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->obd_llogs,
  82                                         mlcd->mlcd_cookies[0].lgc_subsys + 1);
  83                 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
  84                                                 sizeof(*mlcd->mlcd_cookies),
  85                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  86                 if (rc)
  87                         CERROR("error cancelling %d log cookies: rc %d\n",
  88                                (int)(mlcd->mlcd_cookielen /
  89                                      sizeof(*mlcd->mlcd_cookies)), rc);
  90         }
  91
  92         OBD_FREE(mlcd, mlcd->mlcd_size);
  93 }
  94
  95 /* Assumes caller has already pushed us into the kernel context. */
  96 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  97                        struct ptlrpc_request *req, int rc, __u32 op_data)
  98 {
  99         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 100         struct mds_client_data *mcd = med->med_mcd;
 101         struct obd_device *obd = req->rq_export->exp_obd;
 102         int err;
 103         __u64 transno;
 104         loff_t off;
 105         int log_pri = D_HA;
 106         ENTRY;
 107
 108         /* if the export has already been failed, we have no last_rcvd slot */
 109         if (req->rq_export->exp_failed) {
 110                 CERROR("committing transaction for disconnected client\n");
 111                 if (handle)
 112                         GOTO(commit, rc);
 113                 RETURN(rc);
 114         }
 115
 116         if (IS_ERR(handle))
 117                 RETURN(rc);
 118
 119         if (handle == NULL) {
 120                 /* if we're starting our own xaction, use our own inode */
 121                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 122                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 123                 if (IS_ERR(handle)) {
 124                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 125                         RETURN(PTR_ERR(handle));
 126                 }
 127         }
 128
 129         off = med->med_off;
 130
 131         transno = req->rq_reqmsg->transno;
 132         if (rc != 0) {
 133                 LASSERT(transno == 0);
 134         } else if (transno == 0) {
 135                 spin_lock(&mds->mds_transno_lock);
 136                 transno = ++mds->mds_last_transno;
 137                 spin_unlock(&mds->mds_transno_lock);
 138         } else {
 139                 spin_lock(&mds->mds_transno_lock);
 140                 if (transno > mds->mds_last_transno)
 141                         mds->mds_last_transno = transno;
 142                 spin_unlock(&mds->mds_transno_lock);
 143         }
 144         req->rq_repmsg->transno = req->rq_transno = transno;
 145         mcd->mcd_last_transno = cpu_to_le64(transno);
 146         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 147         mcd->mcd_last_result = cpu_to_le32(rc);
 148         mcd->mcd_last_data = cpu_to_le32(op_data);
 149
 150         fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 151                               mds_commit_cb, NULL);
 152         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 153                                   &off, 0);
 154
 155         if (err) {
 156                 log_pri = D_ERROR;
 157                 if (rc == 0)
 158                         rc = err;
 159         }
 160
 161         DEBUG_REQ(log_pri, req,
 162                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 163                   transno, mcd->mcd_uuid, med->med_idx, err);
 164
 165         err = mds_lov_write_objids(obd);
 166         if (err) {
 167                 log_pri = D_ERROR;
 168                 if (rc == 0)
 169                         rc = err;
 170         }
 171         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 172
 173 commit:
 174         err = fsfilt_commit(obd, inode, handle, 0);
 175         if (err) {
 176                 CERROR("error committing transaction: %d\n", err);
 177                 if (!rc)
 178                         rc = err;
 179         }
 180
 181         RETURN(rc);
 182 }
 183
 184 /* this gives the same functionality as the code between
 185  * sys_chmod and inode_setattr
 186  * chown_common and inode_setattr
 187  * utimes and inode_setattr
 188  */
 189 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 190 {
 191         time_t now = LTIME_S(CURRENT_TIME);
 192         struct iattr *attr = &rec->ur_iattr;
 193         unsigned int ia_valid = attr->ia_valid;
 194         int error;
 195         ENTRY;
 196
 197         /* only fix up attrs if the client VFS didn't already */
 198         if (!(ia_valid & ATTR_RAW))
 199                 RETURN(0);
 200
 201         if (!(ia_valid & ATTR_CTIME_SET))
 202                 LTIME_S(attr->ia_ctime) = now;
 203         if (!(ia_valid & ATTR_ATIME_SET))
 204                 LTIME_S(attr->ia_atime) = now;
 205         if (!(ia_valid & ATTR_MTIME_SET))
 206                 LTIME_S(attr->ia_mtime) = now;
 207
 208         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 209                 RETURN(-EPERM);
 210
 211         /* times */
 212         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 213                 if (rec->ur_fsuid != inode->i_uid &&
 214                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 215                         RETURN(error);
 216         }
 217
 218         if (ia_valid & ATTR_SIZE) {
 219                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 220                         RETURN(error);
 221         }
 222
 223         if (ia_valid & ATTR_UID) {
 224                 /* chown */
 225                 error = -EPERM;
 226                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 227                         RETURN(-EPERM);
 228                 if (attr->ia_uid == (uid_t) -1)
 229                         attr->ia_uid = inode->i_uid;
 230                 if (attr->ia_gid == (gid_t) -1)
 231                         attr->ia_gid = inode->i_gid;
 232                 attr->ia_mode = inode->i_mode;
 233                 /*
 234                  * If the user or group of a non-directory has been
 235                  * changed by a non-root user, remove the setuid bit.
 236                  * 19981026 David C Niemi <niemi@tux.org>
 237                  *
 238                  * Changed this to apply to all users, including root,
 239                  * to avoid some races. This is the behavior we had in
 240                  * 2.0. The check for non-root was definitely wrong
 241                  * for 2.2 anyway, as it should have been using
 242                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 243                  */
 244                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 245                     !S_ISDIR(inode->i_mode)) {
 246                         attr->ia_mode &= ~S_ISUID;
 247                         attr->ia_valid |= ATTR_MODE;
 248                 }
 249                 /*
 250                  * Likewise, if the user or group of a non-directory
 251                  * has been changed by a non-root user, remove the
 252                  * setgid bit UNLESS there is no group execute bit
 253                  * (this would be a file marked for mandatory
 254                  * locking).  19981026 David C Niemi <niemi@tux.org>
 255                  *
 256                  * Removed the fsuid check (see the comment above) --
 257                  * 19990830 SD.
 258                  */
 259                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 260                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 261                         attr->ia_mode &= ~S_ISGID;
 262                         attr->ia_valid |= ATTR_MODE;
 263                 }
 264         } else if (ia_valid & ATTR_MODE) {
 265                 int mode = attr->ia_mode;
 266                 /* chmod */
 267                 if (attr->ia_mode == (mode_t) -1)
 268                         attr->ia_mode = inode->i_mode;
 269                 attr->ia_mode =
 270                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 271         }
 272         RETURN(0);
 273 }
 274
 275 void mds_steal_ack_locks(struct ptlrpc_request *req)
 276 {
 277         struct obd_export         *exp = req->rq_export;
 278         struct list_head          *tmp;
 279         struct ptlrpc_reply_state *oldrep;
 280         struct ptlrpc_service     *svc;
 281         unsigned long              flags;
 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"LPX64"\n",
 304                       oldrep->rs_nlocks, oldrep,
 305                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 306                       exp->exp_connection->c_peer.peer_nid);
 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(req2obd(req), &body->fid1, de->d_inode);
 352         mds_pack_inode2body(req2obd(req), 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[2];
 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) && rec->ur_eadata != NULL)
 412                 down(&inode->i_sem);
 413
 414         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 415
 416         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 417         if (IS_ERR(handle))
 418                 GOTO(cleanup, rc = PTR_ERR(handle));
 419
 420         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 421                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 422                        LTIME_S(rec->ur_iattr.ia_mtime),
 423                        LTIME_S(rec->ur_iattr.ia_ctime));
 424         rc = mds_fix_attr(inode, rec);
 425         if (rc)
 426                 GOTO(cleanup, rc);
 427
 428         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 429                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 430                                       (long)&rec->ur_iattr.ia_attr_flags);
 431         else                                            /* setattr */
 432                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 433
 434         if (rc == 0 && S_ISREG(inode->i_mode) && rec->ur_eadata != NULL) {
 435                 rc = fsfilt_set_md(obd, inode, handle,
 436                                    rec->ur_eadata, rec->ur_eadatalen);
 437         }
 438
 439         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 440         mds_pack_inode2fid(obd, &body->fid1, inode);
 441         mds_pack_inode2body(obd, body, inode);
 442
 443         /* Don't return OST-specific attributes if we didn't just set them */
 444         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 445                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 446         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 447                 body->valid |= OBD_MD_FLMTIME;
 448         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 449                 body->valid |= OBD_MD_FLATIME;
 450
 451         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 452                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 453                           rec->ur_eadatalen);
 454                 if (mlcd) {
 455                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 456                                 rec->ur_eadatalen;
 457                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 458                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 459                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 460                                 mlcd->mlcd_cookielen;
 461                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 462                                mlcd->mlcd_cookielen);
 463                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 464                                mlcd->mlcd_eadatalen);
 465                 } else {
 466                         CERROR("unable to allocate log cancel data\n");
 467                 }
 468         }
 469         EXIT;
 470  cleanup:
 471         if (mlcd != NULL)
 472                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 473                                       mds_cancel_cookies_cb, mlcd);
 474         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 475         switch (cleanup_phase) {
 476         case 1:
 477                 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
 478                         up(&inode->i_sem);
 479                 l_dput(de);
 480                 if (locked) {
 481 #ifdef S_PDIROPS
 482                         if (lockh[1].cookie != 0)
 483                                 ldlm_lock_decref(lockh + 1, LCK_CW);
 484 #endif
 485                         if (rc) {
 486                                 ldlm_lock_decref(lockh, LCK_PW);
 487                         } else {
 488                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 489                         }
 490                 }
 491         case 0:
 492                 break;
 493         default:
 494                 LBUG();
 495         }
 496         if (err && !rc)
 497                 rc = err;
 498
 499         req->rq_status = rc;
 500         return 0;
 501 }
 502
 503 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 504                                      struct ptlrpc_request *req)
 505 {
 506         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 507         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 508         struct dentry *parent, *child;
