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         ENTRY;
 918
 919         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 920                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 921                   rec->ur_mode);
 922
 923         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 924                          (obd_timeout + 1) / 4);
 925
 926         if (offset == 3) { /* intent */
 927                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 928                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 929         } else if (offset == 1) { /* non-intent reint */
 930                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 931         } else {
 932                 body = NULL;
 933                 LBUG();
 934         }
 935
 936         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 937
 938         /*
 939          * Step 0: If we are passed a id, then we assume the client already
 940          * opened this file and is only replaying the RPC, so we open the inode
 941          * by fid (at some large expense in security).
 942          */
 943         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 944                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 945                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 946                        rec->ur_name, rec->ur_mode);
 947
 948                 LASSERT(id_ino(rec->ur_id2));
 949
 950                 rc = mds_open_by_id(req, rec->ur_id2, body,
 951                                     rec->ur_flags, rec, rep);
 952                 if (rc != -ENOENT) {
 953                         mds_body_do_reverse_map(med, body);
 954                         RETURN(rc);
 955                 }
 956
 957                 /* We didn't find the correct inode on disk either, so we
 958                  * need to re-create it via a regular replay. */
 959                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 960         } else {
 961                 LASSERT(!id_ino(rec->ur_id2));
 962         }
 963
 964         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 965
 966         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 967                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 968                 RETURN(-ENOMEM);
 969         }
 970
 971         acc_mode = accmode(rec->ur_flags);
 972
 973         /* Step 1: Find and lock the parent */
 974         if (rec->ur_flags & MDS_OPEN_CREAT) {
 975                 /* XXX Well, in fact we only need this lock mode change if
 976                    in addition to O_CREAT, the file does not exist.
 977                    But we do not know if it exists or not yet */
 978                 parent_mode = LCK_PW;
 979         }
 980
 981         if (rec->ur_namelen == 1) {
 982                 /* client (LMV) wants to open the file by id */
 983                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 984                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 985                 if (IS_ERR(dchild)) {
 986                         rc = PTR_ERR(dparent);
 987                         CERROR("child lookup by an id error %d\n", rc);
 988                         GOTO(cleanup, rc);
 989                 }
 990                 goto got_child;
 991         }
 992
 993 restart:
 994         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
 995                                        parent_lockh, &update_mode, rec->ur_name,
 996                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 997         if (IS_ERR(dparent)) {
 998                 rc = PTR_ERR(dparent);
 999                 if (rc != -ENOENT) {
1000                         CERROR("parent lookup for "DLID4" failed, error %d\n",
1001                                OLID4(rec->ur_id1), rc);
1002                 } else {
1003                         /* Just cannot find parent - make it look like
1004                            usual negative lookup to avoid extra MDS RPC */
1005                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1006                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
1007                 }
1008                 GOTO(cleanup, rc);
1009         }
1010         LASSERT(dparent->d_inode != NULL);
1011
1012         cleanup_phase = 1; /* parent dentry and lock */
1013
1014         /* try to retrieve MEA data for this dir */
1015         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1016         if (rc)
1017                 GOTO(cleanup, rc);
1018
1019         if (mea != NULL) {
1020                 /*
1021                  * dir is already splitted, check is requested filename should *
1022                  * live at this MDS or at another one.
