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                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 114                 OBD_FREE(mfd, sizeof *mfd);
 115         }
 116 }
 117
 118 /* Remove the mfd handle so that it cannot be found by open/close again.
 119  * Caller must hold med_open_lock for mfd_list manipulation. */
 120 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
 121 {
 122         class_handle_unhash(&mfd->mfd_handle);
 123         list_del_init(&mfd->mfd_list);
 124         if (decref)
 125                 mds_mfd_put(mfd);
 126 }
 127
 128 /* Caller must hold mds->mds_epoch_sem */
 129 static int mds_alloc_filterdata(struct inode *inode)
 130 {
 131         LASSERT(inode->i_filterdata == NULL);
 132         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 133         if (inode->i_filterdata == NULL)
 134                 return -ENOMEM;
 135         LASSERT(igrab(inode) == inode);
 136         return 0;
 137 }
 138
 139 /* Caller must hold mds->mds_epoch_sem */
 140 static void mds_free_filterdata(struct inode *inode)
 141 {
 142         LASSERT(inode->i_filterdata != NULL);
 143         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 144         inode->i_filterdata = NULL;
 145         iput(inode);
 146 }
 147
 148 /* Write access to a file: executors cause a negative count,
 149  * writers a positive count.  The semaphore is needed to perform
 150  * a check for the sign and then increment or decrement atomically.
 151  *
 152  * This code is closely tied to the allocation of the d_fsdata and the
 153  * MDS epoch, so we use the same semaphore for the whole lot.
 154  *
 155  * We could use a different semaphore for each file, if it ever shows
 156  * up in a profile, which it won't.
 157  *
 158  * epoch argument is nonzero during recovery */
 159 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 160                                 __u64 epoch)
 161 {
 162         int rc = 0;
 163
 164         down(&mds->mds_epoch_sem);
 165
 166         if (atomic_read(&inode->i_writecount) < 0) {
 167                 up(&mds->mds_epoch_sem);
 168                 RETURN(-ETXTBSY);
 169         }
 170
 171
 172         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 173                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 174                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 175                        inode->i_generation);
 176                 goto out;
 177         }
 178
 179         if (inode->i_filterdata == NULL)
 180                 mds_alloc_filterdata(inode);
 181         if (inode->i_filterdata == NULL) {
 182                 rc = -ENOMEM;
 183                 goto out;
 184         }
 185         if (epoch > mds->mds_io_epoch)
 186                 mds->mds_io_epoch = epoch;
 187         else
 188                 mds->mds_io_epoch++;
 189         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 190         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 191                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 192  out:
 193         if (rc == 0)
 194                 atomic_inc(&inode->i_writecount);
 195         up(&mds->mds_epoch_sem);
 196         return rc;
 197 }
 198
 199 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 200  * again -- which is never true if the file is about to be unlinked.  Otherwise
 201  * returns the number of remaining writers. */
 202 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 203                                 struct mds_body *body, int unlinking)
 204 {
 205         int rc = 0;
 206         ENTRY;
 207
 208         down(&mds->mds_epoch_sem);
 209         atomic_dec(&inode->i_writecount);
 210         rc = atomic_read(&inode->i_writecount);
 211         if (rc > 0)
 212                 GOTO(out, rc);
 213 #if 0
 214         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 215                 GOTO(out, rc = EAGAIN);
 216 #endif
 217         mds_free_filterdata(inode);
 218  out:
 219         up(&mds->mds_epoch_sem);
 220         return rc;
 221 }
 222
 223 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 224 {
 225         ENTRY;
 226         down(&mds->mds_epoch_sem);
 227         if (atomic_read(&inode->i_writecount) > 0) {
 228                 up(&mds->mds_epoch_sem);
 229                 RETURN(-ETXTBSY);
 230         }
 231         atomic_dec(&inode->i_writecount);
 232         up(&mds->mds_epoch_sem);
 233         RETURN(0);
 234 }
 235
 236 static void mds_allow_write_access(struct inode *inode)
 237 {
 238         ENTRY;
 239         atomic_inc(&inode->i_writecount);
 240 }
 241
 242 int mds_query_write_access(struct inode *inode)
 243 {
 244         ENTRY;
 245         RETURN(atomic_read(&inode->i_writecount));
 246 }
 247
 248 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 249 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 250                                              struct vfsmount *mnt, int flags,
 251                                              struct ptlrpc_request *req)
 252 {
 253         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 254         struct mds_obd *mds = mds_req2mds(req);
 255         struct mds_file_data *mfd;
 256         struct mds_body *body;
