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
  47 #include "mds_internal.h"
  48
  49 /* Exported function from this file are:
  50  *
  51  * mds_open - called by the intent handler
  52  * mds_close - an rpc handling function
  53  * mds_pin - an rpc handling function - which will go away
  54  * mds_mfd_close - for force closing files when a client dies
  55  */
  56
  57 /*
  58  * MDS file data handling: file data holds a handle for a file opened
  59  * by a client.
  60  */
  61
  62 static void mds_mfd_addref(void *mfdp)
  63 {
  64         struct mds_file_data *mfd = mfdp;
  65
  66         atomic_inc(&mfd->mfd_refcount);
  67         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  68                atomic_read(&mfd->mfd_refcount));
  69 }
  70
  71 struct mds_file_data *mds_mfd_new(void)
  72 {
  73         struct mds_file_data *mfd;
  74
  75         OBD_ALLOC(mfd, sizeof *mfd);
  76         if (mfd == NULL) {
  77                 CERROR("mds: out of memory\n");
  78                 return NULL;
  79         }
  80
  81         atomic_set(&mfd->mfd_refcount, 2);
  82
  83         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  84         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  85
  86         return mfd;
  87 }
  88
  89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  90 {
  91         ENTRY;
  92         LASSERT(handle != NULL);
  93         RETURN(class_handle2object(handle->cookie));
  94 }
  95
  96 static void mds_mfd_put(struct mds_file_data *mfd)
  97 {
  98         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
  99                atomic_read(&mfd->mfd_refcount) - 1);
 100         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 101                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 102         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 103                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 104                 OBD_FREE(mfd, sizeof *mfd);
 105         }
 106 }
 107
 108 static void mds_mfd_destroy(struct mds_file_data *mfd)
 109 {
 110         class_handle_unhash(&mfd->mfd_handle);
 111         mds_mfd_put(mfd);
 112 }
 113
 114 /* Caller must hold mds->mds_epoch_sem */
 115 static int mds_alloc_filterdata(struct inode *inode)
 116 {
 117         LASSERT(LUSTRE_FILTERDATA(inode) == NULL);
 118         OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 119         if (LUSTRE_FILTERDATA(inode) == NULL)
 120                 return -ENOMEM;
 121         LASSERT(igrab(inode) == inode);
 122         return 0;
 123 }
 124
 125 /* Caller must hold mds->mds_epoch_sem */
 126 static void mds_free_filterdata(struct inode *inode)
 127 {
 128         LASSERT(LUSTRE_FILTERDATA(inode) != NULL);
 129         OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
 130         LUSTRE_FILTERDATA(inode) = NULL;
 131         iput(inode);
 132 }
 133
 134 /* Write access to a file: executors cause a negative count,
 135  * writers a positive count.  The semaphore is needed to perform
 136  * a check for the sign and then increment or decrement atomically.
 137  *
 138  * This code is closely tied to the allocation of the d_fsdata and the
 139  * MDS epoch, so we use the same semaphore for the whole lot.
 140  *
 141  * We could use a different semaphore for each file, if it ever shows
 142  * up in a profile, which it won't.
 143  *
 144  * epoch argument is nonzero during recovery */
 145 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 146                                 __u64 epoch)
 147 {
 148         int rc = 0;
 149
 150         down(&mds->mds_epoch_sem);
 151
 152         if (atomic_read(&inode->i_writecount) < 0) {
 153                 up(&mds->mds_epoch_sem);
 154                 RETURN(-ETXTBSY);
 155         }
 156
 157         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 158                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 159                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 160                        inode->i_generation);
 161                 goto out;
 162         }
 163
 164         if (MDS_FILTERDATA(inode) == NULL)
 165                 mds_alloc_filterdata(inode);
 166         if (MDS_FILTERDATA(inode) == NULL) {
 167                 rc = -ENOMEM;
 168                 goto out;
 169         }
 170         if (epoch > mds->mds_io_epoch)
 171                 mds->mds_io_epoch = epoch;
 172         else
 173                 mds->mds_io_epoch++;
 174         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 175         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 176                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 177  out:
 178         if (rc == 0)
 179                 atomic_inc(&inode->i_writecount);
 180         up(&mds->mds_epoch_sem);
 181         return rc;
 182 }
 183
 184 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 185  * again -- which is never true if the file is about to be unlinked.  Otherwise
 186  * returns the number of remaining writers. */
 187 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 188                                 struct mds_body *body, int unlinking)
 189 {
 190         int rc = 0;
 191         ENTRY;
 192
 193         down(&mds->mds_epoch_sem);
 194         atomic_dec(&inode->i_writecount);
 195         rc = atomic_read(&inode->i_writecount);
 196         if (rc > 0)
 197                 GOTO(out, rc);
 198 #if 0
 199         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 200                 GOTO(out, rc = EAGAIN);
 201 #endif
 202         mds_free_filterdata(inode);
 203  out:
 204         up(&mds->mds_epoch_sem);
 205         return rc;
 206 }
 207
 208 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 209 {
 210         ENTRY;
 211         down(&mds->mds_epoch_sem);
 212         if (atomic_read(&inode->i_writecount) > 0) {
 213                 up(&mds->mds_epoch_sem);
 214                 RETURN(-ETXTBSY);
 215         }
 216         atomic_dec(&inode->i_writecount);
 217         up(&mds->mds_epoch_sem);
 218         RETURN(0);
 219 }
 220
 221 static void mds_allow_write_access(struct inode *inode)
 222 {
 223         ENTRY;
 224         atomic_inc(&inode->i_writecount);
 225 }
 226
 227 int mds_query_write_access(struct inode *inode)
 228 {
 229         ENTRY;
 230         RETURN(atomic_read(&inode->i_writecount));
 231 }
 232
 233 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 234 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 235                                              struct vfsmount *mnt, int flags,
 236                                              struct ptlrpc_request *req)
 237 {
 238         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 239         struct mds_obd *mds = mds_req2mds(req);
 240         struct mds_file_data *mfd;
 241         struct mds_body *body;
 242         int rc = 0;
 243         ENTRY;
 244
 245         mfd = mds_mfd_new();
 246         if (mfd == NULL) {
 247                 CERROR("mds: out of memory\n");
 248                 GOTO(cleanup_dentry, rc = -ENOMEM);
 249         }
 250
 251         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 252
 253         if (flags & FMODE_WRITE) {
 254                 /* FIXME: in recovery, need to pass old epoch here */
 255                 rc = mds_get_write_access(mds, dentry->d_inode, 0);
 256                 if (rc)
 257                         GOTO(cleanup_mfd, rc);
 258                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 259         } else if (flags & FMODE_EXEC) {
 260                 rc = mds_deny_write_access(mds, dentry->d_inode);
 261                 if (rc)
 262                         GOTO(cleanup_mfd, rc);
 263         }
 264
 265         dget(dentry);
 266
 267         /* FIXME: invalidate update locks when first open for write comes in */
 268
 269         /* mark the file as open to handle open-unlink. */
 270         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 271         mds_orphan_open_inc(dentry->d_inode);
 272         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 273
 274         mfd->mfd_mode = flags;
 275         mfd->mfd_dentry = dentry;
 276         mfd->mfd_xid = req->rq_xid;
 277
 278         spin_lock(&med->med_open_lock);
 279         list_add(&mfd->mfd_list, &med->med_open_head);
 280         spin_unlock(&med->med_open_lock);
 281         mds_mfd_put(mfd);
 282
 283         body->handle.cookie = mfd->mfd_handle.h_cookie;
 284         RETURN(mfd);
 285 cleanup_mfd:
 286         mds_mfd_put(mfd);
 287         mds_mfd_destroy(mfd);
 288 cleanup_dentry:
 289         return ERR_PTR(rc);
 290 }
 291
 292 static inline void
 293 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 294                     struct lov_desc *desc)
 295 {
 296         int i;
 297
 298         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 299                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 300                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 301         }
 302 }
 303
 304 /* must be called with i_sem held */
 305 int
 306 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
 307                    int offset, struct mds_update_record *rec,
 308                    struct dentry *dchild, void **handle,
 309                    obd_id **ids)
 310 {
 311         struct inode *inode = dchild->d_inode;
 312         struct mds_obd *mds = &obd->u.mds;
 313         struct obd_trans_info oti = { 0 };
 314         struct lov_stripe_md *lsm = NULL;
 315         struct lov_mds_md *lmm = NULL;
 316         int rc, lmm_bufsize, lmm_size;
 317         struct obdo *oa = NULL;
 318         struct mds_body *body;
 319         void *lmm_buf;
 320         ENTRY;
 321
 322         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 323             !(rec->ur_flags & FMODE_WRITE))
 324                 RETURN(0);
 325
 326         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 327
 328         if (!S_ISREG(inode->i_mode))
 329                 RETURN(0);
 330         if (body->valid & OBD_MD_FLEASIZE)
 331                 RETURN(0);
 332
 333         OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 334         if (*ids == NULL)
 335                 RETURN(-ENOMEM);
 336         oti.oti_objid = *ids;
 337
 338         /* replay case */
 339         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 340                 LASSERT(id_ino(rec->ur_id2));
