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/lustre_mds.h>
  44 #include <linux/obd_lov.h>
  45 #include <linux/lustre_fsfilt.h>
  46 #include <linux/lprocfs_status.h>
  47
  48 #include "mds_internal.h"
  49
  50 /* Exported function from this file are:
  51  *
  52  * mds_open - called by the intent handler
  53  * mds_close - an rpc handling function
  54  * mds_pin - an rpc handling function - which will go away
  55  * mds_mfd_close - for force closing files when a client dies
  56  */
  57
  58 /*
  59  * MDS file data handling: file data holds a handle for a file opened
  60  * by a client.
  61  */
  62
  63 static void mds_mfd_addref(void *mfdp)
  64 {
  65         struct mds_file_data *mfd = mfdp;
  66
  67         atomic_inc(&mfd->mfd_refcount);
  68         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  69                atomic_read(&mfd->mfd_refcount));
  70 }
  71
  72 struct mds_file_data *mds_mfd_new(void)
  73 {
  74         struct mds_file_data *mfd;
  75
  76         OBD_ALLOC(mfd, sizeof *mfd);
  77         if (mfd == NULL) {
  78                 CERROR("mds: out of memory\n");
  79                 return NULL;
  80         }
  81
  82         atomic_set(&mfd->mfd_refcount, 2);
  83
  84         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  85         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  86
  87         return mfd;
  88 }
  89
  90 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  91 {
  92         ENTRY;
  93         LASSERT(handle != NULL);
  94         RETURN(class_handle2object(handle->cookie));
  95 }
  96
  97 static void mds_mfd_put(struct mds_file_data *mfd)
  98 {
  99         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 100                atomic_read(&mfd->mfd_refcount) - 1);
 101         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 102                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 103         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 104                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 105                 OBD_FREE(mfd, sizeof *mfd);
 106         }
 107 }
 108
 109 static void mds_mfd_destroy(struct mds_file_data *mfd)
 110 {
 111         class_handle_unhash(&mfd->mfd_handle);
 112         mds_mfd_put(mfd);
 113 }
 114
 115
 116 /* Caller must hold mds->mds_epoch_sem */
 117 static int mds_alloc_filterdata(struct inode *inode)
 118 {
 119         LASSERT(inode->i_filterdata == NULL);
 120         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 121         if (inode->i_filterdata == NULL)
 122                 return -ENOMEM;
 123         LASSERT(igrab(inode) == inode);
 124         return 0;
 125 }
 126
 127 /* Caller must hold mds->mds_epoch_sem */
 128 static void mds_free_filterdata(struct inode *inode)
 129 {
 130         LASSERT(inode->i_filterdata != NULL);
 131         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 132         inode->i_filterdata = NULL;
 133         iput(inode);
 134 }
 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
 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         if (epoch > mds->mds_io_epoch)
 174                 mds->mds_io_epoch = epoch;
 175         else
 176                 mds->mds_io_epoch++;
 177         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 178         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 179                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 180  out:
 181         if (rc == 0)
 182                 atomic_inc(&inode->i_writecount);
 183         up(&mds->mds_epoch_sem);
 184         return rc;
 185 }
 186
 187 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 188  * again -- which is never true if the file is about to be unlinked.  Otherwise
 189  * returns the number of remaining writers. */
 190 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 191                                 struct mds_body *body, int unlinking)
 192 {
 193         int rc = 0;
 194         ENTRY;
 195
 196         down(&mds->mds_epoch_sem);
 197         atomic_dec(&inode->i_writecount);
 198         rc = atomic_read(&inode->i_writecount);
 199         if (rc > 0)
 200                 GOTO(out, rc);
 201 #if 0
 202         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 203                 GOTO(out, rc = EAGAIN);
 204 #endif
 205         mds_free_filterdata(inode);
 206  out:
 207         up(&mds->mds_epoch_sem);
 208         return rc;
 209 }
 210
 211 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 212 {
 213         ENTRY;
 214         down(&mds->mds_epoch_sem);
 215         if (atomic_read(&inode->i_writecount) > 0) {
 216                 up(&mds->mds_epoch_sem);
 217                 RETURN(-ETXTBSY);
 218         }
 219         atomic_dec(&inode->i_writecount);
 220         up(&mds->mds_epoch_sem);
 221         RETURN(0);
 222 }
 223
 224 static void mds_allow_write_access(struct inode *inode)
 225 {
 226         ENTRY;
 227         atomic_inc(&inode->i_writecount);
 228 }
 229
 230 int mds_query_write_access(struct inode *inode)
 231 {
 232         ENTRY;
 233         RETURN(atomic_read(&inode->i_writecount));
 234 }
 235
 236 /* This replaces the VFS mds_dentry_open, it manages mfd and writecount */
 237 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 238                                              struct vfsmount *mnt, int flags,
 239                                              struct ptlrpc_request *req)
 240 {
 241         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 242         struct mds_obd *mds = mds_req2mds(req);
 243         struct mds_file_data *mfd;
 244         struct mds_body *body;
 245         int error;
 246         ENTRY;
 247
 248         mfd = mds_mfd_new();
 249         if (mfd == NULL) {
 250                 CERROR("mds: out of memory\n");
 251                 GOTO(cleanup_dentry, error = -ENOMEM);
 252         }
 253
 254         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 255
 256         if (flags & FMODE_WRITE) {
 257                 /* FIXME: in recovery, need to pass old epoch here */
 258                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 259                 if (error)