1023                  */
1024                 int i;
1025                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1026                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1027                         CDEBUG(D_OTHER,
1028                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
1029                                " should be %lu(%d)\n", obd->obd_name,
1030                                mea->mea_master, dparent->d_inode->i_ino,
1031                                dparent->d_inode->i_generation, rec->ur_name,
1032                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1033                         GOTO(cleanup, rc = -ERESTART);
1034                 }
1035         }
1036
1037         /* Step 2: Lookup the child */
1038         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1039         if (IS_ERR(dchild)) {
1040                 rc = PTR_ERR(dchild);
1041                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1042                 GOTO(cleanup, rc);
1043         }
1044
1045 got_child:
1046         cleanup_phase = 2; /* child dentry */
1047
1048         if (dchild->d_flags & DCACHE_CROSS_REF) {
1049                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1050                 LASSERT(rec->ur_namelen > 1);
1051
1052                 /* we're gonna acquire LOOKUP lock on the child,
1053                  * but we have already locked parent and our order
1054                  * may conflict with enqueue_order_locks(). so,
1055                  * drop parent lock and acquire both the locks in
1056                  * common order. bug 6190 */
1057 #ifdef S_PDIROPS
1058                 if (parent_lockh[1].cookie != 0)
1059                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1060 #endif
1061                 ldlm_lock_decref(parent_lockh, parent_mode);
1062                 l_dput(dchild);
1063                 l_dput(dparent);
1064                 cleanup_phase = 0;
1065
1066                 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1067                                                  parent_lockh, &dparent,
1068                                                  LCK_PR, MDS_INODELOCK_UPDATE,
1069                                                  &update_mode, rec->ur_name,
1070                                                  rec->ur_namelen, child_lockh,
1071                                                  &dchild, LCK_PR,
1072                                                  MDS_INODELOCK_LOOKUP);
1073
1074                 if (rc)
1075                         GOTO(cleanup, rc);
1076
1077 #ifdef S_PDIROPS
1078                 if (parent_lockh[1].cookie != 0)
1079                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1080 #endif
1081                 ldlm_lock_decref(parent_lockh, LCK_PR);
1082
1083                 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1084                         CDEBUG(D_OTHER, "race: name changed (%p)\n",
1085                                dchild->d_inode);
1086                         if (dchild->d_inode)
1087                                 ldlm_lock_decref(child_lockh, LCK_PR);
1088                         l_dput(dchild);
1089                         l_dput(dparent);
1090                         GOTO(restart, rc);
1091                 }
1092
1093                 mds_pack_dentry2body(obd, body, dchild, 1);
1094                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1095                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1096
1097                 if (mea)
1098                         OBD_FREE(mea, mea_size);
1099                 l_dput(dchild);
1100                 l_dput(dparent);
1101                 RETURN(rc);
1102         }
1103
1104         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1105
1106         if (dchild->d_inode)
1107                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1108         else
1109                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1110
1111         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1112         if (dchild->d_inode == NULL) {
1113                 unsigned long ino;
1114                 struct iattr iattr;
1115                 struct inode *inode;
1116
1117                 ino = (unsigned long)id_ino(rec->ur_id2);
1118
1119                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1120                                           update_mode);
1121
1122                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1123                        obd->obd_name, dparent->d_inode->i_ino,
1124                        dparent->d_inode->i_generation, rc);
1125
1126                 if (rc > 0) {
1127                         /* dir got splitted */
1128                         GOTO(cleanup, rc = -ERESTART);
1129                 } else if (rc < 0) {
1130                         /* error happened during spitting */
1131                         GOTO(cleanup, rc);
1132                 }
1133
1134                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1135                         /*
1136                          * dentry is negative and we weren't supposed to create
1137                          * object, get out of here.
1138                          */
1139                         GOTO(cleanup, rc = -ENOENT);
1140                 }
1141
1142                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1143                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1144                                       NULL);
1145                 if (IS_ERR(handle)) {
1146                         rc = PTR_ERR(handle);
1147                         handle = NULL;
1148                         GOTO(cleanup, rc);
1149                 }
1150                 dchild->d_fsdata = (void *) &dp;
1151                 dp.p_ptr = req;
1152                 dp.p_inum = ino;
1153
1154                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1155                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1156                         dchild->d_fsdata = NULL;
1157
1158                 if (rc) {
1159                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1160                         GOTO(cleanup, rc);
1161                 }
1162                 inode = dchild->d_inode;
1163                 if (ino) {
1164                         LASSERT(ino == inode->i_ino);
1165
1166                         /* written as part of setattr */
1167                         inode->i_generation = id_gen(rec->ur_id2);
1168                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1169                                inode->i_ino, inode->i_generation);
1170                 }
1171
1172                 created = 1;
1173                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1174                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1175                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1176
1177                 iattr.ia_uid = rec->ur_fsuid;
1178                 if (dparent->d_inode->i_mode & S_ISGID)
1179                         iattr.ia_gid = dparent->d_inode->i_gid;
1180                 else
1181                         iattr.ia_gid = rec->ur_fsgid;
1182
1183                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1184                         ATTR_MTIME | ATTR_CTIME;
1185
1186                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1187                 if (rc)
1188                         CERROR("error on child setattr: rc = %d\n", rc);
1189
1190                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1191
1192                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1193                 if (rc)
1194                         CERROR("error on parent setattr: rc = %d\n", rc);
1195                 else {
1196                         MD_COUNTER_INCREMENT(obd, create);
1197                 }
1198
1199                 if (ino) {
1200                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1201                                                   handle, rec->ur_id2);
1202                         if (rc) {
1203                                 CERROR("mds_update_inode_sid() failed, "
1204                                        "rc = %d\n", rc);
1205                         }
1206                         id_assign_fid(&body->id1, rec->ur_id2);
1207
1208                         /*
1209                          * make sure, that fid is up-to-date.
