lustre/mds/mds_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *   Author: Phil Schwan <phil@clusterfs.com>
   8  *   Author: Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/version.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/version.h>
  35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  36 # include <linux/buffer_head.h>
  37 # include <linux/workqueue.h>
  38 #else
  39 # include <linux/locks.h>
  40 #endif
  41
  42 #include <linux/obd_class.h>
  43 #include <linux/obd_lov.h>
  44 #include <linux/lustre_fsfilt.h>
  45 #include <linux/lprocfs_status.h>
  46 #include <linux/lustre_gs.h>
  47
  48 #include "mds_internal.h"
  49
  50 /* Exported function from this file are:
  51  *
  52  * mds_open - called by the intent handler
  53  * mds_close - an rpc handling function
  54  * mds_pin - an rpc handling function - which will go away
  55  * mds_mfd_close - for force closing files when a client dies
  56  */
  57
  58 /*
  59  * MDS file data handling: file data holds a handle for a file opened
  60  * by a client.
  61  */
  62
  63 static void mds_mfd_addref(void *mfdp)
  64 {
  65         struct mds_file_data *mfd = mfdp;
  66
  67         atomic_inc(&mfd->mfd_refcount);
  68         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  69                atomic_read(&mfd->mfd_refcount));
  70 }
  71
  72 struct mds_file_data *mds_mfd_new(void)
  73 {
  74         struct mds_file_data *mfd;
  75
  76         OBD_ALLOC(mfd, sizeof *mfd);
  77         if (mfd == NULL) {
  78                 CERROR("mds: out of memory\n");
  79                 return NULL;
  80         }
  81
  82         atomic_set(&mfd->mfd_refcount, 2);
  83
  84         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  85         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  86
  87         return mfd;
  88 }
  89
  90 struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  91 {
  92         ENTRY;
  93         LASSERT(handle != NULL);
  94         RETURN(class_handle2object(handle->cookie));
  95 }
  96
  97 void mds_mfd_put(struct mds_file_data *mfd)
  98 {
  99         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 100                atomic_read(&mfd->mfd_refcount) - 1);
 101         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 102                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 103         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 104                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 105                 OBD_FREE(mfd, sizeof *mfd);
 106         }
 107 }
 108
 109 static void mds_mfd_destroy(struct mds_file_data *mfd)
 110 {
 111         class_handle_unhash(&mfd->mfd_handle);
 112         mds_mfd_put(mfd);
 113 }
 114
 115 /* Caller must hold mds->mds_epoch_sem */
 116 static int mds_alloc_filterdata(struct inode *inode)
 117 {
 118         LASSERT(LUSTRE_FILTERDATA(inode) == NULL);
 119         OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 120         if (LUSTRE_FILTERDATA(inode) == NULL)
 121                 return -ENOMEM;
 122         LASSERT(igrab(inode) == inode);
 123         return 0;
 124 }
 125
 126 /* Caller must hold mds->mds_epoch_sem */
 127 static void mds_free_filterdata(struct inode *inode)
 128 {
 129         LASSERT(LUSTRE_FILTERDATA(inode) != NULL);
 130         OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 131         LUSTRE_FILTERDATA(inode) = NULL;
 132         iput(inode);
 133 }
 134
 135 /* Write access to a file: executors cause a negative count,
 136  * writers a positive count.  The semaphore is needed to perform
 137  * a check for the sign and then increment or decrement atomically.
 138  *
 139  * This code is closely tied to the allocation of the d_fsdata and the
 140  * MDS epoch, so we use the same semaphore for the whole lot.
 141  *
 142  * We could use a different semaphore for each file, if it ever shows
 143  * up in a profile, which it won't.
 144  *
 145  * epoch argument is nonzero during recovery */
 146 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 147                                 __u64 epoch)
 148 {
 149         int rc = 0;
 150
 151         down(&mds->mds_epoch_sem);
 152
 153         if (atomic_read(&inode->i_writecount) < 0) {
 154                 up(&mds->mds_epoch_sem);
 155                 RETURN(-ETXTBSY);
 156         }
 157
 158         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 159                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 160                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 161                        inode->i_generation);
 162                 goto out;
 163         }
 164
 165         if (MDS_FILTERDATA(inode) == NULL)
 166                 mds_alloc_filterdata(inode);
 167         if (MDS_FILTERDATA(inode) == NULL) {
 168                 rc = -ENOMEM;
 169                 goto out;
 170         }
 171         if (epoch > mds->mds_io_epoch) {
 172                 mds->mds_io_epoch = epoch;
 173                 CDEBUG(D_INODE, "repair MDS epoch "LPU64" for ino %lu/%u\n",
 174                        mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 175         } else {
 176                 mds->mds_io_epoch++;
 177                 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 178                        mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 179         }
 180         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 181  out:
 182         if (rc == 0)
 183                 atomic_inc(&inode->i_writecount);
 184         up(&mds->mds_epoch_sem);
 185         return rc;
 186 }
 187
 188 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 189  * again -- which is never true if the file is about to be unlinked.  Otherwise
 190  * returns the number of remaining writers. */
 191 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 192                                 struct mds_body *body, int unlinking)
 193 {
 194         int rc = 0;
 195         ENTRY;
 196
 197         down(&mds->mds_epoch_sem);
 198         atomic_dec(&inode->i_writecount);
 199         rc = atomic_read(&inode->i_writecount);
 200         if (rc > 0)
 201                 GOTO(out, rc);
 202 #if 0
 203         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 204                 GOTO(out, rc = EAGAIN);
 205 #endif
 206         mds_free_filterdata(inode);
 207  out:
 208         up(&mds->mds_epoch_sem);
 209         return rc;
 210 }
 211
 212 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 213 {
 214         ENTRY;
 215         down(&mds->mds_epoch_sem);
 216         if (atomic_read(&inode->i_writecount) > 0) {
 217                 up(&mds->mds_epoch_sem);
 218                 RETURN(-ETXTBSY);
 219         }
 220         atomic_dec(&inode->i_writecount);
 221         up(&mds->mds_epoch_sem);
 222         RETURN(0);
 223 }
 224
 225 static void mds_allow_write_access(struct inode *inode)
 226 {
 227         ENTRY;
 228         atomic_inc(&inode->i_writecount);
 229 }
 230
 231 int mds_query_write_access(struct inode *inode)
 232 {
 233         ENTRY;
 234         RETURN(atomic_read(&inode->i_writecount));
 235 }
 236
 237 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 238 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 239                                              struct vfsmount *mnt, int flags,
 240                                              struct ptlrpc_request *req,
 241                                              struct mds_update_record *rec)
 242 {
 243         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 244         struct mds_obd *mds = mds_req2mds(req);
 245         struct mds_file_data *mfd;
 246         struct mds_body *body;
 247         int rc = 0;
 248         ENTRY;
 249
 250         mfd = mds_mfd_new();
 251         if (mfd == NULL) {
 252                 CERROR("mds: out of memory\n");
 253                 GOTO(cleanup_dentry, rc = -ENOMEM);
 254         }
 255
 256         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 257
 258         if (flags & FMODE_WRITE) {
 259                 /* FIXME: in recovery, need to pass old epoch here */
 260                 rc = mds_get_write_access(mds, dentry->d_inode, rec->ur_ioepoch);
 261                 if (rc)
 262                         GOTO(cleanup_mfd, rc);
 263                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 264         } else if (flags & FMODE_EXEC) {
 265                 rc = mds_deny_write_access(mds, dentry->d_inode);
 266                 if (rc)
 267                         GOTO(cleanup_mfd, rc);
 268         }
 269
 270         dget(dentry);
 271
 272         /* FIXME: invalidate update locks when first open for write comes in */
 273
 274         /* mark the file as open to handle open-unlink. */
 275         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 276         mds_orphan_open_inc(dentry->d_inode);
 277         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 278
 279         mfd->mfd_mode = flags;
 280         mfd->mfd_dentry = dentry;
 281         mfd->mfd_xid = req->rq_xid;
 282
 283         spin_lock(&med->med_open_lock);
 284         list_add(&mfd->mfd_list, &med->med_open_head);
 285         spin_unlock(&med->med_open_lock);
 286         mds_mfd_put(mfd);
 287
 288         body->handle.cookie = mfd->mfd_handle.h_cookie;
 289         RETURN(mfd);
 290 cleanup_mfd:
 291         mds_mfd_put(mfd);
 292         mds_mfd_destroy(mfd);
 293 cleanup_dentry:
 294         return ERR_PTR(rc);
 295 }
 296
 297 static inline void
 298 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 299                     struct lov_desc *desc)
 300 {
 301         int i;
 302
 303         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 304                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 305                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 306         }
 307 }
 308
 309 /* must be called with i_sem held */
 310 int
 311 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
 312                    int offset, struct mds_update_record *rec,
 313                    struct dentry *dchild, void **handle,
 314                    obd_id **ids)
 315 {
 316         struct inode *inode = dchild->d_inode;
 317         struct mds_obd *mds = &obd->u.mds;
 318         struct obd_trans_info oti = { 0 };
 319         struct lov_stripe_md *lsm = NULL;
