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                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 394                         if (lmm == NULL)
 395                                 GOTO(out_oa, rc = -ENOMEM);
 396
 397                         lmm_size = mds->mds_max_mdsize;
 398                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 399                                         lmm, &lmm_size, 1);
 400                         if (rc > 0)
 401                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 402                                                    mds->mds_osc_exp,
 403                                                    0, &lsm, lmm);
 404                         OBD_FREE(lmm, mds->mds_max_mdsize);
 405                         if (rc)
 406                                 GOTO(out_oa, rc);
 407                 }
 408                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 409                 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
 410                 if (rc) {
 411                         int level = D_ERROR;
 412                         if (rc == -ENOSPC)
 413                                 level = D_INODE;
 414                         CDEBUG(level, "error creating objects for "
 415                                       "inode %lu: rc = %d\n",
 416                                inode->i_ino, rc);
 417                         if (rc > 0) {
 418                                 CERROR("obd_create returned invalid "
 419                                        "rc %d\n", rc);
 420                                 rc = -EIO;
 421                         }
 422                         GOTO(out_oa, rc);
 423                 }
 424         } else {
 425                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 426                                    0, &lsm, rec->ur_eadata);
 427                 if (rc) {
 428                         GOTO(out_oa, rc);
 429                 }
 430                 lsm->lsm_object_id = oa->o_id;
 431                 lsm->lsm_object_gr = oa->o_gr;
 432         }
 433         if (inode->i_size) {
 434                 oa->o_size = inode->i_size;
 435                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 436                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 437                 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
 438                 if (rc) {
 439                         CERROR("error setting attrs for inode %lu: rc %d\n",
 440                                inode->i_ino, rc);
 441                         if (rc > 0) {
 442                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 443                                 rc = -EIO;
 444                         }
 445                         GOTO(out_oa, rc);
 446                 }
 447         }
 448
 449         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 450         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 451
 452         LASSERT(lsm && lsm->lsm_object_id);
 453         lmm = NULL;
 454         rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
 455         if (!rec->ur_fid2->id)
 456                 obd_free_memmd(mds->mds_osc_exp, &lsm);
 457         LASSERT(rc >= 0);
 458         lmm_size = rc;
 459         body->eadatasize = rc;
 460         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 461         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 462         lmm_bufsize = req->rq_repmsg->buflens[offset];
 463         LASSERT(lmm_buf);
 464         LASSERT(lmm_bufsize >= lmm_size);
 465
 466         memcpy(lmm_buf, lmm, lmm_size);
 467         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 468  out_oa:
 469         oti_free_cookies(&oti);
 470         obdo_free(oa);
 471  out_ids:
 472         if (rc) {
 473                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 474                 *ids = NULL;
 475         }
 476         RETURN(rc);
 477 }
 478
 479 static void reconstruct_open(struct mds_update_record *rec, int offset,
 480                              struct ptlrpc_request *req,
 481                              struct lustre_handle *child_lockh)
 482 {
 483         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 484         struct mds_client_data *mcd = med->med_mcd;
 485         struct mds_obd *mds = mds_req2mds(req);
 486         struct mds_file_data *mfd;
 487         struct obd_device *obd = req->rq_export->exp_obd;
 488         struct dentry *parent, *child;
 489         struct ldlm_reply *rep;
 490         struct mds_body *body;
 491         int rc;
 492         struct list_head *t;
 493         int put_child = 1;
 494         ENTRY;
 495
 496         LASSERT(offset == 2);                  /* only called via intent */
 497         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 498         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 499
 500         /* copy rc, transno and disp; steal locks */
 501         mds_req_from_mcd(req, mcd);
 502         intent_set_disposition(rep, mcd->mcd_last_data);
 503
 504         /* Only replay if create or open actually happened. */
 505         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 506                 EXIT;
 507                 return; /* error looking up parent or child */
 508         }
 509
 510         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 511         LASSERT(!IS_ERR(parent));
 512
 513         child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 514         LASSERT(!IS_ERR(child));
 515
 516         if (!child->d_inode) {
 517                 GOTO(out_dput, 0); /* child not present to open */
 518         }
 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          */
 523         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 524             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 525             req->rq_status) {
 526                 GOTO(out_dput, 0);
 527         }
 528
 529         /* get lock (write for O_CREAT, read otherwise) */
 530
 531         mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
 532         mds_pack_inode2body(obd, body, child->d_inode);
 533         if (S_ISREG(child->d_inode->i_mode)) {
 534                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 535                                  child->d_inode, 1);
 536
 537                 if (rc)
 538                         LASSERT(rc == req->rq_status);
 539
 540                 /* If we have LOV EA data, the OST holds size, mtime */
 541                 if (!(body->valid & OBD_MD_FLEASIZE))
 542                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 543                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 544         } else {
 545                 /* XXX need to check this case */
 546         }
 547
 548         /* If we're opening a file without an EA, change to a write
 549            lock (unless we already have one). */
 550
 551         /* If we have -EEXIST as the status, and we were asked to create
 552          * exclusively, we can tell we failed because the file already existed.
