lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mdt/mdt_open.c
  37  *
  38  * Lustre Metadata Target (mdt) open/close file handling
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lustre_acl.h>
  46 #include <lustre_mds.h>
  47 #include "mdt_internal.h"
  48
  49 /* we do nothing because we do not have refcount now */
  50 static void mdt_mfd_get(void *mfdp)
  51 {
  52 }
  53
  54 static struct portals_handle_ops mfd_handle_ops = {
  55         .hop_addref = mdt_mfd_get,
  56         .hop_free   = NULL,
  57 };
  58
  59 /* Create a new mdt_file_data struct, initialize it,
  60  * and insert it to global hash table */
  61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  62 {
  63         struct mdt_file_data *mfd;
  64         ENTRY;
  65
  66         OBD_ALLOC_PTR(mfd);
  67         if (mfd != NULL) {
  68                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  69                 mfd->mfd_handle.h_owner = med;
  70                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  71                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  72         }
  73
  74         RETURN(mfd);
  75 }
  76
  77 /*
  78  * Find the mfd pointed to by handle in global hash table.
  79  * In case of replay the handle is obsoleted
  80  * but mfd can be found in mfd list by that handle.
  81  * Callers need to be holding med_open_lock.
  82  */
  83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  84                                      const struct lustre_handle *handle,
  85                                      bool is_replay_or_resent)
  86 {
  87         struct mdt_file_data   *mfd;
  88         ENTRY;
  89
  90         LASSERT(handle != NULL);
  91         mfd = class_handle2object(handle->cookie, med);
  92         /* during dw/setattr replay the mfd can be found by old handle */
  93         if (mfd == NULL && is_replay_or_resent) {
  94                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  95                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  96                                 RETURN(mfd);
  97                 }
  98                 mfd = NULL;
  99         }
 100
 101         RETURN(mfd);
 102 }
 103
 104 /* free mfd */
 105 void mdt_mfd_free(struct mdt_file_data *mfd)
 106 {
 107         LASSERT(cfs_list_empty(&mfd->mfd_list));
 108         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 109 }
 110
 111 static int mdt_create_data(struct mdt_thread_info *info,
 112                            struct mdt_object *p, struct mdt_object *o)
 113 {
 114         struct md_op_spec     *spec = &info->mti_spec;
 115         struct md_attr        *ma   = &info->mti_attr;
 116         int                    rc   = 0;
 117         ENTRY;
 118
 119         if (!md_should_create(spec->sp_cr_flags))
 120                 RETURN(0);
 121
 122         ma->ma_need = MA_INODE | MA_LOV;
 123         ma->ma_valid = 0;
 124         mutex_lock(&o->mot_lov_mutex);
 125         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 126                 if (p != NULL && (fid_is_obf(mdt_object_fid(p)) ||
 127                                   fid_is_dot_lustre(mdt_object_fid(p))))
 128                         GOTO(unlock, rc = -EPERM);
 129
 130                 rc = mdo_create_data(info->mti_env,
 131                                      p ? mdt_object_child(p) : NULL,
 132                                      mdt_object_child(o), spec, ma);
 133                 if (rc == 0)
 134                         rc = mdt_attr_get_complex(info, o, ma);
 135
 136                 if (rc == 0 && ma->ma_valid & MA_LOV)
 137                         o->mot_flags |= MOF_LOV_CREATED;
 138         }
 139 unlock:
 140         mutex_unlock(&o->mot_lov_mutex);
 141         RETURN(rc);
 142 }
 143
 144 static int mdt_ioepoch_opened(struct mdt_object *mo)
 145 {
 146         return mo->mot_ioepoch_count;
 147 }
 148
 149 int mdt_object_is_som_enabled(struct mdt_object *mo)
 150 {
 151         return !mo->mot_ioepoch;
 152 }
 153
 154 /**
 155  * Re-enable Size-on-MDS.
 156  * Call under ->mot_ioepoch_mutex.
 157  */
 158 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 159 {
 160         if (ioepoch == mo->mot_ioepoch) {
 161                 LASSERT(!mdt_ioepoch_opened(mo));
 162                 mo->mot_ioepoch = 0;
 163                 mo->mot_flags = 0;
 164         }
 165 }
 166
 167 /**
 168  * Open the IOEpoch. It is allowed if @writecount is not negative.
 169  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 170  */
 171 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 172                      int created)
 173 {
 174         struct mdt_device *mdt = info->mti_mdt;
 175         int cancel = 0;
 176         int rc = 0;
 177         ENTRY;
 178
 179         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 180             !S_ISREG(lu_object_attr(&o->mot_obj)))
 181                 RETURN(0);
 182
 183         mutex_lock(&o->mot_ioepoch_mutex);
 184         if (mdt_ioepoch_opened(o)) {
 185                 /* Epoch continues even if there is no writers yet. */
 186                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 187                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 188         } else {
 189                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 190                 spin_lock(&mdt->mdt_ioepoch_lock);
 191                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 192                         mdt->mdt_ioepoch = info->mti_replayepoch;
 193
 194                 if (info->mti_replayepoch)
 195                         o->mot_ioepoch = info->mti_replayepoch;
 196                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 197                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 198                 else
 199                         o->mot_ioepoch = mdt->mdt_ioepoch;
 200
 201                 spin_unlock(&mdt->mdt_ioepoch_lock);
 202
 203                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 204                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 205                 if (created)
 206                         o->mot_flags |= MOF_SOM_CREATED;
 207                 cancel = 1;
 208         }
 209         o->mot_ioepoch_count++;
 210         mutex_unlock(&o->mot_ioepoch_mutex);
 211
 212         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 213          * In the truncate case, see mdt_reint_setattr(). */
 214         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 215                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 216                 mdt_lock_reg_init(lh, LCK_EX);
 217                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 218                                      MDT_LOCAL_LOCK);
 219                 if (rc == 0)
 220                         mdt_object_unlock(info, o, lh, 1);
 221         }
 222         RETURN(rc);
 223 }
 224
 225 /**
 226  * Update SOM on-disk attributes.
 227  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 228  * mountid and attributes. If disabling, clean SOM xattr.
 229  * Call under ->mot_ioepoch_mutex.
 230  */
 231 static int mdt_som_attr_set(struct mdt_thread_info *info,
 232                             struct mdt_object *obj, __u64 ioepoch, bool enable)
 233 {
 234         struct md_object        *next = mdt_object_child(obj);
 235         int                      rc;
 236         ENTRY;
 237
 238         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 239                " on "DFID".\n", enable ? "update" : "disabling",
 240                ioepoch, PFID(mdt_object_fid(obj)));
 241
 242         if (enable) {
 243                 struct lu_buf           *buf = &info->mti_buf;
 244                 struct som_attrs        *attrs;
 245                 struct md_attr          *ma = &info->mti_attr;
 246                 struct lu_attr          *la = &ma->ma_attr;
 247                 struct obd_device       *obd = info->mti_mdt->mdt_lut.lut_obd;
 248
 249                 attrs = (struct som_attrs *)info->mti_xattr_buf;
 250                 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
 251
 252                 /* pack SOM attributes */
 253                 memset(attrs, 0, sizeof(*attrs));
 254                 attrs->som_ioepoch = ioepoch;
 255                 attrs->som_mountid = obd->u.obt.obt_mount_count;
 256                 if ((la->la_valid & LA_SIZE) != 0)
 257                         attrs->som_size = la->la_size;
 258                 if ((la->la_valid & LA_BLOCKS) != 0)
 259                         attrs->som_blocks = la->la_blocks;
 260                 lustre_som_swab(attrs);
 261
 262                 /* update SOM attributes */
 263                 buf->lb_buf = attrs;
 264                 buf->lb_len = sizeof(*attrs);
 265                 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
 266         } else {
 267                 /* delete SOM attributes */
 268                 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
 269         }
 270
 271         RETURN(rc);
 272 }
 273
 274 /** Perform the eviction specific actions on ioepoch close. */
 275 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 276                                                 struct mdt_object *o)
 277 {
 278         int rc = 0;
 279
 280         mutex_lock(&o->mot_ioepoch_mutex);
 281         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 282                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 283                o->mot_ioepoch_count);
 284         o->mot_ioepoch_count--;
 285
 286         /* If eviction occured set MOF_SOM_RECOV,
 287          * if no other epoch holders, disable SOM on disk. */
 288         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 289         if (!mdt_ioepoch_opened(o)) {
 290                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 291                 mdt_object_som_enable(o, o->mot_ioepoch);
 292         }
 293         mutex_unlock(&o->mot_ioepoch_mutex);
 294         RETURN(rc);
 295 }
 296
 297 /**
 298  * Perform the replay specific actions on ioepoch close.
 299  * Skip SOM attribute update if obtained and just forget about the inode state
 300  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 301  */
 302 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 303                                               struct mdt_object *o)
 304 {
 305         int rc = MDT_IOEPOCH_CLOSED;
 306         ENTRY;
 307
 308         mutex_lock(&o->mot_ioepoch_mutex);
 309         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 310                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 311         o->mot_ioepoch_count--;
 312
 313         /* Get an info from the replayed request if client is supposed
 314          * to send an Attibute Update, reconstruct @rc if so */
 315         if (info->mti_ioepoch->flags & MF_SOM_AU)
 316                 rc = MDT_IOEPOCH_GETATTR;
 317
 318         if (!mdt_ioepoch_opened(o))
 319                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 320         mutex_unlock(&o->mot_ioepoch_mutex);
 321
 322         RETURN(rc);
 323 }
 324
 325 /**
 326  * Regular file IOepoch close.
 327  * Closes the ioepoch, checks the object state, apply obtained attributes and
 328  * re-enable SOM on the object, if possible. Also checks if the recovery is
 329  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 330  * to obtain SOM attributes under the server-side OST locks.