 320         struct lov_mds_md *lmm = NULL;
 321         int rc, lmm_bufsize, lmm_size;
 322         struct obdo *oa = NULL;
 323         struct mds_body *body;
 324         void *lmm_buf;
 325         ENTRY;
 326
 327         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 328             !(rec->ur_flags & FMODE_WRITE))
 329                 RETURN(0);
 330
 331         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 332
 333         if (!S_ISREG(inode->i_mode))
 334                 RETURN(0);
 335         if (body->valid & OBD_MD_FLEASIZE)
 336                 RETURN(0);
 337
 338         OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 339         if (*ids == NULL)
 340                 RETURN(-ENOMEM);
 341         oti.oti_objid = *ids;
 342
 343         /* replay case */
 344         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 345                 LASSERT(id_ino(rec->ur_id2));
 346                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 347                 lmm_size = rec->ur_eadatalen;
 348                 lmm = rec->ur_eadata;
 349                 LASSERT(lmm);
 350
 351                 if (*handle == NULL)
 352                         *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 353                 if (IS_ERR(*handle)) {
 354                         rc = PTR_ERR(*handle);
 355                         *handle = NULL;
 356                         RETURN(rc);
 357                 }
 358
 359                 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 360
 361                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 362                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 363                 LASSERT(lmm_buf != NULL);
 364                 LASSERT(lmm_bufsize >= lmm_size);
 365
 366                 memcpy(lmm_buf, lmm, lmm_size);
 367                 rc = fsfilt_set_md(obd, inode, *handle, lmm,
 368                                    lmm_size, EA_LOV);
 369                 if (rc)
 370                         CERROR("open replay failed to set md:%d\n", rc);
 371                 RETURN(0);
 372         }
 373
 374         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 375                 GOTO(out_ids, rc = -ENOMEM);
 376
 377         oa = obdo_alloc();
 378         if (oa == NULL)
 379                 RETURN(-ENOMEM);
 380         oa->o_mode = S_IFREG | 0600;
 381         oa->o_id = inode->i_ino;
 382         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 383         oa->o_generation = inode->i_generation;
 384         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 385         oa->o_gid = 0;
 386         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
 387                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 388         oa->o_size = 0;
 389
 390         obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 391                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 392
 393         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 394                 /* check if things like lfs setstripe are sending us the ea */
 395                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 396                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 397                                            mds->mds_dt_exp,
 398                                            0, &lsm, rec->ur_eadata);
 399                         if (rc)
 400                                 GOTO(out_oa, rc);
 401                 } else {
 402                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 403                         if (lmm == NULL)
 404                                 GOTO(out_oa, rc = -ENOMEM);
 405
 406                         lmm_size = mds->mds_max_mdsize;
 407                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 408                                         lmm, &lmm_size, 1, 0);
 409                         if (rc > 0)
 410                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 411                                                    mds->mds_dt_exp,
 412                                                    0, &lsm, lmm);
 413                         OBD_FREE(lmm, mds->mds_max_mdsize);
 414                         if (rc)
 415                                 GOTO(out_oa, rc);
 416                 }
 417
 418                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 419                 oti.oti_flags |= OBD_MODE_CROW;
 420                 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
 421
 422                 if (rc) {
 423                         CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
 424                                "error creating objects for "
 425                                "inode %lu: rc = %d\n",
 426                                inode->i_ino, rc);
 427                         if (rc > 0) {
 428                                 CERROR("obd_create returned invalid "
 429                                        "rc %d\n", rc);
 430                                 rc = -EIO;
 431                         }
 432                         GOTO(out_oa, rc);
 433                 }
 434         } else {
 435                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 436                                    0, &lsm, rec->ur_eadata);
 437                 if (rc)
 438                         GOTO(out_oa, rc);
 439
 440                 lsm->lsm_object_id = oa->o_id;
 441                 lsm->lsm_object_gr = oa->o_gr;
 442         }
 443         if (inode->i_size) {
 444                 oa->o_size = inode->i_size;
 445                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 446                                 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 447
 448                 /* pack lustre id to oss */
 449                 *(obdo_id(oa)) = body->id1;
 450                 oa->o_valid |= OBD_MD_FLIFID;
 451
 452                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti, NULL);
 453                 if (rc) {
 454                         CERROR("error setting attrs for inode %lu: rc %d\n",
 455                                inode->i_ino, rc);
 456                         if (rc > 0) {
 457                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 458                                 rc = -EIO;
 459                         }
 460                         GOTO(out_oa, rc);
 461                 }
 462         }
 463
 464         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 465         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 466
 467         LASSERT(lsm && lsm->lsm_object_id);
 468         lmm = NULL;
 469         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 470         if (!id_ino(rec->ur_id2))
 471                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 472         if (rc < 0) {
 473                 CERROR("cannot pack lsm, err = %d\n", rc);
 474                 GOTO(out_oa, rc);
 475         }
 476
 477         lmm_size = rc;
 478         body->eadatasize = rc;
 479
 480         if (*handle == NULL)
 481                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 482         if (IS_ERR(*handle)) {
 483                 rc = PTR_ERR(*handle);
 484                 *handle = NULL;
 485                 GOTO(out_oa, rc);
 486         }
 487
 488         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 489                            lmm_size, EA_LOV);
 490
 491         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 492         lmm_bufsize = req->rq_repmsg->buflens[offset];
 493         LASSERT(lmm_buf);
 494         LASSERT(lmm_bufsize >= lmm_size);
 495
 496         memcpy(lmm_buf, lmm, lmm_size);
 497         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 498 out_oa:
 499         oti_free_cookies(&oti);
 500         obdo_free(oa);
 501 out_ids:
 502         if (rc) {
 503                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 504                 *ids = NULL;
 505         }
 506         if (lsm)
 507                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 508         RETURN(rc);
 509 }
 510
 511 int
 512 mds_destroy_object(struct obd_device *obd,
 513                    struct inode *inode, int async)
 514 {
 515         struct mds_obd *mds = &obd->u.mds;
 516         struct lov_mds_md *lmm = NULL;
 517         int rc, lmm_size;
 518         ENTRY;
 519
 520         LASSERT(inode != NULL);
 521
 522         if (inode->i_nlink != 0) {
 523                 CDEBUG(D_INODE, "attempt to destroy OSS object when "
 524                        "i_nlink == %d\n", (int)inode->i_nlink);
 525                 RETURN(0);
 526         }
 527
 528         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 529         if (lmm == NULL)
 530                 RETURN(-ENOMEM);
 531
 532         lmm_size = mds->mds_max_mdsize;
 533         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 534         if (rc < 0) {
 535                 CERROR("no stripe info for %lu/%lu inode\n",
 536                        (unsigned long)inode->i_ino,
 537                        (unsigned long)inode->i_generation);
 538                 GOTO(out_free_lmm, rc);
 539         }
 540
 541         if (rc > 0) {
 542                 /* asynchronously unlink objecect on OSS */
 543                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 544                                        NULL, 0, async);
 545                 if (rc) {
 546                         CERROR("error unlinking object on OSS, "
 547                                "err %d\n", rc);
 548                         GOTO(out_free_lmm, rc);
 549                 }
 550         } else {
 551                 CDEBUG(D_INODE, "no stripping info found for inode "
 552                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 553                       (unsigned long)inode->i_generation);
 554         }
 555         EXIT;
 556 out_free_lmm:
 557         OBD_FREE(lmm, mds->mds_max_mdsize);
 558         return rc;
 559 }
 560
 561 static void reconstruct_open(struct mds_update_record *rec, int offset,
 562                              struct ptlrpc_request *req,
 563                              struct lustre_handle *child_lockh)
 564 {
 565         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 566         struct mds_client_data *mcd = med->med_mcd;
