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/jbd.h>
  35 #include <linux/namei.h>
  36 #include <linux/ext3_fs.h>
  37 #include <linux/obd_support.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/obd.h>
  40 #include <linux/lustre_lib.h>
  41 #include <linux/lustre_idl.h>
  42 #include <linux/lustre_mds.h>
  43 #include <linux/lustre_dlm.h>
  44 #include <linux/lustre_log.h>
  45 #include <linux/lustre_fsfilt.h>
  46 #include <linux/lustre_acl.h>
  47 #include <linux/lustre_lite.h>
  48 #include <linux/lustre_smfs.h>
  49 #include "mds_internal.h"
  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 static void mds_cancel_cookies_cb(struct obd_device *obd,
  60                                   __u64 transno, void *cb_data,
  61                                   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_dt_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, mlcd->mlcd_cookielen /
  84                                  sizeof(*mlcd->mlcd_cookies),
  85                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
  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                 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
  91         }
  92
  93         OBD_FREE(mlcd, mlcd->mlcd_size);
  94 }
  95
  96 /* Assumes caller has already pushed us into the kernel context. */
  97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  98                        struct ptlrpc_request *req, int rc, __u32 op_data)
  99 {
 100         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 101         struct obd_device *obd = req->rq_export->exp_obd;
 102         struct mds_client_data *mcd = med->med_mcd;
 103         int err, log_pri = D_HA;
 104         __u64 transno;
 105         loff_t off;
 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(out_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                 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
 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         if (req->rq_reqmsg->opc == MDS_CLOSE) {
 146                 mcd->mcd_last_close_transno = cpu_to_le64(transno);
 147                 mcd->mcd_last_close_xid = cpu_to_le64(req->rq_xid);
 148                 mcd->mcd_last_close_result = cpu_to_le32(rc);
 149                 mcd->mcd_last_close_data = cpu_to_le32(op_data);
 150         } else {
 151                 mcd->mcd_last_transno = cpu_to_le64(transno);
 152                 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 153                 mcd->mcd_last_result = cpu_to_le32(rc);
 154                 mcd->mcd_last_data = cpu_to_le32(op_data);
 155         }
 156
 157         fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
 158                               mds_commit_last_transno_cb, NULL);
 159
 160         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
 161                                   sizeof(*mcd), &off, 0);
 162
 163         if (err) {
 164                 log_pri = D_ERROR;
 165                 if (rc == 0)
 166                         rc = err;
 167         }
 168
 169         DEBUG_REQ(log_pri, req,
 170                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 171                   transno, mcd->mcd_uuid, med->med_idx, err);
 172
 173         err = mds_update_last_fid(obd, handle, 0);
 174         if (err) {
 175                 log_pri = D_ERROR;
 176                 if (rc == 0)
 177                         rc = err;
 178         }
 179
 180         err = mds_dt_write_objids(obd);
 181         if (err) {
 182                 log_pri = D_ERROR;
 183                 if (rc == 0)
 184                         rc = err;
 185         }
 186         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 187
 188         EXIT;
 189 out_commit:
 190         err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
 191                             req->rq_export->exp_sync);
 192         if (err) {
 193                 CERROR("error committing transaction: %d\n", err);
 194                 if (!rc)
 195                         rc = err;
 196         }
 197
 198         return rc;
 199 }
 200
 201 /* this gives the same functionality as the code between
 202  * sys_chmod and inode_setattr
 203  * chown_common and inode_setattr
 204  * utimes and inode_setattr
 205  */
 206 #ifndef ATTR_RAW
 207 /* Just for the case if we have some clients that know about ATTR_RAW */
 208 #define ATTR_RAW 8192
 209 #endif
 210 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 211 {
 212         time_t now = LTIME_S(CURRENT_TIME);
 213         struct iattr *attr = &rec->ur_iattr;
 214         unsigned int ia_valid = attr->ia_valid;
 215         int error;
 216         ENTRY;
 217
 218         /* only fix up attrs if the client VFS didn't already */
 219
 220         if (!(ia_valid & ATTR_RAW))
 221                 RETURN(0);
 222
 223         if (!(ia_valid & ATTR_CTIME_SET))
 224                 LTIME_S(attr->ia_ctime) = now;
 225         if (!(ia_valid & ATTR_ATIME_SET))
 226                 LTIME_S(attr->ia_atime) = now;
 227         if (!(ia_valid & ATTR_MTIME_SET))
 228                 LTIME_S(attr->ia_mtime) = now;
 229
 230         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 231                 RETURN(-EPERM);
 232
 233         /* times */
 234         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 235                 if (rec->ur_fsuid != inode->i_uid &&
 236                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 237                         RETURN(error);
 238         }
 239
 240         if (ia_valid & ATTR_SIZE) {
 241                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 242                         RETURN(error);
 243         }
 244
 245         if (ia_valid & ATTR_UID) {
 246                 /* chown */
 247                 error = -EPERM;
 248                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 249                         RETURN(-EPERM);
 250                 if (attr->ia_uid == (uid_t) -1)
 251                         attr->ia_uid = inode->i_uid;
 252                 if (attr->ia_gid == (gid_t) -1)
 253                         attr->ia_gid = inode->i_gid;
 254                 attr->ia_mode = inode->i_mode;
 255                 /*
 256                  * If the user or group of a non-directory has been
 257                  * changed by a non-root user, remove the setuid bit.
 258                  * 19981026 David C Niemi <niemi@tux.org>
 259                  *
 260                  * Changed this to apply to all users, including root,
 261                  * to avoid some races. This is the behavior we had in
 262                  * 2.0. The check for non-root was definitely wrong
 263                  * for 2.2 anyway, as it should have been using
 264                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 265                  */
 266                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 267                     !S_ISDIR(inode->i_mode)) {
 268                         attr->ia_mode &= ~S_ISUID;
 269                         attr->ia_valid |= ATTR_MODE;
 270                 }
 271                 /*
 272                  * Likewise, if the user or group of a non-directory
 273                  * has been changed by a non-root user, remove the
 274                  * setgid bit UNLESS there is no group execute bit
 275                  * (this would be a file marked for mandatory
 276                  * locking).  19981026 David C Niemi <niemi@tux.org>
 277                  *
 278                  * Removed the fsuid check (see the comment above) --
 279                  * 19990830 SD.
 280                  */
 281                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 282                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 283                         attr->ia_mode &= ~S_ISGID;
 284                         attr->ia_valid |= ATTR_MODE;
 285                 }
 286         } else if (ia_valid & ATTR_MODE) {
 287                 int mode = attr->ia_mode;
 288                 /* chmod */
 289                 if (attr->ia_mode == (mode_t) -1)
 290                         attr->ia_mode = inode->i_mode;
 291                 attr->ia_mode =
 292                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 293         }
 294         RETURN(0);
 295 }
 296
 297 void mds_steal_ack_locks(struct ptlrpc_request *req)
 298 {
 299         struct obd_export         *exp = req->rq_export;
 300         char                       str[PTL_NALFMT_SIZE];
 301         struct list_head          *tmp;
 302         struct ptlrpc_reply_state *oldrep;
 303         struct ptlrpc_service     *svc;
 304         struct llog_create_locks  *lcl;
 305         unsigned long              flags;
 306         int                        i;
 307
 308         /* CAVEAT EMPTOR: spinlock order */
 309         spin_lock_irqsave (&exp->exp_lock, flags);
 310         list_for_each (tmp, &exp->exp_outstanding_replies) {
 311                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 312
 313                 if (oldrep->rs_xid != req->rq_xid)
 314                         continue;
 315
 316                 if (oldrep->rs_msg->opc != req->rq_reqmsg->opc)
 317                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 318                                 "new %d old %d\n", req->rq_xid,
 319                                 req->rq_reqmsg->opc, oldrep->rs_msg->opc);
 320
 321                 svc = oldrep->rs_srv_ni->sni_service;
 322                 spin_lock (&svc->srv_lock);
 323
 324                 list_del_init (&oldrep->rs_exp_list);
 325
 326                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 327                       " o%d NID %s\n", oldrep->rs_nlocks, oldrep,
 328                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg->opc,
 329                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 330
 331                 for (i = 0; i < oldrep->rs_nlocks; i++)
 332                         ptlrpc_save_lock(req,
 333                                          &oldrep->rs_locks[i],
 334                                          oldrep->rs_modes[i]);
 335                 oldrep->rs_nlocks = 0;
 336
 337                 lcl = oldrep->rs_llog_locks;
 338                 oldrep->rs_llog_locks = NULL;
 339                 if (lcl != NULL)
 340                         ptlrpc_save_llog_lock(req, lcl);
 341
 342                 DEBUG_REQ(D_HA, req, "stole locks for");
 343                 ptlrpc_schedule_difficult_reply (oldrep);
 344
 345                 spin_unlock (&svc->srv_lock);
 346                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 347                 return;
 348         }
 349         spin_unlock_irqrestore (&exp->exp_lock, flags);
 350 }
 351
 352 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 353 {
 354         if (req->rq_reqmsg->opc == MDS_CLOSE) {
 355                 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 356                           mcd->mcd_last_close_transno, mcd->mcd_last_close_result);
 357                 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_close_transno;
 358                 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_close_result;
 359         } else {
 360                 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 361                           mcd->mcd_last_transno, mcd->mcd_last_result);
 362                 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 363                 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 364         }
 365
 366         mds_steal_ack_locks(req);
 367 }
 368
 369 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 370                                       int offset, struct ptlrpc_request *req)
 371 {
 372         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 373         struct mds_body *body;
 374         struct dentry *de;
 375
 376         mds_req_from_mcd(req, med->med_mcd);
 377
 378         de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
 379         if (IS_ERR(de)) {
 380                 LASSERT(PTR_ERR(de) == req->rq_status);
 381                 return;
 382         }
 383
 384         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 385         mds_pack_inode2body(req2obd(req), body, de->d_inode, 1);
 386
 387         /* Don't return OST-specific attributes if we didn't just set them */
 388         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 389                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 390         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 391                 body->valid |= OBD_MD_FLMTIME;
 392         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 393                 body->valid |= OBD_MD_FLATIME;
 394
 395         l_dput(de);
 396 }
 397
 398 static int mds_reint_remote_setfacl(struct obd_device *obd,
 399                                     struct mds_export_data *med,
 400                                     struct mds_update_record *rec,
 401                                     struct ptlrpc_request *req)
 402 {
 403         struct rmtacl_upcall_desc desc;
 404         struct dentry   *de;
 405         struct inode    *inode;
 406         struct mds_body *body;
 407         int              rc = 0;
 408         int              repsize[2] = { sizeof(*body), LUSTRE_ACL_SIZE_MAX };
 409         ENTRY;
 410
 411         rc = lustre_pack_reply(req, 2, repsize, NULL);
 412         if (rc)
 413                 RETURN(rc);
 414
 415         de = mds_id2dentry(obd, rec->ur_id1, NULL);
 416         if (IS_ERR(de))
 417                 GOTO(out, rc = PTR_ERR(de));
 418
 419         inode = de->d_inode;
 420         LASSERT(inode);
 421
 422         /* setxattr from remote client:
 423          */
 424         memset(&desc, 0, sizeof(desc));
 425         desc.cmd = (char *) rec->ur_ea2data;
 426         desc.cmdlen = rec->ur_ea2datalen;
 427         desc.res = lustre_msg_buf(req->rq_repmsg, 1, LUSTRE_ACL_SIZE_MAX);
 428         desc.reslen = LUSTRE_ACL_SIZE_MAX;
 429
 430         mds_do_remote_acl_upcall(&desc);
 431         if (desc.upcall_status)
 432                 GOTO(out_put, rc = desc.upcall_status);
 433
 434         if (desc.status < 0)
 435                 desc.status = -desc.status;
 436
 437         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 438         LASSERT(body);
 439
 440         /* client (lmv) will do limited checking upon replied mds_body,
 441          * we pack it as normal, but "steal" field "flags" field to store
 442          * the acl execution status.
 443          */
 444         mds_pack_inode2body(obd, body, inode, 1);
 445         body->flags = desc.status;
 446         mds_body_do_reverse_map(med, body);
 447
 448         EXIT;
 449 out_put:
 450         l_dput(de);
 451 out:
 452         req->rq_status = rc;
 453         return 0;
 454 }
 455
 456 static int mds_get_md_type(char *name)
 457 {
 458         if (!strcmp(name, XATTR_LUSTRE_MDS_LOV_EA))
 459                 RETURN(EA_LOV);
 460         if (!strcmp(name, XATTR_LUSTRE_MDS_MEA_EA))
 461                 RETURN(EA_MEA);
 462         if (!strcmp(name, XATTR_LUSTRE_MDS_SID_EA))
 463                 RETURN(EA_SID);
 464         if (!strcmp(name, XATTR_LUSTRE_MDS_PID_EA))
 465                 RETURN(EA_PID);
 466         if (!strcmp(name, XATTR_LUSTRE_MDS_KEY_EA))
 467                 RETURN(EA_KEY);
 468
 469         RETURN(0);
 470 }
 471
 472 /* In the raw-setattr case, we lock the child inode.
 473  * In the write-back case or if being called from open, the client holds a lock
 474  * already.
