lustre/mds/mds_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *   Author: Phil Schwan <phil@clusterfs.com>
   8  *   Author: Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/version.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/version.h>
  38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  39 # include <linux/buffer_head.h>
  40 # include <linux/workqueue.h>
  41 #else
  42 # include <linux/locks.h>
  43 #endif
  44
  45 #include <obd_class.h>
  46 #include <obd_lov.h>
  47 #include <lustre_fsfilt.h>
  48 #include <lprocfs_status.h>
  49
  50 #include "mds_internal.h"
  51
  52 /* Exported function from this file are:
  53  *
  54  * mds_open - called by the intent handler
  55  * mds_close - an rpc handling function
  56  * mds_pin - an rpc handling function - which will go away
  57  * mds_mfd_close - for force closing files when a client dies
  58  */
  59
  60 /*
  61  * MDS file data handling: file data holds a handle for a file opened
  62  * by a client.
  63  */
  64
  65 static void mds_mfd_addref(void *mfdp)
  66 {
  67         struct mds_file_data *mfd = mfdp;
  68
  69         atomic_inc(&mfd->mfd_refcount);
  70         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  71                atomic_read(&mfd->mfd_refcount));
  72 }
  73
  74 /* Create a new mds_file_data struct.
  75  * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
  76  * one reference for the caller of this function. */
  77 struct mds_file_data *mds_mfd_new(void)
  78 {
  79         struct mds_file_data *mfd;
  80
  81         OBD_ALLOC(mfd, sizeof *mfd);
  82         if (mfd == NULL) {
  83                 CERROR("mds: out of memory\n");
  84                 return NULL;
  85         }
  86
  87         atomic_set(&mfd->mfd_refcount, 2);
  88
  89         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  90         INIT_LIST_HEAD(&mfd->mfd_list);
  91         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  92
  93         return mfd;
  94 }
  95
  96 /* Get a new reference on the mfd pointed to by handle, if handle is still
  97  * valid.  Caller must drop reference with mds_mfd_put(). */
  98 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  99 {
 100         ENTRY;
 101         LASSERT(handle != NULL);
 102         RETURN(class_handle2object(handle->cookie));
 103 }
 104
 105 /* Drop mfd reference, freeing struct if this is the last one. */
 106 static void mds_mfd_put(struct mds_file_data *mfd)
 107 {
 108         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
 109                atomic_read(&mfd->mfd_refcount) - 1);
 110         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 111                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 112         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 113                 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 114         }
 115 }
 116
 117 /* Remove the mfd handle so that it cannot be found by open/close again.
 118  * Caller must hold med_open_lock for mfd_list manipulation. */
 119 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
 120 {
 121         class_handle_unhash(&mfd->mfd_handle);
 122         list_del_init(&mfd->mfd_list);
 123         if (decref)
 124                 mds_mfd_put(mfd);
 125 }
 126
 127 /* Caller must hold mds->mds_epoch_sem */
 128 static int mds_alloc_filterdata(struct inode *inode)
 129 {
 130         LASSERT(inode->i_filterdata == NULL);
 131         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 132         if (inode->i_filterdata == NULL)
 133                 return -ENOMEM;
 134         LASSERT(igrab(inode) == inode);
 135         return 0;
 136 }
 137
 138 /* Caller must hold mds->mds_epoch_sem */
 139 static void mds_free_filterdata(struct inode *inode)
 140 {
 141         LASSERT(inode->i_filterdata != NULL);
 142         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 143         inode->i_filterdata = NULL;
 144         iput(inode);
 145 }
 146
 147 /* Write access to a file: executors cause a negative count,
 148  * writers a positive count.  The semaphore is needed to perform
 149  * a check for the sign and then increment or decrement atomically.
 150  *
 151  * This code is closely tied to the allocation of the d_fsdata and the
 152  * MDS epoch, so we use the same semaphore for the whole lot.
 153  *
 154  * We could use a different semaphore for each file, if it ever shows
 155  * up in a profile, which it won't.
 156  *
 157  * epoch argument is nonzero during recovery */
 158 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 159                                 __u64 epoch)
 160 {
 161         int rc = 0;
 162
 163         down(&mds->mds_epoch_sem);
 164
 165         if (atomic_read(&inode->i_writecount) < 0) {
 166                 up(&mds->mds_epoch_sem);
 167                 RETURN(-ETXTBSY);
 168         }
 169
 170
 171         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 172                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 173                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 174                        inode->i_generation);
 175                 goto out;
 176         }
 177
 178         if (inode->i_filterdata == NULL)
 179                 mds_alloc_filterdata(inode);
 180         if (inode->i_filterdata == NULL) {
 181                 rc = -ENOMEM;
 182                 goto out;
 183         }
 184         if (epoch > mds->mds_io_epoch)
 185                 mds->mds_io_epoch = epoch;
 186         else
 187                 mds->mds_io_epoch++;
 188         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 189         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 190                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 191  out:
 192         if (rc == 0)
 193                 atomic_inc(&inode->i_writecount);
 194         up(&mds->mds_epoch_sem);
 195         return rc;
 196 }
 197
 198 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 199  * again -- which is never true if the file is about to be unlinked.  Otherwise
 200  * returns the number of remaining writers. */
 201 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 202                                 struct mds_body *body, int unlinking)
 203 {
 204         int rc = 0;
 205         ENTRY;
 206
 207         down(&mds->mds_epoch_sem);
 208         atomic_dec(&inode->i_writecount);
 209         rc = atomic_read(&inode->i_writecount);
 210         if (rc > 0)
 211                 GOTO(out, rc);
 212 #if 0
 213         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 214                 GOTO(out, rc = EAGAIN);
 215 #endif
 216         mds_free_filterdata(inode);
 217  out:
 218         up(&mds->mds_epoch_sem);
 219         return rc;
 220 }
 221
 222 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 223 {
 224         ENTRY;
 225         down(&mds->mds_epoch_sem);
 226         if (atomic_read(&inode->i_writecount) > 0) {
 227                 up(&mds->mds_epoch_sem);
 228                 RETURN(-ETXTBSY);
 229         }
 230         atomic_dec(&inode->i_writecount);
 231         up(&mds->mds_epoch_sem);
 232         RETURN(0);
 233 }
 234
 235 static void mds_allow_write_access(struct inode *inode)
 236 {
 237         ENTRY;
 238         atomic_inc(&inode->i_writecount);
 239 }
 240
 241 int mds_query_write_access(struct inode *inode)
 242 {
 243         ENTRY;
 244         RETURN(atomic_read(&inode->i_writecount));
 245 }
 246
