lustre/mds/mds_reint.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/mds_reint.c
   5  *  Lustre Metadata Server (mds) reintegration routines
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #ifndef EXPORT_SYMTAB
  29 # define EXPORT_SYMTAB
  30 #endif
  31 #define DEBUG_SUBSYSTEM S_MDS
  32
  33 #include <linux/fs.h>
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/obd.h>
  37 #include <linux/lustre_lib.h>
  38 #include <linux/lustre_idl.h>
  39 #include <linux/lustre_mds.h>
  40 #include <linux/lustre_dlm.h>
  41 #include <linux/lustre_log.h>
  42 #include <linux/lustre_fsfilt.h>
  43
  44 #include "mds_internal.h"
  45
  46 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  47                    int error)
  48 {
  49         obd_transno_commit_cb(obd, transno, error);
  50 }
  51
  52 struct mds_logcancel_data {
  53         struct lov_mds_md      *mlcd_lmm;
  54         int                     mlcd_size;
  55         int                     mlcd_cookielen;
  56         int                     mlcd_eadatalen;
  57         struct llog_cookie      mlcd_cookies[0];
  58 };
  59
  60
  61 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  62                                   void *cb_data, int error)
  63 {
  64         struct mds_logcancel_data *mlcd = cb_data;
  65         struct lov_stripe_md *lsm = NULL;
  66         struct llog_ctxt *ctxt;
  67         int rc;
  68
  69         obd_transno_commit_cb(obd, transno, error);
  70
  71         CDEBUG(D_HA, "cancelling %d cookies\n",
  72                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  73
  74         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  75                           mlcd->mlcd_eadatalen);
  76         if (rc < 0) {
  77                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  78                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  79                        rc);
  80         } else {
  81                 ///* XXX 0 normally, SENDNOW for debug */);
  82                 ctxt = llog_get_context(&obd->obd_llogs,
  83                                         mlcd->mlcd_cookies[0].lgc_subsys + 1);
  84                 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
  85                                  sizeof(*mlcd->mlcd_cookies),
  86                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
  87                 if (rc)
  88                         CERROR("error cancelling %d log cookies: rc %d\n",
  89                                (int)(mlcd->mlcd_cookielen /
  90                                      sizeof(*mlcd->mlcd_cookies)), rc);
  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 mds_client_data *mcd = med->med_mcd;
 102         struct obd_device *obd = req->rq_export->exp_obd;
 103         int err;
 104         __u64 transno;
 105         loff_t off;
 106         int log_pri = D_HA;
 107         ENTRY;
 108
 109         /* if the export has already been failed, we have no last_rcvd slot */
 110         if (req->rq_export->exp_failed) {
 111                 CERROR("committing transaction for disconnected client\n");
 112                 if (handle)
 113                         GOTO(commit, rc);
 114                 RETURN(rc);
 115         }
 116
 117         if (IS_ERR(handle))
 118                 RETURN(rc);
 119
 120         if (handle == NULL) {
 121                 /* if we're starting our own xaction, use our own inode */
 122                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 123                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 124                 if (IS_ERR(handle)) {
 125                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 126                         RETURN(PTR_ERR(handle));
 127                 }
 128         }
 129
 130         off = med->med_off;
 131
 132         transno = req->rq_reqmsg->transno;
 133         if (rc != 0) {
 134                 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
 135         } else if (transno == 0) {
 136                 spin_lock(&mds->mds_transno_lock);
 137                 transno = ++mds->mds_last_transno;
 138                 spin_unlock(&mds->mds_transno_lock);
 139         } else {
 140                 spin_lock(&mds->mds_transno_lock);
 141                 if (transno > mds->mds_last_transno)
 142                         mds->mds_last_transno = transno;
 143                 spin_unlock(&mds->mds_transno_lock);
 144         }
 145         req->rq_repmsg->transno = req->rq_transno = transno;
 146         mcd->mcd_last_transno = cpu_to_le64(transno);
 147         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 148         mcd->mcd_last_result = cpu_to_le32(rc);
 149         mcd->mcd_last_data = cpu_to_le32(op_data);
 150
 151
 152         fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb,
 153                               transno, handle, mds_commit_cb, NULL);
 154
 155         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 156                                   &off, 0);
 157
 158         if (err) {
 159                 log_pri = D_ERROR;
 160                 if (rc == 0)
 161                         rc = err;
 162         }
 163
 164         DEBUG_REQ(log_pri, req,
 165                   "wrote trans #"LPU64" client %s at idx %u: err = %d",
 166                   transno, mcd->mcd_uuid, med->med_idx, err);
 167
 168         err = mds_lov_write_objids(obd);
 169         if (err) {
 170                 log_pri = D_ERROR;
 171                 if (rc == 0)
 172                         rc = err;
 173         }
 174         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 175
 176 commit:
 177         err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 178         if (err) {
 179                 CERROR("error committing transaction: %d\n", err);
 180                 if (!rc)
 181                         rc = err;
 182         }
 183
 184         RETURN(rc);
 185 }
 186
 187 /* this gives the same functionality as the code between
 188  * sys_chmod and inode_setattr
 189  * chown_common and inode_setattr
 190  * utimes and inode_setattr
 191  */
 192 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 193 {
 194         time_t now = LTIME_S(CURRENT_TIME);
 195         struct iattr *attr = &rec->ur_iattr;
 196         unsigned int ia_valid = attr->ia_valid;
 197         int error;
 198         ENTRY;
 199
 200         /* only fix up attrs if the client VFS didn't already */
 201         if (!(ia_valid & ATTR_RAW))
 202                 RETURN(0);
 203
 204         if (!(ia_valid & ATTR_CTIME_SET))
 205                 LTIME_S(attr->ia_ctime) = now;
 206         if (!(ia_valid & ATTR_ATIME_SET))
 207                 LTIME_S(attr->ia_atime) = now;
 208         if (!(ia_valid & ATTR_MTIME_SET))
 209                 LTIME_S(attr->ia_mtime) = now;
 210
 211         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 212                 RETURN(-EPERM);
 213
 214         /* times */
 215         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 216                 if (rec->_ur_fsuid != inode->i_uid &&
 217                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 218                         RETURN(error);
 219         }
 220
 221         if (ia_valid & ATTR_SIZE) {
 222                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 223                         RETURN(error);
 224         }
 225
 226         if (ia_valid & ATTR_UID) {
 227                 /* chown */
 228                 error = -EPERM;
 229                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 230                         RETURN(-EPERM);
 231                 if (attr->ia_uid == (uid_t) -1)
 232                         attr->ia_uid = inode->i_uid;
 233                 if (attr->ia_gid == (gid_t) -1)
 234                         attr->ia_gid = inode->i_gid;
 235                 attr->ia_mode = inode->i_mode;
 236                 /*
 237                  * If the user or group of a non-directory has been
 238                  * changed by a non-root user, remove the setuid bit.
 239                  * 19981026 David C Niemi <niemi@tux.org>
 240                  *
 241                  * Changed this to apply to all users, including root,
 242                  * to avoid some races. This is the behavior we had in
 243                  * 2.0. The check for non-root was definitely wrong
 244                  * for 2.2 anyway, as it should have been using
 245                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 246                  */
 247                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 248                     !S_ISDIR(inode->i_mode)) {
 249                         attr->ia_mode &= ~S_ISUID;
 250                         attr->ia_valid |= ATTR_MODE;
 251                 }
 252                 /*
 253                  * Likewise, if the user or group of a non-directory
 254                  * has been changed by a non-root user, remove the
 255                  * setgid bit UNLESS there is no group execute bit
 256                  * (this would be a file marked for mandatory
 257                  * locking).  19981026 David C Niemi <niemi@tux.org>
 258                  *
 259                  * Removed the fsuid check (see the comment above) --
 260                  * 19990830 SD.
 261                  */
 262                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 263                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 264                         attr->ia_mode &= ~S_ISGID;
 265                         attr->ia_valid |= ATTR_MODE;
 266                 }
 267         } else if (ia_valid & ATTR_MODE) {
 268                 int mode = attr->ia_mode;
 269                 /* chmod */
 270                 if (attr->ia_mode == (mode_t) -1)
 271                         attr->ia_mode = inode->i_mode;
 272                 attr->ia_mode =
 273                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 274         }
 275         RETURN(0);
 276 }
 277
 278 void mds_steal_ack_locks(struct ptlrpc_request *req)
 279 {
 280         struct obd_export         *exp = req->rq_export;
 281         struct list_head          *tmp;
 282         struct ptlrpc_reply_state *oldrep;
 283         struct ptlrpc_service     *svc;
 284         struct llog_create_locks  *lcl;
 285         unsigned long              flags;
 286         char                       str[PTL_NALFMT_SIZE];
 287         int                        i;
 288
 289         /* CAVEAT EMPTOR: spinlock order */
 290         spin_lock_irqsave (&exp->exp_lock, flags);
 291         list_for_each (tmp, &exp->exp_outstanding_replies) {
 292                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 293
 294                 if (oldrep->rs_xid != req->rq_xid)
 295                         continue;
 296
 297                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 298                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 299                                 "new %d old %d\n", req->rq_xid,
 300                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 301
 302                 svc = oldrep->rs_srv_ni->sni_service;
 303                 spin_lock (&svc->srv_lock);
 304
 305                 list_del_init (&oldrep->rs_exp_list);
 306
 307                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 308                       " o%d NID %s\n",
 309                       oldrep->rs_nlocks, oldrep,
 310                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 311                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 312
 313                 for (i = 0; i < oldrep->rs_nlocks; i++)
 314                         ptlrpc_save_lock(req,
 315                                          &oldrep->rs_locks[i],
 316                                          oldrep->rs_modes[i]);
 317                 oldrep->rs_nlocks = 0;
 318
 319                 lcl = oldrep->rs_llog_locks;
 320                 oldrep->rs_llog_locks = NULL;
 321                 if (lcl != NULL)
 322                         ptlrpc_save_llog_lock(req, lcl);
 323
 324                 DEBUG_REQ(D_HA, req, "stole locks for");
 325                 ptlrpc_schedule_difficult_reply (oldrep);
 326
 327                 spin_unlock (&svc->srv_lock);
 328                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 329                 return;
 330         }
 331         spin_unlock_irqrestore (&exp->exp_lock, flags);
 332 }
 333
 334 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 335 {
 336         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 337                   mcd->mcd_last_transno, mcd->mcd_last_result);
 338         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 339         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 340
 341         mds_steal_ack_locks(req);
 342 }
 343
 344 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 345                                       int offset, struct ptlrpc_request *req)
 346 {
 347         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 348         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 349         struct dentry *de;
 350         struct mds_body *body;
 351
 352         mds_req_from_mcd(req, med->med_mcd);
 353
 354         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 355         if (IS_ERR(de)) {
 356                 LASSERT(PTR_ERR(de) == req->rq_status);
 357                 return;
 358         }
 359
 360         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 361         mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
 362         mds_pack_inode2body(req2obd(req), body, de->d_inode);
 363
 364         /* Don't return OST-specific attributes if we didn't just set them */
 365         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 366                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 367         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 368                 body->valid |= OBD_MD_FLMTIME;
 369         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 370                 body->valid |= OBD_MD_FLATIME;
 371
 372         l_dput(de);
 373 }
 374
 375 /* In the raw-setattr case, we lock the child inode.
 376  * In the write-back case or if being called from open, the client holds a lock
 377  * already.
