)
 728                                 CERROR("rmdir in error path: %d\n", err);
 729                         break;
 730                 default:
 731                         err = vfs_unlink(dir, dchild);
 732                         if (err)
 733                                 CERROR("unlink in error path: %d\n", err);
 734                         break;
 735                 }
 736         } else {
 737                 rc = err;
 738         }
 739         switch (cleanup_phase) {
 740         case 2: /* child dentry */
 741                 l_dput(dchild);
 742         case 1: /* locked parent dentry */
 743                 if (rc) {
 744                         ldlm_lock_decref(&lockh, LCK_PW);
 745                 } else {
 746                         ptlrpc_save_lock (req, &lockh, LCK_PW);
 747                 }
 748                 l_dput(dparent);
 749         case 0:
 750                 break;
 751         default:
 752                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 753                 LBUG();
 754         }
 755         req->rq_status = rc;
 756         return 0;
 757 }
 758
 759 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 760            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 761 {
 762         int i;
 763
 764         for (i = 0; i < RES_NAME_SIZE; i++) {
 765                 /* return 1 here, because enqueue_ordered will skip resources
 766                  * of all zeroes if they're sorted to the end of the list. */
 767                 if (res1->name[i] == 0 && res2->name[i] != 0)
 768                         return 1;
 769                 if (res2->name[i] == 0 && res1->name[i] != 0)
 770                         return 0;
 771
 772                 if (res1->name[i] > res2->name[i])
 773                         return 1;
 774                 if (res1->name[i] < res2->name[i])
 775                         return 0;
 776         }
 777
 778         if (!p1 || !p2)
 779                 return 0;
 780
 781         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 782                 return 1;
 783
 784         return 0;
 785 }
 786
 787 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 788  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 789  * because they take the place of local semaphores.
 790  *
 791  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 792  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 793 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 794                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 795                           ldlm_policy_data_t *p1_policy,
 796                           struct ldlm_res_id *p2_res_id,
 797                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 798                           ldlm_policy_data_t *p2_policy)
 799 {
 800         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 801         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 802         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 803         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 804         int rc, flags;
 805         ENTRY;
 806
 807         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 808
 809         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 810                res_id[1]->name[0]);
 811
 812         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 813                 handles[1] = p1_lockh;
 814                 handles[0] = p2_lockh;
 815                 res_id[1] = p1_res_id;
 816                 res_id[0] = p2_res_id;
 817                 lock_modes[1] = p1_lock_mode;
 818                 lock_modes[0] = p2_lock_mode;
 819                 policies[1] = p1_policy;
 820                 policies[0] = p2_policy;
 821         }
 822
 823         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 824                res_id[0]->name[0], res_id[1]->name[0]);
 825
 826         flags = LDLM_FL_LOCAL_ONLY;
 827         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 828                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 829                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 830                               NULL, 0, NULL, handles[0]);
 831         if (rc != ELDLM_OK)
 832                 RETURN(-EIO);
 833         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 834
 835         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
 836             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
 837                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 838                 ldlm_lock_addref(handles[1], lock_modes[1]);
 839         } else if (res_id[1]->name[0] != 0) {
 840                 flags = LDLM_FL_LOCAL_ONLY;
 841                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 842                                       *res_id[1], LDLM_IBITS, policies[1],
 843                                       lock_modes[1], &flags, mds_blocking_ast,
 844                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 845                                       NULL, handles[1]);
 846                 if (rc != ELDLM_OK) {
 847                         ldlm_lock_decref(handles[0], lock_modes[0]);
 848                         RETURN(-EIO);
 849                 }
 850                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 851         }
 852
 853         RETURN(0);
 854 }
 855
 856 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
 857                            struct lustre_handle *p1_lockh, int p1_lock_mode,
 858                            ldlm_policy_data_t *p1_policy,
 859                            struct ldlm_res_id *p2_res_id,
 860                            struct lustre_handle *p2_lockh, int p2_lock_mode,
 861                            ldlm_policy_data_t *p2_policy,
 862                            struct ldlm_res_id *c1_res_id,
 863                            struct lustre_handle *c1_lockh, int c1_lock_mode,
 864                            ldlm_policy_data_t *c1_policy,
 865                            struct ldlm_res_id *c2_res_id,
 866                            struct lustre_handle *c2_lockh, int c2_lock_mode,
 867                            ldlm_policy_data_t *c2_policy)
 868 {
 869         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
 870                                           c1_res_id, c2_res_id };
 871         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
 872                                                  c1_lockh, c2_lockh };