1210                          */
1211                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1212                 } else {
1213                         rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1214                                                  handle, &body->id1);
1215                         if (rc) {
1216                                 CERROR("mds_alloc_inode_sid() failed, "
1217                                        "rc = %d\n", rc);
1218                         }
1219                 }
1220
1221                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1222                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1223                                            dchild->d_inode, handle,
1224                                            req->rq_export->exp_sync);
1225                         handle = NULL;
1226                 }
1227
1228                 acc_mode = 0;           /* Don't check for permissions */
1229         }
1230         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1231
1232         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1233                  "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1234                  dchild->d_name.name, dchild, dchild->d_inode);
1235
1236         if (S_ISREG(dchild->d_inode->i_mode)) {
1237                 /* Check permissions etc */
1238                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1239                 if (rc != 0)
1240                         GOTO(cleanup, rc);
1241
1242                 /* Can't write to a read-only file */
1243                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1244                         GOTO(cleanup, rc = -EPERM);
1245
1246                 /* An append-only file must be opened in append mode for
1247                  * writing */
1248                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1249                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1250                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1251                         GOTO(cleanup, rc = -EPERM);
1252         }
1253
1254         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1255             (rec->ur_flags & MDS_OPEN_EXCL)) {
1256                 /* File already exists, we didn't just create it, and we
1257                  * were passed O_EXCL; err-or. */
1258                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1259         }
1260
1261         if (S_ISDIR(dchild->d_inode->i_mode)) {
1262                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1263                     rec->ur_flags & FMODE_WRITE) {
1264                         /* we are trying to create or write a exist dir */
1265                         GOTO(cleanup, rc = -EISDIR);
1266                 }
1267                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1268                         GOTO(cleanup, rc = -EACCES);
1269
1270                 /*
1271                  * here was checking for possible GNS mount points to skip their
1272                  * open. I removed it as its detection based only on SUID bit is
1273                  * not reliable. Opening such a dirs should not cause any
1274                  * problems, at least tests show that. --umka
1275                  */
1276         }
1277
1278         /* if we are following a symlink, don't open */
1279         if (S_ISLNK(dchild->d_inode->i_mode))
1280                 GOTO(cleanup_no_trans, rc = 0);
1281
1282         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1283             !S_ISDIR(dchild->d_inode->i_mode))
1284                 GOTO(cleanup, rc = -ENOTDIR);
1285
1286         /* check permission even it's special files */
1287         if (S_ISCHR(dchild->d_inode->i_mode) ||
1288             S_ISBLK(dchild->d_inode->i_mode) ||
1289             S_ISFIFO(dchild->d_inode->i_mode) ||
1290             S_ISSOCK(dchild->d_inode->i_mode)) {
1291                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1292                 if (rc != 0)
1293                         GOTO(cleanup, rc);
1294         }
1295
1296         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1297                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1298                 GOTO(cleanup, rc = -EAGAIN);
1299         }
1300
1301 #if 0
1302         /* We cannot use acc_mode here, because it is zeroed in case of
1303            creating a file, so we get wrong lockmode */
1304         if (accmode(rec->ur_flags) & MAY_WRITE)
1305                 child_mode = LCK_CW;
1306         else if (accmode(rec->ur_flags) & MAY_EXEC)
1307                 child_mode = LCK_PR;
1308         else
1309                 child_mode = LCK_CR;
1310
1311         /* Always returning LOOKUP lock if open succesful to guard
1312          * dentry on client. In case of replay we do not get a lock
1313          * assuming that the caller has it already */