 331  *
 332  * Return value:
 333  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 334  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 335  */
 336 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 337                                         struct mdt_object *o)
 338 {
 339         struct md_attr *tmp_ma;
 340         struct lu_attr *la;
 341         int achange, opened;
 342         int recovery = 0;
 343         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 344         ENTRY;
 345
 346         la = &info->mti_attr.ma_attr;
 347         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 348
 349         mutex_lock(&o->mot_ioepoch_mutex);
 350         o->mot_ioepoch_count--;
 351
 352         tmp_ma = &info->mti_u.som.attr;
 353         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 354         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 355         tmp_ma->ma_som = &info->mti_u.som.data;
 356         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 357         tmp_ma->ma_valid = 0;
 358         rc = mdt_attr_get_complex(info, o, tmp_ma);
 359         if (rc)
 360                 GOTO(error_up, rc);
 361
 362         /* Check the on-disk SOM state. */
 363         if (o->mot_flags & MOF_SOM_RECOV)
 364                 recovery = 1;
 365         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 366                  !(tmp_ma->ma_valid & MA_SOM))
 367                 recovery = 1;
 368
 369         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 370                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 371
 372         opened = mdt_ioepoch_opened(o);
 373         /**
 374          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 375          * Skip the check for file with no LOV  or for unlink files.
 376          */
 377         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 378             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 379                 if (recovery)
 380                         /* If some previous writer was evicted, re-ask the
 381                          * client for attributes. Even if attributes are
 382                          * provided, we cannot believe in them.
 383                          * Another use case is that there is no SOM cache on
 384                          * disk -- first access with SOM or there was an MDS
 385                          * failure. */
 386                         ret = MDT_IOEPOCH_GETATTR;
 387                 else if (o->mot_flags & MOF_SOM_CHANGE)
 388                         /* Some previous writer changed the attribute.
 389                          * Do not believe to the current Size-on-MDS
 390                          * update, re-ask client. */
 391                         ret = MDT_IOEPOCH_GETATTR;
 392                 else if (!(la->la_valid & LA_SIZE) && achange)
 393                         /* Attributes were changed by the last writer
 394                          * only but no Size-on-MDS update is received.*/
 395                         ret = MDT_IOEPOCH_GETATTR;
 396         }
 397
 398         if (achange || ret == MDT_IOEPOCH_GETATTR)
 399                 o->mot_flags |= MOF_SOM_CHANGE;
 400
 401         /* If epoch ends and relable SOM attributes are obtained, update them.
 402          * Create SOM ea for new files even if there is no attributes obtained
 403          * (0-length file). */
 404         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 405                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 406                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 407                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 408                                               MDT_SOM_ENABLE);
 409                         /* Avoid the following setattrs of these attributes,
 410                          * e.g. for atime update. */
 411                         info->mti_attr.ma_valid = 0;
 412                 }
 413                 mdt_object_som_enable(o, o->mot_ioepoch);
 414         }
 415
 416         mutex_unlock(&o->mot_ioepoch_mutex);
 417         /* If recovery is needed, tell the client to perform GETATTR under
 418          * the lock. */
 419         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 420                 struct mdt_body *rep;
 421                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 422                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 423         }
 424
 425         RETURN(rc ? : ret);
 426
 427 error_up:
 428         mutex_unlock(&o->mot_ioepoch_mutex);
 429         return rc;
 430 }
 431
 432 /**
 433  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 434  * - a client closes the IOEpoch;
 435  * - a client eviction occured.
 436  * Return values:
 437  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 438  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 439  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 440  * update is needed.
 441  */
 442 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 443 {
 444         struct ptlrpc_request *req = mdt_info_req(info);
 445         ENTRY;
 446
 447         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 448             !S_ISREG(lu_object_attr(&o->mot_obj)))
 449                 RETURN(0);
 450
 451         LASSERT(o->mot_ioepoch_count);
 452         LASSERT(info->mti_ioepoch == NULL ||
 453                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 454
 455         /* IOEpoch is closed only if client tells about it or eviction occures.
 456          * In the replay case, always close the epoch. */
 457         if (req == NULL)
 458                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 459         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 460                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 461         if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
 462                 RETURN(mdt_ioepoch_close_reg(info, o));
 463         /* IO epoch is not closed. */
 464         RETURN(MDT_IOEPOCH_OPENED);
 465 }
 466
 467 /**
 468  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 469  * for attribute update. It happens if:
 470  * - SOM Attribute Update is obtained;
 471  * - the client failed to obtain it and informs MDS about it;
 472  * - a client eviction occured.
 473  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 474  * cache otherwise.
 475  */
 476 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 477 {
 478         struct ptlrpc_request   *req = mdt_info_req(info);
 479         __u64                    ioepoch = 0;
 480         int                      act = MDT_SOM_ENABLE;
 481         int                      rc = 0;
 482         ENTRY;
 483
 484         LASSERT(!req || info->mti_ioepoch);
 485         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 486             !S_ISREG(lu_object_attr(&o->mot_obj)))
 487                 RETURN(0);
 488
 489         /* No size whereas MF_SOM_CHANGE is set means client failed to
 490          * obtain ost attributes, drop the SOM cache on disk if so. */
 491         if (!req ||
 492             (info->mti_ioepoch &&
 493              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 494              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 495                 act = MDT_SOM_DISABLE;
 496
 497         mutex_lock(&o->mot_ioepoch_mutex);
 498         /* Mark the object it is the recovery state if we failed to obtain
 499          * SOM attributes. */
 500         if (act == MDT_SOM_DISABLE)
 501                 o->mot_flags |= MOF_SOM_RECOV;
 502
 503         if (!mdt_ioepoch_opened(o)) {
 504                 ioepoch =  info->mti_ioepoch ?
 505                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 506
 507                 if (req != NULL
 508                     && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 509                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 510                 mdt_object_som_enable(o, ioepoch);
 511         }
 512         mutex_unlock(&o->mot_ioepoch_mutex);
 513         RETURN(rc);
 514 }
 515
 516 int mdt_write_read(struct mdt_object *o)
 517 {
 518         int rc = 0;
 519         ENTRY;
 520         mutex_lock(&o->mot_ioepoch_mutex);
 521         rc = o->mot_writecount;
 522         mutex_unlock(&o->mot_ioepoch_mutex);
 523         RETURN(rc);
 524 }
 525
 526 int mdt_write_get(struct mdt_object *o)
 527 {
 528         int rc = 0;
 529         ENTRY;
 530         mutex_lock(&o->mot_ioepoch_mutex);
 531         if (o->mot_writecount < 0)
 532                 rc = -ETXTBSY;
 533         else
 534                 o->mot_writecount++;
 535         mutex_unlock(&o->mot_ioepoch_mutex);
 536         RETURN(rc);
 537 }
 538
 539 void mdt_write_put(struct mdt_object *o)
 540 {
 541         ENTRY;
 542         mutex_lock(&o->mot_ioepoch_mutex);
 543         o->mot_writecount--;
 544         mutex_unlock(&o->mot_ioepoch_mutex);
 545         EXIT;
 546 }
 547
 548 static int mdt_write_deny(struct mdt_object *o)
 549 {
 550         int rc = 0;
 551         ENTRY;
 552         mutex_lock(&o->mot_ioepoch_mutex);
 553         if (o->mot_writecount > 0)
 554                 rc = -ETXTBSY;
 555         else
 556                 o->mot_writecount--;
 557         mutex_unlock(&o->mot_ioepoch_mutex);
 558         RETURN(rc);
 559 }
 560
 561 static void mdt_write_allow(struct mdt_object *o)
 562 {
 563         ENTRY;
 564         mutex_lock(&o->mot_ioepoch_mutex);
 565         o->mot_writecount++;
 566         mutex_unlock(&o->mot_ioepoch_mutex);
 567         EXIT;
 568 }
 569
 570 /* there can be no real transaction so prepare the fake one */
 571 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 572 {
 573         struct mdt_device      *mdt = info->mti_mdt;
 574         struct ptlrpc_request  *req = mdt_info_req(info);
 575         struct tg_export_data  *ted;
 576         struct lsd_client_data *lcd;
 577
 578         ENTRY;
 579         /* transaction has occurred already */
 580         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 581                 RETURN_EXIT;
 582
 583         spin_lock(&mdt->mdt_lut.lut_translock);
 584         if (rc != 0) {
 585                 if (info->mti_transno != 0) {
 586                         struct obd_export *exp = req->rq_export;
 587
 588                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
 589                                mdt_obd_name(mdt), info->mti_transno,
 590                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
 591                                rc);
 592                         spin_unlock(&mdt->mdt_lut.lut_translock);
 593                         RETURN_EXIT;
 594                 }
 595         } else if (info->mti_transno == 0) {
 596                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 597         } else {
 598                 /* should be replay */
 599                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 600                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 601         }
 602         spin_unlock(&mdt->mdt_lut.lut_translock);
 603
 604         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 605                info->mti_transno,
 606                req->rq_export->exp_obd->obd_last_committed);
 607
 608         req->rq_transno = info->mti_transno;
 609         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 610
 611         /* update lcd in memory only for resent cases */
 612         ted = &req->rq_export->exp_target_data;
 613         LASSERT(ted);
 614         mutex_lock(&ted->ted_lcd_lock);
 615         lcd = ted->ted_lcd;
 616         if (info->mti_transno < lcd->lcd_last_transno &&
 617             info->mti_transno != 0) {
 618                 /* This should happen during replay. Do not update
 619                  * last rcvd info if replay req transno < last transno,
 620                  * otherwise the following resend(after replay) can not
 621                  * be checked correctly by xid */
 622                 mutex_unlock(&ted->ted_lcd_lock);
 623                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
 624                        mdt_obd_name(mdt), info->mti_transno,
 625                        lcd->lcd_last_transno);
 626                 RETURN_EXIT;
 627         }
 628
 629         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 630             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 631                 if (info->mti_transno != 0)
 632                         lcd->lcd_last_close_transno = info->mti_transno;
 633                 lcd->lcd_last_close_xid = req->rq_xid;
 634                 lcd->lcd_last_close_result = rc;
 635         } else {
 636                 /* VBR: save versions in last_rcvd for reconstruct. */
 637                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 638                 if (pre_versions) {
 639                         lcd->lcd_pre_versions[0] = pre_versions[0];
 640                         lcd->lcd_pre_versions[1] = pre_versions[1];
 641                         lcd->lcd_pre_versions[2] = pre_versions[2];
 642                         lcd->lcd_pre_versions[3] = pre_versions[3];
 643                 }
 644                 if (info->mti_transno != 0)
 645                         lcd->lcd_last_transno = info->mti_transno;
 646
 647                 lcd->lcd_last_xid = req->rq_xid;
 648                 lcd->lcd_last_result = rc;
 649                 lcd->lcd_last_data = info->mti_opdata;
 650         }
 651         mutex_unlock(&ted->ted_lcd_lock);
 652
 653         EXIT;
 654 }
 655
 656 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
 657 {
 658         LASSERT(mfd != NULL);
 659
 660         CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
 661                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
 662
 663         mfd->mfd_mode = mode;
 664 }
 665
 666 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 667                         struct mdt_object *o, __u64 flags, int created)
 668 {
 669         struct ptlrpc_request   *req = mdt_info_req(info);
 670         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 671         struct mdt_file_data    *mfd;
 672         struct md_attr          *ma  = &info->mti_attr;
 673         struct lu_attr          *la  = &ma->ma_attr;
 674         struct mdt_body         *repbody;
 675         int                      rc = 0, isdir, isreg;
 676         ENTRY;
 677
 678         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 679
 680         isreg = S_ISREG(la->la_mode);
 681         isdir = S_ISDIR(la->la_mode);
 682         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 683                 /*
 684                  * No EA, check whether it is will set regEA and dirEA since in
 685                  * above attr get, these size might be zero, so reset it, to
 686                  * retrieve the MD after create obj.
