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 (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 324             !(rec->ur_flags & FMODE_WRITE))
 325                 RETURN(0);
 326
 327         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 328
 329         if (body->valid & OBD_MD_FLEASIZE)
 330                 RETURN(0);
 331
 332         OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 333         if (*ids == NULL)
 334                 RETURN(-ENOMEM);
 335         oti_init(&oti, req);
 336         oti.oti_objid = *ids;
 337
 338         /* replay case */
 339         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 340                 if (rec->ur_fid2->id == 0) {
 341                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 342                         RETURN(-EFAULT);
 343                 }
 344
 345                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 346                 lmm_size = rec->ur_eadatalen;
 347                 lmm = rec->ur_eadata;
 348                 LASSERT(lmm);
 349
 350                 if (*handle == NULL)
 351                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 352                 if (IS_ERR(*handle)) {
 353                         rc = PTR_ERR(*handle);
 354                         *handle = NULL;
 355                         GOTO(out_ids, rc);
 356                 }
 357
 358                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 359
 360                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 361                 if (rc)
 362                         CERROR("open replay failed to set md:%d\n", rc);
 363                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 364                 LASSERT(lmm_buf);
 365                 memcpy(lmm_buf, lmm, lmm_size);
 366
 367                 RETURN(rc);
 368         }
 369
 370         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 371                 GOTO(out_ids, rc = -ENOMEM);
 372
 373         OBDO_ALLOC(oinfo.oi_oa);
 374         if (oinfo.oi_oa == NULL)
 375                 GOTO(out_ids, rc = -ENOMEM);
 376         oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
 377         oinfo.oi_oa->o_gid = 0;
 378         oinfo.oi_oa->o_mode = S_IFREG | 0600;
 379         oinfo.oi_oa->o_id = inode->i_ino;
 380         oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
 381                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
 382         oinfo.oi_oa->o_size = 0;
 383
 384         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 385                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 386
 387         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 388                 /* check if things like lfs setstripe are sending us the ea */
 389                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 390                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 391                                            mds->mds_osc_exp,
 392                                            0, &oinfo.oi_md, rec->ur_eadata);
 393                         if (rc)
 394                                 GOTO(out_oa, rc);
 395                 } else {
 396                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 397                         if (lmm == NULL)
 398                                 GOTO(out_oa, rc = -ENOMEM);
 399
 400                         lmm_size = mds->mds_max_mdsize;
 401                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 402                                         lmm, &lmm_size, 1);
 403                         if (rc > 0)
 404                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 405                                                    mds->mds_osc_exp,
 406                                                    0, &oinfo.oi_md, lmm);
 407                         OBD_FREE(lmm, mds->mds_max_mdsize);
 408                         if (rc)
 409                                 GOTO(out_oa, rc);
 410                 }
 411                 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
 412                                 &oinfo.oi_md, &oti);
 413                 if (rc) {
 414                         int level = D_ERROR;
 415                         if (rc == -ENOSPC)
 416                                 level = D_INODE;
 417                         CDEBUG(level, "error creating objects for "
 418                                       "inode %lu: rc = %d\n",
 419                                inode->i_ino, rc);
 420                         if (rc > 0) {
 421                                 CERROR("obd_create returned invalid "
 422                                        "rc %d\n", rc);
 423                                 rc = -EIO;
 424                         }
 425                         GOTO(out_oa, rc);
 426                 }
 427         } else {
 428                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 429                                    0, &oinfo.oi_md, rec->ur_eadata);
 430                 if (rc) {
 431                         GOTO(out_oa, rc);
 432                 }
 433                 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
 434                 oinfo.oi_md->lsm_object_gr = oinfo.oi_oa->o_gr;
 435         }
 436         if (inode->i_size) {
 437                 oinfo.oi_oa->o_size = inode->i_size;
 438                 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
 439                                 OBD_MD_FLATIME | OBD_MD_FLMTIME |
 440                                 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
 441
 442                 /* pack lustre id to OST */
 443                 oinfo.oi_oa->o_fid = body->fid1.id;
 444                 oinfo.oi_oa->o_generation = body->fid1.generation;
 445                 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
 446
 447                 rc = obd_setattr_rqset(mds->mds_osc_exp, &oinfo, &oti);
 448                 if (rc) {
 449                         CERROR("error setting attrs for inode %lu: rc %d\n",
 450                                inode->i_ino, rc);
 451                         if (rc > 0) {
 452                                 CERROR("obd_setattr_async returned bad rc %d\n",
 453                                        rc);
 454                                 rc = -EIO;
 455                         }
 456                         GOTO(out_oa, rc);
 457                 }
 458         }
 459
 460         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 461         obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
 462
 463         LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
 464         lmm = NULL;
 465         rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
 466         if (rc < 0) {
 467                 CERROR("cannot pack lsm, err = %d\n", rc);
 468                 GOTO(out_oa, rc);
 469         }
 470         lmm_size = rc;
 471         body->eadatasize = rc;
 472
 473         if (*handle == NULL)
 474                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 475         if (IS_ERR(*handle)) {
 476                 rc = PTR_ERR(*handle);
 477                 *handle = NULL;
 478                 GOTO(free_diskmd, rc);
 479         }
 480
