lustre/mdc/mdc_locks.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.sf.net/projects/lustre/
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_MDC
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 # include <linux/pagemap.h>
  30 # include <linux/miscdevice.h>
  31 # include <linux/init.h>
  32 #else
  33 # include <liblustre.h>
  34 #endif
  35
  36 #include <linux/obd_class.h>
  37 #include <linux/lustre_mds.h>
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/lprocfs_status.h>
  40 #include <linux/lustre_acl.h>
  41 #include <linux/lustre_lite.h>
  42 #include "mdc_internal.h"
  43
  44 int it_disposition(struct lookup_intent *it, int flag)
  45 {
  46         return LUSTRE_IT(it)->it_disposition & flag;
  47 }
  48 EXPORT_SYMBOL(it_disposition);
  49
  50 void it_set_disposition(struct lookup_intent *it, int flag)
  51 {
  52         LUSTRE_IT(it)->it_disposition |= flag;
  53 }
  54 EXPORT_SYMBOL(it_set_disposition);
  55
  56 static void mdc_id2mdc_data(struct mdc_op_data *data,
  57                             struct lustre_id *f1,
  58                             struct lustre_id *f2,
  59                             const char *name,
  60                             int namelen, int mode)
  61 {
  62         LASSERT(data);
  63         LASSERT(f1);
  64
  65         data->id1 = *f1;
  66         if (f2)
  67                 data->id2 = *f2;
  68
  69         data->valid = 0;
  70         data->name = name;
  71         data->namelen = namelen;
  72         data->create_mode = mode;
  73         data->mod_time = LTIME_S(CURRENT_TIME);
  74 }
  75
  76 static int it_to_lock_mode(struct lookup_intent *it)
  77 {
  78         /* CREAT needs to be tested before open (both could be set) */
  79         if (it->it_op & IT_CREAT)
  80                 return LCK_PW;
  81         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
  82                               IT_CHDIR))
  83                 return LCK_PR;
  84
  85         LBUG();
  86         RETURN(-EINVAL);
  87 }
  88
  89 int it_open_error(int phase, struct lookup_intent *it)
  90 {
  91         if (it_disposition(it, DISP_OPEN_OPEN)) {
  92                 if (phase == DISP_OPEN_OPEN)
  93                         return LUSTRE_IT(it)->it_status;
  94                 else
  95                         return 0;
  96         }
  97
  98         if (it_disposition(it, DISP_OPEN_CREATE)) {
  99                 if (phase == DISP_OPEN_CREATE)
 100                         return LUSTRE_IT(it)->it_status;
 101                 else
 102                         return 0;
 103         }
 104
 105         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
 106                 if (phase == DISP_LOOKUP_EXECD)
 107                         return LUSTRE_IT(it)->it_status;
 108                 else
 109                         return 0;
 110         }
 111
 112         if (it_disposition(it, DISP_IT_EXECD)) {
 113                 if (phase == DISP_IT_EXECD)
 114                         return LUSTRE_IT(it)->it_status;
 115                 else
 116                         return 0;
 117         }
 118         CERROR("it disp: %X, status: %d\n", LUSTRE_IT(it)->it_disposition,
 119                LUSTRE_IT(it)->it_status);
 120         LBUG();
 121         return 0;
 122 }
 123 EXPORT_SYMBOL(it_open_error);
 124
 125 /* this must be called on a lockh that is known to have a referenced lock */
 126 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
 127 {
 128         struct ldlm_lock *lock;
 129         struct lustre_handle *lockh = (struct lustre_handle *)l;
 130         ENTRY;
 131
 132         if (!*l) {
 133                 EXIT;
 134                 return 0;
 135         }
 136
 137         lock = ldlm_handle2lock(lockh);
 138
 139         LASSERT(lock != NULL);
 140         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 141 #ifdef __KERNEL__
 142         if (lock->l_ast_data && lock->l_ast_data != data) {
 143                 struct inode *new_inode = data;
 144                 struct inode *old_inode = lock->l_ast_data;
 145                 LASSERTF(old_inode->i_state & I_FREEING,
 146                          "Found existing inode %p/%lu/%u state %lu in lock: "
 147                          "setting data to %p/%lu/%u\n", old_inode,
 148                          old_inode->i_ino, old_inode->i_generation,
 149                          old_inode->i_state, new_inode, new_inode->i_ino,
 150                          new_inode->i_generation);
 151         }
 152 #endif
 153         lock->l_ast_data = data;
 154         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 155         LDLM_LOCK_PUT(lock);
 156
 157         EXIT;
 158         return 0;
 159 }
 160 EXPORT_SYMBOL(mdc_set_lock_data);
 161
 162 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id,
 163                       ldlm_iterator_t it, void *data)
 164 {
 165         struct ldlm_res_id res_id = { .name = {0} };
 166         ENTRY;
 167
 168         res_id.name[0] = id_fid(id);
 169         res_id.name[1] = id_group(id);
 170
 171         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
 172                            &res_id, it, data);
 173
 174         EXIT;
 175         return 0;
 176 }
 177
 178 static inline void
 179 mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 180 {
 181         /* Don't hold error requests for replay. */
 182         if (req->rq_replay) {