 260                         GOTO(cleanup_mfd, error);
 261                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 262         } else if (flags & FMODE_EXEC) {
 263                 error = mds_deny_write_access(mds, dentry->d_inode);
 264                 if (error)
 265                         GOTO(cleanup_mfd, error);
 266         }
 267
 268         dget(dentry);
 269
 270         /* Mark the file as open to handle open-unlink. */
 271         mds_open_orphan_inc(dentry->d_inode);
 272
 273         mfd->mfd_mode = flags;
 274         mfd->mfd_dentry = dentry;
 275         mfd->mfd_xid = req->rq_xid;
 276
 277         spin_lock(&med->med_open_lock);
 278         list_add(&mfd->mfd_list, &med->med_open_head);
 279         spin_unlock(&med->med_open_lock);
 280         mds_mfd_put(mfd);
 281
 282         body->handle.cookie = mfd->mfd_handle.h_cookie;
 283
 284         RETURN(mfd);
 285
 286 cleanup_mfd:
 287         mds_mfd_put(mfd);
 288         mds_mfd_destroy(mfd);
 289 cleanup_dentry:
 290         return ERR_PTR(error);
 291 }
 292
 293 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 294                                 struct lov_desc *desc)
 295 {
 296         int i;
 297         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 298                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 299                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 300         }
 301 }
 302
 303 /* Must be called with i_sem held */
 304 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 305                               struct mds_update_record *rec,
 306                               struct mds_obd *mds, struct obd_device *obd,
 307                               struct dentry *dchild, void **handle,
 308                               obd_id **ids)
 309 {
 310         struct obdo *oa;
 311         struct obd_trans_info oti = { 0 };
 312         struct mds_body *body;
 313         struct lov_stripe_md *lsm = NULL;
 314         struct lov_mds_md *lmm = NULL;
 315         struct inode *inode = dchild->d_inode;
 316         void *lmm_buf;
 317         int rc, lmm_bufsize, lmm_size;
 318         ENTRY;
 319
 320         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 321             !(rec->ur_flags & FMODE_WRITE))
 322                 RETURN(0);
 323
 324         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 325
 326         if (!S_ISREG(inode->i_mode))
 327                 RETURN(0);
 328         if (body->valid & OBD_MD_FLEASIZE)
 329                 RETURN(0);
 330
 331         OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 332         if (*ids == NULL)
 333                 RETURN(-ENOMEM);
 334         oti.oti_objid = *ids;
 335
 336         if (*handle == NULL)
 337                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 338         if (IS_ERR(*handle)) {
 339                 rc = PTR_ERR(*handle);
 340                 *handle = NULL;
 341                 GOTO(out_ids, rc);
 342         }
 343
 344         /* replay case */
 345         if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 346                 LASSERT (rec->ur_fid2->id);
 347                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 348                 lmm_size = rec->ur_eadatalen;
 349                 lmm = rec->ur_eadata;
 350                 LASSERT(lmm);
 351
 352                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 353
 354                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 355                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 356                 LASSERT(lmm_buf);
 357                 LASSERT(lmm_bufsize >= lmm_size);
 358                 memcpy(lmm_buf, lmm, lmm_size);
 359                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 360                 if (rc)
 361                         CERROR("open replay failed to set md:%d\n", rc);
 362                 RETURN(0);
 363         }
 364
 365         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 366                 GOTO(out_ids, rc = -ENOMEM);
 367
 368         oa = obdo_alloc();
 369         if (oa == NULL)
 370                 GOTO(out_ids, rc = -ENOMEM);
 371         oa->o_mode = S_IFREG | 0600;
 372         oa->o_id = inode->i_ino;
 373         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 374         oa->o_generation = inode->i_generation;
 375         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 376         oa->o_gid = 0;
 377         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
 378                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 379         oa->o_size = 0;
 380
 381         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
 382                         OBD_MD_FLCTIME);
 383
 384         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 385                 /* check if things like lfs setstripe are sending us the ea */
 386                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 387                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 388                                            mds->mds_osc_exp,
 389                                            0, &lsm, rec->ur_eadata);
 390                         if (rc)
 391                                 GOTO(out_oa, rc);
 392                 }
 393                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 394                 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
 395                 if (rc) {
 396                         int level = D_ERROR;
 397                         if (rc == -ENOSPC)
 398                                 level = D_INODE;
 399                         CDEBUG(level, "error creating objects for "
 400                                       "inode %lu: rc = %d\n",
 401                                inode->i_ino, rc);
 402                         if (rc > 0) {
 403                                 CERROR("obd_create returned invalid "
 404                                        "rc %d\n", rc);
 405                                 rc = -EIO;
 406                         }
 407                         GOTO(out_oa, rc);
 408                 }