 378  *
 379  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 380 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 381                              struct ptlrpc_request *req,
 382                              struct lustre_handle *lh)
 383 {
 384         struct mds_obd *mds = mds_req2mds(req);
 385         struct obd_device *obd = req->rq_export->exp_obd;
 386         struct mds_body *body;
 387         struct dentry *de;
 388         struct inode *inode = NULL;
 389         struct lustre_handle lockh[2] = {{0}, {0}};
 390         int parent_mode;
 391         void *handle = NULL;
 392         struct mds_logcancel_data *mlcd = NULL;
 393         int rc = 0, cleanup_phase = 0, err, locked = 0;
 394         ENTRY;
 395
 396         LASSERT(offset == 0);
 397
 398         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 399                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 400
 401         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 402
 403         MDS_UPDATE_COUNTER(mds, MDS_SETATTR_COUNT);
 404
 405         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 406                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 407                 if (IS_ERR(de))
 408                         GOTO(cleanup, rc = PTR_ERR(de));
 409         } else {
 410                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 411                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
 412                         lockpart |= MDS_INODELOCK_LOOKUP;
 413                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 414                                            lockh, &parent_mode, NULL, 0, lockpart);
 415                 if (IS_ERR(de))
 416                         GOTO(cleanup, rc = PTR_ERR(de));
 417                 locked = 1;
 418         }
 419
 420         cleanup_phase = 1;
 421
 422         inode = de->d_inode;
 423         LASSERT(inode);
 424         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 425             rec->ur_eadata != NULL)
 426                 down(&inode->i_sem);
 427
 428         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 429
 430         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 431         if (IS_ERR(handle))
 432                 GOTO(cleanup, rc = PTR_ERR(handle));
 433
 434         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 435                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 436                        LTIME_S(rec->ur_iattr.ia_mtime),
 437                        LTIME_S(rec->ur_iattr.ia_ctime));
 438         rc = mds_fix_attr(inode, rec);
 439         if (rc)
 440                 GOTO(cleanup, rc);
 441
 442         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 443                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 444                                       (long)&rec->ur_iattr.ia_attr_flags);
 445         else                                            /* setattr */
 446                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 447
 448         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 449             rec->ur_eadata != NULL) {
 450                 struct lov_stripe_md *lsm = NULL;
 451
 452                 rc = ll_permission(inode, MAY_WRITE, NULL);
 453                 if (rc < 0)
 454                         GOTO(cleanup, rc);
 455
 456                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 457                                    mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
 458                 if (rc)
 459                         GOTO(cleanup, rc);
 460
 461                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 462
 463                 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 464                                    rec->ur_eadatalen);
 465                 if (rc)
 466                         GOTO(cleanup, rc);
 467         }
 468
 469         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 470         mds_pack_inode2fid(obd, &body->fid1, inode);
 471         mds_pack_inode2body(obd, body, inode);
 472
 473         /* Don't return OST-specific attributes if we didn't just set them */
 474         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 475                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 476         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 477                 body->valid |= OBD_MD_FLMTIME;
 478         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 479                 body->valid |= OBD_MD_FLATIME;
 480
 481         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 482                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 483                           rec->ur_eadatalen);
 484                 if (mlcd) {
 485                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 486                                 rec->ur_eadatalen;
 487                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 488                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 489                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 490                                 mlcd->mlcd_cookielen;
 491                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 492                                mlcd->mlcd_cookielen);
 493                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 494                                mlcd->mlcd_eadatalen);
 495                 } else {
 496                         CERROR("unable to allocate log cancel data\n");
 497                 }
 498         }
 499         EXIT;
 500  cleanup:
 501         if (mlcd != NULL)
 502                 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
 503                                       handle, mds_cancel_cookies_cb, mlcd);
 504         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 505         switch (cleanup_phase) {
 506         case 1:
 507                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 508                     rec->ur_eadata != NULL)
 509                         up(&inode->i_sem);
 510                 l_dput(de);
 511                 if (locked) {
 512 #ifdef S_PDIROPS
 513                         if (lockh[1].cookie != 0)
 514                                 ldlm_lock_decref(lockh + 1, parent_mode);
 515 #endif
 516                         if (rc) {
 517                                 ldlm_lock_decref(lockh, LCK_PW);
 518                         } else {
 519                                 ptlrpc_save_lock (req, lockh, LCK_PW);
 520                         }
 521                 }
 522         case 0:
 523                 break;
 524         default:
 525                 LBUG();
 526         }
 527         if (err && !rc)
 528                 rc = err;
 529
 530         req->rq_status = rc;
 531         return 0;
 532 }
 533
 534 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 535                                      struct ptlrpc_request *req)
 536 {
 537         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 538         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 539         struct dentry *parent, *child;
 540         struct mds_body *body;
 541         ENTRY;
 542
 543         mds_req_from_mcd(req, med->med_mcd);
 544
 545         if (req->rq_status) {
 546                 EXIT;
 547                 return;
 548         }
 549
 550         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 551         LASSERT(!IS_ERR(parent));
 552         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 553         LASSERT(!IS_ERR(child));
 554         if ((child->d_flags & DCACHE_CROSS_REF)) {
 555                 LASSERTF(child->d_inode == NULL, "BUG 3869");
 556                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 557                 mds_pack_dentry2fid(&body->fid1, child);
 558                 mds_pack_dentry2body(body, child);
 559                 body->valid |= OBD_MD_MDS;
 560         } else if (child->d_inode == NULL) {
 561                 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
 562                           rec->ur_fid1->id, rec->ur_fid1->generation,
 563                           rec->ur_name, rec->ur_mode);
 564                 LASSERTF(child->d_inode != NULL, "BUG 3869");
 565         } else {
 566                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 567                 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
 568                 mds_pack_inode2body(req2obd(req), body, child->d_inode);
 569         }
 570         l_dput(parent);
 571         l_dput(child);
 572         EXIT;
 573 }
 574
 575 static int mds_reint_create(struct mds_update_record *rec, int offset,
 576                             struct ptlrpc_request *req,
 577                             struct lustre_handle *lh)
 578 {
 579         struct dentry *dparent = NULL;
 580         struct mds_obd *mds = mds_req2mds(req);
 581         struct obd_device *obd = req->rq_export->exp_obd;
 582         struct dentry *dchild = NULL;
 583         struct inode *dir = NULL;
 584         void *handle = NULL;
 585         struct lustre_handle lockh[2] = {{0}, {0}};
 586         int parent_mode;
 587         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 588         int created = 0;
 589         struct dentry_params dp;
 590         struct mea *mea = NULL;
 591         int mea_size;
 592         ENTRY;
 593
 594         LASSERT(offset == 0);
 595         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 596
 597         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 598                   rec->ur_fid1->id, rec->ur_fid1->generation,
 599                   rec->ur_name, rec->ur_mode);
 600
 601         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 602
 603
 604         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 605                 GOTO(cleanup, rc = -ESTALE);
 606
 607         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
 608                                         &parent_mode, rec->ur_name,
 609                                         rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 610         if (IS_ERR(dparent)) {
 611                 rc = PTR_ERR(dparent);
 612                 CERROR("parent lookup error %d\n", rc);
 613                 GOTO(cleanup, rc);
 614         }
 615         cleanup_phase = 1; /* locked parent dentry */
 616         dir = dparent->d_inode;
 617         LASSERT(dir);
 618
 619         ldlm_lock_dump_handle(D_OTHER, lockh);
 620
 621         /* try to retrieve MEA data for this dir */
 622         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 623         if (rc)
 624                 GOTO(cleanup, rc);
 625
 626         if (mea != NULL) {
 627                 /* dir is already splitted, check is requested filename
 628                  * should live at this MDS or at another one */
 629                 int i;
 630                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 631                 if (mea->mea_master != mea->mea_fids[i].mds) {
 632                         CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
 633                                " should be %d(%d)\n",
 634                                mea->mea_master, dparent->d_inode->i_ino,
 635                                dparent->d_inode->i_generation, rec->ur_name,
 636                                mea->mea_fids[i].mds, i);
 637                         GOTO(cleanup, rc = -ERESTART);
 638                 }
 639         }
 640
 641         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 642         if (IS_ERR(dchild)) {
 643                 rc = PTR_ERR(dchild);
 644                 CERROR("child lookup error %d\n", rc);
 645                 GOTO(cleanup, rc);
 646         }
 647
 648         cleanup_phase = 2; /* child dentry */
 649
 650         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 651
 652         if (type == S_IFREG || type == S_IFDIR) {
 653                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
 654                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
 655                        obd->obd_name, dparent->d_inode->i_ino,
 656                        dparent->d_inode->i_generation, rc, parent_mode);
 657                 if (rc > 0) {
 658                         /* dir got splitted */
 659                         GOTO(cleanup, rc = -ERESTART);
 660                 } else if (rc < 0) {
 661                         /* error happened during spitting. */
 662                         GOTO(cleanup, rc);
 663                 }
 664         }
 665
 666         if (dir->i_mode & S_ISGID) {
 667                 if (S_ISDIR(rec->ur_mode))
 668                         rec->ur_mode |= S_ISGID;
 669         }
 670
 671         dchild->d_fsdata = (void *)&dp;
 672         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 673         dp.p_ptr = req;
 674
 675         switch (type) {
 676         case S_IFREG:{
 677                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 678                 if (IS_ERR(handle))
 679                         GOTO(cleanup, rc = PTR_ERR(handle));
 680                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 681                 EXIT;
 682                 break;
 683         }
 684         case S_IFDIR:{
 685                 int nstripes = 0;
 686                 int i;
 687
 688                 /* as Peter asked, mkdir() should distribute new directories
 689                  * over the whole cluster in order to distribute namespace
 690                  * processing load. first, we calculate which MDS to use to put
 691                  * new directory's inode in. */
 692                 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
 693                                       rec->ur_flags);
 694                 if (i == mds->mds_num) {
 695                         /* inode will be created locally */
 696                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 697                         if (IS_ERR(handle))
 698                                 GOTO(cleanup, rc = PTR_ERR(handle));
 699
 700                         rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 701
 702                         if (rec->ur_eadata)
 703                                 nstripes = *(u16 *)rec->ur_eadata;
 704
 705                         if (rc == 0 && nstripes) {
 706                                 /* we pass LCK_EX to split routine to
 707                                  * signalthat we have exclusive access
 708                                  * to the directory. simple because
 709                                  * nobody knows it already exists -bzzz */
 710                                 rc = mds_try_to_split_dir(obd, dchild,
 711                                                           NULL, nstripes,
 712                                                           LCK_EX);
 713                                 if (rc > 0) {
 714                                         /* dir got splitted */
 715                                         rc = 0;
 716                                 } else if (rc < 0) {
 717                                         /* an error occured during
 718                                          * splitting. */
 719                                         GOTO(cleanup, rc);
 720                                 }
 721                         }
 722                 } else if (!DENTRY_VALID(dchild)) {
 723                         /* inode will be created on another MDS */
 724                         struct obdo *oa = NULL;
 725                         struct mds_body *body;
 726
 727                         /* first, create that inode */
 728                         oa = obdo_alloc();
 729                         LASSERT(oa != NULL);
 730                         oa->o_mds = i;
 731                         oa->o_easize = 0;
 732                         if (rec->ur_eadata) {
 733                                 /* user asks for creating splitted dir */
 734                                 oa->o_easize = *((u16 *) rec->ur_eadata);
 735                         }
 736
 737                         obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 738                                         OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 739                                         OBD_MD_FLUID | OBD_MD_FLGID);
 740                         oa->o_mode = dir->i_mode;
 741                         CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
 742                                         obd->obd_name, i);
 743                         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 744                                 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
 745                                        obd->obd_name, rec->ur_namelen - 1,
 746                                        rec->ur_name, (unsigned) rec->ur_fid2->id,
 747                                        (unsigned) rec->ur_fid2->generation);
 748                                 oa->o_id = rec->ur_fid2->id;
 749                                 oa->o_generation = rec->ur_fid2->generation;
 750                                 oa->o_flags |= OBD_FL_RECREATE_OBJS;
 751                         }
 752
 753                         rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
 754                         if (rc) {
 755                                 obdo_free(oa);
 756                                 GOTO(cleanup, rc);
 757                         }
 758
 759                         /* now, add new dir entry for it */
 760                         handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 761                         if (IS_ERR(handle)) {
 762                                 obdo_free(oa);
 763                                 GOTO(cleanup, rc = PTR_ERR(handle));
 764                         }
 765                         rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
 766                                                   rec->ur_namelen - 1,
 767                                                   oa->o_id, oa->o_generation,
 768                                                   i);
 769                         LASSERT(rc == 0);
 770
 771                         /* fill reply */
 772                         body = lustre_msg_buf(req->rq_repmsg,
 773                                               offset, sizeof (*body));
 774                         body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 775                         body->fid1.id = oa->o_id;
 776                         body->fid1.mds = i;
 777                         body->fid1.generation = oa->o_generation;
 778                         obdo_free(oa);
 779                 } else {
 780                         /* requested name exists in the directory */
 781                         rc = -EEXIST;
 782                 }
 783                 EXIT;
 784                 break;
 785         }
 786         case S_IFLNK:{
 787                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 788                 if (IS_ERR(handle))
 789                         GOTO(cleanup, rc = PTR_ERR(handle));
 790                 if (rec->ur_tgt == NULL)        /* no target supplied */
 791                         rc = -EINVAL;           /* -EPROTO? */
 792                 else
 793                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 794                 EXIT;
 795                 break;
 796         }
 797         case S_IFCHR:
 798         case S_IFBLK:
 799         case S_IFIFO:
 800         case S_IFSOCK:{
 801                 int rdev = rec->ur_rdev;
 802                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 803                 if (IS_ERR(handle))
 804                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 805                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 806                 EXIT;
 807                 break;
 808         }
 809         default:
 810                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 811                 dchild->d_fsdata = NULL;
 812                 GOTO(cleanup, rc = -EINVAL);
 813         }
 814
 815         /* In case we stored the desired inum in here, we want to clean up. */
 816         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 817                 dchild->d_fsdata = NULL;
 818
 819         if (rc) {
 820                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 821                 GOTO(cleanup, rc);
 822         } else if (dchild->d_inode) {
 823                 struct iattr iattr;
 824                 struct inode *inode = dchild->d_inode;
 825                 struct mds_body *body;
 826
 827                 created = 1;
 828                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 829                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 830                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 831                 iattr.ia_uid = rec->_ur_fsuid;
 832                 if (dir->i_mode & S_ISGID)
 833                         iattr.ia_gid = dir->i_gid;
 834                 else
 835                         iattr.ia_gid = rec->_ur_fsgid;
 836                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 837                         ATTR_MTIME | ATTR_CTIME;
 838
 839                 if (rec->ur_fid2->id) {
 840                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 841                         inode->i_generation = rec->ur_fid2->generation;
 842                         /* Dirtied and committed by the upcoming setattr. */
 843                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 844                                inode->i_ino, inode->i_generation);
 845                 } else {
 846 #if 0
 847                         struct lustre_handle child_ino_lockh;
 848 #endif
 849
 850                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 851                                inode->i_ino, inode->i_generation);
 852
 853 #if 0
 854                         /* The inode we were allocated may have just been freed
 855                          * by an unlink operation.  We take this lock to
 856                          * synchronize against the matching reply-ack-lock taken
 857                          * in unlink, to avoid replay problems if this reply
 858                          * makes it out to the client but the unlink's does not.
