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 the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/version.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/version.h>
  38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  39 # include <linux/buffer_head.h>
  40 # include <linux/workqueue.h>
  41 #else
  42 # include <linux/locks.h>
  43 #endif
  44
  45 #include <obd_class.h>
  46 #include <obd_lov.h>
  47 #include <lustre_fsfilt.h>
  48 #include <lprocfs_status.h>
  49
  50 #include "mds_internal.h"
  51
  52 /* Exported function from this file are:
  53  *
  54  * mds_open - called by the intent handler
  55  * mds_close - an rpc handling function
  56  * mds_pin - an rpc handling function - which will go away
  57  * mds_mfd_close - for force closing files when a client dies
  58  */
  59
  60 /*
  61  * MDS file data handling: file data holds a handle for a file opened
  62  * by a client.
  63  */
  64
  65 static void mds_mfd_addref(void *mfdp)
  66 {
  67         struct mds_file_data *mfd = mfdp;
  68
  69         atomic_inc(&mfd->mfd_refcount);
  70         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  71                atomic_read(&mfd->mfd_refcount));
  72 }
  73
  74 /* Create a new mds_file_data struct.
  75  * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
  76  * one reference for the caller of this function. */
  77 struct mds_file_data *mds_mfd_new(void)
  78 {
  79         struct mds_file_data *mfd;
  80
  81         OBD_ALLOC(mfd, sizeof *mfd);
  82         if (mfd == NULL) {
  83                 CERROR("mds: out of memory\n");
  84                 return NULL;
  85         }
  86
  87         atomic_set(&mfd->mfd_refcount, 2);
  88
  89         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  90         INIT_LIST_HEAD(&mfd->mfd_list);
  91         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  92
  93         return mfd;
  94 }
  95
  96 /* Get a new reference on the mfd pointed to by handle, if handle is still
  97  * valid.  Caller must drop reference with mds_mfd_put(). */
  98 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  99 {
 100         ENTRY;
 101         LASSERT(handle != NULL);
 102         RETURN(class_handle2object(handle->cookie));
 103 }
 104
 105 /* Drop mfd reference, freeing struct if this is the last one. */
 106 static void mds_mfd_put(struct mds_file_data *mfd)
 107 {
 108         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 109                atomic_read(&mfd->mfd_refcount) - 1);
 110         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 111                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 112         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 113                 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 114         }
 115 }
 116
 117 /* Remove the mfd handle so that it cannot be found by open/close again.
 118  * Caller must hold med_open_lock for mfd_list manipulation. */
 119 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
 120 {
 121         class_handle_unhash(&mfd->mfd_handle);
 122         list_del_init(&mfd->mfd_list);
 123         if (decref)
 124                 mds_mfd_put(mfd);
 125 }
 126
 127 /* Caller must hold mds->mds_epoch_sem */
 128 static int mds_alloc_filterdata(struct inode *inode)
 129 {
 130         LASSERT(inode->i_filterdata == NULL);
 131         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 132         if (inode->i_filterdata == NULL)
 133                 return -ENOMEM;
 134         LASSERT(igrab(inode) == inode);
 135         return 0;
 136 }
 137
 138 /* Caller must hold mds->mds_epoch_sem */
 139 static void mds_free_filterdata(struct inode *inode)
 140 {
 141         LASSERT(inode->i_filterdata != NULL);
 142         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 143         inode->i_filterdata = NULL;
 144         iput(inode);
 145 }
 146
 147 /* Write access to a file: executors cause a negative count,
 148  * writers a positive count.  The semaphore is needed to perform
 149  * a check for the sign and then increment or decrement atomically.
 150  *
 151  * This code is closely tied to the allocation of the d_fsdata and the
 152  * MDS epoch, so we use the same semaphore for the whole lot.
 153  *
 154  * We could use a different semaphore for each file, if it ever shows
 155  * up in a profile, which it won't.
 156  *
 157  * epoch argument is nonzero during recovery */
 158 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 159                                 __u64 epoch)
 160 {
 161         int rc = 0;
 162
 163         down(&mds->mds_epoch_sem);
 164
 165         if (atomic_read(&inode->i_writecount) < 0) {
 166                 up(&mds->mds_epoch_sem);
 167                 RETURN(-ETXTBSY);
 168         }
 169
 170
 171         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 172                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 173                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 174                        inode->i_generation);
 175                 goto out;
 176         }
 177
 178         if (inode->i_filterdata == NULL)
 179                 mds_alloc_filterdata(inode);
 180         if (inode->i_filterdata == NULL) {
 181                 rc = -ENOMEM;
 182                 goto out;
 183         }
 184         if (epoch > mds->mds_io_epoch)
 185                 mds->mds_io_epoch = epoch;
 186         else
 187                 mds->mds_io_epoch++;
 188         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 189         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 190                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 191  out:
 192         if (rc == 0)
 193                 atomic_inc(&inode->i_writecount);
 194         up(&mds->mds_epoch_sem);
 195         return rc;
 196 }
 197
 198 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 199  * again -- which is never true if the file is about to be unlinked.  Otherwise
 200  * returns the number of remaining writers. */
 201 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 202                                 struct mds_body *body, int unlinking)
 203 {
 204         int rc = 0;
 205         ENTRY;
 206
 207         down(&mds->mds_epoch_sem);
 208         atomic_dec(&inode->i_writecount);
 209         rc = atomic_read(&inode->i_writecount);
 210         if (rc > 0)
 211                 GOTO(out, rc);
 212 #if 0
 213         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 214                 GOTO(out, rc = EAGAIN);
 215 #endif
 216         mds_free_filterdata(inode);
 217  out:
 218         up(&mds->mds_epoch_sem);
 219         return rc;
 220 }
 221
 222 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 223 {
 224         ENTRY;
 225         down(&mds->mds_epoch_sem);
 226         if (atomic_read(&inode->i_writecount) > 0) {
 227                 up(&mds->mds_epoch_sem);
 228                 RETURN(-ETXTBSY);
 229         }
 230         atomic_dec(&inode->i_writecount);
 231         up(&mds->mds_epoch_sem);
 232         RETURN(0);
 233 }
 234
 235 static void mds_allow_write_access(struct inode *inode)
 236 {
 237         ENTRY;
 238         atomic_inc(&inode->i_writecount);
 239 }
 240
 241 int mds_query_write_access(struct inode *inode)
 242 {
 243         ENTRY;
 244         RETURN(atomic_read(&inode->i_writecount));
 245 }
 246
 247 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 248 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 249                                              struct vfsmount *mnt, int flags,
 250                                              struct ptlrpc_request *req)
 251 {
 252         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 253         struct mds_obd *mds = mds_req2mds(req);
