lustre/mds/mds_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *   Author: Phil Schwan <phil@clusterfs.com>
   8  *   Author: Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/version.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/version.h>
  35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  36 # include <linux/buffer_head.h>
  37 # include <linux/workqueue.h>
  38 #else
  39 # include <linux/locks.h>
  40 #endif
  41
  42 #include <linux/obd_class.h>
  43 #include <linux/obd_lov.h>
  44 #include <linux/lustre_fsfilt.h>
  45 #include <linux/lprocfs_status.h>
  46
  47 #include "mds_internal.h"
  48
  49 /* Exported function from this file are:
  50  *
  51  * mds_open - called by the intent handler
  52  * mds_close - an rpc handling function
  53  * mds_pin - an rpc handling function - which will go away
  54  * mds_mfd_close - for force closing files when a client dies
  55  */
  56
  57 /*
  58  * MDS file data handling: file data holds a handle for a file opened
  59  * by a client.
  60  */
  61
  62 static void mds_mfd_addref(void *mfdp)
  63 {
  64         struct mds_file_data *mfd = mfdp;
  65
  66         atomic_inc(&mfd->mfd_refcount);
  67         CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
  68                atomic_read(&mfd->mfd_refcount));
  69 }
  70
  71 struct mds_file_data *mds_mfd_new(void)
  72 {
  73         struct mds_file_data *mfd;
  74
  75         OBD_ALLOC(mfd, sizeof *mfd);
  76         if (mfd == NULL) {
  77                 CERROR("mds: out of memory\n");
  78                 return NULL;
  79         }
  80
  81         atomic_set(&mfd->mfd_refcount, 2);
  82
  83         INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  84         class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
  85
  86         return mfd;
  87 }
  88
  89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
  90 {
  91         ENTRY;
  92         LASSERT(handle != NULL);
  93         RETURN(class_handle2object(handle->cookie));
  94 }
  95
  96 static void mds_mfd_put(struct mds_file_data *mfd)
  97 {
  98         CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
  99                atomic_read(&mfd->mfd_refcount) - 1);
 100         LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
 101                 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
 102         if (atomic_dec_and_test(&mfd->mfd_refcount)) {
 103                 LASSERT(list_empty(&mfd->mfd_handle.h_link));
 104                 OBD_FREE(mfd, sizeof *mfd);
 105         }
 106 }
 107
 108 static void mds_mfd_destroy(struct mds_file_data *mfd)
 109 {
 110         class_handle_unhash(&mfd->mfd_handle);
 111         mds_mfd_put(mfd);
 112 }
 113
 114
 115 /* Caller must hold mds->mds_epoch_sem */
 116 static int mds_alloc_filterdata(struct inode *inode)
 117 {
 118         LASSERT(inode->i_filterdata == NULL);
 119         OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
 120         if (inode->i_filterdata == NULL)
 121                 return -ENOMEM;
 122         LASSERT(igrab(inode) == inode);
 123         return 0;
 124 }
 125
 126 /* Caller must hold mds->mds_epoch_sem */
 127 static void mds_free_filterdata(struct inode *inode)
 128 {
 129         LASSERT(inode->i_filterdata != NULL);
 130         OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
 131         inode->i_filterdata = NULL;
 132         iput(inode);
 133 }
 134
 135 /* Write access to a file: executors cause a negative count,
 136  * writers a positive count.  The semaphore is needed to perform
 137  * a check for the sign and then increment or decrement atomically.
 138  *
 139  * This code is closely tied to the allocation of the d_fsdata and the
 140  * MDS epoch, so we use the same semaphore for the whole lot.
 141  *
 142  * We could use a different semaphore for each file, if it ever shows
 143  * up in a profile, which it won't.
 144  *
 145  * epoch argument is nonzero during recovery */
 146 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 147                                 __u64 epoch)
 148 {
 149         int rc = 0;
 150
 151         down(&mds->mds_epoch_sem);
 152
 153         if (atomic_read(&inode->i_writecount) < 0) {
 154                 up(&mds->mds_epoch_sem);
 155                 RETURN(-ETXTBSY);
 156         }
 157
 158
 159         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 160                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 161                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 162                        inode->i_generation);
 163                 goto out;
 164         }
 165
 166         if (inode->i_filterdata == NULL)
 167                 mds_alloc_filterdata(inode);
 168         if (inode->i_filterdata == NULL) {
 169                 rc = -ENOMEM;
 170                 goto out;
 171         }
 172         if (epoch > mds->mds_io_epoch)
 173                 mds->mds_io_epoch = epoch;
 174         else
 175                 mds->mds_io_epoch++;
 176         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 177         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 178                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 179  out:
 180         if (rc == 0)
 181                 atomic_inc(&inode->i_writecount);
 182         up(&mds->mds_epoch_sem);
 183         return rc;
 184 }
 185
 186 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 187  * again -- which is never true if the file is about to be unlinked.  Otherwise
 188  * returns the number of remaining writers. */
 189 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 190                                 struct mds_body *body, int unlinking)
 191 {
 192         int rc = 0;
 193         ENTRY;
 194
 195         down(&mds->mds_epoch_sem);
 196         atomic_dec(&inode->i_writecount);
 197         rc = atomic_read(&inode->i_writecount);
 198         if (rc > 0)
 199                 GOTO(out, rc);
 200 #if 0
 201         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 202                 GOTO(out, rc = EAGAIN);
 203 #endif
 204         mds_free_filterdata(inode);
 205  out:
 206         up(&mds->mds_epoch_sem);
 207         return rc;
 208 }
 209
 210 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 211 {
 212         ENTRY;
 213         down(&mds->mds_epoch_sem);
 214         if (atomic_read(&inode->i_writecount) > 0) {
 215                 up(&mds->mds_epoch_sem);
 216                 RETURN(-ETXTBSY);
 217         }
 218         atomic_dec(&inode->i_writecount);
 219         up(&mds->mds_epoch_sem);
 220         RETURN(0);
 221 }
 222
 223 static void mds_allow_write_access(struct inode *inode)
 224 {
 225         ENTRY;
 226         atomic_inc(&inode->i_writecount);
 227 }
 228
 229 int mds_query_write_access(struct inode *inode)
