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/lustre_snap.h>
  40 #include <linux/lprocfs_status.h>
  41 #include "mdc_internal.h"
  42
  43 int it_disposition(struct lookup_intent *it, int flag)
  44 {
  45         return it->d.lustre.it_disposition & flag;
  46 }
  47 EXPORT_SYMBOL(it_disposition);
  48
  49 void it_set_disposition(struct lookup_intent *it, int flag)
  50 {
  51         it->d.lustre.it_disposition |= flag;
  52 }
  53 EXPORT_SYMBOL(it_set_disposition);
  54
  55 static void mdc_fid2mdc_op_data(struct mdc_op_data *data, struct ll_uctxt *ctxt,
  56                                 struct ll_fid *f1, struct ll_fid *f2,
  57                                 const char *name, int namelen, int mode)
  58 {
  59         LASSERT(data);
  60         LASSERT(ctxt);
  61         LASSERT(f1);
  62
  63         data->ctxt = *ctxt;
  64         data->fid1 = *f1;
  65         if (f2)
  66                 data->fid2 = *f2;
  67         else
  68                 memset(&data->fid2, 0, sizeof(data->fid2));
  69         data->name = name;
  70         data->namelen = namelen;
  71         data->create_mode = mode;
  72         data->mod_time = LTIME_S(CURRENT_TIME);
  73 }
  74
  75 static int it_to_lock_mode(struct lookup_intent *it)
  76 {
  77         /* CREAT needs to be tested before open (both could be set) */
  78         if (it->it_op & IT_CREAT)
  79                 return LCK_PW;
  80         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
  81                               IT_CHDIR))
  82                 return LCK_PR;
  83
  84         LBUG();
  85         RETURN(-EINVAL);
  86 }
  87
  88 int it_open_error(int phase, struct lookup_intent *it)
  89 {
  90         if (it_disposition(it, DISP_OPEN_OPEN)) {
  91                 if (phase == DISP_OPEN_OPEN)
  92                         return it->d.lustre.it_status;
  93                 else
  94                         return 0;
  95         }
  96
  97         if (it_disposition(it, DISP_OPEN_CREATE)) {
  98                 if (phase == DISP_OPEN_CREATE)
  99                         return it->d.lustre.it_status;
 100                 else
 101                         return 0;
 102         }
 103
 104         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
 105                 if (phase == DISP_LOOKUP_EXECD)
 106                         return it->d.lustre.it_status;
 107                 else
 108                         return 0;
 109         }
 110
 111         if (it_disposition(it, DISP_IT_EXECD)) {
 112                 if (phase == DISP_IT_EXECD)
 113                         return it->d.lustre.it_status;
 114                 else
 115                         return 0;
 116         }
 117         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 118                it->d.lustre.it_status);
 119         LBUG();
 120         return 0;
 121 }
 122 EXPORT_SYMBOL(it_open_error);
 123
 124 /* this must be called on a lockh that is known to have a referenced lock */
 125 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
 126 {
 127         struct ldlm_lock *lock;
 128         struct lustre_handle *lockh = (struct lustre_handle *)l;
 129         ENTRY;
 130
 131         if (!*l) {
 132                 EXIT;
 133                 return 0;
 134         }
 135
 136         lock = ldlm_handle2lock(lockh);
 137
 138         LASSERT(lock != NULL);
 139         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 140 #ifdef __KERNEL__
 141         if (lock->l_ast_data && lock->l_ast_data != data) {
 142                 struct inode *new_inode = data;
 143                 struct inode *old_inode = lock->l_ast_data;
 144                 LASSERTF(old_inode->i_state & I_FREEING,
 145                          "Found existing inode %p/%lu/%u state %lu in lock: "
 146                          "setting data to %p/%lu/%u\n", old_inode,
 147                          old_inode->i_ino, old_inode->i_generation,
 148                          old_inode->i_state,
 149                          new_inode, new_inode->i_ino, new_inode->i_generation);
 150         }
 151 #endif
 152         lock->l_ast_data = data;
 153         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 154         LDLM_LOCK_PUT(lock);
 155
 156         EXIT;
 157         return 0;
 158 }
 159 EXPORT_SYMBOL(mdc_set_lock_data);
 160
 161 int mdc_change_cbdata(struct obd_export *exp, struct ll_fid *fid,
 162                       ldlm_iterator_t it, void *data)
 163 {
 164         struct ldlm_res_id res_id = { .name = {0} };
 165         ENTRY;
 166
 167         res_id.name[0] = fid->id;
 168         res_id.name[1] = fid->generation;
 169
 170         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace, &res_id, it,
