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 #include <linux/lustre_gs.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 /* Caller must hold mds->mds_epoch_sem */
 116 static int mds_alloc_filterdata(struct inode *inode)
 117 {
 118         LASSERT(LUSTRE_FILTERDATA(inode) == NULL);
 119         OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 120         if (LUSTRE_FILTERDATA(inode) == 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(LUSTRE_FILTERDATA(inode) != NULL);
 130         OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 131         LUSTRE_FILTERDATA(inode) = NULL;
 132         iput(inode);
 133 }
 134
 135 /* Write access to a file: executors cause a negative count,
 136  * writers a positive count.  The semaphore is needed to perform
 137  * a check for the sign and then increment or decrement atomically.
 138  *
 139  * This code is closely tied to the allocation of the d_fsdata and the
 140  * MDS epoch, so we use the same semaphore for the whole lot.
 141  *
 142  * We could use a different semaphore for each file, if it ever shows
 143  * up in a profile, which it won't.
 144  *
 145  * epoch argument is nonzero during recovery */
 146 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 147                                 __u64 epoch)
 148 {
 149         int rc = 0;
 150
 151         down(&mds->mds_epoch_sem);
 152
 153         if (atomic_read(&inode->i_writecount) < 0) {
 154                 up(&mds->mds_epoch_sem);
 155                 RETURN(-ETXTBSY);
 156         }
 157
 158         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 159                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 160                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 161                        inode->i_generation);
 162                 goto out;
 163         }
 164
 165         if (MDS_FILTERDATA(inode) == NULL)
 166                 mds_alloc_filterdata(inode);
 167         if (MDS_FILTERDATA(inode) == NULL) {
 168                 rc = -ENOMEM;
 169                 goto out;
 170         }
 171         if (epoch > mds->mds_io_epoch)
 172                 mds->mds_io_epoch = epoch;
 173         else
 174                 mds->mds_io_epoch++;
 175         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 176         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 177                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 178  out:
 179         if (rc == 0)
 180                 atomic_inc(&inode->i_writecount);
 181         up(&mds->mds_epoch_sem);
 182         return rc;
 183 }
 184
 185 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 186  * again -- which is never true if the file is about to be unlinked.  Otherwise
 187  * returns the number of remaining writers. */
 188 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 189                                 struct mds_body *body, int unlinking)
 190 {
 191         int rc = 0;
 192         ENTRY;
 193
 194         down(&mds->mds_epoch_sem);
 195         atomic_dec(&inode->i_writecount);
 196         rc = atomic_read(&inode->i_writecount);
 197         if (rc > 0)
 198                 GOTO(out, rc);
 199 #if 0
 200         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 201                 GOTO(out, rc = EAGAIN);
 202 #endif
 203         mds_free_filterdata(inode);
 204  out:
 205         up(&mds->mds_epoch_sem);
 206         return rc;
 207 }
 208
 209 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 210 {
 211         ENTRY;
 212         down(&mds->mds_epoch_sem);
 213         if (atomic_read(&inode->i_writecount) > 0) {
 214                 up(&mds->mds_epoch_sem);
 215                 RETURN(-ETXTBSY);
 216         }
 217         atomic_dec(&inode->i_writecount);
 218         up(&mds->mds_epoch_sem);
 219         RETURN(0);
 220 }
 221
 222 static void mds_allow_write_access(struct inode *inode)
 223 {
 224         ENTRY;
 225         atomic_inc(&inode->i_writecount);
 226 }
 227
 228 int mds_query_write_access(struct inode *inode)
 229 {
 230         ENTRY;
 231         RETURN(atomic_read(&inode->i_writecount));
 232 }
 233
 234 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 235 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 236                                              struct vfsmount *mnt, int flags,
 237                                              struct ptlrpc_request *req)
 238 {
 239         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 240         struct mds_obd *mds = mds_req2mds(req);
 241         struct mds_file_data *mfd;
 242         struct mds_body *body;
 243         int rc = 0;
 244         ENTRY;
 245
 246         mfd = mds_mfd_new();
 247         if (mfd == NULL) {
 248                 CERROR("mds: out of memory\n");
 249                 GOTO(cleanup_dentry, rc = -ENOMEM);
 250         }
 251
 252         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 253
 254         if (flags & FMODE_WRITE) {
 255                 /* FIXME: in recovery, need to pass old epoch here */
 256                 rc = mds_get_write_access(mds, dentry->d_inode, 0);
 257                 if (rc)
 258                         GOTO(cleanup_mfd, rc);
 259                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 260         } else if (flags & FMODE_EXEC) {
 261                 rc = mds_deny_write_access(mds, dentry->d_inode);
 262                 if (rc)
 263                         GOTO(cleanup_mfd, rc);
 264         }
 265
 266         dget(dentry);
 267
 268         /* FIXME: invalidate update locks when first open for write comes in */
 269
 270         /* mark the file as open to handle open-unlink. */
 271         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 272         mds_orphan_open_inc(dentry->d_inode);
 273         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 274
 275         mfd->mfd_mode = flags;
 276         mfd->mfd_dentry = dentry;
 277         mfd->mfd_xid = req->rq_xid;
 278
 279         spin_lock(&med->med_open_lock);
 280         list_add(&mfd->mfd_list, &med->med_open_head);
 281         spin_unlock(&med->med_open_lock);
 282         mds_mfd_put(mfd);
 283
 284         body->handle.cookie = mfd->mfd_handle.h_cookie;
 285         RETURN(mfd);
 286 cleanup_mfd:
 287         mds_mfd_put(mfd);
 288         mds_mfd_destroy(mfd);
 289 cleanup_dentry:
 290         return ERR_PTR(rc);
 291 }
 292
 293 static inline void
 294 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 295                     struct lov_desc *desc)
 296 {
 297         int i;
 298
 299         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 300                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 301                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 302         }
 303 }
 304
 305 /* must be called with i_sem held */
 306 int
 307 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
 308                    int offset, struct mds_update_record *rec,
 309                    struct dentry *dchild, void **handle,
 310                    obd_id **ids)
 311 {
 312         struct inode *inode = dchild->d_inode;
 313         struct mds_obd *mds = &obd->u.mds;
 314         struct obd_trans_info oti = { 0 };
 315         struct lov_stripe_md *lsm = NULL;
 316         struct lov_mds_md *lmm = NULL;
 317         int rc, lmm_bufsize, lmm_size;
 318         struct obdo *oa = NULL;
 319         struct mds_body *body;
 320         void *lmm_buf;
 321         ENTRY;
 322
 323         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 324             !(rec->ur_flags & FMODE_WRITE))
 325                 RETURN(0);
 326
 327         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 328
 329         if (!S_ISREG(inode->i_mode))
 330                 RETURN(0);
 331         if (body->valid & OBD_MD_FLEASIZE)
 332                 RETURN(0);
 333
 334         OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 335         if (*ids == NULL)
 336                 RETURN(-ENOMEM);
 337         oti.oti_objid = *ids;
 338
 339         /* replay case */
 340         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 341                 LASSERT(id_ino(rec->ur_id2));
 342                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 343                 lmm_size = rec->ur_eadatalen;
 344                 lmm = rec->ur_eadata;
 345                 LASSERT(lmm);
 346
 347                 if (*handle == NULL)
 348                         *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 349                 if (IS_ERR(*handle)) {
 350                         rc = PTR_ERR(*handle);
 351                         *handle = NULL;