 873         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
 874                               c1_lock_mode, c2_lock_mode };
 875         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
 876                                             c1_policy, c2_policy};
 877         int rc, i, j, sorted, flags;
 878         ENTRY;
 879
 880         CDEBUG(D_DLMTRACE,
 881                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 882                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 883                res_id[3]->name[0]);
 884
 885         /* simple insertion sort - we have at most 4 elements */
 886         for (i = 1; i < 4; i++) {
 887                 j = i - 1;
 888                 dlm_handles[4] = dlm_handles[i];
 889                 res_id[4] = res_id[i];
 890                 lock_modes[4] = lock_modes[i];
 891                 policies[4] = policies[i];
 892
 893                 sorted = 0;
 894                 do {
 895                         if (res_gt(res_id[j], res_id[4], policies[j],
 896                                    policies[4])) {
 897                                 dlm_handles[j + 1] = dlm_handles[j];
 898                                 res_id[j + 1] = res_id[j];
 899                                 lock_modes[j + 1] = lock_modes[j];
 900                                 policies[j + 1] = policies[j];
 901                                 j--;
 902                         } else {
 903                                 sorted = 1;
 904                         }
 905                 } while (j >= 0 && !sorted);
 906
 907                 dlm_handles[j + 1] = dlm_handles[4];
 908                 res_id[j + 1] = res_id[4];
 909                 lock_modes[j + 1] = lock_modes[4];
 910                 policies[j + 1] = policies[4];
 911         }
 912
 913         CDEBUG(D_DLMTRACE,
 914                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 915                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 916                res_id[3]->name[0]);
 917
 918         /* XXX we could send ASTs on all these locks first before blocking? */
 919         for (i = 0; i < 4; i++) {
 920                 flags = 0;
 921                 if (res_id[i]->name[0] == 0)
 922                         break;
 923                 if (i != 0 &&
 924                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
 925                     (policies[i]->l_inodebits.bits &
 926                      policies[i-1]->l_inodebits.bits) ) {
 927                         memcpy(dlm_handles[i], dlm_handles[i-1],
 928                                sizeof(*(dlm_handles[i])));
 929                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
 930                 } else {
 931                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 932                                               *res_id[i], LDLM_IBITS,
 933                                               policies[i],
 934                                               lock_modes[i], &flags,
 935                                               mds_blocking_ast,
 936                                               ldlm_completion_ast, NULL, NULL,
 937                                               NULL, 0, NULL, dlm_handles[i]);
 938                         if (rc != ELDLM_OK)
 939                                 GOTO(out_err, rc = -EIO);
 940                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
 941                 }
 942         }
 943
 944         RETURN(0);
 945 out_err:
 946         while (i-- > 0)
 947                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
 948
 949         return rc;
 950 }
 951
 952 /* In the unlikely case that the child changed while we were waiting
 953  * on the lock, we need to drop the lock on the old child and either:
 954  * - if the child has a lower resource name, then we have to also
 955  *   drop the parent lock and regain the locks in the right order
 956  * - in the rename case, if the child has a lower resource name than one of
 957  *   the other parent/child resources (maxres) we also need to reget the locks
 958  * - if the child has a higher resource name (this is the common case)
 959  *   we can just get the lock on the new child (still in lock order)
 960  *
 961  * Returns 0 if the child did not change or if it changed but could be locked.
 962  * Returns 1 if the child changed and we need to re-lock (no locks held).
 963  * Returns -ve error with a valid dchild (no locks held). */
 964 static int mds_verify_child(struct obd_device *obd,
 965                             struct ldlm_res_id *parent_res_id,
 966                             struct lustre_handle *parent_lockh,
 967                             struct dentry *dparent, int parent_mode,
 968                             struct ldlm_res_id *child_res_id,
 969                             struct lustre_handle *child_lockh,
 970                             struct dentry **dchildp, int child_mode,
 971                             ldlm_policy_data_t *child_policy,
 972                             const char *name, int namelen,
 973                             struct ldlm_res_id *maxres)
 974 {
 975         struct dentry *vchild, *dchild = *dchildp;
 976         int rc = 0, cleanup_phase = 2; /* parent, child locks */
 977         ENTRY;
 978
 979         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
 980         if (IS_ERR(vchild))
 981                 GOTO(cleanup, rc = PTR_ERR(vchild));
 982
 983         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
 984                    (vchild->d_inode != NULL &&
 985                     child_res_id->name[0] == vchild->d_inode->i_ino &&
 986                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
 987                 if (dchild != NULL)
 988                         l_dput(dchild);
 989                 *dchildp = vchild;
 990
 991                 RETURN(0);
 992         }
 993
 994         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
 995                vchild->d_inode, dchild ? dchild->d_inode : 0,
 996                vchild->d_inode ? vchild->d_inode->i_ino : 0,
 997                child_res_id->name[0]);
 998         if (child_res_id->name[0] != 0)
 999                 ldlm_lock_decref(child_lockh, child_mode);
1000         if (dchild)
1001                 l_dput(dchild);
1002
1003         cleanup_phase = 1; /* parent lock only */
1004         *dchildp = dchild = vchild;
1005
