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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mds/mds_open.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  * Author: Mike Shaver <shaver@clusterfs.com>
  42  */
  43
  44 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include <linux/version.h>
  50 #include <linux/module.h>
  51 #include <linux/init.h>
  52 #include <linux/version.h>
  53 #include <linux/buffer_head.h>
  54 #include <linux/workqueue.h>
  55
  56 #include <obd_class.h>
  57 #include <obd_lov.h>
  58 #include <lustre_fsfilt.h>
  59 #include <lprocfs_status.h>
  60
  61 #include "mds_internal.h"
  62
  63 /* Exported function from this file are:
  64  *
  65  * mds_open - called by the intent handler
  66  * mds_close - an rpc handling function
  67  * mds_pin - an rpc handling function - which will go away
  68  * mds_mfd_close - for force closing files when a client dies
  69  */
  70
  71 /*
  72  * MDS file data handling: file data holds a handle for a file opened
  73  * by a client.
  74  */
  75
  76 static void mds_mfd_addref(void *mfdp)
  77 {
  78         struct mds_file_data *mfd = mfdp;
  79
  80         atomic_inc(&mfd->mfd_refcount);
  81         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  82                atomic_read(&mfd->mfd_refcount));
  83 }
  84
  85 /* Create a new mds_file_data struct.
  86  * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
  87  * one reference for the caller of this function. */
  88 struct mds_file_data *mds_mfd_new(void)
  89 {
  90         struct mds_file_data *mfd;
  91
  92         OBD_ALLOC(mfd, sizeof *mfd);
  93         if (mfd == NULL) {
  94                 CERROR("mds: out of memory\n");
  95                 return NULL;
  96         }
  97
  98         atomic_set(&mfd->mfd_refcount, 2);
  99
 100         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
 101         INIT_LIST_HEAD(&mfd->mfd_list);
 102         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
 103
 104         return mfd;
 105 }
 106
 107 /* Get a new reference on the mfd pointed to by handle, if handle is still
 108  * valid.  Caller must drop reference with mds_mfd_put(). */
 109 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
 110 {
 111         ENTRY;
 112         LASSERT(handle != NULL);
 113         RETURN(class_handle2object(handle->cookie));
 114 }
 115
 116 /* Drop mfd reference, freeing struct if this is the last one. */
 117 static void mds_mfd_put(struct mds_file_data *mfd)
 118 {
 119         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 120                atomic_read(&mfd->mfd_refcount) - 1);
 121         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 122                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 123         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 124                 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 125         }
 126 }
 127
 128 /* Remove the mfd handle so that it cannot be found by open/close again.
 129  * Caller must hold med_open_lock for mfd_list manipulation. */
 130 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
 131 {
 132         class_handle_unhash(&mfd->mfd_handle);
 133         list_del_init(&mfd->mfd_list);
 134         if (decref)
 135                 mds_mfd_put(mfd);
 136 }
 137
 138 /* Caller must hold mds->mds_epoch_sem */
 139 static int mds_alloc_filterdata(struct inode *inode)
 140 {
 141         LASSERT(INODE_PRIVATE_DATA(inode) == NULL);
 142         OBD_ALLOC(INODE_PRIVATE_DATA(inode), sizeof(struct mds_filter_data));
 143         if (INODE_PRIVATE_DATA(inode) == NULL)
 144                 return -ENOMEM;
 145         LASSERT(igrab(inode) == inode);
 146         return 0;
 147 }
 148
 149 /* Caller must hold mds->mds_epoch_sem */
 150 static void mds_free_filterdata(struct inode *inode)
 151 {
 152         LASSERT(INODE_PRIVATE_DATA(inode) != NULL);
 153         OBD_FREE(INODE_PRIVATE_DATA(inode), sizeof(struct mds_filter_data));
 154         INODE_PRIVATE_DATA(inode) = NULL;
 155         iput(inode);
 156 }
 157
 158 /* Write access to a file: executors cause a negative count,
 159  * writers a positive count.  The semaphore is needed to perform
 160  * a check for the sign and then increment or decrement atomically.
 161  *
 162  * This code is closely tied to the allocation of the d_fsdata and the
 163  * MDS epoch, so we use the same semaphore for the whole lot.
 164  *
 165  * We could use a different semaphore for each file, if it ever shows
 166  * up in a profile, which it won't.
 167  *
 168  * epoch argument is nonzero during recovery */
 169 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 170                                 __u64 epoch)
 171 {
 172         int rc = 0;
 173
 174         down(&mds->mds_epoch_sem);
 175
 176         if (atomic_read(&inode->i_writecount) < 0) {
 177                 up(&mds->mds_epoch_sem);
 178                 RETURN(-ETXTBSY);
 179         }
 180
 181
 182         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 183                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 184                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 185                        inode->i_generation);
 186                 goto out;
 187         }
 188
 189         if (MDS_FILTERDATA(inode) == NULL)
 190                 mds_alloc_filterdata(inode);
 191         if (MDS_FILTERDATA(inode) == NULL) {
 192                 rc = -ENOMEM;
 193                 goto out;
 194         }
 195         if (epoch > mds->mds_io_epoch)
 196                 mds->mds_io_epoch = epoch;
 197         else
 198                 mds->mds_io_epoch++;
 199         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 200         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 201                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 202  out:
 203         if (rc == 0)
 204                 atomic_inc(&inode->i_writecount);
 205         up(&mds->mds_epoch_sem);
 206         return rc;
 207 }
 208
 209 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 210  * again -- which is never true if the file is about to be unlinked.  Otherwise
 211  * returns the number of remaining writers. */
 212 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 213                                 struct mds_body *body, int unlinking)
 214 {
 215         int rc = 0;
 216         ENTRY;
 217
 218         down(&mds->mds_epoch_sem);
 219         atomic_dec(&inode->i_writecount);
 220         rc = atomic_read(&inode->i_writecount);
 221         if (rc > 0)
 222                 GOTO(out, rc);
 223 #if 0
 224         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 225                 GOTO(out, rc = EAGAIN);
 226 #endif
 227         mds_free_filterdata(inode);
 228  out:
 229         up(&mds->mds_epoch_sem);
 230         return rc;
 231 }
 232
 233 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 234 {
 235         ENTRY;
 236         down(&mds->mds_epoch_sem);
 237         if (atomic_read(&inode->i_writecount) > 0) {
 238                 up(&mds->mds_epoch_sem);
 239                 RETURN(-ETXTBSY);
 240         }
 241         atomic_dec(&inode->i_writecount);
 242         up(&mds->mds_epoch_sem);
 243         RETURN(0);
 244 }
 245
 246 static void mds_allow_write_access(struct inode *inode)
 247 {
 248         ENTRY;
 249         atomic_inc(&inode->i_writecount);
 250 }
 251
 252 int mds_query_write_access(struct inode *inode)
 253 {
 254         ENTRY;
 255         RETURN(atomic_read(&inode->i_writecount));
 256 }
 257
 258 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 259 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 260                                              struct vfsmount *mnt, int flags,
 261                                              struct ptlrpc_request *req)
 262 {
 263         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 264         struct mds_obd *mds = mds_req2mds(req);
 265         struct mds_file_data *mfd;
 266         struct mds_body *body;
 267         int error;
 268         ENTRY;
 269
 270         mfd = mds_mfd_new();
 271         if (mfd == NULL) {
 272                 CERROR("mds: out of memory\n");
 273                 GOTO(cleanup_dentry, error = -ENOMEM);
 274         }
 275
 276         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 277
 278         if (flags & FMODE_WRITE) {
 279                 /* FIXME: in recovery, need to pass old epoch here */
 280                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 281                 if (error)
 282                         GOTO(cleanup_mfd, error);
 283                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 284         } else if (flags & MDS_FMODE_EXEC) {
 285                 error = mds_deny_write_access(mds, dentry->d_inode);
 286                 if (error)
 287                         GOTO(cleanup_mfd, error);
 288         }
 289
 290         dget(dentry);
 291
 292         /* Mark the file as open to handle open-unlink. */
 293         MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
 294         mds_orphan_open_inc(dentry->d_inode);
 295         MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
 296
 297         mfd->mfd_mode = flags;
 298         mfd->mfd_dentry = dentry;
 299         mfd->mfd_xid = req->rq_xid;
 300         body->handle.cookie = mfd->mfd_handle.h_cookie;
 301
 302         if (req->rq_export->exp_disconnected) {
 303                 mds_mfd_unlink(mfd, 0);
 304                 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
 305                 mds_mfd_close(NULL, REQ_REC_OFF, req->rq_export->exp_obd,