 352                         RETURN(rc);
 353                 }
 354
 355                 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 356
 357                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 358                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 359                 LASSERT(lmm_buf != NULL);
 360                 LASSERT(lmm_bufsize >= lmm_size);
 361
 362                 memcpy(lmm_buf, lmm, lmm_size);
 363                 rc = fsfilt_set_md(obd, inode, *handle, lmm,
 364                                    lmm_size, EA_LOV);
 365                 if (rc)
 366                         CERROR("open replay failed to set md:%d\n", rc);
 367                 RETURN(0);
 368         }
 369
 370         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 371                 GOTO(out_ids, rc = -ENOMEM);
 372
 373         oa = obdo_alloc();
 374         if (oa == NULL)
 375                 RETURN(-ENOMEM);
 376         oa->o_mode = S_IFREG | 0600;
 377         oa->o_id = inode->i_ino;
 378         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 379         oa->o_generation = inode->i_generation;
 380         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 381         oa->o_gid = 0;
 382         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
 383                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 384         oa->o_size = 0;
 385
 386         obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 387                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 388
 389         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 390                 /* check if things like lfs setstripe are sending us the ea */
 391                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 392                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 393                                            mds->mds_dt_exp,
 394                                            0, &lsm, rec->ur_eadata);
 395                         if (rc)
 396                                 GOTO(out_oa, rc);
 397                 } else {
 398                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 399                         if (lmm == NULL)
 400                                 GOTO(out_oa, rc = -ENOMEM);
 401
 402                         lmm_size = mds->mds_max_mdsize;
 403                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 404                                         lmm, &lmm_size, 1, 0);
 405                         if (rc > 0)
 406                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 407                                                    mds->mds_dt_exp,
 408                                                    0, &lsm, lmm);
 409                         OBD_FREE(lmm, mds->mds_max_mdsize);
 410                         if (rc)
 411                                 GOTO(out_oa, rc);
 412                 }
 413
 414                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 415                 oti.oti_flags |= OBD_MODE_CROW;
 416                 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
 417
 418                 if (rc) {
 419                         CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
 420                                "error creating objects for "
 421                                "inode %lu: rc = %d\n",
 422                                inode->i_ino, rc);
 423                         if (rc > 0) {
 424                                 CERROR("obd_create returned invalid "
 425                                        "rc %d\n", rc);
 426                                 rc = -EIO;
 427                         }
 428                         GOTO(out_oa, rc);
 429                 }
 430         } else {
 431                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 432                                    0, &lsm, rec->ur_eadata);
 433                 if (rc)
 434                         GOTO(out_oa, rc);
 435
 436                 lsm->lsm_object_id = oa->o_id;
 437                 lsm->lsm_object_gr = oa->o_gr;
 438         }
 439         if (inode->i_size) {
 440                 oa->o_size = inode->i_size;
 441                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 442                                 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 443
 444                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
 445                 if (rc) {
 446                         CERROR("error setting attrs for inode %lu: rc %d\n",
 447                                inode->i_ino, rc);
 448                         if (rc > 0) {
 449                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 450                                 rc = -EIO;
 451                         }
 452                         GOTO(out_oa, rc);
 453                 }
 454         }
 455
 456         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 457         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 458
 459         LASSERT(lsm && lsm->lsm_object_id);
 460         lmm = NULL;
 461         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 462         if (!id_ino(rec->ur_id2))
 463                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 464         if (rc < 0) {
 465                 CERROR("cannot pack lsm, err = %d\n", rc);
 466                 GOTO(out_oa, rc);
 467         }
 468
 469         lmm_size = rc;
 470         body->eadatasize = rc;
 471
 472         if (*handle == NULL)
 473                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 474         if (IS_ERR(*handle)) {
 475                 rc = PTR_ERR(*handle);
 476                 *handle = NULL;
 477                 GOTO(out_oa, rc);
 478         }
 479
 480         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 481                            lmm_size, EA_LOV);
 482
 483         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 484         lmm_bufsize = req->rq_repmsg->buflens[offset];
 485         LASSERT(lmm_buf);
 486         LASSERT(lmm_bufsize >= lmm_size);
 487
 488         memcpy(lmm_buf, lmm, lmm_size);
 489         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 490 out_oa:
 491         oti_free_cookies(&oti);
 492         obdo_free(oa);
 493 out_ids:
 494         if (rc) {
 495                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 496                 *ids = NULL;
 497         }
 498         if (lsm)
 499                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 500         RETURN(rc);
 501 }
 502
 503 int
 504 mds_destroy_object(struct obd_device *obd,
 505                    struct inode *inode, int async)
 506 {
 507         struct mds_obd *mds = &obd->u.mds;
 508         struct lov_mds_md *lmm = NULL;
 509         int rc, lmm_size;
 510         ENTRY;
 511
 512         LASSERT(inode != NULL);
 513
 514         if (inode->i_nlink != 0) {
 515                 CDEBUG(D_INODE, "attempt to destroy OSS object when "
 516                        "i_nlink == %d\n", (int)inode->i_nlink);
 517                 RETURN(0);
 518         }
 519
 520         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 521         if (lmm == NULL)
 522                 RETURN(-ENOMEM);
 523
 524         lmm_size = mds->mds_max_mdsize;
 525         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 526         if (rc < 0) {
 527                 CERROR("no stripe info for %lu/%lu inode\n",
 528                        (unsigned long)inode->i_ino,
 529                        (unsigned long)inode->i_generation);
 530                 GOTO(out_free_lmm, rc);
 531         }
 532
 533         if (rc > 0) {
 534                 /* asynchronously unlink objecect on OSS */
 535                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 536                                        NULL, 0, async);
 537                 if (rc) {
 538                         CERROR("error unlinking object on OSS, "
 539                                "err %d\n", rc);
 540                         GOTO(out_free_lmm, rc);
 541                 }
 542         } else {
 543                 CDEBUG(D_INODE, "no stripping info found for inode "
 544                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 545                       (unsigned long)inode->i_generation);
 546         }
 547         EXIT;
 548 out_free_lmm:
 549         OBD_FREE(lmm, mds->mds_max_mdsize);
 550         return rc;
 551 }
 552
 553 static void reconstruct_open(struct mds_update_record *rec, int offset,
 554                              struct ptlrpc_request *req,