 409         } else {
 410                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 411                                    0, &lsm, rec->ur_eadata);
 412                 if (rc) {
 413                         GOTO(out_oa, rc);
 414                 }
 415                 lsm->lsm_object_id = oa->o_id;
 416                 lsm->lsm_object_gr = oa->o_gr;
 417         }
 418         if (inode->i_size) {
 419                 oa->o_size = inode->i_size;
 420                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 421                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 422                 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
 423                 if (rc) {
 424                         CERROR("error setting attrs for inode %lu: rc %d\n",
 425                                inode->i_ino, rc);
 426                         if (rc > 0) {
 427                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 428                                 rc = -EIO;
 429                         }
 430                         GOTO(out_oa, rc);
 431                 }
 432         }
 433
 434         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 435         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 436
 437         LASSERT(lsm && lsm->lsm_object_id);
 438         lmm = NULL;
 439         rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
 440         if (!rec->ur_fid2->id)
 441                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 442         LASSERT(rc >= 0);
 443         lmm_size = rc;
 444         body->eadatasize = rc;
 445         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 446         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 447         lmm_bufsize = req->rq_repmsg->buflens[offset];
 448         LASSERT(lmm_buf);
 449         LASSERT(lmm_bufsize >= lmm_size);
 450
 451         memcpy(lmm_buf, lmm, lmm_size);
 452         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 453  out_oa:
 454         oti_free_cookies(&oti);
 455         obdo_free(oa);
 456  out_ids:
 457         if (rc) {
 458                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 459                 *ids = NULL;
 460         }
 461         RETURN(rc);
 462 }
 463
 464 static void reconstruct_open(struct mds_update_record *rec, int offset,
 465                              struct ptlrpc_request *req,
 466                              struct lustre_handle *child_lockh)
 467 {
 468         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 469         struct mds_client_data *mcd = med->med_mcd;
 470         struct mds_obd *mds = mds_req2mds(req);
 471         struct mds_file_data *mfd;
 472         struct obd_device *obd = req->rq_export->exp_obd;
 473         struct dentry *parent, *child;
 474         struct ldlm_reply *rep;
 475         struct mds_body *body;
 476         int rc;
 477         struct list_head *t;
 478         int put_child = 1;
 479         ENTRY;
 480
 481         LASSERT(offset == 2);                  /* only called via intent */
 482         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 483         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 484
 485         /* copy rc, transno and disp; steal locks */
 486         mds_req_from_mcd(req, mcd);
 487         intent_set_disposition(rep, mcd->mcd_last_data);
 488
 489         /* Only replay if create or open actually happened. */
 490         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 491                 EXIT;
 492                 return; /* error looking up parent or child */
 493         }
 494
 495         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 496         LASSERT(!IS_ERR(parent));
 497
 498         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 499         LASSERT(!IS_ERR(child));
 500
 501         if (!child->d_inode) {
 502                 GOTO(out_dput, 0); /* child not present to open */
 503         }
 504
 505         /* At this point, we know we have a child. We'll send
 506          * it back _unless_ it not created and open failed.
 507          */
 508         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 509             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 510             req->rq_status) {
 511                 GOTO(out_dput, 0);
 512         }
 513
 514         /* get lock (write for O_CREAT, read otherwise) */
 515
 516         mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
 517         mds_pack_inode2body(obd, body, child->d_inode);
 518         if (S_ISREG(child->d_inode->i_mode)) {
 519                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 520                                  child->d_inode, 1);
 521
 522                 if (rc)
 523                         LASSERT(rc == req->rq_status);
 524
 525                 /* If we have LOV EA data, the OST holds size, mtime */
 526                 if (!(body->valid & OBD_MD_FLEASIZE))
 527                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 528                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 529         } else {
 530                 /* XXX need to check this case */
 531         }
 532
 533         /* If we're opening a file without an EA, change to a write
 534            lock (unless we already have one). */
 535
 536         /* If we have -EEXIST as the status, and we were asked to create
 537          * exclusively, we can tell we failed because the file already existed.
 538          */
 539         if (req->rq_status == -EEXIST &&
 540             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 541              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 542                 GOTO(out_dput, 0);
 543         }
 544
 545         /* If we didn't get as far as trying to open, then some locking thing
 546          * probably went wrong, and we'll just bail here.
 547          */
 548         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 549                 GOTO(out_dput, 0);
 550
 551         /* If we failed, then we must have failed opening, so don't look for
 552          * file descriptor or anything, just give the client the bad news.
