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 = -ERESTART);
 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 = -ERESTART);
 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                 int i;
 629
 630                 /* as Peter asked, mkdir() should distribute new directories
 631                  * over the whole cluster in order to distribute namespace
 632                  * processing load. first, we calculate which MDS to use to
 633                  * put new directory's inode in */
 634                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1);
 635                 if (i == mds->mds_num) {
 636                         /* inode will be created locally */
 637
 638                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 639                         if (IS_ERR(handle))
 640                                 GOTO(cleanup, rc = PTR_ERR(handle));
 641
 642                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 643
 644                         if (rec->ur_eadata)
 645                                 nstripes = *(u16 *)rec->ur_eadata;
 646
 647                         if (rc == 0 && nstripes) {
 648                                 /* FIXME: error handling here */
 649                                 mds_try_to_split_dir(obd, dchild,
 650                                                         NULL, nstripes);
 651                         }
 652                 } else if (!DENTRY_VALID(dchild)) {
 653                         /* inode will be created on another MDS */
 654                         struct obdo *oa = NULL;
 655                         struct mds_body *body;
 656
 657                         /* first, create that inode */
 658                         oa = obdo_alloc();
 659                         LASSERT(oa != NULL);
 660                         oa->o_mds = i;
 661                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 662                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 663                                         OBD_MD_FLUID | OBD_MD_FLGID);
 664                         oa->o_mode = dir->i_mode;
 665                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 666                                         obd->obd_name, i);
 667                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 668                         LASSERT(rc == 0);
 669
 670                         /* now, add new dir entry for it */
 671                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 672                         if (IS_ERR(handle))
 673                                 GOTO(cleanup, rc = PTR_ERR(handle));
 674                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 675                                                   rec->ur_namelen - 1,
 676                                                   oa->o_id, oa->o_generation,
 677                                                   i);
 678                         LASSERT(rc == 0);
 679
 680                         /* fill reply */
 681                         body = lustre_msg_buf(req->rq_repmsg,
 682                                               offset, sizeof (*body));
 683                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 684                         body->fid1.id = oa->o_id;
 685                         body->fid1.mds = i;
 686                         body->fid1.generation = oa->o_generation;
 687                         obdo_free(oa);
 688                 } else {
 689                         /* requested name exists in the directory */
 690                         rc = -EEXIST;
 691                 }
 692                 EXIT;
 693                 break;
 694         }
 695         case S_IFLNK:{
 696                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 697                 if (IS_ERR(handle))
 698                         GOTO(cleanup, rc = PTR_ERR(handle));
 699                 if (rec->ur_tgt == NULL)        /* no target supplied */
 700                         rc = -EINVAL;           /* -EPROTO? */
 701                 else
 702                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 703                 EXIT;
 704                 break;
 705         }
 706         case S_IFCHR:
 707         case S_IFBLK:
 708         case S_IFIFO:
 709         case S_IFSOCK:{
 710                 int rdev = rec->ur_rdev;
 711                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 712                 if (IS_ERR(handle))
 713                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 714                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 715                 EXIT;
 716                 break;
 717         }
 718         default:
 719                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 720                 dchild->d_fsdata = NULL;
 721                 GOTO(cleanup, rc = -EINVAL);
 722         }
 723
 724         /* In case we stored the desired inum in here, we want to clean up. */
 725         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 726                 dchild->d_fsdata = NULL;
 727
 728         if (rc) {
 729                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 730                 GOTO(cleanup, rc);
 731         } else if (dchild->d_inode) {
 732                 struct iattr iattr;
 733                 struct inode *inode = dchild->d_inode;
 734                 struct mds_body *body;
 735
 736                 created = 1;
 737                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 738                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 739                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 740                 iattr.ia_uid = rec->ur_fsuid;
 741                 if (dir->i_mode & S_ISGID)
 742                         iattr.ia_gid = dir->i_gid;
 743                 else
 744                         iattr.ia_gid = rec->ur_fsgid;
 745                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 746                         ATTR_MTIME | ATTR_CTIME;
 747
 748                 if (rec->ur_fid2->id) {
 749                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 750                         inode->i_generation = rec->ur_fid2->generation;
 751                         /* Dirtied and committed by the upcoming setattr. */
 752                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 753                                inode->i_ino, inode->i_generation);
 754                 } else {
 755                         struct lustre_handle child_ino_lockh;
 756
 757                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 758                                inode->i_ino, inode->i_generation);
 759
 760                         /* The inode we were allocated may have just been freed
 761                          * by an unlink operation.  We take this lock to
 762                          * synchronize against the matching reply-ack-lock taken
 763                          * in unlink, to avoid replay problems if this reply
 764                          * makes it out to the client but the unlink's does not.
 765                          * See bug 2029 for more detail.*/
 766                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 767                         if (rc != ELDLM_OK) {
 768                                 CERROR("error locking for unlink/create sync: "
 769                                        "%d\n", rc);
 770                         } else {
 771                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 772                         }
 773                 }
 774
 775                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 776                 if (rc)
 777                         CERROR("error on child setattr: rc = %d\n", rc);
 778
 779                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 780                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 781                 if (rc)
 782                         CERROR("error on parent setattr: rc = %d\n", rc);
 783
 784                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 785                 mds_pack_inode2fid(obd, &body->fid1, inode);
 786                 mds_pack_inode2body(obd, body, inode);
 787         }
 788         EXIT;
 789
 790 cleanup:
 791         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 792
 793         if (rc && created) {
 794                 /* Destroy the file we just created.  This should not need
 795                  * extra journal credits, as we have already modified all of
 796                  * the blocks needed in order to create the file in the first
 797                  * place.
 798                  */
 799                 switch (type) {
 800                 case S_IFDIR:
 801                         err = vfs_rmdir(dir, dchild);
 802                         if (err)
 803                                 CERROR("rmdir in error path: %d\n", err);
 804                         break;
 805                 default:
 806                         err = vfs_unlink(dir, dchild);
 807                         if (err)
 808                                 CERROR("unlink in error path: %d\n", err);
 809                         break;
 810                 }
 811         } else {
 812                 rc = err;
 813         }
 814         switch (cleanup_phase) {
 815         case 2: /* child dentry */
 816                 l_dput(dchild);
 817         case 1: /* locked parent dentry */
 818 #ifdef S_PDIROPS
 819                 if (lockh[1].cookie != 0)
 820                         ldlm_lock_decref(lockh + 1, LCK_CW);
 821 #endif
 822                 if (rc) {
 823                         ldlm_lock_decref(lockh, LCK_PW);
 824                 } else {
 825                         ptlrpc_save_lock (req, lockh, LCK_PW);
 826                 }
 827                 l_dput(dparent);
 828         case 0:
 829                 break;
 830         default:
 831                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 832                 LBUG();
 833         }
 834         if (mea)
 835                 OBD_FREE(mea, mea_size);
 836         req->rq_status = rc;
 837         return 0;
 838 }
 839
 840 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 841            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 842 {
 843         int i;
 844
 845         for (i = 0; i < RES_NAME_SIZE; i++) {
 846                 /* return 1 here, because enqueue_ordered will skip resources
 847                  * of all zeroes if they're sorted to the end of the list. */
 848                 if (res1->name[i] == 0 && res2->name[i] != 0)
 849                         return 1;
 850                 if (res2->name[i] == 0 && res1->name[i] != 0)
 851                         return 0;
 852
 853                 if (res1->name[i] > res2->name[i])
 854                         return 1;
 855                 if (res1->name[i] < res2->name[i])
 856                         return 0;
 857         }
 858
 859         if (!p1 || !p2)
 860                 return 0;
 861
 862         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 863                 return 1;
 864
 865         return 0;
 866 }
 867
 868 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 869  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 870  * because they take the place of local semaphores.