1006         if (dchild->d_inode) {
1007                 int flags = 0;
1008                 child_res_id->name[0] = dchild->d_inode->i_ino;
1009                 child_res_id->name[1] = dchild->d_inode->i_generation;
1010
1011                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1012                     res_gt(maxres, child_res_id, NULL, NULL)) {
1013                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1014                                child_res_id->name[0], parent_res_id->name[0],
1015                                maxres->name[0]);
1016                         GOTO(cleanup, rc = 1);
1017                 }
1018
1019                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1020                                       *child_res_id, LDLM_IBITS, child_policy,
1021                                       child_mode, &flags, mds_blocking_ast,
1022                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1023                                       NULL, child_lockh);
1024                 if (rc != ELDLM_OK)
1025                         GOTO(cleanup, rc = -EIO);
1026
1027         } else {
1028                 memset(child_res_id, 0, sizeof(*child_res_id));
1029         }
1030
1031         EXIT;
1032 cleanup:
1033         if (rc) {
1034                 switch(cleanup_phase) {
1035                 case 2:
1036                         if (child_res_id->name[0] != 0)
1037                                 ldlm_lock_decref(child_lockh, child_mode);
1038                 case 1:
1039                         ldlm_lock_decref(parent_lockh, parent_mode);
1040                 }
1041         }
1042         return rc;
1043 }
1044
1045 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1046                                 struct ll_fid *fid,
1047                                 struct lustre_handle *parent_lockh,
1048                                 struct dentry **dparentp, int parent_mode,
1049                                 __u64 parent_lockpart,
1050                                 char *name, int namelen,
1051                                 struct lustre_handle *child_lockh,
1052                                 struct dentry **dchildp, int child_mode,
1053                                 __u64 child_lockpart)
1054 {
1055         struct ldlm_res_id child_res_id = { .name = {0} };
1056         struct ldlm_res_id parent_res_id = { .name = {0} };
1057         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1058         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1059         struct inode *inode;
1060         int rc = 0, cleanup_phase = 0;
1061         ENTRY;
1062
1063         /* Step 1: Lookup parent */
1064         *dparentp = mds_fid2dentry(mds, fid, NULL);
1065         if (IS_ERR(*dparentp)) {
1066                 rc = PTR_ERR(*dparentp);
1067                 *dparentp = NULL;
1068                 RETURN(rc);
1069         }
1070
1071         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1072                (*dparentp)->d_inode->i_ino, name);
1073
1074         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1075         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1076
1077         cleanup_phase = 1; /* parent dentry */
1078
1079         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1080         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1081         if (IS_ERR(*dchildp)) {
1082                 rc = PTR_ERR(*dchildp);
1083                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1084                 GOTO(cleanup, rc);
1085         }
1086
1087         inode = (*dchildp)->d_inode;
1088         if (inode != NULL)
1089                 inode = igrab(inode);
1090         if (inode == NULL)
1091                 goto retry_locks;
1092
1093         child_res_id.name[0] = inode->i_ino;
1094         child_res_id.name[1] = inode->i_generation;
1095
1096         iput(inode);
1097
1098 retry_locks:
1099         cleanup_phase = 2; /* child dentry */
1100
1101         /* Step 3: Lock parent and child in resource order.  If child doesn't
1102          *         exist, we still have to lock the parent and re-lookup. */
1103         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1104                                    &parent_policy,
1105                                    &child_res_id, child_lockh, child_mode,
1106                                    &child_policy);
1107         if (rc)
1108                 GOTO(cleanup, rc);
1109
1110         if (!(*dchildp)->d_inode)
1111                 cleanup_phase = 3; /* parent lock */
1112         else
1113                 cleanup_phase = 4; /* child lock */
1114
1115         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1116         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1117                               parent_mode, &child_res_id, child_lockh,
1118                               dchildp, child_mode, &child_policy,
1119                               name, namelen, &parent_res_id);
1120         if (rc > 0)
1121                 goto retry_locks;
1122         if (rc < 0) {
1123                 cleanup_phase = 3;
1124                 GOTO(cleanup, rc);
1125         }
1126
1127 cleanup:
1128         if (rc) {
1129                 switch (cleanup_phase) {
1130                 case 4:
1131                         ldlm_lock_decref(child_lockh, child_mode);
1132                 case 3:
1133                         ldlm_lock_decref(parent_lockh, parent_mode);
1134                 case 2:
1135                         l_dput(*dchildp);
1136                 case 1:
1137                         l_dput(*dparentp);
1138                 default: ;
1139                 }
1140         }
1141         return rc;
1142 }
1143
1144 void mds_reconstruct_generic(struct ptlrpc_request *req)
1145 {
1146         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1147
1148         mds_req_from_mcd(req, med->med_mcd);
1149 }
1150
1151 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1152                             struct ptlrpc_request *req,
1153                             struct lustre_handle *lh)
1154 {
1155         struct dentry *dparent, *dchild;
1156         struct mds_obd *mds = mds_req2mds(req);
1157         struct obd_device *obd = req->rq_export->exp_obd;
1158         struct mds_body *body = NULL;
1159         struct inode *child_inode;
1160         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1161         void *handle = NULL;
