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 Cluster File Systems, Inc.
   8  *
   9  *   This file is part of Lustre, http://www.lustre.org.
  10  *
  11  *   Lustre is free software; you can redistribute it and/or
  12  *   modify it under the terms of version 2 of the GNU General Public
  13  *   License as published by the Free Software Foundation.
  14  *
  15  *   Lustre is distributed in the hope that it will be useful,
  16  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  *   GNU General Public License for more details.
  19  *
  20  *   You should have received a copy of the GNU General Public License
  21  *   along with Lustre; if not, write to the Free Software
  22  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  *
  24  *  Author: Peter Braam <braam@clusterfs.com>
  25  *  Author: Andreas Dilger <adilger@clusterfs.com>
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/obd_support.h>
  32 #include <linux/obd_class.h>
  33 #include <linux/obd.h>
  34 #include <linux/lustre_lib.h>
  35 #include <linux/lustre_idl.h>
  36 #include <linux/lustre_mds.h>
  37 #include <linux/lustre_dlm.h>
  38 #include <linux/obd_class.h>
  39
  40 extern inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req);
  41
  42 /* Assumes caller has already pushed us into the kernel context. */
  43 int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  44                          struct ptlrpc_request *req)
  45 {
  46         struct mds_export_data *med = &req->rq_export->exp_mds_data;
  47         struct mds_client_data *mcd = med->med_mcd;
  48         __u64 last_rcvd;
  49         loff_t off;
  50         int rc;
  51
  52         off = MDS_LR_CLIENT + med->med_off * MDS_LR_SIZE;
  53
  54         spin_lock(&mds->mds_last_lock);
  55         last_rcvd = ++mds->mds_last_rcvd;
  56         spin_unlock(&mds->mds_last_lock);
  57         req->rq_repmsg->transno = HTON__u64(last_rcvd);
  58         mcd->mcd_last_rcvd = cpu_to_le64(last_rcvd);
  59         mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
  60         mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
  61
  62         mds_fs_set_last_rcvd(mds, handle);
  63         rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mcd, sizeof(*mcd), &off);
  64         CDEBUG(D_INODE, "wrote trans #"LPD64" for client '%s' at #%d: rc = %d\n",
  65                last_rcvd, mcd->mcd_uuid, med->med_off, rc);
  66
  67         if (rc == sizeof(*mcd))
  68                 rc = 0;
  69         else {
  70                 CERROR("error writing to last_rcvd file: rc = %d\n", rc);
  71                 if (rc >= 0)
  72                         rc = -EIO;
  73         }
  74
  75         return rc;
  76 }
  77
  78 /* In the write-back case, the client holds a lock on a subtree.
  79  * In the intent case, the client holds a lock on the child inode.
  80  * In the pathname case, the client (may) hold a lock on the child inode. */
  81 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
  82                              struct ptlrpc_request *req)
  83 {
  84         struct mds_obd *mds = mds_req2mds(req);
  85         struct obd_device *obd = req->rq_export->exp_obd;
  86         struct dentry *de;
  87         void *handle;
  88         struct lustre_handle child_lockh;
  89         int rc = 0, err;
  90
  91         if (req->rq_reqmsg->bufcount > offset + 1) {
  92                 struct dentry *dir;
  93                 struct lustre_handle dir_lockh;
  94                 char *name;
  95                 int namelen;
  96
  97                 /* a name was supplied by the client; fid1 is the directory */
  98                 dir = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PR,
  99                                             &dir_lockh);
 100                 if (IS_ERR(dir)) {
 101                         LBUG();
 102                         GOTO(out_setattr, rc = PTR_ERR(dir));
 103                 }
 104
 105                 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 106                 namelen = req->rq_reqmsg->buflens[offset + 1] - 1;
 107                 de = mds_name2locked_dentry(obd, dir, NULL, name, namelen,
 108                                             0, &child_lockh, LCK_PR);
 109                 l_dput(dir);
 110                 if (IS_ERR(de)) {
 111                         LBUG();
 112                         GOTO(out_setattr_de, rc = PTR_ERR(de));
 113                 }
 114         } else {
