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 the Lustre file system, http://www.lustre.org
  13  *   Lustre is a trademark of Cluster File Systems, Inc.
  14  *
  15  *   You may have signed or agreed to another license before downloading
  16  *   this software.  If so, you are bound by the terms and conditions
  17  *   of that agreement, and the following does not apply to you.  See the
  18  *   LICENSE file included with this distribution for more information.
  19  *
  20  *   If you did not agree to a different license, then this copy of Lustre
  21  *   is open source software; you can redistribute it and/or modify it
  22  *   under the terms of version 2 of the GNU General Public License as
  23  *   published by the Free Software Foundation.
  24  *
  25  *   In either case, Lustre is distributed in the hope that it will be
  26  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  27  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  28  *   license text for more details.
  29  */
  30
  31 #ifndef EXPORT_SYMTAB
  32 # define EXPORT_SYMTAB
  33 #endif
  34 #define DEBUG_SUBSYSTEM S_MDS
  35
  36 #include <linux/fs.h>
  37 #include <obd_support.h>
  38 #include <obd_class.h>
  39 #include <obd.h>
  40 #include <lustre_lib.h>
  41 #include <lustre/lustre_idl.h>
  42 #include <lustre_mds.h>
  43 #include <lustre_dlm.h>
  44 #include <lustre_fsfilt.h>
  45 #include <lustre_ucache.h>
  46
  47 #include "mds_internal.h"
  48
  49 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
  50                    int error)
  51 {
  52         obd_transno_commit_cb(obd, transno, error);
  53 }
  54
  55 struct mds_logcancel_data {
  56         struct lov_mds_md      *mlcd_lmm;
  57         int                     mlcd_size;
  58         int                     mlcd_cookielen;
  59         int                     mlcd_eadatalen;
  60         struct llog_cookie      mlcd_cookies[0];
  61 };
  62
  63
  64 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
  65                                   void *cb_data, int error)
  66 {
  67         struct mds_logcancel_data *mlcd = cb_data;
  68         struct lov_stripe_md *lsm = NULL;
  69         struct llog_ctxt *ctxt;
  70         int rc;
  71
  72         obd_transno_commit_cb(obd, transno, error);
  73
  74         CDEBUG(D_RPCTRACE, "cancelling %d cookies\n",
  75                (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
  76
  77         rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
  78                           mlcd->mlcd_eadatalen);
  79         if (rc < 0) {
  80                 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
  81                        (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
  82                        rc);
  83         } else {
  84                 ///* XXX 0 normally, SENDNOW for debug */);
  85                 rc = obd_checkmd(obd->u.mds.mds_osc_exp, obd->obd_self_export,
  86                                  lsm);
  87                 if (rc)
  88                         CERROR("Can not revalidate lsm %p \n", lsm);
  89
  90                 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
  91                 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
  92                                                 sizeof(*mlcd->mlcd_cookies),
  93                                  mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
  94                 llog_ctxt_put(ctxt);
  95
  96                 if (rc)
  97                         CERROR("error cancelling %d log cookies: rc %d\n",
  98                                (int)(mlcd->mlcd_cookielen /
  99                                      sizeof(*mlcd->mlcd_cookies)), rc);
 100         }
 101
 102         OBD_FREE(mlcd, mlcd->mlcd_size);
 103 }
 104
 105 /* Assumes caller has already pushed us into the kernel context. */
 106 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
 107                        struct ptlrpc_request *req, int rc, __u32 op_data,
 108                        int force_sync)
 109 {
 110         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 111         struct lsd_client_data *lcd = med->med_lcd;
 112         struct obd_device *obd = req->rq_export->exp_obd;
 113         __u64 transno, prev_transno;
 114         int err;
 115         loff_t off;
 116         int log_pri = D_RPCTRACE;
 117         ENTRY;
 118
 119         if (IS_ERR(handle)) {
 120                 LASSERT(rc != 0);
 121                 RETURN(rc);
 122         }
 123
 124         /* if the export has already been failed, we have no last_rcvd slot */
 125         if (req->rq_export->exp_failed || obd->obd_fail) {
 126                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 127                       req->rq_export->exp_client_uuid.uuid, rc);
 128                 if (rc == 0)
 129                         rc = -ENOTCONN;
 130                 if (handle)
 131                         GOTO(commit, rc);
 132                 RETURN(rc);
 133         }
 134
 135         if (handle == NULL) {
 136                 /* if we're starting our own xaction, use our own inode */
 137                 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
 138                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 139                 if (IS_ERR(handle)) {
 140                         CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
 141                         RETURN(PTR_ERR(handle));
 142                 }
 143         }
 144
 145         off = med->med_lr_off;
 146
 147         transno = lustre_msg_get_transno(req->rq_reqmsg);
 148         if (rc != 0) {
 149                 if (transno != 0) {
 150                         CERROR("%s: replay %s transno "LPU64" failed: rc %d\n",
 151                                obd->obd_name,
 152                                libcfs_nid2str(req->rq_export->exp_connection->c_peer.nid),
 153                                transno, rc);
 154                         transno = 0;
 155                 }
 156         } else if (transno == 0) {
 157                 spin_lock(&mds->mds_transno_lock);
 158                 transno = ++mds->mds_last_transno;
 159                 spin_unlock(&mds->mds_transno_lock);
 160         } else {
 161                 spin_lock(&mds->mds_transno_lock);
 162                 if (transno > mds->mds_last_transno)
 163                         mds->mds_last_transno = transno;
 164                 spin_unlock(&mds->mds_transno_lock);
 165         }
 166
 167         req->rq_transno = transno;
 168         lustre_msg_set_transno(req->rq_repmsg, transno);
 169         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 170                 prev_transno = le64_to_cpu(lcd->lcd_last_close_transno);
 171                 lcd->lcd_last_close_transno = cpu_to_le64(transno);
 172                 lcd->lcd_last_close_xid = cpu_to_le64(req->rq_xid);
 173                 lcd->lcd_last_close_result = cpu_to_le32(rc);
 174                 lcd->lcd_last_close_data = cpu_to_le32(op_data);
 175         } else {
 176                 prev_transno = le64_to_cpu(lcd->lcd_last_transno);
 177                 if (((lustre_msg_get_flags(req->rq_reqmsg) &
 178                       (MSG_RESENT | MSG_REPLAY)) == 0) ||
 179                     (transno > prev_transno)) {
 180                         lcd->lcd_last_transno = cpu_to_le64(transno);
 181                         lcd->lcd_last_xid     = cpu_to_le64(req->rq_xid);
 182                         lcd->lcd_last_result  = cpu_to_le32(rc);
 183                         lcd->lcd_last_data    = cpu_to_le32(op_data);
 184                 }
 185         }
 186         /* update the server data to not lose the greatest transno. Bug 11125 */
 187         if ((transno == 0) && (prev_transno == mds->mds_last_transno))
 188                 mds_update_server_data(obd, 0);
 189
 190         if (off <= 0) {
 191                 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
 192                 err = -EINVAL;
 193         } else {
 194                 struct obd_export *exp = req->rq_export;
 195
 196                 if (!force_sync)
 197                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,transno,
 198                                                           handle, mds_commit_cb,
 199                                                           NULL);
 200
 201                 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, lcd,
 202                                           sizeof(*lcd), &off,
 203                                           force_sync | exp->exp_need_sync);
 204                 if (force_sync)
 205                         mds_commit_cb(obd, transno, NULL, err);
 206         }
 207
 208         if (err) {
 209                 log_pri = D_ERROR;
 210                 if (rc == 0)
 211                         rc = err;
 212         }
 213
 214         DEBUG_REQ(log_pri, req,
 215                   "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
 216                   transno, rc, lcd->lcd_uuid, med->med_lr_idx, err);
 217
 218         err = mds_lov_write_objids(obd);
 219         if (err) {
 220                 log_pri = D_ERROR;
 221                 if (rc == 0)
 222                         rc = err;
 223         }
 224         CDEBUG(log_pri, "wrote objids: err = %d\n", err);
 225
 226 commit:
 227         err = fsfilt_commit(obd, inode, handle, 0);
 228         if (err) {
 229                 CERROR("error committing transaction: %d\n", err);
 230                 if (!rc)
 231                         rc = err;
 232         }
 233
 234         RETURN(rc);
 235 }
 236
 237 /* this gives the same functionality as the code between
 238  * sys_chmod and inode_setattr
 239  * chown_common and inode_setattr
 240  * utimes and inode_setattr
 241  */
 242 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
 243 {
 244         time_t now = CURRENT_SECONDS;
 245         struct iattr *attr = &rec->ur_iattr;
 246         unsigned int ia_valid = attr->ia_valid;
 247         int error;
 248         ENTRY;
 249
 250         if (ia_valid & ATTR_RAW)
 251                 attr->ia_valid &= ~ATTR_RAW;
 252
 253         if (!(ia_valid & ATTR_CTIME_SET))
 254                 LTIME_S(attr->ia_ctime) = now;
 255         else
 256                 attr->ia_valid &= ~ATTR_CTIME_SET;
 257         if (!(ia_valid & ATTR_ATIME_SET))
 258                 LTIME_S(attr->ia_atime) = now;
 259         if (!(ia_valid & ATTR_MTIME_SET))
 260                 LTIME_S(attr->ia_mtime) = now;
 261
 262         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 263                 RETURN((attr->ia_valid & ~ATTR_ATTR_FLAG) ? -EPERM : 0);
 264
 265         /* times */
 266         if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
 267                 if (current->fsuid != inode->i_uid &&
 268                     (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 269                         RETURN(error);
 270         }
 271
 272         if (ia_valid & ATTR_SIZE &&
 273             /* NFSD hack for open(O_CREAT|O_TRUNC)=mknod+truncate (bug 5781) */
 274             !(rec->ur_uc.luc_fsuid == inode->i_uid &&
 275               ia_valid & MDS_OPEN_OWNEROVERRIDE)) {
 276                 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
 277                         RETURN(error);
 278         }
 279
 280         if (ia_valid & (ATTR_UID | ATTR_GID)) {
 281                 /* chown */
 282                 error = -EPERM;
 283                 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 284                         RETURN(-EPERM);
 285                 if (attr->ia_uid == (uid_t) -1)
 286                         attr->ia_uid = inode->i_uid;
 287                 if (attr->ia_gid == (gid_t) -1)
 288                         attr->ia_gid = inode->i_gid;
 289                 if (!(ia_valid & ATTR_MODE))
 290                         attr->ia_mode = inode->i_mode;
 291                 /*
 292                  * If the user or group of a non-directory has been
 293                  * changed by a non-root user, remove the setuid bit.
 294                  * 19981026 David C Niemi <niemi@tux.org>
 295                  *
 296                  * Changed this to apply to all users, including root,
 297                  * to avoid some races. This is the behavior we had in
 298                  * 2.0. The check for non-root was definitely wrong
 299                  * for 2.2 anyway, as it should have been using
 300                  * CAP_FSETID rather than fsuid -- 19990830 SD.