 475  *
 476  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 477 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 478                              struct ptlrpc_request *req, struct lustre_handle *lh)
 479 {
 480         struct mds_obd *mds = mds_req2mds(req);
 481         struct obd_device *obd = req->rq_export->exp_obd;
 482         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 483         struct mds_body *body;
 484         struct dentry *de = NULL;
 485         struct inode *inode = NULL;
 486         struct lustre_handle lockh[2] = {{0}, {0}};
 487         int parent_mode;
 488         void *handle = NULL;
 489         struct mds_logcancel_data *mlcd = NULL;
 490         int rc = 0, cleanup_phase = 0, err;
 491         int repcnt = 1, repsize[2] = { sizeof(*body) };
 492         int locked = 0, do_trunc = 0;
 493         ENTRY;
 494
 495         LASSERT(offset == 1);
 496
 497         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
 498                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
 499                   rec->ur_iattr.ia_valid);
 500
 501         /* remote setfacl need special handling */
 502         if ((rec->ur_iattr.ia_valid & ATTR_EA) &&
 503             !strcmp(rec->ur_eadata, XATTR_NAME_LUSTRE_ACL)) {
 504                 return mds_reint_remote_setfacl(obd, med, rec, req);
 505         }
 506
 507         if (rec->ur_iattr.ia_valid & ATTR_SIZE) {
 508                 repsize[repcnt++] = sizeof(struct lustre_capa);
 509                 do_trunc = 1; /* XXX: ATTR_SIZE will be cleared from ia_valid */
 510         }
 511
 512         rc = lustre_pack_reply(req, repcnt, repsize, NULL);
 513         if (rc)
 514                 RETURN(rc);
 515
 516         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 517         MD_COUNTER_INCREMENT(obd, setattr);
 518
 519         if (med->med_remote) {
 520                 if (rec->ur_iattr.ia_valid & ATTR_GID) {
 521                         CWARN("Deny chgrp from remote client\n");
 522                         GOTO(cleanup, rc = -EPERM);
 523                 }
 524                 if (rec->ur_iattr.ia_valid & ATTR_UID) {
 525                         uid_t uid;
 526
 527                         uid = mds_idmap_lookup_uid(med->med_idmap, 0,
 528                                                    rec->ur_iattr.ia_uid);
 529                         if (uid == MDS_IDMAP_NOTFOUND) {
 530                                 CWARN("Deny chown to uid %u\n",
 531                                       rec->ur_iattr.ia_uid);
 532                                 GOTO(cleanup, rc = -EPERM);
 533                         }
 534                         rec->ur_iattr.ia_uid = uid;
 535                 }
 536         }
 537
 538         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 539                 de = mds_id2dentry(obd, rec->ur_id1, NULL);
 540                 if (IS_ERR(de))
 541                         GOTO(cleanup, rc = PTR_ERR(de));
 542         } else {
 543                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 544                 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
 545                         lockpart |= MDS_INODELOCK_LOOKUP;
 546                 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
 547                                           lockh, &parent_mode, NULL, 0, lockpart);
 548                 if (IS_ERR(de))
 549                         GOTO(cleanup, rc = PTR_ERR(de));
 550                 locked = 1;
 551         }
 552
 553         cleanup_phase = 1;
 554
 555         inode = de->d_inode;
 556         LASSERT(inode);
 557         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 558             rec->ur_eadata != NULL)
 559                 down(&inode->i_sem);
 560
 561         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 562
 563         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 564         if (IS_ERR(handle))
 565                 GOTO(cleanup, rc = PTR_ERR(handle));
 566
 567         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 568                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 569                        LTIME_S(rec->ur_iattr.ia_mtime),
 570                        LTIME_S(rec->ur_iattr.ia_ctime));
 571         rc = mds_fix_attr(inode, rec);
 572         if (rc)
 573                 GOTO(cleanup, rc);
 574
 575         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 576                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 577                                       (long)&rec->ur_iattr.ia_attr_flags);
 578         else                                            /* setattr */
 579                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 580
 581         if (rc == 0) {
 582                 if (rec->ur_iattr.ia_valid & ATTR_EA) {
 583                         int flags = (int) rec->ur_iattr.ia_attr_flags;
 584
 585                         rc = -EOPNOTSUPP;
 586                         if (!med->med_remote && inode->i_op &&
 587                             inode->i_op->setxattr)
 588                                 rc = inode->i_op->setxattr(
 589                                                 de, rec->ur_eadata,
 590                                                 rec->ur_ea2data,
 591                                                 rec->ur_ea2datalen,
 592                                                 flags);
 593                 } else if (rec->ur_iattr.ia_valid & ATTR_EA_RM) {
 594                         rc = -EOPNOTSUPP;
 595                         if (inode->i_op && inode->i_op->removexattr)
 596                                 rc = inode->i_op->removexattr(de,
 597                                                   rec->ur_eadata);
 598                 } else if (rec->ur_iattr.ia_valid & ATTR_EA_CMOBD) {
 599                         char *name;
 600                         int type;
 601
 602                         LASSERT(rec->ur_eadata != NULL);
 603                         LASSERT(rec->ur_ea2data != NULL);
 604                         name = rec->ur_eadata;
 605
 606                         /* XXX: tmp fix for setting LOV EA from CMOBD */
 607                         type = mds_get_md_type(name);
 608
 609                         if (type == EA_LOV) {
 610                                 CDEBUG(D_INFO, "set %s EA for cmobd\n", name);
 611
 612                                 rc = fsfilt_set_md(obd, inode, handle,
 613                                                    rec->ur_ea2data,
 614                                                    rec->ur_ea2datalen,
 615                                                    type);
 616                                 if (rc) {
 617                                         CERROR("fsfilt_set_md() failed, err %d\n",
 618                                                rc);
 619                                         GOTO(cleanup, rc);
 620                                 }
 621                         }
 622                 } else if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 623                            !((rec->ur_iattr.ia_valid & ATTR_KEY) ||
 624                              (rec->ur_iattr.ia_valid & ATTR_MAC))) {
 625
 626                         struct lov_stripe_md *lsm = NULL;
 627                         struct lov_user_md *lum = NULL;
 628
 629                         if (rec->ur_eadata != NULL) {
 630                                 rc = ll_permission(inode, MAY_WRITE, NULL);
 631                                 if (rc < 0)
 632                                         GOTO(cleanup, rc);
 633
 634                                 lum = rec->ur_eadata;
 635
 636                                 /* if lmm_stripe_size is -1 delete default
 637                                  * stripe from dir */
 638                                 if (S_ISDIR(inode->i_mode) &&
 639                                     lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1)){
 640                                         rc = fsfilt_set_md(obd, inode, handle, NULL, 0, EA_LOV);
 641                                         if (rc)
 642                                                 GOTO(cleanup, rc);
 643                                 } else {
 644                                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 645                                                            mds->mds_dt_exp, 0,
 646                                                            &lsm, rec->ur_eadata);
 647                                         if (rc)
 648                                                 GOTO(cleanup, rc);
 649
 650                                         obd_free_memmd(mds->mds_dt_exp, &lsm);
 651                                         rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 652                                                            rec->ur_eadatalen, EA_LOV);
 653                                         if (rc)
 654                                                 GOTO(cleanup, rc);
 655                                 }
 656                         }
 657                 }
 658
 659                 if ((rec->ur_iattr.ia_valid & ATTR_KEY) || (rec->ur_iattr.ia_valid & ATTR_MAC)) {
 660                         void *key;
 661                         int keylen;
 662
 663                         LASSERT(rec->ur_eadatalen || rec->ur_ea3datalen);
 664                         LASSERT(rec->ur_eadata || rec->ur_ea3data);
 665                         key = rec->ur_ea3data ? rec->ur_ea3data : rec->ur_eadata;
 666                         keylen = rec->ur_ea3datalen ? rec->ur_ea3datalen :
 667                                                      rec->ur_eadatalen;
 668                         mds_set_gskey(obd, handle, inode, key, keylen,
 669                                       rec->ur_iattr.ia_valid);
 670                 }
 671         }
 672
 673         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 674         mds_pack_inode2body(obd, body, inode, 1);
 675
 676         /* Don't return OST-specific attributes if we didn't just set them */
 677         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 678                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 679         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 680                 body->valid |= OBD_MD_FLMTIME;
 681         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 682                 body->valid |= OBD_MD_FLATIME;
 683
 684         if (do_trunc) {
 685                 struct lustre_capa capa = {
 686                         .lc_uid   = rec->ur_uc.luc_uid,
 687                         .lc_op    = CAPA_TRUNC,
 688                         .lc_ino   = inode->i_ino,
 689                         .lc_mdsid = mds->mds_num,
 690                 };
 691                 int offset = 1;
 692
 693                 LASSERT(capa.lc_mdsid == mds->mds_num);
 694                 rc = mds_pack_capa(obd, med, NULL, &capa, req,
 695                                    &offset, body);
 696                 if (rc < 0) {
 697                         CERROR("mds_pack_capa: rc = %d\n", rc);
 698                         RETURN(rc);
 699                 }
 700         }
 701
 702         mds_body_do_reverse_map(med, body);
 703
 704         /* The logcookie should be no use anymore, why nobody remove
 705          * following code block?
 706          */
 707         LASSERT(rec->ur_cookielen == 0);
 708         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_dt_obd)) {
 709                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 710                           rec->ur_eadatalen);
 711                 if (mlcd) {
 712                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 713                                 rec->ur_eadatalen;
 714                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 715                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 716                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 717                                 mlcd->mlcd_cookielen;
 718                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 719                                mlcd->mlcd_cookielen);
 720                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 721                                mlcd->mlcd_eadatalen);
 722                 } else {
 723                         CERROR("unable to allocate log cancel data\n");
 724                 }
 725         }
 726         EXIT;
 727  cleanup:
 728         if (mlcd != NULL)
 729                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 730                                       handle, mds_cancel_cookies_cb, mlcd);
 731         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 732         switch (cleanup_phase) {
 733         case 1:
 734                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 735                     rec->ur_eadata != NULL)
 736                         up(&inode->i_sem);
 737                 l_dput(de);
 738                 if (locked) {
 739 #ifdef S_PDIROPS
 740                         if (lockh[1].cookie != 0)
 741                                 ldlm_lock_decref(lockh + 1, parent_mode);
 742 #endif
 743                         if (rc) {
 744                                 ldlm_lock_decref(lockh, LCK_PW);
 745                         } else {
 746                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 747                         }
 748                 }
 749         case 0:
 750                 break;
 751         default:
 752                 LBUG();
 753         }
 754         if (err && !rc)
 755                 rc = err;
 756
 757         req->rq_status = rc;
 758         return 0;
 759 }
 760
 761 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 762                                      struct ptlrpc_request *req)
 763 {
 764         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 765         struct dentry *parent, *child;
 766         struct mds_body *body;
 767         ENTRY;
 768
 769         mds_req_from_mcd(req, med->med_mcd);
 770
 771         if (req->rq_status) {
 772                 EXIT;
 773                 return;
 774         }
 775
 776         parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
 777         LASSERT(!IS_ERR(parent));
 778         child = ll_lookup_one_len(rec->ur_name, parent,
 779                                   rec->ur_namelen - 1);
 780         LASSERT(!IS_ERR(child));
 781         if ((child->d_flags & DCACHE_CROSS_REF)) {
 782                 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
 783                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 784                 mds_pack_dentry2body(req2obd(req), body, child, 1);
 785         } else if (child->d_inode == NULL) {
 786                 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
 787                           OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
 788                 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
 789         } else {
 790                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 791                 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
 792         }
 793         l_dput(parent);
 794         l_dput(child);
 795         EXIT;
 796 }
 797
 798 static int mds_get_default_acl(struct inode *dir, void **pacl)
 799 {
 800         struct dentry de = { .d_inode = dir };
 801         int size, size2;
 802
 803         LASSERT(S_ISDIR(dir->i_mode));
 804
 805         if (!dir->i_op->getxattr)
 806                 return 0;
 807
 808         size = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, NULL, 0);
 809         if (size == 0 || size == -ENODATA || size == -EOPNOTSUPP)
 810                 return 0;
 811         else if (size < 0)
 812                 return size;
 813
 814         OBD_ALLOC(*pacl, size);
 815         if (!*pacl)
 816                 return -ENOMEM;
 817
 818         size2 = dir->i_op->getxattr(&de, XATTR_NAME_ACL_DEFAULT, *pacl, size);
 819         if (size2 != size) {
 820                 /* since we already locked the dir, it should not change
 821                  * between the 2 getxattr calls
 822                  */
 823                 CERROR("2'nd getxattr got %d, expect %d\n", size2, size);
 824                 OBD_FREE(*pacl, size);
 825                 return -EIO;
 826         }
 827
 828         return size;
 829 }
 830
 831 static int mds_reint_create(struct mds_update_record *rec, int offset,
 832                             struct ptlrpc_request *req,
 833                             struct lustre_handle *lh)
 834 {
 835         struct dentry *dparent = NULL;
 836         struct mds_obd *mds = mds_req2mds(req);
 837         struct obd_device *obd = req->rq_export->exp_obd;
 838         struct mds_body *body = NULL;
 839         struct dentry *dchild = NULL;
 840         struct inode *dir = NULL;
 841         void *handle = NULL;
 842         struct lustre_handle lockh[2] = {{0}, {0}};
 843         int parent_mode;
 844         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 845         int created = 0;
 846         struct dentry_params dp;
 847         struct mea *mea = NULL;
 848         int mea_size;
 849         struct lustre_id sid;
 850         __u64 fid;
 851         ENTRY;
 852
 853         LASSERT(offset == 1);
 854
 855         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
 856                         OBD_MDS_DEVICENAME));
 857
 858         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 859                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
 860                   rec->ur_name, rec->ur_mode);
 861
 862         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 863
 864         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 865                 GOTO(cleanup, rc = -ESTALE);
 866
 867         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
 868                                        lockh, &parent_mode, rec->ur_name,
 869                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 870         if (IS_ERR(dparent)) {
 871                 rc = PTR_ERR(dparent);
 872                 CERROR("parent lookup error %d, id "DLID4"\n",
 873                        rc, OLID4(rec->ur_id1));
 874                 GOTO(cleanup, rc);
 875         }
 876         cleanup_phase = 1; /* locked parent dentry */
 877         dir = dparent->d_inode;
 878         LASSERT(dir);
 879
 880         ldlm_lock_dump_handle(D_OTHER, lockh);
 881
 882         /* get parent id: ldlm lock on the parent protects ea */
 883         rc = mds_read_inode_sid(obd, dir, &sid);
 884         if (rc) {
 885                 CERROR("can't read parent id. ino(%lu) rc(%d)\n",
 886                        dir->i_ino, rc);
 887                 GOTO(cleanup, rc);
 888         }
 889
 890         /* try to retrieve MEA data for this dir */
 891         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
 892         if (rc)
 893                 GOTO(cleanup, rc);
 894
 895         if (mea != NULL && mea->mea_count) {
 896                 /*
 897                  * dir is already splitted, check is requested filename should
 898                  * live at this MDS or at another one.
 899                  */
 900                 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 901                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
 902                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 903                                " should be %lu(%d)\n",
 904                                mea->mea_master, dparent->d_inode->i_ino,
 905                                dparent->d_inode->i_generation, rec->ur_name,
 906                                (unsigned long)id_group(&mea->mea_ids[i]), i);
 907                         GOTO(cleanup, rc = -ERESTART);
 908                 }
 909         }
 910
 911         dchild = ll_lookup_one_len(rec->ur_name, dparent,
 912                                    rec->ur_namelen - 1);
 913         if (IS_ERR(dchild)) {
 914                 rc = PTR_ERR(dchild);
 915                 CERROR("Can't find "DLID4"/%s, error %d\n",
 916                        OLID4(rec->ur_id1), rec->ur_name, rc);
 917                 GOTO(cleanup, rc);
 918         }
 919
 920         cleanup_phase = 2; /* child dentry */
 921
 922         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 923
 924         if (type == S_IFREG || type == S_IFDIR) {
 925                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 926                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 927                        obd->obd_name, dparent->d_inode->i_ino,
 928                        dparent->d_inode->i_generation, rc, parent_mode);
 929                 if (rc > 0) {
 930                         /* dir got splitted */
 931                         GOTO(cleanup, rc = -ERESTART);
 932                 } else if (rc < 0) {
 933                         /* error happened during spitting. */
 934                         GOTO(cleanup, rc);
 935                 }
 936         }
 937
 938         if (dir->i_mode & S_ISGID) {
 939                 if (S_ISDIR(rec->ur_mode))
 940                         rec->ur_mode |= S_ISGID;
 941         }
 942
 943         /* for reint case stor ecookie should be zero */
 944         if (rec->ur_flags & MDS_REINT_REQ) {
 945                 LASSERT(id_ino(rec->ur_id1) == 0);
 946                 LASSERT(id_ino(rec->ur_id2) == 0);
 947         }
 948
 949         if (id_fid(rec->ur_id2))
 950                 fid = id_fid(rec->ur_id2);
 951         else
 952                 fid = mds_alloc_fid(obd);
 953
 954         dchild->d_fsdata = (void *)&dp;
 955         dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
 956         dp.p_ptr = req;
 957
 958         dp.p_fid = fid;
 959         dp.p_group = mds->mds_num;
 960
 961         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 962
 963         switch (type) {
 964         case S_IFREG: {
 965                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 966                 if (IS_ERR(handle))
 967                         GOTO(cleanup, rc = PTR_ERR(handle));
 968                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 969
 970                 /* XXX: here we should check what type of EA is in ur_eadata
 971                  * and do appropriate actions. --umka */
 972                 if (rec->ur_eadata && rec->ur_eadatalen &&
 973                     rc == 0 && dchild->d_inode != NULL) {
 974                     if (rec->ur_flags & MDS_REINT_REQ) {
 975                         /* for CMOBD to set lov md info when cmobd reint
 976                          * create */
 977                         CDEBUG(D_INFO, "set lsm %p, len %d to inode %lu \n",
 978                                rec->ur_eadata, rec->ur_eadatalen,
 979                                dchild->d_inode->i_ino);
 980                         rc = fsfilt_set_md(obd, dchild->d_inode, handle,
 981                                            rec->ur_eadata, rec->ur_eadatalen,
 982                                            EA_LOV);
 983                         if (rc) {
 984                                 CERROR("fsfilt_set_md() failed, err %d\n",
 985                                        rc);
 986                                 GOTO(cleanup, rc);
 987                         }
 988                     } else {
 989                         /* assumption: when ur_eadata is not NULL,
 990                          * ur_eadata is crypto key, should fix it later,
 991                          * --wangdi */
 992                         rc = mds_set_gskey(obd, handle, dchild->d_inode,
 993                                            rec->ur_eadata, rec->ur_eadatalen,
 994                                            ATTR_MAC | ATTR_KEY);
 995                         if (rc) {
 996                                 CWARN("mds_set_gskey() failed, err %d\n",
 997                                       rc);
 998                         }
 999                     }
1000                 }
1001                 break;
1002         }
1003         case S_IFDIR: {
1004                 int i;
1005
1006                 /*
1007                  * as Peter asked, mkdir() should distribute new directories
1008                  * over the whole cluster in order to distribute namespace
1009                  * processing load. first, we calculate which MDS to use to put
1010                  * new directory's inode in.