 871  *
 872  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 873  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 874 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 875                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 876                           ldlm_policy_data_t *p1_policy,
 877                           struct ldlm_res_id *p2_res_id,
 878                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 879                           ldlm_policy_data_t *p2_policy)
 880 {
 881         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 882         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 883         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 884         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 885         int rc, flags;
 886         ENTRY;
 887
 888         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 889
 890         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 891                res_id[1]->name[0]);
 892
 893         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 894                 handles[1] = p1_lockh;
 895                 handles[0] = p2_lockh;
 896                 res_id[1] = p1_res_id;
 897                 res_id[0] = p2_res_id;
 898                 lock_modes[1] = p1_lock_mode;
 899                 lock_modes[0] = p2_lock_mode;
 900                 policies[1] = p1_policy;
 901                 policies[0] = p2_policy;
 902         }
 903
 904         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 905                res_id[0]->name[0], res_id[1]->name[0]);
 906
 907         flags = LDLM_FL_LOCAL_ONLY;
 908         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 909                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
 910                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 911                               NULL, 0, NULL, handles[0]);
 912         if (rc != ELDLM_OK)
 913                 RETURN(-EIO);
 914         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 915
 916         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
 917             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
 918                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 919                 ldlm_lock_addref(handles[1], lock_modes[1]);
 920         } else if (res_id[1]->name[0] != 0) {
 921                 flags = LDLM_FL_LOCAL_ONLY;
 922                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 923                                       *res_id[1], LDLM_IBITS, policies[1],
 924                                       lock_modes[1], &flags, mds_blocking_ast,
 925                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 926                                       NULL, handles[1]);
 927                 if (rc != ELDLM_OK) {
 928                         ldlm_lock_decref(handles[0], lock_modes[0]);
 929                         RETURN(-EIO);
 930                 }
 931                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 932         }
 933
 934         RETURN(0);
 935 }
 936
 937 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
 938                            struct lustre_handle *p1_lockh, int p1_lock_mode,
 939                            ldlm_policy_data_t *p1_policy,
 940                            struct ldlm_res_id *p2_res_id,
 941                            struct lustre_handle *p2_lockh, int p2_lock_mode,
 942                            ldlm_policy_data_t *p2_policy,
 943                            struct ldlm_res_id *c1_res_id,
 944                            struct lustre_handle *c1_lockh, int c1_lock_mode,
 945                            ldlm_policy_data_t *c1_policy,
 946                            struct ldlm_res_id *c2_res_id,
 947                            struct lustre_handle *c2_lockh, int c2_lock_mode,
 948                            ldlm_policy_data_t *c2_policy)
 949 {
 950         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
 951                                           c1_res_id, c2_res_id };
 952         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
 953                                                  c1_lockh, c2_lockh };
 954         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
 955                               c1_lock_mode, c2_lock_mode };
 956         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
 957                                             c1_policy, c2_policy};
 958         int rc, i, j, sorted, flags;
 959         ENTRY;
 960
 961         CDEBUG(D_DLMTRACE,
 962                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 963                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 964                res_id[3]->name[0]);
 965
 966         /* simple insertion sort - we have at most 4 elements */
 967         for (i = 1; i < 4; i++) {
 968                 j = i - 1;
 969                 dlm_handles[4] = dlm_handles[i];
 970                 res_id[4] = res_id[i];
 971                 lock_modes[4] = lock_modes[i];
 972                 policies[4] = policies[i];
 973
 974                 sorted = 0;
 975                 do {
 976                         if (res_gt(res_id[j], res_id[4], policies[j],
 977                                    policies[4])) {
 978                                 dlm_handles[j + 1] = dlm_handles[j];
 979                                 res_id[j + 1] = res_id[j];
 980                                 lock_modes[j + 1] = lock_modes[j];
 981                                 policies[j + 1] = policies[j];
 982                                 j--;
 983                         } else {
 984                                 sorted = 1;
 985                         }
 986                 } while (j >= 0 && !sorted);
 987
 988                 dlm_handles[j + 1] = dlm_handles[4];
 989                 res_id[j + 1] = res_id[4];
 990                 lock_modes[j + 1] = lock_modes[4];
 991                 policies[j + 1] = policies[4];
 992         }
 993
 994         CDEBUG(D_DLMTRACE,
 995                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
 996                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
 997                res_id[3]->name[0]);
 998
 999         /* XXX we could send ASTs on all these locks first before blocking? */
1000         for (i = 0; i < 4; i++) {
1001                 flags = 0;
1002                 if (res_id[i]->name[0] == 0)
1003                         break;
1004                 if (i != 0 &&
1005                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1006                     (policies[i]->l_inodebits.bits &
1007                      policies[i-1]->l_inodebits.bits) ) {
1008                         memcpy(dlm_handles[i], dlm_handles[i-1],
1009                                sizeof(*(dlm_handles[i])));
1010                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1011                 } else {
1012                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1013                                               *res_id[i], LDLM_IBITS,
1014                                               policies[i],
1015                                               lock_modes[i], &flags,
1016                                               mds_blocking_ast,
1017                                               ldlm_completion_ast, NULL, NULL,
1018                                               NULL, 0, NULL, dlm_handles[i]);
1019                         if (rc != ELDLM_OK)
1020                                 GOTO(out_err, rc = -EIO);
1021                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1022                 }
1023         }
1024
1025         RETURN(0);
1026 out_err:
1027         while (i-- > 0)
1028                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1029
1030         return rc;
1031 }
1032
1033 /* In the unlikely case that the child changed while we were waiting
1034  * on the lock, we need to drop the lock on the old child and either:
1035  * - if the child has a lower resource name, then we have to also
1036  *   drop the parent lock and regain the locks in the right order
1037  * - in the rename case, if the child has a lower resource name than one of
1038  *   the other parent/child resources (maxres) we also need to reget the locks
1039  * - if the child has a higher resource name (this is the common case)
1040  *   we can just get the lock on the new child (still in lock order)
1041  *
1042  * Returns 0 if the child did not change or if it changed but could be locked.
1043  * Returns 1 if the child changed and we need to re-lock (no locks held).
1044  * Returns -ve error with a valid dchild (no locks held). */
1045 static int mds_verify_child(struct obd_device *obd,
1046                             struct ldlm_res_id *parent_res_id,
1047                             struct lustre_handle *parent_lockh,
1048                             struct dentry *dparent, int parent_mode,
1049                             struct ldlm_res_id *child_res_id,
1050                             struct lustre_handle *child_lockh,
1051                             struct dentry **dchildp, int child_mode,
1052                             ldlm_policy_data_t *child_policy,
1053                             const char *name, int namelen,
1054                             struct ldlm_res_id *maxres)
1055 {
1056         struct dentry *vchild, *dchild = *dchildp;
1057         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1058         ENTRY;
1059
1060         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1061         if (IS_ERR(vchild))
1062                 GOTO(cleanup, rc = PTR_ERR(vchild));
1063
1064         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1065                 if  (child_res_id->name[0] == vchild->d_inum &&
1066                                 child_res_id->name[1] == vchild->d_generation) {
1067                         if (dchild != NULL)
1068                                 l_dput(dchild);
1069                         *dchildp = vchild;
1070                         RETURN(0);
1071                 }
1072                 goto changed;
1073         }
1074
1075         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1076                    (vchild->d_inode != NULL &&
1077                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1078                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1079                 if (dchild != NULL)
1080                         l_dput(dchild);
1081                 *dchildp = vchild;
1082
1083                 RETURN(0);
1084         }
1085
1086 changed:
1087         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1088                vchild->d_inode, dchild ? dchild->d_inode : 0,