 115                 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 116                 if (!de || IS_ERR(de)) {
 117                         LBUG();
 118                         GOTO(out_setattr_de, rc = -ESTALE);
 119                 }
 120         }
 121         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 122
 123         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE,
 124                        de->d_inode->i_sb->s_dev);
 125
 126         handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 127         if (!handle)
 128                 GOTO(out_setattr_de, rc = PTR_ERR(handle));
 129
 130         rc = mds_fs_setattr(mds, de, handle, &rec->ur_iattr);
 131
 132         if (!rc)
 133                 rc = mds_update_last_rcvd(mds, handle, req);
 134
 135         err = mds_fs_commit(mds, de->d_inode, handle);
 136         if (err) {
 137                 CERROR("error on commit: err = %d\n", err);
 138                 if (!rc)
 139                         rc = err;
 140         }
 141
 142         EXIT;
 143 out_setattr_de:
 144         l_dput(de);
 145 out_setattr:
 146         req->rq_status = rc;
 147         return (0);
 148 }
 149
 150 static int mds_reint_recreate(struct mds_update_record *rec, int offset,
 151                               struct ptlrpc_request *req)
 152 {
 153         struct dentry *de = NULL;
 154         struct mds_obd *mds = mds_req2mds(req);
 155         struct dentry *dchild = NULL;
 156         struct inode *dir;
 157         int rc = 0;
 158         ENTRY;
 159
 160         de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 161         if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
 162                 LBUG();
 163                 GOTO(out_create_de, rc = -ESTALE);
 164         }
 165         dir = de->d_inode;
 166         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 167
 168         down(&dir->i_sem);
 169         dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
 170         if (IS_ERR(dchild)) {
 171                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 172                 up(&dir->i_sem);
 173                 LBUG();
 174                 GOTO(out_create_dchild, rc = -ESTALE);
 175         }
 176
 177         if (dchild->d_inode) {
 178                 struct mds_body *body;
 179                 rc = 0;
 180                 body = lustre_msg_buf(req->rq_repmsg, 0);
 181                 body->ino = dchild->d_inode->i_ino;
 182                 body->generation = dchild->d_inode->i_generation;
 183         } else {
 184                 CERROR("child doesn't exist (dir %ld, name %s)\n",
 185                        dir->i_ino, rec->ur_name);
 186                 rc = -ENOENT;
 187                 LBUG();
 188         }
 189
 190 out_create_dchild:
 191         l_dput(dchild);
 192         up(&dir->i_sem);
 193 out_create_de:
 194         l_dput(de);
 195         req->rq_status = rc;
 196         return 0;
 197 }
 198
 199 static int mds_reint_create(struct mds_update_record *rec, int offset,
 200                             struct ptlrpc_request *req)
 201 {
 202         struct dentry *de = NULL;
 203         struct mds_obd *mds = mds_req2mds(req);
 204         struct obd_device *obd = req->rq_export->exp_obd;
 205         struct dentry *dchild = NULL;
 206         struct inode *dir;
 207         void *handle;
 208         struct lustre_handle lockh;
 209         int rc = 0, err, lock_mode, type = rec->ur_mode & S_IFMT;
 210         ENTRY;
 211
 212         /* requests were at offset 2, replies go back at 1 */
 213         if (offset)
 214                 offset = 1;
 215
 216         if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
 217                 LBUG();
 218
 219         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 220
 221         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 222                 GOTO(out_create, rc = -ESTALE);
 223
 224         de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, lock_mode, &lockh);
 225         if (IS_ERR(de)) {
 226                 rc = PTR_ERR(de);
 227                 CERROR("parent lookup error %d\n", rc);
 228                 LBUG();
 229                 GOTO(out_create, rc);
 230         }
 231         dir = de->d_inode;
 232         CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
 233                dir->i_ino, rec->ur_name, rec->ur_mode);
 234
 235         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 236
 237         down(&dir->i_sem);
 238         dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
