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