 306                               mfd, 0, NULL, 0, NULL, 0, NULL);
 307         } else {
 308                 spin_lock(&med->med_open_lock);
 309                 list_add(&mfd->mfd_list, &med->med_open_head);
 310                 spin_unlock(&med->med_open_lock);
 311         }
 312
 313         RETURN(mfd);
 314
 315 cleanup_mfd:
 316         mds_mfd_put(mfd);
 317         mds_mfd_unlink(mfd, 1);
 318 cleanup_dentry:
 319         return ERR_PTR(error);
 320 }
 321
 322 /* Must be called with i_mutex held */
 323 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 324                               struct mds_update_record *rec,
 325                               struct mds_obd *mds, struct obd_device *obd,
 326                               struct dentry *dchild, void **handle,
 327                               struct lov_mds_md **objid)
 328 {
 329         struct inode *inode = dchild->d_inode;
 330         struct obd_trans_info oti = { 0 };
 331         struct lov_mds_md *lmm = NULL;
 332         int rc, lmm_size;
 333         struct mds_body *body;
 334         struct obd_info oinfo = { { { 0 } } };
 335         void *lmm_buf;
 336         ENTRY;
 337
 338         *objid = NULL;
 339
 340         if (!S_ISREG(inode->i_mode))
 341                 RETURN(0);
 342         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 343             !(rec->ur_flags & FMODE_WRITE))
 344                 RETURN(0);
 345
 346         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 347
 348         if (body->valid & OBD_MD_FLEASIZE)
 349                 RETURN(0);
 350
 351         oti_init(&oti, req);
 352
 353         /* replay case */
 354         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 355                 if (rec->ur_fid2->id == 0) {
 356                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 357                         RETURN(-EFAULT);
 358                 }
 359
 360                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 361                 lmm_size = rec->ur_eadatalen;
 362                 lmm = rec->ur_eadata;
 363                 LASSERT(lmm);
 364                 if (lov_mds_md_size(lmm->lmm_stripe_count,
 365                                     lmm->lmm_magic) != lmm_size)
 366                         CWARN("Bad lmm_size during open replay for inode "
 367                               "%lu\n", inode->i_ino);
 368
 369                 if (*handle == NULL) {
 370                         int stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
 371                         *handle = fsfilt_start_log(obd, inode, FSFILT_OP_CREATE,
 372                                                    NULL, stripe_count);
 373                 }
 374                 if (IS_ERR(*handle)) {
 375                         rc = PTR_ERR(*handle);
 376                         *handle = NULL;
 377                         GOTO(out_ids, rc);
 378                 }
 379
 380                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 381                 if (rc)
 382                         CERROR("open replay failed to set md:%d\n", rc);
 383
 384                 /* for replay we not need send lmm to client, this not used now */
 385                 lustre_shrink_reply(req, offset, 0, 1);
 386                 *objid = lmm;
 387
 388                 RETURN(rc);
 389         }
 390
 391         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 392                 GOTO(out_ids, rc = -ENOMEM);
 393
 394         OBDO_ALLOC(oinfo.oi_oa);
 395         if (oinfo.oi_oa == NULL)
 396                 GOTO(out_ids, rc = -ENOMEM);
 397         oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
 398         oinfo.oi_oa->o_gid = 0;
 399         oinfo.oi_oa->o_mode = S_IFREG | 0600;
 400         oinfo.oi_oa->o_id = inode->i_ino;
 401         oinfo.oi_oa->o_gr = 0;
 402         oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
 403                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID| OBD_MD_FLGROUP;
 404         oinfo.oi_oa->o_size = 0;
 405
 406         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 407                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 408         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 409                 /* check if things like lfs setstripe are sending us the ea */
 410                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 411                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 412                                            mds->mds_lov_exp,
 413                                            0, &oinfo.oi_md, rec->ur_eadata);
 414                         if (rc)
 415                                 GOTO(out_oa, rc);
 416                 } else {
 417                         __u32 lmm_sz = mds->mds_max_mdsize;
 418                         OBD_ALLOC(lmm, lmm_sz);
 419                         if (lmm == NULL)
 420                                 GOTO(out_oa, rc = -ENOMEM);
 421
 422                         lmm_size = lmm_sz;
 423                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 424                                         lmm, &lmm_size, 1, 0,
 425                                         req->rq_export->exp_connect_flags);
 426                         if (rc > 0)
 427                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 428                                                    mds->mds_lov_exp,
 429                                                    0, &oinfo.oi_md, lmm);
 430                         OBD_FREE(lmm, lmm_sz);
 431                         if (rc)
 432                                 GOTO(out_oa, rc);
 433                 }
 434                 rc = obd_create(mds->mds_lov_exp, oinfo.oi_oa,
 435                                 &oinfo.oi_md, &oti);
 436                 if (rc) {
 437                         int level = D_ERROR;
 438                         if (rc == -ENOSPC)
 439                                 level = D_INODE;
 440                         CDEBUG_LIMIT(level, "error creating objects for "
 441                                      "inode %lu: rc = %d\n", inode->i_ino, rc);
 442                         if (rc > 0) {
 443                                 CERROR("obd_create returned invalid "
 444                                        "rc %d\n", rc);
 445                                 rc = -EIO;
 446                         }
 447                         GOTO(out_oa, rc);
 448                 }
 449         } else {
 450                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_lov_exp,
 451                                    0, &oinfo.oi_md, rec->ur_eadata);
 452                 if (rc) {
 453                         GOTO(out_oa, rc);
 454                 }
 455                 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
 456         }
 457         if (i_size_read(inode)) {
 458                 oinfo.oi_oa->o_size = i_size_read(inode);
 459                 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
 460                                 OBD_MD_FLATIME | OBD_MD_FLMTIME |
 461                                 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 462
 463                 /* pack lustre id to OST */
 464                 oinfo.oi_oa->o_fid = body->fid1.id;
 465                 oinfo.oi_oa->o_generation = body->fid1.generation;
 466                 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 467                 oinfo.oi_policy.l_extent.start = i_size_read(inode);
 468                 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 469
 470                 rc = obd_punch_rqset(mds->mds_lov_exp, &oinfo, &oti);
 471                 if (rc) {
 472                         CERROR("error setting attrs for inode %lu: rc %d\n",
 473                                inode->i_ino, rc);
 474                         if (rc > 0) {
 475                                 CERROR("obd_setattr_async returned bad rc %d\n",
 476                                        rc);
 477                                 rc = -EIO;
 478                         }
 479                         GOTO(out_oa, rc);
 480                 }
 481         }
 482
 483         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 484         obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
 485
 486         LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
 487         lmm = NULL;
 488         rc = obd_packmd(mds->mds_lov_exp, &lmm, oinfo.oi_md);
 489         if (rc < 0) {
 490                 CERROR("cannot pack lsm, err = %d\n", rc);
 491                 GOTO(out_oa, rc);
 492         }
 493         lmm_size = rc;
 494         body->eadatasize = rc;
 495
 496         if (*handle == NULL)
 497                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 498         if (IS_ERR(*handle)) {
 499                 rc = PTR_ERR(*handle);
 500                 *handle = NULL;
 501                 GOTO(free_diskmd, rc);
 502         }
 503
 504         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 505         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 506         LASSERT(lmm_buf);
 507         memcpy(lmm_buf, lmm, lmm_size);
 508         *objid = lmm_buf; /* save for mds_lov_update_objid */
 509
 510  free_diskmd:
 511         obd_free_diskmd(mds->mds_lov_exp, &lmm);
 512  out_oa:
 513         oti_free_cookies(&oti);
 514         OBDO_FREE(oinfo.oi_oa);
 515  out_ids:
 516         if (oinfo.oi_md)
 517                 obd_free_memmd(mds->mds_lov_exp, &oinfo.oi_md);
 518         RETURN(rc);
 519 }
 520
 521 static void reconstruct_open(struct mds_update_record *rec, int offset,