 247 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 248 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 249                                              struct vfsmount *mnt, int flags,
 250                                              struct ptlrpc_request *req)
 251 {
 252         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 253         struct mds_obd *mds = mds_req2mds(req);
 254         struct mds_file_data *mfd;
 255         struct mds_body *body;
 256         int error;
 257         ENTRY;
 258
 259         mfd = mds_mfd_new();
 260         if (mfd == NULL) {
 261                 CERROR("mds: out of memory\n");
 262                 GOTO(cleanup_dentry, error = -ENOMEM);
 263         }
 264
 265         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 266
 267         if (flags & FMODE_WRITE) {
 268                 /* FIXME: in recovery, need to pass old epoch here */
 269                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 270                 if (error)
 271                         GOTO(cleanup_mfd, error);
 272                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 273         } else if (flags & MDS_FMODE_EXEC) {
 274                 error = mds_deny_write_access(mds, dentry->d_inode);
 275                 if (error)
 276                         GOTO(cleanup_mfd, error);
 277         }
 278
 279         dget(dentry);
 280
 281         /* Mark the file as open to handle open-unlink. */
 282         MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
 283         mds_orphan_open_inc(dentry->d_inode);
 284         MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
 285
 286         mfd->mfd_mode = flags;
 287         mfd->mfd_dentry = dentry;
 288         mfd->mfd_xid = req->rq_xid;
 289
 290         spin_lock(&med->med_open_lock);
 291         list_add(&mfd->mfd_list, &med->med_open_head);
 292         spin_unlock(&med->med_open_lock);
 293
 294         body->handle.cookie = mfd->mfd_handle.h_cookie;
 295
 296         RETURN(mfd);
 297
 298 cleanup_mfd:
 299         mds_mfd_put(mfd);
 300         mds_mfd_unlink(mfd, 1);
 301 cleanup_dentry:
 302         return ERR_PTR(error);
 303 }
 304
 305 /* Must be called with i_mutex held */
 306 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 307                               struct mds_update_record *rec,
 308                               struct mds_obd *mds, struct obd_device *obd,
 309                               struct dentry *dchild, void **handle,
 310                               obd_id **ids)
 311 {
 312         struct inode *inode = dchild->d_inode;
 313         struct obd_trans_info oti = { 0 };
 314         struct lov_mds_md *lmm = NULL;
 315         int rc, lmm_size;
 316         struct mds_body *body;
 317         struct obd_info oinfo = { { { 0 } } };
 318         void *lmm_buf;
 319         ENTRY;
 320
 321         if (!S_ISREG(inode->i_mode))
 322                 RETURN(0);
 323         if (!md_should_create(rec->ur_flags))
 324                 RETURN(0);
 325
 326         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 327
 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_init(&oti, req);
 335         oti.oti_objid = *ids;
 336
 337         /* replay case */
 338         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 339                 if (rec->ur_fid2->id == 0) {
 340                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 341                         RETURN(-EFAULT);
 342                 }
 343
 344                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 345                 lmm_size = rec->ur_eadatalen;
 346                 lmm = rec->ur_eadata;
 347                 LASSERT(lmm);
 348
 349                 if (*handle == NULL)
 350                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 351                 if (IS_ERR(*handle)) {
 352                         rc = PTR_ERR(*handle);
 353                         *handle = NULL;
 354                         GOTO(out_ids, rc);
 355                 }
 356
 357                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 358
 359                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 360                 if (rc)
 361                         CERROR("open replay failed to set md:%d\n", rc);
 362                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 363                 LASSERT(lmm_buf);
 364                 memcpy(lmm_buf, lmm, lmm_size);
 365
 366                 RETURN(rc);
 367         }
 368
 369         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 370                 GOTO(out_ids, rc = -ENOMEM);
 371
 372         OBDO_ALLOC(oinfo.oi_oa);
 373         if (oinfo.oi_oa == NULL)
 374                 GOTO(out_ids, rc = -ENOMEM);
 375         oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
 376         oinfo.oi_oa->o_gid = 0;
 377         oinfo.oi_oa->o_mode = S_IFREG | 0600;
 378         oinfo.oi_oa->o_id = inode->i_ino;
 379         oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
 380                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
 381         oinfo.oi_oa->o_size = 0;
 382
 383         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 384                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 385
 386         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 387                 /* check if things like lfs setstripe are sending us the ea */
 388                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 389                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 390                                            mds->mds_osc_exp,
 391                                            0, &oinfo.oi_md, rec->ur_eadata);
 392                         if (rc)
 393                                 GOTO(out_oa, rc);
 394                 } else {
 395                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 396                         if (lmm == NULL)
 397                                 GOTO(out_oa, rc = -ENOMEM);
 398
 399                         lmm_size = mds->mds_max_mdsize;
 400                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 401                                         lmm, &lmm_size, 1);
 402                         if (rc > 0)
 403                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 404                                                    mds->mds_osc_exp,
 405                                                    0, &oinfo.oi_md, lmm);
 406                         OBD_FREE(lmm, mds->mds_max_mdsize);
 407                         if (rc)
 408                                 GOTO(out_oa, rc);
 409                 }
 410                 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
 411                                 &oinfo.oi_md, &oti);
 412                 if (rc) {
 413                         int level = D_ERROR;
 414                         if (rc == -ENOSPC)
 415                                 level = D_INODE;
 416                         CDEBUG(level, "error creating objects for "
 417                                       "inode %lu: rc = %d\n",
 418                                inode->i_ino, rc);
 419                         if (rc > 0) {
 420                                 CERROR("obd_create returned invalid "
 421                                        "rc %d\n", rc);
 422                                 rc = -EIO;
 423                         }
 424                         GOTO(out_oa, rc);
