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) 2002-2005 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         struct llog_ctxt *ctxt;
  66         int rc;
  67
  68         obd_transno_commit_cb(obd, transno, error);
  69
  70         CDEBUG(D_HA, "cancelling %d cookies\n",
  71                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  72
  73         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  74                           mlcd->mlcd_eadatalen);
  75         if (rc < 0) {
  76                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  77                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  78                        rc);
  79         } else {
  80                 ///* XXX 0 normally, SENDNOW for debug */);
  81                 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
  82                 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
  83                                                 sizeof(*mlcd->mlcd_cookies),
  84                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  85                 if (rc)
  86                         CERROR("error cancelling %d log cookies: rc %d\n",
  87                                (int)(mlcd->mlcd_cookielen /
  88                                      sizeof(*mlcd->mlcd_cookies)), rc);
  89         }
  90
  91         OBD_FREE(mlcd, mlcd->mlcd_size);
  92 }
  93
  94 /* Assumes caller has already pushed us into the kernel context. */
  95 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
  96                        struct ptlrpc_request *req, int rc, __u32 op_data)
  97 {
  98         struct mds_export_data *med = &req->rq_export->exp_mds_data;
  99         struct mds_client_data *mcd = med->med_mcd;
 100         struct obd_device *obd = req->rq_export->exp_obd;
 101         int err;
 102         __u64 transno;
 103         loff_t off;
 104         int log_pri = D_HA;
 105         ENTRY;
 106
 107         /* if the export has already been failed, we have no last_rcvd slot */
 108         if (req->rq_export->exp_failed) {
 109                 CWARN("committing transaction for disconnected client %s\n",
 110                       req->rq_export->exp_client_uuid.uuid);
 111                 if (handle)
 112                         GOTO(commit, rc);
 113                 RETURN(rc);
 114         }
 115
 116         if (IS_ERR(handle))
 117                 RETURN(rc);
 118
 119         if (handle == NULL) {
 120                 /* if we're starting our own xaction, use our own inode */
 121                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 122                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 123                 if (IS_ERR(handle)) {
 124                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 125                         RETURN(PTR_ERR(handle));
 126                 }
 127         }
 128
 129         off = med->med_off;
 130
 131         transno = req->rq_reqmsg->transno;
 132         if (rc != 0) {
 133                 LASSERT(transno == 0);
 134         } else if (transno == 0) {
 135                 spin_lock(&mds->mds_transno_lock);
 136                 transno = ++mds->mds_last_transno;
 137                 spin_unlock(&mds->mds_transno_lock);
 138         } else {
 139                 spin_lock(&mds->mds_transno_lock);
 140                 if (transno > mds->mds_last_transno)
 141                         mds->mds_last_transno = transno;
 142                 spin_unlock(&mds->mds_transno_lock);
 143         }
 144         req->rq_repmsg->transno = req->rq_transno = transno;
 145         mcd->mcd_last_transno = cpu_to_le64(transno);
 146         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
 147         mcd->mcd_last_result = cpu_to_le32(rc);
 148         mcd->mcd_last_data = cpu_to_le32(op_data);
 149
 150         fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
 151                               mds_commit_cb, NULL);
 152         err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
 153                                   &off, 0);
 154
 155         if (err) {
 156                 log_pri = D_ERROR;
 157                 if (rc == 0)
 158                         rc = err;
 159         }
 160
 161         DEBUG_REQ(log_pri, req,
 162                   "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
 163                   transno, rc, mcd->mcd_uuid, med->med_idx, err);
 164
 165         err = mds_lov_write_objids(obd);
 166         if (err) {
 167                 log_pri = D_ERROR;
 168                 if (rc == 0)
 169                         rc = err;
 170         }
 171         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 172
 173 commit:
 174         err = fsfilt_commit(obd, inode, handle, 0);
 175         if (err) {
 176                 CERROR("error committing transaction: %d\n", err);
 177                 if (!rc)
 178                         rc = err;
 179         }
 180
 181         RETURN(rc);
 182 }
 183
 184 /* this gives the same functionality as the code between
 185  * sys_chmod and inode_setattr
 186  * chown_common and inode_setattr
 187  * utimes and inode_setattr
 188  */
 189 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 190 {
 191         time_t now = CURRENT_SECONDS;
 192         struct iattr *attr = &rec->ur_iattr;
 193         unsigned int ia_valid = attr->ia_valid;
 194         int error;
 195         ENTRY;
 196
 197         /* only fix up attrs if the client VFS didn't already */
 198         if (!(ia_valid & ATTR_RAW))
 199                 RETURN(0);
 200
 201         if (!(ia_valid & ATTR_CTIME_SET))
 202                 LTIME_S(attr->ia_ctime) = now;
 203         if (!(ia_valid & ATTR_ATIME_SET))
 204                 LTIME_S(attr->ia_atime) = now;
 205         if (!(ia_valid & ATTR_MTIME_SET))
 206                 LTIME_S(attr->ia_mtime) = now;
 207
 208         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 209                 RETURN(-EPERM);
 210
 211         /* times */
 212         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 213                 if (rec->ur_fsuid != inode->i_uid &&
 214                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 215                         RETURN(error);
 216         }
 217
 218         if (ia_valid & ATTR_SIZE) {
 219                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 220                         RETURN(error);
 221         }
 222
 223         if (ia_valid & ATTR_UID) {
 224                 /* chown */
 225                 error = -EPERM;
 226                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 227                         RETURN(-EPERM);
 228                 if (attr->ia_uid == (uid_t) -1)
 229                         attr->ia_uid = inode->i_uid;
 230                 if (attr->ia_gid == (gid_t) -1)
 231                         attr->ia_gid = inode->i_gid;
 232                 attr->ia_mode = inode->i_mode;
 233                 /*
 234                  * If the user or group of a non-directory has been
 235                  * changed by a non-root user, remove the setuid bit.
 236                  * 19981026 David C Niemi <niemi@tux.org>
 237                  *
 238                  * Changed this to apply to all users, including root,
 239                  * to avoid some races. This is the behavior we had in
 240                  * 2.0. The check for non-root was definitely wrong
 241                  * for 2.2 anyway, as it should have been using
 242                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 243                  */
 244                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 245                     !S_ISDIR(inode->i_mode)) {
 246                         attr->ia_mode &= ~S_ISUID;
 247                         attr->ia_valid |= ATTR_MODE;
 248                 }
 249                 /*
 250                  * Likewise, if the user or group of a non-directory
 251                  * has been changed by a non-root user, remove the
 252                  * setgid bit UNLESS there is no group execute bit
 253                  * (this would be a file marked for mandatory
 254                  * locking).  19981026 David C Niemi <niemi@tux.org>
 255                  *
 256                  * Removed the fsuid check (see the comment above) --
 257                  * 19990830 SD.