 522                              struct ptlrpc_request *req,
 523                              struct lustre_handle *child_lockh)
 524 {
 525         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 526         struct lsd_client_data *lcd = med->med_lcd;
 527         struct mds_obd *mds = mds_req2mds(req);
 528         struct mds_file_data *mfd = NULL;
 529         struct obd_export *exp = req->rq_export;
 530         struct obd_device *obd = exp->exp_obd;
 531         struct dentry *parent, *dchild;
 532         struct ldlm_reply *rep;
 533         struct mds_body *body;
 534         struct list_head *t;
 535         int rc;
 536         int put_child = 1;
 537         ENTRY;
 538
 539         LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
 540         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
 541         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 542
 543         /* copy rc, transno and disp; steal locks */
 544         mds_req_from_lcd(req, lcd);
 545         ldlm_reply_set_disposition(rep, le32_to_cpu(lcd->lcd_last_data));
 546
 547         /* Only replay if create or open actually happened. */
 548         if (!ldlm_reply_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 549                 EXIT;
 550                 return; /* error looking up parent or child */
 551         }
 552
 553         /* If we failed, then we must have failed opening, so don't look for
 554          * file descriptor or anything, just give the client the bad news.
 555          */
 556         if (req->rq_status) {
 557                 EXIT;
 558                 return;
 559         }
 560
 561         /* Now let's see if we have file descriptor present.
 562          * No need to lookup child as it could be already deleted by another
 563          * thread (bug 15010) */
 564         spin_lock(&med->med_open_lock);
 565         list_for_each(t, &med->med_open_head) {
 566                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 567                 if (mfd->mfd_xid == req->rq_xid) {
 568                         mds_mfd_addref(mfd);
 569                         break;
 570                 }
 571                 mfd = NULL;
 572         }
 573         spin_unlock(&med->med_open_lock);
 574
 575         if (mfd && mfd->mfd_dentry && mfd->mfd_dentry->d_inode) {
 576                 dchild = mfd->mfd_dentry;
 577                 put_child = 0;
 578         } else {
 579                 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 580                 if (IS_ERR(parent)) {
 581                         rc = PTR_ERR(parent);
 582                         LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
 583                                       " Evicting client %s with export %s.\n",
 584                                       rec->ur_fid1->id,rec->ur_fid1->generation,
 585                                       rc, obd_uuid2str(&exp->exp_client_uuid),
 586                                       obd_export_nid2str(exp));
 587                         mds_export_evict(exp);
 588                         EXIT;
 589                         return;
 590                 }
 591
 592                 dchild = mds_lookup(obd, rec->ur_name, parent, rec->ur_namelen - 1);
 593                 l_dput(parent);
 594                 if (IS_ERR(dchild)) {
 595                         rc = PTR_ERR(dchild);
 596                         LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
 597                                       " Evicting client %s with export %s.\n",
 598                                       rec->ur_fid1->id,rec->ur_fid1->generation,
 599                                       rc, obd_uuid2str(&exp->exp_client_uuid),
 600                                       obd_export_nid2str(exp));
 601                         mds_export_evict(exp);
 602                         EXIT;
 603                         return;
 604                 }
 605         }
 606
 607         if (!dchild->d_inode)
 608                 GOTO(out_dput, 0); /* child not present to open */
 609
 610         /* At this point, we know we have a child. We'll send
 611          * it back _unless_ it not created and open failed.
 612          */
 613         if (ldlm_reply_disposition(rep, DISP_OPEN_OPEN) &&
 614             !ldlm_reply_disposition(rep, DISP_OPEN_CREATE) &&
 615             req->rq_status) {
 616                 GOTO(out_dput, 0);
 617         }
 618
 619         mds_pack_inode2body(body, dchild->d_inode);
 620         if (S_ISREG(dchild->d_inode->i_mode)) {
 621                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 622                                  body, dchild->d_inode, 1, 0,
 623                                  req->rq_export->exp_connect_flags);
 624
 625                 if (rc)
 626                         LASSERT(rc == req->rq_status);
 627
 628                 /* If we have LOV EA data, the OST holds size, mtime */
 629                 if (!(body->valid & OBD_MD_FLEASIZE))
 630                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 631                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 632         }
 633
 634         if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 635                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 636                                     body->eadatasize, 0);
 637
 638         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 639             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 640                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 641
 642                 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
 643                                   body, acl_off);
 644                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 645                 if (!req->rq_status && rc)
 646                         req->rq_status = rc;
 647         }
 648
 649         /* #warning "XXX fixme" bug 2991 */
 650         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 651          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 652          * to detect a re-open */
 653         if (mfd == NULL) {
 654                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 655 #if LUSTRE_FIX >= 50
 656                         /* Allow file join in beta builds to allow debugging */
 657                         rc = mds_join_file(rec, req, dchild, NULL);
 658                         if (rc)
 659                                 GOTO(out_dput, rc);
 660 #else
 661                         CWARN("file join is not supported in this version of "
 662                               "Lustre\n");
 663                         GOTO(out_dput, req->rq_status = rc = -EOPNOTSUPP);
 664 #endif
 665                 }
 666                 mntget(mds->mds_vfsmnt);
 667                 CERROR("Re-opened file \n");
 668                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 669                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 670                 mntput(mds->mds_vfsmnt);
 671                 if (IS_ERR(mfd)) {
 672                         req->rq_status = PTR_ERR(mfd);
 673                         mfd = NULL;
 674                         CERROR("%s: opening inode "LPU64" failed: rc %d\n",
 675                                req->rq_export->exp_obd->obd_name,
 676                                (__u64)dchild->d_inode->i_ino, req->rq_status);
 677                         GOTO(out_dput, req->rq_status);
 678                 }
 679         } else {
 680                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 681                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 682                        mfd->mfd_handle.h_cookie);
 683         }
 684
 685         if (!ldlm_reply_disposition(rep, DISP_OPEN_LOCK))
 686                 GOTO(out_dput, 0);
 687
 688         /* child_lockh has been set in fixup_handle_for_resent_req called
 689          * in mds_intent_policy for resent request */
 690         if (child_lockh == NULL || !lustre_handle_is_used(child_lockh)) {
 691                 /* the lock is already canceled! clear DISP_OPEN_LOCK */
 692                 ldlm_reply_clear_disposition(rep, DISP_OPEN_LOCK);
 693         }
 694
 695  out_dput:
 696         if (mfd)
 697                 mds_mfd_put(mfd);
 698
 699         if (put_child)
 700                 l_dput(dchild);
 701         EXIT;
 702 }
 703
 704 /* if client disconnects during recovery it may resend opens which were replayed
 705  * on server but their transno less then last_transno on server so they will not
 706  * be detected as reconstructs */
 707 static int open_replay_reconstruct(struct ptlrpc_request *req)
 708 {
 709         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 710         struct mds_file_data *mfd = NULL;
 711         struct list_head *t;
 712
 713         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
 714                 return 0;
 715
 716         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 717                 return 0;
 718
 719         /* if mfd exists then replay was done already */
 720         spin_lock(&med->med_open_lock);
 721         list_for_each(t, &med->med_open_head) {
 722                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 723                 if (mfd->mfd_xid == req->rq_xid) {
 724                         mds_mfd_addref(mfd);
 725                         break;
 726                 }
 727                 mfd = NULL;
 728         }
 729         spin_unlock(&med->med_open_lock);
 730
 731         if (mfd) {
 732                 struct mds_body *body = lustre_msg_buf(req->rq_repmsg,