 425                 }
 426         } else {
 427                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 428                                    0, &oinfo.oi_md, rec->ur_eadata);
 429                 if (rc) {
 430                         GOTO(out_oa, rc);
 431                 }
 432                 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
 433                 oinfo.oi_md->lsm_object_gr = oinfo.oi_oa->o_gr;
 434         }
 435         if (i_size_read(inode)) {
 436                 oinfo.oi_oa->o_size = i_size_read(inode);
 437                 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
 438                                 OBD_MD_FLATIME | OBD_MD_FLMTIME |
 439                                 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 440
 441                 /* pack lustre id to OST */
 442                 oinfo.oi_oa->o_fid = body->fid1.id;
 443                 oinfo.oi_oa->o_generation = body->fid1.generation;
 444                 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 445
 446                 rc = obd_setattr_rqset(mds->mds_osc_exp, &oinfo, &oti);
 447                 if (rc) {
 448                         CERROR("error setting attrs for inode %lu: rc %d\n",
 449                                inode->i_ino, rc);
 450                         if (rc > 0) {
 451                                 CERROR("obd_setattr_async returned bad rc %d\n",
 452                                        rc);
 453                                 rc = -EIO;
 454                         }
 455                         GOTO(out_oa, rc);
 456                 }
 457         }
 458
 459         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 460         obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
 461
 462         LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
 463         lmm = NULL;
 464         rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
 465         if (rc < 0) {
 466                 CERROR("cannot pack lsm, err = %d\n", rc);
 467                 GOTO(out_oa, rc);
 468         }
 469         lmm_size = rc;
 470         body->eadatasize = rc;
 471
 472         if (*handle == NULL)
 473                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 474         if (IS_ERR(*handle)) {
 475                 rc = PTR_ERR(*handle);
 476                 *handle = NULL;
 477                 GOTO(free_diskmd, rc);
 478         }
 479
 480         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 481         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 482         LASSERT(lmm_buf);
 483         memcpy(lmm_buf, lmm, lmm_size);
 484 free_diskmd:
 485         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 486 out_oa:
 487         oti_free_cookies(&oti);
 488         OBDO_FREE(oinfo.oi_oa);
 489 out_ids:
 490         if (rc) {
 491                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 492                 *ids = NULL;
 493         }
 494         if (oinfo.oi_md)
 495                 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
 496         RETURN(rc);
 497 }
 498
 499 static void reconstruct_open(struct mds_update_record *rec, int offset,
 500                              struct ptlrpc_request *req,
 501                              struct lustre_handle *child_lockh)
 502 {
 503         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 504         struct mds_client_data *mcd = med->med_mcd;
 505         struct mds_obd *mds = mds_req2mds(req);
 506         struct mds_file_data *mfd;
 507         struct obd_device *obd = req->rq_export->exp_obd;
 508         struct dentry *parent, *dchild;
 509         struct ldlm_reply *rep;
 510         struct mds_body *body;
 511         int rc;
 512         struct list_head *t;
 513         int put_child = 1;
 514         ENTRY;
 515
 516         LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
 517         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
 518         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 519
 520         /* copy rc, transno and disp; steal locks */
 521         mds_req_from_mcd(req, mcd);
 522         intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data));
 523
 524         /* Only replay if create or open actually happened. */
 525         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 526                 EXIT;
 527                 return; /* error looking up parent or child */
 528         }
 529
 530         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 531         LASSERT(!IS_ERR(parent));
 532
 533         dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 534         LASSERT(!IS_ERR(dchild));
 535
 536         if (!dchild->d_inode)
 537                 GOTO(out_dput, 0); /* child not present to open */
 538
 539         /* At this point, we know we have a child. We'll send
 540          * it back _unless_ it not created and open failed.
 541          */
 542         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 543             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 544             req->rq_status) {
 545                 GOTO(out_dput, 0);
 546         }
 547
 548         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 549         mds_pack_inode2body(body, dchild->d_inode);
 550         if (S_ISREG(dchild->d_inode->i_mode)) {
 551                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 552                                  body, dchild->d_inode, 1);
 553
 554                 if (rc)
 555                         LASSERT(rc == req->rq_status);
 556
 557                 /* If we have LOV EA data, the OST holds size, mtime */
 558                 if (!(body->valid & OBD_MD_FLEASIZE))
 559                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 560                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 561         }
 562
 563         if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 564                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 565                                     body->eadatasize, 0);
 566
 567         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 568             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 569                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 570
 571                 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
 572                                   body, acl_off);
 573                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 574                 if (!req->rq_status && rc)
 575                         req->rq_status = rc;
 576         }
 577
 578         /* If we have -EEXIST as the status, and we were asked to create
 579          * exclusively, we can tell we failed because the file already existed.
 580          */
 581         if (req->rq_status == -EEXIST &&
 582             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 583              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 584                 GOTO(out_dput, 0);
 585         }
 586
 587         /* If we didn't get as far as trying to open, then some locking thing
 588          * probably went wrong, and we'll just bail here.
 589          */
 590         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 591                 GOTO(out_dput, 0);
 592
 593         /* If we failed, then we must have failed opening, so don't look for
 594          * file descriptor or anything, just give the client the bad news.