1162         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1163         ENTRY;
1164
1165         LASSERT(offset == 0 || offset == 2);
1166
1167         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1168                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1169
1170         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1171
1172         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1173                 GOTO(cleanup, rc = -ENOENT);
1174
1175         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1176                                          &parent_lockh, &dparent, LCK_PW,
1177                                          MDS_INODELOCK_UPDATE,
1178                                          rec->ur_name, rec->ur_namelen,
1179                                          &child_lockh, &dchild, LCK_EX,
1180                                          MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1181         if (rc)
1182                 GOTO(cleanup, rc);
1183
1184         cleanup_phase = 1; /* dchild, dparent, locks */
1185
1186         child_inode = dchild->d_inode;
1187         if (child_inode == NULL) {
1188                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1189                        dparent->d_inode->i_ino, rec->ur_name);
1190                 GOTO(cleanup, rc = -ENOENT);
1191         }
1192
1193         cleanup_phase = 2; /* dchild has a lock */
1194
1195         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1196          * reuse (see bug 2029). */
1197         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1198         if (rc != ELDLM_OK)
1199                 GOTO(cleanup, rc);
1200
1201         cleanup_phase = 3; /* child inum lock */
1202
1203         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1204
1205         /* ldlm_reply in buf[0] if called via intent */
1206         if (offset)
1207                 offset = 1;
1208
1209         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1210         LASSERT(body != NULL);
1211
1212         /* If this is the last reference to this inode, get the OBD EA
1213          * data first so the client can destroy OST objects.
1214          * we only do the object removal if no open files remain.
1215          * Nobody can get at this name anymore because of the locks so
1216          * we make decisions here as to whether to remove the inode */
1217         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1218             mds_open_orphan_count(child_inode) == 0) {
1219                 mds_pack_inode2fid(&body->fid1, child_inode);
1220                 mds_pack_inode2body(body, child_inode);
1221                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1222                             child_inode, 1);
1223                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1224                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1225                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1226                 } else {
1227                         log_unlink = 1;
1228                 }
1229         }
1230
1231         /* We have to do these checks ourselves, in case we are making an
1232          * orphan.  The client tells us whether rmdir() or unlink() was called,
1233          * so we need to return appropriate errors (bug 72).
1234          *
1235          * We don't have to check permissions, because vfs_rename (called from
1236          * mds_open_unlink_rename) also calls may_delete. */
1237         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1238                 if (!S_ISDIR(child_inode->i_mode))
1239                         GOTO(cleanup, rc = -ENOTDIR);
1240         } else {
1241                 if (S_ISDIR(child_inode->i_mode))
1242                         GOTO(cleanup, rc = -EISDIR);
1243         }
1244
1245         if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1246             mds_open_orphan_count(child_inode) > 0) {
1247                 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1248                 cleanup_phase = 4; /* transaction */
1249                 GOTO(cleanup, rc);
1250         }
1251
1252         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1253         switch (child_inode->i_mode & S_IFMT) {
1254         case S_IFDIR:
1255                 /* Drop any lingering child directories before we start our
1256                  * transaction, to avoid doing multiple inode dirty/delete
1257                  * in our compound transaction (bug 1321). */
1258                 shrink_dcache_parent(dchild);
1259                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1260                                       NULL);
1261                 if (IS_ERR(handle))
1262                         GOTO(cleanup, rc = PTR_ERR(handle));
1263                 cleanup_phase = 4; /* transaction */
1264                 rc = vfs_rmdir(dparent->d_inode, dchild);
1265                 break;
1266         case S_IFREG: {
1267                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1268                                                         offset + 1, 0);
1269                 handle = fsfilt_start_log(obd, dparent->d_inode,
1270                                           FSFILT_OP_UNLINK, NULL,
1271                                           le32_to_cpu(lmm->lmm_stripe_count));
1272                 if (IS_ERR(handle))
1273                         GOTO(cleanup, rc = PTR_ERR(handle));
1274
1275                 cleanup_phase = 4; /* transaction */
1276                 rc = vfs_unlink(dparent->d_inode, dchild);
1277
1278                 if (!rc && log_unlink)
1279                         if (mds_log_op_unlink(obd, child_inode,
1280                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1281                                 req->rq_repmsg->buflens[offset + 1],
1282                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1283                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1284                                 body->valid |= OBD_MD_FLCOOKIE;
1285                 break;
1286         }
1287         case S_IFLNK:
1288         case S_IFCHR:
1289         case S_IFBLK:
1290         case S_IFIFO:
1291         case S_IFSOCK:
1292                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1293                                       NULL);
1294                 if (IS_ERR(handle))
1295                         GOTO(cleanup, rc = PTR_ERR(handle));
1296                 cleanup_phase = 4; /* transaction */