 230 {
 231         ENTRY;
 232         RETURN(atomic_read(&inode->i_writecount));
 233 }
 234
 235 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 236 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 237                                              struct vfsmount *mnt, int flags,
 238                                              struct ptlrpc_request *req)
 239 {
 240         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 241         struct mds_obd *mds = mds_req2mds(req);
 242         struct mds_file_data *mfd;
 243         struct mds_body *body;
 244         int error;
 245         ENTRY;
 246
 247         mfd = mds_mfd_new();
 248         if (mfd == NULL) {
 249                 CERROR("mds: out of memory\n");
 250                 GOTO(cleanup_dentry, error = -ENOMEM);
 251         }
 252
 253         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 254
 255         if (flags & FMODE_WRITE) {
 256                 /* FIXME: in recovery, need to pass old epoch here */
 257                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 258                 if (error)
 259                         GOTO(cleanup_mfd, error);
 260                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 261         } else if (flags & FMODE_EXEC) {
 262                 error = mds_deny_write_access(mds, dentry->d_inode);
 263                 if (error)
 264                         GOTO(cleanup_mfd, error);
 265         }
 266
 267         dget(dentry);
 268
 269         /* Mark the file as open to handle open-unlink. */
 270
 271         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 272         mds_orphan_open_inc(dentry->d_inode);
 273         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 274
 275         mfd->mfd_mode = flags;
 276         mfd->mfd_dentry = dentry;
 277         mfd->mfd_xid = req->rq_xid;
 278
 279         spin_lock(&med->med_open_lock);
 280         list_add(&mfd->mfd_list, &med->med_open_head);
 281         spin_unlock(&med->med_open_lock);
 282         mds_mfd_put(mfd);
 283
 284         body->handle.cookie = mfd->mfd_handle.h_cookie;
 285
 286         RETURN(mfd);
 287
 288 cleanup_mfd:
 289         mds_mfd_put(mfd);
 290         mds_mfd_destroy(mfd);
 291 cleanup_dentry:
 292         return ERR_PTR(error);
 293 }
 294
 295 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 296                                 struct lov_desc *desc)
 297 {
 298         int i;
 299         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 300                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 301                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 302         }
 303 }
 304
 305 /* Must be called with i_sem 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 obdo *oa = NULL;
 313         struct obd_trans_info oti = { 0 };
 314         struct mds_body *body;
 315         struct lov_stripe_md *lsm = NULL;
 316         struct lov_mds_md *lmm = NULL;
 317         struct inode *inode = dchild->d_inode;
 318         void *lmm_buf;
 319         int rc, lmm_bufsize, lmm_size;
 320         ENTRY;
 321
 322         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 323             !(rec->ur_flags & FMODE_WRITE))
 324                 RETURN(0);
 325
 326         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 327
 328         if (!S_ISREG(inode->i_mode))
 329                 RETURN(0);
 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.oti_objid = *ids;
 337
 338         /* replay case */
 339         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 340                 LASSERT(id_ino(rec->ur_id2));
 341                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 342                 lmm_size = rec->ur_eadatalen;
 343                 lmm = rec->ur_eadata;
 344                 LASSERT(lmm);
 345
 346                 if (*handle == NULL)
 347                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 348                 if (IS_ERR(*handle)) {
 349                         rc = PTR_ERR(*handle);
 350                         *handle = NULL;
 351                         GOTO(out_oa, rc);
 352                 }
 353
 354                 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
 355
 356                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 357                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 358                 LASSERT(lmm_buf);
 359                 LASSERT(lmm_bufsize >= lmm_size);
 360                 memcpy(lmm_buf, lmm, lmm_size);
 361                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 362                 if (rc)
 363                         CERROR("open replay failed to set md:%d\n", rc);
 364                 RETURN(0);
 365         }
 366
 367         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 368                 GOTO(out_ids, rc = -ENOMEM);
 369
 370         oa = obdo_alloc();
 371         if (oa == NULL)
 372                 GOTO(out_ids, rc = -ENOMEM);
 373         oa->o_mode = S_IFREG | 0600;
 374         oa->o_id = inode->i_ino;
 375         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 376         oa->o_generation = inode->i_generation;
 377         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 378         oa->o_gid = 0;
 379         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
 380                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 381         oa->o_size = 0;
 382
 383         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
 384                         OBD_MD_FLCTIME);
 385
 386         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 387                 /* check if things like lfs setstripe are sending us the ea */
 388                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 389                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 390                                            mds->mds_lov_exp,
 391                                            0, &lsm, rec->ur_eadata);
 392                         if (rc)
 393                                 GOTO(out_oa, rc);
 394                 } else {
 395                         /* Per-directory striping default code removed, because
 396                          * it uses the same unnamed EA storage as the directory
 397                          * striping for CMD. -p */
 398                 }
 399                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 400                 rc = obd_create(mds->mds_lov_exp, oa, &lsm, &oti);
 401                 if (rc) {
 402                         int level = D_ERROR;
 403                         if (rc == -ENOSPC)
 404                                 level = D_INODE;
 405                         CDEBUG(level, "error creating objects for "