 254         struct mds_file_data *mfd;
 255         struct mds_body *body;
 256         int error;
 257         ENTRY;
 258
 259         mfd = mds_mfd_new();
 260         if (mfd == NULL) {
 261                 CERROR("mds: out of memory\n");
 262                 GOTO(cleanup_dentry, error = -ENOMEM);
 263         }
 264
 265         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 266
 267         if (flags & FMODE_WRITE) {
 268                 /* FIXME: in recovery, need to pass old epoch here */
 269                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 270                 if (error)
 271                         GOTO(cleanup_mfd, error);
 272                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 273         } else if (flags & MDS_FMODE_EXEC) {
 274                 error = mds_deny_write_access(mds, dentry->d_inode);
 275                 if (error)
 276                         GOTO(cleanup_mfd, error);
 277         }
 278
 279         dget(dentry);
 280
 281         /* Mark the file as open to handle open-unlink. */
 282         MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
 283         mds_orphan_open_inc(dentry->d_inode);
 284         MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
 285
 286         mfd->mfd_mode = flags;
 287         mfd->mfd_dentry = dentry;
 288         mfd->mfd_xid = req->rq_xid;
 289
 290         spin_lock(&med->med_open_lock);
 291         list_add(&mfd->mfd_list, &med->med_open_head);
 292         spin_unlock(&med->med_open_lock);
 293
 294         body->handle.cookie = mfd->mfd_handle.h_cookie;
 295
 296         RETURN(mfd);
 297
 298 cleanup_mfd:
 299         mds_mfd_put(mfd);
 300         mds_mfd_unlink(mfd, 1);
 301 cleanup_dentry:
 302         return ERR_PTR(error);
 303 }
 304
 305 /* Must be called with i_mutex held */
 306 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 307                               struct mds_update_record *rec,
 308                               struct mds_obd *mds, struct obd_device *obd,
 309                               struct dentry *dchild, void **handle,
 310                               struct lov_mds_md **objid)
 311 {
 312         struct inode *inode = dchild->d_inode;
 313         struct obd_trans_info oti = { 0 };
 314         struct lov_mds_md *lmm = NULL;
 315         int rc, lmm_size;
 316         struct mds_body *body;
 317         struct obd_info oinfo = { { { 0 } } };
 318         void *lmm_buf;
 319         ENTRY;
 320
 321         *objid = NULL;
 322
 323         if (!S_ISREG(inode->i_mode))
 324                 RETURN(0);
 325         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 326             !(rec->ur_flags & FMODE_WRITE))
 327                 RETURN(0);
 328
 329         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 330
 331         if (body->valid & OBD_MD_FLEASIZE)
 332                 RETURN(0);
 333
 334         oti_init(&oti, req);
 335
 336         /* replay case */
 337         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 338                 if (rec->ur_fid2->id == 0) {
 339                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 340                         RETURN(-EFAULT);
 341                 }
 342
 343                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 344                 lmm_size = rec->ur_eadatalen;
 345                 lmm = rec->ur_eadata;
 346                 LASSERT(lmm);
 347
 348                 if (*handle == NULL)
 349                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 350                 if (IS_ERR(*handle)) {
 351                         rc = PTR_ERR(*handle);
 352                         *handle = NULL;
 353                         GOTO(out_ids, rc);
 354                 }
 355
 356                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 357                 if (rc)
 358                         CERROR("open replay failed to set md:%d\n", rc);
 359                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 360                 LASSERT(lmm_buf);
 361                 memcpy(lmm_buf, lmm, lmm_size);
 362
 363                 *objid = lmm_buf;
 364                 RETURN(rc);
 365         }
 366
 367         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 368                 GOTO(out_ids, rc = -ENOMEM);
 369
 370         OBDO_ALLOC(oinfo.oi_oa);
 371         if (oinfo.oi_oa == NULL)
 372                 GOTO(out_ids, rc = -ENOMEM);
 373         oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
 374         oinfo.oi_oa->o_gid = 0;
 375         oinfo.oi_oa->o_mode = S_IFREG | 0600;
 376         oinfo.oi_oa->o_id = inode->i_ino;
 377         oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
 378                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
 379         oinfo.oi_oa->o_size = 0;
 380
 381         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 382                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 383
 384         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 385                 /* check if things like lfs setstripe are sending us the ea */
 386                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 387                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 388                                            mds->mds_osc_exp,
 389                                            0, &oinfo.oi_md, rec->ur_eadata);
 390                         if (rc)
 391                                 GOTO(out_oa, rc);
 392                 } else {
 393                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 394                         if (lmm == NULL)
 395                                 GOTO(out_oa, rc = -ENOMEM);
 396
 397                         lmm_size = mds->mds_max_mdsize;
 398                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 399                                         lmm, &lmm_size, 1, 0);
 400                         if (rc > 0)
 401                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 402                                                    mds->mds_osc_exp,
 403                                                    0, &oinfo.oi_md, lmm);
 404                         OBD_FREE(lmm, mds->mds_max_mdsize);
 405                         if (rc)
 406                                 GOTO(out_oa, rc);
 407                 }
 408                 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
 409                                 &oinfo.oi_md, &oti);
 410                 if (rc) {
 411                         int level = D_ERROR;
 412                         if (rc == -ENOSPC)
 413                                 level = D_INODE;
 414                         CDEBUG(level, "error creating objects for "
 415                                       "inode %lu: rc = %d\n",
 416                                inode->i_ino, rc);
 417                         if (rc > 0) {
 418                                 CERROR("obd_create returned invalid "
 419                                        "rc %d\n", rc);
 420                                 rc = -EIO;
 421                         }
 422                         GOTO(out_oa, rc);
 423                 }
 424         } else {
 425                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 426                                    0, &oinfo.oi_md, rec->ur_eadata);
 427                 if (rc) {
 428                         GOTO(out_oa, rc);
 429                 }
 430                 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
 431         }
 432         if (i_size_read(inode)) {
 433                 oinfo.oi_oa->o_size = i_size_read(inode);
 434                 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