 509         struct mds_body *body;
 510
 511         mds_req_from_mcd(req, med->med_mcd);
 512
 513         if (req->rq_status)
 514                 return;
 515
 516         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 517         LASSERT(!IS_ERR(parent));
 518         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 519         LASSERT(!IS_ERR(child));
 520         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 521         mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 522         mds_pack_inode2body(req2obd(req), body, child->d_inode);
 523         l_dput(parent);
 524         l_dput(child);
 525 }
 526
 527 static int mds_reint_create(struct mds_update_record *rec, int offset,
 528                             struct ptlrpc_request *req,
 529                             struct lustre_handle *lh)
 530 {
 531         struct dentry *dparent = NULL;
 532         struct mds_obd *mds = mds_req2mds(req);
 533         struct obd_device *obd = req->rq_export->exp_obd;
 534         struct dentry *dchild = NULL;
 535         struct inode *dir = NULL;
 536         void *handle = NULL;
 537         struct lustre_handle lockh[2];
 538         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 539         int created = 0;
 540         struct dentry_params dp;
 541         struct mea *mea = NULL;
 542         int mea_size;
 543         ENTRY;
 544
 545         LASSERT(offset == 0);
 546         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 547
 548         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 549                   rec->ur_fid1->id, rec->ur_fid1->generation,
 550                   rec->ur_name, rec->ur_mode);
 551
 552         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 553
 554         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 555                 GOTO(cleanup, rc = -ESTALE);
 556
 557         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 558                                         rec->ur_name, rec->ur_namelen - 1,
 559                                         MDS_INODELOCK_UPDATE);
 560         if (IS_ERR(dparent)) {
 561                 rc = PTR_ERR(dparent);
 562                 CERROR("parent lookup error %d\n", rc);
 563                 GOTO(cleanup, rc);
 564         }
 565         cleanup_phase = 1; /* locked parent dentry */
 566         dir = dparent->d_inode;
 567         LASSERT(dir);
 568
 569         ldlm_lock_dump_handle(D_OTHER, lockh);
 570
 571         /* try to retrieve MEA data for this dir */
 572         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 573         if (mea != NULL) {
 574                 /* dir is already splitted, check is requested filename
 575                  * should live at this MDS or at another one */
 576                 int i;
 577                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 578                 if (mea->mea_master != i) {
 579                         CERROR("inapropriate MDS(%d) for %s. should be %d\n",
 580                                 mea->mea_master, rec->ur_name, i);
 581                         GOTO(cleanup, rc = -ESTALE);
 582                 }
 583         }
 584
 585         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 586         if (IS_ERR(dchild)) {
 587                 rc = PTR_ERR(dchild);
 588                 CERROR("child lookup error %d\n", rc);
 589                 GOTO(cleanup, rc);
 590         }
 591
 592         cleanup_phase = 2; /* child dentry */
 593
 594         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 595
 596         if (type == S_IFREG || type == S_IFDIR) {
 597                 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
 598                         if (rc > 0) {
 599                                 /* dir got splitted */
 600                                 GOTO(cleanup, rc = -ESTALE);
 601                         } else {
 602                                 /* error happened during spitting */
 603                                 GOTO(cleanup, rc);
 604                         }
 605                 }
 606         }
 607
 608         if (dir->i_mode & S_ISGID) {
 609                 if (S_ISDIR(rec->ur_mode))
 610                         rec->ur_mode |= S_ISGID;
 611         }
 612
 613         dchild->d_fsdata = (void *)&dp;
 614         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 615         dp.p_ptr = req;
 616
 617         switch (type) {
 618         case S_IFREG:{
 619                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 620                 if (IS_ERR(handle))
 621                         GOTO(cleanup, rc = PTR_ERR(handle));
 622                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 623                 EXIT;
 624                 break;
 625         }
 626         case S_IFDIR:{
 627                 int nstripes = 0;
 628                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 629                 if (IS_ERR(handle))
 630                         GOTO(cleanup, rc = PTR_ERR(handle));
 631
 632                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 633
 634                 if (rec->ur_eadata)
 635                         nstripes = *(u16 *)rec->ur_eadata;
 636
 637 #if 1
 638                 /* this is for current testing yet. after the testing
 639                  * directory will split if size reaches some limite -bzzz */
 640                 if (rc == 0) {
 641 #else
 642                 if (rc == 0 && nstripes) {
 643 #endif
 644                         /* FIXME: error handling here */
 645                         mds_try_to_split_dir(obd, dchild, NULL, nstripes);
 646                 }
 647                 EXIT;
 648                 break;
 649         }
 650         case S_IFLNK:{
 651                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 652                 if (IS_ERR(handle))
 653                         GOTO(cleanup, rc = PTR_ERR(handle));
 654                 if (rec->ur_tgt == NULL)        /* no target supplied */
 655                         rc = -EINVAL;           /* -EPROTO? */
 656                 else
 657                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 658                 EXIT;
 659                 break;
 660         }
 661         case S_IFCHR:
 662         case S_IFBLK:
 663         case S_IFIFO:
 664         case S_IFSOCK:{
 665                 int rdev = rec->ur_rdev;
 666                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 667                 if (IS_ERR(handle))
 668                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 669                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 670                 EXIT;
 671                 break;
 672         }
 673         default:
 674                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 675                 dchild->d_fsdata = NULL;
 676                 GOTO(cleanup, rc = -EINVAL);
 677         }
 678
 679         /* In case we stored the desired inum in here, we want to clean up. */
 680         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 681                 dchild->d_fsdata = NULL;
 682
 683         if (rc) {
 684                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 685                 GOTO(cleanup, rc);
 686         } else {
 687                 struct iattr iattr;
 688                 struct inode *inode = dchild->d_inode;
 689                 struct mds_body *body;
 690
 691                 created = 1;
 692                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 693                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 694                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 695                 iattr.ia_uid = rec->ur_fsuid;
 696                 if (dir->i_mode & S_ISGID)
 697                         iattr.ia_gid = dir->i_gid;
 698                 else
 699                         iattr.ia_gid = rec->ur_fsgid;
 700                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 701                         ATTR_MTIME | ATTR_CTIME;
 702
 703                 if (rec->ur_fid2->id) {
 704                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 705                         inode->i_generation = rec->ur_fid2->generation;
 706                         /* Dirtied and committed by the upcoming setattr. */
 707                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 708                                inode->i_ino, inode->i_generation);
 709                 } else {
 710                         struct lustre_handle child_ino_lockh;
 711
 712                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 713                                inode->i_ino, inode->i_generation);
 714
 715                         /* The inode we were allocated may have just been freed
 716                          * by an unlink operation.  We take this lock to
 717                          * synchronize against the matching reply-ack-lock taken
 718                          * in unlink, to avoid replay problems if this reply
 719                          * makes it out to the client but the unlink's does not.
 720                          * See bug 2029 for more detail.*/
 721                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 722                         if (rc != ELDLM_OK) {
 723                                 CERROR("error locking for unlink/create sync: "
 724                                        "%d\n", rc);
 725                         } else {
 726                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 727                         }
 728                 }
 729
 730                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 731                 if (rc)
 732                         CERROR("error on child setattr: rc = %d\n", rc);
 733
 734                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 735                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 736                 if (rc)
 737                         CERROR("error on parent setattr: rc = %d\n", rc);
 738
 739                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 740                 mds_pack_inode2fid(obd, &body->fid1, inode);
 741                 mds_pack_inode2body(obd, body, inode);
 742         }
 743         EXIT;
 744
 745 cleanup:
 746         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 747
 748         if (rc && created) {
 749                 /* Destroy the file we just created.  This should not need
 750                  * extra journal credits, as we have already modified all of
 751                  * the blocks needed in order to create the file in the first
 752                  * place.
 753                  */
 754                 switch (type) {
 755                 case S_IFDIR:
 756                         err = vfs_rmdir(dir, dchild);
 757                         if (err)
 758                                 CERROR("rmdir in error path: %d\n", err);
 759                         break;
 760                 default:
 761                         err = vfs_unlink(dir, dchild);
 762                         if (err)
 763                                 CERROR("unlink in error path: %d\n", err);
 764                         break;
 765                 }
 766         } else {
 767                 rc = err;
 768         }
 769         switch (cleanup_phase) {
 770         case 2: /* child dentry */
 771                 l_dput(dchild);
 772         case 1: /* locked parent dentry */
 773 #ifdef S_PDIROPS
 774                 if (lockh[1].cookie != 0)
 775                         ldlm_lock_decref(lockh + 1, LCK_CW);
 776 #endif
 777                 if (rc) {
 778                         ldlm_lock_decref(lockh, LCK_PW);
 779                 } else {
 780                         ptlrpc_save_lock (req, lockh, LCK_PW);
 781                 }
 782                 l_dput(dparent);
 783         case 0:
 784                 break;
 785         default:
 786                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 787                 LBUG();
 788         }
 789         if (mea)
 790                 OBD_FREE(mea, mea_size);
 791         req->rq_status = rc;
 792         return 0;
 793 }
 794
 795 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 796            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 797 {
 798         int i;
 799
 800         for (i = 0; i < RES_NAME_SIZE; i++) {
 801                 /* return 1 here, because enqueue_ordered will skip resources
 802                  * of all zeroes if they're sorted to the end of the list. */
 803                 if (res1->name[i] == 0 && res2->name[i] != 0)
 804                         return 1;
 805                 if (res2->name[i] == 0 && res1->name[i] != 0)
 806                         return 0;
 807
 808                 if (res1->name[i] > res2->name[i])
 809                         return 1;
 810                 if (res1->name[i] < res2->name[i])
 811                         return 0;
 812         }
 813
 814         if (!p1 || !p2)
 815                 return 0;
 816
 817         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 818                 return 1;
 819
 820         return 0;
 821 }
 822
 823 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 824  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 825  * because they take the place of local semaphores.
