lustre/mds/mds_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *   Author: Phil Schwan <phil@clusterfs.com>
   8  *   Author: Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/version.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/version.h>
  35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  36 # include <linux/buffer_head.h>
  37 # include <linux/workqueue.h>
  38 #else
  39 # include <linux/locks.h>
  40 #endif
  41
  42 #include <linux/obd_class.h>
  43 #include <linux/lustre_mds.h>
  44 #include <linux/obd_lov.h>
  45 #include <linux/lustre_fsfilt.h>
  46 #include <linux/lprocfs_status.h>
  47
  48 #include "mds_internal.h"
  49
  50 /* Exported function from this file are:
  51  *
  52  * mds_open - called by the intent handler
  53  * mds_close - an rpc handling function
  54  * mds_pin - an rpc handling function - which will go away
  55  * mds_mfd_close - for force closing files when a client dies
  56  */
  57
  58 /*
  59  * MDS file data handling: file data holds a handle for a file opened
  60  * by a client.
  61  */
  62
  63 static void mds_mfd_addref(void *mfdp)
  64 {
  65         struct mds_file_data *mfd = mfdp;
  66
  67         atomic_inc(&mfd->mfd_refcount);
  68         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  69                atomic_read(&mfd->mfd_refcount));
  70 }
  71
  72 struct mds_file_data *mds_mfd_new(void)
  73 {
  74         struct mds_file_data *mfd;
  75
  76         OBD_ALLOC(mfd, sizeof *mfd);
  77         if (mfd == NULL) {
  78                 CERROR("mds: out of memory\n");
  79                 return NULL;
  80         }
  81
  82         atomic_set(&mfd->mfd_refcount, 2);
  83
  84         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  85         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  86
  87         return mfd;
  88 }
  89
  90 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  91 {
  92         ENTRY;
  93         LASSERT(handle != NULL);
  94         RETURN(class_handle2object(handle->cookie));
  95 }
  96
  97 static void mds_mfd_put(struct mds_file_data *mfd)
  98 {
  99         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 100                atomic_read(&mfd->mfd_refcount) - 1);
 101         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 102                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 103         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 104                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 105                 OBD_FREE(mfd, sizeof *mfd);
 106         }
 107 }
 108
 109 static void mds_mfd_destroy(struct mds_file_data *mfd)
 110 {
 111         class_handle_unhash(&mfd->mfd_handle);
 112         mds_mfd_put(mfd);
 113 }
 114
 115
 116 /* Caller must hold mds->mds_epoch_sem */
 117 static int mds_alloc_filterdata(struct inode *inode)
 118 {
 119         LASSERT(inode->i_filterdata == NULL);
 120         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 121         if (inode->i_filterdata == NULL)
 122                 return -ENOMEM;
 123         LASSERT(igrab(inode) == inode);
 124         return 0;
 125 }
 126
 127 /* Caller must hold mds->mds_epoch_sem */
 128 static void mds_free_filterdata(struct inode *inode)
 129 {
 130         LASSERT(inode->i_filterdata != NULL);
 131         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 132         inode->i_filterdata = NULL;
 133         iput(inode);
 134 }
 135
 136 /* Write access to a file: executors cause a negative count,
 137  * writers a positive count.  The semaphore is needed to perform
 138  * a check for the sign and then increment or decrement atomically.
 139  *
 140  * This code is closely tied to the allocation of the d_fsdata and the
 141  * MDS epoch, so we use the same semaphore for the whole lot.
 142  *
 143  * We could use a different semaphore for each file, if it ever shows
 144  * up in a profile, which it won't.
 145  *
 146  * epoch argument is nonzero during recovery */
 147 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 148                                 __u64 epoch)
 149 {
 150         int rc = 0;
 151
 152         down(&mds->mds_epoch_sem);
 153
 154         if (atomic_read(&inode->i_writecount) < 0) {
 155                 up(&mds->mds_epoch_sem);
 156                 RETURN(-ETXTBSY);
 157         }
 158
 159
 160         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 161                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 162                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 163                        inode->i_generation);
 164                 goto out;
 165         }
 166
 167         if (inode->i_filterdata == NULL)
 168                 mds_alloc_filterdata(inode);
 169         if (inode->i_filterdata == NULL) {
 170                 rc = -ENOMEM;
 171                 goto out;
 172         }
 173         if (epoch > mds->mds_io_epoch)
 174                 mds->mds_io_epoch = epoch;
 175         else
 176                 mds->mds_io_epoch++;
 177         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 178         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 179                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 180  out:
 181         if (rc == 0)
 182                 atomic_inc(&inode->i_writecount);
 183         up(&mds->mds_epoch_sem);
 184         return rc;
 185 }
 186
 187 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 188  * again -- which is never true if the file is about to be unlinked.  Otherwise
 189  * returns the number of remaining writers. */
 190 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 191                                 struct mds_body *body, int unlinking)
 192 {
 193         int rc = 0;
 194         ENTRY;
 195
 196         down(&mds->mds_epoch_sem);
 197         atomic_dec(&inode->i_writecount);
 198         rc = atomic_read(&inode->i_writecount);
 199         if (rc > 0)
 200                 GOTO(out, rc);
 201 #if 0
 202         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 203                 GOTO(out, rc = EAGAIN);
 204 #endif
 205         mds_free_filterdata(inode);
 206  out:
 207         up(&mds->mds_epoch_sem);
 208         return rc;
 209 }
 210
 211 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 212 {
 213         ENTRY;
 214         down(&mds->mds_epoch_sem);
 215         if (atomic_read(&inode->i_writecount) > 0) {
 216                 up(&mds->mds_epoch_sem);
 217                 RETURN(-ETXTBSY);
 218         }
 219         atomic_dec(&inode->i_writecount);
 220         up(&mds->mds_epoch_sem);
 221         RETURN(0);
 222 }
 223
 224 static void mds_allow_write_access(struct inode *inode)
 225 {
 226         ENTRY;
 227         atomic_inc(&inode->i_writecount);
 228 }
 229
 230 int mds_query_write_access(struct inode *inode)
 231 {
 232         ENTRY;