 258                  */
 259                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 260                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 261                         attr->ia_mode &= ~S_ISGID;
 262                         attr->ia_valid |= ATTR_MODE;
 263                 }
 264         } else if (ia_valid & ATTR_MODE) {
 265                 int mode = attr->ia_mode;
 266                 /* chmod */
 267                 if (attr->ia_mode == (umode_t)-1)
 268                         mode = inode->i_mode;
 269                 attr->ia_mode =
 270                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 271         }
 272         RETURN(0);
 273 }
 274
 275 void mds_steal_ack_locks(struct ptlrpc_request *req)
 276 {
 277         struct obd_export         *exp = req->rq_export;
 278         struct list_head          *tmp;
 279         struct ptlrpc_reply_state *oldrep;
 280         struct ptlrpc_service     *svc;
 281         unsigned long              flags;
 282         char                       str[PTL_NALFMT_SIZE];
 283         int                        i;
 284
 285         /* CAVEAT EMPTOR: spinlock order */
 286         spin_lock_irqsave (&exp->exp_lock, flags);
 287         list_for_each (tmp, &exp->exp_outstanding_replies) {
 288                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 289
 290                 if (oldrep->rs_xid != req->rq_xid)
 291                         continue;
 292
 293                 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
 294                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 295                                 "new %d old %d\n", req->rq_xid,
 296                                 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
 297
 298                 svc = oldrep->rs_srv_ni->sni_service;
 299                 spin_lock (&svc->srv_lock);
 300
 301                 list_del_init (&oldrep->rs_exp_list);
 302
 303                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 304                       " o%d NID %s\n",
 305                       oldrep->rs_nlocks, oldrep,
 306                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
 307                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 308
 309                 for (i = 0; i < oldrep->rs_nlocks; i++)
 310                         ptlrpc_save_lock(req,
 311                                          &oldrep->rs_locks[i],
 312                                          oldrep->rs_modes[i]);
 313                 oldrep->rs_nlocks = 0;
 314
 315                 DEBUG_REQ(D_HA, req, "stole locks for");
 316                 ptlrpc_schedule_difficult_reply (oldrep);
 317
 318                 spin_unlock (&svc->srv_lock);
 319                 spin_unlock_irqrestore (&exp->exp_lock, flags);
 320                 return;
 321         }
 322         spin_unlock_irqrestore (&exp->exp_lock, flags);
 323 }
 324
 325 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
 326 {
 327         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 328                   mcd->mcd_last_transno, mcd->mcd_last_result);
 329         req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
 330         req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
 331
 332         mds_steal_ack_locks(req);
 333 }
 334
 335 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 336                                       int offset, struct ptlrpc_request *req)
 337 {
 338         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 339         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 340         struct dentry *de;
 341         struct mds_body *body;
 342
 343         mds_req_from_mcd(req, med->med_mcd);
 344
 345         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 346         if (IS_ERR(de)) {
 347                 LASSERT(PTR_ERR(de) == req->rq_status);
 348                 return;
 349         }
 350
 351         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 352         mds_pack_inode2fid(&body->fid1, de->d_inode);
 353         mds_pack_inode2body(body, de->d_inode);
 354
 355         /* Don't return OST-specific attributes if we didn't just set them */
 356         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 357                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 358         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 359                 body->valid |= OBD_MD_FLMTIME;
 360         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 361                 body->valid |= OBD_MD_FLATIME;
 362
 363         l_dput(de);
 364 }
 365
 366 int mds_osc_setattr_async(struct obd_device *obd, struct inode *inode,
 367                           struct lov_mds_md *lmm, int lmm_size,
 368                           struct llog_cookie *logcookies)
 369 {
 370         struct mds_obd *mds = &obd->u.mds;
 371         struct lov_stripe_md *lsm = NULL;
 372         struct obd_trans_info oti = { 0 };
 373         struct obdo *oa = NULL;
 374         int  cleanup_phase = 0, rc = 0;
 375         ENTRY;
 376
 377         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
 378                 GOTO(cleanup, rc);
 379
 380         /* first get memory EA */
 381         oa = obdo_alloc();
 382         if (!oa)
 383                 RETURN(-ENOMEM);
 384
 385         LASSERT(lmm);
 386
 387         cleanup_phase = 1;
 388         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
 389         if (rc < 0) {
 390                 CERROR("Error unpack md %p\n", lmm);
 391                 GOTO(cleanup, rc);
 392         }
 393
 394         cleanup_phase = 2;
 395         /* then fill oa */
 396         oa->o_id = lsm->lsm_object_id;
 397         oa->o_uid = inode->i_uid;
 398         oa->o_gid = inode->i_gid;
 399         oa->o_valid = OBD_MD_FLID | OBD_MD_FLUID | OBD_MD_FLGID;
 400         if (logcookies) {
 401                 oa->o_valid |= OBD_MD_FLCOOKIE;
 402                 oti.oti_logcookies = logcookies;
 403         }
 404
 405         /* do setattr from mds to ost asynchronously */
 406         rc = obd_setattr_async(mds->mds_osc_exp, oa, lsm, &oti);
 407         if (rc)
 408                 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64" on ost error "
 409                        "%d\n", lsm->lsm_object_id, rc);
 410 cleanup:
 411         switch(cleanup_phase) {
 412         case 2:
 413                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 414         case 1:
 415                 obdo_free(oa);
 416         case 0:
 417                 if (logcookies)
 418                         OBD_FREE(logcookies, mds->mds_max_cookiesize);
 419         }
 420
 421         RETURN(rc);
 422 }
 423
 424 /* In the raw-setattr case, we lock the child inode.
 425  * In the write-back case or if being called from open, the client holds a lock
 426  * already.
 427  *
 428  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 429 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 430                              struct ptlrpc_request *req,
 431                              struct lustre_handle *lh)
 432 {
 433         struct mds_obd *mds = mds_req2mds(req);
 434         struct obd_device *obd = req->rq_export->exp_obd;
 435         struct mds_body *body;
 436         struct dentry *de;
 437         struct inode *inode = NULL;
 438         struct lustre_handle lockh;
 439         void *handle = NULL;
 440         struct mds_logcancel_data *mlcd = NULL;
 441         struct lov_mds_md *lmm = NULL;
 442         struct llog_cookie *logcookies = NULL;
 443         int lmm_size = 0, need_lock = 1;
 444         int rc = 0, cleanup_phase = 0, err, locked = 0;
 445         uid_t child_uid = 0;
 446         gid_t child_gid = 0;
 447         ENTRY;
 448
 449         LASSERT(offset == 0);
 450
 451         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 452                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 453
 454         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 455
 456         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
 457             (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
 458                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 459                 if (IS_ERR(de))
 460                         GOTO(cleanup, rc = PTR_ERR(de));
 461                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
 462                         GOTO(cleanup, rc = -EROFS);
 463         } else {
 464                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
 465                                            &lockh, NULL, 0);
 466                 if (IS_ERR(de))
 467                         GOTO(cleanup, rc = PTR_ERR(de));
 468                 locked = 1;
 469         }
 470
 471         cleanup_phase = 1;
 472         inode = de->d_inode;
 473         LASSERT(inode);
 474
 475         /* save uid/gid for quota acq/rel */
 476         child_uid = inode->i_uid;
 477         child_gid = inode->i_gid;
 478
 479         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 480             rec->ur_eadata != NULL) {
 481                 down(&inode->i_sem);
 482                 need_lock = 0;
 483         }
 484
 485         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 486
 487         /* start a log jounal handle if needed*/
 488         if (S_ISREG(inode->i_mode) &&
 489             rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 490                 lmm_size = mds->mds_max_mdsize;
 491                 OBD_ALLOC(lmm, lmm_size);
 492                 if (lmm == NULL)
 493                         GOTO(cleanup, rc = -ENOMEM);
 494
 495                 cleanup_phase = 2;
 496                 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock);
 497                 if (rc < 0)
 498                         GOTO(cleanup, rc);
 499
 500                 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
 501                                           le32_to_cpu(lmm->lmm_stripe_count));
 502         } else {
 503                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 504         }
 505         if (IS_ERR(handle))
 506                 GOTO(cleanup, rc = PTR_ERR(handle));
 507
 508         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 509                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 510                        LTIME_S(rec->ur_iattr.ia_mtime),
 511                        LTIME_S(rec->ur_iattr.ia_ctime));
 512         rc = mds_fix_attr(inode, rec);
 513         if (rc)
 514                 GOTO(cleanup, rc);
 515
 516         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG)    /* ioctl */
 517                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 518                                       (long)&rec->ur_iattr.ia_attr_flags);
 519         else {                                            /* setattr */
 520                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 521                 /* journal chown/chgrp in llog, just like unlink */
 522                 if (rc == 0 && S_ISREG(inode->i_mode) &&
 523                     rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID) && lmm_size){
 524                         OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 525                         if (logcookies == NULL)
 526                                 GOTO(cleanup, rc = -ENOMEM);
 527                         if (mds_log_op_setattr(obd, inode, lmm, lmm_size,
 528                                                logcookies,
 529                                                mds->mds_max_cookiesize) <= 0) {
 530                                 OBD_FREE(logcookies, mds->mds_max_cookiesize);
 531                                 logcookies = NULL;
 532                         }
 533                 }
 534         }
 535
 536         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 537             rec->ur_eadata != NULL) {
 538                 struct lov_stripe_md *lsm = NULL;
 539                 struct lov_user_md *lum = NULL;
 540
 541                 rc = ll_permission(inode, MAY_WRITE, NULL);
 542                 if (rc < 0)
 543                         GOTO(cleanup, rc);
 544
 545                 lum = rec->ur_eadata;
 546                 /* if lmm_stripe_size is -1,  then delete the stripe