 567         struct mds_obd *mds = mds_req2mds(req);
 568         struct mds_file_data *mfd;
 569         struct obd_device *obd = req->rq_export->exp_obd;
 570         struct dentry *parent = NULL, *dchild;
 571         struct ldlm_reply *rep;
 572         struct mds_body *body;
 573         struct list_head *t;
 574         int put_child = 1;
 575         int rc;
 576         ENTRY;
 577
 578         LASSERT(offset == 3); /* only called via intent */
 579         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 580         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 581
 582         /* copy rc, transno and disp; steal locks */
 583         mds_req_from_mcd(req, mcd);
 584         intent_set_disposition(rep, mcd->mcd_last_data);
 585
 586         /* Only replay if create or open actually happened. */
 587         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 588                 EXIT;
 589                 return; /* error looking up parent or child */
 590         }
 591
 592         /* first, we try to open the file by fid. by the time of this
 593          * request, inode can be an orphan and parent can disappear */
 594         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 595                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT|REPLAY)\n",
 596                        OLID4(rec->ur_id2));
 597                 dchild = mds_id2dentry(obd, rec->ur_id2, NULL);
 598         } else if (rec->ur_namelen == 1) {
 599                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 600                        OLID4(rec->ur_id1));
 601                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 602         } else {
 603                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 604                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 605                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 606                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 607                                 rec->ur_namelen - 1);
 608                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 609                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 610         }
 611
 612         if (!dchild->d_inode)
 613                 GOTO(out_dput, 0); /* child not present to open */
 614
 615         /* At this point, we know we have a child. We'll send
 616          * it back _unless_ it not created and open failed.  */
 617         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 618             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 619             req->rq_status)
 620                 GOTO(out_dput, 0);
 621
 622         /* get lock (write for O_CREAT, read otherwise) */
 623         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 624         if (S_ISREG(dchild->d_inode->i_mode)) {
 625                 struct lustre_capa capa = {
 626                         .lc_uid   = rec->ur_uc.luc_uid,
 627                         .lc_op    = capa_op(rec->ur_flags),
 628                         .lc_ino   = dchild->d_inode->i_ino,
 629                         .lc_igen  = dchild->d_inode->i_generation,
 630                         .lc_mdsid = mds->mds_num,
 631                 };
 632                 int off = 7; /* capa offset */
 633
 634                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 635                                  dchild->d_inode, 1, 0);
 636
 637                 if (rc)
 638                         LASSERT(rc == req->rq_status);
 639
 640                 /* If we have LOV EA data, the OST holds size, mtime */
 641                 if (!(body->valid & OBD_MD_FLEASIZE))
 642                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 643                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 644                 rc = mds_pack_capa(obd, med, NULL, &capa, req, &off, body);
 645         }
 646
 647         /* If we have -EEXIST as the status, and we were asked to create
 648          * exclusively, we can tell we failed because the file already existed.
 649          */
 650         if (req->rq_status == -EEXIST &&
 651             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 652              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 653                 GOTO(out_dput, 0);
 654         }
 655
 656         /* If we didn't get as far as trying to open, then some locking thing
 657          * probably went wrong, and we'll just bail here.
 658          */
 659         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 660                 GOTO(out_dput, 0);
 661
 662         /* If we failed, then we must have failed opening, so don't look for
 663          * file descriptor or anything, just give the client the bad news.
 664          */
 665         if (req->rq_status)
 666                 GOTO(out_dput, 0);
 667
 668         mfd = NULL;
 669         list_for_each(t, &med->med_open_head) {
 670                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 671                 if (mfd->mfd_xid == req->rq_xid)
 672                         break;
 673                 mfd = NULL;
 674         }
 675
 676         /* #warning "XXX fixme" bug 2991 */
 677         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 678          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 679          * to detect a re-open */
 680         if (mfd == NULL) {
 681                 mntget(mds->mds_vfsmnt);
 682                 CERROR("Re-opened file \n");
 683                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 684                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req, rec);
 685                 if (!mfd) {
 686                         CERROR("mds: out of memory\n");
 687                         GOTO(out_dput, req->rq_status = -ENOMEM);
 688                 }
 689                 put_child = 0;
 690         } else {
 691                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 692                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 693                        mfd->mfd_handle.h_cookie);
 694         }
 695
 696  out_dput:
 697         if (put_child)
 698                 l_dput(dchild);
 699         if (parent)
 700                 l_dput(parent);
 701         EXIT;
 702 }
 703
 704 /* do NOT or the MAY_*'s, you'll get the weakest */
 705 int accmode(int flags)
 706 {
 707         int res = 0;
 708
 709         if (flags & FMODE_READ)
 710                 res = MAY_READ;
 711         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 712                 res |= MAY_WRITE;
 713         if (flags & FMODE_EXEC)
 714                 res = MAY_EXEC;
 715         return res;
 716 }
 717
 718 /* Handles object creation, actual opening, and I/O epoch */
 719 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 720                            struct mds_body *body, int flags, void **handle,
 721                            struct mds_update_record *rec, struct ldlm_reply *rep,
 722                            int *cancel_update_lock)
 723 {
 724         struct obd_device *obd = req->rq_export->exp_obd;
 725         struct mds_obd *mds = mds_req2mds(req);
 726         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 727         struct mds_file_data *mfd = NULL;
 728         obd_id *ids = NULL;
 729         unsigned mode;
 730         int rc = 0, reply_off;
 731         ENTRY;
 732
 733         *cancel_update_lock = 0;
 734
 735         /* atomically create objects if necessary */
 736         down(&dchild->d_inode->i_sem);
 737         mode = dchild->d_inode->i_mode;
 738
 739         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 740             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 741                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 742                                  dchild->d_inode, 0, 0);
 743                 if (rc) {
 744                         up(&dchild->d_inode->i_sem);
 745                         RETURN(rc);
 746                 }
 747         }
 748
 749         if (rec != NULL) {
 750                 if ((body->valid & OBD_MD_FLEASIZE) &&
 751                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 752                         up(&dchild->d_inode->i_sem);
 753                         RETURN(-EEXIST);
 754                 }
 755
 756                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 757                         /* no EA: create objects */
 758                         rc = mds_create_objects(obd, req, 2, rec,
 759                                                 dchild, handle, &ids);
 760                         if (rc) {
 761                                 CERROR("mds_create_object: rc = %d\n", rc);
 762                                 up(&dchild->d_inode->i_sem);
 763                                 RETURN(rc);
 764                         }
 765                 }
 766
 767                 if (S_ISREG(dchild->d_inode->i_mode) &&
 768                     (body->valid & OBD_MD_FLEASIZE)) {
 769                         rc = mds_revalidate_lov_ea(obd,dchild,req->rq_repmsg,2);
 770                         if (!rc) {
 771                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 772                                                  dchild->d_inode, 0, 0);
 773                                 /* after ost add/delete, lov ea can change
 774                                  * along with size/blocks, which we have
 775                                  * to update as well */
 776                                 body->size = dchild->d_inode->i_size;
 777                                 body->blocks = dchild->d_inode->i_blocks;
 778                                 *cancel_update_lock = 1;
 779                         }
 780                         if (rc) {
 781                                 up(&dchild->d_inode->i_sem);
 782                                 RETURN(rc);
 783                         }
 784                 }
 785         }
 786
 787         reply_off = 3;
 788         rc = mds_pack_acl(req, reply_off, body, dchild->d_inode);
 789         reply_off += 2;
 790
 791         if (rc < 0) {
 792                 CERROR("pack posix acl: rc = %d\n", rc);
 793                 up(&dchild->d_inode->i_sem);
 794                 RETURN(rc);
 795         }
 796