 257         int error;
 258         ENTRY;
 259
 260         mfd = mds_mfd_new();
 261         if (mfd == NULL) {
 262                 CERROR("mds: out of memory\n");
 263                 GOTO(cleanup_dentry, error = -ENOMEM);
 264         }
 265
 266         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 267
 268         if (flags & FMODE_WRITE) {
 269                 /* FIXME: in recovery, need to pass old epoch here */
 270                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 271                 if (error)
 272                         GOTO(cleanup_mfd, error);
 273                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 274         } else if (flags & MDS_FMODE_EXEC) {
 275                 error = mds_deny_write_access(mds, dentry->d_inode);
 276                 if (error)
 277                         GOTO(cleanup_mfd, error);
 278         }
 279
 280         dget(dentry);
 281
 282         /* Mark the file as open to handle open-unlink. */
 283         MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
 284         mds_orphan_open_inc(dentry->d_inode);
 285         MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
 286
 287         mfd->mfd_mode = flags;
 288         mfd->mfd_dentry = dentry;
 289         mfd->mfd_xid = req->rq_xid;
 290
 291         spin_lock(&med->med_open_lock);
 292         list_add(&mfd->mfd_list, &med->med_open_head);
 293         spin_unlock(&med->med_open_lock);
 294
 295         body->handle.cookie = mfd->mfd_handle.h_cookie;
 296
 297         RETURN(mfd);
 298
 299 cleanup_mfd:
 300         mds_mfd_put(mfd);
 301         mds_mfd_unlink(mfd, 1);
 302 cleanup_dentry:
 303         return ERR_PTR(error);
 304 }
 305
 306 /* Must be called with i_mutex held */
 307 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 308                               struct mds_update_record *rec,
 309                               struct mds_obd *mds, struct obd_device *obd,
 310                               struct dentry *dchild, void **handle,
 311                               obd_id **ids)
 312 {
 313         struct inode *inode = dchild->d_inode;
 314         struct obd_trans_info oti = { 0 };
 315         struct lov_mds_md *lmm = NULL;
 316         int rc, lmm_size;
 317         struct mds_body *body;
 318         struct obd_info oinfo = { { { 0 } } };
 319         void *lmm_buf;
 320         ENTRY;
 321
 322         if (!S_ISREG(inode->i_mode))
 323                 RETURN(0);
 324         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 325             !(rec->ur_flags & FMODE_WRITE))
 326                 RETURN(0);
 327
 328         body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
 329
 330         if (body->valid & OBD_MD_FLEASIZE)
 331                 RETURN(0);
 332
 333         OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 334         if (*ids == NULL)
 335                 RETURN(-ENOMEM);
 336         oti_init(&oti, req);
 337         oti.oti_objid = *ids;
 338
 339         /* replay case */
 340         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 341                 if (rec->ur_fid2->id == 0) {
 342                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 343                         RETURN(-EFAULT);
 344                 }
 345
 346                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 347                 lmm_size = rec->ur_eadatalen;
 348                 lmm = rec->ur_eadata;
 349                 LASSERT(lmm);
 350
 351                 if (*handle == NULL)
 352                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 353                 if (IS_ERR(*handle)) {
 354                         rc = PTR_ERR(*handle);
 355                         *handle = NULL;
 356                         GOTO(out_ids, rc);
 357                 }
 358
 359                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 360
 361                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
 362                 if (rc)
 363                         CERROR("open replay failed to set md:%d\n", rc);
 364                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
 365                 LASSERT(lmm_buf);
 366                 memcpy(lmm_buf, lmm, lmm_size);
 367
 368                 RETURN(rc);
 369         }
 370
 371         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 372                 GOTO(out_ids, rc = -ENOMEM);
 373
 374         oinfo.oi_oa = obdo_alloc();
 375         if (oinfo.oi_oa == NULL)
 376                 GOTO(out_ids, rc = -ENOMEM);
 377         oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
 378         oinfo.oi_oa->o_gid = 0;
 379         oinfo.oi_oa->o_mode = S_IFREG | 0600;
 380         oinfo.oi_oa->o_id = inode->i_ino;
 381         oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
 382                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
 383         oinfo.oi_oa->o_size = 0;
 384
 385         obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 386                         OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 387
 388         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 389                 /* check if things like lfs setstripe are sending us the ea */
