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         if (*handle == NULL)
 337                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 338         if (IS_ERR(*handle)) {
 339                 rc = PTR_ERR(*handle);
 340                 *handle = NULL;
 341                 GOTO(out_ids, rc);
 342         }
 343
 344         /* replay case */
 345         if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 346                 LASSERT (rec->ur_fid2->id);
 347                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 348                 lmm_size = rec->ur_eadatalen;
 349                 lmm = rec->ur_eadata;
 350                 LASSERT(lmm);
 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         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 450         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 451         lmm_bufsize = req->rq_repmsg->buflens[offset];
 452         LASSERT(lmm_buf);
 453         LASSERT(lmm_bufsize >= lmm_size);
 454
 455         memcpy(lmm_buf, lmm, lmm_size);
 456         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 457  out_oa:
 458         oti_free_cookies(&oti);
 459         obdo_free(oa);
 460  out_ids:
 461         if (rc) {
 462                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 463                 *ids = NULL;
 464         }
 465         RETURN(rc);
 466 }
 467
 468 static void reconstruct_open(struct mds_update_record *rec, int offset,
 469                              struct ptlrpc_request *req,
 470                              struct lustre_handle *child_lockh)
 471 {
 472         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 473         struct mds_client_data *mcd = med->med_mcd;
 474         struct mds_obd *mds = mds_req2mds(req);
 475         struct mds_file_data *mfd;
 476         struct obd_device *obd = req->rq_export->exp_obd;
 477         struct dentry *parent, *child;
 478         struct ldlm_reply *rep;
 479         struct mds_body *body;
 480         int rc;
 481         struct list_head *t;
 482         int put_child = 1;
 483         ENTRY;
 484
 485         LASSERT(offset == 2);                  /* only called via intent */
 486         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 487         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 488
 489         /* copy rc, transno and disp; steal locks */
 490         mds_req_from_mcd(req, mcd);
 491         intent_set_disposition(rep, mcd->mcd_last_data);
 492
 493         /* Only replay if create or open actually happened. */
 494         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 495                 EXIT;
 496                 return; /* error looking up parent or child */
 497         }
 498
 499         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 500         LASSERTF(!IS_ERR(parent), "fid "LPU64"/%u rc %ld\n", rec->ur_fid1->id,
 501                  rec->ur_fid1->generation, PTR_ERR(parent));
 502
 503         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 504         LASSERTF(!IS_ERR(child), "parent "LPU64"/%u child %s rc %ld\n",
 505                  rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
 506                  PTR_ERR(child));
 507
 508         if (!child->d_inode)
 509                 GOTO(out_dput, 0); /* child not present to open */
 510
 511         /* At this point, we know we have a child. We'll send
 512          * it back _unless_ it not created and open failed.  */
 513         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 514             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 515             req->rq_status)
 516                 GOTO(out_dput, 0);
 517
 518         mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
 519         mds_pack_inode2body(obd, body, child->d_inode);
 520         if (S_ISREG(child->d_inode->i_mode)) {
 521                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 522                                  child->d_inode, 1);
 523
 524                 if (rc)
 525                         LASSERT(rc == req->rq_status);
 526
 527                 /* If we have LOV EA data, the OST holds size, mtime */
 528                 if (!(body->valid & OBD_MD_FLEASIZE))
 529                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 530                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 531         } else {
 532                 /* XXX need to check this case */
 533         }
 534
 535         /* If we're opening a file without an EA, change to a write
 536            lock (unless we already have one). */
 537
 538         /* If we have -EEXIST as the status, and we were asked to create
 539          * exclusively, we can tell we failed because the file already existed.
 540          */
 541         if (req->rq_status == -EEXIST &&
 542             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 543              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 544                 GOTO(out_dput, 0);
 545         }
 546
 547         /* If we didn't get as far as trying to open, then some locking thing
 548          * probably went wrong, and we'll just bail here.
 549          */
 550         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 551                 GOTO(out_dput, 0);
 552
 553         /* If we failed, then we must have failed opening, so don't look for
 554          * file descriptor or anything, just give the client the bad news.