 553          */
 554         if (req->rq_status == -EEXIST &&
 555             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 556              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 557                 GOTO(out_dput, 0);
 558         }
 559
 560         /* If we didn't get as far as trying to open, then some locking thing
 561          * probably went wrong, and we'll just bail here.
 562          */
 563         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 564                 GOTO(out_dput, 0);
 565
 566         /* If we failed, then we must have failed opening, so don't look for
 567          * file descriptor or anything, just give the client the bad news.
 568          */
 569         if (req->rq_status)
 570                 GOTO(out_dput, 0);
 571
 572         mfd = NULL;
 573         list_for_each(t, &med->med_open_head) {
 574                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 575                 if (mfd->mfd_xid == req->rq_xid)
 576                         break;
 577                 mfd = NULL;
 578         }
 579
 580         /* #warning "XXX fixme" bug 2991 */
 581         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 582          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 583          * to detect a re-open */
 584         if (mfd == NULL) {
 585                 mntget(mds->mds_vfsmnt);
 586                 CERROR("Re-opened file \n");
 587                 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
 588                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 589                 if (!mfd) {
 590                         CERROR("mds: out of memory\n");
 591                         GOTO(out_dput, req->rq_status = -ENOMEM);
 592                 }
 593                 put_child = 0;
 594         } else {
 595                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 596                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 597                        mfd->mfd_handle.h_cookie);
 598         }
 599
 600  out_dput:
 601         if (put_child)
 602                 l_dput(child);
 603         l_dput(parent);
 604         EXIT;
 605 }
 606
 607 /* do NOT or the MAY_*'s, you'll get the weakest */
 608 static int accmode(int flags)
 609 {
 610         int res = 0;
 611
 612         if (flags & FMODE_READ)
 613                 res = MAY_READ;
 614         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 615                 res |= MAY_WRITE;
 616         if (flags & FMODE_EXEC)
 617                 res = MAY_EXEC;
 618         return res;
 619 }
 620
 621 /* Handles object creation, actual opening, and I/O epoch */
 622 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 623                            struct mds_body *body, int flags, void **handle,
 624                            struct mds_update_record *rec,struct ldlm_reply *rep)
 625 {
 626         struct mds_obd *mds = mds_req2mds(req);
 627         struct obd_device *obd = req->rq_export->exp_obd;
 628         struct mds_file_data *mfd = NULL;
 629         obd_id *ids = NULL; /* object IDs created */
 630         unsigned mode;
 631         int rc = 0;
 632         ENTRY;
 633
 634         /* atomically create objects if necessary */
 635         down(&dchild->d_inode->i_sem);
 636         mode = dchild->d_inode->i_mode;
 637         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 638             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 639                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 640                                  dchild->d_inode, 0);
 641                 if (rc) {
 642                         up(&dchild->d_inode->i_sem);
 643                         RETURN(rc);
 644                 }
 645         }
 646         if (rec != NULL) {
 647                 /* no EA: create objects */
 648                 rc = mds_create_objects(req, 2, rec, mds, obd,
 649                                         dchild, handle, &ids);
 650                 if (rc) {
 651                         CERROR("mds_create_objects: rc = %d\n", rc);
 652                         up(&dchild->d_inode->i_sem);
 653                         RETURN(rc);
 654                 }
 655                 if (S_ISREG(dchild->d_inode->i_mode) &&
 656                     (body->valid & OBD_MD_FLEASIZE)) {
 657                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 658                                                    req->rq_repmsg, 2);
 659                         if (!rc)
 660                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 661                                                  dchild->d_inode, 0);