1297                 rc = vfs_unlink(dparent->d_inode, dchild);
1298                 break;
1299         default:
1300                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1301                        rec->ur_name);
1302                 LBUG();
1303                 GOTO(cleanup, rc = -EINVAL);
1304         }
1305
1306  cleanup:
1307         if (rc == 0) {
1308                 struct iattr iattr;
1309                 int err;
1310
1311                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1312                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1313                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1314
1315                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1316                 if (err)
1317                         CERROR("error on parent setattr: rc = %d\n", err);
1318         }
1319
1320         switch(cleanup_phase) {
1321         case 4:
1322                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1323                                         rc, 0);
1324                 if (!rc)
1325                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1326                                            "unlinked", 0, NULL);
1327         case 3: /* child ino-reuse lock */
1328                 if (rc && body != NULL) {
1329                         // Don't unlink the OST objects if the MDS unlink failed
1330                         body->valid = 0;
1331                 }
1332                 if (rc)
1333                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1334                 else
1335                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1336         case 2: /* child lock */
1337                 ldlm_lock_decref(&child_lockh, LCK_EX);
1338         case 1: /* child and parent dentry, parent lock */
1339                 if (rc)
1340                         ldlm_lock_decref(&parent_lockh, LCK_PW);
1341                 else
1342                         ptlrpc_save_lock(req, &parent_lockh, LCK_PW);
1343                 l_dput(dchild);
1344                 l_dput(dparent);
1345         case 0:
1346                 break;
1347         default:
1348                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1349                 LBUG();
1350         }
1351         req->rq_status = rc;
1352         return 0;
1353 }
1354
1355 static int mds_reint_link(struct mds_update_record *rec, int offset,
1356                           struct ptlrpc_request *req,
1357                           struct lustre_handle *lh)
1358 {
1359         struct obd_device *obd = req->rq_export->exp_obd;
1360         struct dentry *de_src = NULL;
1361         struct dentry *de_tgt_dir = NULL;
1362         struct dentry *dchild = NULL;
1363         struct mds_obd *mds = mds_req2mds(req);
1364         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1365         struct ldlm_res_id src_res_id = { .name = {0} };
1366         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1367         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1368         ldlm_policy_data_t tgt_dir_policy =
1369                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1370
1371         int rc = 0, cleanup_phase = 0;
1372         ENTRY;
1373
1374         LASSERT(offset == 0);
1375
1376         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1377                   rec->ur_fid1->id, rec->ur_fid1->generation,
1378                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1379
1380         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1381
1382         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1383                 GOTO(cleanup, rc = -ENOENT);
1384
1385         /* Step 1: Lookup the source inode and target directory by FID */
1386         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1387         if (IS_ERR(de_src))
1388                 GOTO(cleanup, rc = PTR_ERR(de_src));
1389
1390         cleanup_phase = 1; /* source dentry */
1391
1392         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1393         if (IS_ERR(de_tgt_dir)) {
1394                 rc = PTR_ERR(de_tgt_dir);
1395                 de_tgt_dir = NULL;
1396                 GOTO(cleanup, rc);
1397         }
1398
1399         cleanup_phase = 2; /* target directory dentry */
1400
1401         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1402                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1403                de_src->d_inode->i_ino);
1404
1405         /* Step 2: Take the two locks */
1406         src_res_id.name[0] = de_src->d_inode->i_ino;
1407         src_res_id.name[1] = de_src->d_inode->i_generation;
1408         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1409         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1410
1411         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1412                                    &src_policy,
1413                                    &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
1414                                    &tgt_dir_policy);
1415         if (rc)
1416                 GOTO(cleanup, rc);
1417
1418         cleanup_phase = 3; /* locks */
1419
1420         /* Step 3: Lookup the child */
1421         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1422         if (IS_ERR(dchild)) {
1423                 rc = PTR_ERR(dchild);
1424                 if (rc != -EPERM && rc != -EACCES)
1425                         CERROR("child lookup error %d\n", rc);
1426                 GOTO(cleanup, rc);
1427         }
1428
1429         cleanup_phase = 4; /* child dentry */
1430
1431         if (dchild->d_inode) {
1432                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1433                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1434                 rc = -EEXIST;
1435                 GOTO(cleanup, rc);
1436         }
1437
1438         /* Step 4: Do it. */
1439         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1440
1441         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1442         if (IS_ERR(handle)) {
1443                 rc = PTR_ERR(handle);
1444                 GOTO(cleanup, rc);
1445         }
1446
1447         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1448         if (rc && rc != -EPERM && rc != -EACCES)
1449                 CERROR("vfs_link error %d\n", rc);
1450 cleanup:
1451         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1452                                 handle, req, rc, 0);