 239         if (IS_ERR(dchild)) {
 240                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 241                 LBUG();
 242                 GOTO(out_create_de, rc = -ESTALE);
 243         }
 244
 245         if (dchild->d_inode) {
 246                 struct mds_body *body;
 247                 struct inode *inode = dchild->d_inode;
 248                 struct lov_mds_md *lmm;
 249
 250                 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
 251                        dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
 252
 253                 body = lustre_msg_buf(req->rq_repmsg, offset);
 254                 mds_pack_inode2fid(&body->fid1, inode);
 255                 mds_pack_inode2body(body, inode);
 256                 if (S_ISREG(inode->i_mode)) {
 257                         lmm = lustre_msg_buf(req->rq_repmsg, offset + 1);
 258                         lmm->lmm_easize = mds->mds_max_mdsize;
 259
 260                         if (mds_fs_get_md(mds, inode, lmm) < 0) {
 261                                 CDEBUG(D_INFO,"No md for %ld: rc %d\n",
 262                                        inode->i_ino, rc);
 263                                 memset(lmm, 0, lmm->lmm_easize);
 264                         } else
 265                                 body->valid |= OBD_MD_FLEASIZE;
 266                 }
 267                 /* now a normal case for intent locking */
 268                 GOTO(out_create_dchild, rc = -EEXIST);
 269         }
 270
 271         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
 272
 273         if (dir->i_mode & S_ISGID) {
 274                 rec->ur_gid = dir->i_gid;
 275                 if (S_ISDIR(rec->ur_mode))
 276                         rec->ur_mode |= S_ISGID;
 277         }
 278
 279         switch (type) {
 280         case S_IFREG:{
 281                 handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
 282                 if (!handle)
 283                         GOTO(out_create_dchild, PTR_ERR(handle));
 284                 rc = vfs_create(dir, dchild, rec->ur_mode);
 285                 EXIT;
 286                 break;
 287         }
 288         case S_IFDIR:{
 289                 handle = mds_fs_start(mds, dir, MDS_FSOP_MKDIR);
 290                 if (!handle)
 291                         GOTO(out_create_dchild, PTR_ERR(handle));
 292                 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
 293                 EXIT;
 294                 break;
 295         }
 296         case S_IFLNK:{
 297                 handle = mds_fs_start(mds, dir, MDS_FSOP_SYMLINK);
 298                 if (!handle)
 299                         GOTO(out_create_dchild, PTR_ERR(handle));
 300                 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
 301                 EXIT;
 302                 break;
 303         }
 304         case S_IFCHR:
 305         case S_IFBLK:
 306         case S_IFIFO:
 307         case S_IFSOCK:{
 308                 int rdev = rec->ur_rdev;
 309                 handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
 310                 if (!handle)
 311                         GOTO(out_create_dchild, PTR_ERR(handle));
 312                 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
 313                 EXIT;
 314                 break;
 315         }
 316         default:
 317                 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
 318                 GOTO(out_create_dchild, rc = -EINVAL);
 319         }
 320
 321         if (rc) {
 322                 CERROR("error during create: %d\n", rc);
 323                 if (rc != -ENOSPC)
 324                         LBUG();
 325                 GOTO(out_create_commit, rc);
 326         } else {
 327                 struct iattr iattr;
 328                 struct inode *inode = dchild->d_inode;
 329                 struct mds_body *body;
 330
 331                 CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
 332                 if (!offset && type == S_IFREG) {
 333                         struct lov_mds_md *lmm;
 334
 335                         lmm = lustre_msg_buf(req->rq_reqmsg, 2);
 336                         rc = mds_fs_set_md(mds, inode, handle, lmm);
 337                         if (rc) {
 338                                 CERROR("error %d setting LOV MD for %ld\n",
 339                                        rc, inode->i_ino);
 340                                 GOTO(out_create_unlink, rc);
 341                         }