 797         rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
 798                             dchild->d_inode);
 799         if (rc < 0) {
 800                 CERROR("mds_pack_gskey: rc = %d\n", rc);
 801                 up(&dchild->d_inode->i_sem);
 802                 RETURN(rc);
 803         }
 804
 805         if (S_ISREG(mode)) {
 806                 struct lustre_capa capa = {
 807                         .lc_uid   = rec->ur_uc.luc_uid,
 808                         .lc_op    = capa_op(rec->ur_flags),
 809                         .lc_ino   = dchild->d_inode->i_ino,
 810                         .lc_igen  = dchild->d_inode->i_generation,
 811                         .lc_mdsid = mds->mds_num,
 812                 };
 813
 814                 rc = mds_pack_capa(obd, med, NULL, &capa, req,
 815                                    &reply_off, body);
 816                 if (rc < 0) {
 817                         CERROR("mds_pack_capa: rc = %d\n", rc);
 818                         up(&dchild->d_inode->i_sem);
 819                         RETURN(rc);
 820                 }
 821         } else {
 822                 reply_off++;
 823         }
 824
 825         /* If the inode has no EA data, then MDSs hold size, mtime */
 826         if (S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) {
 827                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 828                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 829         }
 830
 831         up(&dchild->d_inode->i_sem);
 832
 833         intent_set_disposition(rep, DISP_OPEN_OPEN);
 834         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req, rec);
 835         if (IS_ERR(mfd))
 836                 RETURN(PTR_ERR(mfd));
 837
 838         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 839                mfd->mfd_handle.h_cookie);
 840
 841         if (ids != NULL) {
 842                 mds_dt_update_objids(obd, ids);
 843                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 844         }
 845
 846         RETURN(rc);
 847 }
 848
 849 static int mds_open_by_id(struct ptlrpc_request *req,
 850                           struct lustre_id *id,
 851                           struct mds_body *body, int flags,
 852                           struct mds_update_record *rec,
 853                           struct ldlm_reply *rep)
 854 {
 855         struct mds_obd *mds = mds_req2mds(req);
 856         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 857         struct dentry *dchild;
 858         char idname[LL_ID_NAMELEN];
 859         int idlen = 0, rc, fake;
 860         void *handle = NULL;
 861         ENTRY;
 862
 863         down(&pending_dir->i_sem);
 864         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 865         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 866                                 idlen);
 867         if (IS_ERR(dchild)) {
 868                 up(&pending_dir->i_sem);
 869                 rc = PTR_ERR(dchild);
 870                 CERROR("error looking up %s in PENDING: rc = %d\n",
 871                        idname, rc);
 872                 RETURN(rc);
 873         }
 874
 875         up(&pending_dir->i_sem);
 876         if (dchild->d_inode != NULL) {
 877                 mds_inode_set_orphan(dchild->d_inode);
 878                 mds_pack_inode2body(req2obd(req), body,
 879                                     dchild->d_inode, 1);
 880
 881                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 882                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 883                 CWARN("Orphan %s found and opened in PENDING directory\n",
 884                        idname);
 885                 goto open;
 886         }
 887         l_dput(dchild);
 888
 889         /* we didn't find it in PENDING so it isn't an orphan.  See if it was a
 890          * regular inode that was previously created. */
 891         dchild = mds_id2dentry(req2obd(req), id, NULL);
 892         if (IS_ERR(dchild))
 893                 RETURN(PTR_ERR(dchild));
 894
 895         mds_pack_inode2body(req2obd(req), body,
 896                             dchild->d_inode, 1);
 897
 898         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 899         intent_set_disposition(rep, DISP_LOOKUP_POS);
 900
 901  open:
 902         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep,&fake);
 903         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 904                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 905         /* XXX what do we do here if mds_finish_transno itself failed? */
 906         l_dput(dchild);
 907         RETURN(rc);
 908 }
 909
 910 int mds_pin(struct ptlrpc_request *req, int offset)
 911 {
 912         struct obd_device *obd = req->rq_export->exp_obd;
 913         struct mds_body *request_body, *reply_body;
 914         struct lvfs_run_ctxt saved;
 915         int rc, size = sizeof(*reply_body);
 916         ENTRY;
 917
 918         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 919                                       sizeof(*request_body));
 920
 921         rc = lustre_pack_reply(req, 1, &size, NULL);
 922         if (rc)
 923                 RETURN(rc);
 924         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 925
 926         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 927         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 928                             request_body->flags, NULL, NULL);
 929         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 930
 931         RETURN(rc);
 932 }
 933
 934 /*
 935  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 936  * child_lockh is NULL we just get the lock as a barrier to wait for other
 937  * holders of this lock, and drop it right away again.
 938  */
 939 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 940                        struct lustre_handle *child_lockh)
 941 {
 942         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 943         struct lustre_handle lockh;
 944         int lock_flags = LDLM_FL_ATOMIC_CB;
 945         int rc;
 946         ENTRY;
 947
 948         if (child_lockh == NULL)
 949                 child_lockh = &lockh;
 950
 951         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 952                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 953                               mds_blocking_ast, ldlm_completion_ast, NULL,
 954                               NULL, NULL, 0, NULL, child_lockh);
 955         if (rc != ELDLM_OK)
 956                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 957         else if (child_lockh == &lockh)
 958                 ldlm_lock_decref(child_lockh, LCK_EX);
 959
 960         RETURN(rc);
 961 }
 962
 963 int mds_open(struct mds_update_record *rec, int offset,
 964              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 965 {
 966         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 967         struct obd_device *obd = req->rq_export->exp_obd;
 968         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 969         struct mds_obd *mds = mds_req2mds(req);
 970         struct ldlm_reply *rep = NULL;
 971         struct mds_body *body = NULL;
 972         struct dentry *dchild = NULL, *dparent = NULL;
 973         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 974         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 975         int parent_mode = LCK_PR;
 976         void *handle = NULL;
 977         struct dentry_params dp;
 978         struct mea *mea = NULL;
 979         int mea_size, update_mode;
 980         int child_mode = LCK_PR;
 981         struct lustre_id sid;
 982         __u64 fid = 0;
 983         int cancel_update_lock;
 984         ENTRY;
 985
 986         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 987                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 988                   rec->ur_mode);
 989
 990         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 991                          (obd_timeout + 1) / 4);
 992
 993         if (offset == 3) { /* intent */
 994                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 995                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 996         } else if (offset == 1) { /* non-intent reint */
 997                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 998         } else {
 999                 body = NULL;
1000                 LBUG();
1001         }
1002
1003         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
1004
1005         /*
1006          * Step 0: If we are passed a id, then we assume the client already
1007          * opened this file and is only replaying the RPC, so we open the inode
1008          * by fid (at some large expense in security).
1009          */
1010         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1011                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
1012                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
1013                        rec->ur_name, rec->ur_mode);
1014
1015                 LASSERT(id_ino(rec->ur_id2));
1016
1017                 rc = mds_open_by_id(req, rec->ur_id2, body,
1018                                     rec->ur_flags, rec, rep);
1019                 if (rc != -ENOENT) {
1020                         mds_body_do_reverse_map(med, body);
1021                         RETURN(rc);
1022                 }
1023
1024                 /* We didn't find the correct inode on disk either, so we
1025                  * need to re-create it via a regular replay. */
1026                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
1027         } else {
1028                 LASSERT(!id_ino(rec->ur_id2));
1029         }
1030
1031         LASSERT(offset == 3); /* If we got here, we must be called via intent */
1032
1033         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1034                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1035                 RETURN(-ENOMEM);
1036         }
1037
1038         acc_mode = accmode(rec->ur_flags);
1039
1040         /* Step 1: Find and lock the parent */
1041         if (rec->ur_flags & MDS_OPEN_CREAT) {
1042                 /* XXX Well, in fact we only need this lock mode change if
1043                    in addition to O_CREAT, the file does not exist.
