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 rc = 0;
 251         ENTRY;
 252
 253         mfd = mds_mfd_new();
 254         if (mfd == NULL) {
 255                 CERROR("mds: out of memory\n");
 256                 GOTO(cleanup_dentry, rc = -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                 rc = mds_get_write_access(mds, dentry->d_inode, 0);
 264                 if (rc)
 265                         GOTO(cleanup_mfd, rc);
 266 #ifdef IFILTERDATA_ACTUALLY_USED
 267                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 268 #endif
 269         } else if (flags & FMODE_EXEC) {
 270                 rc = mds_deny_write_access(mds, dentry->d_inode);
 271                 if (rc)
 272                         GOTO(cleanup_mfd, rc);
 273         }
 274
 275         dget(dentry);
 276
 277         /* mark the file as open to handle open-unlink. */
 278         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 279         mds_orphan_open_inc(dentry->d_inode);
 280         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 281
 282         mfd->mfd_mode = flags;
 283         mfd->mfd_dentry = dentry;
 284         mfd->mfd_xid = req->rq_xid;
 285
 286         spin_lock(&med->med_open_lock);
 287         list_add(&mfd->mfd_list, &med->med_open_head);
 288         spin_unlock(&med->med_open_lock);
 289         mds_mfd_put(mfd);
 290
 291         body->handle.cookie = mfd->mfd_handle.h_cookie;
 292         RETURN(mfd);
 293 cleanup_mfd:
 294         mds_mfd_put(mfd);
 295         mds_mfd_destroy(mfd);
 296 cleanup_dentry:
 297         return ERR_PTR(rc);
 298 }
 299
 300 static inline void
 301 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 302                     struct lov_desc *desc)
 303 {
 304         int i;
 305
 306         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 307                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 308                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 309         }
 310 }
 311
 312 /* must be called with i_sem held */
 313 int
 314 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
 315                    int offset, struct mds_update_record *rec,
 316                    struct dentry *dchild, void **handle,
 317                    obd_id **ids)
 318 {
 319         struct inode *inode = dchild->d_inode;
 320         struct mds_obd *mds = &obd->u.mds;
 321         struct obd_trans_info oti = { 0 };
 322         struct lov_stripe_md *lsm = NULL;
 323         struct lov_mds_md *lmm = NULL;
 324         int rc, lmm_bufsize, lmm_size;
 325         struct obdo *oa = NULL;
 326         struct mds_body *body;
 327         void *lmm_buf;
 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                         RETURN(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 != NULL);
 367                 LASSERT(lmm_bufsize >= lmm_size);
 368
 369                 memcpy(lmm_buf, lmm, lmm_size);
 370                 rc = fsfilt_set_md(obd, inode, *handle, lmm,
 371                                    lmm_size, EA_LOV);
 372                 if (rc)
 373                         CERROR("open replay failed to set md:%d\n", rc);
 374                 RETURN(0);
 375         }
 376
 377         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 378                 RETURN(-ENOMEM);
 379
 380         oa = obdo_alloc();
 381         if (oa == NULL)
 382                 RETURN(-ENOMEM);
 383         oa->o_mode = S_IFREG | 0600;
 384         oa->o_id = inode->i_ino;
 385         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 386         oa->o_generation = inode->i_generation;
 387         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 388         oa->o_gid = 0;
 389         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
 390                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 391         oa->o_size = 0;
 392
 393         obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 394                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 395
 396         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 397                 /* check if things like lfs setstripe are sending us the ea */
 398                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 399                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 400                                            mds->mds_dt_exp,
 401                                            0, &lsm, rec->ur_eadata);
 402                         if (rc)
 403                                 GOTO(out_oa, rc);
 404                 } else {
 405                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 406                         if (lmm == NULL)
 407                                 GOTO(out_oa, rc = -ENOMEM);
 408
 409                         lmm_size = mds->mds_max_mdsize;
 410                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 411                                         lmm, &lmm_size, 1, 0);
 412                         if (rc > 0)
 413                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 414                                                    mds->mds_dt_exp,
 415                                                    0, &lsm, lmm);
 416                         OBD_FREE(lmm, mds->mds_max_mdsize);
 417                         if (rc)
 418                                 GOTO(out_oa, rc);
 419                 }
 420
 421                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 422                 oti.oti_flags |= OBD_MODE_CROW;
 423                 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
 424
 425                 if (rc) {
 426                         int level = D_ERROR;
 427                         if (rc == -ENOSPC)
 428                                 level = D_INODE;
 429                         CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
 430                                "error creating objects for "
 431                                "inode %lu: rc = %d\n",
 432                                inode->i_ino, rc);
 433                         if (rc > 0) {
 434                                 CERROR("obd_create returned invalid "
 435                                        "rc %d\n", rc);
 436                                 rc = -EIO;
 437                         }
 438                         GOTO(out_oa, rc);
 439                 }
 440         } else {
 441                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 442                                    0, &lsm, rec->ur_eadata);
 443                 if (rc)
 444                         GOTO(out_oa, rc);
 445
 446                 lsm->lsm_object_id = oa->o_id;
 447                 lsm->lsm_object_gr = oa->o_gr;
 448         }
 449         if (inode->i_size) {
 450                 oa->o_size = inode->i_size;
 451                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 452                                 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 453
 454                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
 455                 if (rc) {
 456                         CERROR("error setting attrs for inode %lu: rc %d\n",