 547                         info from the dir */
 548                 if (S_ISDIR(inode->i_mode) &&
 549                     lum->lmm_stripe_size == (typeof(lum->lmm_stripe_size))(-1)){
 550                         rc = fsfilt_set_md(obd, inode, handle, NULL, 0);
 551                         if (rc)
 552                                 GOTO(cleanup, rc);
 553                 } else {
 554                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 555                                            mds->mds_osc_exp, 0,
 556                                            &lsm, rec->ur_eadata);
 557                         if (rc)
 558                                 GOTO(cleanup, rc);
 559
 560                         obd_free_memmd(mds->mds_osc_exp, &lsm);
 561
 562                         rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 563                                            rec->ur_eadatalen);
 564                         if (rc)
 565                                 GOTO(cleanup, rc);
 566                 }
 567         }
 568
 569         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 570         mds_pack_inode2fid(&body->fid1, inode);
 571         mds_pack_inode2body(body, inode);
 572
 573         /* Don't return OST-specific attributes if we didn't just set them */
 574         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 575                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 576         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 577                 body->valid |= OBD_MD_FLMTIME;
 578         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 579                 body->valid |= OBD_MD_FLATIME;
 580
 581         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 582                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 583                           rec->ur_eadatalen);
 584                 if (mlcd) {
 585                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 586                                 rec->ur_eadatalen;
 587                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 588                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 589                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 590                                 mlcd->mlcd_cookielen;
 591                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 592                                mlcd->mlcd_cookielen);
 593                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 594                                mlcd->mlcd_eadatalen);
 595                 } else {
 596                         CERROR("unable to allocate log cancel data\n");
 597                 }
 598         }
 599         EXIT;
 600  cleanup:
 601         if (mlcd != NULL)
 602                 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 603                                       mds_cancel_cookies_cb, mlcd);
 604         err = mds_finish_transno(mds, inode, handle, req, rc, 0);
 605         /* do mds to ost setattr if needed */
 606         if (!rc && !err && lmm_size)
 607                 mds_osc_setattr_async(obd, inode, lmm, lmm_size, logcookies);
 608
 609         switch (cleanup_phase) {
 610         case 2:
 611                 OBD_FREE(lmm, mds->mds_max_mdsize);
 612         case 1:
 613                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 614                     rec->ur_eadata != NULL)
 615                         up(&inode->i_sem);
 616                 l_dput(de);
 617                 if (locked) {
 618                         if (rc) {
 619                                 ldlm_lock_decref(&lockh, LCK_PW);
 620                         } else {
 621                                 ptlrpc_save_lock (req, &lockh, LCK_PW);
 622                         }
 623                 }
 624         case 0:
 625                 break;
 626         default:
 627                 LBUG();
 628         }
 629         if (err && !rc)
 630                 rc = err;
 631
 632         req->rq_status = rc;
 633
 634         /* trigger dqrel/dqacq for original owner and new owner */
 635         if (rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 636                 mds_adjust_qunit(obd, rec->ur_iattr.ia_uid,
 637                                  rec->ur_iattr.ia_gid, 0, 0, rc);
 638                 mds_adjust_qunit(obd, child_uid, child_gid, 0, 0, rc);
 639         }
 640         return 0;
 641 }
 642
 643 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 644                                      struct ptlrpc_request *req)
 645 {
 646         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 647         struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
 648         struct dentry *parent, *child;
 649         struct mds_body *body;
 650
 651         mds_req_from_mcd(req, med->med_mcd);
 652
 653         if (req->rq_status)
 654                 return;
 655
 656         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 657         LASSERT(!IS_ERR(parent));
 658         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 659         LASSERT(!IS_ERR(child));
 660         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 661         mds_pack_inode2fid(&body->fid1, child->d_inode);
 662         mds_pack_inode2body(body, child->d_inode);
 663         l_dput(parent);
 664         l_dput(child);
 665 }
 666
 667 static int mds_reint_create(struct mds_update_record *rec, int offset,
 668                             struct ptlrpc_request *req,
 669                             struct lustre_handle *lh)
 670 {
 671         struct dentry *dparent = NULL;
 672         struct mds_obd *mds = mds_req2mds(req);
 673         struct obd_device *obd = req->rq_export->exp_obd;
 674         struct dentry *dchild = NULL;
 675         struct inode *dir = NULL;
 676         void *handle = NULL;
 677         struct lustre_handle lockh;
 678         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 679         int created = 0;
 680         uid_t parent_uid = 0;
 681         gid_t parent_gid = 0;
 682         struct dentry_params dp;
 683         ENTRY;
 684
 685         LASSERT(offset == 0);
 686         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 687
 688         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 689                   rec->ur_fid1->id, rec->ur_fid1->generation,
 690                   rec->ur_name, rec->ur_mode);
 691
 692         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 693
 694         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 695                 GOTO(cleanup, rc = -ESTALE);
 696
 697         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
 698                                         rec->ur_name, rec->ur_namelen - 1);
 699         if (IS_ERR(dparent)) {
 700                 rc = PTR_ERR(dparent);
 701                 if (rc != -ENOENT)
 702                         CERROR("parent lookup error %d\n", rc);
 703                 GOTO(cleanup, rc);
 704         }
 705         cleanup_phase = 1; /* locked parent dentry */
 706         dir = dparent->d_inode;
 707         LASSERT(dir);
 708
 709         ldlm_lock_dump_handle(D_OTHER, &lockh);
 710
 711         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 712         if (IS_ERR(dchild)) {
 713                 rc = PTR_ERR(dchild);
 714                 CERROR("child lookup error %d\n", rc);
 715                 GOTO(cleanup, rc);
 716         }
 717
 718         cleanup_phase = 2; /* child dentry */
 719
 720         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 721
 722         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
 723                 if (dchild->d_inode)
 724                         GOTO(cleanup, rc = -EEXIST);
 725                 GOTO(cleanup, rc = -EROFS);
 726         }
 727
 728         if (dir->i_mode & S_ISGID) {
 729                 if (S_ISDIR(rec->ur_mode))
 730                         rec->ur_mode |= S_ISGID;
 731         }
 732
 733         dchild->d_fsdata = (void *)&dp;
 734         dp.p_inum = (unsigned long)rec->ur_fid2->id;
 735         dp.p_ptr = req;
 736
 737         switch (type) {
 738         case S_IFREG:{
 739                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 740                 if (IS_ERR(handle))
 741                         GOTO(cleanup, rc = PTR_ERR(handle));
 742                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 743                 EXIT;
 744                 break;
 745         }
 746         case S_IFDIR:{
 747                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 748                 if (IS_ERR(handle))
 749                         GOTO(cleanup, rc = PTR_ERR(handle));
 750                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 751                 EXIT;
 752                 break;
 753         }
 754         case S_IFLNK:{
 755                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 756                 if (IS_ERR(handle))
 757                         GOTO(cleanup, rc = PTR_ERR(handle));
 758                 if (rec->ur_tgt == NULL)        /* no target supplied */
 759                         rc = -EINVAL;           /* -EPROTO? */
 760                 else
 761                         rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
 762                 EXIT;
 763                 break;
 764         }
 765         case S_IFCHR:
 766         case S_IFBLK:
 767         case S_IFIFO:
 768         case S_IFSOCK:{
 769                 int rdev = rec->ur_rdev;
 770                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 771                 if (IS_ERR(handle))
 772                         GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
 773                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 774                 EXIT;
 775                 break;
 776         }
 777         default:
 778                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 779                 dchild->d_fsdata = NULL;
 780                 GOTO(cleanup, rc = -EINVAL);
 781         }
 782
 783         /* In case we stored the desired inum in here, we want to clean up. */
 784         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 785                 dchild->d_fsdata = NULL;
 786
 787         if (rc) {
 788                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 789                 GOTO(cleanup, rc);
 790         } else {
 791                 struct iattr iattr;
 792                 struct inode *inode = dchild->d_inode;
 793                 struct mds_body *body;
 794
 795                 created = 1;
 796                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 797                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 798                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 799                 iattr.ia_uid = rec->ur_fsuid;
 800                 if (dir->i_mode & S_ISGID)
 801                         iattr.ia_gid = dir->i_gid;
 802                 else
 803                         iattr.ia_gid = rec->ur_fsgid;
 804                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 805                         ATTR_MTIME | ATTR_CTIME;
 806
 807                 if (rec->ur_fid2->id) {
 808                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 809                         inode->i_generation = rec->ur_fid2->generation;
 810                         /* Dirtied and committed by the upcoming setattr. */
 811                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 812                                inode->i_ino, inode->i_generation);
 813                 } else {
 814                         struct lustre_handle child_ino_lockh;
 815
 816                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 817                                inode->i_ino, inode->i_generation);
 818
 819                         /* The inode we were allocated may have just been freed
 820                          * by an unlink operation.  We take this lock to
 821                          * synchronize against the matching reply-ack-lock taken
 822                          * in unlink, to avoid replay problems if this reply
 823                          * makes it out to the client but the unlink's does not.