 406                                       "inode %lu: rc = %d\n",
 407                                inode->i_ino, rc);
 408                         if (rc > 0) {
 409                                 CERROR("obd_create returned invalid "
 410                                        "rc %d\n", rc);
 411                                 rc = -EIO;
 412                         }
 413                         GOTO(out_oa, rc);
 414                 }
 415         } else {
 416                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_lov_exp,
 417                                    0, &lsm, rec->ur_eadata);
 418                 if (rc) {
 419                         GOTO(out_oa, rc);
 420                 }
 421                 lsm->lsm_object_id = oa->o_id;
 422                 lsm->lsm_object_gr = oa->o_gr;
 423         }
 424         if (inode->i_size) {
 425                 oa->o_size = inode->i_size;
 426                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 427                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 428                 rc = obd_setattr(mds->mds_lov_exp, oa, lsm, &oti);
 429                 if (rc) {
 430                         CERROR("error setting attrs for inode %lu: rc %d\n",
 431                                inode->i_ino, rc);
 432                         if (rc > 0) {
 433                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 434                                 rc = -EIO;
 435                         }
 436                         GOTO(out_oa, rc);
 437                 }
 438         }
 439
 440         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 441         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 442
 443         LASSERT(lsm && lsm->lsm_object_id);
 444         lmm = NULL;
 445         rc = obd_packmd(mds->mds_lov_exp, &lmm, lsm);
 446         if (!id_ino(rec->ur_id2))
 447                 obd_free_memmd(mds->mds_lov_exp, &lsm);
 448         LASSERT(rc >= 0);
 449         lmm_size = rc;
 450         body->eadatasize = rc;
 451
 452         if (*handle == NULL)
 453                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 454         if (IS_ERR(*handle)) {
 455                 rc = PTR_ERR(*handle);
 456                 *handle = NULL;
 457                 GOTO(out_ids, rc);
 458         }
 459
 460         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 461         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 462         lmm_bufsize = req->rq_repmsg->buflens[offset];
 463         LASSERT(lmm_buf);
 464         LASSERT(lmm_bufsize >= lmm_size);
 465
 466         memcpy(lmm_buf, lmm, lmm_size);
 467         obd_free_diskmd(mds->mds_lov_exp, &lmm);
 468  out_oa:
 469         oti_free_cookies(&oti);
 470         obdo_free(oa);
 471  out_ids:
 472         if (rc) {
 473                 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
 474                 *ids = NULL;
 475         }
 476         RETURN(rc);
 477 }
 478
 479 static void reconstruct_open(struct mds_update_record *rec, int offset,
 480                              struct ptlrpc_request *req,
 481                              struct lustre_handle *child_lockh)
 482 {
 483         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 484         struct mds_client_data *mcd = med->med_mcd;
 485         struct mds_obd *mds = mds_req2mds(req);
 486         struct mds_file_data *mfd;
 487         struct obd_device *obd = req->rq_export->exp_obd;
 488         struct dentry *parent, *dchild;
 489         struct ldlm_reply *rep;
 490         struct mds_body *body;
 491         struct list_head *t;
 492         int put_child = 1;
 493         int rc;
 494         ENTRY;
 495
 496         LASSERT(offset == 3); /* only called via intent */
 497         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 498         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 499
 500         /* copy rc, transno and disp; steal locks */
 501         mds_req_from_mcd(req, mcd);
 502         intent_set_disposition(rep, mcd->mcd_last_data);
 503
 504         /* Only replay if create or open actually happened. */
 505         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 506                 EXIT;
 507                 return; /* error looking up parent or child */
 508         }
 509
 510         parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 511         LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 512                  OLID4(rec->ur_id1), PTR_ERR(parent));
 513
 514         dchild = ll_lookup_one_len(rec->ur_name, parent,
 515                                    rec->ur_namelen - 1);
 516         LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 517                  OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 518
 519         if (!dchild->d_inode)
 520                 GOTO(out_dput, 0); /* child not present to open */
 521
 522         /* At this point, we know we have a child. We'll send
 523          * it back _unless_ it not created and open failed.  */
 524         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 525             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 526             req->rq_status)
 527                 GOTO(out_dput, 0);
 528
 529         /* get lock (write for O_CREAT, read otherwise) */
 530         mds_pack_inode2body(obd, body, dchild->d_inode, 0);
 531         if (S_ISREG(dchild->d_inode->i_mode)) {
 532                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 533                                  dchild->d_inode, 1);
 534
 535                 if (rc)
 536                         LASSERT(rc == req->rq_status);
 537
 538                 /* If we have LOV EA data, the OST holds size, mtime */
 539                 if (!(body->valid & OBD_MD_FLEASIZE))
 540                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 541                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 542         }
 543
 544         /* If we have -EEXIST as the status, and we were asked to create
 545          * exclusively, we can tell we failed because the file already existed.
 546          */
 547         if (req->rq_status == -EEXIST &&
 548             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 549              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 550                 GOTO(out_dput, 0);
 551         }
 552
 553         /* If we didn't get as far as trying to open, then some locking thing
 554          * probably went wrong, and we'll just bail here.
 555          */
 556         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 557                 GOTO(out_dput, 0);
 558
 559         /* If we failed, then we must have failed opening, so don't look for
 560          * file descriptor or anything, just give the client the bad news.