 435                                 OBD_MD_FLATIME | OBD_MD_FLMTIME |
 436                                 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 437
 438                 /* pack lustre id to OST */
 439                 oinfo.oi_oa->o_fid = body->fid1.id;
 440                 oinfo.oi_oa->o_generation = body->fid1.generation;
 441                 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 442                 oinfo.oi_policy.l_extent.start = i_size_read(inode);
 443                 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 444
 445                 rc = obd_punch_rqset(mds->mds_osc_exp, &oinfo, &oti);
 446                 if (rc) {
 447                         CERROR("error setting attrs for inode %lu: rc %d\n",
 448                                inode->i_ino, rc);
 449                         if (rc > 0) {
 450                                 CERROR("obd_setattr_async returned bad rc %d\n",
 451                                        rc);
 452                                 rc = -EIO;
 453                         }
 454                         GOTO(out_oa, rc);
 455                 }
 456         }
 457
 458         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 459         obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
 460
 461         LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
 462         lmm = NULL;
 463         rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
 464         if (rc < 0) {
 465                 CERROR("cannot pack lsm, err = %d\n", rc);
 466                 GOTO(out_oa, rc);
 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(free_diskmd, rc);
 477         }
 478
 479         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 480         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 481         LASSERT(lmm_buf);
 482         memcpy(lmm_buf, lmm, lmm_size);
 483         *objid = lmm_buf; /* save for mds_lov_update_objid */
 484
 485  free_diskmd:
 486         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 487  out_oa:
 488         oti_free_cookies(&oti);
 489         OBDO_FREE(oinfo.oi_oa);
 490  out_ids:
 491         if (oinfo.oi_md)
 492                 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
 493         RETURN(rc);
 494 }
 495
 496 static void reconstruct_open(struct mds_update_record *rec, int offset,
 497                              struct ptlrpc_request *req,
 498                              struct lustre_handle *child_lockh)
 499 {
 500         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 501         struct lsd_client_data *lcd = med->med_lcd;
 502         struct mds_obd *mds = mds_req2mds(req);
 503         struct mds_file_data *mfd;
 504         struct obd_export *exp = req->rq_export;
 505         struct obd_device *obd = exp->exp_obd;
 506         struct dentry *parent, *dchild;
 507         struct ldlm_reply *rep;
 508         struct mds_body *body;
 509         int rc;
 510         struct list_head *t;
 511         int put_child = 1;
 512         ENTRY;
 513
 514         LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
 515         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
 516         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 517
 518         /* copy rc, transno and disp; steal locks */
 519         mds_req_from_lcd(req, lcd);
 520         intent_set_disposition(rep, le32_to_cpu(lcd->lcd_last_data));
 521
 522         /* Only replay if create or open actually happened. */
 523         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 524                 EXIT;
 525                 return; /* error looking up parent or child */
 526         }
 527
 528         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 529         if (IS_ERR(parent)) {
 530                 rc = PTR_ERR(parent);
 531                 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
 532                               " Evicting client %s with export %s.\n",
 533                               rec->ur_fid1->id, rec->ur_fid1->generation, rc,
 534                               obd_uuid2str(&exp->exp_client_uuid),
 535                               obd_export_nid2str(exp));
 536                 mds_export_evict(exp);
 537                 EXIT;
 538                 return;
 539         }
 540
 541         dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 542         if (IS_ERR(dchild)) {
 543                 rc = PTR_ERR(dchild);
 544                 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
 545                               " Evicting client %s with export %s.\n",
 546                               rec->ur_fid1->id, rec->ur_fid1->generation, rc,
 547                               obd_uuid2str(&exp->exp_client_uuid),
 548                               obd_export_nid2str(exp));
 549                 mds_export_evict(exp);
 550                 EXIT;
 551                 return;
 552         }
 553
 554         if (!dchild->d_inode)
 555                 GOTO(out_dput, 0); /* child not present to open */
 556
 557         /* At this point, we know we have a child. We'll send
 558          * it back _unless_ it not created and open failed.
 559          */
 560         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 561             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 562             req->rq_status) {
 563                 GOTO(out_dput, 0);
 564         }
 565
 566         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 567         mds_pack_inode2body(body, dchild->d_inode);
 568         if (S_ISREG(dchild->d_inode->i_mode)) {
 569                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 570                                  body, dchild->d_inode, 1, 0);
 571
 572                 if (rc)
 573                         LASSERT(rc == req->rq_status);
 574
 575                 /* If we have LOV EA data, the OST holds size, mtime */
 576                 if (!(body->valid & OBD_MD_FLEASIZE))
 577                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 578                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 579         }
 580
 581         if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 582                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 583                                     body->eadatasize, 0);
 584
 585         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 586             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 587                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 588
 589                 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
 590                                   body, acl_off);
 591                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 592                 if (!req->rq_status && rc)
 593                         req->rq_status = rc;
 594         }
 595
 596         /* If we have -EEXIST as the status, and we were asked to create
 597          * exclusively, we can tell we failed because the file already existed.
 598          */
 599         if (req->rq_status == -EEXIST &&
 600             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 601              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 602                 GOTO(out_dput, 0);
 603         }
 604
 605         /* If we didn't get as far as trying to open, then some locking thing
 606          * probably went wrong, and we'll just bail here.
 607          */
 608         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 609                 GOTO(out_dput, 0);
 610
 611         /* If we failed, then we must have failed opening, so don't look for
 612          * file descriptor or anything, just give the client the bad news.
