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