 561          */
 562         if (req->rq_status)
 563                 GOTO(out_dput, 0);
 564
 565         mfd = NULL;
 566         list_for_each(t, &med->med_open_head) {
 567                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 568                 if (mfd->mfd_xid == req->rq_xid)
 569                         break;
 570                 mfd = NULL;
 571         }
 572
 573         /* #warning "XXX fixme" bug 2991 */
 574         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 575          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 576          * to detect a re-open */
 577         if (mfd == NULL) {
 578                 mntget(mds->mds_vfsmnt);
 579                 CERROR("Re-opened file \n");
 580                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 581                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 582                 if (!mfd) {
 583                         CERROR("mds: out of memory\n");
 584                         GOTO(out_dput, req->rq_status = -ENOMEM);
 585                 }
 586                 put_child = 0;
 587         } else {
 588                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 589                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 590                        mfd->mfd_handle.h_cookie);
 591         }
 592
 593  out_dput:
 594         if (put_child)
 595                 l_dput(dchild);
 596         l_dput(parent);
 597         EXIT;
 598 }
 599
 600 /* do NOT or the MAY_*'s, you'll get the weakest */
 601 static int accmode(int flags)
 602 {
 603         int res = 0;
 604
 605         if (flags & FMODE_READ)
 606                 res = MAY_READ;
 607         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 608                 res |= MAY_WRITE;
 609         if (flags & FMODE_EXEC)
 610                 res = MAY_EXEC;
 611         return res;
 612 }
 613
 614 /* Handles object creation, actual opening, and I/O epoch */
 615 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 616                            struct mds_body *body, int flags, void **handle,
 617                            struct mds_update_record *rec,struct ldlm_reply *rep)
 618 {
 619         struct mds_obd *mds = mds_req2mds(req);
 620         struct obd_device *obd = req->rq_export->exp_obd;
 621         struct mds_file_data *mfd = NULL;
 622         obd_id *ids = NULL; /* object IDs created */
 623         unsigned mode;
 624         int rc = 0;
 625         ENTRY;
 626
 627         /* atomically create objects if necessary */
 628         down(&dchild->d_inode->i_sem);
 629         mode = dchild->d_inode->i_mode;
 630         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 631             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 632                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 633                                  dchild->d_inode, 0);
 634                 if (rc) {
 635                         up(&dchild->d_inode->i_sem);
 636                         RETURN(rc);
 637                 }
 638         }
 639         if (rec != NULL) {
 640                 /* no EA: create objects */
 641                 if ((body->valid & OBD_MD_FLEASIZE) &&
 642                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 643                         up(&dchild->d_inode->i_sem);
 644                         RETURN(-EEXIST);
 645                 }
 646                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 647                         /* no EA: create objects */
 648                         rc = mds_create_objects(req, 2, rec, mds, obd,
 649                                                 dchild, handle, &ids);
 650                         if (rc) {
 651                                 CERROR("mds_create_objects: rc = %d\n", rc);
 652                                 up(&dchild->d_inode->i_sem);
 653                                 RETURN(rc);
 654                         }
 655                 }
 656                 if (S_ISREG(dchild->d_inode->i_mode) &&
 657                     (body->valid & OBD_MD_FLEASIZE)) {
 658                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 659                                                    req->rq_repmsg, 2);
 660                         if (!rc)
 661                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 662                                                  dchild->d_inode, 0);
 663                         if (rc) {
 664                                 up(&dchild->d_inode->i_sem);
 665                                 RETURN(rc);
 666                         }
 667                 }
 668         }
 669         /* If the inode has no EA data, then MDSs hold size, mtime */
 670         if (S_ISREG(dchild->d_inode->i_mode) &&
 671             !(body->valid & OBD_MD_FLEASIZE)) {
 672                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 673                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 674         }
 675         up(&dchild->d_inode->i_sem);
 676
 677         intent_set_disposition(rep, DISP_OPEN_OPEN);
 678         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 679         if (IS_ERR(mfd))
 680                 RETURN(PTR_ERR(mfd));
 681
 682         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 683                mfd->mfd_handle.h_cookie);
 684         if (ids != NULL) {
 685                 mds_lov_update_objids(obd, ids);
 686                 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
 687         }
 688         //if (rc)
 689         //        mds_mfd_destroy(mfd);
 690         RETURN(rc);
 691 }
 692
 693 static int mds_open_by_id(struct ptlrpc_request *req,
 694                           struct lustre_id *id,
 695                           struct mds_body *body, int flags,
 696                           struct mds_update_record *rec,
 697                           struct ldlm_reply *rep)
 698 {
 699         struct mds_obd *mds = mds_req2mds(req);
 700         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 701         struct dentry *dchild;
 702         char idname[LL_ID_NAMELEN];
 703         int idlen = 0, rc;
 704         void *handle = NULL;
 705         ENTRY;
 706
 707         down(&pending_dir->i_sem);
 708
 709         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 710
 711         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 712                                 idlen);
 713         if (IS_ERR(dchild)) {
 714                 up(&pending_dir->i_sem);
 715                 rc = PTR_ERR(dchild);
 716                 CERROR("error looking up %s in PENDING: rc = %d\n",
 717                        idname, rc);
 718                 RETURN(rc);
 719         }
 720
 721         if (dchild->d_inode != NULL) {
 722                 up(&pending_dir->i_sem);
 723                 mds_inode_set_orphan(dchild->d_inode);
 724                 mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
 725                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 726                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 727                 CWARN("Orphan %s found and opened in PENDING directory\n",
 728                        idname);
 729                 goto open;
 730         }
 731         dput(dchild);
 732         up(&pending_dir->i_sem);
 733
 734         /*
 735          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 736          * regular inode that was previously created.
 737          */
 738         dchild = mds_id2dentry(req2obd(req), id, NULL);
 739         if (IS_ERR(dchild))
 740                 RETURN(PTR_ERR(dchild));
 741
 742         mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
 743         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 744         intent_set_disposition(rep, DISP_LOOKUP_POS);
 745
 746  open:
 747         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 748         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 749                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 750         /* XXX what do we do here if mds_finish_transno itself failed? */
 751
 752         l_dput(dchild);
 753         RETURN(rc);
 754 }
 755
 756 int mds_pin(struct ptlrpc_request *req, int offset)
 757 {
 758         struct obd_device *obd = req->rq_export->exp_obd;
 759         struct mds_body *request_body, *reply_body;
 760         struct lvfs_run_ctxt saved;
 761         int rc, size = sizeof(*reply_body);
 762         ENTRY;
 763
 764         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 765                                       sizeof(*request_body));
 766
 767         rc = lustre_pack_reply(req, 1, &size, NULL);
 768         if (rc)
 769                 RETURN(rc);
 770         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 771
 772         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 773         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 774                             request_body->flags, NULL, NULL);
 775         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 776
 777         RETURN(rc);
 778 }
 779
 780 /*
 781  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 782  * child_lockh is NULL we just get the lock as a barrier to wait for other
 783  * holders of this lock, and drop it right away again.