 613          */
 614         if (req->rq_status)
 615                 GOTO(out_dput, 0);
 616
 617         mfd = NULL;
 618         spin_lock(&med->med_open_lock);
 619         list_for_each(t, &med->med_open_head) {
 620                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 621                 if (mfd->mfd_xid == req->rq_xid) {
 622                         mds_mfd_addref(mfd);
 623                         break;
 624                 }
 625                 mfd = NULL;
 626         }
 627         spin_unlock(&med->med_open_lock);
 628
 629         /* #warning "XXX fixme" bug 2991 */
 630         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 631          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 632          * to detect a re-open */
 633         if (mfd == NULL) {
 634                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 635                         rc = mds_join_file(rec, req, dchild, NULL);
 636                         if (rc)
 637                                 GOTO(out_dput, rc);
 638                 }
 639                 mntget(mds->mds_vfsmnt);
 640                 CERROR("Re-opened file \n");
 641                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 642                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 643                 mntput(mds->mds_vfsmnt);
 644                 if (IS_ERR(mfd)) {
 645                         req->rq_status = PTR_ERR(mfd);
 646                         mfd = NULL;
 647                         CERROR("%s: opening inode "LPU64" failed: rc %d\n",
 648                                req->rq_export->exp_obd->obd_name,
 649                                (__u64)dchild->d_inode->i_ino, req->rq_status);
 650                         GOTO(out_dput, req->rq_status);
 651                 }
 652         } else {
 653                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 654                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 655                        mfd->mfd_handle.h_cookie);
 656         }
 657
 658         mds_mfd_put(mfd);
 659
 660  out_dput:
 661         if (put_child)
 662                 l_dput(dchild);
 663         l_dput(parent);
 664         EXIT;
 665 }
 666
 667 /* do NOT or the MAY_*'s, you'll get the weakest */
 668 static int accmode(struct inode *inode, int flags)
 669 {
 670         int res = 0;
 671
 672         /* Sadly, NFSD reopens a file repeatedly during operation, so the
 673          * "acc_mode = 0" allowance for newly-created files isn't honoured.
 674          * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
 675          * owner can write to a file even if it is marked readonly to hide
 676          * its brokenness. (bug 5781) */
 677         if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
 678                 return 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 & MDS_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                            struct lustre_handle *lockh)
 694 {
 695         struct mds_obd *mds = mds_req2mds(req);
 696         struct obd_device *obd = req->rq_export->exp_obd;
 697         struct mds_file_data *mfd = NULL;
 698         struct lov_mds_md *lmm = NULL; /* object IDs created */
 699         int rc = 0;
 700         ENTRY;
 701
 702         /* atomically create objects if necessary */
 703         LOCK_INODE_MUTEX(dchild->d_inode);
 704
 705         if (S_ISREG(dchild->d_inode->i_mode) &&
 706             !(body->valid & OBD_MD_FLEASIZE)) {
 707                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 708                                  body, dchild->d_inode, 0, 0);
 709                 if (rc) {
 710                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 711                         RETURN(rc);
 712                 }
 713         }
 714         if (rec != NULL) {
 715                 if ((body->valid & OBD_MD_FLEASIZE) &&
 716                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 717                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 718                         RETURN(-EEXIST);
 719                 }
 720                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 721                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 722                         rc = mds_join_file(rec, req, dchild, lockh);
 723                         if (rc)
 724                                 RETURN(rc);
 725                         LOCK_INODE_MUTEX(dchild->d_inode);
 726                 }
 727                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 728                     !(body->valid & OBD_MD_FLMODEASIZE)) {
 729                         /* no EA: create objects */
 730                         rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
 731                                                 mds, obd, dchild, handle, &lmm);
 732                         if (rc) {
 733                                 CERROR("mds_create_objects: rc = %d\n", rc);
 734                                 UNLOCK_INODE_MUTEX(dchild->d_inode);
 735                                 RETURN(rc);
 736                         }
 737                 }
 738         }
 739         /* If the inode has no EA data, then MDS holds size, mtime */
 740         if (S_ISREG(dchild->d_inode->i_mode) &&
 741             !(body->valid & OBD_MD_FLEASIZE)) {
 742                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 743                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 744         }
 745         UNLOCK_INODE_MUTEX(dchild->d_inode);
 746
 747         if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 748                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 749                                     body->eadatasize, 0);
 750
 751         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 752             rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 753                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 754
 755                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 756                                   dchild->d_inode, req->rq_repmsg,
 757                                   body, acl_off);
 758                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 759                 if (rc)
 760                         RETURN(rc);
 761         }
 762
 763         intent_set_disposition(rep, DISP_OPEN_OPEN);
 764         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 765         if (IS_ERR(mfd))
 766                 RETURN(PTR_ERR(mfd));
 767
 768         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 769                mfd->mfd_handle.h_cookie);
 770
 771         mds_lov_update_objids(obd, lmm);
 772
 773         if (rc) /* coverity[deadcode] */
 774                 mds_mfd_unlink(mfd, 1);
 775
 776         mds_mfd_put(mfd);
 777         RETURN(rc);
 778 }
 779
 780 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 781                            struct mds_body *body, int flags,
 782                            struct mds_update_record *rec,struct ldlm_reply *rep)
 783 {
 784         struct mds_obd *mds = mds_req2mds(req);
 785         struct dentry *dchild;
 786         char fidname[LL_FID_NAMELEN];
 787         int fidlen = 0, rc;
 788         void *handle = NULL;
 789         ENTRY;
 790
 791         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 792         dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 793         if (IS_ERR(dchild)) {
 794                 rc = PTR_ERR(dchild);
 795                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 796                 RETURN(rc);
 797         }
 798
 799         if (dchild->d_inode != NULL) {
 800                 mds_inode_set_orphan(dchild->d_inode);
 801                 CWARN("Orphan %s found and opened in PENDING directory\n",
 802                        fidname);
 803         } else {
 804                 l_dput(dchild);
 805
 806                 /* We didn't find it in PENDING so it isn't an orphan.  See
 807                  * if it was a regular inode that was previously created. */
 808                 dchild = mds_fid2dentry(mds, fid, NULL);
 809                 if (IS_ERR(dchild))
 810                         RETURN(PTR_ERR(dchild));
 811         }
 812
 813         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 814         mds_pack_inode2body(body, dchild->d_inode);
 815         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 816         intent_set_disposition(rep, DISP_LOOKUP_POS);
 817
 818         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
 819         rc = mds_finish_transno(mds, dchild->d_inode, handle,
 820                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
 821         /* XXX what do we do here if mds_finish_transno itself failed? */
 822
 823         l_dput(dchild);