 457                                inode->i_ino, rc);
 458                         if (rc > 0) {
 459                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 460                                 rc = -EIO;
 461                         }
 462                         GOTO(out_oa, rc);
 463                 }
 464         }
 465
 466         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 467         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 468
 469         LASSERT(lsm && lsm->lsm_object_id);
 470         lmm = NULL;
 471         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 472         if (!id_ino(rec->ur_id2))
 473                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 474         if (rc < 0) {
 475                 CERROR("cannot pack lsm, err = %d\n", rc);
 476                 GOTO(out_oa, rc);
 477         }
 478
 479         lmm_size = rc;
 480         body->eadatasize = rc;
 481
 482         if (*handle == NULL)
 483                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 484         if (IS_ERR(*handle)) {
 485                 rc = PTR_ERR(*handle);
 486                 *handle = NULL;
 487                 GOTO(out_oa, rc);
 488         }
 489
 490         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 491                            lmm_size, EA_LOV);
 492
 493         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 494         lmm_bufsize = req->rq_repmsg->buflens[offset];
 495         LASSERT(lmm_buf);
 496         LASSERT(lmm_bufsize >= lmm_size);
 497
 498         memcpy(lmm_buf, lmm, lmm_size);
 499         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 500 out_oa:
 501         oti_free_cookies(&oti);
 502         obdo_free(oa);
 503
 504         if (lsm)
 505                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 506         RETURN(rc);
 507 }
 508
 509 int
 510 mds_destroy_object(struct obd_device *obd,
 511                    struct inode *inode, int async)
 512 {
 513         struct mds_obd *mds = &obd->u.mds;
 514         struct lov_mds_md *lmm = NULL;
 515         int rc, lmm_size;
 516         ENTRY;
 517
 518         LASSERT(inode != NULL);
 519
 520         if (inode->i_nlink != 0) {
 521                 CDEBUG(D_INODE, "attempt to destroy OSS object when "
 522                        "i_nlink == %d\n", (int)inode->i_nlink);
 523                 RETURN(0);
 524         }
 525
 526         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 527         if (lmm == NULL)
 528                 RETURN(-ENOMEM);
 529
 530         lmm_size = mds->mds_max_mdsize;
 531         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 532         if (rc < 0) {
 533                 CERROR("no stripe info for %lu/%lu inode\n",
 534                        (unsigned long)inode->i_ino,
 535                        (unsigned long)inode->i_generation);
 536                 GOTO(out_free_lmm, rc);
 537         }
 538
 539         if (rc > 0) {
 540                 /* asynchronously unlink objecect on OSS */
 541                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 542                                        NULL, 0, async);
 543                 if (rc) {
 544                         CERROR("error unlinking object on OSS, "
 545                                "err %d\n", rc);
 546                         GOTO(out_free_lmm, rc);
 547                 }
 548         } else {
 549                 CDEBUG(D_INODE, "no stripping info found for inode "
 550                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 551                       (unsigned long)inode->i_generation);
 552         }
 553         EXIT;
 554 out_free_lmm:
 555         OBD_FREE(lmm, mds->mds_max_mdsize);
 556         return rc;
 557 }
 558
 559 static void reconstruct_open(struct mds_update_record *rec, int offset,
 560                              struct ptlrpc_request *req,
 561                              struct lustre_handle *child_lockh)
 562 {
 563         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 564         struct mds_client_data *mcd = med->med_mcd;
 565         struct mds_obd *mds = mds_req2mds(req);
 566         struct mds_file_data *mfd;
 567         struct obd_device *obd = req->rq_export->exp_obd;
 568         struct dentry *parent = NULL, *dchild;
 569         struct ldlm_reply *rep;
 570         struct mds_body *body;
 571         struct list_head *t;
 572         int put_child = 1;
 573         int rc;
 574         ENTRY;
 575
 576         LASSERT(offset == 3); /* only called via intent */
 577         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 578         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 579
 580         /* copy rc, transno and disp; steal locks */
 581         mds_req_from_mcd(req, mcd);
 582         intent_set_disposition(rep, mcd->mcd_last_data);
 583
 584         /* Only replay if create or open actually happened. */
 585         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 586                 EXIT;
 587                 return; /* error looking up parent or child */
 588         }
 589
 590         if (rec->ur_namelen == 1) {
 591                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 592                        OLID4(rec->ur_id1));
 593                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 594         } else {
 595                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 596                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 597                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 598                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 599                                 rec->ur_namelen - 1);
 600                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 601                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 602         }
 603
 604         if (!dchild->d_inode)
 605                 GOTO(out_dput, 0); /* child not present to open */
 606
 607         /* At this point, we know we have a child. We'll send
 608          * it back _unless_ it not created and open failed.  */
 609         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 610             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 611             req->rq_status)
 612                 GOTO(out_dput, 0);
 613
 614         /* get lock (write for O_CREAT, read otherwise) */
 615         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 616         if (S_ISREG(dchild->d_inode->i_mode)) {
 617                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 618                                  dchild->d_inode, 1, 0);
 619
 620                 if (rc)
 621                         LASSERT(rc == req->rq_status);
 622
 623                 /* If we have LOV EA data, the OST holds size, mtime */
 624                 if (!(body->valid & OBD_MD_FLEASIZE))
 625                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 626                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 627         }
 628
 629         /* If we have -EEXIST as the status, and we were asked to create
 630          * exclusively, we can tell we failed because the file already existed.
