lustre/mds/mds_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *   Author: Phil Schwan <phil@clusterfs.com>
   8  *   Author: Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/version.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/version.h>
  35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  36 # include <linux/buffer_head.h>
  37 # include <linux/workqueue.h>
  38 #else
  39 # include <linux/locks.h>
  40 #endif
  41
  42 #include <linux/obd_class.h>
  43 #include <linux/obd_lov.h>
  44 #include <linux/lustre_fsfilt.h>
  45 #include <linux/lprocfs_status.h>
  46
  47 #include "mds_internal.h"
  48
  49 /* Exported function from this file are:
  50  *
  51  * mds_open - called by the intent handler
  52  * mds_close - an rpc handling function
  53  * mds_pin - an rpc handling function - which will go away
  54  * mds_mfd_close - for force closing files when a client dies
  55  */
  56
  57 /*
  58  * MDS file data handling: file data holds a handle for a file opened
  59  * by a client.
  60  */
  61
  62 static void mds_mfd_addref(void *mfdp)
  63 {
  64         struct mds_file_data *mfd = mfdp;
  65
  66         atomic_inc(&mfd->mfd_refcount);
  67         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  68                atomic_read(&mfd->mfd_refcount));
  69 }
  70
  71 struct mds_file_data *mds_mfd_new(void)
  72 {
  73         struct mds_file_data *mfd;
  74
  75         OBD_ALLOC(mfd, sizeof *mfd);
  76         if (mfd == NULL) {
  77                 CERROR("mds: out of memory\n");
  78                 return NULL;
  79         }
  80
  81         atomic_set(&mfd->mfd_refcount, 2);
  82
  83         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  84         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  85
  86         return mfd;
  87 }
  88
  89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  90 {
  91         ENTRY;
  92         LASSERT(handle != NULL);
  93         RETURN(class_handle2object(handle->cookie));
  94 }
  95
  96 static void mds_mfd_put(struct mds_file_data *mfd)
  97 {
  98         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
  99                atomic_read(&mfd->mfd_refcount) - 1);
 100         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 101                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 102         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 103                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 104                 OBD_FREE(mfd, sizeof *mfd);
 105         }
 106 }
 107
 108 static void mds_mfd_destroy(struct mds_file_data *mfd)
 109 {
 110         class_handle_unhash(&mfd->mfd_handle);
 111         mds_mfd_put(mfd);
 112 }
 113
 114 #ifdef IFILTERDATA_ACTUALLY_USED
 115 /* Caller must hold mds->mds_epoch_sem */
 116 static int mds_alloc_filterdata(struct inode *inode)
 117 {
 118         LASSERT(inode->i_filterdata == NULL);
 119         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 120         if (inode->i_filterdata == NULL)
 121                 return -ENOMEM;
 122         LASSERT(igrab(inode) == inode);
 123         return 0;
 124 }
 125
 126 /* Caller must hold mds->mds_epoch_sem */
 127 static void mds_free_filterdata(struct inode *inode)
 128 {
 129         LASSERT(inode->i_filterdata != NULL);
 130         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 131         inode->i_filterdata = NULL;
 132         iput(inode);
 133 }
 134 #endif /*IFILTERDATA_ACTUALLY_USED*/
 135
 136 /* Write access to a file: executors cause a negative count,
 137  * writers a positive count.  The semaphore is needed to perform
 138  * a check for the sign and then increment or decrement atomically.
 139  *
 140  * This code is closely tied to the allocation of the d_fsdata and the
 141  * MDS epoch, so we use the same semaphore for the whole lot.
 142  *
 143  * We could use a different semaphore for each file, if it ever shows
 144  * up in a profile, which it won't.
 145  *
 146  * epoch argument is nonzero during recovery */
 147 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 148                                 __u64 epoch)
 149 {
 150         int rc = 0;
 151
 152         down(&mds->mds_epoch_sem);
 153
 154         if (atomic_read(&inode->i_writecount) < 0) {
 155                 up(&mds->mds_epoch_sem);
 156                 RETURN(-ETXTBSY);
 157         }
 158
 159 #ifdef IFILTERDATA_ACTUALLY_USED
 160         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 161                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 162                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 163                        inode->i_generation);
 164                 goto out;
 165         }
 166
 167         if (inode->i_filterdata == NULL)
 168                 mds_alloc_filterdata(inode);
 169         if (inode->i_filterdata == NULL) {
 170                 rc = -ENOMEM;
 171                 goto out;
 172         }
 173 #endif /*IFILTERDATA_ACTUALLY_USED*/
 174         if (epoch > mds->mds_io_epoch)
 175                 mds->mds_io_epoch = epoch;
 176         else
 177                 mds->mds_io_epoch++;
 178 #ifdef IFILTERDATA_ACTUALLY_USED
 179         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 180         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 181                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 182  out:
 183 #endif /*IFILTERDATA_ACTUALLY_USED*/
 184         if (rc == 0)
 185                 atomic_inc(&inode->i_writecount);
 186         up(&mds->mds_epoch_sem);
 187         return rc;
 188 }
 189
 190 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 191  * again -- which is never true if the file is about to be unlinked.  Otherwise
 192  * returns the number of remaining writers. */
 193 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 194                                 struct mds_body *body, int unlinking)
 195 {
 196         int rc = 0;
 197         ENTRY;
 198
 199         down(&mds->mds_epoch_sem);
 200         atomic_dec(&inode->i_writecount);
 201         rc = atomic_read(&inode->i_writecount);
 202         if (rc > 0)
 203                 GOTO(out, rc);
 204 #if 0
 205         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 206                 GOTO(out, rc = EAGAIN);
 207 #endif
 208 #ifdef IFILTERDATA_ACTUALLY_USED
 209         mds_free_filterdata(inode);
 210 #endif
 211  out:
 212         up(&mds->mds_epoch_sem);
 213         return rc;
 214 }
 215
 216 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 217 {
 218         ENTRY;
 219         down(&mds->mds_epoch_sem);
 220         if (atomic_read(&inode->i_writecount) > 0) {
 221                 up(&mds->mds_epoch_sem);
 222                 RETURN(-ETXTBSY);
 223         }
 224         atomic_dec(&inode->i_writecount);
 225         up(&mds->mds_epoch_sem);
 226         RETURN(0);
 227 }
 228
 229 static void mds_allow_write_access(struct inode *inode)
 230 {
 231         ENTRY;
 232         atomic_inc(&inode->i_writecount);
 233 }
 234
 235 int mds_query_write_access(struct inode *inode)
 236 {
 237         ENTRY;
 238         RETURN(atomic_read(&inode->i_writecount));
 239 }
 240
 241 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 242 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 243                                              struct vfsmount *mnt, int flags,
 244                                              struct ptlrpc_request *req)
 245 {
 246         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 247         struct mds_obd *mds = mds_req2mds(req);
 248         struct mds_file_data *mfd;
 249         struct mds_body *body;
 250         int error;
 251         ENTRY;
 252
 253         mfd = mds_mfd_new();
 254         if (mfd == NULL) {
 255                 CERROR("mds: out of memory\n");
 256                 GOTO(cleanup_dentry, error = -ENOMEM);
 257         }
 258
 259         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 260
 261         if (flags & FMODE_WRITE) {
 262                 /* FIXME: in recovery, need to pass old epoch here */