1453         EXIT;
1454
1455         switch (cleanup_phase) {
1456         case 4: /* child dentry */
1457                 l_dput(dchild);
1458         case 3: /* locks */
1459                 if (rc) {
1460                         ldlm_lock_decref(&src_lockh, LCK_EX);
1461                         ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1462                 } else {
1463                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1464                         ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1465                 }
1466         case 2: /* target dentry */
1467                 l_dput(de_tgt_dir);
1468         case 1: /* source dentry */
1469                 l_dput(de_src);
1470         case 0:
1471                 break;
1472         default:
1473                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1474                 LBUG();
1475         }
1476         req->rq_status = rc;
1477         return 0;
1478 }
1479
1480 /*
1481  * add a hard link in the PENDING directory, only used by rename()
1482  */
1483 static int mds_add_link_orphan(struct mds_update_record *rec,
1484                                struct obd_device *obd,
1485                                struct dentry *dentry)
1486 {
1487         struct mds_obd *mds = &obd->u.mds;
1488         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1489         struct dentry *pending_child;
1490         char fidname[LL_FID_NAMELEN];
1491         int fidlen = 0, rc;
1492         ENTRY;
1493
1494         LASSERT(dentry->d_inode);
1495         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
1496
1497         down(&pending_dir->i_sem);
1498         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
1499                             dentry->d_inode->i_generation);
1500
1501         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
1502                mds_open_orphan_count(dentry->d_inode),
1503                rec->ur_name, fidname);
1504
1505         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1506         if (IS_ERR(pending_child))
1507                 GOTO(out_lock, rc = PTR_ERR(pending_child));
1508
1509         if (pending_child->d_inode != NULL) {
1510                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1511                 LASSERT(pending_child->d_inode == dentry->d_inode);
1512                 GOTO(out_dput, rc = 0);
1513         }
1514
1515         lock_kernel();
1516         rc = vfs_link(dentry, pending_dir, pending_child);
1517         unlock_kernel();
1518         if (rc)
1519                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
1520                        rec->ur_name, rc);
1521         else
1522                 mds_inode_set_orphan(dentry->d_inode);
1523 out_dput:
1524         l_dput(pending_child);
1525 out_lock:
1526         up(&pending_dir->i_sem);
1527         RETURN(rc);
1528 }
1529
1530 /* The idea here is that we need to get four locks in the end:
1531  * one on each parent directory, one on each child.  We need to take
1532  * these locks in some kind of order (to avoid deadlocks), and the order
1533  * I selected is "increasing resource number" order.  We need to look up
1534  * the children, however, before we know what the resource number(s) are.
1535  * Thus the following plan:
1536  *
1537  * 1,2. Look up the parents
1538  * 3,4. Look up the children
1539  * 5. Take locks on the parents and children, in order
1540  * 6. Verify that the children haven't changed since they were looked up
1541  *
1542  * If there was a race and the children changed since they were first looked
1543  * up, it is possible that mds_verify_child() will be able to just grab the
1544  * lock on the new child resource (if it has a higher resource than any other)
1545  * but we need to compare against not only its parent, but also against the
1546  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1547  *
1548  * We need the fancy igrab() on the child inodes because we aren't holding a
1549  * lock on the parent after the lookup is done, so dentry->d_inode may change
1550  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1551  */
1552 static int mds_get_parents_children_locked(struct obd_device *obd,
1553                                            struct mds_obd *mds,
1554                                            struct ll_fid *p1_fid,
1555                                            struct dentry **de_srcdirp,
1556                                            struct ll_fid *p2_fid,
1557                                            struct dentry **de_tgtdirp,
1558                                            int parent_mode,
1559                                            const char *old_name, int old_len,
1560                                            struct dentry **de_oldp,
1561                                            const char *new_name, int new_len,
1562                                            struct dentry **de_newp,
1563                                            struct lustre_handle *dlm_handles,
1564                                            int child_mode)
1565 {
1566         struct ldlm_res_id p1_res_id = { .name = {0} };
1567         struct ldlm_res_id p2_res_id = { .name = {0} };
1568         struct ldlm_res_id c1_res_id = { .name = {0} };
1569         struct ldlm_res_id c2_res_id = { .name = {0} };
1570         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1571         /* Only dentry should change, but the inode itself would be
1572            intact otherwise */
1573         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
1574         /* If something is going to be replaced, both dentry and inode locks are
1575            needed */
1576         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
1577                                                         MDS_INODELOCK_UPDATE}};
1578         struct ldlm_res_id *maxres_src, *maxres_tgt;
1579         struct inode *inode;
1580         int rc = 0, cleanup_phase = 0;
1581         ENTRY;
1582
1583         /* Step 1: Lookup the source directory */
1584         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1585         if (IS_ERR(*de_srcdirp))
1586                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1587
1588         cleanup_phase = 1; /* source directory dentry */
1589
1590         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1591         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1592