 555          */
 556         if (req->rq_status)
 557                 GOTO(out_dput, 0);
 558
 559         mfd = NULL;
 560         list_for_each(t, &med->med_open_head) {
 561                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 562                 if (mfd->mfd_xid == req->rq_xid)
 563                         break;
 564                 mfd = NULL;
 565         }
 566
 567         /* #warning "XXX fixme" bug 2991 */
 568         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 569          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 570          * to detect a re-open */
 571         if (mfd == NULL) {
 572                 mntget(mds->mds_vfsmnt);
 573                 CERROR("Re-opened file \n");
 574                 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
 575                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 576                 if (!mfd) {
 577                         CERROR("mds: out of memory\n");
 578                         GOTO(out_dput, req->rq_status = -ENOMEM);
 579                 }
 580                 put_child = 0;
 581         } else {
 582                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 583                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 584                        mfd->mfd_handle.h_cookie);
 585         }
 586
 587  out_dput:
 588         if (put_child)
 589                 l_dput(child);
 590         l_dput(parent);
 591         EXIT;
 592 }
 593
 594 /* do NOT or the MAY_*'s, you'll get the weakest */
 595 static int accmode(int flags)
 596 {
 597         int res = 0;
 598
 599         if (flags & FMODE_READ)
 600                 res = MAY_READ;
 601         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 602                 res |= MAY_WRITE;
 603         if (flags & FMODE_EXEC)
 604                 res = MAY_EXEC;
 605         return res;
 606 }
 607
 608 /* Handles object creation, actual opening, and I/O epoch */
 609 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 610                            struct mds_body *body, int flags, void **handle,
 611                            struct mds_update_record *rec,struct ldlm_reply *rep)
 612 {
 613         struct mds_obd *mds = mds_req2mds(req);
 614         struct obd_device *obd = req->rq_export->exp_obd;
 615         struct mds_file_data *mfd = NULL;
 616         obd_id *ids = NULL; /* object IDs created */
 617         unsigned mode;
 618         int rc = 0;
 619         ENTRY;
 620
 621         /* atomically create objects if necessary */
 622         down(&dchild->d_inode->i_sem);
 623         mode = dchild->d_inode->i_mode;
 624         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 625             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 626                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 627                                  dchild->d_inode, 0);
 628                 if (rc) {
 629                         up(&dchild->d_inode->i_sem);
 630                         RETURN(rc);
 631                 }
 632         }
 633         if (rec != NULL) {
 634                 /* no EA: create objects */
 635                 rc = mds_create_objects(req, 2, rec, mds, obd,
 636                                         dchild, handle, &ids);
 637                 if (rc) {
 638                         CERROR("mds_create_objects: rc = %d\n", rc);
 639                         up(&dchild->d_inode->i_sem);
 640                         RETURN(rc);
 641                 }
 642                 if (S_ISREG(dchild->d_inode->i_mode) &&
 643                     (body->valid & OBD_MD_FLEASIZE)) {
 644                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 645                                                    req->rq_repmsg, 2);
 646                         if (!rc)
 647                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 648                                                  dchild->d_inode, 0);
 649                         if (rc) {
 650                                 up(&dchild->d_inode->i_sem);
 651                                 RETURN(rc);
 652                         }
 653                 }
 654         }
 655         /* If the inode has no EA data, then MDSs hold size, mtime */
 656         if (S_ISREG(dchild->d_inode->i_mode) &&
 657             !(body->valid & OBD_MD_FLEASIZE)) {
 658                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 659                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 660         }
 661         up(&dchild->d_inode->i_sem);
 662
 663         intent_set_disposition(rep, DISP_OPEN_OPEN);
 664         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 665         if (IS_ERR(mfd))
 666                 RETURN(PTR_ERR(mfd));
 667
 668         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 669                mfd->mfd_handle.h_cookie);
 670         if (ids != NULL) {
 671                 mds_lov_update_objids(obd, ids);
 672                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 673         }
 674         //if (rc)
 675         //        mds_mfd_destroy(mfd);
 676         RETURN(rc);
 677 }
 678
 679 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 680                            struct mds_body *body, int flags,
 681                            struct mds_update_record *rec,struct ldlm_reply *rep)
 682 {