 784  */
 785 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 786                        struct lustre_handle *child_lockh)
 787 {
 788         struct ldlm_res_id child_res_id = { .name = { 0, 0, 0, 0 } };
 789         struct lustre_handle lockh;
 790         struct lustre_id sid;
 791         int lock_flags = 0;
 792         int rc;
 793
 794         down(&inode->i_sem);
 795         rc = mds_read_inode_sid(obd, inode, &sid);
 796         up(&inode->i_sem);
 797         if (rc) {
 798                 CERROR("Can't read inode self id, err = %d\n", rc);
 799                 RETURN(rc);
 800         }
 801
 802         child_res_id.name[0] = id_fid(&sid);
 803         child_res_id.name[1] = id_group(&sid);
 804         child_res_id.name[2] = 1;
 805
 806         if (child_lockh == NULL)
 807                 child_lockh = &lockh;
 808
 809         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 810                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 811                               mds_blocking_ast, ldlm_completion_ast, NULL,
 812                               NULL, NULL, 0, NULL, child_lockh);
 813         if (rc != ELDLM_OK)
 814                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 815         else if (child_lockh == &lockh)
 816                 ldlm_lock_decref(child_lockh, LCK_EX);
 817
 818         RETURN(rc);
 819 }
 820
 821 static int is_mount_object(struct dentry *dparent)
 822 {
 823         struct dentry *dchild;
 824
 825         if (!(dparent->d_inode->i_mode & S_ISUID))
 826                 return 0;
 827
 828         dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
 829         if (IS_ERR(dchild) || dchild == NULL)
 830                 return 0;
 831
 832         dput(dchild);
 833         return 1;
 834 }
 835
 836 int mds_open(struct mds_update_record *rec, int offset,
 837              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 838 {
 839         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 840         struct obd_device *obd = req->rq_export->exp_obd;
 841         struct mds_obd *mds = mds_req2mds(req);
 842         struct ldlm_reply *rep = NULL;
 843         struct mds_body *body = NULL;
 844         struct dentry *dchild = NULL, *dparent = NULL;
 845         struct mds_export_data *med;
 846         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 847         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 848         int parent_mode = LCK_PR;
 849         void *handle = NULL;
 850         struct dentry_params dp;
 851         struct mea *mea = NULL;
 852         int mea_size, update_mode;
 853         ENTRY;
 854
 855         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 856                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 857                   rec->ur_mode);
 858
 859         if (offset == 3) { /* intent */
 860                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 861                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 862         } else if (offset == 1) { /* non-intent reint */
 863                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 864         } else {
 865                 body = NULL;
 866                 LBUG();
 867         }
 868
 869         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 870
 871         /*
 872          * Step 0: If we are passed a id, then we assume the client already
 873          * opened this file and is only replaying the RPC, so we open the inode
 874          * by fid (at some large expense in security).
 875          */
 876         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 877                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 878                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 879                        rec->ur_name, rec->ur_mode);
 880
 881                 LASSERT(id_ino(rec->ur_id2));
 882
 883                 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
 884                                     rec, rep);
 885                 if (rc != -ENOENT)
 886                         RETURN(rc);
 887
 888                 /* We didn't find the correct inode on disk either, so we
 889                  * need to re-create it via a regular replay. */
 890                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 891         } else {
 892                 LASSERT(!id_ino(rec->ur_id2));
 893         }
 894
 895         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 896
 897         med = &req->rq_export->exp_mds_data;
 898         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 899                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 900                 RETURN(-ENOMEM);
 901         }
 902
 903         acc_mode = accmode(rec->ur_flags);
 904
 905         /* Step 1: Find and lock the parent */
 906         if (rec->ur_flags & O_CREAT) {
 907                 /* XXX Well, in fact we only need this lock mode change if
 908                    in addition to O_CREAT, the file does not exist.
 909                    But we do not know if it exists or not yet */
 910                 parent_mode = LCK_PW;
 911         }
 912
 913         if (rec->ur_namelen == 1) {
 914                 /* client (LMV) wants to open the file by id */
 915                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 916                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 917                 if (IS_ERR(dchild)) {
 918                         rc = PTR_ERR(dparent);
 919                         CERROR("child lookup by an id error %d\n", rc);
 920                         GOTO(cleanup, rc);
 921                 }
 922                 goto got_child;
 923         }
 924
 925         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
 926                                        parent_lockh, &update_mode, rec->ur_name,
 927                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 928         if (IS_ERR(dparent)) {
 929                 rc = PTR_ERR(dparent);
 930                 if (rc != -ENOENT)
 931                         CERROR("parent lookup for "DLID4" failed, error %d\n",
 932                                OLID4(rec->ur_id1), rc);
 933                 GOTO(cleanup, rc);
 934         }
 935         LASSERT(dparent->d_inode != NULL);
 936
 937         cleanup_phase = 1; /* parent dentry and lock */
 938
 939         /* try to retrieve MEA data for this dir */
 940         rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
 941         if (rc)
 942                 GOTO(cleanup, rc);
 943
 944         if (mea != NULL) {
 945                 /*
 946                  * dir is already splitted, check is requested filename should *
 947                  * live at this MDS or at another one.