 390                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 391                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 392                                            mds->mds_osc_exp,
 393                                            0, &oinfo.oi_md, rec->ur_eadata);
 394                         if (rc)
 395                                 GOTO(out_oa, rc);
 396                 } else {
 397                         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 398                         if (lmm == NULL)
 399                                 GOTO(out_oa, rc = -ENOMEM);
 400
 401                         lmm_size = mds->mds_max_mdsize;
 402                         rc = mds_get_md(obd, dchild->d_parent->d_inode,
 403                                         lmm, &lmm_size, 1);
 404                         if (rc > 0)
 405                                 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 406                                                    mds->mds_osc_exp,
 407                                                    0, &oinfo.oi_md, lmm);
 408                         OBD_FREE(lmm, mds->mds_max_mdsize);
 409                         if (rc)
 410                                 GOTO(out_oa, rc);
 411                 }
 412                 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
 413                                 &oinfo.oi_md, &oti);
 414                 if (rc) {
 415                         int level = D_ERROR;
 416                         if (rc == -ENOSPC)
 417                                 level = D_INODE;
 418                         CDEBUG(level, "error creating objects for "
 419                                       "inode %lu: rc = %d\n",
 420                                inode->i_ino, rc);
 421                         if (rc > 0) {
 422                                 CERROR("obd_create returned invalid "
 423                                        "rc %d\n", rc);
 424                                 rc = -EIO;
 425                         }
 426                         GOTO(out_oa, rc);
 427                 }
 428         } else {
 429                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
 430                                    0, &oinfo.oi_md, rec->ur_eadata);
 431                 if (rc) {
 432                         GOTO(out_oa, rc);
 433                 }
 434                 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
 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(out_oa, 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         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 = mds_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);
 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 dentry_params dp;
 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         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
 884                          (obd_timeout + 1) / 4);
 885
 886         CLASSERT(MAXQUOTAS < 4);
 887         if (offset == DLM_INTENT_REC_OFF) { /* intent */
 888                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 889                                      sizeof(*rep));
 890                 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
 891                                       sizeof(*body));
 892         } else if (offset == REQ_REC_OFF) { /* non-intent reint */
 893                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 894                                       sizeof(*body));
 895                 LBUG(); /* XXX: not supported yet? */
 896         } else {
 897                 body = NULL;
 898                 LBUG();
 899         }
 900
 901         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 902
 903         /* Step 0: If we are passed a fid, then we assume the client already
 904          * opened this file and is only replaying the RPC, so we open the
 905          * inode by fid (at some large expense in security). */
 906         /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
 907         if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
 908              (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
 909             !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
 910                 if (rec->ur_fid2->id == 0) {
 911                         struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
 912                         if (lock) {
 913                                 LDLM_ERROR(lock, "fid2 not set on open replay");
 914                                 LDLM_LOCK_PUT(lock);
 915                         }
 916                         DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
 917                         RETURN(-EFAULT);
 918                 }
 919
 920                 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
 921                                      rec, rep);
 922                 if (rc != -ENOENT) {
 923                         if (req->rq_export->exp_libclient &&
 924                             rec->ur_flags & MDS_OPEN_HAS_EA)
 925                                 RETURN(ENOLCK);
 926
 927                         RETURN(rc);
 928                 }
 929
 930                 /* We didn't find the correct inode on disk either, so we