 662                         if (rc) {
 663                                 up(&dchild->d_inode->i_sem);
 664                                 RETURN(rc);
 665                         }
 666                 }
 667         }
 668         /* If the inode has no EA data, then MDSs hold size, mtime */
 669         if (S_ISREG(dchild->d_inode->i_mode) &&
 670             !(body->valid & OBD_MD_FLEASIZE)) {
 671                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 672                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 673         }
 674         up(&dchild->d_inode->i_sem);
 675
 676         intent_set_disposition(rep, DISP_OPEN_OPEN);
 677         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 678         if (IS_ERR(mfd))
 679                 RETURN(PTR_ERR(mfd));
 680
 681         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 682                mfd->mfd_handle.h_cookie);
 683         if (ids != NULL) {
 684                 mds_lov_update_objids(obd, ids);
 685                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 686         }
 687         //if (rc)
 688         //        mds_mfd_destroy(mfd);
 689         RETURN(rc);
 690 }
 691
 692 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 693                            struct mds_body *body, int flags,
 694                            struct mds_update_record *rec,struct ldlm_reply *rep)
 695 {
 696         struct mds_obd *mds = mds_req2mds(req);
 697         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 698         struct dentry *dchild;
 699         char fidname[LL_FID_NAMELEN];
 700         int fidlen = 0, rc;
 701         void *handle = NULL;
 702         ENTRY;
 703
 704         down(&pending_dir->i_sem);
 705         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 706         dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 707         if (IS_ERR(dchild)) {
 708                 up(&pending_dir->i_sem);
 709                 rc = PTR_ERR(dchild);
 710                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 711                 RETURN(rc);
 712         }
 713
 714         if (dchild->d_inode != NULL) {
 715                 up(&pending_dir->i_sem);
 716                 mds_inode_set_orphan(dchild->d_inode);
 717                 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 718                 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 719                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 720                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 721                 CWARN("Orphan %s found and opened in PENDING directory\n",
 722                        fidname);
 723                 goto open;
 724         }
 725         dput(dchild);
 726         up(&pending_dir->i_sem);
 727
 728         /* We didn't find it in PENDING so it isn't an orphan.  See
 729          * if it was a regular inode that was previously created. */
 730         dchild = mds_fid2dentry(mds, fid, NULL);
 731         if (IS_ERR(dchild))
 732                 RETURN(PTR_ERR(dchild));
 733
 734         mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
 735         mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
 736         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 737         intent_set_disposition(rep, DISP_LOOKUP_POS);
 738
 739  open:
 740         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 741         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 742                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 743         /* XXX what do we do here if mds_finish_transno itself failed? */
 744
 745         l_dput(dchild);
 746         RETURN(rc);
 747 }
 748
 749 int mds_pin(struct ptlrpc_request *req)
 750 {
 751         struct obd_device *obd = req->rq_export->exp_obd;
 752         struct mds_body *request_body, *reply_body;
 753         struct lvfs_run_ctxt saved;
 754         int rc, size = sizeof(*reply_body);
 755         ENTRY;
 756
 757         request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
 758
 759         rc = lustre_pack_reply(req, 1, &size, NULL);
 760         if (rc)
 761                 RETURN(rc);
 762         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 763
 764         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 765         rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
 766                              request_body->flags, NULL, NULL);
 767         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 768
 769         RETURN(rc);