 263                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 264                 if (error)
 265                         GOTO(cleanup_mfd, error);
 266 #ifdef IFILTERDATA_ACTUALLY_USED
 267                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 268 #endif /*IFILTERDATA_ACTUALLY_USED*/
 269         } else if (flags & FMODE_EXEC) {
 270                 error = mds_deny_write_access(mds, dentry->d_inode);
 271                 if (error)
 272                         GOTO(cleanup_mfd, error);
 273         }
 274
 275         dget(dentry);
 276
 277         /* Mark the file as open to handle open-unlink. */
 278
 279         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 280         mds_orphan_open_inc(dentry->d_inode);
 281         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 282
 283         mfd->mfd_mode = flags;
 284         mfd->mfd_dentry = dentry;
 285         mfd->mfd_xid = req->rq_xid;
 286
 287         spin_lock(&med->med_open_lock);
 288         list_add(&mfd->mfd_list, &med->med_open_head);
 289         spin_unlock(&med->med_open_lock);
 290         mds_mfd_put(mfd);
 291
 292         body->handle.cookie = mfd->mfd_handle.h_cookie;
 293
 294         RETURN(mfd);
 295
 296 cleanup_mfd:
 297         mds_mfd_put(mfd);
 298         mds_mfd_destroy(mfd);
 299 cleanup_dentry:
 300         return ERR_PTR(error);
 301 }
 302
 303 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 304                                 struct lov_desc *desc)
 305 {
 306         int i;
 307         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 308                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 309                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 310         }
 311 }
 312
 313 /* Must be called with i_sem held */
 314 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 315                               struct mds_update_record *rec,
 316                               struct mds_obd *mds, struct obd_device *obd,
 317                               struct dentry *dchild, void **handle,
 318                               obd_id **ids)
 319 {
 320         struct obdo *oa = NULL;
 321         struct obd_trans_info oti = { 0 };
 322         struct mds_body *body;
 323         struct lov_stripe_md *lsm = NULL;
 324         struct lov_mds_md *lmm = NULL;
 325         struct inode *inode = dchild->d_inode;
 326         void *lmm_buf;
 327         int rc, lmm_bufsize, lmm_size;
 328         ENTRY;
 329
 330         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 331             !(rec->ur_flags & FMODE_WRITE))
 332                 RETURN(0);
 333
 334         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 335
 336         if (!S_ISREG(inode->i_mode))
 337                 RETURN(0);
 338         if (body->valid & OBD_MD_FLEASIZE)
 339                 RETURN(0);
 340
 341         OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 342         if (*ids == NULL)
 343                 RETURN(-ENOMEM);
 344         oti.oti_objid = *ids;
 345
 346         /* replay case */
 347         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 348                 LASSERT(id_ino(rec->ur_id2));
 349                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 350                 lmm_size = rec->ur_eadatalen;
 351                 lmm = rec->ur_eadata;
 352                 LASSERT(lmm);
 353
 354                 if (*handle == NULL)
 355                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 356                 if (IS_ERR(*handle)) {
 357                         rc = PTR_ERR(*handle);
 358                         *handle = NULL;
 359                         GOTO(out_oa, rc);
 360                 }
 361
 362                 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 363
 364                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 365                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 366                 LASSERT(lmm_buf);
 367                 LASSERT(lmm_bufsize >= lmm_size);
 368                 memcpy(lmm_buf, lmm, lmm_size);
 369                 rc = fsfilt_set_md(obd, inode, *handle, lmm,
 370                                    lmm_size, EA_LOV);
 371                 if (rc)
 372                         CERROR("open replay failed to set md:%d\n", rc);
 373                 RETURN(0);
 374         }
 375
 376         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 377                 GOTO(out_ids, rc = -ENOMEM);
 378
 379         oa = obdo_alloc();
 380         if (oa == NULL)
 381                 GOTO(out_ids, rc = -ENOMEM);
 382         oa->o_mode = S_IFREG | 0600;
 383         oa->o_id = inode->i_ino;
 384         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 385         oa->o_generation = inode->i_generation;
 386         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 387         oa->o_gid = 0;
 388         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
 389                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 390         oa->o_size = 0;
 391
 392         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
 393                         OBD_MD_FLCTIME);
 394
 395         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 396                 /* check if things like lfs setstripe are sending us the ea */
 397                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 398                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 399                                            mds->mds_dt_exp,
 400                                            0, &lsm, rec->ur_eadata);
 401                         if (rc)
 402                                 GOTO(out_oa, rc);
 403                 } else {
 404                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 405                         if (lmm == NULL)
 406                                 GOTO(out_oa, rc = -ENOMEM);
 407
 408                         lmm_size = mds->mds_max_mdsize;
 409                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 410                                         lmm, &lmm_size, 1, 0);
 411                         if (rc > 0)
 412                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 413                                                    mds->mds_dt_exp,
 414                                                    0, &lsm, lmm);
 415                         OBD_FREE(lmm, mds->mds_max_mdsize);
 416                         if (rc)
 417                                 GOTO(out_oa, rc);
 418                 }
 419                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 420                 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
 421                 if (rc) {
 422                         int level = D_ERROR;
 423                         if (rc == -ENOSPC)
 424                                 level = D_INODE;
 425                         CDEBUG(level, "error creating objects for "
 426                                       "inode %lu: rc = %d\n",
 427                                inode->i_ino, rc);
 428                         if (rc > 0) {
 429                                 CERROR("obd_create returned invalid "
 430                                        "rc %d\n", rc);
 431                                 rc = -EIO;
 432                         }
 433                         GOTO(out_oa, rc);
 434                 }
 435         } else {
 436                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 437                                    0, &lsm, rec->ur_eadata);
 438                 if (rc) {
 439                         GOTO(out_oa, rc);
 440                 }
 441                 lsm->lsm_object_id = oa->o_id;
 442                 lsm->lsm_object_gr = oa->o_gr;
 443         }
 444         if (inode->i_size) {
 445                 oa->o_size = inode->i_size;
 446                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 447                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 448                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