 301                  */
 302                 if ((inode->i_mode & S_ISUID) == S_ISUID &&
 303                     !S_ISDIR(inode->i_mode)) {
 304                         attr->ia_mode &= ~S_ISUID;
 305                         attr->ia_valid |= ATTR_MODE;
 306                 }
 307                 /*
 308                  * Likewise, if the user or group of a non-directory
 309                  * has been changed by a non-root user, remove the
 310                  * setgid bit UNLESS there is no group execute bit
 311                  * (this would be a file marked for mandatory
 312                  * locking).  19981026 David C Niemi <niemi@tux.org>
 313                  *
 314                  * Removed the fsuid check (see the comment above) --
 315                  * 19990830 SD.
 316                  */
 317                 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
 318                      (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
 319                         attr->ia_mode &= ~S_ISGID;
 320                         attr->ia_valid |= ATTR_MODE;
 321                 }
 322         } else if (ia_valid & ATTR_MODE) {
 323                 int mode = attr->ia_mode;
 324                 /* chmod */
 325                 if (attr->ia_mode == (umode_t)-1)
 326                         mode = inode->i_mode;
 327                 attr->ia_mode =
 328                         (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
 329         }
 330         RETURN(0);
 331 }
 332
 333 void mds_steal_ack_locks(struct ptlrpc_request *req)
 334 {
 335         struct obd_export         *exp = req->rq_export;
 336         struct list_head          *tmp;
 337         struct ptlrpc_reply_state *oldrep;
 338         struct ptlrpc_service     *svc;
 339         int                        i;
 340
 341         /* CAVEAT EMPTOR: spinlock order */
 342         spin_lock(&exp->exp_lock);
 343         list_for_each (tmp, &exp->exp_outstanding_replies) {
 344                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
 345
 346                 if (oldrep->rs_xid != req->rq_xid)
 347                         continue;
 348
 349                 if (lustre_msg_get_opc(oldrep->rs_msg) !=
 350                     lustre_msg_get_opc(req->rq_reqmsg))
 351                         CERROR ("Resent req xid "LPX64" has mismatched opc: "
 352                                 "new %d old %d\n", req->rq_xid,
 353                                 lustre_msg_get_opc(req->rq_reqmsg),
 354                                 lustre_msg_get_opc(oldrep->rs_msg));
 355
 356                 svc = oldrep->rs_service;
 357                 spin_lock (&svc->srv_lock);
 358
 359                 list_del_init (&oldrep->rs_exp_list);
 360
 361                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 362                       " o%d NID %s\n",
 363                       oldrep->rs_nlocks, oldrep,
 364                       oldrep->rs_xid, oldrep->rs_transno,
 365                       lustre_msg_get_opc(oldrep->rs_msg),
 366                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 367
 368                 for (i = 0; i < oldrep->rs_nlocks; i++)
 369                         ptlrpc_save_lock(req,
 370                                          &oldrep->rs_locks[i],
 371                                          oldrep->rs_modes[i]);
 372                 oldrep->rs_nlocks = 0;
 373
 374                 DEBUG_REQ(D_DLMTRACE, req, "stole locks for");
 375                 ptlrpc_schedule_difficult_reply (oldrep);
 376
 377                 spin_unlock (&svc->srv_lock);
 378                 break;
 379         }
 380         spin_unlock(&exp->exp_lock);
 381 }
 382
 383 void mds_req_from_lcd(struct ptlrpc_request *req, struct lsd_client_data *lcd)
 384 {
 385         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 386                 req->rq_transno = le64_to_cpu(lcd->lcd_last_close_transno);
 387                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
 388                 req->rq_status = le32_to_cpu(lcd->lcd_last_close_result);
 389                 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 390         } else {
 391                 req->rq_transno = le64_to_cpu(lcd->lcd_last_transno);
 392                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
 393                 req->rq_status = le32_to_cpu(lcd->lcd_last_result);
 394                 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 395         }
 396         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
 397                   req->rq_transno, req->rq_status);
 398
 399         mds_steal_ack_locks(req);
 400 }
 401
 402 static void reconstruct_reint_setattr(struct mds_update_record *rec,
 403                                       int offset, struct ptlrpc_request *req)
 404 {
 405         struct obd_export *exp = req->rq_export;
 406         struct mds_export_data *med = &exp->exp_mds_data;
 407         struct mds_obd *obd = &exp->exp_obd->u.mds;
 408         struct dentry *de;
 409         struct mds_body *body;
 410
 411         mds_req_from_lcd(req, med->med_lcd);
 412
 413         de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 414         if (IS_ERR(de)) {
 415                 int rc;
 416                 rc = PTR_ERR(de);
 417                 LCONSOLE_WARN("FID "LPU64"/%u lookup error %d."
 418                               " Evicting client %s with export %s.\n",
 419                               rec->ur_fid1->id, rec->ur_fid1->generation, rc,
 420                               obd_uuid2str(&exp->exp_client_uuid),
 421                               obd_export_nid2str(exp));
 422                 mds_export_evict(exp);
 423                 EXIT;
 424                 return;
 425         }
 426
 427         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
 428         mds_pack_inode2fid(&body->fid1, de->d_inode);
 429         mds_pack_inode2body(body, de->d_inode);
 430
 431         /* Don't return OST-specific attributes if we didn't just set them */
 432         if (rec->ur_iattr.ia_valid & ATTR_SIZE)
 433                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 434         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 435                 body->valid |= OBD_MD_FLMTIME;
 436         if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 437                 body->valid |= OBD_MD_FLATIME;
 438
 439         l_dput(de);
 440 }
 441
 442 int mds_osc_setattr_async(struct obd_device *obd, struct inode *inode,
 443                           struct lov_mds_md *lmm, int lmm_size,
 444                           struct llog_cookie *logcookies, struct ll_fid *fid)
 445 {
 446         struct mds_obd *mds = &obd->u.mds;
 447         struct obd_trans_info oti = { 0 };
 448         struct obd_info oinfo = { { { 0 } } };
 449         int rc;
 450         ENTRY;
 451
 452         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
 453                 RETURN(0);
 454
 455         /* first get memory EA */
 456         OBDO_ALLOC(oinfo.oi_oa);
 457         if (!oinfo.oi_oa)
 458                 RETURN(-ENOMEM);
 459
 460         LASSERT(lmm);
 461
 462         rc = obd_unpackmd(mds->mds_osc_exp, &oinfo.oi_md, lmm, lmm_size);
 463         if (rc < 0) {
 464                 CERROR("Error unpack md %p for inode %lu\n", lmm, inode->i_ino);
 465                 GOTO(out, rc);
 466         }
 467
 468         rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, oinfo.oi_md);
 469         if (rc) {
 470                 CERROR("Error revalidate lsm %p \n", oinfo.oi_md);
 471                 GOTO(out, rc);
 472         }
 473
 474         /* then fill oa */
 475         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 476         oinfo.oi_oa->o_valid |= OBD_MD_FLID;
 477         oinfo.oi_oa->o_id = oinfo.oi_md->lsm_object_id;
 478         if (logcookies) {
 479                 oinfo.oi_oa->o_valid |= OBD_MD_FLCOOKIE;
 480                 oti.oti_logcookies = logcookies;
 481         }
 482
 483         LASSERT(fid != NULL);
 484         oinfo.oi_oa->o_fid = fid->id;
 485         oinfo.oi_oa->o_generation = fid->generation;
 486         oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 487
 488         /* do async setattr from mds to ost not waiting for responses. */
 489         rc = obd_setattr_async(mds->mds_osc_exp, &oinfo, &oti, NULL);
 490         if (rc)
 491                 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64
 492                        " on ost error %d\n", oinfo.oi_md->lsm_object_id, rc);
 493 out:
 494         if (oinfo.oi_md)
 495                 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
 496         OBDO_FREE(oinfo.oi_oa);
 497         RETURN(rc);
 498 }
 499
 500 /* In the raw-setattr case, we lock the child inode.
 501  * In the write-back case or if being called from open, the client holds a lock
 502  * already.
 503  *
 504  * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
 505 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
 506                              struct ptlrpc_request *req,
 507                              struct lustre_handle *lh)
 508 {
 509         unsigned int ia_valid = rec->ur_iattr.ia_valid;
 510         struct mds_obd *mds = mds_req2mds(req);
 511         struct obd_device *obd = req->rq_export->exp_obd;
 512         struct mds_body *body;
 513         struct dentry *de;
 514         struct inode *inode = NULL;
 515         struct lustre_handle lockh;
 516         void *handle = NULL;
 517         struct mds_logcancel_data *mlcd = NULL;
 518         struct lov_mds_md *lmm = NULL;
 519         struct llog_cookie *logcookies = NULL;
 520         int lmm_size = 0, need_lock = 1, cookie_size = 0;
 521         int rc = 0, cleanup_phase = 0, err, locked = 0, sync = 0;
 522         unsigned int qcids[MAXQUOTAS] = { 0, 0 };
 523         unsigned int qpids[MAXQUOTAS] = { rec->ur_iattr.ia_uid,
 524                                           rec->ur_iattr.ia_gid };
 525         ENTRY;
 526
 527         LASSERT(offset == REQ_REC_OFF);
 528         offset = REPLY_REC_OFF;
 529
 530         DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
 531                   rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
 532         OBD_COUNTER_INCREMENT(obd, setattr);
 533
 534         MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
 535
 536         if (rec->ur_dlm)
 537                 ldlm_request_cancel(req, rec->ur_dlm, 0);
 538
 539         if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
 540             (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
 541                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 542                 if (IS_ERR(de))
 543                         GOTO(cleanup, rc = PTR_ERR(de));
 544                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
 545                         GOTO(cleanup, rc = -EROFS);
 546         } else {
 547                 __u64 lockpart = MDS_INODELOCK_UPDATE;
 548                 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
 549                         lockpart |= MDS_INODELOCK_LOOKUP;
 550
 551                 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
 552                                            &lockh, NULL, 0, lockpart);
 553                 if (IS_ERR(de))
 554                         GOTO(cleanup, rc = PTR_ERR(de));
 555                 locked = 1;
 556         }
 557
 558         cleanup_phase = 1;
 559         inode = de->d_inode;
 560         LASSERT(inode);
 561
 562         if ((rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) ||
 563             (rec->ur_iattr.ia_valid & ATTR_SIZE)) {
 564                 /* Check write access for the O_TRUNC case */
 565                 if (mds_query_write_access(inode) < 0)
 566                         GOTO(cleanup, rc = -ETXTBSY);
 567         }
 568
 569         /* save uid/gid for quota acq/rel */
 570         qcids[USRQUOTA] = inode->i_uid;
 571         qcids[GRPQUOTA] = inode->i_gid;
 572
 573         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 574             rec->ur_eadata != NULL) {
 575                 LOCK_INODE_MUTEX(inode);
 576                 need_lock = 0;
 577         }
 578
 579         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
 580
 581         /* start a log jounal handle if needed */
 582         if (S_ISREG(inode->i_mode) &&
 583             rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 584                 lmm_size = mds->mds_max_mdsize;
 585                 OBD_ALLOC(lmm, lmm_size);
 586                 if (lmm == NULL)
 587                         GOTO(cleanup, rc = -ENOMEM);
 588
 589                 cleanup_phase = 2;
 590                 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock, 0);
 591                 if (rc < 0)
 592                         GOTO(cleanup, rc);
 593                 rc = 0;
 594
 595                 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
 596                                           le32_to_cpu(lmm->lmm_stripe_count));
 597         } else {
 598                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
 599         }
 600         if (IS_ERR(handle))
 601                 GOTO(cleanup, rc = PTR_ERR(handle));
 602
 603         if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 604                 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
 605                        LTIME_S(rec->ur_iattr.ia_mtime),
 606                        LTIME_S(rec->ur_iattr.ia_ctime));
 607         rc = mds_fix_attr(inode, rec);
 608         if (rc)
 609                 GOTO(cleanup, rc);
 610
 611         if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) {  /* ioctl */
 612                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
 613                                       (long)&rec->ur_flags);
 614         } else if (rec->ur_iattr.ia_valid) {            /* setattr */
 615                 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
 616                 /* journal chown/chgrp in llog, just like unlink */
 617                 if (rc == 0 && lmm_size){
 618                         cookie_size = mds_get_cookie_size(obd, lmm);
 619                         OBD_ALLOC(logcookies, cookie_size);
 620                         if (logcookies == NULL)
 621                                 GOTO(cleanup, rc = -ENOMEM);
 622
 623                         if (mds_log_op_setattr(obd, inode, lmm, lmm_size,
 624                                                logcookies, cookie_size) <= 0) {
 625                                 OBD_FREE(logcookies, cookie_size);
 626                                 logcookies = NULL;
 627                         }
 628                 }
 629         }
 630
 631         if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 632             rec->ur_eadata != NULL) {
 633                 struct lov_stripe_md *lsm = NULL;
 634                 struct lov_user_md *lum = NULL;
 635
 636                 rc = ll_permission(inode, MAY_WRITE, NULL);
 637                 if (rc < 0)
 638                         GOTO(cleanup, rc);
 639
 640                 lum = rec->ur_eadata;
 641                 /* if { size, offset, count } = { 0, -1, 0 } (i.e. all default
 642                  * values specified) then delete default striping from dir. */
 643                 if (S_ISDIR(inode->i_mode) &&
 644                     ((lum->lmm_stripe_size == 0 &&
 645                       lum->lmm_stripe_offset ==
 646                       (typeof(lum->lmm_stripe_offset))(-1) &&
 647                       lum->lmm_stripe_count == 0) ||
 648                     /* lmm_stripe_size == -1 is deprecated in 1.4.6 */
 649                     lum->lmm_stripe_size ==
 650                     (typeof(lum->lmm_stripe_size))(-1))){
 651                         rc = fsfilt_set_md(obd, inode, handle, NULL, 0, "lov");
