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 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                 /* pack lustre id to oss */
 445                 *(obdo_id(oa)) = body->id1;
 446                 oa->o_valid |= OBD_MD_FLIFID;
 447
 448                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti, NULL);
 449                 if (rc) {
 450                         CERROR("error setting attrs for inode %lu: rc %d\n",
 451                                inode->i_ino, rc);
 452                         if (rc > 0) {
 453                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 454                                 rc = -EIO;
 455                         }
 456                         GOTO(out_oa, rc);
 457                 }
 458         }
 459
 460         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 461         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 462
 463         LASSERT(lsm && lsm->lsm_object_id);
 464         lmm = NULL;
 465         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 466         if (!id_ino(rec->ur_id2))
 467                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 468         if (rc < 0) {
 469                 CERROR("cannot pack lsm, err = %d\n", rc);
 470                 GOTO(out_oa, rc);
 471         }
 472
 473         lmm_size = rc;
 474         body->eadatasize = rc;
 475
 476         if (*handle == NULL)
 477                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 478         if (IS_ERR(*handle)) {
 479                 rc = PTR_ERR(*handle);
 480                 *handle = NULL;
 481                 GOTO(out_oa, rc);
 482         }
 483
 484         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 485                            lmm_size, EA_LOV);
 486
 487         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 488         lmm_bufsize = req->rq_repmsg->buflens[offset];
 489         LASSERT(lmm_buf);
 490         LASSERT(lmm_bufsize >= lmm_size);
 491
 492         memcpy(lmm_buf, lmm, lmm_size);
 493         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 494 out_oa:
 495         oti_free_cookies(&oti);
 496         obdo_free(oa);
 497 out_ids:
 498         if (rc) {
 499                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 500                 *ids = NULL;
 501         }
 502         if (lsm)
 503                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 504         RETURN(rc);
 505 }
 506
 507 int
 508 mds_destroy_object(struct obd_device *obd,
 509                    struct inode *inode, int async)
 510 {
 511         struct mds_obd *mds = &obd->u.mds;
 512         struct lov_mds_md *lmm = NULL;
 513         int rc, lmm_size;
 514         ENTRY;
 515
 516         LASSERT(inode != NULL);
 517
 518         if (inode->i_nlink != 0) {
 519                 CDEBUG(D_INODE, "attempt to destroy OSS object when "
 520                        "i_nlink == %d\n", (int)inode->i_nlink);
 521                 RETURN(0);
 522         }
 523
 524         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 525         if (lmm == NULL)
 526                 RETURN(-ENOMEM);
 527
 528         lmm_size = mds->mds_max_mdsize;
 529         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 530         if (rc < 0) {
 531                 CERROR("no stripe info for %lu/%lu inode\n",
 532                        (unsigned long)inode->i_ino,
 533                        (unsigned long)inode->i_generation);
 534                 GOTO(out_free_lmm, rc);
 535         }
 536
 537         if (rc > 0) {
 538                 /* asynchronously unlink objecect on OSS */
 539                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 540                                        NULL, 0, async);
 541                 if (rc) {
 542                         CERROR("error unlinking object on OSS, "
 543                                "err %d\n", rc);
 544                         GOTO(out_free_lmm, rc);
 545                 }
 546         } else {
 547                 CDEBUG(D_INODE, "no stripping info found for inode "
 548                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 549                       (unsigned long)inode->i_generation);
 550         }
 551         EXIT;
 552 out_free_lmm:
 553         OBD_FREE(lmm, mds->mds_max_mdsize);
 554         return rc;
 555 }
 556
 557 static void reconstruct_open(struct mds_update_record *rec, int offset,
 558                              struct ptlrpc_request *req,
 559                              struct lustre_handle *child_lockh)
 560 {
 561         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 562         struct mds_client_data *mcd = med->med_mcd;
 563         struct mds_obd *mds = mds_req2mds(req);
 564         struct mds_file_data *mfd;
 565         struct obd_device *obd = req->rq_export->exp_obd;
 566         struct dentry *parent = NULL, *dchild;
 567         struct ldlm_reply *rep;
 568         struct mds_body *body;
 569         struct list_head *t;
 570         int put_child = 1;
 571         int rc;
 572         ENTRY;
 573
 574         LASSERT(offset == 3); /* only called via intent */
 575         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 576         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 577
 578         /* copy rc, transno and disp; steal locks */
 579         mds_req_from_mcd(req, mcd);
 580         intent_set_disposition(rep, mcd->mcd_last_data);
 581
 582         /* Only replay if create or open actually happened. */
 583         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 584                 EXIT;
 585                 return; /* error looking up parent or child */
 586         }
 587
 588         /* first, we try to open the file by fid. by the time of this
 589          * request, inode can be an orphan and parent can disappear */
 590         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 591                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT|REPLAY)\n",
 592                        OLID4(rec->ur_id2));
 593                 dchild = mds_id2dentry(obd, rec->ur_id2, NULL);
 594         } else if (rec->ur_namelen == 1) {
 595                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 596                        OLID4(rec->ur_id1));
 597                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 598         } else {
 599                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 600                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 601                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 602                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 603                                 rec->ur_namelen - 1);
 604                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 605                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 606         }
 607
 608         if (!dchild->d_inode)
 609                 GOTO(out_dput, 0); /* child not present to open */
 610
 611         /* At this point, we know we have a child. We'll send
 612          * it back _unless_ it not created and open failed.  */
 613         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 614             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 615             req->rq_status)
 616                 GOTO(out_dput, 0);
 617
 618         /* get lock (write for O_CREAT, read otherwise) */
 619         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 620         if (S_ISREG(dchild->d_inode->i_mode)) {
 621                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 622                                  dchild->d_inode, 1, 0);
 623
 624                 if (rc)
 625                         LASSERT(rc == req->rq_status);
 626
 627                 /* If we have LOV EA data, the OST holds size, mtime */
 628                 if (!(body->valid & OBD_MD_FLEASIZE))
 629                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 630                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 631         }
 632
 633         /* If we have -EEXIST as the status, and we were asked to create
 634          * exclusively, we can tell we failed because the file already existed.
 635          */
 636         if (req->rq_status == -EEXIST &&
 637             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 638              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 639                 GOTO(out_dput, 0);
 640         }
 641
 642         /* If we didn't get as far as trying to open, then some locking thing
 643          * probably went wrong, and we'll just bail here.