 931                  * need to re-create it via a regular replay. */
 932                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
 933                         DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
 934                         RETURN(-EFAULT);
 935                 }
 936         } else if (rec->ur_fid2->id) {
 937                 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
 938                           rec->ur_fid2->id, rec->ur_fid2->generation);
 939                 RETURN(-EFAULT);
 940         }
 941
 942         /* If we got here, we must be called via intent */
 943         LASSERT(offset == DLM_INTENT_REC_OFF);
 944
 945         med = &req->rq_export->exp_mds_data;
 946         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 947                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 948                 RETURN(-ENOMEM);
 949         }
 950
 951         /* Step 1: Find and lock the parent */
 952         if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
 953                 parent_mode = LCK_EX;
 954         dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
 955                                         &parent_lockh, MDS_INODELOCK_UPDATE);
 956         if (IS_ERR(dparent)) {
 957                 rc = PTR_ERR(dparent);
 958                 if (rc != -ENOENT) {
 959                         CERROR("parent "LPU64"/%u lookup error %d\n",
 960                                rec->ur_fid1->id, rec->ur_fid1->generation, rc);
 961                 } else {
 962                         /* Just cannot find parent - make it look like
 963                          * usual negative lookup to avoid extra MDS RPC */
 964                         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 965                         intent_set_disposition(rep, DISP_LOOKUP_NEG);
 966                 }
 967                 GOTO(cleanup, rc);
 968         }
 969         LASSERT(dparent->d_inode != NULL);
 970
 971         cleanup_phase = 1; /* parent dentry and lock */
 972
 973         if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
 974                 dchild = dget(dparent);
 975                 cleanup_phase = 2; /* child dentry */
 976                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
 977                 GOTO(found_child, rc);
 978         }
 979
 980         /* Step 2: Lookup the child */
 981         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 982         if (IS_ERR(dchild)) {
 983                 rc = PTR_ERR(dchild);
 984                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 985                 GOTO(cleanup, rc);
 986         }
 987
 988         cleanup_phase = 2; /* child dentry */
 989
 990         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 991         if (dchild->d_inode)
 992                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 993         else
 994                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 995
 996         /*Step 3: If the child was negative, and we're supposed to, create it.*/
 997         if (dchild->d_inode == NULL) {
 998                 unsigned long ino = rec->ur_fid2->id;
 999                 struct iattr iattr;
1000                 struct inode *inode;
1001
1002                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1003                         /* It's negative and we weren't supposed to create it */
1004                         GOTO(cleanup, rc = -ENOENT);
1005                 }
1006
1007                 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1008                         GOTO(cleanup, rc = -EROFS);
1009
1010                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1011                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1012                                       NULL);
1013                 if (IS_ERR(handle)) {
1014                         rc = PTR_ERR(handle);
1015                         handle = NULL;
1016                         GOTO(cleanup, rc);
1017                 }
1018                 dchild->d_fsdata = (void *) &dp;
1019                 dp.p_ptr = req;
1020                 dp.p_inum = ino;
1021
1022                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1023                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1024                         dchild->d_fsdata = NULL;
1025
1026                 if (rc) {
1027                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1028                         GOTO(cleanup, rc);
1029                 }
1030                 inode = dchild->d_inode;
1031                 if (ino) {
1032                         LASSERT(ino == inode->i_ino);
1033                         /* Written as part of setattr */
1034                         inode->i_generation = rec->ur_fid2->generation;
1035                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1036                                inode->i_ino, inode->i_generation);
1037                 }
1038
1039                 created = 1;
1040                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1041                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1042                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1043