 824         RETURN(rc);
 825 }
 826
 827 int mds_pin(struct ptlrpc_request *req, int offset)
 828 {
 829         struct obd_device *obd = req->rq_export->exp_obd;
 830         struct mds_body *reqbody, *repbody;
 831         struct lvfs_run_ctxt saved;
 832         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 833         ENTRY;
 834
 835         reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
 836
 837         rc = lustre_pack_reply(req, 2, size, NULL);
 838         if (rc)
 839                 RETURN(rc);
 840
 841         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 842                                  sizeof(*repbody));
 843
 844         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 845         rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
 846                              NULL);
 847         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 848
 849         RETURN(rc);
 850 }
 851
 852 /*  Get an internal lock on the inode number (but not generation) to sync
 853  *  new inode creation with inode unlink (bug 2029).  If child_lockh is NULL
 854  *  we just get the lock as a barrier to wait for other holders of this lock,
 855  *  and drop it right away again. */
 856 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 857                        struct lustre_handle *child_lockh)
 858 {
 859         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 860         struct lustre_handle lockh;
 861         int lock_flags = LDLM_FL_ATOMIC_CB;
 862         int rc;
 863
 864         if (child_lockh == NULL)
 865                 child_lockh = &lockh;
 866
 867         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
 868                                     LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 869                                     ldlm_blocking_ast, ldlm_completion_ast,
 870                                     NULL, NULL, 0, NULL, child_lockh);
 871         if (rc != ELDLM_OK)
 872                 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
 873         else if (child_lockh == &lockh)
 874                 ldlm_lock_decref(child_lockh, LCK_EX);
 875
 876         RETURN(rc);
 877 }
 878
 879 int mds_open(struct mds_update_record *rec, int offset,
 880              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 881 {
 882         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 883         struct obd_device *obd = req->rq_export->exp_obd;
 884         struct mds_obd *mds = mds_req2mds(req);
 885         struct ldlm_reply *rep = NULL;
 886         struct mds_body *body = NULL;
 887         struct dentry *dchild = NULL, *dparent = NULL;
 888         struct mds_export_data *med;
 889         struct lustre_handle parent_lockh;
 890         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 891         int parent_mode = LCK_CR;
 892         void *handle = NULL;
 893         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
 894         unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
 895         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
 896         int child_mode = LCK_CR;
 897         /* Always returning LOOKUP lock if open succesful to guard
 898            dentry on client. */
 899         ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
 900         struct ldlm_res_id child_res_id = { .name = {0}};
 901         int lock_flags = 0;
 902         int rec_pending = 0;
 903         unsigned int gid = current->fsgid;
 904         ENTRY;
 905
 906         mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
 907
 908         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 909                          (obd_timeout + 1) / 4);
 910
 911         CLASSERT(MAXQUOTAS < 4);
 912         if (offset == DLM_INTENT_REC_OFF) { /* intent */
 913                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 914                                      sizeof(*rep));
 915                 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
 916                                       sizeof(*body));
 917         } else if (offset == REQ_REC_OFF) { /* non-intent reint */
 918                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 919                                       sizeof(*body));
 920                 LBUG(); /* XXX: not supported yet? */
 921         } else {
 922                 body = NULL;
 923                 LBUG();
 924         }
 925
 926         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 927
 928         /* Step 0: If we are passed a fid, then we assume the client already
 929          * opened this file and is only replaying the RPC, so we open the
 930          * inode by fid (at some large expense in security). */
 931         /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
 932         if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
 933              (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
 934             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 935                 if (rec->ur_fid2->id == 0) {
 936                         struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
 937                         if (lock) {
 938                                 LDLM_ERROR(lock, "fid2 not set on open replay");
 939                                 LDLM_LOCK_PUT(lock);
 940                         }
 941                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 942                         RETURN(-EFAULT);
 943                 }
 944
 945                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 946                                      rec, rep);
 947                 if (rc != -ENOENT) {
 948                         if (req->rq_export->exp_libclient &&
 949                             rec->ur_flags & MDS_OPEN_HAS_EA)
 950                                 RETURN(0);
 951
 952                         RETURN(rc);
 953                 }
 954
 955                 /* We didn't find the correct inode on disk either, so we
 956                  * need to re-create it via a regular replay. */
 957                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 958                         DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
 959                         RETURN(-EFAULT);
 960                 }
 961         } else if (rec->ur_fid2->id) {
 962                 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
 963                           rec->ur_fid2->id, rec->ur_fid2->generation);
 964                 RETURN(-EFAULT);
 965         }
 966
 967         /* If we got here, we must be called via intent */
 968         LASSERT(offset == DLM_INTENT_REC_OFF);
 969
 970         med = &req->rq_export->exp_mds_data;
 971         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 972                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 973                 RETURN(-ENOMEM);
 974         }
 975
 976         /* Step 1: Find and lock the parent */
 977         if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
 978                 parent_mode = LCK_EX;
 979         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 980                                         &parent_lockh, rec->ur_name,
 981                                         rec->ur_namelen - 1,
 982                                         MDS_INODELOCK_UPDATE);
 983         if (IS_ERR(dparent)) {
 984                 rc = PTR_ERR(dparent);
 985                 if (rc != -ENOENT) {
 986                         CERROR("parent "LPU64"/%u lookup error %d\n",
 987                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 988                 } else {
 989                         /* Just cannot find parent - make it look like
 990                          * usual negative lookup to avoid extra MDS RPC */
 991                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 992                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
 993                 }
 994                 GOTO(cleanup, rc);
 995         }
 996         LASSERT(dparent->d_inode != NULL);
 997
 998         cleanup_phase = 1; /* parent dentry and lock */
 999
1000         if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
1001                 dchild = dget(dparent);
1002                 cleanup_phase = 2; /* child dentry */
1003                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1004                 GOTO(found_child, rc);
1005         }
1006
1007         /* Step 2: Lookup the child */
1008
1009         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1010             (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) {
1011                 /* hack for nfsd with no_subtree_check, it will use anon
1012                  * dentry w/o filename to open the file. the anon dentry's
1013                  * parent was set to itself, so rec->ur_fid1 is the file.