 687                  */
 688                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 689                                                        &RMF_MDT_MD,
 690                                                        RCL_SERVER);
 691                 /* in replay case, p == NULL */
 692                 rc = mdt_create_data(info, p, o);
 693                 if (rc)
 694                         RETURN(rc);
 695         }
 696
 697         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 698                ma->ma_valid, ma->ma_lmm_size);
 699
 700         if (ma->ma_valid & MA_LOV) {
 701                 LASSERT(ma->ma_lmm_size != 0);
 702                 repbody->eadatasize = ma->ma_lmm_size;
 703                 if (isdir)
 704                         repbody->valid |= OBD_MD_FLDIREA;
 705                 else
 706                         repbody->valid |= OBD_MD_FLEASIZE;
 707         }
 708
 709         if (flags & FMODE_WRITE) {
 710                 rc = mdt_write_get(o);
 711                 if (rc == 0) {
 712                         mdt_ioepoch_open(info, o, created);
 713                         repbody->ioepoch = o->mot_ioepoch;
 714                 }
 715         } else if (flags & MDS_FMODE_EXEC) {
 716                 rc = mdt_write_deny(o);
 717         }
 718         if (rc)
 719                 RETURN(rc);
 720
 721         rc = mo_open(info->mti_env, mdt_object_child(o),
 722                      created ? flags | MDS_OPEN_CREATED : flags);
 723         if (rc)
 724                 GOTO(err_out, rc);
 725
 726         mfd = mdt_mfd_new(med);
 727         if (mfd == NULL)
 728                 GOTO(err_out, rc = -ENOMEM);
 729
 730         /*
 731          * Keep a reference on this object for this open, and is
 732          * released by mdt_mfd_close().
 733          */
 734         mdt_object_get(info->mti_env, o);
 735         mfd->mfd_object = o;
 736         mfd->mfd_xid = req->rq_xid;
 737
 738         /*
 739          * @flags is always not zero. At least it should be FMODE_READ,
 740          * FMODE_WRITE or MDS_FMODE_EXEC.
 741          */
 742         LASSERT(flags != 0);
 743
 744         /* Open handling. */
 745         mdt_mfd_set_mode(mfd, flags);
 746
 747         atomic_inc(&o->mot_open_count);
 748         if (flags & MDS_OPEN_LEASE)
 749                 atomic_inc(&o->mot_lease_count);
 750
 751         /* replay handle */
 752         if (req_is_replay(req)) {
 753                 struct mdt_file_data *old_mfd;
 754                 /* Check wheather old cookie already exist in
 755                  * the list, becasue when do recovery, client
 756                  * might be disconnected from server, and
 757                  * restart replay, so there maybe some orphan
 758                  * mfd here, we should remove them */
 759                 LASSERT(info->mti_rr.rr_handle != NULL);
 760                 spin_lock(&med->med_open_lock);
 761                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 762                 if (old_mfd != NULL) {
 763                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 764                                "cookie = "LPX64"\n", mfd,
 765                                PFID(mdt_object_fid(mfd->mfd_object)),
 766                                info->mti_rr.rr_handle->cookie);
 767                         class_handle_unhash(&old_mfd->mfd_handle);
 768                         cfs_list_del_init(&old_mfd->mfd_list);
 769                         spin_unlock(&med->med_open_lock);
 770                         /* no attr update for that close */
 771                         la->la_valid = 0;
 772                         ma->ma_valid |= MA_FLAGS;
 773                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 774                         mdt_mfd_close(info, old_mfd);
 775                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 776                         ma->ma_valid &= ~MA_FLAGS;
 777                 } else {
 778                         spin_unlock(&med->med_open_lock);
 779                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 780                                "cookie = "LPX64"\n",
 781                                PFID(mdt_object_fid(mfd->mfd_object)),
 782                                info->mti_rr.rr_handle->cookie);
 783                 }
 784
 785                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 786                        info->mti_rr.rr_handle->cookie);
 787
 788                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 789         }
 790
 791         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 792
 793         if (req->rq_export->exp_disconnected) {
 794                 spin_lock(&med->med_open_lock);
 795                 class_handle_unhash(&mfd->mfd_handle);
 796                 cfs_list_del_init(&mfd->mfd_list);
 797                 spin_unlock(&med->med_open_lock);
 798                 mdt_mfd_close(info, mfd);
 799         } else {
 800                 spin_lock(&med->med_open_lock);
 801                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 802                 spin_unlock(&med->med_open_lock);
 803         }
 804
 805         mdt_empty_transno(info, rc);
 806
 807         RETURN(rc);
 808
 809 err_out:
 810         if (flags & FMODE_WRITE)
 811                         /* XXX We also need to close io epoch here.
 812                          * See LU-1220 - green */
 813                 mdt_write_put(o);
 814         else if (flags & FMODE_EXEC)
 815                 mdt_write_allow(o);
 816         return rc;
 817 }
 818
 819 int mdt_finish_open(struct mdt_thread_info *info,
 820                     struct mdt_object *p, struct mdt_object *o,
 821                     __u64 flags, int created, struct ldlm_reply *rep)
 822 {
 823         struct ptlrpc_request   *req = mdt_info_req(info);
 824         struct obd_export       *exp = req->rq_export;
 825         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 826         struct md_attr          *ma  = &info->mti_attr;
 827         struct lu_attr          *la  = &ma->ma_attr;
 828         struct mdt_file_data    *mfd;
 829         struct mdt_body         *repbody;
 830         int                      rc = 0;
 831         int                      isreg, isdir, islnk;
 832         cfs_list_t              *t;
 833         ENTRY;
 834
 835         LASSERT(ma->ma_valid & MA_INODE);
 836
 837         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 838
 839         isreg = S_ISREG(la->la_mode);
 840         isdir = S_ISDIR(la->la_mode);
 841         islnk = S_ISLNK(la->la_mode);
 842         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 843
 844         /* LU-2275, simulate broken behaviour (esp. prevalent in
 845          * pre-2.4 servers where a very strange reply is sent on error
 846          * that looks like it was actually almost succesful and a failure at the
 847          * same time */
 848         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 849                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 850                                                DISP_LOOKUP_NEG |
 851                                                DISP_LOOKUP_POS);
 852
 853                 if (flags & MDS_OPEN_LOCK)
 854                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 855
 856                 RETURN(-ENOENT);
 857         }
 858
 859         if (exp_connect_rmtclient(exp)) {
 860                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 861
 862                 rc = mdt_pack_remote_perm(info, o, buf);
 863                 if (rc) {
 864                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 865                         repbody->aclsize = 0;
 866                 } else {
 867                         repbody->valid |= OBD_MD_FLRMTPERM;
 868                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 869                 }
 870         }
 871 #ifdef CONFIG_FS_POSIX_ACL
 872         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 873                 const struct lu_env *env = info->mti_env;
 874                 struct md_object *next = mdt_object_child(o);
 875                 struct lu_buf *buf = &info->mti_buf;
 876
 877                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 878                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 879                                                    RCL_SERVER);
 880                 if (buf->lb_len > 0) {
 881                         rc = mo_xattr_get(env, next, buf,
 882                                           XATTR_NAME_ACL_ACCESS);
 883                         if (rc < 0) {
 884                                 if (rc == -ENODATA) {
 885                                         repbody->aclsize = 0;
 886                                         repbody->valid |= OBD_MD_FLACL;
 887                                         rc = 0;
 888                                 } else if (rc == -EOPNOTSUPP) {
 889                                         rc = 0;
 890                                 } else {
 891                                         CERROR("got acl size: %d\n", rc);
 892                                 }
 893                         } else {
 894                                 repbody->aclsize = rc;
 895                                 repbody->valid |= OBD_MD_FLACL;
 896                                 rc = 0;
 897                         }
 898                 }
 899         }
 900 #endif
 901
 902         if (info->mti_mdt->mdt_lut.lut_mds_capa &&
 903             exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
 904                 struct lustre_capa *capa;
 905
 906                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 907                 LASSERT(capa);
 908                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 909                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 910                 if (rc)
 911                         RETURN(rc);
 912                 repbody->valid |= OBD_MD_FLMDSCAPA;
 913         }
 914         if (info->mti_mdt->mdt_lut.lut_oss_capa &&
 915             exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
 916             S_ISREG(lu_object_attr(&o->mot_obj))) {
 917                 struct lustre_capa *capa;
 918
 919                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 920                 LASSERT(capa);
 921                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 922                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 923                 if (rc)
 924                         RETURN(rc);
 925                 repbody->valid |= OBD_MD_FLOSSCAPA;
 926         }
 927
 928         /*
 929          * If we are following a symlink, don't open; and do not return open
 930          * handle for special nodes as client required.