 233         RETURN(atomic_read(&inode->i_writecount));
 234 }
 235
 236 /* This replaces the VFS mds_dentry_open, it manages mfd and writecount */
 237 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 238                                              struct vfsmount *mnt, int flags,
 239                                              struct ptlrpc_request *req)
 240 {
 241         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 242         struct mds_obd *mds = mds_req2mds(req);
 243         struct mds_file_data *mfd;
 244         struct mds_body *body;
 245         int error;
 246         ENTRY;
 247
 248         mfd = mds_mfd_new();
 249         if (mfd == NULL) {
 250                 CERROR("mds: out of memory\n");
 251                 GOTO(cleanup_dentry, error = -ENOMEM);
 252         }
 253
 254         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 255
 256         if (flags & FMODE_WRITE) {
 257                 /* FIXME: in recovery, need to pass old epoch here */
 258                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 259                 if (error)
 260                         GOTO(cleanup_mfd, error);
 261                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 262         } else if (flags & FMODE_EXEC) {
 263                 error = mds_deny_write_access(mds, dentry->d_inode);
 264                 if (error)
 265                         GOTO(cleanup_mfd, error);
 266         }
 267
 268         dget(dentry);
 269
 270         /* Mark the file as open to handle open-unlink. */
 271         mds_open_orphan_inc(dentry->d_inode);
 272
 273         mfd->mfd_mode = flags;
 274         mfd->mfd_dentry = dentry;
 275         mfd->mfd_xid = req->rq_xid;
 276
 277         spin_lock(&med->med_open_lock);
 278         list_add(&mfd->mfd_list, &med->med_open_head);
 279         spin_unlock(&med->med_open_lock);
 280         mds_mfd_put(mfd);
 281
 282         body->handle.cookie = mfd->mfd_handle.h_cookie;
 283
 284         RETURN(mfd);
 285
 286 cleanup_mfd:
 287         mds_mfd_put(mfd);
 288         mds_mfd_destroy(mfd);
 289 cleanup_dentry:
 290         return ERR_PTR(error);
 291 }
 292
 293 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 294                                 struct lov_desc *desc)
 295 {
 296         int i;
 297         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 298                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 299                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 300         }
 301 }
 302
 303 /* Must be called with i_sem held */
 304 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 305                               struct mds_update_record *rec,
 306                               struct mds_obd *mds, struct obd_device *obd,
 307                               struct dentry *dchild, void **handle,
 308                               obd_id **ids)
 309 {
 310         struct obdo *oa;
 311         struct obd_trans_info oti = { 0 };
 312         struct mds_body *body;
 313         struct lov_stripe_md *lsm = NULL;
 314         struct lov_mds_md *lmm = NULL;
 315         struct inode *inode = dchild->d_inode;
 316         void *lmm_buf;
 317         int rc, lmm_bufsize, lmm_size;
 318         ENTRY;
 319
 320         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 321             !(rec->ur_flags & FMODE_WRITE))
 322                 RETURN(0);
 323
 324         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 325
 326         if (!S_ISREG(inode->i_mode))
 327                 RETURN(0);
 328         if (body->valid & OBD_MD_FLEASIZE)
 329                 RETURN(0);
 330
 331         OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 332         if (*ids == NULL)
 333                 RETURN(-ENOMEM);
 334         oti.oti_objid = *ids;
 335
 336         /* replay case */
 337         if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 338                 LASSERT (rec->ur_fid2->id);
 339                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 340                 lmm_size = rec->ur_eadatalen;
 341                 lmm = rec->ur_eadata;
 342                 LASSERT(lmm);
 343
 344                 if (*handle == NULL)
 345                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 346                 if (IS_ERR(*handle)) {
 347                         rc = PTR_ERR(*handle);
 348                         *handle = NULL;
 349                         GOTO(out_ids, rc);
 350                 }
 351
 352                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 353
 354                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 355                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 356                 LASSERT(lmm_buf);
 357                 LASSERT(lmm_bufsize >= lmm_size);
 358                 memcpy(lmm_buf, lmm, lmm_size);
 359                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 360                 if (rc)
 361                         CERROR("open replay failed to set md:%d\n", rc);
 362                 RETURN(0);
 363         }
 364
 365         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 366                 GOTO(out_ids, rc = -ENOMEM);
 367
 368         oa = obdo_alloc();
 369         if (oa == NULL)
 370                 GOTO(out_ids, rc = -ENOMEM);
 371         oa->o_mode = S_IFREG | 0600;
 372         oa->o_id = inode->i_ino;
 373         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 374         oa->o_generation = inode->i_generation;
 375         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 376         oa->o_gid = 0;
 377         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
 378                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 379         oa->o_size = 0;
 380
 381         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
 382                         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, &lsm, rec->ur_eadata);
 390                         if (rc)
 391                                 GOTO(out_oa, rc);
 392                 } else {
 393                         /* Per-directory striping default code removed, because
 394                          * it uses the same unnamed EA storage as the directory
 395                          * striping for CMD. -p */
 396                 }
 397                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 398                 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
 399                 if (rc) {
 400                         int level = D_ERROR;
 401                         if (rc == -ENOSPC)
 402                                 level = D_INODE;
 403                         CDEBUG(level, "error creating objects for "
 404                                       "inode %lu: rc = %d\n",
 405                                inode->i_ino, rc);
 406                         if (rc > 0) {