 859                          * See bug 2029 for more detail.*/
 860                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 861                         if (rc != ELDLM_OK) {
 862                                 CERROR("error locking for unlink/create sync: "
 863                                        "%d\n", rc);
 864                         } else {
 865                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 866                         }
 867 #endif
 868                 }
 869
 870                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 871                 if (rc)
 872                         CERROR("error on child setattr: rc = %d\n", rc);
 873
 874                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 875                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 876                 if (rc)
 877                         CERROR("error on parent setattr: rc = %d\n", rc);
 878                 else
 879                         MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
 880
 881                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 882                 mds_pack_inode2fid(obd, &body->fid1, inode);
 883                 mds_pack_inode2body(obd, body, inode);
 884         }
 885         EXIT;
 886
 887 cleanup:
 888         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 889
 890         if (rc && created) {
 891                 /* Destroy the file we just created. This should not need extra
 892                  * journal credits, as we have already modified all of the
 893                  * blocks needed in order to create the file in the first
 894                  * place. */
 895                 switch (type) {
 896                 case S_IFDIR:
 897                         err = vfs_rmdir(dir, dchild);
 898                         if (err)
 899                                 CERROR("rmdir in error path: %d\n", err);
 900                         break;
 901                 default:
 902                         err = vfs_unlink(dir, dchild);
 903                         if (err)
 904                                 CERROR("unlink in error path: %d\n", err);
 905                         break;
 906                 }
 907         } else {
 908                 rc = err;
 909         }
 910         switch (cleanup_phase) {
 911         case 2: /* child dentry */
 912                 l_dput(dchild);
 913         case 1: /* locked parent dentry */
 914 #ifdef S_PDIROPS
 915                 if (lockh[1].cookie != 0)
 916                         ldlm_lock_decref(lockh + 1, parent_mode);
 917 #endif
 918                 if (rc) {
 919                         ldlm_lock_decref(lockh, LCK_PW);
 920                 } else {
 921                         ptlrpc_save_lock(req, lockh, LCK_PW);
 922                 }
 923                 l_dput(dparent);
 924         case 0:
 925                 break;
 926         default:
 927                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 928                 LBUG();
 929         }
 930         if (mea)
 931                 OBD_FREE(mea, mea_size);
 932         req->rq_status = rc;
 933         return 0;
 934 }
 935
 936 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
 937            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
 938 {
 939         int i;
 940
 941         for (i = 0; i < RES_NAME_SIZE; i++) {
 942                 /* return 1 here, because enqueue_ordered will skip resources
 943                  * of all zeroes if they're sorted to the end of the list. */
 944                 if (res1->name[i] == 0 && res2->name[i] != 0)
 945                         return 1;
 946                 if (res2->name[i] == 0 && res1->name[i] != 0)
 947                         return 0;
 948
 949                 if (res1->name[i] > res2->name[i])
 950                         return 1;
 951                 if (res1->name[i] < res2->name[i])
 952                         return 0;
 953         }
 954
 955         if (!p1 || !p2)
 956                 return 0;
 957
 958         if (memcmp(p1, p2, sizeof(*p1)) < 0)
 959                 return 1;
 960
 961         return 0;
 962 }
 963
 964 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 965  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 966  * because they take the place of local semaphores.
 967  *
 968  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 969  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 970 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 971                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 972                           ldlm_policy_data_t *p1_policy,
 973                           struct ldlm_res_id *p2_res_id,
 974                           struct lustre_handle *p2_lockh, int p2_lock_mode,
 975                           ldlm_policy_data_t *p2_policy)
 976 {
 977         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 978         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 979         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 980         ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
 981         int rc, flags;
 982         ENTRY;
 983
 984         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 985
 986         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
 987                res_id[1]->name[0]);
 988
 989         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
 990                 handles[1] = p1_lockh;
 991                 handles[0] = p2_lockh;
 992                 res_id[1] = p1_res_id;
 993                 res_id[0] = p2_res_id;
 994                 lock_modes[1] = p1_lock_mode;
 995                 lock_modes[0] = p2_lock_mode;
 996                 policies[1] = p1_policy;
 997                 policies[0] = p2_policy;
 998         }
 999
1000         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1001                res_id[0]->name[0], res_id[1]->name[0]);
1002
1003         flags = LDLM_FL_LOCAL_ONLY;
1004         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1005                               LDLM_IBITS, policies[0], lock_modes[0], &flags,
1006                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1007                               NULL, 0, NULL, handles[0]);
1008         if (rc != ELDLM_OK)
1009                 RETURN(-EIO);
1010         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1011
1012         if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1013             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1014                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1015                 ldlm_lock_addref(handles[1], lock_modes[1]);
1016         } else if (res_id[1]->name[0] != 0) {
1017                 flags = LDLM_FL_LOCAL_ONLY;
1018                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1019                                       *res_id[1], LDLM_IBITS, policies[1],
1020                                       lock_modes[1], &flags, mds_blocking_ast,
1021                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1022                                       NULL, handles[1]);
1023                 if (rc != ELDLM_OK) {
1024                         ldlm_lock_decref(handles[0], lock_modes[0]);
1025                         RETURN(-EIO);
1026                 }
1027                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1028         }
1029
1030         RETURN(0);
1031 }
1032
1033 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1034                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1035                            ldlm_policy_data_t *p1_policy,
1036                            struct ldlm_res_id *p2_res_id,
1037                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1038                            ldlm_policy_data_t *p2_policy,
1039                            struct ldlm_res_id *c1_res_id,
1040                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1041                            ldlm_policy_data_t *c1_policy,
1042                            struct ldlm_res_id *c2_res_id,
1043                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1044                            ldlm_policy_data_t *c2_policy)
1045 {
1046         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1047                                           c1_res_id, c2_res_id };
1048         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1049                                                  c1_lockh, c2_lockh };
1050         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1051                               c1_lock_mode, c2_lock_mode };
1052         ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1053                                             c1_policy, c2_policy};
1054         int rc, i, j, sorted, flags;
1055         ENTRY;
1056
1057         CDEBUG(D_DLMTRACE,
1058                "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1059                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1060                res_id[3]->name[0]);
1061
1062         /* simple insertion sort - we have at most 4 elements */
1063         for (i = 1; i < 4; i++) {
1064                 j = i - 1;
1065                 dlm_handles[4] = dlm_handles[i];
1066                 res_id[4] = res_id[i];
1067                 lock_modes[4] = lock_modes[i];
1068                 policies[4] = policies[i];
1069
1070                 sorted = 0;
1071                 do {
1072                         if (res_gt(res_id[j], res_id[4], policies[j],
1073                                    policies[4])) {
1074                                 dlm_handles[j + 1] = dlm_handles[j];
1075                                 res_id[j + 1] = res_id[j];
1076                                 lock_modes[j + 1] = lock_modes[j];
1077                                 policies[j + 1] = policies[j];
1078                                 j--;
1079                         } else {
1080                                 sorted = 1;
1081                         }
1082                 } while (j >= 0 && !sorted);
1083
1084                 dlm_handles[j + 1] = dlm_handles[4];
1085                 res_id[j + 1] = res_id[4];
1086                 lock_modes[j + 1] = lock_modes[4];
1087                 policies[j + 1] = policies[4];
1088         }
1089
1090         CDEBUG(D_DLMTRACE,
1091                "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1092                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1093                res_id[3]->name[0]);
1094
1095         /* XXX we could send ASTs on all these locks first before blocking? */
1096         for (i = 0; i < 4; i++) {
1097                 flags = 0;
1098                 if (res_id[i]->name[0] == 0)
1099                         break;
1100                 if (i != 0 &&
1101                     !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1102                     (policies[i]->l_inodebits.bits &
1103                      policies[i-1]->l_inodebits.bits) ) {
1104                         memcpy(dlm_handles[i], dlm_handles[i-1],
1105                                sizeof(*(dlm_handles[i])));
1106                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1107                 } else {
1108                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1109                                               *res_id[i], LDLM_IBITS,
1110                                               policies[i],
1111                                               lock_modes[i], &flags,
1112                                               mds_blocking_ast,
1113                                               ldlm_completion_ast, NULL, NULL,
1114                                               NULL, 0, NULL, dlm_handles[i]);
1115                         if (rc != ELDLM_OK)
1116                                 GOTO(out_err, rc = -EIO);
1117                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1118                 }
1119         }
1120
1121         RETURN(0);
1122 out_err:
1123         while (i-- > 0)
1124                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1125
1126         return rc;
1127 }
1128
1129 /* In the unlikely case that the child changed while we were waiting
1130  * on the lock, we need to drop the lock on the old child and either:
1131  * - if the child has a lower resource name, then we have to also
1132  *   drop the parent lock and regain the locks in the right order
1133  * - in the rename case, if the child has a lower resource name than one of
1134  *   the other parent/child resources (maxres) we also need to reget the locks
1135  * - if the child has a higher resource name (this is the common case)
1136  *   we can just get the lock on the new child (still in lock order)
1137  *
1138  * Returns 0 if the child did not change or if it changed but could be locked.
1139  * Returns 1 if the child changed and we need to re-lock (no locks held).