 826  *
 827  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 828  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 829 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 830                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 831                           ldlm_policy_data_t *p1_policy,
 832                           struct ldlm_res_id *p2_res_id,
 833                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 834                           ldlm_policy_data_t *p2_policy)
 835 {
 836         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 837         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 838         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 839         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 840         int rc, flags;
 841         ENTRY;
 842
 843         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 844
 845         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 846                res_id[1]->name[0]);
 847
 848         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 849                 handles[1] = p1_lockh;
 850                 handles[0] = p2_lockh;
 851                 res_id[1] = p1_res_id;
 852                 res_id[0] = p2_res_id;
 853                 lock_modes[1] = p1_lock_mode;
 854                 lock_modes[0] = p2_lock_mode;
 855                 policies[1] = p1_policy;
 856                 policies[0] = p2_policy;
 857         }
 858
 859         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 860                res_id[0]->name[0], res_id[1]->name[0]);
 861
 862         flags = LDLM_FL_LOCAL_ONLY;
 863         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 864                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 865                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 866                               NULL, 0, NULL, handles[0]);
 867         if (rc != ELDLM_OK)
 868                 RETURN(-EIO);
 869         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 870
 871         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
 872             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
 873                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 874                 ldlm_lock_addref(handles[1], lock_modes[1]);
 875         } else if (res_id[1]->name[0] != 0) {
 876                 flags = LDLM_FL_LOCAL_ONLY;
 877                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 878                                       *res_id[1], LDLM_IBITS, policies[1],
 879                                       lock_modes[1], &flags, mds_blocking_ast,
 880                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 881                                       NULL, handles[1]);
 882                 if (rc != ELDLM_OK) {
 883                         ldlm_lock_decref(handles[0], lock_modes[0]);
 884                         RETURN(-EIO);
 885                 }
 886                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 887         }
 888
 889         RETURN(0);
 890 }
 891
 892 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
 893                            struct lustre_handle *p1_lockh, int p1_lock_mode,
 894                            ldlm_policy_data_t *p1_policy,
 895                            struct ldlm_res_id *p2_res_id,
 896                            struct lustre_handle *p2_lockh, int p2_lock_mode,
 897                            ldlm_policy_data_t *p2_policy,
 898                            struct ldlm_res_id *c1_res_id,
 899                            struct lustre_handle *c1_lockh, int c1_lock_mode,
 900                            ldlm_policy_data_t *c1_policy,
 901                            struct ldlm_res_id *c2_res_id,
 902                            struct lustre_handle *c2_lockh, int c2_lock_mode,
 903                            ldlm_policy_data_t *c2_policy)
 904 {
 905         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
 906                                           c1_res_id, c2_res_id };
 907         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
 908                                                  c1_lockh, c2_lockh };
 909         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
 910                               c1_lock_mode, c2_lock_mode };
 911         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
 912                                             c1_policy, c2_policy};
 913         int rc, i, j, sorted, flags;
 914         ENTRY;
 915
 916         CDEBUG(D_DLMTRACE,
 917                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 918                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 919                res_id[3]->name[0]);
 920
 921         /* simple insertion sort - we have at most 4 elements */
 922         for (i = 1; i < 4; i++) {
 923                 j = i - 1;
 924                 dlm_handles[4] = dlm_handles[i];
 925                 res_id[4] = res_id[i];
 926                 lock_modes[4] = lock_modes[i];
 927                 policies[4] = policies[i];
 928
 929                 sorted = 0;
 930                 do {
 931                         if (res_gt(res_id[j], res_id[4], policies[j],
 932                                    policies[4])) {
 933                                 dlm_handles[j + 1] = dlm_handles[j];
 934                                 res_id[j + 1] = res_id[j];
 935                                 lock_modes[j + 1] = lock_modes[j];
 936                                 policies[j + 1] = policies[j];
 937                                 j--;
 938                         } else {
 939                                 sorted = 1;
 940                         }
 941                 } while (j >= 0 && !sorted);
 942
 943                 dlm_handles[j + 1] = dlm_handles[4];
 944                 res_id[j + 1] = res_id[4];
 945                 lock_modes[j + 1] = lock_modes[4];
 946                 policies[j + 1] = policies[4];
 947         }
 948
 949         CDEBUG(D_DLMTRACE,
 950                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 951                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 952                res_id[3]->name[0]);
 953
 954         /* XXX we could send ASTs on all these locks first before blocking? */
 955         for (i = 0; i < 4; i++) {
 956                 flags = 0;
 957                 if (res_id[i]->name[0] == 0)
 958                         break;
 959                 if (i != 0 &&
 960                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
 961                     (policies[i]->l_inodebits.bits &
 962                      policies[i-1]->l_inodebits.bits) ) {
 963                         memcpy(dlm_handles[i], dlm_handles[i-1],
 964                                sizeof(*(dlm_handles[i])));
 965                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
 966                 } else {
 967                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 968                                               *res_id[i], LDLM_IBITS,
 969                                               policies[i],
 970                                               lock_modes[i], &flags,
 971                                               mds_blocking_ast,
 972                                               ldlm_completion_ast, NULL, NULL,
 973                                               NULL, 0, NULL, dlm_handles[i]);
 974                         if (rc != ELDLM_OK)
 975                                 GOTO(out_err, rc = -EIO);
 976                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
 977                 }
 978         }
 979
 980         RETURN(0);
 981 out_err:
 982         while (i-- > 0)
 983                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
 984
 985         return rc;
 986 }
 987
 988 /* In the unlikely case that the child changed while we were waiting
 989  * on the lock, we need to drop the lock on the old child and either:
 990  * - if the child has a lower resource name, then we have to also
 991  *   drop the parent lock and regain the locks in the right order
 992  * - in the rename case, if the child has a lower resource name than one of
 993  *   the other parent/child resources (maxres) we also need to reget the locks
 994  * - if the child has a higher resource name (this is the common case)
 995  *   we can just get the lock on the new child (still in lock order)
 996  *
 997  * Returns 0 if the child did not change or if it changed but could be locked.
 998  * Returns 1 if the child changed and we need to re-lock (no locks held).
 999  * Returns -ve error with a valid dchild (no locks held). */
1000 static int mds_verify_child(struct obd_device *obd,
1001                             struct ldlm_res_id *parent_res_id,
1002                             struct lustre_handle *parent_lockh,
1003                             struct dentry *dparent, int parent_mode,
1004                             struct ldlm_res_id *child_res_id,
1005                             struct lustre_handle *child_lockh,
1006                             struct dentry **dchildp, int child_mode,
1007                             ldlm_policy_data_t *child_policy,
1008                             const char *name, int namelen,
1009                             struct ldlm_res_id *maxres)
1010 {
1011         struct dentry *vchild, *dchild = *dchildp;
1012         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1013         ENTRY;
1014
1015         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1016         if (IS_ERR(vchild))
1017                 GOTO(cleanup, rc = PTR_ERR(vchild));
1018
1019         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1020                 if  (child_res_id->name[0] == vchild->d_inum &&
1021                                 child_res_id->name[1] == vchild->d_generation) {
1022                         if (dchild != NULL)
1023                                 l_dput(dchild);
1024                         *dchildp = vchild;
1025                         RETURN(0);
1026                 }
1027                 goto changed;
1028         }
1029
1030         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1031                    (vchild->d_inode != NULL &&
1032                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1033                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1034                 if (dchild != NULL)
1035                         l_dput(dchild);
1036                 *dchildp = vchild;
1037
1038                 RETURN(0);
1039         }
1040
1041 changed:
1042         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1043                vchild->d_inode, dchild ? dchild->d_inode : 0,
1044                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1045                child_res_id->name[0]);
1046         if (child_res_id->name[0] != 0)
1047                 ldlm_lock_decref(child_lockh, child_mode);
1048         if (dchild)
1049                 l_dput(dchild);
1050
1051         cleanup_phase = 1; /* parent lock only */
1052         *dchildp = dchild = vchild;
1053
1054         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1055                 int flags = 0;
1056                 if (dchild->d_inode) {
1057                         child_res_id->name[0] = dchild->d_inode->i_ino;
1058                         child_res_id->name[1] = dchild->d_inode->i_generation;
1059                 } else {
1060                         child_res_id->name[0] = dchild->d_inum;
1061                         child_res_id->name[1] = dchild->d_generation;
1062                 }
1063
1064                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1065                     res_gt(maxres, child_res_id, NULL, NULL)) {
1066                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1067                                child_res_id->name[0], parent_res_id->name[0],
1068                                maxres->name[0]);
1069                         GOTO(cleanup, rc = 1);
1070                 }
1071
1072                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1073                                       *child_res_id, LDLM_IBITS, child_policy,
1074                                       child_mode, &flags, mds_blocking_ast,
1075                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1076                                       NULL, child_lockh);
1077                 if (rc != ELDLM_OK)
1078                         GOTO(cleanup, rc = -EIO);
1079
1080         } else {
1081                 memset(child_res_id, 0, sizeof(*child_res_id));
1082         }
1083
1084         EXIT;
1085 cleanup:
1086         if (rc) {
1087                 switch(cleanup_phase) {
1088                 case 2:
1089                         if (child_res_id->name[0] != 0)
1090                                 ldlm_lock_decref(child_lockh, child_mode);
1091                 case 1:
1092                         ldlm_lock_decref(parent_lockh, parent_mode);
1093                 }
1094         }
1095         return rc;
1096 }
1097
1098 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1099                                 struct ll_fid *fid,
1100                                 struct lustre_handle *parent_lockh,
1101                                 struct dentry **dparentp, int parent_mode,
1102                                 __u64 parent_lockpart,