 407                                 CERROR("obd_create returned invalid "
 408                                        "rc %d\n", rc);
 409                                 rc = -EIO;
 410                         }
 411                         GOTO(out_oa, rc);
 412                 }
 413         } else {
 414                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 415                                    0, &lsm, rec->ur_eadata);
 416                 if (rc) {
 417                         GOTO(out_oa, rc);
 418                 }
 419                 lsm->lsm_object_id = oa->o_id;
 420                 lsm->lsm_object_gr = oa->o_gr;
 421         }
 422         if (inode->i_size) {
 423                 oa->o_size = inode->i_size;
 424                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 425                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 426                 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
 427                 if (rc) {
 428                         CERROR("error setting attrs for inode %lu: rc %d\n",
 429                                inode->i_ino, rc);
 430                         if (rc > 0) {
 431                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 432                                 rc = -EIO;
 433                         }
 434                         GOTO(out_oa, rc);
 435                 }
 436         }
 437
 438         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 439         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 440
 441         LASSERT(lsm && lsm->lsm_object_id);
 442         lmm = NULL;
 443         rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
 444         if (!rec->ur_fid2->id)
 445                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 446         LASSERT(rc >= 0);
 447         lmm_size = rc;
 448         body->eadatasize = rc;
 449
 450         if (*handle == NULL)
 451                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 452         if (IS_ERR(*handle)) {
 453                 rc = PTR_ERR(*handle);
 454                 *handle = NULL;
 455                 GOTO(out_ids, rc);
 456         }
 457
 458         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 459         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 460         lmm_bufsize = req->rq_repmsg->buflens[offset];
 461         LASSERT(lmm_buf);
 462         LASSERT(lmm_bufsize >= lmm_size);
 463
 464         memcpy(lmm_buf, lmm, lmm_size);
 465         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 466  out_oa:
 467         oti_free_cookies(&oti);
 468         obdo_free(oa);
 469  out_ids:
 470         if (rc) {
 471                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 472                 *ids = NULL;
 473         }
 474         RETURN(rc);
 475 }
 476
 477 static void reconstruct_open(struct mds_update_record *rec, int offset,
 478                              struct ptlrpc_request *req,
 479                              struct lustre_handle *child_lockh)
 480 {
 481         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 482         struct mds_client_data *mcd = med->med_mcd;
 483         struct mds_obd *mds = mds_req2mds(req);
 484         struct mds_file_data *mfd;
 485         struct obd_device *obd = req->rq_export->exp_obd;
 486         struct dentry *parent, *child;
 487         struct ldlm_reply *rep;
 488         struct mds_body *body;
 489         int rc;
 490         struct list_head *t;
 491         int put_child = 1;
 492         ENTRY;
 493
 494         LASSERT(offset == 3); /* only called via intent */
 495         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 496         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 497
 498         /* copy rc, transno and disp; steal locks */
 499         mds_req_from_mcd(req, mcd);
 500         intent_set_disposition(rep, mcd->mcd_last_data);
 501
 502         /* Only replay if create or open actually happened. */
 503         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 504                 EXIT;
 505                 return; /* error looking up parent or child */
 506         }
 507
 508         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 509         LASSERTF(!IS_ERR(parent), "fid "LPU64"/%u rc %ld\n", rec->ur_fid1->id,
 510                  rec->ur_fid1->generation, PTR_ERR(parent));
 511
 512         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 513         LASSERTF(!IS_ERR(child), "parent "LPU64"/%u child %s rc %ld\n",
 514                  rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
 515                  PTR_ERR(child));
 516
 517         if (!child->d_inode)
 518                 GOTO(out_dput, 0); /* child not present to open */
 519
 520         /* At this point, we know we have a child. We'll send
 521          * it back _unless_ it not created and open failed.  */
 522         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 523             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 524             req->rq_status)
 525                 GOTO(out_dput, 0);
 526
 527         mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
 528         mds_pack_inode2body(obd, body, child->d_inode);
 529         if (S_ISREG(child->d_inode->i_mode)) {
 530                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 531                                  child->d_inode, 1);
 532
 533                 if (rc)
 534                         LASSERT(rc == req->rq_status);
 535
 536                 /* If we have LOV EA data, the OST holds size, mtime */
 537                 if (!(body->valid & OBD_MD_FLEASIZE))
 538                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 539                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 540         } else {
 541                 /* XXX need to check this case */
 542         }
 543
 544         /* If we're opening a file without an EA, change to a write
 545            lock (unless we already have one). */
 546
 547         /* If we have -EEXIST as the status, and we were asked to create
 548          * exclusively, we can tell we failed because the file already existed.
 549          */
 550         if (req->rq_status == -EEXIST &&
 551             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 552              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 553                 GOTO(out_dput, 0);
 554         }
 555
 556         /* If we didn't get as far as trying to open, then some locking thing
 557          * probably went wrong, and we'll just bail here.
 558          */
 559         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 560                 GOTO(out_dput, 0);
 561
 562         /* If we failed, then we must have failed opening, so don't look for
 563          * file descriptor or anything, just give the client the bad news.