 631          */
 632         if (req->rq_status == -EEXIST &&
 633             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 634              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 635                 GOTO(out_dput, 0);
 636         }
 637
 638         /* If we didn't get as far as trying to open, then some locking thing
 639          * probably went wrong, and we'll just bail here.
 640          */
 641         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 642                 GOTO(out_dput, 0);
 643
 644         /* If we failed, then we must have failed opening, so don't look for
 645          * file descriptor or anything, just give the client the bad news.
 646          */
 647         if (req->rq_status)
 648                 GOTO(out_dput, 0);
 649
 650         mfd = NULL;
 651         list_for_each(t, &med->med_open_head) {
 652                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 653                 if (mfd->mfd_xid == req->rq_xid)
 654                         break;
 655                 mfd = NULL;
 656         }
 657
 658         /* #warning "XXX fixme" bug 2991 */
 659         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 660          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 661          * to detect a re-open */
 662         if (mfd == NULL) {
 663                 mntget(mds->mds_vfsmnt);
 664                 CERROR("Re-opened file \n");
 665                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 666                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 667                 if (!mfd) {
 668                         CERROR("mds: out of memory\n");
 669                         GOTO(out_dput, req->rq_status = -ENOMEM);
 670                 }
 671                 put_child = 0;
 672         } else {
 673                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 674                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 675                        mfd->mfd_handle.h_cookie);
 676         }
 677
 678  out_dput:
 679         if (put_child)
 680                 l_dput(dchild);
 681         if (parent)
 682                 l_dput(parent);
 683         EXIT;
 684 }
 685
 686 /* do NOT or the MAY_*'s, you'll get the weakest */
 687 static int accmode(int flags)
 688 {
 689         int res = 0;
 690
 691         if (flags & FMODE_READ)
 692                 res = MAY_READ;
 693         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 694                 res |= MAY_WRITE;
 695         if (flags & FMODE_EXEC)
 696                 res = MAY_EXEC;
 697         return res;
 698 }
 699
 700 /* Handles object creation, actual opening, and I/O epoch */
 701 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 702                            struct mds_body *body, int flags, void **handle,
 703                            struct mds_update_record *rec, struct ldlm_reply *rep)
 704 {
 705         struct obd_device *obd = req->rq_export->exp_obd;
 706         struct mds_obd *mds = mds_req2mds(req);
 707         struct mds_file_data *mfd = NULL;
 708         obd_id *ids = NULL;
 709         unsigned mode;
 710         int rc = 0;
 711         ENTRY;
 712
 713         /* atomically create objects if necessary */
 714         down(&dchild->d_inode->i_sem);
 715         mode = dchild->d_inode->i_mode;
 716
 717         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 718             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 719                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 720                                  dchild->d_inode, 0, 0);
 721                 if (rc) {
 722                         up(&dchild->d_inode->i_sem);
 723                         RETURN(rc);
 724                 }
 725         }
 726
 727         if (rec != NULL) {
 728                 if ((body->valid & OBD_MD_FLEASIZE) &&
 729                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 730                         up(&dchild->d_inode->i_sem);
 731                         RETURN(-EEXIST);
 732                 }
 733
 734                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 735                         /* no EA: create objects */
 736                         rc = mds_create_objects(obd, req, 2, rec,
 737                                                 dchild, handle, &ids);
 738                         if (rc) {
 739                                 CERROR("mds_create_object: rc = %d\n", rc);
 740                                 up(&dchild->d_inode->i_sem);
 741                                 RETURN(rc);
 742                         }
 743                 }
 744
 745                 if (S_ISREG(dchild->d_inode->i_mode) &&
 746                     (body->valid & OBD_MD_FLEASIZE)) {
 747                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 748                                                    req->rq_repmsg, 2);
 749                         if (!rc)
 750                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 751                                                  dchild->d_inode, 0, 0);
 752                         if (rc) {
 753                                 up(&dchild->d_inode->i_sem);
 754                                 RETURN(rc);
 755                         }
 756                 }
 757         }
 758
 759         rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
 760         if (rc < 0) {
 761                 CERROR("mds_pack_acl: rc = %d\n", rc);
 762                 up(&dchild->d_inode->i_sem);
 763                 RETURN(rc);
 764         }
 765
 766         /* If the inode has no EA data, then MDSs hold size, mtime */
 767         if (S_ISREG(dchild->d_inode->i_mode) &&
 768             !(body->valid & OBD_MD_FLEASIZE)) {
 769                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 770                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 771         }
 772
 773         up(&dchild->d_inode->i_sem);
 774
 775         intent_set_disposition(rep, DISP_OPEN_OPEN);
 776         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 777         if (IS_ERR(mfd))
 778                 RETURN(PTR_ERR(mfd));
 779
 780         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 781                mfd->mfd_handle.h_cookie);
 782
 783         if (ids != NULL) {
 784                 mds_dt_update_objids(obd, ids);
 785                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 786         }
 787
 788         RETURN(rc);
 789 }
 790
 791 static int mds_open_by_id(struct ptlrpc_request *req,
 792                           struct lustre_id *id,
 793                           struct mds_body *body, int flags,