 931          */
 932         if (islnk || (!isreg && !isdir &&
 933             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 934                 lustre_msg_set_transno(req->rq_repmsg, 0);
 935                 RETURN(0);
 936         }
 937
 938         /*
 939          * We need to return the existing object's fid back, so it is done here,
 940          * after preparing the reply.
 941          */
 942         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 943                 RETURN(-EEXIST);
 944
 945         /* This can't be done earlier, we need to return reply body */
 946         if (isdir) {
 947                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 948                         /* We are trying to create or write an existing dir. */
 949                         RETURN(-EISDIR);
 950                 }
 951         } else if (flags & MDS_OPEN_DIRECTORY)
 952                 RETURN(-ENOTDIR);
 953
 954         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 955                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 956                 RETURN(-EAGAIN);
 957         }
 958
 959         mfd = NULL;
 960         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 961                 spin_lock(&med->med_open_lock);
 962                 cfs_list_for_each(t, &med->med_open_head) {
 963                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 964                         if (mfd->mfd_xid == req->rq_xid)
 965                                 break;
 966                         mfd = NULL;
 967                 }
 968                 spin_unlock(&med->med_open_lock);
 969
 970                 if (mfd != NULL) {
 971                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 972                         /*set repbody->ea_size for resent case*/
 973                         if (ma->ma_valid & MA_LOV) {
 974                                 LASSERT(ma->ma_lmm_size != 0);
 975                                 repbody->eadatasize = ma->ma_lmm_size;
 976                                 if (isdir)
 977                                         repbody->valid |= OBD_MD_FLDIREA;
 978                                 else
 979                                         repbody->valid |= OBD_MD_FLEASIZE;
 980                         }
 981                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 982                         RETURN(0);
 983                 }
 984         }
 985
 986         rc = mdt_mfd_open(info, p, o, flags, created);
 987         if (!rc)
 988                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 989
 990         RETURN(rc);
 991 }
 992
 993 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 994                              struct lsd_client_data *lcd);
 995
 996 void mdt_reconstruct_open(struct mdt_thread_info *info,
 997                           struct mdt_lock_handle *lhc)
 998 {
 999         const struct lu_env *env = info->mti_env;
1000         struct mdt_device       *mdt  = info->mti_mdt;
1001         struct req_capsule      *pill = info->mti_pill;
1002         struct ptlrpc_request   *req  = mdt_info_req(info);
1003         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
1004         struct lsd_client_data  *lcd  = ted->ted_lcd;
1005         struct md_attr          *ma   = &info->mti_attr;
1006         struct mdt_reint_record *rr   = &info->mti_rr;
1007         __u64                   flags = info->mti_spec.sp_cr_flags;
1008         struct ldlm_reply       *ldlm_rep;
1009         struct mdt_object       *parent;
1010         struct mdt_object       *child;
1011         struct mdt_body         *repbody;
1012         int                      rc;
1013         ENTRY;
1014
1015         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1016         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1017         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1018
1019         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1020         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1021                                                RCL_SERVER);
1022         ma->ma_need = MA_INODE | MA_HSM;
1023         if (ma->ma_lmm_size > 0)
1024                 ma->ma_need |= MA_LOV;
1025
1026         ma->ma_valid = 0;
1027
1028         mdt_req_from_lcd(req, lcd);
1029         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1030
1031         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1032                ldlm_rep->lock_policy_res1, req->rq_status);
1033
1034         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1035             req->rq_status != 0)
1036                 /* We did not create successfully, return error to client. */
1037                 GOTO(out, rc = req->rq_status);
1038
1039         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1040                 struct obd_export *exp = req->rq_export;
1041                 /*
1042                  * We failed after creation, but we do not know in which step
1043                  * we failed. So try to check the child object.
1044                  */
1045                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1046                 if (IS_ERR(parent)) {
1047                         rc = PTR_ERR(parent);
1048                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1049                                       " Evicting client %s with export %s.\n",
1050                                       PFID(rr->rr_fid1), rc,
1051                                       obd_uuid2str(&exp->exp_client_uuid),
1052                                       obd_export_nid2str(exp));
1053                         mdt_export_evict(exp);
1054                         RETURN_EXIT;
1055                 }
1056                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1057                 if (IS_ERR(child)) {
1058                         rc = PTR_ERR(child);
1059                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1060                                       " Evicting client %s with export %s.\n",
1061                                       PFID(mdt_object_fid(child)), rc,
1062                                       obd_uuid2str(&exp->exp_client_uuid),
1063                                       obd_export_nid2str(exp));
1064                         mdt_object_put(env, parent);
1065                         mdt_export_evict(exp);
1066                         RETURN_EXIT;
1067                 }
1068
1069                 if (unlikely(mdt_object_remote(child))) {
1070                         /* the child object was created on remote server */
1071                         if (!mdt_is_dne_client(exp)) {
1072                                 /* Return -EIO for old client */
1073                                 mdt_object_put(env, parent);
1074                                 mdt_object_put(env, child);
1075                                 GOTO(out, rc = -EIO);
1076                         }
1077                         repbody->fid1 = *rr->rr_fid2;
1078                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1079                         rc = 0;
1080                 } else {
1081                         if (mdt_object_exists(child)) {
1082                                 mdt_set_capainfo(info, 1, rr->rr_fid2,
1083                                                  BYPASS_CAPA);
1084                                 rc = mdt_attr_get_complex(info, child, ma);
1085                                 if (rc == 0)
1086                                         rc = mdt_finish_open(info, parent,
1087                                                              child, flags,
1088                                                              1, ldlm_rep);
1089                         } else {
1090                                 /* the child does not exist, we should do
1091                                  * regular open */
1092                                 mdt_object_put(env, parent);
1093                                 mdt_object_put(env, child);
1094                                 GOTO(regular_open, 0);
1095                         }
1096                 }
1097                 mdt_object_put(env, parent);
1098                 mdt_object_put(env, child);
1099                 GOTO(out, rc);
1100         } else {
1101 regular_open:
1102                 /* We did not try to create, so we are a pure open */
1103                 rc = mdt_reint_open(info, lhc);
1104         }
1105
1106         EXIT;
1107 out:
1108         req->rq_status = rc;
1109         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1110         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1111 }
1112
1113 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1114 {
1115         __u64                    flags = info->mti_spec.sp_cr_flags;
1116         struct mdt_reint_record *rr = &info->mti_rr;
1117         struct md_attr          *ma = &info->mti_attr;
1118         struct mdt_object       *o;
1119         int                      rc;
1120         ENTRY;
1121
1122         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1123         if (IS_ERR(o))
1124                 RETURN(rc = PTR_ERR(o));
1125
1126         if (unlikely(mdt_object_remote(o))) {
1127                 /* the child object was created on remote server */
1128                 struct mdt_body *repbody;
1129
1130                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1131                                                 DISP_LOOKUP_EXECD |
1132                                                 DISP_LOOKUP_POS));
1133                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1134                 repbody->fid1 = *rr->rr_fid2;
1135                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1136                 rc = 0;
1137         } else {
1138                 if (mdt_object_exists(o)) {
1139                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1140                                                         DISP_LOOKUP_EXECD |
1141                                                         DISP_LOOKUP_POS));
1142
1143                         rc = mdt_attr_get_complex(info, o, ma);
1144                         if (rc == 0)
1145                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
1146                                                      rep);
1147                 } else {
1148                         rc = -ENOENT;
1149                 }
1150         }
1151
1152         mdt_object_put(info->mti_env, o);
1153         RETURN(rc);
1154 }
1155
1156 /* lock object for open */
1157 static int mdt_object_open_lock(struct mdt_thread_info *info,
1158                                 struct mdt_object *obj,
1159                                 struct mdt_lock_handle *lhc,
1160                                 __u64 *ibits)
1161 {
1162         struct md_attr  *ma = &info->mti_attr;
1163         __u64            open_flags = info->mti_spec.sp_cr_flags;
1164         ldlm_mode_t      lm = LCK_CR;
1165         bool             acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1166         bool             try_layout = false;
1167         bool             create_layout = false;
1168         int              rc = 0;
1169         ENTRY;
1170
1171         *ibits = 0;
1172         mdt_lock_handle_init(lhc);
1173
1174         if (req_is_replay(mdt_info_req(info)))
1175                 RETURN(0);
1176
1177         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1178                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1179                     md_should_create(open_flags))
1180                         create_layout = true;
1181                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1182                     ma->ma_need & MA_LOV)
1183                         try_layout = true;
1184         }
1185
1186         if (acq_lease) {
1187                 /* lease open, acquire write mode of open sem */
1188                 down_write(&obj->mot_open_sem);
1189
1190                 /* Lease exists and ask for new lease */
1191                 if (atomic_read(&obj->mot_lease_count) > 0) {
1192                         /* only exclusive open is supported, so lease
1193                          * are conflicted to each other */
1194                         GOTO(out, rc = -EBUSY);
1195                 }
1196
1197                 /* Lease must be with open lock */
1198                 if (!(open_flags & MDS_OPEN_LOCK)) {
1199                         CERROR("Request lease for file:"DFID ", but open lock "
1200                                 "is missed, open_flags = "LPO64".\n",
1201                                 PFID(mdt_object_fid(obj)), open_flags);
1202                         GOTO(out, rc = -EPROTO);
1203                 }
1204
1205                 /* XXX: only exclusive open is supported. */
1206                 lm = LCK_EX;
1207                 *ibits = MDS_INODELOCK_OPEN;
1208
1209                 /* never grant LCK_EX layout lock to client */
1210                 try_layout = false;
1211         } else { /* normal open */
1212                 /* normal open holds read mode of open sem */
1213                 down_read(&obj->mot_open_sem);
1214
1215                 if (open_flags & MDS_OPEN_LOCK) {
1216                         if (open_flags & FMODE_WRITE)
1217                                 lm = LCK_CW;
1218                         else if (open_flags & MDS_FMODE_EXEC)
1219                                 lm = LCK_PR;
1220                         else
1221                                 lm = LCK_CR;
1222
1223                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1224                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1225                         if (open_flags & FMODE_WRITE)
1226                                 lm = LCK_CW;
1227                         else
1228                                 lm = LCK_CR;
1229
1230                         /* revoke lease */
1231                         *ibits = MDS_INODELOCK_OPEN;
1232                         try_layout = false;
1233
1234                         lhc = &info->mti_lh[MDT_LH_LOCAL];
1235                 }
1236                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1237                         PFID(mdt_object_fid(obj)),
1238                         atomic_read(&obj->mot_open_count), lm);
1239         }
1240
1241         mdt_lock_reg_init(lhc, lm);
1242
1243         /* one problem to return layout lock on open is that it may result
1244          * in too many layout locks cached on the client side. */
1245         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1246                 /* return lookup lock to validate inode at the client side,
1247                  * this is pretty important otherwise mdt will return layout
1248                  * lock for each open.