 171                            data);
 172
 173         EXIT;
 174         return 0;
 175 }
 176
 177 #if CONFIG_SNAPFS
 178 int mdc_set_clone_info(struct obd_import *climp, struct lustre_msg *msg,
 179                        int offset)
 180 {
 181         struct client_obd *cli_obd = &climp->imp_obd->u.cli;
 182         struct clonefs_info *cl_info;
 183         ENTRY;
 184
 185         cl_info = (struct clonefs_info *)lustre_msg_buf(msg, offset,
 186                                                          sizeof (*cl_info));
 187         memcpy(cl_info, cli_obd->cl_clone_info, sizeof(*cl_info));
 188         RETURN(0);
 189 }
 190 #endif
 191
 192 /* We always reserve enough space in the reply packet for a stripe MD, because
 193  * we don't know in advance the file type. */
 194 int mdc_enqueue(struct obd_export *exp,
 195                 int lock_type,
 196                 struct lookup_intent *it,
 197                 int lock_mode,
 198                 struct mdc_op_data *data,
 199                 struct lustre_handle *lockh,
 200                 void *lmm,
 201                 int lmmsize,
 202                 ldlm_completion_callback cb_completion,
 203                 ldlm_blocking_callback cb_blocking,
 204                 void *cb_data)
 205 {
 206         struct ptlrpc_request *req;
 207         struct obd_device *obddev = class_exp2obd(exp);
 208         struct ldlm_res_id res_id =
 209                 { .name = {data->fid1.id, data->fid1.generation} };
 210         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 211         int size[6] = {sizeof(struct ldlm_request), sizeof(struct ldlm_intent)};
 212         int rc, flags = LDLM_FL_HAS_INTENT;
 213         int repsize[4] = {sizeof(struct ldlm_reply),
 214                           sizeof(struct mds_body),
 215                           obddev->u.cli.cl_max_mds_easize,
 216                           obddev->u.cli.cl_max_mds_cookiesize};
 217         struct ldlm_reply *dlm_rep;
 218         struct ldlm_intent *lit;
 219         struct ldlm_request *lockreq;
 220         void *eadata;
 221         unsigned long irqflags;
 222         int   reply_buffers = 0;
 223         ENTRY;
 224
 225 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 226 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 227
 228         if (it->it_op & IT_OPEN) {
 229                 it->it_create_mode |= S_IFREG;
 230                 it->it_create_mode &= ~current->fs->umask;
 231
 232                 size[2] = sizeof(struct mds_rec_create);
 233                 size[3] = data->namelen + 1;
 234                 size[4] = obddev->u.cli.cl_max_mds_easize;
 235                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE,
 236                                       5, size, NULL);
 237                 if (!req)
 238                         RETURN(-ENOMEM);
 239
 240                 spin_lock_irqsave (&req->rq_lock, irqflags);
 241                 req->rq_replay = 1;
 242                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 243
 244                 /* pack the intent */
 245                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 246                 lit->opc = (__u64)it->it_op;
 247
 248                 /* pack the intended request */
 249                 mdc_open_pack(req->rq_reqmsg, 2, data, it->it_create_mode, 0,
 250                               it->it_flags, lmm, lmmsize);
 251                 /* get ready for the reply */
 252                 reply_buffers = 3;
 253                 req->rq_replen = lustre_msg_size(3, repsize);
 254         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 255                 int valid = OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
 256                 size[2] = sizeof(struct mds_body);
 257                 size[3] = data->namelen + 1;
 258
 259                 if (it->it_op & IT_GETATTR)
 260                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 261 #if CONFIG_SNAPFS
 262                 size[4] = sizeof(struct clonefs_info);
 263                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 5,
 264                                       size, NULL);
 265 #else
 266                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 4,
 267                                       size, NULL);
 268
 269 #endif
 270                 if (!req)
 271                         RETURN(-ENOMEM);
 272
 273                 /* pack the intent */
 274                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 275                 lit->opc = (__u64)it->it_op;
 276
 277                 /* pack the intended request */
 278                 mdc_getattr_pack(req->rq_reqmsg, valid, 2, it->it_flags, data);
 279 #if CONFIG_SNAPFS