1044                    But we do not know if it exists or not yet */
1045                 parent_mode = LCK_PW;
1046         }
1047
1048         if (rec->ur_namelen == 1) {
1049                 /* client (LMV) wants to open the file by id */
1050                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
1051                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
1052                 if (IS_ERR(dchild)) {
1053                         rc = PTR_ERR(dparent);
1054                         CERROR("child lookup by an id error %d\n", rc);
1055                         GOTO(cleanup, rc);
1056                 }
1057                 goto got_child;
1058         }
1059
1060 restart:
1061         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1062                                        parent_lockh, &update_mode, rec->ur_name,
1063                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1064         if (IS_ERR(dparent)) {
1065                 rc = PTR_ERR(dparent);
1066                 if (rc != -ENOENT) {
1067                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1068                                OLID4(rec->ur_id1), rc);
1069                 } else {
1070                         /* Just cannot find parent - make it look like
1071                            usual negative lookup to avoid extra MDS RPC */
1072                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1073                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1074                 }
1075                 GOTO(cleanup, rc);
1076         }
1077         LASSERT(dparent->d_inode != NULL);
1078
1079         cleanup_phase = 1; /* parent dentry and lock */
1080
1081         /* try to retrieve MEA data for this dir */
1082         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1083         if (rc)
1084                 GOTO(cleanup, rc);
1085
1086         if (mea != NULL && mea->mea_count) {
1087                 /*
1088                  * dir is already splitted, check is requested filename should *
1089                  * live at this MDS or at another one.
1090                  */
1091                 int i;
1092                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1093                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1094                         CDEBUG(D_OTHER,
1095                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1096                                " should be %lu(%d)\n", obd->obd_name,
1097                                mea->mea_master, dparent->d_inode->i_ino,
1098                                dparent->d_inode->i_generation, rec->ur_name,
1099                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1100                         GOTO(cleanup, rc = -ERESTART);
1101                 }
1102         }
1103
1104         /* Step 2: Lookup the child */
1105         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1106         if (IS_ERR(dchild)) {
1107                 rc = PTR_ERR(dchild);
1108                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1109                 GOTO(cleanup, rc);
1110         }
1111
1112 got_child:
1113         cleanup_phase = 2; /* child dentry */
1114
1115         if (dchild->d_flags & DCACHE_CROSS_REF) {
1116                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1117                 LASSERT(rec->ur_namelen > 1);
1118
1119                 /* we're gonna acquire LOOKUP lock on the child,
1120                  * but we have already locked parent and our order
1121                  * may conflict with enqueue_order_locks(). so,
1122                  * drop parent lock and acquire both the locks in
1123                  * common order. bug 6190 */
1124 #ifdef S_PDIROPS
1125                 if (parent_lockh[1].cookie != 0)
1126                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1127 #endif
1128                 ldlm_lock_decref(parent_lockh, parent_mode);
1129                 l_dput(dchild);
1130                 l_dput(dparent);
1131                 cleanup_phase = 0;
1132
1133                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1134                                                  parent_lockh, &dparent,
1135                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1136                                                  &update_mode, rec->ur_name,
1137                                                  rec->ur_namelen, child_lockh,
1138                                                  &dchild, LCK_PR,
1139                                                  MDS_INODELOCK_LOOKUP);
1140
1141                 if (rc)
1142                         GOTO(cleanup, rc);
1143
1144 #ifdef S_PDIROPS
1145                 if (parent_lockh[1].cookie != 0)
1146                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1147 #endif
1148                 ldlm_lock_decref(parent_lockh, LCK_PR);
1149
1150                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1151                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1152                                dchild->d_inode);
1153                         if (dchild->d_inode)
1154                                 ldlm_lock_decref(child_lockh, LCK_PR);
1155                         l_dput(dchild);
1156                         l_dput(dparent);
1157                         GOTO(restart, rc);
1158                 }
1159
1160                 mds_pack_dentry2body(obd, body, dchild, 1);
1161                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1162                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1163
1164                 if (mea)
1165                         OBD_FREE(mea, mea_size);
1166                 l_dput(dchild);
1167                 l_dput(dparent);
1168                 RETURN(rc);
1169         }
1170
1171         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1172
1173         if (dchild->d_inode)
1174                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1175         else
1176                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1177
1178         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1179         if (dchild->d_inode == NULL) {
1180                 unsigned long ino;
1181                 struct iattr iattr;
1182                 struct inode *inode;
1183
1184                 /* get parent id: ldlm lock on the parent protects ea */
1185                 rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
1186                 if (rc) {
1187                         CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
1188                                 dparent->d_inode->i_ino, rc);
1189                         GOTO(cleanup, rc);
1190                 }
1191
1192                 ino = (unsigned long)id_ino(rec->ur_id2);
1193
1194                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1195                                           update_mode);
1196
1197                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1198                        obd->obd_name, dparent->d_inode->i_ino,
1199                        dparent->d_inode->i_generation, rc);
1200
1201                 if (rc > 0) {
1202                         /* dir got splitted */
1203                         GOTO(cleanup, rc = -ERESTART);
1204                 } else if (rc < 0) {
1205                         /* error happened during spitting */
1206                         GOTO(cleanup, rc);
1207                 }
1208
1209                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1210                         /*
1211                          * dentry is negative and we weren't supposed to create
1212                          * object, get out of here.
1213                          */
1214                         GOTO(cleanup, rc = -ENOENT);
1215                 }
1216
1217                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1218                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1219                                       NULL);
1220                 if (IS_ERR(handle)) {
1221                         rc = PTR_ERR(handle);
1222                         handle = NULL;
1223                         GOTO(cleanup, rc);
1224                 }
1225                 if (id_fid(rec->ur_id2))
1226                         fid = id_fid(rec->ur_id2);
1227                 else
1228                         fid = mds_alloc_fid(obd);
1229
1230                 dchild->d_fsdata = (void *) &dp;
1231                 dp.p_ptr = req;
1232                 dp.p_inum = ino;
1233
1234                 dp.p_fid = fid;
1235                 dp.p_group = mds->mds_num;
1236
1237                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1238                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1239                         dchild->d_fsdata = NULL;
1240
1241                 if (rc) {
1242                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1243                         GOTO(cleanup, rc);
1244                 }
1245
1246                 inode = dchild->d_inode;
1247
1248                 if (ino) {
1249                         LASSERT(ino == inode->i_ino);
1250
1251                         /* written as part of setattr */
1252                         inode->i_generation = id_gen(rec->ur_id2);
1253                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1254                                inode->i_ino, inode->i_generation);
1255                 }
1256
1257                 created = 1;
1258                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1259                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1260                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1261
1262                 iattr.ia_uid = rec->ur_fsuid;
1263                 if (dparent->d_inode->i_mode & S_ISGID)
1264                         iattr.ia_gid = dparent->d_inode->i_gid;
1265                 else
1266                         iattr.ia_gid = rec->ur_fsgid;
1267
1268                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1269                         ATTR_MTIME | ATTR_CTIME;
1270
1271                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1272                 if (rc)
1273                         CERROR("error on child setattr: rc = %d\n", rc);
1274
1275                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1276
1277                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1278                 if (rc)
1279                         CERROR("error on parent setattr: rc = %d\n", rc);
1280                 else {
1281                         MD_COUNTER_INCREMENT(obd, create);
1282                 }
1283
1284                 mds_inode2id(obd, &body->id1, inode, fid);
1285                 mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
1286