 555                              struct lustre_handle *child_lockh)
 556 {
 557         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 558         struct mds_client_data *mcd = med->med_mcd;
 559         struct mds_obd *mds = mds_req2mds(req);
 560         struct mds_file_data *mfd;
 561         struct obd_device *obd = req->rq_export->exp_obd;
 562         struct dentry *parent = NULL, *dchild;
 563         struct ldlm_reply *rep;
 564         struct mds_body *body;
 565         struct list_head *t;
 566         int put_child = 1;
 567         int rc;
 568         ENTRY;
 569
 570         LASSERT(offset == 3); /* only called via intent */
 571         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 572         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 573
 574         /* copy rc, transno and disp; steal locks */
 575         mds_req_from_mcd(req, mcd);
 576         intent_set_disposition(rep, mcd->mcd_last_data);
 577
 578         /* Only replay if create or open actually happened. */
 579         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 580                 EXIT;
 581                 return; /* error looking up parent or child */
 582         }
 583
 584         if (rec->ur_namelen == 1) {
 585                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 586                        OLID4(rec->ur_id1));
 587                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 588         } else {
 589                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 590                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 591                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 592                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 593                                 rec->ur_namelen - 1);
 594                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 595                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 596         }
 597
 598         if (!dchild->d_inode)
 599                 GOTO(out_dput, 0); /* child not present to open */
 600
 601         /* At this point, we know we have a child. We'll send
 602          * it back _unless_ it not created and open failed.  */
 603         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 604             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 605             req->rq_status)
 606                 GOTO(out_dput, 0);
 607
 608         /* get lock (write for O_CREAT, read otherwise) */
 609         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 610         if (S_ISREG(dchild->d_inode->i_mode)) {
 611                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 612                                  dchild->d_inode, 1, 0);
 613
 614                 if (rc)
 615                         LASSERT(rc == req->rq_status);
 616
 617                 /* If we have LOV EA data, the OST holds size, mtime */
 618                 if (!(body->valid & OBD_MD_FLEASIZE))
 619                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 620                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 621         }
 622
 623         /* If we have -EEXIST as the status, and we were asked to create
 624          * exclusively, we can tell we failed because the file already existed.
 625          */
 626         if (req->rq_status == -EEXIST &&
 627             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 628              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 629                 GOTO(out_dput, 0);
 630         }
 631
 632         /* If we didn't get as far as trying to open, then some locking thing
 633          * probably went wrong, and we'll just bail here.
 634          */
 635         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 636                 GOTO(out_dput, 0);
 637
 638         /* If we failed, then we must have failed opening, so don't look for
 639          * file descriptor or anything, just give the client the bad news.
 640          */
 641         if (req->rq_status)
 642                 GOTO(out_dput, 0);
 643
 644         mfd = NULL;
 645         list_for_each(t, &med->med_open_head) {
 646                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 647                 if (mfd->mfd_xid == req->rq_xid)
 648                         break;
 649                 mfd = NULL;
 650         }
 651
 652         /* #warning "XXX fixme" bug 2991 */
 653         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 654          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 655          * to detect a re-open */
 656         if (mfd == NULL) {
 657                 mntget(mds->mds_vfsmnt);
 658                 CERROR("Re-opened file \n");
 659                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 660                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 661                 if (!mfd) {
 662                         CERROR("mds: out of memory\n");
 663                         GOTO(out_dput, req->rq_status = -ENOMEM);
 664                 }
 665                 put_child = 0;
 666         } else {
 667                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 668                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 669                        mfd->mfd_handle.h_cookie);
 670         }
 671
 672  out_dput:
 673         if (put_child)
 674                 l_dput(dchild);
 675         if (parent)
 676                 l_dput(parent);
 677         EXIT;
 678 }
 679
 680 /* do NOT or the MAY_*'s, you'll get the weakest */
 681 static int accmode(int flags)
 682 {
 683         int res = 0;
 684
 685         if (flags & FMODE_READ)
 686                 res = MAY_READ;
 687         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 688                 res |= MAY_WRITE;
 689         if (flags & FMODE_EXEC)
 690                 res = MAY_EXEC;
 691         return res;
 692 }
 693
 694 /* Handles object creation, actual opening, and I/O epoch */
 695 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 696                            struct mds_body *body, int flags, void **handle,
 697                            struct mds_update_record *rec, struct ldlm_reply *rep)
 698 {
 699         struct obd_device *obd = req->rq_export->exp_obd;
 700         struct mds_obd *mds = mds_req2mds(req);
 701         struct mds_file_data *mfd = NULL;
 702         obd_id *ids = NULL;
 703         unsigned mode;
 704         int rc = 0, reply_off;
 705         ENTRY;
 706
 707         /* atomically create objects if necessary */
 708         down(&dchild->d_inode->i_sem);
 709         mode = dchild->d_inode->i_mode;
 710
 711         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 712             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 713                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 714                                  dchild->d_inode, 0, 0);
 715                 if (rc) {
 716                         up(&dchild->d_inode->i_sem);
 717                         RETURN(rc);
 718                 }
 719         }
 720
 721         if (rec != NULL) {
 722                 if ((body->valid & OBD_MD_FLEASIZE) &&
 723                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 724                         up(&dchild->d_inode->i_sem);
 725                         RETURN(-EEXIST);
 726                 }
 727
 728                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 729                         /* no EA: create objects */
 730                         rc = mds_create_objects(obd, req, 2, rec,
 731                                                 dchild, handle, &ids);
 732                         if (rc) {
 733                                 CERROR("mds_create_object: rc = %d\n", rc);
 734                                 up(&dchild->d_inode->i_sem);
 735                                 RETURN(rc);
 736                         }
 737                 }
 738
 739                 if (S_ISREG(dchild->d_inode->i_mode) &&
 740                     (body->valid & OBD_MD_FLEASIZE)) {
 741                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 742                                                    req->rq_repmsg, 2);
 743                         if (!rc)
 744                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 745                                                  dchild->d_inode, 0, 0);
 746                         if (rc) {
 747                                 up(&dchild->d_inode->i_sem);
 748                                 RETURN(rc);
 749                         }
 750                 }
 751         }
 752
 753         reply_off = 3;