 481         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 482         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 483         LASSERT(lmm_buf);
 484         memcpy(lmm_buf, lmm, lmm_size);
 485 free_diskmd:
 486         obd_free_diskmd(mds->mds_osc_exp, &lmm);
 487 out_oa:
 488         oti_free_cookies(&oti);
 489         OBDO_FREE(oinfo.oi_oa);
 490 out_ids:
 491         if (rc) {
 492                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 493                 *ids = NULL;
 494         }
 495         if (oinfo.oi_md)
 496                 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
 497         RETURN(rc);
 498 }
 499
 500 static void reconstruct_open(struct mds_update_record *rec, int offset,
 501                              struct ptlrpc_request *req,
 502                              struct lustre_handle *child_lockh)
 503 {
 504         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 505         struct mds_client_data *mcd = med->med_mcd;
 506         struct mds_obd *mds = mds_req2mds(req);
 507         struct mds_file_data *mfd;
 508         struct obd_device *obd = req->rq_export->exp_obd;
 509         struct dentry *parent, *dchild;
 510         struct ldlm_reply *rep;
 511         struct mds_body *body;
 512         int rc;
 513         struct list_head *t;
 514         int put_child = 1;
 515         ENTRY;
 516
 517         LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
 518         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
 519         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 520
 521         /* copy rc, transno and disp; steal locks */
 522         mds_req_from_mcd(req, mcd);
 523         intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data));
 524
 525         /* Only replay if create or open actually happened. */
 526         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 527                 EXIT;
 528                 return; /* error looking up parent or child */
 529         }
 530
 531         parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 532         LASSERT(!IS_ERR(parent));
 533
 534         dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
 535         LASSERT(!IS_ERR(dchild));
 536
 537         if (!dchild->d_inode)
 538                 GOTO(out_dput, 0); /* child not present to open */
 539
 540         /* At this point, we know we have a child. We'll send
 541          * it back _unless_ it not created and open failed.
 542          */
 543         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 544             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 545             req->rq_status) {
 546                 GOTO(out_dput, 0);
 547         }
 548
 549         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 550         mds_pack_inode2body(body, dchild->d_inode);
 551         if (S_ISREG(dchild->d_inode->i_mode)) {
 552                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 553                                  body, dchild->d_inode, 1);
 554
 555                 if (rc)
 556                         LASSERT(rc == req->rq_status);
 557
 558                 /* If we have LOV EA data, the OST holds size, mtime */
 559                 if (!(body->valid & OBD_MD_FLEASIZE))
 560                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 561                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 562         }
 563
 564         if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 565                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 566                                     body->eadatasize, 0);
 567
 568         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 569             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 570                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 571
 572                 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
 573                                   body, acl_off);
 574                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 575                 if (!req->rq_status && rc)
 576                         req->rq_status = rc;
 577         }
 578
 579         /* If we have -EEXIST as the status, and we were asked to create
 580          * exclusively, we can tell we failed because the file already existed.
 581          */
 582         if (req->rq_status == -EEXIST &&
 583             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 584              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 585                 GOTO(out_dput, 0);
 586         }
 587
 588         /* If we didn't get as far as trying to open, then some locking thing
 589          * probably went wrong, and we'll just bail here.
 590          */
 591         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 592                 GOTO(out_dput, 0);
 593
 594         /* If we failed, then we must have failed opening, so don't look for
 595          * file descriptor or anything, just give the client the bad news.
 596          */
 597         if (req->rq_status)
 598                 GOTO(out_dput, 0);
 599
 600         mfd = NULL;
 601         spin_lock(&med->med_open_lock);
 602         list_for_each(t, &med->med_open_head) {
 603                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 604                 if (mfd->mfd_xid == req->rq_xid) {
 605                         mds_mfd_addref(mfd);
 606                         break;
 607                 }
 608                 mfd = NULL;
 609         }
 610         spin_unlock(&med->med_open_lock);
 611
 612         /* #warning "XXX fixme" bug 2991 */
 613         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 614          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 615          * to detect a re-open */
 616         if (mfd == NULL) {
 617                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 618                         rc = mds_join_file(rec, req, dchild, NULL);
 619                         if (rc)
 620                                 GOTO(out_dput, rc);
 621                 }
 622                 mntget(mds->mds_vfsmnt);
 623                 CERROR("Re-opened file \n");
 624                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 625                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 626                 if (!mfd) {
 627                         CERROR("mds: out of memory\n");
 628                         GOTO(out_dput, req->rq_status = -ENOMEM);
 629                 }
 630                 put_child = 0;
 631         } else {
 632                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 633                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 634                        mfd->mfd_handle.h_cookie);
 635         }
 636
 637         mds_mfd_put(mfd);
 638
 639  out_dput:
 640         if (put_child)
 641                 l_dput(dchild);
 642         l_dput(parent);
 643         EXIT;
 644 }
 645
 646 /* do NOT or the MAY_*'s, you'll get the weakest */
 647 static int accmode(struct inode *inode, int flags)
 648 {
 649         int res = 0;
 650
 651         /* Sadly, NFSD reopens a file repeatedly during operation, so the
 652          * "acc_mode = 0" allowance for newly-created files isn't honoured.