 770 }
 771
 772 /*  Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
 773  *  If child_lockh is NULL we just get the lock as a barrier to wait for
 774  *  other holders of this lock, and drop it right away again. */
 775 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 776                        struct lustre_handle *child_lockh)
 777 {
 778         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 779         struct lustre_handle lockh;
 780         int lock_flags = 0;
 781         int rc;
 782
 783         if (child_lockh == NULL)
 784                 child_lockh = &lockh;
 785
 786         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 787                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 788                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 789                               NULL, 0, NULL, child_lockh);
 790         if (rc != ELDLM_OK)
 791                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 792         else if (child_lockh == &lockh)
 793                 ldlm_lock_decref(child_lockh, LCK_EX);
 794
 795         RETURN(rc);
 796 }
 797
 798 static int is_mount_object(struct dentry *dparent)
 799 {
 800         struct dentry *dchild;
 801
 802         if (!(dparent->d_inode->i_mode & S_ISUID))
 803                 return 0;
 804
 805         dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
 806         if (IS_ERR(dchild) || dchild == NULL)
 807                 return 0;
 808
 809         dput(dchild);
 810         return 1;
 811 }
 812
 813 int mds_open(struct mds_update_record *rec, int offset,
 814              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 815 {
 816         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 817         struct obd_device *obd = req->rq_export->exp_obd;
 818         struct mds_obd *mds = mds_req2mds(req);
 819         struct ldlm_reply *rep = NULL;
 820         struct mds_body *body = NULL;
 821         struct dentry *dchild = NULL, *dparent = NULL;
 822         struct mds_export_data *med;
 823         struct lustre_handle parent_lockh[2];
 824         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 825         int parent_mode = LCK_PR;
 826         void *handle = NULL;
 827         struct dentry_params dp;
 828         struct mea *mea = NULL;
 829         int mea_size, update_mode;
 830         ENTRY;
 831
 832         parent_lockh[0].cookie = 0;
 833         parent_lockh[1].cookie = 0;
 834
 835         if (offset == 2) { /* intent */
 836                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 837                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 838         } else if (offset == 0) { /* non-intent reint */
 839                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 840         } else {
 841                 body = NULL;
 842                 LBUG();
 843         }
 844
 845         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 846
 847         /* Step 0: If we are passed a fid, then we assume the client already
 848          * opened this file and is only replaying the RPC, so we open the
 849          * inode by fid (at some large expense in security). */
 850         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 851                 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
 852                           rep->lock_policy_res1);
 853
 854                 LASSERT(rec->ur_fid2->id);
 855
 856                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 857                                      rec, rep);
 858                 if (rc != -ENOENT)
 859                         RETURN(rc);
 860                 /* We didn't find the correct inode on disk either, so we
 861                  * need to re-create it via a regular replay. */
 862                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 863         } else {
 864                 LASSERT(!rec->ur_fid2->id);
 865         }
 866
 867         LASSERT(offset == 2); /* If we got here, we must be called via intent */
 868
 869         med = &req->rq_export->exp_mds_data;
 870         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 871                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 872                 RETURN(-ENOMEM);
 873         }
 874
 875         acc_mode = accmode(rec->ur_flags);
 876
 877         /* Step 1: Find and lock the parent */
 878         if (rec->ur_flags & O_CREAT) {
 879                 /* XXX Well, in fact we only need this lock mode change if
 880                    in addition to O_CREAT, the file does not exist.