 754         rc = mds_pack_acl(req, &reply_off, body, dchild->d_inode);
 755
 756         if (rc < 0) {
 757                 CERROR("pack posix acl: rc = %d\n", rc);
 758                 up(&dchild->d_inode->i_sem);
 759                 RETURN(rc);
 760         }
 761
 762         rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
 763                             dchild->d_inode);
 764         if (rc < 0) {
 765                 CERROR("mds_pack_gskey: rc = %d\n", rc);
 766                 up(&dchild->d_inode->i_sem);
 767                 RETURN(rc);
 768         }
 769         /* If the inode has no EA data, then MDSs hold size, mtime */
 770         if (S_ISREG(dchild->d_inode->i_mode) &&
 771             !(body->valid & OBD_MD_FLEASIZE)) {
 772                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 773                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 774         }
 775
 776         up(&dchild->d_inode->i_sem);
 777
 778         intent_set_disposition(rep, DISP_OPEN_OPEN);
 779         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 780         if (IS_ERR(mfd))
 781                 RETURN(PTR_ERR(mfd));
 782
 783         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 784                mfd->mfd_handle.h_cookie);
 785
 786         if (ids != NULL) {
 787                 mds_dt_update_objids(obd, ids);
 788                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 789         }
 790
 791         RETURN(rc);
 792 }
 793
 794 static int mds_open_by_id(struct ptlrpc_request *req,
 795                           struct lustre_id *id,
 796                           struct mds_body *body, int flags,
 797                           struct mds_update_record *rec,
 798                           struct ldlm_reply *rep)
 799 {
 800         struct mds_obd *mds = mds_req2mds(req);
 801         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 802         struct dentry *dchild;
 803         char idname[LL_ID_NAMELEN];
 804         int idlen = 0, rc;
 805         void *handle = NULL;
 806         ENTRY;
 807
 808         down(&pending_dir->i_sem);
 809         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 810         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 811                                 idlen);
 812         if (IS_ERR(dchild)) {
 813                 up(&pending_dir->i_sem);
 814                 rc = PTR_ERR(dchild);
 815                 CERROR("error looking up %s in PENDING: rc = %d\n",
 816                        idname, rc);
 817                 RETURN(rc);
 818         }
 819
 820         up(&pending_dir->i_sem);
 821         if (dchild->d_inode != NULL) {
 822                 mds_inode_set_orphan(dchild->d_inode);
 823                 mds_pack_inode2body(req2obd(req), body,
 824                                     dchild->d_inode, 1);
 825
 826                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 827                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 828                 CWARN("Orphan %s found and opened in PENDING directory\n",
 829                        idname);
 830                 goto open;
 831         }
 832         l_dput(dchild);
 833
 834         /*
 835          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 836          * regular inode that was previously created.
 837          */
 838         dchild = mds_id2dentry(req2obd(req), id, NULL);
 839         if (IS_ERR(dchild))
 840                 RETURN(PTR_ERR(dchild));
 841
 842         mds_pack_inode2body(req2obd(req), body,
 843                             dchild->d_inode, 1);
 844
 845         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 846         intent_set_disposition(rep, DISP_LOOKUP_POS);
 847
 848  open:
 849         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 850         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 851                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 852         /* XXX what do we do here if mds_finish_transno itself failed? */
 853         l_dput(dchild);
 854         RETURN(rc);
 855 }
 856
 857 int mds_pin(struct ptlrpc_request *req, int offset)
 858 {
 859         struct obd_device *obd = req->rq_export->exp_obd;
 860         struct mds_body *request_body, *reply_body;
 861         struct lvfs_run_ctxt saved;
 862         int rc, size = sizeof(*reply_body);
 863         ENTRY;
 864
 865         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 866                                       sizeof(*request_body));
 867
 868         rc = lustre_pack_reply(req, 1, &size, NULL);
 869         if (rc)
 870                 RETURN(rc);
 871         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 872
 873         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 874         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 875                             request_body->flags, NULL, NULL);
 876         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 877
 878         RETURN(rc);
 879 }
 880
 881 /*
 882  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 883  * child_lockh is NULL we just get the lock as a barrier to wait for other
 884  * holders of this lock, and drop it right away again.
 885  */
 886 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 887                        struct lustre_handle *child_lockh)
 888 {
 889         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 890         struct lustre_handle lockh;
 891         int lock_flags = LDLM_FL_ATOMIC_CB;
 892         int rc;
 893         ENTRY;
 894
 895         if (child_lockh == NULL)
 896                 child_lockh = &lockh;
 897
 898         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 899                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 900                               mds_blocking_ast, ldlm_completion_ast, NULL,
 901                               NULL, NULL, 0, NULL, child_lockh);
 902         if (rc != ELDLM_OK)
 903                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 904         else if (child_lockh == &lockh)
 905                 ldlm_lock_decref(child_lockh, LCK_EX);
 906
 907         RETURN(rc);
 908 }
 909
 910 int mds_open(struct mds_update_record *rec, int offset,
 911              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 912 {
 913         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 914         struct obd_device *obd = req->rq_export->exp_obd;
 915         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 916         struct mds_obd *mds = mds_req2mds(req);
 917         struct ldlm_reply *rep = NULL;
 918         struct mds_body *body = NULL;
 919         struct dentry *dchild = NULL, *dparent = NULL;
 920         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 921         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 922         int parent_mode = LCK_PR;
 923         void *handle = NULL;
 924         struct dentry_params dp;
 925         struct mea *mea = NULL;
 926         int mea_size, update_mode;
 927         int child_mode = LCK_PR;
 928         struct lustre_id sid;
 929         __u64 fid = 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 && mea->mea_count) {
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                 /* get parent id: ldlm lock on the parent protects ea */
1131                 rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
1132                 if (rc) {
1133                         CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
1134                                 dparent->d_inode->i_ino, rc);
1135                         GOTO(cleanup, rc);
1136                 }
1137
1138                 ino = (unsigned long)id_ino(rec->ur_id2);
1139
1140                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1141                                           update_mode);
1142
1143                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1144                        obd->obd_name, dparent->d_inode->i_ino,
1145                        dparent->d_inode->i_generation, rc);
1146
1147                 if (rc > 0) {
1148                         /* dir got splitted */
1149                         GOTO(cleanup, rc = -ERESTART);
1150                 } else if (rc < 0) {
1151                         /* error happened during spitting */
1152                         GOTO(cleanup, rc);
1153                 }
1154
1155                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1156                         /*
1157                          * dentry is negative and we weren't supposed to create
1158                          * object, get out of here.