1103                                 char *name, int namelen,
1104                                 struct lustre_handle *child_lockh,
1105                                 struct dentry **dchildp, int child_mode,
1106                                 __u64 child_lockpart)
1107 {
1108         struct ldlm_res_id child_res_id = { .name = {0} };
1109         struct ldlm_res_id parent_res_id = { .name = {0} };
1110         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1111         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1112         struct inode *inode;
1113         int rc = 0, cleanup_phase = 0;
1114         ENTRY;
1115
1116         /* Step 1: Lookup parent */
1117         *dparentp = mds_fid2dentry(mds, fid, NULL);
1118         if (IS_ERR(*dparentp))
1119                 RETURN(rc = PTR_ERR(*dparentp));
1120         LASSERT((*dparentp)->d_inode);
1121
1122         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1123                (*dparentp)->d_inode->i_ino, name);
1124
1125         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1126         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1127 #ifdef S_PDIROPS
1128         parent_lockh[1].cookie = 0;
1129         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1130                 /* lock just dir { ino, generation } to flush client cache */
1131                 if (parent_mode == LCK_PW) {
1132                         struct ldlm_res_id res_id = { .name = {0} };
1133                         ldlm_policy_data_t policy;
1134                         int flags = 0;
1135                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1136                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1137                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1138                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1139                                               res_id, LDLM_IBITS,
1140                                               &policy, LCK_CW, &flags,
1141                                               mds_blocking_ast,
1142                                               ldlm_completion_ast, NULL, NULL,
1143                                               NULL, 0, NULL, parent_lockh+1);
1144                         if (rc != ELDLM_OK)
1145                                 RETURN(-ENOLCK);
1146                 }
1147
1148                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1149                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1150                        (*dparentp)->d_inode->i_ino,
1151                        (*dparentp)->d_inode->i_generation,
1152                        parent_res_id.name[2]);
1153         }
1154 #endif
1155
1156         cleanup_phase = 1; /* parent dentry */
1157
1158         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1159         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1160         if (IS_ERR(*dchildp)) {
1161                 rc = PTR_ERR(*dchildp);
1162                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1163                 GOTO(cleanup, rc);
1164         }
1165
1166         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1167                 /* inode lives on another MDS: return * mds/ino/gen
1168                  * and LOOKUP lock. drop possible UPDATE lock! */
1169                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1170                 child_res_id.name[0] = (*dchildp)->d_inum;
1171                 child_res_id.name[1] = (*dchildp)->d_generation;
1172                 goto retry_locks;
1173         }
1174
1175         inode = (*dchildp)->d_inode;
1176         if (inode != NULL)
1177                 inode = igrab(inode);
1178         if (inode == NULL)
1179                 goto retry_locks;
1180
1181         child_res_id.name[0] = inode->i_ino;
1182         child_res_id.name[1] = inode->i_generation;
1183
1184         iput(inode);
1185
1186 retry_locks:
1187         cleanup_phase = 2; /* child dentry */
1188
1189         /* Step 3: Lock parent and child in resource order.  If child doesn't
1190          *         exist, we still have to lock the parent and re-lookup. */
1191         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1192                                    &parent_policy, &child_res_id, child_lockh,
1193                                    child_mode, &child_policy);
1194         if (rc)
1195                 GOTO(cleanup, rc);
1196
1197         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1198                 cleanup_phase = 4; /* child lock */
1199         else
1200                 cleanup_phase = 3; /* parent lock */
1201
1202         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1203         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1204                               parent_mode, &child_res_id, child_lockh,
1205                               dchildp, child_mode, &child_policy,
1206                               name, namelen, &parent_res_id);
1207         if (rc > 0)
1208                 goto retry_locks;
1209         if (rc < 0) {
1210                 cleanup_phase = 3;
1211                 GOTO(cleanup, rc);
1212         }
1213
1214 cleanup:
1215         if (rc) {
1216                 switch (cleanup_phase) {
1217                 case 4:
1218                         ldlm_lock_decref(child_lockh, child_mode);
1219                 case 3:
1220                         ldlm_lock_decref(parent_lockh, parent_mode);
1221                 case 2:
1222                         l_dput(*dchildp);
1223                 case 1:
1224 #ifdef S_PDIROPS
1225                         if (parent_lockh[1].cookie)
1226                                 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1227 #endif
1228                         l_dput(*dparentp);
1229                 default: ;
1230                 }
1231         }
1232         return rc;
1233 }
1234
1235 void mds_reconstruct_generic(struct ptlrpc_request *req)
1236 {
1237         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1238
1239         mds_req_from_mcd(req, med->med_mcd);
1240 }
1241
1242 int mds_create_local_dentry(struct mds_update_record *rec,
1243                            struct obd_device *obd)
1244 {
1245         struct mds_obd *mds = &obd->u.mds;
1246         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1247         int fidlen = 0, rc, cleanup_phase = 0;
1248         struct dentry *new_child = NULL;
1249         char *fidname = rec->ur_name;
1250         struct dentry *child = NULL;
1251         struct lustre_handle lockh;
1252         unsigned mode;
1253         void *handle;
1254         ENTRY;
1255
1256         down(&fids_dir->i_sem);
1257         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1258         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1259                         fidname, (unsigned) rec->ur_fid1->id,
1260                         (unsigned) rec->ur_fid1->generation);
1261
1262         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1263         up(&fids_dir->i_sem);
1264         if (IS_ERR(new_child)) {
1265                 CERROR("can't lookup %s: %d\n", fidname,
1266                                 (int) PTR_ERR(new_child));
1267                 GOTO(cleanup, rc = PTR_ERR(new_child));
1268         }
1269         cleanup_phase = 1;
1270
1271         if (new_child->d_inode != NULL) {
1272                 /* nice. we've already have local dentry! */
1273                 CERROR("found dentry in FIDS/: %u/%u\n",
1274                        (unsigned) new_child->d_inode->i_ino,
1275                        (unsigned) new_child->d_inode->i_generation);
1276                 rec->ur_fid1->id = fids_dir->i_ino;
1277                 rec->ur_fid1->generation = fids_dir->i_generation;
1278                 rec->ur_namelen = fidlen + 1;
1279                 GOTO(cleanup, rc = 0);
1280         }
1281
1282         /* new, local dentry will be added soon. we need no aliases here */
1283         d_drop(new_child);
1284
1285         child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
1286                                       &lockh, NULL, 0, MDS_INODELOCK_UPDATE);
1287         if (IS_ERR(child)) {
1288                 CERROR("can't get victim\n");
1289                 GOTO(cleanup, rc = PTR_ERR(child));
1290         }
1291         cleanup_phase = 2;
1292
1293         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1294         if (IS_ERR(handle))
1295                 GOTO(cleanup, rc = PTR_ERR(handle));
1296
1297         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1298                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1299                                   mds->mds_num);
1300         if (rc)
1301                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1302                        (unsigned long) child->d_inode->i_ino,
1303                        (unsigned long) child->d_inode->i_generation, rc);
1304         else
1305                 mark_inode_dirty(child->d_inode);
1306         fsfilt_commit(obd, fids_dir, handle, 0);
1307
1308         rec->ur_fid1->id = fids_dir->i_ino;
1309         rec->ur_fid1->generation = fids_dir->i_generation;
1310         rec->ur_namelen = fidlen + 1;
1311
1312 cleanup:
1313         switch(cleanup_phase) {
1314                 case 2:
1315                         ldlm_lock_decref(&lockh, LCK_EX);
1316                         dput(child);
1317                 case 1:
1318                         dput(new_child);
1319                 case 0:
1320                        break;
1321         }
1322         RETURN(rc);
1323 }
1324
1325 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1326                                         struct ptlrpc_request *slave)
1327 {
1328         struct lov_mds_md *eadata;
1329         void *cookie, *cookie2;
1330         struct mds_body *body2;
1331         struct mds_body *body;
1332         void *ea, *ea2;
1333         ENTRY;
1334
1335         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1336         LASSERT(body != NULL);
1337
1338         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1339         LASSERT(body2 != NULL);
1340
1341         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1342                 CWARN("empty reply\n");
1343                 RETURN(0);
1344         }
1345
1346         memcpy(body2, body, sizeof(*body));
1347         body2->valid &= ~OBD_MD_FLCOOKIE;
1348
1349         if (!(body->valid & OBD_MD_FLEASIZE))
1350                 RETURN(0);
1351
1352         if (body->eadatasize == 0) {
1353                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1354                 RETURN(0);
1355         }
1356
1357         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1358
1359         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1360         LASSERT(ea != NULL);
1361
1362         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1363         LASSERT(ea2 != NULL);
1364
1365         memcpy(ea2, ea, body->eadatasize);
1366
1367         if (body->valid & OBD_MD_FLCOOKIE) {
1368                 LASSERT(master->rq_repmsg->buflens[2] >=
1369                                 slave->rq_repmsg->buflens[2]);
1370                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1371                                 slave->rq_repmsg->buflens[2]);
1372                 LASSERT(cookie != NULL);
1373
1374                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1375                                 master->rq_repmsg->buflens[2]);
1376                 LASSERT(cookie2 != NULL);
1377                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1378                 body2->valid |= OBD_MD_FLCOOKIE;
1379         }
1380         RETURN(0);
1381 }
1382
1383 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1384                                    struct ptlrpc_request *req,
1385                                    struct lustre_handle *parent_lockh,
1386                                    struct dentry *dparent,
1387                                    struct lustre_handle *child_lockh,
1388                                    struct dentry *dchild)
1389 {
1390         struct mds_obd *mds = mds_req2mds(req);
1391         struct mdc_op_data op_data;
1392         int rc = 0, cleanup_phase = 0;
1393         struct ptlrpc_request *request = NULL;
1394         ENTRY;
1395
1396         LASSERT(offset == 0 || offset == 2);
1397
1398         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1399                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1400         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)\n",
1401                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1402                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1403
1404         /* time to drop i_nlink on remote MDS */
1405         op_data.fid1.mds = dchild->d_mdsnum;
1406         op_data.fid1.id = dchild->d_inum;
1407         op_data.fid1.generation = dchild->d_generation;
1408         op_data.create_mode = rec->ur_mode;
1409         op_data.namelen = 0;
1410         op_data.name = NULL;
1411         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1412         cleanup_phase = 2;
1413         if (request) {
1414                 mds_copy_unlink_reply(req, request);
1415                 ptlrpc_req_finished(request);
1416         }
1417         if (rc == 0)
1418                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1419         req->rq_status = rc;