 183                 unsigned long irqflags;
 184                 spin_lock_irqsave(&req->rq_lock, irqflags);
 185                 req->rq_replay = 0;
 186                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
 187         }
 188         if (rc && req->rq_transno != 0) {
 189                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 190                 LBUG();
 191         }
 192 }
 193
 194 /* We always reserve enough space in the reply packet for a stripe MD, because
 195  * we don't know in advance the file type. */
 196 int mdc_enqueue(struct obd_export *exp,
 197                 int lock_type,
 198                 struct lookup_intent *it,
 199                 int lock_mode,
 200                 struct mdc_op_data *data,
 201                 struct lustre_handle *lockh,
 202                 void *lmm,
 203                 int lmmsize,
 204                 ldlm_completion_callback cb_completion,
 205                 ldlm_blocking_callback cb_blocking,
 206                 void *cb_data)
 207 {
 208         struct ptlrpc_request *req;
 209         struct ldlm_res_id res_id = {
 210                 .name = {id_fid(&data->id1), id_group(&data->id1)}
 211         };
 212         struct obd_device *obddev = class_exp2obd(exp);
 213         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 214         struct ldlm_intent *lit;
 215         struct ldlm_request *lockreq;
 216         int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
 217                           [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
 218                           [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
 219         int repsize[5] = {sizeof(struct ldlm_reply),
 220                           sizeof(struct mds_body),
 221                           obddev->u.cli.cl_max_mds_easize};
 222         int req_buffers = 3, reply_buffers = 0;
 223         int rc, flags = LDLM_FL_HAS_INTENT;
 224         struct ldlm_reply *dlm_rep = NULL;
 225         void *eadata;
 226         unsigned long irqflags;
 227         ENTRY;
 228
 229 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 230 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 231
 232         reqsize[0] = mdc_get_secdesc_size();
 233
 234         if (it->it_op & IT_OPEN) {
 235                 it->it_create_mode |= S_IFREG;
 236                 it->it_create_mode &= ~current->fs->umask;
 237
 238                 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
 239                 reqsize[req_buffers++] = data->namelen + 1;
 240                 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
 241
 242                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 243                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 244                 if (!req)
 245                         RETURN(-ENOMEM);
 246
 247                 spin_lock_irqsave (&req->rq_lock, irqflags);
 248                 req->rq_replay = 1;
 249                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 250
 251                 /* pack the intent */
 252                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 253                                      sizeof (*lit));
 254                 lit->opc = (__u64)it->it_op;
 255
 256                 /* pack the intended request */
 257                 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
 258                               it->it_create_mode, 0, it->it_flags, lmm, lmmsize);
 259                 /* get ready for the reply */
 260                 repsize[3] = 4;
 261                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 262                 reply_buffers = 5;
 263                 req->rq_replen = lustre_msg_size(5, repsize);
 264         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 265                 __u64 valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE |
 266                             OBD_MD_FLACL_ACCESS;
 267
 268                 reqsize[req_buffers++] = sizeof(struct mds_body);
 269                 reqsize[req_buffers++] = data->namelen + 1;
 270
 271                 if (it->it_op & IT_GETATTR)
 272                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 273
 274                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 275                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 276
 277                 if (!req)
 278                         RETURN(-ENOMEM);
 279
 280                 /* pack the intent */
 281                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 282                                      sizeof (*lit));
 283                 lit->opc = (__u64)it->it_op;
 284
 285                 /* pack the intended request */
 286                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 287                                  valid, it->it_flags, data);
 288
 289                 /* get ready for the reply */
 290                 repsize[3] = 4;
 291                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 292                 reply_buffers = 5;
 293                 req->rq_replen = lustre_msg_size(5, repsize);
 294         } else if (it->it_op == IT_READDIR) {
 295                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 296                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 297                                       LDLM_ENQUEUE, 2, reqsize, NULL);
