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_fsfilt.h>
  42
  43 #include "mds_internal.h"
  44
  45 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  46                    int error)
  47 {
  48         obd_transno_commit_cb(obd, transno, error);
  49 }
  50
  51 struct mds_logcancel_data {
  52         struct lov_mds_md      *mlcd_lmm;
  53         int                     mlcd_size;
  54         int                     mlcd_cookielen;
  55         int                     mlcd_eadatalen;
  56         struct llog_cookie      mlcd_cookies[0];
  57 };
  58
  59
  60 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  61                                   void *cb_data, int error)
  62 {
  63         struct mds_logcancel_data *mlcd = cb_data;
  64         struct lov_stripe_md *lsm = NULL;
  65 #ifdef ENABLE_ORPHANS
  66         struct llog_ctxt *ctxt;
  67 #endif
  68         int rc;
  69
  70         obd_transno_commit_cb(obd, transno, error);
  71
  72         CDEBUG(D_HA, "cancelling %d cookies\n",
  73                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  74
  75         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  76                           mlcd->mlcd_eadatalen);
  77         if (rc < 0) {
  78                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  79                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  80                        rc);
  81         } else {
  82 #ifdef ENABLE_ORPHANS
  83                 ///* XXX 0 normally, SENDNOW for debug */);
  84                 ctxt = llog_get_context(obd, mlcd->mlcd_cookies[0].lgc_subsys + 1);
  85                 rc = llog_cancel(ctxt, lsm,
  86                                          mlcd->mlcd_cookielen /
  87                                          sizeof(*mlcd->mlcd_cookies),
  88                                          mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  89                 if (rc)
  90                         CERROR("error cancelling %d log cookies: rc %d\n",
  91                                (int)(mlcd->mlcd_cookielen /
  92                                      sizeof(*mlcd->mlcd_cookies)), rc);
  93 #endif
  94         }
  95
  96         OBD_FREE(mlcd, mlcd->mlcd_size);
  97 }
  98
  99 /* Assumes caller has already pushed us into the kernel context. */
 100 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
 101                        struct ptlrpc_request *req, int rc, __u32 op_data)
 102 {
 103         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 104         struct mds_client_data *mcd = med->med_mcd;
 105         struct obd_device *obd = req->rq_export->exp_obd;
 106         int err;
 107         __u64 transno;
 108         loff_t off;
 109         int log_pri = D_HA;
 110         ENTRY;
 111
 112         /* if the export has already been failed, we have no last_rcvd slot */
 113         if (req->rq_export->exp_failed) {
 114                 CERROR("committing transaction for disconnected client\n");
 115                 if (handle)
 116                         GOTO(commit, rc);
 117                 RETURN(rc);
 118         }
 119
 120         if (IS_ERR(handle))
 121                 RETURN(rc);
 122
 123         if (handle == NULL) {
 124                 /* if we're starting our own xaction, use our own inode */
 125                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 126                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 127                 if (IS_ERR(handle)) {
 128                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 129                         RETURN(PTR_ERR(handle));
 130                 }
 131         }
 132
 133         off = med->med_off;
 134
 135         transno = req->rq_reqmsg->transno;
 136         if (transno == 0) {
 137                 spin_lock(&mds->mds_transno_lock);
 138                 transno = ++mds->mds_last_transno;
 139                 spin_unlock(&mds->mds_transno_lock);
 140         } else {
 141                 spin_lock(&mds->mds_transno_lock);
 142                 if (transno > mds->mds_last_transno)
 143                         mds->mds_last_transno = transno;
 144                 spin_unlock(&mds->mds_transno_lock);
 145         }
 146         req->rq_repmsg->transno = req->rq_transno = transno;
 147         mcd->mcd_last_transno = cpu_to_le64(transno);
 148         mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
 149         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 150         mcd->mcd_last_result = cpu_to_le32(rc);
 151         mcd->mcd_last_data = cpu_to_le32(op_data);
 152
 153         fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 154                               mds_commit_cb, NULL);
 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\n",
 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, 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_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 279 {
 280         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 281                   mcd->mcd_last_transno, mcd->mcd_last_result);
 282         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 283         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 284
 285         if (req->rq_export->exp_outstanding_reply)
 286                 mds_steal_ack_locks(req->rq_export, req);
 287 }
 288
 289 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 290                                       int offset, struct ptlrpc_request *req)
 291 {
 292         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 293         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 294         struct dentry *de;
 295         struct mds_body *body;
 296
 297         mds_req_from_mcd(req, med->med_mcd);
 298
 299         if (req->rq_export->exp_outstanding_reply)
 300                 mds_steal_ack_locks(req->rq_export, req);
 301
 302         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 303         if (IS_ERR(de)) {
 304                 LASSERT(PTR_ERR(de) == req->rq_status);
 305                 return;
 306         }
 307
 308         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 309         mds_pack_inode2fid(&body->fid1, de->d_inode);
 310         mds_pack_inode2body(body, de->d_inode);
 311
 312         /* Don't return OST-specific attributes if we didn't just set them */
 313         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 314                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 315         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 316                 body->valid |= OBD_MD_FLMTIME;
 317         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 318                 body->valid |= OBD_MD_FLATIME;
 319
 320         l_dput(de);
 321 }
 322
 323 /* In the raw-setattr case, we lock the child inode.
 324  * In the write-back case or if being called from open, the client holds a lock
 325  * already.