 881                    But we do not know if it exists or not yet */
 882                 parent_mode = LCK_PW;
 883         }
 884
 885         if (rec->ur_namelen == 1) {
 886                 /* client (LMV) wants to open the file by fid */
 887                 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
 888                                 (unsigned) rec->ur_fid1->mds,
 889                                 (unsigned) rec->ur_fid1->id,
 890                                 (unsigned) rec->ur_fid1->generation);
 891                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 892                 if (IS_ERR(dchild)) {
 893                         rc = PTR_ERR(dparent);
 894                         CERROR("child lookup by a fid error %d\n", rc);
 895                         GOTO(cleanup, rc);
 896                 }
 897                 goto got_child;
 898         }
 899
 900         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 901                                         parent_lockh, &update_mode, rec->ur_name,
 902                                         rec->ur_namelen - 1,
 903                                         MDS_INODELOCK_UPDATE);
 904         if (IS_ERR(dparent)) {
 905                 rc = PTR_ERR(dparent);
 906                 CERROR("parent lookup error %d\n", rc);
 907                 GOTO(cleanup, rc);
 908         }
 909         LASSERT(dparent->d_inode != NULL);
 910
 911         cleanup_phase = 1; /* parent dentry and lock */
 912
 913         /* try to retrieve MEA data for this dir */
 914         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 915         if (rc)
 916                 GOTO(cleanup, rc);
 917
 918         if (mea != NULL) {
 919                 /* dir is already splitted, check is requested filename
 920                  * should live at this MDS or at another one */
 921                 int i;
 922                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 923                 if (mea->mea_master != i) {
 924                         CERROR("inapropriate MDS(%d) for %s. should be %d\n",
 925                                 mea->mea_master, rec->ur_name, i);
 926                         GOTO(cleanup, rc = -ERESTART);
 927                 }
 928         }
 929
 930         /* Step 2: Lookup the child */
 931         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 932         if (IS_ERR(dchild)) {
 933                 rc = PTR_ERR(dchild);
 934                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 935                 GOTO(cleanup, rc);
 936         }
 937
 938 got_child:
 939         cleanup_phase = 2; /* child dentry */
 940
 941         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 942
 943         if (dchild->d_flags & DCACHE_CROSS_REF) {
 944                 struct ldlm_res_id res_id = { . name = {0} };
 945                 ldlm_policy_data_t policy;
 946                 int flags = 0;
 947                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 948                        (unsigned long) dchild->d_mdsnum,
 949                        (unsigned long) dchild->d_inum,
 950                        (unsigned long) dchild->d_generation);
 951                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 952                 body->fid1.id = dchild->d_inum;
 953                 body->fid1.mds = dchild->d_mdsnum;
 954                 body->fid1.generation = dchild->d_generation;
 955                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 956                 res_id.name[0] = dchild->d_inum;
 957                 res_id.name[1] = dchild->d_generation;
 958                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 959                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 960                                       res_id, LDLM_IBITS, &policy,
 961                                       LCK_PR, &flags,
 962                                       mds_blocking_ast,
 963                                       ldlm_completion_ast, NULL, NULL,
 964                                       NULL, 0, NULL, child_lockh);
 965 #ifdef S_PDIROPS
 966                 if (parent_lockh[1].cookie != 0)
 967                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 968 #endif
 969                 ldlm_lock_decref(parent_lockh, parent_mode);
 970                 if (mea)
 971                         OBD_FREE(mea, mea_size);
 972                 l_dput(dchild);
 973                 l_dput(dparent);
 974                 RETURN(rc);
 975         }
 976
 977         if (dchild->d_inode)
 978                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 979         else
 980                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 981
 982         /*Step 3: If the child was negative, and we're supposed to, create it.*/
 983         if (dchild->d_inode == NULL) {
 984                 unsigned long ino = rec->ur_fid2->id;
 985                 struct iattr iattr;
 986                 struct inode *inode;
 987
 988                 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
 989                         if (rc > 0) {
 990                                 /* dir got splitted */
 991                                 GOTO(cleanup, rc = -ERESTART);
 992                         } else {
 993                                 /* error happened during spitting */
 994                                 GOTO(cleanup, rc);
 995                         }
 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