 595          */
 596         if (req->rq_status)
 597                 GOTO(out_dput, 0);
 598
 599         mfd = NULL;
 600         spin_lock(&med->med_open_lock);
 601         list_for_each(t, &med->med_open_head) {
 602                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 603                 if (mfd->mfd_xid == req->rq_xid) {
 604                         mds_mfd_addref(mfd);
 605                         break;
 606                 }
 607                 mfd = NULL;
 608         }
 609         spin_unlock(&med->med_open_lock);
 610
 611         /* #warning "XXX fixme" bug 2991 */
 612         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 613          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 614          * to detect a re-open */
 615         if (mfd == NULL) {
 616                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 617                         rc = mds_join_file(rec, req, dchild, NULL);
 618                         if (rc)
 619                                 GOTO(out_dput, rc);
 620                 }
 621                 mntget(mds->mds_vfsmnt);
 622                 CERROR("Re-opened file \n");
 623                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 624                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 625                 if (!mfd) {
 626                         CERROR("mds: out of memory\n");
 627                         GOTO(out_dput, req->rq_status = -ENOMEM);
 628                 }
 629                 put_child = 0;
 630         } else {
 631                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 632                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 633                        mfd->mfd_handle.h_cookie);
 634         }
 635
 636         mds_mfd_put(mfd);
 637
 638  out_dput:
 639         if (put_child)
 640                 l_dput(dchild);
 641         l_dput(parent);
 642         EXIT;
 643 }
 644
 645 /* do NOT or the MAY_*'s, you'll get the weakest */
 646 static int accmode(struct inode *inode, int flags)
 647 {
 648         int res = 0;
 649
 650         /* Sadly, NFSD reopens a file repeatedly during operation, so the
 651          * "acc_mode = 0" allowance for newly-created files isn't honoured.
 652          * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
 653          * owner can write to a file even if it is marked readonly to hide
 654          * its brokenness. (bug 5781) */
 655         if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
 656                 return 0;
 657
 658         if (flags & FMODE_READ)
 659                 res = MAY_READ;
 660         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 661                 res |= MAY_WRITE;
 662         if (flags & MDS_FMODE_EXEC)
 663                 res = MAY_EXEC;
 664         return res;
 665 }
 666
 667 /* Handles object creation, actual opening, and I/O epoch */
 668 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 669                            struct mds_body *body, int flags, void **handle,
 670                            struct mds_update_record *rec,struct ldlm_reply *rep,
 671                            struct lustre_handle *lockh)
 672 {
 673         struct mds_obd *mds = mds_req2mds(req);
 674         struct obd_device *obd = req->rq_export->exp_obd;
 675         struct mds_file_data *mfd = NULL;
 676         obd_id *ids = NULL; /* object IDs created */
 677         int rc = 0;
 678         ENTRY;
 679
 680         /* atomically create objects if necessary */
 681         LOCK_INODE_MUTEX(dchild->d_inode);
 682
 683         if (S_ISREG(dchild->d_inode->i_mode) &&
 684             !(body->valid & OBD_MD_FLEASIZE)) {
 685                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 686                                  body, dchild->d_inode, 0);
 687                 if (rc) {
 688                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 689                         RETURN(rc);
 690                 }
 691         }
 692         if (rec != NULL) {
 693                 if ((body->valid & OBD_MD_FLEASIZE) &&
 694                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 695                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 696                         RETURN(-EEXIST);
 697                 }
 698                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 699                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 700                         rc = mds_join_file(rec, req, dchild, lockh);
 701                         if (rc)
 702                                 RETURN(rc);
 703                         LOCK_INODE_MUTEX(dchild->d_inode);
 704                 }
 705                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 706                     !(body->valid & OBD_MD_FLMODEASIZE)) {
 707                         /* no EA: create objects */
 708                         rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
 709                                                 mds, obd, dchild, handle, &ids);
 710                         if (rc) {
 711                                 CERROR("mds_create_objects: rc = %d\n", rc);
 712                                 UNLOCK_INODE_MUTEX(dchild->d_inode);
 713                                 RETURN(rc);
 714                         }
 715                 }
 716         }
 717         /* If the inode has no EA data, then MDS holds size, mtime */
 718         if (S_ISREG(dchild->d_inode->i_mode) &&
 719             !(body->valid & OBD_MD_FLEASIZE)) {
 720                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 721                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 722         }
 723         UNLOCK_INODE_MUTEX(dchild->d_inode);
 724
 725         if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 726                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 727                                     body->eadatasize, 0);
 728
 729         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 730             rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 731                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 732
 733                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 734                                   dchild->d_inode, req->rq_repmsg,
 735                                   body, acl_off);
 736                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 737                 if (rc)
 738                         RETURN(rc);
 739         }
 740
 741         intent_set_disposition(rep, DISP_OPEN_OPEN);
 742         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 743         if (IS_ERR(mfd))
 744                 RETURN(PTR_ERR(mfd));
 745
 746         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 747                mfd->mfd_handle.h_cookie);
 748
 749         if (ids != NULL) {
 750                 mds_lov_update_objids(obd, ids);
 751                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 752         }
 753         if (rc) /* coverity[deadcode] */
 754                 mds_mfd_unlink(mfd, 1);
 755
 756         mds_mfd_put(mfd);
 757         RETURN(rc);
 758 }
 759
 760 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 761                            struct mds_body *body, int flags,
 762                            struct mds_update_record *rec,struct ldlm_reply *rep)
 763 {
 764         struct mds_obd *mds = mds_req2mds(req);
 765         struct dentry *dchild;
 766         char fidname[LL_FID_NAMELEN];
 767         int fidlen = 0, rc;
 768         void *handle = NULL;
 769         ENTRY;
 770
 771         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 772         dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 773         if (IS_ERR(dchild)) {
 774                 rc = PTR_ERR(dchild);
 775                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 776                 RETURN(rc);
 777         }
 778
 779         if (dchild->d_inode != NULL) {
 780                 mds_inode_set_orphan(dchild->d_inode);
 781                 CWARN("Orphan %s found and opened in PENDING directory\n",
 782                        fidname);
 783         } else {
 784                 l_dput(dchild);
 785
 786                 /* We didn't find it in PENDING so it isn't an orphan.  See
 787                  * if it was a regular inode that was previously created. */
 788                 dchild = mds_fid2dentry(mds, fid, NULL);
 789                 if (IS_ERR(dchild))
 790                         RETURN(PTR_ERR(dchild));
 791         }
 792
 793         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 794         mds_pack_inode2body(body, dchild->d_inode);
 795         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 796         intent_set_disposition(rep, DISP_LOOKUP_POS);
 797
 798         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