 326  *
 327  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 328 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 329                              struct ptlrpc_request *req,
 330                              struct lustre_handle *lh)
 331 {
 332         struct mds_obd *mds = mds_req2mds(req);
 333         struct obd_device *obd = req->rq_export->exp_obd;
 334         struct mds_body *body;
 335         struct dentry *de;
 336         struct inode *inode = NULL;
 337         struct lustre_handle lockh;
 338         void *handle = NULL;
 339         struct mds_logcancel_data *mlcd = NULL;
 340         int rc = 0, cleanup_phase = 0, err, locked = 0;
 341         ENTRY;
 342
 343         LASSERT(offset == 0);
 344
 345         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 346
 347         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
 348                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 349                 if (IS_ERR(de))
 350                         GOTO(cleanup, rc = PTR_ERR(de));
 351         } else {
 352                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 353                                            &lockh, NULL, 0);
 354                 if (IS_ERR(de))
 355                         GOTO(cleanup, rc = PTR_ERR(de));
 356                 locked = 1;
 357         }
 358
 359         cleanup_phase = 1;
 360         inode = de->d_inode;
 361         LASSERT(inode);
 362         if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
 363                 down(&inode->i_sem);
 364
 365         CDEBUG(D_INODE, "ino %lu\n", inode->i_ino);
 366         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 367
 368         handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 369         if (IS_ERR(handle))
 370                 GOTO(cleanup, rc = PTR_ERR(handle));
 371
 372         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 373                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 374                        LTIME_S(rec->ur_iattr.ia_mtime),
 375                        LTIME_S(rec->ur_iattr.ia_ctime));
 376         rc = mds_fix_attr(inode, rec);
 377         if (rc)
 378                 GOTO(cleanup, rc);
 379
 380         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 381                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 382                                       (long)&rec->ur_iattr.ia_attr_flags);
 383         else                                            /* setattr */
 384                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 385
 386         if (rc == 0 && S_ISREG(inode->i_mode) && rec->ur_eadata != NULL) {
 387                 rc = fsfilt_set_md(obd, inode, handle,
 388                                    rec->ur_eadata, rec->ur_eadatalen);
 389         }
 390
 391         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 392         mds_pack_inode2fid(&body->fid1, inode);
 393         mds_pack_inode2body(body, inode);
 394
 395         /* Don't return OST-specific attributes if we didn't just set them */
 396         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 397                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 398         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 399                 body->valid |= OBD_MD_FLMTIME;
 400         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 401                 body->valid |= OBD_MD_FLATIME;
 402
 403         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 404                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 405                           rec->ur_eadatalen);
 406                 if (mlcd) {
 407                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 408                                 rec->ur_eadatalen;
 409                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 410                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 411                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 412                                 mlcd->mlcd_cookielen;
 413                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 414                                mlcd->mlcd_cookielen);
 415                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 416                                mlcd->mlcd_eadatalen);
 417                 } else {
 418                         CERROR("unable to allocate log cancel data\n");
 419                 }
 420         }
 421         EXIT;
 422  cleanup:
 423         if (mlcd != NULL)
 424                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 425                                       mds_cancel_cookies_cb, mlcd);
 426         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 427         switch (cleanup_phase) {
 428         case 1:
 429                 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
 430                         down(&inode->i_sem);
 431                 l_dput(de);
 432                 if (locked) {
 433                         if (rc) {
 434                                 ldlm_lock_decref(&lockh, LCK_PW);
 435                         } else {
 436                                 ldlm_put_lock_into_req(req, &lockh, LCK_PW);
 437                         }
 438                 }
 439         case 0:
 440                 break;
 441         default:
 442                 LBUG();
 443         }
 444         if (err && !rc)
 445                 rc = err;
 446
 447         req->rq_status = rc;
 448         return 0;
 449 }
 450
 451 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 452                                      struct ptlrpc_request *req)
 453 {
 454         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 455         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 456         struct dentry *parent, *child;
 457         struct mds_body *body;
 458
 459         mds_req_from_mcd(req, med->med_mcd);
 460
 461         if (req->rq_status)
 462                 return;
 463
 464         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 465         LASSERT(!IS_ERR(parent));
 466         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 467         LASSERT(!IS_ERR(child));
 468         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 469         mds_pack_inode2fid(&body->fid1, child->d_inode);
 470         mds_pack_inode2body(body, child->d_inode);
 471         l_dput(parent);
 472         l_dput(child);
 473 }
 474
 475 static int mds_reint_create(struct mds_update_record *rec, int offset,
 476                             struct ptlrpc_request *req,
 477                             struct lustre_handle *lh)
 478 {
 479         struct dentry *dparent = NULL;
 480         struct mds_obd *mds = mds_req2mds(req);
 481         struct obd_device *obd = req->rq_export->exp_obd;
 482         struct dentry *dchild = NULL;
 483         struct inode *dir = NULL;
 484         void *handle = NULL;
 485         struct lustre_handle lockh;
 486         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 487         int created = 0;
 488         struct dentry_params dp;
 489         ENTRY;
 490
 491         LASSERT(offset == 0);
 492         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 493
 494         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 495
 496         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 497                 GOTO(cleanup, rc = -ESTALE);
 498
 499         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