 653          * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
 654          * owner can write to a file even if it is marked readonly to hide
 655          * its brokenness. (bug 5781) */
 656         if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
 657                 return 0;
 658
 659         if (flags & FMODE_READ)
 660                 res = MAY_READ;
 661         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 662                 res |= MAY_WRITE;
 663         if (flags & MDS_FMODE_EXEC)
 664                 res = MAY_EXEC;
 665         return res;
 666 }
 667
 668 /* Handles object creation, actual opening, and I/O epoch */
 669 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 670                            struct mds_body *body, int flags, void **handle,
 671                            struct mds_update_record *rec,struct ldlm_reply *rep,
 672                            struct lustre_handle *lockh)
 673 {
 674         struct mds_obd *mds = mds_req2mds(req);
 675         struct obd_device *obd = req->rq_export->exp_obd;
 676         struct mds_file_data *mfd = NULL;
 677         obd_id *ids = NULL; /* object IDs created */
 678         int rc = 0;
 679         ENTRY;
 680
 681         /* atomically create objects if necessary */
 682         LOCK_INODE_MUTEX(dchild->d_inode);
 683
 684         if (S_ISREG(dchild->d_inode->i_mode) &&
 685             !(body->valid & OBD_MD_FLEASIZE)) {
 686                 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
 687                                  body, dchild->d_inode, 0);
 688                 if (rc) {
 689                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 690                         RETURN(rc);
 691                 }
 692         }
 693         if (rec != NULL) {
 694                 if ((body->valid & OBD_MD_FLEASIZE) &&
 695                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 696                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 697                         RETURN(-EEXIST);
 698                 }
 699                 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 700                         UNLOCK_INODE_MUTEX(dchild->d_inode);
 701                         rc = mds_join_file(rec, req, dchild, lockh);
 702                         if (rc)
 703                                 RETURN(rc);
 704                         LOCK_INODE_MUTEX(dchild->d_inode);
 705                 }
 706                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 707                     !(body->valid & OBD_MD_FLMODEASIZE)) {
 708                         /* no EA: create objects */
 709                         rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
 710                                                 mds, obd, dchild, handle, &ids);
 711                         if (rc) {
 712                                 CERROR("mds_create_objects: rc = %d\n", rc);
 713                                 UNLOCK_INODE_MUTEX(dchild->d_inode);
 714                                 RETURN(rc);
 715                         }
 716                 }
 717         }
 718         /* If the inode has no EA data, then MDS holds size, mtime */
 719         if (S_ISREG(dchild->d_inode->i_mode) &&
 720             !(body->valid & OBD_MD_FLEASIZE)) {
 721                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 722                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 723         }
 724         UNLOCK_INODE_MUTEX(dchild->d_inode);
 725
 726         if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
 727                 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
 728                                     body->eadatasize, 0);
 729
 730         if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
 731             rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 732                 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
 733
 734                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 735                                   dchild->d_inode, req->rq_repmsg,
 736                                   body, acl_off);
 737                 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
 738                 if (rc)
 739                         RETURN(rc);
 740         }
 741
 742         intent_set_disposition(rep, DISP_OPEN_OPEN);
 743         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 744         if (IS_ERR(mfd))
 745                 RETURN(PTR_ERR(mfd));
 746
 747         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 748                mfd->mfd_handle.h_cookie);
 749
 750         if (ids != NULL) {
 751                 mds_lov_update_objids(obd, ids);
 752                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 753         }
 754         if (rc) /* coverity[deadcode] */
 755                 mds_mfd_unlink(mfd, 1);
 756
 757         mds_mfd_put(mfd);
 758         RETURN(rc);
 759 }
 760
 761 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
 762                            struct mds_body *body, int flags,
 763                            struct mds_update_record *rec,struct ldlm_reply *rep)
 764 {
 765         struct mds_obd *mds = mds_req2mds(req);
 766         struct dentry *dchild;
 767         char fidname[LL_FID_NAMELEN];
 768         int fidlen = 0, rc;
 769         void *handle = NULL;
 770         ENTRY;
 771
 772         fidlen = ll_fid2str(fidname, fid->id, fid->generation);
 773         dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 774         if (IS_ERR(dchild)) {
 775                 rc = PTR_ERR(dchild);
 776                 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
 777                 RETURN(rc);
 778         }
 779
 780         if (dchild->d_inode != NULL) {
 781                 mds_inode_set_orphan(dchild->d_inode);
 782                 CWARN("Orphan %s found and opened in PENDING directory\n",
 783                        fidname);
 784         } else {
 785                 l_dput(dchild);
 786
 787                 /* We didn't find it in PENDING so it isn't an orphan.  See
 788                  * if it was a regular inode that was previously created. */
 789                 dchild = mds_fid2dentry(mds, fid, NULL);
 790                 if (IS_ERR(dchild))
 791                         RETURN(PTR_ERR(dchild));
 792         }
 793
 794         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
 795         mds_pack_inode2body(body, dchild->d_inode);
 796         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 797         intent_set_disposition(rep, DISP_LOOKUP_POS);
 798
 799         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
 800         rc = mds_finish_transno(mds, dchild->d_inode, handle,
 801                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
 802         /* XXX what do we do here if mds_finish_transno itself failed? */
 803