1159                          */
1160                         GOTO(cleanup, rc = -ENOENT);
1161                 }
1162
1163                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1164                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1165                                       NULL);
1166                 if (IS_ERR(handle)) {
1167                         rc = PTR_ERR(handle);
1168                         handle = NULL;
1169                         GOTO(cleanup, rc);
1170                 }
1171                 if (ino)
1172                         fid = id_fid(rec->ur_id2);
1173                 else
1174                         fid = mds_alloc_fid(obd);
1175                 dchild->d_fsdata = (void *) &dp;
1176                 dp.p_ptr = req;
1177                 dp.p_inum = ino;
1178                 dp.p_fid = fid;
1179                 dp.p_group = mds->mds_num;
1180
1181                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1182                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1183                         dchild->d_fsdata = NULL;
1184
1185                 if (rc) {
1186                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1187                         GOTO(cleanup, rc);
1188                 }
1189                 inode = dchild->d_inode;
1190                 if (ino) {
1191                         LASSERT(ino == inode->i_ino);
1192
1193                         /* written as part of setattr */
1194                         inode->i_generation = id_gen(rec->ur_id2);
1195                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1196                                inode->i_ino, inode->i_generation);
1197                 }
1198
1199                 created = 1;
1200                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1201                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1202                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1203
1204                 iattr.ia_uid = rec->ur_fsuid;
1205                 if (dparent->d_inode->i_mode & S_ISGID)
1206                         iattr.ia_gid = dparent->d_inode->i_gid;
1207                 else
1208                         iattr.ia_gid = rec->ur_fsgid;
1209
1210                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1211                         ATTR_MTIME | ATTR_CTIME;
1212
1213                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1214                 if (rc)
1215                         CERROR("error on child setattr: rc = %d\n", rc);
1216
1217                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1218
1219                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1220                 if (rc)
1221                         CERROR("error on parent setattr: rc = %d\n", rc);
1222                 else {
1223                         MD_COUNTER_INCREMENT(obd, create);
1224                 }
1225
1226                 mds_inode2id(obd, &body->id1, inode, fid);
1227                 mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
1228
1229                 if ((rec->ur_flags & MDS_OPEN_HAS_KEY) ||
1230                      mds->mds_crypto_type == MKS_TYPE) {
1231                         rc = mds_set_gskey(obd, handle, dchild->d_inode,
1232                                            rec->ur_ea2data, rec->ur_ea2datalen,
1233                                            ATTR_KEY | ATTR_MAC);
1234                         if (rc) {
1235                                 CERROR("error in set gs key rc %d\n", rc);
1236                         }
1237                 }
1238
1239                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1240                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1241                                            dchild->d_inode, handle,
1242                                            req->rq_export->exp_sync);
1243                         handle = NULL;
1244                 }
1245
1246                 acc_mode = 0;           /* Don't check for permissions */
1247         }
1248         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1249
1250         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1251                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1252                  dchild->d_name.name, dchild, dchild->d_inode);
1253
1254         if (S_ISREG(dchild->d_inode->i_mode)) {
1255                 /* Check permissions etc */
1256                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1257                 if (rc != 0)
1258                         GOTO(cleanup, rc);
1259
1260                 /* Can't write to a read-only file */
1261                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1262                         GOTO(cleanup, rc = -EPERM);
1263
1264                 /* An append-only file must be opened in append mode for
1265                  * writing */
1266                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1267                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1268                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1269                         GOTO(cleanup, rc = -EPERM);
1270         }
1271
1272         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1273             (rec->ur_flags & MDS_OPEN_EXCL)) {
1274                 /* File already exists, we didn't just create it, and we
1275                  * were passed O_EXCL; err-or. */
1276                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1277         }
1278
1279         if (S_ISDIR(dchild->d_inode->i_mode)) {
1280                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1281                     rec->ur_flags & FMODE_WRITE) {
1282                         /* we are trying to create or write a exist dir */
1283                         GOTO(cleanup, rc = -EISDIR);
1284                 }
1285                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1286                         GOTO(cleanup, rc = -EACCES);
1287
1288                 /*
1289                  * here was checking for possible GNS mount points to skip their
1290                  * open. I removed it as its detection based only on SUID bit is
1291                  * not reliable. Opening such a dirs should not cause any
1292                  * problems, at least tests show that. --umka
1293                  */
1294         }
1295
1296         /* if we are following a symlink, don't open */
1297         if (S_ISLNK(dchild->d_inode->i_mode))
1298                 GOTO(cleanup_no_trans, rc = 0);
1299
1300         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1301             !S_ISDIR(dchild->d_inode->i_mode))
1302                 GOTO(cleanup, rc = -ENOTDIR);
1303
1304         /* check permission even it's special files */
1305         if (S_ISCHR(dchild->d_inode->i_mode) ||
1306             S_ISBLK(dchild->d_inode->i_mode) ||
1307             S_ISFIFO(dchild->d_inode->i_mode) ||
1308             S_ISSOCK(dchild->d_inode->i_mode)) {
1309                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1310                 if (rc != 0)
1311                         GOTO(cleanup, rc);
1312         }
1313
1314         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1315                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1316                 GOTO(cleanup, rc = -EAGAIN);
1317         }
1318
1319 #if 0
1320         /* We cannot use acc_mode here, because it is zeroed in case of
1321            creating a file, so we get wrong lockmode */
1322         if (accmode(rec->ur_flags) & MAY_WRITE)
1323                 child_mode = LCK_CW;
1324         else if (accmode(rec->ur_flags) & MAY_EXEC)
1325                 child_mode = LCK_PR;
1326         else
1327                 child_mode = LCK_CR;
1328
1329         /* Always returning LOOKUP lock if open succesful to guard
1330          * dentry on client. In case of replay we do not get a lock
1331          * assuming that the caller has it already */
1332         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1333                 struct ldlm_res_id child_res_id = { .name = {0}};
1334                 ldlm_policy_data_t policy;
1335                 int lock_flags = LDLM_FL_ATOMIC_CB;
1336
1337                 /* LOOKUP lock will protect dentry on client -bzzz */
1338                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1339                                                 MDS_INODELOCK_OPEN;
1340
1341                 child_res_id.name[0] = id_fid(&body->id1);
1342                 child_res_id.name[1] = id_group(&body->id1);
1343
1344                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1345                                       child_res_id, LDLM_IBITS, &policy,
1346                                       child_mode, &lock_flags,
1347                                       mds_blocking_ast, ldlm_completion_ast,
1348                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1349                 if (rc != ELDLM_OK)
1350                         GOTO(cleanup, rc);
1351
1352                 cleanup_phase = 3;
1353         }
1354 #else
1355 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1356 #warning "disable opencache lock for CMD2"
1357 #endif
1358
1359         /* Step 5: mds_open it */
1360         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1361                              rec, rep);
1362         if (rc)
1363                 GOTO(cleanup, rc);
1364
1365         /* if this is a writer, we have to invalidate client's
1366          * update locks in order to make sure they don't use
1367          * isize/iblocks from mds anymore.