 342                 }
 343
 344                 iattr.ia_atime = rec->ur_time;
 345                 iattr.ia_ctime = rec->ur_time;
 346                 iattr.ia_mtime = rec->ur_time;
 347                 iattr.ia_uid = rec->ur_uid;
 348                 iattr.ia_gid = rec->ur_gid;
 349                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
 350                         ATTR_MTIME | ATTR_CTIME;
 351
 352                 rc = mds_fs_setattr(mds, dchild, handle, &iattr);
 353                 if (rc) {
 354                         CERROR("error on setattr: rc = %d\n", rc);
 355                         /* XXX should we abort here in case of error? */
 356                 }
 357
 358                 rc = mds_update_last_rcvd(mds, handle, req);
 359                 if (rc) {
 360                         CERROR("error on mds_update_last_rcvd: rc = %d\n", rc);
 361                         /* XXX should we abort here in case of error? */
 362                 }
 363
 364                 body = lustre_msg_buf(req->rq_repmsg, offset);
 365                 body->ino = inode->i_ino;
 366                 body->generation = inode->i_generation;
 367                 body->valid = OBD_MD_FLID | OBD_MD_FLGENER;
 368         }
 369         EXIT;
 370 out_create_commit:
 371         err = mds_fs_commit(mds, dir, handle);
 372         if (err) {
 373                 CERROR("error on commit: err = %d\n", err);
 374                 if (!rc)
 375                         rc = err;
 376         }
 377 out_create_dchild:
 378         l_dput(dchild);
 379         ldlm_lock_decref(&lockh, lock_mode);
 380 out_create_de:
 381         up(&dir->i_sem);
 382         l_dput(de);
 383 out_create:
 384         req->rq_status = rc;
 385         return 0;
 386
 387 out_create_unlink:
 388         /* Destroy the file we just created.  This should not need extra
 389          * journal credits, as we have already modified all of the blocks
 390          * needed in order to create the file in the first place.
 391          */
 392         switch (type) {
 393         case S_IFDIR:
 394                 err = vfs_rmdir(dir, dchild);
 395                 if (err)
 396                         CERROR("failed rmdir in error path: rc = %d\n", err);
 397                 break;
 398         default:
 399                 err = vfs_unlink(dir, dchild);
 400                 if (err)
 401                         CERROR("failed unlink in error path: rc = %d\n", err);
 402                 break;
 403         }
 404
 405         goto out_create_commit;
 406 }
 407
 408 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
 409                             struct ptlrpc_request *req)
 410 {
 411         struct dentry *de = NULL;
 412         struct dentry *dchild = NULL;
 413         struct mds_obd *mds = mds_req2mds(req);
 414         struct obd_device *obd = req->rq_export->exp_obd;
 415         char *name;
 416         struct inode *dir, *inode;
 417         struct lustre_handle lockh, child_lockh;
 418         void *handle;
 419         int namelen, lock_mode, err, rc = 0;
 420         ENTRY;
 421
 422         /* a name was supplied by the client; fid1 is the directory */
 423         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_PW : LCK_PW;
 424         de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, lock_mode, &lockh);
 425         if (IS_ERR(de)) {
 426                 LBUG();
 427                 RETURN(PTR_ERR(de));
 428         }
 429
 430         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 431                 GOTO(out_unlink, rc = -ENOENT);
 432
 433         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 434         namelen = req->rq_reqmsg->buflens[offset + 1] - 1;
 435 #warning FIXME: if mds_name2locked_dentry decrefs this lock, we must not
 436         memcpy(&child_lockh, &lockh, sizeof(child_lockh));
 437         dchild = mds_name2locked_dentry(obd, de, NULL, name, namelen,
 438                                         LCK_EX, &child_lockh, lock_mode);
 439
 440         if (IS_ERR(dchild)) {
 441                 LBUG();
 442                 GOTO(out_unlink, rc = PTR_ERR(dchild));
 443         }
 444
 445         dir = de->d_inode;
 446         inode = dchild->d_inode;
 447         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 448
 449         if (!inode) {
 450                 CDEBUG(D_INODE, "child doesn't exist (dir %ld, name %s\n",
 451                        dir->i_ino, rec->ur_name);