 652                         if (rc)
 653                                 GOTO(cleanup, rc);
 654                 } else {
 655                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 656                                            mds->mds_osc_exp, 0,
 657                                            &lsm, rec->ur_eadata);
 658                         if (rc)
 659                                 GOTO(cleanup, rc);
 660
 661                         obd_free_memmd(mds->mds_osc_exp, &lsm);
 662
 663                         rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
 664                                            rec->ur_eadatalen, "lov");
 665                         if (rc)
 666                                 GOTO(cleanup, rc);
 667                 }
 668         }
 669
 670         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
 671         mds_pack_inode2fid(&body->fid1, inode);
 672         mds_pack_inode2body(body, inode);
 673
 674         /* don't return OST-specific attributes if we didn't just set them. */
 675         if (ia_valid & ATTR_SIZE)
 676                 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 677         if (ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
 678                 body->valid |= OBD_MD_FLMTIME;
 679         if (ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
 680                 body->valid |= OBD_MD_FLATIME;
 681
 682         if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
 683                 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
 684                           rec->ur_eadatalen);
 685                 if (mlcd) {
 686                         mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
 687                                 rec->ur_eadatalen;
 688                         mlcd->mlcd_eadatalen = rec->ur_eadatalen;
 689                         mlcd->mlcd_cookielen = rec->ur_cookielen;
 690                         mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
 691                                 mlcd->mlcd_cookielen;
 692                         memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
 693                                mlcd->mlcd_cookielen);
 694                         memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
 695                                mlcd->mlcd_eadatalen);
 696                 } else {
 697                         CERROR("unable to allocate log cancel data\n");
 698                 }
 699         }
 700         EXIT;
 701  cleanup:
 702         if (mlcd != NULL)
 703                 sync = fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
 704                                              mds_cancel_cookies_cb, mlcd);
 705         err = mds_finish_transno(mds, inode, handle, req, rc, 0, sync);
 706         /* do mds to ost setattr if needed */
 707         if (!rc && !err && lmm_size)
 708                 mds_osc_setattr_async(obd, inode, lmm, lmm_size,
 709                                       logcookies, rec->ur_fid1);
 710
 711         switch (cleanup_phase) {
 712         case 2:
 713                 OBD_FREE(lmm, mds->mds_max_mdsize);
 714                 if (logcookies)
 715                         OBD_FREE(logcookies, cookie_size);
 716         case 1:
 717                 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
 718                     rec->ur_eadata != NULL)
 719                         UNLOCK_INODE_MUTEX(inode);
 720                 l_dput(de);
 721                 if (locked) {
 722                         if (rc) {
 723                                 ldlm_lock_decref(&lockh, LCK_EX);
 724                         } else {
 725                                 ptlrpc_save_lock (req, &lockh, LCK_EX);
 726                         }
 727                 }
 728         case 0:
 729                 break;
 730         default:
 731                 LBUG();
 732         }
 733         if (err && !rc)
 734                 rc = err;
 735
 736         req->rq_status = rc;
 737
 738         /* trigger dqrel/dqacq for original owner and new owner */
 739         if (ia_valid & (ATTR_UID | ATTR_GID))
 740                 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
 741                               FSFILT_OP_SETATTR);
 742
 743         return 0;
 744 }
 745
 746 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
 747                                      struct ptlrpc_request *req)
 748 {
 749         struct obd_export *exp = req->rq_export;
 750         struct mds_export_data *med = &exp->exp_mds_data;
 751         struct mds_obd *obd = &exp->exp_obd->u.mds;
 752         struct dentry *parent, *child;
 753         struct mds_body *body;
 754         int rc;
 755
 756         mds_req_from_lcd(req, med->med_lcd);
 757
 758         if (req->rq_status)
 759                 return;
 760
 761         parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
 762         if (IS_ERR(parent)) {
 763                 rc = PTR_ERR(parent);
 764                 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
 765                               " Evicting client %s with export %s.\n",
 766                               rec->ur_fid1->id, rec->ur_fid1->generation, rc,
 767                               obd_uuid2str(&exp->exp_client_uuid),
 768                               obd_export_nid2str(exp));
 769                 mds_export_evict(exp);
 770                 EXIT;
 771                 return;
 772         }
 773         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 774         if (IS_ERR(child)) {
 775                 rc = PTR_ERR(child);
 776                 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
 777                               " Evicting client %s with export %s.\n",
 778                               rec->ur_fid1->id, rec->ur_fid1->generation, rc,
 779                               obd_uuid2str(&exp->exp_client_uuid),
 780                               obd_export_nid2str(exp));
 781                 mds_export_evict(exp);
 782                 EXIT;
 783                 return;
 784         }
 785
 786         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
 787         mds_pack_inode2fid(&body->fid1, child->d_inode);
 788         mds_pack_inode2body(body, child->d_inode);
 789
 790         l_dput(parent);
 791         l_dput(child);
 792 }
 793
 794 static int mds_reint_create(struct mds_update_record *rec, int offset,
 795                             struct ptlrpc_request *req,
 796                             struct lustre_handle *lh)
 797 {
 798         struct dentry *dparent = NULL;
 799         struct mds_obd *mds = mds_req2mds(req);
 800         struct obd_device *obd = req->rq_export->exp_obd;
 801         struct dentry *dchild = NULL;
 802         struct inode *dir = NULL;
 803         void *handle = NULL;
 804         struct lustre_handle lockh;
 805         int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
 806         int created = 0;
 807         unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
 808         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
 809         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
 810         int rec_pending = 0;
 811         unsigned int gid = current->fsgid;
 812         ENTRY;
 813
 814         LASSERT(offset == REQ_REC_OFF);
 815         offset = REPLY_REC_OFF;
 816
 817         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
 818                         LUSTRE_MDS_NAME));
 819
 820         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
 821                   rec->ur_fid1->id, rec->ur_fid1->generation,
 822                   rec->ur_name, rec->ur_mode);
 823
 824         MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
 825
 826         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 827                 GOTO(cleanup, rc = -ESTALE);
 828
 829         if (rec->ur_dlm)
 830                 ldlm_request_cancel(req, rec->ur_dlm, 0);
 831
 832         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX, &lockh,
 833                                         rec->ur_name, rec->ur_namelen - 1,
 834                                         MDS_INODELOCK_UPDATE);
 835         if (IS_ERR(dparent)) {
 836                 rc = PTR_ERR(dparent);
 837                 CDEBUG(D_DENTRY, "parent "LPU64"/%u lookup error %d\n",
 838                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 839                 GOTO(cleanup, rc);
 840         }
 841         cleanup_phase = 1; /* locked parent dentry */
 842         dir = dparent->d_inode;
 843         LASSERT(dir);
 844
 845         ldlm_lock_dump_handle(D_OTHER, &lockh);
 846
 847         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 848         if (IS_ERR(dchild)) {
 849                 rc = PTR_ERR(dchild);
 850                 CDEBUG(D_DENTRY, "child lookup error %d\n", rc);
 851                 GOTO(cleanup, rc);
 852         }
 853
 854         cleanup_phase = 2; /* child dentry */
 855
 856         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
 857
 858         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
 859                 if (dchild->d_inode)
 860                         GOTO(cleanup, rc = -EEXIST);
 861                 GOTO(cleanup, rc = -EROFS);
 862         }
 863
 864         if (dir->i_mode & S_ISGID && S_ISDIR(rec->ur_mode))
 865                 rec->ur_mode |= S_ISGID;
 866
 867         dchild->d_fsdata = (void *)&dp;
 868         dp.ldp_inum = (unsigned long)rec->ur_fid2->id;
 869         dp.ldp_ptr = req;
 870
 871         if (dir->i_mode & S_ISGID)
 872                 gid = dir->i_gid;
 873         else
 874                 gid = current->fsgid;
 875
 876         /* we try to get enough quota to write here, and let ldiskfs
 877          * decide if it is out of quota or not b=14783 */
 878         lquota_chkquota(mds_quota_interface_ref, obd,
 879                         current->fsuid, gid, 1, &rec_pending);
 880
 881         switch (type) {
 882         case S_IFREG:{
 883                 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
 884                 if (IS_ERR(handle))
 885                         GOTO(cleanup, rc = PTR_ERR(handle));
 886                 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
 887                 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
 888                 EXIT;
 889                 break;
 890         }
 891         case S_IFDIR:{
 892                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
 893                 if (IS_ERR(handle))
 894                         GOTO(cleanup, rc = PTR_ERR(handle));
 895                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 896                 mds_counter_incr(req->rq_export, LPROC_MDS_MKDIR);
 897                 EXIT;
 898                 break;
 899         }
 900         case S_IFLNK:{
 901                 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
 902                 if (IS_ERR(handle))
 903                         GOTO(cleanup, rc = PTR_ERR(handle));
 904                 if (rec->ur_tgt == NULL)        /* no target supplied */
 905                         rc = -EINVAL;           /* -EPROTO? */
 906                 else
 907                         rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
 908                 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
 909                 EXIT;
 910                 break;
 911         }
 912         case S_IFCHR:
 913         case S_IFBLK:
 914         case S_IFIFO:
 915         case S_IFSOCK:{
 916                 int rdev = rec->ur_rdev;
 917                 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
 918                 if (IS_ERR(handle))
 919                         GOTO(cleanup, rc = PTR_ERR(handle));
 920                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 921                 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
 922                 EXIT;
 923                 break;
 924         }
 925         default:
 926                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 927                 dchild->d_fsdata = NULL;
 928                 GOTO(cleanup, rc = -EINVAL);
 929         }
 930
 931         /* In case we stored the desired inum in here, we want to clean up. */
 932         if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
 933                 dchild->d_fsdata = NULL;
 934
 935         if (rc) {
 936                 CDEBUG(D_INODE, "error during create: %d\n", rc);
 937                 GOTO(cleanup, rc);
 938         } else {
 939                 struct iattr iattr;
 940                 struct inode *inode = dchild->d_inode;
 941                 struct mds_body *body;
 942
 943                 created = 1;
 944                 LTIME_S(iattr.ia_atime) = rec->ur_time;
 945                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
 946                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
 947                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
 948                 iattr.ia_gid = gid;
 949                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 950                         ATTR_MTIME | ATTR_CTIME;
 951
 952                 if (rec->ur_fid2->id) {
 953                         LASSERT(rec->ur_fid2->id == inode->i_ino);
 954                         inode->i_generation = rec->ur_fid2->generation;
 955                         /* Dirtied and committed by the upcoming setattr. */
 956                         CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
 957                                inode->i_ino, inode->i_generation);
 958                 } else {
 959                         CDEBUG(D_INODE, "created ino %lu with gen %x\n",
 960                                inode->i_ino, inode->i_generation);
 961                 }
 962
 963                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
 964                 if (rc)
 965                         CERROR("error on child setattr: rc = %d\n", rc);
 966
 967                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
 968                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
 969                 if (rc)
 970                         CERROR("error on parent setattr: rc = %d\n", rc);
 971
 972                 if (S_ISDIR(inode->i_mode)) {
 973                         struct lov_mds_md lmm;
 974                         int lmm_size = sizeof(lmm);
 975                         rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0);
 976                         if (rc > 0) {
 977                                 LOCK_INODE_MUTEX(inode);
 978                                 rc = fsfilt_set_md(obd, inode, handle,
 979                                                    &lmm, lmm_size, "lov");
 980                                 UNLOCK_INODE_MUTEX(inode);
 981                         }
 982                         if (rc)
 983                                 CERROR("error on copy stripe info: rc = %d\n",
 984                                         rc);
 985                 }
 986
 987                 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
 988                 mds_pack_inode2fid(&body->fid1, inode);
 989                 mds_pack_inode2body(body, inode);
 990         }
 991         EXIT;
 992
 993 cleanup:
 994         err = mds_finish_transno(mds, dir, handle, req, rc, 0, 0);
 995         if (rec_pending)
 996                 lquota_pending_commit(mds_quota_interface_ref, obd,
 997                                       current->fsuid, gid, 1);
 998
 999         if (rc && created) {
1000                 /* Destroy the file we just created.  This should not need
1001                  * extra journal credits, as we have already modified all of
1002                  * the blocks needed in order to create the file in the first
1003                  * place.