1287                 if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) {
1288                         rc = mds_set_gskey(obd, handle, dchild->d_inode,
1289                                            rec->ur_ea2data, rec->ur_ea2datalen,
1290                                            ATTR_KEY | ATTR_MAC);
1291                         if (rc) {
1292                                 CERROR("error in set gs key rc %d\n", rc);
1293                         }
1294                 }
1295
1296                 rc = fsfilt_commit(obd, dchild->d_inode->i_sb, dchild->d_inode, handle, 0);
1297                 handle = NULL;
1298                 acc_mode = 0;           /* Don't check for permissions */
1299         }
1300         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1301
1302         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1303                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1304                  dchild->d_name.name, dchild, dchild->d_inode);
1305
1306         if (S_ISREG(dchild->d_inode->i_mode)) {
1307                 /* Check permissions etc */
1308                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1309                 if (rc != 0)
1310                         GOTO(cleanup, rc);
1311
1312                 /* Can't write to a read-only file */
1313                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1314                         GOTO(cleanup, rc = -EPERM);
1315
1316                 /* An append-only file must be opened in append mode for
1317                  * writing */
1318                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1319                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1320                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1321                         GOTO(cleanup, rc = -EPERM);
1322         }
1323
1324         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1325             (rec->ur_flags & MDS_OPEN_EXCL)) {
1326                 /* File already exists, we didn't just create it, and we
1327                  * were passed O_EXCL; err-or. */
1328                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1329         }
1330
1331         if (S_ISDIR(dchild->d_inode->i_mode)) {
1332                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1333                     rec->ur_flags & FMODE_WRITE) {
1334                         /* we are trying to create or write a exist dir */
1335                         GOTO(cleanup, rc = -EISDIR);
1336                 }
1337                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1338                         GOTO(cleanup, rc = -EACCES);
1339
1340                 /*
1341                  * here was checking for possible GNS mount points to skip their
1342                  * open. I removed it as its detection based only on SUID bit is
1343                  * not reliable. Opening such a dirs should not cause any
1344                  * problems, at least tests show that. --umka
1345                  */
1346         }
1347
1348         /* if we are following a symlink, don't open */
1349         if (S_ISLNK(dchild->d_inode->i_mode))
1350                 GOTO(cleanup_no_trans, rc = 0);
1351
1352         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1353             !S_ISDIR(dchild->d_inode->i_mode))
1354                 GOTO(cleanup, rc = -ENOTDIR);
1355
1356         /* check permission even it's special files */
1357         if (S_ISCHR(dchild->d_inode->i_mode) ||
1358             S_ISBLK(dchild->d_inode->i_mode) ||
1359             S_ISFIFO(dchild->d_inode->i_mode) ||
1360             S_ISSOCK(dchild->d_inode->i_mode)) {
1361                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1362                 if (rc != 0)
1363                         GOTO(cleanup, rc);
1364         }
1365
1366         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1367                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1368                 GOTO(cleanup, rc = -EAGAIN);
1369         }
1370
1371 #if 0
1372         /* We cannot use acc_mode here, because it is zeroed in case of
1373            creating a file, so we get wrong lockmode */
1374         if (accmode(rec->ur_flags) & MAY_WRITE)
1375                 child_mode = LCK_CW;
1376         else if (accmode(rec->ur_flags) & MAY_EXEC)
1377                 child_mode = LCK_PR;
1378         else
1379                 child_mode = LCK_CR;
1380
1381         /* Always returning LOOKUP lock if open succesful to guard
1382          * dentry on client. In case of replay we do not get a lock
1383          * assuming that the caller has it already */
1384         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1385                 struct ldlm_res_id child_res_id = { .name = {0}};
1386                 ldlm_policy_data_t policy;
1387                 int lock_flags = LDLM_FL_ATOMIC_CB;
1388
1389                 /* LOOKUP lock will protect dentry on client -bzzz */
1390                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1391                                                 MDS_INODELOCK_OPEN;
1392
1393                 child_res_id.name[0] = id_fid(&body->id1);
1394                 child_res_id.name[1] = id_group(&body->id1);
1395
1396                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1397                                       child_res_id, LDLM_IBITS, &policy,
1398                                       child_mode, &lock_flags,
1399                                       mds_blocking_ast, ldlm_completion_ast,
1400                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1401                 if (rc != ELDLM_OK)
1402                         GOTO(cleanup, rc);
1403
1404                 cleanup_phase = 3;
1405         }
1406 #else
1407 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1408 #endif
1409
1410         /* Step 5: mds_open it */
1411         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1412                              rec, rep, &cancel_update_lock);
1413         if (rc)
1414                 GOTO(cleanup, rc);
1415
1416         /* reintegration case */
1417         if ((rec->ur_flags & MDS_REINT_REQ)) {
1418                 rc = mds_fidmap_add(obd, &body->id1);
1419                 if (rc < 0) {
1420                         CERROR("can't create fid->ino mapping, err %d\n",
1421                                rc);
1422                 } else {
1423                         rc = 0;
1424                 }
1425         }
1426
1427         /* if this is a writer, we have to invalidate client's
1428          * update locks in order to make sure they don't use
1429          * isize/iblocks from mds anymore.
1430          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1431          * XXX: optimization is to do this for first writer only */
1432         if ((accmode(rec->ur_flags) & MAY_WRITE) || cancel_update_lock) {
1433                 struct ldlm_res_id child_res_id = { .name = {0}};
1434                 ldlm_policy_data_t sz_policy;
1435                 struct lustre_handle sz_lockh;
1436                 int lock_flags = LDLM_FL_ATOMIC_CB;
1437
1438                 child_res_id.name[0] = id_fid(&body->id1);
1439                 child_res_id.name[1] = id_group(&body->id1);
1440                 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1441
1442                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1443                                       child_res_id, LDLM_IBITS, &sz_policy,
1444                                       LCK_PW, &lock_flags, mds_blocking_ast,
1445                                       ldlm_completion_ast, NULL, NULL, NULL,
1446                                       0, NULL, &sz_lockh);
1447                 if (rc == ELDLM_OK)
1448                         ldlm_lock_decref(&sz_lockh, LCK_PW);
1449                 else
1450                         CERROR("can't invalidate client's update locks\n");
1451         }
1452
1453         EXIT;
1454 cleanup:
1455         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1456                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1457
1458 cleanup_no_trans:
1459         switch (cleanup_phase) {
1460         case 3:
1461                 if (rc) {
1462                         ldlm_lock_decref(child_lockh, child_mode);
1463                         child_lockh->cookie = 0;
1464                 }
1465         case 2:
1466                 if (rc && created) {
1467                         int err = vfs_unlink(dparent->d_inode, dchild);
1468                         if (err) {
1469                                 CERROR("unlink(%.*s) in error path: %d\n",
1470                                        dchild->d_name.len, dchild->d_name.name,
1471                                        err);
1472                         }
1473                 } else if (created) {
1474                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1475                 }
1476                 /* audit stuff for OPEN */
1477                 if (offset == 3) {
1478                         mds_audit(req, dchild, rec->ur_name,
1479                                   rec->ur_namelen - 1, AUDIT_OPEN, rc);
1480                 }
1481
1482                 l_dput(dchild);
1483         case 1:
1484                 if (dparent == NULL)
1485                         break;
1486
1487                 l_dput(dparent);
1488 #ifdef S_PDIROPS
1489                 if (parent_lockh[1].cookie != 0)
1490                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1491 #endif
1492                 if (rc)
1493                         ldlm_lock_decref(parent_lockh, parent_mode);
1494                 else
1495                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1496         }
1497         if (mea)
1498                 OBD_FREE(mea, mea_size);
1499         if (rc == 0)
1500                 atomic_inc(&mds->mds_open_count);
1501
1502         mds_body_do_reverse_map(med, body);
1503
1504         /*
1505          * If we have not taken the "open" lock, we may not return 0 here,
1506          * because caller expects 0 to mean "lock is taken", and it needs
1507          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1508          * client. Later caller should rewrite the return value back to zero
1509          * if it to be used any further.
1510          */
1511         if ((cleanup_phase != 3) && !rc)
1512                 rc = ENOLCK;
1513
1514         RETURN(rc);
1515 }
1516
1517 /* Close a "file descriptor" and possibly unlink an orphan from the
1518  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1519  *
1520  * If we are being called from mds_disconnect() because the client has
1521  * disappeared, then req == NULL and we do not update last_rcvd because
1522  * there is nothing that could be recovered by the client at this stage
1523  * (it will not even _have_ an entry in last_rcvd anymore).