 452                 /* going to out_unlink_cancel causes an LBUG, don't know why */
 453                 GOTO(out_unlink_dchild, rc = -ENOENT);
 454         }
 455
 456         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
 457
 458         switch (rec->ur_mode /* & S_IFMT ? */) {
 459         case S_IFDIR:
 460                 handle = mds_fs_start(mds, dir, MDS_FSOP_RMDIR);
 461                 if (!handle)
 462                         GOTO(out_unlink_cancel, rc = PTR_ERR(handle));
 463                 rc = vfs_rmdir(dir, dchild);
 464                 break;
 465         case S_IFREG:
 466                 /* get OBD EA data first so client can also destroy object */
 467                 if ((inode->i_mode & S_IFMT) == S_IFREG && offset) {
 468                         struct mds_body *body;
 469                         struct lov_mds_md *lmm;
 470
 471                         body = lustre_msg_buf(req->rq_repmsg, 1);
 472                         mds_pack_inode2fid(&body->fid1, inode);
 473                         mds_pack_inode2body(body, inode);
 474
 475                         lmm = lustre_msg_buf(req->rq_repmsg, 2);
 476                         lmm->lmm_easize = mds->mds_max_mdsize;
 477                         if ((rc = mds_fs_get_md(mds, inode, lmm)) < 0) {
 478                                 CDEBUG(D_INFO, "No md for ino %ld: rc = %d\n",
 479                                        inode->i_ino, rc);
 480                                 memset(lmm, 0, lmm->lmm_easize);
 481                         } else
 482                                 body->valid |= OBD_MD_FLEASIZE;
 483                 }
 484                 /* no break */
 485         case S_IFLNK:
 486         case S_IFCHR:
 487         case S_IFBLK:
 488         case S_IFIFO:
 489         case S_IFSOCK:
 490                 handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
 491                 if (!handle)
 492                         GOTO(out_unlink_cancel, rc = PTR_ERR(handle));
 493                 rc = vfs_unlink(dir, dchild);
 494                 break;
 495         default:
 496                 CERROR("bad file type %o unlinking %s\n", rec->ur_mode, name);
 497                 handle = NULL;
 498                 LBUG();
 499                 GOTO(out_unlink_cancel, rc = -EINVAL);
 500         }
 501
 502         if (!rc)
 503                 rc = mds_update_last_rcvd(mds, handle, req);
 504         err = mds_fs_commit(mds, dir, handle);
 505         if (err) {
 506                 CERROR("error on commit: err = %d\n", err);
 507                 if (!rc)
 508                         rc = err;
 509         }
 510
 511         EXIT;
 512
 513 out_unlink_cancel:
 514         ldlm_lock_decref(&child_lockh, LCK_EX);
 515         err = ldlm_cli_cancel(&child_lockh);
 516         if (err < 0) {
 517                 CERROR("failed to cancel child inode lock: err = %d\n", err);
 518                 if (!rc)
 519                         rc = -ENOLCK;   /*XXX translate LDLM lock error */
 520         }
 521 out_unlink_dchild:
 522         l_dput(dchild);
 523         up(&dir->i_sem);
 524 out_unlink:
 525         ldlm_lock_decref(&lockh, lock_mode);
 526         l_dput(de);
 527         req->rq_status = rc;
 528         return 0;
 529 }
 530
 531 static int mds_reint_link(struct mds_update_record *rec, int offset,
 532                           struct ptlrpc_request *req)
 533 {
 534         struct obd_device *obd = req->rq_export->exp_obd;
 535         struct dentry *de_src = NULL;
 536         struct dentry *de_tgt_dir = NULL;
 537         struct dentry *dchild = NULL;
 538         struct mds_obd *mds = mds_req2mds(req);
 539         struct lustre_handle *handle, tgtlockh, srclockh;
 540         int lock_mode;
 541         __u64 res_id[3] = { 0 };
 542         int flags = 0;
 543         int rc = 0, err;
 544
 545         ENTRY;
 546         de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 547         if (IS_ERR(de_src) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) {
 548                 GOTO(out_link, rc = -ESTALE);
 549         }
 550
 551         /* plan to change the link count on this inode: write lock */
 552         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_PW : LCK_PW;
 553         res_id[0] = de_src->d_inode->i_ino;
 554
 555         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 556                              NULL, 0, lock_mode, &srclockh);
 557         if (rc == 0) {