 824                          * See bug 2029 for more detail.*/
 825                         rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
 826                         if (rc != ELDLM_OK) {
 827                                 CERROR("error locking for unlink/create sync: "
 828                                        "%d\n", rc);
 829                         } else {
 830                                 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
 831                         }
 832                 }
 833
 834                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 835                 if (rc)
 836                         CERROR("error on child setattr: rc = %d\n", rc);
 837
 838                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 839                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 840                 if (rc)
 841                         CERROR("error on parent setattr: rc = %d\n", rc);
 842
 843                 if (S_ISDIR(inode->i_mode)) {
 844                         struct lov_mds_md lmm;
 845                         int lmm_size = sizeof(lmm);
 846                         rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1);
 847                         if (rc > 0) {
 848                                 down(&inode->i_sem);
 849                                 rc = fsfilt_set_md(obd, inode, handle,
 850                                                    &lmm, lmm_size);
 851                                 up(&inode->i_sem);
 852                         }
 853                         if (rc)
 854                                 CERROR("error on copy stripe info: rc = %d\n",
 855                                         rc);
 856                 }
 857
 858                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
 859                 mds_pack_inode2fid(&body->fid1, inode);
 860                 mds_pack_inode2body(body, inode);
 861         }
 862         EXIT;
 863
 864 cleanup:
 865         err = mds_finish_transno(mds, dir, handle, req, rc, 0);
 866
 867         if (rc && created) {
 868                 /* Destroy the file we just created.  This should not need
 869                  * extra journal credits, as we have already modified all of
 870                  * the blocks needed in order to create the file in the first
 871                  * place.
 872                  */
 873                 switch (type) {
 874                 case S_IFDIR:
 875                         err = vfs_rmdir(dir, dchild);
 876                         if (err)
 877                                 CERROR("rmdir in error path: %d\n", err);
 878                         break;
 879                 default:
 880                         err = vfs_unlink(dir, dchild);
 881                         if (err)
 882                                 CERROR("unlink in error path: %d\n", err);
 883                         break;
 884                 }
 885         } else if (created) {
 886                 /* save uid/gid of create inode and parent */
 887                 parent_uid = dir->i_uid;
 888                 parent_gid = dir->i_gid;
 889         } else {
 890                 rc = err;
 891         }
 892
 893         switch (cleanup_phase) {
 894         case 2: /* child dentry */
 895                 l_dput(dchild);
 896         case 1: /* locked parent dentry */
 897                 if (rc) {
 898                         ldlm_lock_decref(&lockh, LCK_PW);
 899                 } else {
 900                         ptlrpc_save_lock (req, &lockh, LCK_PW);
 901                 }
 902                 l_dput(dparent);
 903         case 0:
 904                 break;
 905         default:
 906                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
 907                 LBUG();
 908         }
 909         req->rq_status = rc;
 910
 911         /* trigger dqacq on the owner of child and parent */
 912         mds_adjust_qunit(obd, current->fsuid, current->fsgid,
 913                          parent_uid, parent_gid, rc);
 914         return 0;
 915 }
 916
 917 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
 918 {
 919         int i;
 920
 921         for (i = 0; i < RES_NAME_SIZE; i++) {
 922                 /* return 1 here, because enqueue_ordered will skip resources
 923                  * of all zeroes if they're sorted to the end of the list. */
 924                 if (res1->name[i] == 0 && res2->name[i] != 0)
 925                         return 1;
 926                 if (res2->name[i] == 0 && res1->name[i] != 0)
 927                         return 0;
 928
 929                 if (res1->name[i] > res2->name[i])
 930                         return 1;
 931                 if (res1->name[i] < res2->name[i])
 932                         return 0;
 933         }
 934         return 0;
 935 }
 936
 937 /* This function doesn't use ldlm_match_or_enqueue because we're always called
 938  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
 939  * because they take the place of local semaphores.
 940  *
 941  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
 942  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
 943 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
 944                           struct lustre_handle *p1_lockh, int p1_lock_mode,
 945                           struct ldlm_res_id *p2_res_id,
 946                           struct lustre_handle *p2_lockh, int p2_lock_mode)
 947 {
 948         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
 949         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
 950         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
 951         int rc, flags;
 952         ENTRY;
 953
 954         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
 955
 956         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
 957                res_id[0]->name[0], res_id[1]->name[0]);
 958
 959         if (res_gt(p1_res_id, p2_res_id)) {
 960                 handles[1] = p1_lockh;
 961                 handles[0] = p2_lockh;
 962                 res_id[1] = p1_res_id;
 963                 res_id[0] = p2_res_id;
 964                 lock_modes[1] = p1_lock_mode;
 965                 lock_modes[0] = p2_lock_mode;
 966         }
 967
 968         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
 969                res_id[0]->name[0], res_id[1]->name[0]);
 970
 971         flags = LDLM_FL_LOCAL_ONLY;
 972         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
 973                               LDLM_PLAIN, NULL, lock_modes[0], &flags,
 974                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 975                               NULL, 0, NULL, handles[0]);
 976         if (rc != ELDLM_OK)
 977                 RETURN(-EIO);
 978         ldlm_lock_dump_handle(D_OTHER, handles[0]);
 979
 980         if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0) {
 981                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
 982                 ldlm_lock_addref(handles[1], lock_modes[1]);
 983         } else if (res_id[1]->name[0] != 0) {
 984                 flags = LDLM_FL_LOCAL_ONLY;
 985                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 986                                       *res_id[1], LDLM_PLAIN, NULL,
 987                                       lock_modes[1], &flags, mds_blocking_ast,
 988                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
 989                                       NULL, handles[1]);
 990                 if (rc != ELDLM_OK) {
 991                         ldlm_lock_decref(handles[0], lock_modes[0]);
 992                         RETURN(-EIO);
 993                 }
 994                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
 995         }
 996
 997         RETURN(0);
 998 }
 999
1000 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1001                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1002                            struct ldlm_res_id *p2_res_id,
1003                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1004                            struct ldlm_res_id *c1_res_id,
1005                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1006                            struct ldlm_res_id *c2_res_id,
1007                            struct lustre_handle *c2_lockh, int c2_lock_mode)
1008 {
1009         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1010                                           c1_res_id, c2_res_id };
1011         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1012                                                  c1_lockh, c2_lockh };
1013         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1014                               c1_lock_mode, c2_lock_mode };
1015         int rc, i, j, sorted, flags;
1016         ENTRY;
1017
1018         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1019                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1020                res_id[3]->name[0]);
1021
1022         /* simple insertion sort - we have at most 4 elements */
1023         for (i = 1; i < 4; i++) {
1024                 j = i - 1;
1025                 dlm_handles[4] = dlm_handles[i];
1026                 res_id[4] = res_id[i];
1027                 lock_modes[4] = lock_modes[i];
1028
1029                 sorted = 0;
1030                 do {
1031                         if (res_gt(res_id[j], res_id[4])) {
1032                                 dlm_handles[j + 1] = dlm_handles[j];
1033                                 res_id[j + 1] = res_id[j];
1034                                 lock_modes[j + 1] = lock_modes[j];
1035                                 j--;
1036                         } else {
1037                                 sorted = 1;
1038                         }
1039                 } while (j >= 0 && !sorted);
1040
1041                 dlm_handles[j + 1] = dlm_handles[4];
1042                 res_id[j + 1] = res_id[4];
1043                 lock_modes[j + 1] = lock_modes[4];
1044         }
1045
1046         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1047                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1048                res_id[3]->name[0]);
1049
1050         /* XXX we could send ASTs on all these locks first before blocking? */
1051         for (i = 0; i < 4; i++) {
1052                 flags = 0;
1053                 if (res_id[i]->name[0] == 0)
1054                         break;
1055                 if (i != 0 &&
1056                     memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) == 0) {
1057                         memcpy(dlm_handles[i], dlm_handles[i-1],
1058                                sizeof(*(dlm_handles[i])));
1059                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1060                 } else {
1061                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1062                                               *res_id[i], LDLM_PLAIN, NULL,
1063                                               lock_modes[i], &flags,
1064                                               mds_blocking_ast,
1065                                               ldlm_completion_ast, NULL, NULL,
1066                                               NULL, 0, NULL, dlm_handles[i]);
1067                         if (rc != ELDLM_OK)
1068                                 GOTO(out_err, rc = -EIO);
1069                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1070                 }
1071         }
1072
1073         RETURN(0);
1074 out_err:
1075         while (i-- > 0)
1076                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1077
1078         return rc;
1079 }
1080
1081 /* In the unlikely case that the child changed while we were waiting
1082  * on the lock, we need to drop the lock on the old child and either:
1083  * - if the child has a lower resource name, then we have to also
1084  *   drop the parent lock and regain the locks in the right order
1085  * - in the rename case, if the child has a lower resource name than one of
1086  *   the other parent/child resources (maxres) we also need to reget the locks
1087  * - if the child has a higher resource name (this is the common case)
1088  *   we can just get the lock on the new child (still in lock order)
1089  *
1090  * Returns 0 if the child did not change or if it changed but could be locked.