1089                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1090                child_res_id->name[0]);
1091         if (child_res_id->name[0] != 0)
1092                 ldlm_lock_decref(child_lockh, child_mode);
1093         if (dchild)
1094                 l_dput(dchild);
1095
1096         cleanup_phase = 1; /* parent lock only */
1097         *dchildp = dchild = vchild;
1098
1099         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1100                 int flags = 0;
1101                 if (dchild->d_inode) {
1102                         child_res_id->name[0] = dchild->d_inode->i_ino;
1103                         child_res_id->name[1] = dchild->d_inode->i_generation;
1104                 } else {
1105                         child_res_id->name[0] = dchild->d_inum;
1106                         child_res_id->name[1] = dchild->d_generation;
1107                 }
1108
1109                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1110                     res_gt(maxres, child_res_id, NULL, NULL)) {
1111                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1112                                child_res_id->name[0], parent_res_id->name[0],
1113                                maxres->name[0]);
1114                         GOTO(cleanup, rc = 1);
1115                 }
1116
1117                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1118                                       *child_res_id, LDLM_IBITS, child_policy,
1119                                       child_mode, &flags, mds_blocking_ast,
1120                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1121                                       NULL, child_lockh);
1122                 if (rc != ELDLM_OK)
1123                         GOTO(cleanup, rc = -EIO);
1124
1125         } else {
1126                 memset(child_res_id, 0, sizeof(*child_res_id));
1127         }
1128
1129         EXIT;
1130 cleanup:
1131         if (rc) {
1132                 switch(cleanup_phase) {
1133                 case 2:
1134                         if (child_res_id->name[0] != 0)
1135                                 ldlm_lock_decref(child_lockh, child_mode);
1136                 case 1:
1137                         ldlm_lock_decref(parent_lockh, parent_mode);
1138                 }
1139         }
1140         return rc;
1141 }
1142
1143 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1144                                 struct ll_fid *fid,
1145                                 struct lustre_handle *parent_lockh,
1146                                 struct dentry **dparentp, int parent_mode,
1147                                 __u64 parent_lockpart,
1148                                 char *name, int namelen,
1149                                 struct lustre_handle *child_lockh,
1150                                 struct dentry **dchildp, int child_mode,
1151                                 __u64 child_lockpart)
1152 {
1153         struct ldlm_res_id child_res_id = { .name = {0} };
1154         struct ldlm_res_id parent_res_id = { .name = {0} };
1155         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1156         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1157         struct inode *inode;
1158         int rc = 0, cleanup_phase = 0;
1159         ENTRY;
1160
1161         /* Step 1: Lookup parent */
1162         *dparentp = mds_fid2dentry(mds, fid, NULL);
1163         if (IS_ERR(*dparentp))
1164                 RETURN(rc = PTR_ERR(*dparentp));
1165         LASSERT((*dparentp)->d_inode);
1166
1167         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1168                (*dparentp)->d_inode->i_ino, name);
1169
1170         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1171         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1172 #ifdef S_PDIROPS
1173         parent_lockh[1].cookie = 0;
1174         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1175                 /* lock just dir { ino, generation } to flush client cache */
1176                 if (parent_mode == LCK_PW) {
1177                         struct ldlm_res_id res_id = { .name = {0} };
1178                         ldlm_policy_data_t policy;
1179                         int flags = 0;
1180                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1181                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1182                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1183                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1184                                               res_id, LDLM_IBITS,
1185                                               &policy, LCK_CW, &flags,
1186                                               mds_blocking_ast,
1187                                               ldlm_completion_ast, NULL, NULL,
1188                                               NULL, 0, NULL, parent_lockh+1);
1189                         if (rc != ELDLM_OK)
1190                                 RETURN(-ENOLCK);
1191                 }
1192
1193                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1194                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1195                        (*dparentp)->d_inode->i_ino,
1196                        (*dparentp)->d_inode->i_generation,
1197                        parent_res_id.name[2]);
1198         }
1199 #endif
1200
1201         cleanup_phase = 1; /* parent dentry */
1202
1203         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1204         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1205         if (IS_ERR(*dchildp)) {
1206                 rc = PTR_ERR(*dchildp);
1207                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1208                 GOTO(cleanup, rc);
1209         }
1210
1211         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1212                 /* inode lives on another MDS: return * mds/ino/gen
1213                  * and LOOKUP lock. drop possible UPDATE lock! */
1214                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1215                 child_res_id.name[0] = (*dchildp)->d_inum;
1216                 child_res_id.name[1] = (*dchildp)->d_generation;
1217                 goto retry_locks;
1218         }
1219
1220         inode = (*dchildp)->d_inode;
1221         if (inode != NULL)
1222                 inode = igrab(inode);
1223         if (inode == NULL)
1224                 goto retry_locks;
1225
1226         child_res_id.name[0] = inode->i_ino;
1227         child_res_id.name[1] = inode->i_generation;
1228
1229         iput(inode);
1230
1231 retry_locks:
1232         cleanup_phase = 2; /* child dentry */
1233
1234         /* Step 3: Lock parent and child in resource order.  If child doesn't
1235          *         exist, we still have to lock the parent and re-lookup. */
1236         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1237                                    &parent_policy, &child_res_id, child_lockh,
1238                                    child_mode, &child_policy);
1239         if (rc)
1240                 GOTO(cleanup, rc);
1241
1242         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1243                 cleanup_phase = 4; /* child lock */
1244         else
1245                 cleanup_phase = 3; /* parent lock */
1246
1247         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1248         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1249                               parent_mode, &child_res_id, child_lockh,
1250                               dchildp, child_mode, &child_policy,
1251                               name, namelen, &parent_res_id);
1252         if (rc > 0)
1253                 goto retry_locks;
1254         if (rc < 0) {
1255                 cleanup_phase = 3;
1256                 GOTO(cleanup, rc);
1257         }
1258
1259 cleanup:
1260         if (rc) {
1261                 switch (cleanup_phase) {
1262                 case 4:
1263                         ldlm_lock_decref(child_lockh, child_mode);
1264                 case 3:
1265                         ldlm_lock_decref(parent_lockh, parent_mode);
1266                 case 2:
1267                         l_dput(*dchildp);
1268                 case 1:
1269 #ifdef S_PDIROPS
1270                         if (parent_lockh[1].cookie)
1271                                 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1272 #endif
1273                         l_dput(*dparentp);
1274                 default: ;
1275                 }
1276         }
1277         return rc;
1278 }
1279
1280 void mds_reconstruct_generic(struct ptlrpc_request *req)
1281 {
1282         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1283
1284         mds_req_from_mcd(req, med->med_mcd);
1285 }
1286
1287 int mds_create_local_dentry(struct mds_update_record *rec,
1288                            struct obd_device *obd)
1289 {
1290         struct mds_obd *mds = &obd->u.mds;
1291         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1292         int fidlen = 0, rc, cleanup_phase = 0;
1293         struct dentry *new_child = NULL;
1294         char *fidname = rec->ur_name;
1295         struct dentry *child = NULL;
1296         struct lustre_handle lockh;
1297         void *handle;
1298         ENTRY;
1299
1300         down(&fids_dir->i_sem);
1301         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1302         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1303                         fidname, (unsigned) rec->ur_fid1->id,
1304                         (unsigned) rec->ur_fid1->generation);
1305
1306         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1307         up(&fids_dir->i_sem);
1308         if (IS_ERR(new_child)) {
1309                 CERROR("can't lookup %s: %d\n", fidname,
1310                                 (int) PTR_ERR(new_child));
1311                 GOTO(cleanup, rc = PTR_ERR(new_child));
1312         }
1313         cleanup_phase = 1;
1314
1315         if (new_child->d_inode != NULL) {
1316                 /* nice. we've already have local dentry! */
1317                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1318                        (unsigned) new_child->d_inode->i_ino,
1319                        (unsigned) new_child->d_inode->i_generation);
1320                 rec->ur_fid1->id = fids_dir->i_ino;
1321                 rec->ur_fid1->generation = fids_dir->i_generation;
1322                 rec->ur_namelen = fidlen + 1;
1323                 GOTO(cleanup, rc = 0);
1324         }
1325
1326         /* new, local dentry will be added soon. we need no aliases here */
1327         d_drop(new_child);