 553          */
 554         if (req->rq_status)
 555                 GOTO(out_dput, 0);
 556
 557         mfd = NULL;
 558         list_for_each(t, &med->med_open_head) {
 559                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 560                 if (mfd->mfd_xid == req->rq_xid)
 561                         break;
 562                 mfd = NULL;
 563         }
 564
 565         /* #warning "XXX fixme" bug 2991 */
 566         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 567          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 568          * to detect a re-open */
 569         if (mfd == NULL) {
 570                 mntget(mds->mds_vfsmnt);
 571                 CERROR("Re-opened file \n");
 572                 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
 573                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 574                 if (!mfd) {
 575                         CERROR("mds: out of memory\n");
 576                         GOTO(out_dput, req->rq_status = -ENOMEM);
 577                 }
 578                 put_child = 0;
 579         } else {
 580                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 581                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 582                        mfd->mfd_handle.h_cookie);
 583         }
 584
 585  out_dput:
 586         if (put_child)
 587                 l_dput(child);
 588         l_dput(parent);
 589         EXIT;
 590 }
 591
 592 /* do NOT or the MAY_*'s, you'll get the weakest */
 593 static int accmode(int flags)
 594 {
 595         int res = 0;
 596
 597         if (flags & FMODE_READ)
 598                 res = MAY_READ;
 599         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 600                 res |= MAY_WRITE;
 601         if (flags & FMODE_EXEC)
 602                 res = MAY_EXEC;
 603         return res;
 604 }
 605
 606 /* Handles object creation, actual opening, and I/O epoch */
 607 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 608                            struct mds_body *body, int flags, void **handle,
 609                            struct mds_update_record *rec,struct ldlm_reply *rep)
 610 {
 611         struct mds_obd *mds = mds_req2mds(req);
 612         struct obd_device *obd = req->rq_export->exp_obd;
 613         struct mds_file_data *mfd = NULL;
 614         obd_id *ids = NULL; /* object IDs created */
 615         unsigned mode;
 616         int rc = 0;
 617         ENTRY;
 618
 619         /* atomically create objects if necessary */
 620         down(&dchild->d_inode->i_sem);
 621         mode = dchild->d_inode->i_mode;
 622         if ((S_ISREG(mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 623             (S_ISDIR(mode) && (body->valid & OBD_MD_FLDIREA))) {
 624                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 625                                  dchild->d_inode, 0);
 626                 if (rc) {
 627                         up(&dchild->d_inode->i_sem);
 628                         RETURN(rc);
 629                 }
 630         }
 631         if (rec != NULL) {
 632                 /* no EA: create objects */
 633                 rc = mds_create_objects(req, 2, rec, mds, obd,
 634                                         dchild, handle, &ids);
 635                 if (rc) {
 636                         CERROR("mds_create_objects: rc = %d\n", rc);
 637                         up(&dchild->d_inode->i_sem);
 638                         RETURN(rc);
 639                 }
 640         }
 641         /* If the inode has EA data, then OSTs hold size, mtime */
 642         if (S_ISREG(dchild->d_inode->i_mode) &&
 643             !(body->valid & OBD_MD_FLEASIZE)) {
 644                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 645                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 646         }
 647         up(&dchild->d_inode->i_sem);
 648
 649         intent_set_disposition(rep, DISP_OPEN_OPEN);
 650         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 651         if (IS_ERR(mfd))
 652                 RETURN(PTR_ERR(mfd));
 653
 654         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 655                mfd->mfd_handle.h_cookie);
 656         if (ids != NULL) {
 657                 mds_lov_update_objids(obd, ids);
 658                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 659         }
 660         //if (rc)
 661         //        mds_mfd_destroy(mfd);
 662         RETURN(rc);
 663 }
 664
 665 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 666                            struct mds_body *body, int flags,
 667                            struct mds_update_record *rec,struct ldlm_reply *rep)
 668 {
 669         struct mds_obd *mds = mds_req2mds(req);
 670         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 671         struct dentry *dchild;
 672         char fidname[LL_FID_NAMELEN];
 673         int fidlen = 0, rc;
 674         void *handle = NULL;
 675         ENTRY;
 676
 677         down(&pending_dir->i_sem);
 678         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 679         dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 680         if (IS_ERR(dchild)) {
 681                 up(&pending_dir->i_sem);
 682                 rc = PTR_ERR(dchild);
 683                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 684                 RETURN(rc);
 685         }
 686
 687         if (dchild->d_inode != NULL) {
 688                 up(&pending_dir->i_sem);
 689                 mds_inode_set_orphan(dchild->d_inode);
 690                 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 691                 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 692                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 693                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 694                 CWARN("Orphan %s found and opened in PENDING directory\n",
 695                        fidname);
 696                 goto open;
 697         }
 698         dput(dchild);
 699         up(&pending_dir->i_sem);
 700
 701         /* We didn't find it in PENDING so it isn't an orphan.  See
 702          * if it was a regular inode that was previously created. */
 703         dchild = mds_fid2dentry(mds, fid, NULL);
 704         if (IS_ERR(dchild))
 705                 RETURN(PTR_ERR(dchild));
 706
 707         mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 708         mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 709         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 710         intent_set_disposition(rep, DISP_LOOKUP_POS);
 711
 712  open:
 713         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 714         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 715                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 716         /* XXX what do we do here if mds_finish_transno itself failed? */
 717
 718         l_dput(dchild);
 719         RETURN(rc);
 720 }
 721
 722 int mds_pin(struct ptlrpc_request *req)
 723 {
 724         struct obd_device *obd = req->rq_export->exp_obd;
 725         struct mds_body *request_body, *reply_body;
 726         struct lvfs_run_ctxt saved;
 727         int rc, size = sizeof(*reply_body);
 728         ENTRY;
 729
 730         request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
 731
 732         rc = lustre_pack_reply(req, 1, &size, NULL);
 733         if (rc)
 734                 RETURN(rc);
 735         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 736
 737         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 738         rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
 739                              request_body->flags, NULL, NULL);
 740         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 741
 742         RETURN(rc);
 743 }
 744
 745 /*  Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
 746  *  If child_lockh is NULL we just get the lock as a barrier to wait for
 747  *  other holders of this lock, and drop it right away again. */