1004                  */
1005                 switch (type) {
1006                 case S_IFDIR:
1007                         err = vfs_rmdir(dir, dchild);
1008                         if (err)
1009                                 CERROR("rmdir in error path: %d\n", err);
1010                         break;
1011                 default:
1012                         err = vfs_unlink(dir, dchild);
1013                         if (err)
1014                                 CERROR("unlink in error path: %d\n", err);
1015                         break;
1016                 }
1017         } else if (created) {
1018                 /* The inode we were allocated may have just been freed
1019                  * by an unlink operation.  We take this lock to
1020                  * synchronize against the matching reply-ack-lock taken
1021                  * in unlink, to avoid replay problems if this reply
1022                  * makes it out to the client but the unlink's does not.
1023                  * See bug 2029 for more detail.*/
1024                 mds_lock_new_child(obd, dchild->d_inode, NULL);
1025                 /* save uid/gid of create inode and parent */
1026                 qpids[USRQUOTA] = dir->i_uid;
1027                 qpids[GRPQUOTA] = dir->i_gid;
1028         } else {
1029                 rc = err;
1030         }
1031
1032         switch (cleanup_phase) {
1033         case 2: /* child dentry */
1034                 l_dput(dchild);
1035         case 1: /* locked parent dentry */
1036                 if (rc) {
1037                         ldlm_lock_decref(&lockh, LCK_EX);
1038                 } else {
1039                         ptlrpc_save_lock (req, &lockh, LCK_EX);
1040                 }
1041                 l_dput(dparent);
1042         case 0:
1043                 break;
1044         default:
1045                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1046                 LBUG();
1047         }
1048         req->rq_status = rc;
1049
1050         /* trigger dqacq on the owner of child and parent */
1051         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1052                       FSFILT_OP_CREATE);
1053         return 0;
1054 }
1055
1056 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1057            ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1058 {
1059         int i;
1060
1061         for (i = 0; i < RES_NAME_SIZE; i++) {
1062                 /* return 1 here, because enqueue_ordered will skip resources
1063                  * of all zeroes if they're sorted to the end of the list. */
1064                 if (res1->name[i] == 0 && res2->name[i] != 0)
1065                         return 1;
1066                 if (res2->name[i] == 0 && res1->name[i] != 0)
1067                         return 0;
1068
1069                 if (res1->name[i] > res2->name[i])
1070                         return 1;
1071                 if (res1->name[i] < res2->name[i])
1072                         return 0;
1073         }
1074         if (!p1 || !p2)
1075                 return 0;
1076         if (memcmp(p1, p2, sizeof(*p1)) < 0)
1077                 return 1;
1078         return 0;
1079 }
1080
1081 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1082  * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1083  * because they take the place of local semaphores.
1084  *
1085  * One or two locks are taken in numerical order.  A res_id->name[0] of 0 means
1086  * no lock is taken for that res_id.  Must be at least one non-zero res_id. */
1087 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1088                           struct lustre_handle *p1_lockh, int p1_lock_mode,
1089                           ldlm_policy_data_t *p1_policy,
1090                           struct ldlm_res_id *p2_res_id,
1091                           struct lustre_handle *p2_lockh, int p2_lock_mode,
1092                           ldlm_policy_data_t *p2_policy)
1093 {
1094         struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1095         struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1096         int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1097         ldlm_policy_data_t *policies[2] = {p1_policy, p2_policy};
1098         int rc, flags;
1099         ENTRY;
1100
1101         LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1102
1103         CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1104                res_id[0]->name[0], res_id[1]->name[0]);
1105
1106         if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1107                 handles[1] = p1_lockh;
1108                 handles[0] = p2_lockh;
1109                 res_id[1] = p1_res_id;
1110                 res_id[0] = p2_res_id;
1111                 lock_modes[1] = p1_lock_mode;
1112                 lock_modes[0] = p2_lock_mode;
1113                 policies[1] = p1_policy;
1114                 policies[0] = p2_policy;
1115         }
1116
1117         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1118                res_id[0]->name[0], res_id[1]->name[0]);
1119
1120         flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1121         rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[0],
1122                                     LDLM_IBITS, policies[0], lock_modes[0],
1123                                     &flags, ldlm_blocking_ast,
1124                                     ldlm_completion_ast, NULL, NULL, 0,
1125                                     NULL, handles[0]);
1126         if (rc != ELDLM_OK)
1127                 RETURN(-EIO);
1128         ldlm_lock_dump_handle(D_OTHER, handles[0]);
1129
1130         if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0 &&
1131             (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1132                 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1133                 ldlm_lock_addref(handles[1], lock_modes[1]);
1134         } else if (res_id[1]->name[0] != 0) {
1135                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1136                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[1],
1137                                             LDLM_IBITS, policies[1],
1138                                             lock_modes[1], &flags,
1139                                             ldlm_blocking_ast,
1140                                             ldlm_completion_ast, NULL, NULL,
1141                                             0, NULL, handles[1]);
1142                 if (rc != ELDLM_OK) {
1143                         ldlm_lock_decref(handles[0], lock_modes[0]);
1144                         RETURN(-EIO);
1145                 }
1146                 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1147         }
1148
1149         RETURN(0);
1150 }
1151
1152 static inline int res_eq(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
1153 {
1154         return !memcmp(res1, res2, sizeof(*res1));
1155 }
1156
1157 static inline void
1158 try_to_aggregate_locks(struct ldlm_res_id *res1, ldlm_policy_data_t *p1,
1159                         struct ldlm_res_id *res2, ldlm_policy_data_t *p2)
1160 {
1161         if (!res_eq(res1, res2))
1162                 return;
1163         /* XXX: any additional inodebits (to current LOOKUP and UPDATE)
1164          * should be taken with great care here */
1165         p1->l_inodebits.bits |= p2->l_inodebits.bits;
1166 }
1167
1168 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1169                            struct lustre_handle *p1_lockh, int p1_lock_mode,
1170                            ldlm_policy_data_t *p1_policy,
1171                            struct ldlm_res_id *p2_res_id,
1172                            struct lustre_handle *p2_lockh, int p2_lock_mode,
1173                            ldlm_policy_data_t *p2_policy,
1174                            struct ldlm_res_id *c1_res_id,
1175                            struct lustre_handle *c1_lockh, int c1_lock_mode,
1176                            ldlm_policy_data_t *c1_policy,
1177                            struct ldlm_res_id *c2_res_id,
1178                            struct lustre_handle *c2_lockh, int c2_lock_mode,
1179                            ldlm_policy_data_t *c2_policy)
1180 {
1181         struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1182                                           c1_res_id, c2_res_id };
1183         struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1184                                                  c1_lockh, c2_lockh };
1185         int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1186                               c1_lock_mode, c2_lock_mode };
1187         ldlm_policy_data_t *policies[5] = {p1_policy, p2_policy,
1188                                            c1_policy, c2_policy};
1189         int rc, i, j, sorted, flags;
1190         ENTRY;
1191
1192         CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1193                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1194                res_id[3]->name[0]);
1195
1196         /* simple insertion sort - we have at most 4 elements */
1197         for (i = 1; i < 4; i++) {
1198                 j = i - 1;
1199                 dlm_handles[4] = dlm_handles[i];
1200                 res_id[4] = res_id[i];
1201                 lock_modes[4] = lock_modes[i];
1202                 policies[4] = policies[i];
1203
1204                 sorted = 0;
1205                 do {
1206                         if (res_gt(res_id[j], res_id[4], policies[j],
1207                                    policies[4])) {
1208                                 dlm_handles[j + 1] = dlm_handles[j];
1209                                 res_id[j + 1] = res_id[j];
1210                                 lock_modes[j + 1] = lock_modes[j];
1211                                 policies[j + 1] = policies[j];
1212                                 j--;
1213                         } else {
1214                                 sorted = 1;
1215                         }
1216                 } while (j >= 0 && !sorted);
1217
1218                 dlm_handles[j + 1] = dlm_handles[4];
1219                 res_id[j + 1] = res_id[4];
1220                 lock_modes[j + 1] = lock_modes[4];
1221                 policies[j + 1] = policies[4];
1222         }
1223
1224         CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1225                res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1226                res_id[3]->name[0]);
1227
1228         /* XXX we could send ASTs on all these locks first before blocking? */
1229         for (i = 0; i < 4; i++) {
1230                 flags = LDLM_FL_ATOMIC_CB;
1231                 if (res_id[i]->name[0] == 0)
1232                         break;
1233                 if (i && res_eq(res_id[i], res_id[i-1])) {
1234                         memcpy(dlm_handles[i], dlm_handles[i-1],
1235                                sizeof(*(dlm_handles[i])));
1236                         ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1237                 } else {
1238                         /* we need to enqueue locks with different inodebits
1239                          * at once, because otherwise concurrent thread can
1240                          * hit the windown between these two locks and we'll
1241                          * get to deadlock. see bug 10360. note also, that it
1242                          * is impossible to have >2 equal res. */
1243                         if (i < 3)
1244                                 try_to_aggregate_locks(res_id[i], policies[i],
1245                                                        res_id[i+1], policies[i+1]);
1246                         rc = ldlm_cli_enqueue_local(obd->obd_namespace,
1247                                                     res_id[i], LDLM_IBITS,
1248                                                     policies[i], lock_modes[i],
1249                                                     &flags, ldlm_blocking_ast,
1250                                                     ldlm_completion_ast, NULL,
1251                                                     NULL, 0, NULL,
1252                                                     dlm_handles[i]);
1253                         if (rc != ELDLM_OK)
1254                                 GOTO(out_err, rc = -EIO);
1255                         ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1256                 }
1257         }
1258
1259         RETURN(0);
1260 out_err:
1261         while (i-- > 0)
1262                 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1263
1264         return rc;
1265 }
1266
1267 /* In the unlikely case that the child changed while we were waiting
1268  * on the lock, we need to drop the lock on the old child and either:
1269  * - if the child has a lower resource name, then we have to also
1270  *   drop the parent lock and regain the locks in the right order
1271  * - in the rename case, if the child has a lower resource name than one of
1272  *   the other parent/child resources (maxres) we also need to reget the locks
1273  * - if the child has a higher resource name (this is the common case)
1274  *   we can just get the lock on the new child (still in lock order)
1275  *
1276  * Returns 0 if the child did not change or if it changed but could be locked.