 733                                                        DLM_REPLY_REC_OFF,
 734                                                        sizeof(*body));
 735                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
 736
 737                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 738                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 739                        mfd->mfd_handle.h_cookie);
 740                 mds_mfd_put(mfd);
 741                 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
 742                 lustre_msg_set_transno(req->rq_repmsg,
 743                                        lustre_msg_get_transno(req->rq_reqmsg));
 744                 lustre_msg_set_status(req->rq_repmsg, 0);
 745                 return 1;
 746         }
 747         return 0;
 748 }
 749
 750 /* do NOT or the MAY_*'s, you'll get the weakest */
 751 static int accmode(struct inode *inode, int flags)
 752 {
 753         int res = 0;
 754
 755         /* Sadly, NFSD reopens a file repeatedly during operation, so the
 756          * "acc_mode = 0" allowance for newly-created files isn't honoured.
 757          * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
 758          * owner can write to a file even if it is marked readonly to hide
 759          * its brokenness. (bug 5781) */
 760         if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
 761                 return 0;
 762
 763         if (flags & FMODE_READ)
 764                 res = MAY_READ;
 765         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 766                 res |= MAY_WRITE;
 767         if (flags & MDS_FMODE_EXEC)
 768                 res = MAY_EXEC;
 769         return res;
 770 }
 771
 772 /* Handles object creation, actual opening, and I/O epoch */
 773 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 774                            struct mds_body *body, int flags, void **handle,
 775                            struct mds_update_record *rec,struct ldlm_reply *rep,
 776                            struct lustre_handle *lockh)
 777 {
 778         struct mds_obd *mds = mds_req2mds(req);
 779         struct obd_device *obd = req->rq_export->exp_obd;
 780         struct mds_file_data *mfd = NULL;
 781         struct lov_mds_md *lmm = NULL; /* object IDs created */
 782         int rc = 0;
 783         ENTRY;
 784
 785         /* atomically create objects if necessary */
 786         LOCK_INODE_MUTEX(dchild->d_inode);
 787
 788         if (S_ISREG(dchild->d_inode->i_mode) &&
 789             !(body->valid & OBD_MD_FLEASIZE)) {
 790                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 791                                  body, dchild->d_inode, 0, 0,
 792                                  req->rq_export->exp_connect_flags);
 793                 if (rc) {
 794                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 795                         RETURN(rc);
 796                 }
 797         }
 798         if (rec != NULL) {
 799                 if ((body->valid & OBD_MD_FLEASIZE) &&
 800                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 801                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 802                         RETURN(-EEXIST);
 803                 }
 804                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 805 #if LUSTRE_FIX >= 50
 806                         /* Allow file join in beta builds to allow debugging */
 807                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 808                         rc = mds_join_file(rec, req, dchild, lockh);
 809                         if (rc)
 810                                 RETURN(rc);
 811                         LOCK_INODE_MUTEX(dchild->d_inode);
 812 #else
 813                         CWARN("file join is not supported in this version of "
 814                               "Lustre\n");
 815                         RETURN(-EOPNOTSUPP);
 816 #endif
 817                 }
 818                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 819                     !(body->valid & OBD_MD_FLMODEASIZE)) {
 820                         /* split open transactions here */
 821                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_SPLIT_OPEN, 10);
 822                         /* no EA: create objects */
 823                         rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
 824                                                 mds, obd, dchild, handle, &lmm);
 825                         if (rc) {
 826                                 CERROR("mds_create_objects: rc = %d\n", rc);
 827                                 UNLOCK_INODE_MUTEX(dchild->d_inode);
 828                                 RETURN(rc);
 829                         }
 830                 }
 831         }
 832         /* If the inode has no EA data, then MDS holds size, mtime */
 833         if (S_ISREG(dchild->d_inode->i_mode) &&
 834             !(body->valid & OBD_MD_FLEASIZE)) {
 835                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 836                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 837         }
 838         UNLOCK_INODE_MUTEX(dchild->d_inode);
 839
 840         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 841             rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 842                 int acl_off = DLM_REPLY_REC_OFF + 2;
 843
 844                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 845                                   dchild->d_inode, req->rq_repmsg,
 846                                   body, acl_off);
 847                 if (rc)
 848                         RETURN(rc);
 849         }
 850
 851         if ((rc = mds_lov_prepare_objids(obd,lmm)) != 0)
 852                 RETURN(rc);
 853
 854         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 855         if (IS_ERR(mfd))
 856                 RETURN(PTR_ERR(mfd));
 857
 858         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 859                mfd->mfd_handle.h_cookie);
 860
 861         mds_lov_update_objids(obd, lmm);
 862
 863         if (rc)
 864                 mds_mfd_unlink(mfd, 1);
 865
 866         mds_mfd_put(mfd);
 867         RETURN(rc);
 868 }
 869
 870 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 871                            struct mds_body *body, int flags,
 872                            struct mds_update_record *rec,struct ldlm_reply *rep)
 873 {
 874         struct obd_device *obd = req->rq_export->exp_obd;
 875         struct mds_obd *mds = mds_req2mds(req);
 876         struct dentry *dchild, *dparent = NULL;
 877         char fidname[LL_FID_NAMELEN];
 878         int fidlen = 0, rc;
 879         void *handle = NULL;
 880         struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
 881         ENTRY;
 882
 883         ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
 884         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 885         dchild = mds_lookup(obd, fidname, mds->mds_pending_dir, fidlen);
 886         if (IS_ERR(dchild)) {
 887                 rc = PTR_ERR(dchild);
 888                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 889                 RETURN(rc);
 890         }
 891
 892         if (dchild->d_inode != NULL) {
 893                 mds_inode_set_orphan(dchild->d_inode);
 894                 CWARN("Orphan %s found and opened in PENDING directory\n",
 895                        fidname);
 896         } else {
 897                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
 898                 l_dput(dchild);
 899
 900                 /* We didn't find it in PENDING so it isn't an orphan.  See
 901                  * if it was a regular inode that was previously created. */
 902                 dchild = mds_fid2dentry(mds, fid, NULL);
 903                 if (IS_ERR(dchild))
 904                         RETURN(PTR_ERR(dchild));
 905                 /**
 906                  * bug19224
 907                  * this can be replay of partially committed open|create,
 908                  * the create itself was committed while LOV EA weren't
 909                  * We need to set versions again if conditions are:
 910                  * - this is replay
 911                  * - the transaction is greater than last_committed
 912                  * - this was open|create
 913                  * - there was real create so parent pre_version was saved
 914                  */
 915                 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
 916                     (lustre_msg_get_transno(req->rq_reqmsg) >
 917                      req->rq_export->exp_last_committed) &&
 918                     (rec->ur_flags & MDS_OPEN_CREAT) &&
 919                     (pre_versions && pre_versions[0] != 0)) {
 920                         /* need parent to set version */
 921                         dparent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 922                         if (IS_ERR(dparent)) {
 923                                 CERROR("Can't find parent for open replay\n");
 924                                 l_dput(dchild);
 925                                 RETURN(PTR_ERR(dparent));
 926                         }
 927                         /* though file was created, the versions were not
 928                          * changed yet, need to replay that too */
 929                         inodes[0] = dparent->d_inode;
 930                         inodes[1] = dchild->d_inode;
 931                 }
 932         }
 933
 934         mds_pack_inode2body(body, dchild->d_inode);