 748 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 749                        struct lustre_handle *child_lockh)
 750 {
 751         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 752         struct lustre_handle lockh;
 753         int lock_flags = 0;
 754         int rc;
 755
 756         if (child_lockh == NULL)
 757                 child_lockh = &lockh;
 758
 759         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 760                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 761                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 762                               NULL, 0, NULL, child_lockh);
 763         if (rc != ELDLM_OK)
 764                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 765         else if (child_lockh == &lockh)
 766                 ldlm_lock_decref(child_lockh, LCK_EX);
 767
 768         RETURN(rc);
 769 }
 770
 771 int mds_open(struct mds_update_record *rec, int offset,
 772              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 773 {
 774         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 775         struct obd_device *obd = req->rq_export->exp_obd;
 776         struct mds_obd *mds = mds_req2mds(req);
 777         struct ldlm_reply *rep = NULL;
 778         struct mds_body *body = NULL;
 779         struct dentry *dchild = NULL, *dparent = NULL;
 780         struct mds_export_data *med;
 781         struct lustre_handle parent_lockh[2];
 782         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 783         int parent_mode = LCK_PR;
 784         void *handle = NULL;
 785         struct dentry_params dp;
 786         struct mea *mea = NULL;
 787         int mea_size, update_mode;
 788         ENTRY;
 789
 790         parent_lockh[0].cookie = 0;
 791         parent_lockh[1].cookie = 0;
 792
 793         if (offset == 2) { /* intent */
 794                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 795                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 796         } else if (offset == 0) { /* non-intent reint */
 797                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 798         } else {
 799                 body = NULL;
 800                 LBUG();
 801         }
 802
 803         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 804
 805         /* Step 0: If we are passed a fid, then we assume the client already
 806          * opened this file and is only replaying the RPC, so we open the
 807          * inode by fid (at some large expense in security). */
 808         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 809                 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
 810                           rep->lock_policy_res1);
 811
 812                 LASSERT(rec->ur_fid2->id);
 813
 814                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 815                                      rec, rep);
 816                 if (rc != -ENOENT)
 817                         RETURN(rc);
 818                 /* We didn't find the correct inode on disk either, so we
 819                  * need to re-create it via a regular replay. */
 820                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 821         } else {
 822                 LASSERT(!rec->ur_fid2->id);
 823         }
 824
 825         LASSERT(offset == 2); /* If we got here, we must be called via intent */
 826
 827         med = &req->rq_export->exp_mds_data;
 828         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 829                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 830                 RETURN(-ENOMEM);
 831         }
 832
 833         acc_mode = accmode(rec->ur_flags);
 834
 835         /* Step 1: Find and lock the parent */
 836         if (rec->ur_flags & O_CREAT) {
 837                 /* XXX Well, in fact we only need this lock mode change if
 838                    in addition to O_CREAT, the file does not exist.
 839                    But we do not know if it exists or not yet */
 840                 parent_mode = LCK_PW;
 841         }
 842
 843         if (rec->ur_namelen == 1) {
 844                 /* client (LMV) wants to open the file by fid */
 845                 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
 846                                 (unsigned) rec->ur_fid1->mds,
 847                                 (unsigned) rec->ur_fid1->id,
 848                                 (unsigned) rec->ur_fid1->generation);
 849                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 850                 if (IS_ERR(dchild)) {
 851                         rc = PTR_ERR(dparent);
 852                         CERROR("child lookup by a fid error %d\n", rc);
 853                         GOTO(cleanup, rc);
 854                 }
 855                 goto got_child;
 856         }
 857
 858         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 859                                         parent_lockh, &update_mode, rec->ur_name,
 860                                         rec->ur_namelen - 1,
 861                                         MDS_INODELOCK_UPDATE);
 862         if (IS_ERR(dparent)) {
 863                 rc = PTR_ERR(dparent);
 864                 CERROR("parent lookup error %d\n", rc);
 865                 GOTO(cleanup, rc);
 866         }
 867         LASSERT(dparent->d_inode != NULL);
 868
 869         cleanup_phase = 1; /* parent dentry and lock */
 870
 871         /* try to retrieve MEA data for this dir */
 872         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 873         if (rc)
 874                 GOTO(cleanup, rc);
 875
 876         if (mea != NULL) {
 877                 /* dir is already splitted, check is requested filename
 878                  * should live at this MDS or at another one */
 879                 int i;
 880                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 881                 if (mea->mea_master != i) {
 882                         CERROR("inapropriate MDS(%d) for %s. should be %d\n",
 883                                 mea->mea_master, rec->ur_name, i);
 884                         GOTO(cleanup, rc = -ERESTART);
 885                 }
 886         }
 887
 888         /* Step 2: Lookup the child */
 889         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 890         if (IS_ERR(dchild)) {
 891                 rc = PTR_ERR(dchild);
 892                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 893                 GOTO(cleanup, rc);
 894         }
 895
 896 got_child:
 897         cleanup_phase = 2; /* child dentry */
 898
 899         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 900
 901         if (dchild->d_flags & DCACHE_CROSS_REF) {
 902                 struct ldlm_res_id res_id = { . name = {0} };
 903                 ldlm_policy_data_t policy;
 904                 int flags = 0;
 905                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 906                        (unsigned long) dchild->d_mdsnum,
 907                        (unsigned long) dchild->d_inum,
 908                        (unsigned long) dchild->d_generation);
 909                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 910                 body->fid1.id = dchild->d_inum;
 911                 body->fid1.mds = dchild->d_mdsnum;
 912                 body->fid1.generation = dchild->d_generation;
 913                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 914                 res_id.name[0] = dchild->d_inum;
 915                 res_id.name[1] = dchild->d_generation;
 916                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 917                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 918                                       res_id, LDLM_IBITS, &policy,
 919                                       LCK_PR, &flags,
 920                                       mds_blocking_ast,
 921                                       ldlm_completion_ast, NULL, NULL,
 922                                       NULL, 0, NULL, child_lockh);
 923 #ifdef S_PDIROPS
 924                 if (parent_lockh[1].cookie != 0)