 280                 mdc_set_clone_info(class_exp2cliimp(exp), req->rq_reqmsg, 4);
 281 #endif
 282                 /* get ready for the reply */
 283                 reply_buffers = 3;
 284                 req->rq_replen = lustre_msg_size(3, repsize);
 285         } else if (it->it_op == IT_READDIR) {
 286                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 287 #if CONFIG_SNAPFS
 288                 size[1] = sizeof(struct clonefs_info);
 289                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 2,
 290                                       size, NULL);
 291 #else
 292                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
 293                                       size, NULL);
 294
 295 #endif
 296                 if (!req)
 297                         RETURN(-ENOMEM);
 298 #if CONFIG_SNAPFS
 299                 mdc_set_clone_info(class_exp2cliimp(exp), req->rq_reqmsg, 1);
 300 #endif
 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                 size[2] = sizeof(struct mds_body);
 306                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 307                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 3,
 308                                       size, NULL);
 309                 if (!req)
 310                         RETURN(-ENOMEM);
 311
 312                 /* pack the intended request */
 313                 mdc_getattr_pack(req->rq_reqmsg, 0,  2, 0, data);
 314
 315                 /* pack the intent */
 316                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 317                 lit->opc = (__u64)it->it_op;
 318
 319                 /* get ready for the reply */
 320                 reply_buffers = 3;
 321                 req->rq_replen = lustre_msg_size(3, repsize);
 322         } else {
 323                 LBUG();
 324                 RETURN(-EINVAL);
 325         }
 326         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 327         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 328                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 329                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 330                               lockh);
 331         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 332
 333         /* Similarly, if we're going to replay this request, we don't want to
 334          * actually get a lock, just perform the intent. */
 335         if (req->rq_transno || req->rq_replay) {
 336                 lockreq = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*lockreq));
 337                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 338         }
 339
 340         /* This can go when we're sure that this can never happen */
 341         LASSERT(rc != -ENOENT);
 342         if (rc == ELDLM_LOCK_ABORTED) {
 343                 lock_mode = 0;
 344                 memset(lockh, 0, sizeof(*lockh));
 345                 rc = 0;
 346         } else if (rc != 0) {
 347                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 348                 LASSERT (rc < 0);
 349                 ptlrpc_req_finished(req);
 350                 RETURN(rc);
 351         } else { /* rc = 0 */
 352                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 353                 LASSERT(lock);
 354
 355                 /* If the server gave us back a different lock mode, we should
 356                  * fix up our variables. */
 357                 if (lock->l_req_mode != lock_mode) {
 358                         ldlm_lock_addref(lockh, lock->l_req_mode);
 359                         ldlm_lock_decref(lockh, lock_mode);
 360                         lock_mode = lock->l_req_mode;
 361                 }
 362
 363                 ldlm_lock_allow_match(lock);
 364                 LDLM_LOCK_PUT(lock);
 365         }
 366
 367         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 368         LASSERT(dlm_rep != NULL);           /* checked by ldlm_cli_enqueue() */
 369         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 370
 371         it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
 372         it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
 373         it->d.lustre.it_lock_mode = lock_mode;
 374         it->d.lustre.it_data = req;
 375
 376         if (it->d.lustre.it_status < 0 && req->rq_replay) {
 377                 LASSERT(req->rq_transno == 0);
 378                 /* Don't hold error requests for replay. */
 379                 spin_lock(&req->rq_lock);
 380                 req->rq_replay = 0;
 381                 spin_unlock(&req->rq_lock);
 382         }
 383
 384         DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