1420
1421 #ifdef S_PDIROPS
1422         if (parent_lockh[1].cookie != 0)
1423                 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1424 #endif
1425         ldlm_lock_decref(child_lockh, LCK_EX);
1426         if (rc)
1427                 ldlm_lock_decref(parent_lockh, LCK_PW);
1428         else
1429                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1430         l_dput(dchild);
1431         l_dput(dparent);
1432
1433         return 0;
1434 }
1435
1436 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1437                             struct ptlrpc_request *req,
1438                             struct lustre_handle *lh)
1439 {
1440         struct dentry *dparent, *dchild;
1441         struct mds_obd *mds = mds_req2mds(req);
1442         struct obd_device *obd = req->rq_export->exp_obd;
1443         struct mds_body *body = NULL;
1444         struct inode *child_inode;
1445         struct lustre_handle parent_lockh[2], child_lockh, child_reuse_lockh;
1446         char fidname[LL_FID_NAMELEN];
1447         void *handle = NULL;
1448         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1449         int unlink_by_fid = 0;
1450         ENTRY;
1451
1452         LASSERT(offset == 0 || offset == 2);
1453
1454         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1455                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1456
1457         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1458
1459         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1460                 GOTO(cleanup, rc = -ENOENT);
1461
1462         if (rec->ur_namelen == 1) {
1463                 /* this is request to drop i_nlink on local inode */
1464                 unlink_by_fid = 1;
1465                 rec->ur_name = fidname;
1466                 rc = mds_create_local_dentry(rec, obd);
1467                 LASSERT(rc == 0);
1468         }
1469         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1470                                          parent_lockh, &dparent, LCK_PW,
1471                                          MDS_INODELOCK_UPDATE,
1472                                          rec->ur_name, rec->ur_namelen,
1473                                          &child_lockh, &dchild, LCK_EX,
1474                                          MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1475         if (rc)
1476                 GOTO(cleanup, rc);
1477
1478         if (dchild->d_flags & DCACHE_CROSS_REF) {
1479                 /* we should have parent lock only here */
1480                 LASSERT(unlink_by_fid == 0);
1481                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1482                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1483                                         dparent, &child_lockh, dchild);
1484                 RETURN(0);
1485         }
1486
1487         cleanup_phase = 1; /* dchild, dparent, locks */
1488
1489         dget(dchild);
1490         child_inode = dchild->d_inode;
1491         if (child_inode == NULL) {
1492                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1493                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1494                 GOTO(cleanup, rc = -ENOENT);
1495         }
1496
1497         cleanup_phase = 2; /* dchild has a lock */
1498
1499         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1500          * reuse (see bug 2029). */
1501         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1502         if (rc != ELDLM_OK)
1503                 GOTO(cleanup, rc);
1504
1505         cleanup_phase = 3; /* child inum lock */
1506
1507         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1508
1509         /* ldlm_reply in buf[0] if called via intent */
1510         if (offset)
1511                 offset = 1;
1512
1513         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1514         LASSERT(body != NULL);
1515
1516         /* If this is the last reference to this inode, get the OBD EA
1517          * data first so the client can destroy OST objects.
1518          * we only do the object removal if no open files remain.
1519          * Nobody can get at this name anymore because of the locks so
1520          * we make decisions here as to whether to remove the inode */
1521         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1522             mds_open_orphan_count(child_inode) == 0) {
1523                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1524                 mds_pack_inode2body(obd, body, child_inode);
1525                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1526                             child_inode, 1);
1527                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1528                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1529                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1530                 } else {
1531                         log_unlink = 1;
1532                 }
1533         }
1534
1535         /* We have to do these checks ourselves, in case we are making an
1536          * orphan.  The client tells us whether rmdir() or unlink() was called,
1537          * so we need to return appropriate errors (bug 72).
1538          *
1539          * We don't have to check permissions, because vfs_rename (called from
1540          * mds_open_unlink_rename) also calls may_delete. */
1541         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1542                 if (!S_ISDIR(child_inode->i_mode))
1543                         GOTO(cleanup, rc = -ENOTDIR);
1544         } else {
1545                 if (S_ISDIR(child_inode->i_mode))
1546                         GOTO(cleanup, rc = -EISDIR);
1547         }
1548
1549         if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1550             mds_open_orphan_count(child_inode) > 0) {
1551                 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1552                 cleanup_phase = 4; /* transaction */
1553                 GOTO(cleanup, rc);
1554         }
1555
1556         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1557         switch (child_inode->i_mode & S_IFMT) {
1558         case S_IFDIR:
1559                 /* Drop any lingering child directories before we start our
1560                  * transaction, to avoid doing multiple inode dirty/delete
1561                  * in our compound transaction (bug 1321). */
1562                 shrink_dcache_parent(dchild);
1563                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1564                                       NULL);
1565                 if (IS_ERR(handle))
1566                         GOTO(cleanup, rc = PTR_ERR(handle));
1567                 cleanup_phase = 4; /* transaction */
1568                 rc = vfs_rmdir(dparent->d_inode, dchild);
1569                 break;
1570         case S_IFREG: {
1571 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1572                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1573                                                         offset + 1, 0);
1574                 handle = fsfilt_start_log(obd, dparent->d_inode,
1575                                           FSFILT_OP_UNLINK, NULL,
1576                                           le32_to_cpu(lmm->lmm_stripe_count));
1577                 if (IS_ERR(handle))
1578                         GOTO(cleanup, rc = PTR_ERR(handle));
1579
1580                 cleanup_phase = 4; /* transaction */
1581                 rc = vfs_unlink(dparent->d_inode, dchild);
1582
1583                 if (!rc && log_unlink)
1584                         if (mds_log_op_unlink(obd, child_inode,
1585                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1586                                 req->rq_repmsg->buflens[offset + 1],
1587                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1588                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1589                                 body->valid |= OBD_MD_FLCOOKIE;
1590                 break;
1591         }
1592         case S_IFLNK:
1593         case S_IFCHR:
1594         case S_IFBLK:
1595         case S_IFIFO:
1596         case S_IFSOCK:
1597                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1598                                       NULL);
1599                 if (IS_ERR(handle))
1600                         GOTO(cleanup, rc = PTR_ERR(handle));
1601                 cleanup_phase = 4; /* transaction */
1602                 rc = vfs_unlink(dparent->d_inode, dchild);
1603                 break;
1604         default:
1605                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1606                        rec->ur_name);
1607                 LBUG();
1608                 GOTO(cleanup, rc = -EINVAL);
1609         }
1610
1611  cleanup:
1612         if (rc == 0) {
1613                 struct iattr iattr;
1614                 int err;
1615
1616                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1617                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1618                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1619
1620                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1621                 if (err)
1622                         CERROR("error on parent setattr: rc = %d\n", err);
1623         }
1624
1625         switch(cleanup_phase) {
1626         case 4:
1627                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1628                                         rc, 0);
1629                 if (!rc)
1630                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1631                                            "unlinked", 0, NULL);
1632         case 3: /* child ino-reuse lock */
1633                 if (rc && body != NULL) {
1634                         // Don't unlink the OST objects if the MDS unlink failed
1635                         body->valid = 0;
1636                 }
1637                 if (rc)
1638                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1639                 else
1640                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1641         case 2: /* child lock */
1642                 ldlm_lock_decref(&child_lockh, LCK_EX);
1643         case 1: /* child and parent dentry, parent lock */
1644 #ifdef S_PDIROPS
1645                 if (parent_lockh[1].cookie != 0)
1646                         ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1647 #endif
1648                 if (rc)
1649                         ldlm_lock_decref(parent_lockh, LCK_PW);
1650                 else
1651                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1652                 l_dput(dchild);
1653                 l_dput(dchild);
1654                 l_dput(dparent);
1655         case 0:
1656                 break;
1657         default:
1658                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1659                 LBUG();
1660         }
1661         req->rq_status = rc;
1662         return 0;
1663 }
1664
1665 /*
1666  * to service requests from remote MDS to increment i_nlink
1667  */
1668 static int mds_reint_link_acquire(struct mds_update_record *rec,
1669                                  int offset, struct ptlrpc_request *req,
1670                                  struct lustre_handle *lh)
1671 {
1672         struct obd_device *obd = req->rq_export->exp_obd;
1673         struct ldlm_res_id src_res_id = { .name = {0} };
1674         struct lustre_handle *handle = NULL, src_lockh;
1675         struct mds_obd *mds = mds_req2mds(req);
1676         int rc = 0, cleanup_phase = 0;
1677         struct dentry *de_src = NULL;
1678         ldlm_policy_data_t policy;
1679         int flags = 0;
1680         ENTRY;
1681
1682         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1683                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1684                   (unsigned) rec->ur_fid1->id,
1685                   (unsigned) rec->ur_fid1->generation);
1686
1687         /* Step 1: Lookup the source inode and target directory by FID */
1688         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1689         if (IS_ERR(de_src))
1690                 GOTO(cleanup, rc = PTR_ERR(de_src));
1691         cleanup_phase = 1; /* source dentry */
1692
1693         src_res_id.name[0] = de_src->d_inode->i_ino;
1694         src_res_id.name[1] = de_src->d_inode->i_generation;
1695         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1696
1697         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1698                         src_res_id, LDLM_IBITS, &policy,
1699                         LCK_EX, &flags, mds_blocking_ast,
1700                         ldlm_completion_ast, NULL, NULL,
1701                         NULL, 0, NULL, &src_lockh);
1702         if (rc != ELDLM_OK)
1703                 GOTO(cleanup, rc = -ENOLCK);
1704         cleanup_phase = 2; /* lock */
1705
1706         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1707
1708         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1709         if (IS_ERR(handle)) {
1710                 rc = PTR_ERR(handle);
1711                 GOTO(cleanup, rc);
1712         }
1713         de_src->d_inode->i_nlink++;
1714         mark_inode_dirty(de_src->d_inode);
1715
1716 cleanup:
1717         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1718                                         handle, req, rc, 0);
1719         EXIT;