 925                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 926 #endif
 927                 ldlm_lock_decref(parent_lockh, parent_mode);
 928                 if (mea)
 929                         OBD_FREE(mea, mea_size);
 930                 l_dput(dchild);
 931                 l_dput(dparent);
 932                 RETURN(rc);
 933         }
 934
 935         if (dchild->d_inode)
 936                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 937         else
 938                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 939
 940         /*Step 3: If the child was negative, and we're supposed to, create it.*/
 941         if (dchild->d_inode == NULL) {
 942                 unsigned long ino = rec->ur_fid2->id;
 943                 struct iattr iattr;
 944                 struct inode *inode;
 945
 946                 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
 947                         if (rc > 0) {
 948                                 /* dir got splitted */
 949                                 GOTO(cleanup, rc = -ERESTART);
 950                         } else {
 951                                 /* error happened during spitting */
 952                                 GOTO(cleanup, rc);
 953                         }
 954                 }
 955
 956                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 957                         /* It's negative and we weren't supposed to create it */
 958                         GOTO(cleanup, rc = -ENOENT);
 959                 }
 960
 961                 intent_set_disposition(rep, DISP_OPEN_CREATE);
 962                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
 963                                       NULL);
 964                 if (IS_ERR(handle)) {
 965                         rc = PTR_ERR(handle);
 966                         handle = NULL;
 967                         GOTO(cleanup, rc);
 968                 }
 969                 dchild->d_fsdata = (void *) &dp;
 970                 dp.p_ptr = req;
 971                 dp.p_inum = ino;
 972
 973                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
 974                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
 975                         dchild->d_fsdata = NULL;
 976
 977                 if (rc) {
 978                         CDEBUG(D_INODE, "error during create: %d\n", rc);
 979                         GOTO(cleanup, rc);
 980                 }
 981                 inode = dchild->d_inode;
 982                 if (ino) {
 983                         LASSERT(ino == inode->i_ino);
 984                         /* Written as part of setattr */
 985                         inode->i_generation = rec->ur_fid2->generation;
 986                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 987                                inode->i_ino, inode->i_generation);
 988                 }
 989
 990                 created = 1;
 991                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
 992                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
 993                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
 994
 995                 iattr.ia_uid = rec->_ur_fsuid;
 996                 if (dparent->d_inode->i_mode & S_ISGID)
 997                         iattr.ia_gid = dparent->d_inode->i_gid;
 998                 else
 999                         iattr.ia_gid = rec->_ur_fsgid;
1000
1001                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1002                         ATTR_MTIME | ATTR_CTIME;
1003
1004                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1005                 if (rc)
1006                         CERROR("error on child setattr: rc = %d\n", rc);
1007
1008                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1009
1010                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1011                 if (rc)
1012                         CERROR("error on parent setattr: rc = %d\n", rc);
1013
1014                 acc_mode = 0;           /* Don't check for permissions */
1015         }
1016
1017         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1018
1019         mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1020         mds_pack_inode2body(obd, body, dchild->d_inode);
1021
1022         if (S_ISREG(dchild->d_inode->i_mode)) {
1023                 /* Check permissions etc */
1024                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1025                 if (rc != 0)
1026                         GOTO(cleanup, rc);
1027
1028                 /* Can't write to a read-only file */
1029                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1030                         GOTO(cleanup, rc = -EPERM);
1031
1032                 /* An append-only file must be opened in append mode for
1033                  * writing */
1034                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1035                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1036                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1037                         GOTO(cleanup, rc = -EPERM);
1038         }
1039
1040         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1041             (rec->ur_flags & MDS_OPEN_EXCL)) {
1042                 /* File already exists, we didn't just create it, and we
1043                  * were passed O_EXCL; err-or. */
1044                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1045         }
1046
1047         /* if we are following a symlink, don't open */
1048         if (S_ISLNK(dchild->d_inode->i_mode))
1049                 GOTO(cleanup, rc = 0);
1050
1051         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1052             !S_ISDIR(dchild->d_inode->i_mode))
1053                 GOTO(cleanup, rc = -ENOTDIR);
1054
1055         if (S_ISDIR(dchild->d_inode->i_mode)) {
1056                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1057                     rec->ur_flags & FMODE_WRITE) {
1058                         /*we are tryying to create or write a exist dir*/
1059                         GOTO(cleanup, rc = -EISDIR);
1060                 }
1061                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1062                         GOTO(cleanup, rc = -EACCES);
1063                 }
1064         }
1065
1066         if (rc == 0) {
1067                 struct ldlm_res_id res_id = { . name = {0} };
1068                 ldlm_policy_data_t policy;
1069                 int flags = 0;
1070                 res_id.name[0] = dchild->d_inode->i_ino;
1071                 res_id.name[1] = dchild->d_inode->i_generation;
1072                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1073                                                 MDS_INODELOCK_UPDATE;
1074                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1075                                       res_id, LDLM_IBITS, &policy,
1076                                       LCK_PR, &flags,
1077                                       mds_blocking_ast,
1078                                       ldlm_completion_ast, NULL, NULL,
1079                                       NULL, 0, NULL, child_lockh);
1080                 if (rc == 0)
1081                         cleanup_phase = 3;
1082         }
1083
1084         /* Step 5: mds_open it */
1085         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1086                              rep);
1087         GOTO(cleanup, rc);
1088
1089  cleanup:
1090         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1091                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1092
1093         switch (cleanup_phase) {
1094         case 3:
1095                 if (rc)
1096                         ldlm_lock_decref(child_lockh, LCK_PR);
1097         case 2:
1098                 if (rc && created) {
1099                         int err = vfs_unlink(dparent->d_inode, dchild);
1100                         if (err) {
1101                                 CERROR("unlink(%*s) in error path: %d\n",