1140  * Returns -ve error with a valid dchild (no locks held). */
1141 static int mds_verify_child(struct obd_device *obd,
1142                             struct ldlm_res_id *parent_res_id,
1143                             struct lustre_handle *parent_lockh,
1144                             struct dentry *dparent, int parent_mode,
1145                             struct ldlm_res_id *child_res_id,
1146                             struct lustre_handle *child_lockh,
1147                             struct dentry **dchildp, int child_mode,
1148                             ldlm_policy_data_t *child_policy,
1149                             const char *name, int namelen,
1150                             struct ldlm_res_id *maxres)
1151 {
1152         struct dentry *vchild, *dchild = *dchildp;
1153         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1154         ENTRY;
1155
1156         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1157         if (IS_ERR(vchild))
1158                 GOTO(cleanup, rc = PTR_ERR(vchild));
1159
1160         if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1161                 if  (child_res_id->name[0] == vchild->d_inum &&
1162                                 child_res_id->name[1] == vchild->d_generation) {
1163                         if (dchild != NULL)
1164                                 l_dput(dchild);
1165                         *dchildp = vchild;
1166                         RETURN(0);
1167                 }
1168                 goto changed;
1169         }
1170
1171         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1172                    (vchild->d_inode != NULL &&
1173                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1174                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1175                 if (dchild != NULL)
1176                         l_dput(dchild);
1177                 *dchildp = vchild;
1178
1179                 RETURN(0);
1180         }
1181
1182 changed:
1183         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1184                vchild->d_inode, dchild ? dchild->d_inode : 0,
1185                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1186                child_res_id->name[0]);
1187         if (child_res_id->name[0] != 0)
1188                 ldlm_lock_decref(child_lockh, child_mode);
1189         if (dchild)
1190                 l_dput(dchild);
1191
1192         cleanup_phase = 1; /* parent lock only */
1193         *dchildp = dchild = vchild;
1194
1195         if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1196                 int flags = 0;
1197                 if (dchild->d_inode) {
1198                         child_res_id->name[0] = dchild->d_inode->i_ino;
1199                         child_res_id->name[1] = dchild->d_inode->i_generation;
1200                 } else {
1201                         child_res_id->name[0] = dchild->d_inum;
1202                         child_res_id->name[1] = dchild->d_generation;
1203                 }
1204
1205                 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1206                     res_gt(maxres, child_res_id, NULL, NULL)) {
1207                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1208                                child_res_id->name[0], parent_res_id->name[0],
1209                                maxres->name[0]);
1210                         GOTO(cleanup, rc = 1);
1211                 }
1212
1213                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1214                                       *child_res_id, LDLM_IBITS, child_policy,
1215                                       child_mode, &flags, mds_blocking_ast,
1216                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1217                                       NULL, child_lockh);
1218                 if (rc != ELDLM_OK)
1219                         GOTO(cleanup, rc = -EIO);
1220
1221         } else {
1222                 memset(child_res_id, 0, sizeof(*child_res_id));
1223         }
1224
1225         EXIT;
1226 cleanup:
1227         if (rc) {
1228                 switch(cleanup_phase) {
1229                 case 2:
1230                         if (child_res_id->name[0] != 0)
1231                                 ldlm_lock_decref(child_lockh, child_mode);
1232                 case 1:
1233                         ldlm_lock_decref(parent_lockh, parent_mode);
1234                 }
1235         }
1236         return rc;
1237 }
1238
1239 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1240                                 struct ll_fid *fid,
1241                                 struct lustre_handle *parent_lockh,
1242                                 struct dentry **dparentp, int parent_mode,
1243                                 __u64 parent_lockpart, int *update_mode,
1244                                 char *name, int namelen,
1245                                 struct lustre_handle *child_lockh,
1246                                 struct dentry **dchildp, int child_mode,
1247                                 __u64 child_lockpart, void *clone_info)
1248 {
1249         struct ldlm_res_id child_res_id = { .name = {0} };
1250         struct ldlm_res_id parent_res_id = { .name = {0} };
1251         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1252         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1253         struct inode *inode;
1254         int rc = 0, cleanup_phase = 0;
1255         ENTRY;
1256
1257         /* Step 1: Lookup parent */
1258         *dparentp = mds_fid2dentry(mds, fid, NULL);
1259         if (IS_ERR(*dparentp)) {
1260                 rc = PTR_ERR(*dparentp);
1261                 *dparentp = NULL;
1262                 RETURN(rc);
1263         }
1264
1265         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1266                (*dparentp)->d_inode->i_ino, name);
1267
1268         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1269         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1270 #ifdef S_PDIROPS
1271         parent_lockh[1].cookie = 0;
1272         if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1273                 struct ldlm_res_id res_id = { .name = {0} };
1274                 ldlm_policy_data_t policy;
1275                 int flags = 0;
1276                 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1277                 if (*update_mode) {
1278                         res_id.name[0] = (*dparentp)->d_inode->i_ino;
1279                         res_id.name[1] = (*dparentp)->d_inode->i_generation;
1280                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1281                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1282                                               res_id, LDLM_IBITS, &policy,
1283                                               *update_mode, &flags,
1284                                               mds_blocking_ast,
1285                                               ldlm_completion_ast,
1286                                               NULL, NULL, NULL, 0, NULL,
1287                                               parent_lockh + 1);
1288                         if (rc != ELDLM_OK)
1289                                 RETURN(-ENOLCK);
1290                 }
1291
1292                 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1293                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1294                        (*dparentp)->d_inode->i_ino,
1295                        (*dparentp)->d_inode->i_generation,
1296                        parent_res_id.name[2]);
1297         }
1298 #endif
1299
1300         cleanup_phase = 1; /* parent dentry */
1301 #ifdef CONFIG_SNAPFS
1302         if (clone_info) {
1303                 /*FIXME is there any other FUNC will use d_fsdata,
1304                  *excepet creating inode according inum*/
1305                 (*dparentp)->d_fsdata = clone_info;
1306         }
1307 #endif
1308
1309         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1310         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1311         if (IS_ERR(*dchildp)) {
1312                 rc = PTR_ERR(*dchildp);
1313                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1314                 GOTO(cleanup, rc);
1315         }
1316
1317         if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1318                 /* inode lives on another MDS: return * mds/ino/gen
1319                  * and LOOKUP lock. drop possible UPDATE lock! */
1320                 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1321                 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1322                 child_res_id.name[0] = (*dchildp)->d_inum;
1323                 child_res_id.name[1] = (*dchildp)->d_generation;
1324                 goto retry_locks;
1325         }
1326
1327         inode = (*dchildp)->d_inode;
1328         if (inode != NULL)
1329                 inode = igrab(inode);
1330         if (inode == NULL)
1331                 goto retry_locks;
1332
1333         child_res_id.name[0] = inode->i_ino;
1334         child_res_id.name[1] = inode->i_generation;
1335
1336         iput(inode);
1337
1338 retry_locks:
1339         cleanup_phase = 2; /* child dentry */
1340
1341         /* Step 3: Lock parent and child in resource order.  If child doesn't
1342          *         exist, we still have to lock the parent and re-lookup. */
1343         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1344                                    &parent_policy, &child_res_id, child_lockh,
1345                                    child_mode, &child_policy);
1346         if (rc)
1347                 GOTO(cleanup, rc);
1348
1349         if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1350                 cleanup_phase = 4; /* child lock */
1351         else
1352                 cleanup_phase = 3; /* parent lock */
1353
1354         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1355         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1356                               parent_mode, &child_res_id, child_lockh,
1357                               dchildp, child_mode, &child_policy,
1358                               name, namelen, &parent_res_id);
1359         if (rc > 0)
1360                 goto retry_locks;
1361         if (rc < 0) {
1362                 cleanup_phase = 3;
1363                 GOTO(cleanup, rc);
1364         }
1365
1366 cleanup:
1367         if (rc) {
1368                 switch (cleanup_phase) {
1369                 case 4:
1370                         ldlm_lock_decref(child_lockh, child_mode);
1371                 case 3:
1372                         ldlm_lock_decref(parent_lockh, parent_mode);
1373                 case 2:
1374                         l_dput(*dchildp);
1375                 case 1:
1376 #ifdef S_PDIROPS
1377                         if (parent_lockh[1].cookie)
1378                                 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1379 #endif
1380                         l_dput(*dparentp);
1381                 default: ;
1382                 }
1383         }
1384         return rc;
1385 }
1386
1387 void mds_reconstruct_generic(struct ptlrpc_request *req)
1388 {
1389         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1390
1391         mds_req_from_mcd(req, med->med_mcd);
1392 }
1393
1394 int mds_create_local_dentry(struct mds_update_record *rec,
1395                            struct obd_device *obd)
1396 {
1397         struct mds_obd *mds = &obd->u.mds;
1398         struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1399         int fidlen = 0, rc, cleanup_phase = 0;
1400         struct dentry *new_child = NULL;
1401         char *fidname = rec->ur_name;
1402         struct dentry *child = NULL;
1403         struct lustre_handle lockh[2] = {{0}, {0}};
1404         void *handle;
1405         ENTRY;
1406
1407         down(&fids_dir->i_sem);
1408         fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1409         CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1410                         fidname, (unsigned) rec->ur_fid1->id,
1411                         (unsigned) rec->ur_fid1->generation);
1412
1413         new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1414         up(&fids_dir->i_sem);
1415         if (IS_ERR(new_child)) {
1416                 CERROR("can't lookup %s: %d\n", fidname,
1417                                 (int) PTR_ERR(new_child));
1418                 GOTO(cleanup, rc = PTR_ERR(new_child));
1419         }
1420         cleanup_phase = 1;
1421
1422         if (new_child->d_inode != NULL) {
1423                 /* nice. we've already have local dentry! */
1424                 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1425                        (unsigned) new_child->d_inode->i_ino,
1426                        (unsigned) new_child->d_inode->i_generation);
1427                 rec->ur_fid1->id = fids_dir->i_ino;
1428                 rec->ur_fid1->generation = fids_dir->i_generation;
1429                 rec->ur_namelen = fidlen + 1;
1430                 GOTO(cleanup, rc = 0);
1431         }
1432
1433         /* new, local dentry will be added soon. we need no aliases here */
1434         d_drop(new_child);
1435
1436         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1437                 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1438         } else {
1439                 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1440                                               LCK_EX, lockh, NULL, NULL, 0,
1441                                               MDS_INODELOCK_UPDATE);
1442         }
1443
1444         if (IS_ERR(child)) {
1445                 rc = PTR_ERR(child);
1446                 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1447                         CERROR("can't get victim: %d\n", rc);
1448                 GOTO(cleanup, rc);
1449         }
1450         cleanup_phase = 2;