 449                 if (rc) {
 450                         CERROR("error setting attrs for inode %lu: rc %d\n",
 451                                inode->i_ino, rc);
 452                         if (rc > 0) {
 453                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 454                                 rc = -EIO;
 455                         }
 456                         GOTO(out_oa, rc);
 457                 }
 458         }
 459
 460         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 461         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 462
 463         LASSERT(lsm && lsm->lsm_object_id);
 464         lmm = NULL;
 465         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 466         if (!id_ino(rec->ur_id2))
 467                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 468         LASSERT(rc >= 0);
 469         lmm_size = rc;
 470         body->eadatasize = rc;
 471
 472         if (*handle == NULL)
 473                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 474         if (IS_ERR(*handle)) {
 475                 rc = PTR_ERR(*handle);
 476                 *handle = NULL;
 477                 GOTO(out_ids, rc);
 478         }
 479
 480         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 481                            lmm_size, EA_LOV);
 482
 483         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 484         lmm_bufsize = req->rq_repmsg->buflens[offset];
 485         LASSERT(lmm_buf);
 486         LASSERT(lmm_bufsize >= lmm_size);
 487
 488         memcpy(lmm_buf, lmm, lmm_size);
 489         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 490  out_oa:
 491         oti_free_cookies(&oti);
 492         obdo_free(oa);
 493  out_ids:
 494         if (rc) {
 495                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 496                 *ids = NULL;
 497         }
 498         if(lsm)
 499                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 500         RETURN(rc);
 501 }
 502
 503 int
 504 mds_destroy_objects(struct obd_device *obd,
 505                     struct inode *inode, int async)
 506 {
 507         struct mds_obd *mds = &obd->u.mds;
 508         struct lov_mds_md *lmm = NULL;
 509         int rc, lmm_size;
 510         ENTRY;
 511
 512         LASSERT(inode != NULL);
 513
 514         if (inode->i_nlink != 0) {
 515                 CWARN("attempt to destroy OSS object when "
 516                       "i_nlink == %d\n", (int)inode->i_nlink);
 517                 RETURN(0);
 518         }
 519
 520         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 521         if (lmm == NULL)
 522                 RETURN(-ENOMEM);
 523
 524         lmm_size = mds->mds_max_mdsize;
 525         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 526         if (rc < 0) {
 527                 CERROR("no stripe info for %lu/%lu inode\n",
 528                        (unsigned long)inode->i_ino,
 529                        (unsigned long)inode->i_generation);
 530                 GOTO(out_free_lmm, rc);
 531         }
 532
 533         if (rc > 0) {
 534                 /* asynchronously unlink objecect on OSS */
 535                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 536                                        NULL, 0, async);
 537                 if (rc) {
 538                         CERROR("error unlinking object on OSS, "
 539                                "err %d\n", rc);
 540                         GOTO(out_free_lmm, rc);
 541                 }
 542         } else {
 543                 CDEBUG(D_INODE, "no stripping info found for inode "
 544                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 545                       (unsigned long)inode->i_generation);
 546         }
 547         EXIT;
 548 out_free_lmm:
 549         OBD_FREE(lmm, mds->mds_max_mdsize);
 550         return rc;
 551 }
 552
 553 static void reconstruct_open(struct mds_update_record *rec, int offset,
 554                              struct ptlrpc_request *req,
 555                              struct lustre_handle *child_lockh)
 556 {
 557         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 558         struct mds_client_data *mcd = med->med_mcd;
 559         struct mds_obd *mds = mds_req2mds(req);
 560         struct mds_file_data *mfd;
 561         struct obd_device *obd = req->rq_export->exp_obd;
 562         struct dentry *parent = NULL, *dchild;
 563         struct ldlm_reply *rep;
 564         struct mds_body *body;
 565         struct list_head *t;
 566         int put_child = 1;
 567         int rc;
 568         ENTRY;
 569
 570         LASSERT(offset == 3); /* only called via intent */
 571         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 572         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 573
 574         /* copy rc, transno and disp; steal locks */
 575         mds_req_from_mcd(req, mcd);
 576         intent_set_disposition(rep, mcd->mcd_last_data);
 577
 578         /* Only replay if create or open actually happened. */
 579         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 580                 EXIT;
 581                 return; /* error looking up parent or child */
 582         }
 583
 584         if (rec->ur_namelen == 1) {
 585                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 586                        OLID4(rec->ur_id1));
 587                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 588         } else {
 589                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 590                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 591                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 592                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 593                                 rec->ur_namelen - 1);
 594                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 595                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 596         }
 597
 598         if (!dchild->d_inode)
 599                 GOTO(out_dput, 0); /* child not present to open */
 600
 601         /* At this point, we know we have a child. We'll send
 602          * it back _unless_ it not created and open failed.  */
 603         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 604             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 605             req->rq_status)
 606                 GOTO(out_dput, 0);
 607
 608         /* get lock (write for O_CREAT, read otherwise) */
 609         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 610         if (S_ISREG(dchild->d_inode->i_mode)) {
 611                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 612                                  dchild->d_inode, 1, 0);
 613
 614                 if (rc)
 615                         LASSERT(rc == req->rq_status);
 616
 617                 /* If we have LOV EA data, the OST holds size, mtime */
 618                 if (!(body->valid & OBD_MD_FLEASIZE))
 619                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 620                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 621         }
 622
 623         /* If we have -EEXIST as the status, and we were asked to create
 624          * exclusively, we can tell we failed because the file already existed.
 625          */
 626         if (req->rq_status == -EEXIST &&
 627             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 628              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 629                 GOTO(out_dput, 0);
 630         }
 631
 632         /* If we didn't get as far as trying to open, then some locking thing
 633          * probably went wrong, and we'll just bail here.
 634          */
 635         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 636                 GOTO(out_dput, 0);
 637
 638         /* If we failed, then we must have failed opening, so don't look for
 639          * file descriptor or anything, just give the client the bad news.