 794                           struct mds_update_record *rec,
 795                           struct ldlm_reply *rep)
 796 {
 797         struct mds_obd *mds = mds_req2mds(req);
 798         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 799         struct dentry *dchild;
 800         char idname[LL_ID_NAMELEN];
 801         int idlen = 0, rc;
 802         void *handle = NULL;
 803         ENTRY;
 804
 805         down(&pending_dir->i_sem);
 806         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 807         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 808                                 idlen);
 809         if (IS_ERR(dchild)) {
 810                 up(&pending_dir->i_sem);
 811                 rc = PTR_ERR(dchild);
 812                 CERROR("error looking up %s in PENDING: rc = %d\n",
 813                        idname, rc);
 814                 RETURN(rc);
 815         }
 816
 817         up(&pending_dir->i_sem);
 818         if (dchild->d_inode != NULL) {
 819                 mds_inode_set_orphan(dchild->d_inode);
 820                 mds_pack_inode2body(req2obd(req), body,
 821                                     dchild->d_inode, 1);
 822
 823                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 824                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 825                 CWARN("Orphan %s found and opened in PENDING directory\n",
 826                        idname);
 827                 goto open;
 828         }
 829         l_dput(dchild);
 830
 831         /*
 832          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 833          * regular inode that was previously created.
 834          */
 835         dchild = mds_id2dentry(req2obd(req), id, NULL);
 836         if (IS_ERR(dchild))
 837                 RETURN(PTR_ERR(dchild));
 838
 839         mds_pack_inode2body(req2obd(req), body,
 840                             dchild->d_inode, 1);
 841
 842         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 843         intent_set_disposition(rep, DISP_LOOKUP_POS);
 844
 845  open:
 846         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 847         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 848                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 849         /* XXX what do we do here if mds_finish_transno itself failed? */
 850         l_dput(dchild);
 851         RETURN(rc);
 852 }
 853
 854 int mds_pin(struct ptlrpc_request *req, int offset)
 855 {
 856         struct obd_device *obd = req->rq_export->exp_obd;
 857         struct mds_body *request_body, *reply_body;
 858         struct lvfs_run_ctxt saved;
 859         int rc, size = sizeof(*reply_body);
 860         ENTRY;
 861
 862         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 863                                       sizeof(*request_body));
 864
 865         rc = lustre_pack_reply(req, 1, &size, NULL);
 866         if (rc)
 867                 RETURN(rc);
 868         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 869
 870         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 871         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 872                             request_body->flags, NULL, NULL);
 873         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 874
 875         RETURN(rc);
 876 }
 877
 878 /*
 879  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 880  * child_lockh is NULL we just get the lock as a barrier to wait for other
 881  * holders of this lock, and drop it right away again.
 882  */
 883 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 884                        struct lustre_handle *child_lockh)
 885 {
 886         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 887         struct lustre_handle lockh;
 888         int lock_flags = 0;
 889         int rc;
 890         ENTRY;
 891
 892         if (child_lockh == NULL)
 893                 child_lockh = &lockh;
 894
 895         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 896                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 897                               mds_blocking_ast, ldlm_completion_ast, NULL,
 898                               NULL, NULL, 0, NULL, child_lockh);
 899         if (rc != ELDLM_OK)
 900                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 901         else if (child_lockh == &lockh)
 902                 ldlm_lock_decref(child_lockh, LCK_EX);
 903
 904         RETURN(rc);
 905 }
 906
 907 int mds_open(struct mds_update_record *rec, int offset,
 908              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 909 {
 910         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 911         struct obd_device *obd = req->rq_export->exp_obd;
 912         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 913         struct mds_obd *mds = mds_req2mds(req);
 914         struct ldlm_reply *rep = NULL;
 915         struct mds_body *body = NULL;
 916         struct dentry *dchild = NULL, *dparent = NULL;
 917         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 918         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 919         int parent_mode = LCK_PR;
 920         void *handle = NULL;
 921         struct dentry_params dp;
 922         struct mea *mea = NULL;
 923         int mea_size, update_mode;
 924         int child_mode = LCK_PR;
 925         /* Always returning LOOKUP lock if open succesful to guard
 926            dentry on client. */
 927         ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
 928         struct ldlm_res_id child_res_id = { .name = {0}};
 929         int lock_flags = 0;
 930         ENTRY;
 931
 932         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 933                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 934                   rec->ur_mode);
 935
 936         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 937                          (obd_timeout + 1) / 4);
 938
 939         if (offset == 3) { /* intent */
 940                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 941                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 942         } else if (offset == 1) { /* non-intent reint */
 943                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 944         } else {
 945                 body = NULL;
 946                 LBUG();
 947         }
 948
 949         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 950
 951         /*
 952          * Step 0: If we are passed a id, then we assume the client already
 953          * opened this file and is only replaying the RPC, so we open the inode
 954          * by fid (at some large expense in security).