 799         rc = mds_finish_transno(mds, dchild->d_inode, handle,
 800                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
 801         /* XXX what do we do here if mds_finish_transno itself failed? */
 802
 803         l_dput(dchild);
 804         RETURN(rc);
 805 }
 806
 807 int mds_pin(struct ptlrpc_request *req, int offset)
 808 {
 809         struct obd_device *obd = req->rq_export->exp_obd;
 810         struct mds_body *reqbody, *repbody;
 811         struct lvfs_run_ctxt saved;
 812         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 813         ENTRY;
 814
 815         reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
 816
 817         rc = lustre_pack_reply(req, 2, size, NULL);
 818         if (rc)
 819                 RETURN(rc);
 820         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 821                                  sizeof(*repbody));
 822
 823         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 824         rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
 825                              NULL);
 826         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 827
 828         RETURN(rc);
 829 }
 830
 831 /*  Get an internal lock on the inode number (but not generation) to sync
 832  *  new inode creation with inode unlink (bug 2029).  If child_lockh is NULL
 833  *  we just get the lock as a barrier to wait for other holders of this lock,
 834  *  and drop it right away again. */
 835 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 836                        struct lustre_handle *child_lockh)
 837 {
 838         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 839         struct lustre_handle lockh;
 840         int lock_flags = LDLM_FL_ATOMIC_CB;
 841         int rc;
 842
 843         if (child_lockh == NULL)
 844                 child_lockh = &lockh;
 845
 846         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
 847                                     LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 848                                     ldlm_blocking_ast, ldlm_completion_ast,
 849                                     NULL, NULL, 0, NULL, child_lockh);
 850         if (rc != ELDLM_OK)
 851                 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
 852         else if (child_lockh == &lockh)
 853                 ldlm_lock_decref(child_lockh, LCK_EX);
 854
 855         RETURN(rc);
 856 }
 857
 858 int mds_open(struct mds_update_record *rec, int offset,
 859              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 860 {
 861         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 862         struct obd_device *obd = req->rq_export->exp_obd;
 863         struct mds_obd *mds = mds_req2mds(req);
 864         struct ldlm_reply *rep = NULL;
 865         struct mds_body *body = NULL;
 866         struct dentry *dchild = NULL, *dparent = NULL;
 867         struct mds_export_data *med;
 868         struct lustre_handle parent_lockh;
 869         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 870         int parent_mode = LCK_CR;
 871         void *handle = NULL;
 872         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
 873         unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
 874         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
 875         int child_mode = LCK_CR;
 876         /* Always returning LOOKUP lock if open succesful to guard
 877            dentry on client. */
 878         ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
 879         struct ldlm_res_id child_res_id = { .name = {0}};
 880         int lock_flags = 0;
 881         ENTRY;
 882
 883         mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
 884
 885         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 886                          (obd_timeout + 1) / 4);
 887
 888         CLASSERT(MAXQUOTAS < 4);
 889         if (offset == DLM_INTENT_REC_OFF) { /* intent */
 890                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 891                                      sizeof(*rep));
 892                 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
 893                                       sizeof(*body));
 894         } else if (offset == REQ_REC_OFF) { /* non-intent reint */
 895                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 896                                       sizeof(*body));
 897                 LBUG(); /* XXX: not supported yet? */
 898         } else {
 899                 body = NULL;
 900                 LBUG();
 901         }
 902
 903         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 904
 905         /* Step 0: If we are passed a fid, then we assume the client already
 906          * opened this file and is only replaying the RPC, so we open the
 907          * inode by fid (at some large expense in security). */
 908         /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
 909         if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
 910              (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
 911             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 912                 if (rec->ur_fid2->id == 0) {
 913                         struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
 914                         if (lock) {
 915                                 LDLM_ERROR(lock, "fid2 not set on open replay");
 916                                 LDLM_LOCK_PUT(lock);
 917                         }
 918                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 919                         RETURN(-EFAULT);
 920                 }
 921
 922                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 923                                      rec, rep);
 924                 if (rc != -ENOENT) {
 925                         if (req->rq_export->exp_libclient &&
 926                             rec->ur_flags & MDS_OPEN_HAS_EA)
 927                                 RETURN(0);
 928
 929                         RETURN(rc);
 930                 }
 931
 932                 /* We didn't find the correct inode on disk either, so we
 933                  * need to re-create it via a regular replay. */
 934                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 935                         DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
 936                         RETURN(-EFAULT);
 937                 }
 938         } else if (rec->ur_fid2->id) {
 939                 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
 940                           rec->ur_fid2->id, rec->ur_fid2->generation);
 941                 RETURN(-EFAULT);
 942         }
 943
 944         /* If we got here, we must be called via intent */
 945         LASSERT(offset == DLM_INTENT_REC_OFF);
 946
 947         med = &req->rq_export->exp_mds_data;
 948         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 949                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 950                 RETURN(-ENOMEM);
 951         }
 952
 953         /* Step 1: Find and lock the parent */
 954         if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
 955                 parent_mode = LCK_EX;
 956         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 957                                         &parent_lockh, MDS_INODELOCK_UPDATE);
 958         if (IS_ERR(dparent)) {
 959                 rc = PTR_ERR(dparent);
 960                 if (rc != -ENOENT) {
 961                         CERROR("parent "LPU64"/%u lookup error %d\n",
 962                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 963                 } else {
 964                         /* Just cannot find parent - make it look like
 965                          * usual negative lookup to avoid extra MDS RPC */
 966                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 967                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
 968                 }
 969                 GOTO(cleanup, rc);
 970         }
 971         LASSERT(dparent->d_inode != NULL);
 972
 973         cleanup_phase = 1; /* parent dentry and lock */
 974
 975         if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 976                 dchild = dget(dparent);
 977                 cleanup_phase = 2; /* child dentry */
 978                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
 979                 GOTO(found_child, rc);
 980         }
 981
 982         /* Step 2: Lookup the child */
 983
 984         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
 985             (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) {
 986                 /* hack for nfsd with no_subtree_check, it will use anon
 987                  * dentry w/o filename to open the file. the anon dentry's
 988                  * parent was set to itself, so rec->ur_fid1 is the file.