1011                  */
1012
1013                 /* XXX: here we order mds_choose_mdsnum() to use local mdsnum
1014                  * for reint requests. This should be gone when real flushing on
1015                  * LMV is fixed. --umka */
1016                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
1017                                       rec->ur_flags, &req->rq_peer, dir,
1018                                       (rec->ur_flags & MDS_REINT_REQ) ? 1 : 0);
1019
1020                 if (i == mds->mds_num) {
1021                         /* inode will be created locally */
1022                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1023                         if (IS_ERR(handle))
1024                                 GOTO(cleanup, rc = PTR_ERR(handle));
1025
1026                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
1027                         if (rc) {
1028                                 CDEBUG(D_OTHER,
1029                                        "Can't create dir \"%s\", rc = %d\n",
1030                                        dchild->d_name.name, rc);
1031                                 GOTO(cleanup, rc);
1032                         }
1033
1034                 } else if (!DENTRY_VALID(dchild)) {
1035                         /* inode will be created on another MDS */
1036                         struct obdo *oa = NULL;
1037                         void *acl = NULL;
1038                         int acl_size;
1039
1040                         /* first, create that inode */
1041                         oa = obdo_alloc();
1042                         if (!oa)
1043                                 GOTO(cleanup, rc = -ENOMEM);
1044
1045                         oa->o_mds = i;
1046                         oa->o_easize = 0;
1047
1048                         if (rec->ur_eadata) {
1049                                 /* user asks for creating splitted dir */
1050                                 oa->o_easize = *((u16 *) rec->ur_eadata);
1051                         }
1052
1053                         obdo_from_inode(oa, dir, OBD_MD_FLATIME |
1054                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1055
1056                         /* adjust the uid/gid/mode bits */
1057                         oa->o_mode = rec->ur_mode;
1058                         oa->o_uid = current->fsuid;
1059                         oa->o_gid = (dir->i_mode & S_ISGID) ?
1060                                      dir->i_gid : current->fsgid;
1061
1062                         /* letting remote MDS know that this is reint
1063                          * request. */
1064                         if (rec->ur_flags & MDS_REINT_REQ)
1065                                 oa->o_flags |= OBD_FL_REINT;
1066
1067                         /* transfer parent id to remote inode */
1068                         memcpy(obdo_id(oa), &sid, sizeof(sid));
1069                         oa->o_valid |= OBD_MD_FLTYPE | OBD_MD_FLUID |
1070                                        OBD_MD_FLGID | OBD_MD_FLIFID;
1071
1072                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
1073                                obd->obd_name, i);
1074
1075                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1076                                 /*
1077                                  * here inode number and generation are
1078                                  * important, as this is replay request and we
1079                                  * need them to check if such an object is
1080                                  * already created.
1081                                  */
1082                                 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
1083                                        obd->obd_name, rec->ur_namelen - 1,
1084                                        rec->ur_name, (unsigned)id_ino(rec->ur_id2),
1085                                        (unsigned)id_gen(rec->ur_id2));
1086                                 oa->o_id = id_ino(rec->ur_id2);
1087                                 oa->o_fid = id_fid(rec->ur_id2);
1088                                 oa->o_generation = id_gen(rec->ur_id2);
1089                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
1090                                 LASSERT(oa->o_fid != 0);
1091                         }
1092
1093                         /* obtain default ACL */
1094                         acl_size = mds_get_default_acl(dir, &acl);
1095                         if (acl_size < 0) {
1096                                 obdo_free(oa);
1097                                 GOTO(cleanup, rc = -ENOMEM);
1098                         }
1099
1100                         /*
1101                          * before obd_create() is called, o_fid is not known if
1102                          * this is not recovery of cause.
1103                          */
1104                         rc = obd_create(mds->mds_md_exp, oa, acl, acl_size,
1105                                         NULL, NULL);
1106
1107                         if (acl)
1108                                 OBD_FREE(acl, acl_size);
1109
1110                         if (rc) {
1111                                 CERROR("can't create remote inode: %d\n", rc);
1112                                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
1113                                           id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1114                                           rec->ur_name, rec->ur_mode);
1115                                 obdo_free(oa);
1116                                 GOTO(cleanup, rc);
1117                         }
1118
1119                         LASSERT(oa->o_fid != 0);
1120
1121                         /* now, add new dir entry for it */
1122                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1123                         if (IS_ERR(handle)) {
1124                                 obdo_free(oa);
1125                                 GOTO(cleanup, rc = PTR_ERR(handle));
1126                         }
1127
1128                         /* creating local dentry for remote inode. */
1129                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
1130                                                   rec->ur_namelen - 1, oa->o_id,
1131                                                   oa->o_generation, i, oa->o_fid);
1132
1133                         if (rc) {
1134                                 CERROR("Can't create local entry %*s for "
1135                                        "remote inode.\n", rec->ur_namelen - 1,
1136                                         rec->ur_name);
1137                                 GOTO(cleanup, rc);
1138                         }
1139
1140                         /* fill reply */
1141                         body->valid |= OBD_MD_FLID | OBD_MD_MDS | OBD_MD_FID;
1142
1143                         obdo2id(&body->id1, oa);
1144                         obdo_free(oa);
1145                 } else {
1146                         /* requested name exists in the directory */
1147                         rc = -EEXIST;
1148                 }
1149                 break;
1150         }
1151         case S_IFLNK:{
1152                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
1153                 if (IS_ERR(handle))
1154                         GOTO(cleanup, rc = PTR_ERR(handle));
1155                 if (rec->ur_tgt == NULL)        /* no target supplied */
1156                         rc = -EINVAL;           /* -EPROTO? */
1157                 else
1158                         rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
1159                 break;
1160         }
1161         case S_IFCHR:
1162         case S_IFBLK:
1163         case S_IFIFO:
1164         case S_IFSOCK:{
1165                 int rdev = rec->ur_rdev;
1166                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
1167                 if (IS_ERR(handle))
1168                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
1169                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
1170                 break;
1171         }
1172         default:
1173                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
1174                 dchild->d_fsdata = NULL;
1175                 GOTO(cleanup, rc = -EINVAL);
1176         }
1177
1178         /* In case we stored the desired inum in here, we want to clean up. */
1179         if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
1180                 dchild->d_fsdata = NULL;
1181
1182         if (rc) {
1183                 CDEBUG(D_INODE, "error during create: %d\n", rc);
1184                 GOTO(cleanup, rc);
1185         } else if (dchild->d_inode) {
1186                 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1187                 struct inode *inode = dchild->d_inode;
1188                 struct iattr iattr;
1189
1190                 created = 1;
1191                 iattr.ia_uid = rec->ur_fsuid;
1192                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1193                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1194                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1195
1196                 if (dir->i_mode & S_ISGID)
1197                         iattr.ia_gid = dir->i_gid;
1198                 else
1199                         iattr.ia_gid = rec->ur_fsgid;
1200
1201                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1202                         ATTR_MTIME | ATTR_CTIME;
1203
1204                 if (id_ino(rec->ur_id2)) {
1205                         LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
1206                         inode->i_generation = id_gen(rec->ur_id2);
1207                         /* dirtied and committed by the upcoming setattr. */
1208                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1209                                inode->i_ino, inode->i_generation);
1210                 }
1211                 mds_inode2id(obd, &body->id1, dchild->d_inode, fid);
1212                 mds_update_inode_ids(obd, inode, handle, &body->id1, &sid);
1213
1214                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1215                 if (rc)
1216                         CERROR("error on child setattr: rc = %d\n", rc);
1217
1218                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1219                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1220                 if (rc)
1221                         CERROR("error on parent setattr: rc = %d\n", rc);
1222                 else
1223                         MD_COUNTER_INCREMENT(obd, create);
1224
1225                 /* take care of default stripe inheritance */
1226                 if (type == S_IFDIR) {
1227                         struct lov_mds_md lmm;
1228                         int lmm_size = sizeof(lmm), nstripes = 0;
1229
1230                         rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0);
1231                         if (rc > 0) {
1232                                 down(&inode->i_sem);
1233                                 rc = fsfilt_set_md(obd, inode, handle,
1234                                                    &lmm, lmm_size, EA_LOV);
1235                                 up(&inode->i_sem);
1236                         }
1237                         if (rc) {
1238                                 CERROR("error on copy stripe info: rc = %d\n",
1239                                        rc);
1240                                 rc = 0;
1241                         }
1242
1243                         if (rec->ur_eadata)
1244                                 nstripes = *(u16 *)rec->ur_eadata;
1245
1246                         if (nstripes) {
1247                                 /*
1248                                  * we pass LCK_EX to split routine to signal,
1249                                  * that we have exclusive access to the
1250                                  * directory. Simple because nobody knows it
1251                                  * already exists -bzzz
1252                                  */
1253                                 rc = mds_try_to_split_dir(obd, dchild,
1254                                                           NULL, nstripes,
1255                                                           LCK_EX);
1256                                 if (rc > 0) {
1257                                         /* dir got splitted */
1258                                         rc = 0;
1259                                 } else if (rc < 0) {
1260                                         /* an error occured during
1261                                          * splitting. */
1262                                         GOTO(cleanup, rc);
1263                                 }
1264                         }
1265
1266                 }
1267
1268                 mds_pack_inode2body(obd, body, inode, 1);
1269                 mds_body_do_reverse_map(med, body);
1270
1271                 if (rec->ur_flags & MDS_REINT_REQ) {
1272                         LASSERT(body != NULL);
1273                         rc = mds_fidmap_add(obd, &body->id1);
1274                         if (rc < 0) {
1275                                 CERROR("can't create fid->ino mapping, "
1276                                        "err %d\n", rc);
1277                         } else {
1278                                 rc = 0;
1279                         }
1280                 }
1281         }
1282
1283         EXIT;
1284 cleanup:
1285         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
1286
1287         if (rc && created) {
1288                 /* Destroy the file we just created. This should not need extra
1289                  * journal credits, as we have already modified all of the
1290                  * blocks needed in order to create the file in the first
1291                  * place. */
1292                 switch (type) {
1293                 case S_IFDIR:
1294                         err = vfs_rmdir(dir, dchild);
1295                         if (err)
1296                                 CERROR("rmdir in error path: %d\n", err);
1297                         break;
1298                 default:
1299                         err = vfs_unlink(dir, dchild);
1300                         if (err)
1301                                 CERROR("unlink in error path: %d\n", err);
1302                         break;
1303                 }
1304         } else if (created) {
1305                 /* The inode we were allocated may have just been freed
1306                  * by an unlink operation.  We take this lock to
1307                  * synchronize against the matching reply-ack-lock taken
1308                  * in unlink, to avoid replay problems if this reply
1309                  * makes it out to the client but the unlink's does not.
1310                  * See bug 2029 for more detail.*/
1311                 mds_lock_new_child(obd, dchild->d_inode, NULL);
1312         } else {
1313                 rc = err;
1314         }
1315         switch (cleanup_phase) {
1316         case 2: /* child dentry */
1317                 if (rc == -EACCES)
1318                         mds_audit_perm(req, dchild->d_inode, AUDIT_CREATE);
1319                 l_dput(dchild);
1320         case 1: /* locked parent dentry */
1321 #ifdef S_PDIROPS
1322                 if (lockh[1].cookie != 0)
1323                         ldlm_lock_decref(lockh + 1, parent_mode);
1324 #endif
1325                 if (rc) {
1326                         ldlm_lock_decref(lockh, LCK_PW);
1327                 } else {
1328                         ptlrpc_save_lock(req, lockh, LCK_PW);
1329                 }
1330                 l_dput(dparent);
1331         case 0:
1332                 break;
1333         default:
1334                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1335                 LBUG();
1336         }
1337         if (mea)
1338                 OBD_FREE(mea, mea_size);
1339         req->rq_status = rc;
1340         return 0;
1341 }
1342
1343 static inline int
1344 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1345        ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1346 {
1347         int i;
1348
1349         for (i = 0; i < RES_NAME_SIZE; i++) {
1350                 /*
1351                  * this is needed to make zeroed res_id entries to be put at the
1352                  * end of list in *ordered_locks() .
1353                  */
1354                 if (res1->name[i] == 0 && res2->name[i] != 0)
1355                         return 1;
1356                 if (res2->name[i] == 0 && res1->name[i] != 0)
1357                         return 0;
1358                 if (res1->name[i] > res2->name[i])
1359                         return 1;
1360                 if (res1->name[i] < res2->name[i])
1361                         return 0;
1362         }
1363
1364         if (!p1 || !p2)
1365                 return 0;
1366
1367         if (memcmp(p1, p2, sizeof(*p1)) < 0)
1368                 return 1;
1369
1370         return 0;
1371 }
1372
1373 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1374  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1375  * because they take the place of local semaphores.
1376  *
1377  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
1378  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
1379 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1380                           struct lustre_handle *p1_lockh, int p1_lock_mode,
1381                           ldlm_policy_data_t *p1_policy,
1382                           struct ldlm_res_id *p2_res_id,
1383                           struct lustre_handle *p2_lockh, int p2_lock_mode,
1384                           ldlm_policy_data_t *p2_policy)
1385 {
1386         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1387         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1388         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1389         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1390         int rc, flags;
1391         ENTRY;
1392
1393         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1394
1395         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1396                res_id[0]->name[0], res_id[1]->name[0]);
1397
1398         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1399                 handles[1] = p1_lockh;
1400                 handles[0] = p2_lockh;
1401                 res_id[1] = p1_res_id;
1402                 res_id[0] = p2_res_id;
1403                 lock_modes[1] = p1_lock_mode;
1404                 lock_modes[0] = p2_lock_mode;
1405                 policies[1] = p1_policy;
1406                 policies[0] = p2_policy;
1407         }
1408
1409         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1410                res_id[0]->name[0], res_id[1]->name[0]);
1411
1412         flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1413         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1414                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1415                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1416                               NULL, 0, NULL, handles[0]);
1417         if (rc != ELDLM_OK)
1418                 RETURN(-EIO);
1419         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1420
1421         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1422             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1423                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1424                 ldlm_lock_addref(handles[1], lock_modes[1]);
1425         } else if (res_id[1]->name[0] != 0) {
1426                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1427                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1428                                       *res_id[1], LDLM_IBITS, policies[1],
1429                                       lock_modes[1], &flags, mds_blocking_ast,
1430                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1431                                       NULL, handles[1]);
1432                 if (rc != ELDLM_OK) {
1433                         ldlm_lock_decref(handles[0], lock_modes[0]);
1434                         RETURN(-EIO);
1435                 }
1436                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1437         }
1438
1439         RETURN(0);
1440 }
1441
1442 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1443                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1444                            ldlm_policy_data_t *p1_policy,
1445                            struct ldlm_res_id *p2_res_id,
1446                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1447                            ldlm_policy_data_t *p2_policy,
1448                            struct ldlm_res_id *c1_res_id,
1449                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1450                            ldlm_policy_data_t *c1_policy,
1451                            struct ldlm_res_id *c2_res_id,
1452                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1453                            ldlm_policy_data_t *c2_policy)
1454 {
1455         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1456                                           c1_res_id, c2_res_id };
1457         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1458                                                  c1_lockh, c2_lockh };
1459         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1460                               c1_lock_mode, c2_lock_mode };
1461         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1462                                             c1_policy, c2_policy};
1463         int rc, i, j, sorted, flags;
1464         ENTRY;
1465
1466         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1467                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1468                res_id[3]->name[0]);
1469
1470         /*
1471          * simple insertion sort - we have at most 4 elements. Note, that zeroed
1472          * res_id should be at the end of list after sorting is finished.