1277  * Returns 1 if the child changed and we need to re-lock (no locks held).
1278  * Returns -ve error with a valid dchild (no locks held). */
1279 static int mds_verify_child(struct obd_device *obd,
1280                             struct ldlm_res_id *parent_res_id,
1281                             struct lustre_handle *parent_lockh,
1282                             struct dentry *dparent, int parent_mode,
1283                             struct ldlm_res_id *child_res_id,
1284                             struct lustre_handle *child_lockh,
1285                             struct dentry **dchildp, int child_mode,
1286                             ldlm_policy_data_t *child_policy,
1287                             const char *name, int namelen,
1288                             struct ldlm_res_id *maxres)
1289 {
1290         struct dentry *vchild, *dchild = *dchildp;
1291         int rc = 0, cleanup_phase = 2; /* parent, child locks */
1292         ENTRY;
1293
1294         vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1295         if (IS_ERR(vchild))
1296                 GOTO(cleanup, rc = PTR_ERR(vchild));
1297
1298         if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1299                    (vchild->d_inode != NULL &&
1300                     child_res_id->name[0] == vchild->d_inode->i_ino &&
1301                     child_res_id->name[1] == vchild->d_inode->i_generation))) {
1302                 if (dchild != NULL)
1303                         l_dput(dchild);
1304                 *dchildp = vchild;
1305
1306                 RETURN(0);
1307         }
1308
1309         CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1310                vchild->d_inode, dchild ? dchild->d_inode : 0,
1311                vchild->d_inode ? vchild->d_inode->i_ino : 0,
1312                child_res_id->name[0]);
1313         if (child_res_id->name[0] != 0)
1314                 ldlm_lock_decref(child_lockh, child_mode);
1315         if (dchild)
1316                 l_dput(dchild);
1317
1318         cleanup_phase = 1; /* parent lock only */
1319         *dchildp = dchild = vchild;
1320
1321         if (dchild->d_inode) {
1322                 int flags = LDLM_FL_ATOMIC_CB;
1323                 child_res_id->name[0] = dchild->d_inode->i_ino;
1324                 child_res_id->name[1] = dchild->d_inode->i_generation;
1325
1326                 /* Make sure that we don't try to re-enqueue a lock on the
1327                  * same resource if it happens that the source is renamed to
1328                  * the target by another thread (bug 9974, thanks racer :-) */
1329                 if (!res_gt(child_res_id, parent_res_id, NULL, NULL) ||
1330                     !res_gt(child_res_id, maxres, NULL, NULL)) {
1331                         CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1332                                child_res_id->name[0], parent_res_id->name[0],
1333                                maxres->name[0]);
1334                         GOTO(cleanup, rc = 1);
1335                 }
1336
1337                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1338                                             LDLM_IBITS, child_policy,
1339                                             child_mode, &flags,
1340                                             ldlm_blocking_ast,
1341                                             ldlm_completion_ast, NULL,
1342                                             NULL, 0, NULL, child_lockh);
1343                 if (rc != ELDLM_OK)
1344                         GOTO(cleanup, rc = -EIO);
1345         } else {
1346                 memset(child_res_id, 0, sizeof(*child_res_id));
1347         }
1348
1349         EXIT;
1350 cleanup:
1351         if (rc) {
1352                 switch(cleanup_phase) {
1353                 case 2:
1354                         if (child_res_id->name[0] != 0)
1355                                 ldlm_lock_decref(child_lockh, child_mode);
1356                 case 1:
1357                         ldlm_lock_decref(parent_lockh, parent_mode);
1358                 }
1359         }
1360         return rc;
1361 }
1362
1363 #define INODE_CTIME_AGE (10)
1364 #define INODE_CTIME_OLD(inode) (LTIME_S(inode->i_ctime) +               \
1365                                 INODE_CTIME_AGE < CURRENT_SECONDS)
1366
1367 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1368                                 struct ll_fid *fid,
1369                                 struct lustre_handle *parent_lockh,
1370                                 struct dentry **dparentp, int parent_mode,
1371                                 __u64 parent_lockpart,
1372                                 char *name, int namelen,
1373                                 struct lustre_handle *child_lockh,
1374                                 struct dentry **dchildp, int child_mode,
1375                                 __u64 child_lockpart)
1376 {
1377         struct ldlm_res_id child_res_id = { .name = {0} };
1378         struct ldlm_res_id parent_res_id = { .name = {0} };
1379         ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1380         ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1381         struct inode *inode;
1382         int rc = 0, cleanup_phase = 0;
1383         ENTRY;
1384
1385         /* Step 1: Lookup parent */
1386         *dparentp = mds_fid2dentry(mds, fid, NULL);
1387         if (IS_ERR(*dparentp)) {
1388                 rc = PTR_ERR(*dparentp);
1389                 *dparentp = NULL;
1390                 RETURN(rc);
1391         }
1392
1393         CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1394                (*dparentp)->d_inode->i_ino, name);
1395
1396         parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1397         parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1398
1399         cleanup_phase = 1; /* parent dentry */
1400
1401         /* Step 2: Lookup child (without DLM lock, to get resource name) */
1402         *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1403         if (IS_ERR(*dchildp)) {
1404                 rc = PTR_ERR(*dchildp);
1405                 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1406                 GOTO(cleanup, rc);
1407         }
1408
1409         cleanup_phase = 2; /* child dentry */
1410         inode = (*dchildp)->d_inode;
1411         if (inode != NULL) {
1412                 if (is_bad_inode(inode)) {
1413                         CERROR("bad inode returned %lu/%u\n",
1414                                inode->i_ino, inode->i_generation);
1415                         GOTO(cleanup, rc = -ENOENT);
1416                 }
1417                 inode = igrab(inode);
1418         }
1419         if (inode == NULL)
1420                 goto retry_locks;
1421
1422         child_res_id.name[0] = inode->i_ino;
1423         child_res_id.name[1] = inode->i_generation;
1424
1425         /* If we want a LCK_CR for a directory, and this directory has not been
1426            changed for some time, we return not only a LOOKUP lock, but also an
1427            UPDATE lock to have negative dentry starts working for this dir.
1428            Also we apply same logic to non-directories. If the file is rarely
1429            changed - we return both locks and this might save us RPC on
1430            later STAT. */
1431         if ((child_mode & (LCK_CR|LCK_PR|LCK_CW)) && INODE_CTIME_OLD(inode))
1432                 child_policy.l_inodebits.bits |= MDS_INODELOCK_UPDATE;
1433
1434         iput(inode);
1435
1436 retry_locks:
1437         cleanup_phase = 2; /* child dentry */
1438
1439         /* Step 3: Lock parent and child in resource order.  If child doesn't
1440          *         exist, we still have to lock the parent and re-lookup. */
1441         rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1442                                    &parent_policy,
1443                                    &child_res_id, child_lockh, child_mode,
1444                                    &child_policy);
1445         if (rc)
1446                 GOTO(cleanup, rc);
1447
1448         if (!(*dchildp)->d_inode)
1449                 cleanup_phase = 3; /* parent lock */
1450         else
1451                 cleanup_phase = 4; /* child lock */
1452
1453         /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1454         rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1455                               parent_mode, &child_res_id, child_lockh, dchildp,
1456                               child_mode,&child_policy, name, namelen, &parent_res_id);
1457         if (rc > 0)
1458                 goto retry_locks;
1459         if (rc < 0) {
1460                 cleanup_phase = 2;
1461                 GOTO(cleanup, rc);
1462         }
1463
1464 cleanup:
1465         if (rc) {
1466                 switch (cleanup_phase) {
1467                 case 4:
1468                         ldlm_lock_decref(child_lockh, child_mode);
1469                 case 3:
1470                         ldlm_lock_decref(parent_lockh, parent_mode);
1471                 case 2:
1472                         l_dput(*dchildp);
1473                 case 1:
1474                         l_dput(*dparentp);
1475                 default: ;
1476                 }
1477         }
1478         return rc;
1479 }
1480
1481 void mds_reconstruct_generic(struct ptlrpc_request *req)
1482 {
1483         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1484
1485         mds_req_from_lcd(req, med->med_lcd);
1486 }
1487
1488 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1489  * we instead link the inode into the PENDING directory until it is
1490  * finally released.  We can't simply call mds_reint_rename() or some
1491  * part thereof, because we don't have the inode to check for link
1492  * count/open status until after it is locked.
1493  *
1494  * For lock ordering, caller must get child->i_mutex first, then
1495  * pending->i_mutex before starting journal transaction.
