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 #ifdef IFILTERDATA_ACTUALLY_USED
 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 #endif /*IFILTERDATA_ACTUALLY_USED*/
 135
 136 /* Write access to a file: executors cause a negative count,
 137  * writers a positive count.  The semaphore is needed to perform
 138  * a check for the sign and then increment or decrement atomically.
 139  *
 140  * This code is closely tied to the allocation of the d_fsdata and the
 141  * MDS epoch, so we use the same semaphore for the whole lot.
 142  *
 143  * We could use a different semaphore for each file, if it ever shows
 144  * up in a profile, which it won't.
 145  *
 146  * epoch argument is nonzero during recovery */
 147 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
 148                                 __u64 epoch)
 149 {
 150         int rc = 0;
 151
 152         down(&mds->mds_epoch_sem);
 153
 154         if (atomic_read(&inode->i_writecount) < 0) {
 155                 up(&mds->mds_epoch_sem);
 156                 RETURN(-ETXTBSY);
 157         }
 158
 159 #ifdef IFILTERDATA_ACTUALLY_USED
 160         if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
 161                 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
 162                        MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
 163                        inode->i_generation);
 164                 goto out;
 165         }
 166
 167         if (inode->i_filterdata == NULL)
 168                 mds_alloc_filterdata(inode);
 169         if (inode->i_filterdata == NULL) {
 170                 rc = -ENOMEM;
 171                 goto out;
 172         }
 173 #endif /*IFILTERDATA_ACTUALLY_USED*/
 174         if (epoch > mds->mds_io_epoch)
 175                 mds->mds_io_epoch = epoch;
 176         else
 177                 mds->mds_io_epoch++;
 178 #ifdef IFILTERDATA_ACTUALLY_USED
 179         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
 180         CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
 181                mds->mds_io_epoch, inode->i_ino, inode->i_generation);
 182  out:
 183 #endif /*IFILTERDATA_ACTUALLY_USED*/
 184         if (rc == 0)
 185                 atomic_inc(&inode->i_writecount);
 186         up(&mds->mds_epoch_sem);
 187         return rc;
 188 }
 189
 190 /* Returns EAGAIN if the client needs to get size and/or cookies and close
 191  * again -- which is never true if the file is about to be unlinked.  Otherwise
 192  * returns the number of remaining writers. */
 193 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
 194                                 struct mds_body *body, int unlinking)
 195 {
 196         int rc = 0;
 197         ENTRY;
 198
 199         down(&mds->mds_epoch_sem);
 200         atomic_dec(&inode->i_writecount);
 201         rc = atomic_read(&inode->i_writecount);
 202         if (rc > 0)
 203                 GOTO(out, rc);
 204 #if 0
 205         if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
 206                 GOTO(out, rc = EAGAIN);
 207 #endif
 208 #ifdef IFILTERDATA_ACTUALLY_USED
 209         mds_free_filterdata(inode);
 210 #endif
 211  out:
 212         up(&mds->mds_epoch_sem);
 213         return rc;
 214 }
 215
 216 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
 217 {
 218         ENTRY;
 219         down(&mds->mds_epoch_sem);
 220         if (atomic_read(&inode->i_writecount) > 0) {
 221                 up(&mds->mds_epoch_sem);
 222                 RETURN(-ETXTBSY);
 223         }
 224         atomic_dec(&inode->i_writecount);
 225         up(&mds->mds_epoch_sem);
 226         RETURN(0);
 227 }
 228
 229 static void mds_allow_write_access(struct inode *inode)
 230 {
 231         ENTRY;
 232         atomic_inc(&inode->i_writecount);
 233 }
 234
 235 int mds_query_write_access(struct inode *inode)
 236 {
 237         ENTRY;
 238         RETURN(atomic_read(&inode->i_writecount));
 239 }
 240
 241 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 242 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 243                                              struct vfsmount *mnt, int flags,
 244                                              struct ptlrpc_request *req)
 245 {
 246         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 247         struct mds_obd *mds = mds_req2mds(req);
 248         struct mds_file_data *mfd;
 249         struct mds_body *body;
 250         int error;
 251         ENTRY;
 252
 253         mfd = mds_mfd_new();
 254         if (mfd == NULL) {
 255                 CERROR("mds: out of memory\n");
 256                 GOTO(cleanup_dentry, error = -ENOMEM);
 257         }
 258
 259         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 260
 261         if (flags & FMODE_WRITE) {
 262                 /* FIXME: in recovery, need to pass old epoch here */
 263                 error = mds_get_write_access(mds, dentry->d_inode, 0);
 264                 if (error)
 265                         GOTO(cleanup_mfd, error);
 266 #ifdef IFILTERDATA_ACTUALLY_USED
 267                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
 268 #endif /*IFILTERDATA_ACTUALLY_USED*/
 269         } else if (flags & FMODE_EXEC) {
 270                 error = mds_deny_write_access(mds, dentry->d_inode);
 271                 if (error)
 272                         GOTO(cleanup_mfd, error);
 273         }
 274
 275         dget(dentry);
 276
 277         /* Mark the file as open to handle open-unlink. */
 278
 279         DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 280         mds_orphan_open_inc(dentry->d_inode);
 281         UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 282
 283         mfd->mfd_mode = flags;
 284         mfd->mfd_dentry = dentry;
 285         mfd->mfd_xid = req->rq_xid;
 286
 287         spin_lock(&med->med_open_lock);
 288         list_add(&mfd->mfd_list, &med->med_open_head);
 289         spin_unlock(&med->med_open_lock);
 290         mds_mfd_put(mfd);
 291
 292         body->handle.cookie = mfd->mfd_handle.h_cookie;
 293
 294         RETURN(mfd);
 295
 296 cleanup_mfd:
 297         mds_mfd_put(mfd);
 298         mds_mfd_destroy(mfd);
 299 cleanup_dentry:
 300         return ERR_PTR(error);
 301 }
 302
 303 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 304                                 struct lov_desc *desc)
 305 {
 306         int i;
 307         for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
 308                 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
 309                         le64_to_cpu(lmm->lmm_objects[i].l_object_id);
 310         }
 311 }
 312
 313 /* Must be called with i_sem held */
 314 static int mds_create_objects(struct ptlrpc_request *req, int offset,
 315                               struct mds_update_record *rec,
 316                               struct mds_obd *mds, struct obd_device *obd,
 317                               struct dentry *dchild, void **handle,