1473          */
1474         for (i = 1; i < 4; i++) {
1475                 j = i - 1;
1476                 dlm_handles[4] = dlm_handles[i];
1477                 res_id[4] = res_id[i];
1478                 lock_modes[4] = lock_modes[i];
1479                 policies[4] = policies[i];
1480
1481                 sorted = 0;
1482                 do {
1483                         if (res_gt(res_id[j], res_id[4], policies[j],
1484                                    policies[4])) {
1485                                 dlm_handles[j + 1] = dlm_handles[j];
1486                                 res_id[j + 1] = res_id[j];
1487                                 lock_modes[j + 1] = lock_modes[j];
1488                                 policies[j + 1] = policies[j];
1489                                 j--;
1490                         } else {
1491                                 sorted = 1;
1492                         }
1493                 } while (j >= 0 && !sorted);
1494
1495                 dlm_handles[j + 1] = dlm_handles[4];
1496                 res_id[j + 1] = res_id[4];
1497                 lock_modes[j + 1] = lock_modes[4];
1498                 policies[j + 1] = policies[4];
1499         }
1500
1501         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1502                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1503                res_id[3]->name[0]);
1504
1505         /* XXX we could send ASTs on all these locks first before blocking? */
1506         for (i = 0; i < 4; i++) {
1507                 flags = LDLM_FL_ATOMIC_CB;
1508
1509                 /*
1510                  * nevertheless zeroed res_ids should be at the end of list, and
1511                  * could use break here, I think, that it is more correctly for
1512                  * clear understanding of code to have continue here, as it
1513                  * clearly means, that zeroed res_id should be skipped and does
1514                  * not mean, that if we meet zeroed res_id we should stop
1515                  * locking loop.
1516                  */
1517                 if (res_id[i]->name[0] == 0)
1518                         continue;
1519
1520                 if (i != 0 &&
1521                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1522                     (policies[i]->l_inodebits.bits &
1523                      policies[i-1]->l_inodebits.bits) ) {
1524                         memcpy(dlm_handles[i], dlm_handles[i-1],
1525                                sizeof(*(dlm_handles[i])));
1526                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1527                 } else {
1528                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1529                                               *res_id[i], LDLM_IBITS,
1530                                               policies[i],
1531                                               lock_modes[i], &flags,
1532                                               mds_blocking_ast,
1533                                               ldlm_completion_ast, NULL, NULL,
1534                                               NULL, 0, NULL, dlm_handles[i]);
1535                         if (rc != ELDLM_OK)
1536                                 GOTO(out_err, rc = -EIO);
1537                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1538                 }
1539         }
1540
1541         RETURN(0);
1542 out_err:
1543         while (i-- > 0)
1544                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1545
1546         return rc;
1547 }
1548
1549 /* In the unlikely case that the child changed while we were waiting
1550  * on the lock, we need to drop the lock on the old child and either:
1551  * - if the child has a lower resource name, then we have to also
1552  *   drop the parent lock and regain the locks in the right order
1553  * - in the rename case, if the child has a lower resource name than one of
1554  *   the other parent/child resources (maxres) we also need to reget the locks
1555  * - if the child has a higher resource name (this is the common case)
1556  *   we can just get the lock on the new child (still in lock order)
1557  *
1558  * Returns 0 if the child did not change or if it changed but could be locked.
1559  * Returns 1 if the child changed and we need to re-lock (no locks held).
1560  * Returns -ve error with a valid dchild (no locks held). */
1561 static int mds_verify_child(struct obd_device *obd,
1562                             struct ldlm_res_id *parent_res_id,
1563                             struct lustre_handle *parent_lockh,
1564                             struct dentry *dparent, int parent_mode,
1565                             struct ldlm_res_id *child_res_id,
1566                             struct lustre_handle *child_lockh,
1567                             struct dentry **dchildp, int child_mode,
1568                             ldlm_policy_data_t *child_policy,
1569                             const char *name, int namelen,
1570                             struct ldlm_res_id *maxres,
1571                             unsigned long child_ino, __u32 child_gen)
1572 {
1573         struct lustre_id sid;
1574         struct dentry *vchild, *dchild = *dchildp;
1575         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1576         ENTRY;
1577
1578         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1579         if (IS_ERR(vchild))
1580                 GOTO(cleanup, rc = PTR_ERR(vchild));
1581
1582         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1583                 if (child_gen == vchild->d_generation &&
1584                     child_ino == vchild->d_inum) {
1585                         if (dchild)
1586                                 l_dput(dchild);
1587                         *dchildp = vchild;
1588                         RETURN(0);
1589                 }
1590                 goto changed;
1591         }
1592
1593         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1594                    (vchild->d_inode != NULL &&
1595                     child_gen == vchild->d_inode->i_generation &&
1596                     child_ino == vchild->d_inode->i_ino))) {
1597                 if (dchild)
1598                         l_dput(dchild);
1599                 *dchildp = vchild;
1600                 RETURN(0);
1601         }
1602
1603 changed:
1604         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1605                vchild->d_inode, dchild ? dchild->d_inode : 0,
1606                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1607                child_res_id->name[0]);
1608
1609         if (child_res_id->name[0] != 0)
1610                 ldlm_lock_decref(child_lockh, child_mode);
1611         if (dchild)
1612                 l_dput(dchild);
1613
1614         cleanup_phase = 1; /* parent lock only */
1615         *dchildp = dchild = vchild;
1616
1617         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1618                 int flags = LDLM_FL_ATOMIC_CB;
1619
1620                 if (dchild->d_inode) {
1621                         down(&dchild->d_inode->i_sem);
1622                         rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1623                         up(&dchild->d_inode->i_sem);
1624                         if (rc) {
1625                                 CERROR("Can't read inode self id, inode %lu,"
1626                                        " rc %d\n",  dchild->d_inode->i_ino, rc);
1627                                 GOTO(cleanup, rc);
1628                         }
1629                         child_res_id->name[0] = id_fid(&sid);
1630                         child_res_id->name[1] = id_group(&sid);
1631                 } else {
1632                         child_res_id->name[0] = dchild->d_fid;
1633                         child_res_id->name[1] = dchild->d_mdsnum;
1634                 }
1635
1636                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1637                     res_gt(maxres, child_res_id, NULL, NULL)) {
1638                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1639                                child_res_id->name[0], parent_res_id->name[0],
1640                                maxres->name[0]);
1641                         GOTO(cleanup, rc = 1);
1642                 }
1643
1644                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1645                                       *child_res_id, LDLM_IBITS, child_policy,
1646                                       child_mode, &flags, mds_blocking_ast,
1647                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1648                                       NULL, child_lockh);
1649                 if (rc != ELDLM_OK)
1650                         GOTO(cleanup, rc = -EIO);
1651
1652         } else {
1653                 memset(child_res_id, 0, sizeof(*child_res_id));
1654         }
1655
1656         EXIT;
1657 cleanup:
1658         if (rc) {
1659                 switch(cleanup_phase) {
1660                 case 2:
1661                         if (child_res_id->name[0] != 0)
1662                                 ldlm_lock_decref(child_lockh, child_mode);
1663                 case 1:
1664                         ldlm_lock_decref(parent_lockh, parent_mode);
1665                 }
1666         }
1667         return rc;
1668 }
1669
1670 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1671                                 struct lustre_id *id,
1672                                 struct lustre_handle *parent_lockh,
1673                                 struct dentry **dparentp, int parent_mode,
1674                                 __u64 parent_lockpart, int *update_mode,
1675                                 char *name, int namelen,
1676                                 struct lustre_handle *child_lockh,
1677                                 struct dentry **dchildp, int child_mode,
1678                                 __u64 child_lockpart)
1679 {
1680         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1681         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1682         struct ldlm_res_id parent_res_id = { .name = {0} };
1683         struct ldlm_res_id child_res_id = { .name = {0} };
1684         unsigned long child_ino = 0; __u32 child_gen = 0;
1685         int rc = 0, cleanup_phase = 0;
1686         struct lustre_id sid;
1687         struct inode *inode;
1688         ENTRY;
1689
1690         /* Step 1: Lookup parent */
1691         *dparentp = mds_id2dentry(obd, id, NULL);
1692         if (IS_ERR(*dparentp)) {
1693                 rc = PTR_ERR(*dparentp);
1694                 *dparentp = NULL;
1695                 RETURN(rc);
1696         }
1697
1698         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1699                (*dparentp)->d_inode->i_ino, name);
1700
1701         parent_res_id.name[0] = id_fid(id);
1702         parent_res_id.name[1] = id_group(id);
1703
1704 #ifdef S_PDIROPS
1705         parent_lockh[1].cookie = 0;
1706         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1707                 struct ldlm_res_id res_id = { .name = {0} };
1708                 ldlm_policy_data_t policy;
1709                 int flags = LDLM_FL_ATOMIC_CB;
1710
1711                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1712                 if (*update_mode) {
1713                         res_id.name[0] = id_fid(id);
1714                         res_id.name[1] = id_group(id);
1715                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1716
1717                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1718                                               res_id, LDLM_IBITS, &policy,
1719                                               *update_mode, &flags,
1720                                               mds_blocking_ast,
1721                                               ldlm_completion_ast,
1722                                               NULL, NULL, NULL, 0, NULL,
1723                                               parent_lockh + 1);
1724                         if (rc != ELDLM_OK)
1725                                 RETURN(-ENOLCK);
1726                 }
1727
1728                 parent_res_id.name[2] = full_name_hash((unsigned char *)name,
1729                                                        namelen - 1);
1730
1731                 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1732                        (unsigned long)id_fid(id), (unsigned long)id_group(id),
1733                        parent_res_id.name[2]);
1734         }
1735 #endif
1736
1737         cleanup_phase = 1; /* parent dentry */
1738
1739         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1740         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1741         if (IS_ERR(*dchildp)) {
1742                 rc = PTR_ERR(*dchildp);
1743                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1744                 GOTO(cleanup, rc);
1745         }
1746
1747         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1748                 /*
1749                  * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1750                  * lock. Drop possible UPDATE lock!
1751                  */
1752                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1753                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1754
1755                 child_res_id.name[0] = (*dchildp)->d_fid;
1756                 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1757                 child_ino = (*dchildp)->d_inum;
1758                 child_gen = (*dchildp)->d_generation;
1759                 goto retry_locks;
1760         }
1761
1762         inode = (*dchildp)->d_inode;
1763         if (inode != NULL)
1764                 inode = igrab(inode);
1765         if (inode == NULL)
1766                 goto retry_locks;
1767
1768         down(&inode->i_sem);
1769         rc = mds_read_inode_sid(obd, inode, &sid);
1770         up(&inode->i_sem);
1771         if (rc) {
1772                 CERROR("Can't read inode self id, inode %lu, "
1773                        "rc %d\n", inode->i_ino, rc);
1774                 iput(inode);
1775                 GOTO(cleanup, rc);
1776         }
1777
1778         child_ino = inode->i_ino;
1779         child_gen = inode->i_generation;
1780         child_res_id.name[0] = id_fid(&sid);
1781         child_res_id.name[1] = id_group(&sid);
1782         iput(inode);
1783
1784 retry_locks:
1785         cleanup_phase = 2; /* child dentry */
1786
1787         /* Step 3: Lock parent and child in resource order.  If child doesn't
1788          * exist, we still have to lock the parent and re-lookup. */
1789         rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1790                                    &parent_policy, &child_res_id, child_lockh,
1791                                    child_mode, &child_policy);
1792         if (rc)
1793                 GOTO(cleanup, rc);
1794
1795         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1796                 cleanup_phase = 4; /* child lock */
1797         else
1798                 cleanup_phase = 3; /* parent lock */
1799
1800         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1801         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1802                               parent_mode, &child_res_id, child_lockh,
1803                               dchildp, child_mode, &child_policy,
1804                               name, namelen, &parent_res_id, child_ino,
1805                               child_gen);
1806         if (rc > 0)
1807                 goto retry_locks;
1808         if (rc < 0) {
1809                 cleanup_phase = 3;
1810                 GOTO(cleanup, rc);
1811         }
1812
1813         EXIT;
1814 cleanup:
1815         if (rc) {
1816                 switch (cleanup_phase) {
1817                 case 4:
1818                         ldlm_lock_decref(child_lockh, child_mode);
1819                 case 3:
1820                         ldlm_lock_decref(parent_lockh, parent_mode);
1821                 case 2:
1822                         l_dput(*dchildp);
1823                 case 1:
1824 #ifdef S_PDIROPS
1825                         if (parent_lockh[1].cookie)
1826                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1827 #endif
1828                         l_dput(*dparentp);
1829                 }
1830         }
1831         return rc;
1832 }
1833
1834 void mds_reconstruct_generic(struct ptlrpc_request *req)
1835 {
1836         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1837         mds_req_from_mcd(req, med->med_mcd);
1838 }
1839
1840 /* If we are unlinking an open file/dir (i.e. creating an orphan) then we
1841  * instead link the inode into the PENDING directory until it is finally
1842  * released. We can't simply call mds_reint_rename() or some part thereof,
1843  * because we don't have the inode to check for link count/open status until
1844  * after it is locked.
1845  *
1846  * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1847  * before starting journal transaction.