1014                  * And in MDC it cannot derive the dentry's parent dentry,
1015                  * hence the file's name, so we hack here in MDS,
1016                  * refer to bug 13030. */
1017                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1018         } else {
1019                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1020                                            rec->ur_namelen - 1);
1021         }
1022         if (IS_ERR(dchild)) {
1023                 rc = PTR_ERR(dchild);
1024                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1025                 GOTO(cleanup, rc);
1026         }
1027
1028         cleanup_phase = 2; /* child dentry */
1029
1030         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1031         if (dchild->d_inode)
1032                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1033         else
1034                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1035
1036         /*Step 3: If the child was negative, and we're supposed to, create it.*/
1037         if (dchild->d_inode == NULL) {
1038                 unsigned long ino = rec->ur_fid2->id;
1039                 struct iattr iattr;
1040                 struct inode *inode;
1041
1042                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1043                         /* It's negative and we weren't supposed to create it */
1044                         GOTO(cleanup, rc = -ENOENT);
1045                 }
1046
1047                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1048                         GOTO(cleanup, rc = -EROFS);
1049
1050                 if (dparent->d_inode->i_mode & S_ISGID)
1051                         gid = dparent->d_inode->i_gid;
1052                 else
1053                         gid = current->fsgid;
1054
1055                 /* we try to get enough quota to write here, and let ldiskfs
1056                  * decide if it is out of quota or not b=14783 */
1057                 lquota_chkquota(mds_quota_interface_ref, obd,
1058                                 current->fsuid, gid, 1, &rec_pending);
1059
1060                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1061                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1062                                       NULL);
1063                 if (IS_ERR(handle)) {
1064                         rc = PTR_ERR(handle);
1065                         handle = NULL;
1066                         GOTO(cleanup, rc);
1067                 }
1068                 dchild->d_fsdata = (void *) &dp;
1069                 dp.ldp_ptr = req;
1070                 dp.ldp_inum = ino;
1071
1072                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1073                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1074                         dchild->d_fsdata = NULL;
1075
1076                 if (rc) {
1077                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1078                         GOTO(cleanup, rc);
1079                 }
1080                 inode = dchild->d_inode;
1081                 if (ino) {
1082                         LASSERT(ino == inode->i_ino);
1083                         /* Written as part of setattr */
1084                         inode->i_generation = rec->ur_fid2->generation;
1085                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1086                                inode->i_ino, inode->i_generation);
1087                 }
1088
1089                 created = 1;
1090                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1091                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1092                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1093
1094                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
1095                 iattr.ia_gid = gid;
1096
1097                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1098                         ATTR_MTIME | ATTR_CTIME;
1099
1100                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1101                 if (rc)
1102                         CERROR("error on child setattr: rc = %d\n", rc);
1103
1104                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1105
1106                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1107                 if (rc)
1108                         CERROR("error on parent setattr: rc = %d\n", rc);
1109
1110                 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1111                 handle = NULL;
1112                 acc_mode = 0;           /* Don't check for permissions */
1113         } else {
1114                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1115         }
1116
1117         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1118                  "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1119                  dchild->d_name.name, dchild, dchild->d_inode,
1120                  dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1121
1122 found_child:
1123         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1124         mds_pack_inode2body(body, dchild->d_inode);
1125
1126         if (S_ISREG(dchild->d_inode->i_mode)) {
1127                 /* Check permissions etc */
1128                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1129                 if (rc != 0)
1130                         GOTO(cleanup, rc);
1131
1132                 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1133                     (acc_mode & MAY_WRITE))
1134                         GOTO(cleanup, rc = -EROFS);
1135
1136                 /* An append-only file must be opened in append mode for
1137                  * writing */
1138                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1139                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1140                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1141                         GOTO(cleanup, rc = -EPERM);
1142         }
1143
1144         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1145             (rec->ur_flags & MDS_OPEN_EXCL)) {
1146                 /* File already exists, we didn't just create it, and we
1147                  * were passed O_EXCL; err-or. */
1148                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1149         }
1150
1151         /* if we are following a symlink, don't open */
1152         if (S_ISLNK(dchild->d_inode->i_mode))
1153                 GOTO(cleanup_no_trans, rc = 0);
1154
1155         if (S_ISDIR(dchild->d_inode->i_mode)) {
1156                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1157                     rec->ur_flags & FMODE_WRITE) {
1158                         /* we are trying to create or write a exist dir */
1159                         GOTO(cleanup, rc = -EISDIR);
1160                 }
1161                 if (rec->ur_flags & MDS_FMODE_EXEC) {
1162                         /* we are trying to exec a directory */
1163                         GOTO(cleanup, rc = -EACCES);
1164                 }
1165                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1166                         intent_set_disposition(rep, DISP_OPEN_OPEN);
1167                         GOTO(cleanup, rc = -EACCES);
1168                 }
1169         } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1170                 GOTO(cleanup, rc = -ENOTDIR);
1171         }
1172
1173         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1174                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1175                 GOTO(cleanup, rc = -EAGAIN);
1176         }
1177
1178         /* Obtain OPEN lock as well */
1179         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1180
1181         /* We cannot use acc_mode here, because it is zeroed in case of
1182            creating a file, so we get wrong lockmode */
1183         if (rec->ur_flags & FMODE_WRITE)
1184                 child_mode = LCK_CW;
1185         else if (rec->ur_flags & MDS_FMODE_EXEC)
1186                 child_mode = LCK_PR;
1187         else
1188                 child_mode = LCK_CR;
1189
1190         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1191              (rec->ur_flags & MDS_OPEN_LOCK)) {
1192                 /* In case of replay we do not get a lock assuming that the
1193                    caller has it already */
1194                 child_res_id.name[0] = dchild->d_inode->i_ino;
1195                 child_res_id.name[1] = dchild->d_inode->i_generation;
1196
1197                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1198                                             LDLM_IBITS, &policy, child_mode,
1199                                             &lock_flags, ldlm_blocking_ast,
1200                                             ldlm_completion_ast, NULL, NULL,
1201                                             0, NULL, child_lockh);
1202                 if (rc != ELDLM_OK)
1203                         GOTO(cleanup, rc);
1204
1205                 /* Let mds_intent_policy know that we have a lock to return */
1206                 intent_set_disposition(rep, DISP_OPEN_LOCK);
1207                 cleanup_phase = 3;
1208         }
1209
1210         if (!S_ISREG(dchild->d_inode->i_mode) &&
1211             !S_ISDIR(dchild->d_inode->i_mode) &&
1212             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1213                 /* If client supports this, do not return open handle for
1214                  * special device nodes */
1215                 GOTO(cleanup_no_trans, rc = 0);
1216         }
1217
1218         /* Step 5: mds_open it */