1328
1329         child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
1330                                       &lockh, NULL, 0, MDS_INODELOCK_UPDATE);
1331         if (IS_ERR(child)) {
1332                 CERROR("can't get victim\n");
1333                 GOTO(cleanup, rc = PTR_ERR(child));
1334         }
1335         cleanup_phase = 2;
1336
1337         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1338         if (IS_ERR(handle))
1339                 GOTO(cleanup, rc = PTR_ERR(handle));
1340
1341         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1342                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1343                                   mds->mds_num);
1344         if (rc)
1345                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1346                        (unsigned long) child->d_inode->i_ino,
1347                        (unsigned long) child->d_inode->i_generation, rc);
1348         else {
1349                 if (S_ISDIR(child->d_inode->i_mode)) {
1350                         fids_dir->i_nlink++;
1351                         mark_inode_dirty(fids_dir);
1352                 }
1353                 mark_inode_dirty(child->d_inode);
1354         }
1355         fsfilt_commit(obd, fids_dir, handle, 0);
1356
1357         rec->ur_fid1->id = fids_dir->i_ino;
1358         rec->ur_fid1->generation = fids_dir->i_generation;
1359         rec->ur_namelen = fidlen + 1;
1360
1361 cleanup:
1362         switch(cleanup_phase) {
1363                 case 2:
1364                         ldlm_lock_decref(&lockh, LCK_EX);
1365                         dput(child);
1366                 case 1:
1367                         dput(new_child);
1368                 case 0:
1369                        break;
1370         }
1371         RETURN(rc);
1372 }
1373
1374 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1375                                         struct ptlrpc_request *slave)
1376 {
1377         void *cookie, *cookie2;
1378         struct mds_body *body2;
1379         struct mds_body *body;
1380         void *ea, *ea2;
1381         ENTRY;
1382
1383         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1384         LASSERT(body != NULL);
1385
1386         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1387         LASSERT(body2 != NULL);
1388
1389         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1390                 RETURN(0);
1391         }
1392
1393         memcpy(body2, body, sizeof(*body));
1394         body2->valid &= ~OBD_MD_FLCOOKIE;
1395
1396         if (!(body->valid & OBD_MD_FLEASIZE))
1397                 RETURN(0);
1398
1399         if (body->eadatasize == 0) {
1400                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1401                 RETURN(0);
1402         }
1403
1404         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1405
1406         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1407         LASSERT(ea != NULL);
1408
1409         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1410         LASSERT(ea2 != NULL);
1411
1412         memcpy(ea2, ea, body->eadatasize);
1413
1414         if (body->valid & OBD_MD_FLCOOKIE) {
1415                 LASSERT(master->rq_repmsg->buflens[2] >=
1416                                 slave->rq_repmsg->buflens[2]);
1417                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1418                                 slave->rq_repmsg->buflens[2]);
1419                 LASSERT(cookie != NULL);
1420
1421                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1422                                 master->rq_repmsg->buflens[2]);
1423                 LASSERT(cookie2 != NULL);
1424                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1425                 body2->valid |= OBD_MD_FLCOOKIE;
1426         }
1427         RETURN(0);
1428 }
1429
1430 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1431                                    struct ptlrpc_request *req,
1432                                    struct lustre_handle *parent_lockh,
1433                                    struct dentry *dparent,
1434                                    struct lustre_handle *child_lockh,
1435                                    struct dentry *dchild)
1436 {
1437         struct mds_obd *mds = mds_req2mds(req);
1438         struct mdc_op_data op_data;
1439         int rc = 0, cleanup_phase = 0;
1440         struct ptlrpc_request *request = NULL;
1441         ENTRY;
1442
1443         LASSERT(offset == 0 || offset == 2);
1444
1445         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)\n",
1446                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1447                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1448
1449         /* time to drop i_nlink on remote MDS */
1450         op_data.fid1.mds = dchild->d_mdsnum;
1451         op_data.fid1.id = dchild->d_inum;
1452         op_data.fid1.generation = dchild->d_generation;
1453         op_data.create_mode = rec->ur_mode;
1454         op_data.namelen = 0;
1455         op_data.name = NULL;
1456         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1457         cleanup_phase = 2;
1458         if (request) {
1459                 mds_copy_unlink_reply(req, request);
1460                 ptlrpc_req_finished(request);
1461         }
1462         if (rc == 0)
1463                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1464         req->rq_status = rc;
1465
1466 #ifdef S_PDIROPS
1467         if (parent_lockh[1].cookie != 0)
1468                 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1469 #endif
1470         ldlm_lock_decref(child_lockh, LCK_EX);
1471         if (rc)
1472                 ldlm_lock_decref(parent_lockh, LCK_PW);
1473         else
1474                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1475         l_dput(dchild);
1476         l_dput(dparent);
1477
1478         return 0;
1479 }
1480
1481 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1482                             struct ptlrpc_request *req,
1483                             struct lustre_handle *lh)
1484 {
1485         struct dentry *dparent, *dchild;
1486         struct mds_obd *mds = mds_req2mds(req);
1487         struct obd_device *obd = req->rq_export->exp_obd;
1488         struct mds_body *body = NULL;
1489         struct inode *child_inode;
1490         struct lustre_handle parent_lockh[2], child_lockh, child_reuse_lockh;
1491         char fidname[LL_FID_NAMELEN];
1492         void *handle = NULL;
1493         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1494         int unlink_by_fid = 0;
1495         ENTRY;
1496
1497         LASSERT(offset == 0 || offset == 2);
1498
1499         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1500                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1501
1502         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1503
1504         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1505                 GOTO(cleanup, rc = -ENOENT);
1506
1507         if (rec->ur_namelen == 1) {
1508                 /* this is request to drop i_nlink on local inode */
1509                 unlink_by_fid = 1;
1510                 rec->ur_name = fidname;
1511                 rc = mds_create_local_dentry(rec, obd);
1512                 LASSERT(rc == 0);
1513         }
1514         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1515                                          parent_lockh, &dparent, LCK_PW,
1516                                          MDS_INODELOCK_UPDATE,
1517                                          rec->ur_name, rec->ur_namelen,
1518                                          &child_lockh, &dchild, LCK_EX,
1519                                          MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1520         if (rc)
1521                 GOTO(cleanup, rc);
1522
1523         if (dchild->d_flags & DCACHE_CROSS_REF) {
1524                 /* we should have parent lock only here */
1525                 LASSERT(unlink_by_fid == 0);
1526                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1527                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1528                                              dparent, &child_lockh, dchild);
1529                 RETURN(0);
1530         }
1531
1532         cleanup_phase = 1; /* dchild, dparent, locks */
1533
1534         dget(dchild);
1535         child_inode = dchild->d_inode;
1536         if (child_inode == NULL) {
1537                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1538                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1539                 GOTO(cleanup, rc = -ENOENT);
1540         }
1541
1542         cleanup_phase = 2; /* dchild has a lock */
1543
1544         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1545          * reuse (see bug 2029). */
1546         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1547         if (rc != ELDLM_OK)
1548                 GOTO(cleanup, rc);
1549
1550         cleanup_phase = 3; /* child inum lock */
1551
1552         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1553
1554         /* ldlm_reply in buf[0] if called via intent */
1555         if (offset)
1556                 offset = 1;
1557
1558         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1559         LASSERT(body != NULL);
1560
1561         /* If this is the last reference to this inode, get the OBD EA
1562          * data first so the client can destroy OST objects.
1563          * we only do the object removal if no open files remain.
1564          * Nobody can get at this name anymore because of the locks so
1565          * we make decisions here as to whether to remove the inode */
1566         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1567             mds_open_orphan_count(child_inode) == 0) {
1568                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1569                 mds_pack_inode2body(obd, body, child_inode);
1570                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1571                             child_inode, 1);
1572                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1573                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1574                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1575                 } else {
1576                         log_unlink = 1;
1577                 }
1578         }
1579
1580         /* We have to do these checks ourselves, in case we are making an
1581          * orphan.  The client tells us whether rmdir() or unlink() was called,
1582          * so we need to return appropriate errors (bug 72).