 500                                         rec->ur_name, rec->ur_namelen - 1);
 501         if (IS_ERR(dparent)) {
 502                 rc = PTR_ERR(dparent);
 503                 CERROR("parent lookup error %d\n", rc);
 504                 GOTO(cleanup, rc);
 505         }
 506         cleanup_phase = 1; /* locked parent dentry */
 507         dir = dparent->d_inode;
 508         LASSERT(dir);
 509         CDEBUG(D_INODE, "parent ino %lu creating name %s mode %o\n",
 510                dir->i_ino, rec->ur_name, rec->ur_mode);
 511
 512         ldlm_lock_dump_handle(D_OTHER, &lockh);
 513
 514         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 515         if (IS_ERR(dchild)) {
 516                 rc = PTR_ERR(dchild);
 517                 CERROR("child lookup error %d\n", rc);
 518                 GOTO(cleanup, rc);
 519         }
 520
 521         cleanup_phase = 2; /* child dentry */
 522
 523         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 524
 525         if (dir->i_mode & S_ISGID) {
 526                 if (S_ISDIR(rec->ur_mode))
 527                         rec->ur_mode |= S_ISGID;
 528         }
 529
 530         dchild->d_fsdata = (void *)&dp;
 531         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 532         dp.p_ptr = req;
 533
 534         switch (type) {
 535         case S_IFREG:{
 536                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 537                 if (IS_ERR(handle))
 538                         GOTO(cleanup, rc = PTR_ERR(handle));
 539                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 540                 EXIT;
 541                 break;
 542         }
 543         case S_IFDIR:{
 544                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 545                 if (IS_ERR(handle))
 546                         GOTO(cleanup, rc = PTR_ERR(handle));
 547                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 548                 EXIT;
 549                 break;
 550         }
 551         case S_IFLNK:{
 552                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 553                 if (IS_ERR(handle))
 554                         GOTO(cleanup, rc = PTR_ERR(handle));
 555                 if (rec->ur_tgt == NULL)        /* no target supplied */
 556                         rc = -EINVAL;           /* -EPROTO? */
 557                 else
 558                         rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 559                 EXIT;
 560                 break;
 561         }
 562         case S_IFCHR:
 563         case S_IFBLK:
 564         case S_IFIFO:
 565         case S_IFSOCK:{
 566                 int rdev = rec->ur_rdev;
 567                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 568                 if (IS_ERR(handle))
 569                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 570                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 571                 EXIT;
 572                 break;
 573         }
 574         default:
 575                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 576                 dchild->d_fsdata = NULL;
 577                 GOTO(cleanup, rc = -EINVAL);
 578         }
 579
 580         /* In case we stored the desired inum in here, we want to clean up. */
 581         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 582                 dchild->d_fsdata = NULL;
 583
 584         if (rc) {
 585                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 586                 GOTO(cleanup, rc);
 587         } else {
 588                 struct iattr iattr;
 589                 struct inode *inode = dchild->d_inode;
 590                 struct mds_body *body;
 591
 592                 created = 1;
 593                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 594                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 595                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 596                 iattr.ia_uid = rec->ur_fsuid;
 597                 if (dir->i_mode & S_ISGID)
 598                         iattr.ia_gid = dir->i_gid;
 599                 else
 600                         iattr.ia_gid = rec->ur_fsgid;
 601                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 602                         ATTR_MTIME | ATTR_CTIME;
 603
 604                 if (rec->ur_fid2->id) {
 605                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 606                         inode->i_generation = rec->ur_fid2->generation;
 607                         /* Dirtied and committed by the upcoming setattr. */
 608                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 609                                inode->i_ino, inode->i_generation);
 610                 } else {
 611                         struct lustre_handle child_ino_lockh;
 612                         struct ldlm_res_id child_res_id =
 613                              { .name = { inode->i_ino, 0 } };
 614                         int lock_flags = 0;
 615
 616                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 617                                inode->i_ino, inode->i_generation);
 618
 619                         /* The inode we were allocated may have just been freed
 620                          * by an unlink operation.  We take this lock to
 621                          * synchronize against the matching reply-ack-lock taken
 622                          * in unlink, to avoid replay problems if this reply
 623                          * makes it out to the client but the unlink's does not.
 624                          * See bug 2029 for more detail.*/
 625                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 626                                               NULL, child_res_id, LDLM_PLAIN,
 627                                               NULL, 0, LCK_EX, &lock_flags,
 628                                               ldlm_completion_ast,
 629                                               mds_blocking_ast, NULL,
 630                                               &child_ino_lockh);
 631                         if (rc != ELDLM_OK) {
 632                                 CERROR("error locking for unlink/create sync: "
 633                                        "%d\n", rc);
 634                         } else {
 635                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 636                         }
 637                 }
 638
 639                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 640                 if (rc)
 641                         CERROR("error on child setattr: rc = %d\n", rc);
 642
 643                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 644                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 645                 if (rc)
 646                         CERROR("error on parent setattr: rc = %d\n", rc);
 647
 648                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 649                 mds_pack_inode2fid(&body->fid1, inode);
 650                 mds_pack_inode2body(body, inode);
 651         }
 652         EXIT;
 653
 654 cleanup:
 655         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 656
 657         if (rc && created) {
 658                 /* Destroy the file we just created.  This should not need
 659                  * extra journal credits, as we have already modified all of
 660                  * the blocks needed in order to create the file in the first
 661                  * place.