1056                 acc_mode = 0;           /* Don't check for permissions */
1057         }
1058
1059         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1060
1061         mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1062         mds_pack_inode2body(obd, body, dchild->d_inode);
1063
1064         if (S_ISREG(dchild->d_inode->i_mode)) {
1065                 /* Check permissions etc */
1066                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1067                 if (rc != 0)
1068                         GOTO(cleanup, rc);
1069
1070                 /* Can't write to a read-only file */
1071                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1072                         GOTO(cleanup, rc = -EPERM);
1073
1074                 /* An append-only file must be opened in append mode for
1075                  * writing */
1076                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1077                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1078                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1079                         GOTO(cleanup, rc = -EPERM);
1080         }
1081
1082         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1083             (rec->ur_flags & MDS_OPEN_EXCL)) {
1084                 /* File already exists, we didn't just create it, and we
1085                  * were passed O_EXCL; err-or. */
1086                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1087         }
1088
1089         /* if we are following a symlink, don't open */
1090         if (S_ISLNK(dchild->d_inode->i_mode))
1091                 GOTO(cleanup, rc = 0);
1092
1093         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1094             !S_ISDIR(dchild->d_inode->i_mode))
1095                 GOTO(cleanup, rc = -ENOTDIR);
1096
1097         if (S_ISDIR(dchild->d_inode->i_mode)) {
1098                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1099                     rec->ur_flags & FMODE_WRITE) {
1100                         /*we are trying to create or write a exist dir*/
1101                         GOTO(cleanup, rc = -EISDIR);
1102                 }
1103                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1104                         GOTO(cleanup, rc = -EACCES);
1105                 }
1106                 if (is_mount_object(dchild)) {
1107                         CERROR("Found possible GNS mount object %*s; not "
1108                                "opening.\n", dchild->d_name.len,
1109                                dchild->d_name.name);
1110                         GOTO(cleanup, rc = 0); // success, but don't really open
1111                 }
1112         }
1113
1114         if (rc == 0) {
1115                 struct ldlm_res_id res_id = { . name = {0} };
1116                 ldlm_policy_data_t policy;
1117                 int flags = 0;
1118                 res_id.name[0] = dchild->d_inode->i_ino;
1119                 res_id.name[1] = dchild->d_inode->i_generation;
1120                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1121                                                 MDS_INODELOCK_UPDATE;
1122                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1123                                       res_id, LDLM_IBITS, &policy,
1124                                       LCK_PR, &flags,
1125                                       mds_blocking_ast,
1126                                       ldlm_completion_ast, NULL, NULL,
1127                                       NULL, 0, NULL, child_lockh);
1128                 if (rc == 0)
1129                         cleanup_phase = 3;
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 3:
1143                 if (rc)
1144                         ldlm_lock_decref(child_lockh, LCK_PR);
1145         case 2:
1146                 if (rc && created) {
1147                         int err = vfs_unlink(dparent->d_inode, dchild);
1148                         if (err) {
1149                                 CERROR("unlink(%*s) in error path: %d\n",
1150                                        dchild->d_name.len, dchild->d_name.name,
1151                                        err);
1152                         }
1153                 } else if (created) {
1154                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1155                 }
1156                 l_dput(dchild);
1157         case 1:
1158                 if (dparent == NULL)
1159                         break;
1160
1161                 l_dput(dparent);
1162 #ifdef S_PDIROPS
1163                 if (parent_lockh[1].cookie != 0)
1164                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1165 #endif
1166                 if (rc)
1167                         ldlm_lock_decref(parent_lockh, parent_mode);
1168                 else
1169                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1170         }
1171         if (rc == 0)
1172                 atomic_inc(&mds->mds_open_count);
1173         if (mea)
1174                 OBD_FREE(mea, mea_size);
1175         if ((cleanup_phase != 3) && !rc)
1176                 rc = ENOLCK;
1177         RETURN(rc);
1178 }
1179
1180 /* Close a "file descriptor" and possibly unlink an orphan from the
1181  * PENDING directory.
1182  *
1183  * If we are being called from mds_disconnect() because the client has
1184  * disappeared, then req == NULL and we do not update last_rcvd because
1185  * there is nothing that could be recovered by the client at this stage
1186  * (it will not even _have_ an entry in last_rcvd anymore).