 955          */
 956         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 957                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 958                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 959                        rec->ur_name, rec->ur_mode);
 960
 961                 LASSERT(id_ino(rec->ur_id2));
 962
 963                 rc = mds_open_by_id(req, rec->ur_id2, body,
 964                                     rec->ur_flags, rec, rep);
 965                 if (rc != -ENOENT) {
 966                         mds_body_do_reverse_map(med, body);
 967                         RETURN(rc);
 968                 }
 969
 970                 /* We didn't find the correct inode on disk either, so we
 971                  * need to re-create it via a regular replay. */
 972                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 973         } else {
 974                 LASSERT(!id_ino(rec->ur_id2));
 975         }
 976
 977         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 978
 979         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 980                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 981                 RETURN(-ENOMEM);
 982         }
 983
 984         acc_mode = accmode(rec->ur_flags);
 985
 986         /* Step 1: Find and lock the parent */
 987         if (rec->ur_flags & MDS_OPEN_CREAT) {
 988                 /* XXX Well, in fact we only need this lock mode change if
 989                    in addition to O_CREAT, the file does not exist.
 990                    But we do not know if it exists or not yet */
 991                 parent_mode = LCK_PW;
 992         }
 993
 994         if (rec->ur_namelen == 1) {
 995                 /* client (LMV) wants to open the file by id */
 996                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 997                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 998                 if (IS_ERR(dchild)) {
 999                         rc = PTR_ERR(dparent);
1000                         CERROR("child lookup by an id error %d\n", rc);
1001                         GOTO(cleanup, rc);
1002                 }
1003                 goto got_child;
1004         }
1005
1006 restart:
1007         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1008                                        parent_lockh, &update_mode, rec->ur_name,
1009                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1010         if (IS_ERR(dparent)) {
1011                 rc = PTR_ERR(dparent);
1012                 if (rc != -ENOENT) {
1013                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1014                                OLID4(rec->ur_id1), rc);
1015                 } else {
1016                         /* Just cannot find parent - make it look like
1017                            usual negative lookup to avoid extra MDS RPC */
1018                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1019                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1020                 }
1021                 GOTO(cleanup, rc);
1022         }
1023         LASSERT(dparent->d_inode != NULL);
1024
1025         cleanup_phase = 1; /* parent dentry and lock */
1026
1027         /* try to retrieve MEA data for this dir */
1028         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1029         if (rc)
1030                 GOTO(cleanup, rc);
1031
1032         if (mea != NULL) {
1033                 /*
1034                  * dir is already splitted, check is requested filename should *
1035                  * live at this MDS or at another one.
1036                  */
1037                 int i;
1038                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1039                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1040                         CDEBUG(D_OTHER,
1041                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1042                                " should be %lu(%d)\n", obd->obd_name,
1043                                mea->mea_master, dparent->d_inode->i_ino,
1044                                dparent->d_inode->i_generation, rec->ur_name,
1045                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1046                         GOTO(cleanup, rc = -ERESTART);
1047                 }
1048         }
1049
1050         /* Step 2: Lookup the child */
1051         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1052         if (IS_ERR(dchild)) {
1053                 rc = PTR_ERR(dchild);
1054                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1055                 GOTO(cleanup, rc);
1056         }
1057
1058 got_child:
1059         cleanup_phase = 2; /* child dentry */
1060
1061         if (dchild->d_flags & DCACHE_CROSS_REF) {
1062                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1063                 LASSERT(rec->ur_namelen > 1);
1064
1065                 /* we're gonna acquire LOOKUP lock on the child,
1066                  * but we have already locked parent and our order
1067                  * may conflict with enqueue_order_locks(). so,
1068                  * drop parent lock and acquire both the locks in
1069                  * common order. bug 6190 */
1070 #ifdef S_PDIROPS
1071                 if (parent_lockh[1].cookie != 0)
1072                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1073 #endif
1074                 ldlm_lock_decref(parent_lockh, parent_mode);
1075                 l_dput(dchild);
1076                 l_dput(dparent);
1077                 cleanup_phase = 0;
1078
1079                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1080                                                  parent_lockh, &dparent,
1081                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1082                                                  &update_mode, rec->ur_name,
1083                                                  rec->ur_namelen, child_lockh,
1084                                                  &dchild, LCK_PR,
1085                                                  MDS_INODELOCK_LOOKUP);
1086
1087                 if (rc)
1088                         GOTO(cleanup, rc);
1089
1090 #ifdef S_PDIROPS
1091                 if (parent_lockh[1].cookie != 0)
1092                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1093 #endif
1094                 ldlm_lock_decref(parent_lockh, LCK_PR);
1095
1096                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1097                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1098                                dchild->d_inode);
1099                         if (dchild->d_inode)
1100                                 ldlm_lock_decref(child_lockh, LCK_PR);
1101                         l_dput(dchild);
1102                         l_dput(dparent);
1103                         GOTO(restart, rc);
1104                 }
1105
1106                 mds_pack_dentry2body(obd, body, dchild, 1);
1107                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1108                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1109
1110                 if (mea)
1111                         OBD_FREE(mea, mea_size);
1112                 l_dput(dchild);
1113                 l_dput(dparent);
1114                 RETURN(rc);
1115         }
1116
1117         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1118
1119         if (dchild->d_inode)
1120                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1121         else
1122                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1123
1124         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1125         if (dchild->d_inode == NULL) {
1126                 unsigned long ino;
1127                 struct iattr iattr;
1128                 struct inode *inode;
1129
1130                 ino = (unsigned long)id_ino(rec->ur_id2);
1131
1132                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1133                                           update_mode);
1134
1135                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1136                        obd->obd_name, dparent->d_inode->i_ino,
1137                        dparent->d_inode->i_generation, rc);
1138
1139                 if (rc > 0) {
1140                         /* dir got splitted */
1141                         GOTO(cleanup, rc = -ERESTART);
1142                 } else if (rc < 0) {
1143                         /* error happened during spitting */
1144                         GOTO(cleanup, rc);
1145                 }
1146
1147                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1148                         /*
1149                          * dentry is negative and we weren't supposed to create
1150                          * object, get out of here.