 640          */
 641         if (req->rq_status)
 642                 GOTO(out_dput, 0);
 643
 644         mfd = NULL;
 645         list_for_each(t, &med->med_open_head) {
 646                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 647                 if (mfd->mfd_xid == req->rq_xid)
 648                         break;
 649                 mfd = NULL;
 650         }
 651
 652         /* #warning "XXX fixme" bug 2991 */
 653         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 654          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 655          * to detect a re-open */
 656         if (mfd == NULL) {
 657                 mntget(mds->mds_vfsmnt);
 658                 CERROR("Re-opened file \n");
 659                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 660                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 661                 if (!mfd) {
 662                         CERROR("mds: out of memory\n");
 663                         GOTO(out_dput, req->rq_status = -ENOMEM);
 664                 }
 665                 put_child = 0;
 666         } else {
 667                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 668                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 669                        mfd->mfd_handle.h_cookie);
 670         }
 671
 672  out_dput:
 673         if (put_child)
 674                 l_dput(dchild);
 675         if (parent)
 676                 l_dput(parent);
 677         EXIT;
 678 }
 679
 680 /* do NOT or the MAY_*'s, you'll get the weakest */
 681 static int accmode(int flags)
 682 {
 683         int res = 0;
 684
 685         if (flags & FMODE_READ)
 686                 res = MAY_READ;
 687         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 688                 res |= MAY_WRITE;
 689         if (flags & FMODE_EXEC)
 690                 res = MAY_EXEC;
 691         return res;
 692 }
 693
 694 /* Handles object creation, actual opening, and I/O epoch */
 695 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 696                            struct mds_body *body, int flags, void **handle,
 697                            struct mds_update_record *rec, struct ldlm_reply *rep)
 698 {
 699         struct mds_obd *mds = mds_req2mds(req);
 700         struct obd_device *obd = req->rq_export->exp_obd;
 701         struct mds_file_data *mfd = NULL;
 702         obd_id *ids = NULL; /* object IDs created */
 703         unsigned mode;
 704         int rc = 0;
 705         ENTRY;
 706
 707         /* atomically create objects if necessary */
 708         down(&dchild->d_inode->i_sem);
 709         mode = dchild->d_inode->i_mode;
 710         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 711             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 712                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 713                                  dchild->d_inode, 0, 0);
 714                 if (rc) {
 715                         up(&dchild->d_inode->i_sem);
 716                         RETURN(rc);
 717                 }
 718         }
 719         if (rec != NULL) {
 720                 /* no EA: create objects */
 721                 if ((body->valid & OBD_MD_FLEASIZE) &&
 722                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 723                         up(&dchild->d_inode->i_sem);
 724                         RETURN(-EEXIST);
 725                 }
 726                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 727                         /* no EA: create objects */
 728                         rc = mds_create_objects(req, 2, rec, mds, obd,
 729                                                 dchild, handle, &ids);
 730                         if (rc) {
 731                                 CERROR("mds_create_objects: rc = %d\n", rc);
 732                                 up(&dchild->d_inode->i_sem);
 733                                 RETURN(rc);
 734                         }
 735                 }
 736                 if (S_ISREG(dchild->d_inode->i_mode) &&
 737                     (body->valid & OBD_MD_FLEASIZE)) {
 738                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 739                                                    req->rq_repmsg, 2);
 740                         if (!rc)
 741                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 742                                                  dchild->d_inode, 0, 0);
 743                         if (rc) {
 744                                 up(&dchild->d_inode->i_sem);
 745                                 RETURN(rc);
 746                         }
 747                 }
 748         }
 749         rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
 750         if (rc < 0) {
 751                 CERROR("mds_pack_acl: rc = %d\n", rc);
 752                 up(&dchild->d_inode->i_sem);
 753                 RETURN(rc);
 754         }
 755
 756         /* If the inode has no EA data, then MDSs hold size, mtime */
 757         if (S_ISREG(dchild->d_inode->i_mode) &&
 758             !(body->valid & OBD_MD_FLEASIZE)) {
 759                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 760                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 761         }
 762         up(&dchild->d_inode->i_sem);
 763
 764         intent_set_disposition(rep, DISP_OPEN_OPEN);
 765         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 766         if (IS_ERR(mfd))
 767                 RETURN(PTR_ERR(mfd));
 768
 769         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 770                mfd->mfd_handle.h_cookie);
 771         if (ids != NULL) {
 772                 mds_dt_update_objids(obd, ids);
 773                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 774         }
 775         //if (rc)
 776         //        mds_mfd_destroy(mfd);
 777         RETURN(rc);
 778 }
 779
 780 static int mds_open_by_id(struct ptlrpc_request *req,
 781                           struct lustre_id *id,
 782                           struct mds_body *body, int flags,
 783                           struct mds_update_record *rec,
 784                           struct ldlm_reply *rep)
 785 {
 786         struct mds_obd *mds = mds_req2mds(req);
 787         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 788         struct dentry *dchild;
 789         char idname[LL_ID_NAMELEN];
 790         int idlen = 0, rc;
 791         void *handle = NULL;
 792         ENTRY;
 793
 794         down(&pending_dir->i_sem);
 795         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 796         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 797                                 idlen);
 798         if (IS_ERR(dchild)) {
 799                 up(&pending_dir->i_sem);
 800                 rc = PTR_ERR(dchild);
 801                 CERROR("error looking up %s in PENDING: rc = %d\n",
 802                        idname, rc);
 803                 RETURN(rc);
 804         }
 805
 806         up(&pending_dir->i_sem);
 807         if (dchild->d_inode != NULL) {
 808                 mds_inode_set_orphan(dchild->d_inode);
 809                 mds_pack_inode2body(req2obd(req), body,
 810                                     dchild->d_inode, 1);
 811
 812                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 813                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 814                 CWARN("Orphan %s found and opened in PENDING directory\n",
 815                        idname);
 816                 goto open;
 817         }
 818         l_dput(dchild);
 819
 820         /*
 821          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 822          * regular inode that was previously created.
 823          */
 824         dchild = mds_id2dentry(req2obd(req), id, NULL);
 825         if (IS_ERR(dchild))
 826                 RETURN(PTR_ERR(dchild));
 827
 828         mds_pack_inode2body(req2obd(req), body,
 829                             dchild->d_inode, 1);
 830
 831         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 832         intent_set_disposition(rep, DISP_LOOKUP_POS);
 833
 834  open:
 835         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 836         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 837                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 838         /* XXX what do we do here if mds_finish_transno itself failed? */
 839         l_dput(dchild);
 840         RETURN(rc);
 841 }
 842
 843 int mds_pin(struct ptlrpc_request *req, int offset)
 844 {
 845         struct obd_device *obd = req->rq_export->exp_obd;
 846         struct mds_body *request_body, *reply_body;
 847         struct lvfs_run_ctxt saved;
 848         int rc, size = sizeof(*reply_body);
 849         ENTRY;
 850
 851         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 852                                       sizeof(*request_body));
 853
 854         rc = lustre_pack_reply(req, 1, &size, NULL);
 855         if (rc)
 856                 RETURN(rc);
 857         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 858
 859         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 860         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 861                             request_body->flags, NULL, NULL);
 862         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 863
 864         RETURN(rc);
 865 }
 866
 867 /*
 868  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 869  * child_lockh is NULL we just get the lock as a barrier to wait for other
 870  * holders of this lock, and drop it right away again.