1187  *
1188  * Returns EAGAIN if the client needs to get more data and re-close. */
1189 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1190                   struct mds_file_data *mfd, int unlink_orphan)
1191 {
1192         struct inode *inode = mfd->mfd_dentry->d_inode;
1193         char fidname[LL_FID_NAMELEN];
1194         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1195         struct dentry *pending_child = NULL;
1196         struct mds_obd *mds = &obd->u.mds;
1197         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1198         void *handle = NULL;
1199         struct mds_body *request_body = NULL, *reply_body = NULL;
1200         struct dentry_params dp;
1201         struct iattr iattr = { 0 };
1202         ENTRY;
1203
1204         if (req && req->rq_reqmsg != NULL)
1205                 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1206                                               sizeof(*request_body));
1207         if (req && req->rq_repmsg != NULL)
1208                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1209                                             sizeof(*reply_body));
1210
1211         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1212
1213         last_orphan = mds_open_orphan_dec_test(inode) &&
1214                 mds_inode_is_orphan(inode);
1215
1216         /* this is half of the actual "close" */
1217         if (mfd->mfd_mode & FMODE_WRITE) {
1218                 rc = mds_put_write_access(mds, inode, request_body,
1219                                           last_orphan && unlink_orphan);
1220         } else if (mfd->mfd_mode & FMODE_EXEC) {
1221                 mds_allow_write_access(inode);
1222         }
1223
1224         if (last_orphan && unlink_orphan) {
1225                 struct lov_mds_md *lmm = NULL;
1226                 int stripe_count = 0;
1227                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1228
1229                 if (obd->obd_recovering) {
1230                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1231                                " is over\n", fidname);
1232                         GOTO(out, rc);
1233                 }
1234
1235                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1236
1237                 /* Sadly, there is no easy way to save pending_child from
1238                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1239                  * but normally it will still be in the dcache. */
1240                 pending_dir = mds->mds_pending_dir->d_inode;
1241                 down(&pending_dir->i_sem);
1242                 cleanup_phase = 1; /* up(i_sem) when finished */
1243                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1244                                                fidlen);
1245                 if (IS_ERR(pending_child))
1246                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1247                 LASSERT(pending_child->d_inode != NULL);
1248
1249                 cleanup_phase = 2; /* dput(pending_child) when finished */
1250                 if (req != NULL && req->rq_repmsg != NULL) {
1251                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1252                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1253                 }
1254
1255                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1256                                           NULL, stripe_count);
1257                 if (IS_ERR(handle)) {
1258                         rc = PTR_ERR(handle);
1259                         handle = NULL;
1260                         GOTO(cleanup, rc);
1261                 }
1262
1263                 if (req != NULL && req->rq_repmsg != NULL &&
1264                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1265                     mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1266                                       req->rq_repmsg->buflens[1],
1267                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1268                                       req->rq_repmsg->buflens[2]) > 0) {
1269                         reply_body->valid |= OBD_MD_FLCOOKIE;
1270                 }
1271
1272                 pending_child->d_fsdata = (void *) &dp;
1273                 dp.p_inum = 0;
1274                 dp.p_ptr = req;
1275                 if (S_ISDIR(pending_child->d_inode->i_mode))
1276                         rc = vfs_rmdir(pending_dir, pending_child);
1277                 else
1278                         rc = vfs_unlink(pending_dir, pending_child);
1279                 if (rc)
1280                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1281
1282                 goto out; /* Don't bother updating attrs on unlinked inode */
1283         }
1284
1285         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1286                 /* Update the on-disk attributes if this was the last write
1287                  * close, and all information was provided (i.e., rc == 0)
1288                  *
1289                  * XXX this should probably be abstracted with mds_reint_setattr
1290                  */
1291
1292 #if 0
1293                 if (request_body->valid & OBD_MD_FLMTIME &&
1294                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1295                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1296                         iattr.ia_valid |= ATTR_MTIME;
1297                 }
1298                 if (request_body->valid & OBD_MD_FLCTIME &&
1299                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1300                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1301                         iattr.ia_valid |= ATTR_CTIME;
1302                 }
1303
1304                 /* XXX can't set block count with fsfilt_setattr (!) */
1305                 if (request_body->valid & OBD_MD_FLSIZE) {
1306                         iattr.ia_valid |= ATTR_SIZE;
1307                         iattr.ia_size = request_body->size;
1308                 }
1309                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1310                         iattr.ia_valid |= ATTR_BLOCKS;
1311                         iattr.ia_blocks = request_body->blocks
1312                 } */
1313
1314 #endif
1315         }
1316
1317         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1318                 /* Only start a transaction to write out only the atime if
1319                  * it is more out-of-date than the specified limit.  If we
1320                  * are already going to write out the atime then do it anyway.