1496  *
1497  * returns 1 on success
1498  * returns 0 if we lost a race and didn't make a new link
1499  * returns negative on error
1500  */
1501 static int mds_orphan_add_link(struct mds_update_record *rec,
1502                                struct obd_device *obd, struct dentry *dentry)
1503 {
1504         struct mds_obd *mds = &obd->u.mds;
1505         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1506         struct inode *inode = dentry->d_inode;
1507         struct dentry *pending_child;
1508         char fidname[LL_FID_NAMELEN];
1509         int fidlen = 0, rc, mode;
1510         ENTRY;
1511
1512         LASSERT(inode != NULL);
1513         LASSERT(!mds_inode_is_orphan(inode));
1514 #ifndef HAVE_I_ALLOC_SEM
1515         LASSERT(TRYLOCK_INODE_MUTEX(inode) == 0);
1516 #endif
1517         LASSERT(TRYLOCK_INODE_MUTEX(pending_dir) == 0);
1518
1519         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1520
1521         CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1522                mds_orphan_open_count(inode), inode->i_nlink,
1523                S_ISDIR(inode->i_mode) ? "dir" :
1524                 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1525
1526         if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1527                 RETURN(0);
1528
1529         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1530         if (IS_ERR(pending_child))
1531                 RETURN(PTR_ERR(pending_child));
1532
1533         if (pending_child->d_inode != NULL) {
1534                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1535                 LASSERT(pending_child->d_inode == inode);
1536                 GOTO(out_dput, rc = 0);
1537         }
1538
1539         /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1540          * for linking and return real mode back then -bzzz */
1541         mode = inode->i_mode;
1542         inode->i_mode = S_IFREG;
1543         rc = vfs_link(dentry, pending_dir, pending_child);
1544         if (rc)
1545                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1546                        rec->ur_name, rc);
1547         else
1548                 mds_inode_set_orphan(inode);
1549
1550         /* return mode and correct i_nlink if inode is directory */
1551         inode->i_mode = mode;
1552         LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1553                  S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1554                  inode->i_nlink);
1555         if (S_ISDIR(mode)) {
1556                 inode->i_nlink++;
1557                 pending_dir->i_nlink++;
1558                 mark_inode_dirty(inode);
1559                 mark_inode_dirty(pending_dir);
1560         }
1561
1562         GOTO(out_dput, rc = 1);
1563 out_dput:
1564         l_dput(pending_child);
1565         RETURN(rc);
1566 }
1567
1568 int mds_get_cookie_size(struct obd_device *obd, struct lov_mds_md *lmm)
1569 {
1570         int count = le32_to_cpu(lmm->lmm_stripe_count);
1571         int real_csize = count * sizeof(struct llog_cookie);
1572         return real_csize;
1573 }
1574
1575 void mds_shrink_reply(struct obd_device *obd, struct ptlrpc_request *req,
1576                       struct mds_body *body, int md_off)
1577 {
1578         int cookie_size = 0, md_size = 0;
1579
1580         if (body && body->valid & OBD_MD_FLEASIZE) {
1581                 md_size = body->eadatasize;
1582         }
1583         if (body && body->valid & OBD_MD_FLCOOKIE) {
1584                 LASSERT(body->valid & OBD_MD_FLEASIZE);
1585                 cookie_size = mds_get_cookie_size(obd, lustre_msg_buf(
1586                                                   req->rq_repmsg, md_off, 0));
1587         }
1588
1589         CDEBUG(D_INFO, "Shrink to md_size %d cookie_size %d \n", md_size,
1590                cookie_size);
1591
1592         lustre_shrink_reply(req, md_off, md_size, 1);
1593
1594         lustre_shrink_reply(req, md_off + (md_size > 0), cookie_size, 0);
1595 }
1596
1597 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1598                             struct ptlrpc_request *req,
1599                             struct lustre_handle *lh)
1600 {
1601         struct dentry *dparent = NULL, *dchild;
1602         struct mds_obd *mds = mds_req2mds(req);
1603         struct obd_device *obd = req->rq_export->exp_obd;
1604         struct mds_body *body = NULL;
1605         struct inode *child_inode = NULL;
1606         struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1607         void *handle = NULL;
1608         int rc = 0, cleanup_phase = 0;
1609         unsigned int qcids[MAXQUOTAS] = { 0, 0 };
1610         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1611         ENTRY;
1612
1613         LASSERT(offset == REQ_REC_OFF); /*  || offset == DLM_INTENT_REC_OFF); */
1614         offset = REPLY_REC_OFF;
1615
1616         DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1617                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1618
1619         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1620
1621         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1622                 GOTO(cleanup, rc = -ENOENT);
1623
1624         if (rec->ur_dlm)
1625                 ldlm_request_cancel(req, rec->ur_dlm, 0);
1626
1627         rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1628                                          &parent_lockh, &dparent, LCK_EX,
1629                                          MDS_INODELOCK_UPDATE,
1630                                          rec->ur_name, rec->ur_namelen,
1631                                          &child_lockh, &dchild, LCK_EX,
1632                                          MDS_INODELOCK_FULL);
1633         if (rc)
1634                 GOTO(cleanup, rc);
1635
1636         cleanup_phase = 1; /* dchild, dparent, locks */
1637
1638         dget(dchild);
1639         child_inode = dchild->d_inode;
1640         if (child_inode == NULL) {
1641                 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1642                        dparent->d_inode->i_ino, rec->ur_name);
1643                 GOTO(cleanup, rc = -ENOENT);
1644         }
1645
1646         /* save uid/gid for quota acquire/release */
1647         qcids[USRQUOTA] = child_inode->i_uid;
1648         qcids[GRPQUOTA] = child_inode->i_gid;
1649         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1650         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1651
1652         cleanup_phase = 2; /* dchild has a lock */
1653
1654         /* We have to do these checks ourselves, in case we are making an
1655          * orphan.  The client tells us whether rmdir() or unlink() was called,
1656          * so we need to return appropriate errors (bug 72). */
1657         if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1658                 if (!S_ISDIR(child_inode->i_mode))
1659                         GOTO(cleanup, rc = -ENOTDIR);
1660         } else {
1661                 if (S_ISDIR(child_inode->i_mode))
1662                         GOTO(cleanup, rc = -EISDIR);
1663         }
1664
1665         /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1666         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1667                 GOTO(cleanup, rc = -EROFS);
1668
1669         /* Step 3: Get a lock on the ino to sync with creation WRT inode
1670          * reuse (see bug 2029). */
1671         rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1672         if (rc != ELDLM_OK)
1673                 GOTO(cleanup, rc);
1674
1675         cleanup_phase = 3; /* child inum lock */
1676
1677         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1678
1679         /* ldlm_reply in buf[0] if called via intent */
1680         if (offset == DLM_INTENT_REC_OFF)
1681                 offset = DLM_REPLY_REC_OFF;
1682
1683         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1684         LASSERT(body != NULL);
1685
1686         /* child orphan sem protects orphan_dec_test && is_orphan race */
1687         MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1688         cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1689
1690         /* If this is potentially the last reference to this inode, get the
1691          * OBD EA data first so the client can destroy OST objects.  We
1692          * only do the object removal later if no open files/links remain. */
1693         if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1694             child_inode->i_nlink == 1) {
1695                 if (mds_orphan_open_count(child_inode) > 0) {
1696                         /* need to lock pending_dir before transaction */
1697                         LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1698                         cleanup_phase = 5; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
1699                 } else if (S_ISREG(child_inode->i_mode)) {
1700                         mds_pack_inode2fid(&body->fid1, child_inode);
1701                         mds_pack_inode2body(body, child_inode);
1702                         mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1703                                     child_inode, MDS_PACK_MD_LOCK, 0);
1704                 }
1705         }
1706
1707         /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1708         switch (child_inode->i_mode & S_IFMT) {
1709         case S_IFDIR:
1710                 /* Drop any lingering child directories before we start our
1711                  * transaction, to avoid doing multiple inode dirty/delete
1712                  * in our compound transaction (bug 1321). */
1713                 shrink_dcache_parent(dchild);
1714                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1715                                       NULL);
1716                 if (IS_ERR(handle))
1717                         GOTO(cleanup, rc = PTR_ERR(handle));
1718                 rc = vfs_rmdir(dparent->d_inode, dchild);
1719                 mds_counter_incr(req->rq_export, LPROC_MDS_RMDIR);
1720                 break;
1721         case S_IFREG: {
1722                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1723                                                         offset + 1, 0);
1724                 handle = fsfilt_start_log(obd, dparent->d_inode,
1725                                           FSFILT_OP_UNLINK, NULL,
1726                                           le32_to_cpu(lmm->lmm_stripe_count));
1727                 if (IS_ERR(handle))
1728                         GOTO(cleanup, rc = PTR_ERR(handle));
1729                 rc = vfs_unlink(dparent->d_inode, dchild);
1730                 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
1731                 break;
1732         }
1733         case S_IFLNK:
1734         case S_IFCHR:
1735         case S_IFBLK:
1736         case S_IFIFO:
1737         case S_IFSOCK:
1738                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1739                                       NULL);
1740                 if (IS_ERR(handle))
1741                         GOTO(cleanup, rc = PTR_ERR(handle));
1742                 rc = vfs_unlink(dparent->d_inode, dchild);
1743                 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
1744                 break;
1745         default:
1746                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1747                        rec->ur_name);
1748                 LBUG();
1749                 GOTO(cleanup, rc = -EINVAL);
1750         }
1751
1752         if (rc == 0 && child_inode->i_nlink == 0) {
1753                 if (mds_orphan_open_count(child_inode) > 0)
1754                         rc = mds_orphan_add_link(rec, obd, dchild);
1755
1756                 if (rc == 1)
1757                         GOTO(cleanup, rc = 0);
1758
1759                 if (!S_ISREG(child_inode->i_mode))
1760                         GOTO(cleanup, rc);
1761
1762                 if (!(body->valid & OBD_MD_FLEASIZE)) {
1763                         body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1764                                        OBD_MD_FLATIME | OBD_MD_FLMTIME);
1765                 } else if (mds_log_op_unlink(obd,
1766                                 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1767                                 lustre_msg_buflen(req->rq_repmsg, offset + 1),
1768                                 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1769                                 lustre_msg_buflen(req->rq_repmsg, offset+2)) >
1770                            0) {
1771                         body->valid |= OBD_MD_FLCOOKIE;
1772                 }
1773         }
1774
1775         GOTO(cleanup, rc);
1776 cleanup:
1777         if (rc == 0) {
1778                 struct iattr iattr;
1779                 int err;
1780
1781                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1782                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1783                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1784
1785                 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1786                 if (err)
1787                         CERROR("error on parent setattr: rc = %d\n", err);
1788         }
1789
1790         rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
1791                                 handle, req, rc, 0, 0);
1792         if (!rc)
1793                 (void)obd_set_info_async(mds->mds_osc_exp, sizeof(KEY_UNLINKED),
1794                                          KEY_UNLINKED, 0, NULL, NULL);
1795         switch(cleanup_phase) {
1796         case 5: /* pending_dir semaphore */
1797                 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1798         case 4: /* child inode semaphore */
1799                 MDS_UP_READ_ORPHAN_SEM(child_inode);
1800         case 3: /* child ino-reuse lock */
1801                 if (rc && body != NULL) {
1802                         // Don't unlink the OST objects if the MDS unlink failed
1803                         body->valid = 0;
1804                 }
1805                 if (rc)
1806                         ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1807                 else