1524  *
1525  * Returns EAGAIN if the client needs to get more data and re-close. */
1526 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1527                   struct obd_device *obd, struct mds_file_data *mfd,
1528                   int unlink_orphan)
1529 {
1530         struct inode *inode = mfd->mfd_dentry->d_inode;
1531         char idname[LL_ID_NAMELEN];
1532         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1533         struct dentry *pending_child = NULL;
1534         struct mds_obd *mds = &obd->u.mds;
1535         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1536         void *handle = NULL;
1537         struct mds_body *request_body = NULL, *reply_body = NULL;
1538         struct dentry_params dp;
1539         struct iattr iattr = { 0 };
1540         struct llog_create_locks *lcl = NULL;
1541         ENTRY;
1542
1543         if (req && req->rq_reqmsg != NULL)
1544                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1545                                               sizeof(*request_body));
1546         if (req && req->rq_repmsg != NULL)
1547                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1548                                             sizeof(*reply_body));
1549
1550         if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
1551                 /* we set i_size/i_blocks here, nobody will see
1552                  * them until all write references are dropped.
1553                  * btw, we hold one reference */
1554                 LASSERT(mfd->mfd_mode & FMODE_WRITE);
1555                 LASSERT(request_body->valid & OBD_MD_FLEPOCH);
1556                 LASSERT(MDS_FILTERDATA(inode));
1557                 if (MDS_FILTERDATA(inode)->io_epoch != request_body->io_epoch)
1558                         CDEBUG(D_ERROR, "try to update attr. for old epoch "
1559                                LPD64" while current "LPD64"\n",
1560                                MDS_FILTERDATA(inode)->io_epoch,
1561                                request_body->io_epoch);
1562                 i_size_write(inode, request_body->size);
1563                 inode->i_blocks = request_body->blocks;
1564                 LTIME_S(inode->i_mtime) = (request_body->mtime);
1565
1566                 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1567                 iattr.ia_size = inode->i_size;
1568                 iattr.ia_valid |= ATTR_SIZE|ATTR_MTIME;
1569                 mds_inode_unset_attrs_old(inode);
1570         }
1571
1572         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1573         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1574                idname, inode->i_nlink, mds_orphan_open_count(inode));
1575
1576         last_orphan = mds_orphan_open_dec_test(inode) &&
1577                 mds_inode_is_orphan(inode);
1578         UP_WRITE_I_ALLOC_SEM(inode);
1579
1580         /* this is half of the actual "close" */
1581         if (mfd->mfd_mode & FMODE_WRITE) {
1582                 rc = mds_put_write_access(mds, inode, request_body,
1583                                           last_orphan && unlink_orphan);
1584         } else if (mfd->mfd_mode & FMODE_EXEC) {
1585                 mds_allow_write_access(inode);
1586         }
1587
1588         if (last_orphan && unlink_orphan) {
1589                 struct lov_mds_md *lmm = NULL;
1590                 int stripe_count = 0;
1591                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1592
1593                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1594
1595                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1596                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1597                         CERROR("found \"orphan\" %s %s with link count %d\n",
1598                                S_ISREG(inode->i_mode) ? "file" : "dir",
1599                                idname, inode->i_nlink);
1600                 /*
1601                  * sadly, there is no easy way to save pending_child from
1602                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1603                  * normally it will still be in the dcache.
1604                  */
1605                 down(&pending_dir->i_sem);
1606                 cleanup_phase = 1; /* up(i_sem) when finished */
1607                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1608                                                idlen);
1609                 if (IS_ERR(pending_child))
1610                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1611                 if (pending_child->d_inode == NULL) {
1612                         CERROR("orphan %s has been removed\n", idname);
1613                         GOTO(cleanup, rc = 0);
1614                 }
1615
1616                 cleanup_phase = 2; /* dput(pending_child) when finished */
1617                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1618                         rc = vfs_rmdir(pending_dir, pending_child);
1619                         if (rc)
1620                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1621                                        idname, rc);
1622                         goto out;
1623                 }
1624
1625                 if (req != NULL && req->rq_repmsg != NULL) {
1626                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1627                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1628                 }
1629
1630                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1631                                           NULL, stripe_count);
1632                 if (IS_ERR(handle)) {
1633                         rc = PTR_ERR(handle);
1634                         handle = NULL;
1635                         GOTO(cleanup, rc);
1636                 }
1637
1638                 pending_child->d_fsdata = (void *) &dp;
1639                 dp.p_inum = 0;
1640                 dp.p_ptr = req;
1641                 rc = vfs_unlink(pending_dir, pending_child);
1642                 if (rc)
1643                         CERROR("error unlinking orphan %s: rc %d\n",
1644                                idname, rc);
1645
1646                 if (req != NULL && req->rq_repmsg != NULL &&
1647                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1648                     mds_log_op_unlink(obd, pending_child->d_inode,
1649                                       lmm, req->rq_repmsg->buflens[1],
1650                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1651                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1652                         reply_body->valid |= OBD_MD_FLCOOKIE;
1653                 }
1654
1655                 rc = mds_destroy_object(obd, inode, 1);
1656                 if (rc) {
1657                         CERROR("cannot destroy OSS object on close, err %d\n",
1658                                rc);
1659                         rc = 0;
1660                 }
1661
1662                 goto out; /* Don't bother updating attrs on unlinked inode */
1663         } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
1664                 /* last writer closed file - let's update i_size/i_blocks */
1665                 mds_validate_size(obd, inode, request_body, NULL, &iattr);
1666         }
1667
1668 #if 0
1669         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1670                 /* Update the on-disk attributes if this was the last write
1671                  * close, and all information was provided (i.e., rc == 0)
1672                  *
1673                  * XXX this should probably be abstracted with mds_reint_setattr
1674                  */
1675
1676                 if (request_body->valid & OBD_MD_FLMTIME &&
1677                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1678                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1679                         iattr.ia_valid |= ATTR_MTIME;
1680                 }
1681                 if (request_body->valid & OBD_MD_FLCTIME &&
1682                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1683                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1684                         iattr.ia_valid |= ATTR_CTIME;
1685                 }
1686
1687                 /* XXX can't set block count with fsfilt_setattr (!) */
1688                 if (request_body->valid & OBD_MD_FLSIZE) {
1689                         iattr.ia_valid |= ATTR_SIZE;
1690                         iattr.ia_size = request_body->size;
1691                 }
1692                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1693                         iattr.ia_valid |= ATTR_BLOCKS;
1694                         iattr.ia_blocks = request_body->blocks
1695                 } */
1696
1697         }
1698 #endif
1699         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1700                 /* Only start a transaction to write out only the atime if
1701                  * it is more out-of-date than the specified limit.  If we
1702                  * are already going to write out the atime then do it anyway.