 948                  */
 949                 int i;
 950                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 951                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
 952                         CDEBUG(D_OTHER,
 953                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
 954                                " should be %lu(%d)\n", obd->obd_name,
 955                                mea->mea_master, dparent->d_inode->i_ino,
 956                                dparent->d_inode->i_generation, rec->ur_name,
 957                                (unsigned long)id_group(&mea->mea_ids[i]), i);
 958                         GOTO(cleanup, rc = -ERESTART);
 959                 }
 960         }
 961
 962         /* Step 2: Lookup the child */
 963         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 964         if (IS_ERR(dchild)) {
 965                 rc = PTR_ERR(dchild);
 966                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 967                 GOTO(cleanup, rc);
 968         }
 969
 970 got_child:
 971         cleanup_phase = 2; /* child dentry */
 972
 973         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 974
 975         if (dchild->d_flags & DCACHE_CROSS_REF) {
 976                 struct ldlm_res_id res_id = { . name = {0} };
 977                 ldlm_policy_data_t policy;
 978                 int flags = 0;
 979
 980                 mds_pack_dentry2body(obd, body, dchild, 1);
 981                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 982
 983                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
 984                        OLID4(&body->id1));
 985
 986                 res_id.name[0] = dchild->d_fid;
 987                 res_id.name[1] = dchild->d_mdsnum;
 988
 989                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 990                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 991                                       res_id, LDLM_IBITS, &policy,
 992                                       LCK_PR, &flags, mds_blocking_ast,
 993                                       ldlm_completion_ast, NULL, NULL,
 994                                       NULL, 0, NULL, child_lockh);
 995 #ifdef S_PDIROPS
 996                 if (parent_lockh[1].cookie != 0)
 997                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 998 #endif
 999                 ldlm_lock_decref(parent_lockh, parent_mode);
1000                 if (mea)
1001                         OBD_FREE(mea, mea_size);
1002                 l_dput(dchild);
1003                 l_dput(dparent);
1004                 RETURN(rc);
1005         }
1006
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;
1015                 struct iattr iattr;
1016                 struct inode *inode;
1017
1018                 ino = (unsigned long)id_ino(rec->ur_id2);
1019
1020                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1021                                           update_mode);
1022
1023                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1024                        obd->obd_name, dparent->d_inode->i_ino,
1025                        dparent->d_inode->i_generation, rc);
1026
1027                 if (rc > 0) {
1028                         /* dir got splitted */
1029                         GOTO(cleanup, rc = -ERESTART);
1030                 } else if (rc < 0) {
1031                         /* error happened during spitting */
1032                         GOTO(cleanup, rc);
1033                 }
1034
1035                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1036                         /*
1037                          * dentry is negative and we weren't supposed to create
1038                          * object, get out of here.
1039                          */
1040                         GOTO(cleanup, rc = -ENOENT);
1041                 }
1042
1043                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1044                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1045                                       NULL);
1046                 if (IS_ERR(handle)) {
1047                         rc = PTR_ERR(handle);
1048                         handle = NULL;
1049                         GOTO(cleanup, rc);
1050                 }
1051                 dchild->d_fsdata = (void *) &dp;
1052                 dp.p_ptr = req;
1053                 dp.p_inum = ino;
1054
1055                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1056                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1057                         dchild->d_fsdata = NULL;
1058
1059                 if (rc) {
1060                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1061                         GOTO(cleanup, rc);
1062                 }
1063                 inode = dchild->d_inode;
1064                 if (ino) {
1065                         LASSERT(ino == inode->i_ino);
1066
1067                         /* written as part of setattr */
1068                         inode->i_generation = id_gen(rec->ur_id2);
1069                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1070                                inode->i_ino, inode->i_generation);
1071                 }
1072
1073                 created = 1;
1074                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1075                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1076                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1077
1078                 iattr.ia_uid = rec->ur_fsuid;
1079                 if (dparent->d_inode->i_mode & S_ISGID)
1080                         iattr.ia_gid = dparent->d_inode->i_gid;
1081                 else
1082                         iattr.ia_gid = rec->ur_fsgid;
1083
1084                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1085                         ATTR_MTIME | ATTR_CTIME;
1086
1087                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1088                 if (rc)
1089                         CERROR("error on child setattr: rc = %d\n", rc);
1090
1091                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1092
1093                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1094                 if (rc)
1095                         CERROR("error on parent setattr: rc = %d\n", rc);
1096                 else {
1097                         MD_COUNTER_INCREMENT(obd, create);
1098                 }
1099
1100                 down(&dchild->d_inode->i_sem);
1101                 if (ino) {
1102                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1103                                                   handle, rec->ur_id2);
1104                         if (rc) {
1105                                 CERROR("mds_update_inode_sid() failed, "
1106                                        "rc = %d\n", rc);
1107                         }
1108                         id_assign_fid(&body->id1, rec->ur_id2);
1109
1110                         /*
1111                          * make sure, that fid is up-to-date.
1112                          */
1113                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1114                 } else {
1115                         rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1116                                                  handle, &body->id1);
1117                         if (rc) {
1118                                 CERROR("mds_alloc_inode_sid() failed, "
1119                                        "rc = %d\n", rc);
1120                         }
1121                 }
1122                 up(&dchild->d_inode->i_sem);
1123
1124                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1125                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1126                                            dchild->d_inode, handle, 0);
1127                         handle = NULL;
1128                 }
1129
1130                 acc_mode = 0;           /* Don't check for permissions */
1131
1132                 /*
1133                  * we do not read fid from EA here, because it is already
1134                  * updated and thus we avoid not needed IO, locking, etc.