1102                                        dchild->d_name.len, dchild->d_name.name,
1103                                        err);
1104                         }
1105                 } else if (created) {
1106                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1107                 }
1108                 l_dput(dchild);
1109         case 1:
1110                 if (dparent == NULL)
1111                         break;
1112
1113                 l_dput(dparent);
1114 #ifdef S_PDIROPS
1115                 if (parent_lockh[1].cookie != 0)
1116                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1117 #endif
1118                 if (rc)
1119                         ldlm_lock_decref(parent_lockh, parent_mode);
1120                 else
1121                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1122         }
1123         if (rc == 0)
1124                 atomic_inc(&mds->mds_open_count);
1125         if (mea)
1126                 OBD_FREE(mea, mea_size);
1127         if ((cleanup_phase != 3) && !rc)
1128                 rc = ENOLCK;
1129         RETURN(rc);
1130 }
1131
1132 /* Close a "file descriptor" and possibly unlink an orphan from the
1133  * PENDING directory.
1134  *
1135  * If we are being called from mds_disconnect() because the client has
1136  * disappeared, then req == NULL and we do not update last_rcvd because
1137  * there is nothing that could be recovered by the client at this stage
1138  * (it will not even _have_ an entry in last_rcvd anymore).
1139  *
1140  * Returns EAGAIN if the client needs to get more data and re-close. */
1141 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1142                   struct mds_file_data *mfd, int unlink_orphan)
1143 {
1144         struct inode *inode = mfd->mfd_dentry->d_inode;
1145         char fidname[LL_FID_NAMELEN];
1146         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1147         struct dentry *pending_child = NULL;
1148         struct mds_obd *mds = &obd->u.mds;
1149         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1150         void *handle = NULL;
1151         struct mds_body *request_body = NULL, *reply_body = NULL;
1152         struct dentry_params dp;
1153         struct iattr iattr = { 0 };
1154         ENTRY;
1155
1156         if (req && req->rq_reqmsg != NULL)
1157                 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1158                                               sizeof(*request_body));
1159         if (req && req->rq_repmsg != NULL)
1160                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1161                                             sizeof(*reply_body));
1162
1163         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1164
1165         last_orphan = mds_open_orphan_dec_test(inode) &&
1166                 mds_inode_is_orphan(inode);
1167
1168         /* this is half of the actual "close" */
1169         if (mfd->mfd_mode & FMODE_WRITE) {
1170                 rc = mds_put_write_access(mds, inode, request_body,
1171                                           last_orphan && unlink_orphan);
1172         } else if (mfd->mfd_mode & FMODE_EXEC) {
1173                 mds_allow_write_access(inode);
1174         }
1175
1176         if (last_orphan && unlink_orphan) {
1177                 struct lov_mds_md *lmm = NULL;
1178                 int stripe_count = 0;
1179                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1180
1181                 if (obd->obd_recovering) {
1182                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1183                                " is over\n", fidname);
1184                         GOTO(out, rc);
1185                 }
1186
1187                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1188
1189                 /* Sadly, there is no easy way to save pending_child from
1190                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1191                  * but normally it will still be in the dcache. */
1192                 pending_dir = mds->mds_pending_dir->d_inode;
1193                 down(&pending_dir->i_sem);
1194                 cleanup_phase = 1; /* up(i_sem) when finished */
1195                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1196                                                fidlen);
1197                 if (IS_ERR(pending_child))
1198                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1199                 LASSERT(pending_child->d_inode != NULL);
1200
1201                 cleanup_phase = 2; /* dput(pending_child) when finished */
1202                 if (req != NULL && req->rq_repmsg != NULL) {
1203                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1204                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1205                 }
1206
1207                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1208                                           NULL, stripe_count);
1209                 if (IS_ERR(handle)) {
1210                         rc = PTR_ERR(handle);
1211                         handle = NULL;
1212                         GOTO(cleanup, rc);
1213                 }
1214
1215                 if (req != NULL && req->rq_repmsg != NULL &&
1216                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1217                     mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1218                                       req->rq_repmsg->buflens[1],
1219                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1220                                       req->rq_repmsg->buflens[2]) > 0) {
1221                         reply_body->valid |= OBD_MD_FLCOOKIE;
1222                 }
1223
1224                 pending_child->d_fsdata = (void *) &dp;
1225                 dp.p_inum = 0;
1226                 dp.p_ptr = req;
1227                 if (S_ISDIR(pending_child->d_inode->i_mode))
1228                         rc = vfs_rmdir(pending_dir, pending_child);
1229                 else
1230                         rc = vfs_unlink(pending_dir, pending_child);
1231                 if (rc)
1232                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1233
1234                 goto out; /* Don't bother updating attrs on unlinked inode */
1235         }
1236
1237         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1238                 /* Update the on-disk attributes if this was the last write
1239                  * close, and all information was provided (i.e., rc == 0)
1240                  *
1241                  * XXX this should probably be abstracted with mds_reint_setattr
1242                  */
1243
1244 #if 0
1245                 if (request_body->valid & OBD_MD_FLMTIME &&
1246                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1247                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1248                         iattr.ia_valid |= ATTR_MTIME;
1249                 }
1250                 if (request_body->valid & OBD_MD_FLCTIME &&
1251                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1252                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1253                         iattr.ia_valid |= ATTR_CTIME;
1254                 }
1255
1256                 /* XXX can't set block count with fsfilt_setattr (!) */
1257                 if (request_body->valid & OBD_MD_FLSIZE) {
1258                         iattr.ia_valid |= ATTR_SIZE;
1259                         iattr.ia_size = request_body->size;
1260                 }
1261                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1262                         iattr.ia_valid |= ATTR_BLOCKS;
1263                         iattr.ia_blocks = request_body->blocks
1264                 } */
1265
1266 #endif
1267         }
1268
1269         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1270                 /* Only start a transaction to write out only the atime if
1271                  * it is more out-of-date than the specified limit.  If we
1272                  * are already going to write out the atime then do it anyway.