1044                 iattr.ia_uid = current->fsuid;  /* set by push_ctxt already */
1045                 if (dparent->d_inode->i_mode & S_ISGID)
1046                         iattr.ia_gid = dparent->d_inode->i_gid;
1047                 else
1048                         iattr.ia_gid = current->fsgid;
1049
1050                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1051                         ATTR_MTIME | ATTR_CTIME;
1052
1053                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1054                 if (rc)
1055                         CERROR("error on child setattr: rc = %d\n", rc);
1056
1057                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1058
1059                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1060                 if (rc)
1061                         CERROR("error on parent setattr: rc = %d\n", rc);
1062
1063                 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1064                 handle = NULL;
1065                 acc_mode = 0;           /* Don't check for permissions */
1066         } else {
1067                 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1068         }
1069
1070         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1071                  "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1072                  dchild->d_name.name, dchild, dchild->d_inode,
1073                  dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1074
1075 found_child:
1076         mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1077         mds_pack_inode2body(body, dchild->d_inode);
1078
1079         if (S_ISREG(dchild->d_inode->i_mode)) {
1080                 /* Check permissions etc */
1081                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1082                 if (rc != 0)
1083                         GOTO(cleanup, rc);
1084
1085                 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1086                     (acc_mode & MAY_WRITE))
1087                         GOTO(cleanup, rc = -EROFS);
1088
1089                 /* An append-only file must be opened in append mode for
1090                  * writing */
1091                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1092                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1093                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1094                         GOTO(cleanup, rc = -EPERM);
1095         }
1096
1097         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1098             (rec->ur_flags & MDS_OPEN_EXCL)) {
1099                 /* File already exists, we didn't just create it, and we
1100                  * were passed O_EXCL; err-or. */
1101                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1102         }
1103
1104         /* if we are following a symlink, don't open */
1105         if (S_ISLNK(dchild->d_inode->i_mode))
1106                 GOTO(cleanup_no_trans, rc = 0);
1107
1108         if (S_ISDIR(dchild->d_inode->i_mode)) {
1109                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1110                     rec->ur_flags & FMODE_WRITE) {
1111                         /* we are trying to create or write a exist dir */
1112                         GOTO(cleanup, rc = -EISDIR);
1113                 }
1114                 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1115                         intent_set_disposition(rep, DISP_OPEN_OPEN);
1116                         GOTO(cleanup, rc = -EACCES);
1117                 }
1118         } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1119                 GOTO(cleanup, rc = -ENOTDIR);
1120         }
1121
1122         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1123                 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1124                 GOTO(cleanup, rc = -EAGAIN);
1125         }
1126
1127         /* Obtain OPEN lock as well */
1128         policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1129
1130         /* We cannot use acc_mode here, because it is zeroed in case of
1131            creating a file, so we get wrong lockmode */
1132         if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1133                 child_mode = LCK_CW;
1134         else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1135                 child_mode = LCK_PR;
1136         else
1137                 child_mode = LCK_CR;
1138
1139         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1140              (rec->ur_flags & MDS_OPEN_LOCK)) {
1141                 /* In case of replay we do not get a lock assuming that the
1142                    caller has it already */
1143                 child_res_id.name[0] = dchild->d_inode->i_ino;
1144                 child_res_id.name[1] = dchild->d_inode->i_generation;
1145
1146                 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1147                                             LDLM_IBITS, &policy, child_mode,
1148                                             &lock_flags, ldlm_blocking_ast,
1149                                             ldlm_completion_ast, NULL, NULL,