 318                               obd_id **ids)
 319 {
 320         struct obdo *oa = NULL;
 321         struct obd_trans_info oti = { 0 };
 322         struct mds_body *body;
 323         struct lov_stripe_md *lsm = NULL;
 324         struct lov_mds_md *lmm = NULL;
 325         struct inode *inode = dchild->d_inode;
 326         void *lmm_buf;
 327         int rc, lmm_bufsize, lmm_size;
 328         ENTRY;
 329
 330         if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
 331             !(rec->ur_flags & FMODE_WRITE))
 332                 RETURN(0);
 333
 334         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 335
 336         if (!S_ISREG(inode->i_mode))
 337                 RETURN(0);
 338         if (body->valid & OBD_MD_FLEASIZE)
 339                 RETURN(0);
 340
 341         OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 342         if (*ids == NULL)
 343                 RETURN(-ENOMEM);
 344         oti.oti_objid = *ids;
 345
 346         /* replay case */
 347         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 348                 LASSERT(id_ino(rec->ur_id2));
 349                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 350                 lmm_size = rec->ur_eadatalen;
 351                 lmm = rec->ur_eadata;
 352                 LASSERT(lmm);
 353
 354                 if (*handle == NULL)
 355                         *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
 356                 if (IS_ERR(*handle)) {
 357                         rc = PTR_ERR(*handle);
 358                         *handle = NULL;
 359                         GOTO(out_oa, rc);
 360                 }
 361
 362                 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 363
 364                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 365                 lmm_bufsize = req->rq_repmsg->buflens[offset];
 366                 LASSERT(lmm_buf);
 367                 LASSERT(lmm_bufsize >= lmm_size);
 368                 memcpy(lmm_buf, lmm, lmm_size);
 369                 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 370                 if (rc)
 371                         CERROR("open replay failed to set md:%d\n", rc);
 372                 RETURN(0);
 373         }
 374
 375         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
 376                 GOTO(out_ids, rc = -ENOMEM);
 377
 378         oa = obdo_alloc();
 379         if (oa == NULL)
 380                 GOTO(out_ids, rc = -ENOMEM);
 381         oa->o_mode = S_IFREG | 0600;
 382         oa->o_id = inode->i_ino;
 383         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 384         oa->o_generation = inode->i_generation;
 385         oa->o_uid = 0; /* must have 0 uid / gid on OST */
 386         oa->o_gid = 0;
 387         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
 388                 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
 389         oa->o_size = 0;
 390
 391         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
 392                         OBD_MD_FLCTIME);
 393
 394         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
 395                 /* check if things like lfs setstripe are sending us the ea */
 396                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
 397                         rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
 398                                            mds->mds_dt_exp,
 399                                            0, &lsm, rec->ur_eadata);
 400                         if (rc)
 401                                 GOTO(out_oa, rc);
 402                 } else {
 403                         /* Per-directory striping default code removed, because
 404                          * it uses the same unnamed EA storage as the directory
 405                          * striping for CMD. -p */
 406                 }
 407                 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
 408                 rc = obd_create(mds->mds_dt_exp, oa, &lsm, &oti);
 409                 if (rc) {
 410                         int level = D_ERROR;
 411                         if (rc == -ENOSPC)
 412                                 level = D_INODE;
 413                         CDEBUG(level, "error creating objects for "
 414                                       "inode %lu: rc = %d\n",
 415                                inode->i_ino, rc);
 416                         if (rc > 0) {
 417                                 CERROR("obd_create returned invalid "
 418                                        "rc %d\n", rc);
 419                                 rc = -EIO;
 420                         }
 421                         GOTO(out_oa, rc);
 422                 }
 423         } else {
 424                 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
 425                                    0, &lsm, rec->ur_eadata);
 426                 if (rc) {
 427                         GOTO(out_oa, rc);
 428                 }
 429                 lsm->lsm_object_id = oa->o_id;
 430                 lsm->lsm_object_gr = oa->o_gr;
 431         }
 432         if (inode->i_size) {
 433                 oa->o_size = inode->i_size;
 434                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
 435                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
 436                 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
 437                 if (rc) {
 438                         CERROR("error setting attrs for inode %lu: rc %d\n",
 439                                inode->i_ino, rc);
 440                         if (rc > 0) {
 441                                 CERROR("obd_setattr returned bad rc %d\n", rc);
 442                                 rc = -EIO;
 443                         }
 444                         GOTO(out_oa, rc);
 445                 }
 446         }
 447
 448         body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
 449         obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
 450
 451         LASSERT(lsm && lsm->lsm_object_id);
 452         lmm = NULL;
 453         rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
 454         if (!id_ino(rec->ur_id2))
 455                 obd_free_memmd(mds->mds_dt_exp, &lsm);
 456         LASSERT(rc >= 0);
 457         lmm_size = rc;
 458         body->eadatasize = rc;
 459
 460         if (*handle == NULL)
 461                 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
 462         if (IS_ERR(*handle)) {
 463                 rc = PTR_ERR(*handle);
 464                 *handle = NULL;
 465                 GOTO(out_ids, rc);
 466         }
 467
 468         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
 469         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
 470         lmm_bufsize = req->rq_repmsg->buflens[offset];
 471         LASSERT(lmm_buf);
 472         LASSERT(lmm_bufsize >= lmm_size);
 473
 474         memcpy(lmm_buf, lmm, lmm_size);