 935         ldlm_reply_set_disposition(rep, DISP_LOOKUP_POS);
 936         ldlm_reply_set_disposition(rep, DISP_OPEN_OPEN);
 937
 938         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
 939         rc = mds_finish_transno(mds, inodes, handle,
 940                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
 941         /* XXX what do we do here if mds_finish_transno itself failed? */
 942
 943         l_dput(dparent);
 944         l_dput(dchild);
 945         RETURN(rc);
 946 }
 947
 948 int mds_pin(struct ptlrpc_request *req, int offset)
 949 {
 950         struct obd_device *obd = req->rq_export->exp_obd;
 951         struct mds_body *reqbody, *repbody;
 952         struct lvfs_run_ctxt saved;
 953         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 954         ENTRY;
 955
 956         reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
 957
 958         rc = lustre_pack_reply(req, 2, size, NULL);
 959         if (rc)
 960                 RETURN(rc);
 961
 962         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 963                                  sizeof(*repbody));
 964
 965         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 966         rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
 967                              NULL);
 968         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 969
 970         RETURN(rc);
 971 }
 972
 973 /*  Get an internal lock on the inode number (but not generation) to sync
 974  *  new inode creation with inode unlink (bug 2029).  If child_lockh is NULL
 975  *  we just get the lock as a barrier to wait for other holders of this lock,
 976  *  and drop it right away again. */
 977 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 978                        struct lustre_handle *child_lockh)
 979 {
 980         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 981         struct lustre_handle lockh;
 982         int lock_flags = LDLM_FL_ATOMIC_CB;
 983         int rc;
 984
 985         if (child_lockh == NULL)
 986                 child_lockh = &lockh;
 987
 988         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
 989                                     LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 990                                     ldlm_blocking_ast, ldlm_completion_ast,
 991                                     NULL, NULL, 0, NULL, child_lockh);
 992         if (rc != ELDLM_OK)
 993                 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
 994         else if (child_lockh == &lockh)
 995                 ldlm_lock_decref(child_lockh, LCK_EX);
 996
 997         RETURN(rc);
 998 }
 999
1000 int mds_open(struct mds_update_record *rec, int offset,
1001              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
1002 {
1003         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
1004         struct obd_device *obd = req->rq_export->exp_obd;
1005         struct mds_obd *mds = mds_req2mds(req);
1006         struct ldlm_reply *rep = NULL;
1007         struct mds_body *body = NULL;
1008         struct dentry *dchild = NULL, *dparent = NULL;
1009         struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
1010         struct mds_export_data *med;
1011         struct lustre_handle parent_lockh;
1012         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
1013         int parent_mode = LCK_CR;
1014         void *handle = NULL;
1015         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1016         unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
1017         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1018         unsigned int ids[MAXQUOTAS] = { 0, 0 };
1019         int child_mode = LCK_CR;
1020         /* Always returning LOOKUP lock if open succesful to guard
1021            dentry on client. */
1022         int lock_flags = 0;
1023         int quota_pending[2] = {0, 0};
1024         int use_parent, need_open_lock;
1025         unsigned int gid = current->fsgid;
1026         ENTRY;
1027
1028         mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
1029
1030         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
1031                          (obd_timeout + 1) / 4);
1032
1033         CLASSERT(MAXQUOTAS < 4);
1034         if (offset == DLM_INTENT_REC_OFF) { /* intent */
1035                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
1036                                      sizeof(*rep));
1037                 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1038                                       sizeof(*body));
1039         } else if (offset == REQ_REC_OFF) { /* non-intent reint */
1040                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1041                                       sizeof(*body));
1042                 LBUG(); /* XXX: not supported yet? */
1043         } else {
1044                 body = NULL;
1045                 LBUG();
1046         }
1047
1048         /* check the open resent|replay case */
1049         if (open_replay_reconstruct(req))
1050                 RETURN(0);
1051
1052         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
1053
1054         /* Step 0: If we are passed a fid, then we assume the client already
1055          * opened this file and is only replaying the RPC, so we open the
1056          * inode by fid (at some large expense in security). */
1057         /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
1058         if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1059              (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
1060             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
1061                 if (rec->ur_fid2->id == 0) {
1062                         struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
1063                         if (lock) {
1064                                 LDLM_ERROR(lock, "fid2 not set on open replay");
1065                                 LDLM_LOCK_PUT(lock);
1066                         }
1067                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
1068                         RETURN(-EFAULT);
1069                 }
1070
1071                 /** check there is no stale orphan with same inode number */
1072                 if (rec->ur_flags & MDS_OPEN_CREAT) {
1073                         rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1074                         if (rc)
1075                                 RETURN(rc);
1076                 }
1077
1078                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
1079                                      rec, rep);
1080                 if (rc != -ENOENT) {
1081                         if (req->rq_export->exp_libclient &&
1082                             rec->ur_flags & MDS_OPEN_HAS_EA)
1083                                 RETURN(0);
1084
1085                         RETURN(rc);
1086                 }
1087
1088                 /* We didn't find the correct inode on disk either, so we
1089                  * need to re-create it via a regular replay. */
1090                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1091                         DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
1092                         RETURN(-EFAULT);
1093                 }
1094         } else if (rec->ur_fid2->id) {
1095                 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
1096                           rec->ur_fid2->id, rec->ur_fid2->generation);
1097                 RETURN(-EFAULT);
1098         }
1099
1100         /* If we got here, we must be called via intent */
1101         LASSERT(offset == DLM_INTENT_REC_OFF);
1102
1103         med = &req->rq_export->exp_mds_data;
1104         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1105                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1106                 RETURN(-ENOMEM);
1107         }
1108
1109         if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
1110                 parent_mode = LCK_EX;
1111
1112         /* We cannot use acc_mode here, because it is zeroed in case of
1113            creating a file, so we get wrong lockmode */
1114         if (rec->ur_flags & FMODE_WRITE)
1115                 child_mode = LCK_CW;
1116         else if (rec->ur_flags & MDS_FMODE_EXEC)
1117                 child_mode = LCK_PR;
1118         else
1119                 child_mode = LCK_CR;
1120
1121         /* join file and nfsd can't need lookup dchild as use parent for it */
1122         use_parent = (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1123                      (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) ||
1124                      (rec->ur_flags & MDS_OPEN_JOIN_FILE);
1125
1126         need_open_lock = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1127                            (rec->ur_flags & MDS_OPEN_LOCK);
1128
1129         /* Try to lock both parent and child first. If child is not found,
1130          * return only locked parent.  This is enough to prevent other
1131          * threads from changing this directory until creation is finished. */
1132         rc = mds_get_parent_child_locked(obd, &obd->u.mds,
1133                                          rec->ur_fid1,
1134                                          &parent_lockh,
1135                                          &dparent, parent_mode,
1136                                          MDS_INODELOCK_UPDATE,
1137                                          use_parent ? NULL : rec->ur_name,
1138                                          rec->ur_namelen,
1139                                          (rec->ur_flags & MDS_OPEN_LOCK) ?