 644          */
 645         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 646                 GOTO(out_dput, 0);
 647
 648         /* If we failed, then we must have failed opening, so don't look for
 649          * file descriptor or anything, just give the client the bad news.
 650          */
 651         if (req->rq_status)
 652                 GOTO(out_dput, 0);
 653
 654         mfd = NULL;
 655         list_for_each(t, &med->med_open_head) {
 656                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 657                 if (mfd->mfd_xid == req->rq_xid)
 658                         break;
 659                 mfd = NULL;
 660         }
 661
 662         /* #warning "XXX fixme" bug 2991 */
 663         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 664          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 665          * to detect a re-open */
 666         if (mfd == NULL) {
 667                 mntget(mds->mds_vfsmnt);
 668                 CERROR("Re-opened file \n");
 669                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 670                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 671                 if (!mfd) {
 672                         CERROR("mds: out of memory\n");
 673                         GOTO(out_dput, req->rq_status = -ENOMEM);
 674                 }
 675                 put_child = 0;
 676         } else {
 677                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 678                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 679                        mfd->mfd_handle.h_cookie);
 680         }
 681
 682  out_dput:
 683         if (put_child)
 684                 l_dput(dchild);
 685         if (parent)
 686                 l_dput(parent);
 687         EXIT;
 688 }
 689
 690 /* do NOT or the MAY_*'s, you'll get the weakest */
 691 int accmode(int flags)
 692 {
 693         int res = 0;
 694
 695         if (flags & FMODE_READ)
 696                 res = MAY_READ;
 697         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 698                 res |= MAY_WRITE;
 699         if (flags & FMODE_EXEC)
 700                 res = MAY_EXEC;
 701         return res;
 702 }
 703
 704 /* Handles object creation, actual opening, and I/O epoch */
 705 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 706                            struct mds_body *body, int flags, void **handle,
 707                            struct mds_update_record *rec, struct ldlm_reply *rep)
 708 {
 709         struct obd_device *obd = req->rq_export->exp_obd;
 710         struct mds_obd *mds = mds_req2mds(req);
 711         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 712         struct mds_file_data *mfd = NULL;
 713         obd_id *ids = NULL;
 714         unsigned mode;
 715         int rc = 0, reply_off;
 716         ENTRY;
 717
 718         /* atomically create objects if necessary */
 719         down(&dchild->d_inode->i_sem);
 720         mode = dchild->d_inode->i_mode;
 721
 722         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 723             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 724                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 725                                  dchild->d_inode, 0, 0);
 726                 if (rc) {
 727                         up(&dchild->d_inode->i_sem);
 728                         RETURN(rc);
 729                 }
 730         }
 731
 732         if (rec != NULL) {
 733                 if ((body->valid & OBD_MD_FLEASIZE) &&
 734                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 735                         up(&dchild->d_inode->i_sem);
 736                         RETURN(-EEXIST);
 737                 }
 738
 739                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 740                         /* no EA: create objects */
 741                         rc = mds_create_objects(obd, req, 2, rec,
 742                                                 dchild, handle, &ids);
 743                         if (rc) {
 744                                 CERROR("mds_create_object: rc = %d\n", rc);
 745                                 up(&dchild->d_inode->i_sem);
 746                                 RETURN(rc);
 747                         }
 748                 }
 749
 750                 if (S_ISREG(dchild->d_inode->i_mode) &&
 751                     (body->valid & OBD_MD_FLEASIZE)) {
 752                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 753                                                    req->rq_repmsg, 2);
 754                         if (!rc)
 755                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 756                                                  dchild->d_inode, 0, 0);
 757                         if (rc) {
 758                                 up(&dchild->d_inode->i_sem);
 759                                 RETURN(rc);
 760                         }
 761                 }
 762         }
 763
 764         reply_off = 3;
 765         rc = mds_pack_acl(req, reply_off, body, dchild->d_inode);
 766         reply_off += 2;
 767
 768         if (rc < 0) {
 769                 CERROR("pack posix acl: rc = %d\n", rc);
 770                 up(&dchild->d_inode->i_sem);
 771                 RETURN(rc);
 772         }
 773
 774         rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
 775                             dchild->d_inode);
 776         if (rc < 0) {
 777                 CERROR("mds_pack_gskey: rc = %d\n", rc);
 778                 up(&dchild->d_inode->i_sem);
 779                 RETURN(rc);
 780         }
 781
 782         if (S_ISREG(mode)) {
 783                 struct lustre_capa capa = {
 784                         .lc_uid   = rec->ur_uc.luc_uid,
 785                         .lc_op    = capa_op(rec->ur_flags),
 786                         .lc_ino   = dchild->d_inode->i_ino,
 787                         .lc_mdsid = mds->mds_num,
 788                 };
 789
 790                 rc = mds_pack_capa(obd, med, NULL, &capa, req,
 791                                    &reply_off, body);
 792                 if (rc < 0) {
 793                         CERROR("mds_pack_capa: rc = %d\n", rc);
 794                         up(&dchild->d_inode->i_sem);
 795                         RETURN(rc);
 796                 }
 797         } else {
 798                 reply_off++;
 799         }
 800
 801         /* If the inode has no EA data, then MDSs hold size, mtime */
 802         if (S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) {
 803                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 804                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 805         }
 806
 807         up(&dchild->d_inode->i_sem);
 808
 809         intent_set_disposition(rep, DISP_OPEN_OPEN);
 810         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 811         if (IS_ERR(mfd))
 812                 RETURN(PTR_ERR(mfd));
 813
 814         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 815                mfd->mfd_handle.h_cookie);
 816
 817         if (ids != NULL) {
 818                 mds_dt_update_objids(obd, ids);
 819                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 820         }
 821
 822         RETURN(rc);
 823 }
 824
 825 static int mds_open_by_id(struct ptlrpc_request *req,
 826                           struct lustre_id *id,
 827                           struct mds_body *body, int flags,
 828                           struct mds_update_record *rec,
 829                           struct ldlm_reply *rep)
 830 {
 831         struct mds_obd *mds = mds_req2mds(req);
 832         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 833         struct dentry *dchild;
 834         char idname[LL_ID_NAMELEN];
 835         int idlen = 0, rc;
 836         void *handle = NULL;
 837         ENTRY;
 838
 839         down(&pending_dir->i_sem);
 840         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 841         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 842                                 idlen);
 843         if (IS_ERR(dchild)) {
 844                 up(&pending_dir->i_sem);
 845                 rc = PTR_ERR(dchild);
 846                 CERROR("error looking up %s in PENDING: rc = %d\n",
 847                        idname, rc);
 848                 RETURN(rc);
 849         }
 850
 851         up(&pending_dir->i_sem);
 852         if (dchild->d_inode != NULL) {
 853                 mds_inode_set_orphan(dchild->d_inode);
 854                 mds_pack_inode2body(req2obd(req), body,
 855                                     dchild->d_inode, 1);
 856
 857                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 858                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 859                 CWARN("Orphan %s found and opened in PENDING directory\n",