1368          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1369          * XXX: optimization is to do this for first writer only */
1370         if (accmode(rec->ur_flags) & MAY_WRITE) {
1371                 struct ldlm_res_id child_res_id = { .name = {0}};
1372                 ldlm_policy_data_t sz_policy;
1373                 struct lustre_handle sz_lockh;
1374                 int lock_flags = LDLM_FL_ATOMIC_CB;
1375
1376                 child_res_id.name[0] = id_fid(&body->id1);
1377                 child_res_id.name[1] = id_group(&body->id1);
1378                 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1379
1380                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1381                                       child_res_id, LDLM_IBITS, &sz_policy,
1382                                       LCK_PW, &lock_flags, mds_blocking_ast,
1383                                       ldlm_completion_ast, NULL, NULL, NULL,
1384                                       0, NULL, &sz_lockh);
1385                 if (rc == ELDLM_OK)
1386                         ldlm_lock_decref(&sz_lockh, LCK_PW);
1387                 else
1388                         CERROR("can't invalidate client's update locks\n");
1389         }
1390
1391         EXIT;
1392 cleanup:
1393         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1394                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1395
1396 cleanup_no_trans:
1397         switch (cleanup_phase) {
1398         case 3:
1399                 if (rc) {
1400                         ldlm_lock_decref(child_lockh, child_mode);
1401                         child_lockh->cookie = 0;
1402                 }
1403         case 2:
1404                 if (rc && created) {
1405                         int err = vfs_unlink(dparent->d_inode, dchild);
1406                         if (err) {
1407                                 CERROR("unlink(%.*s) in error path: %d\n",
1408                                        dchild->d_name.len, dchild->d_name.name,
1409                                        err);
1410                         }
1411                 } else if (created) {
1412                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1413                 }
1414                 l_dput(dchild);
1415         case 1:
1416                 if (dparent == NULL)
1417                         break;
1418
1419                 l_dput(dparent);
1420 #ifdef S_PDIROPS
1421                 if (parent_lockh[1].cookie != 0)
1422                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1423 #endif
1424                 if (rc)
1425                         ldlm_lock_decref(parent_lockh, parent_mode);
1426                 else
1427                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1428         }
1429         if (mea)
1430                 OBD_FREE(mea, mea_size);
1431         if (rc == 0)
1432                 atomic_inc(&mds->mds_open_count);
1433
1434         mds_body_do_reverse_map(med, body);
1435
1436         /*
1437          * If we have not taken the "open" lock, we may not return 0 here,
1438          * because caller expects 0 to mean "lock is taken", and it needs
1439          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1440          * client. Later caller should rewrite the return value back to zero
1441          * if it to be used any further.
1442          */
1443         if ((cleanup_phase != 3) && !rc)
1444                 rc = ENOLCK;
1445
1446         RETURN(rc);
1447 }
1448
1449 /* Close a "file descriptor" and possibly unlink an orphan from the
1450  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1451  *
1452  * If we are being called from mds_disconnect() because the client has
1453  * disappeared, then req == NULL and we do not update last_rcvd because
1454  * there is nothing that could be recovered by the client at this stage
1455  * (it will not even _have_ an entry in last_rcvd anymore).
1456  *
1457  * Returns EAGAIN if the client needs to get more data and re-close. */
1458 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1459                   struct obd_device *obd, struct mds_file_data *mfd,
1460                   int unlink_orphan)
1461 {
1462         struct inode *inode = mfd->mfd_dentry->d_inode;
1463         char idname[LL_ID_NAMELEN];
1464         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1465         struct dentry *pending_child = NULL;
1466         struct mds_obd *mds = &obd->u.mds;
1467         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1468         void *handle = NULL;
1469         struct mds_body *request_body = NULL, *reply_body = NULL;
1470         struct dentry_params dp;
1471         struct iattr iattr = { 0 };
1472         struct llog_create_locks *lcl = NULL;
1473         ENTRY;
1474
1475         if (req && req->rq_reqmsg != NULL)
1476                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1477                                               sizeof(*request_body));
1478         if (req && req->rq_repmsg != NULL)
1479                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1480                                             sizeof(*reply_body));
1481
1482         if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
1483                 /* we set i_size/i_blocks here, nobody will see
1484                  * them until all write references are dropped.
1485                  * btw, we hold one reference */
1486                 LASSERT(mfd->mfd_mode & FMODE_WRITE);
1487                 i_size_write(inode, request_body->size);
1488                 inode->i_blocks = request_body->blocks;
1489                 iattr.ia_size = inode->i_size;
1490                 iattr.ia_valid |= ATTR_SIZE;
1491                 mds_inode_unset_attrs_old(inode);
1492         }
1493
1494         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1495         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1496                idname, inode->i_nlink, mds_orphan_open_count(inode));
1497
1498         last_orphan = mds_orphan_open_dec_test(inode) &&
1499                 mds_inode_is_orphan(inode);
1500         UP_WRITE_I_ALLOC_SEM(inode);
1501
1502         /* this is half of the actual "close" */
1503         if (mfd->mfd_mode & FMODE_WRITE) {
1504                 rc = mds_put_write_access(mds, inode, request_body,
1505                                           last_orphan && unlink_orphan);
1506         } else if (mfd->mfd_mode & FMODE_EXEC) {
1507                 mds_allow_write_access(inode);
1508         }
1509
1510         if (last_orphan && unlink_orphan) {
1511                 struct lov_mds_md *lmm = NULL;
1512                 int stripe_count = 0;
1513                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1514
1515                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1516
1517                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1518                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1519                         CERROR("found \"orphan\" %s %s with link count %d\n",
1520                                S_ISREG(inode->i_mode) ? "file" : "dir",
1521                                idname, inode->i_nlink);
1522                 /*
1523                  * sadly, there is no easy way to save pending_child from
1524                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1525                  * normally it will still be in the dcache.