1140                                                 child_lockh : NULL,
1141                                          &dchild, child_mode,
1142                                          MDS_INODELOCK_LOOKUP |
1143                                          MDS_INODELOCK_OPEN);
1144
1145         if (rc) {
1146                 if (rc != -ENOENT) {
1147                         CERROR("parent "LPU64"/%u lookup/take lock error %d\n",
1148                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
1149                 } else {
1150                         /* Just cannot find parent - make it look like
1151                          * usual negative lookup to avoid extra MDS RPC */
1152                         ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
1153                         ldlm_reply_set_disposition(rep, DISP_LOOKUP_NEG);
1154                 }
1155                 GOTO(cleanup, rc);
1156         }
1157         LASSERT(dparent->d_inode != NULL);
1158
1159         cleanup_phase = 1; /* parent dentry and lock */
1160
1161         if (use_parent)
1162                 dchild = dget(dparent);
1163
1164         if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
1165                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1166                 GOTO(found_child, rc);
1167         }
1168
1169         cleanup_phase = 2; /* child dentry */
1170
1171         ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
1172         if (dchild->d_inode)
1173                 ldlm_reply_set_disposition(rep, DISP_LOOKUP_POS);
1174         else
1175                 ldlm_reply_set_disposition(rep, DISP_LOOKUP_NEG);
1176
1177         /*Step 3: If the child was negative, and we're supposed to, create it.*/
1178         if (dchild->d_inode == NULL) {
1179                 unsigned long ino = rec->ur_fid2->id;
1180                 struct iattr iattr;
1181                 struct inode *inode;
1182
1183                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1184                         /* It's negative and we weren't supposed to create it */
1185                         GOTO(cleanup, rc = -ENOENT);
1186                 }
1187
1188                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1189                         GOTO(cleanup, rc = -EROFS);
1190
1191                 /** check there is no stale orphan with same inode number */
1192                 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1193                 if (rc)
1194                         GOTO(cleanup, rc);
1195
1196                 /* version recovery check */
1197                 rc = mds_version_get_check(req, dparent->d_inode, 0);
1198                 if (rc)
1199                         GOTO(cleanup_no_trans, rc);
1200
1201                 if (dparent->d_inode->i_mode & S_ISGID)
1202                         gid = dparent->d_inode->i_gid;
1203                 else
1204                         gid = current->fsgid;
1205
1206                 /* we try to get enough quota to write here, and let ldiskfs
1207                  * decide if it is out of quota or not b=14783
1208                  * FIXME: after CMD is used, pointer to obd_trans_info* couldn't
1209                  * be NULL, b=14840 */
1210                 ids[0] = current->fsuid;
1211                 ids[1] = gid;
1212                 lquota_chkquota(mds_quota_interface_ref, req->rq_export,
1213                                 ids[0], ids[1], 1, quota_pending,
1214                                 NULL, NULL, 0);
1215
1216                 ldlm_reply_set_disposition(rep, DISP_OPEN_CREATE);
1217                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1218                                       NULL);
1219                 if (IS_ERR(handle)) {
1220                         rc = PTR_ERR(handle);
1221                         handle = NULL;
1222                         GOTO(cleanup, rc);
1223                 }
1224                 dchild->d_fsdata = (void *) &dp;
1225                 dp.ldp_ptr = req;
1226                 dp.ldp_inum = ino;
1227
1228                 LOCK_INODE_MUTEX(dparent->d_inode);
1229                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1230                 UNLOCK_INODE_MUTEX(dparent->d_inode);
1231                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1232                         dchild->d_fsdata = NULL;
1233
1234                 if (rc) {
1235                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1236                         GOTO(cleanup, rc);
1237                 }
1238                 inode = dchild->d_inode;
1239                 created = 1;
1240                 if (ino) {
1241                         if (ino != inode->i_ino) {
1242                                 /* FID support is needed to replay this
1243                                  * correctly. Now fail gracefully like there is
1244                                  * version mismatch */
1245                                 if (req->rq_export->exp_delayed)
1246                                         rc = -EOVERFLOW;
1247                                 else
1248                                         rc = -EFAULT;
1249                                 GOTO(cleanup, rc);
1250                         }
1251                         /* Written as part of setattr */
1252                         inode->i_generation = rec->ur_fid2->generation;
1253                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1254                                inode->i_ino, inode->i_generation);
1255                 }
1256
1257                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1258                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1259                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1260
1261                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
1262                 iattr.ia_gid = gid;
1263
1264                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1265                         ATTR_MTIME | ATTR_CTIME;
1266
1267                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1268                 if (rc)
1269                         CERROR("error on child setattr: rc = %d\n", rc);
1270
1271                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1272
1273                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1274                 if (rc)
1275                         CERROR("error on parent setattr: rc = %d\n", rc);
1276
1277                 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1278                 handle = NULL;
1279                 acc_mode = 0;           /* Don't check for permissions */
1280         } else {
1281                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1282                 /* Child previously existed so the lookup and lock is already
1283                  * done, so no further locking is needed. */
1284                 /* for nfs and join - we need two locks for same fid, but
1285                  * with different mode */
1286                 if (need_open_lock && !use_parent)  {
1287                         ldlm_reply_set_disposition(rep, DISP_OPEN_LOCK);
1288                         need_open_lock = 0;
1289                 }
1290         }
1291
1292         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1293                  "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1294                  dchild->d_name.name, dchild, dchild->d_inode,
1295                  dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1296
1297 found_child:
1298         mds_pack_inode2body(body, dchild->d_inode);
1299         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1300             (rec->ur_flags & MDS_OPEN_EXCL)) {
1301                 /* File already exists, we didn't just create it, and we
1302                  * were passed O_EXCL; err-or. */
1303                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1304         }
1305
1306         /* if we are following special file, don't open */
1307         if (!S_ISREG(dchild->d_inode->i_mode) &&
1308             !S_ISDIR(dchild->d_inode->i_mode))
1309                 GOTO(cleanup_no_trans, rc = 0);
1310
1311         ldlm_reply_set_disposition(rep, DISP_OPEN_OPEN);
1312         if (S_ISREG(dchild->d_inode->i_mode)) {
1313                 /* Check permissions etc */
1314                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1315                 if (rc != 0)
1316                         GOTO(cleanup, rc);
1317