 662                  */
 663                 switch (type) {
 664                 case S_IFDIR:
 665                         err = vfs_rmdir(dir, dchild);
 666                         if (err)
 667                                 CERROR("rmdir in error path: %d\n", err);
 668                         break;
 669                 default:
 670                         err = vfs_unlink(dir, dchild);
 671                         if (err)
 672                                 CERROR("unlink in error path: %d\n", err);
 673                         break;
 674                 }
 675         } else {
 676                 rc = err;
 677         }
 678         switch (cleanup_phase) {
 679         case 2: /* child dentry */
 680                 l_dput(dchild);
 681         case 1: /* locked parent dentry */
 682                 if (rc) {
 683                         ldlm_lock_decref(&lockh, LCK_PW);
 684                 } else {
 685                         ldlm_put_lock_into_req(req, &lockh, LCK_PW);
 686                 }
 687                 l_dput(dparent);
 688         case 0:
 689                 break;
 690         default:
 691                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 692                 LBUG();
 693         }
 694         req->rq_status = rc;
 695         return 0;
 696 }
 697
 698 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 699  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 700  * because they take the place of local semaphores.
 701  *
 702  * Two locks are taken in numerical order */
 703 int enqueue_ordered_locks(int lock_mode, struct obd_device *obd,
 704                           struct ldlm_res_id *p1_res_id,
 705                           struct ldlm_res_id *p2_res_id,
 706                           struct lustre_handle *p1_lockh,
 707                           struct lustre_handle *p2_lockh)
 708 {
 709         struct ldlm_res_id res_id[2];
 710         struct lustre_handle *handles[2] = {p1_lockh, p2_lockh};
 711         int rc, flags;
 712         ENTRY;
 713
 714         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 715
 716         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
 717                p1_res_id[0].name[0], p2_res_id[0].name[0]);
 718
 719         if (p1_res_id->name[0] < p2_res_id->name[0]) {
 720                 handles[0] = p1_lockh;
 721                 handles[1] = p2_lockh;
 722                 res_id[0] = *p1_res_id;
 723                 res_id[1] = *p2_res_id;
 724         } else {
 725                 handles[1] = p1_lockh;
 726                 handles[0] = p2_lockh;
 727                 res_id[1] = *p1_res_id;
 728                 res_id[0] = *p2_res_id;
 729         }
 730
 731         CDEBUG(D_INFO, "lock order: "LPU64"/"LPU64"\n",
 732                p1_res_id[0].name[0], p2_res_id[0].name[0]);
 733
 734         flags = LDLM_FL_LOCAL_ONLY;
 735         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL, res_id[0],
 736                               LDLM_PLAIN, NULL, 0, lock_mode, &flags,
 737                               ldlm_completion_ast, mds_blocking_ast, NULL,
 738                               handles[0]);
 739         if (rc != ELDLM_OK)
 740                 RETURN(-EIO);
 741         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 742
 743         if (memcmp(&res_id[0], &res_id[1], sizeof(res_id[0])) == 0) {
 744                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 745                 ldlm_lock_addref(handles[1], lock_mode);
 746         } else {
 747                 flags = LDLM_FL_LOCAL_ONLY;
 748                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 749                                       res_id[1], LDLM_PLAIN, NULL, 0, lock_mode,
 750                                       &flags, ldlm_completion_ast,
 751                                       mds_blocking_ast, NULL, handles[1]);
 752                 if (rc != ELDLM_OK) {
 753                         ldlm_lock_decref(handles[0], lock_mode);
 754                         RETURN(-EIO);
 755                 }
 756         }
 757         ldlm_lock_dump_handle(D_OTHER, handles[1]);
 758
 759         RETURN(0);
 760 }
 761
 762 void mds_reconstruct_generic(struct ptlrpc_request *req)
 763 {
 764         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 765
 766         mds_req_from_mcd(req, med->med_mcd);
 767 }
 768
 769 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
 770                             struct ptlrpc_request *req,
 771                             struct lustre_handle *lh)
 772 {
 773         struct dentry *dparent = NULL;
 774         struct dentry *dchild = NULL;
 775         struct mds_obd *mds = mds_req2mds(req);
 776         struct obd_device *obd = req->rq_export->exp_obd;
 777         struct mds_body *body = NULL;
 778         struct inode *child_inode;
 779         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
 780         void *handle = NULL;
 781         struct ldlm_res_id child_res_id = { .name = {0} };
 782         int rc = 0, flags = 0, log_unlink = 0, cleanup_phase = 0;
 783         ENTRY;
 784
 785         LASSERT(offset == 0 || offset == 2);
 786
 787         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
 788
 789         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 790                 GOTO(cleanup, rc = -ENOENT);
 791
 792         /* Step 1: Lookup the parent by FID */
 793         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 794                                         &parent_lockh, rec->ur_name,
 795                                         rec->ur_namelen - 1);
 796         if (IS_ERR(dparent))
 797                 GOTO(cleanup, rc = PTR_ERR(dparent));
 798         LASSERT(dparent->d_inode);
 799
 800         cleanup_phase = 1; /* Have parent dentry lock */
 801
 802         /* Step 2: Lookup the child */
 803         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 804         if (IS_ERR(dchild))
 805                 GOTO(cleanup, rc = PTR_ERR(dchild));
 806
 807         cleanup_phase = 2; /* child dentry */
 808
 809         child_inode = dchild->d_inode;
 810         if (child_inode == NULL) {
 811                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
 812                        dparent->d_inode->i_ino, rec->ur_name);
 813                 GOTO(cleanup, rc = -ENOENT);