1526                  */
1527                 down(&pending_dir->i_sem);
1528                 cleanup_phase = 1; /* up(i_sem) when finished */
1529                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1530                                                idlen);
1531                 if (IS_ERR(pending_child))
1532                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1533                 if (pending_child->d_inode == NULL) {
1534                         CERROR("orphan %s has been removed\n", idname);
1535                         GOTO(cleanup, rc = 0);
1536                 }
1537
1538                 cleanup_phase = 2; /* dput(pending_child) when finished */
1539                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1540                         rc = vfs_rmdir(pending_dir, pending_child);
1541                         if (rc)
1542                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1543                                        idname, rc);
1544                         goto out;
1545                 }
1546
1547                 if (req != NULL && req->rq_repmsg != NULL) {
1548                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1549                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1550                 }
1551
1552                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1553                                           NULL, stripe_count);
1554                 if (IS_ERR(handle)) {
1555                         rc = PTR_ERR(handle);
1556                         handle = NULL;
1557                         GOTO(cleanup, rc);
1558                 }
1559
1560                 pending_child->d_fsdata = (void *) &dp;
1561                 dp.p_inum = 0;
1562                 dp.p_ptr = req;
1563                 rc = vfs_unlink(pending_dir, pending_child);
1564                 if (rc)
1565                         CERROR("error unlinking orphan %s: rc %d\n",
1566                                idname, rc);
1567
1568                 if (req != NULL && req->rq_repmsg != NULL &&
1569                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1570                     mds_log_op_unlink(obd, pending_child->d_inode,
1571                                       lmm, req->rq_repmsg->buflens[1],
1572                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1573                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1574                         reply_body->valid |= OBD_MD_FLCOOKIE;
1575                 }
1576
1577                 rc = mds_destroy_object(obd, inode, 1);
1578                 if (rc) {
1579                         CERROR("cannot destroy OSS object on close, err %d\n",
1580                                rc);
1581                         rc = 0;
1582                 }
1583
1584                 goto out; /* Don't bother updating attrs on unlinked inode */
1585         } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
1586                 /* last writer closed file - let's update i_size/i_blocks */
1587                 mds_validate_size(obd, inode, request_body, &iattr);
1588         }
1589
1590 #if 0
1591         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1592                 /* Update the on-disk attributes if this was the last write
1593                  * close, and all information was provided (i.e., rc == 0)
1594                  *
1595                  * XXX this should probably be abstracted with mds_reint_setattr
1596                  */
1597
1598                 if (request_body->valid & OBD_MD_FLMTIME &&
1599                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1600                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1601                         iattr.ia_valid |= ATTR_MTIME;
1602                 }
1603                 if (request_body->valid & OBD_MD_FLCTIME &&
1604                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1605                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1606                         iattr.ia_valid |= ATTR_CTIME;
1607                 }
1608
1609                 /* XXX can't set block count with fsfilt_setattr (!) */
1610                 if (request_body->valid & OBD_MD_FLSIZE) {
1611                         iattr.ia_valid |= ATTR_SIZE;
1612                         iattr.ia_size = request_body->size;
1613                 }
1614                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1615                         iattr.ia_valid |= ATTR_BLOCKS;
1616                         iattr.ia_blocks = request_body->blocks
1617                 } */
1618
1619         }
1620 #endif
1621         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1622                 /* Only start a transaction to write out only the atime if
1623                  * it is more out-of-date than the specified limit.  If we
1624                  * are already going to write out the atime then do it anyway.
1625                  * */
1626                 if ((request_body->atime >
1627                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1628                     (iattr.ia_valid != 0 &&
1629                      request_body->atime > LTIME_S(inode->i_atime))) {
1630                         LTIME_S(iattr.ia_atime) = request_body->atime;
1631                         iattr.ia_valid |= ATTR_ATIME;
1632                 }
1633         }
1634
1635         if (iattr.ia_valid != 0) {
1636                 if (handle == NULL)
1637                         handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1638                 if (IS_ERR(handle))
1639                         GOTO(cleanup, rc = PTR_ERR(handle));
1640                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1641                 if (rc)
1642                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1643         }
1644 out:
1645         /* If other clients have this file open for write, rc will be > 0 */
1646         if (rc > 0)
1647                 rc = 0;
1648         if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
1649                         && !mds_inode_is_orphan(inode)
1650                         && atomic_read(&inode->i_writecount) == 0
1651                         && inode->i_nlink != 0) {
1652                 CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
1653                        inode->i_ino, inode->i_generation, inode->i_nlink);
1654         }
1655         l_dput(mfd->mfd_dentry);
1656         mds_mfd_destroy(mfd);
1657
1658  cleanup:
1659         atomic_dec(&mds->mds_open_count);
1660         if (req != NULL && reply_body != NULL) {
1661                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1662         } else if (handle) {
1663                 int err, force_sync = 0;
1664
1665                 if (req && req->rq_export)
1666                         force_sync = req->rq_export->exp_sync;
1667
1668                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1669                                     force_sync);
1670                 if (err) {
1671                         CERROR("error committing close: %d\n", err);
1672                         if (!rc)
1673                                 rc = err;
1674                 }
1675         }
1676
1677         switch (cleanup_phase) {
1678         case 2:
1679                 if (lcl != NULL)
1680                         ptlrpc_save_llog_lock(req, lcl);
1681                 dput(pending_child);
1682         case 1:
1683                 up(&pending_dir->i_sem);
1684         }
1685         RETURN(rc);
1686 }
1687
1688 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1689                                     struct ldlm_lock_desc *new, void *data,
1690                                     int flag)
1691 {
1692         struct lustre_handle lockh = { 0 };
1693         int rc;
1694         ENTRY;
1695
1696         switch (flag) {
1697         case LDLM_CB_BLOCKING:
1698                 ldlm_lock2handle(lock, &lockh);
1699                 rc = ldlm_cli_cancel(&lockh);
1700                 if (rc != ELDLM_OK)
1701                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
1702                 break;
1703         case LDLM_CB_CANCELING: {
1704                 break;
1705         }
1706         default:
1707                 LBUG();
1708         }
1709
1710         RETURN(0);
1711 }
1712 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1713 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1714
1715 int mds_validate_size(struct obd_device *obd, struct inode *inode,
1716                       struct mds_body *body, struct iattr *iattr)