 558                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 559                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 560                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 561                                       &flags, ldlm_completion_ast,
 562                                       mds_blocking_ast, NULL, 0, &srclockh);
 563                 if (rc != ELDLM_OK) {
 564                         CERROR("lock enqueue: err: %d\n", rc);
 565                         GOTO(out_link_src_put, rc = -EIO);
 566                 }
 567         } else
 568                 ldlm_lock_dump((void *)(unsigned long)srclockh.addr);
 569
 570         de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
 571         if (IS_ERR(de_tgt_dir)) {
 572                 GOTO(out_link_src, rc = -ESTALE);
 573         }
 574
 575         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_PW : LCK_PW;
 576         res_id[0] = de_tgt_dir->d_inode->i_ino;
 577
 578         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 579                              NULL, 0, lock_mode, &tgtlockh);
 580         if (rc == 0) {
 581                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 582                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 583                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 584                                       &flags, ldlm_completion_ast,
 585                                       mds_blocking_ast, NULL, 0, &tgtlockh);
 586                 if (rc != ELDLM_OK) {
 587                         CERROR("lock enqueue: err: %d\n", rc);
 588                         GOTO(out_link_tgt_dir_put, rc = -EIO);
 589                 }
 590         } else
 591                 ldlm_lock_dump((void *)(unsigned long)tgtlockh.addr);
 592
 593         down(&de_tgt_dir->d_inode->i_sem);
 594         dchild = lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen - 1);
 595         if (IS_ERR(dchild)) {
 596                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 597                 GOTO(out_link_tgt_dir, rc = -ESTALE);
 598         }
 599
 600         if (dchild->d_inode) {
 601                 struct inode *inode = dchild->d_inode;
 602                 /* in intent case ship back attributes to client */
 603                 if (offset) {
 604                         struct mds_body *body =
 605                                 lustre_msg_buf(req->rq_repmsg, 1);
 606
 607                         mds_pack_inode2fid(&body->fid1, inode);
 608                         mds_pack_inode2body(body, inode);
 609                         if (S_ISREG(inode->i_mode)) {
 610                                 struct lov_mds_md *lmm;
 611
 612                                 lmm = lustre_msg_buf(req->rq_repmsg, 2);
 613                                 lmm->lmm_easize = mds->mds_max_mdsize;
 614                                 if ((rc = mds_fs_get_md(mds, inode, lmm)) < 0) {
 615                                         CDEBUG(D_INFO,"No md for %ld: rc %d\n",
 616                                                inode->i_ino, rc);
 617                                         memset(lmm, 0, lmm->lmm_easize);
 618                                 } else
 619                                         body->valid |= OBD_MD_FLEASIZE;
 620                         }
 621                 }
 622                 CERROR("child exists (dir %ld, name %s\n",
 623                        de_tgt_dir->d_inode->i_ino, rec->ur_name);
 624                 GOTO(out_link_dchild, rc = -EEXIST);
 625         }
 626
 627         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
 628                        de_src->d_inode->i_sb->s_dev);
 629
 630         handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
 631         if (!handle)
 632                 GOTO(out_link_dchild, rc = PTR_ERR(handle));
 633
 634         rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
 635         if (rc)
 636                 CERROR("link error %d\n", rc);
 637         if (!rc)
 638                 rc = mds_update_last_rcvd(mds, handle, req);
 639
 640         err = mds_fs_commit(mds, de_tgt_dir->d_inode, handle);
 641         if (err) {
 642                 CERROR("error on commit: err = %d\n", err);
 643                 if (!rc)
 644                         rc = err;
 645         }
 646         EXIT;
 647
 648
 649
 650 out_link_dchild:
 651         l_dput(dchild);
 652 out_link_tgt_dir:
 653         ldlm_lock_decref(&tgtlockh, lock_mode);
 654 out_link_tgt_dir_put:
 655         up(&de_tgt_dir->d_inode->i_sem);