 814         }
 815
 816         DEBUG_REQ(D_INODE, req, "parent ino %lu, child ino %lu",
 817                   dparent->d_inode->i_ino, child_inode->i_ino);
 818
 819         /* Step 3: Get a lock on the child */
 820         child_res_id.name[0] = child_inode->i_ino;
 821         child_res_id.name[1] = child_inode->i_generation;
 822
 823         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 824                               child_res_id, LDLM_PLAIN, NULL, 0, LCK_EX,
 825                               &flags, ldlm_completion_ast, mds_blocking_ast,
 826                               NULL, &child_lockh);
 827         if (rc != ELDLM_OK)
 828                 GOTO(cleanup, rc);
 829
 830         cleanup_phase = 3; /* child lock */
 831
 832         /* Step 4: Get a lock on the ino to sync with creation WRT inode
 833          * reuse (see bug 2029). */
 834         child_res_id.name[1] = 0;
 835
 836         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 837                               child_res_id, LDLM_PLAIN, NULL, 0, LCK_EX,
 838                               &flags, ldlm_completion_ast, mds_blocking_ast,
 839                               NULL, &child_reuse_lockh);
 840         if (rc != ELDLM_OK)
 841                 GOTO(cleanup, rc);
 842
 843         cleanup_phase = 4; /* child lock */
 844
 845         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
 846
 847         /* ldlm_reply in buf[0] if called via intent */
 848         if (offset)
 849                 offset = 1;
 850
 851         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 852         LASSERT(body != NULL);
 853
 854         /* If this is the last reference to this inode, get the OBD EA
 855          * data first so the client can destroy OST objects.
 856          * we only do the object removal if no open files remain.
 857          * Nobody can get at this name anymore because of the locks so
 858          * we make decisions here as to whether to remove the inode */
 859         if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
 860             mds_open_orphan_count(child_inode) == 0) {
 861                 mds_pack_inode2fid(&body->fid1, child_inode);
 862                 mds_pack_inode2body(body, child_inode);
 863                 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
 864                             child_inode, 1);
 865                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 866                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 867                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 868                 } else {
 869                         log_unlink = 1;
 870                 }
 871         }
 872
 873         /* We have to do these checks ourselves, in case we are making an
 874          * orphan.  The client tells us whether rmdir() or unlink() was called,
 875          * so we need to return appropriate errors (bug 72).
 876          *
 877          * We don't have to check permissions, because vfs_rename (called from
 878          * mds_open_unlink_rename) also calls may_delete. */
 879         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
 880                 if (!S_ISDIR(child_inode->i_mode))
 881                         GOTO(cleanup, rc = -ENOTDIR);
 882         } else {
 883                 if (S_ISDIR(child_inode->i_mode))
 884                         GOTO(cleanup, rc = -EISDIR);
 885         }
 886
 887         if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
 888             mds_open_orphan_count(child_inode) > 0) {
 889                 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
 890                 cleanup_phase = 5; /* transaction */
 891                 GOTO(cleanup, rc);
 892         }
 893
 894         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
 895         switch (child_inode->i_mode & S_IFMT) {
 896         case S_IFDIR:
 897                 /* Drop any lingering child directories before we start our
 898                  * transaction, to avoid doing multiple inode dirty/delete
 899                  * in our compound transaction (bug 1321). */
 900                 shrink_dcache_parent(dchild);
 901                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
 902                                       NULL);
 903                 if (IS_ERR(handle))
 904                         GOTO(cleanup, rc = PTR_ERR(handle));
 905                 cleanup_phase = 5; /* transaction */
 906                 rc = vfs_rmdir(dparent->d_inode, dchild);
 907                 break;
 908         case S_IFREG: {
 909                 handle = fsfilt_start(obd, dparent->d_inode,
 910                                       FSFILT_OP_UNLINK_LOG, NULL);
 911                 if (IS_ERR(handle))
 912                         GOTO(cleanup, rc = PTR_ERR(handle));
 913
 914                 cleanup_phase = 5; /* transaction */
 915                 rc = vfs_unlink(dparent->d_inode, dchild);
 916
 917 #ifdef ENABLE_ORPHANS
 918
 919                 if (!rc && log_unlink)
 920                         if (mds_log_op_unlink(obd, child_inode, req->rq_repmsg,
 921                                               offset + 1) > 0)
 922                                 body->valid |= OBD_MD_FLCOOKIE;
 923 #endif
 924                 break;
 925         }
 926         case S_IFLNK:
 927         case S_IFCHR:
 928         case S_IFBLK:
 929         case S_IFIFO:
 930         case S_IFSOCK:
 931                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
 932                                       NULL);
 933                 if (IS_ERR(handle))
 934                         GOTO(cleanup, rc = PTR_ERR(handle));
 935                 cleanup_phase = 5;
 936                 rc = vfs_unlink(dparent->d_inode, dchild);
 937                 break;
 938         default:
 939                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
 940                        rec->ur_name);
 941                 LBUG();
 942                 GOTO(cleanup, rc = -EINVAL);
 943         }
 944
 945  cleanup:
 946         if (rc == 0) {
 947                 struct iattr iattr;
 948                 int err;
 949
 950                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 951                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 952                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 953