1720         switch (cleanup_phase) {
1721                 case 2:
1722                         if (rc)
1723                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1724                         else
1725                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1726                 case 1:
1727                         l_dput(de_src);
1728                 case 0:
1729                         break;
1730                 default:
1731                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1732                         LBUG();
1733         }
1734         req->rq_status = rc;
1735         return 0;
1736 }
1737
1738 /*
1739  * request to link to foreign inode:
1740  *  - acquire i_nlinks on this inode
1741  *  - add dentry
1742  */
1743 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1744                                     int offset, struct ptlrpc_request *req,
1745                                     struct lustre_handle *lh)
1746 {
1747         struct lustre_handle *handle = NULL, tgt_dir_lockh[2];
1748         struct obd_device *obd = req->rq_export->exp_obd;
1749         struct dentry *de_tgt_dir = NULL;
1750         struct mds_obd *mds = mds_req2mds(req);
1751         int rc = 0, cleanup_phase = 0;
1752         struct mdc_op_data op_data;
1753         struct ptlrpc_request *request = NULL;
1754         ENTRY;
1755
1756 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1757         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1758                   (unsigned) rec->ur_fid2->mds,
1759                   (unsigned) rec->ur_fid2->id,
1760                   (unsigned) rec->ur_fid2->generation,
1761                   rec->ur_namelen - 1, rec->ur_name,
1762                   (unsigned) rec->ur_fid1->mds,
1763                   (unsigned) rec->ur_fid1->id,
1764                   (unsigned)rec->ur_fid1->generation);
1765
1766         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1767                                            tgt_dir_lockh, rec->ur_name,
1768                                            rec->ur_namelen - 1,
1769                                            MDS_INODELOCK_UPDATE);
1770         if (IS_ERR(de_tgt_dir))
1771                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1772         cleanup_phase = 1;
1773
1774         op_data.fid1 = *(rec->ur_fid1);
1775         op_data.namelen = 0;
1776         op_data.name = NULL;
1777         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1778         LASSERT(rc == 0);
1779         cleanup_phase = 2;
1780         if (request)
1781                 ptlrpc_req_finished(request);
1782
1783         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1784
1785         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1786         if (IS_ERR(handle)) {
1787                 rc = PTR_ERR(handle);
1788                 GOTO(cleanup, rc);
1789         }
1790
1791         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1792                                   rec->ur_namelen - 1, rec->ur_fid1->id,
1793                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
1794         cleanup_phase = 3;
1795
1796 cleanup:
1797         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1798                                 handle, req, rc, 0);
1799         EXIT;
1800
1801         switch (cleanup_phase) {
1802                 case 3:
1803                         if (rc) {
1804                                 /* FIXME: drop i_nlink on remote inode here */
1805                                 CERROR("MUST drop drop i_nlink here\n");
1806                         }
1807                 case 2:
1808                 case 1:
1809                         if (rc) {
1810                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1811 #ifdef S_PDIROPS
1812                                 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
1813 #endif
1814                         } else {
1815                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1816 #ifdef S_PDIROPS
1817                                 ptlrpc_save_lock(req, tgt_dir_lockh + 1, LCK_CW);
1818 #endif
1819                         }
1820                         l_dput(de_tgt_dir);
1821                         break;
1822                 default:
1823                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1824                         LBUG();
1825         }
1826         req->rq_status = rc;
1827         return 0;
1828 }
1829
1830 static int mds_reint_link(struct mds_update_record *rec, int offset,
1831                           struct ptlrpc_request *req,
1832                           struct lustre_handle *lh)
1833 {
1834         struct obd_device *obd = req->rq_export->exp_obd;
1835         struct dentry *de_src = NULL;
1836         struct dentry *de_tgt_dir = NULL;
1837         struct dentry *dchild = NULL;
1838         struct mds_obd *mds = mds_req2mds(req);
1839         struct lustre_handle *handle = NULL, tgt_dir_lockh[2], src_lockh;
1840         struct ldlm_res_id src_res_id = { .name = {0} };
1841         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1842         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1843         ldlm_policy_data_t tgt_dir_policy =
1844                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1845
1846         int rc = 0, cleanup_phase = 0;
1847         ENTRY;
1848
1849         LASSERT(offset == 0);
1850
1851         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1852                   rec->ur_fid1->id, rec->ur_fid1->generation,
1853                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1854
1855         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1856
1857         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1858                 GOTO(cleanup, rc = -ENOENT);
1859
1860         if (rec->ur_fid1->mds != mds->mds_num) {
1861                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
1862                 RETURN(rc);
1863         }
1864
1865         if (rec->ur_namelen == 1) {
1866                 rc = mds_reint_link_acquire(rec, offset, req, lh);
1867                 RETURN(rc);
1868         }
1869
1870         /* Step 1: Lookup the source inode and target directory by FID */
1871         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1872         if (IS_ERR(de_src))
1873                 GOTO(cleanup, rc = PTR_ERR(de_src));
1874
1875         cleanup_phase = 1; /* source dentry */
1876
1877         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1878         if (IS_ERR(de_tgt_dir))
1879                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1880
1881         cleanup_phase = 2; /* target directory dentry */
1882
1883         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1884                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1885                de_src->d_inode->i_ino);
1886
1887         /* Step 2: Take the two locks */
1888         src_res_id.name[0] = de_src->d_inode->i_ino;
1889         src_res_id.name[1] = de_src->d_inode->i_generation;
1890         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1891         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1892 #ifdef S_PDIROPS
1893         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
1894                 int flags = 0;
1895                 /* Get a temp lock on just ino, gen to flush client cache */
1896                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1897                                       tgt_dir_res_id, LDLM_IBITS, &src_policy,
1898                                       LCK_CW, &flags, mds_blocking_ast,
1899                                       ldlm_completion_ast, NULL, NULL,
1900                                       NULL, 0, NULL, tgt_dir_lockh + 1);
1901                 if (rc != ELDLM_OK)
1902                         GOTO(cleanup, rc = -ENOLCK);
1903
1904                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
1905                                                         rec->ur_namelen - 1);
1906                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1907                        de_tgt_dir->d_inode->i_ino,
1908                        de_tgt_dir->d_inode->i_generation,
1909                        tgt_dir_res_id.name[2]);
1910         }
1911 #endif
1912         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1913                                    &src_policy,
1914                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
1915                                    &tgt_dir_policy);
1916         if (rc)
1917                 GOTO(cleanup, rc);
1918
1919         cleanup_phase = 3; /* locks */
1920
1921         /* Step 3: Lookup the child */
1922         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1923         if (IS_ERR(dchild)) {
1924                 rc = PTR_ERR(dchild);
1925                 if (rc != -EPERM && rc != -EACCES)
1926                         CERROR("child lookup error %d\n", rc);
1927                 GOTO(cleanup, rc);
1928         }
1929
1930         cleanup_phase = 4; /* child dentry */
1931
1932         if (dchild->d_inode) {
1933                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1934                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1935                 rc = -EEXIST;
1936                 GOTO(cleanup, rc);
1937         }
1938
1939         /* Step 4: Do it. */
1940         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1941
1942         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1943         if (IS_ERR(handle)) {
1944                 rc = PTR_ERR(handle);
1945                 GOTO(cleanup, rc);
1946         }
1947
1948         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1949         if (rc && rc != -EPERM && rc != -EACCES)
1950                 CERROR("vfs_link error %d\n", rc);
1951 cleanup:
1952         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1953                                 handle, req, rc, 0);
1954         EXIT;
1955
1956         switch (cleanup_phase) {
1957         case 4: /* child dentry */
1958                 l_dput(dchild);
1959         case 3: /* locks */
1960                 if (rc) {
1961                         ldlm_lock_decref(&src_lockh, LCK_EX);
1962                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1963                 } else {
1964                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1965                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1966                 }
1967         case 2: /* target dentry */
1968 #ifdef S_PDIROPS
1969                 if (tgt_dir_lockh[1].cookie)
1970                         ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
1971 #endif
1972                 if (de_tgt_dir)
1973                         l_dput(de_tgt_dir);
1974         case 1: /* source dentry */
1975                 l_dput(de_src);
1976         case 0:
1977                 break;
1978         default:
1979                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1980                 LBUG();
1981         }
1982         req->rq_status = rc;
1983         return 0;
1984 }
1985
1986 /*
1987  * add a hard link in the PENDING directory, only used by rename()
1988  */
1989 static int mds_add_link_orphan(struct mds_update_record *rec,
1990                                struct obd_device *obd,
1991                                struct dentry *dentry)
1992 {
1993         struct mds_obd *mds = &obd->u.mds;
1994         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1995         struct dentry *pending_child;
1996         char fidname[LL_FID_NAMELEN];
1997         int fidlen = 0, rc;
1998         ENTRY;
1999
2000         LASSERT(dentry->d_inode);
2001         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2002
2003         down(&pending_dir->i_sem);
2004         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2005                             dentry->d_inode->i_generation);
2006
2007         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2008                mds_open_orphan_count(dentry->d_inode),
2009                rec->ur_name, fidname);
2010
2011         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2012         if (IS_ERR(pending_child))
2013                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2014
2015         if (pending_child->d_inode != NULL) {
2016                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2017                 LASSERT(pending_child->d_inode == dentry->d_inode);
2018                 GOTO(out_dput, rc = 0);
2019         }
2020
2021         lock_kernel();
2022         rc = vfs_link(dentry, pending_dir, pending_child);
2023         unlock_kernel();
2024         if (rc)
2025                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2026                        rec->ur_name, rc);
2027         else
2028                 mds_inode_set_orphan(dentry->d_inode);
2029 out_dput:
2030         l_dput(pending_child);
2031 out_lock:
2032         up(&pending_dir->i_sem);
2033         RETURN(rc);
2034 }
2035
2036 /* The idea here is that we need to get four locks in the end:
2037  * one on each parent directory, one on each child.  We need to take
2038  * these locks in some kind of order (to avoid deadlocks), and the order
2039  * I selected is "increasing resource number" order.  We need to look up
2040  * the children, however, before we know what the resource number(s) are.