 804         l_dput(dchild);
 805         RETURN(rc);
 806 }
 807
 808 int mds_pin(struct ptlrpc_request *req, int offset)
 809 {
 810         struct obd_device *obd = req->rq_export->exp_obd;
 811         struct mds_body *reqbody, *repbody;
 812         struct lvfs_run_ctxt saved;
 813         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 814         ENTRY;
 815
 816         reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
 817
 818         rc = lustre_pack_reply(req, 2, size, NULL);
 819         if (rc)
 820                 RETURN(rc);
 821         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 822                                  sizeof(*repbody));
 823
 824         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 825         rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
 826                              NULL);
 827         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 828
 829         RETURN(rc);
 830 }
 831
 832 /*  Get an internal lock on the inode number (but not generation) to sync
 833  *  new inode creation with inode unlink (bug 2029).  If child_lockh is NULL
 834  *  we just get the lock as a barrier to wait for other holders of this lock,
 835  *  and drop it right away again. */
 836 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 837                        struct lustre_handle *child_lockh)
 838 {
 839         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 840         struct lustre_handle lockh;
 841         int lock_flags = LDLM_FL_ATOMIC_CB;
 842         int rc;
 843
 844         if (child_lockh == NULL)
 845                 child_lockh = &lockh;
 846
 847         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
 848                                     LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 849                                     ldlm_blocking_ast, ldlm_completion_ast,
 850                                     NULL, NULL, 0, NULL, child_lockh);
 851         if (rc != ELDLM_OK)
 852                 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
 853         else if (child_lockh == &lockh)
 854                 ldlm_lock_decref(child_lockh, LCK_EX);
 855
 856         RETURN(rc);
 857 }
 858
 859 int mds_open(struct mds_update_record *rec, int offset,
 860              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 861 {
 862         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 863         struct obd_device *obd = req->rq_export->exp_obd;
 864         struct mds_obd *mds = mds_req2mds(req);
 865         struct ldlm_reply *rep = NULL;
 866         struct mds_body *body = NULL;
 867         struct dentry *dchild = NULL, *dparent = NULL;
 868         struct mds_export_data *med;
 869         struct lustre_handle parent_lockh;
 870         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 871         int parent_mode = LCK_CR;
 872         void *handle = NULL;
 873         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
 874         unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
 875         unsigned int qpids[MAXQUOTAS] = { 0, 0 };
 876         int child_mode = LCK_CR;
 877         /* Always returning LOOKUP lock if open succesful to guard
 878            dentry on client. */
 879         ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
 880         struct ldlm_res_id child_res_id = { .name = {0}};
 881         int lock_flags = 0;
 882         ENTRY;
 883
 884         mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
 885
 886         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 887                          (obd_timeout + 1) / 4);
 888
 889         CLASSERT(MAXQUOTAS < 4);
 890         if (offset == DLM_INTENT_REC_OFF) { /* intent */
 891                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 892                                      sizeof(*rep));
 893                 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
 894                                       sizeof(*body));
 895         } else if (offset == REQ_REC_OFF) { /* non-intent reint */
 896                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 897                                       sizeof(*body));
 898                 LBUG(); /* XXX: not supported yet? */
 899         } else {
 900                 body = NULL;
 901                 LBUG();
 902         }
 903
 904         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 905
 906         /* Step 0: If we are passed a fid, then we assume the client already
 907          * opened this file and is only replaying the RPC, so we open the
 908          * inode by fid (at some large expense in security). */
 909         /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
 910         if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
 911              (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
 912             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 913                 if (rec->ur_fid2->id == 0) {
 914                         struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
 915                         if (lock) {
 916                                 LDLM_ERROR(lock, "fid2 not set on open replay");
 917                                 LDLM_LOCK_PUT(lock);
 918                         }
 919                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 920                         RETURN(-EFAULT);
 921                 }
 922
 923                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 924                                      rec, rep);
 925                 if (rc != -ENOENT) {
 926                         if (req->rq_export->exp_libclient &&
 927                             rec->ur_flags & MDS_OPEN_HAS_EA)
 928                                 RETURN(0);
 929
 930                         RETURN(rc);
 931                 }
 932
 933                 /* We didn't find the correct inode on disk either, so we
 934                  * need to re-create it via a regular replay. */
 935                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 936                         DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
 937                         RETURN(-EFAULT);
 938                 }
 939         } else if (rec->ur_fid2->id) {
 940                 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
 941                           rec->ur_fid2->id, rec->ur_fid2->generation);
 942                 RETURN(-EFAULT);
 943         }
 944
 945         /* If we got here, we must be called via intent */
 946         LASSERT(offset == DLM_INTENT_REC_OFF);
 947
 948         med = &req->rq_export->exp_mds_data;
 949         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 950                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 951                 RETURN(-ENOMEM);
 952         }
 953
 954         /* Step 1: Find and lock the parent */