 954                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 955                 if (err)
 956                         CERROR("error on parent setattr: rc = %d\n", err);
 957         }
 958
 959         switch(cleanup_phase) {
 960         case 5:
 961                 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
 962                                         rc, 0);
 963                 if (!rc)
 964                         (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
 965                                            "unlinked", 0, NULL);
 966         case 4: /* child ino-reuse lock */
 967                 if (rc && body != NULL) {
 968                         // Don't unlink the OST objects if the MDS unlink failed
 969                         body->valid = 0;
 970                 }
 971                 if (rc)
 972                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
 973                 else
 974                         ldlm_put_lock_into_req(req, &child_reuse_lockh, LCK_EX);
 975         case 3: /* child lock */
 976                 ldlm_lock_decref(&child_lockh, LCK_EX);
 977         case 2: /* child dentry */
 978                 l_dput(dchild);
 979         case 1: /* parent dentry and lock */
 980                 if (rc) {
 981                         ldlm_lock_decref(&parent_lockh, LCK_PW);
 982                 } else {
 983                         ldlm_put_lock_into_req(req, &parent_lockh, LCK_PW);
 984                 }
 985                 l_dput(dparent);
 986         case 0:
 987                 break;
 988         default:
 989                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 990                 LBUG();
 991         }
 992         req->rq_status = rc;
 993         return 0;
 994 }
 995
 996 static int mds_reint_link(struct mds_update_record *rec, int offset,
 997                           struct ptlrpc_request *req,
 998                           struct lustre_handle *lh)
 999 {
1000         struct obd_device *obd = req->rq_export->exp_obd;
1001         struct dentry *de_src = NULL;
1002         struct dentry *de_tgt_dir = NULL;
1003         struct dentry *dchild = NULL;
1004         struct mds_obd *mds = mds_req2mds(req);
1005         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1006         struct ldlm_res_id src_res_id = { .name = {0} };
1007         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1008         int lock_mode = 0, rc = 0, cleanup_phase = 0;
1009         ENTRY;
1010
1011         LASSERT(offset == 0);
1012
1013         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1014
1015         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1016                 GOTO(cleanup, rc = -ENOENT);
1017
1018         /* Step 1: Lookup the source inode and target directory by FID */
1019         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1020         if (IS_ERR(de_src))
1021                 GOTO(cleanup, rc = PTR_ERR(de_src));
1022
1023         cleanup_phase = 1; /* source dentry */
1024
1025         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1026         if (IS_ERR(de_tgt_dir))
1027                 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1028
1029         cleanup_phase = 2; /* target directory dentry */
1030
1031         CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1032                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1033                de_src->d_inode->i_ino);
1034
1035         /* Step 2: Take the two locks */
1036         lock_mode = LCK_EX;
1037         src_res_id.name[0] = de_src->d_inode->i_ino;
1038         src_res_id.name[1] = de_src->d_inode->i_generation;
1039         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1040         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1041
1042         rc = enqueue_ordered_locks(LCK_EX, obd, &src_res_id, &tgt_dir_res_id,
1043                                    &src_lockh, &tgt_dir_lockh);
1044         if (rc != ELDLM_OK)
1045                 GOTO(cleanup, rc = -EIO);
1046
1047         cleanup_phase = 3; /* locks */
1048
1049         /* Step 3: Lookup the child */
1050         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1051         if (IS_ERR(dchild)) {
1052                 rc = PTR_ERR(dchild);
1053                 if (rc != -EPERM && rc != -EACCES)
1054                         CERROR("child lookup error %d\n", rc);
1055                 GOTO(cleanup, rc);
1056         }
1057
1058         cleanup_phase = 4; /* child dentry */
1059
1060         if (dchild->d_inode) {
1061                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1062                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1063                 rc = -EEXIST;
1064                 GOTO(cleanup, rc);
1065         }
1066
1067         /* Step 4: Do it. */
1068         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1069
1070         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1071         if (IS_ERR(handle)) {
1072                 rc = PTR_ERR(handle);
1073                 GOTO(cleanup, rc);
1074         }
1075
1076         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1077         if (rc && rc != -EPERM && rc != -EACCES)
1078                 CERROR("vfs_link error %d\n", rc);
1079 cleanup:
1080         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1081                                 handle, req, rc, 0);
1082         EXIT;
1083
1084         switch (cleanup_phase) {
1085         case 4: /* child dentry */
1086                 l_dput(dchild);
1087         case 3: /* locks */
1088                 if (rc) {
1089                         ldlm_lock_decref(&src_lockh, lock_mode);
1090                         ldlm_lock_decref(&tgt_dir_lockh, lock_mode);
1091                 } else {
1092                         ldlm_put_lock_into_req(req, &src_lockh, lock_mode);
1093                         ldlm_put_lock_into_req(req, &tgt_dir_lockh, lock_mode);
1094                 }
1095         case 2: /* target dentry */
1096                 l_dput(de_tgt_dir);
1097         case 1: /* source dentry */
1098                 l_dput(de_src);
1099         case 0:
1100                 break;
1101         default:
1102                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1103                 LBUG();