 860                        idname);
 861                 goto open;
 862         }
 863         l_dput(dchild);
 864
 865         /* we didn't find it in PENDING so it isn't an orphan.  See if it was a
 866          * regular inode that was previously created. */
 867         dchild = mds_id2dentry(req2obd(req), id, NULL);
 868         if (IS_ERR(dchild))
 869                 RETURN(PTR_ERR(dchild));
 870
 871         mds_pack_inode2body(req2obd(req), body,
 872                             dchild->d_inode, 1);
 873
 874         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 875         intent_set_disposition(rep, DISP_LOOKUP_POS);
 876
 877  open:
 878         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 879         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 880                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 881         /* XXX what do we do here if mds_finish_transno itself failed? */
 882         l_dput(dchild);
 883         RETURN(rc);
 884 }
 885
 886 int mds_pin(struct ptlrpc_request *req, int offset)
 887 {
 888         struct obd_device *obd = req->rq_export->exp_obd;
 889         struct mds_body *request_body, *reply_body;
 890         struct lvfs_run_ctxt saved;
 891         int rc, size = sizeof(*reply_body);
 892         ENTRY;
 893
 894         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 895                                       sizeof(*request_body));
 896
 897         rc = lustre_pack_reply(req, 1, &size, NULL);
 898         if (rc)
 899                 RETURN(rc);
 900         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 901
 902         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 903         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 904                             request_body->flags, NULL, NULL);
 905         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 906
 907         RETURN(rc);
 908 }
 909
 910 /*
 911  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 912  * child_lockh is NULL we just get the lock as a barrier to wait for other
 913  * holders of this lock, and drop it right away again.
 914  */
 915 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 916                        struct lustre_handle *child_lockh)
 917 {
 918         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 919         struct lustre_handle lockh;
 920         int lock_flags = LDLM_FL_ATOMIC_CB;
 921         int rc;
 922         ENTRY;
 923
 924         if (child_lockh == NULL)
 925                 child_lockh = &lockh;
 926
 927         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 928                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 929                               mds_blocking_ast, ldlm_completion_ast, NULL,
 930                               NULL, NULL, 0, NULL, child_lockh);
 931         if (rc != ELDLM_OK)
 932                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 933         else if (child_lockh == &lockh)
 934                 ldlm_lock_decref(child_lockh, LCK_EX);
 935
 936         RETURN(rc);
 937 }
 938
 939 int mds_open(struct mds_update_record *rec, int offset,
 940              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 941 {
 942         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 943         struct obd_device *obd = req->rq_export->exp_obd;
 944         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 945         struct mds_obd *mds = mds_req2mds(req);
 946         struct ldlm_reply *rep = NULL;
 947         struct mds_body *body = NULL;
 948         struct dentry *dchild = NULL, *dparent = NULL;
 949         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 950         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 951         int parent_mode = LCK_PR;
 952         void *handle = NULL;
 953         struct dentry_params dp;
 954         struct mea *mea = NULL;
 955         int mea_size, update_mode;
 956         int child_mode = LCK_PR;
 957         struct lustre_id sid;
 958         __u64 fid = 0;
 959         ENTRY;
 960
 961         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 962                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 963                   rec->ur_mode);
 964
 965         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 966                          (obd_timeout + 1) / 4);
 967
 968         if (offset == 3) { /* intent */
 969                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 970                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 971         } else if (offset == 1) { /* non-intent reint */
 972                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 973         } else {
 974                 body = NULL;
 975                 LBUG();
 976         }
 977
 978         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 979
 980         /*
 981          * Step 0: If we are passed a id, then we assume the client already
 982          * opened this file and is only replaying the RPC, so we open the inode
 983          * by fid (at some large expense in security).
 984          */
 985         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 986                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 987                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 988                        rec->ur_name, rec->ur_mode);
 989
 990                 LASSERT(id_ino(rec->ur_id2));
 991
 992                 rc = mds_open_by_id(req, rec->ur_id2, body,
 993                                     rec->ur_flags, rec, rep);
 994                 if (rc != -ENOENT) {
 995                         mds_body_do_reverse_map(med, body);
 996                         RETURN(rc);
 997                 }
 998
 999                 /* We didn't find the correct inode on disk either, so we
1000                  * need to re-create it via a regular replay. */
1001                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
1002         } else {
1003                 LASSERT(!id_ino(rec->ur_id2));
1004         }
1005
1006         LASSERT(offset == 3); /* If we got here, we must be called via intent */
1007
1008         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1009                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1010                 RETURN(-ENOMEM);
1011         }
1012
1013         acc_mode = accmode(rec->ur_flags);
1014
1015         /* Step 1: Find and lock the parent */
1016         if (rec->ur_flags & MDS_OPEN_CREAT) {
1017                 /* XXX Well, in fact we only need this lock mode change if
1018                    in addition to O_CREAT, the file does not exist.
1019                    But we do not know if it exists or not yet */
1020                 parent_mode = LCK_PW;
1021         }
1022
1023         if (rec->ur_namelen == 1) {
1024                 /* client (LMV) wants to open the file by id */
1025                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
1026                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
1027                 if (IS_ERR(dchild)) {
1028                         rc = PTR_ERR(dparent);
1029                         CERROR("child lookup by an id error %d\n", rc);
1030                         GOTO(cleanup, rc);
1031                 }
1032                 goto got_child;
1033         }
1034
1035 restart:
1036         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1037                                        parent_lockh, &update_mode, rec->ur_name,
1038                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1039         if (IS_ERR(dparent)) {
1040                 rc = PTR_ERR(dparent);
1041                 if (rc != -ENOENT) {
1042                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1043                                OLID4(rec->ur_id1), rc);
1044                 } else {
1045                         /* Just cannot find parent - make it look like
1046                            usual negative lookup to avoid extra MDS RPC */
1047                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1048                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1049                 }
1050                 GOTO(cleanup, rc);
1051         }
1052         LASSERT(dparent->d_inode != NULL);
1053
1054         cleanup_phase = 1; /* parent dentry and lock */
1055
1056         /* try to retrieve MEA data for this dir */
1057         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1058         if (rc)
1059                 GOTO(cleanup, rc);
1060
1061         if (mea != NULL && mea->mea_count) {
1062                 /*
1063                  * dir is already splitted, check is requested filename should *
1064                  * live at this MDS or at another one.