1091  * Returns 1 if the child changed and we need to re-lock (no locks held).
1092  * Returns -ve error with a valid dchild (no locks held). */
1093 static int mds_verify_child(struct obd_device *obd,
1094                             struct ldlm_res_id *parent_res_id,
1095                             struct lustre_handle *parent_lockh,
1096                             struct dentry *dparent, int parent_mode,
1097                             struct ldlm_res_id *child_res_id,
1098                             struct lustre_handle *child_lockh,
1099                             struct dentry **dchildp, int child_mode,
1100                             const char *name, int namelen,
1101                             struct ldlm_res_id *maxres)
1102 {
1103         struct dentry *vchild, *dchild = *dchildp;
1104         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1105         ENTRY;
1106
1107         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1108         if (IS_ERR(vchild))
1109                 GOTO(cleanup, rc = PTR_ERR(vchild));
1110
1111         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1112                    (vchild->d_inode != NULL &&
1113                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1114                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1115                 if (dchild != NULL)
1116                         l_dput(dchild);
1117                 *dchildp = vchild;
1118
1119                 RETURN(0);
1120         }
1121
1122         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1123                vchild->d_inode, dchild ? dchild->d_inode : 0,
1124                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1125                child_res_id->name[0]);
1126         if (child_res_id->name[0] != 0)
1127                 ldlm_lock_decref(child_lockh, child_mode);
1128         if (dchild)
1129                 l_dput(dchild);
1130
1131         cleanup_phase = 1; /* parent lock only */
1132         *dchildp = dchild = vchild;
1133
1134         if (dchild->d_inode) {
1135                 int flags = 0;
1136                 child_res_id->name[0] = dchild->d_inode->i_ino;
1137                 child_res_id->name[1] = dchild->d_inode->i_generation;
1138
1139                 if (res_gt(parent_res_id, child_res_id) ||
1140                     res_gt(maxres, child_res_id)) {
1141                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1142                                child_res_id->name[0], parent_res_id->name[0],
1143                                maxres->name[0]);
1144                         GOTO(cleanup, rc = 1);
1145                 }
1146
1147                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1148                                       *child_res_id, LDLM_PLAIN, NULL,
1149                                       child_mode, &flags, mds_blocking_ast,
1150                                       ldlm_completion_ast, NULL, NULL, NULL, 0,
1151                                       NULL, child_lockh);
1152                 if (rc != ELDLM_OK)
1153                         GOTO(cleanup, rc = -EIO);
1154         } else {
1155                 memset(child_res_id, 0, sizeof(*child_res_id));
1156         }
1157
1158         EXIT;
1159 cleanup:
1160         if (rc) {
1161                 switch(cleanup_phase) {
1162                 case 2:
1163                         if (child_res_id->name[0] != 0)
1164                                 ldlm_lock_decref(child_lockh, child_mode);
1165                 case 1:
1166                         ldlm_lock_decref(parent_lockh, parent_mode);
1167                 }
1168         }
1169         return rc;
1170 }
1171
1172 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1173                                 struct ll_fid *fid,
1174                                 struct lustre_handle *parent_lockh,
1175                                 struct dentry **dparentp, int parent_mode,
1176                                 char *name, int namelen,
1177                                 struct lustre_handle *child_lockh,
1178                                 struct dentry **dchildp, int child_mode)
1179 {
1180         struct ldlm_res_id child_res_id = { .name = {0} };
1181         struct ldlm_res_id parent_res_id = { .name = {0} };
1182         struct inode *inode;
1183         int rc = 0, cleanup_phase = 0;
1184         ENTRY;
1185
1186         /* Step 1: Lookup parent */
1187         *dparentp = mds_fid2dentry(mds, fid, NULL);
1188         if (IS_ERR(*dparentp)) {
1189                 rc = PTR_ERR(*dparentp);
1190                 *dparentp = NULL;
1191                 RETURN(rc);
1192         }
1193
1194         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1195                (*dparentp)->d_inode->i_ino, name);
1196
1197         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1198         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1199
1200         cleanup_phase = 1; /* parent dentry */
1201
1202         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1203         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1204         if (IS_ERR(*dchildp)) {
1205                 rc = PTR_ERR(*dchildp);
1206                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1207                 GOTO(cleanup, rc);
1208         }
1209
1210         cleanup_phase = 2; /* child dentry */
1211         inode = (*dchildp)->d_inode;
1212         if (inode != NULL) {
1213                 if (is_bad_inode(inode)) {
1214                         CERROR("bad inode returned %lu/%u\n",
1215                                inode->i_ino, inode->i_generation);
1216                         GOTO(cleanup, rc = -ENOENT);
1217                 }
1218                 inode = igrab(inode);
1219         }
1220         if (inode == NULL)
1221                 goto retry_locks;
1222
1223         child_res_id.name[0] = inode->i_ino;
1224         child_res_id.name[1] = inode->i_generation;
1225         iput(inode);
1226
1227 retry_locks:
1228         cleanup_phase = 2; /* child dentry */
1229
1230         /* Step 3: Lock parent and child in resource order.  If child doesn't
1231          *         exist, we still have to lock the parent and re-lookup. */
1232         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1233                                    &child_res_id, child_lockh, child_mode);
1234         if (rc)
1235                 GOTO(cleanup, rc);
1236
1237         if (!(*dchildp)->d_inode)
1238                 cleanup_phase = 3; /* parent lock */
1239         else
1240                 cleanup_phase = 4; /* child lock */
1241
1242         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1243         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1244                               parent_mode, &child_res_id, child_lockh, dchildp,
1245                               child_mode, name, namelen, &parent_res_id);
1246         if (rc > 0)
1247                 goto retry_locks;
1248         if (rc < 0) {
1249                 cleanup_phase = 3;
1250                 GOTO(cleanup, rc);
1251         }
1252
1253 cleanup:
1254         if (rc) {
1255                 switch (cleanup_phase) {
1256                 case 4:
1257                         ldlm_lock_decref(child_lockh, child_mode);
1258                 case 3:
1259                         ldlm_lock_decref(parent_lockh, parent_mode);
1260                 case 2:
1261                         l_dput(*dchildp);
1262                 case 1:
1263                         l_dput(*dparentp);
1264                 default: ;
1265                 }
1266         }
1267         return rc;
1268 }
1269
1270 void mds_reconstruct_generic(struct ptlrpc_request *req)
1271 {
1272         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1273
1274         mds_req_from_mcd(req, med->med_mcd);
1275 }
1276
1277 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1278  * we instead link the inode into the PENDING directory until it is
1279  * finally released.  We can't simply call mds_reint_rename() or some
1280  * part thereof, because we don't have the inode to check for link
1281  * count/open status until after it is locked.
1282  *
1283  * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1284  * before starting journal transaction.