 298
 299                 if (!req)
 300                         RETURN(-ENOMEM);
 301                 /* get ready for the reply */
 302                 reply_buffers = 1;
 303                 req->rq_replen = lustre_msg_size(1, repsize);
 304         } else if (it->it_op == IT_UNLINK) {
 305                 reqsize[req_buffers++] = sizeof(struct mds_body);
 306                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 307                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 308                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 309                 if (!req)
 310                         RETURN(-ENOMEM);
 311
 312                 /* pack the intended request */
 313                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 314                                  0, 0, data);
 315
 316                 /* pack the intent */
 317                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 318                                      sizeof (*lit));
 319                 lit->opc = (__u64)it->it_op;
 320
 321                 /* get ready for the reply */
 322                 reply_buffers = 3;
 323                 req->rq_replen = lustre_msg_size(3, repsize);
 324         } else {
 325                 LBUG();
 326                 RETURN(-EINVAL);
 327         }
 328
 329         mdc_pack_secdesc(req, reqsize[0]);
 330
 331         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 332         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 333                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 334                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 335                               lockh);
 336         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 337
 338         /* Similarly, if we're going to replay this request, we don't want to
 339          * actually get a lock, just perform the intent. */
 340         if (req->rq_transno || req->rq_replay) {
 341                 lockreq = lustre_msg_buf(req->rq_reqmsg,
 342                                          MDS_REQ_INTENT_LOCKREQ_OFF,
 343                                          sizeof (*lockreq));
 344                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 345         }
 346
 347         /* This can go when we're sure that this can never happen */
 348         LASSERT(rc != -ENOENT);
 349         /* We need dlm_rep to be assigned this early, to check lock mode of
 350            returned lock from request to avoid possible race with lock
 351            conversion */
 352         if (rc == ELDLM_LOCK_ABORTED || !rc) {
 353                 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 354                 LASSERT(dlm_rep != NULL);   /* checked by ldlm_cli_enqueue() */
 355         }
 356         if (rc == ELDLM_LOCK_ABORTED) {
 357                 lock_mode = 0;
 358                 memset(lockh, 0, sizeof(*lockh));
 359                 rc = 0;
 360         } else if (rc != 0) {
 361                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 362                 LASSERTF(rc < 0, "rc = %d\n", rc);
 363                 mdc_clear_replay_flag(req, rc);
 364                 ptlrpc_req_finished(req);
 365                 RETURN(rc);
 366         } else { /* rc = 0 */
 367                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 368                 LASSERT(lock);
 369
 370                 /* If the server gave us back a different lock mode, we should
 371                  * fix up our variables. */
 372                 if (dlm_rep->lock_desc.l_req_mode != lock_mode) {
 373                         ldlm_lock_addref(lockh, dlm_rep->lock_desc.l_req_mode);
 374                         ldlm_lock_decref(lockh, lock_mode);
 375                         lock_mode = dlm_rep->lock_desc.l_req_mode;
 376                 }
 377
 378                 ldlm_lock_allow_match(lock);
 379                 LDLM_LOCK_PUT(lock);
 380         }
 381
 382         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 383
 384         LUSTRE_IT(it)->it_disposition = (int) dlm_rep->lock_policy_res1;
 385         LUSTRE_IT(it)->it_status = (int) dlm_rep->lock_policy_res2;
 386         LUSTRE_IT(it)->it_lock_mode = lock_mode;
 387         LUSTRE_IT(it)->it_data = req;
 388
 389         if (LUSTRE_IT(it)->it_status < 0 && req->rq_replay)
 390                 mdc_clear_replay_flag(req, LUSTRE_IT(it)->it_status);
 391
 392         DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
 393                   LUSTRE_IT(it)->it_disposition, LUSTRE_IT(it)->it_status);
 394
 395         /* We know what to expect, so we do any byte flipping required here */
 396         LASSERT(reply_buffers == 5 || reply_buffers == 4 ||
 397                 reply_buffers == 3 || reply_buffers == 1);
 398         if (reply_buffers >= 3) {
 399                 struct mds_body *body;
 400
 401                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 402                                           lustre_swab_mds_body);
 403                 if (body == NULL) {
 404                         CERROR ("Can't swab mds_body\n");
 405                         RETURN (-EPROTO);
 406                 }
 407
 408                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 409                         /* The eadata is opaque; just check that it is there.