 341                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 342                 lmm_size = rec->ur_eadatalen;
 343                 lmm = rec->ur_eadata;
 344                 LASSERT(lmm);
 345
 346                 if (*handle == NULL)
 347                         *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 348                 if (IS_ERR(*handle)) {
 349                         rc = PTR_ERR(*handle);
 350                         *handle = NULL;
 351                         RETURN(rc);
 352                 }
 353
 354                 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 355
 356                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 357                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 358                 LASSERT(lmm_buf != NULL);
 359                 LASSERT(lmm_bufsize >= lmm_size);
 360
 361                 memcpy(lmm_buf, lmm, lmm_size);
 362                 rc = fsfilt_set_md(obd, inode, *handle, lmm,
 363                                    lmm_size, EA_LOV);
 364                 if (rc)
 365                         CERROR("open replay failed to set md:%d\n", rc);
 366                 RETURN(0);
 367         }
 368
 369         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 370                 GOTO(out_ids, rc = -ENOMEM);
 371
 372         oa = obdo_alloc();
 373         if (oa == NULL)
 374                 RETURN(-ENOMEM);
 375         oa->o_mode = S_IFREG | 0600;
 376         oa->o_id = inode->i_ino;
 377         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 378         oa->o_generation = inode->i_generation;
 379         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 380         oa->o_gid = 0;
 381         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
 382                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 383         oa->o_size = 0;
 384
 385         obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 386                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 387
 388         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 389                 /* check if things like lfs setstripe are sending us the ea */
 390                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 391                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 392                                            mds->mds_dt_exp,
 393                                            0, &lsm, rec->ur_eadata);
 394                         if (rc)
 395                                 GOTO(out_oa, rc);
 396                 } else {
 397                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 398                         if (lmm == NULL)
 399                                 GOTO(out_oa, rc = -ENOMEM);
 400
 401                         lmm_size = mds->mds_max_mdsize;
 402                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 403                                         lmm, &lmm_size, 1, 0);
 404                         if (rc > 0)
 405                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 406                                                    mds->mds_dt_exp,
 407                                                    0, &lsm, lmm);
 408                         OBD_FREE(lmm, mds->mds_max_mdsize);
 409                         if (rc)
 410                                 GOTO(out_oa, rc);
 411                 }
 412
 413                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 414                 oti.oti_flags |= OBD_MODE_CROW;
 415                 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
 416
 417                 if (rc) {
 418                         CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
 419                                "error creating objects for "
 420                                "inode %lu: rc = %d\n",
 421                                inode->i_ino, rc);
 422                         if (rc > 0) {
 423                                 CERROR("obd_create returned invalid "
 424                                        "rc %d\n", rc);
 425                                 rc = -EIO;
 426                         }
 427                         GOTO(out_oa, rc);
 428                 }
 429         } else {
 430                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 431                                    0, &lsm, rec->ur_eadata);
 432                 if (rc)
 433                         GOTO(out_oa, rc);
 434
 435                 lsm->lsm_object_id = oa->o_id;
 436                 lsm->lsm_object_gr = oa->o_gr;
 437         }
 438         if (inode->i_size) {
 439                 oa->o_size = inode->i_size;
 440                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 441                                 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 442
 443                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
 444                 if (rc) {
 445                         CERROR("error setting attrs for inode %lu: rc %d\n",
 446                                inode->i_ino, rc);
 447                         if (rc > 0) {
 448                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 449                                 rc = -EIO;
 450                         }
 451                         GOTO(out_oa, rc);
 452                 }
 453         }
 454
 455         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 456         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 457
 458         LASSERT(lsm && lsm->lsm_object_id);
 459         lmm = NULL;
 460         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 461         if (!id_ino(rec->ur_id2))
 462                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 463         if (rc < 0) {
 464                 CERROR("cannot pack lsm, err = %d\n", rc);
 465                 GOTO(out_oa, rc);
 466         }
 467
 468         lmm_size = rc;
 469         body->eadatasize = rc;
 470
 471         if (*handle == NULL)
 472                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 473         if (IS_ERR(*handle)) {
 474                 rc = PTR_ERR(*handle);
 475                 *handle = NULL;
 476                 GOTO(out_oa, rc);
 477         }
 478
 479         rc = fsfilt_set_md(obd, inode, *handle, lmm,
 480                            lmm_size, EA_LOV);
 481
 482         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 483         lmm_bufsize = req->rq_repmsg->buflens[offset];
 484         LASSERT(lmm_buf);
 485         LASSERT(lmm_bufsize >= lmm_size);
 486
 487         memcpy(lmm_buf, lmm, lmm_size);
 488         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 489 out_oa:
 490         oti_free_cookies(&oti);
 491         obdo_free(oa);
 492 out_ids:
 493         if (rc) {
 494                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 495                 *ids = NULL;
 496         }
 497         if (lsm)
 498                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 499         RETURN(rc);
 500 }
 501
 502 int
 503 mds_destroy_object(struct obd_device *obd,
 504                    struct inode *inode, int async)
 505 {
 506         struct mds_obd *mds = &obd->u.mds;
 507         struct lov_mds_md *lmm = NULL;
 508         int rc, lmm_size;
 509         ENTRY;
 510
 511         LASSERT(inode != NULL);
 512
 513         if (inode->i_nlink != 0) {
 514                 CDEBUG(D_INODE, "attempt to destroy OSS object when "
 515                        "i_nlink == %d\n", (int)inode->i_nlink);
 516                 RETURN(0);
 517         }
 518
 519         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 520         if (lmm == NULL)
 521                 RETURN(-ENOMEM);
 522
 523         lmm_size = mds->mds_max_mdsize;
 524         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
 525         if (rc < 0) {
 526                 CERROR("no stripe info for %lu/%lu inode\n",
 527                        (unsigned long)inode->i_ino,
 528                        (unsigned long)inode->i_generation);
 529                 GOTO(out_free_lmm, rc);
 530         }
 531
 532         if (rc > 0) {
 533                 /* asynchronously unlink objecect on OSS */
 534                 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
 535                                        NULL, 0, async);
 536                 if (rc) {
 537                         CERROR("error unlinking object on OSS, "
 538                                "err %d\n", rc);
 539                         GOTO(out_free_lmm, rc);
 540                 }
 541         } else {
 542                 CDEBUG(D_INODE, "no stripping info found for inode "
 543                       "%lu/%lu\n", (unsigned long)inode->i_ino,
 544                       (unsigned long)inode->i_generation);
 545         }
 546         EXIT;
 547 out_free_lmm:
 548         OBD_FREE(lmm, mds->mds_max_mdsize);
 549         return rc;
 550 }
 551
 552 static void reconstruct_open(struct mds_update_record *rec, int offset,
 553                              struct ptlrpc_request *req,
 554                              struct lustre_handle *child_lockh)
 555 {
 556         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 557         struct mds_client_data *mcd = med->med_mcd;
 558         struct mds_obd *mds = mds_req2mds(req);
 559         struct mds_file_data *mfd;
 560         struct obd_device *obd = req->rq_export->exp_obd;
 561         struct dentry *parent = NULL, *dchild;
 562         struct ldlm_reply *rep;
 563         struct mds_body *body;
 564         struct list_head *t;
 565         int put_child = 1;
 566         int rc;
 567         ENTRY;
 568
 569         LASSERT(offset == 3); /* only called via intent */
 570         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 571         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 572
 573         /* copy rc, transno and disp; steal locks */
 574         mds_req_from_mcd(req, mcd);
 575         intent_set_disposition(rep, mcd->mcd_last_data);
 576
 577         /* Only replay if create or open actually happened. */
 578         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 579                 EXIT;
 580                 return; /* error looking up parent or child */
 581         }
 582
 583         if (rec->ur_namelen == 1) {
 584                 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