 564          */
 565         if (req->rq_status)
 566                 GOTO(out_dput, 0);
 567
 568         mfd = NULL;
 569         list_for_each(t, &med->med_open_head) {
 570                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 571                 if (mfd->mfd_xid == req->rq_xid)
 572                         break;
 573                 mfd = NULL;
 574         }
 575
 576         /* #warning "XXX fixme" bug 2991 */
 577         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 578          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 579          * to detect a re-open */
 580         if (mfd == NULL) {
 581                 mntget(mds->mds_vfsmnt);
 582                 CERROR("Re-opened file \n");
 583                 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
 584                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 585                 if (!mfd) {
 586                         CERROR("mds: out of memory\n");
 587                         GOTO(out_dput, req->rq_status = -ENOMEM);
 588                 }
 589                 put_child = 0;
 590         } else {
 591                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 592                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 593                        mfd->mfd_handle.h_cookie);
 594         }
 595
 596  out_dput:
 597         if (put_child)
 598                 l_dput(child);
 599         l_dput(parent);
 600         EXIT;
 601 }
 602
 603 /* do NOT or the MAY_*'s, you'll get the weakest */
 604 static int accmode(int flags)
 605 {
 606         int res = 0;
 607
 608         if (flags & FMODE_READ)
 609                 res = MAY_READ;
 610         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 611                 res |= MAY_WRITE;
 612         if (flags & FMODE_EXEC)
 613                 res = MAY_EXEC;
 614         return res;
 615 }
 616
 617 /* Handles object creation, actual opening, and I/O epoch */
 618 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 619                            struct mds_body *body, int flags, void **handle,
 620                            struct mds_update_record *rec,struct ldlm_reply *rep)
 621 {
 622         struct mds_obd *mds = mds_req2mds(req);
 623         struct obd_device *obd = req->rq_export->exp_obd;
 624         struct mds_file_data *mfd = NULL;
 625         obd_id *ids = NULL; /* object IDs created */
 626         unsigned mode;
 627         int rc = 0;
 628         ENTRY;
 629
 630         /* atomically create objects if necessary */
 631         down(&dchild->d_inode->i_sem);
 632         mode = dchild->d_inode->i_mode;
 633         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 634             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 635                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 636                                  dchild->d_inode, 0);
 637                 if (rc) {
 638                         up(&dchild->d_inode->i_sem);
 639                         RETURN(rc);
 640                 }
 641         }
 642         if (rec != NULL) {
 643                 /* no EA: create objects */
 644                 rc = mds_create_objects(req, 2, rec, mds, obd,
 645                                         dchild, handle, &ids);
 646                 if (rc) {
 647                         CERROR("mds_create_objects: rc = %d\n", rc);
 648                         up(&dchild->d_inode->i_sem);
 649                         RETURN(rc);
 650                 }
 651                 if (S_ISREG(dchild->d_inode->i_mode) &&
 652                     (body->valid & OBD_MD_FLEASIZE)) {
 653                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 654                                                    req->rq_repmsg, 2);
 655                         if (!rc)
 656                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 657                                                  dchild->d_inode, 0);
 658                         if (rc) {
 659                                 up(&dchild->d_inode->i_sem);
 660                                 RETURN(rc);
 661                         }
 662                 }
 663         }
 664         /* If the inode has no EA data, then MDSs hold size, mtime */
 665         if (S_ISREG(dchild->d_inode->i_mode) &&
 666             !(body->valid & OBD_MD_FLEASIZE)) {
 667                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 668                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 669         }
 670         up(&dchild->d_inode->i_sem);
 671
 672         intent_set_disposition(rep, DISP_OPEN_OPEN);
 673         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 674         if (IS_ERR(mfd))
 675                 RETURN(PTR_ERR(mfd));
 676
 677         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 678                mfd->mfd_handle.h_cookie);
 679         if (ids != NULL) {
 680                 mds_lov_update_objids(obd, ids);
 681                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 682         }
 683         //if (rc)
 684         //        mds_mfd_destroy(mfd);
 685         RETURN(rc);
 686 }
 687
 688 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 689                            struct mds_body *body, int flags,
 690                            struct mds_update_record *rec,struct ldlm_reply *rep)
 691 {
 692         struct mds_obd *mds = mds_req2mds(req);
 693         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 694         struct dentry *dchild;
 695         char fidname[LL_FID_NAMELEN];
 696         int fidlen = 0, rc;
 697         void *handle = NULL;
 698         ENTRY;
 699
 700         down(&pending_dir->i_sem);
 701         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 702         dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 703         if (IS_ERR(dchild)) {
 704                 up(&pending_dir->i_sem);
 705                 rc = PTR_ERR(dchild);
 706                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 707                 RETURN(rc);
 708         }
 709
 710         if (dchild->d_inode != NULL) {
 711                 up(&pending_dir->i_sem);
 712                 mds_inode_set_orphan(dchild->d_inode);
 713                 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 714                 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 715                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 716                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 717                 CWARN("Orphan %s found and opened in PENDING directory\n",
 718                        fidname);
 719                 goto open;
 720         }
 721         dput(dchild);
 722         up(&pending_dir->i_sem);
 723
 724         /* We didn't find it in PENDING so it isn't an orphan.  See
 725          * if it was a regular inode that was previously created. */
 726         dchild = mds_fid2dentry(mds, fid, NULL);
 727         if (IS_ERR(dchild))
 728                 RETURN(PTR_ERR(dchild));
 729
 730         mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 731         mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 732         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 733         intent_set_disposition(rep, DISP_LOOKUP_POS);
 734
 735  open:
 736         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 737         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 738                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 739         /* XXX what do we do here if mds_finish_transno itself failed? */
 740
 741         l_dput(dchild);
 742         RETURN(rc);
 743 }
 744
 745 int mds_pin(struct ptlrpc_request *req, int offset)
 746 {
 747         struct obd_device *obd = req->rq_export->exp_obd;
 748         struct mds_body *request_body, *reply_body;
 749         struct lvfs_run_ctxt saved;
 750         int rc, size = sizeof(*reply_body);
 751         ENTRY;
 752
 753         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 754                                       sizeof(*request_body));
 755
 756         rc = lustre_pack_reply(req, 1, &size, NULL);
 757         if (rc)
 758                 RETURN(rc);
 759         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 760
 761         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 762         rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
 763                              request_body->flags, NULL, NULL);