 410                          * Eventually, obd_unpackmd() will check the contents */
 411                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 412                                                     NULL);
 413                         if (eadata == NULL) {
 414                                 CERROR ("Missing/short eadata\n");
 415                                 RETURN (-EPROTO);
 416                         }
 417                         if (it->it_op & IT_OPEN) {
 418                                 void *replayea;
 419
 420                                 replayea = lustre_msg_buf(req->rq_reqmsg,
 421                                                           MDS_REQ_INTENT_REC_OFF + 2,
 422                                                           body->eadatasize);
 423                                 LASSERT(replayea);
 424                                 memcpy(replayea, eadata, body->eadatasize);
 425
 426                                 LASSERT(req->rq_reqmsg->bufcount == 6);
 427                                 req->rq_reqmsg->buflens[5] = body->eadatasize;
 428                                 /* If this isn't the last buffer, we might
 429                                  * have to shift other data around. */
 430                         }
 431                 }
 432         }
 433
 434         RETURN(rc);
 435 }
 436 EXPORT_SYMBOL(mdc_enqueue);
 437
 438 /*
 439  * This long block is all about fixing up the lock and request state
 440  * so that it is correct as of the moment _before_ the operation was
 441  * applied; that way, the VFS will think that everything is normal and
 442  * call Lustre's regular VFS methods.
 443  *
 444  * If we're performing a creation, that means that unless the creation
 445  * failed with EEXIST, we should fake up a negative dentry.
 446  *
 447  * For everything else, we want to lookup to succeed.
 448  *
 449  * One additional note: if CREATE or OPEN succeeded, we add an extra
 450  * reference to the request because we need to keep it around until
 451  * ll_create/ll_open gets called.
 452  *
 453  * The server will return to us, in it_disposition, an indication of
 454  * exactly what d.lustre->it_status refers to.
 455  *
 456  * If DISP_OPEN_OPEN is set, then d.lustre->it_status refers to the open() call,
 457  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 458  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 459  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 460  * was successful.
 461  *
 462  * Else, if DISP_LOOKUP_EXECD then d.lustre->it_status is the rc of the
 463  * child lookup.
 464  */
 465 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid,
 466                     const char *name, int len, void *lmm, int lmmsize,
 467                     struct lustre_id *cid, struct lookup_intent *it,
 468                     int lookup_flags, struct ptlrpc_request **reqp,
 469                     ldlm_blocking_callback cb_blocking)
 470 {
 471         struct lustre_handle lockh;
 472         struct ptlrpc_request *request;
 473         struct mds_body *mds_body;
 474         struct lustre_handle old_lock;
 475         struct ldlm_lock *lock;
 476         int rc = 0;
 477         ENTRY;
 478         LASSERT(it);
 479
 480         CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
 481                len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
 482
 483         if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 484                     it->it_op == IT_CHDIR)) {
 485                 /* We could just return 1 immediately, but since we should only
 486                  * be called in revalidate_it if we already have a lock, let's
 487                  * verify that. */
 488                 struct ldlm_res_id res_id = {.name = {id_fid(cid),
 489                                                       id_group(cid)}};
 490                 struct lustre_handle lockh;
 491                 ldlm_policy_data_t policy;
 492                 int mode;
 493
 494                 /* For the GETATTR case, ll_revalidate_it issues two separate
 495                    queries - for LOOKUP and for UPDATE lock because it cannot
 496                    check them together - we might have those two bits to be
 497                    present in two separate granted locks */
 498                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 499                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 500
 501                 mode = LCK_PR;
 502                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 503                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 504                                      LDLM_IBITS, &policy, mode,
 505                                      &lockh);
 506
 507                 if (!rc) {
 508                         mode = LCK_CR;
 509                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 510                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 511                                              LDLM_IBITS, &policy, mode,
 512                                              &lockh);
 513                 }
 514                 if (!rc) {
 515                         mode = LCK_PW;
 516                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 517                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 518                                              LDLM_IBITS, &policy, mode,
 519                                              &lockh);
 520                 }
 521                 if (!rc) {
 522                         mode = LCK_CW;
 523                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 524                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 525                                              LDLM_IBITS, &policy, mode,
 526                                              &lockh);
 527                 }
 528                 if (rc) {
 529                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 530                                sizeof(lockh));
 531                         LUSTRE_IT(it)->it_lock_mode = mode;
 532                 }
 533
 534                 /* Only return failure if it was not GETATTR by cid (from
 535                    inode_revalidate) */
 536                 if (rc || name)
 537                         RETURN(rc);
 538         }
 539
 540         /* lookup_it may be called only after revalidate_it has run, because
 541          * revalidate_it cannot return errors, only zero.  Returning zero causes
 542          * this call to lookup, which *can* return an error.