 989                  * And in MDC it cannot derive the dentry's parent dentry,
 990                  * hence the file's name, so we hack here in MDS,
 991                  * refer to bug 13030. */
 992                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 993         } else {
 994                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
 995                                            rec->ur_namelen - 1);
 996         }
 997         if (IS_ERR(dchild)) {
 998                 rc = PTR_ERR(dchild);
 999                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1000                 GOTO(cleanup, rc);
1001         }
1002
1003         cleanup_phase = 2; /* child dentry */
1004
1005         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1006         if (dchild->d_inode)
1007                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1008         else
1009                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1010
1011         /*Step 3: If the child was negative, and we're supposed to, create it.*/
1012         if (dchild->d_inode == NULL) {
1013                 unsigned long ino = rec->ur_fid2->id;
1014                 struct iattr iattr;
1015                 struct inode *inode;
1016
1017                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1018                         /* It's negative and we weren't supposed to create it */
1019                         GOTO(cleanup, rc = -ENOENT);
1020                 }
1021
1022                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1023                         GOTO(cleanup, rc = -EROFS);
1024
1025                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1026                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1027                                       NULL);
1028                 if (IS_ERR(handle)) {
1029                         rc = PTR_ERR(handle);
1030                         handle = NULL;
1031                         GOTO(cleanup, rc);
1032                 }
1033                 dchild->d_fsdata = (void *) &dp;
1034                 dp.ldp_ptr = req;
1035                 dp.ldp_inum = ino;
1036
1037                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1038                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1039                         dchild->d_fsdata = NULL;
1040
1041                 if (rc) {
1042                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1043                         GOTO(cleanup, rc);
1044                 }
1045                 inode = dchild->d_inode;
1046                 if (ino) {
1047                         LASSERT(ino == inode->i_ino);
1048                         /* Written as part of setattr */
1049                         inode->i_generation = rec->ur_fid2->generation;
1050                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1051                                inode->i_ino, inode->i_generation);
1052                 }
1053
1054                 created = 1;
1055                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1056                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1057                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1058
1059                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
1060                 if (dparent->d_inode->i_mode & S_ISGID)
1061                         iattr.ia_gid = dparent->d_inode->i_gid;
1062                 else
1063                         iattr.ia_gid = current->fsgid;
1064
1065                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1066                         ATTR_MTIME | ATTR_CTIME;
1067
1068                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1069                 if (rc)
1070                         CERROR("error on child setattr: rc = %d\n", rc);
1071
1072                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1073
1074                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1075                 if (rc)
1076                         CERROR("error on parent setattr: rc = %d\n", rc);
1077
1078                 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1079                 handle = NULL;
1080                 acc_mode = 0;           /* Don't check for permissions */
1081         } else {
1082                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1083         }
1084
1085         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1086                  "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1087                  dchild->d_name.name, dchild, dchild->d_inode,
1088                  dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1089
1090 found_child:
1091         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1092         mds_pack_inode2body(body, dchild->d_inode);
1093
1094         if (S_ISREG(dchild->d_inode->i_mode)) {
1095                 /* Check permissions etc */
1096                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1097                 if (rc != 0)
1098                         GOTO(cleanup, rc);
1099
1100                 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1101                     (acc_mode & MAY_WRITE))
1102                         GOTO(cleanup, rc = -EROFS);
1103
1104                 /* An append-only file must be opened in append mode for
1105                  * writing */
1106                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1107                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1108                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1109                         GOTO(cleanup, rc = -EPERM);
1110         }
1111
1112         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1113             (rec->ur_flags & MDS_OPEN_EXCL)) {
1114                 /* File already exists, we didn't just create it, and we
1115                  * were passed O_EXCL; err-or. */
1116                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1117         }
1118
1119         /* if we are following a symlink, don't open */
1120         if (S_ISLNK(dchild->d_inode->i_mode))
1121                 GOTO(cleanup_no_trans, rc = 0);
1122
1123         if (S_ISDIR(dchild->d_inode->i_mode)) {
1124                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1125                     rec->ur_flags & FMODE_WRITE) {
1126                         /* we are trying to create or write a exist dir */
1127                         GOTO(cleanup, rc = -EISDIR);
1128                 }
1129                 if (rec->ur_flags & MDS_FMODE_EXEC) {
1130                         /* we are trying to exec a directory */
1131                         GOTO(cleanup, rc = -EACCES);
1132                 }
1133                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1134                         intent_set_disposition(rep, DISP_OPEN_OPEN);
1135                         GOTO(cleanup, rc = -EACCES);
1136                 }
1137         } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1138                 GOTO(cleanup, rc = -ENOTDIR);
1139         }
1140
1141         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1142                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1143                 GOTO(cleanup, rc = -EAGAIN);
1144         }
1145
1146         /* Obtain OPEN lock as well */
1147         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1148
1149         /* We cannot use acc_mode here, because it is zeroed in case of
1150            creating a file, so we get wrong lockmode */
1151         if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1152                 child_mode = LCK_CW;
1153         else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1154                 child_mode = LCK_PR;
1155         else
1156                 child_mode = LCK_CR;
1157
1158         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1159              (rec->ur_flags & MDS_OPEN_LOCK)) {
1160                 /* In case of replay we do not get a lock assuming that the
1161                    caller has it already */
1162                 child_res_id.name[0] = dchild->d_inode->i_ino;
1163                 child_res_id.name[1] = dchild->d_inode->i_generation;
1164
1165                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1166                                             LDLM_IBITS, &policy, child_mode,
1167                                             &lock_flags, ldlm_blocking_ast,
1168                                             ldlm_completion_ast, NULL, NULL,
1169                                             0, NULL, child_lockh);
1170                 if (rc != ELDLM_OK)
1171                         GOTO(cleanup, rc);
1172
1173                 /* Let mds_intent_policy know that we have a lock to return */
1174                 intent_set_disposition(rep, DISP_OPEN_LOCK);
1175                 cleanup_phase = 3;
1176         }
1177
1178         if (!S_ISREG(dchild->d_inode->i_mode) &&
1179             !S_ISDIR(dchild->d_inode->i_mode) &&
1180             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1181                 /* If client supports this, do not return open handle for
1182                  * special device nodes */
1183                 GOTO(cleanup_no_trans, rc = 0);
1184         }
1185
1186         /* Step 5: mds_open it */
1187         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1188                              rep, &parent_lockh);