 585                        OLID4(rec->ur_id1));
 586                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 587         } else {
 588                 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 589                 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 590                                 OLID4(rec->ur_id1), PTR_ERR(parent));
 591                 dchild = ll_lookup_one_len(rec->ur_name, parent,
 592                                 rec->ur_namelen - 1);
 593                 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 594                                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 595         }
 596
 597         if (!dchild->d_inode)
 598                 GOTO(out_dput, 0); /* child not present to open */
 599
 600         /* At this point, we know we have a child. We'll send
 601          * it back _unless_ it not created and open failed.  */
 602         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 603             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 604             req->rq_status)
 605                 GOTO(out_dput, 0);
 606
 607         /* get lock (write for O_CREAT, read otherwise) */
 608         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 609         if (S_ISREG(dchild->d_inode->i_mode)) {
 610                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 611                                  dchild->d_inode, 1, 0);
 612
 613                 if (rc)
 614                         LASSERT(rc == req->rq_status);
 615
 616                 /* If we have LOV EA data, the OST holds size, mtime */
 617                 if (!(body->valid & OBD_MD_FLEASIZE))
 618                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 619                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 620         }
 621
 622         /* If we have -EEXIST as the status, and we were asked to create
 623          * exclusively, we can tell we failed because the file already existed.
 624          */
 625         if (req->rq_status == -EEXIST &&
 626             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 627              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 628                 GOTO(out_dput, 0);
 629         }
 630
 631         /* If we didn't get as far as trying to open, then some locking thing
 632          * probably went wrong, and we'll just bail here.
 633          */
 634         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 635                 GOTO(out_dput, 0);
 636
 637         /* If we failed, then we must have failed opening, so don't look for
 638          * file descriptor or anything, just give the client the bad news.
 639          */
 640         if (req->rq_status)
 641                 GOTO(out_dput, 0);
 642
 643         mfd = NULL;
 644         list_for_each(t, &med->med_open_head) {
 645                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 646                 if (mfd->mfd_xid == req->rq_xid)
 647                         break;
 648                 mfd = NULL;
 649         }
 650
 651         /* #warning "XXX fixme" bug 2991 */
 652         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 653          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 654          * to detect a re-open */
 655         if (mfd == NULL) {
 656                 mntget(mds->mds_vfsmnt);
 657                 CERROR("Re-opened file \n");
 658                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 659                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 660                 if (!mfd) {
 661                         CERROR("mds: out of memory\n");
 662                         GOTO(out_dput, req->rq_status = -ENOMEM);
 663                 }
 664                 put_child = 0;
 665         } else {
 666                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 667                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 668                        mfd->mfd_handle.h_cookie);
 669         }
 670
 671  out_dput:
 672         if (put_child)
 673                 l_dput(dchild);
 674         if (parent)
 675                 l_dput(parent);
 676         EXIT;
 677 }
 678
 679 /* do NOT or the MAY_*'s, you'll get the weakest */
 680 static int accmode(int flags)
 681 {
 682         int res = 0;
 683
 684         if (flags & FMODE_READ)
 685                 res = MAY_READ;
 686         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 687                 res |= MAY_WRITE;
 688         if (flags & FMODE_EXEC)
 689                 res = MAY_EXEC;
 690         return res;
 691 }
 692
 693 /* Handles object creation, actual opening, and I/O epoch */
 694 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 695                            struct mds_body *body, int flags, void **handle,
 696                            struct mds_update_record *rec, struct ldlm_reply *rep)
 697 {
 698         struct obd_device *obd = req->rq_export->exp_obd;
 699         struct mds_obd *mds = mds_req2mds(req);
 700         struct mds_file_data *mfd = NULL;
 701         obd_id *ids = NULL;
 702         unsigned mode;
 703         int rc = 0;
 704         ENTRY;
 705
 706         /* atomically create objects if necessary */
 707         down(&dchild->d_inode->i_sem);
 708         mode = dchild->d_inode->i_mode;
 709
 710         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 711             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 712                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 713                                  dchild->d_inode, 0, 0);
 714                 if (rc) {
 715                         up(&dchild->d_inode->i_sem);
 716                         RETURN(rc);
 717                 }
 718         }
 719
 720         if (rec != NULL) {
 721                 if ((body->valid & OBD_MD_FLEASIZE) &&
 722                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 723                         up(&dchild->d_inode->i_sem);
 724                         RETURN(-EEXIST);
 725                 }
 726
 727                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 728                         /* no EA: create objects */
 729                         rc = mds_create_objects(obd, req, 2, rec,
 730                                                 dchild, handle, &ids);
 731                         if (rc) {
 732                                 CERROR("mds_create_object: rc = %d\n", rc);
 733                                 up(&dchild->d_inode->i_sem);
 734                                 RETURN(rc);
 735                         }
 736                 }
 737
 738                 if (S_ISREG(dchild->d_inode->i_mode) &&
 739                     (body->valid & OBD_MD_FLEASIZE)) {
 740                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 741                                                    req->rq_repmsg, 2);
 742                         if (!rc)
 743                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 744                                                  dchild->d_inode, 0, 0);
 745                         if (rc) {
 746                                 up(&dchild->d_inode->i_sem);
 747                                 RETURN(rc);
 748                         }
 749                 }
 750         }
 751         rc = mds_pack_acl(req, 3, body, dchild->d_inode);
 752         if (rc < 0) {
 753                 CERROR("pack posix acl: rc = %d\n", rc);
 754                 up(&dchild->d_inode->i_sem);
 755                 RETURN(rc);
 756         }
 757
 758         /* If the inode has no EA data, then MDSs hold size, mtime */
 759         if (S_ISREG(dchild->d_inode->i_mode) &&
 760             !(body->valid & OBD_MD_FLEASIZE)) {
 761                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 762                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 763         }
 764
 765         up(&dchild->d_inode->i_sem);
 766
 767         intent_set_disposition(rep, DISP_OPEN_OPEN);
 768         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 769         if (IS_ERR(mfd))
 770                 RETURN(PTR_ERR(mfd));
 771
 772         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 773                mfd->mfd_handle.h_cookie);
 774
 775         if (ids != NULL) {
 776                 mds_dt_update_objids(obd, ids);
 777                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 778         }
 779
 780         RETURN(rc);
 781 }
 782
 783 static int mds_open_by_id(struct ptlrpc_request *req,
 784                           struct lustre_id *id,
 785                           struct mds_body *body, int flags,
 786                           struct mds_update_record *rec,
 787                           struct ldlm_reply *rep)
 788 {
 789         struct mds_obd *mds = mds_req2mds(req);
 790         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 791         struct dentry *dchild;
 792         char idname[LL_ID_NAMELEN];
 793         int idlen = 0, rc;
 794         void *handle = NULL;
 795         ENTRY;
 796
 797         down(&pending_dir->i_sem);
 798         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 799         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 800                                 idlen);
 801         if (IS_ERR(dchild)) {
 802                 up(&pending_dir->i_sem);
 803                 rc = PTR_ERR(dchild);
 804                 CERROR("error looking up %s in PENDING: rc = %d\n",
 805                        idname, rc);
 806                 RETURN(rc);
 807         }
 808
 809         up(&pending_dir->i_sem);
 810         if (dchild->d_inode != NULL) {
 811                 mds_inode_set_orphan(dchild->d_inode);
 812                 mds_pack_inode2body(req2obd(req), body,
 813                                     dchild->d_inode, 1);
 814
 815                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 816                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 817                 CWARN("Orphan %s found and opened in PENDING directory\n",
 818                        idname);
 819                 goto open;
 820         }
 821         l_dput(dchild);
 822
 823         /*
 824          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 825          * regular inode that was previously created.