1285  *
1286  * returns 1 on success
1287  * returns 0 if we lost a race and didn't make a new link
1288  * returns negative on error
1289  */
1290 static int mds_orphan_add_link(struct mds_update_record *rec,
1291                                struct obd_device *obd, struct dentry *dentry)
1292 {
1293         struct mds_obd *mds = &obd->u.mds;
1294         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1295         struct inode *inode = dentry->d_inode;
1296         struct dentry *pending_child;
1297         char fidname[LL_FID_NAMELEN];
1298         int fidlen = 0, rc, mode;
1299         ENTRY;
1300
1301         LASSERT(inode != NULL);
1302         LASSERT(!mds_inode_is_orphan(inode));
1303 #ifndef HAVE_I_ALLOC_SEM
1304         LASSERT(down_trylock(&inode->i_sem) != 0);
1305 #endif
1306         LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1307
1308         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1309
1310         CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1311                mds_orphan_open_count(inode), inode->i_nlink,
1312                S_ISDIR(inode->i_mode) ? "dir" :
1313                 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1314
1315         if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1316                 RETURN(0);
1317
1318         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1319         if (IS_ERR(pending_child))
1320                 RETURN(PTR_ERR(pending_child));
1321
1322         if (pending_child->d_inode != NULL) {
1323                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1324                 LASSERT(pending_child->d_inode == inode);
1325                 GOTO(out_dput, rc = 0);
1326         }
1327
1328         /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1329          * for linking and return real mode back then -bzzz */
1330         mode = inode->i_mode;
1331         inode->i_mode = S_IFREG;
1332         rc = vfs_link(dentry, pending_dir, pending_child);
1333         if (rc)
1334                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1335                        rec->ur_name, rc);
1336         else
1337                 mds_inode_set_orphan(inode);
1338
1339         /* return mode and correct i_nlink if inode is directory */
1340         inode->i_mode = mode;
1341         LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1342                  S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1343                  inode->i_nlink);
1344         if (S_ISDIR(mode)) {
1345                 inode->i_nlink++;
1346                 pending_dir->i_nlink++;
1347                 mark_inode_dirty(inode);
1348                 mark_inode_dirty(pending_dir);
1349         }
1350
1351         GOTO(out_dput, rc = 1);
1352 out_dput:
1353         l_dput(pending_child);
1354         RETURN(rc);
1355 }
1356
1357 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1358                             struct ptlrpc_request *req,
1359                             struct lustre_handle *lh)
1360 {
1361         struct dentry *dparent = NULL, *dchild;
1362         struct mds_obd *mds = mds_req2mds(req);
1363         struct obd_device *obd = req->rq_export->exp_obd;
1364         struct mds_body *body = NULL;
1365         struct inode *child_inode = NULL;
1366         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1367         void *handle = NULL;
1368         int rc = 0, cleanup_phase = 0;
1369         uid_t child_uid = 0, parent_uid = 0;
1370         gid_t child_gid = 0, parent_gid = 0;
1371         ENTRY;
1372
1373         LASSERT(offset == 0 || offset == 2);
1374
1375         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1376                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1377
1378         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1379
1380         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1381                 GOTO(cleanup, rc = -ENOENT);
1382
1383         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1384                                          &parent_lockh, &dparent, LCK_PW,
1385                                          rec->ur_name, rec->ur_namelen,
1386                                          &child_lockh, &dchild, LCK_EX);
1387         if (rc)
1388                 GOTO(cleanup, rc);
1389
1390         cleanup_phase = 1; /* dchild, dparent, locks */
1391
1392         dget(dchild);
1393         child_inode = dchild->d_inode;
1394         if (child_inode == NULL) {
1395                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1396                        dparent->d_inode->i_ino, rec->ur_name);
1397                 GOTO(cleanup, rc = -ENOENT);
1398         }
1399
1400         /* save uid/gid for quota acquire/release */
1401         child_uid = child_inode->i_uid;
1402         child_gid = child_inode->i_gid;
1403         parent_uid = dparent->d_inode->i_uid;
1404         parent_gid = dparent->d_inode->i_gid;
1405
1406         cleanup_phase = 2; /* dchild has a lock */
1407
1408         /* We have to do these checks ourselves, in case we are making an
1409          * orphan.  The client tells us whether rmdir() or unlink() was called,
1410          * so we need to return appropriate errors (bug 72). */
1411         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1412                 if (!S_ISDIR(child_inode->i_mode))
1413                         GOTO(cleanup, rc = -ENOTDIR);
1414         } else {
1415                 if (S_ISDIR(child_inode->i_mode))
1416                         GOTO(cleanup, rc = -EISDIR);
1417         }
1418
1419         /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1420         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1421                 GOTO(cleanup, rc = -EROFS);
1422
1423         /* Step 3: Get a lock on the ino to sync with creation WRT inode
1424          * reuse (see bug 2029). */
1425         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1426         if (rc != ELDLM_OK)
1427                 GOTO(cleanup, rc);
1428
1429         cleanup_phase = 3; /* child inum lock */
1430
1431         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1432
1433         /* ldlm_reply in buf[0] if called via intent */
1434         if (offset)
1435                 offset = 1;
1436
1437         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1438         LASSERT(body != NULL);
1439
1440         /* child orphan sem protects orphan_dec_test && is_orphan race */
1441         MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1442         cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1443
1444         /* If this is potentially the last reference to this inode, get the
1445          * OBD EA data first so the client can destroy OST objects.  We
1446          * only do the object removal later if no open files/links remain. */
1447         if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1448             child_inode->i_nlink == 1) {
1449                 if (mds_orphan_open_count(child_inode) > 0) {
1450                         /* need to lock pending_dir before transaction */
1451                         down(&mds->mds_pending_dir->d_inode->i_sem);
1452                         cleanup_phase = 5; /* up(&pending_dir->i_sem) */
1453                 } else if (S_ISREG(child_inode->i_mode)) {
1454                         mds_pack_inode2fid(&body->fid1, child_inode);
1455                         mds_pack_inode2body(body, child_inode);
1456                         mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1457                                     child_inode, MDS_PACK_MD_LOCK);
1458                 }
1459         }
1460
1461         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1462         switch (child_inode->i_mode & S_IFMT) {
1463         case S_IFDIR:
1464                 /* Drop any lingering child directories before we start our
1465                  * transaction, to avoid doing multiple inode dirty/delete
1466                  * in our compound transaction (bug 1321). */
1467                 shrink_dcache_parent(dchild);
1468                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1469                                       NULL);
1470                 if (IS_ERR(handle))
1471                         GOTO(cleanup, rc = PTR_ERR(handle));
1472                 rc = vfs_rmdir(dparent->d_inode, dchild);
1473                 break;
1474         case S_IFREG: {
1475                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1476                                                         offset + 1, 0);
1477                 handle = fsfilt_start_log(obd, dparent->d_inode,
1478                                           FSFILT_OP_UNLINK, NULL,
1479                                           le32_to_cpu(lmm->lmm_stripe_count));
1480                 if (IS_ERR(handle))
1481                         GOTO(cleanup, rc = PTR_ERR(handle));
1482                 rc = vfs_unlink(dparent->d_inode, dchild);
1483                 break;
1484         }
1485         case S_IFLNK:
1486         case S_IFCHR:
1487         case S_IFBLK:
1488         case S_IFIFO:
1489         case S_IFSOCK:
1490                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1491                                       NULL);
1492                 if (IS_ERR(handle))
1493                         GOTO(cleanup, rc = PTR_ERR(handle));
1494                 rc = vfs_unlink(dparent->d_inode, dchild);
1495                 break;
1496         default:
1497                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1498                        rec->ur_name);
1499                 LBUG();
1500                 GOTO(cleanup, rc = -EINVAL);
1501         }
1502
1503         if (rc == 0 && child_inode->i_nlink == 0) {
1504                 if (mds_orphan_open_count(child_inode) > 0)
1505                         rc = mds_orphan_add_link(rec, obd, dchild);
1506
1507                 if (rc == 1)
1508                         GOTO(cleanup, rc = 0);
1509
1510                 if (!S_ISREG(child_inode->i_mode))
1511                         GOTO(cleanup, rc);
1512
1513                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1514                         body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1515                                        OBD_MD_FLATIME | OBD_MD_FLMTIME);
1516                 } else if (mds_log_op_unlink(obd, child_inode,
1517                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1518                                         req->rq_repmsg->buflens[offset + 1],
1519                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1520                                         req->rq_repmsg->buflens[offset+2]) > 0){
1521                         body->valid |= OBD_MD_FLCOOKIE;
1522                 }
1523         }
1524
1525         GOTO(cleanup, rc);
1526 cleanup:
1527         if (rc == 0) {
1528                 struct iattr iattr;
1529                 int err;
1530
1531                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1532                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1533                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1534
1535                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1536                 if (err)
1537                         CERROR("error on parent setattr: rc = %d\n", err);
1538         }
1539
1540         rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
1541                                 handle, req, rc, 0);
1542         if (!rc)
1543                 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1544                                    "unlinked", 0, NULL);
1545         switch(cleanup_phase) {
1546         case 5: /* pending_dir semaphore */
1547                 up(&mds->mds_pending_dir->d_inode->i_sem);
1548         case 4: /* child inode semaphore */
1549                 MDS_UP_READ_ORPHAN_SEM(child_inode);