1189         GOTO(cleanup, rc);
1190
1191  cleanup:
1192         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1193                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
1194
1195  cleanup_no_trans:
1196         switch (cleanup_phase) {
1197         case 3:
1198                 if (rc)
1199                         /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1200                          * mds_intent_policy won't try to return any locks */
1201                         ldlm_lock_decref(child_lockh, child_mode);
1202         case 2:
1203                 if (rc && created) {
1204                         int err = vfs_unlink(dparent->d_inode, dchild);
1205                         if (err) {
1206                                 CERROR("unlink(%.*s) in error path: %d\n",
1207                                        dchild->d_name.len, dchild->d_name.name,
1208                                        err);
1209                         }
1210                 } else if (created) {
1211                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1212                         /* save uid/gid for quota acquire/release */
1213                         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1214                         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1215                 }
1216                 l_dput(dchild);
1217         case 1:
1218                 if (dparent == NULL)
1219                         break;
1220
1221                 l_dput(dparent);
1222                 if (rc)
1223                         ldlm_lock_decref(&parent_lockh, parent_mode);
1224                 else
1225                         ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1226         }
1227         /* trigger dqacq on the owner of child and parent */
1228         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1229                       FSFILT_OP_CREATE);
1230
1231         /* If we have not taken the "open" lock, we may not return 0 here,
1232            because caller expects 0 to mean "lock is taken", and it needs
1233            nonzero return here for caller to return EDLM_LOCK_ABORTED to
1234            client. Later caller should rewrite the return value back to zero
1235            if it to be used any further
1236          */
1237         RETURN(rc);
1238 }
1239
1240 /* Close a "file descriptor" and possibly unlink an orphan from the
1241  * PENDING directory.  Caller must hold child->i_mutex, this drops it.
1242  *
1243  * If we are being called from mds_disconnect() because the client has
1244  * disappeared, then req == NULL and we do not update last_rcvd because
1245  * there is nothing that could be recovered by the client at this stage
1246  * (it will not even _have_ an entry in last_rcvd anymore).
1247  *
1248  * Returns EAGAIN if the client needs to get more data and re-close. */
1249 int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd,
1250                   struct mds_file_data *mfd, int unlink_orphan,
1251                   struct lov_mds_md *lmm, int lmm_size,
1252                   struct llog_cookie *logcookies, int cookies_size,
1253                   __u64 *valid)
1254 {
1255         struct inode *inode = mfd->mfd_dentry->d_inode;
1256         char fidname[LL_FID_NAMELEN];
1257         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1258         struct dentry *pending_child = NULL;
1259         struct mds_obd *mds = &obd->u.mds;
1260         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1261         void *handle = NULL;
1262         struct mds_body *request_body = NULL, *reply_body = NULL;
1263         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1264         struct iattr iattr = { 0 };
1265         ENTRY;
1266
1267         if (req && req->rq_reqmsg != NULL)
1268                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1269                                               sizeof(*request_body));
1270         if (req && req->rq_repmsg != NULL)
1271                 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1272                                             sizeof(*reply_body));
1273
1274         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1275
1276         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1277                inode->i_nlink, mds_orphan_open_count(inode));
1278
1279         last_orphan = mds_orphan_open_dec_test(inode) &&
1280                 mds_inode_is_orphan(inode);
1281         MDS_UP_WRITE_ORPHAN_SEM(inode);
1282
1283         /* this is half of the actual "close" */
1284         if (mfd->mfd_mode & FMODE_WRITE) {
1285                 rc = mds_put_write_access(mds, inode, request_body,
1286                                           last_orphan && unlink_orphan);
1287         } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1288                 mds_allow_write_access(inode);
1289         }
1290
1291         if (last_orphan && unlink_orphan) {
1292                 int stripe_count = 0;
1293                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1294
1295                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1296
1297                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1298                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1299                         CERROR("found \"orphan\" %s %s with link count %d\n",
1300                                S_ISREG(inode->i_mode) ? "file" : "dir",
1301                                fidname, inode->i_nlink);
1302
1303                 /* Sadly, there is no easy way to save pending_child from
1304                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1305                  * but normally it will still be in the dcache. */
1306                 LOCK_INODE_MUTEX(pending_dir);
1307                 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1308                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1309                                                fidlen);
1310                 if (IS_ERR(pending_child))
1311                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1312                 LASSERT(pending_child->d_inode != NULL);
1313
1314                 cleanup_phase = 2; /* dput(pending_child) when finished */
1315                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1316                         rc = vfs_rmdir(pending_dir, pending_child);
1317                         if (rc)
1318                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1319                                        fidname,rc);
1320                         goto out;
1321                 }
1322
1323                 if (lmm != NULL) {
1324                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1325                 }
1326
1327                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1328                                           NULL, stripe_count);
1329                 if (IS_ERR(handle)) {
1330                         rc = PTR_ERR(handle);
1331                         handle = NULL;
1332                         GOTO(cleanup, rc);
1333                 }
1334
1335                 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1336                     mds_log_op_unlink(obd, lmm, lmm_size,
1337                                       logcookies, cookies_size) > 0) {
1338                         *valid |= OBD_MD_FLCOOKIE;
1339                 }
1340
1341                 dp.ldp_inum = 0;
1342                 dp.ldp_ptr = req;
1343                 pending_child->d_fsdata = (void *) &dp;
1344                 rc = vfs_unlink(pending_dir, pending_child);
1345                 if (rc)
1346                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1347
1348                 goto out; /* Don't bother updating attrs on unlinked inode */
1349         }
1350
1351 #if 0
1352         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1353                 /* Update the on-disk attributes if this was the last write
1354                  * close, and all information was provided (i.e., rc == 0)
1355                  *
1356                  * XXX this should probably be abstracted with mds_reint_setattr
1357                  */
1358
1359                 if (request_body->valid & OBD_MD_FLMTIME &&
1360                     LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1361                         LTIME_S(iattr.ia_mtime) = request_body->mtime;
1362                         iattr.ia_valid |= ATTR_MTIME;
1363                 }
1364                 if (request_body->valid & OBD_MD_FLCTIME &&
1365                     LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1366                         LTIME_S(iattr.ia_ctime) = request_body->ctime;
1367                         iattr.ia_valid |= ATTR_CTIME;
1368                 }
1369
1370                 /* XXX can't set block count with fsfilt_setattr (!) */
1371                 if (request_body->valid & OBD_MD_FLSIZE) {
1372                         iattr.ia_valid |= ATTR_SIZE;
1373                         iattr.ia_size = request_body->size;
1374                 }
1375                 /* iattr.ia_blocks = request_body->blocks */
1376
1377         }
1378 #endif
1379         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1380                 /* Only start a transaction to write out only the atime if
1381                  * it is more out-of-date than the specified limit.  If we
1382                  * are already going to write out the atime then do it anyway.