1848  *
1849  * returns 1 on success
1850  * returns 0 if we lost a race and didn't make a new link
1851  * returns negative on error
1852  */
1853 static int mds_orphan_add_link(struct mds_update_record *rec,
1854                                struct obd_device *obd, struct dentry *dentry)
1855 {
1856         struct mds_obd *mds = &obd->u.mds;
1857         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1858         struct inode *inode = dentry->d_inode;
1859         struct dentry *pending_child;
1860         char idname[LL_ID_NAMELEN];
1861         int idlen = 0, rc, mode;
1862         ENTRY;
1863
1864         LASSERT(inode != NULL);
1865         LASSERT(!mds_inode_is_orphan(inode));
1866 #ifndef HAVE_I_ALLOC_SEM
1867         LASSERT(down_trylock(&inode->i_sem) != 0);
1868 #endif
1869         LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1870
1871         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1872
1873         CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1874                mds_orphan_open_count(inode), inode->i_nlink,
1875                S_ISDIR(inode->i_mode) ? "dir" :
1876                S_ISREG(inode->i_mode) ? "file" : "other",
1877                rec->ur_name, idname);
1878
1879         if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1880                 RETURN(0);
1881
1882         pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1883         if (IS_ERR(pending_child))
1884                 RETURN(PTR_ERR(pending_child));
1885
1886         if (pending_child->d_inode != NULL) {
1887                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1888                 LASSERT(pending_child->d_inode == inode);
1889                 GOTO(out_dput, rc = 0);
1890         }
1891
1892         /*
1893          * link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG for
1894          * linking and return real mode back then -bzzz
1895          */
1896         mode = inode->i_mode;
1897         inode->i_mode = S_IFREG;
1898         rc = vfs_link(dentry, pending_dir, pending_child);
1899         if (rc)
1900                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1901                        rec->ur_name, rc);
1902         else
1903                 mds_inode_set_orphan(inode);
1904
1905         /* return mode and correct i_nlink if inode is directory */
1906         inode->i_mode = mode;
1907         LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1908                  S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1909                  inode->i_nlink);
1910         if (S_ISDIR(mode)) {
1911                 i_nlink_inc(inode);
1912                 i_nlink_inc(pending_dir);
1913                 mark_inode_dirty(inode);
1914                 mark_inode_dirty(pending_dir);
1915         }
1916
1917         GOTO(out_dput, rc = 1);
1918 out_dput:
1919         l_dput(pending_child);
1920         return rc;
1921 }
1922
1923 int mds_create_local_dentry(struct mds_update_record *rec,
1924                             struct obd_device *obd)
1925 {
1926         struct mds_obd *mds = &obd->u.mds;
1927         struct inode *id_dir = mds->mds_id_dir->d_inode;
1928         int idlen = 0, rc, cleanup_phase = 0;
1929         struct dentry *new_child = NULL;
1930         char *idname = rec->ur_name;
1931         struct dentry *child = NULL;
1932         struct lustre_handle lockh[2] = {{0}, {0}};
1933         struct lustre_id sid;
1934         void *handle;
1935         ENTRY;
1936
1937         down(&id_dir->i_sem);
1938         idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1939                           id_gen(rec->ur_id1));
1940
1941         CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1942                idname, OLID4(rec->ur_id1));
1943
1944         new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1945                                       idlen);
1946         up(&id_dir->i_sem);
1947         if (IS_ERR(new_child)) {
1948                 CERROR("can't lookup %s: %d\n", idname,
1949                        (int) PTR_ERR(new_child));
1950                 GOTO(cleanup, rc = PTR_ERR(new_child));
1951         }
1952         cleanup_phase = 1;
1953
1954         down(&id_dir->i_sem);
1955         rc = mds_read_inode_sid(obd, id_dir, &sid);
1956         up(&id_dir->i_sem);
1957         if (rc) {
1958                 CERROR("Can't read inode self id, inode %lu, "
1959                        "rc %d\n", id_dir->i_ino, rc);
1960                 GOTO(cleanup, rc);
1961         }
1962
1963         if (new_child->d_inode != NULL) {
1964                 /* nice. we've already have local dentry! */
1965                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1966                        (unsigned)new_child->d_inode->i_ino,
1967                        (unsigned)new_child->d_inode->i_generation);
1968
1969                 id_ino(rec->ur_id1) = id_dir->i_ino;
1970                 id_gen(rec->ur_id1) = id_dir->i_generation;
1971                 rec->ur_namelen = idlen + 1;
1972
1973                 id_fid(rec->ur_id1) = id_fid(&sid);
1974                 id_group(rec->ur_id1) = id_group(&sid);
1975
1976                 GOTO(cleanup, rc = 0);
1977         }
1978
1979         /* new, local dentry will be added soon. we need no aliases here */
1980         d_drop(new_child);
1981
1982         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1983                 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1984         } else {
1985                 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1986                                              LCK_EX, lockh, NULL, NULL, 0,
1987                                              MDS_INODELOCK_UPDATE);
1988         }
1989
1990         if (IS_ERR(child)) {
1991                 rc = PTR_ERR(child);
1992                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1993                         CERROR("can't get victim: %d\n", rc);
1994                 GOTO(cleanup, rc);
1995         }
1996         cleanup_phase = 2;
1997
1998         handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1999         if (IS_ERR(handle))
2000                 GOTO(cleanup, rc = PTR_ERR(handle));
2001
2002         rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
2003                                   idlen, id_ino(rec->ur_id1),
2004                                   id_gen(rec->ur_id1), mds->mds_num,
2005                                   id_fid(rec->ur_id1));
2006         if (rc)
2007                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
2008                        (unsigned long)child->d_inode->i_ino,
2009                        (unsigned long)child->d_inode->i_generation, rc);
2010         else {
2011                 if (S_ISDIR(child->d_inode->i_mode)) {
2012                         i_nlink_inc(id_dir);
2013                         mark_inode_dirty(id_dir);
2014                 }
2015                 mark_inode_dirty(child->d_inode);
2016         }
2017         fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
2018
2019         id_ino(rec->ur_id1) = id_dir->i_ino;
2020         id_gen(rec->ur_id1) = id_dir->i_generation;
2021         rec->ur_namelen = idlen + 1;
2022
2023         id_fid(rec->ur_id1) = id_fid(&sid);
2024         id_group(rec->ur_id1) = id_group(&sid);
2025
2026         EXIT;
2027 cleanup:
2028         switch(cleanup_phase) {
2029                 case 2:
2030                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
2031                                 ldlm_lock_decref(lockh, LCK_EX);
2032                         dput(child);
2033                 case 1:
2034                         dput(new_child);
2035                 case 0:
2036                        break;
2037         }
2038         return rc;
2039 }
2040
2041 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
2042                                  struct ptlrpc_request *slave)
2043 {
2044         void *cookie, *cookie2;
2045         struct mds_body *body2;
2046         struct mds_body *body;
2047         void *ea, *ea2;
2048         ENTRY;
2049
2050         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
2051         LASSERT(body != NULL);
2052
2053         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
2054         LASSERT(body2 != NULL);
2055
2056         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
2057                 RETURN(0);
2058
2059         memcpy(body2, body, sizeof(*body));
2060         body2->valid &= ~OBD_MD_FLCOOKIE;
2061
2062         if (!(body->valid & OBD_MD_FLEASIZE) &&
2063             !(body->valid & OBD_MD_FLDIREA))
2064                 RETURN(0);
2065
2066         if (body->eadatasize == 0) {
2067                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
2068                 RETURN(0);
2069         }
2070
2071         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
2072
2073         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
2074         LASSERT(ea != NULL);
2075
2076         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
2077         LASSERT(ea2 != NULL);
2078
2079         memcpy(ea2, ea, body->eadatasize);
2080
2081         if (body->valid & OBD_MD_FLCOOKIE) {
2082                 LASSERT(master->rq_repmsg->buflens[2] >=
2083                                 slave->rq_repmsg->buflens[2]);
2084                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
2085                                 slave->rq_repmsg->buflens[2]);
2086                 LASSERT(cookie != NULL);
2087
2088                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
2089                                 master->rq_repmsg->buflens[2]);
2090                 LASSERT(cookie2 != NULL);
2091                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
2092                 body2->valid |= OBD_MD_FLCOOKIE;
2093         }
2094         RETURN(0);
2095 }
2096
2097 static int mds_reint_unlink_remote(struct mds_update_record *rec,
2098                                    int offset, struct ptlrpc_request *req,
2099                                    struct lustre_handle *parent_lockh,
2100                                    int update_mode, struct dentry *dparent,
2101                                    struct lustre_handle *child_lockh,
2102                                    struct dentry *dchild)
2103 {
2104         struct obd_device *obd = req->rq_export->exp_obd;
2105         struct mds_obd *mds = mds_req2mds(req);
2106         struct ptlrpc_request *request = NULL;
2107         int rc = 0, cleanup_phase = 0;
2108         struct mdc_op_data *op_data;
2109         void *handle;
2110         ENTRY;
2111
2112         LASSERT(offset == 1 || offset == 3);
2113
2114         /* time to drop i_nlink on remote MDS */
2115         OBD_ALLOC(op_data, sizeof(*op_data));
2116         if (op_data == NULL)
2117                 RETURN(-ENOMEM);
2118
2119         memset(op_data, 0, sizeof(*op_data));
2120         mds_pack_dentry2id(obd, &op_data->id1, dchild, 1);
2121         op_data->create_mode = rec->ur_mode;
2122
2123         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
2124                   rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2125
2126         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2127                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
2128                           rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data->id1));
2129         }
2130
2131         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2132                 op_data->create_mode |= MDS_MODE_REPLAY;
2133
2134         rc = md_unlink(mds->mds_md_exp, op_data, &request);
2135         OBD_FREE(op_data, sizeof(*op_data));
2136         cleanup_phase = 2;
2137
2138         if (request) {
2139                 if (rc == 0)
2140                         mds_copy_unlink_reply(req, request);
2141                 ptlrpc_req_finished(request);
2142         }
2143
2144         if (rc == 0) {
2145                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2146                                       NULL);
2147                 if (IS_ERR(handle))
2148                         GOTO(cleanup, rc = PTR_ERR(handle));
2149                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
2150                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
2151                                         rc, 0);
2152         }
2153         EXIT;
2154 cleanup:
2155         req->rq_status = rc;
2156
2157 #ifdef S_PDIROPS
2158         if (parent_lockh[1].cookie != 0)
2159                 ldlm_lock_decref(parent_lockh + 1, update_mode);
2160 #endif
2161         ldlm_lock_decref(child_lockh, LCK_EX);
2162         if (rc)
2163                 ldlm_lock_decref(parent_lockh, LCK_PW);
2164         else
2165                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2166         l_dput(dchild);
2167         l_dput(dparent);
2168
2169         return 0;
2170 }
2171
2172 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
2173                             struct ptlrpc_request *req, struct lustre_handle *lh)
2174 {
2175         struct dentry *dparent = NULL, *dchild;
2176         struct mds_obd *mds = mds_req2mds(req);
2177         struct obd_device *obd = req->rq_export->exp_obd;
2178         struct mds_body *body = NULL;
2179         struct inode *child_inode = NULL;
2180         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2181         struct lustre_handle child_lockh = {0};
2182         struct lustre_handle child_reuse_lockh = {0};
2183         struct lustre_handle *slave_lockh = NULL;
2184         char idname[LL_ID_NAMELEN];
2185         struct llog_create_locks *lcl = NULL;
2186         void *handle = NULL;
2187         int rc = 0, cleanup_phase = 0;
2188         int unlink_by_id = 0;
2189         int update_mode;
2190         ENTRY;
2191
2192         LASSERT(offset == 1 || offset == 3);
2193
2194         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
2195                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2196                   rec->ur_name);
2197
2198         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2199
2200         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2201                 DEBUG_REQ(D_HA, req, "unlink replay");
2202                 LASSERT(offset == 1); /* should not come from intent */
2203                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2204                        lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
2205                        req->rq_repmsg->buflens[2]);
2206         }
2207
2208         MD_COUNTER_INCREMENT(obd, unlink);
2209
2210         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
2211                 GOTO(cleanup, rc = -ENOENT);
2212
2213         if (rec->ur_namelen == 1) {
2214                 /* this is request to drop i_nlink on local inode */
2215                 unlink_by_id = 1;
2216                 rec->ur_name = idname;
2217                 rc = mds_create_local_dentry(rec, obd);
2218                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
2219                         DEBUG_REQ(D_HA, req,
2220                                   "drop nlink on inode "DLID4" (replay)",
2221                                   OLID4(rec->ur_id1));
2222                         req->rq_status = 0;
2223                         RETURN(0);
2224                 }
2225         }
2226
2227         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
2228                 /* master mds for directory asks slave removing inode is already
2229                  * locked */
2230                 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
2231                                                LCK_PW, parent_lockh,
2232                                                &update_mode, rec->ur_name,
2233                                                rec->ur_namelen,
2234                                                MDS_INODELOCK_UPDATE);
2235                 if (IS_ERR(dparent))
2236                         GOTO(cleanup, rc = PTR_ERR(dparent));
2237                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
2238                                            rec->ur_namelen - 1);
2239                 if (IS_ERR(dchild))
2240                         GOTO(cleanup, rc = PTR_ERR(dchild));
2241                 child_lockh.cookie = 0;
2242                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
2243                 LASSERT(dchild->d_inode != NULL);
2244                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
2245         } else {
2246                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
2247                                                  parent_lockh, &dparent,
2248                                                  LCK_PW, MDS_INODELOCK_UPDATE,
2249                                                  &update_mode, rec->ur_name,
2250                                                  rec->ur_namelen, &child_lockh,
2251                                                  &dchild, LCK_EX,
2252                                                  (MDS_INODELOCK_LOOKUP |
2253                                                   MDS_INODELOCK_UPDATE));
2254         }
2255         if (rc)
2256                 GOTO(cleanup, rc);
2257
2258         if (dchild->d_flags & DCACHE_CROSS_REF) {
2259                 /* we should have parent lock only here */
2260                 LASSERT(unlink_by_id == 0);
2261                 LASSERT(dchild->d_mdsnum != mds->mds_num);
2262                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
2263                                         update_mode, dparent, &child_lockh, dchild);
2264                 RETURN(0);
2265         }
2266
2267         cleanup_phase = 1; /* dchild, dparent, locks */
2268
2269         dget(dchild);
2270         child_inode = dchild->d_inode;
2271         if (child_inode == NULL) {
2272                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
2273                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
2274                 GOTO(cleanup, rc = -ENOENT);
2275         }
2276
2277         cleanup_phase = 2; /* dchild has a lock */
2278
2279         /* We have to do these checks ourselves, in case we are making an
2280          * orphan.  The client tells us whether rmdir() or unlink() was called,
2281          * so we need to return appropriate errors (bug 72).
2282          *
2283          * We don't have to check permissions, because vfs_rename (called from
2284          * mds_open_unlink_rename) also calls may_delete. */
2285         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
2286                 if (!S_ISDIR(child_inode->i_mode))
2287                         GOTO(cleanup, rc = -ENOTDIR);
2288         } else {
2289                 if (S_ISDIR(child_inode->i_mode))
2290                         GOTO(cleanup, rc = -EISDIR);
2291         }
2292
2293         /* handle splitted dir */
2294         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
2295         if (rc)
2296                 GOTO(cleanup, rc);
2297
2298         /* Step 4: Get a lock on the ino to sync with creation WRT inode
2299          * reuse (see bug 2029). */
2300         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
2301         if (rc != ELDLM_OK)
2302                 GOTO(cleanup, rc);
2303         cleanup_phase = 3; /* child inum lock */
2304
2305         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2306
2307         /* ldlm_reply in buf[0] if called via intent */
2308         if (offset == 3)
2309                 offset = 1;
2310         else
2311                 offset = 0;
2312
2313         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2314         LASSERT(body != NULL);
2315
2316         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2317         DOWN_READ_I_ALLOC_SEM(child_inode);
2318         cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2319
2320         /* If this is potentially the last reference to this inode, get the
2321          * OBD EA data first so the client can destroy OST objects.  We
2322          * only do the object removal later if no open files/links remain. */
2323         if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2324             child_inode->i_nlink == 1) {
2325                 if (mds_orphan_open_count(child_inode) > 0) {
2326                         /* need to lock pending_dir before transaction */
2327                         down(&mds->mds_pending_dir->d_inode->i_sem);
2328                         cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2329                 } else if (S_ISREG(child_inode->i_mode)) {
2330                         mds_pack_inode2body(obd, body, child_inode, 0);
2331                         mds_pack_md(obd, req->rq_repmsg, offset + 1,
2332                                     body, child_inode, MDS_PACK_MD_LOCK, 0);
2333                 }
2334         }
2335
2336         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2337         switch (child_inode->i_mode & S_IFMT) {
2338         case S_IFDIR:
2339                 /* Drop any lingering child directories before we start our
2340                  * transaction, to avoid doing multiple inode dirty/delete
2341                  * in our compound transaction (bug 1321). */
2342                 shrink_dcache_parent(dchild);
2343                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2344                                       NULL);
2345                 if (IS_ERR(handle))
2346                         GOTO(cleanup, rc = PTR_ERR(handle));
2347                 rc = vfs_rmdir(dparent->d_inode, dchild);
2348                 break;
2349         case S_IFREG: {
2350 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2351                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2352                                                         offset + 1, 0);
2353                 handle = fsfilt_start_log(obd, dparent->d_inode,
2354                                           FSFILT_OP_UNLINK, NULL,
2355                                           le32_to_cpu(lmm->lmm_stripe_count));
2356                 if (IS_ERR(handle))
2357                         GOTO(cleanup, rc = PTR_ERR(handle));
2358                 rc = vfs_unlink(dparent->d_inode, dchild);
2359                 break;
2360         }
2361         case S_IFLNK:
2362         case S_IFCHR:
2363         case S_IFBLK:
2364         case S_IFIFO:
2365         case S_IFSOCK:
2366                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2367                                       NULL);
2368                 if (IS_ERR(handle))
2369                         GOTO(cleanup, rc = PTR_ERR(handle));
2370                 rc = vfs_unlink(dparent->d_inode, dchild);
2371                 break;
2372         default:
2373                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2374                        rec->ur_name);
2375                 LBUG();
2376                 GOTO(cleanup, rc = -EINVAL);
2377         }
2378
2379         if (rc == 0 && child_inode->i_nlink == 0) {
2380                 if (mds_orphan_open_count(child_inode) > 0)
2381                         rc = mds_orphan_add_link(rec, obd, dchild);
2382
2383                 if (rc == 1)
2384                         GOTO(cleanup, rc = 0);
2385
2386                 if (!S_ISREG(child_inode->i_mode))
2387                         GOTO(cleanup, rc);
2388
2389                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2390                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2391                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2392                 } else if (mds_log_op_unlink(obd, child_inode,
2393                                              lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2394                                              req->rq_repmsg->buflens[offset + 1],