1583          *
1584          * We don't have to check permissions, because vfs_rename (called from
1585          * mds_open_unlink_rename) also calls may_delete. */
1586         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1587                 if (!S_ISDIR(child_inode->i_mode))
1588                         GOTO(cleanup, rc = -ENOTDIR);
1589         } else {
1590                 if (S_ISDIR(child_inode->i_mode))
1591                         GOTO(cleanup, rc = -EISDIR);
1592         }
1593
1594         if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1595             mds_open_orphan_count(child_inode) > 0) {
1596                 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1597                 cleanup_phase = 4; /* transaction */
1598                 GOTO(cleanup, rc);
1599         }
1600
1601         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1602         switch (child_inode->i_mode & S_IFMT) {
1603         case S_IFDIR:
1604                 /* Drop any lingering child directories before we start our
1605                  * transaction, to avoid doing multiple inode dirty/delete
1606                  * in our compound transaction (bug 1321). */
1607                 shrink_dcache_parent(dchild);
1608                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1609                                       NULL);
1610                 if (IS_ERR(handle))
1611                         GOTO(cleanup, rc = PTR_ERR(handle));
1612                 cleanup_phase = 4; /* transaction */
1613                 rc = vfs_rmdir(dparent->d_inode, dchild);
1614                 break;
1615         case S_IFREG: {
1616 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1617                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1618                                                         offset + 1, 0);
1619                 handle = fsfilt_start_log(obd, dparent->d_inode,
1620                                           FSFILT_OP_UNLINK, NULL,
1621                                           le32_to_cpu(lmm->lmm_stripe_count));
1622                 if (IS_ERR(handle))
1623                         GOTO(cleanup, rc = PTR_ERR(handle));
1624
1625                 cleanup_phase = 4; /* transaction */
1626                 rc = vfs_unlink(dparent->d_inode, dchild);
1627
1628                 if (!rc && log_unlink)
1629                         if (mds_log_op_unlink(obd, child_inode,
1630                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1631                                 req->rq_repmsg->buflens[offset + 1],
1632                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1633                                 req->rq_repmsg->buflens[offset + 2]) > 0)
1634                                 body->valid |= OBD_MD_FLCOOKIE;
1635                 break;
1636         }
1637         case S_IFLNK:
1638         case S_IFCHR:
1639         case S_IFBLK:
1640         case S_IFIFO:
1641         case S_IFSOCK:
1642                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1643                                       NULL);
1644                 if (IS_ERR(handle))
1645                         GOTO(cleanup, rc = PTR_ERR(handle));
1646                 cleanup_phase = 4; /* transaction */
1647                 rc = vfs_unlink(dparent->d_inode, dchild);
1648                 break;
1649         default:
1650                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1651                        rec->ur_name);
1652                 LBUG();
1653                 GOTO(cleanup, rc = -EINVAL);
1654         }
1655
1656  cleanup:
1657         if (rc == 0) {
1658                 struct iattr iattr;
1659                 int err;
1660
1661                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1662                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1663                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1664
1665                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1666                 if (err)
1667                         CERROR("error on parent setattr: rc = %d\n", err);
1668         }
1669
1670         switch(cleanup_phase) {
1671         case 4:
1672                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1673                                         rc, 0);
1674                 if (!rc)
1675                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1676                                            "unlinked", 0, NULL);
1677         case 3: /* child ino-reuse lock */
1678                 if (rc && body != NULL) {
1679                         // Don't unlink the OST objects if the MDS unlink failed
1680                         body->valid = 0;
1681                 }
1682                 if (rc)
1683                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1684                 else
1685                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1686         case 2: /* child lock */
1687                 ldlm_lock_decref(&child_lockh, LCK_EX);
1688         case 1: /* child and parent dentry, parent lock */
1689 #ifdef S_PDIROPS
1690                 if (parent_lockh[1].cookie != 0)
1691                         ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1692 #endif
1693                 if (rc)
1694                         ldlm_lock_decref(parent_lockh, LCK_PW);
1695                 else
1696                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1697                 l_dput(dchild);
1698                 l_dput(dchild);
1699                 l_dput(dparent);
1700         case 0:
1701                 break;
1702         default:
1703                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1704                 LBUG();
1705         }
1706         req->rq_status = rc;
1707         return 0;
1708 }
1709
1710 /*
1711  * to service requests from remote MDS to increment i_nlink
1712  */
1713 static int mds_reint_link_acquire(struct mds_update_record *rec,
1714                                  int offset, struct ptlrpc_request *req,
1715                                  struct lustre_handle *lh)
1716 {
1717         struct obd_device *obd = req->rq_export->exp_obd;
1718         struct ldlm_res_id src_res_id = { .name = {0} };
1719         struct lustre_handle *handle = NULL, src_lockh;
1720         struct mds_obd *mds = mds_req2mds(req);
1721         int rc = 0, cleanup_phase = 0;
1722         struct dentry *de_src = NULL;
1723         ldlm_policy_data_t policy;
1724         int flags = 0;
1725         ENTRY;
1726
1727         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1728                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1729                   (unsigned) rec->ur_fid1->id,
1730                   (unsigned) rec->ur_fid1->generation);
1731
1732         /* Step 1: Lookup the source inode and target directory by FID */
1733         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1734         if (IS_ERR(de_src))
1735                 GOTO(cleanup, rc = PTR_ERR(de_src));
1736         cleanup_phase = 1; /* source dentry */
1737
1738         src_res_id.name[0] = de_src->d_inode->i_ino;
1739         src_res_id.name[1] = de_src->d_inode->i_generation;
1740         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1741
1742         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1743                         src_res_id, LDLM_IBITS, &policy,
1744                         LCK_EX, &flags, mds_blocking_ast,
1745                         ldlm_completion_ast, NULL, NULL,
1746                         NULL, 0, NULL, &src_lockh);
1747         if (rc != ELDLM_OK)
1748                 GOTO(cleanup, rc = -ENOLCK);
1749         cleanup_phase = 2; /* lock */
1750
1751         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1752
1753         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1754         if (IS_ERR(handle)) {
1755                 rc = PTR_ERR(handle);
1756                 GOTO(cleanup, rc);
1757         }
1758         de_src->d_inode->i_nlink++;
1759         mark_inode_dirty(de_src->d_inode);
1760
1761 cleanup:
1762         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1763                                         handle, req, rc, 0);
1764         EXIT;
1765         switch (cleanup_phase) {
1766                 case 2:
1767                         if (rc)
1768                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1769                         else
1770                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1771                 case 1:
1772                         l_dput(de_src);
1773                 case 0:
1774                         break;
1775                 default:
1776                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1777                         LBUG();
1778         }
1779         req->rq_status = rc;
1780         return 0;
1781 }
1782
1783 /*
1784  * request to link to foreign inode:
1785  *  - acquire i_nlinks on this inode
1786  *  - add dentry
1787  */
1788 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1789                                     int offset, struct ptlrpc_request *req,
1790                                     struct lustre_handle *lh)
1791 {
1792         struct lustre_handle *handle = NULL, tgt_dir_lockh[2];
1793         struct obd_device *obd = req->rq_export->exp_obd;
1794         struct dentry *de_tgt_dir = NULL;
1795         struct mds_obd *mds = mds_req2mds(req);
1796         int rc = 0, cleanup_phase = 0;
1797         struct mdc_op_data op_data;
1798         struct ptlrpc_request *request = NULL;
1799         ENTRY;
1800
1801 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1802         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1803                   (unsigned) rec->ur_fid2->mds,
1804                   (unsigned) rec->ur_fid2->id,
1805                   (unsigned) rec->ur_fid2->generation,
1806                   rec->ur_namelen - 1, rec->ur_name,
1807                   (unsigned) rec->ur_fid1->mds,
1808                   (unsigned) rec->ur_fid1->id,
1809                   (unsigned)rec->ur_fid1->generation);
1810
1811         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1812                                            tgt_dir_lockh, rec->ur_name,
1813                                            rec->ur_namelen - 1,
1814                                            MDS_INODELOCK_UPDATE);
1815         if (IS_ERR(de_tgt_dir))
1816                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1817         cleanup_phase = 1;
1818
1819         op_data.fid1 = *(rec->ur_fid1);
1820         op_data.namelen = 0;
1821         op_data.name = NULL;
1822         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1823         LASSERT(rc == 0);
1824         cleanup_phase = 2;
1825         if (request)
1826                 ptlrpc_req_finished(request);
1827
1828         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1829
1830         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1831         if (IS_ERR(handle)) {
1832                 rc = PTR_ERR(handle);
1833                 GOTO(cleanup, rc);
1834         }
1835
1836         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1837                                   rec->ur_namelen - 1, rec->ur_fid1->id,
1838                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
1839         cleanup_phase = 3;
1840
1841 cleanup:
1842         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1843                                 handle, req, rc, 0);
1844         EXIT;
1845
1846         switch (cleanup_phase) {
1847                 case 3:
1848                         if (rc) {
1849                                 /* FIXME: drop i_nlink on remote inode here */
1850                                 CERROR("MUST drop drop i_nlink here\n");
1851                         }
1852                 case 2:
1853                 case 1:
1854                         if (rc) {
1855                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1856 #ifdef S_PDIROPS