1151                          */
1152                         GOTO(cleanup, rc = -ENOENT);
1153                 }
1154
1155                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1156                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1157                                       NULL);
1158                 if (IS_ERR(handle)) {
1159                         rc = PTR_ERR(handle);
1160                         handle = NULL;
1161                         GOTO(cleanup, rc);
1162                 }
1163                 dchild->d_fsdata = (void *) &dp;
1164                 dp.p_ptr = req;
1165                 dp.p_inum = ino;
1166
1167                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1168                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1169                         dchild->d_fsdata = NULL;
1170
1171                 if (rc) {
1172                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1173                         GOTO(cleanup, rc);
1174                 }
1175                 inode = dchild->d_inode;
1176                 if (ino) {
1177                         LASSERT(ino == inode->i_ino);
1178
1179                         /* written as part of setattr */
1180                         inode->i_generation = id_gen(rec->ur_id2);
1181                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1182                                inode->i_ino, inode->i_generation);
1183                 }
1184
1185                 created = 1;
1186                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1187                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1188                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1189
1190                 iattr.ia_uid = rec->ur_fsuid;
1191                 if (dparent->d_inode->i_mode & S_ISGID)
1192                         iattr.ia_gid = dparent->d_inode->i_gid;
1193                 else
1194                         iattr.ia_gid = rec->ur_fsgid;
1195
1196                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1197                         ATTR_MTIME | ATTR_CTIME;
1198
1199                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1200                 if (rc)
1201                         CERROR("error on child setattr: rc = %d\n", rc);
1202
1203                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1204
1205                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1206                 if (rc)
1207                         CERROR("error on parent setattr: rc = %d\n", rc);
1208                 else {
1209                         MD_COUNTER_INCREMENT(obd, create);
1210                 }
1211
1212                 if (ino) {
1213                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1214                                                   handle, rec->ur_id2);
1215                         if (rc) {
1216                                 CERROR("mds_update_inode_sid() failed, "
1217                                        "rc = %d\n", rc);
1218                         }
1219                         id_assign_fid(&body->id1, rec->ur_id2);
1220
1221                         /*
1222                          * make sure, that fid is up-to-date.
1223                          */
1224                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1225                 } else {
1226                         rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1227                                                  handle, &body->id1);
1228                         if (rc) {
1229                                 CERROR("mds_alloc_inode_sid() failed, "
1230                                        "rc = %d\n", rc);
1231                         }
1232                 }
1233
1234                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1235                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1236                                            dchild->d_inode, handle,
1237                                            req->rq_export->exp_sync);
1238                         handle = NULL;
1239                 }
1240
1241                 acc_mode = 0;           /* Don't check for permissions */
1242         }
1243         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1244
1245         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1246                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1247                  dchild->d_name.name, dchild, dchild->d_inode);
1248
1249         if (S_ISREG(dchild->d_inode->i_mode)) {
1250                 /* Check permissions etc */
1251                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1252                 if (rc != 0)
1253                         GOTO(cleanup, rc);
1254
1255                 /* Can't write to a read-only file */
1256                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1257                         GOTO(cleanup, rc = -EPERM);
1258
1259                 /* An append-only file must be opened in append mode for
1260                  * writing */
1261                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1262                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1263                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1264                         GOTO(cleanup, rc = -EPERM);
1265         }
1266
1267         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1268             (rec->ur_flags & MDS_OPEN_EXCL)) {
1269                 /* File already exists, we didn't just create it, and we
1270                  * were passed O_EXCL; err-or. */
1271                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1272         }
1273
1274         if (S_ISDIR(dchild->d_inode->i_mode)) {
1275                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1276                     rec->ur_flags & FMODE_WRITE) {
1277                         /* we are trying to create or write a exist dir */
1278                         GOTO(cleanup, rc = -EISDIR);
1279                 }
1280                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1281                         GOTO(cleanup, rc = -EACCES);
1282
1283                 /*
1284                  * here was checking for possible GNS mount points to skip their
1285                  * open. I removed it as its detection based only on SUID bit is
1286                  * not reliable. Opening such a dirs should not cause any
1287                  * problems, at least tests show that. --umka
1288                  */
1289         }
1290
1291         /* if we are following a symlink, don't open */
1292         if (S_ISLNK(dchild->d_inode->i_mode))
1293                 GOTO(cleanup_no_trans, rc = 0);
1294
1295         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1296             !S_ISDIR(dchild->d_inode->i_mode))
1297                 GOTO(cleanup, rc = -ENOTDIR);
1298