 475         obd_free_diskmd(mds->mds_dt_exp, &lmm);
 476  out_oa:
 477         oti_free_cookies(&oti);
 478         obdo_free(oa);
 479  out_ids:
 480         if (rc) {
 481                 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
 482                 *ids = NULL;
 483         }
 484         RETURN(rc);
 485 }
 486
 487 static void reconstruct_open(struct mds_update_record *rec, int offset,
 488                              struct ptlrpc_request *req,
 489                              struct lustre_handle *child_lockh)
 490 {
 491         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 492         struct mds_client_data *mcd = med->med_mcd;
 493         struct mds_obd *mds = mds_req2mds(req);
 494         struct mds_file_data *mfd;
 495         struct obd_device *obd = req->rq_export->exp_obd;
 496         struct dentry *parent, *dchild;
 497         struct ldlm_reply *rep;
 498         struct mds_body *body;
 499         struct list_head *t;
 500         int put_child = 1;
 501         int rc;
 502         ENTRY;
 503
 504         LASSERT(offset == 3); /* only called via intent */
 505         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 506         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 507
 508         /* copy rc, transno and disp; steal locks */
 509         mds_req_from_mcd(req, mcd);
 510         intent_set_disposition(rep, mcd->mcd_last_data);
 511
 512         /* Only replay if create or open actually happened. */
 513         if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
 514                 EXIT;
 515                 return; /* error looking up parent or child */
 516         }
 517
 518         parent = mds_id2dentry(obd, rec->ur_id1, NULL);
 519         LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
 520                  OLID4(rec->ur_id1), PTR_ERR(parent));
 521
 522         dchild = ll_lookup_one_len(rec->ur_name, parent,
 523                                    rec->ur_namelen - 1);
 524         LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
 525                  OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 526
 527         if (!dchild->d_inode)
 528                 GOTO(out_dput, 0); /* child not present to open */
 529
 530         /* At this point, we know we have a child. We'll send
 531          * it back _unless_ it not created and open failed.  */
 532         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
 533             !intent_disposition(rep, DISP_OPEN_CREATE) &&
 534             req->rq_status)
 535                 GOTO(out_dput, 0);
 536
 537         /* get lock (write for O_CREAT, read otherwise) */
 538         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
 539         if (S_ISREG(dchild->d_inode->i_mode)) {
 540                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 541                                  dchild->d_inode, 1);
 542
 543                 if (rc)
 544                         LASSERT(rc == req->rq_status);
 545
 546                 /* If we have LOV EA data, the OST holds size, mtime */
 547                 if (!(body->valid & OBD_MD_FLEASIZE))
 548                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 549                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 550         }
 551
 552         /* If we have -EEXIST as the status, and we were asked to create
 553          * exclusively, we can tell we failed because the file already existed.
 554          */
 555         if (req->rq_status == -EEXIST &&
 556             ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
 557              (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
 558                 GOTO(out_dput, 0);
 559         }
 560
 561         /* If we didn't get as far as trying to open, then some locking thing
 562          * probably went wrong, and we'll just bail here.
 563          */
 564         if (!intent_disposition(rep, DISP_OPEN_OPEN))
 565                 GOTO(out_dput, 0);
 566
 567         /* If we failed, then we must have failed opening, so don't look for
 568          * file descriptor or anything, just give the client the bad news.
 569          */
 570         if (req->rq_status)
 571                 GOTO(out_dput, 0);
 572
 573         mfd = NULL;
 574         list_for_each(t, &med->med_open_head) {
 575                 mfd = list_entry(t, struct mds_file_data, mfd_list);
 576                 if (mfd->mfd_xid == req->rq_xid)
 577                         break;
 578                 mfd = NULL;
 579         }
 580
 581         /* #warning "XXX fixme" bug 2991 */
 582         /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
 583          * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
 584          * to detect a re-open */
 585         if (mfd == NULL) {
 586                 mntget(mds->mds_vfsmnt);
 587                 CERROR("Re-opened file \n");
 588                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
 589                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
 590                 if (!mfd) {
 591                         CERROR("mds: out of memory\n");
 592                         GOTO(out_dput, req->rq_status = -ENOMEM);
 593                 }
 594                 put_child = 0;
 595         } else {
 596                 body->handle.cookie = mfd->mfd_handle.h_cookie;
 597                 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
 598                        mfd->mfd_handle.h_cookie);
 599         }
 600
 601  out_dput:
 602         if (put_child)
 603                 l_dput(dchild);
 604         l_dput(parent);
 605         EXIT;
 606 }
 607
 608 /* do NOT or the MAY_*'s, you'll get the weakest */
 609 static int accmode(int flags)
 610 {
 611         int res = 0;
 612
 613         if (flags & FMODE_READ)
 614                 res = MAY_READ;
 615         if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 616                 res |= MAY_WRITE;
 617         if (flags & FMODE_EXEC)
 618                 res = MAY_EXEC;
 619         return res;
 620 }
 621
 622 /* Handles object creation, actual opening, and I/O epoch */
 623 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
 624                            struct mds_body *body, int flags, void **handle,
 625                            struct mds_update_record *rec,struct ldlm_reply *rep)
 626 {
 627         struct mds_obd *mds = mds_req2mds(req);
 628         struct obd_device *obd = req->rq_export->exp_obd;
 629         struct mds_file_data *mfd = NULL;
 630         obd_id *ids = NULL; /* object IDs created */
 631         unsigned mode;
 632         int rc = 0;
 633         ENTRY;
 634
 635         /* atomically create objects if necessary */
 636         down(&dchild->d_inode->i_sem);
 637         mode = dchild->d_inode->i_mode;
 638         if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
 639             (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