 826          */
 827         dchild = mds_id2dentry(req2obd(req), id, NULL);
 828         if (IS_ERR(dchild))
 829                 RETURN(PTR_ERR(dchild));
 830
 831         mds_pack_inode2body(req2obd(req), body,
 832                             dchild->d_inode, 1);
 833
 834         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 835         intent_set_disposition(rep, DISP_LOOKUP_POS);
 836
 837  open:
 838         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 839         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 840                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 841         /* XXX what do we do here if mds_finish_transno itself failed? */
 842         l_dput(dchild);
 843         RETURN(rc);
 844 }
 845
 846 int mds_pin(struct ptlrpc_request *req, int offset)
 847 {
 848         struct obd_device *obd = req->rq_export->exp_obd;
 849         struct mds_body *request_body, *reply_body;
 850         struct lvfs_run_ctxt saved;
 851         int rc, size = sizeof(*reply_body);
 852         ENTRY;
 853
 854         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 855                                       sizeof(*request_body));
 856
 857         rc = lustre_pack_reply(req, 1, &size, NULL);
 858         if (rc)
 859                 RETURN(rc);
 860         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 861
 862         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 863         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 864                             request_body->flags, NULL, NULL);
 865         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 866
 867         RETURN(rc);
 868 }
 869
 870 /*
 871  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 872  * child_lockh is NULL we just get the lock as a barrier to wait for other
 873  * holders of this lock, and drop it right away again.
 874  */
 875 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 876                        struct lustre_handle *child_lockh)
 877 {
 878         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 879         struct lustre_handle lockh;
 880         int lock_flags = LDLM_FL_ATOMIC_CB;
 881         int rc;
 882         ENTRY;
 883
 884         if (child_lockh == NULL)
 885                 child_lockh = &lockh;
 886
 887         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 888                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 889                               mds_blocking_ast, ldlm_completion_ast, NULL,
 890                               NULL, NULL, 0, NULL, child_lockh);
 891         if (rc != ELDLM_OK)
 892                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 893         else if (child_lockh == &lockh)
 894                 ldlm_lock_decref(child_lockh, LCK_EX);
 895
 896         RETURN(rc);
 897 }
 898
 899 int mds_open(struct mds_update_record *rec, int offset,
 900              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 901 {
 902         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 903         struct obd_device *obd = req->rq_export->exp_obd;
 904         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
 905         struct mds_obd *mds = mds_req2mds(req);
 906         struct ldlm_reply *rep = NULL;
 907         struct mds_body *body = NULL;
 908         struct dentry *dchild = NULL, *dparent = NULL;
 909         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 910         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 911         int parent_mode = LCK_PR;
 912         void *handle = NULL;
 913         struct dentry_params dp;
 914         struct mea *mea = NULL;
 915         int mea_size, update_mode;
 916         int child_mode = LCK_PR;
 917         __u64 fid = 0;
 918         ENTRY;
 919
 920         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 921                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 922                   rec->ur_mode);
 923
 924         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 925                          (obd_timeout + 1) / 4);
 926
 927         if (offset == 3) { /* intent */
 928                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 929                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 930         } else if (offset == 1) { /* non-intent reint */
 931                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 932         } else {
 933                 body = NULL;
 934                 LBUG();
 935         }
 936
 937         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 938
 939         /*
 940          * Step 0: If we are passed a id, then we assume the client already
 941          * opened this file and is only replaying the RPC, so we open the inode
 942          * by fid (at some large expense in security).
 943          */
 944         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 945                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 946                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 947                        rec->ur_name, rec->ur_mode);
 948
 949                 LASSERT(id_ino(rec->ur_id2));
 950
 951                 rc = mds_open_by_id(req, rec->ur_id2, body,
 952                                     rec->ur_flags, rec, rep);
 953                 if (rc != -ENOENT) {
 954                         mds_body_do_reverse_map(med, body);
 955                         RETURN(rc);
 956                 }
 957
 958                 /* We didn't find the correct inode on disk either, so we
 959                  * need to re-create it via a regular replay. */
 960                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 961         } else {
 962                 LASSERT(!id_ino(rec->ur_id2));
 963         }
 964
 965         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 966
 967         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 968                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 969                 RETURN(-ENOMEM);
 970         }
 971
 972         acc_mode = accmode(rec->ur_flags);
 973
 974         /* Step 1: Find and lock the parent */
 975         if (rec->ur_flags & MDS_OPEN_CREAT) {
 976                 /* XXX Well, in fact we only need this lock mode change if
 977                    in addition to O_CREAT, the file does not exist.
 978                    But we do not know if it exists or not yet */
 979                 parent_mode = LCK_PW;
 980         }
 981
 982         if (rec->ur_namelen == 1) {
 983                 /* client (LMV) wants to open the file by id */
 984                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 985                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 986                 if (IS_ERR(dchild)) {
 987                         rc = PTR_ERR(dparent);
 988                         CERROR("child lookup by an id error %d\n", rc);
 989                         GOTO(cleanup, rc);
 990                 }
 991                 goto got_child;
 992         }
 993
 994 restart:
 995         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
 996                                        parent_lockh, &update_mode, rec->ur_name,
 997                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 998         if (IS_ERR(dparent)) {
 999                 rc = PTR_ERR(dparent);
1000                 if (rc != -ENOENT) {
1001                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1002                                OLID4(rec->ur_id1), rc);
1003                 } else {
1004                         /* Just cannot find parent - make it look like
1005                            usual negative lookup to avoid extra MDS RPC */
1006                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1007                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1008                 }
1009                 GOTO(cleanup, rc);
1010         }
1011         LASSERT(dparent->d_inode != NULL);
1012
1013         cleanup_phase = 1; /* parent dentry and lock */
1014
1015         /* try to retrieve MEA data for this dir */
1016         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1017         if (rc)
1018                 GOTO(cleanup, rc);
1019
1020         if (mea != NULL) {
1021                 /*
1022                  * dir is already splitted, check is requested filename should *
1023                  * live at this MDS or at another one.