1857                                 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
1858 #endif
1859                         } else {
1860                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1861 #ifdef S_PDIROPS
1862                                 ptlrpc_save_lock(req, tgt_dir_lockh + 1, LCK_CW);
1863 #endif
1864                         }
1865                         l_dput(de_tgt_dir);
1866                         break;
1867                 default:
1868                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1869                         LBUG();
1870         }
1871         req->rq_status = rc;
1872         return 0;
1873 }
1874
1875 static int mds_reint_link(struct mds_update_record *rec, int offset,
1876                           struct ptlrpc_request *req,
1877                           struct lustre_handle *lh)
1878 {
1879         struct obd_device *obd = req->rq_export->exp_obd;
1880         struct dentry *de_src = NULL;
1881         struct dentry *de_tgt_dir = NULL;
1882         struct dentry *dchild = NULL;
1883         struct mds_obd *mds = mds_req2mds(req);
1884         struct lustre_handle *handle = NULL, tgt_dir_lockh[2], src_lockh;
1885         struct ldlm_res_id src_res_id = { .name = {0} };
1886         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1887         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1888         ldlm_policy_data_t tgt_dir_policy =
1889                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1890
1891         int rc = 0, cleanup_phase = 0;
1892         ENTRY;
1893
1894         LASSERT(offset == 0);
1895
1896         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1897                   rec->ur_fid1->id, rec->ur_fid1->generation,
1898                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1899
1900         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1901
1902         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1903                 GOTO(cleanup, rc = -ENOENT);
1904
1905         if (rec->ur_fid1->mds != mds->mds_num) {
1906                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
1907                 RETURN(rc);
1908         }
1909
1910         if (rec->ur_namelen == 1) {
1911                 rc = mds_reint_link_acquire(rec, offset, req, lh);
1912                 RETURN(rc);
1913         }
1914
1915         /* Step 1: Lookup the source inode and target directory by FID */
1916         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1917         if (IS_ERR(de_src))
1918                 GOTO(cleanup, rc = PTR_ERR(de_src));
1919
1920         cleanup_phase = 1; /* source dentry */
1921
1922         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1923         if (IS_ERR(de_tgt_dir))
1924                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1925
1926         cleanup_phase = 2; /* target directory dentry */
1927
1928         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1929                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1930                de_src->d_inode->i_ino);
1931
1932         /* Step 2: Take the two locks */
1933         src_res_id.name[0] = de_src->d_inode->i_ino;
1934         src_res_id.name[1] = de_src->d_inode->i_generation;
1935         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1936         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1937 #ifdef S_PDIROPS
1938         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
1939                 int flags = 0;
1940                 /* Get a temp lock on just ino, gen to flush client cache */
1941                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1942                                       tgt_dir_res_id, LDLM_IBITS, &src_policy,
1943                                       LCK_CW, &flags, mds_blocking_ast,
1944                                       ldlm_completion_ast, NULL, NULL,
1945                                       NULL, 0, NULL, tgt_dir_lockh + 1);
1946                 if (rc != ELDLM_OK)
1947                         GOTO(cleanup, rc = -ENOLCK);
1948
1949                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
1950                                                         rec->ur_namelen - 1);
1951                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1952                        de_tgt_dir->d_inode->i_ino,
1953                        de_tgt_dir->d_inode->i_generation,
1954                        tgt_dir_res_id.name[2]);
1955         }
1956 #endif
1957         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1958                                    &src_policy,
1959                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
1960                                    &tgt_dir_policy);
1961         if (rc)
1962                 GOTO(cleanup, rc);
1963
1964         cleanup_phase = 3; /* locks */
1965
1966         /* Step 3: Lookup the child */
1967         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1968         if (IS_ERR(dchild)) {
1969                 rc = PTR_ERR(dchild);
1970                 if (rc != -EPERM && rc != -EACCES)
1971                         CERROR("child lookup error %d\n", rc);
1972                 GOTO(cleanup, rc);
1973         }
1974
1975         cleanup_phase = 4; /* child dentry */
1976
1977         if (dchild->d_inode) {
1978                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1979                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1980                 rc = -EEXIST;
1981                 GOTO(cleanup, rc);
1982         }
1983
1984         /* Step 4: Do it. */
1985         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1986
1987         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1988         if (IS_ERR(handle)) {
1989                 rc = PTR_ERR(handle);
1990                 GOTO(cleanup, rc);
1991         }
1992
1993         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1994         if (rc && rc != -EPERM && rc != -EACCES)
1995                 CERROR("vfs_link error %d\n", rc);
1996 cleanup:
1997         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1998                                 handle, req, rc, 0);
1999         EXIT;
2000
2001         switch (cleanup_phase) {
2002         case 4: /* child dentry */
2003                 l_dput(dchild);
2004         case 3: /* locks */
2005                 if (rc) {
2006                         ldlm_lock_decref(&src_lockh, LCK_EX);
2007                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2008                 } else {
2009                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2010                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2011                 }
2012         case 2: /* target dentry */
2013 #ifdef S_PDIROPS
2014                 if (tgt_dir_lockh[1].cookie)
2015                         ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
2016 #endif
2017                 if (de_tgt_dir)
2018                         l_dput(de_tgt_dir);
2019         case 1: /* source dentry */
2020                 l_dput(de_src);
2021         case 0:
2022                 break;
2023         default:
2024                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2025                 LBUG();
2026         }
2027         req->rq_status = rc;
2028         return 0;
2029 }
2030
2031 /*
2032  * add a hard link in the PENDING directory, only used by rename()
2033  */
2034 static int mds_add_link_orphan(struct mds_update_record *rec,
2035                                struct obd_device *obd,
2036                                struct dentry *dentry)
2037 {
2038         struct mds_obd *mds = &obd->u.mds;
2039         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2040         struct dentry *pending_child;
2041         char fidname[LL_FID_NAMELEN];
2042         int fidlen = 0, rc;
2043         ENTRY;
2044
2045         LASSERT(dentry->d_inode);
2046         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2047
2048         down(&pending_dir->i_sem);
2049         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2050                             dentry->d_inode->i_generation);
2051
2052         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2053                mds_open_orphan_count(dentry->d_inode),
2054                rec->ur_name, fidname);
2055
2056         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2057         if (IS_ERR(pending_child))
2058                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2059
2060         if (pending_child->d_inode != NULL) {
2061                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2062                 LASSERT(pending_child->d_inode == dentry->d_inode);
2063                 GOTO(out_dput, rc = 0);
2064         }
2065
2066         lock_kernel();
2067         rc = vfs_link(dentry, pending_dir, pending_child);
2068         unlock_kernel();
2069         if (rc)
2070                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2071                        rec->ur_name, rc);
2072         else
2073                 mds_inode_set_orphan(dentry->d_inode);
2074 out_dput:
2075         l_dput(pending_child);
2076 out_lock:
2077         up(&pending_dir->i_sem);
2078         RETURN(rc);
2079 }
2080
2081 /* The idea here is that we need to get four locks in the end:
2082  * one on each parent directory, one on each child.  We need to take
2083  * these locks in some kind of order (to avoid deadlocks), and the order
2084  * I selected is "increasing resource number" order.  We need to look up
2085  * the children, however, before we know what the resource number(s) are.
2086  * Thus the following plan:
2087  *
2088  * 1,2. Look up the parents
2089  * 3,4. Look up the children
2090  * 5. Take locks on the parents and children, in order
2091  * 6. Verify that the children haven't changed since they were looked up
2092  *
2093  * If there was a race and the children changed since they were first looked
2094  * up, it is possible that mds_verify_child() will be able to just grab the
2095  * lock on the new child resource (if it has a higher resource than any other)
2096  * but we need to compare against not only its parent, but also against the
2097  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2098  *
2099  * We need the fancy igrab() on the child inodes because we aren't holding a
2100  * lock on the parent after the lookup is done, so dentry->d_inode may change
2101  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2102  */
2103 static int mds_get_parents_children_locked(struct obd_device *obd,
2104                                            struct mds_obd *mds,
2105                                            struct ll_fid *p1_fid,
2106                                            struct dentry **de_srcdirp,
2107                                            struct ll_fid *p2_fid,
2108                                            struct dentry **de_tgtdirp,
2109                                            int parent_mode,
2110                                            const char *old_name, int old_len,
2111                                            struct dentry **de_oldp,
2112                                            const char *new_name, int new_len,
2113                                            struct dentry **de_newp,
2114                                            struct lustre_handle *dlm_handles,
2115                                            int child_mode)
2116 {
2117         struct ldlm_res_id p1_res_id = { .name = {0} };
2118         struct ldlm_res_id p2_res_id = { .name = {0} };
2119         struct ldlm_res_id c1_res_id = { .name = {0} };
2120         struct ldlm_res_id c2_res_id = { .name = {0} };
2121         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2122         /* Only dentry should change, but the inode itself would be
2123            intact otherwise */
2124         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2125         /* If something is going to be replaced, both dentry and inode locks are