1065                  */
1066                 int i;
1067                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1068                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1069                         CDEBUG(D_OTHER,
1070                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1071                                " should be %lu(%d)\n", obd->obd_name,
1072                                mea->mea_master, dparent->d_inode->i_ino,
1073                                dparent->d_inode->i_generation, rec->ur_name,
1074                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1075                         GOTO(cleanup, rc = -ERESTART);
1076                 }
1077         }
1078
1079         /* Step 2: Lookup the child */
1080         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1081         if (IS_ERR(dchild)) {
1082                 rc = PTR_ERR(dchild);
1083                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1084                 GOTO(cleanup, rc);
1085         }
1086
1087 got_child:
1088         cleanup_phase = 2; /* child dentry */
1089
1090         if (dchild->d_flags & DCACHE_CROSS_REF) {
1091                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1092                 LASSERT(rec->ur_namelen > 1);
1093
1094                 /* we're gonna acquire LOOKUP lock on the child,
1095                  * but we have already locked parent and our order
1096                  * may conflict with enqueue_order_locks(). so,
1097                  * drop parent lock and acquire both the locks in
1098                  * common order. bug 6190 */
1099 #ifdef S_PDIROPS
1100                 if (parent_lockh[1].cookie != 0)
1101                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1102 #endif
1103                 ldlm_lock_decref(parent_lockh, parent_mode);
1104                 l_dput(dchild);
1105                 l_dput(dparent);
1106                 cleanup_phase = 0;
1107
1108                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1109                                                  parent_lockh, &dparent,
1110                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1111                                                  &update_mode, rec->ur_name,
1112                                                  rec->ur_namelen, child_lockh,
1113                                                  &dchild, LCK_PR,
1114                                                  MDS_INODELOCK_LOOKUP);
1115
1116                 if (rc)
1117                         GOTO(cleanup, rc);
1118
1119 #ifdef S_PDIROPS
1120                 if (parent_lockh[1].cookie != 0)
1121                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1122 #endif
1123                 ldlm_lock_decref(parent_lockh, LCK_PR);
1124
1125                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1126                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1127                                dchild->d_inode);
1128                         if (dchild->d_inode)
1129                                 ldlm_lock_decref(child_lockh, LCK_PR);
1130                         l_dput(dchild);
1131                         l_dput(dparent);
1132                         GOTO(restart, rc);
1133                 }
1134
1135                 mds_pack_dentry2body(obd, body, dchild, 1);
1136                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1137                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1138
1139                 if (mea)
1140                         OBD_FREE(mea, mea_size);
1141                 l_dput(dchild);
1142                 l_dput(dparent);
1143                 RETURN(rc);
1144         }
1145
1146         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1147
1148         if (dchild->d_inode)
1149                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1150         else
1151                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1152
1153         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1154         if (dchild->d_inode == NULL) {
1155                 unsigned long ino;
1156                 struct iattr iattr;
1157                 struct inode *inode;
1158
1159                 /* get parent id: ldlm lock on the parent protects ea */
1160                 rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
1161                 if (rc) {
1162                         CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
1163                                 dparent->d_inode->i_ino, rc);
1164                         GOTO(cleanup, rc);
1165                 }
1166
1167                 ino = (unsigned long)id_ino(rec->ur_id2);
1168
1169                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1170                                           update_mode);
1171
1172                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1173                        obd->obd_name, dparent->d_inode->i_ino,
1174                        dparent->d_inode->i_generation, rc);
1175
1176                 if (rc > 0) {
1177                         /* dir got splitted */
1178                         GOTO(cleanup, rc = -ERESTART);
1179                 } else if (rc < 0) {
1180                         /* error happened during spitting */
1181                         GOTO(cleanup, rc);
1182                 }
1183
1184                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1185                         /*
1186                          * dentry is negative and we weren't supposed to create
1187                          * object, get out of here.