 871  */
 872 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 873                        struct lustre_handle *child_lockh)
 874 {
 875         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 876         struct lustre_handle lockh;
 877         int lock_flags = 0;
 878         int rc;
 879         ENTRY;
 880
 881         if (child_lockh == NULL)
 882                 child_lockh = &lockh;
 883
 884         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 885                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 886                               mds_blocking_ast, ldlm_completion_ast, NULL,
 887                               NULL, NULL, 0, NULL, child_lockh);
 888         if (rc != ELDLM_OK)
 889                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 890         else if (child_lockh == &lockh)
 891                 ldlm_lock_decref(child_lockh, LCK_EX);
 892
 893         RETURN(rc);
 894 }
 895
 896 int mds_open(struct mds_update_record *rec, int offset,
 897              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 898 {
 899         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 900         struct obd_device *obd = req->rq_export->exp_obd;
 901         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 902         struct mds_obd *mds = mds_req2mds(req);
 903         struct ldlm_reply *rep = NULL;
 904         struct mds_body *body = NULL;
 905         struct dentry *dchild = NULL, *dparent = NULL;
 906         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 907         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 908         int parent_mode = LCK_PR;
 909         void *handle = NULL;
 910         struct dentry_params dp;
 911         struct mea *mea = NULL;
 912         int mea_size, update_mode;
 913         int child_mode = LCK_PR;
 914         /* Always returning LOOKUP lock if open succesful to guard
 915            dentry on client. */
 916         ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
 917         struct ldlm_res_id child_res_id = { .name = {0}};
 918         int lock_flags = 0;
 919         ENTRY;
 920
 921         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 922                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 923                   rec->ur_mode);
 924
 925         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 926                          (obd_timeout + 1) / 4);
 927
 928         if (offset == 3) { /* intent */
 929                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 930                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 931         } else if (offset == 1) { /* non-intent reint */
 932                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 933         } else {
 934                 body = NULL;
 935                 LBUG();
 936         }
 937
 938         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 939
 940         /*
 941          * Step 0: If we are passed a id, then we assume the client already
 942          * opened this file and is only replaying the RPC, so we open the inode
 943          * by fid (at some large expense in security).
 944          */
 945         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 946                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 947                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 948                        rec->ur_name, rec->ur_mode);
 949
 950                 LASSERT(id_ino(rec->ur_id2));
 951
 952                 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
 953                                     rec, rep);
 954                 if (rc != -ENOENT) {
 955                         mds_body_do_reverse_map(med, body);
 956                         RETURN(rc);
 957                 }
 958
 959                 /* We didn't find the correct inode on disk either, so we
 960                  * need to re-create it via a regular replay. */
 961                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 962         } else {
 963                 LASSERT(!id_ino(rec->ur_id2));
 964         }
 965
 966         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 967
 968         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 969                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 970                 RETURN(-ENOMEM);
 971         }
 972
 973         acc_mode = accmode(rec->ur_flags);
 974
 975         /* Step 1: Find and lock the parent */
 976         if (rec->ur_flags & MDS_OPEN_CREAT) {
 977                 /* XXX Well, in fact we only need this lock mode change if
 978                    in addition to O_CREAT, the file does not exist.
 979                    But we do not know if it exists or not yet */
 980                 parent_mode = LCK_PW;
 981         }
 982
 983         if (rec->ur_namelen == 1) {
 984                 /* client (LMV) wants to open the file by id */
 985                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 986                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 987                 if (IS_ERR(dchild)) {
 988                         rc = PTR_ERR(dparent);
 989                         CERROR("child lookup by an id error %d\n", rc);
 990                         GOTO(cleanup, rc);
 991                 }
 992                 goto got_child;
 993         }
 994
 995 restart:
 996         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
 997                                        parent_lockh, &update_mode, rec->ur_name,
 998                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 999         if (IS_ERR(dparent)) {
1000                 rc = PTR_ERR(dparent);
1001                 if (rc != -ENOENT) {
1002                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1003                                OLID4(rec->ur_id1), rc);
1004                 } else {
1005                         /* Just cannot find parent - make it look like
1006                            usual negative lookup to avoid extra MDS RPC */
1007                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1008                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1009                 }
1010                 GOTO(cleanup, rc);
1011         }
1012         LASSERT(dparent->d_inode != NULL);
1013
1014         cleanup_phase = 1; /* parent dentry and lock */
1015
1016         /* try to retrieve MEA data for this dir */
1017         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1018         if (rc)
1019                 GOTO(cleanup, rc);
1020
1021         if (mea != NULL) {
1022                 /*
1023                  * dir is already splitted, check is requested filename should *
1024                  * live at this MDS or at another one.
1025                  */
1026                 int i;
1027                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1028                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1029                         CDEBUG(D_OTHER,
1030                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1031                                " should be %lu(%d)\n", obd->obd_name,
1032                                mea->mea_master, dparent->d_inode->i_ino,
1033                                dparent->d_inode->i_generation, rec->ur_name,
1034                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1035                         GOTO(cleanup, rc = -ERESTART);
1036                 }
1037         }
1038
1039         /* Step 2: Lookup the child */
1040         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1041         if (IS_ERR(dchild)) {
1042                 rc = PTR_ERR(dchild);
1043                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1044                 GOTO(cleanup, rc);
1045         }
1046
1047 got_child:
1048         cleanup_phase = 2; /* child dentry */
1049
1050         if (dchild->d_flags & DCACHE_CROSS_REF) {
1051                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1052                 LASSERT(rec->ur_namelen > 1);
1053
1054                 /* we're gonna acquire LOOKUP lock on the child,
1055                  * but we have already locked parent and our order
1056                  * may conflict with enqueue_order_locks(). so,
1057                  * drop parent lock and acquire both the locks in
1058                  * common order. bug 6190 */
1059 #ifdef S_PDIROPS
1060                 if (parent_lockh[1].cookie != 0)
1061                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1062 #endif
1063                 ldlm_lock_decref(parent_lockh, parent_mode);
1064                 l_dput(dchild);
1065                 l_dput(dparent);
1066                 cleanup_phase = 0;
1067
1068                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1069                                                  parent_lockh, &dparent,
1070                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1071                                                  &update_mode, rec->ur_name,
1072                                                  rec->ur_namelen, child_lockh,
1073                                                  &dchild, LCK_PR,
1074                                                  MDS_INODELOCK_LOOKUP);
1075
1076                 if (rc)
1077                         GOTO(cleanup, rc);
1078
1079 #ifdef S_PDIROPS
1080                 if (parent_lockh[1].cookie != 0)
1081                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1082 #endif
1083                 ldlm_lock_decref(parent_lockh, LCK_PR);
1084
1085                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1086                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1087                                dchild->d_inode);
1088                         if (dchild->d_inode)
1089                                 ldlm_lock_decref(child_lockh, LCK_PR);
1090                         l_dput(dchild);
1091                         l_dput(dparent);
1092                         GOTO(restart, rc);
1093                 }
1094
1095                 mds_pack_dentry2body(obd, body, dchild, 1);
1096                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1097                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1098
1099                 if (mea)
1100                         OBD_FREE(mea, mea_size);
1101                 l_dput(dchild);
1102                 l_dput(dparent);
1103                 RETURN(rc);
1104         }
1105
1106         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1107
1108         if (dchild->d_inode)
1109                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1110         else
1111                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1112
1113         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1114         if (dchild->d_inode == NULL) {
1115                 unsigned long ino;
1116                 struct iattr iattr;
1117                 struct inode *inode;
1118
1119                 ino = (unsigned long)id_ino(rec->ur_id2);
1120
1121                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1122                                           update_mode);
1123
1124                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1125                        obd->obd_name, dparent->d_inode->i_ino,
1126                        dparent->d_inode->i_generation, rc);
1127
1128                 if (rc > 0) {
1129                         /* dir got splitted */
1130                         GOTO(cleanup, rc = -ERESTART);
1131                 } else if (rc < 0) {
1132                         /* error happened during spitting */
1133                         GOTO(cleanup, rc);
1134                 }
1135
1136                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1137                         /*
1138                          * dentry is negative and we weren't supposed to create
1139                          * object, get out of here.