 640                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 641                                  dchild->d_inode, 0);
 642                 if (rc) {
 643                         up(&dchild->d_inode->i_sem);
 644                         RETURN(rc);
 645                 }
 646         }
 647         if (rec != NULL) {
 648                 /* no EA: create objects */
 649                 if ((body->valid & OBD_MD_FLEASIZE) &&
 650                     (rec->ur_flags & MDS_OPEN_HAS_EA)) {
 651                         up(&dchild->d_inode->i_sem);
 652                         RETURN(-EEXIST);
 653                 }
 654                 if (!(body->valid & OBD_MD_FLEASIZE)) {
 655                         /* no EA: create objects */
 656                         rc = mds_create_objects(req, 2, rec, mds, obd,
 657                                                 dchild, handle, &ids);
 658                         if (rc) {
 659                                 CERROR("mds_create_objects: rc = %d\n", rc);
 660                                 up(&dchild->d_inode->i_sem);
 661                                 RETURN(rc);
 662                         }
 663                 }
 664                 if (S_ISREG(dchild->d_inode->i_mode) &&
 665                     (body->valid & OBD_MD_FLEASIZE)) {
 666                         rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
 667                                                    req->rq_repmsg, 2);
 668                         if (!rc)
 669                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
 670                                                  dchild->d_inode, 0);
 671                         if (rc) {
 672                                 up(&dchild->d_inode->i_sem);
 673                                 RETURN(rc);
 674                         }
 675                 }
 676         }
 677         rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
 678         if (rc < 0) {
 679                 CERROR("mds_pack_acl: rc = %d\n", rc);
 680                 up(&dchild->d_inode->i_sem);
 681                 RETURN(rc);
 682         }
 683
 684         /* If the inode has no EA data, then MDSs hold size, mtime */
 685         if (S_ISREG(dchild->d_inode->i_mode) &&
 686             !(body->valid & OBD_MD_FLEASIZE)) {
 687                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 688                                 OBD_MD_FLATIME | OBD_MD_FLMTIME);
 689         }
 690         up(&dchild->d_inode->i_sem);
 691
 692         intent_set_disposition(rep, DISP_OPEN_OPEN);
 693         mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
 694         if (IS_ERR(mfd))
 695                 RETURN(PTR_ERR(mfd));
 696
 697         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
 698                mfd->mfd_handle.h_cookie);
 699         if (ids != NULL) {
 700                 mds_dt_update_objids(obd, ids);
 701                 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
 702         }
 703         //if (rc)
 704         //        mds_mfd_destroy(mfd);
 705         RETURN(rc);
 706 }
 707
 708 static int mds_open_by_id(struct ptlrpc_request *req,
 709                           struct lustre_id *id,
 710                           struct mds_body *body, int flags,
 711                           struct mds_update_record *rec,
 712                           struct ldlm_reply *rep)
 713 {
 714         struct mds_obd *mds = mds_req2mds(req);
 715         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
 716         struct dentry *dchild;
 717         char idname[LL_ID_NAMELEN];
 718         int idlen = 0, rc;
 719         void *handle = NULL;
 720         ENTRY;
 721
 722         down(&pending_dir->i_sem);
 723
 724         idlen = ll_id2str(idname, id_ino(id), id_gen(id));
 725
 726         dchild = lookup_one_len(idname, mds->mds_pending_dir,
 727                                 idlen);
 728         if (IS_ERR(dchild)) {
 729                 up(&pending_dir->i_sem);
 730                 rc = PTR_ERR(dchild);
 731                 CERROR("error looking up %s in PENDING: rc = %d\n",
 732                        idname, rc);
 733                 RETURN(rc);
 734         }
 735
 736         if (dchild->d_inode != NULL) {
 737                 up(&pending_dir->i_sem);
 738                 mds_inode_set_orphan(dchild->d_inode);
 739                 mds_pack_inode2body(req2obd(req), body,
 740                                     dchild->d_inode, 1);
 741
 742                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 743                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 744                 CWARN("Orphan %s found and opened in PENDING directory\n",
 745                        idname);
 746                 goto open;
 747         }
 748         dput(dchild);
 749         up(&pending_dir->i_sem);
 750
 751         /*
 752          * we didn't find it in PENDING so it isn't an orphan.  See if it was a
 753          * regular inode that was previously created.
 754          */
 755         dchild = mds_id2dentry(req2obd(req), id, NULL);
 756         if (IS_ERR(dchild))
 757                 RETURN(PTR_ERR(dchild));
 758
 759         mds_pack_inode2body(req2obd(req), body,
 760                             dchild->d_inode, 1);
 761
 762         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 763         intent_set_disposition(rep, DISP_LOOKUP_POS);
 764
 765  open:
 766         rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
 767         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
 768                                 req, rc, rep ? rep->lock_policy_res1 : 0);
 769         /* XXX what do we do here if mds_finish_transno itself failed? */
 770
 771         l_dput(dchild);
 772         RETURN(rc);
 773 }
 774
 775 int mds_pin(struct ptlrpc_request *req, int offset)
 776 {
 777         struct obd_device *obd = req->rq_export->exp_obd;
 778         struct mds_body *request_body, *reply_body;
 779         struct lvfs_run_ctxt saved;
 780         int rc, size = sizeof(*reply_body);
 781         ENTRY;
 782
 783         request_body = lustre_msg_buf(req->rq_reqmsg, offset,
 784                                       sizeof(*request_body));
 785
 786         rc = lustre_pack_reply(req, 1, &size, NULL);
 787         if (rc)
 788                 RETURN(rc);
 789         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 790
 791         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 792         rc = mds_open_by_id(req, &request_body->id1, reply_body,
 793                             request_body->flags, NULL, NULL);
 794         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 795
 796         RETURN(rc);
 797 }
 798
 799 /*
 800  * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
 801  * child_lockh is NULL we just get the lock as a barrier to wait for other
 802  * holders of this lock, and drop it right away again.