 683         struct mds_obd *mds = mds_req2mds(req);
 684         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 685         struct dentry *dchild;
 686         char fidname[LL_FID_NAMELEN];
 687         int fidlen = 0, rc;
 688         void *handle = NULL;
 689         ENTRY;
 690
 691         down(&pending_dir->i_sem);
 692         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 693         dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 694         if (IS_ERR(dchild)) {
 695                 up(&pending_dir->i_sem);
 696                 rc = PTR_ERR(dchild);
 697                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 698                 RETURN(rc);
 699         }
 700
 701         if (dchild->d_inode != NULL) {
 702                 up(&pending_dir->i_sem);
 703                 mds_inode_set_orphan(dchild->d_inode);
 704                 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 705                 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 706                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 707                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 708                 CWARN("Orphan %s found and opened in PENDING directory\n",
 709                        fidname);
 710                 goto open;
 711         }
 712         dput(dchild);
 713         up(&pending_dir->i_sem);
 714
 715         /* We didn't find it in PENDING so it isn't an orphan.  See
 716          * if it was a regular inode that was previously created. */
 717         dchild = mds_fid2dentry(mds, fid, NULL);
 718         if (IS_ERR(dchild))
 719                 RETURN(PTR_ERR(dchild));
 720
 721         mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 722         mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 723         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 724         intent_set_disposition(rep, DISP_LOOKUP_POS);
 725
 726  open:
 727         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 728         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 729                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 730         /* XXX what do we do here if mds_finish_transno itself failed? */
 731
 732         l_dput(dchild);
 733         RETURN(rc);
 734 }
 735
 736 int mds_pin(struct ptlrpc_request *req)
 737 {
 738         struct obd_device *obd = req->rq_export->exp_obd;
 739         struct mds_body *request_body, *reply_body;
 740         struct lvfs_run_ctxt saved;
 741         int rc, size = sizeof(*reply_body);
 742         ENTRY;
 743
 744         request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
 745
 746         rc = lustre_pack_reply(req, 1, &size, NULL);
 747         if (rc)
 748                 RETURN(rc);
 749         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 750
 751         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 752         rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
 753                              request_body->flags, NULL, NULL);
 754         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 755
 756         RETURN(rc);
 757 }
 758
 759 /*  Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
 760  *  If child_lockh is NULL we just get the lock as a barrier to wait for
 761  *  other holders of this lock, and drop it right away again. */
 762 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 763                        struct lustre_handle *child_lockh)
 764 {
 765         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 766         struct lustre_handle lockh;
 767         int lock_flags = 0;
 768         int rc;
 769
 770         if (child_lockh == NULL)
 771                 child_lockh = &lockh;
 772
 773         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 774                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 775                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 776                               NULL, 0, NULL, child_lockh);
 777         if (rc != ELDLM_OK)
 778                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 779         else if (child_lockh == &lockh)
 780                 ldlm_lock_decref(child_lockh, LCK_EX);
 781
 782         RETURN(rc);
 783 }
 784
 785 static int is_mount_object(struct dentry *dparent)
 786 {
 787         struct dentry *dchild;
 788
 789         if (!(dparent->d_inode->i_mode & S_ISUID))
 790                 return 0;
 791
 792         dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
 793         if (IS_ERR(dchild) || dchild == NULL)
 794                 return 0;
 795
 796         dput(dchild);
 797         return 1;
 798 }
 799
 800 int mds_open(struct mds_update_record *rec, int offset,
 801              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 802 {
 803         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 804         struct obd_device *obd = req->rq_export->exp_obd;
 805         struct mds_obd *mds = mds_req2mds(req);
 806         struct ldlm_reply *rep = NULL;
 807         struct mds_body *body = NULL;
 808         struct dentry *dchild = NULL, *dparent = NULL;
 809         struct mds_export_data *med;
 810         struct lustre_handle parent_lockh[2];
 811         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 812         int parent_mode = LCK_PR;
 813         void *handle = NULL;
 814         struct dentry_params dp;
 815         struct mea *mea = NULL;