1150                                             0, NULL, child_lockh);
1151                 if (rc != ELDLM_OK)
1152                         GOTO(cleanup, rc);
1153
1154                 cleanup_phase = 3;
1155         }
1156
1157         if (!S_ISREG(dchild->d_inode->i_mode) &&
1158             !S_ISDIR(dchild->d_inode->i_mode) &&
1159             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1160                 /* If client supports this, do not return open handle for
1161                  * special device nodes */
1162                 GOTO(cleanup_no_trans, rc = 0);
1163         }
1164
1165         /* Step 5: mds_open it */
1166         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1167                              rep, &parent_lockh);
1168         GOTO(cleanup, rc);
1169
1170  cleanup:
1171         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1172                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1173
1174  cleanup_no_trans:
1175         switch (cleanup_phase) {
1176         case 3:
1177                 if (rc)
1178                         ldlm_lock_decref(child_lockh, child_mode);
1179         case 2:
1180                 if (rc && created) {
1181                         int err = vfs_unlink(dparent->d_inode, dchild);
1182                         if (err) {
1183                                 CERROR("unlink(%.*s) in error path: %d\n",
1184                                        dchild->d_name.len, dchild->d_name.name,
1185                                        err);
1186                         }
1187                 } else if (created) {
1188                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1189                         /* save uid/gid for quota acquire/release */
1190                         qpids[USRQUOTA] = dparent->d_inode->i_uid;
1191                         qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1192                 }
1193                 l_dput(dchild);
1194         case 1:
1195                 if (dparent == NULL)
1196                         break;
1197
1198                 l_dput(dparent);
1199                 if (rc)
1200                         ldlm_lock_decref(&parent_lockh, parent_mode);
1201                 else
1202                         ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1203         }
1204         /* If we have not taken the "open" lock, we may not return 0 here,
1205            because caller expects 0 to mean "lock is taken", and it needs
1206            nonzero return here for caller to return EDLM_LOCK_ABORTED to
1207            client. Later caller should rewrite the return value back to zero
1208            if it to be used any further
1209          */
1210         if ((cleanup_phase != 3) && !rc)
1211                 rc = ENOLCK;
1212
1213         /* trigger dqacq on the owner of child and parent */
1214         lquota_adjust(quota_interface, obd, qcids, qpids, rc, FSFILT_OP_CREATE);
1215         RETURN(rc);
1216 }
1217
1218 /* Close a "file descriptor" and possibly unlink an orphan from the
1219  * PENDING directory.  Caller must hold child->i_mutex, this drops it.
1220  *
1221  * If we are being called from mds_disconnect() because the client has
1222  * disappeared, then req == NULL and we do not update last_rcvd because
1223  * there is nothing that could be recovered by the client at this stage
1224  * (it will not even _have_ an entry in last_rcvd anymore).
1225  *
1226  * Returns EAGAIN if the client needs to get more data and re-close. */
1227 int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd,
1228                   struct mds_file_data *mfd, int unlink_orphan)
1229 {
1230         struct inode *inode = mfd->mfd_dentry->d_inode;
1231         char fidname[LL_FID_NAMELEN];
1232         int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1233         struct dentry *pending_child = NULL;
1234         struct mds_obd *mds = &obd->u.mds;
1235         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1236         void *handle = NULL;
1237         struct mds_body *request_body = NULL, *reply_body = NULL;
1238         struct dentry_params dp;
1239         struct iattr iattr = { 0 };
1240         ENTRY;
1241
1242         if (req && req->rq_reqmsg != NULL)
1243                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1244                                               sizeof(*request_body));
1245         if (req && req->rq_repmsg != NULL)
1246                 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1247                                             sizeof(*reply_body));
1248
1249         fidlen = mds_fid2str(fidname, inode->i_ino, inode->i_generation);
1250
1251         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1252                inode->i_nlink, mds_orphan_open_count(inode));
1253
1254         last_orphan = mds_orphan_open_dec_test(inode) &&
1255                 mds_inode_is_orphan(inode);
1256         MDS_UP_WRITE_ORPHAN_SEM(inode);
1257
1258         /* this is half of the actual "close" */
1259         if (mfd->mfd_mode & FMODE_WRITE) {
1260                 rc = mds_put_write_access(mds, inode, request_body,
1261                                           last_orphan && unlink_orphan);