1024                  */
1025                 int i;
1026                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1027                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1028                         CDEBUG(D_OTHER,
1029                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1030                                " should be %lu(%d)\n", obd->obd_name,
1031                                mea->mea_master, dparent->d_inode->i_ino,
1032                                dparent->d_inode->i_generation, rec->ur_name,
1033                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1034                         GOTO(cleanup, rc = -ERESTART);
1035                 }
1036         }
1037
1038         /* Step 2: Lookup the child */
1039         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1040         if (IS_ERR(dchild)) {
1041                 rc = PTR_ERR(dchild);
1042                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1043                 GOTO(cleanup, rc);
1044         }
1045
1046 got_child:
1047         cleanup_phase = 2; /* child dentry */
1048
1049         if (dchild->d_flags & DCACHE_CROSS_REF) {
1050                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1051                 LASSERT(rec->ur_namelen > 1);
1052
1053                 /* we're gonna acquire LOOKUP lock on the child,
1054                  * but we have already locked parent and our order
1055                  * may conflict with enqueue_order_locks(). so,
1056                  * drop parent lock and acquire both the locks in
1057                  * common order. bug 6190 */
1058 #ifdef S_PDIROPS
1059                 if (parent_lockh[1].cookie != 0)
1060                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1061 #endif
1062                 ldlm_lock_decref(parent_lockh, parent_mode);
1063                 l_dput(dchild);
1064                 l_dput(dparent);
1065                 cleanup_phase = 0;
1066
1067                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1068                                                  parent_lockh, &dparent,
1069                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1070                                                  &update_mode, rec->ur_name,
1071                                                  rec->ur_namelen, child_lockh,
1072                                                  &dchild, LCK_PR,
1073                                                  MDS_INODELOCK_LOOKUP);
1074
1075                 if (rc)
1076                         GOTO(cleanup, rc);
1077
1078 #ifdef S_PDIROPS
1079                 if (parent_lockh[1].cookie != 0)
1080                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1081 #endif
1082                 ldlm_lock_decref(parent_lockh, LCK_PR);
1083
1084                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1085                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1086                                dchild->d_inode);
1087                         if (dchild->d_inode)
1088                                 ldlm_lock_decref(child_lockh, LCK_PR);
1089                         l_dput(dchild);
1090                         l_dput(dparent);
1091                         GOTO(restart, rc);
1092                 }
1093
1094                 mds_pack_dentry2body(obd, body, dchild, 1);
1095                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1096                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1097
1098                 if (mea)
1099                         OBD_FREE(mea, mea_size);
1100                 l_dput(dchild);
1101                 l_dput(dparent);
1102                 RETURN(rc);
1103         }
1104
1105         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1106
1107         if (dchild->d_inode)
1108                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1109         else
1110                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1111
1112         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1113         if (dchild->d_inode == NULL) {
1114                 unsigned long ino;
1115                 struct iattr iattr;
1116                 struct inode *inode;
1117
1118                 ino = (unsigned long)id_ino(rec->ur_id2);
1119
1120                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1121                                           update_mode);
1122
1123                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1124                        obd->obd_name, dparent->d_inode->i_ino,
1125                        dparent->d_inode->i_generation, rc);
1126
1127                 if (rc > 0) {
1128                         /* dir got splitted */
1129                         GOTO(cleanup, rc = -ERESTART);
1130                 } else if (rc < 0) {
1131                         /* error happened during spitting */
1132                         GOTO(cleanup, rc);
1133                 }
1134
1135                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1136                         /*
1137                          * dentry is negative and we weren't supposed to create
1138                          * object, get out of here.
1139                          */
1140                         GOTO(cleanup, rc = -ENOENT);
1141                 }
1142
1143                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1144                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1145                                       NULL);
1146                 if (IS_ERR(handle)) {
1147                         rc = PTR_ERR(handle);
1148                         handle = NULL;
1149                         GOTO(cleanup, rc);
1150                 }
1151                 fid = mds_alloc_fid(obd);
1152                 dchild->d_fsdata = (void *) &dp;
1153                 dp.p_ptr = req;
1154                 dp.p_inum = ino;
1155                 dp.p_fid = fid;
1156                 dp.p_group = mds->mds_num;
1157
1158                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1159                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1160                         dchild->d_fsdata = NULL;
1161
1162                 if (rc) {
1163                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1164                         GOTO(cleanup, rc);
1165                 }
1166                 inode = dchild->d_inode;
1167                 if (ino) {
1168                         LASSERT(ino == inode->i_ino);
1169
1170                         /* written as part of setattr */
1171                         inode->i_generation = id_gen(rec->ur_id2);
1172                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1173                                inode->i_ino, inode->i_generation);
1174                 }
1175
1176                 created = 1;
1177                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1178                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1179                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1180
1181                 iattr.ia_uid = rec->ur_fsuid;
1182                 if (dparent->d_inode->i_mode & S_ISGID)
1183                         iattr.ia_gid = dparent->d_inode->i_gid;
1184                 else
1185                         iattr.ia_gid = rec->ur_fsgid;
1186
1187                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1188                         ATTR_MTIME | ATTR_CTIME;
1189
1190                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1191                 if (rc)
1192                         CERROR("error on child setattr: rc = %d\n", rc);
1193
1194                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1195
1196                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1197                 if (rc)
1198                         CERROR("error on parent setattr: rc = %d\n", rc);
1199                 else {
1200                         MD_COUNTER_INCREMENT(obd, create);
1201                 }
1202
1203                 if (ino) {
1204                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1205                                                   handle, rec->ur_id2);
1206                         if (rc) {
1207                                 CERROR("mds_update_inode_sid() failed, "
1208                                        "rc = %d\n", rc);
1209                         }
1210                         id_assign_fid(&body->id1, rec->ur_id2);
1211
1212                         /*
1213                          * make sure, that fid is up-to-date.