 543          *
 544          * We only want to execute the request associated with the intent one
 545          * time, however, so don't send the request again.  Instead, skip past
 546          * this and use the request from revalidate.  In this case, revalidate
 547          * never dropped its reference, so the refcounts are all OK */
 548         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 549                 struct mdc_op_data *op_data;
 550
 551                 OBD_ALLOC(op_data, sizeof(*op_data));
 552                 if (op_data == NULL)
 553                         RETURN(-ENOMEM);
 554
 555                 mdc_id2mdc_data(op_data, pid, cid, name, len, 0);
 556
 557                 if (name != NULL)
 558                         op_data->valid |= OBD_MD_FID;
 559
 560                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 561                                  op_data, &lockh, lmm, lmmsize,
 562                                  ldlm_completion_ast, cb_blocking, NULL);
 563                 OBD_FREE(op_data, sizeof(*op_data));
 564                 if (rc < 0)
 565                         RETURN(rc);
 566
 567                 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, sizeof(lockh));
 568         }
 569         request = *reqp = LUSTRE_IT(it)->it_data;
 570         LASSERT(request != NULL);
 571
 572         /* If we're doing an IT_OPEN which did not result in an actual
 573          * successful open, then we need to remove the bit which saves this
 574          * request for unconditional replay.
 575          *
 576          * It's important that we do this first!  Otherwise we might exit the
 577          * function without doing so, and try to replay a failed create (bug
 578          * 3440) */
 579         if (it->it_op & IT_OPEN && request->rq_replay &&
 580             (!it_disposition(it, DISP_OPEN_OPEN) || LUSTRE_IT(it)->it_status != 0))
 581                 mdc_clear_replay_flag(request, LUSTRE_IT(it)->it_status);
 582
 583         if (!it_disposition(it, DISP_IT_EXECD)) {
 584                 /* The server failed before it even started executing the
 585                  * intent, i.e. because it couldn't unpack the request. */
 586                 LASSERT(LUSTRE_IT(it)->it_status != 0);
 587                 RETURN(LUSTRE_IT(it)->it_status);
 588         }
 589         rc = it_open_error(DISP_IT_EXECD, it);
 590         if (rc)
 591                 RETURN(rc);
 592
 593         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 594         LASSERT(mds_body != NULL);      /* mdc_enqueue checked */
 595         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 596
 597         /* If we were revalidating a fid/name pair, mark the intent in case we
 598          * fail and get called again from lookup */
 599         if (cid != NULL) {
 600                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 601                 /* Also: did we find the same inode? */
 602
 603                 /* we have to compare all the fields but type, because MDS can
 604                  * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
 605                 if (!id_equal(cid, &mds_body->id1))
 606                         RETURN(-ESTALE);
 607         }
 608
 609         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 610         if (rc)
 611                 RETURN(rc);
 612
 613         /*
 614          * keep requests around for the multiple phases of the call this shows
 615          * the DISP_XX must guarantee we make it into the call.
 616          */
 617         if (it_disposition(it, DISP_OPEN_CREATE) &&
 618             !it_open_error(DISP_OPEN_CREATE, it))
 619                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 620         if (it_disposition(it, DISP_OPEN_OPEN) &&
 621             !it_open_error(DISP_OPEN_OPEN, it))
 622                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 623
 624         if (it->it_op & IT_CREAT) {
 625                 /* XXX this belongs in ll_create_it */
 626         } else if (it->it_op == IT_OPEN) {
 627                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 628         } else {
 629                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 630         }
 631
 632         /*
 633          * if we already have a matching lock, then cancel the new one. We have
 634          * to set the data here instead of in mdc_enqueue, because we need to
 635          * use the child's inode as the l_ast_data to match, and that's not
 636          * available until intent_finish has performed the iget().)
 637          */
 638         lock = ldlm_handle2lock(&lockh);
 639         if (lock) {
 640                 ldlm_policy_data_t policy = lock->l_policy_data;
 641                 LDLM_DEBUG(lock, "matching against this");
 642                 LDLM_LOCK_PUT(lock);
 643
 644                 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
 645                          id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
 646                          "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
 647                          "response res_id: %lu/%lu.\n",
 648                          (unsigned long)lock->l_resource->lr_name.name[0],
 649                          (unsigned long)lock->l_resource->lr_name.name[1],
 650                          (unsigned long)id_fid(&mds_body->id1),
 651                          (unsigned long)id_group(&mds_body->id1));
 652
 653                 memcpy(&old_lock, &lockh, sizeof(lockh));
 654                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 655                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 656                         ldlm_lock_decref_and_cancel(&lockh,
 657                                                     LUSTRE_IT(it)->it_lock_mode);
 658                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 659                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 660                                sizeof(lockh));
 661                 }
 662         }
 663         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 664                len, name, ldlm_it2str(it->it_op), LUSTRE_IT(it)->it_status,
 665                LUSTRE_IT(it)->it_disposition, rc);
 666
 667         RETURN(rc);
 668 }
 669 EXPORT_SYMBOL(mdc_intent_lock);