1249                  * However this is a double-edged sword because changing
1250                  * permission will revoke huge # of LOOKUP locks. */
1251                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1252                 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1253                                          MDT_CROSS_LOCK)) {
1254                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1255                         if (*ibits != 0)
1256                                 rc = mdt_object_lock(info, obj, lhc, *ibits,
1257                                                 MDT_CROSS_LOCK);
1258                 }
1259         } else if (*ibits != 0) {
1260                 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1261         }
1262
1263         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1264                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1265                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1266
1267         /* will change layout, revoke layout locks by enqueuing EX lock. */
1268         if (rc == 0 && create_layout) {
1269                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1270
1271                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1272                         ", open_flags = "LPO64"\n",
1273                         PFID(mdt_object_fid(obj)), open_flags);
1274
1275                 LASSERT(!try_layout);
1276                 mdt_lock_handle_init(ll);
1277                 mdt_lock_reg_init(ll, LCK_EX);
1278                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1279                                         MDT_LOCAL_LOCK);
1280
1281                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1282         }
1283
1284         /* Check if there is any other open handles after acquiring
1285          * open lock. At this point, caching open handles have been revoked
1286          * by open lock.
1287          * XXX: Now only exclusive open is supported. Need to check the
1288          * type of open for generic lease support. */
1289         if (rc == 0 && acq_lease) {
1290                 struct ptlrpc_request *req = mdt_info_req(info);
1291                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1292                 struct mdt_file_data *mfd;
1293                 bool is_replay_or_resent;
1294                 int open_count = 0;
1295
1296                 /* For lease: application can open a file and then apply lease,
1297                  * @handle contains original open handle in that case.
1298                  * In recovery, open REQ will be replayed and the lease REQ may
1299                  * be resent that means the open handle is already stale, so we
1300                  * need to fix it up here by finding new handle. */
1301                 is_replay_or_resent = req_is_replay(req) ||
1302                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1303
1304                 /* if the request is _not_ a replay request, rr_handle
1305                  * may be used to hold an openhandle which is issuing the
1306                  * lease request, so that this openhandle doesn't count. */
1307                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1308                                      is_replay_or_resent);
1309                 if (mfd != NULL)
1310                         ++open_count;
1311
1312                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1313                         PFID(mdt_object_fid(obj)),
1314                         atomic_read(&obj->mot_open_count), open_count);
1315
1316                 if (atomic_read(&obj->mot_open_count) > open_count)
1317                         GOTO(out, rc = -EBUSY);
1318         }
1319         GOTO(out, rc);
1320
1321 out:
1322         RETURN(rc);
1323 }
1324
1325 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1326                                    struct mdt_object *obj,
1327                                    struct mdt_lock_handle *lhc,
1328                                    __u64 ibits, int rc)
1329 {
1330         __u64 open_flags = info->mti_spec.sp_cr_flags;
1331         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1332         ENTRY;
1333
1334         if (req_is_replay(mdt_info_req(info)))
1335                 RETURN_EXIT;
1336
1337         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1338          * return to client side. */
1339         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1340                 mdt_object_unlock(info, obj, ll, 1);
1341
1342         ll = &info->mti_lh[MDT_LH_LAYOUT];
1343         /* Release local layout lock, layout was created */
1344         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1345                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1346                 mdt_object_unlock(info, obj, ll, 1);
1347         }
1348
1349         if (open_flags & MDS_OPEN_LEASE)
1350                 up_write(&obj->mot_open_sem);
1351         else
1352                 up_read(&obj->mot_open_sem);
1353
1354         /* Cross-ref case, the lock should be returned to the client */
1355         if (ibits == 0 || rc == -EREMOTE)
1356                 RETURN_EXIT;
1357
1358         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1359                 /* for the open request, the lock will only return to client
1360                  * if open or layout lock is granted. */
1361                 rc = 1;
1362         }
1363
1364         if (rc != 0) {
1365                 struct ldlm_reply       *ldlm_rep;
1366
1367                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1368                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1369                 mdt_object_unlock(info, obj, lhc, 1);
1370         }
1371         RETURN_EXIT;
1372 }
1373
1374 /**
1375  * Check release is permitted for the current HSM flags.
1376  */
1377 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1378 {
1379         if (!(ma->ma_valid & MA_HSM))
1380                 return false;
1381
1382         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1383                 return false;
1384
1385         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1386                 return false;
1387
1388         return true;
1389 }
1390
1391 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1392                          struct mdt_lock_handle *lhc)
1393 {
1394         const struct lu_env     *env   = info->mti_env;
1395         struct mdt_device       *mdt   = info->mti_mdt;
1396         __u64                    flags = info->mti_spec.sp_cr_flags;
1397         struct mdt_reint_record *rr    = &info->mti_rr;
1398         struct md_attr          *ma    = &info->mti_attr;
1399         struct mdt_object       *parent= NULL;
1400         struct mdt_object       *o;
1401         int                      rc;
1402         __u64                    ibits = 0;
1403         ENTRY;
1404
1405         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1406                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1407                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1408                         if (IS_ERR(parent)) {
1409                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1410                                        " for anonymous created %ld, try to"
1411                                        " use server-side parent.\n",
1412                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1413                                 parent = NULL;
1414                         }
1415                 }
1416                 if (parent == NULL)
1417                         ma->ma_need |= MA_PFID;
1418         }
1419
1420         o = mdt_object_find(env, mdt, rr->rr_fid2);
1421         if (IS_ERR(o))
1422                 RETURN(rc = PTR_ERR(o));
1423
1424         if (mdt_object_remote(o)) {
1425                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1426                        mdt_obd_name(info->mti_mdt),
1427                        PFID(rr->rr_fid2));
1428                 GOTO(out, rc = -EREMOTE);
1429         } else if (!mdt_object_exists(o)) {
1430                 mdt_set_disposition(info, rep,
1431                                     DISP_IT_EXECD |
1432                                     DISP_LOOKUP_EXECD |
1433                                     DISP_LOOKUP_NEG);
1434                 GOTO(out, rc = -ENOENT);
1435         }
1436
1437         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1438
1439         if (flags & MDS_OPEN_RELEASE)
1440                 ma->ma_need |= MA_HSM;
1441         rc = mdt_attr_get_complex(info, o, ma);
1442         if (rc)
1443                 GOTO(out, rc);
1444
1445         /* If a release request, check file flags are fine and ask for an
1446          * exclusive open access. */
1447         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1448                 GOTO(out, rc = -EPERM);
1449
1450         rc = mdt_object_open_lock(info, o, lhc, &ibits);
1451         if (rc)
1452                 GOTO(out_unlock, rc);
1453
1454         if (ma->ma_valid & MA_PFID) {
1455                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1456                 if (IS_ERR(parent)) {
1457                         CDEBUG(D_INODE, "Fail to find parent "DFID
1458                                " for anonymous created %ld, try to"
1459                                " use system default.\n",
1460                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1461                         parent = NULL;
1462                 }
1463         }
1464
1465         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1466         if (!rc) {
1467                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1468                 if (flags & MDS_OPEN_LOCK)
1469                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1470                 if (flags & MDS_OPEN_LEASE)
1471                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1472         }
1473         GOTO(out_unlock, rc);
1474
1475 out_unlock:
1476         mdt_object_open_unlock(info, o, lhc, ibits, rc);
1477 out:
1478         mdt_object_put(env, o);
1479         if (parent != NULL)
1480                 mdt_object_put(env, parent);
1481         return rc;
1482 }
1483
1484 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1485 static int mdt_cross_open(struct mdt_thread_info *info,
1486                           const struct lu_fid *parent_fid,
1487                           const struct lu_fid *fid,
1488                           struct ldlm_reply *rep, __u32 flags)
1489 {
1490         struct md_attr    *ma = &info->mti_attr;
1491         struct mdt_object *o;
1492         int                rc;
1493         ENTRY;
1494
1495         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1496         if (IS_ERR(o))
1497                 RETURN(rc = PTR_ERR(o));
1498
1499         if (mdt_object_remote(o)) {
1500                 /* Something is wrong here, the object is on another MDS! */
1501                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1502                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1503                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1504                                 &o->mot_obj,
1505                                 "Object isn't on this server! FLD error?\n");
1506                 rc = -EFAULT;
1507         } else {
1508                 if (mdt_object_exists(o)) {
1509                         /* Do permission check for cross-open. */
1510                         rc = mo_permission(info->mti_env, NULL,
1511                                            mdt_object_child(o),
1512                                            NULL, flags | MDS_OPEN_CROSS);
1513                         if (rc)
1514                                 goto out;
1515
1516                         mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1517                         rc = mdt_attr_get_complex(info, o, ma);
1518                         if (rc != 0)
1519                                 GOTO(out, rc);
1520
1521                         /* Do not create lov object if the fid is opened
1522                          * under OBF */
1523                         if (S_ISREG(ma->ma_attr.la_mode) &&
1524                             !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1525                             fid_is_obf(parent_fid))
1526                                 GOTO(out, rc = -EPERM);
1527
1528                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1529                 } else {
1530                         /*
1531                          * Something is wrong here. lookup was positive but
1532                          * there is no object!