1214                          */
1215                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1216                 } else {
1217                         rc = mds_set_inode_sid(obd, dchild->d_inode,
1218                                                handle, &body->id1, fid);
1219                         if (rc) {
1220                                 CERROR("mds_set_inode_sid() failed, "
1221                                        "rc = %d\n", rc);
1222                         }
1223                 }
1224
1225                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1226                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1227                                            dchild->d_inode, handle,
1228                                            req->rq_export->exp_sync);
1229                         handle = NULL;
1230                 }
1231
1232                 acc_mode = 0;           /* Don't check for permissions */
1233         }
1234         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1235
1236         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1237                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1238                  dchild->d_name.name, dchild, dchild->d_inode);
1239
1240         if (S_ISREG(dchild->d_inode->i_mode)) {
1241                 /* Check permissions etc */
1242                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1243                 if (rc != 0)
1244                         GOTO(cleanup, rc);
1245
1246                 /* Can't write to a read-only file */
1247                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1248                         GOTO(cleanup, rc = -EPERM);
1249
1250                 /* An append-only file must be opened in append mode for
1251                  * writing */
1252                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1253                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1254                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1255                         GOTO(cleanup, rc = -EPERM);
1256         }
1257
1258         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1259             (rec->ur_flags & MDS_OPEN_EXCL)) {
1260                 /* File already exists, we didn't just create it, and we
1261                  * were passed O_EXCL; err-or. */
1262                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1263         }
1264
1265         if (S_ISDIR(dchild->d_inode->i_mode)) {
1266                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1267                     rec->ur_flags & FMODE_WRITE) {
1268                         /* we are trying to create or write a exist dir */
1269                         GOTO(cleanup, rc = -EISDIR);
1270                 }
1271                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1272                         GOTO(cleanup, rc = -EACCES);
1273
1274                 /*
1275                  * here was checking for possible GNS mount points to skip their
1276                  * open. I removed it as its detection based only on SUID bit is
1277                  * not reliable. Opening such a dirs should not cause any
1278                  * problems, at least tests show that. --umka
1279                  */
1280         }
1281
1282         /* if we are following a symlink, don't open */
1283         if (S_ISLNK(dchild->d_inode->i_mode))
1284                 GOTO(cleanup_no_trans, rc = 0);
1285
1286         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1287             !S_ISDIR(dchild->d_inode->i_mode))
1288                 GOTO(cleanup, rc = -ENOTDIR);
1289
1290         /* check permission even it's special files */
1291         if (S_ISCHR(dchild->d_inode->i_mode) ||
1292             S_ISBLK(dchild->d_inode->i_mode) ||
1293             S_ISFIFO(dchild->d_inode->i_mode) ||
1294             S_ISSOCK(dchild->d_inode->i_mode)) {
1295                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1296                 if (rc != 0)
1297                         GOTO(cleanup, rc);
1298         }
1299
1300         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1301                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1302                 GOTO(cleanup, rc = -EAGAIN);
1303         }
1304
1305 #if 0
1306         /* We cannot use acc_mode here, because it is zeroed in case of
1307            creating a file, so we get wrong lockmode */
1308         if (accmode(rec->ur_flags) & MAY_WRITE)
1309                 child_mode = LCK_CW;
1310         else if (accmode(rec->ur_flags) & MAY_EXEC)
1311                 child_mode = LCK_PR;
1312         else
1313                 child_mode = LCK_CR;
1314
1315         /* Always returning LOOKUP lock if open succesful to guard
1316          * dentry on client. In case of replay we do not get a lock
1317          * assuming that the caller has it already */
1318         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1319                 struct ldlm_res_id child_res_id = { .name = {0}};
1320                 ldlm_policy_data_t policy;
1321                 int lock_flags = LDLM_FL_ATOMIC_CB;
1322
1323                 /* LOOKUP lock will protect dentry on client -bzzz */
1324                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1325                                                 MDS_INODELOCK_OPEN;
1326
1327                 child_res_id.name[0] = id_fid(&body->id1);
1328                 child_res_id.name[1] = id_group(&body->id1);
1329
1330                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1331                                       child_res_id, LDLM_IBITS, &policy,
1332                                       child_mode, &lock_flags,
1333                                       mds_blocking_ast, ldlm_completion_ast,
1334                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1335                 if (rc != ELDLM_OK)
1336                         GOTO(cleanup, rc);
1337
1338                 cleanup_phase = 3;
1339         }
1340 #else
1341 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1342 #warning "disable opencache lock for CMD2"
1343 #endif
1344
1345         /* Step 5: mds_open it */
1346         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1347                              rec, rep);
1348         if (rc)
1349                 GOTO(cleanup, rc);
1350
1351         /* if this is a writer, we have to invalidate client's
1352          * update locks in order to make sure they don't use
1353          * isize/iblocks from mds anymore.
1354          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1355          * XXX: optimization is to do this for first writer only */
1356         if (accmode(rec->ur_flags) & MAY_WRITE) {
1357                 struct ldlm_res_id child_res_id = { .name = {0}};
1358                 ldlm_policy_data_t sz_policy;
1359                 struct lustre_handle sz_lockh;
1360                 int lock_flags = 0;
1361
1362                 child_res_id.name[0] = id_fid(&body->id1);
1363                 child_res_id.name[1] = id_group(&body->id1);
1364                 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1365
1366                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1367                                       child_res_id, LDLM_IBITS, &sz_policy,
1368                                       LCK_PW, &lock_flags, mds_blocking_ast,
1369                                       ldlm_completion_ast, NULL, NULL, NULL,
1370                                       0, NULL, &sz_lockh);
1371                 if (rc == ELDLM_OK)
1372                         ldlm_lock_decref(&sz_lockh, LCK_PW);
1373                 else
1374                         CERROR("can't invalidate client's update locks\n");
1375         }
1376
1377         EXIT;
1378 cleanup:
1379         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1380                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1381
1382 cleanup_no_trans:
1383         switch (cleanup_phase) {
1384         case 3:
1385                 if (rc) {
1386                         ldlm_lock_decref(child_lockh, child_mode);
1387                         child_lockh->cookie = 0;
1388                 }
1389         case 2:
1390                 if (rc && created) {
1391                         int err = vfs_unlink(dparent->d_inode, dchild);
1392                         if (err) {
1393                                 CERROR("unlink(%.*s) in error path: %d\n",
1394                                        dchild->d_name.len, dchild->d_name.name,
1395                                        err);
1396                         }
1397                 } else if (created) {
1398                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1399                 }
1400                 l_dput(dchild);
1401         case 1:
1402                 if (dparent == NULL)
1403                         break;
1404
1405                 l_dput(dparent);
1406 #ifdef S_PDIROPS
1407                 if (parent_lockh[1].cookie != 0)
1408                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1409 #endif
1410                 if (rc)
1411                         ldlm_lock_decref(parent_lockh, parent_mode);
1412                 else
1413                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1414         }
1415         if (mea)
1416                 OBD_FREE(mea, mea_size);
1417         if (rc == 0)
1418                 atomic_inc(&mds->mds_open_count);
1419
1420         mds_body_do_reverse_map(med, body);
1421
1422         /*
1423          * If we have not taken the "open" lock, we may not return 0 here,
1424          * because caller expects 0 to mean "lock is taken", and it needs
1425          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1426          * client. Later caller should rewrite the return value back to zero
1427          * if it to be used any further.
1428          */
1429         if ((cleanup_phase != 3) && !rc)
1430                 rc = ENOLCK;
1431
1432         RETURN(rc);
1433 }
1434
1435 /* Close a "file descriptor" and possibly unlink an orphan from the
1436  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1437  *
1438  * If we are being called from mds_disconnect() because the client has
1439  * disappeared, then req == NULL and we do not update last_rcvd because
1440  * there is nothing that could be recovered by the client at this stage
1441  * (it will not even _have_ an entry in last_rcvd anymore).