1262         } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1263                 mds_allow_write_access(inode);
1264         }
1265
1266         if (last_orphan && unlink_orphan) {
1267                 struct lov_mds_md *lmm = NULL;
1268                 int stripe_count = 0;
1269                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1270
1271                 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1272
1273                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1274                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1275                         CERROR("found \"orphan\" %s %s with link count %d\n",
1276                                S_ISREG(inode->i_mode) ? "file" : "dir",
1277                                fidname, inode->i_nlink);
1278
1279                 /* Sadly, there is no easy way to save pending_child from
1280                  * mds_reint_unlink() into mfd, so we need to re-lookup,
1281                  * but normally it will still be in the dcache. */
1282                 LOCK_INODE_MUTEX(pending_dir);
1283                 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1284                 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1285                                                fidlen);
1286                 if (IS_ERR(pending_child))
1287                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1288                 LASSERT(pending_child->d_inode != NULL);
1289
1290                 cleanup_phase = 2; /* dput(pending_child) when finished */
1291                 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1292                         rc = vfs_rmdir(pending_dir, pending_child);
1293                         if (rc)
1294                                 CERROR("error unlinking orphan dir %s: rc %d\n",
1295                                        fidname,rc);
1296                         goto out;
1297                 }
1298
1299                 if (req != NULL && req->rq_repmsg != NULL) {
1300                         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF+1,0);
1301                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1302                 }
1303
1304                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1305                                           NULL, stripe_count);
1306                 if (IS_ERR(handle)) {
1307                         rc = PTR_ERR(handle);
1308                         handle = NULL;
1309                         GOTO(cleanup, rc);
1310                 }
1311
1312                 if (req != NULL && req->rq_repmsg != NULL &&
1313                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1314                     mds_log_op_unlink(obd, lmm,
1315                              lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF+1),
1316                              lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF+2, 0),
1317                              lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF+2))
1318                              > 0) {
1319                         reply_body->valid |= OBD_MD_FLCOOKIE;
1320                 }
1321
1322                 pending_child->d_fsdata = (void *) &dp;
1323                 dp.p_inum = 0;
1324                 dp.p_ptr = req;
1325                 rc = vfs_unlink(pending_dir, pending_child);
1326                 if (rc)
1327                         CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1328
1329                 goto out; /* Don't bother updating attrs on unlinked inode */
1330         }
1331
1332 #if 0
1333         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1334                 /* Update the on-disk attributes if this was the last write
1335                  * close, and all information was provided (i.e., rc == 0)
1336                  *
1337                  * XXX this should probably be abstracted with mds_reint_setattr
1338                  */
1339
1340                 if (request_body->valid & OBD_MD_FLMTIME &&
1341                     LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1342                         LTIME_S(iattr.ia_mtime) = request_body->mtime;
1343                         iattr.ia_valid |= ATTR_MTIME;
1344                 }
1345                 if (request_body->valid & OBD_MD_FLCTIME &&
1346                     LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1347                         LTIME_S(iattr.ia_ctime) = request_body->ctime;
1348                         iattr.ia_valid |= ATTR_CTIME;
1349                 }
1350
1351                 /* XXX can't set block count with fsfilt_setattr (!) */
1352                 if (request_body->valid & OBD_MD_FLSIZE) {
1353                         iattr.ia_valid |= ATTR_SIZE;
1354                         iattr.ia_size = request_body->size;
1355                 }
1356                 /* iattr.ia_blocks = request_body->blocks */
1357
1358         }
1359 #endif
1360         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1361                 /* Only start a transaction to write out only the atime if
1362                  * it is more out-of-date than the specified limit.  If we
1363                  * are already going to write out the atime then do it anyway.