1451
1452         handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1453         if (IS_ERR(handle))
1454                 GOTO(cleanup, rc = PTR_ERR(handle));
1455
1456         rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1457                                   rec->ur_fid1->id, rec->ur_fid1->generation,
1458                                   mds->mds_num);
1459         if (rc)
1460                 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1461                        (unsigned long) child->d_inode->i_ino,
1462                        (unsigned long) child->d_inode->i_generation, rc);
1463         else {
1464                 if (S_ISDIR(child->d_inode->i_mode)) {
1465                         fids_dir->i_nlink++;
1466                         mark_inode_dirty(fids_dir);
1467                 }
1468                 mark_inode_dirty(child->d_inode);
1469         }
1470         fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1471
1472         rec->ur_fid1->id = fids_dir->i_ino;
1473         rec->ur_fid1->generation = fids_dir->i_generation;
1474         rec->ur_namelen = fidlen + 1;
1475
1476 cleanup:
1477         switch(cleanup_phase) {
1478                 case 2:
1479                         if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1480                                 ldlm_lock_decref(lockh, LCK_EX);
1481                         dput(child);
1482                 case 1:
1483                         dput(new_child);
1484                 case 0:
1485                        break;
1486         }
1487         RETURN(rc);
1488 }
1489
1490 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1491                                         struct ptlrpc_request *slave)
1492 {
1493         void *cookie, *cookie2;
1494         struct mds_body *body2;
1495         struct mds_body *body;
1496         void *ea, *ea2;
1497         ENTRY;
1498
1499         body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1500         LASSERT(body != NULL);
1501
1502         body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1503         LASSERT(body2 != NULL);
1504
1505         if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1506                 RETURN(0);
1507         }
1508
1509         memcpy(body2, body, sizeof(*body));
1510         body2->valid &= ~OBD_MD_FLCOOKIE;
1511
1512         if (!(body->valid & OBD_MD_FLEASIZE) &&
1513             !(body->valid & OBD_MD_FLDIREA))
1514                 RETURN(0);
1515
1516         if (body->eadatasize == 0) {
1517                 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1518                 RETURN(0);
1519         }
1520
1521         LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1522
1523         ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1524         LASSERT(ea != NULL);
1525
1526         ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1527         LASSERT(ea2 != NULL);
1528
1529         memcpy(ea2, ea, body->eadatasize);
1530
1531         if (body->valid & OBD_MD_FLCOOKIE) {
1532                 LASSERT(master->rq_repmsg->buflens[2] >=
1533                                 slave->rq_repmsg->buflens[2]);
1534                 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1535                                 slave->rq_repmsg->buflens[2]);
1536                 LASSERT(cookie != NULL);
1537
1538                 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1539                                 master->rq_repmsg->buflens[2]);
1540                 LASSERT(cookie2 != NULL);
1541                 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1542                 body2->valid |= OBD_MD_FLCOOKIE;
1543         }
1544         RETURN(0);
1545 }
1546
1547 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1548                                    struct ptlrpc_request *req,
1549                                    struct lustre_handle *parent_lockh,
1550                                    int update_mode,
1551                                    struct dentry *dparent,
1552                                    struct lustre_handle *child_lockh,
1553                                    struct dentry *dchild)
1554 {
1555         struct obd_device *obd = req->rq_export->exp_obd;
1556         struct mds_obd *mds = mds_req2mds(req);
1557         struct mdc_op_data op_data;
1558         int rc = 0, cleanup_phase = 0;
1559         struct ptlrpc_request *request = NULL;
1560         void *handle;
1561         ENTRY;
1562
1563         LASSERT(offset == 0 || offset == 2);
1564
1565         DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1566                   rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1567                   (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1568         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1569                 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1570                           rec->ur_namelen - 1, rec->ur_name,
1571                           (unsigned)dchild->d_mdsnum,
1572                           (unsigned) dchild->d_inum,
1573                           (unsigned) dchild->d_generation);
1574
1575         /* time to drop i_nlink on remote MDS */
1576         memset(&op_data, 0, sizeof(op_data));
1577         op_data.fid1.mds = dchild->d_mdsnum;
1578         op_data.fid1.id = dchild->d_inum;
1579         op_data.fid1.generation = dchild->d_generation;
1580         op_data.create_mode = rec->ur_mode;
1581         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1582                 op_data.create_mode |= MDS_MODE_REPLAY;
1583         rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1584         cleanup_phase = 2;
1585         if (request) {
1586                 if (rc == 0)
1587                         mds_copy_unlink_reply(req, request);
1588                 ptlrpc_req_finished(request);
1589         }
1590         if (rc == 0) {
1591                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1592                                       NULL);
1593                 if (IS_ERR(handle))
1594                         GOTO(cleanup, rc = PTR_ERR(handle));
1595                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1596                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1597                                         rc, 0);
1598         }
1599 cleanup:
1600         req->rq_status = rc;
1601
1602 #ifdef S_PDIROPS
1603         if (parent_lockh[1].cookie != 0)
1604                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1605 #endif
1606         ldlm_lock_decref(child_lockh, LCK_EX);
1607         if (rc)
1608                 ldlm_lock_decref(parent_lockh, LCK_PW);
1609         else
1610                 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1611         l_dput(dchild);
1612         l_dput(dparent);
1613
1614         return 0;
1615 }
1616
1617 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1618                             struct ptlrpc_request *req,
1619                             struct lustre_handle *lh)
1620 {
1621         struct dentry *dparent, *dchild;
1622         struct mds_obd *mds = mds_req2mds(req);
1623         struct obd_device *obd = req->rq_export->exp_obd;
1624         struct mds_body *body = NULL;
1625         struct inode *child_inode;
1626         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1627         struct lustre_handle child_lockh = {0};
1628 #if 0
1629         struct lustre_handle child_reuse_lockh = {0};
1630 #endif
1631         struct lustre_handle * slave_lockh = NULL;
1632         struct llog_create_locks *lcl = NULL;
1633         char fidname[LL_FID_NAMELEN];
1634         void *handle = NULL;
1635         int rc = 0, log_unlink = 0, cleanup_phase = 0;
1636         int unlink_by_fid = 0;
1637         int update_mode;
1638         ENTRY;
1639
1640         LASSERT(offset == 0 || offset == 2);
1641
1642         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1643                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1644
1645         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1646
1647         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1648                 DEBUG_REQ(D_HA, req, "unlink replay\n");
1649                 memcpy(lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1650                        lustre_msg_buf(req->rq_reqmsg, 2, 0),
1651                        req->rq_repmsg->buflens[offset + 2]);
1652         }
1653
1654         MDS_UPDATE_COUNTER(mds, MDS_UNLINK_COUNT);
1655
1656         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1657                 GOTO(cleanup, rc = -ENOENT);
1658
1659         if (rec->ur_namelen == 1) {
1660                 /* this is request to drop i_nlink on local inode */
1661                 unlink_by_fid = 1;
1662                 rec->ur_name = fidname;
1663                 rc = mds_create_local_dentry(rec, obd);
1664                 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1665                         DEBUG_REQ(D_HA, req,
1666                                   "drop nlink on inode %u/%u/%u (replay)",
1667                                   (unsigned) rec->ur_fid1->mds,
1668                                   (unsigned) rec->ur_fid1->id,
1669                                   (unsigned) rec->ur_fid1->generation);
1670                         req->rq_status = 0;
1671                         RETURN(0);
1672                 }
1673         }
1674
1675         if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1676                 /* master mds for directory asks slave removing
1677                  * inode is already locked */
1678                 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1679                                                LCK_PW, parent_lockh,
1680                                                &update_mode, rec->ur_name,
1681                                                rec->ur_namelen,
1682                                                MDS_INODELOCK_UPDATE);
1683                 if (IS_ERR(dparent))
1684                         GOTO(cleanup, rc = PTR_ERR(dparent));
1685                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1686                                            rec->ur_namelen - 1);
1687                 if (IS_ERR(dchild))
1688                         GOTO(cleanup, rc = PTR_ERR(dchild));
1689                 child_lockh.cookie = 0;
1690                 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1691                 LASSERT(dchild->d_inode != NULL);
1692                 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1693         } else {
1694                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1695                                                  parent_lockh, &dparent,
1696                                                  LCK_PW, MDS_INODELOCK_UPDATE,
1697                                                  &update_mode, rec->ur_name,
1698                                                  rec->ur_namelen, &child_lockh,
1699                                                  &dchild, LCK_EX,
1700                                                  MDS_INODELOCK_LOOKUP |
1701                                                  MDS_INODELOCK_UPDATE, NULL);
1702         }
1703         if (rc)
1704                 GOTO(cleanup, rc);
1705
1706         if (dchild->d_flags & DCACHE_CROSS_REF) {
1707                 /* we should have parent lock only here */
1708                 LASSERT(unlink_by_fid == 0);
1709                 LASSERT(dchild->d_mdsnum != mds->mds_num);
1710                 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1711                                         update_mode, dparent, &child_lockh, dchild);
1712                 RETURN(0);
1713         }
1714
1715         cleanup_phase = 1; /* dchild, dparent, locks */
1716
1717         dget(dchild);
1718         child_inode = dchild->d_inode;
1719         if (child_inode == NULL) {
1720                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1721                        dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1722                 GOTO(cleanup, rc = -ENOENT);
1723         }
1724
1725         cleanup_phase = 2; /* dchild has a lock */
1726
1727         /* We have to do these checks ourselves, in case we are making an
1728          * orphan.  The client tells us whether rmdir() or unlink() was called,
1729          * so we need to return appropriate errors (bug 72).
1730          *
1731          * We don't have to check permissions, because vfs_rename (called from
1732          * mds_open_unlink_rename) also calls may_delete. */
1733         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1734                 if (!S_ISDIR(child_inode->i_mode))
1735                         GOTO(cleanup, rc = -ENOTDIR);
1736         } else {
1737                 if (S_ISDIR(child_inode->i_mode))
1738                         GOTO(cleanup, rc = -EISDIR);
1739         }
1740
1741         /* handle splitted dir */
1742         rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1743         if (rc)
1744                 GOTO(cleanup, rc);
1745
1746 #if 0
1747         /* Step 4: Get a lock on the ino to sync with creation WRT inode
1748          * reuse (see bug 2029). */
1749         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1750         if (rc != ELDLM_OK)
1751                 GOTO(cleanup, rc);
1752 #endif
1753         cleanup_phase = 3; /* child inum lock */
1754
1755         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1756
1757         /* ldlm_reply in buf[0] if called via intent */
1758         if (offset)
1759                 offset = 1;
1760
1761         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1762         LASSERT(body != NULL);
1763
1764         /* If this is the last reference to this inode, get the OBD EA
1765          * data first so the client can destroy OST objects.
1766          * we only do the object removal if no open files remain.