1442  *
1443  * Returns EAGAIN if the client needs to get more data and re-close. */
1444 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1445                   struct obd_device *obd, struct mds_file_data *mfd,
1446                   int unlink_orphan)
1447 {
1448         struct inode *inode = mfd->mfd_dentry->d_inode;
1449         char idname[LL_ID_NAMELEN];
1450         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1451         struct dentry *pending_child = NULL;
1452         struct mds_obd *mds = &obd->u.mds;
1453         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1454         void *handle = NULL;
1455         struct mds_body *request_body = NULL, *reply_body = NULL;
1456         struct dentry_params dp;
1457         struct iattr iattr = { 0 };
1458         struct llog_create_locks *lcl = NULL;
1459         ENTRY;
1460
1461         if (req && req->rq_reqmsg != NULL)
1462                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1463                                               sizeof(*request_body));
1464         if (req && req->rq_repmsg != NULL)
1465                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1466                                             sizeof(*reply_body));
1467
1468         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1469         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1470                idname, inode->i_nlink, mds_orphan_open_count(inode));
1471
1472         last_orphan = mds_orphan_open_dec_test(inode) &&
1473                 mds_inode_is_orphan(inode);
1474         UP_WRITE_I_ALLOC_SEM(inode);
1475
1476         /* this is half of the actual "close" */
1477         if (mfd->mfd_mode & FMODE_WRITE) {
1478                 rc = mds_put_write_access(mds, inode, request_body,
1479                                           last_orphan && unlink_orphan);
1480         } else if (mfd->mfd_mode & FMODE_EXEC) {
1481                 mds_allow_write_access(inode);
1482         }
1483
1484         if (last_orphan && unlink_orphan) {
1485                 struct lov_mds_md *lmm = NULL;
1486                 int stripe_count = 0;
1487                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1488
1489                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1490
1491                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1492                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1493                         CERROR("found \"orphan\" %s %s with link count %d\n",
1494                                S_ISREG(inode->i_mode) ? "file" : "dir",
1495                                idname, inode->i_nlink);
1496                 /*
1497                  * sadly, there is no easy way to save pending_child from
1498                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1499                  * normally it will still be in the dcache.
1500                  */
1501                 down(&pending_dir->i_sem);
1502                 cleanup_phase = 1; /* up(i_sem) when finished */
1503                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1504                                                idlen);
1505                 if (IS_ERR(pending_child))
1506                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1507                 if (pending_child->d_inode == NULL) {
1508                         CERROR("orphan %s has been removed\n", idname);
1509                         GOTO(cleanup, rc = 0);
1510                 }
1511
1512                 cleanup_phase = 2; /* dput(pending_child) when finished */
1513                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1514                         rc = vfs_rmdir(pending_dir, pending_child);
1515                         if (rc)
1516                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1517                                        idname, rc);
1518                         goto out;
1519                 }
1520
1521                 if (req != NULL && req->rq_repmsg != NULL) {
1522                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1523                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1524                 }
1525
1526                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1527                                           NULL, stripe_count);
1528                 if (IS_ERR(handle)) {
1529                         rc = PTR_ERR(handle);
1530                         handle = NULL;
1531                         GOTO(cleanup, rc);
1532                 }
1533
1534                 pending_child->d_fsdata = (void *) &dp;
1535                 dp.p_inum = 0;
1536                 dp.p_ptr = req;
1537                 rc = vfs_unlink(pending_dir, pending_child);
1538                 if (rc)
1539                         CERROR("error unlinking orphan %s: rc %d\n",
1540                                idname, rc);
1541
1542                 if (req != NULL && req->rq_repmsg != NULL &&
1543                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1544                     mds_log_op_unlink(obd, pending_child->d_inode,
1545                                       lmm, req->rq_repmsg->buflens[1],
1546                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1547                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1548                         reply_body->valid |= OBD_MD_FLCOOKIE;
1549                 }
1550
1551                 rc = mds_destroy_object(obd, inode, 1);
1552                 if (rc) {
1553                         CERROR("cannot destroy OSS object on close, err %d\n",
1554                                rc);
1555                         rc = 0;
1556                 }
1557
1558                 goto out; /* Don't bother updating attrs on unlinked inode */
1559         } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0 && request_body) {
1560                 /* last writer closed file - let's update i_size/i_blocks */
1561                 if (request_body->valid & OBD_MD_FLSIZE) {
1562                         LASSERT(request_body->valid & OBD_MD_FLBLOCKS);
1563                         CDEBUG(D_OTHER, "update size "LPD64" for "DLID4
1564                                ", epoch "LPD64"\n", inode->i_size,
1565                                OLID4(&request_body->id1),
1566                                request_body->io_epoch);
1567                         iattr.ia_size = inode->i_size;
1568                         iattr.ia_valid |= ATTR_SIZE;
1569                 }
1570         }
1571
1572 #if 0
1573         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1574                 /* Update the on-disk attributes if this was the last write
1575                  * close, and all information was provided (i.e., rc == 0)
1576                  *
1577                  * XXX this should probably be abstracted with mds_reint_setattr
1578                  */
1579
1580                 if (request_body->valid & OBD_MD_FLMTIME &&
1581                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1582                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1583                         iattr.ia_valid |= ATTR_MTIME;
1584                 }
1585                 if (request_body->valid & OBD_MD_FLCTIME &&
1586                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1587                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1588                         iattr.ia_valid |= ATTR_CTIME;
1589                 }
1590
1591                 /* XXX can't set block count with fsfilt_setattr (!) */
1592                 if (request_body->valid & OBD_MD_FLSIZE) {
1593                         iattr.ia_valid |= ATTR_SIZE;
1594                         iattr.ia_size = request_body->size;
1595                 }
1596                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1597                         iattr.ia_valid |= ATTR_BLOCKS;
1598                         iattr.ia_blocks = request_body->blocks
1599                 } */
1600
1601         }
1602 #endif
1603         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1604                 /* Only start a transaction to write out only the atime if
1605                  * it is more out-of-date than the specified limit.  If we
1606                  * are already going to write out the atime then do it anyway.
1607                  * */
1608                 if ((request_body->atime >
1609                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1610                     (iattr.ia_valid != 0 &&
1611                      request_body->atime > LTIME_S(inode->i_atime))) {
1612                         LTIME_S(iattr.ia_atime) = request_body->atime;
1613                         iattr.ia_valid |= ATTR_ATIME;
1614                 }
1615         }
1616
1617         if (iattr.ia_valid != 0) {
1618                 if (handle == NULL)
1619                         handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1620                 if (IS_ERR(handle))
1621                         GOTO(cleanup, rc = PTR_ERR(handle));
1622                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1623                 if (rc)
1624                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1625         }
1626 out:
1627         /* If other clients have this file open for write, rc will be > 0 */
1628         if (rc > 0)
1629                 rc = 0;
1630         l_dput(mfd->mfd_dentry);
1631         mds_mfd_destroy(mfd);
1632
1633  cleanup:
1634         atomic_dec(&mds->mds_open_count);
1635         if (req != NULL && reply_body != NULL) {
1636                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1637         } else if (handle) {
1638                 int err, force_sync = 0;
1639
1640                 if (req && req->rq_export)
1641                         force_sync = req->rq_export->exp_sync;
1642
1643                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1644                                     force_sync);
1645                 if (err) {
1646                         CERROR("error committing close: %d\n", err);
1647                         if (!rc)
1648                                 rc = err;
1649                 }
1650         }
1651
1652         switch (cleanup_phase) {
1653         case 2:
1654                 if (lcl != NULL)
1655                         ptlrpc_save_llog_lock(req, lcl);
1656                 dput(pending_child);
1657         case 1:
1658                 up(&pending_dir->i_sem);
1659         }
1660         RETURN(rc);
1661 }
1662
1663 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1664                                     struct ldlm_lock_desc *new, void *data,
1665                                     int flag)
1666 {
1667         struct lustre_handle lockh = { 0 };
1668         int rc;
1669         ENTRY;
1670
1671         switch (flag) {
1672         case LDLM_CB_BLOCKING:
1673                 ldlm_lock2handle(lock, &lockh);
1674                 rc = ldlm_cli_cancel(&lockh);
1675                 if (rc != ELDLM_OK)
1676                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
1677                 break;
1678         case LDLM_CB_CANCELING: {
1679                 break;
1680         }
1681         default:
1682                 LBUG();
1683         }
1684
1685         RETURN(0);
1686 }
1687 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1688 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1689
1690 int mds_validate_size(struct obd_device *obd, struct mds_body *body,
1691                       struct mds_file_data *mfd)
1692 {
1693         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1694         struct inode *inode = mfd->mfd_dentry->d_inode;
1695         struct lustre_handle lockh = { 0 };
1696         struct lov_stripe_md *lsm = NULL;
1697         int rc, len, flags;
1698         void *lmm = NULL;
1699         ENTRY;
1700
1701         /* we update i_size/i_blocks only for regular files */
1702         if (!S_ISREG(inode->i_mode))
1703                 RETURN(0);
1704
1705         /* we update i_size/i_blocks only for writers */
1706         if (!(mfd->mfd_mode & FMODE_WRITE))
1707                 RETURN(0);
1708
1709         /* we like when client reports actual i_size/i_blocks himself */
1710         if (body->valid & OBD_MD_FLSIZE) {
1711                 LASSERT(body->valid & OBD_MD_FLBLOCKS);
1712                 CDEBUG(D_OTHER, "client reports "LPD64"/"LPD64" for "DLID4"\n",
1713                        body->size, body->blocks, OLID4(&body->id1));
1714                 RETURN(0);
1715         }
1716
1717         /* we shouldn't fetch size from OSTes during recovery - deadlock */
1718         if (obd->obd_recovering)
1719                 RETURN(0);
1720
1721         DOWN_READ_I_ALLOC_SEM(inode);
1722         if (atomic_read(&inode->i_writecount) > 1
1723                         || mds_inode_is_orphan(inode)) {
1724                 /* there is no need to update i_size/i_blocks on orphans.