1318                 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1319                     (acc_mode & MAY_WRITE))
1320                         GOTO(cleanup, rc = -EROFS);
1321
1322                 /* An append-only file must be opened in append mode for
1323                  * writing */
1324                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1325                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1326                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1327                         GOTO(cleanup, rc = -EPERM);
1328         } else {
1329                 if (S_ISDIR(dchild->d_inode->i_mode)) {
1330                         if (rec->ur_flags & MDS_OPEN_CREAT ||
1331                             rec->ur_flags & FMODE_WRITE) {
1332                                 /* we are trying to create or write a exist dir*/
1333                                 GOTO(cleanup, rc = -EISDIR);
1334                         }
1335                         if (rec->ur_flags & MDS_FMODE_EXEC) {
1336                                 /* we are trying to exec a directory */
1337                                 GOTO(cleanup, rc = -EACCES);
1338                         }
1339                         if (ll_permission(dchild->d_inode, acc_mode, NULL))
1340                                 GOTO(cleanup, rc = -EACCES);
1341                 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1342                         GOTO(cleanup, rc = -ENOTDIR);
1343                 }
1344         }
1345
1346         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1347                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1348                 GOTO(cleanup, rc = -EAGAIN);
1349         }
1350
1351         if (need_open_lock) {
1352                 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN } };
1353                 struct ldlm_res_id child_res_id;
1354
1355                 /* In case of replay we do not get a lock assuming that the
1356                    caller has it already */
1357                 child_res_id.name[0] = dchild->d_inode->i_ino;
1358                 child_res_id.name[1] = dchild->d_inode->i_generation;
1359
1360                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1361                                             LDLM_IBITS, &policy, child_mode,
1362                                             &lock_flags, ldlm_blocking_ast,
1363                                             ldlm_completion_ast, NULL, NULL,
1364                                             0, NULL, child_lockh);
1365                 if (rc != ELDLM_OK)
1366                         GOTO(cleanup, rc);
1367
1368                 /* Let mds_intent_policy know that we have a lock to return */
1369                 ldlm_reply_set_disposition(rep, DISP_OPEN_LOCK);
1370         }
1371
1372         if (!S_ISREG(dchild->d_inode->i_mode) &&
1373             !S_ISDIR(dchild->d_inode->i_mode) &&
1374             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1375                 /* If client supports this, do not return open handle for
1376                  * special device nodes */
1377                 GOTO(cleanup_no_trans, rc = 0);
1378         }
1379
1380         /* Step 5: mds_open it */
1381         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1382                              rep, &parent_lockh);
1383         GOTO(cleanup, rc);
1384
1385  cleanup:
1386         inodes[0] = (!created || IS_ERR(dparent)) ? NULL : dparent->d_inode;
1387         inodes[1] = (created && dchild) ? dchild->d_inode : NULL;
1388         rc = mds_finish_transno(mds, inodes, handle, req, rc,
1389                                 rep ? rep->lock_policy_res1 : 0, 0);
1390
1391  cleanup_no_trans:
1392         if (quota_pending[0] || quota_pending[1])
1393                 lquota_pending_commit(mds_quota_interface_ref, obd,
1394                                       ids[0], ids[1], quota_pending);
1395         switch (cleanup_phase) {
1396         case 2:
1397                 if (rc && created) {
1398                         int err;
1399                         LOCK_INODE_MUTEX(dparent->d_inode);
1400                         err = ll_vfs_unlink(dparent->d_inode, dchild,
1401                                             mds->mds_vfsmnt);
1402                         UNLOCK_INODE_MUTEX(dparent->d_inode);
1403                         if (err) {
1404                                 CERROR("unlink(%.*s) in error path: %d\n",
1405                                        dchild->d_name.len, dchild->d_name.name,
1406                                        err);
1407                         }
1408                 } else if (created) {
1409                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1410                         /* save uid/gid for quota acquire/release */
1411                         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1412                         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1413                 }
1414         case 1:
1415                 if (dchild) {
1416                         l_dput(dchild);
1417                         /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1418                          * mds_intent_policy won't try to return any locks */
1419                         if (rc && child_lockh->cookie)
1420                                 ldlm_lock_decref(child_lockh, child_mode);
1421                 }
1422                 if (dparent == NULL)
1423                         break;
1424
1425                 l_dput(dparent);
1426                 if (rc)
1427                         ldlm_lock_decref(&parent_lockh, parent_mode);
1428                 else
1429                         ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1430         }
1431         /* trigger dqacq on the owner of child and parent */
1432         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1433                       FSFILT_OP_CREATE);
1434
1435         RETURN(rc);
1436 }
1437
1438 /* Close a "file descriptor" and possibly unlink an orphan from the
1439  * PENDING directory.  Caller must hold child->i_mutex, this drops it.
1440  *
1441  * If we are being called from mds_disconnect() because the client has
1442  * disappeared, then req == NULL and we do not update last_rcvd because
1443  * there is nothing that could be recovered by the client at this stage
1444  * (it will not even _have_ an entry in last_rcvd anymore). */
1445 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1446                   struct obd_device *obd, struct mds_file_data *mfd,
1447                   int unlink_orphan, struct lov_mds_md *lmm, int lmm_size,
1448                   struct llog_cookie *logcookies, int cookies_size,
1449                   __u64 *valid)
1450 {
1451         struct inode *inode = mfd->mfd_dentry->d_inode;
1452         char fidname[LL_FID_NAMELEN];
1453         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1454         struct dentry *pending_child = NULL;
1455         struct mds_obd *mds = &obd->u.mds;
1456         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1457         void *handle = NULL;
1458         struct mds_body *request_body = NULL, *reply_body = NULL;
1459         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1460         struct iattr iattr = { 0 };
1461         ENTRY;
1462
1463         if (req && req->rq_reqmsg != NULL)
1464                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1465                                               sizeof(*request_body));
1466         if (req && req->rq_repmsg != NULL)
1467                 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1468                                             sizeof(*reply_body));
1469
1470         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1471
1472         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1473                inode->i_nlink, mds_orphan_open_count(inode));
1474
1475         last_orphan = (mds_orphan_open_dec_test(inode) &&
1476                        mds_inode_is_orphan(inode) &&
1477                        !obd->obd_recovering);
1478
1479         /* this is half of the actual "close" */
1480         if (mfd->mfd_mode & FMODE_WRITE) {
1481                 rc = mds_put_write_access(mds, inode, request_body,
1482                                           last_orphan && unlink_orphan);
1483         } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1484                 mds_allow_write_access(inode);
1485         }
1486         /* here writecount change also needs protection from orphan write sem.