1219         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1220                              rep, &parent_lockh);
1221         GOTO(cleanup, rc);
1222
1223  cleanup:
1224         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1225                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
1226
1227  cleanup_no_trans:
1228         if (rec_pending)
1229                 lquota_pending_commit(mds_quota_interface_ref, obd,
1230                                       current->fsuid, gid, 1);
1231         switch (cleanup_phase) {
1232         case 3:
1233                 if (rc)
1234                         /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1235                          * mds_intent_policy won't try to return any locks */
1236                         ldlm_lock_decref(child_lockh, child_mode);
1237         case 2:
1238                 if (rc && created) {
1239                         int err = vfs_unlink(dparent->d_inode, dchild);
1240                         if (err) {
1241                                 CERROR("unlink(%.*s) in error path: %d\n",
1242                                        dchild->d_name.len, dchild->d_name.name,
1243                                        err);
1244                         }
1245                 } else if (created) {
1246                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1247                         /* save uid/gid for quota acquire/release */
1248                         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1249                         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1250                 }
1251                 l_dput(dchild);
1252         case 1:
1253                 if (dparent == NULL)
1254                         break;
1255
1256                 l_dput(dparent);
1257                 if (rc)
1258                         ldlm_lock_decref(&parent_lockh, parent_mode);
1259                 else
1260                         ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1261         }
1262         /* trigger dqacq on the owner of child and parent */
1263         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1264                       FSFILT_OP_CREATE);
1265         RETURN(rc);
1266 }
1267
1268 /* Close a "file descriptor" and possibly unlink an orphan from the
1269  * PENDING directory.  Caller must hold child->i_mutex, this drops it.
1270  *
1271  * If we are being called from mds_disconnect() because the client has
1272  * disappeared, then req == NULL and we do not update last_rcvd because
1273  * there is nothing that could be recovered by the client at this stage
1274  * (it will not even _have_ an entry in last_rcvd anymore).
1275  *
1276  * Returns EAGAIN if the client needs to get more data and re-close. */
1277 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1278                   struct obd_device *obd, struct mds_file_data *mfd,
1279                   int unlink_orphan, struct lov_mds_md *lmm, int lmm_size,
1280                   struct llog_cookie *logcookies, int cookies_size,
1281                   __u64 *valid)
1282 {
1283         struct inode *inode = mfd->mfd_dentry->d_inode;
1284         char fidname[LL_FID_NAMELEN];
1285         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1286         struct dentry *pending_child = NULL;
1287         struct mds_obd *mds = &obd->u.mds;
1288         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1289         void *handle = NULL;
1290         struct mds_body *request_body = NULL, *reply_body = NULL;
1291         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1292         struct iattr iattr = { 0 };
1293         ENTRY;
1294
1295         if (req && req->rq_reqmsg != NULL)
1296                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1297                                               sizeof(*request_body));
1298         if (req && req->rq_repmsg != NULL)
1299                 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1300                                             sizeof(*reply_body));
1301
1302         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1303
1304         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1305                inode->i_nlink, mds_orphan_open_count(inode));
1306
1307         last_orphan = mds_orphan_open_dec_test(inode) &&
1308                       mds_inode_is_orphan(inode);
1309
1310         /* this is half of the actual "close" */
1311         if (mfd->mfd_mode & FMODE_WRITE) {
1312                 rc = mds_put_write_access(mds, inode, request_body,
1313                                           last_orphan && unlink_orphan);
1314         } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1315                 mds_allow_write_access(inode);
1316         }
1317         /* here writecount change also needs protection from orphan write sem.
1318          * so drop orphan write sem after mds_put_write_access, bz 12888. */
1319         MDS_UP_WRITE_ORPHAN_SEM(inode);
1320
1321         if (last_orphan && unlink_orphan) {
1322                 int stripe_count = 0;
1323                 /* mds_put_write_access must have succeeded */
1324                 LASSERTF(rc == 0, "inode %lu/%u: rc %d",
1325                          inode->i_ino, inode->i_generation, rc);
1326
1327                 CDEBUG(D_INODE, "destroying orphan object %s\n", fidname);
1328
1329                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1330                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1331                         CERROR("found \"orphan\" %s %s with link count %d\n",
1332                                S_ISREG(inode->i_mode) ? "file" : "dir",
1333                                fidname, inode->i_nlink);
1334
1335                 /* Sadly, there is no easy way to save pending_child from
1336                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1337                  * but normally it will still be in the dcache. */
1338                 LOCK_INODE_MUTEX(pending_dir);
1339                 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1340                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1341                                                fidlen);
1342                 if (IS_ERR(pending_child))
1343                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1344                 LASSERT(pending_child->d_inode != NULL);
1345
1346                 cleanup_phase = 2; /* dput(pending_child) when finished */
1347                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1348                         rc = vfs_rmdir(pending_dir, pending_child);
1349                         if (rc)
1350                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1351                                        fidname,rc);
1352                         goto out;
1353                 }
1354
1355                 if (lmm != NULL) {
1356                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1357                 }
1358
1359                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1360                                           NULL, stripe_count);
1361                 if (IS_ERR(handle)) {
1362                         rc = PTR_ERR(handle);
1363                         handle = NULL;
1364                         GOTO(cleanup, rc);
1365                 }
1366                 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1367                     mds_log_op_unlink(obd, lmm, lmm_size,
1368                                       logcookies, cookies_size) > 0) {
1369                         *valid |= OBD_MD_FLCOOKIE;
1370                 }
1371
1372                 dp.ldp_inum = 0;
1373                 dp.ldp_ptr = req;
1374                 pending_child->d_fsdata = (void *) &dp;
1375                 rc = vfs_unlink(pending_dir, pending_child);
1376                 if (rc)
1377                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1378
1379                 goto out; /* Don't bother updating attrs on unlinked inode */
1380         }
1381
1382 #if 0
1383         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1384                 /* Update the on-disk attributes if this was the last write
1385                  * close, and all information was provided (i.e., rc == 0)
1386                  *
1387                  * XXX this should probably be abstracted with mds_reint_setattr
1388                  */
1389
1390                 if (request_body->valid & OBD_MD_FLMTIME &&
1391                     LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1392                         LTIME_S(iattr.ia_mtime) = request_body->mtime;
1393                         iattr.ia_valid |= ATTR_MTIME;
1394                 }
1395                 if (request_body->valid & OBD_MD_FLCTIME &&
1396                     LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1397                         LTIME_S(iattr.ia_ctime) = request_body->ctime;
1398                         iattr.ia_valid |= ATTR_CTIME;
1399                 }
1400
1401                 /* XXX can't set block count with fsfilt_setattr (!) */
1402                 if (request_body->valid & OBD_MD_FLSIZE) {
1403                         iattr.ia_valid |= ATTR_SIZE;
1404                         iattr.ia_size = request_body->size;
1405                 }
1406                 /* iattr.ia_blocks = request_body->blocks */
1407
1408         }
1409 #endif
1410         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1411                 /* Only start a transaction to write out only the atime if
1412                  * it is more out-of-date than the specified limit.  If we
1413                  * are already going to write out the atime then do it anyway.