1104         }
1105         req->rq_status = rc;
1106         return 0;
1107 }
1108
1109 /*
1110  * add a hard link in the PENDING directory, only used by rename()
1111  */
1112 static int mds_add_link_orphan(struct mds_update_record *rec,
1113                                struct obd_device *obd,
1114                                struct dentry *dentry)
1115 {
1116         struct mds_obd *mds = &obd->u.mds;
1117         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1118         struct dentry *pending_child;
1119         char fidname[LL_FID_NAMELEN];
1120         int fidlen = 0, rc;
1121         ENTRY;
1122
1123         LASSERT(dentry->d_inode);
1124         LASSERT(!mds_inode_is_orphan(dentry->d_inode));
1125
1126         down(&pending_dir->i_sem);
1127         fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
1128                             dentry->d_inode->i_generation);
1129
1130         CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
1131                mds_open_orphan_count(dentry->d_inode),
1132                rec->ur_name, fidname);
1133
1134         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1135         if (IS_ERR(pending_child))
1136                 GOTO(out_lock, rc = PTR_ERR(pending_child));
1137
1138         if (pending_child->d_inode != NULL) {
1139                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1140                 LASSERT(pending_child->d_inode == dentry->d_inode);
1141                 GOTO(out_dput, rc = 0);
1142         }
1143
1144         lock_kernel();
1145         rc = vfs_link(dentry, pending_dir, pending_child);
1146         unlock_kernel();
1147         if (rc)
1148                 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
1149                        rec->ur_name, rc);
1150         else
1151                 mds_inode_set_orphan(dentry->d_inode);
1152 out_dput:
1153         l_dput(pending_child);
1154 out_lock:
1155         up(&pending_dir->i_sem);
1156         RETURN(rc);
1157 }
1158
1159 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1160                             struct ptlrpc_request *req,
1161                             struct lustre_handle *lockh)
1162 {
1163         struct obd_device *obd = req->rq_export->exp_obd;
1164         struct dentry *de_srcdir = NULL;
1165         struct dentry *de_tgtdir = NULL;
1166         struct dentry *de_old = NULL;
1167         struct dentry *de_new = NULL;
1168         struct mds_obd *mds = mds_req2mds(req);
1169         struct lustre_handle dlm_handles[4];
1170         struct ldlm_res_id p1_res_id = { .name = {0} };
1171         struct ldlm_res_id p2_res_id = { .name = {0} };
1172         struct ldlm_res_id c1_res_id = { .name = {0} };
1173         struct ldlm_res_id c2_res_id = { .name = {0} };
1174         struct mds_body *body = NULL;
1175         int rc = 0, lock_count = 3, flags = LDLM_FL_LOCAL_ONLY;
1176         int cleanup_phase = 0;
1177         void *handle = NULL;
1178         ENTRY;
1179
1180         LASSERT(offset == 0);
1181
1182         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1183
1184         de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1185         if (IS_ERR(de_srcdir))
1186                 GOTO(cleanup, rc = PTR_ERR(de_srcdir));
1187
1188         cleanup_phase = 1; /* source directory dentry */
1189
1190         de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1191         if (IS_ERR(de_tgtdir))
1192                 GOTO(cleanup, rc = PTR_ERR(de_tgtdir));
1193
1194         cleanup_phase = 2; /* target directory dentry */
1195
1196         /* The idea here is that we need to get four locks in the end:
1197          * one on each parent directory, one on each child.  We need to take
1198          * these locks in some kind of order (to avoid deadlocks), and the order
1199          * I selected is "increasing resource number" order.  We need to take
1200          * the locks on the parent directories, however, before we can lookup
1201          * the children.  Thus the following plan:
1202          *
1203          * 1. Take locks on the parent(s), in order
1204          * 2. Lookup the children
1205          * 3. Take locks on the children, in order
1206          * 4. Execute the rename
1207          */
1208
1209         /* Step 1: Take locks on the parent(s), in order */
1210         p1_res_id.name[0] = de_srcdir->d_inode->i_ino;
1211         p1_res_id.name[1] = de_srcdir->d_inode->i_generation;
1212
1213         p2_res_id.name[0] = de_tgtdir->d_inode->i_ino;
1214         p2_res_id.name[1] = de_tgtdir->d_inode->i_generation;
1215
1216         rc = enqueue_ordered_locks(LCK_EX, obd, &p1_res_id, &p2_res_id,
1217                                    &(dlm_handles[0]), &(dlm_handles[1]));
1218         if (rc != ELDLM_OK)
1219                 GOTO(cleanup, rc);
1220
1221         cleanup_phase = 3; /* parent locks */
1222
1223         /* Step 2: Lookup the children */
1224         de_old = ll_lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen-1);
1225         if (IS_ERR(de_old)) {
1226                 CERROR("old child lookup error (%*s): %ld\n",
1227                        rec->ur_namelen - 1, rec->ur_name, PTR_ERR(de_old));
1228                 GOTO(cleanup, rc = PTR_ERR(de_old));
1229         }
1230
1231         cleanup_phase = 4; /* original name dentry */
1232
1233         if (de_old->d_inode == NULL)
1234                 GOTO(cleanup, rc = -ENOENT);
1235
1236         /* sanity check for src inode */
1237         if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1238             de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
1239                 GOTO(cleanup, rc = -EINVAL);
1240
1241         de_new = ll_lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
1242         if (IS_ERR(de_new)) {
1243                 CERROR("new child lookup error (%*s): %ld\n",
1244                        rec->ur_tgtlen - 1, rec->ur_tgt, PTR_ERR(de_new));
1245                 GOTO(cleanup, rc = PTR_ERR(de_new));
1246         }
1247
1248         cleanup_phase = 5; /* target dentry */
1249
1250         /* sanity check for dest inode */
1251         if (de_new->d_inode &&
1252             (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1253             de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
1254                 GOTO(cleanup, rc = -EINVAL);