1135                  */
1136                 body->valid |= OBD_MD_FID;
1137                 mds_pack_inode2body(obd, body, dchild->d_inode, 0);
1138         } else {
1139                 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1140         }
1141         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1142                  "dchild %*s (%p) inode %p\n", dchild->d_name.len,
1143                  dchild->d_name.name, dchild, dchild->d_inode);
1144
1145         if (S_ISREG(dchild->d_inode->i_mode)) {
1146                 /* Check permissions etc */
1147                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1148                 if (rc != 0)
1149                         GOTO(cleanup, rc);
1150
1151                 /* Can't write to a read-only file */
1152                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1153                         GOTO(cleanup, rc = -EPERM);
1154
1155                 /* An append-only file must be opened in append mode for
1156                  * writing */
1157                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1158                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1159                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1160                         GOTO(cleanup, rc = -EPERM);
1161         }
1162
1163         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1164             (rec->ur_flags & MDS_OPEN_EXCL)) {
1165                 /* File already exists, we didn't just create it, and we
1166                  * were passed O_EXCL; err-or. */
1167                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1168         }
1169
1170         /* if we are following a symlink, don't open */
1171         if (S_ISLNK(dchild->d_inode->i_mode))
1172                 GOTO(cleanup, rc = 0);
1173
1174         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1175             !S_ISDIR(dchild->d_inode->i_mode))
1176                 GOTO(cleanup, rc = -ENOTDIR);
1177
1178         if (S_ISDIR(dchild->d_inode->i_mode)) {
1179                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1180                     rec->ur_flags & FMODE_WRITE) {
1181                         /*we are trying to create or write a exist dir*/
1182                         GOTO(cleanup, rc = -EISDIR);
1183                 }
1184                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1185                         GOTO(cleanup, rc = -EACCES);
1186
1187                 if (is_mount_object(dchild)) {
1188                         CERROR("Found possible GNS mount object %*s; not "
1189                                "opening.\n", dchild->d_name.len,
1190                                dchild->d_name.name);
1191                         GOTO(cleanup, rc = 0); // success, but don't really open
1192                 }
1193         }
1194
1195         /* Step 5: mds_open it */
1196         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1197                              rec, rep);
1198         GOTO(cleanup, rc);
1199
1200  cleanup:
1201         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1202                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1203
1204         switch (cleanup_phase) {
1205         case 2:
1206                 if (rc && created) {
1207                         int err = vfs_unlink(dparent->d_inode, dchild);
1208                         if (err) {
1209                                 CERROR("unlink(%*s) in error path: %d\n",
1210                                        dchild->d_name.len, dchild->d_name.name,
1211                                        err);
1212                         }
1213                 } else if (created) {
1214                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1215                 }
1216                 l_dput(dchild);
1217         case 1:
1218                 if (dparent == NULL)
1219                         break;
1220
1221                 l_dput(dparent);
1222 #ifdef S_PDIROPS
1223                 if (parent_lockh[1].cookie != 0)
1224                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1225 #endif
1226                 if (rc)
1227                         ldlm_lock_decref(parent_lockh, parent_mode);
1228                 else
1229                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1230         }
1231         if (mea)
1232                 OBD_FREE(mea, mea_size);
1233         RETURN(rc);
1234 }
1235
1236 /* Close a "file descriptor" and possibly unlink an orphan from the
1237  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1238  *
1239  * If we are being called from mds_disconnect() because the client has
1240  * disappeared, then req == NULL and we do not update last_rcvd because
1241  * there is nothing that could be recovered by the client at this stage
1242  * (it will not even _have_ an entry in last_rcvd anymore).
1243  *
1244  * Returns EAGAIN if the client needs to get more data and re-close. */
1245 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1246                   struct obd_device *obd, struct mds_file_data *mfd,
1247                   int unlink_orphan)
1248 {
1249         struct inode *inode = mfd->mfd_dentry->d_inode;
1250         char idname[LL_ID_NAMELEN];
1251         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1252         struct dentry *pending_child = NULL;
1253         struct mds_obd *mds = &obd->u.mds;
1254         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1255         void *handle = NULL;
1256         struct mds_body *request_body = NULL, *reply_body = NULL;
1257         struct dentry_params dp;
1258         struct iattr iattr = { 0 };
1259         struct llog_create_locks *lcl = NULL;
1260         ENTRY;
1261
1262         if (req && req->rq_reqmsg != NULL)
1263                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1264                                               sizeof(*request_body));
1265         if (req && req->rq_repmsg != NULL)
1266                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1267                                             sizeof(*reply_body));
1268
1269         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1270         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1271                idname, inode->i_nlink, mds_orphan_open_count(inode));
1272
1273         last_orphan = mds_orphan_open_dec_test(inode) &&
1274                 mds_inode_is_orphan(inode);
1275         UP_WRITE_I_ALLOC_SEM(inode);
1276
1277         /* this is half of the actual "close" */
1278         if (mfd->mfd_mode & FMODE_WRITE) {
1279                 rc = mds_put_write_access(mds, inode, request_body,
1280                                           last_orphan && unlink_orphan);
1281         } else if (mfd->mfd_mode & FMODE_EXEC) {
1282                 mds_allow_write_access(inode);
1283         }
1284
1285         if (last_orphan && unlink_orphan) {
1286                 struct lov_mds_md *lmm = NULL;
1287                 int stripe_count = 0;
1288                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1289
1290                 if (obd->obd_recovering) {
1291                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1292                                " is over\n", idname);
1293                         GOTO(out, rc);
1294                 }
1295
1296                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
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                                idname, 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, but
1306                  * normally it will still be in the dcache.
1307                  */
1308                 down(&pending_dir->i_sem);
1309                 cleanup_phase = 1; /* up(i_sem) when finished */
1310                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1311                                                idlen);
1312                 if (IS_ERR(pending_child))
1313                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1314                 LASSERT(pending_child->d_inode != NULL);
1315
1316                 cleanup_phase = 2; /* dput(pending_child) when finished */
1317                 if (req != NULL && req->rq_repmsg != NULL) {
1318                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1319                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1320                 }
1321
1322                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1323                                           NULL, stripe_count);
1324                 if (IS_ERR(handle)) {
1325                         rc = PTR_ERR(handle);
1326                         handle = NULL;
1327                         GOTO(cleanup, rc);
1328                 }
1329
1330                 pending_child->d_fsdata = (void *) &dp;
1331                 dp.p_inum = 0;
1332                 dp.p_ptr = req;
1333                 if (S_ISDIR(pending_child->d_inode->i_mode))
1334                         rc = vfs_rmdir(pending_dir, pending_child);
1335                 else
1336                         rc = vfs_unlink(pending_dir, pending_child);
1337                 if (rc)
1338                         CERROR("error unlinking orphan %s: rc %d\n",
1339                                idname, rc);
1340
1341                 if (req != NULL && req->rq_repmsg != NULL &&
1342                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1343                     mds_log_op_unlink(obd, pending_child->d_inode,
1344                                       lmm, req->rq_repmsg->buflens[1],
1345                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1346                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1347                         reply_body->valid |= OBD_MD_FLCOOKIE;
1348                 }
1349
1350                 goto out; /* Don't bother updating attrs on unlinked inode */
1351         }
1352
1353 #if 0
1354         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1355                 /* Update the on-disk attributes if this was the last write
1356                  * close, and all information was provided (i.e., rc == 0)
1357                  *
1358                  * XXX this should probably be abstracted with mds_reint_setattr
1359                  */
1360
1361                 if (request_body->valid & OBD_MD_FLMTIME &&
1362                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1363                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1364                         iattr.ia_valid |= ATTR_MTIME;
1365                 }
1366                 if (request_body->valid & OBD_MD_FLCTIME &&
1367                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1368                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1369                         iattr.ia_valid |= ATTR_CTIME;
1370                 }
1371
1372                 /* XXX can't set block count with fsfilt_setattr (!) */
1373                 if (request_body->valid & OBD_MD_FLSIZE) {
1374                         iattr.ia_valid |= ATTR_SIZE;
1375                         iattr.ia_size = request_body->size;
1376                 }
1377                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1378                         iattr.ia_valid |= ATTR_BLOCKS;
1379                         iattr.ia_blocks = request_body->blocks
1380                 } */
1381
1382         }
1383 #endif
1384         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1385                 /* Only start a transaction to write out only the atime if
1386                  * it is more out-of-date than the specified limit.  If we
1387                  * are already going to write out the atime then do it anyway.