 816         int mea_size, update_mode;
 817         ENTRY;
 818
 819         DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %*s mode %o",
 820                   rec->ur_fid1->id, rec->ur_fid1->generation,
 821                   rec->ur_namelen - 1, rec->ur_name, rec->ur_mode);
 822
 823         parent_lockh[0].cookie = 0;
 824         parent_lockh[1].cookie = 0;
 825
 826         if (offset == 2) { /* intent */
 827                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 828                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 829         } else if (offset == 0) { /* non-intent reint */
 830                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 831         } else {
 832                 body = NULL;
 833                 LBUG();
 834         }
 835
 836         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 837
 838         MDS_UPDATE_COUNTER(mds, MDS_OPEN_COUNT);
 839
 840         /* Step 0: If we are passed a fid, then we assume the client already
 841          * opened this file and is only replaying the RPC, so we open the
 842          * inode by fid (at some large expense in security). */
 843         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 844                 DEBUG_REQ(D_HA, req, "open replay");
 845                 CDEBUG(D_HA, "open fid "LPU64"/%u name %*s mode %o\n",
 846                           rec->ur_fid2->id, rec->ur_fid2->generation,
 847                           rec->ur_namelen - 1, rec->ur_name, rec->ur_mode);
 848
 849                 LASSERT(rec->ur_fid2->id);
 850
 851                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 852                                      rec, rep);
 853                 if (rc != -ENOENT)
 854                         RETURN(rc);
 855                 /* We didn't find the correct inode on disk either, so we
 856                  * need to re-create it via a regular replay. */
 857                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 858         } else {
 859                 LASSERT(!rec->ur_fid2->id);
 860         }
 861
 862         LASSERT(offset == 2); /* If we got here, we must be called via intent */
 863
 864         med = &req->rq_export->exp_mds_data;
 865         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 866                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 867                 RETURN(-ENOMEM);
 868         }
 869
 870         acc_mode = accmode(rec->ur_flags);
 871
 872         /* Step 1: Find and lock the parent */
 873         if (rec->ur_flags & O_CREAT) {
 874                 /* XXX Well, in fact we only need this lock mode change if
 875                    in addition to O_CREAT, the file does not exist.
 876                    But we do not know if it exists or not yet */
 877                 parent_mode = LCK_PW;
 878         }
 879
 880         if (rec->ur_namelen == 1) {
 881                 /* client (LMV) wants to open the file by fid */
 882                 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
 883                                 (unsigned) rec->ur_fid1->mds,
 884                                 (unsigned) rec->ur_fid1->id,
 885                                 (unsigned) rec->ur_fid1->generation);
 886                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 887                 if (IS_ERR(dchild)) {
 888                         rc = PTR_ERR(dparent);
 889                         CERROR("child lookup by a fid error %d\n", rc);
 890                         GOTO(cleanup, rc);
 891                 }
 892                 goto got_child;
 893         }
 894
 895         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 896                                         parent_lockh, &update_mode, rec->ur_name,
 897                                         rec->ur_namelen - 1,
 898                                         MDS_INODELOCK_UPDATE);
 899         if (IS_ERR(dparent)) {
 900                 rc = PTR_ERR(dparent);
 901                 CERROR("parent lookup error %d\n", rc);
 902                 GOTO(cleanup, rc);
 903         }
 904         LASSERT(dparent->d_inode != NULL);
 905
 906         cleanup_phase = 1; /* parent dentry and lock */
 907
 908         /* try to retrieve MEA data for this dir */
 909         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 910         if (rc)
 911                 GOTO(cleanup, rc);
 912
 913         if (mea != NULL) {
 914                 /* dir is already splitted, check is requested filename
 915                  * should live at this MDS or at another one */
 916                 int i;
 917                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 918                 if (mea->mea_master != mea->mea_fids[i].mds) {
 919                         CDEBUG(D_OTHER,
 920                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
 921                                " should be %d(%d)\n", obd->obd_name,
 922                                mea->mea_master, dparent->d_inode->i_ino,
 923                                dparent->d_inode->i_generation, rec->ur_name,
 924                                mea->mea_fids[i].mds, i);
 925                         GOTO(cleanup, rc = -ERESTART);
 926                 }
 927         }
 928
 929         /* Step 2: Lookup the child */
 930         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 931         if (IS_ERR(dchild)) {