 764         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 765
 766         RETURN(rc);
 767 }
 768
 769 /*  Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
 770  *  If child_lockh is NULL we just get the lock as a barrier to wait for
 771  *  other holders of this lock, and drop it right away again. */
 772 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 773                        struct lustre_handle *child_lockh)
 774 {
 775         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 776         struct lustre_handle lockh;
 777         int lock_flags = 0;
 778         int rc;
 779
 780         if (child_lockh == NULL)
 781                 child_lockh = &lockh;
 782
 783         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 784                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 785                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 786                               NULL, 0, NULL, child_lockh);
 787         if (rc != ELDLM_OK)
 788                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 789         else if (child_lockh == &lockh)
 790                 ldlm_lock_decref(child_lockh, LCK_EX);
 791
 792         RETURN(rc);
 793 }
 794
 795 static int is_mount_object(struct dentry *dparent)
 796 {
 797         struct dentry *dchild;
 798
 799         if (!(dparent->d_inode->i_mode & S_ISUID))
 800                 return 0;
 801
 802         dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
 803         if (IS_ERR(dchild) || dchild == NULL)
 804                 return 0;
 805
 806         dput(dchild);
 807         return 1;
 808 }
 809
 810 int mds_open(struct mds_update_record *rec, int offset,
 811              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 812 {
 813         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 814         struct obd_device *obd = req->rq_export->exp_obd;
 815         struct mds_obd *mds = mds_req2mds(req);
 816         struct ldlm_reply *rep = NULL;
 817         struct mds_body *body = NULL;
 818         struct dentry *dchild = NULL, *dparent = NULL;
 819         struct mds_export_data *med;
 820         struct lustre_handle parent_lockh[2];
 821         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 822         int parent_mode = LCK_PR;
 823         void *handle = NULL;
 824         struct dentry_params dp;
 825         struct mea *mea = NULL;
 826         int mea_size, update_mode;
 827         ENTRY;
 828
 829         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %*s mode %o",
 830                   rec->ur_fid1->id, rec->ur_fid1->generation,
 831                   rec->ur_namelen - 1, rec->ur_name, rec->ur_mode);
 832
 833         parent_lockh[0].cookie = 0;
 834         parent_lockh[1].cookie = 0;
 835
 836         if (offset == 3) { /* intent */
 837                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 838                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 839         } else if (offset == 1) { /* non-intent reint */
 840                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 841         } else {
 842                 body = NULL;
 843                 LBUG();
 844         }
 845
 846         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 847
 848         MDS_UPDATE_COUNTER(mds, MDS_OPEN_COUNT);
 849
 850         /* Step 0: If we are passed a fid, then we assume the client already
 851          * opened this file and is only replaying the RPC, so we open the
 852          * inode by fid (at some large expense in security). */
 853         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 854                 DEBUG_REQ(D_HA, req, "open replay");
 855                 CDEBUG(D_HA, "open fid "LPU64"/%u name %*s mode %o\n",
 856                           rec->ur_fid2->id, rec->ur_fid2->generation,
 857                           rec->ur_namelen - 1, rec->ur_name, rec->ur_mode);
 858
 859                 LASSERT(rec->ur_fid2->id);
 860
 861                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 862                                      rec, rep);
 863                 if (rc != -ENOENT)
 864                         RETURN(rc);
 865                 /* We didn't find the correct inode on disk either, so we
 866                  * need to re-create it via a regular replay. */
 867                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 868         } else {
 869                 LASSERT(!rec->ur_fid2->id);
 870         }
 871
 872         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 873
 874         med = &req->rq_export->exp_mds_data;
 875         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 876                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 877                 RETURN(-ENOMEM);
 878         }
 879
 880         acc_mode = accmode(rec->ur_flags);
 881
 882         /* Step 1: Find and lock the parent */
 883         if (rec->ur_flags & O_CREAT) {
 884                 /* XXX Well, in fact we only need this lock mode change if
 885                    in addition to O_CREAT, the file does not exist.
 886                    But we do not know if it exists or not yet */
 887                 parent_mode = LCK_PW;
 888         }
 889
 890         if (rec->ur_namelen == 1) {
 891                 /* client (LMV) wants to open the file by fid */
 892                 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
 893                                 (unsigned) rec->ur_fid1->mds,
 894                                 (unsigned) rec->ur_fid1->id,
 895                                 (unsigned) rec->ur_fid1->generation);
 896                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 897                 if (IS_ERR(dchild)) {
 898                         rc = PTR_ERR(dparent);
 899                         CERROR("child lookup by a fid error %d\n", rc);
 900                         GOTO(cleanup, rc);
 901                 }
 902                 goto got_child;
 903         }
 904
 905         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 906                                         parent_lockh, &update_mode, rec->ur_name,
 907                                         rec->ur_namelen - 1,
 908                                         MDS_INODELOCK_UPDATE);
 909         if (IS_ERR(dparent)) {
 910                 rc = PTR_ERR(dparent);
 911                 CERROR("parent lookup error %d\n", rc);
 912                 GOTO(cleanup, rc);
 913         }
 914         LASSERT(dparent->d_inode != NULL);
 915
 916         cleanup_phase = 1; /* parent dentry and lock */
 917
 918         /* try to retrieve MEA data for this dir */
 919         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 920         if (rc)
 921                 GOTO(cleanup, rc);
 922
 923         if (mea != NULL) {
 924                 /* dir is already splitted, check is requested filename
 925                  * should live at this MDS or at another one */
 926                 int i;
 927                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 928                 if (mea->mea_master != mea->mea_fids[i].mds) {
 929                         CDEBUG(D_OTHER,
 930                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
 931                                " should be %d(%d)\n", obd->obd_name,
 932                                mea->mea_master, dparent->d_inode->i_ino,
 933                                dparent->d_inode->i_generation, rec->ur_name,
 934                                mea->mea_fids[i].mds, i);
 935                         GOTO(cleanup, rc = -ERESTART);
 936                 }
 937         }
 938
 939         /* Step 2: Lookup the child */
 940         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 941         if (IS_ERR(dchild)) {
 942                 rc = PTR_ERR(dchild);
 943                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 944                 GOTO(cleanup, rc);
 945         }
 946
 947 got_child:
 948         cleanup_phase = 2; /* child dentry */
 949
 950         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 951
 952         if (dchild->d_flags & DCACHE_CROSS_REF) {
 953                 struct ldlm_res_id res_id = { . name = {0} };
 954                 ldlm_policy_data_t policy;
 955                 int flags = 0;
 956                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 957                        (unsigned long) dchild->d_mdsnum,
 958                        (unsigned long) dchild->d_inum,