1593         /* Step 2: Lookup the target directory */
1594         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1595                 *de_tgtdirp = dget(*de_srcdirp);
1596         } else {
1597                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1598                 if (IS_ERR(*de_tgtdirp)) {
1599                         rc = PTR_ERR(*de_tgtdirp);
1600                         *de_tgtdirp = NULL;
1601                         GOTO(cleanup, rc);
1602                 }
1603         }
1604
1605         cleanup_phase = 2; /* target directory dentry */
1606
1607         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1608         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1609
1610         /* Step 3: Lookup the source child entry */
1611         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1612         if (IS_ERR(*de_oldp)) {
1613                 rc = PTR_ERR(*de_oldp);
1614                 CERROR("old child lookup error (%*s): %d\n",
1615                        old_len - 1, old_name, rc);
1616                 GOTO(cleanup, rc);
1617         }
1618
1619         cleanup_phase = 3; /* original name dentry */
1620
1621         inode = (*de_oldp)->d_inode;
1622         if (inode != NULL)
1623                 inode = igrab(inode);
1624         if (inode == NULL)
1625                 GOTO(cleanup, rc = -ENOENT);
1626
1627         c1_res_id.name[0] = inode->i_ino;
1628         c1_res_id.name[1] = inode->i_generation;
1629         iput(inode);
1630
1631         /* Step 4: Lookup the target child entry */
1632         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1633         if (IS_ERR(*de_newp)) {
1634                 rc = PTR_ERR(*de_newp);
1635                 CERROR("new child lookup error (%*s): %d\n",
1636                        old_len - 1, old_name, rc);
1637                 GOTO(cleanup, rc);
1638         }
1639
1640         cleanup_phase = 4; /* target dentry */
1641
1642         inode = (*de_newp)->d_inode;
1643         if (inode != NULL)
1644                 inode = igrab(inode);
1645         if (inode == NULL)
1646                 goto retry_locks;
1647
1648         c2_res_id.name[0] = inode->i_ino;
1649         c2_res_id.name[1] = inode->i_generation;
1650
1651         iput(inode);
1652
1653 retry_locks:
1654         /* Step 5: Take locks on the parents and child(ren) */
1655         maxres_src = &p1_res_id;
1656         maxres_tgt = &p2_res_id;
1657         cleanup_phase = 4; /* target dentry */
1658
1659         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
1660                 maxres_src = &c1_res_id;
1661         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
1662                 maxres_tgt = &c2_res_id;
1663
1664         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1665                                     &p_policy,
1666                                     &p2_res_id, &dlm_handles[1], parent_mode,
1667                                     &p_policy,
1668                                     &c1_res_id, &dlm_handles[2], child_mode,
1669                                     &c1_policy,
1670                                     &c2_res_id, &dlm_handles[3], child_mode,
1671                                     &c2_policy);
1672         if (rc)
1673                 GOTO(cleanup, rc);
1674
1675         cleanup_phase = 6; /* parent and child(ren) locks */
1676
1677         /* Step 6a: Re-lookup source child to verify it hasn't changed */
1678         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1679                               parent_mode, &c1_res_id, &dlm_handles[2],
1680                               de_oldp, child_mode, &c1_policy, old_name,old_len,
1681                               maxres_tgt);
1682         if (rc) {
1683                 if (c2_res_id.name[0] != 0)
1684                         ldlm_lock_decref(&dlm_handles[3], child_mode);
1685                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1686                 cleanup_phase = 4;
1687                 if (rc > 0)
1688                         goto retry_locks;
1689                 GOTO(cleanup, rc);
1690         }
1691
1692         if ((*de_oldp)->d_inode == NULL)
1693                 GOTO(cleanup, rc = -ENOENT);
1694
1695         /* Step 6b: Re-lookup target child to verify it hasn't changed */
1696         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
1697                               parent_mode, &c2_res_id, &dlm_handles[3],
1698                               de_newp, child_mode, &c2_policy, new_name,
1699                               new_len, maxres_src);
1700         if (rc) {
1701                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1702                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1703                 cleanup_phase = 4;
1704                 if (rc > 0)
1705                         goto retry_locks;
1706                 GOTO(cleanup, rc);
1707         }
1708
1709         EXIT;
1710 cleanup:
1711         if (rc) {
1712                 switch (cleanup_phase) {
1713                 case 6: /* child lock(s) */
1714                         if (c2_res_id.name[0] != 0)
1715                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
1716                         if (c1_res_id.name[0] != 0)
1717                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1718                 case 5: /* parent locks */
1719                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
1720                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
1721                 case 4: /* target dentry */
1722                         l_dput(*de_newp);
1723                 case 3: /* source dentry */
1724                         l_dput(*de_oldp);
1725                 case 2: /* target directory dentry */
1726                         l_dput(*de_tgtdirp);
1727                 case 1: /* source directry dentry */
1728                         l_dput(*de_srcdirp);
1729                 }
1730         }
1731
1732         return rc;
1733 }
1734
1735 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1736                             struct ptlrpc_request *req,
1737                             struct lustre_handle *lockh)
1738 {
1739         struct obd_device *obd = req->rq_export->exp_obd;
1740         struct dentry *de_srcdir = NULL;
1741         struct dentry *de_tgtdir = NULL;
1742         struct dentry *de_old = NULL;
1743         struct dentry *de_new = NULL;
1744         struct mds_obd *mds = mds_req2mds(req);