1314         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1315                 struct ldlm_res_id child_res_id = { .name = {0}};
1316                 ldlm_policy_data_t policy;
1317                 int lock_flags = LDLM_FL_ATOMIC_CB;
1318
1319                 /* LOOKUP lock will protect dentry on client -bzzz */
1320                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1321                                                 MDS_INODELOCK_OPEN;
1322
1323                 child_res_id.name[0] = id_fid(&body->id1);
1324                 child_res_id.name[1] = id_group(&body->id1);
1325
1326                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1327                                       child_res_id, LDLM_IBITS, &policy,
1328                                       child_mode, &lock_flags,
1329                                       mds_blocking_ast, ldlm_completion_ast,
1330                                       NULL, NULL, NULL, 0, NULL, child_lockh);
1331                 if (rc != ELDLM_OK)
1332                         GOTO(cleanup, rc);
1333
1334                 cleanup_phase = 3;
1335         }
1336 #else
1337 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1338 #warning "disable opencache lock for CMD2"
1339 #endif
1340
1341         /* Step 5: mds_open it */
1342         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1343                              rec, rep);
1344         if (rc)
1345                 GOTO(cleanup, rc);
1346
1347         /* if this is a writer, we have to invalidate client's
1348          * update locks in order to make sure they don't use
1349          * isize/iblocks from mds anymore.
1350          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1351          * XXX: optimization is to do this for first writer only */
1352         if (accmode(rec->ur_flags) & MAY_WRITE) {
1353                 struct ldlm_res_id child_res_id = { .name = {0}};
1354                 ldlm_policy_data_t sz_policy;
1355                 struct lustre_handle sz_lockh;
1356                 int lock_flags = 0;
1357
1358                 child_res_id.name[0] = id_fid(&body->id1);
1359                 child_res_id.name[1] = id_group(&body->id1);
1360                 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1361
1362                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1363                                       child_res_id, LDLM_IBITS, &sz_policy,
1364                                       LCK_PW, &lock_flags, mds_blocking_ast,
1365                                       ldlm_completion_ast, NULL, NULL, NULL,
1366                                       0, NULL, &sz_lockh);
1367                 if (rc == ELDLM_OK)
1368                         ldlm_lock_decref(&sz_lockh, LCK_PW);
1369                 else
1370                         CERROR("can't invalidate client's update locks\n");
1371         }
1372
1373         EXIT;
1374 cleanup:
1375         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1376                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1377
1378 cleanup_no_trans:
1379         switch (cleanup_phase) {
1380         case 3:
1381                 if (rc) {
1382                         ldlm_lock_decref(child_lockh, child_mode);
1383                         child_lockh->cookie = 0;
1384                 }
1385         case 2:
1386                 if (rc && created) {
1387                         int err = vfs_unlink(dparent->d_inode, dchild);
1388                         if (err) {
1389                                 CERROR("unlink(%.*s) in error path: %d\n",
1390                                        dchild->d_name.len, dchild->d_name.name,
1391                                        err);
1392                         }
1393                 } else if (created) {
1394                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1395                 }
1396                 l_dput(dchild);
1397         case 1:
1398                 if (dparent == NULL)
1399                         break;
1400
1401                 l_dput(dparent);
1402 #ifdef S_PDIROPS
1403                 if (parent_lockh[1].cookie != 0)
1404                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1405 #endif
1406                 if (rc)
1407                         ldlm_lock_decref(parent_lockh, parent_mode);
1408                 else
1409                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1410         }
1411         if (mea)
1412                 OBD_FREE(mea, mea_size);
1413         if (rc == 0)
1414                 atomic_inc(&mds->mds_open_count);
1415
1416         mds_body_do_reverse_map(med, body);
1417
1418         /*
1419          * If we have not taken the "open" lock, we may not return 0 here,
1420          * because caller expects 0 to mean "lock is taken", and it needs
1421          * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1422          * client. Later caller should rewrite the return value back to zero
1423          * if it to be used any further.