1550         case 3: /* child ino-reuse lock */
1551                 if (rc && body != NULL) {
1552                         // Don't unlink the OST objects if the MDS unlink failed
1553                         body->valid = 0;
1554                 }
1555                 if (rc)
1556                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1557                 else
1558                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1559         case 2: /* child lock */
1560                 ldlm_lock_decref(&child_lockh, LCK_EX);
1561         case 1: /* child and parent dentry, parent lock */
1562                 if (rc)
1563                         ldlm_lock_decref(&parent_lockh, LCK_PW);
1564                 else
1565                         ptlrpc_save_lock(req, &parent_lockh, LCK_PW);
1566                 l_dput(dchild);
1567                 l_dput(dchild);
1568                 l_dput(dparent);
1569         case 0:
1570                 break;
1571         default:
1572                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1573                 LBUG();
1574         }
1575         req->rq_status = rc;
1576
1577         /* trigger dqrel on the owner of child and parent */
1578         mds_adjust_qunit(obd, child_uid, child_gid, parent_uid, parent_gid, rc);
1579         return 0;
1580 }
1581
1582 static int mds_reint_link(struct mds_update_record *rec, int offset,
1583                           struct ptlrpc_request *req,
1584                           struct lustre_handle *lh)
1585 {
1586         struct obd_device *obd = req->rq_export->exp_obd;
1587         struct dentry *de_src = NULL;
1588         struct dentry *de_tgt_dir = NULL;
1589         struct dentry *dchild = NULL;
1590         struct mds_obd *mds = mds_req2mds(req);
1591         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1592         struct ldlm_res_id src_res_id = { .name = {0} };
1593         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1594         int rc = 0, cleanup_phase = 0;
1595         ENTRY;
1596
1597         LASSERT(offset == 0);
1598
1599         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1600                   rec->ur_fid1->id, rec->ur_fid1->generation,
1601                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1602
1603         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1604
1605         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1606                 GOTO(cleanup, rc = -ENOENT);
1607
1608         /* Step 1: Lookup the source inode and target directory by FID */
1609         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1610         if (IS_ERR(de_src))
1611                 GOTO(cleanup, rc = PTR_ERR(de_src));
1612
1613         cleanup_phase = 1; /* source dentry */
1614
1615         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1616         if (IS_ERR(de_tgt_dir)) {
1617                 rc = PTR_ERR(de_tgt_dir);
1618                 de_tgt_dir = NULL;
1619                 GOTO(cleanup, rc);
1620         }
1621
1622         cleanup_phase = 2; /* target directory dentry */
1623
1624         CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
1625                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1626                de_src->d_inode->i_ino);
1627
1628         /* Step 2: Take the two locks */
1629         src_res_id.name[0] = de_src->d_inode->i_ino;
1630         src_res_id.name[1] = de_src->d_inode->i_generation;
1631         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1632         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1633
1634         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1635                                    &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX);
1636         if (rc)
1637                 GOTO(cleanup, rc);
1638
1639         cleanup_phase = 3; /* locks */
1640
1641         /* Step 3: Lookup the child */
1642         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1643         if (IS_ERR(dchild)) {
1644                 rc = PTR_ERR(dchild);
1645                 if (rc != -EPERM && rc != -EACCES)
1646                         CERROR("child lookup error %d\n", rc);
1647                 GOTO(cleanup, rc);
1648         }
1649
1650         cleanup_phase = 4; /* child dentry */
1651
1652         if (dchild->d_inode) {
1653                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1654                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1655                 rc = -EEXIST;
1656                 GOTO(cleanup, rc);
1657         }
1658
1659         /* Step 4: Do it. */
1660         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1661
1662         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1663                 GOTO(cleanup, rc = -EROFS);
1664
1665         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1666         if (IS_ERR(handle)) {
1667                 rc = PTR_ERR(handle);
1668                 GOTO(cleanup, rc);
1669         }
1670
1671         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1672         if (rc && rc != -EPERM && rc != -EACCES)
1673                 CERROR("vfs_link error %d\n", rc);
1674 cleanup:
1675         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1676                                 handle, req, rc, 0);
1677         EXIT;
1678
1679         switch (cleanup_phase) {
1680         case 4: /* child dentry */
1681                 l_dput(dchild);
1682         case 3: /* locks */
1683                 if (rc) {
1684                         ldlm_lock_decref(&src_lockh, LCK_EX);
1685                         ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1686                 } else {
1687                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1688                         ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1689                 }
1690         case 2: /* target dentry */
1691                 l_dput(de_tgt_dir);
1692         case 1: /* source dentry */
1693                 l_dput(de_src);
1694         case 0:
1695                 break;
1696         default:
1697                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1698                 LBUG();
1699         }
1700         req->rq_status = rc;
1701         return 0;
1702 }
1703
1704 /* The idea here is that we need to get four locks in the end:
1705  * one on each parent directory, one on each child.  We need to take
1706  * these locks in some kind of order (to avoid deadlocks), and the order
1707  * I selected is "increasing resource number" order.  We need to look up
1708  * the children, however, before we know what the resource number(s) are.
1709  * Thus the following plan:
1710  *
1711  * 1,2. Look up the parents
1712  * 3,4. Look up the children
1713  * 5. Take locks on the parents and children, in order
1714  * 6. Verify that the children haven't changed since they were looked up
1715  *
1716  * If there was a race and the children changed since they were first looked
1717  * up, it is possible that mds_verify_child() will be able to just grab the
1718  * lock on the new child resource (if it has a higher resource than any other)
1719  * but we need to compare against not only its parent, but also against the
1720  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1721  *
1722  * We need the fancy igrab() on the child inodes because we aren't holding a
1723  * lock on the parent after the lookup is done, so dentry->d_inode may change
1724  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1725  */
1726 static int mds_get_parents_children_locked(struct obd_device *obd,
1727                                            struct mds_obd *mds,
1728                                            struct ll_fid *p1_fid,
1729                                            struct dentry **de_srcdirp,
1730                                            struct ll_fid *p2_fid,
1731                                            struct dentry **de_tgtdirp,
1732                                            int parent_mode,
1733                                            const char *old_name, int old_len,
1734                                            struct dentry **de_oldp,
1735                                            const char *new_name, int new_len,
1736                                            struct dentry **de_newp,
1737                                            struct lustre_handle *dlm_handles,
1738                                            int child_mode)
1739 {
1740         struct ldlm_res_id p1_res_id = { .name = {0} };
1741         struct ldlm_res_id p2_res_id = { .name = {0} };
1742         struct ldlm_res_id c1_res_id = { .name = {0} };
1743         struct ldlm_res_id c2_res_id = { .name = {0} };
1744         struct ldlm_res_id *maxres_src, *maxres_tgt;
1745         struct inode *inode;
1746         int rc = 0, cleanup_phase = 0;
1747         ENTRY;
1748
1749         /* Step 1: Lookup the source directory */
1750         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1751         if (IS_ERR(*de_srcdirp))
1752                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1753
1754         cleanup_phase = 1; /* source directory dentry */
1755
1756         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1757         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1758
1759         /* Step 2: Lookup the target directory */
1760         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1761                 *de_tgtdirp = dget(*de_srcdirp);
1762         } else {
1763                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1764                 if (IS_ERR(*de_tgtdirp)) {
1765                         rc = PTR_ERR(*de_tgtdirp);
1766                         *de_tgtdirp = NULL;
1767                         GOTO(cleanup, rc);
1768                 }
1769         }
1770
1771         cleanup_phase = 2; /* target directory dentry */
1772
1773         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1774         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1775
1776         /* Step 3: Lookup the source child entry */
1777         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1778         if (IS_ERR(*de_oldp)) {
1779                 rc = PTR_ERR(*de_oldp);
1780                 CERROR("old child lookup error (%.*s): %d\n",
1781                        old_len - 1, old_name, rc);
1782                 GOTO(cleanup, rc);
1783         }
1784
1785         cleanup_phase = 3; /* original name dentry */
1786
1787         inode = (*de_oldp)->d_inode;
1788         if (inode != NULL)
1789                 inode = igrab(inode);
1790         if (inode == NULL)
1791                 GOTO(cleanup, rc = -ENOENT);
1792
1793         c1_res_id.name[0] = inode->i_ino;
1794         c1_res_id.name[1] = inode->i_generation;
1795
1796         iput(inode);
1797
1798         /* Step 4: Lookup the target child entry */
1799         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1800         if (IS_ERR(*de_newp)) {
1801                 rc = PTR_ERR(*de_newp);
1802                 CERROR("new child lookup error (%.*s): %d\n",
1803                        old_len - 1, old_name, rc);
1804                 GOTO(cleanup, rc);
1805         }
1806
1807         cleanup_phase = 4; /* target dentry */
1808
1809         inode = (*de_newp)->d_inode;
1810         if (inode != NULL)
1811                 inode = igrab(inode);
1812         if (inode == NULL)
1813                 goto retry_locks;
1814
1815         c2_res_id.name[0] = inode->i_ino;
1816         c2_res_id.name[1] = inode->i_generation;
1817         iput(inode);
1818
1819 retry_locks:
1820         /* Step 5: Take locks on the parents and child(ren) */
1821         maxres_src = &p1_res_id;
1822         maxres_tgt = &p2_res_id;
1823         cleanup_phase = 4; /* target dentry */
1824
1825         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id))
1826                 maxres_src = &c1_res_id;