 803  */
 804 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
 805                        struct lustre_handle *child_lockh)
 806 {
 807         struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
 808         struct lustre_handle lockh;
 809         int lock_flags = 0;
 810         int rc;
 811
 812         if (child_lockh == NULL)
 813                 child_lockh = &lockh;
 814
 815         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
 816                               LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
 817                               mds_blocking_ast, ldlm_completion_ast, NULL,
 818                               NULL, NULL, 0, NULL, child_lockh);
 819         if (rc != ELDLM_OK)
 820                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 821         else if (child_lockh == &lockh)
 822                 ldlm_lock_decref(child_lockh, LCK_EX);
 823
 824         RETURN(rc);
 825 }
 826
 827 static int is_mount_object(struct dentry *dparent)
 828 {
 829         struct dentry *dchild;
 830
 831         if (!(dparent->d_inode->i_mode & S_ISUID))
 832                 return 0;
 833
 834         dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
 835         if (IS_ERR(dchild) || dchild == NULL)
 836                 return 0;
 837
 838         dput(dchild);
 839         return 1;
 840 }
 841
 842 int mds_open(struct mds_update_record *rec, int offset,
 843              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 844 {
 845         /* XXX ALLOCATE _something_ - 464 bytes on stack here */
 846         struct obd_device *obd = req->rq_export->exp_obd;
 847         struct mds_obd *mds = mds_req2mds(req);
 848         struct ldlm_reply *rep = NULL;
 849         struct mds_body *body = NULL;
 850         struct dentry *dchild = NULL, *dparent = NULL;
 851         struct mds_export_data *med;
 852         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 853         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
 854         int parent_mode = LCK_PR;
 855         void *handle = NULL;
 856         struct dentry_params dp;
 857         struct mea *mea = NULL;
 858         int mea_size, update_mode;
 859         ENTRY;
 860
 861         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
 862                   OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
 863                   rec->ur_mode);
 864
 865         if (offset == 3) { /* intent */
 866                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 867                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 868         } else if (offset == 1) { /* non-intent reint */
 869                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 870         } else {
 871                 body = NULL;
 872                 LBUG();
 873         }
 874
 875         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 876
 877         /*
 878          * Step 0: If we are passed a id, then we assume the client already
 879          * opened this file and is only replaying the RPC, so we open the inode
 880          * by fid (at some large expense in security).
 881          */
 882         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 883                 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
 884                        OLID4(rec->ur_id2), rec->ur_namelen - 1,
 885                        rec->ur_name, rec->ur_mode);
 886
 887                 LASSERT(id_ino(rec->ur_id2));
 888
 889                 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
 890                                     rec, rep);
 891                 if (rc != -ENOENT)
 892                         RETURN(rc);
 893
 894                 /* We didn't find the correct inode on disk either, so we
 895                  * need to re-create it via a regular replay. */
 896                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
 897         } else {
 898                 LASSERT(!id_ino(rec->ur_id2));
 899         }
 900
 901         LASSERT(offset == 3); /* If we got here, we must be called via intent */
 902
 903         med = &req->rq_export->exp_mds_data;
 904         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 905                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 906                 RETURN(-ENOMEM);
 907         }
 908
 909         acc_mode = accmode(rec->ur_flags);
 910
 911         /* Step 1: Find and lock the parent */
 912         if (rec->ur_flags & O_CREAT) {
 913                 /* XXX Well, in fact we only need this lock mode change if
 914                    in addition to O_CREAT, the file does not exist.
 915                    But we do not know if it exists or not yet */
 916                 parent_mode = LCK_PW;
 917         }
 918
 919         if (rec->ur_namelen == 1) {
 920                 /* client (LMV) wants to open the file by id */
 921                 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
 922                 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
 923                 if (IS_ERR(dchild)) {
 924                         rc = PTR_ERR(dparent);
 925                         CERROR("child lookup by an id error %d\n", rc);
 926                         GOTO(cleanup, rc);
 927                 }
 928                 goto got_child;
 929         }
 930
 931         dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
 932                                        parent_lockh, &update_mode, rec->ur_name,
 933                                        rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
 934         if (IS_ERR(dparent)) {
 935                 rc = PTR_ERR(dparent);
 936                 if (rc != -ENOENT)
 937                         CERROR("parent lookup for "DLID4" failed, error %d\n",
 938                                OLID4(rec->ur_id1), rc);
 939                 GOTO(cleanup, rc);
 940         }
 941         LASSERT(dparent->d_inode != NULL);
 942
 943         cleanup_phase = 1; /* parent dentry and lock */
 944
 945         /* try to retrieve MEA data for this dir */
 946         rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
 947         if (rc)
 948                 GOTO(cleanup, rc);
 949
 950         if (mea != NULL) {
 951                 /*
 952                  * dir is already splitted, check is requested filename should *
 953                  * live at this MDS or at another one.
 954                  */
 955                 int i;
 956                 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
 957                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
 958                         CDEBUG(D_OTHER,
 959                                "%s: inapropriate MDS(%d) for %lu/%u:%s."
 960                                " should be %lu(%d)\n", obd->obd_name,
 961                                mea->mea_master, dparent->d_inode->i_ino,
 962                                dparent->d_inode->i_generation, rec->ur_name,
 963                                (unsigned long)id_group(&mea->mea_ids[i]), i);
 964                         GOTO(cleanup, rc = -ERESTART);
 965                 }
 966         }
 967
 968         /* Step 2: Lookup the child */
 969         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
 970         if (IS_ERR(dchild)) {
 971                 rc = PTR_ERR(dchild);
 972                 dchild = NULL; /* don't confuse mds_finish_transno() below */
 973                 GOTO(cleanup, rc);
 974         }
 975
 976 got_child:
 977         cleanup_phase = 2; /* child dentry */
 978
 979         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 980
 981         if (dchild->d_flags & DCACHE_CROSS_REF) {
 982                 struct ldlm_res_id res_id = { . name = {0} };
 983                 ldlm_policy_data_t policy;
 984                 int flags = 0;
 985
 986                 mds_pack_dentry2body(obd, body, dchild, 1);
 987                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 988
 989                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
 990                        OLID4(&body->id1));
 991
 992                 res_id.name[0] = dchild->d_fid;
 993                 res_id.name[1] = dchild->d_mdsnum;
 994
 995                 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
 996                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 997                                       res_id, LDLM_IBITS, &policy,
 998                                       LCK_PR, &flags, mds_blocking_ast,
 999                                       ldlm_completion_ast, NULL, NULL,
1000                                       NULL, 0, NULL, child_lockh);
1001 #ifdef S_PDIROPS
1002                 if (parent_lockh[1].cookie != 0)
1003                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1004 #endif
1005                 ldlm_lock_decref(parent_lockh, parent_mode);
1006                 if (mea)
1007                         OBD_FREE(mea, mea_size);
1008                 l_dput(dchild);
1009                 l_dput(dparent);
1010                 RETURN(rc);
1011         }
1012
1013         if (dchild->d_inode)
1014                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1015         else
1016                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1017
1018         /* Step 3: If the child was negative, and we're supposed to, create it.*/
1019         if (dchild->d_inode == NULL) {
1020                 unsigned long ino;
1021                 struct iattr iattr;
1022                 struct inode *inode;
1023
1024                 ino = (unsigned long)id_ino(rec->ur_id2);
1025
1026                 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1027                                           update_mode);
1028
1029                 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1030                        obd->obd_name, dparent->d_inode->i_ino,
1031                        dparent->d_inode->i_generation, rc);
1032
1033                 if (rc > 0) {
1034                         /* dir got splitted */
1035                         GOTO(cleanup, rc = -ERESTART);
1036                 } else if (rc < 0) {
1037                         /* error happened during spitting */
1038                         GOTO(cleanup, rc);
1039                 }
1040
1041                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1042                         /*
1043                          * dentry is negative and we weren't supposed to create
1044                          * object, get out of here.