 955         if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
 956                 parent_mode = LCK_EX;
 957         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 958                                         &parent_lockh, MDS_INODELOCK_UPDATE);
 959         if (IS_ERR(dparent)) {
 960                 rc = PTR_ERR(dparent);
 961                 if (rc != -ENOENT) {
 962                         CERROR("parent "LPU64"/%u lookup error %d\n",
 963                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 964                 } else {
 965                         /* Just cannot find parent - make it look like
 966                          * usual negative lookup to avoid extra MDS RPC */
 967                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 968                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
 969                 }
 970                 GOTO(cleanup, rc);
 971         }
 972         LASSERT(dparent->d_inode != NULL);
 973
 974         cleanup_phase = 1; /* parent dentry and lock */
 975
 976         if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 977                 dchild = dget(dparent);
 978                 cleanup_phase = 2; /* child dentry */
 979                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
 980                 GOTO(found_child, rc);
 981         }
 982
 983         /* Step 2: Lookup the child */
 984
 985         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
 986             (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) {
 987                 /* hack for nfsd with no_subtree_check, it will use anon
 988                  * dentry w/o filename to open the file. the anon dentry's
 989                  * parent was set to itself, so rec->ur_fid1 is the file.
 990                  * And in MDC it cannot derive the dentry's parent dentry,
 991                  * hence the file's name, so we hack here in MDS,
 992                  * refer to bug 13030. */
 993                 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
 994         } else {
 995                 dchild = ll_lookup_one_len(rec->ur_name, dparent,
 996                                            rec->ur_namelen - 1);
 997         }
 998         if (IS_ERR(dchild)) {
 999                 rc = PTR_ERR(dchild);
1000                 dchild = NULL; /* don't confuse mds_finish_transno() below */
1001                 GOTO(cleanup, rc);
1002         }
1003
1004         cleanup_phase = 2; /* child dentry */
1005
1006         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1007         if (dchild->d_inode)
1008                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1009         else
1010                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1011
1012         /*Step 3: If the child was negative, and we're supposed to, create it.*/
1013         if (dchild->d_inode == NULL) {
1014                 unsigned long ino = rec->ur_fid2->id;
1015                 struct iattr iattr;
1016                 struct inode *inode;
1017
1018                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1019                         /* It's negative and we weren't supposed to create it */
1020                         GOTO(cleanup, rc = -ENOENT);
1021                 }
1022
1023                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1024                         GOTO(cleanup, rc = -EROFS);
1025
1026                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1027                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1028                                       NULL);
1029                 if (IS_ERR(handle)) {
1030                         rc = PTR_ERR(handle);
1031                         handle = NULL;
1032                         GOTO(cleanup, rc);
1033                 }
1034                 dchild->d_fsdata = (void *) &dp;
1035                 dp.ldp_ptr = req;
1036                 dp.ldp_inum = ino;
1037
1038                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1039                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1040                         dchild->d_fsdata = NULL;
1041
1042                 if (rc) {
1043                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1044                         GOTO(cleanup, rc);
1045                 }
1046                 inode = dchild->d_inode;
1047                 if (ino) {
1048                         LASSERT(ino == inode->i_ino);
1049                         /* Written as part of setattr */
1050                         inode->i_generation = rec->ur_fid2->generation;
1051                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1052                                inode->i_ino, inode->i_generation);
1053                 }
1054
1055                 created = 1;
1056                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1057                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1058                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1059
1060                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
1061                 if (dparent->d_inode->i_mode & S_ISGID)
1062                         iattr.ia_gid = dparent->d_inode->i_gid;
1063                 else
1064                         iattr.ia_gid = current->fsgid;
1065
1066                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1067                         ATTR_MTIME | ATTR_CTIME;
1068
1069                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1070                 if (rc)
1071                         CERROR("error on child setattr: rc = %d\n", rc);
1072
1073                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1074
1075                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1076                 if (rc)
1077                         CERROR("error on parent setattr: rc = %d\n", rc);
1078
1079                 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1080                 handle = NULL;
1081                 acc_mode = 0;           /* Don't check for permissions */
1082         } else {
1083                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1084         }
1085
1086         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1087                  "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1088                  dchild->d_name.name, dchild, dchild->d_inode,
1089                  dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1090
1091 found_child:
1092         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1093         mds_pack_inode2body(body, dchild->d_inode);
1094
1095         if (S_ISREG(dchild->d_inode->i_mode)) {
1096                 /* Check permissions etc */
1097                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1098                 if (rc != 0)
1099                         GOTO(cleanup, rc);