1140                          */
1141                         GOTO(cleanup, rc = -ENOENT);
1142                 }
1143
1144                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1145                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1146                                       NULL);
1147                 if (IS_ERR(handle)) {
1148                         rc = PTR_ERR(handle);
1149                         handle = NULL;
1150                         GOTO(cleanup, rc);
1151                 }
1152                 dchild->d_fsdata = (void *) &dp;
1153                 dp.p_ptr = req;
1154                 dp.p_inum = ino;
1155
1156                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1157                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1158                         dchild->d_fsdata = NULL;
1159
1160                 if (rc) {
1161                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1162                         GOTO(cleanup, rc);
1163                 }
1164                 inode = dchild->d_inode;
1165                 if (ino) {
1166                         LASSERT(ino == inode->i_ino);
1167
1168                         /* written as part of setattr */
1169                         inode->i_generation = id_gen(rec->ur_id2);
1170                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1171                                inode->i_ino, inode->i_generation);
1172                 }
1173
1174                 created = 1;
1175                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1176                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1177                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1178
1179                 iattr.ia_uid = rec->ur_fsuid;
1180                 if (dparent->d_inode->i_mode & S_ISGID)
1181                         iattr.ia_gid = dparent->d_inode->i_gid;
1182                 else
1183                         iattr.ia_gid = rec->ur_fsgid;
1184
1185                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1186                         ATTR_MTIME | ATTR_CTIME;
1187
1188                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1189                 if (rc)
1190                         CERROR("error on child setattr: rc = %d\n", rc);
1191
1192                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1193
1194                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1195                 if (rc)
1196                         CERROR("error on parent setattr: rc = %d\n", rc);
1197                 else {
1198                         MD_COUNTER_INCREMENT(obd, create);
1199                 }
1200
1201                 down(&dchild->d_inode->i_sem);
1202                 if (ino) {
1203                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1204                                                   handle, rec->ur_id2);
1205                         if (rc) {
1206                                 CERROR("mds_update_inode_sid() failed, "
1207                                        "rc = %d\n", rc);
1208                         }
1209                         id_assign_fid(&body->id1, rec->ur_id2);
1210
1211                         /*
1212                          * make sure, that fid is up-to-date.
1213                          */
1214                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1215                 } else {
1216                         rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1217                                                  handle, &body->id1);
1218                         if (rc) {
1219                                 CERROR("mds_alloc_inode_sid() failed, "
1220                                        "rc = %d\n", rc);
1221                         }
1222                 }
1223                 up(&dchild->d_inode->i_sem);
1224
1225                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1226                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1227                                            dchild->d_inode, handle,
1228                                            req->rq_export->exp_sync);
1229                         handle = NULL;
1230                 }
1231
1232                 acc_mode = 0;           /* Don't check for permissions */
1233         }
1234         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1235
1236         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1237                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1238                  dchild->d_name.name, dchild, dchild->d_inode);
1239
1240         if (S_ISREG(dchild->d_inode->i_mode)) {
1241                 /* Check permissions etc */
1242                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1243                 if (rc != 0)
1244                         GOTO(cleanup, rc);
1245
1246                 /* Can't write to a read-only file */
1247                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1248                         GOTO(cleanup, rc = -EPERM);
1249
1250                 /* An append-only file must be opened in append mode for
1251                  * writing */
1252                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1253                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1254                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1255                         GOTO(cleanup, rc = -EPERM);
1256         }
1257
1258         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1259             (rec->ur_flags & MDS_OPEN_EXCL)) {
1260                 /* File already exists, we didn't just create it, and we
1261                  * were passed O_EXCL; err-or. */
1262                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1263         }
1264
1265         if (S_ISDIR(dchild->d_inode->i_mode)) {
1266                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1267                     rec->ur_flags & FMODE_WRITE) {
1268                         /* we are trying to create or write a exist dir */
1269                         GOTO(cleanup, rc = -EISDIR);
1270                 }
1271                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1272                         GOTO(cleanup, rc = -EACCES);
1273
1274                 /*
1275                  * here was checking for possible GNS mount points to skip their
1276                  * open. I removed it as its detection based only on SUID bit is
1277                  * not reliable. Opening such a dirs should not cause any
1278                  * problems, at least tests show that. --umka
1279                  */
1280         }
1281
1282         /* if we are following a symlink, don't open */
1283         if (S_ISLNK(dchild->d_inode->i_mode))
1284                 GOTO(cleanup_no_trans, rc = 0);
1285
1286         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1287             !S_ISDIR(dchild->d_inode->i_mode))
1288                 GOTO(cleanup, rc = -ENOTDIR);
1289
1290         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1291                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1292                 GOTO(cleanup, rc = -EAGAIN);
1293         }
1294
1295         /* Obtain OPEN lock as well */
1296         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1297
1298         /* We cannot use acc_mode here, because it is zeroed in case of
1299            creating a file, so we get wrong lockmode */
1300         if (accmode(rec->ur_flags) & MAY_WRITE)