 932                 rc = PTR_ERR(dchild);
 933                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 934                 GOTO(cleanup, rc);
 935         }
 936
 937 got_child:
 938         cleanup_phase = 2; /* child dentry */
 939
 940         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 941
 942         if (dchild->d_flags & DCACHE_CROSS_REF) {
 943                 struct ldlm_res_id res_id = { . name = {0} };
 944                 ldlm_policy_data_t policy;
 945                 int flags = 0;
 946                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 947                        (unsigned long) dchild->d_mdsnum,
 948                        (unsigned long) dchild->d_inum,
 949                        (unsigned long) dchild->d_generation);
 950                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 951                 body->fid1.id = dchild->d_inum;
 952                 body->fid1.mds = dchild->d_mdsnum;
 953                 body->fid1.generation = dchild->d_generation;
 954                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 955                 res_id.name[0] = dchild->d_inum;
 956                 res_id.name[1] = dchild->d_generation;
 957                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 958                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 959                                       res_id, LDLM_IBITS, &policy,
 960                                       LCK_PR, &flags,
 961                                       mds_blocking_ast,
 962                                       ldlm_completion_ast, NULL, NULL,
 963                                       NULL, 0, NULL, child_lockh);
 964 #ifdef S_PDIROPS
 965                 if (parent_lockh[1].cookie != 0)
 966                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 967 #endif
 968                 ldlm_lock_decref(parent_lockh, parent_mode);
 969                 if (mea)
 970                         OBD_FREE(mea, mea_size);
 971                 l_dput(dchild);
 972                 l_dput(dparent);
 973                 RETURN(rc);
 974         }
 975
 976         if (dchild->d_inode)
 977                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 978         else
 979                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 980
 981         /*Step 3: If the child was negative, and we're supposed to, create it.*/
 982         if (dchild->d_inode == NULL) {
 983                 unsigned long ino = rec->ur_fid2->id;
 984                 struct iattr iattr;
 985                 struct inode *inode;
 986                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, update_mode);
 987                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
 988                        obd->obd_name, dparent->d_inode->i_ino,
 989                        dparent->d_inode->i_generation, rc);
 990                 if (rc > 0) {
 991                         /* dir got splitted */
 992                         GOTO(cleanup, rc = -ERESTART);
 993                 } else if (rc < 0) {
 994                         /* error happened during spitting */
 995                         GOTO(cleanup, rc);
 996                 }
 997
 998                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 999                         /* It's negative and we weren't supposed to create it */
1000                         GOTO(cleanup, rc = -ENOENT);
1001                 }
1002
1003                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1004                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1005                                       NULL);
1006                 if (IS_ERR(handle)) {
1007                         rc = PTR_ERR(handle);
1008                         handle = NULL;
1009                         GOTO(cleanup, rc);
1010                 }
1011                 dchild->d_fsdata = (void *) &dp;
1012                 dp.p_ptr = req;
1013                 dp.p_inum = ino;
1014
1015                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1016                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1017                         dchild->d_fsdata = NULL;
1018
1019                 if (rc) {
1020                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1021                         GOTO(cleanup, rc);
1022                 }
1023                 inode = dchild->d_inode;
1024                 if (ino) {
1025                         LASSERT(ino == inode->i_ino);
1026                         /* Written as part of setattr */
1027                         inode->i_generation = rec->ur_fid2->generation;
1028                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1029                                inode->i_ino, inode->i_generation);
1030                 }
1031
1032                 created = 1;
1033                 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
1034                 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1035                 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1036
1037                 iattr.ia_uid = rec->_ur_fsuid;
1038                 if (dparent->d_inode->i_mode & S_ISGID)
1039                         iattr.ia_gid = dparent->d_inode->i_gid;
1040                 else
1041                         iattr.ia_gid = rec->_ur_fsgid;
1042