2395                                              lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2396                                              req->rq_repmsg->buflens[offset + 2],
2397                                              &lcl) > 0){
2398                         body->valid |= OBD_MD_FLCOOKIE;
2399                 }
2400
2401                 rc = mds_destroy_object(obd, child_inode, 1);
2402                 if (rc) {
2403                         CERROR("can't remove OST object, err %d\n",
2404                                rc);
2405                 }
2406
2407                 if (child_inode->i_nlink == 0)
2408                         mds_fidmap_del(obd, &body->id1);
2409         }
2410
2411         GOTO(cleanup, rc);
2412
2413 cleanup:
2414         if (rc == 0) {
2415                 struct iattr iattr;
2416                 int err;
2417
2418                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2419                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2420                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2421
2422                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2423                 if (err)
2424                         CERROR("error on parent setattr: rc = %d\n", err);
2425         }
2426         rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2427                                 handle, req, rc, 0);
2428         if (!rc)
2429                 (void)obd_set_info(mds->mds_dt_exp, strlen("unlinked"),
2430                                    "unlinked", 0, NULL);
2431
2432         switch(cleanup_phase) {
2433         case 5: /* pending_dir semaphore */
2434                 up(&mds->mds_pending_dir->d_inode->i_sem);
2435         case 4: /* child inode semaphore */
2436                 UP_READ_I_ALLOC_SEM(child_inode);
2437                  /* handle splitted dir */
2438                 if (rc == 0) {
2439                         /* master directory can be non-empty or something else ... */
2440                         mds_unlink_slave_objs(obd, dchild);
2441                 }
2442                 if (lcl != NULL)
2443                         ptlrpc_save_llog_lock(req, lcl);
2444         case 3: /* child ino-reuse lock */
2445                 if (rc && body != NULL) {
2446                         // Don't unlink the OST objects if the MDS unlink failed
2447                         body->valid = 0;
2448                 }
2449                 if (rc)
2450                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2451                 else
2452                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2453         case 2: /* child lock */
2454                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2455                 if (child_lockh.cookie)
2456                         ldlm_lock_decref(&child_lockh, LCK_EX);
2457         case 1: /* child and parent dentry, parent lock */
2458 #ifdef S_PDIROPS
2459                 if (parent_lockh[1].cookie != 0)
2460                         ldlm_lock_decref(parent_lockh + 1, update_mode);
2461 #endif
2462                 if (rc)
2463                         ldlm_lock_decref(parent_lockh, LCK_PW);
2464                 else
2465                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2466                 if (dchild->d_inode && rc && (dchild->d_inode->i_nlink == 0 ||
2467                                 mds_inode_is_orphan(dchild->d_inode)))
2468                         CDEBUG(D_ERROR, "unlink, but return %d\n", rc);
2469
2470                 /* catching failed permissions check for audit */
2471                 if (rc == -EACCES)
2472                         mds_audit_perm(req, dchild->d_inode, AUDIT_UNLINK);
2473
2474                 l_dput(dchild);
2475                 l_dput(dchild);
2476                 l_dput(dparent);
2477         case 0:
2478                 break;
2479         default:
2480                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2481                 LBUG();
2482         }
2483         req->rq_status = rc;
2484         return 0;
2485 }
2486
2487 /*
2488  * to service requests from remote MDS to increment i_nlink
2489  */
2490 static int mds_reint_link_acquire(struct mds_update_record *rec,
2491                                   int offset, struct ptlrpc_request *req,
2492                                   struct lustre_handle *lh)
2493 {
2494         struct obd_device *obd = req->rq_export->exp_obd;
2495         struct ldlm_res_id src_res_id = { .name = {0} };
2496         struct lustre_handle *handle = NULL, src_lockh = {0};
2497         struct mds_obd *mds = mds_req2mds(req);
2498         int rc = 0, cleanup_phase = 0;
2499         struct dentry *de_src = NULL;
2500         ldlm_policy_data_t policy;
2501         int flags = LDLM_FL_ATOMIC_CB;
2502         ENTRY;
2503
2504         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2505                   obd->obd_name, OLID4(rec->ur_id1));
2506
2507         /* Step 1: Lookup the source inode and target directory by ID */
2508         de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2509         if (IS_ERR(de_src))
2510                 GOTO(cleanup, rc = PTR_ERR(de_src));
2511         cleanup_phase = 1; /* source dentry */
2512
2513         src_res_id.name[0] = id_fid(rec->ur_id1);
2514         src_res_id.name[1] = id_group(rec->ur_id1);
2515         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2516
2517         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2518                               src_res_id, LDLM_IBITS, &policy,
2519                               LCK_EX, &flags, mds_blocking_ast,
2520                               ldlm_completion_ast, NULL, NULL,
2521                               NULL, 0, NULL, &src_lockh);
2522         if (rc != ELDLM_OK)
2523                 GOTO(cleanup, rc = -ENOLCK);
2524         cleanup_phase = 2; /* lock */
2525
2526         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2527
2528         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2529         if (IS_ERR(handle)) {
2530                 rc = PTR_ERR(handle);
2531                 GOTO(cleanup, rc);
2532         }
2533         i_nlink_inc(de_src->d_inode);
2534         mark_inode_dirty(de_src->d_inode);
2535
2536         EXIT;
2537 cleanup:
2538         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2539                                 handle, req, rc, 0);
2540         switch (cleanup_phase) {
2541                 case 2:
2542                         if (rc)
2543                                 ldlm_lock_decref(&src_lockh, LCK_EX);
2544                         else
2545                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2546                 case 1:
2547                         l_dput(de_src);
2548                 case 0:
2549                         break;
2550                 default:
2551                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2552                         LBUG();
2553         }
2554         req->rq_status = rc;
2555         return 0;
2556 }
2557
2558 /*
2559  * request to link to foreign inode:
2560  *  - acquire i_nlinks on this inode
2561  *  - add dentry
2562  */
2563 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2564                                     int offset, struct ptlrpc_request *req,
2565                                     struct lustre_handle *lh)
2566 {
2567         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2568         struct obd_device *obd = req->rq_export->exp_obd;
2569         struct dentry *de_tgt_dir = NULL;
2570         struct mds_obd *mds = mds_req2mds(req);
2571         int rc = 0, cleanup_phase = 0;
2572         struct mdc_op_data *op_data;
2573         struct ptlrpc_request *request = NULL;
2574         int update_mode;
2575         ENTRY;
2576
2577         DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2578                   ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2579                   OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2580                   OLID4(rec->ur_id1));
2581
2582         de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2583                                           tgt_dir_lockh, &update_mode,
2584                                           rec->ur_name, rec->ur_namelen - 1,
2585                                           MDS_INODELOCK_UPDATE);
2586         if (IS_ERR(de_tgt_dir))
2587                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2588         cleanup_phase = 1;
2589
2590         OBD_ALLOC(op_data, sizeof(*op_data));
2591         if (op_data == NULL)
2592                 GOTO(cleanup, rc = -ENOMEM);
2593
2594         memset(op_data, 0, sizeof(*op_data));
2595         op_data->id1 = *(rec->ur_id1);
2596         rc = md_link(mds->mds_md_exp, op_data, &request);
2597         OBD_FREE(op_data, sizeof(*op_data));
2598
2599         if (request)
2600                 ptlrpc_req_finished(request);
2601         if (rc)
2602                 GOTO(cleanup, rc);
2603
2604         cleanup_phase = 2;
2605
2606         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2607
2608         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2609         if (IS_ERR(handle)) {
2610                 rc = PTR_ERR(handle);
2611                 GOTO(cleanup, rc);
2612         }
2613
2614         cleanup_phase = 3;
2615
2616         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2617                                   rec->ur_namelen - 1, id_ino(rec->ur_id1),
2618                                   id_gen(rec->ur_id1), id_group(rec->ur_id1),
2619                                   id_fid(rec->ur_id1));
2620         EXIT;
2621 cleanup:
2622         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2623                                 handle, req, rc, 0);
2624
2625         switch (cleanup_phase) {
2626                 case 3:
2627                         if (rc) {
2628                                 OBD_ALLOC(op_data, sizeof(*op_data));
2629                                 if (op_data != NULL) {
2630                                         request = NULL;
2631                                         memset(op_data, 0, sizeof(*op_data));
2632
2633                                         op_data->id1 = *(rec->ur_id1);
2634                                         op_data->create_mode = rec->ur_mode;
2635
2636                                         rc = md_unlink(mds->mds_md_exp, op_data, &request);
2637                                         OBD_FREE(op_data, sizeof(*op_data));
2638                                         if (request)
2639                                                 ptlrpc_req_finished(request);
2640                                         if (rc) {
2641                                                 CERROR("error %d while dropping i_nlink on "
2642                                                        "remote inode\n", rc);
2643                                         }
2644                                 } else {
2645                                         CERROR("rc %d prevented dropping i_nlink on "
2646                                                "remote inode\n", -ENOMEM);
2647                                 }
2648                         }
2649                 case 2:
2650                 case 1:
2651                         if (rc) {
2652                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2653 #ifdef S_PDIROPS
2654                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2655 #endif
2656                         } else {
2657                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2658 #ifdef S_PDIROPS
2659                                 ptlrpc_save_lock(req, tgt_dir_lockh + 1, update_mode);
2660 #endif
2661                         }
2662                         l_dput(de_tgt_dir);
2663                         break;
2664                 default:
2665                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2666                         LBUG();
2667         }
2668         req->rq_status = rc;
2669         return 0;
2670 }
2671
2672 static int mds_reint_link(struct mds_update_record *rec, int offset,
2673                           struct ptlrpc_request *req, struct lustre_handle *lh)
2674 {
2675         struct obd_device *obd = req->rq_export->exp_obd;
2676         struct dentry *de_src = NULL;
2677         struct dentry *de_tgt_dir = NULL;
2678         struct dentry *dchild = NULL;
2679         struct mds_obd *mds = mds_req2mds(req);
2680         struct lustre_handle *handle = NULL;
2681         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2682         struct ldlm_res_id src_res_id = { .name = {0} };
2683         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2684         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2685         ldlm_policy_data_t tgt_dir_policy =
2686                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2687         int rc = 0, cleanup_phase = 0;
2688 #ifdef S_PDIROPS
2689         int update_mode = 0;
2690 #endif
2691         ENTRY;
2692
2693         LASSERT(offset == 1);
2694
2695         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2696                   id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2697                   id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2698                   rec->ur_name);
2699
2700         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2701         MD_COUNTER_INCREMENT(obd, link);
2702
2703         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2704                 GOTO(cleanup, rc = -ENOENT);
2705
2706         if (id_group(rec->ur_id1) != mds->mds_num) {
2707                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2708                 RETURN(rc);
2709         }
2710
2711         if (rec->ur_namelen == 1) {
2712                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2713                 RETURN(rc);
2714         }
2715
2716         /* Step 1: Lookup the source inode and target directory by ID */
2717         de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2718         if (IS_ERR(de_src))
2719                 GOTO(cleanup, rc = PTR_ERR(de_src));
2720
2721         cleanup_phase = 1; /* source dentry */
2722
2723         de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2724         if (IS_ERR(de_tgt_dir)) {
2725                 rc = PTR_ERR(de_tgt_dir);
2726                 de_tgt_dir = NULL;
2727                 GOTO(cleanup, rc);
2728         }
2729
2730         cleanup_phase = 2; /* target directory dentry */
2731
2732         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2733                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2734                rec->ur_name, de_src->d_inode->i_ino);
2735
2736         /* Step 2: Take the two locks */
2737         src_res_id.name[0] = id_fid(rec->ur_id1);
2738         src_res_id.name[1] = id_group(rec->ur_id1);
2739         tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2740         tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2741
2742 #ifdef S_PDIROPS
2743         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2744                 int flags = LDLM_FL_ATOMIC_CB;
2745                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2746                 if (update_mode) {
2747                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2748                                               tgt_dir_res_id, LDLM_IBITS,
2749                                               &src_policy, update_mode, &flags,
2750                                               mds_blocking_ast,
2751                                               ldlm_completion_ast, NULL, NULL,
2752                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2753                         if (rc != ELDLM_OK)
2754                                 GOTO(cleanup, rc = -ENOLCK);
2755                 }
2756
2757                 tgt_dir_res_id.name[2] = full_name_hash((unsigned char *)rec->ur_name,
2758                                                         rec->ur_namelen - 1);
2759                 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2760                        (unsigned long)id_fid(rec->ur_id2),
2761                        (unsigned long)id_group(rec->ur_id2),
2762                        tgt_dir_res_id.name[2]);
2763         }
2764 #endif
2765         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2766                                    &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2767                                    LCK_EX, &tgt_dir_policy);
2768         if (rc)
2769                 GOTO(cleanup, rc);
2770
2771         cleanup_phase = 3; /* locks */
2772
2773         /* Step 3: Lookup the child */
2774         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2775                                    rec->ur_namelen - 1);
2776         if (IS_ERR(dchild)) {
2777                 rc = PTR_ERR(dchild);
2778                 if (rc != -EPERM && rc != -EACCES)
2779                         CERROR("child lookup error %d\n", rc);
2780                 GOTO(cleanup, rc);
2781         }
2782
2783         cleanup_phase = 4; /* child dentry */
2784
2785         if (dchild->d_inode) {
2786                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2787                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2788                 rc = -EEXIST;
2789                 GOTO(cleanup, rc);
2790         }
2791
2792         /* Step 4: Do it. */
2793         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2794
2795         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2796         if (IS_ERR(handle)) {
2797                 rc = PTR_ERR(handle);
2798                 GOTO(cleanup, rc);
2799         }
2800
2801         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2802         if (rc && rc != -EPERM && rc != -EACCES)
2803                 CERROR("vfs_link error %d\n", rc);
2804 cleanup:
2805         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2806                                 handle, req, rc, 0);
2807         EXIT;
2808
2809         switch (cleanup_phase) {
2810         case 4: /* child dentry */
2811                 l_dput(dchild);
2812         case 3: /* locks */
2813                 if (rc) {
2814                         ldlm_lock_decref(&src_lockh, LCK_EX);
2815                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2816                 } else {
2817                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2818                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2819                 }
2820         case 2: /* target dentry */
2821 #ifdef S_PDIROPS
2822                 if (tgt_dir_lockh[1].cookie && update_mode)
2823                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2824 #endif
2825                 if (de_tgt_dir)
2826                         l_dput(de_tgt_dir);
2827         case 1: /* source dentry */
2828                 l_dput(de_src);
2829         case 0:
2830                 break;
2831         default:
2832                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2833                 LBUG();
2834         }
2835         req->rq_status = rc;
2836         return 0;
2837 }
2838
2839 /* The idea here is that we need to get four locks in the end:
2840  * one on each parent directory, one on each child.  We need to take
2841  * these locks in some kind of order (to avoid deadlocks), and the order
2842  * I selected is "increasing resource number" order.  We need to look up
2843  * the children, however, before we know what the resource number(s) are.
2844  * Thus the following plan:
2845  *
2846  * 1,2. Look up the parents
2847  * 3,4. Look up the children
2848  * 5. Take locks on the parents and children, in order
2849  * 6. Verify that the children haven't changed since they were looked up
2850  *
2851  * If there was a race and the children changed since they were first looked
2852  * up, it is possible that mds_verify_child() will be able to just grab the
2853  * lock on the new child resource (if it has a higher resource than any other)
2854  * but we need to compare against not only its parent, but also against the
2855  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2856  *
2857  * We need the fancy igrab() on the child inodes because we aren't holding a
2858  * lock on the parent after the lookup is done, so dentry->d_inode may change
2859  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2860  */
2861 static int mds_get_parents_children_locked(struct obd_device *obd,
2862                                            struct mds_obd *mds,
2863                                            struct lustre_id *p1_id,
2864                                            struct dentry **de_srcdirp,
2865                                            struct lustre_id *p2_id,
2866                                            struct dentry **de_tgtdirp,
2867                                            int parent_mode,
2868                                            const char *old_name, int old_len,
2869                                            struct dentry **de_oldp,
2870                                            const char *new_name, int new_len,
2871                                            struct dentry **de_newp,
2872                                            struct lustre_handle *dlm_handles,
2873                                            int child_mode)
2874 {
2875         struct ldlm_res_id p1_res_id = { .name = {0} };
2876         struct ldlm_res_id p2_res_id = { .name = {0} };
2877         struct ldlm_res_id c1_res_id = { .name = {0} };
2878         struct ldlm_res_id c2_res_id = { .name = {0} };
2879         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2880         /* Only dentry should disappear, but the inode itself would be
2881            intact otherwise. */
2882         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2883         /* If something is going to be replaced, both dentry and inode locks are
2884            needed */
2885         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2886         struct ldlm_res_id *maxres_src, *maxres_tgt;
2887         struct inode *inode;
2888         int rc = 0, cleanup_phase = 0;
2889         __u32 child_gen1 = 0;
2890         __u32 child_gen2 = 0;
2891         unsigned long child_ino1 = 0;
2892         unsigned long child_ino2 = 0;
2893         ENTRY;
2894
2895         /* Step 1: Lookup the source directory */
2896         *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2897         if (IS_ERR(*de_srcdirp))
2898                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2899
2900         cleanup_phase = 1; /* source directory dentry */
2901
2902         p1_res_id.name[0] = id_fid(p1_id);
2903         p1_res_id.name[1] = id_group(p1_id);
2904
2905         /* Step 2: Lookup the target directory */
2906         if (id_equal_stc(p1_id, p2_id)) {
2907                 *de_tgtdirp = dget(*de_srcdirp);
2908         } else {
2909                 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2910                 if (IS_ERR(*de_tgtdirp)) {
2911                         rc = PTR_ERR(*de_tgtdirp);
2912                         *de_tgtdirp = NULL;
2913                         GOTO(cleanup, rc);
2914                 }
2915         }
2916
2917         cleanup_phase = 2; /* target directory dentry */
2918
2919         p2_res_id.name[0] = id_fid(p2_id);
2920         p2_res_id.name[1] = id_group(p2_id);
2921
2922 #ifdef S_PDIROPS
2923         dlm_handles[5].cookie = 0;
2924         dlm_handles[6].cookie = 0;
2925
2926         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2927                 /*
2928                  * get a temp lock on just fid, group to flush client cache and
2929                  * to protect dirs from concurrent splitting.