1301                 child_mode = LCK_CW;
1302         else if (accmode(rec->ur_flags) & MAY_EXEC)
1303                 child_mode = LCK_PR;
1304         else
1305                 child_mode = LCK_CR;
1306
1307         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1308                 struct lustre_id sid;
1309
1310                 down(&dchild->d_inode->i_sem);
1311                 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1312                 up(&dchild->d_inode->i_sem);
1313                 if (rc) {
1314                         CERROR("Can't read inode self id, "
1315                                "inode %lu, rc %d\n",
1316                                dchild->d_inode->i_ino, rc);
1317                         GOTO(cleanup, rc);
1318                 }
1319
1320                 /* In case of replay we do not get a lock assuming that the
1321                    caller has it already */
1322                 child_res_id.name[0] = id_fid(&sid);
1323                 child_res_id.name[1] = id_group(&sid);
1324
1325                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1326                                       child_res_id, LDLM_IBITS, &policy,
1327                                       child_mode, &lock_flags,
1328                                       mds_blocking_ast, ldlm_completion_ast,
1329                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1330                 if (rc != ELDLM_OK)
1331                         GOTO(cleanup, rc);
1332
1333                 cleanup_phase = 3;
1334         }
1335
1336         /* Step 5: mds_open it */
1337         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1338                              rec, rep);
1339
1340         EXIT;
1341 cleanup:
1342         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1343                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1344
1345 cleanup_no_trans:
1346         switch (cleanup_phase) {
1347         case 3:
1348                 if (rc) {
1349                         ldlm_lock_decref(child_lockh, child_mode);
1350                         child_lockh->cookie = 0;
1351                 }
1352         case 2:
1353                 if (rc && created) {
1354                         int err = vfs_unlink(dparent->d_inode, dchild);
1355                         if (err) {
1356                                 CERROR("unlink(%.*s) in error path: %d\n",
1357                                        dchild->d_name.len, dchild->d_name.name,
1358                                        err);
1359                         }
1360                 } else if (created) {
1361                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1362                 }
1363                 l_dput(dchild);
1364         case 1:
1365                 if (dparent == NULL)
1366                         break;
1367
1368                 l_dput(dparent);
1369 #ifdef S_PDIROPS
1370                 if (parent_lockh[1].cookie != 0)
1371                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1372 #endif
1373                 if (rc)
1374                         ldlm_lock_decref(parent_lockh, parent_mode);
1375                 else
1376                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1377         }
1378         if (mea)
1379                 OBD_FREE(mea, mea_size);
1380         if (rc == 0)
1381                 atomic_inc(&mds->mds_open_count);
1382
1383         mds_body_do_reverse_map(med, body);
1384
1385         /*
1386          * If we have not taken the "open" lock, we may not return 0 here,
1387          * because caller expects 0 to mean "lock is taken", and it needs
1388          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1389          * client. Later caller should rewrite the return value back to zero
1390          * if it to be used any further.
1391          */
1392         if ((cleanup_phase != 3) && !rc)
1393                 rc = ENOLCK;
1394
1395         RETURN(rc);
1396 }
1397
1398 /* Close a "file descriptor" and possibly unlink an orphan from the
1399  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1400  *
1401  * If we are being called from mds_disconnect() because the client has
1402  * disappeared, then req == NULL and we do not update last_rcvd because
1403  * there is nothing that could be recovered by the client at this stage
1404  * (it will not even _have_ an entry in last_rcvd anymore).
1405  *
1406  * Returns EAGAIN if the client needs to get more data and re-close. */
1407 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1408                   struct obd_device *obd, struct mds_file_data *mfd,
1409                   int unlink_orphan)
1410 {
1411         struct inode *inode = mfd->mfd_dentry->d_inode;
1412         char idname[LL_ID_NAMELEN];
1413         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1414         struct dentry *pending_child = NULL;
1415         struct mds_obd *mds = &obd->u.mds;
1416         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1417         void *handle = NULL;
1418         struct mds_body *request_body = NULL, *reply_body = NULL;
1419         struct dentry_params dp;
1420         struct iattr iattr = { 0 };
1421         struct llog_create_locks *lcl = NULL;
1422         ENTRY;
1423
1424         if (req && req->rq_reqmsg != NULL)
1425                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1426                                               sizeof(*request_body));
1427         if (req && req->rq_repmsg != NULL)
1428                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1429                                             sizeof(*reply_body));
1430
1431         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1432         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1433                idname, inode->i_nlink, mds_orphan_open_count(inode));
1434
1435         last_orphan = mds_orphan_open_dec_test(inode) &&
1436                 mds_inode_is_orphan(inode);
1437         UP_WRITE_I_ALLOC_SEM(inode);
1438
1439         /* this is half of the actual "close" */
1440         if (mfd->mfd_mode & FMODE_WRITE) {
1441                 rc = mds_put_write_access(mds, inode, request_body,
1442                                           last_orphan && unlink_orphan);
1443         } else if (mfd->mfd_mode & FMODE_EXEC) {
1444                 mds_allow_write_access(inode);
1445         }
1446
1447         if (last_orphan && unlink_orphan) {
1448                 struct lov_mds_md *lmm = NULL;
1449                 int stripe_count = 0;
1450                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1451
1452                 if (obd->obd_recovering) {
1453                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1454                                " is over\n", idname);
1455                         GOTO(out, rc);
1456                 }
1457
1458                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1459
1460                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1461                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1462                         CERROR("found \"orphan\" %s %s with link count %d\n",
1463                                S_ISREG(inode->i_mode) ? "file" : "dir",
1464                                idname, inode->i_nlink);
1465                 /*
1466                  * sadly, there is no easy way to save pending_child from
1467                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1468                  * normally it will still be in the dcache.