1767          * Nobody can get at this name anymore because of the locks so
1768          * we make decisions here as to whether to remove the inode */
1769         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1770             mds_open_orphan_count(child_inode) == 0) {
1771                 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1772                 mds_pack_inode2body(obd, body, child_inode);
1773                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1774                             child_inode, 1);
1775                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1776                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1777                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1778                 } else {
1779                         log_unlink = 1;
1780                 }
1781         }
1782
1783         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1784         switch (child_inode->i_mode & S_IFMT) {
1785         case S_IFDIR:
1786                 /* Drop any lingering child directories before we start our
1787                  * transaction, to avoid doing multiple inode dirty/delete
1788                  * in our compound transaction (bug 1321). */
1789                 shrink_dcache_parent(dchild);
1790                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1791                                       NULL);
1792                 if (IS_ERR(handle))
1793                         GOTO(cleanup, rc = PTR_ERR(handle));
1794                 cleanup_phase = 4; /* transaction */
1795                 rc = vfs_rmdir(dparent->d_inode, dchild);
1796                 break;
1797         case S_IFREG: {
1798 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1799                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1800                                                         offset + 1, 0);
1801                 handle = fsfilt_start_log(obd, dparent->d_inode,
1802                                           FSFILT_OP_UNLINK, NULL,
1803                                           le32_to_cpu(lmm->lmm_stripe_count));
1804                 if (IS_ERR(handle))
1805                         GOTO(cleanup, rc = PTR_ERR(handle));
1806
1807                 cleanup_phase = 4; /* transaction */
1808                 rc = vfs_unlink(dparent->d_inode, dchild);
1809
1810                 if (!rc && log_unlink)
1811                         if (mds_log_op_unlink(obd, child_inode,
1812                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1813                                 req->rq_repmsg->buflens[offset + 1],
1814                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1815                                 req->rq_repmsg->buflens[offset + 2], &lcl) > 0)
1816                                 body->valid |= OBD_MD_FLCOOKIE;
1817                 break;
1818         }
1819         case S_IFLNK:
1820         case S_IFCHR:
1821         case S_IFBLK:
1822         case S_IFIFO:
1823         case S_IFSOCK:
1824                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1825                                       NULL);
1826                 if (IS_ERR(handle))
1827                         GOTO(cleanup, rc = PTR_ERR(handle));
1828                 cleanup_phase = 4; /* transaction */
1829                 rc = vfs_unlink(dparent->d_inode, dchild);
1830                 break;
1831         default:
1832                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1833                        rec->ur_name);
1834                 LBUG();
1835                 GOTO(cleanup, rc = -EINVAL);
1836         }
1837
1838  cleanup:
1839         if (rc == 0) {
1840                 struct iattr iattr;
1841                 int err;
1842
1843                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1844                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1845                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1846
1847                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1848                 if (err)
1849                         CERROR("error on parent setattr: rc = %d\n", err);
1850         }
1851
1852         switch(cleanup_phase) {
1853         case 4:
1854                 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1855                 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1856                 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1857                                 mds_open_orphan_count(dchild->d_inode) > 0) {
1858
1859                         /* filesystem is really going to destroy an inode
1860                          * we have to delay this till inode is opened -bzzz */
1861                         mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1862                 }
1863                 /* handle splitted dir */
1864                 if (rc == 0) {
1865                         /* master directory can be non-empty or something else ... */
1866                         mds_unlink_slave_objs(obd, dchild);
1867                 }
1868                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1869                                         rc, 0);
1870                 if (!rc)
1871                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1872                                            "unlinked", 0, NULL);
1873                 if (lcl != NULL)
1874                         ptlrpc_save_llog_lock(req, lcl);
1875         case 3: /* child ino-reuse lock */
1876 #if 0
1877                 if (rc && body != NULL) {
1878                         // Don't unlink the OST objects if the MDS unlink failed
1879                         body->valid = 0;
1880                 }
1881                 if (rc)
1882                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1883                 else
1884                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1885 #endif
1886         case 2: /* child lock */
1887                 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1888                 if (child_lockh.cookie)
1889                         ldlm_lock_decref(&child_lockh, LCK_EX);
1890         case 1: /* child and parent dentry, parent lock */
1891 #ifdef S_PDIROPS
1892                 if (parent_lockh[1].cookie != 0)
1893                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1894 #endif
1895                 if (rc)
1896                         ldlm_lock_decref(parent_lockh, LCK_PW);
1897                 else
1898                         ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1899                 l_dput(dchild);
1900                 l_dput(dchild);
1901                 l_dput(dparent);
1902         case 0:
1903                 break;
1904         default:
1905                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1906                 LBUG();
1907         }
1908         req->rq_status = rc;
1909         return 0;
1910 }
1911
1912 /*
1913  * to service requests from remote MDS to increment i_nlink
1914  */
1915 static int mds_reint_link_acquire(struct mds_update_record *rec,
1916                                  int offset, struct ptlrpc_request *req,
1917                                  struct lustre_handle *lh)
1918 {
1919         struct obd_device *obd = req->rq_export->exp_obd;
1920         struct ldlm_res_id src_res_id = { .name = {0} };
1921         struct lustre_handle *handle = NULL, src_lockh = {0};
1922         struct mds_obd *mds = mds_req2mds(req);
1923         int rc = 0, cleanup_phase = 0;
1924         struct dentry *de_src = NULL;
1925         ldlm_policy_data_t policy;
1926         int flags = 0;
1927         ENTRY;
1928
1929         DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1930                   obd->obd_name, (unsigned) rec->ur_fid1->mds,
1931                   (unsigned) rec->ur_fid1->id,
1932                   (unsigned) rec->ur_fid1->generation);
1933
1934         /* Step 1: Lookup the source inode and target directory by FID */
1935         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1936         if (IS_ERR(de_src))
1937                 GOTO(cleanup, rc = PTR_ERR(de_src));
1938         cleanup_phase = 1; /* source dentry */
1939
1940         src_res_id.name[0] = de_src->d_inode->i_ino;
1941         src_res_id.name[1] = de_src->d_inode->i_generation;
1942         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1943
1944         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1945                         src_res_id, LDLM_IBITS, &policy,
1946                         LCK_EX, &flags, mds_blocking_ast,
1947                         ldlm_completion_ast, NULL, NULL,
1948                         NULL, 0, NULL, &src_lockh);
1949         if (rc != ELDLM_OK)
1950                 GOTO(cleanup, rc = -ENOLCK);
1951         cleanup_phase = 2; /* lock */
1952
1953         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1954
1955         handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1956         if (IS_ERR(handle)) {
1957                 rc = PTR_ERR(handle);
1958                 GOTO(cleanup, rc);
1959         }
1960         de_src->d_inode->i_nlink++;
1961         mark_inode_dirty(de_src->d_inode);
1962
1963 cleanup:
1964         rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1965                                         handle, req, rc, 0);
1966         EXIT;
1967         switch (cleanup_phase) {
1968                 case 2:
1969                         if (rc)
1970                                 ldlm_lock_decref(&src_lockh, LCK_EX);
1971                         else
1972                                 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1973                 case 1:
1974                         l_dput(de_src);
1975                 case 0:
1976                         break;
1977                 default:
1978                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1979                         LBUG();
1980         }
1981         req->rq_status = rc;
1982         return 0;
1983 }
1984
1985 /*
1986  * request to link to foreign inode:
1987  *  - acquire i_nlinks on this inode
1988  *  - add dentry
1989  */
1990 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1991                                     int offset, struct ptlrpc_request *req,
1992                                     struct lustre_handle *lh)
1993 {
1994         struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1995         struct obd_device *obd = req->rq_export->exp_obd;
1996         struct dentry *de_tgt_dir = NULL;
1997         struct mds_obd *mds = mds_req2mds(req);
1998         int rc = 0, cleanup_phase = 0;
1999         struct mdc_op_data op_data;
2000         struct ptlrpc_request *request = NULL;
2001         int update_mode;
2002         ENTRY;
2003
2004 #define fmt     "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
2005         DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
2006                   (unsigned) rec->ur_fid2->mds,
2007                   (unsigned) rec->ur_fid2->id,
2008                   (unsigned) rec->ur_fid2->generation,
2009                   rec->ur_namelen - 1, rec->ur_name,
2010                   (unsigned) rec->ur_fid1->mds,
2011                   (unsigned) rec->ur_fid1->id,
2012                   (unsigned)rec->ur_fid1->generation);
2013
2014         de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
2015                                            tgt_dir_lockh, &update_mode,
2016                                            rec->ur_name, rec->ur_namelen - 1,
2017                                            MDS_INODELOCK_UPDATE);
2018         if (IS_ERR(de_tgt_dir))
2019                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2020         cleanup_phase = 1;
2021
2022         op_data.fid1 = *(rec->ur_fid1);
2023         op_data.namelen = 0;
2024         op_data.name = NULL;
2025         rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2026         if (rc)
2027                 GOTO(cleanup, rc);
2028
2029         cleanup_phase = 2;
2030         if (request)
2031                 ptlrpc_req_finished(request);
2032
2033         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2034
2035         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2036         if (IS_ERR(handle)) {
2037                 rc = PTR_ERR(handle);
2038                 GOTO(cleanup, rc);
2039         }
2040
2041         rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2042                                   rec->ur_namelen - 1, rec->ur_fid1->id,
2043                                   rec->ur_fid1->generation, rec->ur_fid1->mds);
2044         cleanup_phase = 3;
2045
2046 cleanup:
2047         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2048                                 handle, req, rc, 0);
2049         EXIT;
2050
2051         switch (cleanup_phase) {
2052                 case 3:
2053                         if (rc) {
2054                                 /* FIXME: drop i_nlink on remote inode here */
2055                                 CERROR("MUST drop drop i_nlink here\n");
2056                         }
2057                 case 2:
2058                 case 1:
2059                         if (rc) {
2060                                 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2061 #ifdef S_PDIROPS
2062                                 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2063 #endif
2064                         } else {
2065                                 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2066 #ifdef S_PDIROPS
2067                                 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2068 #endif
2069                         }
2070                         l_dput(de_tgt_dir);
2071                         break;
2072                 default:
2073                         CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2074                         LBUG();
2075         }
2076         req->rq_status = rc;
2077         return 0;
2078 }
2079
2080 static int mds_reint_link(struct mds_update_record *rec, int offset,
2081                           struct ptlrpc_request *req,
2082                           struct lustre_handle *lh)
2083 {
2084         struct obd_device *obd = req->rq_export->exp_obd;
2085         struct dentry *de_src = NULL;
2086         struct dentry *de_tgt_dir = NULL;
2087         struct dentry *dchild = NULL;
2088         struct mds_obd *mds = mds_req2mds(req);
2089         struct lustre_handle *handle = NULL;
2090         struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2091         struct ldlm_res_id src_res_id = { .name = {0} };
2092         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2093         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2094         ldlm_policy_data_t tgt_dir_policy =
2095                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2096         int rc = 0, cleanup_phase = 0;
2097         int update_mode = 0;
2098         ENTRY;
2099
2100         LASSERT(offset == 0);
2101
2102         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2103                   rec->ur_fid1->id, rec->ur_fid1->generation,
2104                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2105
2106         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2107
2108         MDS_UPDATE_COUNTER(mds, MDS_LINK_COUNT);
2109
2110 //      memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2111         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2112                 GOTO(cleanup, rc = -ENOENT);
2113
2114         if (rec->ur_fid1->mds != mds->mds_num) {
2115                 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2116                 RETURN(rc);
2117         }
2118
2119         if (rec->ur_namelen == 1) {
2120                 rc = mds_reint_link_acquire(rec, offset, req, lh);
2121                 RETURN(rc);
2122         }
2123
2124         /* Step 1: Lookup the source inode and target directory by FID */
2125         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2126         if (IS_ERR(de_src))
2127                 GOTO(cleanup, rc = PTR_ERR(de_src));
2128
2129         cleanup_phase = 1; /* source dentry */
2130
2131         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2132         if (IS_ERR(de_tgt_dir)) {
2133                 rc = PTR_ERR(de_tgt_dir);
2134                 de_tgt_dir = NULL;
2135                 GOTO(cleanup, rc);
2136         }
2137
2138         cleanup_phase = 2; /* target directory dentry */
2139
2140         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2141                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2142                de_src->d_inode->i_ino);
2143
2144         /* Step 2: Take the two locks */
2145         src_res_id.name[0] = de_src->d_inode->i_ino;
2146         src_res_id.name[1] = de_src->d_inode->i_generation;
2147         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2148         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2149 #ifdef S_PDIROPS
2150         if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2151                 int flags = 0;
2152                 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2153                 if (update_mode) {
2154                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2155                                               tgt_dir_res_id, LDLM_IBITS,
2156                                               &src_policy, update_mode, &flags,
2157                                               mds_blocking_ast,
2158                                               ldlm_completion_ast, NULL, NULL,
2159                                               NULL, 0, NULL, tgt_dir_lockh + 1);
2160                         if (rc != ELDLM_OK)
2161                                 GOTO(cleanup, rc = -ENOLCK);
2162                 }
2163
2164                 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2165                                                         rec->ur_namelen - 1);
2166                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2167                        de_tgt_dir->d_inode->i_ino,
2168                        de_tgt_dir->d_inode->i_generation,
2169                        tgt_dir_res_id.name[2]);
2170         }
2171 #endif
2172         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2173                                    &src_policy,
2174                                    &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2175                                    &tgt_dir_policy);
2176         if (rc)
2177                 GOTO(cleanup, rc);
2178
2179         cleanup_phase = 3; /* locks */
2180
2181         /* Step 3: Lookup the child */
2182         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2183         if (IS_ERR(dchild)) {
2184                 rc = PTR_ERR(dchild);
2185                 if (rc != -EPERM && rc != -EACCES)
2186                         CERROR("child lookup error %d\n", rc);
2187                 GOTO(cleanup, rc);
2188         }
2189
2190         cleanup_phase = 4; /* child dentry */
2191
2192         if (dchild->d_inode) {
2193                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2194                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
2195                 rc = -EEXIST;
2196                 GOTO(cleanup, rc);
2197         }
2198
2199         /* Step 4: Do it. */
2200         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2201
2202         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2203         if (IS_ERR(handle)) {
2204                 rc = PTR_ERR(handle);
2205                 GOTO(cleanup, rc);
2206         }
2207
2208         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2209         if (rc && rc != -EPERM && rc != -EACCES)
2210                 CERROR("vfs_link error %d\n", rc);
2211 cleanup:
2212         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2213                                 handle, req, rc, 0);
2214         EXIT;
2215
2216         switch (cleanup_phase) {
2217         case 4: /* child dentry */
2218                 l_dput(dchild);
2219         case 3: /* locks */
2220                 if (rc) {
2221                         ldlm_lock_decref(&src_lockh, LCK_EX);
2222                         ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2223                 } else {
2224                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2225                         ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2226                 }
2227         case 2: /* target dentry */
2228 #ifdef S_PDIROPS
2229                 if (tgt_dir_lockh[1].cookie && update_mode)
2230                         ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2231 #endif
2232                 if (de_tgt_dir)
2233                         l_dput(de_tgt_dir);
2234         case 1: /* source dentry */
2235                 l_dput(de_src);
2236         case 0:
2237                 break;
2238         default:
2239                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2240                 LBUG();
2241         }
2242         req->rq_status = rc;
2243         return 0;
2244 }
2245
2246 /*
2247  * add a hard link in the PENDING directory, only used by rename()
2248  */
2249 static int mds_add_link_orphan(struct mds_update_record *rec,
2250                                struct obd_device *obd,
2251                                struct dentry *dentry)
2252 {
2253         struct mds_obd *mds = &obd->u.mds;
2254         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2255         struct dentry *pending_child;
2256         char fidname[LL_FID_NAMELEN];
2257         int fidlen = 0, rc;
2258         ENTRY;
2259
2260         LASSERT(dentry->d_inode);
2261         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2262
2263         down(&pending_dir->i_sem);
2264         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2265                             dentry->d_inode->i_generation);
2266
2267         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2268                mds_open_orphan_count(dentry->d_inode),
2269                rec->ur_name, fidname);
2270
2271         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2272         if (IS_ERR(pending_child))
2273                 GOTO(out_lock, rc = PTR_ERR(pending_child));
2274
2275         if (pending_child->d_inode != NULL) {
2276                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2277                 LASSERT(pending_child->d_inode == dentry->d_inode);
2278                 GOTO(out_dput, rc = 0);
2279         }
2280
2281         lock_kernel();
2282         rc = vfs_link(dentry, pending_dir, pending_child);
2283         unlock_kernel();
2284         if (rc)
2285                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2286                        rec->ur_name, rc);
2287         else
2288                 mds_inode_set_orphan(dentry->d_inode);
2289 out_dput:
2290         l_dput(pending_child);
2291 out_lock:
2292         up(&pending_dir->i_sem);
2293         RETURN(rc);
2294 }
2295
2296 /* The idea here is that we need to get four locks in the end:
2297  * one on each parent directory, one on each child.  We need to take
2298  * these locks in some kind of order (to avoid deadlocks), and the order
2299  * I selected is "increasing resource number" order.  We need to look up
2300  * the children, however, before we know what the resource number(s) are.