2930                  */
2931                 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2932                                            LCK_PW, &p_policy, &p2_res_id,
2933                                            &dlm_handles[6], LCK_PW, &p_policy);
2934                 if (rc != ELDLM_OK)
2935                         GOTO(cleanup, rc);
2936
2937                 p1_res_id.name[2] = full_name_hash((unsigned char *)old_name,
2938                                                    old_len - 1);
2939                 p2_res_id.name[2] = full_name_hash((unsigned char *)new_name,
2940                                                    new_len - 1);
2941
2942                 CDEBUG(D_INFO, "take locks on "
2943                        LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2944                        p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2945                        p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2946         }
2947         cleanup_phase = 3;
2948 #endif
2949
2950         /* Step 3: Lookup the source child entry */
2951         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2952                                      old_len - 1);
2953         if (IS_ERR(*de_oldp)) {
2954                 rc = PTR_ERR(*de_oldp);
2955                 CERROR("old child lookup error (%.*s): %d\n",
2956                        old_len - 1, old_name, rc);
2957                 GOTO(cleanup, rc);
2958         }
2959
2960         cleanup_phase = 4; /* original name dentry */
2961
2962         inode = (*de_oldp)->d_inode;
2963         if (inode != NULL) {
2964                 struct lustre_id sid;
2965
2966                 inode = igrab(inode);
2967                 if (inode == NULL)
2968                         GOTO(cleanup, rc = -ENOENT);
2969
2970                 down(&inode->i_sem);
2971                 rc = mds_read_inode_sid(obd, inode, &sid);
2972                 up(&inode->i_sem);
2973                 if (rc) {
2974                         CERROR("Can't read inode self id, inode %lu, "
2975                                "rc %d\n", inode->i_ino, rc);
2976                         iput(inode);
2977                         GOTO(cleanup, rc);
2978                 }
2979
2980                 child_ino1 = inode->i_ino;
2981                 child_gen1 = inode->i_generation;
2982                 c1_res_id.name[0] = id_fid(&sid);
2983                 c1_res_id.name[1] = id_group(&sid);
2984                 iput(inode);
2985         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2986                 child_ino1 = (*de_oldp)->d_inum;
2987                 child_gen1 = (*de_oldp)->d_generation;
2988                 c1_res_id.name[0] = (*de_oldp)->d_fid;
2989                 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2990         } else {
2991                 GOTO(cleanup, rc = -ENOENT);
2992         }
2993
2994         /* Step 4: Lookup the target child entry */
2995         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2996                                      new_len - 1);
2997         if (IS_ERR(*de_newp)) {
2998                 rc = PTR_ERR(*de_newp);
2999                 CERROR("new child lookup error (%.*s): %d\n",
3000                        old_len - 1, old_name, rc);
3001                 GOTO(cleanup, rc);
3002         }
3003
3004         cleanup_phase = 5; /* target dentry */
3005
3006         inode = (*de_newp)->d_inode;
3007         if (inode != NULL) {
3008                 struct lustre_id sid;
3009
3010                 inode = igrab(inode);
3011                 if (inode == NULL)
3012                         goto retry_locks;
3013
3014                 down(&inode->i_sem);
3015                 rc = mds_read_inode_sid(obd, inode, &sid);
3016                 up(&inode->i_sem);
3017                 if (rc) {
3018                         CERROR("Can't read inode self id, inode %lu, "
3019                                "rc %d\n", inode->i_ino, rc);
3020                         GOTO(cleanup, rc);
3021                 }
3022
3023                 child_ino2 = inode->i_ino;
3024                 child_gen2 = inode->i_generation;
3025                 c2_res_id.name[0] = id_fid(&sid);
3026                 c2_res_id.name[1] = id_group(&sid);
3027                 iput(inode);
3028         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
3029                 child_ino2 = (*de_newp)->d_inum;
3030                 child_gen2 = (*de_newp)->d_generation;
3031                 c2_res_id.name[0] = (*de_newp)->d_fid;
3032                 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
3033         }
3034
3035 retry_locks:
3036         /* Step 5: Take locks on the parents and child(ren) */
3037         maxres_src = &p1_res_id;
3038         maxres_tgt = &p2_res_id;
3039         cleanup_phase = 5; /* target dentry */
3040
3041         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
3042                 maxres_src = &c1_res_id;
3043         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
3044                 maxres_tgt = &c2_res_id;
3045
3046         rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
3047                                     &p_policy,
3048                                     &p2_res_id, &dlm_handles[1], parent_mode,
3049                                     &p_policy,
3050                                     &c1_res_id, &dlm_handles[2], child_mode,
3051                                     &c1_policy,
3052                                     &c2_res_id, &dlm_handles[3], child_mode,
3053                                     &c2_policy);
3054         if (rc)
3055                 GOTO(cleanup, rc);
3056
3057         cleanup_phase = 6; /* parent and child(ren) locks */
3058
3059         /* Step 6a: Re-lookup source child to verify it hasn't changed */
3060         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
3061                               parent_mode, &c1_res_id, &dlm_handles[2],
3062                               de_oldp, child_mode, &c1_policy, old_name, old_len,
3063                               maxres_tgt, child_ino1, child_gen1);
3064         if (rc) {
3065                 if (c2_res_id.name[0] != 0)
3066                         ldlm_lock_decref(&dlm_handles[3], child_mode);
3067                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3068                 cleanup_phase = 5;
3069                 if (rc > 0)
3070                         goto retry_locks;
3071                 GOTO(cleanup, rc);
3072         }
3073
3074         if (!DENTRY_VALID(*de_oldp))
3075                 GOTO(cleanup, rc = -ENOENT);
3076
3077         /* Step 6b: Re-lookup target child to verify it hasn't changed */
3078         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
3079                               parent_mode, &c2_res_id, &dlm_handles[3],
3080                               de_newp, child_mode, &c2_policy, new_name,
3081                               new_len, maxres_src, child_ino2, child_gen2);
3082         if (rc) {
3083                 ldlm_lock_decref(&dlm_handles[2], child_mode);
3084                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3085                 cleanup_phase = 5;
3086                 if (rc > 0)
3087                         goto retry_locks;
3088                 GOTO(cleanup, rc);
3089         }
3090
3091         EXIT;
3092 cleanup:
3093         if (rc) {
3094                 switch (cleanup_phase) {
3095                 case 6: /* child lock(s) */
3096                         if (c2_res_id.name[0] != 0)
3097                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
3098                         if (c1_res_id.name[0] != 0)
3099                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
3100                         if (dlm_handles[1].cookie != 0)
3101                                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
3102                         if (dlm_handles[0].cookie != 0)
3103                                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
3104                 case 5: /* target dentry */
3105                         l_dput(*de_newp);
3106                 case 4: /* source dentry */
3107                         l_dput(*de_oldp);
3108                 case 3:
3109 #ifdef S_PDIROPS
3110                         if (dlm_handles[5].cookie != 0)
3111                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3112                         if (dlm_handles[6].cookie != 0)
3113                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3114 #endif
3115                 case 2: /* target directory dentry */
3116                         l_dput(*de_tgtdirp);
3117                 case 1: /* source directry dentry */
3118                         l_dput(*de_srcdirp);
3119                 }
3120         }
3121
3122         return rc;
3123 }
3124
3125 /*
3126  * checks if dentry can be removed. This function also handles cross-ref
3127  * dentries.
3128  */
3129 static int mds_check_for_rename(struct obd_device *obd,
3130                                 struct dentry *dentry)
3131 {
3132         struct mds_obd *mds = &obd->u.mds;
3133         struct lustre_handle *rlockh;
3134         struct ptlrpc_request *req;
3135         struct mdc_op_data *op_data;
3136         struct lookup_intent it;
3137         int handle_size, rc = 0;
3138         ENTRY;
3139
3140         LASSERT(dentry != NULL);
3141
3142         if (dentry->d_inode) {
3143                 if (S_ISDIR(dentry->d_inode->i_mode) &&
3144                     !mds_is_dir_empty(obd, dentry))
3145                         rc = -ENOTEMPTY;
3146         } else {
3147                 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
3148                 handle_size = sizeof(struct lustre_handle);
3149
3150                 OBD_ALLOC(rlockh, handle_size);
3151                 if (rlockh == NULL)
3152                         RETURN(-ENOMEM);
3153
3154                 memset(rlockh, 0, handle_size);
3155                 OBD_ALLOC(op_data, sizeof(*op_data));
3156                 if (op_data == NULL) {
3157                         OBD_FREE(rlockh, handle_size);
3158                         RETURN(-ENOMEM);
3159                 }
3160                 memset(op_data, 0, sizeof(*op_data));
3161                 mds_pack_dentry2id(obd, &op_data->id1, dentry, 1);
3162
3163                 it.it_op = IT_UNLINK;
3164                 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
3165                 if (!it.d.fs_data)
3166                         RETURN(-ENOMEM);
3167                 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
3168                                 op_data, rlockh, NULL, 0, ldlm_completion_ast,
3169                                 mds_blocking_ast, NULL);
3170                 OBD_FREE(op_data, sizeof(*op_data));
3171
3172
3173                 if (rc) {
3174                         OBD_FREE(it.d.fs_data,
3175                                  sizeof(struct lustre_intent_data));
3176                         RETURN(rc);
3177                 }
3178                 if (rlockh->cookie != 0)
3179                         ldlm_lock_decref(rlockh, LCK_EX);
3180
3181                 if (LUSTRE_IT(&it)->it_data) {
3182                         req = (struct ptlrpc_request *)LUSTRE_IT(&it)->it_data;
3183                         ptlrpc_req_finished(req);
3184                 }
3185
3186                 if (LUSTRE_IT(&it)->it_status)
3187                         rc = LUSTRE_IT(&it)->it_status;
3188                 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
3189                 OBD_FREE(rlockh, handle_size);
3190         }
3191         RETURN(rc);
3192 }
3193
3194 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
3195                                 struct ptlrpc_request *req, struct lustre_id *id,
3196                                 struct dentry *de_dir, struct dentry *de)
3197 {
3198         struct obd_device *obd = req->rq_export->exp_obd;
3199         struct mds_obd *mds = mds_req2mds(req);
3200         void *handle = NULL;
3201         int rc = 0;
3202         ENTRY;
3203
3204         if (de->d_inode) {
3205                 /*
3206                  * name exists and points to local inode try to unlink this name
3207                  * and create new one.
3208                  */
3209                 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
3210                        obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
3211                        (unsigned long)de->d_inode->i_generation);
3212
3213                 /* checking if we can remove local dentry. */
3214                 rc = mds_check_for_rename(obd, de);
3215                 if (rc)
3216                         GOTO(cleanup, rc);
3217
3218                 handle = fsfilt_start(obd, de_dir->d_inode,
3219                                       FSFILT_OP_RENAME, NULL);
3220                 if (IS_ERR(handle))
3221                         GOTO(cleanup, rc = PTR_ERR(handle));
3222                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3223                 if (rc)
3224                         GOTO(cleanup, rc);
3225         } else if (de->d_flags & DCACHE_CROSS_REF) {
3226                 CDEBUG(D_OTHER, "%s: %s points to remote inode %lu/%lu\n",
3227                        obd->obd_name, rec->ur_tgt, (unsigned long)de->d_mdsnum,
3228                         (unsigned long)de->d_fid);
3229
3230                 /* checking if we can remove local dentry. */
3231                 rc = mds_check_for_rename(obd, de);
3232                 if (rc)
3233                         GOTO(cleanup, rc);
3234
3235                 /*
3236                  * to be fully POSIX compatible, we should add one more check:
3237                  *
3238                  * if de_new is subdir of dir rec->ur_id1. If so - return
3239                  * -EINVAL.
3240                  *
3241                  * I do not know how to implement it right now, because
3242                  * inodes/dentries for new and old names lie on different MDS,
3243                  * so add this notice here just to make it visible for the rest
3244                  * of developers and do not forget about. And when this check
3245                  * will be added, del_cross_ref should gone, that is local
3246                  * dentry is able to be removed if all checks passed. --umka
3247                  */
3248
3249                 handle = fsfilt_start(obd, de_dir->d_inode,
3250                                       FSFILT_OP_RENAME, NULL);
3251                 if (IS_ERR(handle))
3252                         GOTO(cleanup, rc = PTR_ERR(handle));
3253                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
3254                 if (rc)
3255                         GOTO(cleanup, rc);
3256         } else {
3257                 /* name doesn't exist. the simplest case. */
3258                 handle = fsfilt_start(obd, de_dir->d_inode,
3259                                       FSFILT_OP_LINK, NULL);
3260                 if (IS_ERR(handle))
3261                         GOTO(cleanup, rc = PTR_ERR(handle));
3262         }
3263
3264         rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
3265                                   rec->ur_tgtlen - 1, id_ino(id),
3266                                   id_gen(id), id_group(id), id_fid(id));
3267         if (rc) {
3268                 CERROR("add_dir_entry() returned error %d\n", rc);
3269                 GOTO(cleanup, rc);
3270         }
3271
3272         EXIT;
3273 cleanup:
3274         rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3275                                 handle, req, rc, 0);
3276
3277         return rc;
3278 }
3279
3280 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
3281                                 struct ptlrpc_request *req, struct dentry *de_dir,
3282                                 struct dentry *de)
3283 {
3284         struct obd_device *obd = req->rq_export->exp_obd;
3285         struct mds_obd *mds = mds_req2mds(req);
3286         void *handle = NULL;
3287         int rc = 0;
3288         ENTRY;
3289
3290         handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
3291         if (IS_ERR(handle))
3292                 GOTO(cleanup, rc = PTR_ERR(handle));
3293         rc = fsfilt_del_dir_entry(obd, de);
3294         d_drop(de);
3295
3296         EXIT;
3297 cleanup:
3298         rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
3299                                 handle, req, rc, 0);
3300         return rc;
3301 }
3302
3303 static int mds_reint_rename_create_name(struct mds_update_record *rec,
3304                                         int offset, struct ptlrpc_request *req)
3305 {
3306         struct lustre_handle parent_lockh[2] = {{0}, {0}};
3307         struct obd_device *obd = req->rq_export->exp_obd;
3308         struct mds_obd *mds = mds_req2mds(req);
3309         struct lustre_handle child_lockh = {0};
3310         struct dentry *de_tgtdir = NULL;
3311         struct dentry *de_new = NULL;
3312         int cleanup_phase = 0;
3313         int update_mode, rc = 0;
3314         struct lustre_id ids[2]; /* sid, pid */
3315         struct obd_export *set_exp;
3316         ENTRY;
3317
3318         /*
3319          * another MDS executing rename operation has asked us to create target
3320          * name. such a creation should destroy existing target name.