1100
1101                 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1102                     (acc_mode & MAY_WRITE))
1103                         GOTO(cleanup, rc = -EROFS);
1104
1105                 /* An append-only file must be opened in append mode for
1106                  * writing */
1107                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1108                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1109                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1110                         GOTO(cleanup, rc = -EPERM);
1111         }
1112
1113         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1114             (rec->ur_flags & MDS_OPEN_EXCL)) {
1115                 /* File already exists, we didn't just create it, and we
1116                  * were passed O_EXCL; err-or. */
1117                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1118         }
1119
1120         /* if we are following a symlink, don't open */
1121         if (S_ISLNK(dchild->d_inode->i_mode))
1122                 GOTO(cleanup_no_trans, rc = 0);
1123
1124         if (S_ISDIR(dchild->d_inode->i_mode)) {
1125                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1126                     rec->ur_flags & FMODE_WRITE) {
1127                         /* we are trying to create or write a exist dir */
1128                         GOTO(cleanup, rc = -EISDIR);
1129                 }
1130                 if (rec->ur_flags & MDS_FMODE_EXEC) {
1131                         /* we are trying to exec a directory */
1132                         GOTO(cleanup, rc = -EACCES);
1133                 }
1134                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1135                         intent_set_disposition(rep, DISP_OPEN_OPEN);
1136                         GOTO(cleanup, rc = -EACCES);
1137                 }
1138         } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1139                 GOTO(cleanup, rc = -ENOTDIR);
1140         }
1141
1142         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1143                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1144                 GOTO(cleanup, rc = -EAGAIN);
1145         }
1146
1147         /* Obtain OPEN lock as well */
1148         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1149
1150         /* We cannot use acc_mode here, because it is zeroed in case of
1151            creating a file, so we get wrong lockmode */
1152         if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1153                 child_mode = LCK_CW;
1154         else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1155                 child_mode = LCK_PR;
1156         else
1157                 child_mode = LCK_CR;
1158
1159         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1160              (rec->ur_flags & MDS_OPEN_LOCK)) {
1161                 /* In case of replay we do not get a lock assuming that the
1162                    caller has it already */
1163                 child_res_id.name[0] = dchild->d_inode->i_ino;
1164                 child_res_id.name[1] = dchild->d_inode->i_generation;
1165
1166                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1167                                             LDLM_IBITS, &policy, child_mode,
1168                                             &lock_flags, ldlm_blocking_ast,
1169                                             ldlm_completion_ast, NULL, NULL,
1170                                             0, NULL, child_lockh);
1171                 if (rc != ELDLM_OK)
1172                         GOTO(cleanup, rc);
1173
1174                 /* Let mds_intent_policy know that we have a lock to return */
1175                 intent_set_disposition(rep, DISP_OPEN_LOCK);
1176                 cleanup_phase = 3;
1177         }
1178
1179         if (!S_ISREG(dchild->d_inode->i_mode) &&
1180             !S_ISDIR(dchild->d_inode->i_mode) &&
1181             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1182                 /* If client supports this, do not return open handle for
1183                  * special device nodes */
1184                 GOTO(cleanup_no_trans, rc = 0);
1185         }
1186
1187         /* Step 5: mds_open it */
1188         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1189                              rep, &parent_lockh);
1190         GOTO(cleanup, rc);
1191
1192  cleanup:
1193         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1194                                 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
1195
1196  cleanup_no_trans:
1197         switch (cleanup_phase) {
1198         case 3:
1199                 if (rc)
1200                         /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1201                          * mds_intent_policy won't try to return any locks */
1202                         ldlm_lock_decref(child_lockh, child_mode);
1203         case 2:
1204                 if (rc && created) {
1205                         int err = vfs_unlink(dparent->d_inode, dchild);
1206                         if (err) {
1207                                 CERROR("unlink(%.*s) in error path: %d\n",
1208                                        dchild->d_name.len, dchild->d_name.name,
1209                                        err);
1210                         }
1211                 } else if (created) {
1212                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1213                         /* save uid/gid for quota acquire/release */
1214                         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1215                         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1216                 }
1217                 l_dput(dchild);
1218         case 1:
1219                 if (dparent == NULL)
1220                         break;
1221
1222                 l_dput(dparent);
1223                 if (rc)
1224                         ldlm_lock_decref(&parent_lockh, parent_mode);
1225                 else
1226                         ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1227         }
1228         /* trigger dqacq on the owner of child and parent */
1229         lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1230                       FSFILT_OP_CREATE);
1231
1232         /* If we have not taken the "open" lock, we may not return 0 here,
1233            because caller expects 0 to mean "lock is taken", and it needs
1234            nonzero return here for caller to return EDLM_LOCK_ABORTED to
1235            client. Later caller should rewrite the return value back to zero
1236            if it to be used any further
1237          */
1238         RETURN(rc);
1239 }
1240
1241 /* Close a "file descriptor" and possibly unlink an orphan from the
1242  * PENDING directory.  Caller must hold child->i_mutex, this drops it.