1045                          */
1046                         GOTO(cleanup, rc = -ENOENT);
1047                 }
1048
1049                 intent_set_disposition(rep, DISP_OPEN_CREATE);
1050                 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1051                                       NULL);
1052                 if (IS_ERR(handle)) {
1053                         rc = PTR_ERR(handle);
1054                         handle = NULL;
1055                         GOTO(cleanup, rc);
1056                 }
1057                 dchild->d_fsdata = (void *) &dp;
1058                 dp.p_ptr = req;
1059                 dp.p_inum = ino;
1060
1061                 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1062                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1063                         dchild->d_fsdata = NULL;
1064
1065                 if (rc) {
1066                         CDEBUG(D_INODE, "error during create: %d\n", rc);
1067                         GOTO(cleanup, rc);
1068                 }
1069                 inode = dchild->d_inode;
1070                 if (ino) {
1071                         LASSERT(ino == inode->i_ino);
1072
1073                         /* written as part of setattr */
1074                         inode->i_generation = id_gen(rec->ur_id2);
1075                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1076                                inode->i_ino, inode->i_generation);
1077                 }
1078
1079                 created = 1;
1080                 LTIME_S(iattr.ia_atime) = rec->ur_time;
1081                 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1082                 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1083
1084                 iattr.ia_uid = rec->ur_fsuid;
1085                 if (dparent->d_inode->i_mode & S_ISGID)
1086                         iattr.ia_gid = dparent->d_inode->i_gid;
1087                 else
1088                         iattr.ia_gid = rec->ur_fsgid;
1089
1090                 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1091                         ATTR_MTIME | ATTR_CTIME;
1092
1093                 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1094                 if (rc)
1095                         CERROR("error on child setattr: rc = %d\n", rc);
1096
1097                 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1098
1099                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1100                 if (rc)
1101                         CERROR("error on parent setattr: rc = %d\n", rc);
1102                 else {
1103                         MD_COUNTER_INCREMENT(obd, create);
1104                 }
1105
1106                 down(&dchild->d_inode->i_sem);
1107                 if (ino) {
1108                         rc = mds_update_inode_sid(obd, dchild->d_inode,
1109                                                   handle, rec->ur_id2);
1110                         if (rc) {
1111                                 CERROR("mds_update_inode_sid() failed, "
1112                                        "rc = %d\n", rc);
1113                         }
1114                         id_assign_fid(&body->id1, rec->ur_id2);
1115
1116                         /*
1117                          * make sure, that fid is up-to-date.
1118                          */
1119                         mds_set_last_fid(obd, id_fid(rec->ur_id2));
1120                 } else {
1121                         rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1122                                                  handle, &body->id1);
1123                         if (rc) {
1124                                 CERROR("mds_alloc_inode_sid() failed, "
1125                                        "rc = %d\n", rc);
1126                         }
1127                 }
1128                 up(&dchild->d_inode->i_sem);
1129
1130                 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1131                         rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1132                                            dchild->d_inode, handle, 0);
1133                         handle = NULL;
1134                 }
1135
1136                 acc_mode = 0;           /* Don't check for permissions */
1137         }
1138         mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1139
1140         LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1141                  "dchild %*s (%p) inode %p\n", dchild->d_name.len,
1142                  dchild->d_name.name, dchild, dchild->d_inode);
1143
1144         if (S_ISREG(dchild->d_inode->i_mode)) {
1145                 /* Check permissions etc */
1146                 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1147                 if (rc != 0)
1148                         GOTO(cleanup, rc);
1149
1150                 /* Can't write to a read-only file */
1151                 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1152                         GOTO(cleanup, rc = -EPERM);
1153
1154                 /* An append-only file must be opened in append mode for
1155                  * writing */
1156                 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1157                     ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1158                      (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1159                         GOTO(cleanup, rc = -EPERM);
1160         }
1161
1162         if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1163             (rec->ur_flags & MDS_OPEN_EXCL)) {
1164                 /* File already exists, we didn't just create it, and we
1165                  * were passed O_EXCL; err-or. */
1166                 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1167         }
1168
1169         /* if we are following a symlink, don't open */
1170         if (S_ISLNK(dchild->d_inode->i_mode))
1171                 GOTO(cleanup, rc = 0);
1172
1173         if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1174             !S_ISDIR(dchild->d_inode->i_mode))
1175                 GOTO(cleanup, rc = -ENOTDIR);
1176
1177         if (S_ISDIR(dchild->d_inode->i_mode)) {
1178                 if (rec->ur_flags & MDS_OPEN_CREAT ||
1179                     rec->ur_flags & FMODE_WRITE) {
1180                         /*we are trying to create or write a exist dir*/
1181                         GOTO(cleanup, rc = -EISDIR);
1182                 }
1183                 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1184                         GOTO(cleanup, rc = -EACCES);
1185
1186                 if (is_mount_object(dchild)) {
1187                         CERROR("Found possible GNS mount object %*s; not "
1188                                "opening.\n", dchild->d_name.len,
1189                                dchild->d_name.name);
1190                         GOTO(cleanup, rc = 0); // success, but don't really open
1191                 }
1192         }
1193
1194         /* Step 5: mds_open it */
1195         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1196                              rec, rep);
1197         GOTO(cleanup, rc);
1198
1199  cleanup:
1200         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1201                                 req, rc, rep ? rep->lock_policy_res1 : 0);
1202
1203         switch (cleanup_phase) {
1204         case 2:
1205                 if (rc && created) {
1206                         int err = vfs_unlink(dparent->d_inode, dchild);
1207                         if (err) {
1208                                 CERROR("unlink(%*s) in error path: %d\n",
1209                                        dchild->d_name.len, dchild->d_name.name,
1210                                        err);
1211                         }
1212                 } else if (created) {
1213                         mds_lock_new_child(obd, dchild->d_inode, NULL);
1214                 }
1215                 l_dput(dchild);
1216         case 1:
1217                 if (dparent == NULL)
1218                         break;
1219
1220                 l_dput(dparent);
1221 #ifdef S_PDIROPS
1222                 if (parent_lockh[1].cookie != 0)
1223                         ldlm_lock_decref(parent_lockh + 1, update_mode);
1224 #endif
1225                 if (rc)
1226                         ldlm_lock_decref(parent_lockh, parent_mode);
1227                 else
1228                         ptlrpc_save_lock (req, parent_lockh, parent_mode);
1229         }
1230         if (mea)
1231                 OBD_FREE(mea, mea_size);
1232         RETURN(rc);
1233 }
1234
1235 /* Close a "file descriptor" and possibly unlink an orphan from the
1236  * PENDING directory.  Caller must hold child->i_sem, this drops it.