1424          */
1425         if ((cleanup_phase != 3) && !rc)
1426                 rc = ENOLCK;
1427
1428         RETURN(rc);
1429 }
1430
1431 /* Close a "file descriptor" and possibly unlink an orphan from the
1432  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1433  *
1434  * If we are being called from mds_disconnect() because the client has
1435  * disappeared, then req == NULL and we do not update last_rcvd because
1436  * there is nothing that could be recovered by the client at this stage
1437  * (it will not even _have_ an entry in last_rcvd anymore).
1438  *
1439  * Returns EAGAIN if the client needs to get more data and re-close. */
1440 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1441                   struct obd_device *obd, struct mds_file_data *mfd,
1442                   int unlink_orphan)
1443 {
1444         struct inode *inode = mfd->mfd_dentry->d_inode;
1445         char idname[LL_ID_NAMELEN];
1446         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1447         struct dentry *pending_child = NULL;
1448         struct mds_obd *mds = &obd->u.mds;
1449         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1450         void *handle = NULL;
1451         struct mds_body *request_body = NULL, *reply_body = NULL;
1452         struct dentry_params dp;
1453         struct iattr iattr = { 0 };
1454         struct llog_create_locks *lcl = NULL;
1455         ENTRY;
1456
1457         if (req && req->rq_reqmsg != NULL)
1458                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1459                                               sizeof(*request_body));
1460         if (req && req->rq_repmsg != NULL)
1461                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1462                                             sizeof(*reply_body));
1463
1464         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1465         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1466                idname, inode->i_nlink, mds_orphan_open_count(inode));
1467
1468         last_orphan = mds_orphan_open_dec_test(inode) &&
1469                 mds_inode_is_orphan(inode);
1470         UP_WRITE_I_ALLOC_SEM(inode);
1471
1472         /* this is half of the actual "close" */
1473         if (mfd->mfd_mode & FMODE_WRITE) {
1474                 rc = mds_put_write_access(mds, inode, request_body,
1475                                           last_orphan && unlink_orphan);
1476         } else if (mfd->mfd_mode & FMODE_EXEC) {
1477                 mds_allow_write_access(inode);
1478         }
1479
1480         if (last_orphan && unlink_orphan) {
1481                 struct lov_mds_md *lmm = NULL;
1482                 int stripe_count = 0;
1483                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1484
1485                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1486
1487                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1488                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1489                         CERROR("found \"orphan\" %s %s with link count %d\n",
1490                                S_ISREG(inode->i_mode) ? "file" : "dir",
1491                                idname, inode->i_nlink);
1492                 /*
1493                  * sadly, there is no easy way to save pending_child from
1494                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1495                  * normally it will still be in the dcache.
1496                  */
1497                 down(&pending_dir->i_sem);
1498                 cleanup_phase = 1; /* up(i_sem) when finished */
1499                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1500                                                idlen);
1501                 if (IS_ERR(pending_child))
1502                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1503                 if (pending_child->d_inode == NULL) {
1504                         CERROR("orphan %s has been removed\n", idname);
1505                         GOTO(cleanup, rc = 0);
1506                 }
1507
1508                 cleanup_phase = 2; /* dput(pending_child) when finished */
1509                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1510                         rc = vfs_rmdir(pending_dir, pending_child);
1511                         if (rc)
1512                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1513                                        idname, rc);
1514                         goto out;
1515                 }
1516
1517                 if (req != NULL && req->rq_repmsg != NULL) {
1518                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1519                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1520                 }
1521
1522                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1523                                           NULL, stripe_count);
1524                 if (IS_ERR(handle)) {
1525                         rc = PTR_ERR(handle);
1526                         handle = NULL;
1527                         GOTO(cleanup, rc);
1528                 }
1529
1530                 pending_child->d_fsdata = (void *) &dp;
1531                 dp.p_inum = 0;
1532                 dp.p_ptr = req;
1533                 rc = vfs_unlink(pending_dir, pending_child);