1243  *
1244  * If we are being called from mds_disconnect() because the client has
1245  * disappeared, then req == NULL and we do not update last_rcvd because
1246  * there is nothing that could be recovered by the client at this stage
1247  * (it will not even _have_ an entry in last_rcvd anymore).
1248  *
1249  * Returns EAGAIN if the client needs to get more data and re-close. */
1250 int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd,
1251                   struct mds_file_data *mfd, int unlink_orphan,
1252                   struct lov_mds_md *lmm, int lmm_size,
1253                   struct llog_cookie *logcookies, int cookies_size,
1254                   __u64 *valid)
1255 {
1256         struct inode *inode = mfd->mfd_dentry->d_inode;
1257         char fidname[LL_FID_NAMELEN];
1258         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1259         struct dentry *pending_child = NULL;
1260         struct mds_obd *mds = &obd->u.mds;
1261         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1262         void *handle = NULL;
1263         struct mds_body *request_body = NULL, *reply_body = NULL;
1264         struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1265         struct iattr iattr = { 0 };
1266         ENTRY;
1267
1268         if (req && req->rq_reqmsg != NULL)
1269                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1270                                               sizeof(*request_body));
1271         if (req && req->rq_repmsg != NULL)
1272                 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1273                                             sizeof(*reply_body));
1274
1275         fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1276
1277         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1278                inode->i_nlink, mds_orphan_open_count(inode));
1279
1280         last_orphan = mds_orphan_open_dec_test(inode) &&
1281                 mds_inode_is_orphan(inode);
1282         MDS_UP_WRITE_ORPHAN_SEM(inode);
1283
1284         /* this is half of the actual "close" */
1285         if (mfd->mfd_mode & FMODE_WRITE) {
1286                 rc = mds_put_write_access(mds, inode, request_body,
1287                                           last_orphan && unlink_orphan);
1288         } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1289                 mds_allow_write_access(inode);
1290         }
1291
1292         if (last_orphan && unlink_orphan) {
1293                 int stripe_count = 0;
1294                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1295
1296                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1297
1298                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1299                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1300                         CERROR("found \"orphan\" %s %s with link count %d\n",
1301                                S_ISREG(inode->i_mode) ? "file" : "dir",
1302                                fidname, inode->i_nlink);
1303
1304                 /* Sadly, there is no easy way to save pending_child from
1305                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1306                  * but normally it will still be in the dcache. */
1307                 LOCK_INODE_MUTEX(pending_dir);
1308                 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1309                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1310                                                fidlen);
1311                 if (IS_ERR(pending_child))
1312                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1313                 LASSERT(pending_child->d_inode != NULL);
1314
1315                 cleanup_phase = 2; /* dput(pending_child) when finished */
1316                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1317                         rc = vfs_rmdir(pending_dir, pending_child);
1318                         if (rc)
1319                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1320                                        fidname,rc);
1321                         goto out;
1322                 }
1323
1324                 if (lmm != NULL) {
1325                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1326                 }
1327
1328                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1329                                           NULL, stripe_count);
1330                 if (IS_ERR(handle)) {
1331                         rc = PTR_ERR(handle);
1332                         handle = NULL;
1333                         GOTO(cleanup, rc);
1334                 }
1335
1336                 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1337                     mds_log_op_unlink(obd, lmm, lmm_size,
1338                                       logcookies, cookies_size) > 0) {
1339                         *valid |= OBD_MD_FLCOOKIE;
1340                 }
1341
1342                 dp.ldp_inum = 0;
1343                 dp.ldp_ptr = req;
1344                 pending_child->d_fsdata = (void *) &dp;
1345                 rc = vfs_unlink(pending_dir, pending_child);
1346                 if (rc)
1347                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1348
1349                 goto out; /* Don't bother updating attrs on unlinked inode */
1350         }
1351
1352 #if 0
1353         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1354                 /* Update the on-disk attributes if this was the last write
1355                  * close, and all information was provided (i.e., rc == 0)
1356                  *
1357                  * XXX this should probably be abstracted with mds_reint_setattr
1358                  */
1359
1360                 if (request_body->valid & OBD_MD_FLMTIME &&
1361                     LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1362                         LTIME_S(iattr.ia_mtime) = request_body->mtime;
1363                         iattr.ia_valid |= ATTR_MTIME;
1364                 }
1365                 if (request_body->valid & OBD_MD_FLCTIME &&
1366                     LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1367                         LTIME_S(iattr.ia_ctime) = request_body->ctime;
1368                         iattr.ia_valid |= ATTR_CTIME;
1369                 }
1370
1371                 /* XXX can't set block count with fsfilt_setattr (!) */
1372                 if (request_body->valid & OBD_MD_FLSIZE) {
1373                         iattr.ia_valid |= ATTR_SIZE;
1374                         iattr.ia_size = request_body->size;
1375                 }
1376                 /* iattr.ia_blocks = request_body->blocks */
1377
1378         }
1379 #endif
1380         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1381                 /* Only start a transaction to write out only the atime if
1382                  * it is more out-of-date than the specified limit.  If we
1383                  * are already going to write out the atime then do it anyway.