1469                  */
1470                 down(&pending_dir->i_sem);
1471                 cleanup_phase = 1; /* up(i_sem) when finished */
1472                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1473                                                idlen);
1474                 if (IS_ERR(pending_child))
1475                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1476                 LASSERT(pending_child->d_inode != NULL);
1477
1478                 cleanup_phase = 2; /* dput(pending_child) when finished */
1479                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1480                         rc = vfs_rmdir(pending_dir, pending_child);
1481                         if (rc)
1482                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1483                                        idname, rc);
1484                         goto out;
1485                 }
1486
1487                 if (req != NULL && req->rq_repmsg != NULL) {
1488                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1489                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1490                 }
1491
1492                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1493                                           NULL, stripe_count);
1494                 if (IS_ERR(handle)) {
1495                         rc = PTR_ERR(handle);
1496                         handle = NULL;
1497                         GOTO(cleanup, rc);
1498                 }
1499
1500                 pending_child->d_fsdata = (void *) &dp;
1501                 dp.p_inum = 0;
1502                 dp.p_ptr = req;
1503                 rc = vfs_unlink(pending_dir, pending_child);
1504                 if (rc)
1505                         CERROR("error unlinking orphan %s: rc %d\n",
1506                                idname, rc);
1507
1508                 if (req != NULL && req->rq_repmsg != NULL &&
1509                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1510                     mds_log_op_unlink(obd, pending_child->d_inode,
1511                                       lmm, req->rq_repmsg->buflens[1],
1512                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1513                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1514                         reply_body->valid |= OBD_MD_FLCOOKIE;
1515                 }
1516
1517                 rc = mds_destroy_objects(obd, inode, 1);
1518                 if (rc) {
1519                         CERROR("cannot destroy OSS object on close, err %d\n",
1520                                rc);
1521                         rc = 0;
1522                 }
1523
1524                 goto out; /* Don't bother updating attrs on unlinked inode */
1525         }
1526
1527 #if 0
1528         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1529                 /* Update the on-disk attributes if this was the last write
1530                  * close, and all information was provided (i.e., rc == 0)
1531                  *
1532                  * XXX this should probably be abstracted with mds_reint_setattr
1533                  */
1534
1535                 if (request_body->valid & OBD_MD_FLMTIME &&
1536                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1537                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1538                         iattr.ia_valid |= ATTR_MTIME;
1539                 }
1540                 if (request_body->valid & OBD_MD_FLCTIME &&
1541                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1542                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1543                         iattr.ia_valid |= ATTR_CTIME;
1544                 }
1545
1546                 /* XXX can't set block count with fsfilt_setattr (!) */
1547                 if (request_body->valid & OBD_MD_FLSIZE) {
1548                         iattr.ia_valid |= ATTR_SIZE;
1549                         iattr.ia_size = request_body->size;
1550                 }
1551                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1552                         iattr.ia_valid |= ATTR_BLOCKS;
1553                         iattr.ia_blocks = request_body->blocks
1554                 } */
1555
1556         }
1557 #endif
1558         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1559                 /* Only start a transaction to write out only the atime if
1560                  * it is more out-of-date than the specified limit.  If we
1561                  * are already going to write out the atime then do it anyway.
1562                  * */
1563                 if ((request_body->atime >
1564                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1565                     (iattr.ia_valid != 0 &&
1566                      request_body->atime > LTIME_S(inode->i_atime))) {
1567                         LTIME_S(iattr.ia_atime) = request_body->atime;
1568                         iattr.ia_valid |= ATTR_ATIME;
1569                 }
1570         }
1571
1572         if (iattr.ia_valid != 0) {
1573                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1574                 if (IS_ERR(handle))
1575                         GOTO(cleanup, rc = PTR_ERR(handle));
1576                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1577                 if (rc)
1578                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1579         }
1580 out:
1581         /* If other clients have this file open for write, rc will be > 0 */
1582         if (rc > 0)
1583                 rc = 0;
1584         l_dput(mfd->mfd_dentry);
1585         mds_mfd_destroy(mfd);
1586
1587  cleanup:
1588         atomic_dec(&mds->mds_open_count);
1589         if (req != NULL && reply_body != NULL) {
1590                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1591         } else if (handle) {
1592                 int err, force_sync = 0;
1593
1594                 if (req && req->rq_export)
1595                         force_sync = req->rq_export->exp_sync;
1596
1597                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1598                                     force_sync);
1599                 if (err) {
1600                         CERROR("error committing close: %d\n", err);
1601                         if (!rc)
1602                                 rc = err;
1603                 }
1604         }
1605
1606         switch (cleanup_phase) {
1607         case 2:
1608                 if (lcl != NULL)
1609                         ptlrpc_save_llog_lock(req, lcl);
1610                 dput(pending_child);
1611         case 1:
1612                 up(&pending_dir->i_sem);
1613         }
1614         RETURN(rc);
1615 }
1616
1617 int mds_close(struct ptlrpc_request *req, int offset)
1618 {
1619         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1620         struct obd_device *obd = req->rq_export->exp_obd;
1621         struct mds_body *body;
1622         struct mds_file_data *mfd;
1623         struct lvfs_run_ctxt saved;
1624         struct inode *inode;
1625         int rc, repsize[3] = {sizeof(struct mds_body),
1626                               obd->u.mds.mds_max_mdsize,
1627                               obd->u.mds.mds_max_cookiesize};
1628         ENTRY;
1629         MD_COUNTER_INCREMENT(obd, close);
1630
1631         rc = lustre_pack_reply(req, 3, repsize, NULL);
1632         if (rc) {
1633                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1634                 req->rq_status = rc;
1635         } else {
1636                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1637         }
1638
1639         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1640                 DEBUG_REQ(D_HA, req, "close replay");
1641                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1642                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1643                        req->rq_repmsg->buflens[2]);
1644         }
1645
1646         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1647                                   lustre_swab_mds_body);
1648         if (body == NULL) {
1649                 CERROR("Can't unpack body\n");
1650                 req->rq_status = -EFAULT;
1651                 RETURN(-EFAULT);
1652         }
1653
1654         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1655                 /* do some stuff */ ;
1656
1657         mfd = mds_handle2mfd(&body->handle);
1658         if (mfd == NULL) {
1659                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1660                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1661                 req->rq_status = -ESTALE;
1662                 RETURN(-ESTALE);
1663         }
1664
1665         inode = mfd->mfd_dentry->d_inode;
1666         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1667         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1668         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1669                 struct mds_body *rep_body;
1670
1671                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1672                                           sizeof (*rep_body));
1673                 LASSERT(rep_body != NULL);
1674
1675                 mds_pack_inode2body(obd, rep_body, inode,
1676                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1677
1678                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1679                             inode, MDS_PACK_MD_LOCK, 0);
1680         }
1681         spin_lock(&med->med_open_lock);
1682         list_del(&mfd->mfd_list);
1683         spin_unlock(&med->med_open_lock);
1684
1685         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1686         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1687         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1688
1689         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1690                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1691                 req->rq_status = -ENOMEM;
1692                 mds_mfd_put(mfd);
1693                 RETURN(-ENOMEM);
1694         }
1695
1696         mds_mfd_put(mfd);
1697         RETURN(0);
1698 }
1699
1700 int mds_done_writing(struct ptlrpc_request *req, int offset)
1701 {
1702         struct mds_body *body;
1703         int rc, size = sizeof(struct mds_body);
1704         ENTRY;
1705
1706         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1707
1708         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1709                                   lustre_swab_mds_body);
1710         if (body == NULL) {
1711                 CERROR("Can't unpack body\n");
1712                 req->rq_status = -EFAULT;
1713                 RETURN(-EFAULT);
1714         }
1715
1716         rc = lustre_pack_reply(req, 1, &size, NULL);
1717         if (rc) {
1718                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1719                 req->rq_status = rc;
1720         }
1721
1722         RETURN(0);
1723 }