1534                 if (rc)
1535                         CERROR("error unlinking orphan %s: rc %d\n",
1536                                idname, rc);
1537
1538                 if (req != NULL && req->rq_repmsg != NULL &&
1539                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1540                     mds_log_op_unlink(obd, pending_child->d_inode,
1541                                       lmm, req->rq_repmsg->buflens[1],
1542                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1543                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1544                         reply_body->valid |= OBD_MD_FLCOOKIE;
1545                 }
1546
1547                 rc = mds_destroy_object(obd, inode, 1);
1548                 if (rc) {
1549                         CERROR("cannot destroy OSS object on close, err %d\n",
1550                                rc);
1551                         rc = 0;
1552                 }
1553
1554                 goto out; /* Don't bother updating attrs on unlinked inode */
1555         } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0 && request_body) {
1556                 /* last writer closed file - let's update i_size/i_blocks */
1557                 if (request_body->valid & OBD_MD_FLSIZE) {
1558                         LASSERT(request_body->valid & OBD_MD_FLBLOCKS);
1559                         CDEBUG(D_OTHER, "update size "LPD64" for "DLID4
1560                                ", epoch "LPD64"\n", inode->i_size,
1561                                OLID4(&request_body->id1),
1562                                request_body->io_epoch);
1563                         iattr.ia_size = inode->i_size;
1564                         iattr.ia_valid |= ATTR_SIZE;
1565                 }
1566         }
1567
1568 #if 0
1569         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1570                 /* Update the on-disk attributes if this was the last write
1571                  * close, and all information was provided (i.e., rc == 0)
1572                  *
1573                  * XXX this should probably be abstracted with mds_reint_setattr
1574                  */
1575
1576                 if (request_body->valid & OBD_MD_FLMTIME &&
1577                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1578                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1579                         iattr.ia_valid |= ATTR_MTIME;
1580                 }
1581                 if (request_body->valid & OBD_MD_FLCTIME &&
1582                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1583                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1584                         iattr.ia_valid |= ATTR_CTIME;
1585                 }
1586
1587                 /* XXX can't set block count with fsfilt_setattr (!) */
1588                 if (request_body->valid & OBD_MD_FLSIZE) {
1589                         iattr.ia_valid |= ATTR_SIZE;
1590                         iattr.ia_size = request_body->size;
1591                 }
1592                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1593                         iattr.ia_valid |= ATTR_BLOCKS;
1594                         iattr.ia_blocks = request_body->blocks
1595                 } */
1596
1597         }
1598 #endif
1599         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1600                 /* Only start a transaction to write out only the atime if
1601                  * it is more out-of-date than the specified limit.  If we
1602                  * are already going to write out the atime then do it anyway.
1603                  * */
1604                 if ((request_body->atime >
1605                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1606                     (iattr.ia_valid != 0 &&
1607                      request_body->atime > LTIME_S(inode->i_atime))) {
1608                         LTIME_S(iattr.ia_atime) = request_body->atime;
1609                         iattr.ia_valid |= ATTR_ATIME;
1610                 }
1611         }
1612
1613         if (iattr.ia_valid != 0) {
1614                 if (handle == NULL)
1615                         handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1616                 if (IS_ERR(handle))
1617                         GOTO(cleanup, rc = PTR_ERR(handle));
1618                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1619                 if (rc)
1620                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1621         }
1622 out:
1623         /* If other clients have this file open for write, rc will be > 0 */
1624         if (rc > 0)
1625                 rc = 0;
1626         l_dput(mfd->mfd_dentry);
1627         mds_mfd_destroy(mfd);
1628
1629  cleanup:
1630         atomic_dec(&mds->mds_open_count);
1631         if (req != NULL && reply_body != NULL) {
1632                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1633         } else if (handle) {
1634                 int err, force_sync = 0;
1635
1636                 if (req && req->rq_export)
1637                         force_sync = req->rq_export->exp_sync;
1638
1639                 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1640                                     force_sync);
1641                 if (err) {
1642                         CERROR("error committing close: %d\n", err);