2041  * Thus the following plan:
2042  *
2043  * 1,2. Look up the parents
2044  * 3,4. Look up the children
2045  * 5. Take locks on the parents and children, in order
2046  * 6. Verify that the children haven't changed since they were looked up
2047  *
2048  * If there was a race and the children changed since they were first looked
2049  * up, it is possible that mds_verify_child() will be able to just grab the
2050  * lock on the new child resource (if it has a higher resource than any other)
2051  * but we need to compare against not only its parent, but also against the
2052  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2053  *
2054  * We need the fancy igrab() on the child inodes because we aren't holding a
2055  * lock on the parent after the lookup is done, so dentry->d_inode may change
2056  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2057  */
2058 static int mds_get_parents_children_locked(struct obd_device *obd,
2059                                            struct mds_obd *mds,
2060                                            struct ll_fid *p1_fid,
2061                                            struct dentry **de_srcdirp,
2062                                            struct ll_fid *p2_fid,
2063                                            struct dentry **de_tgtdirp,
2064                                            int parent_mode,
2065                                            const char *old_name, int old_len,
2066                                            struct dentry **de_oldp,
2067                                            const char *new_name, int new_len,
2068                                            struct dentry **de_newp,
2069                                            struct lustre_handle *dlm_handles,
2070                                            int child_mode)
2071 {
2072         struct ldlm_res_id p1_res_id = { .name = {0} };
2073         struct ldlm_res_id p2_res_id = { .name = {0} };
2074         struct ldlm_res_id c1_res_id = { .name = {0} };
2075         struct ldlm_res_id c2_res_id = { .name = {0} };
2076         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2077         /* Only dentry should change, but the inode itself would be
2078            intact otherwise */
2079         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2080         /* If something is going to be replaced, both dentry and inode locks are
2081            needed */
2082         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2083                                                         MDS_INODELOCK_UPDATE}};
2084         struct ldlm_res_id *maxres_src, *maxres_tgt;
2085         struct inode *inode;
2086         int rc = 0, cleanup_phase = 0;
2087         ENTRY;
2088
2089         /* Step 1: Lookup the source directory */
2090         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2091         if (IS_ERR(*de_srcdirp))
2092                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2093
2094         cleanup_phase = 1; /* source directory dentry */
2095
2096         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2097         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2098
2099         /* Step 2: Lookup the target directory */
2100         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2101                 *de_tgtdirp = dget(*de_srcdirp);
2102         } else {
2103                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2104                 if (IS_ERR(*de_tgtdirp))
2105                         GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
2106         }
2107
2108         cleanup_phase = 2; /* target directory dentry */
2109
2110         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2111         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2112
2113 #ifdef S_PDIROPS
2114         dlm_handles[5].cookie = 0;
2115         dlm_handles[6].cookie = 0;
2116         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2117                 /* Get a temp lock on just ino, gen to flush client cache */
2118                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2119                                            LCK_CW, &p_policy, &p2_res_id,
2120                                            &(dlm_handles[6]),LCK_CW,&p_policy);
2121                 if (rc != ELDLM_OK)
2122                         GOTO(cleanup, rc);
2123
2124                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2125                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2126                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2127                        (*de_srcdirp)->d_inode->i_ino,
2128                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2129                        (*de_tgtdirp)->d_inode->i_ino,
2130                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2131         }
2132 #endif
2133
2134         /* Step 3: Lookup the source child entry */
2135         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2136         if (IS_ERR(*de_oldp)) {
2137                 rc = PTR_ERR(*de_oldp);
2138                 CERROR("old child lookup error (%*s): %d\n",
2139                        old_len - 1, old_name, rc);
2140                 GOTO(cleanup, rc);
2141         }
2142
2143         cleanup_phase = 3; /* original name dentry */
2144
2145         inode = (*de_oldp)->d_inode;
2146         if (inode != NULL)
2147                 inode = igrab(inode);
2148         if (inode == NULL)
2149                 GOTO(cleanup, rc = -ENOENT);
2150
2151         c1_res_id.name[0] = inode->i_ino;
2152         c1_res_id.name[1] = inode->i_generation;
2153         iput(inode);
2154
2155         /* Step 4: Lookup the target child entry */
2156         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2157         if (IS_ERR(*de_newp)) {
2158                 rc = PTR_ERR(*de_newp);
2159                 CERROR("new child lookup error (%*s): %d\n",
2160                        old_len - 1, old_name, rc);
2161                 GOTO(cleanup, rc);
2162         }
2163
2164         cleanup_phase = 4; /* target dentry */
2165
2166         inode = (*de_newp)->d_inode;
2167         if (inode != NULL)
2168                 inode = igrab(inode);
2169         if (inode == NULL)
2170                 goto retry_locks;
2171
2172         c2_res_id.name[0] = inode->i_ino;
2173         c2_res_id.name[1] = inode->i_generation;
2174
2175         iput(inode);
2176
2177 retry_locks:
2178         /* Step 5: Take locks on the parents and child(ren) */
2179         maxres_src = &p1_res_id;
2180         maxres_tgt = &p2_res_id;
2181         cleanup_phase = 4; /* target dentry */
2182
2183         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2184                 maxres_src = &c1_res_id;
2185         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2186                 maxres_tgt = &c2_res_id;
2187
2188         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2189                                     &p_policy,
2190                                     &p2_res_id, &dlm_handles[1], parent_mode,
2191                                     &p_policy,
2192                                     &c1_res_id, &dlm_handles[2], child_mode,
2193                                     &c1_policy,
2194                                     &c2_res_id, &dlm_handles[3], child_mode,
2195                                     &c2_policy);
2196         if (rc)
2197                 GOTO(cleanup, rc);
2198
2199         cleanup_phase = 6; /* parent and child(ren) locks */
2200
2201         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2202         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2203                               parent_mode, &c1_res_id, &dlm_handles[2],
2204                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2205                               maxres_tgt);
2206         if (rc) {
2207                 if (c2_res_id.name[0] != 0)
2208                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2209                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2210                 cleanup_phase = 4;
2211                 if (rc > 0)
2212                         goto retry_locks;
2213                 GOTO(cleanup, rc);
2214         }
2215
2216         if ((*de_oldp)->d_inode == NULL)
2217                 GOTO(cleanup, rc = -ENOENT);
2218
2219         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2220         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2221                               parent_mode, &c2_res_id, &dlm_handles[3],
2222                               de_newp, child_mode, &c2_policy, new_name,
2223                               new_len, maxres_src);
2224         if (rc) {
2225                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2226                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2227                 cleanup_phase = 4;
2228                 if (rc > 0)
2229                         goto retry_locks;
2230                 GOTO(cleanup, rc);
2231         }
2232
2233         EXIT;
2234 cleanup:
2235         if (rc) {
2236                 switch (cleanup_phase) {
2237                 case 6: /* child lock(s) */
2238                         if (c2_res_id.name[0] != 0)
2239                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2240                         if (c1_res_id.name[0] != 0)
2241                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2242                 case 5: /* parent locks */
2243                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2244                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2245                 case 4: /* target dentry */
2246                         l_dput(*de_newp);
2247                 case 3: /* source dentry */
2248                         l_dput(*de_oldp);
2249                 case 2: /* target directory dentry */
2250                         l_dput(*de_tgtdirp);
2251                 case 1: /* source directry dentry */
2252                         l_dput(*de_srcdirp);
2253                 }
2254         }
2255
2256         return rc;
2257 }
2258
2259 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2260                                         int offset, struct ptlrpc_request *req)
2261 {
2262         struct obd_device *obd = req->rq_export->exp_obd;
2263         struct dentry *de_srcdir = NULL;
2264         struct dentry *de_new = NULL;
2265         struct mds_obd *mds = mds_req2mds(req);
2266         struct lustre_handle parent_lockh[2];
2267         struct lustre_handle child_lockh;
2268         int cleanup_phase = 0;
2269         void *handle = NULL;
2270         int rc = 0;
2271         ENTRY;
2272
2273         /* another MDS executing rename operation has asked us
2274          * to create target name. such a creation should destroy
2275          * existing target name */
2276
2277         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2278                         obd->obd_name, rec->ur_tgt,
2279                         (unsigned long) rec->ur_fid1->mds,
2280                         (unsigned long) rec->ur_fid1->id,
2281                         (unsigned long) rec->ur_fid1->generation);
2282
2283         /* first, lookup the target */
2284         child_lockh.cookie = 0;
2285         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2286                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2287                                          rec->ur_tgt, rec->ur_tgtlen,
2288                                          &child_lockh, &de_new, LCK_EX,
2289                                          MDS_INODELOCK_LOOKUP);
2290         if (rc)
2291                 GOTO(cleanup, rc);
2292
2293         cleanup_phase = 1;
2294
2295         LASSERT(de_srcdir);
2296         LASSERT(de_srcdir->d_inode);
2297         LASSERT(de_new);
2298
2299         if (de_new->d_inode) {
2300                 /* name exists and points to local inode
2301                  * try to unlink this name and create new one */
2302                 CERROR("%s: %s points to local inode %lu/%lu\n",
2303                        obd->obd_name, rec->ur_tgt,
2304                        (unsigned long) de_new->d_inode->i_ino,
2305                        (unsigned long) de_new->d_inode->i_generation);
2306                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2307                                       FSFILT_OP_RENAME, NULL);
2308                 if (IS_ERR(handle))
2309                         GOTO(cleanup, rc = PTR_ERR(handle));
2310                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2311                 if (rc)
2312                         GOTO(cleanup, rc);
2313         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2314                 /* name exists adn points to remove inode */
2315                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2316                        obd->obd_name, rec->ur_tgt,