 656         l_dput(de_tgt_dir);
 657 out_link_src:
 658         ldlm_lock_decref(&srclockh, lock_mode);
 659 out_link_src_put:
 660         l_dput(de_src);
 661 out_link:
 662         req->rq_status = rc;
 663         return 0;
 664 }
 665
 666 static int mds_reint_rename(struct mds_update_record *rec, int offset,
 667                             struct ptlrpc_request *req)
 668 {
 669         struct obd_device *obd = req->rq_export->exp_obd;
 670         struct dentry *de_srcdir = NULL;
 671         struct dentry *de_tgtdir = NULL;
 672         struct dentry *de_old = NULL;
 673         struct dentry *de_new = NULL;
 674         struct mds_obd *mds = mds_req2mds(req);
 675         struct lustre_handle tgtlockh, srclockh, oldhandle;
 676         int flags, lock_mode, rc = 0, err;
 677         void *handle;
 678         __u64 res_id[3] = { 0 };
 679         ENTRY;
 680
 681         de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 682         if (IS_ERR(de_srcdir))
 683                 GOTO(out_rename, rc = -ESTALE);
 684
 685         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_PW : LCK_PW;
 686         res_id[0] = de_srcdir->d_inode->i_ino;
 687
 688         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 689                              NULL, 0, lock_mode, &srclockh);
 690         if (rc == 0) {
 691                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 692                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 693                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 694                                       &flags, ldlm_completion_ast,
 695                                       mds_blocking_ast, NULL, 0, &srclockh);
 696                 if (rc != ELDLM_OK) {
 697                         CERROR("lock enqueue: err: %d\n", rc);
 698                         GOTO(out_rename_srcput, rc = -EIO);
 699                 }
 700         } else
 701                 ldlm_lock_dump((void *)(unsigned long)srclockh.addr);
 702
 703         de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
 704         if (IS_ERR(de_tgtdir))
 705                 GOTO(out_rename_srcdir, rc = -ESTALE);
 706
 707         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_PW : LCK_PW;
 708         res_id[0] = de_tgtdir->d_inode->i_ino;
 709
 710         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 711                              NULL, 0, lock_mode, &tgtlockh);
 712         if (rc == 0) {
 713                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 714                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 715                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 716                                       &flags, ldlm_completion_ast,
 717                                       mds_blocking_ast, NULL, 0, &tgtlockh);
 718                 if (rc != ELDLM_OK) {
 719                         CERROR("lock enqueue: err: %d\n", rc);
 720                         GOTO(out_rename_tgtput, rc = -EIO);
 721                 }
 722         } else
 723                 ldlm_lock_dump((void *)(unsigned long)tgtlockh.addr);
 724
 725         double_lock(de_tgtdir, de_srcdir);
 726
 727         de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
 728         if (IS_ERR(de_old)) {
 729                 CERROR("old child lookup error (%*s): %ld\n",
 730                        rec->ur_namelen - 1, rec->ur_name, PTR_ERR(de_old));
 731                 GOTO(out_rename_tgtdir, rc = -ENOENT);
 732         }
 733
 734         de_new = lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
 735         if (IS_ERR(de_new)) {
 736                 CERROR("new child lookup error (%*s): %ld\n",
 737                        rec->ur_tgtlen - 1, rec->ur_tgt, PTR_ERR(de_new));
 738                 GOTO(out_rename_deold, rc = -ENOENT);
 739         }
 740
 741         /* in intent case ship back attributes to client */
 742         if (offset) {
 743                 struct mds_body *body = lustre_msg_buf(req->rq_repmsg, 1);
 744                 struct inode *inode = de_new->d_inode;
 745
 746                 if (!inode) {
 747                         body->valid = 0;
 748                 } else {
 749                         mds_pack_inode2fid(&body->fid1, inode);
 750                         mds_pack_inode2body(body, inode);
 751                         if (S_ISREG(inode->i_mode)) {
 752                                 struct lov_mds_md *lmm;
 753
 754                                 lmm = lustre_msg_buf(req->rq_repmsg, 2);