2126            needed */
2127         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2128                                                         MDS_INODELOCK_UPDATE}};
2129         struct ldlm_res_id *maxres_src, *maxres_tgt;
2130         struct inode *inode;
2131         int rc = 0, cleanup_phase = 0;
2132         ENTRY;
2133
2134         /* Step 1: Lookup the source directory */
2135         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2136         if (IS_ERR(*de_srcdirp))
2137                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2138
2139         cleanup_phase = 1; /* source directory dentry */
2140
2141         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2142         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2143
2144         /* Step 2: Lookup the target directory */
2145         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2146                 *de_tgtdirp = dget(*de_srcdirp);
2147         } else {
2148                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2149                 if (IS_ERR(*de_tgtdirp))
2150                         GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
2151         }
2152
2153         cleanup_phase = 2; /* target directory dentry */
2154
2155         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2156         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2157
2158 #ifdef S_PDIROPS
2159         dlm_handles[5].cookie = 0;
2160         dlm_handles[6].cookie = 0;
2161         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2162                 /* Get a temp lock on just ino, gen to flush client cache */
2163                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2164                                            LCK_CW, &p_policy, &p2_res_id,
2165                                            &(dlm_handles[6]),LCK_CW,&p_policy);
2166                 if (rc != ELDLM_OK)
2167                         GOTO(cleanup, rc);
2168
2169                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2170                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2171                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2172                        (*de_srcdirp)->d_inode->i_ino,
2173                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2174                        (*de_tgtdirp)->d_inode->i_ino,
2175                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2176         }
2177 #endif
2178
2179         /* Step 3: Lookup the source child entry */
2180         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2181         if (IS_ERR(*de_oldp)) {
2182                 rc = PTR_ERR(*de_oldp);
2183                 CERROR("old child lookup error (%*s): %d\n",
2184                        old_len - 1, old_name, rc);
2185                 GOTO(cleanup, rc);
2186         }
2187
2188         cleanup_phase = 3; /* original name dentry */
2189
2190         inode = (*de_oldp)->d_inode;
2191         if (inode != NULL) {
2192                 inode = igrab(inode);
2193                 if (inode == NULL)
2194                         GOTO(cleanup, rc = -ENOENT);
2195
2196                 c1_res_id.name[0] = inode->i_ino;
2197                 c1_res_id.name[1] = inode->i_generation;
2198                 iput(inode);
2199         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2200                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2201                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2202         }
2203
2204         /* Step 4: Lookup the target child entry */
2205         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2206         if (IS_ERR(*de_newp)) {
2207                 rc = PTR_ERR(*de_newp);
2208                 CERROR("new child lookup error (%*s): %d\n",
2209                        old_len - 1, old_name, rc);
2210                 GOTO(cleanup, rc);
2211         }
2212
2213         cleanup_phase = 4; /* target dentry */
2214
2215         inode = (*de_newp)->d_inode;
2216         if (inode != NULL) {
2217                 inode = igrab(inode);
2218                 if (inode == NULL)
2219                         goto retry_locks;
2220
2221                 c2_res_id.name[0] = inode->i_ino;
2222                 c2_res_id.name[1] = inode->i_generation;
2223                 iput(inode);
2224         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2225                 c2_res_id.name[0] = (*de_newp)->d_inum;
2226                 c2_res_id.name[1] = (*de_newp)->d_generation;
2227         }
2228
2229 retry_locks:
2230         /* Step 5: Take locks on the parents and child(ren) */
2231         maxres_src = &p1_res_id;
2232         maxres_tgt = &p2_res_id;
2233         cleanup_phase = 4; /* target dentry */
2234
2235         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2236                 maxres_src = &c1_res_id;
2237         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2238                 maxres_tgt = &c2_res_id;
2239
2240         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2241                                     &p_policy,
2242                                     &p2_res_id, &dlm_handles[1], parent_mode,
2243                                     &p_policy,
2244                                     &c1_res_id, &dlm_handles[2], child_mode,
2245                                     &c1_policy,
2246                                     &c2_res_id, &dlm_handles[3], child_mode,
2247                                     &c2_policy);
2248         if (rc)
2249                 GOTO(cleanup, rc);
2250
2251         cleanup_phase = 6; /* parent and child(ren) locks */
2252
2253         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2254         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2255                               parent_mode, &c1_res_id, &dlm_handles[2],
2256                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2257                               maxres_tgt);
2258         if (rc) {
2259                 if (c2_res_id.name[0] != 0)
2260                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2261                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2262                 cleanup_phase = 4;
2263                 if (rc > 0)
2264                         goto retry_locks;
2265                 GOTO(cleanup, rc);
2266         }
2267
2268         if (!DENTRY_VALID(*de_oldp))
2269                 GOTO(cleanup, rc = -ENOENT);
2270
2271         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2272         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2273                               parent_mode, &c2_res_id, &dlm_handles[3],
2274                               de_newp, child_mode, &c2_policy, new_name,
2275                               new_len, maxres_src);
2276         if (rc) {
2277                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2278                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2279                 cleanup_phase = 4;
2280                 if (rc > 0)
2281                         goto retry_locks;
2282                 GOTO(cleanup, rc);
2283         }
2284
2285         EXIT;
2286 cleanup:
2287         if (rc) {
2288                 switch (cleanup_phase) {
2289                 case 6: /* child lock(s) */
2290                         if (c2_res_id.name[0] != 0)
2291                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2292                         if (c1_res_id.name[0] != 0)
2293                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2294                 case 5: /* parent locks */
2295                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2296                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2297                 case 4: /* target dentry */
2298                         l_dput(*de_newp);
2299                 case 3: /* source dentry */
2300                         l_dput(*de_oldp);
2301                 case 2: /* target directory dentry */
2302                         l_dput(*de_tgtdirp);
2303                 case 1: /* source directry dentry */
2304                         l_dput(*de_srcdirp);
2305                 }
2306         }
2307
2308         return rc;
2309 }
2310
2311 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2312                                         int offset, struct ptlrpc_request *req)
2313 {
2314         struct obd_device *obd = req->rq_export->exp_obd;
2315         struct dentry *de_srcdir = NULL;
2316         struct dentry *de_new = NULL;
2317         struct mds_obd *mds = mds_req2mds(req);
2318         struct lustre_handle parent_lockh[2];
2319         struct lustre_handle child_lockh;
2320         int cleanup_phase = 0;
2321         void *handle = NULL;
2322         int rc = 0;
2323         ENTRY;
2324
2325         /* another MDS executing rename operation has asked us
2326          * to create target name. such a creation should destroy
2327          * existing target name */
2328
2329         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2330                         obd->obd_name, rec->ur_tgt,
2331                         (unsigned long) rec->ur_fid1->mds,
2332                         (unsigned long) rec->ur_fid1->id,
2333                         (unsigned long) rec->ur_fid1->generation);
2334
2335         /* first, lookup the target */
2336         child_lockh.cookie = 0;
2337         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2338                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2339                                          rec->ur_tgt, rec->ur_tgtlen,
2340                                          &child_lockh, &de_new, LCK_EX,
2341                                          MDS_INODELOCK_LOOKUP);
2342         if (rc)
2343                 GOTO(cleanup, rc);
2344
2345         cleanup_phase = 1;
2346
2347         LASSERT(de_srcdir);
2348         LASSERT(de_srcdir->d_inode);
2349         LASSERT(de_new);
2350
2351         if (de_new->d_inode) {
2352                 /* name exists and points to local inode
2353                  * try to unlink this name and create new one */
2354                 CERROR("%s: %s points to local inode %lu/%lu\n",
2355                        obd->obd_name, rec->ur_tgt,
2356                        (unsigned long) de_new->d_inode->i_ino,
2357                        (unsigned long) de_new->d_inode->i_generation);
2358                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2359                                       FSFILT_OP_RENAME, NULL);
2360                 if (IS_ERR(handle))
2361                         GOTO(cleanup, rc = PTR_ERR(handle));
2362                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2363                 if (rc)
2364                         GOTO(cleanup, rc);
2365         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2366                 /* name exists adn points to remove inode */
2367                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2368                        obd->obd_name, rec->ur_tgt,
2369                        (unsigned long) de_new->d_mdsnum,
2370                        (unsigned long) de_new->d_inum,
2371                        (unsigned long) de_new->d_generation);
2372         } else {
2373                 /* name doesn't exist. the simplest case */
2374                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2375                                       FSFILT_OP_LINK, NULL);
2376                 if (IS_ERR(handle))
2377                         GOTO(cleanup, rc = PTR_ERR(handle));
2378         }
2379
2380         cleanup_phase = 2;
2381         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2382                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2383                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2384         if (rc)
2385                 CERROR("add_dir_entry() returned error %d\n", rc);
2386 cleanup:
2387         EXIT;