1643                         if (!rc)
1644                                 rc = err;
1645                 }
1646         }
1647
1648         switch (cleanup_phase) {
1649         case 2:
1650                 if (lcl != NULL)
1651                         ptlrpc_save_llog_lock(req, lcl);
1652                 dput(pending_child);
1653         case 1:
1654                 up(&pending_dir->i_sem);
1655         }
1656         RETURN(rc);
1657 }
1658
1659 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1660                                     struct ldlm_lock_desc *new, void *data,
1661                                     int flag)
1662 {
1663         struct lustre_handle lockh = { 0 };
1664         int rc;
1665         ENTRY;
1666
1667         switch (flag) {
1668         case LDLM_CB_BLOCKING:
1669                 ldlm_lock2handle(lock, &lockh);
1670                 rc = ldlm_cli_cancel(&lockh);
1671                 if (rc != ELDLM_OK)
1672                         CERROR("ldlm_cli_cancel failed: %d\n", rc);
1673                 break;
1674         case LDLM_CB_CANCELING: {
1675                 break;
1676         }
1677         default:
1678                 LBUG();
1679         }
1680
1681         RETURN(0);
1682 }
1683 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1684 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1685
1686 int mds_validate_size(struct obd_device *obd, struct mds_body *body,
1687                       struct mds_file_data *mfd)
1688 {
1689         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1690         struct inode *inode = mfd->mfd_dentry->d_inode;
1691         struct lustre_handle lockh = { 0 };
1692         struct lov_stripe_md *lsm = NULL;
1693         int rc, len, flags;
1694         void *lmm = NULL;
1695         ENTRY;
1696
1697         /* we update i_size/i_blocks only for regular files */
1698         if (!S_ISREG(inode->i_mode))
1699                 RETURN(0);
1700
1701         /* we update i_size/i_blocks only for writers */
1702         if (!(mfd->mfd_mode & FMODE_WRITE))
1703                 RETURN(0);
1704
1705         /* we like when client reports actual i_size/i_blocks himself */
1706         if (body->valid & OBD_MD_FLSIZE) {
1707                 LASSERT(body->valid & OBD_MD_FLBLOCKS);
1708                 CDEBUG(D_OTHER, "client reports "LPD64"/"LPD64" for "DLID4"\n",
1709                        body->size, body->blocks, OLID4(&body->id1));
1710                 RETURN(0);
1711         }
1712
1713         /* we shouldn't fetch size from OSTes during recovery - deadlock */
1714         if (obd->obd_recovering)
1715                 RETURN(0);
1716
1717         DOWN_READ_I_ALLOC_SEM(inode);
1718         if (atomic_read(&inode->i_writecount) > 1
1719                         || mds_inode_is_orphan(inode)) {
1720                 /* there is no need to update i_size/i_blocks on orphans.
1721                  * also, if this is not last writer, then it doesn't make
1722                  * sense to fetch i_size/i_blocks from OSSes */
1723                 UP_READ_I_ALLOC_SEM(inode);
1724                 RETURN(0);
1725         }
1726         UP_READ_I_ALLOC_SEM(inode);
1727
1728         /* 1: client didn't send actual i_size/i_blocks
1729          * 2: we seem to be last writer
1730          * 3: the file doesn't look like to be disappeared
1731          * conclusion: we're gonna fetch them from OSSes */
1732
1733         down(&inode->i_sem);
1734         len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1735         up(&inode->i_sem);
1736
1737         if (len < 0) {
1738                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1739                 GOTO(cleanup, rc = len);
1740         } else if (len == 0) {
1741                 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1742                 GOTO(cleanup, rc = 0);
1743         }
1744
1745         OBD_ALLOC(lmm, len);
1746         if (lmm == NULL) {
1747                 CERROR("can't allocate memory\n");
1748                 GOTO(cleanup, rc = -ENOMEM);
1749         }
1750
1751         down(&inode->i_sem);
1752         rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1753         up(&inode->i_sem);
1754
1755         if (rc < 0) {
1756                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1757                 GOTO(cleanup, rc);
1758         }
1759
1760         rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1761         if (rc < 0) {
1762                 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1763                 GOTO(cleanup, rc);
1764         }
1765
1766         CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1767
1768         flags = LDLM_FL_HAS_INTENT;
1769         rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1770                          LCK_PR, &flags, mds_extent_lock_callback,
1771                          ldlm_completion_ast, NULL, NULL,
1772                          sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1773         if (rc != 0) {
1774                 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1775                 GOTO(cleanup, rc);
1776         }
1777
1778         body->size = lov_merge_size(lsm, 0);
1779         body->blocks = lov_merge_blocks(lsm);
1780         body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1781