1703                  * */
1704                 if ((request_body->atime >
1705                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1706                     (iattr.ia_valid != 0 &&
1707                      request_body->atime > LTIME_S(inode->i_atime))) {
1708                         LTIME_S(iattr.ia_atime) = request_body->atime;
1709                         iattr.ia_valid |= ATTR_ATIME;
1710                 }
1711         }
1712
1713         if (iattr.ia_valid != 0) {
1714                 if (handle == NULL)
1715                         handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1716                 if (IS_ERR(handle))
1717                         GOTO(cleanup, rc = PTR_ERR(handle));
1718                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1719                 if (rc)
1720                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1721         }
1722 out:
1723         /* If other clients have this file open for write, rc will be > 0 */
1724         if (rc > 0)
1725                 rc = 0;
1726         if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
1727                         && !mds_inode_is_orphan(inode)
1728                         && atomic_read(&inode->i_writecount) == 0
1729                         && inode->i_nlink != 0) {
1730                 CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
1731                        inode->i_ino, inode->i_generation, inode->i_nlink);
1732         }
1733         l_dput(mfd->mfd_dentry);
1734         mds_mfd_destroy(mfd);
1735
1736  cleanup:
1737         atomic_dec(&mds->mds_open_count);
1738         if (req != NULL && reply_body != NULL) {
1739                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1740         } else if (handle) {
1741                 int err, force_sync = 0;
1742
1743                 if (req && req->rq_export)
1744                         force_sync = req->rq_export->exp_sync;
1745
1746                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1747                                     force_sync);
1748                 if (err) {
1749                         CERROR("error committing close: %d\n", err);
1750                         if (!rc)
1751                                 rc = err;
1752                 }
1753         }
1754
1755         switch (cleanup_phase) {
1756         case 2:
1757                 if (lcl != NULL)
1758                         ptlrpc_save_llog_lock(req, lcl);
1759                 dput(pending_child);
1760         case 1:
1761                 up(&pending_dir->i_sem);
1762         }
1763         RETURN(rc);
1764 }
1765
1766 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1767                                     struct ldlm_lock_desc *new, void *data,
1768                                     int flag)
1769 {
1770         struct lustre_handle lockh = { 0 };
1771         int rc;
1772         ENTRY;
1773
1774         switch (flag) {
1775         case LDLM_CB_BLOCKING:
1776                 ldlm_lock2handle(lock, &lockh);
1777                 rc = ldlm_cli_cancel(&lockh);
1778                 if (rc != ELDLM_OK)
1779                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
1780                 break;
1781         case LDLM_CB_CANCELING: {
1782                 break;
1783         }
1784         default:
1785                 LBUG();
1786         }
1787
1788         RETURN(0);
1789 }
1790 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1791 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1792 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
1793
1794 int mds_validate_size(struct obd_device *obd, struct inode *inode,
1795                       struct mds_body *body, struct lov_stripe_md *lsm,
1796                       struct iattr *iattr)
1797 {
1798         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1799         struct lustre_handle lockh = { 0 };
1800         int rc, len = 0, flags;
1801         void *lmm = NULL;
1802         ENTRY;
1803
1804         /* we update i_size/i_blocks only for regular files */
1805         if (!S_ISREG(inode->i_mode))
1806                 RETURN(0);
1807
1808         /* we shouldn't fetch size from OSTes during recovery - deadlock */
1809         if (obd->obd_recovering) {
1810                 CERROR("size-on-mds has no support on OST yet\n");
1811                 RETURN(0);
1812         }
1813
1814         /* if nobody modified attrs. we're lucky */
1815         if (!mds_inode_has_old_attrs(inode))
1816                 RETURN(0);
1817
1818         /* 1: client didn't send actual i_size/i_blocks
1819          * 2: we seem to be last writer
1820          * 3: the file doesn't look like to be disappeared
1821          * conclusion: we're gonna fetch them from OSSes */
1822
1823         if (lsm)
1824                 goto have_lsm;
1825
1826         down(&inode->i_sem);
1827         len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1828         up(&inode->i_sem);
1829
1830         if (len < 0) {
1831                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1832                 GOTO(cleanup, rc = len);
1833         } else if (len == 0) {
1834                 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1835                 GOTO(cleanup, rc = 0);
1836         }
1837
1838         OBD_ALLOC(lmm, len);
1839         if (lmm == NULL) {
1840                 CERROR("can't allocate memory\n");
1841                 GOTO(cleanup, rc = -ENOMEM);
1842         }
1843
1844         down(&inode->i_sem);
1845         rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1846         up(&inode->i_sem);
1847
1848         if (rc < 0) {
1849                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1850                 GOTO(cleanup, rc);
1851         }
1852
1853         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1854         if (rc < 0) {
1855                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1856                 GOTO(cleanup, rc);
1857         }
1858
1859  have_lsm:
1860
1861         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1862
1863         flags = LDLM_FL_HAS_INTENT;
1864         rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1865                          LCK_PR, &flags, mds_extent_lock_callback,
1866                          ldlm_completion_ast, NULL, NULL,
1867                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1868         if (rc == -ENOENT) {
1869                 /* while we were enqueueing lock on OST, another thread
1870                  * unlinked the file and started OST object destoying.
1871                  * it's safe to return 0 here */
1872                 GOTO(cleanup, rc = 0);
1873         } else if (rc != 0) {
1874                 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1875                 GOTO(cleanup, rc);
1876         }
1877
1878         if (body)
1879                 CDEBUG(D_INODE, ""DLID4"\n", OLID4(&body->id1));
1880
1881         CDEBUG(D_INODE, "LOV reports "LPD64"/%lu [%s%s%s]\n",
1882                inode->i_size, inode->i_blocks,
1883                atomic_read(&inode->i_writecount) > 1 ? "U" : "",
1884                mds_inode_has_old_attrs(inode) ? "D" : "",
1885                mds_inode_is_orphan(inode) ? "O" : "");
1886
1887         i_size_write(inode, lov_merge_size(lsm, 0));
1888         inode->i_blocks = lov_merge_blocks(lsm);
1889         LTIME_S(inode->i_mtime) = lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1890         iattr->ia_size = inode->i_size;
1891         LTIME_S(iattr->ia_mtime) = LTIME_S(inode->i_mtime);
1892         iattr->ia_valid |= ATTR_SIZE | ATTR_MTIME;
1893
1894         DOWN_WRITE_I_ALLOC_SEM(inode);
1895         mds_inode_unset_attrs_old(inode);
1896         UP_WRITE_I_ALLOC_SEM(inode);
1897
1898         obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1899
1900 cleanup:
1901         if (lmm != NULL) {
1902                 if (lsm != NULL)
1903                         obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1904                 OBD_FREE(lmm, len);
1905         }
1906         RETURN(rc);
1907 }
1908
1909 int mds_close(struct ptlrpc_request *req, int offset)
1910 {
1911         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1912         struct obd_device *obd = req->rq_export->exp_obd;
1913         struct mds_body *body;
1914         struct mds_file_data *mfd;
1915         struct lvfs_run_ctxt saved;
1916         struct inode *inode;
1917         int rc, repsize[3] = {sizeof(struct mds_body),
1918                               obd->u.mds.mds_max_mdsize,
1919                               obd->u.mds.mds_max_cookiesize};
1920         ENTRY;
1921         MD_COUNTER_INCREMENT(obd, close);
1922
1923         rc = lustre_pack_reply(req, 3, repsize, NULL);
1924         if (rc) {
1925                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1926                 req->rq_status = rc;
1927         } else {
1928                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1929         }
1930
1931         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1932                 DEBUG_REQ(D_HA, req, "close replay");
1933                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1934                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1935                        req->rq_repmsg->buflens[2]);
1936         }
1937
1938         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1939                                   lustre_swab_mds_body);
1940         if (body == NULL) {
1941                 CERROR("Can't unpack body\n");
1942                 req->rq_status = -EFAULT;
1943                 RETURN(-EFAULT);
1944         }
1945
1946         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1947                 /* do some stuff */ ;
1948
1949         mfd = mds_handle2mfd(&body->handle);
1950         if (mfd == NULL) {
1951                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1952                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1953                 req->rq_status = -ESTALE;
1954                 RETURN(-ESTALE);
1955         }
1956
1957         inode = mfd->mfd_dentry->d_inode;
1958
1959         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1960         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1961
1962         if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
1963                 /* the client modified data through the handle
1964                  * we need to care about attrs. -bzzz */
1965                 mds_inode_set_attrs_old(inode);
1966         }
1967
1968         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1969                 struct mds_body *rep_body;
1970
1971                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1972                                           sizeof (*rep_body));
1973                 LASSERT(rep_body != NULL);
1974
1975                 mds_pack_inode2body(obd, rep_body, inode,
1976                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1977
1978                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1979                             inode, MDS_PACK_MD_LOCK, 0);
1980         }
1981         spin_lock(&med->med_open_lock);
1982         list_del(&mfd->mfd_list);
1983         spin_unlock(&med->med_open_lock);
1984
1985         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1986         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1987         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1988
1989         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1990                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1991                 req->rq_status = -ENOMEM;
1992                 mds_mfd_put(mfd);
1993                 RETURN(-ENOMEM);
1994         }
1995
1996         mds_mfd_put(mfd);
1997         RETURN(0);
1998 }
1999
2000
2001 int mds_done_writing(struct ptlrpc_request *req, int offset)
2002 {
2003         struct mds_body *body;
2004         int rc, size = sizeof(struct mds_body);
2005         ENTRY;
2006
2007         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2008
2009         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
2010                                   lustre_swab_mds_body);
2011         if (body == NULL) {
2012                 CERROR("Can't unpack body\n");
2013                 req->rq_status = -EFAULT;
2014                 RETURN(-EFAULT);
2015         }
2016
2017         rc = lustre_pack_reply(req, 1, &size, NULL);
2018         if (rc) {
2019                 CERROR("lustre_pack_reply: rc = %d\n", rc);
2020                 req->rq_status = rc;
2021         }
2022
2023         RETURN(0);
2024 }