1237  *
1238  * If we are being called from mds_disconnect() because the client has
1239  * disappeared, then req == NULL and we do not update last_rcvd because
1240  * there is nothing that could be recovered by the client at this stage
1241  * (it will not even _have_ an entry in last_rcvd anymore).
1242  *
1243  * Returns EAGAIN if the client needs to get more data and re-close. */
1244 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1245                   struct obd_device *obd, struct mds_file_data *mfd,
1246                   int unlink_orphan)
1247 {
1248         struct inode *inode = mfd->mfd_dentry->d_inode;
1249         char idname[LL_ID_NAMELEN];
1250         int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1251         struct dentry *pending_child = NULL;
1252         struct mds_obd *mds = &obd->u.mds;
1253         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1254         void *handle = NULL;
1255         struct mds_body *request_body = NULL, *reply_body = NULL;
1256         struct dentry_params dp;
1257         struct iattr iattr = { 0 };
1258         struct llog_create_locks *lcl = NULL;
1259         ENTRY;
1260
1261         if (req && req->rq_reqmsg != NULL)
1262                 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1263                                               sizeof(*request_body));
1264         if (req && req->rq_repmsg != NULL)
1265                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1266                                             sizeof(*reply_body));
1267
1268         idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1269         CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1270                idname, inode->i_nlink, mds_orphan_open_count(inode));
1271
1272         last_orphan = mds_orphan_open_dec_test(inode) &&
1273                 mds_inode_is_orphan(inode);
1274         UP_WRITE_I_ALLOC_SEM(inode);
1275
1276         /* this is half of the actual "close" */
1277         if (mfd->mfd_mode & FMODE_WRITE) {
1278                 rc = mds_put_write_access(mds, inode, request_body,
1279                                           last_orphan && unlink_orphan);
1280         } else if (mfd->mfd_mode & FMODE_EXEC) {
1281                 mds_allow_write_access(inode);
1282         }
1283
1284         if (last_orphan && unlink_orphan) {
1285                 struct lov_mds_md *lmm = NULL;
1286                 int stripe_count = 0;
1287                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1288
1289                 if (obd->obd_recovering) {
1290                         CDEBUG(D_HA, "not remove orphan %s until recovery"
1291                                " is over\n", idname);
1292                         GOTO(out, rc);
1293                 }
1294
1295                 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1296
1297                 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1298                     (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1299                         CERROR("found \"orphan\" %s %s with link count %d\n",
1300                                S_ISREG(inode->i_mode) ? "file" : "dir",
1301                                idname, inode->i_nlink);
1302                 /*
1303                  * sadly, there is no easy way to save pending_child from
1304                  * mds_reint_unlink() into mfd, so we need to re-lookup, but
1305                  * normally it will still be in the dcache.
1306                  */
1307                 down(&pending_dir->i_sem);
1308                 cleanup_phase = 1; /* up(i_sem) when finished */
1309                 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1310                                                idlen);
1311                 if (IS_ERR(pending_child))
1312                         GOTO(cleanup, rc = PTR_ERR(pending_child));
1313                 LASSERT(pending_child->d_inode != NULL);
1314
1315                 cleanup_phase = 2; /* dput(pending_child) when finished */
1316                 if (req != NULL && req->rq_repmsg != NULL) {
1317                         lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1318                         stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1319                 }
1320
1321                 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1322                                           NULL, stripe_count);
1323                 if (IS_ERR(handle)) {
1324                         rc = PTR_ERR(handle);
1325                         handle = NULL;
1326                         GOTO(cleanup, rc);
1327                 }
1328
1329                 pending_child->d_fsdata = (void *) &dp;
1330                 dp.p_inum = 0;
1331                 dp.p_ptr = req;
1332                 if (S_ISDIR(pending_child->d_inode->i_mode))
1333                         rc = vfs_rmdir(pending_dir, pending_child);
1334                 else
1335                         rc = vfs_unlink(pending_dir, pending_child);
1336                 if (rc)
1337                         CERROR("error unlinking orphan %s: rc %d\n",
1338                                idname, rc);
1339
1340                 if (req != NULL && req->rq_repmsg != NULL &&
1341                     (reply_body->valid & OBD_MD_FLEASIZE) &&
1342                     mds_log_op_unlink(obd, pending_child->d_inode,
1343                                       lmm, req->rq_repmsg->buflens[1],
1344                                       lustre_msg_buf(req->rq_repmsg, 2, 0),
1345                                       req->rq_repmsg->buflens[2], &lcl) > 0) {
1346                         reply_body->valid |= OBD_MD_FLCOOKIE;
1347                 }
1348
1349                 goto out; /* Don't bother updating attrs on unlinked inode */
1350         }
1351
1352 #if 0
1353         if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1354                 /* Update the on-disk attributes if this was the last write
1355                  * close, and all information was provided (i.e., rc == 0)
1356                  *
1357                  * XXX this should probably be abstracted with mds_reint_setattr
1358                  */
1359
1360                 if (request_body->valid & OBD_MD_FLMTIME &&
1361                     LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1362                         LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1363                         iattr.ia_valid |= ATTR_MTIME;
1364                 }
1365                 if (request_body->valid & OBD_MD_FLCTIME &&
1366                     LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1367                         LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1368                         iattr.ia_valid |= ATTR_CTIME;
1369                 }
1370
1371                 /* XXX can't set block count with fsfilt_setattr (!) */
1372                 if (request_body->valid & OBD_MD_FLSIZE) {
1373                         iattr.ia_valid |= ATTR_SIZE;
1374                         iattr.ia_size = request_body->size;
1375                 }
1376                 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1377                         iattr.ia_valid |= ATTR_BLOCKS;
1378                         iattr.ia_blocks = request_body->blocks
1379                 } */
1380
1381         }
1382 #endif
1383         if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1384                 /* Only start a transaction to write out only the atime if
1385                  * it is more out-of-date than the specified limit.  If we
1386                  * are already going to write out the atime then do it anyway.