1533                          */
1534                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1535                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1536                         rc = -EFAULT;
1537                 }
1538         }
1539 out:
1540         mdt_object_put(info->mti_env, o);
1541         RETURN(rc);
1542 }
1543
1544 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1545 {
1546         struct mdt_device       *mdt = info->mti_mdt;
1547         struct ptlrpc_request   *req = mdt_info_req(info);
1548         struct mdt_object       *parent;
1549         struct mdt_object       *child;
1550         struct mdt_lock_handle  *lh;
1551         struct ldlm_reply       *ldlm_rep;
1552         struct mdt_body         *repbody;
1553         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1554         struct md_attr          *ma = &info->mti_attr;
1555         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1556         __u64                    ibits = 0;
1557         struct mdt_reint_record *rr = &info->mti_rr;
1558         struct lu_name          *lname;
1559         int                      result, rc;
1560         int                      created = 0;
1561         __u32                    msg_flags;
1562         ENTRY;
1563
1564         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1565                                (obd_timeout + 1) / 4);
1566
1567         mdt_counter_incr(req, LPROC_MDT_OPEN);
1568         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1569
1570         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1571         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1572                                                RCL_SERVER);
1573         ma->ma_need = MA_INODE;
1574         if (ma->ma_lmm_size > 0)
1575                 ma->ma_need |= MA_LOV;
1576
1577         ma->ma_valid = 0;
1578
1579         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1580         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1581
1582         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1583                 CERROR("file join is not supported anymore.\n");
1584                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1585         }
1586         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1587
1588         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1589             info->mti_spec.u.sp_ea.eadata == NULL)
1590                 GOTO(out, result = err_serious(-EINVAL));
1591
1592         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1593                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1594                PFID(rr->rr_fid1), rr->rr_name,
1595                PFID(rr->rr_fid2), create_flags,
1596                ma->ma_attr.la_mode, msg_flags);
1597         if (info->mti_cross_ref) {
1598                 /* This is cross-ref open */
1599                 mdt_set_disposition(info, ldlm_rep,
1600                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1601                              DISP_LOOKUP_POS));
1602                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1603                                         ldlm_rep, create_flags);
1604                 GOTO(out, result);
1605         } else if (req_is_replay(req) ||
1606             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1607                 /* This is a replay request or from liblustre with ea. */
1608                 result = mdt_open_by_fid(info, ldlm_rep);
1609
1610                 if (result != -ENOENT) {
1611                         if (req->rq_export->exp_libclient &&
1612                             create_flags & MDS_OPEN_HAS_EA)
1613                                 GOTO(out, result = 0);
1614                         GOTO(out, result);
1615                 }
1616                 /* We didn't find the correct object, so we need to re-create it
1617                  * via a regular replay. */
1618                 if (!(create_flags & MDS_OPEN_CREAT)) {
1619                         DEBUG_REQ(D_ERROR, req,
1620                                   "OPEN & CREAT not in open replay/by_fid.");
1621                         GOTO(out, result = -EFAULT);
1622                 }
1623                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1624         } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) ||
1625                    (create_flags & MDS_OPEN_BY_FID)) {
1626                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1627                 /* If result is 0 then open by FID has found the file
1628                  * and there is nothing left for us to do here.  More
1629                  * generally if it is anything other than -ENOENT or
1630                  * -EREMOTE then we return that now.  If -ENOENT and
1631                  * MDS_OPEN_CREAT is set then we must create the file
1632                  * below.  If -EREMOTE then we need to return a LOOKUP
1633                  * lock to the client, which we do below.  Hence this
1634                  * odd looking condition.  See LU-2523. */
1635                 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1636                     result != -EREMOTE)
1637                         GOTO(out, result);
1638
1639                 if (unlikely(rr->rr_namelen == 0))
1640                         GOTO(out, result = -EINVAL);
1641
1642                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1643         }
1644
1645         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1646                 GOTO(out, result = err_serious(-ENOMEM));
1647
1648         mdt_set_disposition(info, ldlm_rep,
1649                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1650
1651         lh = &info->mti_lh[MDT_LH_PARENT];
1652         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1653                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1654
1655         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1656                                       MDS_INODELOCK_UPDATE);
1657         if (IS_ERR(parent))
1658                 GOTO(out, result = PTR_ERR(parent));
1659
1660         /* get and check version of parent */
1661         result = mdt_version_get_check(info, parent, 0);
1662         if (result)
1663                 GOTO(out_parent, result);
1664
1665         fid_zero(child_fid);
1666
1667         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1668         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1669                             lname, child_fid, &info->mti_spec);
1670         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1671                  "looking for "DFID"/%s, result fid="DFID"\n",
1672                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1673
1674         if (result != 0 && result != -ENOENT && result != -ESTALE)
1675                 GOTO(out_parent, result);
1676
1677         if (result == -ENOENT || result == -ESTALE) {
1678                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1679                 if (result == -ESTALE) {
1680                         /*
1681                          * -ESTALE means the parent is a dead(unlinked) dir, so
1682                          * it should return -ENOENT to in accordance with the
1683                          * original mds implementaion.
1684                          */
1685                         GOTO(out_parent, result = -ENOENT);
1686                 }
1687                 if (!(create_flags & MDS_OPEN_CREAT))
1688                         GOTO(out_parent, result);
1689                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1690                         GOTO(out_parent, result = -EROFS);
1691                 *child_fid = *info->mti_rr.rr_fid2;
1692                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1693                          PFID(child_fid));
1694                 /* In the function below, .hs_keycmp resolves to
1695                  * lu_obj_hop_keycmp() */
1696                 /* coverity[overrun-buffer-val] */
1697                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1698         } else {
1699                 /*
1700                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1701                  * return FID back in that case.
1702                  */
1703                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1704                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1705         }
1706         if (IS_ERR(child))
1707                 GOTO(out_parent, result = PTR_ERR(child));
1708
1709         /** check version of child  */
1710         rc = mdt_version_get_check(info, child, 1);
1711         if (rc)
1712                 GOTO(out_child, result = rc);
1713
1714         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1715         if (result == -ENOENT) {
1716                 /* Create under OBF and .lustre is not permitted */
1717                 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1718                         GOTO(out_child, result = -EPERM);
1719
1720                 /* save versions in reply */
1721                 mdt_version_get_save(info, parent, 0);
1722                 mdt_version_get_save(info, child, 1);
1723
1724                 /* version of child will be changed */
1725                 info->mti_mos = child;
1726
1727                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1728                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1729
1730                 /* Let lower layers know what is lock mode on directory. */
1731                 info->mti_spec.sp_cr_mode =
1732                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1733
1734                 /*
1735                  * Do not perform lookup sanity check. We know that name does
1736                  * not exist.
1737                  */
1738                 info->mti_spec.sp_cr_lookup = 0;
1739                 info->mti_spec.sp_feat = &dt_directory_features;
1740
1741                 result = mdo_create(info->mti_env,
1742                                     mdt_object_child(parent),
1743                                     lname,
1744                                     mdt_object_child(child),
1745                                     &info->mti_spec,
1746                                     &info->mti_attr);
1747                 if (result == -ERESTART) {
1748                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1749                         GOTO(out_child, result);
1750                 } else {
1751
1752                         /* XXX: we should call this once, see few lines below */
1753                         if (result == 0)
1754                                 result = mdt_attr_get_complex(info, child, ma);
1755
1756                         if (result != 0)
1757                                 GOTO(out_child, result);
1758                 }
1759                 created = 1;
1760         } else {
1761                 /*
1762                  * The object is on remote node, return its FID for remote open.
1763                  */
1764                 if (mdt_object_remote(child)) {
1765                         /*
1766                          * Check if this lock already was sent to client and
1767                          * this is resent case. For resent case do not take lock
1768                          * again, use what is already granted.