1827         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id))
1828                 maxres_tgt = &c2_res_id;
1829
1830         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1831                                     &p2_res_id, &dlm_handles[1], parent_mode,
1832                                     &c1_res_id, &dlm_handles[2], child_mode,
1833                                     &c2_res_id, &dlm_handles[3], child_mode);
1834         if (rc)
1835                 GOTO(cleanup, rc);
1836
1837         cleanup_phase = 6; /* parent and child(ren) locks */
1838
1839         /* Step 6a: Re-lookup source child to verify it hasn't changed */
1840         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1841                               parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
1842                               child_mode, old_name, old_len, maxres_tgt);
1843         if (rc) {
1844                 if (c2_res_id.name[0] != 0)
1845                         ldlm_lock_decref(&dlm_handles[3], child_mode);
1846                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1847                 cleanup_phase = 4;
1848                 if (rc > 0)
1849                         goto retry_locks;
1850                 GOTO(cleanup, rc);
1851         }
1852
1853         if ((*de_oldp)->d_inode == NULL)
1854                 GOTO(cleanup, rc = -ENOENT);
1855
1856         /* Step 6b: Re-lookup target child to verify it hasn't changed */
1857         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
1858                               parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
1859                               child_mode, new_name, new_len, maxres_src);
1860         if (rc) {
1861                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1862                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1863                 cleanup_phase = 4;
1864                 if (rc > 0)
1865                         goto retry_locks;
1866                 GOTO(cleanup, rc);
1867         }
1868
1869         EXIT;
1870 cleanup:
1871         if (rc) {
1872                 switch (cleanup_phase) {
1873                 case 6: /* child lock(s) */
1874                         if (c2_res_id.name[0] != 0)
1875                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
1876                         if (c1_res_id.name[0] != 0)
1877                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
1878                 case 5: /* parent locks */
1879                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
1880                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
1881                 case 4: /* target dentry */
1882                         l_dput(*de_newp);
1883                 case 3: /* source dentry */
1884                         l_dput(*de_oldp);
1885                 case 2: /* target directory dentry */
1886                         l_dput(*de_tgtdirp);
1887                 case 1: /* source directry dentry */
1888                         l_dput(*de_srcdirp);
1889                 }
1890         }
1891
1892         return rc;
1893 }
1894
1895 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1896                             struct ptlrpc_request *req,
1897                             struct lustre_handle *lockh)
1898 {
1899         struct obd_device *obd = req->rq_export->exp_obd;
1900         struct dentry *de_srcdir = NULL;
1901         struct dentry *de_tgtdir = NULL;
1902         struct dentry *de_old = NULL;
1903         struct dentry *de_new = NULL;
1904         struct inode *old_inode = NULL, *new_inode = NULL;
1905         struct mds_obd *mds = mds_req2mds(req);
1906         struct lustre_handle dlm_handles[4];
1907         struct mds_body *body = NULL;
1908         struct lov_mds_md *lmm = NULL;
1909         int rc = 0, lock_count = 3, cleanup_phase = 0;
1910         void *handle = NULL;
1911         ENTRY;
1912
1913         LASSERT(offset == 0);
1914
1915         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
1916                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
1917                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
1918
1919         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1920
1921         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
1922                                              rec->ur_fid2, &de_tgtdir, LCK_PW,
1923                                              rec->ur_name, rec->ur_namelen,
1924                                              &de_old, rec->ur_tgt,
1925                                              rec->ur_tgtlen, &de_new,
1926                                              dlm_handles, LCK_EX);
1927         if (rc)
1928                 GOTO(cleanup, rc);
1929
1930         cleanup_phase = 1; /* parent(s), children, locks */
1931
1932         old_inode = de_old->d_inode;
1933         new_inode = de_new->d_inode;
1934
1935         if (new_inode != NULL)
1936                 lock_count = 4;
1937
1938         /* sanity check for src inode */
1939         if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
1940             old_inode->i_ino == de_tgtdir->d_inode->i_ino)
1941                 GOTO(cleanup, rc = -EINVAL);
1942
1943         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1944                 GOTO(cleanup, rc = -EROFS);
1945
1946         if (new_inode == NULL)
1947                 goto no_unlink;
1948
1949         igrab(new_inode);
1950         cleanup_phase = 2; /* iput(new_inode) when finished */
1951
1952         /* sanity check for dest inode */
1953         if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
1954             new_inode->i_ino == de_tgtdir->d_inode->i_ino)
1955                 GOTO(cleanup, rc = -EINVAL);
1956
1957         if (old_inode == new_inode)
1958                 GOTO(cleanup, rc = 0);
1959
1960         /* if we are about to remove the target at first, pass the EA of
1961          * that inode to client to perform and cleanup on OST */
1962         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1963         LASSERT(body != NULL);
1964
1965         /* child orphan sem protects orphan_dec_test && is_orphan race */
1966         MDS_DOWN_READ_ORPHAN_SEM(new_inode);
1967         cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1968
1969         if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
1970             new_inode->i_nlink == 1) {
1971                 if (mds_orphan_open_count(new_inode) > 0) {
1972                         /* need to lock pending_dir before transaction */
1973                         down(&mds->mds_pending_dir->d_inode->i_sem);
1974                         cleanup_phase = 4; /* up(&pending_dir->i_sem) */
1975                 } else if (S_ISREG(new_inode->i_mode)) {
1976                         mds_pack_inode2fid(&body->fid1, new_inode);
1977                         mds_pack_inode2body(body, new_inode);
1978                         mds_pack_md(obd, req->rq_repmsg, 1, body, new_inode,
1979                                     MDS_PACK_MD_LOCK);
1980                 }
1981         }
1982
1983 no_unlink:
1984         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
1985                        de_srcdir->d_inode->i_sb);
1986
1987 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
1988         /* Check if we are moving old entry into its child. 2.6 does not
1989            check for this in vfs_rename() anymore */
1990         if (is_subdir(de_new, de_old))
1991                 GOTO(cleanup, rc = -EINVAL);
1992 #endif
1993
1994         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1995         handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
1996                                   NULL, le32_to_cpu(lmm->lmm_stripe_count));
1997
1998         if (IS_ERR(handle))
1999                 GOTO(cleanup, rc = PTR_ERR(handle));
2000
2001         lock_kernel();
2002         de_old->d_fsdata = req;
2003         de_new->d_fsdata = req;
2004
2005         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2006         unlock_kernel();
2007
2008         if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2009                 if (mds_orphan_open_count(new_inode) > 0)
2010                         rc = mds_orphan_add_link(rec, obd, de_new);
2011
2012                 if (rc == 1)
2013                         GOTO(cleanup, rc = 0);
2014
2015                 if (!S_ISREG(new_inode->i_mode))
2016                         GOTO(cleanup, rc);
2017
2018                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2019                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2020                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2021                 } else if (mds_log_op_unlink(obd, new_inode,
2022                                              lustre_msg_buf(req->rq_repmsg,1,0),
2023                                              req->rq_repmsg->buflens[1],
2024                                              lustre_msg_buf(req->rq_repmsg,2,0),
2025                                              req->rq_repmsg->buflens[2]) > 0) {
2026                         body->valid |= OBD_MD_FLCOOKIE;
2027                 }
2028         }
2029
2030         GOTO(cleanup, rc);
2031 cleanup:
2032         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2033                                 handle, req, rc, 0);
2034
2035         switch (cleanup_phase) {
2036         case 4:
2037                 up(&mds->mds_pending_dir->d_inode->i_sem);
2038         case 3:
2039                 MDS_UP_READ_ORPHAN_SEM(new_inode);
2040         case 2:
2041                 iput(new_inode);
2042         case 1:
2043                 if (rc) {
2044                         if (lock_count == 4)
2045                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2046                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2047                         ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2048                         ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2049                 } else {
2050                         if (lock_count == 4)
2051                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2052                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2053                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2054                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2055                 }
2056                 l_dput(de_new);
2057                 l_dput(de_old);
2058                 l_dput(de_tgtdir);
2059                 l_dput(de_srcdir);
2060         case 0:
2061                 break;
2062         default:
2063                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2064                 LBUG();
2065         }
2066         req->rq_status = rc;
2067         return 0;
2068 }
2069
2070 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2071                            struct ptlrpc_request *, struct lustre_handle *);
2072
2073 static mds_reinter reinters[REINT_MAX] = {
2074         [REINT_SETATTR] mds_reint_setattr,
2075         [REINT_CREATE] mds_reint_create,
2076         [REINT_LINK] mds_reint_link,
2077         [REINT_UNLINK] mds_reint_unlink,
2078         [REINT_RENAME] mds_reint_rename,
2079         [REINT_OPEN] mds_open
2080 };
2081
2082 int mds_reint_rec(struct mds_update_record *rec, int offset,
2083                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2084 {
2085         struct obd_device *obd = req->rq_export->exp_obd;
2086         struct obd_run_ctxt saved;
2087         int rc;
2088         ENTRY;
2089
2090         /* checked by unpacker */
2091         LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2092
2093         push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2094         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2095         pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2096
2097         RETURN(rc);
2098 }