 755                                 lmm->lmm_easize = mds->mds_max_mdsize;
 756                                 if ((rc = mds_fs_get_md(mds, inode, lmm)) < 0) {
 757                                         CDEBUG(D_INFO,"No md for %ld: rc %d\n",
 758                                                inode->i_ino, rc);
 759                                         memset(lmm, 0, lmm->lmm_easize);
 760                                 } else
 761                                         body->valid |= OBD_MD_FLEASIZE;
 762                         }
 763                 }
 764         }
 765
 766         OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
 767                        de_srcdir->d_inode->i_sb->s_dev);
 768
 769         handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
 770         if (!handle)
 771                 GOTO(out_rename_denew, rc = PTR_ERR(handle));
 772         lock_kernel();
 773         rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new,
 774                         NULL);
 775         unlock_kernel();
 776
 777         if (!rc)
 778                 rc = mds_update_last_rcvd(mds, handle, req);
 779
 780         err = mds_fs_commit(mds, de_tgtdir->d_inode, handle);
 781         if (err) {
 782                 CERROR("error on commit: err = %d\n", err);
 783                 if (!rc)
 784                         rc = err;
 785         }
 786         EXIT;
 787
 788       out_rename_denew:
 789         l_dput(de_new);
 790       out_rename_deold:
 791         if (!rc) {
 792                 res_id[0] = de_old->d_inode->i_ino;
 793                 /* Take an exclusive lock on the resource that we're
 794                  * about to free, to force everyone to drop their
 795                  * locks. */
 796                 LDLM_DEBUG_NOLOCK("getting EX lock res "LPU64, res_id[0]);
 797                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 798                                       res_id, LDLM_PLAIN, NULL, 0, LCK_EX,
 799                                       &flags, ldlm_completion_ast,
 800                                       mds_blocking_ast, NULL, 0, &oldhandle);
 801                 if (rc)
 802                         CERROR("failed to get child inode lock (child ino "
 803                                LPD64" dir ino %ld)\n",
 804                                res_id[0], de_old->d_inode->i_ino);
 805         }
 806
 807         l_dput(de_old);
 808
 809         if (!rc) {
 810                 ldlm_lock_decref(&oldhandle, LCK_EX);
 811                 rc = ldlm_cli_cancel(&oldhandle);
 812                 if (rc < 0)
 813                         CERROR("failed to cancel child inode lock ino "
 814                                LPD64": %d\n", res_id[0], rc);
 815         }
 816 out_rename_tgtdir:
 817         double_up(&de_srcdir->d_inode->i_sem, &de_tgtdir->d_inode->i_sem);
 818         ldlm_lock_decref(&tgtlockh, lock_mode);
 819 out_rename_tgtput:
 820         l_dput(de_tgtdir);
 821 out_rename_srcdir:
 822         ldlm_lock_decref(&srclockh, lock_mode);
 823 out_rename_srcput:
 824         l_dput(de_srcdir);
 825 out_rename:
 826         req->rq_status = rc;
 827         return 0;
 828 }
 829
 830 typedef int (*mds_reinter) (struct mds_update_record *, int offset,
 831                             struct ptlrpc_request *);
 832
 833 static mds_reinter reinters[REINT_MAX + 1] = {
 834         [REINT_SETATTR] mds_reint_setattr,
 835         [REINT_CREATE] mds_reint_create,
 836         [REINT_UNLINK] mds_reint_unlink,
 837         [REINT_LINK] mds_reint_link,
 838         [REINT_RENAME] mds_reint_rename,
 839         [REINT_RECREATE] mds_reint_recreate,
 840 };
 841
 842 int mds_reint_rec(struct mds_update_record *rec, int offset,
 843                   struct ptlrpc_request *req)
 844 {
 845         struct mds_obd *mds = mds_req2mds(req);
 846         struct obd_run_ctxt saved;
 847         struct obd_ucred uc;
 848
 849         int rc;
 850
 851         if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
 852                 CERROR("opcode %d not valid\n", rec->ur_opcode);
 853                 rc = req->rq_status = -EINVAL;
 854                 RETURN(rc);
 855         }
 856
 857         uc.ouc_fsuid = rec->ur_fsuid;
 858         uc.ouc_fsgid = rec->ur_fsgid;
 859
 860         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 861         rc = reinters[rec->ur_opcode] (rec, offset, req);
 862         pop_ctxt(&saved);
 863
 864         return rc;
 865 }