3321          */
3322         CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
3323                obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
3324
3325         /* first, lookup the target */
3326         rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
3327                                          &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
3328                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
3329                                          &child_lockh, &de_new, LCK_EX,
3330                                          MDS_INODELOCK_LOOKUP);
3331         if (rc)
3332                 GOTO(cleanup, rc);
3333
3334         /* get parent id: ldlm lock on the parent protects ea */
3335         rc = mds_read_inode_sid(obd, de_tgtdir->d_inode, &ids[1]);
3336
3337         if (rc)
3338                 GOTO(cleanup, rc);
3339         cleanup_phase = 1;
3340
3341         LASSERT(de_tgtdir);
3342         LASSERT(de_tgtdir->d_inode);
3343         LASSERT(de_new);
3344
3345         rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
3346                                   de_tgtdir, de_new);
3347         if (rc)
3348                 GOTO(cleanup, rc);
3349
3350         ids[0] = *(rec->ur_id1);
3351         if (id_group(ids) == mds->mds_num)
3352                 set_exp = req->rq_export;
3353         else
3354                 set_exp = mds->mds_md_exp;
3355         rc = obd_set_info(set_exp, strlen("ids"), "ids",
3356                           sizeof(struct lustre_id) * 2, ids);
3357
3358         EXIT;
3359 cleanup:
3360
3361         if (cleanup_phase == 1) {
3362 #ifdef S_PDIROPS
3363                 if (parent_lockh[1].cookie != 0)
3364                         ldlm_lock_decref(parent_lockh + 1, update_mode);
3365 #endif
3366                 ldlm_lock_decref(parent_lockh, LCK_PW);
3367                 if (child_lockh.cookie != 0)
3368                         ldlm_lock_decref(&child_lockh, LCK_EX);
3369                 l_dput(de_new);
3370                 l_dput(de_tgtdir);
3371         }
3372
3373         req->rq_status = rc;
3374         return 0;
3375 }
3376
3377 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3378                                       struct ptlrpc_request *req)
3379 {
3380         struct obd_device *obd = req->rq_export->exp_obd;
3381         struct ptlrpc_request *req2 = NULL;
3382         struct dentry *de_srcdir = NULL;
3383         struct dentry *de_old = NULL;
3384         struct mds_obd *mds = mds_req2mds(req);
3385         struct lustre_handle parent_lockh[2] = {{0}, {0}};
3386         struct lustre_handle child_lockh = {0};
3387         struct mdc_op_data *op_data;
3388         int update_mode, rc = 0;
3389         ENTRY;
3390
3391         CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3392                obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3393
3394         OBD_ALLOC(op_data, sizeof(*op_data));
3395         if (op_data == NULL)
3396                 RETURN(-ENOMEM);
3397         memset(op_data, 0, sizeof(*op_data));
3398
3399         rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3400                                          &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3401                                          &update_mode, rec->ur_name,
3402                                          rec->ur_namelen, &child_lockh, &de_old,
3403                                          LCK_EX, MDS_INODELOCK_LOOKUP);
3404         if (rc) {
3405                 OBD_FREE(op_data, sizeof(*op_data));
3406                 RETURN(rc);
3407         }
3408
3409         LASSERT(de_srcdir);
3410         LASSERT(de_srcdir->d_inode);
3411         LASSERT(de_old);
3412
3413         /*
3414          * we already know the target should be created on another MDS so, we
3415          * have to request that MDS to do it.
3416          */
3417
3418         /* prepare source id */
3419         if (de_old->d_flags & DCACHE_CROSS_REF) {
3420                 LASSERT(de_old->d_inode == NULL);
3421                 CDEBUG(D_OTHER, "request to move remote name\n");
3422                 mds_pack_dentry2id(obd, &op_data->id1, de_old, 1);
3423         } else if (de_old->d_inode == NULL) {
3424                 /* oh, source doesn't exist */
3425                 OBD_FREE(op_data, sizeof(*op_data));
3426                 GOTO(cleanup, rc = -ENOENT);
3427         } else {
3428                 struct lustre_id sid;
3429                 struct inode *inode = de_old->d_inode;
3430
3431                 LASSERT(inode != NULL);
3432                 CDEBUG(D_OTHER, "request to move local name\n");
3433                 id_ino(&op_data->id1) = inode->i_ino;
3434                 id_group(&op_data->id1) = mds->mds_num;
3435                 id_gen(&op_data->id1) = inode->i_generation;
3436
3437                 down(&inode->i_sem);
3438                 rc = mds_read_inode_sid(obd, inode, &sid);
3439                 up(&inode->i_sem);
3440                 if (rc) {
3441                         CERROR("Can't read inode self id, "
3442                                "inode %lu, rc = %d\n",
3443                                inode->i_ino, rc);
3444                         GOTO(cleanup, rc);
3445                 }
3446
3447                 id_fid(&op_data->id1) = id_fid(&sid);
3448         }
3449
3450         op_data->id2 = *rec->ur_id2;
3451         rc = md_rename(mds->mds_md_exp, op_data, NULL, 0,
3452                        rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3453         OBD_FREE(op_data, sizeof(*op_data));
3454
3455         if (rc)
3456                 GOTO(cleanup, rc);
3457
3458         rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3459                                   de_old);
3460
3461         EXIT;
3462 cleanup:
3463         if (req2)
3464                 ptlrpc_req_finished(req2);
3465
3466 #ifdef S_PDIROPS
3467         if (parent_lockh[1].cookie != 0)
3468                 ldlm_lock_decref(parent_lockh + 1, update_mode);
3469 #endif
3470         ldlm_lock_decref(parent_lockh, LCK_PW);
3471         if (child_lockh.cookie != 0)
3472                 ldlm_lock_decref(&child_lockh, LCK_EX);
3473
3474         l_dput(de_old);
3475         l_dput(de_srcdir);
3476
3477         req->rq_status = rc;
3478         return 0;
3479 }
3480
3481 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3482                             struct ptlrpc_request *req, struct lustre_handle *lockh)
3483 {
3484         struct obd_device *obd = req->rq_export->exp_obd;
3485         struct dentry *de_srcdir = NULL;
3486         struct dentry *de_tgtdir = NULL;
3487         struct dentry *de_old = NULL;
3488         struct dentry *de_new = NULL;
3489         struct inode *old_inode = NULL, *new_inode = NULL;
3490         struct mds_obd *mds = mds_req2mds(req);
3491         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3492         struct mds_body *body = NULL;
3493         struct llog_create_locks *lcl = NULL;
3494         struct lov_mds_md *lmm = NULL;
3495         int rc = 0, cleanup_phase = 0;
3496         struct lustre_id ids[2];  /* sid, pid */
3497         void *handle = NULL;
3498         ENTRY;
3499
3500         LASSERT(offset == 1);
3501
3502         DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3503                   OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3504                   rec->ur_tgt);
3505
3506         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3507
3508         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3509                 DEBUG_REQ(D_HA, req, "rename replay");
3510                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3511                        lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3512                        req->rq_repmsg->buflens[2]);
3513         }
3514
3515         MD_COUNTER_INCREMENT(obd, rename);
3516
3517         if (rec->ur_namelen == 1) {
3518                 rc = mds_reint_rename_create_name(rec, offset, req);
3519                 RETURN(rc);
3520         }
3521
3522         /* check if new name should be located on remote target. */
3523         if (id_group(rec->ur_id2) != mds->mds_num) {
3524                 rc = mds_reint_rename_to_remote(rec, offset, req);
3525                 RETURN(rc);
3526         }
3527
3528         rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3529                                              rec->ur_id2, &de_tgtdir, LCK_PW,
3530                                              rec->ur_name, rec->ur_namelen,
3531                                              &de_old, rec->ur_tgt,
3532                                              rec->ur_tgtlen, &de_new,
3533                                              dlm_handles, LCK_EX);
3534         if (rc)
3535                 GOTO(cleanup, rc);
3536
3537         cleanup_phase = 1; /* parent(s), children, locks */
3538         old_inode = de_old->d_inode;
3539         new_inode = de_new->d_inode;
3540
3541         /* sanity check for src inode */
3542         if (de_old->d_flags & DCACHE_CROSS_REF) {
3543                 LASSERT(de_old->d_inode == NULL);
3544
3545                 /*
3546                  * in the case of cross-ref dir, we can perform this check only
3547                  * if child and parent lie on the same mds. This is because
3548                  * otherwise they can have the same inode numbers.
3549                  */
3550                 if (de_old->d_mdsnum == mds->mds_num) {
3551                         if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3552                             de_old->d_inum == de_tgtdir->d_inode->i_ino)
3553                                 GOTO(cleanup, rc = -EINVAL);
3554                 }
3555         } else {
3556                 LASSERT(de_old->d_inode != NULL);
3557                 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3558                     de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3559                         GOTO(cleanup, rc = -EINVAL);
3560         }
3561
3562         /* sanity check for dest inode */
3563         if (de_new->d_flags & DCACHE_CROSS_REF) {
3564                 LASSERT(new_inode == NULL);
3565
3566                 /* the same check about target dentry. */
3567                 if (de_new->d_mdsnum == mds->mds_num) {
3568                         if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3569                             de_new->d_inum == de_tgtdir->d_inode->i_ino)
3570                                 GOTO(cleanup, rc = -EINVAL);
3571                 }
3572
3573                 /*
3574                  * regular files usualy do not have ->rename() implemented. But
3575                  * we handle only this case when @de_new is cross-ref entry,
3576                  * because in other cases it will be handled by vfs_rename().
3577                  */
3578                 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3579                     !de_old->d_inode->i_op->rename))
3580                         GOTO(cleanup, rc = -EPERM);
3581         } else {
3582                 if (new_inode &&
3583                     (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3584                      new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3585                         GOTO(cleanup, rc = -EINVAL);
3586
3587         }
3588
3589         /*
3590          * check if inodes point to each other. This should be checked before
3591          * is_subdir() check, as for the same entries it will think that they
3592          * are subdirs.
3593          */
3594         if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3595             !(de_new->d_flags & DCACHE_CROSS_REF) &&
3596             old_inode == new_inode)
3597                 GOTO(cleanup, rc = 0);
3598
3599 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3600         /*
3601          * check if we are moving old entry into its child. 2.6 does not check
3602          * for this in vfs_rename() anymore.
3603          */
3604         if (is_subdir(de_new, de_old))
3605                 GOTO(cleanup, rc = -EINVAL);
3606 #endif
3607
3608         /*
3609          * if we are about to remove the target at first, pass the EA of that
3610          * inode to client to perform and cleanup on OST.
3611          */
3612         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3613         LASSERT(body != NULL);
3614
3615         /* get new parent id: ldlm lock on the parent protects ea */
3616         rc = mds_read_inode_sid(obd, de_tgtdir->d_inode, &ids[1]);
3617         if (rc)
3618                 GOTO(cleanup, rc);
3619
3620         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3621         if (new_inode)
3622                 DOWN_READ_I_ALLOC_SEM(new_inode);
3623
3624         cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3625
3626         if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3627             new_inode->i_nlink == 2) ||
3628             new_inode->i_nlink == 1)) {
3629                 if (mds_orphan_open_count(new_inode) > 0) {
3630                         /* need to lock pending_dir before transaction */
3631                         down(&mds->mds_pending_dir->d_inode->i_sem);
3632                         cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3633                 } else if (S_ISREG(new_inode->i_mode)) {
3634                         mds_pack_inode2body(obd, body, new_inode, 0);
3635                         mds_pack_md(obd, req->rq_repmsg, 1, body,
3636                                     new_inode, MDS_PACK_MD_LOCK, 0);
3637                  }
3638         }
3639
3640         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3641                        de_srcdir->d_inode->i_sb);
3642
3643         if (de_old->d_flags & DCACHE_CROSS_REF) {
3644                 struct lustre_id old_id;
3645                 struct obd_export *set_exp;
3646
3647
3648                 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3649
3650                 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3651                                           de_tgtdir, de_new);
3652                 if (rc)
3653                         GOTO(cleanup, rc);
3654
3655                 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3656                                           de_old);
3657                 if (rc)
3658                         GOTO(cleanup, rc);
3659
3660                 ids[0] = old_id;
3661                 if (id_group(ids) == mds->mds_num)
3662                         set_exp = req->rq_export;
3663                 else
3664                         set_exp = mds->mds_md_exp;
3665                 rc = obd_set_info(set_exp, strlen("ids"), "ids",
3666                                   sizeof(struct lustre_id) * 2, ids);
3667
3668                 GOTO(cleanup, rc);
3669         }
3670
3671         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3672         handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3673                                   NULL, le32_to_cpu(lmm->lmm_stripe_count));
3674
3675         if (IS_ERR(handle))
3676                 GOTO(cleanup, rc = PTR_ERR(handle));
3677
3678         lock_kernel();
3679         de_old->d_fsdata = req;
3680         de_new->d_fsdata = req;
3681         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3682         unlock_kernel();
3683
3684         if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3685                 if (mds_orphan_open_count(new_inode) > 0)
3686                         rc = mds_orphan_add_link(rec, obd, de_new);
3687
3688                 if (rc == 1)
3689                         GOTO(cleanup, rc = 0);
3690
3691                 if (!S_ISREG(new_inode->i_mode))
3692                         GOTO(cleanup, rc);
3693
3694                 if (!(body->valid & OBD_MD_FLEASIZE)) {
3695                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3696                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
3697                 } else if (mds_log_op_unlink(obd, new_inode,
3698                                              lustre_msg_buf(req->rq_repmsg,1,0),
3699                                              req->rq_repmsg->buflens[1],
3700                                              lustre_msg_buf(req->rq_repmsg,2,0),
3701                                              req->rq_repmsg->buflens[2],
3702                                              &lcl) > 0) {
3703                         body->valid |= OBD_MD_FLCOOKIE;
3704                 }
3705
3706                 rc = mds_destroy_object(obd, old_inode, 1);
3707                 if (rc) {
3708                         CERROR("can't remove OST object, err %d\n",
3709                                rc);
3710                 }
3711         }
3712
3713         if (rc == 0)
3714                rc = mds_update_inode_ids(obd, de_old->d_inode,
3715                                          handle, NULL, &ids[1]);
3716
3717         EXIT;
3718 cleanup:
3719         rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3720                                 handle, req, rc, 0);
3721
3722         switch (cleanup_phase) {
3723         case 3:
3724                 up(&mds->mds_pending_dir->d_inode->i_sem);
3725         case 2:
3726                 if (new_inode)
3727                         UP_READ_I_ALLOC_SEM(new_inode);
3728         case 1:
3729 #ifdef S_PDIROPS
3730                 if (dlm_handles[5].cookie != 0)
3731                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3732                 if (dlm_handles[6].cookie != 0)
3733                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3734 #endif
3735                 if (lcl != NULL)
3736                         ptlrpc_save_llog_lock(req, lcl);
3737
3738                 if (rc) {
3739                         if (dlm_handles[3].cookie != 0)
3740                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3741                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3742                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3743                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3744                 } else {
3745                         if (dlm_handles[3].cookie != 0)
3746                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3747                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3748                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3749                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3750                 }
3751                 l_dput(de_new);
3752                 l_dput(de_old);
3753                 l_dput(de_tgtdir);
3754                 l_dput(de_srcdir);
3755         case 0:
3756                 break;
3757         default:
3758                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3759                 LBUG();
3760         }
3761         req->rq_status = rc;
3762         return 0;
3763 }
3764
3765 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3766                            struct ptlrpc_request *, struct lustre_handle *);
3767
3768 static mds_reinter reinters[REINT_MAX + 1] = {
3769         [REINT_SETATTR] mds_reint_setattr,
3770         [REINT_CREATE] mds_reint_create,
3771         [REINT_LINK] mds_reint_link,
3772         [REINT_UNLINK] mds_reint_unlink,
3773         [REINT_RENAME] mds_reint_rename,
3774         [REINT_OPEN] mds_open
3775 };
3776
3777 int mds_reint_rec(struct mds_update_record *rec, int offset,
3778                   struct ptlrpc_request *req, struct lustre_handle *lockh)
3779 {
3780         struct obd_device *obd = req->rq_export->exp_obd;
3781         struct lvfs_run_ctxt saved;
3782         int rc;
3783
3784         /* checked by unpacker */
3785         LASSERT(rec->ur_opcode <= REINT_MAX &&
3786                 reinters[rec->ur_opcode] != NULL);
3787
3788         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3789         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3790         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3791
3792         return rc;
3793 }