1273                  * */
1274                 if ((LTIME_S(request_body->atime) >
1275                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1276                     (iattr.ia_valid != 0 &&
1277                      LTIME_S(request_body->atime) > LTIME_S(inode->i_atime))) {
1278                         LTIME_S(iattr.ia_atime) = LTIME_S(request_body->atime);
1279                         iattr.ia_valid |= ATTR_ATIME;
1280                 }
1281         }
1282
1283         if (iattr.ia_valid != 0) {
1284                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1285                 if (IS_ERR(handle))
1286                         GOTO(cleanup, rc = PTR_ERR(handle));
1287                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1288                 if (rc)
1289                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1290         }
1291 out:
1292         /* If other clients have this file open for write, rc will be > 0 */
1293         if (rc > 0)
1294                 rc = 0;
1295         l_dput(mfd->mfd_dentry);
1296         mds_mfd_destroy(mfd);
1297
1298  cleanup:
1299         atomic_dec(&mds->mds_open_count);
1300         if (req != NULL && reply_body != NULL) {
1301                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1302         } else if (handle) {
1303                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1304                 if (err) {
1305                         CERROR("error committing close: %d\n", err);
1306                         if (!rc)
1307                                 rc = err;
1308                 }
1309         }
1310
1311         switch (cleanup_phase) {
1312         case 2:
1313                 dput(pending_child);
1314         case 1:
1315                 up(&pending_dir->i_sem);
1316         }
1317         RETURN(rc);
1318 }
1319
1320 int mds_close(struct ptlrpc_request *req)
1321 {
1322         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1323         struct obd_device *obd = req->rq_export->exp_obd;
1324         struct mds_body *body;
1325         struct mds_file_data *mfd;
1326         struct lvfs_run_ctxt saved;
1327         struct inode *inode;
1328         int rc, repsize[3] = {sizeof(struct mds_body),
1329                               obd->u.mds.mds_max_mdsize,
1330                               obd->u.mds.mds_max_cookiesize};
1331         ENTRY;
1332
1333         rc = lustre_pack_reply(req, 3, repsize, NULL);
1334         if (rc) {
1335                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1336                 req->rq_status = rc;
1337         } else {
1338                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1339         }
1340
1341         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1342         if (body == NULL) {
1343                 CERROR("Can't unpack body\n");
1344                 req->rq_status = -EFAULT;
1345                 RETURN(-EFAULT);
1346         }
1347
1348         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1349                 /* do some stuff */ ;
1350
1351         mfd = mds_handle2mfd(&body->handle);
1352         if (mfd == NULL) {
1353                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1354                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1355                 req->rq_status = -ESTALE;
1356                 RETURN(-ESTALE);
1357         }
1358
1359         inode = mfd->mfd_dentry->d_inode;
1360         if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1361                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1362                 LASSERT(body != NULL);
1363
1364                 mds_pack_inode2fid(obd, &body->fid1, inode);
1365                 mds_pack_inode2body(obd, body, inode);
1366                 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1367         }
1368         spin_lock(&med->med_open_lock);
1369         list_del(&mfd->mfd_list);
1370         spin_unlock(&med->med_open_lock);
1371
1372         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1373         req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1374         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1375
1376         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1377                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1378                 req->rq_status = -ENOMEM;
1379                 mds_mfd_put(mfd);
1380                 RETURN(-ENOMEM);
1381         }
1382
1383         mds_mfd_put(mfd);
1384         RETURN(0);
1385 }
1386
1387 int mds_done_writing(struct ptlrpc_request *req)
1388 {
1389         struct mds_body *body;
1390         int rc, size = sizeof(struct mds_body);
1391         ENTRY;
1392
1393         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1394
1395         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1396         if (body == NULL) {
1397                 CERROR("Can't unpack body\n");
1398                 req->rq_status = -EFAULT;
1399                 RETURN(-EFAULT);
1400         }
1401
1402         rc = lustre_pack_reply(req, 1, &size, NULL);
1403         if (rc) {
1404                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1405                 req->rq_status = rc;
1406         }
1407
1408         RETURN(0);
1409 }