1769                          */
1770                         LASSERT(lhc != NULL);
1771
1772                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1773                                 struct ldlm_lock *lock;
1774
1775                                 LASSERT(msg_flags & MSG_RESENT);
1776
1777                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1778                                 if (!lock) {
1779                                         CERROR("Invalid lock handle "LPX64"\n",
1780                                                lhc->mlh_reg_lh.cookie);
1781                                         LBUG();
1782                                 }
1783                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1784                                                         &lock->l_resource->lr_name));
1785                                 LDLM_LOCK_PUT(lock);
1786                                 rc = 0;
1787                         } else {
1788                                 mdt_lock_handle_init(lhc);
1789                                 mdt_lock_reg_init(lhc, LCK_PR);
1790
1791                                 rc = mdt_object_lock(info, child, lhc,
1792                                                      MDS_INODELOCK_LOOKUP,
1793                                                      MDT_CROSS_LOCK);
1794                         }
1795                         repbody->fid1 = *mdt_object_fid(child);
1796                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1797                         if (rc != 0)
1798                                 result = rc;
1799                         else
1800                                 result = -EREMOTE;
1801                         GOTO(out_child, result);
1802                 } else {
1803                         if (mdt_object_exists(child)) {
1804                                 /* We have to get attr & LOV EA & HSM for this
1805                                  * object */
1806                                 ma->ma_need |= MA_HSM;
1807                                 result = mdt_attr_get_complex(info, child, ma);
1808                         } else {
1809                                 /*object non-exist!!!*/
1810                                 LBUG();
1811                         }
1812                 }
1813         }
1814
1815         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1816
1817         /* get openlock if this is not replay and if a client requested it */
1818         if (!req_is_replay(req)) {
1819                 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1820                 if (rc != 0)
1821                         GOTO(out_child_unlock, result = rc);
1822                 else if (create_flags & MDS_OPEN_LOCK)
1823                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1824         }
1825
1826         /* Try to open it now. */
1827         rc = mdt_finish_open(info, parent, child, create_flags,
1828                              created, ldlm_rep);
1829         if (rc) {
1830                 result = rc;
1831                 /* openlock will be released if mdt_finish_open failed */
1832                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1833
1834                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1835                         CERROR("%s: cannot open volatile file "DFID", orphan "
1836                                "file will be left in PENDING directory until "
1837                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1838                                PFID(mdt_object_fid(child)), rc);
1839                         GOTO(out_child_unlock, result);
1840                 }
1841
1842                 if (created) {
1843                         ma->ma_need = 0;
1844                         ma->ma_valid = 0;
1845                         ma->ma_cookie_size = 0;
1846                         rc = mdo_unlink(info->mti_env,
1847                                         mdt_object_child(parent),
1848                                         mdt_object_child(child),
1849                                         lname,
1850                                         &info->mti_attr, 0);
1851                         if (rc != 0)
1852                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1853                                        mdt_obd_name(info->mti_mdt),
1854                                        PFID(mdt_object_fid(child)), rc);
1855                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1856                 }
1857         }
1858         EXIT;
1859 out_child_unlock:
1860         mdt_object_open_unlock(info, child, lhc, ibits, result);
1861 out_child:
1862         mdt_object_put(info->mti_env, child);
1863 out_parent:
1864         mdt_object_unlock_put(info, parent, lh, result || !created);
1865 out:
1866         if (result)
1867                 lustre_msg_set_transno(req->rq_repmsg, 0);
1868         return result;
1869 }
1870
1871 /**
1872  * Create an orphan object use local root.
1873  */
1874 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1875                                           struct mdt_device *mdt,
1876                                           const struct lu_fid *fid,
1877                                           struct md_attr *attr, fmode_t fmode)
1878 {
1879         const struct lu_env *env = info->mti_env;
1880         struct md_op_spec *spec = &info->mti_spec;
1881         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1882         struct mdt_object *obj = NULL;
1883         struct mdt_object *local_root;
1884         static const char name[] = "i_am_nobody";
1885         struct lu_name *lname;
1886         struct lu_ucred *uc;
1887         cfs_cap_t uc_cap_save;
1888         int rc;
1889         ENTRY;
1890
1891         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1892         if (rc != 0)
1893                 RETURN(ERR_PTR(rc));
1894
1895         local_root = mdt_object_find(env, mdt, local_root_fid);
1896         if (IS_ERR(local_root))
1897                 RETURN(local_root);
1898
1899         obj = mdt_object_new(env, mdt, fid);
1900         if (IS_ERR(obj))
1901                 GOTO(out, rc = PTR_ERR(obj));
1902
1903         spec->sp_cr_lookup = 0;
1904         spec->sp_feat = &dt_directory_features;
1905         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1906         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1907         if (attr->ma_valid & MA_LOV) {
1908                 spec->u.sp_ea.eadata = attr->ma_lmm;
1909                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1910                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1911         } else {
1912                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1913         }
1914
1915         lname = mdt_name(env, (char *)name, sizeof(name) - 1);
1916
1917         uc = lu_ucred(env);
1918         uc_cap_save = uc->uc_cap;
1919         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1920         rc = mdo_create(env, mdt_object_child(local_root), lname,
1921                         mdt_object_child(obj), spec, attr);
1922         uc->uc_cap = uc_cap_save;
1923         if (rc < 0) {
1924                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1925                        mdt_obd_name(mdt), PFID(fid), rc);
1926                 GOTO(out, rc);
1927         }
1928
1929         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1930         if (rc < 0)
1931                 CERROR("%s: cannot open volatile file "DFID", orphan "
1932                        "file will be left in PENDING directory until "
1933                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1934                        PFID(fid), rc);
1935         GOTO(out, rc);
1936
1937 out:
1938         if (rc < 0) {
1939                 if (!IS_ERR(obj))
1940                         mdt_object_put(env, obj);
1941                 obj = ERR_PTR(rc);
1942         }
1943         mdt_object_put(env, local_root);
1944         return obj;
1945 }
1946
1947 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1948                            struct md_attr *ma)
1949 {
1950         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1951         struct close_data      *data;
1952         struct ldlm_lock       *lease;
1953         struct mdt_object      *orphan;
1954         struct md_attr         *orp_ma;
1955         struct lu_buf          *buf;
1956         bool                    lease_broken;
1957         int                     rc;
1958         int                     rc2;
1959         ENTRY;
1960
1961         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1962                 RETURN(-EROFS);
1963
1964         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1965         if (data == NULL)
1966                 RETURN(-EPROTO);
1967
1968         lease = ldlm_handle2lock(&data->cd_handle);
1969         if (lease == NULL)
1970                 RETURN(-ESTALE);
1971
1972         /* try to hold open_sem so that nobody else can open the file */
1973         if (!down_write_trylock(&o->mot_open_sem)) {
1974                 ldlm_lock_cancel(lease);
1975                 GOTO(out_reprocess, rc = -EBUSY);
1976         }
1977
1978         /* Check if the lease open lease has already canceled */
1979         lock_res_and_lock(lease);
1980         lease_broken = ldlm_is_cancel(lease);
1981         unlock_res_and_lock(lease);
1982
1983         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1984                    PFID(mdt_object_fid(o)), lease_broken);
1985
1986         /* Cancel server side lease. Client side counterpart should
1987          * have been cancelled. It's okay to cancel it now as we've
1988          * held mot_open_sem. */
1989         ldlm_lock_cancel(lease);
1990
1991         if (lease_broken) /* don't perform release task */
1992                 GOTO(out_unlock, rc = -ESTALE);
1993
1994         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1995                 GOTO(out_unlock, rc = -EINVAL);
1996
1997         /* ma_need was set before but it seems fine to change it in order to
1998          * avoid modifying the one from RPC */
1999         ma->ma_need = MA_HSM;
2000         rc = mdt_attr_get_complex(info, o, ma);
2001         if (rc != 0)
2002                 GOTO(out_unlock, rc);
2003
2004         if (!mdt_hsm_release_allow(ma))
2005                 GOTO(out_unlock, rc = -EPERM);
2006
2007         /* already released? */
2008         if (ma->ma_hsm.mh_flags & HS_RELEASED)
2009                 GOTO(out_unlock, rc = 0);
2010
2011         /* Compare on-disk and packed data_version */
2012         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2013                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2014                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2015                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2016                 GOTO(out_unlock, rc = -EPERM);
2017         }
2018
2019         ma->ma_valid = MA_INODE;
2020         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2021         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2022         if (rc < 0)
2023                 GOTO(out_unlock, rc);
2024
2025         ma->ma_need = MA_INODE | MA_LOV;
2026         rc = mdt_attr_get_complex(info, o, ma);
2027         if (rc < 0)
2028                 GOTO(out_unlock, rc);
2029
2030         if (!(ma->ma_valid & MA_LOV)) {
2031                 /* Even empty file are released */
2032                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2033                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2034                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2035                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2036                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2037         } else {
2038                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2039                  * ma_lmm as lov_user_md, then it will be confused by union of
2040                  * layout_gen and stripe_offset. */
2041                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2042                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2043                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2044                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2045                 else
2046                         GOTO(out_unlock, rc = -EINVAL);
2047         }
2048
2049         /* Set file as released */
2050         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2051
2052         /* Hopefully it's not used in this call path */
2053         orp_ma = &info->mti_u.som.attr;
2054         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2055         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2056         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2057         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2058         orp_ma->ma_lmm = ma->ma_lmm;
2059         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2060         orp_ma->ma_valid = MA_INODE | MA_LOV;
2061         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2062                                  FMODE_WRITE);
2063         if (IS_ERR(orphan)) {
2064                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2065                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2066                        PTR_ERR(orphan));
2067                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2068         }
2069
2070         /* Set up HSM attribute for orphan object */
2071         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2072         buf = &info->mti_buf;
2073         buf->lb_buf = info->mti_xattr_buf;
2074         buf->lb_len = sizeof(struct hsm_attrs);
2075         ma->ma_hsm.mh_flags |= HS_RELEASED;
2076         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2077         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2078
2079         mdt_lock_reg_init(lh, LCK_EX);
2080         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2081                              MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2082         if (rc != 0)
2083                 GOTO(out_close, rc);
2084
2085         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2086                           XATTR_NAME_HSM, 0);
2087
2088         if (rc == 0)
2089                 /* Swap layout with orphan object */
2090                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2091                                      mdt_object_child(orphan),
2092                                      SWAP_LAYOUTS_MDS_HSM);
2093
2094         /* Release exclusive LL */
2095         mdt_object_unlock(info, o, lh, 1);
2096
2097         EXIT;
2098
2099 out_close:
2100         /* Close orphan object anyway */
2101         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2102                        FMODE_WRITE);
2103         if (rc2 < 0)
2104                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2105                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2106         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2107                         "object closed\n");
2108         mdt_object_put(info->mti_env, orphan);
2109
2110 out_unlock:
2111         up_write(&o->mot_open_sem);
2112
2113         if (rc == 0) { /* already released */
2114                 struct mdt_body *repbody;
2115                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2116                 LASSERT(repbody != NULL);
2117                 repbody->valid |= OBD_MD_FLRELEASED;
2118         }
2119
2120 out_reprocess:
2121         ldlm_reprocess_all(lease->l_resource);
2122         LDLM_LOCK_PUT(lease);
2123
2124         ma->ma_valid = 0;
2125         ma->ma_need = 0;
2126
2127         return rc;
2128 }
2129
2130 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2131                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2132
2133 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2134 {
2135         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2136 }
2137
2138 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2139 {
2140         struct mdt_object *o = mfd->mfd_object;
2141         struct md_object *next = mdt_object_child(o);
2142         struct md_attr *ma = &info->mti_attr;
2143         int ret = MDT_IOEPOCH_CLOSED;
2144         int rc = 0;
2145         __u64 mode;
2146         ENTRY;
2147
2148         mode = mfd->mfd_mode;
2149
2150         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2151                 rc = mdt_hsm_release(info, o, ma);
2152                 if (rc < 0) {
2153                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2154                                 mdt_obd_name(info->mti_mdt),
2155                                 PFID(mdt_object_fid(o)), rc);
2156                         /* continue to close even error occurred. */
2157                 }
2158         }
2159
2160         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2161                 mdt_write_put(o);
2162                 ret = mdt_ioepoch_close(info, o);
2163         } else if (mode & MDS_FMODE_EXEC) {
2164                 mdt_write_allow(o);
2165         } else if (mode & MDS_FMODE_EPOCH) {
2166                 ret = mdt_ioepoch_close(info, o);
2167         } else if (mode & MDS_FMODE_SOM) {
2168                 ret = mdt_som_au_close(info, o);
2169         }
2170
2171         /* Update atime on close only. */
2172         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2173             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2174                 /* Set the atime only. */
2175                 ma->ma_valid = MA_INODE;
2176                 ma->ma_attr.la_valid = LA_ATIME;
2177                 rc = mo_attr_set(info->mti_env, next, ma);
2178         }
2179
2180         /* If file data is modified, add the dirty flag. */
2181         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2182                 rc = mdt_add_dirty_flag(info, o, ma);
2183
2184         ma->ma_need |= MA_INODE;
2185         ma->ma_valid &= ~MA_INODE;
2186
2187         if (!MFD_CLOSED(mode))
2188                 rc = mo_close(info->mti_env, next, ma, mode);
2189
2190         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2191                 struct mdt_export_data *med;
2192
2193                 /* The IOepoch is still opened or SOM update is needed.