1387                  * */
1388                 if ((request_body->atime >
1389                      LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1390                     (iattr.ia_valid != 0 &&
1391                      request_body->atime > LTIME_S(inode->i_atime))) {
1392                         LTIME_S(iattr.ia_atime) = request_body->atime;
1393                         iattr.ia_valid |= ATTR_ATIME;
1394                 }
1395         }
1396
1397         if (iattr.ia_valid != 0) {
1398                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1399                 if (IS_ERR(handle))
1400                         GOTO(cleanup, rc = PTR_ERR(handle));
1401                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1402                 if (rc)
1403                         CERROR("error in setattr(%s): rc %d\n", idname, rc);
1404         }
1405 out:
1406         /* If other clients have this file open for write, rc will be > 0 */
1407         if (rc > 0)
1408                 rc = 0;
1409         l_dput(mfd->mfd_dentry);
1410         mds_mfd_destroy(mfd);
1411
1412  cleanup:
1413         if (req != NULL && reply_body != NULL) {
1414                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1415         } else if (handle) {
1416                 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1417                 if (err) {
1418                         CERROR("error committing close: %d\n", err);
1419                         if (!rc)
1420                                 rc = err;
1421                 }
1422         }
1423
1424         switch (cleanup_phase) {
1425         case 2:
1426                 if (lcl != NULL)
1427                         ptlrpc_save_llog_lock(req, lcl);
1428                 dput(pending_child);
1429         case 1:
1430                 up(&pending_dir->i_sem);
1431         }
1432         RETURN(rc);
1433 }
1434
1435 int mds_close(struct ptlrpc_request *req, int offset)
1436 {
1437         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1438         struct obd_device *obd = req->rq_export->exp_obd;
1439         struct mds_body *body;
1440         struct mds_file_data *mfd;
1441         struct lvfs_run_ctxt saved;
1442         struct inode *inode;
1443         int rc, repsize[3] = {sizeof(struct mds_body),
1444                               obd->u.mds.mds_max_mdsize,
1445                               obd->u.mds.mds_max_cookiesize};
1446         ENTRY;
1447         MD_COUNTER_INCREMENT(obd, close);
1448
1449         rc = lustre_pack_reply(req, 3, repsize, NULL);
1450         if (rc) {
1451                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1452                 req->rq_status = rc;
1453         } else {
1454                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1455         }
1456
1457         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1458                 DEBUG_REQ(D_HA, req, "close replay\n");
1459                 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1460                        lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1461                        req->rq_repmsg->buflens[2]);
1462         }
1463
1464
1465         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1466                                   lustre_swab_mds_body);
1467         if (body == NULL) {
1468                 CERROR("Can't unpack body\n");
1469                 req->rq_status = -EFAULT;
1470                 RETURN(-EFAULT);
1471         }
1472
1473         if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1474                 /* do some stuff */ ;
1475
1476         mfd = mds_handle2mfd(&body->handle);
1477         if (mfd == NULL) {
1478                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1479                           ": cookie "LPX64, body->id1.li_stc.u.e3s.l3s_ino,
1480                           body->handle.cookie);
1481                 req->rq_status = -ESTALE;
1482                 RETURN(-ESTALE);
1483         }
1484
1485         inode = mfd->mfd_dentry->d_inode;
1486         /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1487         DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1488         if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1489                 struct mds_body *rep_body;
1490
1491                 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1492                                           sizeof (*rep_body));
1493                 LASSERT(rep_body != NULL);
1494
1495                 mds_pack_inode2body(obd, rep_body, inode,
1496                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1497
1498                 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1499                             inode, MDS_PACK_MD_LOCK);
1500         }
1501         spin_lock(&med->med_open_lock);
1502         list_del(&mfd->mfd_list);
1503         spin_unlock(&med->med_open_lock);
1504
1505         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1506         req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1507         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1508
1509         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1510                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1511                 req->rq_status = -ENOMEM;
1512                 mds_mfd_put(mfd);
1513                 RETURN(-ENOMEM);
1514         }
1515
1516         mds_mfd_put(mfd);
1517         RETURN(0);
1518 }
1519
1520 int mds_done_writing(struct ptlrpc_request *req, int offset)
1521 {
1522         struct mds_body *body;
1523         int rc, size = sizeof(struct mds_body);
1524         ENTRY;
1525
1526         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1527
1528         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1529                                   lustre_swab_mds_body);
1530         if (body == NULL) {
1531                 CERROR("Can't unpack body\n");
1532                 req->rq_status = -EFAULT;
1533                 RETURN(-EFAULT);
1534         }
1535
1536         rc = lustre_pack_reply(req, 1, &size, NULL);
1537         if (rc) {
1538                 CERROR("lustre_pack_reply: rc = %d\n", rc);
1539                 req->rq_status = rc;
1540         }
1541
1542         RETURN(0);
1543 }