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