1299         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1300                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1301                 GOTO(cleanup, rc = -EAGAIN);
1302         }
1303
1304 #warning "disable opencache lock for CMD2"
1305 #if 0
1306         /* Obtain OPEN lock as well */
1307         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1308
1309         /* We cannot use acc_mode here, because it is zeroed in case of
1310            creating a file, so we get wrong lockmode */
1311         if (accmode(rec->ur_flags) & MAY_WRITE)
1312                 child_mode = LCK_CW;
1313         else if (accmode(rec->ur_flags) & MAY_EXEC)
1314                 child_mode = LCK_PR;
1315         else
1316                 child_mode = LCK_CR;
1317
1318         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1319                 /* In case of replay we do not get a lock assuming that the
1320                    caller has it already */
1321                 child_res_id.name[0] = id_fid(&body->id1);
1322                 child_res_id.name[1] = id_group(&body->id1);
1323
1324                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1325                                       child_res_id, LDLM_IBITS, &policy,
1326                                       child_mode, &lock_flags,
1327                                       mds_blocking_ast, ldlm_completion_ast,
1328                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1329                 if (rc != ELDLM_OK)
1330                         GOTO(cleanup, rc);
1331
1332                 cleanup_phase = 3;
1333         }
1334 #endif
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                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1453
1454                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1455                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1456                         CERROR("found \"orphan\" %s %s with link count %d\n",
1457                                S_ISREG(inode->i_mode) ? "file" : "dir",
1458                                idname, inode->i_nlink);
1459                 /*
1460                  * sadly, there is no easy way to save pending_child from
1461                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1462                  * normally it will still be in the dcache.
1463                  */
1464                 down(&pending_dir->i_sem);
1465                 cleanup_phase = 1; /* up(i_sem) when finished */
1466                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1467                                                idlen);
1468                 if (IS_ERR(pending_child))
1469                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1470                 if (pending_child->d_inode == NULL) {
1471                         CERROR("orphan %s has been removed\n", idname);
1472                         GOTO(cleanup, rc = 0);
1473                 }
1474
1475                 cleanup_phase = 2; /* dput(pending_child) when finished */
1476                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1477                         rc = vfs_rmdir(pending_dir, pending_child);
1478                         if (rc)
1479                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1480                                        idname, rc);
1481                         goto out;
1482                 }
1483
1484                 if (req != NULL && req->rq_repmsg != NULL) {
1485                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1486                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1487                 }
1488
1489                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1490                                           NULL, stripe_count);
1491                 if (IS_ERR(handle)) {
1492                         rc = PTR_ERR(handle);
1493                         handle = NULL;
1494                         GOTO(cleanup, rc);
1495                 }
1496
1497                 pending_child->d_fsdata = (void *) &dp;
1498                 dp.p_inum = 0;
1499                 dp.p_ptr = req;
1500                 rc = vfs_unlink(pending_dir, pending_child);
1501                 if (rc)
1502                         CERROR("error unlinking orphan %s: rc %d\n",
1503                                idname, rc);
1504
1505                 if (req != NULL && req->rq_repmsg != NULL &&
1506                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1507                     mds_log_op_unlink(obd, pending_child->d_inode,
1508                                       lmm, req->rq_repmsg->buflens[1],
1509                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1510                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1511                         reply_body->valid |= OBD_MD_FLCOOKIE;
1512                 }
1513
1514                 rc = mds_destroy_object(obd, inode, 1);
1515                 if (rc) {
1516                         CERROR("cannot destroy OSS object on close, err %d\n",
1517                                rc);
1518                         rc = 0;
1519                 }
1520
1521                 goto out; /* Don't bother updating attrs on unlinked inode */
1522         }
1523
1524 #if 0
1525         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1526                 /* Update the on-disk attributes if this was the last write
1527                  * close, and all information was provided (i.e., rc == 0)
1528                  *
1529                  * XXX this should probably be abstracted with mds_reint_setattr
1530                  */
1531
1532                 if (request_body->valid & OBD_MD_FLMTIME &&
1533                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1534                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1535                         iattr.ia_valid |= ATTR_MTIME;
1536                 }
1537                 if (request_body->valid & OBD_MD_FLCTIME &&
1538                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1539                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1540                         iattr.ia_valid |= ATTR_CTIME;
1541                 }
1542
1543                 /* XXX can't set block count with fsfilt_setattr (!) */
1544                 if (request_body->valid & OBD_MD_FLSIZE) {
1545                         iattr.ia_valid |= ATTR_SIZE;
1546                         iattr.ia_size = request_body->size;
1547                 }
1548                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1549                         iattr.ia_valid |= ATTR_BLOCKS;
1550                         iattr.ia_blocks = request_body->blocks
1551                 } */
1552
1553         }
1554 #endif
1555         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1556                 /* Only start a transaction to write out only the atime if
1557                  * it is more out-of-date than the specified limit.  If we
1558                  * are already going to write out the atime then do it anyway.