1321                  * */
1322                 if ((LTIME_S(request_body->atime) >
1323                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1324                     (iattr.ia_valid != 0 &&
1325                      LTIME_S(request_body->atime) > LTIME_S(inode->i_atime))) {
1326                         LTIME_S(iattr.ia_atime) = LTIME_S(request_body->atime);
1327                         iattr.ia_valid |= ATTR_ATIME;
1328                 }
1329         }
1330
1331         if (iattr.ia_valid != 0) {
1332                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1333                 if (IS_ERR(handle))
1334                         GOTO(cleanup, rc = PTR_ERR(handle));
1335                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1336                 if (rc)
1337                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1338         }
1339 out:
1340         /* If other clients have this file open for write, rc will be > 0 */
1341         if (rc > 0)
1342                 rc = 0;
1343         l_dput(mfd->mfd_dentry);
1344         mds_mfd_destroy(mfd);
1345
1346  cleanup:
1347         atomic_dec(&mds->mds_open_count);
1348         if (req != NULL && reply_body != NULL) {
1349                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1350         } else if (handle) {
1351                 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1352                 if (err) {
1353                         CERROR("error committing close: %d\n", err);
1354                         if (!rc)
1355                                 rc = err;
1356                 }
1357         }
1358
1359         switch (cleanup_phase) {
1360         case 2:
1361                 dput(pending_child);
1362         case 1:
1363                 up(&pending_dir->i_sem);
1364         }
1365         RETURN(rc);
1366 }
1367
1368 int mds_close(struct ptlrpc_request *req)
1369 {
1370         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1371         struct obd_device *obd = req->rq_export->exp_obd;
1372         struct mds_body *body;
1373         struct mds_file_data *mfd;
1374         struct lvfs_run_ctxt saved;
1375         struct inode *inode;
1376         int rc, repsize[3] = {sizeof(struct mds_body),
1377                               obd->u.mds.mds_max_mdsize,
1378                               obd->u.mds.mds_max_cookiesize};
1379         ENTRY;
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         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1390         if (body == NULL) {
1391                 CERROR("Can't unpack body\n");
1392                 req->rq_status = -EFAULT;
1393                 RETURN(-EFAULT);
1394         }
1395
1396         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1397                 /* do some stuff */ ;
1398
1399         mfd = mds_handle2mfd(&body->handle);
1400         if (mfd == NULL) {
1401                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1402                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1403                 req->rq_status = -ESTALE;
1404                 RETURN(-ESTALE);
1405         }
1406
1407         inode = mfd->mfd_dentry->d_inode;
1408         if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1409                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1410                 LASSERT(body != NULL);
1411
1412                 mds_pack_inode2fid(obd, &body->fid1, inode);
1413                 mds_pack_inode2body(obd, body, inode);
1414                 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1415         }
1416         spin_lock(&med->med_open_lock);
1417         list_del(&mfd->mfd_list);
1418         spin_unlock(&med->med_open_lock);
1419
1420         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1421         req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1422         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1423
1424         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1425                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1426                 req->rq_status = -ENOMEM;
1427                 mds_mfd_put(mfd);
1428                 RETURN(-ENOMEM);
1429         }
1430
1431         mds_mfd_put(mfd);
1432         RETURN(0);
1433 }
1434
1435 int mds_done_writing(struct ptlrpc_request *req)
1436 {
1437         struct mds_body *body;
1438         int rc, size = sizeof(struct mds_body);
1439         ENTRY;
1440
1441         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1442
1443         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1444         if (body == NULL) {
1445                 CERROR("Can't unpack body\n");
1446                 req->rq_status = -EFAULT;
1447                 RETURN(-EFAULT);
1448         }
1449
1450         rc = lustre_pack_reply(req, 1, &size, NULL);
1451         if (rc) {
1452                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1453                 req->rq_status = rc;
1454         }
1455
1456         RETURN(0);
1457 }