 959                        (unsigned long) dchild->d_generation);
 960                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 961                 body->fid1.id = dchild->d_inum;
 962                 body->fid1.mds = dchild->d_mdsnum;
 963                 body->fid1.generation = dchild->d_generation;
 964                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 965                 res_id.name[0] = dchild->d_inum;
 966                 res_id.name[1] = dchild->d_generation;
 967                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 968                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 969                                       res_id, LDLM_IBITS, &policy,
 970                                       LCK_PR, &flags,
 971                                       mds_blocking_ast,
 972                                       ldlm_completion_ast, NULL, NULL,
 973                                       NULL, 0, NULL, child_lockh);
 974 #ifdef S_PDIROPS
 975                 if (parent_lockh[1].cookie != 0)
 976                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 977 #endif
 978                 ldlm_lock_decref(parent_lockh, parent_mode);
 979                 if (mea)
 980                         OBD_FREE(mea, mea_size);
 981                 l_dput(dchild);
 982                 l_dput(dparent);
 983                 RETURN(rc);
 984         }
 985
 986         if (dchild->d_inode)
 987                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 988         else
 989                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 990
 991         /*Step 3: If the child was negative, and we're supposed to, create it.*/
 992         if (dchild->d_inode == NULL) {
 993                 unsigned long ino = rec->ur_fid2->id;
 994                 struct iattr iattr;
 995                 struct inode *inode;
 996                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, update_mode);
 997                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
 998                        obd->obd_name, dparent->d_inode->i_ino,
 999                        dparent->d_inode->i_generation, rc);
1000                 if (rc > 0) {
1001                         /* dir got splitted */
1002                         GOTO(cleanup, rc = -ERESTART);
1003                 } else if (rc < 0) {
1004                         /* error happened during spitting */
1005                         GOTO(cleanup, rc);
1006                 }
1007
1008                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1009                         /* It's negative and we weren't supposed to create it */
1010                         GOTO(cleanup, rc = -ENOENT);
1011                 }
1012
1013                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1014                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1015                                       NULL);
1016                 if (IS_ERR(handle)) {
1017                         rc = PTR_ERR(handle);
1018                         handle = NULL;
1019                         GOTO(cleanup, rc);
1020                 }
1021                 dchild->d_fsdata = (void *) &dp;
1022                 dp.p_ptr = req;
1023                 dp.p_inum = ino;
1024
1025                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1026                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1027                         dchild->d_fsdata = NULL;
1028
1029                 if (rc) {
1030                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1031                         GOTO(cleanup, rc);
1032                 }
1033                 inode = dchild->d_inode;
1034                 if (ino) {
1035                         LASSERT(ino == inode->i_ino);
1036                         /* Written as part of setattr */
1037                         inode->i_generation = rec->ur_fid2->generation;
1038                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1039                                inode->i_ino, inode->i_generation);
1040                 }
1041
1042                 created = 1;
1043                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1044                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1045                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1046
1047                 iattr.ia_uid = rec->_ur_fsuid;
1048                 if (dparent->d_inode->i_mode & S_ISGID)
1049                         iattr.ia_gid = dparent->d_inode->i_gid;
1050                 else
1051                         iattr.ia_gid = rec->_ur_fsgid;
1052
1053                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1054                         ATTR_MTIME | ATTR_CTIME;
1055
1056                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1057                 if (rc)
1058                         CERROR("error on child setattr: rc = %d\n", rc);
1059
1060                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1061
1062                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1063                 if (rc)
1064                         CERROR("error on parent setattr: rc = %d\n", rc);
1065                 else
1066                         MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
1067
1068                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1069                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1070                                            dchild->d_inode, handle, 0);
1071                         handle = NULL;
1072                 }
1073
1074                 acc_mode = 0;           /* Don't check for permissions */
1075         }
1076
1077         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1078
1079         mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1080         mds_pack_inode2body(obd, body, dchild->d_inode);
1081
1082         if (S_ISREG(dchild->d_inode->i_mode)) {
1083                 /* Check permissions etc */
1084                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1085                 if (rc != 0)
1086                         GOTO(cleanup, rc);
1087
1088                 /* Can't write to a read-only file */
1089                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1090                         GOTO(cleanup, rc = -EPERM);
1091
1092                 /* An append-only file must be opened in append mode for
1093                  * writing */
1094                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1095                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1096                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1097                         GOTO(cleanup, rc = -EPERM);
1098         }
1099
1100         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1101             (rec->ur_flags & MDS_OPEN_EXCL)) {
1102                 /* File already exists, we didn't just create it, and we
1103                  * were passed O_EXCL; err-or. */
1104                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1105         }
1106
1107         /* if we are following a symlink, don't open */
1108         if (S_ISLNK(dchild->d_inode->i_mode))
1109                 GOTO(cleanup, rc = 0);
1110
1111         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1112             !S_ISDIR(dchild->d_inode->i_mode))
1113                 GOTO(cleanup, rc = -ENOTDIR);
1114
1115         if (S_ISDIR(dchild->d_inode->i_mode)) {
1116                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1117                     rec->ur_flags & FMODE_WRITE) {
1118                         /*we are trying to create or write a exist dir*/
1119                         GOTO(cleanup, rc = -EISDIR);
1120                 }
1121                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1122                         GOTO(cleanup, rc = -EACCES);
1123                 }
1124                 if (is_mount_object(dchild)) {
1125                         CERROR("Found possible GNS mount object %*s; not "
1126                                "opening.\n", dchild->d_name.len,
1127                                dchild->d_name.name);
1128                         GOTO(cleanup, rc = 0); // success, but don't really open
1129                 }
1130         }
1131
1132         /* Step 5: mds_open it */
1133         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1134                              rep);
1135         GOTO(cleanup, rc);
1136