1043                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1044                         ATTR_MTIME | ATTR_CTIME;
1045
1046                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1047                 if (rc)
1048                         CERROR("error on child setattr: rc = %d\n", rc);
1049
1050                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1051
1052                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1053                 if (rc)
1054                         CERROR("error on parent setattr: rc = %d\n", rc);
1055                 else
1056                         MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
1057
1058                 acc_mode = 0;           /* Don't check for permissions */
1059         }
1060
1061         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1062
1063         mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1064         mds_pack_inode2body(obd, body, dchild->d_inode);
1065
1066         if (S_ISREG(dchild->d_inode->i_mode)) {
1067                 /* Check permissions etc */
1068                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1069                 if (rc != 0)
1070                         GOTO(cleanup, rc);
1071
1072                 /* Can't write to a read-only file */
1073                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1074                         GOTO(cleanup, rc = -EPERM);
1075
1076                 /* An append-only file must be opened in append mode for
1077                  * writing */
1078                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1079                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1080                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1081                         GOTO(cleanup, rc = -EPERM);
1082         }
1083
1084         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1085             (rec->ur_flags & MDS_OPEN_EXCL)) {
1086                 /* File already exists, we didn't just create it, and we
1087                  * were passed O_EXCL; err-or. */
1088                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1089         }
1090
1091         /* if we are following a symlink, don't open */
1092         if (S_ISLNK(dchild->d_inode->i_mode))
1093                 GOTO(cleanup, rc = 0);
1094
1095         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1096             !S_ISDIR(dchild->d_inode->i_mode))
1097                 GOTO(cleanup, rc = -ENOTDIR);
1098
1099         if (S_ISDIR(dchild->d_inode->i_mode)) {
1100                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1101                     rec->ur_flags & FMODE_WRITE) {
1102                         /*we are trying to create or write a exist dir*/
1103                         GOTO(cleanup, rc = -EISDIR);
1104                 }
1105                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1106                         GOTO(cleanup, rc = -EACCES);
1107                 }
1108                 if (is_mount_object(dchild)) {
1109                         CERROR("Found possible GNS mount object %*s; not "
1110                                "opening.\n", dchild->d_name.len,
1111                                dchild->d_name.name);
1112                         GOTO(cleanup, rc = 0); // success, but don't really open
1113                 }
1114         }
1115
1116         /* Step 5: mds_open it */
1117         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1118                              rep);
1119         GOTO(cleanup, rc);
1120
1121  cleanup:
1122         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1123                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1124
1125         switch (cleanup_phase) {
1126         case 2:
1127                 if (rc && created) {
1128                         int err = vfs_unlink(dparent->d_inode, dchild);
1129                         if (err) {
1130                                 CERROR("unlink(%*s) in error path: %d\n",
1131                                        dchild->d_name.len, dchild->d_name.name,
1132                                        err);
1133                         }
1134                 } else if (created) {
1135 #if 0
1136                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1137 #endif
1138                 }
1139                 l_dput(dchild);
1140         case 1:
1141                 if (dparent == NULL)
1142                         break;
1143
1144                 l_dput(dparent);
1145 #ifdef S_PDIROPS
1146                 if (parent_lockh[1].cookie != 0)
1147                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1148 #endif
1149                 if (rc)
1150                         ldlm_lock_decref(parent_lockh, parent_mode);
1151                 else
1152                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1153         }
1154         if (mea)
1155                 OBD_FREE(mea, mea_size);
1156         RETURN(rc);
1157 }
1158
1159 /* Close a "file descriptor" and possibly unlink an orphan from the
1160  * PENDING directory.
1161  *
1162  * If we are being called from mds_disconnect() because the client has
1163  * disappeared, then req == NULL and we do not update last_rcvd because
1164  * there is nothing that could be recovered by the client at this stage
1165  * (it will not even _have_ an entry in last_rcvd anymore).