1188                          */
1189                         GOTO(cleanup, rc = -ENOENT);
1190                 }
1191
1192                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1193                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1194                                       NULL);
1195                 if (IS_ERR(handle)) {
1196                         rc = PTR_ERR(handle);
1197                         handle = NULL;
1198                         GOTO(cleanup, rc);
1199                 }
1200                 if (id_fid(rec->ur_id2))
1201                         fid = id_fid(rec->ur_id2);
1202                 else
1203                         fid = mds_alloc_fid(obd);
1204
1205                 dchild->d_fsdata = (void *) &dp;
1206                 dp.p_ptr = req;
1207                 dp.p_inum = ino;
1208
1209                 dp.p_fid = fid;
1210                 dp.p_group = mds->mds_num;
1211
1212                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1213                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1214                         dchild->d_fsdata = NULL;
1215
1216                 if (rc) {
1217                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1218                         GOTO(cleanup, rc);
1219                 }
1220
1221                 inode = dchild->d_inode;
1222
1223                 if (ino) {
1224                         LASSERT(ino == inode->i_ino);
1225
1226                         /* written as part of setattr */
1227                         inode->i_generation = id_gen(rec->ur_id2);
1228                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1229                                inode->i_ino, inode->i_generation);
1230                 }
1231
1232                 created = 1;
1233                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1234                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1235                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1236
1237                 iattr.ia_uid = rec->ur_fsuid;
1238                 if (dparent->d_inode->i_mode & S_ISGID)
1239                         iattr.ia_gid = dparent->d_inode->i_gid;
1240                 else
1241                         iattr.ia_gid = rec->ur_fsgid;
1242
1243                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1244                         ATTR_MTIME | ATTR_CTIME;
1245
1246                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1247                 if (rc)
1248                         CERROR("error on child setattr: rc = %d\n", rc);
1249
1250                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1251
1252                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1253                 if (rc)
1254                         CERROR("error on parent setattr: rc = %d\n", rc);
1255                 else {
1256                         MD_COUNTER_INCREMENT(obd, create);
1257                 }
1258
1259                 mds_inode2id(obd, &body->id1, inode, fid);
1260                 mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
1261
1262                 if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) {
1263                         rc = mds_set_gskey(obd, handle, dchild->d_inode,
1264                                            rec->ur_ea2data, rec->ur_ea2datalen,
1265                                            ATTR_KEY | ATTR_MAC);
1266                         if (rc) {
1267                                 CERROR("error in set gs key rc %d\n", rc);
1268                         }
1269                 }
1270
1271                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1272                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1273                                            dchild->d_inode, handle,
1274                                            req->rq_export->exp_sync);
1275                         handle = NULL;
1276                 }
1277
1278                 acc_mode = 0;           /* Don't check for permissions */
1279         }
1280         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1281
1282         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1283                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1284                  dchild->d_name.name, dchild, dchild->d_inode);
1285
1286         if (S_ISREG(dchild->d_inode->i_mode)) {
1287                 /* Check permissions etc */
1288                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1289                 if (rc != 0)
1290                         GOTO(cleanup, rc);
1291
1292                 /* Can't write to a read-only file */
1293                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1294                         GOTO(cleanup, rc = -EPERM);
1295
1296                 /* An append-only file must be opened in append mode for
1297                  * writing */
1298                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1299                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1300                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1301                         GOTO(cleanup, rc = -EPERM);
1302         }
1303
1304         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1305             (rec->ur_flags & MDS_OPEN_EXCL)) {
1306                 /* File already exists, we didn't just create it, and we
1307                  * were passed O_EXCL; err-or. */
1308                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1309         }
1310
1311         if (S_ISDIR(dchild->d_inode->i_mode)) {
1312                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1313                     rec->ur_flags & FMODE_WRITE) {
1314                         /* we are trying to create or write a exist dir */
1315                         GOTO(cleanup, rc = -EISDIR);
1316                 }
1317                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1318                         GOTO(cleanup, rc = -EACCES);
1319
1320                 /*
1321                  * here was checking for possible GNS mount points to skip their
1322                  * open. I removed it as its detection based only on SUID bit is
1323                  * not reliable. Opening such a dirs should not cause any
1324                  * problems, at least tests show that. --umka
1325                  */
1326         }
1327
1328         /* if we are following a symlink, don't open */
1329         if (S_ISLNK(dchild->d_inode->i_mode))
1330                 GOTO(cleanup_no_trans, rc = 0);
1331
1332         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1333             !S_ISDIR(dchild->d_inode->i_mode))
1334                 GOTO(cleanup, rc = -ENOTDIR);
1335
1336         /* check permission even it's special files */
1337         if (S_ISCHR(dchild->d_inode->i_mode) ||
1338             S_ISBLK(dchild->d_inode->i_mode) ||
1339             S_ISFIFO(dchild->d_inode->i_mode) ||
1340             S_ISSOCK(dchild->d_inode->i_mode)) {
1341                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1342                 if (rc != 0)
1343                         GOTO(cleanup, rc);
1344         }
1345
1346         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1347                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1348                 GOTO(cleanup, rc = -EAGAIN);
1349         }
1350
1351 #if 0
1352         /* We cannot use acc_mode here, because it is zeroed in case of
1353            creating a file, so we get wrong lockmode */
1354         if (accmode(rec->ur_flags) & MAY_WRITE)
1355                 child_mode = LCK_CW;
1356         else if (accmode(rec->ur_flags) & MAY_EXEC)
1357                 child_mode = LCK_PR;
1358         else
1359                 child_mode = LCK_CR;
1360
1361         /* Always returning LOOKUP lock if open succesful to guard
1362          * dentry on client. In case of replay we do not get a lock
1363          * assuming that the caller has it already */
1364         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1365                 struct ldlm_res_id child_res_id = { .name = {0}};
1366                 ldlm_policy_data_t policy;
1367                 int lock_flags = LDLM_FL_ATOMIC_CB;
1368
1369                 /* LOOKUP lock will protect dentry on client -bzzz */
1370                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1371                                                 MDS_INODELOCK_OPEN;
1372
1373                 child_res_id.name[0] = id_fid(&body->id1);
1374                 child_res_id.name[1] = id_group(&body->id1);
1375
1376                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1377                                       child_res_id, LDLM_IBITS, &policy,
1378                                       child_mode, &lock_flags,
1379                                       mds_blocking_ast, ldlm_completion_ast,
1380                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1381                 if (rc != ELDLM_OK)
1382                         GOTO(cleanup, rc);
1383
1384                 cleanup_phase = 3;
1385         }
1386 #else
1387 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1388 #warning "disable opencache lock for CMD2"
1389 #endif
1390
1391         /* Step 5: mds_open it */
1392         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1393                              rec, rep);
1394         if (rc)
1395                 GOTO(cleanup, rc);
1396
1397         /* reintegration case */
1398         if ((rec->ur_flags & MDS_REINT_REQ)) {
1399                 rc = mds_fidmap_add(obd, &body->id1);
1400                 if (rc < 0) {
1401                         CERROR("can't create fid->ino mapping, err %d\n",
1402                                rc);
1403                 } else {
1404                         rc = 0;
1405                 }
1406         }
1407
1408         /* if this is a writer, we have to invalidate client's
1409          * update locks in order to make sure they don't use
1410          * isize/iblocks from mds anymore.