1808                         ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1809         case 2: /* child lock */
1810                 ldlm_lock_decref(&child_lockh, LCK_EX);
1811         case 1: /* child and parent dentry, parent lock */
1812                 if (rc)
1813                         ldlm_lock_decref(&parent_lockh, LCK_EX);
1814                 else
1815                         ptlrpc_save_lock(req, &parent_lockh, LCK_EX);
1816                 l_dput(dchild);
1817                 l_dput(dchild);
1818                 l_dput(dparent);
1819         case 0:
1820                 break;
1821         default:
1822                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1823                 LBUG();
1824         }
1825         req->rq_status = rc;
1826
1827         mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1828
1829         /* trigger dqrel on the owner of child and parent */
1830         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1831                       FSFILT_OP_UNLINK);
1832         return 0;
1833 }
1834
1835 static int mds_reint_link(struct mds_update_record *rec, int offset,
1836                           struct ptlrpc_request *req,
1837                           struct lustre_handle *lh)
1838 {
1839         struct obd_device *obd = req->rq_export->exp_obd;
1840         struct dentry *de_src = NULL;
1841         struct dentry *de_tgt_dir = NULL;
1842         struct dentry *dchild = NULL;
1843         struct mds_obd *mds = mds_req2mds(req);
1844         struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1845         struct ldlm_res_id src_res_id = { .name = {0} };
1846         struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1847         ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1848         ldlm_policy_data_t tgt_dir_policy =
1849                                        {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1850         int rc = 0, cleanup_phase = 0;
1851         ENTRY;
1852
1853         LASSERT(offset == REQ_REC_OFF);
1854
1855         DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1856                   rec->ur_fid1->id, rec->ur_fid1->generation,
1857                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1858
1859         mds_counter_incr(req->rq_export, LPROC_MDS_LINK);
1860
1861         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1862
1863         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1864                 GOTO(cleanup, rc = -ENOENT);
1865
1866         if (rec->ur_dlm)
1867                 ldlm_request_cancel(req, rec->ur_dlm, 0);
1868
1869         /* Step 1: Lookup the source inode and target directory by FID */
1870         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1871         if (IS_ERR(de_src))
1872                 GOTO(cleanup, rc = PTR_ERR(de_src));
1873
1874         cleanup_phase = 1; /* source dentry */
1875
1876         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1877         if (IS_ERR(de_tgt_dir)) {
1878                 rc = PTR_ERR(de_tgt_dir);
1879                 de_tgt_dir = NULL;
1880                 GOTO(cleanup, rc);
1881         }
1882
1883         cleanup_phase = 2; /* target directory dentry */
1884
1885         CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
1886                de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1887                de_src->d_inode->i_ino);
1888
1889         /* Step 2: Take the two locks */
1890         src_res_id.name[0] = de_src->d_inode->i_ino;
1891         src_res_id.name[1] = de_src->d_inode->i_generation;
1892         tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1893         tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1894
1895         rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1896                                    &src_policy,
1897                                    &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
1898                                    &tgt_dir_policy);
1899         if (rc)
1900                 GOTO(cleanup, rc);
1901
1902         cleanup_phase = 3; /* locks */
1903
1904         if (mds_inode_is_orphan(de_src->d_inode)) {
1905                 CDEBUG(D_INODE, "an attempt to link an orphan inode %lu/%u\n",
1906                        de_src->d_inode->i_ino,
1907                        de_src->d_inode->i_generation);
1908                 GOTO(cleanup, rc = -ENOENT);
1909         }
1910
1911         /* Step 3: Lookup the child */
1912         dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1913         if (IS_ERR(dchild)) {
1914                 rc = PTR_ERR(dchild);
1915                 if (rc != -EPERM && rc != -EACCES && rc != -ENAMETOOLONG)
1916                         CERROR("child lookup error %d\n", rc);
1917                 GOTO(cleanup, rc);
1918         }
1919
1920         cleanup_phase = 4; /* child dentry */
1921
1922         if (dchild->d_inode) {
1923                 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1924                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
1925                 rc = -EEXIST;
1926                 GOTO(cleanup, rc);
1927         }
1928
1929         /* Step 4: Do it. */
1930         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1931
1932         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1933                 GOTO(cleanup, rc = -EROFS);
1934
1935         handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1936         if (IS_ERR(handle))
1937                 GOTO(cleanup, rc = PTR_ERR(handle));
1938
1939         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1940         if (rc && rc != -EPERM && rc != -EACCES)
1941                 CERROR("vfs_link error %d\n", rc);
1942 cleanup:
1943         rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1944                                 handle, req, rc, 0, 0);
1945         EXIT;
1946
1947         switch (cleanup_phase) {
1948         case 4: /* child dentry */
1949                 l_dput(dchild);
1950         case 3: /* locks */
1951                 if (rc) {
1952                         ldlm_lock_decref(&src_lockh, LCK_EX);
1953                         ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1954                 } else {
1955                         ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1956                         ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1957                 }
1958         case 2: /* target dentry */
1959                 l_dput(de_tgt_dir);
1960         case 1: /* source dentry */
1961                 l_dput(de_src);
1962         case 0:
1963                 break;
1964         default:
1965                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1966                 LBUG();
1967         }
1968         req->rq_status = rc;
1969         return 0;
1970 }
1971
1972 /* The idea here is that we need to get four locks in the end:
1973  * one on each parent directory, one on each child.  We need to take
1974  * these locks in some kind of order (to avoid deadlocks), and the order
1975  * I selected is "increasing resource number" order.  We need to look up
1976  * the children, however, before we know what the resource number(s) are.
1977  * Thus the following plan:
1978  *
1979  * 1,2. Look up the parents
1980  * 3,4. Look up the children
1981  * 5. Take locks on the parents and children, in order
1982  * 6. Verify that the children haven't changed since they were looked up
1983  *
1984  * If there was a race and the children changed since they were first looked
1985  * up, it is possible that mds_verify_child() will be able to just grab the
1986  * lock on the new child resource (if it has a higher resource than any other)
1987  * but we need to compare against not only its parent, but also against the
1988  * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1989  *
1990  * We need the fancy igrab() on the child inodes because we aren't holding a
1991  * lock on the parent after the lookup is done, so dentry->d_inode may change
1992  * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1993  */
1994 int mds_get_parents_children_locked(struct obd_device *obd,
1995                                     struct mds_obd *mds,
1996                                     struct ll_fid *p1_fid,
1997                                     struct dentry **de_srcdirp,
1998                                     struct ll_fid *p2_fid,
1999                                     struct dentry **de_tgtdirp,
2000                                     int parent_mode,
2001                                     const char *old_name, int old_len,
2002                                     struct dentry **de_oldp,
2003                                     const char *new_name, int new_len,
2004                                     struct dentry **de_newp,
2005                                     struct lustre_handle *dlm_handles,
2006                                     int child_mode)
2007 {
2008         struct ldlm_res_id p1_res_id = { .name = {0} };
2009         struct ldlm_res_id p2_res_id = { .name = {0} };
2010         struct ldlm_res_id c1_res_id = { .name = {0} };
2011         struct ldlm_res_id c2_res_id = { .name = {0} };
2012         ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2013         /* Only dentry should disappear, but the inode itself would be
2014            intact otherwise. */
2015         ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2016         /* If something is going to be replaced, both dentry and inode locks are           needed */
2017         ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2018         struct ldlm_res_id *maxres_src, *maxres_tgt;
2019         struct inode *inode;
2020         int rc = 0, cleanup_phase = 0;
2021         ENTRY;
2022
2023         /* Step 1: Lookup the source directory */
2024         *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2025         if (IS_ERR(*de_srcdirp))
2026                 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2027
2028         cleanup_phase = 1; /* source directory dentry */
2029
2030         p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2031         p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2032
2033         /* Step 2: Lookup the target directory */
2034         if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2035                 *de_tgtdirp = dget(*de_srcdirp);
2036         } else {
2037                 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2038                 if (IS_ERR(*de_tgtdirp)) {
2039                         rc = PTR_ERR(*de_tgtdirp);
2040                         *de_tgtdirp = NULL;
2041                         GOTO(cleanup, rc);
2042                 }
2043         }
2044
2045         cleanup_phase = 2; /* target directory dentry */
2046
2047         p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2048         p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2049
2050         /* Step 3: Lookup the source child entry */
2051         *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2052         if (IS_ERR(*de_oldp)) {
2053                 rc = PTR_ERR(*de_oldp);
2054                 CDEBUG(D_INODE, "old child lookup error (%.*s): rc %d\n",
2055                        old_len - 1, old_name, rc);
2056                 GOTO(cleanup, rc);
2057         }
2058
2059         cleanup_phase = 3; /* original name dentry */
2060
2061         inode = (*de_oldp)->d_inode;
2062         if (inode != NULL)
2063                 inode = igrab(inode);
2064         if (inode == NULL)
2065                 GOTO(cleanup, rc = -ENOENT);
2066
2067         c1_res_id.name[0] = inode->i_ino;
2068         c1_res_id.name[1] = inode->i_generation;
2069
2070         iput(inode);
2071
2072         /* Step 4: Lookup the target child entry */
2073         if (!new_name)
2074                 GOTO(retry_locks, rc);
2075         *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2076         if (IS_ERR(*de_newp)) {
2077                 rc = PTR_ERR(*de_newp);
2078                 CDEBUG(D_DENTRY, "new child lookup error (%.*s): rc %d\n",
2079                        old_len - 1, old_name, rc);
2080                 GOTO(cleanup, rc);
2081         }
2082
2083         cleanup_phase = 4; /* target dentry */
2084
2085         inode = (*de_newp)->d_inode;
2086         if (inode != NULL)
2087                 inode = igrab(inode);
2088         if (inode == NULL)
2089                 goto retry_locks;
2090
2091         c2_res_id.name[0] = inode->i_ino;
2092         c2_res_id.name[1] = inode->i_generation;
2093         iput(inode);
2094
2095 retry_locks:
2096         /* Step 5: Take locks on the parents and child(ren) */
2097         maxres_src = &p1_res_id;
2098         maxres_tgt = &p2_res_id;
2099         cleanup_phase = 4; /* target dentry */
2100
2101         if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id,NULL,NULL))
2102                 maxres_src = &c1_res_id;
2103         if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id,NULL,NULL))
2104                 maxres_tgt = &c2_res_id;
2105
2106         rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2107                                     &p_policy,
2108                                     &p2_res_id, &dlm_handles[1], parent_mode,
2109                                     &p_policy,
2110                                     &c1_res_id, &dlm_handles[2], child_mode,
2111                                     &c1_policy,
2112                                     &c2_res_id, &dlm_handles[3], child_mode,
2113                                     &c2_policy);
2114         if (rc)
2115                 GOTO(cleanup, rc);
2116
2117         cleanup_phase = 6; /* parent and child(ren) locks */
2118
2119         /* Step 6a: Re-lookup source child to verify it hasn't changed */
2120         rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2121                               parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
2122                               child_mode, &c1_policy, old_name, old_len,
2123                               maxres_tgt);
2124         if (rc) {