1717 {
1718         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1719         struct lustre_handle lockh = { 0 };
1720         struct lov_stripe_md *lsm = NULL;
1721         int rc, len, flags;
1722         void *lmm = NULL;
1723         ENTRY;
1724
1725         /* we update i_size/i_blocks only for regular files */
1726         if (!S_ISREG(inode->i_mode))
1727                 RETURN(0);
1728
1729         /* we shouldn't fetch size from OSTes during recovery - deadlock */
1730         if (obd->obd_recovering) {
1731                 CERROR("size-on-mds has no support on OST yet\n");
1732                 RETURN(0);
1733         }
1734
1735         /* if nobody modified attrs. we're lucky */
1736         if (!mds_inode_has_old_attrs(inode))
1737                 RETURN(0);
1738
1739         /* 1: client didn't send actual i_size/i_blocks
1740          * 2: we seem to be last writer
1741          * 3: the file doesn't look like to be disappeared
1742          * conclusion: we're gonna fetch them from OSSes */
1743
1744         down(&inode->i_sem);
1745         len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1746         up(&inode->i_sem);
1747
1748         if (len < 0) {
1749                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1750                 GOTO(cleanup, rc = len);
1751         } else if (len == 0) {
1752                 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1753                 GOTO(cleanup, rc = 0);
1754         }
1755
1756         OBD_ALLOC(lmm, len);
1757         if (lmm == NULL) {
1758                 CERROR("can't allocate memory\n");
1759                 GOTO(cleanup, rc = -ENOMEM);
1760         }
1761
1762         down(&inode->i_sem);
1763         rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1764         up(&inode->i_sem);
1765
1766         if (rc < 0) {
1767                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1768                 GOTO(cleanup, rc);
1769         }
1770
1771         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1772         if (rc < 0) {
1773                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1774                 GOTO(cleanup, rc);
1775         }
1776
1777         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1778
1779         flags = LDLM_FL_HAS_INTENT;
1780         rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1781                          LCK_PR, &flags, mds_extent_lock_callback,
1782                          ldlm_completion_ast, NULL, NULL,
1783                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1784         if (rc == -ENOENT) {
1785                 /* while we were enqueueing lock on OST, another thread
1786                  * unlinked the file and started OST object destoying.
1787                  * it's safe to return 0 here */
1788                 GOTO(cleanup, rc = 0);
1789         } else if (rc != 0) {
1790                 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1791                 GOTO(cleanup, rc);
1792         }
1793
1794         CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n",
1795                inode->i_size, inode->i_blocks, OLID4(&body->id1),
1796                atomic_read(&inode->i_writecount) > 1 ? "U" : "",
1797                mds_inode_has_old_attrs(inode) ? "D" : "",
1798                mds_inode_is_orphan(inode) ? "O" : "");
1799
1800         i_size_write(inode, lov_merge_size(lsm, 0));
1801         inode->i_blocks = lov_merge_blocks(lsm);
1802         iattr->ia_size = inode->i_size;
1803         iattr->ia_valid |= ATTR_SIZE;
1804         DOWN_WRITE_I_ALLOC_SEM(inode);
1805         mds_inode_unset_attrs_old(inode);
1806         UP_WRITE_I_ALLOC_SEM(inode);
1807
1808         obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1809
1810 cleanup:
1811         if (lsm != NULL)
1812                 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1813         if (lmm != NULL)
1814                 OBD_FREE(lmm, len);
1815         RETURN(rc);
1816 }
1817
1818 int mds_close(struct ptlrpc_request *req, int offset)
1819 {
1820         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1821         struct obd_device *obd = req->rq_export->exp_obd;
1822         struct mds_body *body;
1823         struct mds_file_data *mfd;
1824         struct lvfs_run_ctxt saved;
1825         struct inode *inode;
1826         int rc, repsize[3] = {sizeof(struct mds_body),
1827                               obd->u.mds.mds_max_mdsize,
1828                               obd->u.mds.mds_max_cookiesize};
1829         ENTRY;
1830         MD_COUNTER_INCREMENT(obd, close);
1831
1832         rc = lustre_pack_reply(req, 3, repsize, NULL);
1833         if (rc) {
1834                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1835                 req->rq_status = rc;
1836         } else {
1837                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1838         }
1839
1840         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1841                 DEBUG_REQ(D_HA, req, "close replay");
1842                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1843                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1844                        req->rq_repmsg->buflens[2]);
1845         }
1846
1847         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1848                                   lustre_swab_mds_body);
1849         if (body == NULL) {
1850                 CERROR("Can't unpack body\n");
1851                 req->rq_status = -EFAULT;
1852                 RETURN(-EFAULT);
1853         }
1854
1855         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1856                 /* do some stuff */ ;
1857
1858         mfd = mds_handle2mfd(&body->handle);
1859         if (mfd == NULL) {
1860                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1861                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1862                 req->rq_status = -ESTALE;
1863                 RETURN(-ESTALE);
1864         }
1865
1866         inode = mfd->mfd_dentry->d_inode;
1867
1868         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1869         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1870
1871         if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
1872                 /* the client modified data through the handle
1873                  * we need to care about attrs. -bzzz */
1874                 mds_inode_set_attrs_old(inode);
1875         }
1876
1877         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1878                 struct mds_body *rep_body;
1879
1880                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1881                                           sizeof (*rep_body));
1882                 LASSERT(rep_body != NULL);
1883
1884                 mds_pack_inode2body(obd, rep_body, inode,
1885                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1886
1887                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1888                             inode, MDS_PACK_MD_LOCK, 0);
1889         }
1890         spin_lock(&med->med_open_lock);
1891         list_del(&mfd->mfd_list);
1892         spin_unlock(&med->med_open_lock);
1893
1894         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1895         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1896         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1897
1898         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1899                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1900                 req->rq_status = -ENOMEM;
1901                 mds_mfd_put(mfd);
1902                 RETURN(-ENOMEM);
1903         }
1904
1905         mds_mfd_put(mfd);
1906         RETURN(0);
1907 }
1908
1909
1910 int mds_done_writing(struct ptlrpc_request *req, int offset)
1911 {
1912         struct mds_body *body;
1913         int rc, size = sizeof(struct mds_body);
1914         ENTRY;
1915
1916         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1917
1918         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1919                                   lustre_swab_mds_body);
1920         if (body == NULL) {
1921                 CERROR("Can't unpack body\n");
1922                 req->rq_status = -EFAULT;
1923                 RETURN(-EFAULT);
1924         }
1925
1926         rc = lustre_pack_reply(req, 1, &size, NULL);
1927         if (rc) {
1928                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1929                 req->rq_status = rc;
1930         }
1931
1932         RETURN(0);
1933 }