1782         CDEBUG(D_OTHER, "LOV reports "LPD64"/"LPD64" for "DLID4"\n",
1783                         body->size, body->blocks, OLID4(&body->id1));
1784
1785         obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1786
1787 cleanup:
1788         if (lsm != NULL)
1789                 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1790         if (lmm != NULL)
1791                 OBD_FREE(lmm, len);
1792         RETURN(rc);
1793 }
1794
1795 int mds_close(struct ptlrpc_request *req, int offset)
1796 {
1797         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1798         struct obd_device *obd = req->rq_export->exp_obd;
1799         struct mds_body *body;
1800         struct mds_file_data *mfd;
1801         struct lvfs_run_ctxt saved;
1802         struct inode *inode;
1803         int rc, repsize[3] = {sizeof(struct mds_body),
1804                               obd->u.mds.mds_max_mdsize,
1805                               obd->u.mds.mds_max_cookiesize};
1806         ENTRY;
1807         MD_COUNTER_INCREMENT(obd, close);
1808
1809         rc = lustre_pack_reply(req, 3, repsize, NULL);
1810         if (rc) {
1811                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1812                 req->rq_status = rc;
1813         } else {
1814                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1815         }
1816
1817         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1818                 DEBUG_REQ(D_HA, req, "close replay");
1819                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1820                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1821                        req->rq_repmsg->buflens[2]);
1822         }
1823
1824         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1825                                   lustre_swab_mds_body);
1826         if (body == NULL) {
1827                 CERROR("Can't unpack body\n");
1828                 req->rq_status = -EFAULT;
1829                 RETURN(-EFAULT);
1830         }
1831
1832         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1833                 /* do some stuff */ ;
1834
1835         mfd = mds_handle2mfd(&body->handle);
1836         if (mfd == NULL) {
1837                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1838                           ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1839                 req->rq_status = -ESTALE;
1840                 RETURN(-ESTALE);
1841         }
1842
1843         rc = mds_validate_size(obd, body, mfd);
1844         LASSERT(rc == 0);
1845
1846         inode = mfd->mfd_dentry->d_inode;
1847
1848         if (mfd->mfd_mode & FMODE_WRITE) {
1849                 /* we set i_size/i_blocks here, nobody will see
1850                  * them until all write references are dropped.
1851                  * btw, we hold one reference */
1852                 if (body->valid & OBD_MD_FLSIZE)
1853                         i_size_write(inode, body->size);
1854                 if (body->valid & OBD_MD_FLBLOCKS)
1855                         inode->i_blocks = body->blocks;
1856         }
1857
1858         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1859         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1860         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1861                 struct mds_body *rep_body;
1862
1863                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1864                                           sizeof (*rep_body));
1865                 LASSERT(rep_body != NULL);
1866
1867                 mds_pack_inode2body(obd, rep_body, inode,
1868                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1869
1870                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1871                             inode, MDS_PACK_MD_LOCK, 0);
1872         }
1873         spin_lock(&med->med_open_lock);
1874         list_del(&mfd->mfd_list);
1875         spin_unlock(&med->med_open_lock);
1876
1877         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1878         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1879         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1880
1881         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1882                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1883                 req->rq_status = -ENOMEM;
1884                 mds_mfd_put(mfd);
1885                 RETURN(-ENOMEM);
1886         }
1887
1888         mds_mfd_put(mfd);
1889         RETURN(0);
1890 }
1891
1892 int mds_done_writing(struct ptlrpc_request *req, int offset)
1893 {
1894         struct mds_body *body;
1895         int rc, size = sizeof(struct mds_body);
1896         ENTRY;
1897
1898         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1899
1900         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1901                                   lustre_swab_mds_body);
1902         if (body == NULL) {
1903                 CERROR("Can't unpack body\n");
1904                 req->rq_status = -EFAULT;
1905                 RETURN(-EFAULT);
1906         }
1907
1908         rc = lustre_pack_reply(req, 1, &size, NULL);
1909         if (rc) {
1910                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1911                 req->rq_status = rc;
1912         }
1913
1914         RETURN(0);
1915 }