2125                 if (c2_res_id.name[0] != 0)
2126                         ldlm_lock_decref(&dlm_handles[3], child_mode);
2127                 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2128                 cleanup_phase = 4;
2129                 if (rc > 0)
2130                         goto retry_locks;
2131                 GOTO(cleanup, rc);
2132         }
2133
2134         if ((*de_oldp)->d_inode == NULL)
2135                 GOTO(cleanup, rc = -ENOENT);
2136
2137         if (!new_name)
2138                 GOTO(cleanup, rc);
2139         /* Step 6b: Re-lookup target child to verify it hasn't changed */
2140         rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2141                               parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
2142                               child_mode, &c2_policy, new_name, new_len,
2143                               maxres_src);
2144         if (rc) {
2145                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2146                 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2147                 cleanup_phase = 4;
2148                 if (rc > 0)
2149                         goto retry_locks;
2150                 GOTO(cleanup, rc);
2151         }
2152
2153         EXIT;
2154 cleanup:
2155         if (rc) {
2156                 switch (cleanup_phase) {
2157                 case 6: /* child lock(s) */
2158                         if (c2_res_id.name[0] != 0)
2159                                 ldlm_lock_decref(&dlm_handles[3], child_mode);
2160                         if (c1_res_id.name[0] != 0)
2161                                 ldlm_lock_decref(&dlm_handles[2], child_mode);
2162                 case 5: /* parent locks */
2163                         ldlm_lock_decref(&dlm_handles[1], parent_mode);
2164                         ldlm_lock_decref(&dlm_handles[0], parent_mode);
2165                 case 4: /* target dentry */
2166                         l_dput(*de_newp);
2167                 case 3: /* source dentry */
2168                         l_dput(*de_oldp);
2169                 case 2: /* target directory dentry */
2170                         l_dput(*de_tgtdirp);
2171                 case 1: /* source directry dentry */
2172                         l_dput(*de_srcdirp);
2173                 }
2174         }
2175
2176         return rc;
2177 }
2178
2179 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2180                             struct ptlrpc_request *req,
2181                             struct lustre_handle *lockh)
2182 {
2183         struct obd_device *obd = req->rq_export->exp_obd;
2184         struct dentry *de_srcdir = NULL;
2185         struct dentry *de_tgtdir = NULL;
2186         struct dentry *de_old = NULL;
2187         struct dentry *de_new = NULL;
2188         struct inode *old_inode = NULL, *new_inode = NULL;
2189         struct mds_obd *mds = mds_req2mds(req);
2190         struct lustre_handle dlm_handles[4];
2191         struct mds_body *body = NULL;
2192         struct lov_mds_md *lmm = NULL;
2193         int rc = 0, lock_count = 3, cleanup_phase = 0;
2194         void *handle = NULL;
2195         unsigned int qcids[MAXQUOTAS] = { 0, 0 };
2196         unsigned int qpids[4] = { 0, 0, 0, 0 };
2197         ENTRY;
2198
2199         LASSERT(offset == REQ_REC_OFF);
2200         offset = REPLY_REC_OFF;
2201
2202         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2203                   rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2204                   rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2205
2206         mds_counter_incr(req->rq_export, LPROC_MDS_RENAME);
2207
2208         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2209
2210         if (rec->ur_dlm)
2211                 ldlm_request_cancel(req, rec->ur_dlm, 0);
2212
2213         rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2214                                              rec->ur_fid2, &de_tgtdir, LCK_EX,
2215                                              rec->ur_name, rec->ur_namelen,
2216                                              &de_old, rec->ur_tgt,
2217                                              rec->ur_tgtlen, &de_new,
2218                                              dlm_handles, LCK_EX);
2219         if (rc)
2220                 GOTO(cleanup, rc);
2221
2222         cleanup_phase = 1; /* parent(s), children, locks */
2223
2224         old_inode = de_old->d_inode;
2225         new_inode = de_new->d_inode;
2226
2227         if (new_inode != NULL)
2228                 lock_count = 4;
2229
2230         /* sanity check for src inode */
2231         if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
2232             old_inode->i_ino == de_tgtdir->d_inode->i_ino)
2233                 GOTO(cleanup, rc = -EINVAL);
2234
2235         if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2236                 GOTO(cleanup, rc = -EROFS);
2237
2238         if (new_inode == NULL)
2239                 goto no_unlink;
2240
2241         igrab(new_inode);
2242         cleanup_phase = 2; /* iput(new_inode) when finished */
2243
2244         /* sanity check for dest inode */
2245         if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
2246             new_inode->i_ino == de_tgtdir->d_inode->i_ino)
2247                 GOTO(cleanup, rc = -EINVAL);
2248
2249         if (old_inode == new_inode)
2250                 GOTO(cleanup, rc = 0);
2251
2252         /* save uids/gids for qunit acquire/release */
2253         qcids[USRQUOTA] = old_inode->i_uid;
2254         qcids[GRPQUOTA] = old_inode->i_gid;
2255         qpids[USRQUOTA] = de_tgtdir->d_inode->i_uid;
2256         qpids[GRPQUOTA] = de_tgtdir->d_inode->i_gid;
2257         qpids[2] = de_srcdir->d_inode->i_uid;
2258         qpids[3] = de_srcdir->d_inode->i_gid;
2259
2260         /* if we are about to remove the target at first, pass the EA of
2261          * that inode to client to perform and cleanup on OST */
2262         body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
2263         LASSERT(body != NULL);
2264
2265         /* child orphan sem protects orphan_dec_test && is_orphan race */
2266         MDS_DOWN_READ_ORPHAN_SEM(new_inode);
2267         cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
2268
2269         if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
2270             new_inode->i_nlink == 1) {
2271                 if (mds_orphan_open_count(new_inode) > 0) {
2272                         /* need to lock pending_dir before transaction */
2273                         LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2274                         cleanup_phase = 4; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
2275                 } else if (S_ISREG(new_inode->i_mode)) {
2276                         mds_pack_inode2fid(&body->fid1, new_inode);
2277                         mds_pack_inode2body(body, new_inode);
2278                         mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
2279                                     new_inode, MDS_PACK_MD_LOCK, 0);
2280                 }
2281         }
2282
2283 no_unlink:
2284         OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_RENAME_WRITE,
2285                        de_srcdir->d_inode->i_sb);
2286
2287 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
2288         /* Check if we are moving old entry into its child. 2.6 does not
2289            check for this in vfs_rename() anymore */
2290         if (is_subdir(de_new, de_old))
2291                 GOTO(cleanup, rc = -EINVAL);
2292 #endif
2293
2294         lmm = lustre_msg_buf(req->rq_repmsg, offset + 1, 0);
2295         handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
2296                                   NULL, le32_to_cpu(lmm->lmm_stripe_count));
2297
2298         if (IS_ERR(handle))
2299                 GOTO(cleanup, rc = PTR_ERR(handle));
2300
2301         lock_kernel();
2302         de_old->d_fsdata = req;
2303         de_new->d_fsdata = req;
2304
2305         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2306         unlock_kernel();
2307
2308         if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2309                 if (mds_orphan_open_count(new_inode) > 0)
2310                         rc = mds_orphan_add_link(rec, obd, de_new);
2311
2312                 if (rc == 1)
2313                         GOTO(cleanup, rc = 0);
2314
2315                 if (!S_ISREG(new_inode->i_mode))
2316                         GOTO(cleanup, rc);
2317
2318                 if (!(body->valid & OBD_MD_FLEASIZE)) {
2319                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2320                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
2321                 } else if (mds_log_op_unlink(obd,
2322                                              lustre_msg_buf(req->rq_repmsg,
2323                                                             offset + 1, 0),
2324                                              lustre_msg_buflen(req->rq_repmsg,
2325                                                                offset + 1),
2326                                              lustre_msg_buf(req->rq_repmsg,
2327                                                             offset + 2, 0),
2328                                              lustre_msg_buflen(req->rq_repmsg,
2329                                                                offset + 2))
2330                            > 0) {
2331                         body->valid |= OBD_MD_FLCOOKIE;
2332                 }
2333         }
2334
2335         GOTO(cleanup, rc);
2336 cleanup:
2337         rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2338                                 handle, req, rc, 0, 0);
2339
2340         switch (cleanup_phase) {
2341         case 4:
2342                 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2343         case 3:
2344                 MDS_UP_READ_ORPHAN_SEM(new_inode);
2345         case 2:
2346                 iput(new_inode);
2347         case 1:
2348                 if (rc) {
2349                         if (lock_count == 4)
2350                                 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2351                         ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2352                         ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
2353                         ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
2354                 } else {
2355                         if (lock_count == 4)
2356                                 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2357                         ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2358                         ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_EX);
2359                         ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_EX);
2360                 }
2361                 l_dput(de_new);
2362                 l_dput(de_old);
2363                 l_dput(de_tgtdir);
2364                 l_dput(de_srcdir);
2365         case 0:
2366                 break;
2367         default:
2368                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2369                 LBUG();
2370         }
2371         req->rq_status = rc;
2372
2373         /* acquire/release qunit */
2374         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
2375                       FSFILT_OP_RENAME);
2376         return 0;
2377 }
2378
2379 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2380                            struct ptlrpc_request *, struct lustre_handle *);
2381
2382 static mds_reinter reinters[REINT_MAX] = {
2383         [REINT_SETATTR] mds_reint_setattr,
2384         [REINT_CREATE] mds_reint_create,
2385         [REINT_LINK] mds_reint_link,
2386         [REINT_UNLINK] mds_reint_unlink,
2387         [REINT_RENAME] mds_reint_rename,
2388         [REINT_OPEN] mds_open
2389 };
2390
2391 int mds_reint_rec(struct mds_update_record *rec, int offset,
2392                   struct ptlrpc_request *req, struct lustre_handle *lockh)
2393 {
2394         struct obd_device *obd = req->rq_export->exp_obd;
2395         struct mds_obd *mds = &obd->u.mds;
2396         struct lvfs_run_ctxt saved;
2397         int rc;
2398 #ifdef CRAY_XT3
2399         gid_t fsgid = rec->ur_uc.luc_fsgid;
2400 #endif
2401         ENTRY;
2402
2403 #ifdef CRAY_XT3
2404         if (req->rq_uid != LNET_UID_ANY) {
2405                 /* non-root local cluster client
2406                  * NB root's creds are believed... */
2407                 LASSERT (req->rq_uid != 0);
2408                 rec->ur_uc.luc_fsuid = req->rq_uid;
2409                 rec->ur_uc.luc_cap = 0;
2410         }
2411 #endif
2412
2413         /* get group info of this user */
2414         rec->ur_uc.luc_uce = upcall_cache_get_entry(mds->mds_group_hash,
2415                                                     rec->ur_uc.luc_fsuid,
2416                                                     rec->ur_uc.luc_fsgid, 2,
2417                                                     &rec->ur_uc.luc_suppgid1);
2418
2419         if (IS_ERR(rec->ur_uc.luc_uce)) {
2420                 rc = PTR_ERR(rec->ur_uc.luc_uce);
2421                 rec->ur_uc.luc_uce = NULL;
2422                 RETURN(rc);
2423         }
2424
2425         /* checked by unpacker */
2426         LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2427
2428 #ifdef CRAY_XT3
2429         if (rec->ur_uc.luc_uce)
2430                 rec->ur_uc.luc_fsgid = rec->ur_uc.luc_uce->ue_primary;
2431 #endif
2432
2433         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2434
2435 #ifdef CRAY_XT3
2436         if (rec->ur_uc.luc_uce && fsgid != rec->ur_uc.luc_fsgid &&
2437             in_group_p(fsgid)) {
2438                 rec->ur_uc.luc_fsgid = fsgid;
2439                 current->fsgid = saved.luc.luc_fsgid = fsgid;
2440         }
2441 #endif
2442
2443         rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2444         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2445
2446         upcall_cache_put_entry(mds->mds_group_hash, rec->ur_uc.luc_uce);
2447         RETURN(rc);
2448 }