2317                        (unsigned long) de_new->d_mdsnum,
2318                        (unsigned long) de_new->d_inum,
2319                        (unsigned long) de_new->d_generation);
2320         } else {
2321                 /* name doesn't exist. the simplest case */
2322                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2323                                       FSFILT_OP_LINK, NULL);
2324                 if (IS_ERR(handle))
2325                         GOTO(cleanup, rc = PTR_ERR(handle));
2326         }
2327
2328         cleanup_phase = 2;
2329         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2330                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2331                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2332         if (rc)
2333                 CERROR("add_dir_entry() returned error %d\n", rc);
2334 cleanup:
2335         EXIT;
2336         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2337                                 handle, req, rc, 0);
2338         switch(cleanup_phase) {
2339                 case 2:
2340                 case 1:
2341 #ifdef S_PDIROPS
2342                         if (parent_lockh[1].cookie != 0)
2343                                 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2344 #endif
2345                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2346                         if (child_lockh.cookie != 0)
2347                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2348                         l_dput(de_new);
2349                         l_dput(de_srcdir);
2350                         break;
2351                 default:
2352                         LBUG();
2353         }
2354
2355         req->rq_status = rc;
2356
2357         RETURN(0);
2358 }
2359
2360 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2361                                       struct ptlrpc_request *req)
2362 {
2363         struct obd_device *obd = req->rq_export->exp_obd;
2364         struct ptlrpc_request *req2 = NULL;
2365         struct dentry *de_srcdir = NULL;
2366         struct dentry *de_old = NULL;
2367         struct mds_obd *mds = mds_req2mds(req);
2368         struct lustre_handle parent_lockh[2];
2369         struct lustre_handle child_lockh;
2370         struct mdc_op_data opdata;
2371         int cleanup_phase = 0;
2372         void *handle = NULL;
2373         int rc = 0;
2374         ENTRY;
2375
2376         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2377                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2378         memset(&opdata, 0, sizeof(opdata));
2379
2380         child_lockh.cookie = 0;
2381         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2382                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2383                                          rec->ur_name, rec->ur_namelen,
2384                                          &child_lockh, &de_old, LCK_EX,
2385                                          MDS_INODELOCK_LOOKUP);
2386         LASSERT(rc == 0);
2387         LASSERT(de_srcdir);
2388         LASSERT(de_srcdir->d_inode);
2389         LASSERT(de_old);
2390
2391         /* we already know the target should be created on another MDS
2392          * so, we have to request that MDS to do it */
2393
2394         /* prepare source fid */
2395         if (de_old->d_flags & DCACHE_CROSS_REF) {
2396                 LASSERT(de_old->d_inode == NULL);
2397                 CDEBUG(D_OTHER, "request to move remote name\n");
2398                 opdata.fid1.mds = de_old->d_mdsnum;
2399                 opdata.fid1.id = de_old->d_inum;
2400                 opdata.fid1.generation = de_old->d_generation;
2401         } else if (de_old->d_inode == NULL) {
2402                 /* oh, source doesn't exist */
2403                 GOTO(cleanup, rc = -ENOENT);
2404         } else {
2405                 LASSERT(de_old->d_inode != NULL);
2406                 CDEBUG(D_OTHER, "request to move local name\n");
2407                 opdata.fid1.mds = mds->mds_num;
2408                 opdata.fid1.id = de_old->d_inode->i_ino;
2409                 opdata.fid1.generation = de_old->d_inode->i_generation;
2410         }
2411
2412         opdata.fid2 = *(rec->ur_fid2);
2413         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2414                        rec->ur_tgtlen - 1, &req2);
2415
2416         if (rc)
2417                 GOTO(cleanup, rc);
2418
2419         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2420         if (IS_ERR(handle))
2421                 GOTO(cleanup, rc = PTR_ERR(handle));
2422         rc = fsfilt_del_dir_entry(obd, de_old);
2423         d_drop(de_old);
2424
2425 cleanup:
2426         EXIT;
2427         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2428                                 handle, req, rc, 0);
2429         if (req2)
2430                 ptlrpc_req_finished(req2);
2431
2432 #ifdef S_PDIROPS
2433         if (parent_lockh[1].cookie != 0)
2434                 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2435 #endif
2436         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2437         if (child_lockh.cookie != 0)
2438                 ldlm_lock_decref(&child_lockh, LCK_EX);
2439
2440         l_dput(de_old);
2441         l_dput(de_srcdir);
2442
2443         req->rq_status = rc;
2444         RETURN(0);
2445
2446 }
2447
2448 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2449                             struct ptlrpc_request *req,
2450                             struct lustre_handle *lockh)
2451 {
2452         struct obd_device *obd = req->rq_export->exp_obd;
2453         struct dentry *de_srcdir = NULL;
2454         struct dentry *de_tgtdir = NULL;
2455         struct dentry *de_old = NULL;
2456         struct dentry *de_new = NULL;
2457         struct mds_obd *mds = mds_req2mds(req);
2458         struct lustre_handle dlm_handles[7];
2459         struct mds_body *body = NULL;
2460         int rc = 0, lock_count = 3;
2461         int cleanup_phase = 0;
2462         void *handle = NULL;
2463         ENTRY;
2464
2465         LASSERT(offset == 0);
2466
2467         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2468                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2469                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2470
2471         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2472
2473         if (rec->ur_namelen == 1) {
2474                 rc = mds_reint_rename_create_name(rec, offset, req);
2475                 RETURN(rc);
2476         }
2477
2478         if (rec->ur_fid2->mds != mds->mds_num) {
2479                 rc = mds_reint_rename_to_remote(rec, offset, req);
2480                 RETURN(rc);
2481         }
2482
2483         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2484                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2485                                              rec->ur_name, rec->ur_namelen,
2486                                              &de_old, rec->ur_tgt,
2487                                              rec->ur_tgtlen, &de_new,
2488                                              dlm_handles, LCK_EX);
2489         if (rc)
2490                 GOTO(cleanup, rc);
2491
2492         cleanup_phase = 1; /* parent(s), children, locks */
2493
2494         if (de_new->d_inode)
2495                 lock_count = 4;
2496
2497         /* sanity check for src inode */
2498         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2499             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2500                 GOTO(cleanup, rc = -EINVAL);
2501
2502         /* sanity check for dest inode */
2503         if (de_new->d_inode &&
2504             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2505              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2506                 GOTO(cleanup, rc = -EINVAL);
2507
2508         if (de_old->d_inode == de_new->d_inode) {
2509                 GOTO(cleanup, rc = 0);
2510         }
2511
2512         /* if we are about to remove the target at first, pass the EA of
2513          * that inode to client to perform and cleanup on OST */
2514         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2515         LASSERT(body != NULL);
2516
2517         if (de_new->d_inode &&
2518             S_ISREG(de_new->d_inode->i_mode) &&
2519             de_new->d_inode->i_nlink == 1 &&
2520             mds_open_orphan_count(de_new->d_inode) == 0) {
2521                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2522                 mds_pack_inode2body(obd, body, de_new->d_inode);
2523                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2524                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2525                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2526                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2527                 } else {
2528                         /* XXX need log unlink? */
2529                 }
2530         }
2531
2532         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2533                        de_srcdir->d_inode->i_sb);
2534
2535         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2536         if (IS_ERR(handle))
2537                 GOTO(cleanup, rc = PTR_ERR(handle));
2538
2539         /* FIXME need adjust the journal block count? */
2540         /* if the target should be moved to PENDING, we at first increase the
2541          * link and later vfs_rename() will decrease the link count again */
2542         if (de_new->d_inode &&
2543             S_ISREG(de_new->d_inode->i_mode) &&
2544             de_new->d_inode->i_nlink == 1 &&
2545             mds_open_orphan_count(de_new->d_inode) > 0) {
2546                 rc = mds_add_link_orphan(rec, obd, de_new);
2547                 if (rc)
2548                         GOTO(cleanup, rc);
2549         }
2550
2551         lock_kernel();
2552         de_old->d_fsdata = req;
2553         de_new->d_fsdata = req;
2554         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2555         unlock_kernel();
2556
2557         GOTO(cleanup, rc);
2558 cleanup:
2559         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2560                                 handle, req, rc, 0);
2561         switch (cleanup_phase) {
2562         case 1:
2563 #ifdef S_PDIROPS
2564                 if (dlm_handles[5].cookie != 0)
2565                         ldlm_lock_decref(&(dlm_handles[5]), LCK_CW);
2566                 if (dlm_handles[6].cookie != 0)
2567                         ldlm_lock_decref(&(dlm_handles[6]), LCK_CW);
2568 #endif
2569                 if (rc) {
2570                         if (lock_count == 4)
2571                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2572                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2573                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2574                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2575                 } else {
2576                         if (lock_count == 4)
2577                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2578                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2579                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2580                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2581                 }
2582                 l_dput(de_new);
2583                 l_dput(de_old);
2584                 l_dput(de_tgtdir);
2585                 l_dput(de_srcdir);
2586         case 0:
2587                 break;
2588         default:
2589                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2590                 LBUG();
2591         }
2592         req->rq_status = rc;
2593         return 0;
2594 }
2595
2596 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2597                            struct ptlrpc_request *, struct lustre_handle *);
2598
2599 static mds_reinter reinters[REINT_MAX + 1] = {
2600         [REINT_SETATTR] mds_reint_setattr,
2601         [REINT_CREATE] mds_reint_create,
2602         [REINT_LINK] mds_reint_link,
2603         [REINT_UNLINK] mds_reint_unlink,
2604         [REINT_RENAME] mds_reint_rename,
2605         [REINT_OPEN] mds_open
2606 };
2607
2608 int mds_reint_rec(struct mds_update_record *rec, int offset,
2609                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2610 {
2611         struct obd_device *obd = req->rq_export->exp_obd;
2612         struct obd_run_ctxt saved;
2613         int rc;
2614
2615         /* checked by unpacker */
2616         LASSERT(rec->ur_opcode <= REINT_MAX &&
2617                 reinters[rec->ur_opcode] != NULL);
2618
2619         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2620         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2621         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2622
2623         return rc;
2624 }