1137  cleanup:
1138         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1139                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1140
1141         switch (cleanup_phase) {
1142         case 2:
1143                 if (rc && created) {
1144                         int err = vfs_unlink(dparent->d_inode, dchild);
1145                         if (err) {
1146                                 CERROR("unlink(%*s) in error path: %d\n",
1147                                        dchild->d_name.len, dchild->d_name.name,
1148                                        err);
1149                         }
1150                 } else if (created) {
1151 #if 0
1152                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1153 #endif
1154                 }
1155                 l_dput(dchild);
1156         case 1:
1157                 if (dparent == NULL)
1158                         break;
1159
1160                 l_dput(dparent);
1161 #ifdef S_PDIROPS
1162                 if (parent_lockh[1].cookie != 0)
1163                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1164 #endif
1165                 if (rc)
1166                         ldlm_lock_decref(parent_lockh, parent_mode);
1167                 else
1168                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1169         }
1170         if (mea)
1171                 OBD_FREE(mea, mea_size);
1172         RETURN(rc);
1173 }
1174
1175 /* Close a "file descriptor" and possibly unlink an orphan from the
1176  * PENDING directory.
1177  *
1178  * If we are being called from mds_disconnect() because the client has
1179  * disappeared, then req == NULL and we do not update last_rcvd because
1180  * there is nothing that could be recovered by the client at this stage
1181  * (it will not even _have_ an entry in last_rcvd anymore).
1182  *
1183  * Returns EAGAIN if the client needs to get more data and re-close. */
1184 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1185                   struct obd_device *obd, struct mds_file_data *mfd,
1186                   int unlink_orphan)
1187 {
1188         struct inode *inode = mfd->mfd_dentry->d_inode;
1189         char fidname[LL_FID_NAMELEN];
1190         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1191         struct dentry *pending_child = NULL;
1192         struct mds_obd *mds = &obd->u.mds;
1193         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1194         void *handle = NULL;
1195         struct mds_body *request_body = NULL, *reply_body = NULL;
1196         struct dentry_params dp;
1197         struct iattr iattr = { 0 };
1198         struct llog_create_locks *lcl = NULL;
1199         ENTRY;
1200
1201         if (req && req->rq_reqmsg != NULL)
1202                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1203                                               sizeof(*request_body));
1204         if (req && req->rq_repmsg != NULL)
1205                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1206                                             sizeof(*reply_body));
1207
1208         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1209
1210         last_orphan = mds_open_orphan_dec_test(inode) &&
1211                 mds_inode_is_orphan(inode);
1212
1213         /* this is half of the actual "close" */
1214         if (mfd->mfd_mode & FMODE_WRITE) {
1215                 rc = mds_put_write_access(mds, inode, request_body,
1216                                           last_orphan && unlink_orphan);
1217         } else if (mfd->mfd_mode & FMODE_EXEC) {
1218                 mds_allow_write_access(inode);
1219         }
1220
1221         if (last_orphan && unlink_orphan) {
1222                 struct lov_mds_md *lmm = NULL;
1223                 int stripe_count = 0;
1224                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1225
1226                 if (obd->obd_recovering) {
1227                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1228                                " is over\n", fidname);
1229                         GOTO(out, rc);
1230                 }
1231
1232                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1233
1234                 /* Sadly, there is no easy way to save pending_child from
1235                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1236                  * but normally it will still be in the dcache. */
1237                 pending_dir = mds->mds_pending_dir->d_inode;
1238                 down(&pending_dir->i_sem);
1239                 cleanup_phase = 1; /* up(i_sem) when finished */
1240                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1241                                                fidlen);
1242                 if (IS_ERR(pending_child))
1243                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1244                 LASSERT(pending_child->d_inode != NULL);
1245
1246                 cleanup_phase = 2; /* dput(pending_child) when finished */
1247                 if (req != NULL && req->rq_repmsg != NULL) {
1248                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1249                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1250                 }
1251
1252                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1253                                           NULL, stripe_count);
1254                 if (IS_ERR(handle)) {
1255                         rc = PTR_ERR(handle);
1256                         handle = NULL;
1257                         GOTO(cleanup, rc);
1258                 }
1259
1260                 pending_child->d_fsdata = (void *) &dp;
1261                 dp.p_inum = 0;
1262                 dp.p_ptr = req;
1263                 if (S_ISDIR(pending_child->d_inode->i_mode))
1264                         rc = vfs_rmdir(pending_dir, pending_child);
1265                 else
1266                         rc = vfs_unlink(pending_dir, pending_child);
1267                 if (rc)
1268                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1269
1270                 if (req != NULL && req->rq_repmsg != NULL &&
1271                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1272                     mds_log_op_unlink(obd, pending_child->d_inode,
1273                                       lmm, req->rq_repmsg->buflens[1],
1274                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1275                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1276                         reply_body->valid |= OBD_MD_FLCOOKIE;
1277                 }
1278
1279                 goto out; /* Don't bother updating attrs on unlinked inode */
1280         }
1281
1282         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1283                 /* Update the on-disk attributes if this was the last write
1284                  * close, and all information was provided (i.e., rc == 0)
1285                  *
1286                  * XXX this should probably be abstracted with mds_reint_setattr
1287                  */
1288
1289 #if 0
1290                 if (request_body->valid & OBD_MD_FLMTIME &&
1291                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1292                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1293                         iattr.ia_valid |= ATTR_MTIME;
1294                 }
1295                 if (request_body->valid & OBD_MD_FLCTIME &&
1296                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1297                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1298                         iattr.ia_valid |= ATTR_CTIME;
1299                 }
1300
1301                 /* XXX can't set block count with fsfilt_setattr (!) */
1302                 if (request_body->valid & OBD_MD_FLSIZE) {
1303                         iattr.ia_valid |= ATTR_SIZE;
1304                         iattr.ia_size = request_body->size;
1305                 }
1306                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1307                         iattr.ia_valid |= ATTR_BLOCKS;
1308                         iattr.ia_blocks = request_body->blocks
1309                 } */
1310
1311 #endif
1312         }
1313
1314         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1315                 /* Only start a transaction to write out only the atime if
1316                  * it is more out-of-date than the specified limit.  If we
1317                  * are already going to write out the atime then do it anyway.