1725                  * also, if this is not last writer, then it doesn't make
1726                  * sense to fetch i_size/i_blocks from OSSes */
1727                 UP_READ_I_ALLOC_SEM(inode);
1728                 RETURN(0);
1729         }
1730         UP_READ_I_ALLOC_SEM(inode);
1731
1732         /* 1: client didn't send actual i_size/i_blocks
1733          * 2: we seem to be last writer
1734          * 3: the file doesn't look like to be disappeared
1735          * conclusion: we're gonna fetch them from OSSes */
1736
1737         down(&inode->i_sem);
1738         len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1739         up(&inode->i_sem);
1740
1741         if (len < 0) {
1742                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1743                 GOTO(cleanup, rc = len);
1744         } else if (len == 0) {
1745                 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1746                 GOTO(cleanup, rc = 0);
1747         }
1748
1749         OBD_ALLOC(lmm, len);
1750         if (lmm == NULL) {
1751                 CERROR("can't allocate memory\n");
1752                 GOTO(cleanup, rc = -ENOMEM);
1753         }
1754
1755         down(&inode->i_sem);
1756         rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1757         up(&inode->i_sem);
1758
1759         if (rc < 0) {
1760                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1761                 GOTO(cleanup, rc);
1762         }
1763
1764         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1765         if (rc < 0) {
1766                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1767                 GOTO(cleanup, rc);
1768         }
1769
1770         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1771
1772         flags = LDLM_FL_HAS_INTENT;
1773         rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1774                          LCK_PR, &flags, mds_extent_lock_callback,
1775                          ldlm_completion_ast, NULL, NULL,
1776                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1777         if (rc != 0) {
1778                 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1779                 GOTO(cleanup, rc);
1780         }
1781
1782         body->size = lov_merge_size(lsm, 0);
1783         body->blocks = lov_merge_blocks(lsm);
1784         body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1785
1786         CDEBUG(D_OTHER, "LOV reports "LPD64"/"LPD64" for "DLID4"\n",
1787                         body->size, body->blocks, OLID4(&body->id1));
1788
1789         obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1790
1791 cleanup:
1792         if (lsm != NULL)
1793                 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1794         if (lmm != NULL)
1795                 OBD_FREE(lmm, len);
1796         RETURN(rc);
1797 }
1798
1799 int mds_close(struct ptlrpc_request *req, int offset)
1800 {
1801         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1802         struct obd_device *obd = req->rq_export->exp_obd;
1803         struct mds_body *body;
1804         struct mds_file_data *mfd;
1805         struct lvfs_run_ctxt saved;
1806         struct inode *inode;
1807         int rc, repsize[3] = {sizeof(struct mds_body),
1808                               obd->u.mds.mds_max_mdsize,
1809                               obd->u.mds.mds_max_cookiesize};
1810         ENTRY;
1811         MD_COUNTER_INCREMENT(obd, close);
1812
1813         rc = lustre_pack_reply(req, 3, repsize, NULL);
1814         if (rc) {
1815                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1816                 req->rq_status = rc;
1817         } else {
1818                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1819         }
1820
1821         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1822                 DEBUG_REQ(D_HA, req, "close replay");
1823                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1824                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1825                        req->rq_repmsg->buflens[2]);
1826         }
1827
1828         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1829                                   lustre_swab_mds_body);
1830         if (body == NULL) {
1831                 CERROR("Can't unpack body\n");
1832                 req->rq_status = -EFAULT;
1833                 RETURN(-EFAULT);
1834         }
1835
1836         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1837                 /* do some stuff */ ;
1838
1839         mfd = mds_handle2mfd(&body->handle);
1840         if (mfd == NULL) {
1841                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1842                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1843                 req->rq_status = -ESTALE;
1844                 RETURN(-ESTALE);
1845         }
1846
1847         rc = mds_validate_size(obd, body, mfd);
1848         LASSERT(rc == 0);
1849
1850         inode = mfd->mfd_dentry->d_inode;
1851
1852         if (mfd->mfd_mode & FMODE_WRITE) {
1853                 /* we set i_size/i_blocks here, nobody will see
1854                  * them until all write references are dropped.
1855                  * btw, we hold one reference */
1856                 if (body->valid & OBD_MD_FLSIZE)
1857                         i_size_write(inode, body->size);
1858                 if (body->valid & OBD_MD_FLBLOCKS)
1859                         inode->i_blocks = body->blocks;
1860         }
1861
1862         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1863         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1864         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1865                 struct mds_body *rep_body;
1866
1867                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1868                                           sizeof (*rep_body));
1869                 LASSERT(rep_body != NULL);
1870
1871                 mds_pack_inode2body(obd, rep_body, inode,
1872                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1873
1874                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1875                             inode, MDS_PACK_MD_LOCK, 0);
1876         }
1877         spin_lock(&med->med_open_lock);
1878         list_del(&mfd->mfd_list);
1879         spin_unlock(&med->med_open_lock);
1880
1881         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1882         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1883         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1884
1885         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1886                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1887                 req->rq_status = -ENOMEM;
1888                 mds_mfd_put(mfd);
1889                 RETURN(-ENOMEM);
1890         }
1891
1892         mds_mfd_put(mfd);
1893         RETURN(0);
1894 }
1895
1896 int mds_done_writing(struct ptlrpc_request *req, int offset)
1897 {
1898         struct mds_body *body;
1899         int rc, size = sizeof(struct mds_body);
1900         ENTRY;
1901
1902         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1903
1904         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1905                                   lustre_swab_mds_body);
1906         if (body == NULL) {
1907                 CERROR("Can't unpack body\n");
1908                 req->rq_status = -EFAULT;
1909                 RETURN(-EFAULT);
1910         }
1911
1912         rc = lustre_pack_reply(req, 1, &size, NULL);
1913         if (rc) {
1914                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1915                 req->rq_status = rc;
1916         }
1917
1918         RETURN(0);
1919 }