1388                  * */
1389                 if ((request_body->atime >
1390                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1391                     (iattr.ia_valid != 0 &&
1392                      request_body->atime > LTIME_S(inode->i_atime))) {
1393                         LTIME_S(iattr.ia_atime) = request_body->atime;
1394                         iattr.ia_valid |= ATTR_ATIME;
1395                 }
1396         }
1397
1398         if (iattr.ia_valid != 0) {
1399                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1400                 if (IS_ERR(handle))
1401                         GOTO(cleanup, rc = PTR_ERR(handle));
1402                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1403                 if (rc)
1404                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1405         }
1406 out:
1407         /* If other clients have this file open for write, rc will be > 0 */
1408         if (rc > 0)
1409                 rc = 0;
1410         l_dput(mfd->mfd_dentry);
1411         mds_mfd_destroy(mfd);
1412
1413  cleanup:
1414         if (req != NULL && reply_body != NULL) {
1415                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1416         } else if (handle) {
1417                 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1418                 if (err) {
1419                         CERROR("error committing close: %d\n", err);
1420                         if (!rc)
1421                                 rc = err;
1422                 }
1423         }
1424
1425         switch (cleanup_phase) {
1426         case 2:
1427                 if (lcl != NULL)
1428                         ptlrpc_save_llog_lock(req, lcl);
1429                 dput(pending_child);
1430         case 1:
1431                 up(&pending_dir->i_sem);
1432         }
1433         RETURN(rc);
1434 }
1435
1436 int mds_close(struct ptlrpc_request *req, int offset)
1437 {
1438         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1439         struct obd_device *obd = req->rq_export->exp_obd;
1440         struct mds_body *body;
1441         struct mds_file_data *mfd;
1442         struct lvfs_run_ctxt saved;
1443         struct inode *inode;
1444         int rc, repsize[3] = {sizeof(struct mds_body),
1445                               obd->u.mds.mds_max_mdsize,
1446                               obd->u.mds.mds_max_cookiesize};
1447         ENTRY;
1448         MD_COUNTER_INCREMENT(obd, close);
1449
1450         rc = lustre_pack_reply(req, 3, repsize, NULL);
1451         if (rc) {
1452                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1453                 req->rq_status = rc;
1454         } else {
1455                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1456         }
1457
1458         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1459                 DEBUG_REQ(D_HA, req, "close replay\n");
1460                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1461                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1462                        req->rq_repmsg->buflens[2]);
1463         }
1464
1465
1466         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1467                                   lustre_swab_mds_body);
1468         if (body == NULL) {
1469                 CERROR("Can't unpack body\n");
1470                 req->rq_status = -EFAULT;
1471                 RETURN(-EFAULT);
1472         }
1473
1474         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1475                 /* do some stuff */ ;
1476
1477         mfd = mds_handle2mfd(&body->handle);
1478         if (mfd == NULL) {
1479                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1480                           ": cookie "LPX64, body->id1.li_stc.u.e3s.l3s_ino,
1481                           body->handle.cookie);
1482                 req->rq_status = -ESTALE;
1483                 RETURN(-ESTALE);
1484         }
1485
1486         inode = mfd->mfd_dentry->d_inode;
1487         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1488         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1489         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1490                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1491                 LASSERT(body != NULL);
1492
1493                 mds_pack_inode2body(obd, body, inode, 0);
1494                 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, MDS_PACK_MD_LOCK);
1495         }
1496         spin_lock(&med->med_open_lock);
1497         list_del(&mfd->mfd_list);
1498         spin_unlock(&med->med_open_lock);
1499
1500         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1501         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1502         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1503
1504         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1505                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1506                 req->rq_status = -ENOMEM;
1507                 mds_mfd_put(mfd);
1508                 RETURN(-ENOMEM);
1509         }
1510
1511         mds_mfd_put(mfd);
1512         RETURN(0);
1513 }
1514
1515 int mds_done_writing(struct ptlrpc_request *req, int offset)
1516 {
1517         struct mds_body *body;
1518         int rc, size = sizeof(struct mds_body);
1519         ENTRY;
1520
1521         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1522
1523         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1524                                   lustre_swab_mds_body);
1525         if (body == NULL) {
1526                 CERROR("Can't unpack body\n");
1527                 req->rq_status = -EFAULT;
1528                 RETURN(-EFAULT);
1529         }
1530
1531         rc = lustre_pack_reply(req, 1, &size, NULL);
1532         if (rc) {
1533                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1534                 req->rq_status = rc;
1535         }
1536
1537         RETURN(0);
1538 }