2301  * Thus the following plan:
2302  *
2303  * 1,2. Look up the parents
2304  * 3,4. Look up the children
2305  * 5. Take locks on the parents and children, in order
2306  * 6. Verify that the children haven't changed since they were looked up
2307  *
2308  * If there was a race and the children changed since they were first looked
2309  * up, it is possible that mds_verify_child() will be able to just grab the
2310  * lock on the new child resource (if it has a higher resource than any other)
2311  * but we need to compare against not only its parent, but also against the
2312  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2313  *
2314  * We need the fancy igrab() on the child inodes because we aren't holding a
2315  * lock on the parent after the lookup is done, so dentry->d_inode may change
2316  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2317  */
2318 static int mds_get_parents_children_locked(struct obd_device *obd,
2319                                            struct mds_obd *mds,
2320                                            struct ll_fid *p1_fid,
2321                                            struct dentry **de_srcdirp,
2322                                            struct ll_fid *p2_fid,
2323                                            struct dentry **de_tgtdirp,
2324                                            int parent_mode,
2325                                            const char *old_name, int old_len,
2326                                            struct dentry **de_oldp,
2327                                            const char *new_name, int new_len,
2328                                            struct dentry **de_newp,
2329                                            struct lustre_handle *dlm_handles,
2330                                            int child_mode)
2331 {
2332         struct ldlm_res_id p1_res_id = { .name = {0} };
2333         struct ldlm_res_id p2_res_id = { .name = {0} };
2334         struct ldlm_res_id c1_res_id = { .name = {0} };
2335         struct ldlm_res_id c2_res_id = { .name = {0} };
2336         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2337         /* Only dentry should change, but the inode itself would be
2338            intact otherwise */
2339         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2340         /* If something is going to be replaced, both dentry and inode locks are
2341            needed */
2342         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2343                                                         MDS_INODELOCK_UPDATE}};
2344         struct ldlm_res_id *maxres_src, *maxres_tgt;
2345         struct inode *inode;
2346         int rc = 0, cleanup_phase = 0;
2347         ENTRY;
2348
2349         /* Step 1: Lookup the source directory */
2350         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2351         if (IS_ERR(*de_srcdirp))
2352                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2353
2354         cleanup_phase = 1; /* source directory dentry */
2355
2356         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2357         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2358
2359         /* Step 2: Lookup the target directory */
2360         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2361                 *de_tgtdirp = dget(*de_srcdirp);
2362         } else {
2363                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2364                 if (IS_ERR(*de_tgtdirp)) {
2365                         rc = PTR_ERR(*de_tgtdirp);
2366                         *de_tgtdirp = NULL;
2367                         GOTO(cleanup, rc);
2368                 }
2369         }
2370
2371         cleanup_phase = 2; /* target directory dentry */
2372
2373         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2374         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2375
2376 #ifdef S_PDIROPS
2377         dlm_handles[5].cookie = 0;
2378         dlm_handles[6].cookie = 0;
2379         if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2380                 /* Get a temp lock on just ino, gen to flush client cache and
2381                  * to protect dirs from concurrent splitting */
2382                 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2383                                            LCK_PW, &p_policy, &p2_res_id,
2384                                            &(dlm_handles[6]),LCK_PW,&p_policy);
2385                 if (rc != ELDLM_OK)
2386                         GOTO(cleanup, rc);
2387                 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2388                 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2389                 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2390                        (*de_srcdirp)->d_inode->i_ino,
2391                        (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2392                        (*de_tgtdirp)->d_inode->i_ino,
2393                        (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2394         }
2395         cleanup_phase = 3;
2396 #endif
2397
2398         /* Step 3: Lookup the source child entry */
2399         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2400         if (IS_ERR(*de_oldp)) {
2401                 rc = PTR_ERR(*de_oldp);
2402                 CERROR("old child lookup error (%*s): %d\n",
2403                        old_len - 1, old_name, rc);
2404                 GOTO(cleanup, rc);
2405         }
2406
2407         cleanup_phase = 4; /* original name dentry */
2408
2409         inode = (*de_oldp)->d_inode;
2410         if (inode != NULL) {
2411                 inode = igrab(inode);
2412                 if (inode == NULL)
2413                         GOTO(cleanup, rc = -ENOENT);
2414
2415                 c1_res_id.name[0] = inode->i_ino;
2416                 c1_res_id.name[1] = inode->i_generation;
2417                 iput(inode);
2418         } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2419                 c1_res_id.name[0] = (*de_oldp)->d_inum;
2420                 c1_res_id.name[1] = (*de_oldp)->d_generation;
2421         }
2422
2423         /* Step 4: Lookup the target child entry */
2424         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2425         if (IS_ERR(*de_newp)) {
2426                 rc = PTR_ERR(*de_newp);
2427                 CERROR("new child lookup error (%*s): %d\n",
2428                        old_len - 1, old_name, rc);
2429                 GOTO(cleanup, rc);
2430         }
2431
2432         cleanup_phase = 5; /* target dentry */
2433
2434         inode = (*de_newp)->d_inode;
2435         if (inode != NULL) {
2436                 inode = igrab(inode);
2437                 if (inode == NULL)
2438                         goto retry_locks;
2439
2440                 c2_res_id.name[0] = inode->i_ino;
2441                 c2_res_id.name[1] = inode->i_generation;
2442                 iput(inode);
2443         } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2444                 c2_res_id.name[0] = (*de_newp)->d_inum;
2445                 c2_res_id.name[1] = (*de_newp)->d_generation;
2446         }
2447
2448 retry_locks:
2449         /* Step 5: Take locks on the parents and child(ren) */
2450         maxres_src = &p1_res_id;
2451         maxres_tgt = &p2_res_id;
2452         cleanup_phase = 5; /* target dentry */
2453
2454         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2455                 maxres_src = &c1_res_id;
2456         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2457                 maxres_tgt = &c2_res_id;
2458
2459         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2460                                     &p_policy,
2461                                     &p2_res_id, &dlm_handles[1], parent_mode,
2462                                     &p_policy,
2463                                     &c1_res_id, &dlm_handles[2], child_mode,
2464                                     &c1_policy,
2465                                     &c2_res_id, &dlm_handles[3], child_mode,
2466                                     &c2_policy);
2467         if (rc)
2468                 GOTO(cleanup, rc);
2469
2470         cleanup_phase = 6; /* parent and child(ren) locks */
2471
2472         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2473         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2474                               parent_mode, &c1_res_id, &dlm_handles[2],
2475                               de_oldp, child_mode, &c1_policy, old_name,old_len,
2476                               maxres_tgt);
2477         if (rc) {
2478                 if (c2_res_id.name[0] != 0)
2479                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2480                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2481                 cleanup_phase = 5;
2482                 if (rc > 0)
2483                         goto retry_locks;
2484                 GOTO(cleanup, rc);
2485         }
2486
2487         if (!DENTRY_VALID(*de_oldp))
2488                 GOTO(cleanup, rc = -ENOENT);
2489
2490         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2491         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2492                               parent_mode, &c2_res_id, &dlm_handles[3],
2493                               de_newp, child_mode, &c2_policy, new_name,
2494                               new_len, maxres_src);
2495         if (rc) {
2496                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2497                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2498                 cleanup_phase = 5;
2499                 if (rc > 0)
2500                         goto retry_locks;
2501                 GOTO(cleanup, rc);
2502         }
2503
2504         EXIT;
2505 cleanup:
2506         if (rc) {
2507                 switch (cleanup_phase) {
2508                 case 6: /* child lock(s) */
2509                         if (c2_res_id.name[0] != 0)
2510                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2511                         if (c1_res_id.name[0] != 0)
2512                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2513                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2514                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2515                 case 5: /* target dentry */
2516                         l_dput(*de_newp);
2517                 case 4: /* source dentry */
2518                         l_dput(*de_oldp);
2519                 case 3:
2520 #ifdef S_PDIROPS
2521                         if (dlm_handles[5].cookie != 0)
2522                                 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2523                         if (dlm_handles[6].cookie != 0)
2524                                 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2525 #endif
2526                 case 2: /* target directory dentry */
2527                         l_dput(*de_tgtdirp);
2528                 case 1: /* source directry dentry */
2529                         l_dput(*de_srcdirp);
2530                 }
2531         }
2532
2533         return rc;
2534 }
2535
2536 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2537                                         int offset, struct ptlrpc_request *req)
2538 {
2539         struct obd_device *obd = req->rq_export->exp_obd;
2540         struct dentry *de_srcdir = NULL;
2541         struct dentry *de_new = NULL;
2542         struct mds_obd *mds = mds_req2mds(req);
2543         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2544         struct lustre_handle child_lockh = {0};
2545         int cleanup_phase = 0;
2546         void *handle = NULL;
2547         int update_mode, rc = 0;
2548         ENTRY;
2549
2550         /* another MDS executing rename operation has asked us
2551          * to create target name. such a creation should destroy
2552          * existing target name */
2553
2554         CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2555                         obd->obd_name, rec->ur_tgt,
2556                         (unsigned long) rec->ur_fid1->mds,
2557                         (unsigned long) rec->ur_fid1->id,
2558                         (unsigned long) rec->ur_fid1->generation);
2559
2560         /* first, lookup the target */
2561         child_lockh.cookie = 0;
2562         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2563                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2564                                          &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2565                                          &child_lockh, &de_new, LCK_EX,
2566                                          MDS_INODELOCK_LOOKUP, NULL);
2567         if (rc)
2568                 GOTO(cleanup, rc);
2569
2570         cleanup_phase = 1;
2571
2572         LASSERT(de_srcdir);
2573         LASSERT(de_srcdir->d_inode);
2574         LASSERT(de_new);
2575
2576         if (de_new->d_inode) {
2577                 /* name exists and points to local inode
2578                  * try to unlink this name and create new one */
2579                 CERROR("%s: %s points to local inode %lu/%lu\n",
2580                        obd->obd_name, rec->ur_tgt,
2581                        (unsigned long) de_new->d_inode->i_ino,
2582                        (unsigned long) de_new->d_inode->i_generation);
2583                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2584                                       FSFILT_OP_RENAME, NULL);
2585                 if (IS_ERR(handle))
2586                         GOTO(cleanup, rc = PTR_ERR(handle));
2587                 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2588                 if (rc)
2589                         GOTO(cleanup, rc);
2590         } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2591                 /* name exists adn points to remove inode */
2592                 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2593                        obd->obd_name, rec->ur_tgt,