1745         struct lustre_handle dlm_handles[4];
1746         struct mds_body *body = NULL;
1747         int rc = 0, lock_count = 3;
1748         int cleanup_phase = 0;
1749         void *handle = NULL;
1750         ENTRY;
1751
1752         LASSERT(offset == 0);
1753
1754         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
1755                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
1756                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
1757
1758         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1759
1760         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
1761                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
1762                                              rec->ur_name, rec->ur_namelen,
1763                                              &de_old, rec->ur_tgt,
1764                                              rec->ur_tgtlen, &de_new,
1765                                              dlm_handles, LCK_EX);
1766         if (rc)
1767                 GOTO(cleanup, rc);
1768
1769         cleanup_phase = 1; /* parent(s), children, locks */
1770
1771         if (de_new->d_inode)
1772                 lock_count = 4;
1773
1774         /* sanity check for src inode */
1775         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1776             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
1777                 GOTO(cleanup, rc = -EINVAL);
1778
1779         /* sanity check for dest inode */
1780         if (de_new->d_inode &&
1781             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1782              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
1783                 GOTO(cleanup, rc = -EINVAL);
1784
1785         if (de_old->d_inode == de_new->d_inode) {
1786                 GOTO(cleanup, rc = 0);
1787         }
1788
1789         /* if we are about to remove the target at first, pass the EA of
1790          * that inode to client to perform and cleanup on OST */
1791         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1792         LASSERT(body != NULL);
1793
1794         if (de_new->d_inode &&
1795             S_ISREG(de_new->d_inode->i_mode) &&
1796             de_new->d_inode->i_nlink == 1 &&
1797             mds_open_orphan_count(de_new->d_inode) == 0) {
1798                 mds_pack_inode2fid(&body->fid1, de_new->d_inode);
1799                 mds_pack_inode2body(body, de_new->d_inode);
1800                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
1801                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1802                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1803                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1804                 } else {
1805                         /* XXX need log unlink? */
1806                 }
1807         }
1808
1809         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
1810                        de_srcdir->d_inode->i_sb);
1811
1812         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
1813         if (IS_ERR(handle))
1814                 GOTO(cleanup, rc = PTR_ERR(handle));
1815
1816         /* FIXME need adjust the journal block count? */
1817         /* if the target should be moved to PENDING, we at first increase the
1818          * link and later vfs_rename() will decrease the link count again */
1819         if (de_new->d_inode &&
1820             S_ISREG(de_new->d_inode->i_mode) &&
1821             de_new->d_inode->i_nlink == 1 &&
1822             mds_open_orphan_count(de_new->d_inode) > 0) {
1823                 rc = mds_add_link_orphan(rec, obd, de_new);
1824                 if (rc)
1825                         GOTO(cleanup, rc);
1826         }
1827
1828         lock_kernel();
1829         de_old->d_fsdata = req;
1830         de_new->d_fsdata = req;
1831         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
1832         unlock_kernel();
1833
1834         GOTO(cleanup, rc);
1835 cleanup:
1836         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
1837                                 handle, req, rc, 0);
1838         switch (cleanup_phase) {
1839         case 1:
1840                 if (rc) {
1841                         if (lock_count == 4)
1842                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
1843                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
1844                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
1845                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
1846                 } else {
1847                         if (lock_count == 4)
1848                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
1849                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
1850                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
1851                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
1852                 }
1853                 l_dput(de_new);
1854                 l_dput(de_old);
1855                 l_dput(de_tgtdir);
1856                 l_dput(de_srcdir);
1857         case 0:
1858                 break;
1859         default:
1860                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1861                 LBUG();
1862         }
1863         req->rq_status = rc;
1864         return 0;
1865 }
1866
1867 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
1868                            struct ptlrpc_request *, struct lustre_handle *);
1869
1870 static mds_reinter reinters[REINT_MAX + 1] = {
1871         [REINT_SETATTR] mds_reint_setattr,
1872         [REINT_CREATE] mds_reint_create,
1873         [REINT_LINK] mds_reint_link,
1874         [REINT_UNLINK] mds_reint_unlink,
1875         [REINT_RENAME] mds_reint_rename,
1876         [REINT_OPEN] mds_open
1877 };
1878
1879 int mds_reint_rec(struct mds_update_record *rec, int offset,
1880                   struct ptlrpc_request *req, struct lustre_handle *lockh)
1881 {
1882         struct obd_device *obd = req->rq_export->exp_obd;
1883         struct obd_run_ctxt saved;
1884         int rc;
1885         ENTRY;
1886
1887         /* checked by unpacker */
1888         LASSERT(rec->ur_opcode <= REINT_MAX &&
1889                 reinters[rec->ur_opcode] != NULL);
1890
1891         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1892         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
1893         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1894
1895         RETURN(rc);
1896 }