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