1487          * so drop orphan write sem after mds_put_write_access, bz 12888. */
1488         MDS_UP_WRITE_ORPHAN_SEM(inode);
1489
1490         if (last_orphan && unlink_orphan) {
1491                 int stripe_count = 0;
1492                 /* mds_put_write_access must have succeeded */
1493                 LASSERTF(rc == 0, "inode %lu/%u: rc %d",
1494                          inode->i_ino, inode->i_generation, rc);
1495
1496                 CDEBUG(D_INODE, "destroying orphan object %s\n", fidname);
1497
1498                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1499                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1500                         CERROR("found \"orphan\" %s %s with link count %d\n",
1501                                S_ISREG(inode->i_mode) ? "file" : "dir",
1502                                fidname, inode->i_nlink);
1503
1504                 /* Sadly, there is no easy way to save pending_child from
1505                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1506                  * but normally it will still be in the dcache. */
1507                 LOCK_INODE_MUTEX(pending_dir);
1508                 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1509                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1510                                                fidlen);
1511                 if (IS_ERR(pending_child))
1512                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1513                 LASSERT(pending_child->d_inode != NULL);
1514
1515                 cleanup_phase = 2; /* dput(pending_child) when finished */
1516                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1517                         rc = ll_vfs_rmdir(pending_dir, pending_child,
1518                                           mds->mds_vfsmnt);
1519                         if (rc)
1520                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1521                                        fidname,rc);
1522                         goto out;
1523                 }
1524
1525                 if (lmm != NULL) {
1526                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1527                 }
1528
1529                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1530                                           NULL, stripe_count);
1531                 if (IS_ERR(handle)) {
1532                         rc = PTR_ERR(handle);
1533                         handle = NULL;
1534                         GOTO(cleanup, rc);
1535                 }
1536                 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1537                     mds_log_op_unlink(obd, lmm, lmm_size,
1538                                       logcookies, cookies_size) > 0) {
1539                         *valid |= OBD_MD_FLCOOKIE;
1540                 }
1541
1542                 dp.ldp_inum = 0;
1543                 dp.ldp_ptr = req;
1544                 pending_child->d_fsdata = (void *) &dp;
1545                 rc = ll_vfs_unlink(pending_dir, pending_child, mds->mds_vfsmnt);
1546                 if (rc)
1547                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1548
1549                 goto out; /* Don't bother updating attrs on unlinked inode */
1550         }
1551
1552         if (request_body != NULL) {
1553                 /* Only start a transaction to write out only the atime if it
1554                  * is more out-of-date than the specified limit.  If we are
1555                  * already going to write out the atime then do it anyway. */
1556                 if ((request_body->valid & OBD_MD_FLATIME) &&
1557                     ((LTIME_S(iattr.ia_atime) >
1558                       LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1559                      (iattr.ia_valid != 0 &&
1560                       LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))) {
1561                         LTIME_S(iattr.ia_atime) = request_body->atime;
1562                         iattr.ia_valid |= ATTR_ATIME;
1563                 }
1564
1565                 /* Store a rough estimate of the file size on the MDS for
1566                  * tools like e2scan and HSM that are just using this for *
1567                  * rough decision making and will get the proper size later.
1568                  * * This is NOT guaranteed to be correct with multiple *
1569                  * writers, but is only needed until SOM is done. b=11063 */
1570                 if (S_ISREG(inode->i_mode) &&
1571                     (request_body->valid & OBD_MD_FLSIZE) &&
1572                     (iattr.ia_valid != 0)) {
1573                         iattr.ia_size = request_body->size;
1574                         iattr.ia_valid |= ATTR_SIZE;
1575                 }
1576         }
1577
1578         if (iattr.ia_valid != 0) {
1579                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1580                 if (IS_ERR(handle)) {
1581                         rc = PTR_ERR(handle);
1582                         handle = NULL;
1583                         GOTO(cleanup, rc);
1584                 }
1585                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1586                 if (rc)
1587                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1588         }
1589 out:
1590         /* If other clients have this file open for write, rc will be > 0 */
1591         if (rc > 0)
1592                 rc = 0;
1593
1594  cleanup:
1595         l_dput(mfd->mfd_dentry);
1596         mds_mfd_put(mfd);
1597         if (req != NULL && reply_body != NULL) {
1598                 rc = mds_finish_transno(mds, NULL, handle, req, rc, 0, 0);
1599         } else if (handle) {
1600                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1601                 if (err) {
1602                         CERROR("error committing close: %d\n", err);
1603                         if (!rc)
1604                                 rc = err;
1605                 }
1606         }
1607
1608         switch (cleanup_phase) {
1609         case 2:
1610                 dput(pending_child);
1611         case 1:
1612                 UNLOCK_INODE_MUTEX(pending_dir);
1613         }
1614         RETURN(rc);
1615 }
1616
1617 int mds_close(struct ptlrpc_request *req, int offset)
1618 {
1619         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1620         struct obd_device *obd = req->rq_export->exp_obd;
1621         struct mds_body *body;
1622         struct mds_file_data *mfd;
1623         struct lvfs_run_ctxt saved;
1624         struct inode *inode;
1625         int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1626                                sizeof(struct mds_body),
1627                                obd->u.mds.mds_max_mdsize,
1628                                obd->u.mds.mds_max_cookiesize };
1629         struct mds_body *reply_body;
1630         struct lov_mds_md *lmm;
1631         int lmm_size;
1632         struct llog_cookie *logcookies;
1633         int cookies_size;
1634         ENTRY;
1635
1636         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1637                                   lustre_swab_mds_body);
1638         if (body == NULL) {
1639                 CERROR("Can't unpack body\n");
1640                 req->rq_status = -EFAULT;
1641                 GOTO(cleanup, rc = -EFAULT);
1642         }
1643         /*XXX need indicase - close is need to return LOV EA */
1644         body->valid |= OBD_MD_FLEASIZE;
1645
1646         rc = lustre_pack_reply(req, 4, repsize, NULL);
1647         if (rc)
1648                 req->rq_status = rc;
1649                 /* continue on to drop local open even if we can't send reply */
1650         else
1651                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1652
1653         CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n",
1654                req->rq_replen,
1655                obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1656         mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1657
1658         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1659                 /* do some stuff */ ;
1660
1661         spin_lock(&med->med_open_lock);
1662         mfd = mds_handle2mfd(&body->handle);
1663         if (mfd == NULL) {
1664                 spin_unlock(&med->med_open_lock);
1665                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1666                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1667                 req->rq_status = -ESTALE;
1668                 GOTO(cleanup, rc = -ESTALE);
1669         }
1670         /* Remove mfd handle so it can't be found again.  We consume mfd_list
1671          * reference here, but still have mds_handle2mfd ref until mfd_close. */
1672         mds_mfd_unlink(mfd, 1);
1673         spin_unlock(&med->med_open_lock);
1674
1675         inode = mfd->mfd_dentry->d_inode;
1676         /* child orphan sem protects orphan_dec_test && is_orphan race */
1677         MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1678         if (!obd->obd_recovering &&
1679             mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1680                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1681                                       sizeof(*body));
1682                 LASSERT(body != NULL);
1683
1684                 mds_pack_inode2body(body, inode);
1685                 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1686                             MDS_PACK_MD_LOCK, 0,
1687                             req->rq_export->exp_connect_flags);
1688         }
1689
1690         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1691         reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1692         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1693         lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1694         logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1695         cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1696         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1697                                        lmm, lmm_size, logcookies, cookies_size,
1698                                        &reply_body->valid);
1699         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1700
1701 cleanup:
1702         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1703                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1704                 req->rq_status = -ENOMEM;
1705                 RETURN(-ENOMEM);
1706         }
1707
1708         RETURN(rc);
1709 }
1710
1711 int mds_done_writing(struct ptlrpc_request *req, int offset)
1712 {
1713         struct mds_body *body;
1714         int rc, size[2] = { sizeof(struct ptlrpc_body),
1715                             sizeof(struct mds_body) };
1716         ENTRY;
1717
1718         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1719
1720         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1721                                   lustre_swab_mds_body);
1722         if (body == NULL) {
1723                 CERROR("Can't unpack body\n");
1724                 req->rq_status = -EFAULT;
1725                 RETURN(-EFAULT);
1726         }
1727
1728         rc = lustre_pack_reply(req, 2, size, NULL);
1729         if (rc)
1730                 req->rq_status = rc;
1731
1732         RETURN(0);
1733 }