2388         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2389                                 handle, req, rc, 0);
2390         switch(cleanup_phase) {
2391                 case 2:
2392                 case 1:
2393 #ifdef S_PDIROPS
2394                         if (parent_lockh[1].cookie != 0)
2395                                 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2396 #endif
2397                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2398                         if (child_lockh.cookie != 0)
2399                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2400                         l_dput(de_new);
2401                         l_dput(de_srcdir);
2402                         break;
2403                 default:
2404                         LBUG();
2405         }
2406
2407         req->rq_status = rc;
2408
2409         RETURN(0);
2410 }
2411
2412 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2413                                       struct ptlrpc_request *req)
2414 {
2415         struct obd_device *obd = req->rq_export->exp_obd;
2416         struct ptlrpc_request *req2 = NULL;
2417         struct dentry *de_srcdir = NULL;
2418         struct dentry *de_old = NULL;
2419         struct mds_obd *mds = mds_req2mds(req);
2420         struct lustre_handle parent_lockh[2];
2421         struct lustre_handle child_lockh;
2422         struct mdc_op_data opdata;
2423         int cleanup_phase = 0;
2424         void *handle = NULL;
2425         int rc = 0;
2426         ENTRY;
2427
2428         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2429                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2430         memset(&opdata, 0, sizeof(opdata));
2431
2432         child_lockh.cookie = 0;
2433         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2434                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2435                                          rec->ur_name, rec->ur_namelen,
2436                                          &child_lockh, &de_old, LCK_EX,
2437                                          MDS_INODELOCK_LOOKUP);
2438         LASSERT(rc == 0);
2439         LASSERT(de_srcdir);
2440         LASSERT(de_srcdir->d_inode);
2441         LASSERT(de_old);
2442
2443         /* we already know the target should be created on another MDS
2444          * so, we have to request that MDS to do it */
2445
2446         /* prepare source fid */
2447         if (de_old->d_flags & DCACHE_CROSS_REF) {
2448                 LASSERT(de_old->d_inode == NULL);
2449                 CDEBUG(D_OTHER, "request to move remote name\n");
2450                 opdata.fid1.mds = de_old->d_mdsnum;
2451                 opdata.fid1.id = de_old->d_inum;
2452                 opdata.fid1.generation = de_old->d_generation;
2453         } else if (de_old->d_inode == NULL) {
2454                 /* oh, source doesn't exist */
2455                 GOTO(cleanup, rc = -ENOENT);
2456         } else {
2457                 LASSERT(de_old->d_inode != NULL);
2458                 CDEBUG(D_OTHER, "request to move local name\n");
2459                 opdata.fid1.mds = mds->mds_num;
2460                 opdata.fid1.id = de_old->d_inode->i_ino;
2461                 opdata.fid1.generation = de_old->d_inode->i_generation;
2462         }
2463
2464         opdata.fid2 = *(rec->ur_fid2);
2465         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2466                        rec->ur_tgtlen - 1, &req2);
2467
2468         if (rc)
2469                 GOTO(cleanup, rc);
2470
2471         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2472         if (IS_ERR(handle))
2473                 GOTO(cleanup, rc = PTR_ERR(handle));
2474         rc = fsfilt_del_dir_entry(obd, de_old);
2475         d_drop(de_old);
2476
2477 cleanup:
2478         EXIT;
2479         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2480                                 handle, req, rc, 0);
2481         if (req2)
2482                 ptlrpc_req_finished(req2);
2483
2484 #ifdef S_PDIROPS
2485         if (parent_lockh[1].cookie != 0)
2486                 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2487 #endif
2488         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2489         if (child_lockh.cookie != 0)
2490                 ldlm_lock_decref(&child_lockh, LCK_EX);
2491
2492         l_dput(de_old);
2493         l_dput(de_srcdir);
2494
2495         req->rq_status = rc;
2496         RETURN(0);
2497
2498 }
2499
2500 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2501                             struct ptlrpc_request *req,
2502                             struct lustre_handle *lockh)
2503 {
2504         struct obd_device *obd = req->rq_export->exp_obd;
2505         struct dentry *de_srcdir = NULL;
2506         struct dentry *de_tgtdir = NULL;
2507         struct dentry *de_old = NULL;
2508         struct dentry *de_new = NULL;
2509         struct mds_obd *mds = mds_req2mds(req);
2510         struct lustre_handle dlm_handles[7];
2511         struct mds_body *body = NULL;
2512         int rc = 0, lock_count = 3;
2513         int cleanup_phase = 0;
2514         void *handle = NULL;
2515         ENTRY;
2516
2517         LASSERT(offset == 0);
2518
2519         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2520                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2521                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2522
2523         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2524
2525         if (rec->ur_namelen == 1) {
2526                 rc = mds_reint_rename_create_name(rec, offset, req);
2527                 RETURN(rc);
2528         }
2529
2530         if (rec->ur_fid2->mds != mds->mds_num) {
2531                 rc = mds_reint_rename_to_remote(rec, offset, req);
2532                 RETURN(rc);
2533         }
2534
2535         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2536                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2537                                              rec->ur_name, rec->ur_namelen,
2538                                              &de_old, rec->ur_tgt,
2539                                              rec->ur_tgtlen, &de_new,
2540                                              dlm_handles, LCK_EX);
2541         if (rc)
2542                 GOTO(cleanup, rc);
2543
2544         cleanup_phase = 1; /* parent(s), children, locks */
2545
2546         if (de_new->d_inode)
2547                 lock_count = 4;
2548
2549         /* sanity check for src inode */
2550         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2551             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2552                 GOTO(cleanup, rc = -EINVAL);
2553
2554         /* sanity check for dest inode */
2555         if (de_new->d_inode &&
2556             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2557              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2558                 GOTO(cleanup, rc = -EINVAL);
2559
2560         if (de_old->d_inode == de_new->d_inode) {
2561                 GOTO(cleanup, rc = 0);
2562         }
2563
2564         /* if we are about to remove the target at first, pass the EA of
2565          * that inode to client to perform and cleanup on OST */
2566         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2567         LASSERT(body != NULL);
2568
2569         if (de_new->d_inode &&
2570             S_ISREG(de_new->d_inode->i_mode) &&
2571             de_new->d_inode->i_nlink == 1 &&
2572             mds_open_orphan_count(de_new->d_inode) == 0) {
2573                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2574                 mds_pack_inode2body(obd, body, de_new->d_inode);
2575                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2576                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2577                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2578                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2579                 } else {
2580                         /* XXX need log unlink? */
2581                 }
2582         }
2583
2584         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2585                        de_srcdir->d_inode->i_sb);
2586
2587         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2588         if (IS_ERR(handle))
2589                 GOTO(cleanup, rc = PTR_ERR(handle));
2590
2591         /* FIXME need adjust the journal block count? */
2592         /* if the target should be moved to PENDING, we at first increase the
2593          * link and later vfs_rename() will decrease the link count again */
2594         if (de_new->d_inode &&
2595             S_ISREG(de_new->d_inode->i_mode) &&
2596             de_new->d_inode->i_nlink == 1 &&
2597             mds_open_orphan_count(de_new->d_inode) > 0) {
2598                 rc = mds_add_link_orphan(rec, obd, de_new);
2599                 if (rc)
2600                         GOTO(cleanup, rc);
2601         }
2602
2603         lock_kernel();
2604         de_old->d_fsdata = req;
2605         de_new->d_fsdata = req;
2606         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2607         unlock_kernel();
2608
2609         GOTO(cleanup, rc);
2610 cleanup:
2611         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2612                                 handle, req, rc, 0);
2613         switch (cleanup_phase) {
2614         case 1:
2615 #ifdef S_PDIROPS
2616                 if (dlm_handles[5].cookie != 0)
2617                         ldlm_lock_decref(&(dlm_handles[5]), LCK_CW);
2618                 if (dlm_handles[6].cookie != 0)
2619                         ldlm_lock_decref(&(dlm_handles[6]), LCK_CW);
2620 #endif
2621                 if (rc) {
2622                         if (lock_count == 4)
2623                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2624                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2625                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2626                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2627                 } else {
2628                         if (lock_count == 4)
2629                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2630                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2631                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2632                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2633                 }
2634                 l_dput(de_new);
2635                 l_dput(de_old);
2636                 l_dput(de_tgtdir);
2637                 l_dput(de_srcdir);
2638         case 0:
2639                 break;
2640         default:
2641                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2642                 LBUG();
2643         }
2644         req->rq_status = rc;
2645         return 0;
2646 }
2647
2648 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2649                            struct ptlrpc_request *, struct lustre_handle *);
2650
2651 static mds_reinter reinters[REINT_MAX + 1] = {
2652         [REINT_SETATTR] mds_reint_setattr,
2653         [REINT_CREATE] mds_reint_create,
2654         [REINT_LINK] mds_reint_link,
2655         [REINT_UNLINK] mds_reint_unlink,
2656         [REINT_RENAME] mds_reint_rename,
2657         [REINT_OPEN] mds_open
2658 };
2659
2660 int mds_reint_rec(struct mds_update_record *rec, int offset,
2661                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2662 {
2663         struct obd_device *obd = req->rq_export->exp_obd;
2664         struct obd_run_ctxt saved;
2665         int rc;
2666
2667         /* checked by unpacker */
2668         LASSERT(rec->ur_opcode <= REINT_MAX &&
2669                 reinters[rec->ur_opcode] != NULL);
2670
2671         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2672         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2673         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2674
2675         return rc;
2676 }