1166  *
1167  * Returns EAGAIN if the client needs to get more data and re-close. */
1168 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1169                   struct mds_file_data *mfd, int unlink_orphan)
1170 {
1171         struct inode *inode = mfd->mfd_dentry->d_inode;
1172         char fidname[LL_FID_NAMELEN];
1173         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1174         struct dentry *pending_child = NULL;
1175         struct mds_obd *mds = &obd->u.mds;
1176         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1177         void *handle = NULL;
1178         struct mds_body *request_body = NULL, *reply_body = NULL;
1179         struct dentry_params dp;
1180         struct iattr iattr = { 0 };
1181         struct llog_create_locks *lcl = NULL;
1182         ENTRY;
1183
1184         if (req && req->rq_reqmsg != NULL)
1185                 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1186                                               sizeof(*request_body));
1187         if (req && req->rq_repmsg != NULL)
1188                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1189                                             sizeof(*reply_body));
1190
1191         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1192
1193         last_orphan = mds_open_orphan_dec_test(inode) &&
1194                 mds_inode_is_orphan(inode);
1195
1196         /* this is half of the actual "close" */
1197         if (mfd->mfd_mode & FMODE_WRITE) {
1198                 rc = mds_put_write_access(mds, inode, request_body,
1199                                           last_orphan && unlink_orphan);
1200         } else if (mfd->mfd_mode & FMODE_EXEC) {
1201                 mds_allow_write_access(inode);
1202         }
1203
1204         if (last_orphan && unlink_orphan) {
1205                 struct lov_mds_md *lmm = NULL;
1206                 int stripe_count = 0;
1207                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1208
1209                 if (obd->obd_recovering) {
1210                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1211                                " is over\n", fidname);
1212                         GOTO(out, rc);
1213                 }
1214
1215                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1216
1217                 /* Sadly, there is no easy way to save pending_child from
1218                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1219                  * but normally it will still be in the dcache. */
1220                 pending_dir = mds->mds_pending_dir->d_inode;
1221                 down(&pending_dir->i_sem);
1222                 cleanup_phase = 1; /* up(i_sem) when finished */
1223                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1224                                                fidlen);
1225                 if (IS_ERR(pending_child))
1226                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1227                 LASSERT(pending_child->d_inode != NULL);
1228
1229                 cleanup_phase = 2; /* dput(pending_child) when finished */
1230                 if (req != NULL && req->rq_repmsg != NULL) {
1231                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1232                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1233                 }
1234
1235                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1236                                           NULL, stripe_count);
1237                 if (IS_ERR(handle)) {
1238                         rc = PTR_ERR(handle);
1239                         handle = NULL;
1240                         GOTO(cleanup, rc);
1241                 }
1242
1243                 pending_child->d_fsdata = (void *) &dp;
1244                 dp.p_inum = 0;
1245                 dp.p_ptr = req;
1246                 if (S_ISDIR(pending_child->d_inode->i_mode))
1247                         rc = vfs_rmdir(pending_dir, pending_child);
1248                 else
1249                         rc = vfs_unlink(pending_dir, pending_child);
1250                 if (rc)
1251                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1252
1253                 if (req != NULL && req->rq_repmsg != NULL &&
1254                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1255                     mds_log_op_unlink(obd, pending_child->d_inode,
1256                                       lmm, req->rq_repmsg->buflens[1],
1257                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1258                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1259                         reply_body->valid |= OBD_MD_FLCOOKIE;
1260                 }
1261
1262                 goto out; /* Don't bother updating attrs on unlinked inode */
1263         }
1264
1265         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1266                 /* Update the on-disk attributes if this was the last write
1267                  * close, and all information was provided (i.e., rc == 0)
1268                  *
1269                  * XXX this should probably be abstracted with mds_reint_setattr
1270                  */
1271
1272 #if 0
1273                 if (request_body->valid & OBD_MD_FLMTIME &&
1274                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1275                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1276                         iattr.ia_valid |= ATTR_MTIME;
1277                 }
1278                 if (request_body->valid & OBD_MD_FLCTIME &&
1279                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1280                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1281                         iattr.ia_valid |= ATTR_CTIME;
1282                 }
1283
1284                 /* XXX can't set block count with fsfilt_setattr (!) */
1285                 if (request_body->valid & OBD_MD_FLSIZE) {
1286                         iattr.ia_valid |= ATTR_SIZE;
1287                         iattr.ia_size = request_body->size;
1288                 }
1289                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1290                         iattr.ia_valid |= ATTR_BLOCKS;
1291                         iattr.ia_blocks = request_body->blocks
1292                 } */
1293
1294 #endif
1295         }
1296
1297         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1298                 /* Only start a transaction to write out only the atime if
1299                  * it is more out-of-date than the specified limit.  If we
1300                  * are already going to write out the atime then do it anyway.