1318                  * */
1319                 if ((request_body->atime >
1320                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1321                     (iattr.ia_valid != 0 &&
1322                      request_body->atime > LTIME_S(inode->i_atime))) {
1323                         LTIME_S(iattr.ia_atime) = request_body->atime;
1324                         iattr.ia_valid |= ATTR_ATIME;
1325                 }
1326         }
1327
1328         if (iattr.ia_valid != 0) {
1329                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1330                 if (IS_ERR(handle))
1331                         GOTO(cleanup, rc = PTR_ERR(handle));
1332                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1333                 if (rc)
1334                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1335         }
1336 out:
1337         /* If other clients have this file open for write, rc will be > 0 */
1338         if (rc > 0)
1339                 rc = 0;
1340         l_dput(mfd->mfd_dentry);
1341         mds_mfd_destroy(mfd);
1342
1343  cleanup:
1344         if (req != NULL && reply_body != NULL) {
1345                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1346         } else if (handle) {
1347                 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1348                 if (err) {
1349                         CERROR("error committing close: %d\n", err);
1350                         if (!rc)
1351                                 rc = err;
1352                 }
1353         }
1354
1355         switch (cleanup_phase) {
1356         case 2:
1357                 if (lcl != NULL)
1358                         ptlrpc_save_llog_lock(req, lcl);
1359                 dput(pending_child);
1360         case 1:
1361                 up(&pending_dir->i_sem);
1362         }
1363         RETURN(rc);
1364 }
1365
1366 int mds_close(struct ptlrpc_request *req, int offset)
1367 {
1368         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1369         struct obd_device *obd = req->rq_export->exp_obd;
1370         struct mds_body *body;
1371         struct mds_file_data *mfd;
1372         struct lvfs_run_ctxt saved;
1373         struct inode *inode;
1374         int rc, repsize[3] = {sizeof(struct mds_body),
1375                               obd->u.mds.mds_max_mdsize,
1376                               obd->u.mds.mds_max_cookiesize};
1377         ENTRY;
1378
1379         MDS_UPDATE_COUNTER((&obd->u.mds), MDS_CLOSE_COUNT);
1380
1381         rc = lustre_pack_reply(req, 3, repsize, NULL);
1382         if (rc) {
1383                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1384                 req->rq_status = rc;
1385         } else {
1386                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1387         }
1388
1389         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1390                 DEBUG_REQ(D_HA, req, "close replay\n");
1391                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1392                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1393                        req->rq_repmsg->buflens[2]);
1394         }
1395
1396
1397         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1398                                   lustre_swab_mds_body);
1399         if (body == NULL) {
1400                 CERROR("Can't unpack body\n");
1401                 req->rq_status = -EFAULT;
1402                 RETURN(-EFAULT);
1403         }
1404
1405         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1406                 /* do some stuff */ ;
1407
1408         mfd = mds_handle2mfd(&body->handle);
1409         if (mfd == NULL) {
1410                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1411                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1412                 req->rq_status = -ESTALE;
1413                 RETURN(-ESTALE);
1414         }
1415
1416         inode = mfd->mfd_dentry->d_inode;
1417         if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1418                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1419                 LASSERT(body != NULL);
1420
1421                 mds_pack_inode2fid(obd, &body->fid1, inode);
1422                 mds_pack_inode2body(obd, body, inode);
1423                 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1424         }
1425         spin_lock(&med->med_open_lock);
1426         list_del(&mfd->mfd_list);
1427         spin_unlock(&med->med_open_lock);
1428
1429         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1430         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1431         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1432
1433         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1434                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1435                 req->rq_status = -ENOMEM;
1436                 mds_mfd_put(mfd);
1437                 RETURN(-ENOMEM);
1438         }
1439
1440         mds_mfd_put(mfd);
1441         RETURN(0);
1442 }
1443
1444 int mds_done_writing(struct ptlrpc_request *req, int offset)
1445 {
1446         struct mds_body *body;
1447         int rc, size = sizeof(struct mds_body);
1448         ENTRY;
1449
1450         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1451
1452         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1453                                   lustre_swab_mds_body);
1454         if (body == NULL) {
1455                 CERROR("Can't unpack body\n");
1456                 req->rq_status = -EFAULT;
1457                 RETURN(-EFAULT);
1458         }
1459
1460         rc = lustre_pack_reply(req, 1, &size, NULL);
1461         if (rc) {
1462                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1463                 req->rq_status = rc;
1464         }
1465
1466         RETURN(0);
1467 }