1383                  * */
1384                 LTIME_S(iattr.ia_atime) = request_body->atime;
1385                 if ((LTIME_S(iattr.ia_atime) >
1386                      LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1387                     (iattr.ia_valid != 0 &&
1388                      LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1389                         iattr.ia_valid |= ATTR_ATIME;
1390         }
1391
1392         if (iattr.ia_valid != 0) {
1393                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1394                 if (IS_ERR(handle)) {
1395                         rc = PTR_ERR(handle);
1396                         handle = NULL;
1397                         GOTO(cleanup, rc);
1398                 }
1399                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1400                 if (rc)
1401                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1402         }
1403 out:
1404         /* If other clients have this file open for write, rc will be > 0 */
1405         if (rc > 0)
1406                 rc = 0;
1407         l_dput(mfd->mfd_dentry);
1408         mds_mfd_put(mfd);
1409
1410  cleanup:
1411         if (req != NULL && reply_body != NULL) {
1412                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0);
1413         } else if (handle) {
1414                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1415                 if (err) {
1416                         CERROR("error committing close: %d\n", err);
1417                         if (!rc)
1418                                 rc = err;
1419                 }
1420         }
1421
1422         switch (cleanup_phase) {
1423         case 2:
1424                 dput(pending_child);
1425         case 1:
1426                 UNLOCK_INODE_MUTEX(pending_dir);
1427         }
1428         RETURN(rc);
1429 }
1430
1431 int mds_close(struct ptlrpc_request *req, int offset)
1432 {
1433         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1434         struct obd_device *obd = req->rq_export->exp_obd;
1435         struct mds_body *body;
1436         struct mds_file_data *mfd;
1437         struct lvfs_run_ctxt saved;
1438         struct inode *inode;
1439         int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1440                                sizeof(struct mds_body),
1441                                obd->u.mds.mds_max_mdsize,
1442                                obd->u.mds.mds_max_cookiesize };
1443         struct mds_body *reply_body;
1444         struct lov_mds_md *lmm;
1445         int lmm_size;
1446         struct llog_cookie *logcookies;
1447         int cookies_size;
1448         ENTRY;
1449
1450         rc = lustre_pack_reply(req, 4, repsize, NULL);
1451         if (rc) {
1452                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1453                 req->rq_status = rc;
1454                 /* continue on to drop local open even if we can't send reply */
1455         } else {
1456                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1457         }
1458
1459         CDEBUG(D_HA, "close req->rep_len %d mdsize %d cookiesize %d\n",
1460                req->rq_replen,
1461                obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1462         mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1463
1464         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1465                                   lustre_swab_mds_body);
1466         if (body == NULL) {
1467                 CERROR("Can't unpack body\n");
1468                 req->rq_status = -EFAULT;
1469                 RETURN(-EFAULT);
1470         }
1471
1472         spin_lock(&med->med_open_lock);
1473         mfd = mds_handle2mfd(&body->handle);
1474         if (mfd == NULL) {
1475                 spin_unlock(&med->med_open_lock);
1476                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1477                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1478                 req->rq_status = -ESTALE;
1479                 RETURN(-ESTALE);
1480         }
1481         /* Remove mfd handle so it can't be found again.  We consume mfd_list
1482          * reference here, but still have mds_handle2mfd ref until mfd_close. */
1483         mds_mfd_unlink(mfd, 1);
1484         spin_unlock(&med->med_open_lock);
1485
1486         inode = mfd->mfd_dentry->d_inode;
1487         /* child orphan sem protects orphan_dec_test && is_orphan race */
1488         MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1489         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1490                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1491                                       sizeof(*body));
1492                 LASSERT(body != NULL);
1493
1494                 mds_pack_inode2fid(&body->fid1, inode);
1495                 mds_pack_inode2body(body, inode);
1496                 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1497                             MDS_PACK_MD_LOCK);
1498         }
1499
1500         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1501         reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1502         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1503         lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1504         logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1505         cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1506         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1507                                        lmm, lmm_size, logcookies, cookies_size,
1508                                        &reply_body->valid);
1509         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1510
1511         mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1512         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1513                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1514                 req->rq_status = -ENOMEM;
1515                 RETURN(-ENOMEM);
1516         }
1517
1518         RETURN(rc);
1519 }
1520
1521 int mds_done_writing(struct ptlrpc_request *req, int offset)
1522 {
1523         struct mds_body *body;
1524         int rc, size[2] = { sizeof(struct ptlrpc_body),
1525                             sizeof(struct mds_body) };
1526         ENTRY;
1527
1528         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1529
1530         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1531                                   lustre_swab_mds_body);
1532         if (body == NULL) {
1533                 CERROR("Can't unpack body\n");
1534                 req->rq_status = -EFAULT;
1535                 RETURN(-EFAULT);
1536         }
1537
1538         rc = lustre_pack_reply(req, 2, size, NULL);
1539         if (rc) {
1540                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1541                 req->rq_status = rc;
1542         }
1543
1544         RETURN(0);
1545 }