1255
1256         /* Step 3: Take locks on the children */
1257         c1_res_id.name[0] = de_old->d_inode->i_ino;
1258         c1_res_id.name[1] = de_old->d_inode->i_generation;
1259         if (de_new->d_inode == NULL) {
1260                 flags = LDLM_FL_LOCAL_ONLY;
1261                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
1262                                       c1_res_id, LDLM_PLAIN, NULL, 0, LCK_EX,
1263                                       &flags, ldlm_completion_ast,
1264                                       mds_blocking_ast, NULL,
1265                                       &(dlm_handles[2]));
1266                 lock_count = 3;
1267         } else {
1268                 c2_res_id.name[0] = de_new->d_inode->i_ino;
1269                 c2_res_id.name[1] = de_new->d_inode->i_generation;
1270                 rc = enqueue_ordered_locks(LCK_EX, obd, &c1_res_id, &c2_res_id,
1271                                            &(dlm_handles[2]),
1272                                            &(dlm_handles[3]));
1273                 lock_count = 4;
1274         }
1275         if (rc != ELDLM_OK)
1276                 GOTO(cleanup, rc);
1277
1278         cleanup_phase = 6; /* child locks */
1279
1280         /* if we are about to remove the target at first, pass the EA of
1281          * that inode to client to perform and cleanup on OST */
1282         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1283         LASSERT(body != NULL);
1284
1285         if (de_new->d_inode &&
1286             S_ISREG(de_new->d_inode->i_mode) &&
1287             de_new->d_inode->i_nlink == 1 &&
1288             mds_open_orphan_count(de_new->d_inode) == 0) {
1289                 mds_pack_inode2fid(&body->fid1, de_new->d_inode);
1290                 mds_pack_inode2body(body, de_new->d_inode);
1291                 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
1292                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1293                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1294                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1295                 } else {
1296                         /* XXX need log unlink? */
1297                 }
1298         }
1299
1300         /* Step 4: Execute the rename */
1301         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,de_srcdir->d_inode->i_sb);
1302
1303         handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
1304         if (IS_ERR(handle))
1305                 GOTO(cleanup, rc = PTR_ERR(handle));
1306
1307         /* FIXME need adjust the journal block count? */
1308         /* if the target should be moved to PENDING, we at first increase the
1309          * link and later vfs_rename() will decrease the link count again */
1310         if (de_new->d_inode &&
1311             S_ISREG(de_new->d_inode->i_mode) &&
1312             de_new->d_inode->i_nlink == 1 &&
1313             mds_open_orphan_count(de_new->d_inode) > 0) {
1314                 rc = mds_add_link_orphan(rec, obd, de_new);
1315                 if (rc)
1316                         GOTO(cleanup, rc);
1317         }
1318
1319         lock_kernel();
1320         de_old->d_fsdata = req;
1321         de_new->d_fsdata = req;
1322         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
1323         unlock_kernel();
1324
1325         EXIT;
1326 cleanup:
1327         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
1328                                 handle, req, rc, 0);
1329         switch (cleanup_phase) {
1330         case 6: /* child locks */
1331                 if (rc) {
1332                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
1333                         if (lock_count == 4)
1334                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
1335                 } else {
1336                         ldlm_put_lock_into_req(req, &(dlm_handles[2]), LCK_EX);
1337                         if (lock_count == 4)
1338                                 ldlm_put_lock_into_req(req,
1339                                                 &(dlm_handles[3]), LCK_EX);
1340                 }
1341         case 5: /* target dentry */
1342                 l_dput(de_new);
1343         case 4: /* source dentry */
1344                 l_dput(de_old);
1345         case 3: /* parent locks */
1346                 if (rc) {
1347                         ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
1348                         ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
1349                 } else {
1350                         ldlm_put_lock_into_req(req, &(dlm_handles[0]), LCK_EX);
1351                         ldlm_put_lock_into_req(req, &(dlm_handles[1]), LCK_EX);
1352                 }
1353         case 2: /* target directory dentry */
1354                 l_dput(de_tgtdir);
1355         case 1: /* source directry dentry */
1356                 l_dput(de_srcdir);
1357         case 0:
1358                 break;
1359         default:
1360                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1361                 LBUG();
1362         }
1363         req->rq_status = rc;
1364         return 0;
1365 }
1366
1367 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
1368                            struct ptlrpc_request *, struct lustre_handle *);
1369
1370 static mds_reinter reinters[REINT_MAX + 1] = {
1371         [REINT_SETATTR] mds_reint_setattr,
1372         [REINT_CREATE] mds_reint_create,
1373         [REINT_UNLINK] mds_reint_unlink,
1374         [REINT_LINK] mds_reint_link,
1375         [REINT_RENAME] mds_reint_rename,
1376         [REINT_OPEN] mds_open
1377 };
1378
1379 int mds_reint_rec(struct mds_update_record *rec, int offset,
1380                   struct ptlrpc_request *req, struct lustre_handle *lockh)
1381 {
1382         struct obd_device *obd = req->rq_export->exp_obd;
1383         struct obd_run_ctxt saved;
1384         int rc;
1385         ENTRY;
1386
1387         /* checked by unpacker */
1388         LASSERT(rec->ur_opcode <= REINT_MAX &&
1389                 reinters[rec->ur_opcode] != NULL);
1390
1391         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1392         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
1393         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1394
1395         RETURN(rc);
1396 }