1559                  * */
1560                 if ((request_body->atime >
1561                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1562                     (iattr.ia_valid != 0 &&
1563                      request_body->atime > LTIME_S(inode->i_atime))) {
1564                         LTIME_S(iattr.ia_atime) = request_body->atime;
1565                         iattr.ia_valid |= ATTR_ATIME;
1566                 }
1567         }
1568
1569         if (iattr.ia_valid != 0) {
1570                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1571                 if (IS_ERR(handle))
1572                         GOTO(cleanup, rc = PTR_ERR(handle));
1573                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1574                 if (rc)
1575                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1576         }
1577 out:
1578         /* If other clients have this file open for write, rc will be > 0 */
1579         if (rc > 0)
1580                 rc = 0;
1581         l_dput(mfd->mfd_dentry);
1582         mds_mfd_destroy(mfd);
1583
1584  cleanup:
1585         atomic_dec(&mds->mds_open_count);
1586         if (req != NULL && reply_body != NULL) {
1587                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1588         } else if (handle) {
1589                 int err, force_sync = 0;
1590
1591                 if (req && req->rq_export)
1592                         force_sync = req->rq_export->exp_sync;
1593
1594                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1595                                     force_sync);
1596                 if (err) {
1597                         CERROR("error committing close: %d\n", err);
1598                         if (!rc)
1599                                 rc = err;
1600                 }
1601         }
1602
1603         switch (cleanup_phase) {
1604         case 2:
1605                 if (lcl != NULL)
1606                         ptlrpc_save_llog_lock(req, lcl);
1607                 dput(pending_child);
1608         case 1:
1609                 up(&pending_dir->i_sem);
1610         }
1611         RETURN(rc);
1612 }
1613
1614 int mds_close(struct ptlrpc_request *req, int offset)
1615 {
1616         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1617         struct obd_device *obd = req->rq_export->exp_obd;
1618         struct mds_body *body;
1619         struct mds_file_data *mfd;
1620         struct lvfs_run_ctxt saved;
1621         struct inode *inode;
1622         int rc, repsize[3] = {sizeof(struct mds_body),
1623                               obd->u.mds.mds_max_mdsize,
1624                               obd->u.mds.mds_max_cookiesize};
1625         ENTRY;
1626         MD_COUNTER_INCREMENT(obd, close);
1627
1628         rc = lustre_pack_reply(req, 3, repsize, NULL);
1629         if (rc) {
1630                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1631                 req->rq_status = rc;
1632         } else {
1633                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1634         }
1635
1636         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1637                 DEBUG_REQ(D_HA, req, "close replay");
1638                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1639                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1640                        req->rq_repmsg->buflens[2]);
1641         }
1642
1643         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1644                                   lustre_swab_mds_body);
1645         if (body == NULL) {
1646                 CERROR("Can't unpack body\n");
1647                 req->rq_status = -EFAULT;
1648                 RETURN(-EFAULT);
1649         }
1650
1651         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1652                 /* do some stuff */ ;
1653
1654         mfd = mds_handle2mfd(&body->handle);
1655         if (mfd == NULL) {
1656                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1657                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1658                 req->rq_status = -ESTALE;
1659                 RETURN(-ESTALE);
1660         }
1661
1662         inode = mfd->mfd_dentry->d_inode;
1663         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1664         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1665         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1666                 struct mds_body *rep_body;
1667
1668                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1669                                           sizeof (*rep_body));
1670                 LASSERT(rep_body != NULL);
1671
1672                 mds_pack_inode2body(obd, rep_body, inode,
1673                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1674
1675                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1676                             inode, MDS_PACK_MD_LOCK, 0);
1677         }
1678         spin_lock(&med->med_open_lock);
1679         list_del(&mfd->mfd_list);
1680         spin_unlock(&med->med_open_lock);
1681
1682         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1683         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1684         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1685
1686         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1687                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1688                 req->rq_status = -ENOMEM;
1689                 mds_mfd_put(mfd);
1690                 RETURN(-ENOMEM);
1691         }
1692
1693         mds_mfd_put(mfd);
1694         RETURN(0);
1695 }
1696
1697 int mds_done_writing(struct ptlrpc_request *req, int offset)
1698 {
1699         struct mds_body *body;
1700         int rc, size = sizeof(struct mds_body);
1701         ENTRY;
1702
1703         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1704
1705         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1706                                   lustre_swab_mds_body);
1707         if (body == NULL) {
1708                 CERROR("Can't unpack body\n");
1709                 req->rq_status = -EFAULT;
1710                 RETURN(-EFAULT);
1711         }
1712
1713         rc = lustre_pack_reply(req, 1, &size, NULL);
1714         if (rc) {
1715                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1716                 req->rq_status = rc;
1717         }
1718
1719         RETURN(0);
1720 }