1414                  * */
1415                 LTIME_S(iattr.ia_atime) = request_body->atime;
1416                 if ((LTIME_S(iattr.ia_atime) >
1417                      LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1418                     (iattr.ia_valid != 0 &&
1419                      LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1420                         iattr.ia_valid |= ATTR_ATIME;
1421         }
1422
1423         if (iattr.ia_valid != 0) {
1424                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1425                 if (IS_ERR(handle)) {
1426                         rc = PTR_ERR(handle);
1427                         handle = NULL;
1428                         GOTO(cleanup, rc);
1429                 }
1430                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1431                 if (rc)
1432                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1433         }
1434 out:
1435         /* If other clients have this file open for write, rc will be > 0 */
1436         if (rc > 0)
1437                 rc = 0;
1438         l_dput(mfd->mfd_dentry);
1439         mds_mfd_put(mfd);
1440
1441  cleanup:
1442         if (req != NULL && reply_body != NULL) {
1443                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0);
1444         } else if (handle) {
1445                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1446                 if (err) {
1447                         CERROR("error committing close: %d\n", err);
1448                         if (!rc)
1449                                 rc = err;
1450                 }
1451         }
1452
1453         switch (cleanup_phase) {
1454         case 2:
1455                 dput(pending_child);
1456         case 1:
1457                 UNLOCK_INODE_MUTEX(pending_dir);
1458         }
1459         RETURN(rc);
1460 }
1461
1462 int mds_close(struct ptlrpc_request *req, int offset)
1463 {
1464         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1465         struct obd_device *obd = req->rq_export->exp_obd;
1466         struct mds_body *body;
1467         struct mds_file_data *mfd;
1468         struct lvfs_run_ctxt saved;
1469         struct inode *inode;
1470         int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1471                                sizeof(struct mds_body),
1472                                obd->u.mds.mds_max_mdsize,
1473                                obd->u.mds.mds_max_cookiesize };
1474         struct mds_body *reply_body;
1475         struct lov_mds_md *lmm;
1476         int lmm_size;
1477         struct llog_cookie *logcookies;
1478         int cookies_size;
1479         ENTRY;
1480
1481         rc = lustre_pack_reply(req, 4, repsize, NULL);
1482         if (rc)
1483                 req->rq_status = rc;
1484                 /* continue on to drop local open even if we can't send reply */
1485         else
1486                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1487
1488         CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n",
1489                req->rq_replen,
1490                obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1491         mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1492
1493         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1494                                   lustre_swab_mds_body);
1495         if (body == NULL) {
1496                 CERROR("Can't unpack body\n");
1497                 req->rq_status = -EFAULT;
1498                 RETURN(-EFAULT);
1499         }
1500
1501         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1502                 /* do some stuff */ ;
1503
1504         spin_lock(&med->med_open_lock);
1505         mfd = mds_handle2mfd(&body->handle);
1506         if (mfd == NULL) {
1507                 spin_unlock(&med->med_open_lock);
1508                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1509                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1510                 req->rq_status = -ESTALE;
1511                 RETURN(-ESTALE);
1512         }
1513         /* Remove mfd handle so it can't be found again.  We consume mfd_list
1514          * reference here, but still have mds_handle2mfd ref until mfd_close. */
1515         mds_mfd_unlink(mfd, 1);
1516         spin_unlock(&med->med_open_lock);
1517
1518         inode = mfd->mfd_dentry->d_inode;
1519         /* child orphan sem protects orphan_dec_test && is_orphan race */
1520         MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1521         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1522                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1523                                       sizeof(*body));
1524                 LASSERT(body != NULL);
1525
1526                 mds_pack_inode2fid(&body->fid1, inode);
1527                 mds_pack_inode2body(body, inode);
1528                 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1529                             MDS_PACK_MD_LOCK, 0);
1530         }
1531
1532         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1533         reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1534         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1535         lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1536         logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1537         cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1538         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1539                                        lmm, lmm_size, logcookies, cookies_size,
1540                                        &reply_body->valid);
1541         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1542
1543         mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1544         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1545                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1546                 req->rq_status = -ENOMEM;
1547                 RETURN(-ENOMEM);
1548         }
1549
1550         RETURN(rc);
1551 }
1552
1553 int mds_done_writing(struct ptlrpc_request *req, int offset)
1554 {
1555         struct mds_body *body;
1556         int rc, size[2] = { sizeof(struct ptlrpc_body),
1557                             sizeof(struct mds_body) };
1558         ENTRY;
1559
1560         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1561
1562         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1563                                   lustre_swab_mds_body);
1564         if (body == NULL) {
1565                 CERROR("Can't unpack body\n");
1566                 req->rq_status = -EFAULT;
1567                 RETURN(-EFAULT);
1568         }
1569
1570         rc = lustre_pack_reply(req, 2, size, NULL);
1571         if (rc)
1572                 req->rq_status = rc;
1573
1574         RETURN(0);
1575 }