1301                  * */
1302                 if ((LTIME_S(request_body->atime) >
1303                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1304                     (iattr.ia_valid != 0 &&
1305                      LTIME_S(request_body->atime) > LTIME_S(inode->i_atime))) {
1306                         LTIME_S(iattr.ia_atime) = LTIME_S(request_body->atime);
1307                         iattr.ia_valid |= ATTR_ATIME;
1308                 }
1309         }
1310
1311         if (iattr.ia_valid != 0) {
1312                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1313                 if (IS_ERR(handle))
1314                         GOTO(cleanup, rc = PTR_ERR(handle));
1315                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1316                 if (rc)
1317                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1318         }
1319 out:
1320         /* If other clients have this file open for write, rc will be > 0 */
1321         if (rc > 0)
1322                 rc = 0;
1323         l_dput(mfd->mfd_dentry);
1324         mds_mfd_destroy(mfd);
1325
1326  cleanup:
1327         if (req != NULL && reply_body != NULL) {
1328                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1329         } else if (handle) {
1330                 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1331                 if (err) {
1332                         CERROR("error committing close: %d\n", err);
1333                         if (!rc)
1334                                 rc = err;
1335                 }
1336         }
1337
1338         switch (cleanup_phase) {
1339         case 2:
1340                 if (lcl != NULL)
1341                         ptlrpc_save_llog_lock(req, lcl);
1342                 dput(pending_child);
1343         case 1:
1344                 up(&pending_dir->i_sem);
1345         }
1346         RETURN(rc);
1347 }
1348
1349 int mds_close(struct ptlrpc_request *req)
1350 {
1351         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1352         struct obd_device *obd = req->rq_export->exp_obd;
1353         struct mds_body *body;
1354         struct mds_file_data *mfd;
1355         struct lvfs_run_ctxt saved;
1356         struct inode *inode;
1357         int rc, repsize[3] = {sizeof(struct mds_body),
1358                               obd->u.mds.mds_max_mdsize,
1359                               obd->u.mds.mds_max_cookiesize};
1360         ENTRY;
1361
1362         MDS_UPDATE_COUNTER((&obd->u.mds), MDS_CLOSE_COUNT);
1363
1364         rc = lustre_pack_reply(req, 3, repsize, NULL);
1365         if (rc) {
1366                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1367                 req->rq_status = rc;
1368         } else {
1369                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1370         }
1371
1372         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1373                 DEBUG_REQ(D_HA, req, "close replay\n");
1374                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1375                        lustre_msg_buf(req->rq_reqmsg, 1, 0),
1376                        req->rq_repmsg->buflens[2]);
1377         }
1378
1379
1380         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1381         if (body == NULL) {
1382                 CERROR("Can't unpack body\n");
1383                 req->rq_status = -EFAULT;
1384                 RETURN(-EFAULT);
1385         }
1386
1387         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1388                 /* do some stuff */ ;
1389
1390         mfd = mds_handle2mfd(&body->handle);
1391         if (mfd == NULL) {
1392                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1393                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1394                 req->rq_status = -ESTALE;
1395                 RETURN(-ESTALE);
1396         }
1397
1398         inode = mfd->mfd_dentry->d_inode;
1399         if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1400                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1401                 LASSERT(body != NULL);
1402
1403                 mds_pack_inode2fid(obd, &body->fid1, inode);
1404                 mds_pack_inode2body(obd, body, inode);
1405                 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1406         }
1407         spin_lock(&med->med_open_lock);
1408         list_del(&mfd->mfd_list);
1409         spin_unlock(&med->med_open_lock);
1410
1411         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1412         req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1413         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1414
1415         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1416                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1417                 req->rq_status = -ENOMEM;
1418                 mds_mfd_put(mfd);
1419                 RETURN(-ENOMEM);
1420         }
1421
1422         mds_mfd_put(mfd);
1423         RETURN(0);
1424 }
1425
1426 int mds_done_writing(struct ptlrpc_request *req)
1427 {
1428         struct mds_body *body;
1429         int rc, size = sizeof(struct mds_body);
1430         ENTRY;
1431
1432         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1433
1434         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1435         if (body == NULL) {
1436                 CERROR("Can't unpack body\n");
1437                 req->rq_status = -EFAULT;
1438                 RETURN(-EFAULT);
1439         }
1440
1441         rc = lustre_pack_reply(req, 1, &size, NULL);
1442         if (rc) {
1443                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1444                 req->rq_status = rc;
1445         }
1446
1447         RETURN(0);
1448 }