1411          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1412          * XXX: optimization is to do this for first writer only */
1413         if (accmode(rec->ur_flags) & MAY_WRITE) {
1414                 struct ldlm_res_id child_res_id = { .name = {0}};
1415                 ldlm_policy_data_t sz_policy;
1416                 struct lustre_handle sz_lockh;
1417                 int lock_flags = LDLM_FL_ATOMIC_CB;
1418
1419                 child_res_id.name[0] = id_fid(&body->id1);
1420                 child_res_id.name[1] = id_group(&body->id1);
1421                 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1422
1423                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1424                                       child_res_id, LDLM_IBITS, &sz_policy,
1425                                       LCK_PW, &lock_flags, mds_blocking_ast,
1426                                       ldlm_completion_ast, NULL, NULL, NULL,
1427                                       0, NULL, &sz_lockh);
1428                 if (rc == ELDLM_OK)
1429                         ldlm_lock_decref(&sz_lockh, LCK_PW);
1430                 else
1431                         CERROR("can't invalidate client's update locks\n");
1432         }
1433
1434         EXIT;
1435 cleanup:
1436         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1437                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1438
1439 cleanup_no_trans:
1440         switch (cleanup_phase) {
1441         case 3:
1442                 if (rc) {
1443                         ldlm_lock_decref(child_lockh, child_mode);
1444                         child_lockh->cookie = 0;
1445                 }
1446         case 2:
1447                 if (rc && created) {
1448                         int err = vfs_unlink(dparent->d_inode, dchild);
1449                         if (err) {
1450                                 CERROR("unlink(%.*s) in error path: %d\n",
1451                                        dchild->d_name.len, dchild->d_name.name,
1452                                        err);
1453                         }
1454                 } else if (created) {
1455                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1456                 }
1457                 /* audit stuff for OPEN */
1458                 if (offset == 3 && (dchild->d_inode || dparent)) {
1459                         struct lustre_id au_id;
1460                         struct inode * au_inode = dchild->d_inode;
1461
1462                         if (au_inode == NULL) {
1463                                 au_inode = dparent->d_inode;
1464                                 au_id = *(rec->ur_id1);
1465                                 mds_audit_open(req, &au_id, au_inode,
1466                                                rec->ur_name, rec->ur_namelen,
1467                                                rc);
1468                         } else {
1469                                 if (fid == 0)
1470                                         mds_read_inode_sid(obd, au_inode, &au_id);
1471                                 else
1472                                         mds_inode2id(obd, &au_id, au_inode, fid);
1473                                 mds_audit_open(req, &au_id, au_inode,
1474                                                NULL, 0, rc);
1475                         }
1476                 }
1477
1478                 l_dput(dchild);
1479         case 1:
1480                 if (dparent == NULL)
1481                         break;
1482
1483                 l_dput(dparent);
1484 #ifdef S_PDIROPS
1485                 if (parent_lockh[1].cookie != 0)
1486                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1487 #endif
1488                 if (rc)
1489                         ldlm_lock_decref(parent_lockh, parent_mode);
1490                 else
1491                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1492         }
1493         if (mea)
1494                 OBD_FREE(mea, mea_size);
1495         if (rc == 0)
1496                 atomic_inc(&mds->mds_open_count);
1497
1498         mds_body_do_reverse_map(med, body);
1499
1500         /*
1501          * If we have not taken the "open" lock, we may not return 0 here,
1502          * because caller expects 0 to mean "lock is taken", and it needs
1503          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1504          * client. Later caller should rewrite the return value back to zero
1505          * if it to be used any further.
1506          */
1507         if ((cleanup_phase != 3) && !rc)
1508                 rc = ENOLCK;
1509
1510         RETURN(rc);
1511 }
1512
1513 /* Close a "file descriptor" and possibly unlink an orphan from the
1514  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1515  *
1516  * If we are being called from mds_disconnect() because the client has
1517  * disappeared, then req == NULL and we do not update last_rcvd because
1518  * there is nothing that could be recovered by the client at this stage
1519  * (it will not even _have_ an entry in last_rcvd anymore).
1520  *
1521  * Returns EAGAIN if the client needs to get more data and re-close. */
1522 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1523                   struct obd_device *obd, struct mds_file_data *mfd,
1524                   int unlink_orphan)
1525 {
1526         struct inode *inode = mfd->mfd_dentry->d_inode;
1527         char idname[LL_ID_NAMELEN];
1528         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1529         struct dentry *pending_child = NULL;
1530         struct mds_obd *mds = &obd->u.mds;
1531         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1532         void *handle = NULL;
1533         struct mds_body *request_body = NULL, *reply_body = NULL;
1534         struct dentry_params dp;
1535         struct iattr iattr = { 0 };
1536         struct llog_create_locks *lcl = NULL;
1537         ENTRY;
1538
1539         if (req && req->rq_reqmsg != NULL)
1540                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1541                                               sizeof(*request_body));
1542         if (req && req->rq_repmsg != NULL)
1543                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1544                                             sizeof(*reply_body));
1545
1546         if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
1547                 /* we set i_size/i_blocks here, nobody will see
1548                  * them until all write references are dropped.
1549                  * btw, we hold one reference */
1550                 LASSERT(mfd->mfd_mode & FMODE_WRITE);
1551                 i_size_write(inode, request_body->size);
1552                 inode->i_blocks = request_body->blocks;
1553                 iattr.ia_size = inode->i_size;
1554                 iattr.ia_valid |= ATTR_SIZE;
1555                 mds_inode_unset_attrs_old(inode);
1556         }
1557
1558         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1559         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1560                idname, inode->i_nlink, mds_orphan_open_count(inode));
1561
1562         last_orphan = mds_orphan_open_dec_test(inode) &&
1563                 mds_inode_is_orphan(inode);
1564         UP_WRITE_I_ALLOC_SEM(inode);
1565
1566         /* this is half of the actual "close" */
1567         if (mfd->mfd_mode & FMODE_WRITE) {
1568                 rc = mds_put_write_access(mds, inode, request_body,
1569                                           last_orphan && unlink_orphan);
1570         } else if (mfd->mfd_mode & FMODE_EXEC) {
1571                 mds_allow_write_access(inode);
1572         }
1573
1574         if (last_orphan && unlink_orphan) {
1575                 struct lov_mds_md *lmm = NULL;
1576                 int stripe_count = 0;
1577                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1578
1579                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1580
1581                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1582                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1583                         CERROR("found \"orphan\" %s %s with link count %d\n",
1584                                S_ISREG(inode->i_mode) ? "file" : "dir",
1585                                idname, inode->i_nlink);
1586                 /*
1587                  * sadly, there is no easy way to save pending_child from
1588                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1589                  * normally it will still be in the dcache.