2194                  * Put mfd back into the list. */
2195                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2196                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2197                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2198
2199                 LASSERT(mdt_info_req(info));
2200                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2201                 spin_lock(&med->med_open_lock);
2202                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2203                 class_handle_hash_back(&mfd->mfd_handle);
2204                 spin_unlock(&med->med_open_lock);
2205
2206                 if (ret == MDT_IOEPOCH_OPENED) {
2207                         ret = 0;
2208                 } else {
2209                         ret = -EAGAIN;
2210                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2211                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2212                 }
2213         } else {
2214                 /* adjust open and lease count */
2215                 if (mode & MDS_OPEN_LEASE) {
2216                         LASSERT(atomic_read(&o->mot_lease_count) > 0);
2217                         atomic_dec(&o->mot_lease_count);
2218                 }
2219                 LASSERT(atomic_read(&o->mot_open_count) > 0);
2220                 atomic_dec(&o->mot_open_count);
2221
2222                 mdt_mfd_free(mfd);
2223                 mdt_object_put(info->mti_env, o);
2224         }
2225
2226         RETURN(rc ? rc : ret);
2227 }
2228
2229 int mdt_close(struct tgt_session_info *tsi)
2230 {
2231         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2232         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2233         struct mdt_export_data *med;
2234         struct mdt_file_data   *mfd;
2235         struct mdt_object      *o;
2236         struct md_attr         *ma = &info->mti_attr;
2237         struct mdt_body        *repbody = NULL;
2238         int rc, ret = 0;
2239         ENTRY;
2240
2241         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2242         /* Close may come with the Size-on-MDS update. Unpack it. */
2243         rc = mdt_close_unpack(info);
2244         if (rc)
2245                 GOTO(out, rc = err_serious(rc));
2246
2247         LASSERT(info->mti_ioepoch);
2248
2249         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2250                              info->mti_mdt->mdt_max_mdsize);
2251         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2252                              info->mti_mdt->mdt_max_cookiesize);
2253         rc = req_capsule_server_pack(info->mti_pill);
2254         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2255                 mdt_client_compatibility(info);
2256                 if (rc == 0)
2257                         mdt_fix_reply(info);
2258                 mdt_exit_ucred(info);
2259                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2260         }
2261
2262         /* Continue to close handle even if we can not pack reply */
2263         if (rc == 0) {
2264                 repbody = req_capsule_server_get(info->mti_pill,
2265                                                  &RMF_MDT_BODY);
2266                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2267                                                     &RMF_MDT_MD);
2268                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2269                                                        &RMF_MDT_MD,
2270                                                        RCL_SERVER);
2271                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2272                                                        &RMF_LOGCOOKIES);
2273                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2274                                                           &RMF_LOGCOOKIES,
2275                                                           RCL_SERVER);
2276                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2277                 repbody->eadatasize = 0;
2278                 repbody->aclsize = 0;
2279         } else {
2280                 rc = err_serious(rc);
2281         }
2282
2283         med = &req->rq_export->exp_mdt_data;
2284         spin_lock(&med->med_open_lock);
2285         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2286                              req_is_replay(req));
2287         if (mdt_mfd_closed(mfd)) {
2288                 spin_unlock(&med->med_open_lock);
2289                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2290                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2291                        info->mti_ioepoch->handle.cookie);
2292                 /** not serious error since bug 3633 */
2293                 rc = -ESTALE;
2294         } else {
2295                 class_handle_unhash(&mfd->mfd_handle);
2296                 cfs_list_del_init(&mfd->mfd_list);
2297                 spin_unlock(&med->med_open_lock);
2298
2299                 /* Do not lose object before last unlink. */
2300                 o = mfd->mfd_object;
2301                 mdt_object_get(info->mti_env, o);
2302                 ret = mdt_mfd_close(info, mfd);
2303                 if (repbody != NULL)
2304                         rc = mdt_handle_last_unlink(info, o, ma);
2305                 mdt_empty_transno(info, rc);
2306                 mdt_object_put(info->mti_env, o);
2307         }
2308         if (repbody != NULL) {
2309                 mdt_client_compatibility(info);
2310                 rc = mdt_fix_reply(info);
2311         }
2312
2313         mdt_exit_ucred(info);
2314         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2315                 GOTO(out, rc = err_serious(-ENOMEM));
2316
2317         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2318                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2319                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2320 out:
2321         mdt_thread_info_fini(info);
2322         RETURN(rc ? rc : ret);
2323 }
2324
2325 /**
2326  * DONE_WRITING rpc handler.
2327  *
2328  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2329  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2330  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2331  * skip attributes and reconstruct the reply here.
2332  */
2333 int mdt_done_writing(struct tgt_session_info *tsi)
2334 {
2335         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2336         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2337         struct mdt_body         *repbody = NULL;
2338         struct mdt_export_data  *med;
2339         struct mdt_file_data    *mfd;
2340         int rc;
2341         ENTRY;
2342
2343         rc = req_capsule_server_pack(tsi->tsi_pill);
2344         if (rc)
2345                 GOTO(out, rc = err_serious(rc));
2346
2347         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2348         repbody->eadatasize = 0;
2349         repbody->aclsize = 0;
2350
2351         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2352         rc = mdt_close_unpack(info);
2353         if (rc)
2354                 GOTO(out, rc = err_serious(rc));
2355
2356         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2357                 mdt_exit_ucred(info);
2358                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2359         }
2360
2361         med = &info->mti_exp->exp_mdt_data;
2362         spin_lock(&med->med_open_lock);
2363         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2364                              req_is_replay(req));
2365         if (mfd == NULL) {
2366                 spin_unlock(&med->med_open_lock);
2367                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2368                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
2369                        PFID(info->mti_rr.rr_fid1),
2370                        info->mti_ioepoch->handle.cookie,
2371                        info->mti_ioepoch->ioepoch);
2372                 /* If this is a replay, reconstruct the transno. */
2373                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2374                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2375                              -EAGAIN : 0;
2376                         mdt_empty_transno(info, rc);
2377                 } else
2378                         rc = -ESTALE;
2379                 GOTO(error_ucred, rc);
2380         }
2381
2382         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2383                 mfd->mfd_mode == MDS_FMODE_TRUNC);
2384         class_handle_unhash(&mfd->mfd_handle);
2385         cfs_list_del_init(&mfd->mfd_list);
2386         spin_unlock(&med->med_open_lock);
2387
2388         /* Set EPOCH CLOSE flag if not set by client. */
2389         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2390         info->mti_attr.ma_valid = 0;
2391
2392         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2393         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2394         if (info->mti_attr.ma_lmm == NULL)
2395                 GOTO(error_ucred, rc = -ENOMEM);
2396
2397         rc = mdt_mfd_close(info, mfd);
2398
2399         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2400         mdt_empty_transno(info, rc);
2401 error_ucred:
2402         mdt_exit_ucred(info);
2403 out:
2404         mdt_thread_info_fini(info);
2405         RETURN(rc);
2406 }