 385                   it->d.lustre.it_disposition, it->d.lustre.it_status);
 386
 387         /* We know what to expect, so we do any byte flipping required here */
 388         LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
 389         if (reply_buffers >= 3) {
 390                 struct mds_body *body;
 391
 392                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 393                                            lustre_swab_mds_body);
 394                 if (body == NULL) {
 395                         CERROR ("Can't swab mds_body\n");
 396                         RETURN (-EPROTO);
 397                 }
 398
 399                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 400                         void *replayea;
 401                         /* The eadata is opaque; just check that it is
 402                          * there.  Eventually, obd_unpackmd() will check
 403                          * the contents */
 404                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 405                                                     NULL);
 406                         if (eadata == NULL) {
 407                                 CERROR ("Missing/short eadata\n");
 408                                 RETURN (-EPROTO);
 409                         }
 410                         if (it->it_op & IT_OPEN) {
 411                                 replayea = lustre_msg_buf(req->rq_reqmsg, 4,
 412                                                           obddev->u.cli.cl_max_mds_easize);
 413                                 LASSERT(replayea);
 414                                 memcpy(replayea, eadata, body->eadatasize);
 415                         }
 416                 }
 417         }
 418
 419         RETURN(rc);
 420 }
 421 EXPORT_SYMBOL(mdc_enqueue);
 422
 423 /*
 424  * This long block is all about fixing up the lock and request state
 425  * so that it is correct as of the moment _before_ the operation was
 426  * applied; that way, the VFS will think that everything is normal and
 427  * call Lustre's regular VFS methods.
 428  *
 429  * If we're performing a creation, that means that unless the creation
 430  * failed with EEXIST, we should fake up a negative dentry.
 431  *
 432  * For everything else, we want to lookup to succeed.
 433  *
 434  * One additional note: if CREATE or OPEN succeeded, we add an extra
 435  * reference to the request because we need to keep it around until
 436  * ll_create/ll_open gets called.
 437  *
 438  * The server will return to us, in it_disposition, an indication of
 439  * exactly what d.lustre.it_status refers to.
 440  *
 441  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 442  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 443  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 444  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 445  * was successful.
 446  *
 447  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 448  * child lookup.
 449  */
 450 int mdc_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
 451                     struct ll_fid *pfid, const char *name, int len,
 452                     void *lmm, int lmmsize, struct ll_fid *cfid,
 453                     struct lookup_intent *it, int lookup_flags,
 454                     struct ptlrpc_request **reqp,
 455                     ldlm_blocking_callback cb_blocking)
 456 {
 457         struct lustre_handle lockh;
 458         struct ptlrpc_request *request;
 459         int rc = 0;
 460         struct mds_body *mds_body;
 461         struct lustre_handle old_lock;
 462         struct ldlm_lock *lock;
 463         ENTRY;
 464         LASSERT(it);
 465
 466         CDEBUG(D_DLMTRACE, "name: %*s in %ld, intent: %s\n", len, name,
 467                pfid ? (unsigned long) pfid->id : 0 , ldlm_it2str(it->it_op));
 468
 469         if (cfid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 470                      it->it_op == IT_CHDIR)) {
 471                 /* We could just return 1 immediately, but since we should only
 472                  * be called in revalidate_it if we already have a lock, let's
 473                  * verify that. */
 474                 struct ldlm_res_id res_id ={.name = {cfid->id,
 475                                                      cfid->generation}};
 476                 struct lustre_handle lockh;
 477                 ldlm_policy_data_t policy;
 478                 int mode = LCK_PR;
 479
 480                 /* For the GETATTR case, ll_revalidate_it issues two separate
 481                    queries - for LOOKUP and for UPDATE lock because if cannot
 482                    check them together - we might have those two bits to be