1590                  */
1591                 down(&pending_dir->i_sem);
1592                 cleanup_phase = 1; /* up(i_sem) when finished */
1593                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1594                                                idlen);
1595                 if (IS_ERR(pending_child))
1596                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1597                 if (pending_child->d_inode == NULL) {
1598                         CERROR("orphan %s has been removed\n", idname);
1599                         GOTO(cleanup, rc = 0);
1600                 }
1601
1602                 cleanup_phase = 2; /* dput(pending_child) when finished */
1603                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1604                         rc = vfs_rmdir(pending_dir, pending_child);
1605                         if (rc)
1606                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1607                                        idname, rc);
1608                         goto out;
1609                 }
1610
1611                 if (req != NULL && req->rq_repmsg != NULL) {
1612                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1613                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1614                 }
1615
1616                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1617                                           NULL, stripe_count);
1618                 if (IS_ERR(handle)) {
1619                         rc = PTR_ERR(handle);
1620                         handle = NULL;
1621                         GOTO(cleanup, rc);
1622                 }
1623
1624                 pending_child->d_fsdata = (void *) &dp;
1625                 dp.p_inum = 0;
1626                 dp.p_ptr = req;
1627                 rc = vfs_unlink(pending_dir, pending_child);
1628                 if (rc)
1629                         CERROR("error unlinking orphan %s: rc %d\n",
1630                                idname, rc);
1631
1632                 if (req != NULL && req->rq_repmsg != NULL &&
1633                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1634                     mds_log_op_unlink(obd, pending_child->d_inode,
1635                                       lmm, req->rq_repmsg->buflens[1],
1636                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1637                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1638                         reply_body->valid |= OBD_MD_FLCOOKIE;
1639                 }
1640
1641                 rc = mds_destroy_object(obd, inode, 1);
1642                 if (rc) {
1643                         CERROR("cannot destroy OSS object on close, err %d\n",
1644                                rc);
1645                         rc = 0;
1646                 }
1647
1648                 goto out; /* Don't bother updating attrs on unlinked inode */
1649         } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
1650                 /* last writer closed file - let's update i_size/i_blocks */
1651                 mds_validate_size(obd, inode, request_body, &iattr);
1652         }
1653
1654 #if 0
1655         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1656                 /* Update the on-disk attributes if this was the last write
1657                  * close, and all information was provided (i.e., rc == 0)
1658                  *
1659                  * XXX this should probably be abstracted with mds_reint_setattr
1660                  */
1661
1662                 if (request_body->valid & OBD_MD_FLMTIME &&
1663                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1664                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1665                         iattr.ia_valid |= ATTR_MTIME;
1666                 }
1667                 if (request_body->valid & OBD_MD_FLCTIME &&
1668                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1669                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1670                         iattr.ia_valid |= ATTR_CTIME;
1671                 }
1672
1673                 /* XXX can't set block count with fsfilt_setattr (!) */
1674                 if (request_body->valid & OBD_MD_FLSIZE) {
1675                         iattr.ia_valid |= ATTR_SIZE;
1676                         iattr.ia_size = request_body->size;
1677                 }
1678                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1679                         iattr.ia_valid |= ATTR_BLOCKS;
1680                         iattr.ia_blocks = request_body->blocks
1681                 } */
1682
1683         }
1684 #endif
1685         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1686                 /* Only start a transaction to write out only the atime if
1687                  * it is more out-of-date than the specified limit.  If we
1688                  * are already going to write out the atime then do it anyway.
1689                  * */
1690                 if ((request_body->atime >
1691                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1692                     (iattr.ia_valid != 0 &&
1693                      request_body->atime > LTIME_S(inode->i_atime))) {
1694                         LTIME_S(iattr.ia_atime) = request_body->atime;
1695                         iattr.ia_valid |= ATTR_ATIME;
1696                 }
1697         }
1698
1699         if (iattr.ia_valid != 0) {
1700                 if (handle == NULL)
1701                         handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1702                 if (IS_ERR(handle))
1703                         GOTO(cleanup, rc = PTR_ERR(handle));
1704                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1705                 if (rc)
1706                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1707         }
1708 out:
1709         /* If other clients have this file open for write, rc will be > 0 */
1710         if (rc > 0)
1711                 rc = 0;
1712         if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
1713                         && !mds_inode_is_orphan(inode)
1714                         && atomic_read(&inode->i_writecount) == 0
1715                         && inode->i_nlink != 0) {
1716                 CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
1717                        inode->i_ino, inode->i_generation, inode->i_nlink);
1718         }
1719         l_dput(mfd->mfd_dentry);
1720         mds_mfd_destroy(mfd);
1721
1722  cleanup:
1723         atomic_dec(&mds->mds_open_count);
1724         if (req != NULL && reply_body != NULL) {
1725                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1726         } else if (handle) {
1727                 int err, force_sync = 0;
1728
1729                 if (req && req->rq_export)
1730                         force_sync = req->rq_export->exp_sync;
1731
1732                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1733                                     force_sync);
1734                 if (err) {
1735                         CERROR("error committing close: %d\n", err);
1736                         if (!rc)
1737                                 rc = err;
1738                 }
1739         }
1740
1741         switch (cleanup_phase) {
1742         case 2:
1743                 if (lcl != NULL)
1744                         ptlrpc_save_llog_lock(req, lcl);
1745                 dput(pending_child);
1746         case 1:
1747                 up(&pending_dir->i_sem);
1748         }
1749         RETURN(rc);
1750 }
1751
1752 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1753                                     struct ldlm_lock_desc *new, void *data,
1754                                     int flag)
1755 {
1756         struct lustre_handle lockh = { 0 };
1757         int rc;
1758         ENTRY;
1759
1760         switch (flag) {
1761         case LDLM_CB_BLOCKING:
1762                 ldlm_lock2handle(lock, &lockh);
1763                 rc = ldlm_cli_cancel(&lockh);
1764                 if (rc != ELDLM_OK)
1765                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
1766                 break;
1767         case LDLM_CB_CANCELING: {
1768                 break;
1769         }
1770         default:
1771                 LBUG();
1772         }
1773
1774         RETURN(0);
1775 }
1776 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1777 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1778
1779 int mds_validate_size(struct obd_device *obd, struct inode *inode,
1780                       struct mds_body *body, struct iattr *iattr)
1781 {
1782         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1783         struct lustre_handle lockh = { 0 };
1784         struct lov_stripe_md *lsm = NULL;
1785         int rc, len, flags;
1786         void *lmm = NULL;
1787         ENTRY;
1788
1789         /* we update i_size/i_blocks only for regular files */
1790         if (!S_ISREG(inode->i_mode))
1791                 RETURN(0);
1792
1793         /* we shouldn't fetch size from OSTes during recovery - deadlock */
1794         if (obd->obd_recovering) {
1795                 CERROR("size-on-mds has no support on OST yet\n");
1796                 RETURN(0);
1797         }
1798
1799         /* if nobody modified attrs. we're lucky */
1800         if (!mds_inode_has_old_attrs(inode))
1801                 RETURN(0);
1802
1803         /* 1: client didn't send actual i_size/i_blocks
1804          * 2: we seem to be last writer
1805          * 3: the file doesn't look like to be disappeared
1806          * conclusion: we're gonna fetch them from OSSes */
1807
1808         down(&inode->i_sem);
1809         len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1810         up(&inode->i_sem);
1811
1812         if (len < 0) {
1813                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1814                 GOTO(cleanup, rc = len);
1815         } else if (len == 0) {
1816                 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1817                 GOTO(cleanup, rc = 0);
1818         }
1819
1820         OBD_ALLOC(lmm, len);
1821         if (lmm == NULL) {
1822                 CERROR("can't allocate memory\n");
1823                 GOTO(cleanup, rc = -ENOMEM);
1824         }
1825
1826         down(&inode->i_sem);
1827         rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1828         up(&inode->i_sem);
1829
1830         if (rc < 0) {
1831                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1832                 GOTO(cleanup, rc);
1833         }
1834
1835         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1836         if (rc < 0) {
1837                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1838                 GOTO(cleanup, rc);
1839         }
1840
1841         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1842
1843         flags = LDLM_FL_HAS_INTENT;
1844         rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1845                          LCK_PR, &flags, mds_extent_lock_callback,
1846                          ldlm_completion_ast, NULL, NULL,
1847                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1848         if (rc == -ENOENT) {
1849                 /* while we were enqueueing lock on OST, another thread
1850                  * unlinked the file and started OST object destoying.