1364                  * */
1365                 LTIME_S(iattr.ia_atime) = request_body->atime;
1366                 if ((LTIME_S(iattr.ia_atime) >
1367                      LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1368                     (iattr.ia_valid != 0 &&
1369                      LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1370                         iattr.ia_valid |= ATTR_ATIME;
1371         }
1372
1373         if (iattr.ia_valid != 0) {
1374                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1375                 if (IS_ERR(handle)) {
1376                         rc = PTR_ERR(handle);
1377                         handle = NULL;
1378                         GOTO(cleanup, rc);
1379                 }
1380                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1381                 if (rc)
1382                         CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1383         }
1384 out:
1385         /* If other clients have this file open for write, rc will be > 0 */
1386         if (rc > 0)
1387                 rc = 0;
1388         l_dput(mfd->mfd_dentry);
1389         mds_mfd_put(mfd);
1390
1391  cleanup:
1392         if (req != NULL && reply_body != NULL) {
1393                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1394         } else if (handle) {
1395                 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1396                 if (err) {
1397                         CERROR("error committing close: %d\n", err);
1398                         if (!rc)
1399                                 rc = err;
1400                 }
1401         }
1402
1403         switch (cleanup_phase) {
1404         case 2:
1405                 dput(pending_child);
1406         case 1:
1407                 UNLOCK_INODE_MUTEX(pending_dir);
1408         }
1409         RETURN(rc);
1410 }
1411
1412 int mds_close(struct ptlrpc_request *req, int offset)
1413 {
1414         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1415         struct obd_device *obd = req->rq_export->exp_obd;
1416         struct mds_body *body;
1417         struct mds_file_data *mfd;
1418         struct lvfs_run_ctxt saved;
1419         struct inode *inode;
1420         int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1421                                sizeof(struct mds_body),
1422                                obd->u.mds.mds_max_mdsize,
1423                                obd->u.mds.mds_max_cookiesize };
1424         ENTRY;
1425
1426         rc = lustre_pack_reply(req, 4, repsize, NULL);
1427         if (rc) {
1428                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1429                 req->rq_status = rc;
1430                 /* continue on to drop local open even if we can't send reply */
1431         } else {
1432                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1433         }
1434
1435         CDEBUG(D_HA, "close req->rep_len %d mdsize %d cookiesize %d\n",
1436                req->rq_replen,
1437                obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1438
1439         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1440                                   lustre_swab_mds_body);
1441         if (body == NULL) {
1442                 CERROR("Can't unpack body\n");
1443                 req->rq_status = -EFAULT;
1444                 RETURN(-EFAULT);
1445         }
1446
1447         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1448                 /* do some stuff */ ;
1449
1450         spin_lock(&med->med_open_lock);
1451         mfd = mds_handle2mfd(&body->handle);
1452         if (mfd == NULL) {
1453                 spin_unlock(&med->med_open_lock);
1454                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1455                           ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1456                 req->rq_status = -ESTALE;
1457                 RETURN(-ESTALE);
1458         }
1459         /* Remove mfd handle so it can't be found again.  We consume mfd_list
1460          * reference here, but still have mds_handle2mfd ref until mfd_close. */
1461         mds_mfd_unlink(mfd, 1);
1462         spin_unlock(&med->med_open_lock);
1463
1464         inode = mfd->mfd_dentry->d_inode;
1465         /* child orphan sem protects orphan_dec_test && is_orphan race */
1466         MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1467         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1468                 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1469                                       sizeof(*body));
1470                 LASSERT(body != NULL);
1471
1472                 mds_pack_inode2fid(&body->fid1, inode);
1473                 mds_pack_inode2body(body, inode);
1474                 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1475                             MDS_PACK_MD_LOCK);
1476         }
1477
1478         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1479         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1480         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1481
1482         mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1483         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1484                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1485                 req->rq_status = -ENOMEM;
1486                 RETURN(-ENOMEM);
1487         }
1488
1489         RETURN(rc);
1490 }
1491
1492 int mds_done_writing(struct ptlrpc_request *req, int offset)
1493 {
1494         struct mds_body *body;
1495         int rc, size[2] = { sizeof(struct ptlrpc_body),
1496                             sizeof(struct mds_body) };
1497         ENTRY;
1498
1499         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1500
1501         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1502                                   lustre_swab_mds_body);
1503         if (body == NULL) {
1504                 CERROR("Can't unpack body\n");
1505                 req->rq_status = -EFAULT;
1506                 RETURN(-EFAULT);
1507         }
1508
1509         rc = lustre_pack_reply(req, 2, size, NULL);
1510         if (rc) {
1511                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1512                 req->rq_status = rc;
1513         }
1514
1515         RETURN(0);
1516 }