2594                        (unsigned long) de_new->d_mdsnum,
2595                        (unsigned long) de_new->d_inum,
2596                        (unsigned long) de_new->d_generation);
2597         } else {
2598                 /* name doesn't exist. the simplest case */
2599                 handle = fsfilt_start(obd, de_srcdir->d_inode,
2600                                       FSFILT_OP_LINK, NULL);
2601                 if (IS_ERR(handle))
2602                         GOTO(cleanup, rc = PTR_ERR(handle));
2603         }
2604
2605         cleanup_phase = 2;
2606         rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2607                         rec->ur_tgtlen - 1, rec->ur_fid1->id,
2608                         rec->ur_fid1->generation, rec->ur_fid1->mds);
2609         if (rc)
2610                 CERROR("add_dir_entry() returned error %d\n", rc);
2611 cleanup:
2612         EXIT;
2613         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2614                                 handle, req, rc, 0);
2615         switch(cleanup_phase) {
2616                 case 2:
2617                 case 1:
2618 #ifdef S_PDIROPS
2619                         if (parent_lockh[1].cookie != 0)
2620                                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2621 #endif
2622                         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2623                         if (child_lockh.cookie != 0)
2624                                 ldlm_lock_decref(&child_lockh, LCK_EX);
2625                         l_dput(de_new);
2626                         l_dput(de_srcdir);
2627                         break;
2628                 default:
2629                         LBUG();
2630         }
2631
2632         req->rq_status = rc;
2633
2634         RETURN(0);
2635 }
2636
2637 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2638                                       struct ptlrpc_request *req)
2639 {
2640         struct obd_device *obd = req->rq_export->exp_obd;
2641         struct ptlrpc_request *req2 = NULL;
2642         struct dentry *de_srcdir = NULL;
2643         struct dentry *de_old = NULL;
2644         struct mds_obd *mds = mds_req2mds(req);
2645         struct lustre_handle parent_lockh[2] = {{0}, {0}};
2646         struct lustre_handle child_lockh = {0};
2647         struct mdc_op_data opdata;
2648         void *handle = NULL;
2649         int update_mode, rc = 0;
2650         ENTRY;
2651
2652         CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2653                obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2654         memset(&opdata, 0, sizeof(opdata));
2655
2656         child_lockh.cookie = 0;
2657         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2658                                          &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2659                                          &update_mode, rec->ur_name,
2660                                          rec->ur_namelen, &child_lockh, &de_old,
2661                                          LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2662         LASSERT(rc == 0);
2663         LASSERT(de_srcdir);
2664         LASSERT(de_srcdir->d_inode);
2665         LASSERT(de_old);
2666
2667         /* we already know the target should be created on another MDS
2668          * so, we have to request that MDS to do it */
2669
2670         /* prepare source fid */
2671         if (de_old->d_flags & DCACHE_CROSS_REF) {
2672                 LASSERT(de_old->d_inode == NULL);
2673                 CDEBUG(D_OTHER, "request to move remote name\n");
2674                 opdata.fid1.mds = de_old->d_mdsnum;
2675                 opdata.fid1.id = de_old->d_inum;
2676                 opdata.fid1.generation = de_old->d_generation;
2677         } else if (de_old->d_inode == NULL) {
2678                 /* oh, source doesn't exist */
2679                 GOTO(cleanup, rc = -ENOENT);
2680         } else {
2681                 LASSERT(de_old->d_inode != NULL);
2682                 CDEBUG(D_OTHER, "request to move local name\n");
2683                 opdata.fid1.mds = mds->mds_num;
2684                 opdata.fid1.id = de_old->d_inode->i_ino;
2685                 opdata.fid1.generation = de_old->d_inode->i_generation;
2686         }
2687
2688         opdata.fid2 = *(rec->ur_fid2);
2689         rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2690                        rec->ur_tgtlen - 1, &req2);
2691
2692         if (rc)
2693                 GOTO(cleanup, rc);
2694
2695         handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2696         if (IS_ERR(handle))
2697                 GOTO(cleanup, rc = PTR_ERR(handle));
2698         rc = fsfilt_del_dir_entry(obd, de_old);
2699         d_drop(de_old);
2700
2701 cleanup:
2702         EXIT;
2703         rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2704                                 handle, req, rc, 0);
2705         if (req2)
2706                 ptlrpc_req_finished(req2);
2707
2708 #ifdef S_PDIROPS
2709         if (parent_lockh[1].cookie != 0)
2710                 ldlm_lock_decref(&parent_lockh[1], update_mode);
2711 #endif
2712         ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2713         if (child_lockh.cookie != 0)
2714                 ldlm_lock_decref(&child_lockh, LCK_EX);
2715
2716         l_dput(de_old);
2717         l_dput(de_srcdir);
2718
2719         req->rq_status = rc;
2720         RETURN(0);
2721
2722 }
2723
2724 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2725                             struct ptlrpc_request *req, struct lustre_handle *lockh)
2726 {
2727         struct obd_device *obd = req->rq_export->exp_obd;
2728         struct dentry *de_srcdir = NULL;
2729         struct dentry *de_tgtdir = NULL;
2730         struct dentry *de_old = NULL;
2731         struct dentry *de_new = NULL;
2732         struct mds_obd *mds = mds_req2mds(req);
2733         struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2734         struct mds_body *body = NULL;
2735         struct llog_create_locks *lcl = NULL;
2736         int rc = 0, log_unlink = 0, lock_count = 3;
2737         int cleanup_phase = 0;
2738         void *handle = NULL;
2739         ENTRY;
2740
2741         LASSERT(offset == 0);
2742
2743         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2744                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2745                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2746
2747         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2748
2749         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2750                 DEBUG_REQ(D_HA, req, "rename replay\n");
2751                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2752                        lustre_msg_buf(req->rq_reqmsg, 3, 0),
2753                        req->rq_repmsg->buflens[2]);
2754         }
2755
2756         MDS_UPDATE_COUNTER(mds, MDS_RENAME_COUNT);
2757
2758         if (rec->ur_namelen == 1) {
2759                 rc = mds_reint_rename_create_name(rec, offset, req);
2760                 RETURN(rc);
2761         }
2762
2763         if (rec->ur_fid2->mds != mds->mds_num) {
2764                 rc = mds_reint_rename_to_remote(rec, offset, req);
2765                 RETURN(rc);
2766         }
2767
2768         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2769                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
2770                                              rec->ur_name, rec->ur_namelen,
2771                                              &de_old, rec->ur_tgt,
2772                                              rec->ur_tgtlen, &de_new,
2773                                              dlm_handles, LCK_EX);
2774         if (rc)
2775                 GOTO(cleanup, rc);
2776
2777         cleanup_phase = 1; /* parent(s), children, locks */
2778
2779         if (de_new->d_inode)
2780                 lock_count = 4;
2781
2782         /* sanity check for src inode */
2783         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2784             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2785                 GOTO(cleanup, rc = -EINVAL);
2786
2787         /* sanity check for dest inode */
2788         if (de_new->d_inode &&
2789             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2790              de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2791                 GOTO(cleanup, rc = -EINVAL);
2792
2793         if (de_old->d_inode == de_new->d_inode)
2794                 GOTO(cleanup, rc = 0);
2795
2796         /* if we are about to remove the target at first, pass the EA of
2797          * that inode to client to perform and cleanup on OST */
2798         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2799         LASSERT(body != NULL);
2800
2801         if (de_new->d_inode &&
2802             S_ISREG(de_new->d_inode->i_mode) &&
2803             de_new->d_inode->i_nlink == 1 &&
2804             mds_open_orphan_count(de_new->d_inode) == 0) {
2805                 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2806                 mds_pack_inode2body(obd, body, de_new->d_inode);
2807                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2808                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2809                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2810                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2811                 } else {
2812                         log_unlink = 1;
2813                 }
2814         }
2815
2816         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2817                        de_srcdir->d_inode->i_sb);
2818
2819         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2820         if (IS_ERR(handle))
2821                 GOTO(cleanup, rc = PTR_ERR(handle));
2822
2823         /* FIXME need adjust the journal block count? */
2824         /* if the target should be moved to PENDING, we at first increase the
2825          * link and later vfs_rename() will decrease the link count again */
2826         if (de_new->d_inode &&
2827             S_ISREG(de_new->d_inode->i_mode) &&
2828             de_new->d_inode->i_nlink == 1 &&
2829             mds_open_orphan_count(de_new->d_inode) > 0) {
2830                 rc = mds_add_link_orphan(rec, obd, de_new);
2831                 if (rc)
2832                         GOTO(cleanup, rc);
2833         }
2834
2835         lock_kernel();
2836         de_old->d_fsdata = req;
2837         de_new->d_fsdata = req;
2838         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2839         unlock_kernel();
2840
2841         if (!rc && log_unlink) {
2842                 if (mds_log_op_unlink(obd, de_tgtdir->d_inode,
2843                                 lustre_msg_buf(req->rq_repmsg, 1, 0),
2844                                 req->rq_repmsg->buflens[1],
2845                                 lustre_msg_buf(req->rq_repmsg, 2, 0),
2846                                 req->rq_repmsg->buflens[2], &lcl) > 0) {
2847                         body->valid |= OBD_MD_FLCOOKIE;
2848                 }
2849         }
2850
2851         GOTO(cleanup, rc);
2852 cleanup:
2853         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2854                                 handle, req, rc, 0);
2855         switch (cleanup_phase) {
2856         case 1:
2857 #ifdef S_PDIROPS
2858                 if (dlm_handles[5].cookie != 0)
2859                         ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2860                 if (dlm_handles[6].cookie != 0)
2861                         ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2862 #endif
2863                 if (lcl != NULL)
2864                         ptlrpc_save_llog_lock(req, lcl);
2865
2866                 if (rc) {
2867                         if (lock_count == 4)
2868                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2869                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2870                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2871                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2872                 } else {
2873                         if (lock_count == 4)
2874                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2875                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2876                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2877                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2878                 }
2879                 l_dput(de_new);
2880                 l_dput(de_old);
2881                 l_dput(de_tgtdir);
2882                 l_dput(de_srcdir);
2883         case 0:
2884                 break;
2885         default:
2886                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2887                 LBUG();
2888         }
2889         req->rq_status = rc;
2890         return 0;
2891 }
2892
2893 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2894                            struct ptlrpc_request *, struct lustre_handle *);
2895
2896 static mds_reinter reinters[REINT_MAX + 1] = {
2897         [REINT_SETATTR] mds_reint_setattr,
2898         [REINT_CREATE] mds_reint_create,
2899         [REINT_LINK] mds_reint_link,
2900         [REINT_UNLINK] mds_reint_unlink,
2901         [REINT_RENAME] mds_reint_rename,
2902         [REINT_OPEN] mds_open
2903 };
2904
2905 int mds_reint_rec(struct mds_update_record *rec, int offset,
2906                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2907 {
2908         struct obd_device *obd = req->rq_export->exp_obd;
2909         struct lvfs_run_ctxt saved;
2910         int rc;
2911
2912         /* checked by unpacker */
2913         LASSERT(rec->ur_opcode <= REINT_MAX &&
2914                 reinters[rec->ur_opcode] != NULL);
2915
2916         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2917         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2918         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2919
2920         return rc;
2921 }