1384                  * */
1385                 LTIME_S(iattr.ia_atime) = request_body->atime;
1386                 if ((LTIME_S(iattr.ia_atime) >
1387                      LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1388                     (iattr.ia_valid != 0 &&
1389                      LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1390                         iattr.ia_valid |= ATTR_ATIME;
1391         }
1392
1393         if (iattr.ia_valid != 0) {
1394                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1395                 if (IS_ERR(handle)) {
1396                         rc = PTR_ERR(handle);
1397                         handle = NULL;
1398                         GOTO(cleanup, rc);
1399                 }
1400                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1401                 if (rc)
1402                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1403         }
1404 out:
1405         /* If other clients have this file open for write, rc will be > 0 */
1406         if (rc > 0)
1407                 rc = 0;
1408         l_dput(mfd->mfd_dentry);
1409         mds_mfd_put(mfd);
1410
1411  cleanup:
1412         if (req != NULL && reply_body != NULL) {
1413                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0);
1414         } else if (handle) {
1415                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1416                 if (err) {
1417                         CERROR("error committing close: %d\n", err);
1418                         if (!rc)
1419                                 rc = err;
1420                 }
1421         }
1422
1423         switch (cleanup_phase) {
1424         case 2:
1425                 dput(pending_child);
1426         case 1:
1427                 UNLOCK_INODE_MUTEX(pending_dir);
1428         }
1429         RETURN(rc);
1430 }
1431
1432 int mds_close(struct ptlrpc_request *req, int offset)
1433 {
1434         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1435         struct obd_device *obd = req->rq_export->exp_obd;
1436         struct mds_body *body;
1437         struct mds_file_data *mfd;
1438         struct lvfs_run_ctxt saved;
1439         struct inode *inode;
1440         int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1441                                sizeof(struct mds_body),
1442                                obd->u.mds.mds_max_mdsize,
1443                                obd->u.mds.mds_max_cookiesize };
1444         struct mds_body *reply_body;
1445         struct lov_mds_md *lmm;
1446         int lmm_size;
1447         struct llog_cookie *logcookies;
1448         int cookies_size;
1449         ENTRY;
1450
1451         rc = lustre_pack_reply(req, 4, repsize, NULL);
1452         if (rc) {
1453                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1454                 req->rq_status = rc;
1455                 /* continue on to drop local open even if we can't send reply */
1456         } else {
1457                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1458         }
1459
1460         CDEBUG(D_HA, "close req->rep_len %d mdsize %d cookiesize %d\n",
1461                req->rq_replen,
1462                obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1463         mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1464
1465         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1466                                   lustre_swab_mds_body);
1467         if (body == NULL) {
1468                 CERROR("Can't unpack body\n");
1469                 req->rq_status = -EFAULT;
1470                 RETURN(-EFAULT);
1471         }
1472
1473         spin_lock(&med->med_open_lock);
1474         mfd = mds_handle2mfd(&body->handle);
1475         if (mfd == NULL) {
1476                 spin_unlock(&med->med_open_lock);
1477                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1478                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1479                 req->rq_status = -ESTALE;
1480                 RETURN(-ESTALE);
1481         }
1482         /* Remove mfd handle so it can't be found again.  We consume mfd_list
1483          * reference here, but still have mds_handle2mfd ref until mfd_close. */
1484         mds_mfd_unlink(mfd, 1);
1485         spin_unlock(&med->med_open_lock);
1486
1487         inode = mfd->mfd_dentry->d_inode;
1488         /* child orphan sem protects orphan_dec_test && is_orphan race */
1489         MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1490         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1491                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1492                                       sizeof(*body));
1493                 LASSERT(body != NULL);
1494
1495                 mds_pack_inode2fid(&body->fid1, inode);
1496                 mds_pack_inode2body(body, inode);
1497                 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1498                             MDS_PACK_MD_LOCK);
1499         }
1500
1501         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1502         reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1503         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1504         lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1505         logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1506         cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1507         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1508                                        lmm, lmm_size, logcookies, cookies_size,
1509                                        &reply_body->valid);
1510         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1511
1512         mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1513         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1514                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1515                 req->rq_status = -ENOMEM;
1516                 RETURN(-ENOMEM);
1517         }
1518
1519         RETURN(rc);
1520 }
1521
1522 int mds_done_writing(struct ptlrpc_request *req, int offset)
1523 {
1524         struct mds_body *body;
1525         int rc, size[2] = { sizeof(struct ptlrpc_body),
1526                             sizeof(struct mds_body) };
1527         ENTRY;
1528
1529         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1530
1531         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1532                                   lustre_swab_mds_body);
1533         if (body == NULL) {
1534                 CERROR("Can't unpack body\n");
1535                 req->rq_status = -EFAULT;
1536                 RETURN(-EFAULT);
1537         }
1538
1539         rc = lustre_pack_reply(req, 2, size, NULL);
1540         if (rc) {
1541                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1542                 req->rq_status = rc;
1543         }
1544
1545         RETURN(0);
1546 }