1851                  * it's safe to return 0 here */
1852                 GOTO(cleanup, rc = 0);
1853         } else if (rc != 0) {
1854                 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1855                 GOTO(cleanup, rc);
1856         }
1857
1858         CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n",
1859                inode->i_size, inode->i_blocks, OLID4(&body->id1),
1860                atomic_read(&inode->i_writecount) > 1 ? "U" : "",
1861                mds_inode_has_old_attrs(inode) ? "D" : "",
1862                mds_inode_is_orphan(inode) ? "O" : "");
1863
1864         i_size_write(inode, lov_merge_size(lsm, 0));
1865         inode->i_blocks = lov_merge_blocks(lsm);
1866         iattr->ia_size = inode->i_size;
1867         iattr->ia_valid |= ATTR_SIZE;
1868         DOWN_WRITE_I_ALLOC_SEM(inode);
1869         mds_inode_unset_attrs_old(inode);
1870         UP_WRITE_I_ALLOC_SEM(inode);
1871
1872         obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1873
1874 cleanup:
1875         if (lsm != NULL)
1876                 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1877         if (lmm != NULL)
1878                 OBD_FREE(lmm, len);
1879         RETURN(rc);
1880 }
1881
1882 int mds_close(struct ptlrpc_request *req, int offset)
1883 {
1884         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1885         struct obd_device *obd = req->rq_export->exp_obd;
1886         struct mds_body *body;
1887         struct mds_file_data *mfd;
1888         struct lvfs_run_ctxt saved;
1889         struct inode *inode;
1890         int rc, repsize[3] = {sizeof(struct mds_body),
1891                               obd->u.mds.mds_max_mdsize,
1892                               obd->u.mds.mds_max_cookiesize};
1893         ENTRY;
1894         MD_COUNTER_INCREMENT(obd, close);
1895
1896         rc = lustre_pack_reply(req, 3, repsize, NULL);
1897         if (rc) {
1898                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1899                 req->rq_status = rc;
1900         } else {
1901                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1902         }
1903
1904         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1905                 DEBUG_REQ(D_HA, req, "close replay");
1906                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1907                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1908                        req->rq_repmsg->buflens[2]);
1909         }
1910
1911         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1912                                   lustre_swab_mds_body);
1913         if (body == NULL) {
1914                 CERROR("Can't unpack body\n");
1915                 req->rq_status = -EFAULT;
1916                 RETURN(-EFAULT);
1917         }
1918
1919         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1920                 /* do some stuff */ ;
1921
1922         mfd = mds_handle2mfd(&body->handle);
1923         if (mfd == NULL) {
1924                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1925                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1926                 req->rq_status = -ESTALE;
1927                 RETURN(-ESTALE);
1928         }
1929
1930         inode = mfd->mfd_dentry->d_inode;
1931
1932         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1933         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1934
1935         if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
1936                 /* the client modified data through the handle
1937                  * we need to care about attrs. -bzzz */
1938                 mds_inode_set_attrs_old(inode);
1939         }
1940
1941         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1942                 struct mds_body *rep_body;
1943
1944                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1945                                           sizeof (*rep_body));
1946                 LASSERT(rep_body != NULL);
1947
1948                 mds_pack_inode2body(obd, rep_body, inode,
1949                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1950
1951                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1952                             inode, MDS_PACK_MD_LOCK, 0);
1953         }
1954         spin_lock(&med->med_open_lock);
1955         list_del(&mfd->mfd_list);
1956         spin_unlock(&med->med_open_lock);
1957
1958         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1959         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1960         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1961
1962         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1963                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1964                 req->rq_status = -ENOMEM;
1965                 mds_mfd_put(mfd);
1966                 RETURN(-ENOMEM);
1967         }
1968
1969         mds_mfd_put(mfd);
1970         RETURN(0);
1971 }
1972
1973
1974 int mds_done_writing(struct ptlrpc_request *req, int offset)
1975 {
1976         struct mds_body *body;
1977         int rc, size = sizeof(struct mds_body);
1978         ENTRY;
1979
1980         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1981
1982         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1983                                   lustre_swab_mds_body);
1984         if (body == NULL) {
1985                 CERROR("Can't unpack body\n");
1986                 req->rq_status = -EFAULT;
1987                 RETURN(-EFAULT);
1988         }
1989
1990         rc = lustre_pack_reply(req, 1, &size, NULL);
1991         if (rc) {
1992                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1993                 req->rq_status = rc;
1994         }
1995
1996         RETURN(0);
1997 }