 483                    present in two separate granted locks */
 484                 policy.l_inodebits.bits =
 485                                  (it->it_op == IT_GETATTR)?MDS_INODELOCK_UPDATE:
 486                                                            MDS_INODELOCK_LOOKUP;
 487                 mode = LCK_PR;
 488                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 489                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 490                                      LDLM_IBITS, &policy, LCK_PR, &lockh);
 491                 if (!rc) {
 492                         mode = LCK_PW;
 493                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 494                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 495                                              LDLM_IBITS, &policy, LCK_PW,
 496                                              &lockh);
 497                 }
 498                 if (rc) {
 499                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 500                                sizeof(lockh));
 501                         it->d.lustre.it_lock_mode = mode;
 502                 }
 503                 RETURN(rc);
 504         }
 505
 506         /* lookup_it may be called only after revalidate_it has run, because
 507          * revalidate_it cannot return errors, only zero.  Returning zero causes
 508          * this call to lookup, which *can* return an error.
 509          *
 510          * We only want to execute the request associated with the intent one
 511          * time, however, so don't send the request again.  Instead, skip past
 512          * this and use the request from revalidate.  In this case, revalidate
 513          * never dropped its reference, so the refcounts are all OK */
 514         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 515                 struct mdc_op_data op_data;
 516                 mdc_fid2mdc_op_data(&op_data, uctxt, pfid, cfid, name, len, 0);
 517
 518                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 519                                  &op_data, &lockh, lmm, lmmsize,
 520                                  ldlm_completion_ast, cb_blocking, NULL);
 521                 if (rc < 0)
 522                         RETURN(rc);
 523                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 524         }
 525         request = *reqp = it->d.lustre.it_data;
 526         LASSERT(request != NULL);
 527
 528         if (!it_disposition(it, DISP_IT_EXECD)) {
 529                 /* The server failed before it even started executing the
 530                  * intent, i.e. because it couldn't unpack the request. */
 531                 LASSERT(it->d.lustre.it_status != 0);
 532                 RETURN(it->d.lustre.it_status);
 533         }
 534         rc = it_open_error(DISP_IT_EXECD, it);
 535         if (rc)
 536                 RETURN(rc);
 537
 538         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 539         LASSERT(mds_body != NULL);           /* mdc_enqueue checked */
 540         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 541
 542         /* If we were revalidating a fid/name pair, mark the intent in
 543          * case we fail and get called again from lookup */
 544         if (cfid != NULL) {
 545                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 546                 /* Also: did we find the same inode? */
 547                 /* we have to compare all the fields but type, because
 548                  * MDS can return mds/ino/generation triple if inode
 549                  * lives on another MDS -bzzz */
 550                 if (cfid->generation != mds_body->fid1.generation ||
 551                                 cfid->id != mds_body->fid1.id ||
 552                                 cfid->mds != mds_body->fid1.mds)
 553                         RETURN(-ESTALE);
 554         }
 555
 556         /* If we're doing an IT_OPEN which did not result in an actual
 557          * successful open, then we need to remove the bit which saves
 558          * this request for unconditional replay. */
 559         if (it->it_op & IT_OPEN) {
 560                 if (!it_disposition(it, DISP_OPEN_OPEN) ||
 561                     it->d.lustre.it_status != 0) {
 562                         unsigned long irqflags;
 563
 564                         spin_lock_irqsave(&request->rq_lock, irqflags);
 565                         request->rq_replay = 0;
 566                         spin_unlock_irqrestore(&request->rq_lock, irqflags);
 567                 }
 568         }
 569
 570         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 571         if (rc)
 572                 RETURN(rc);
 573
 574         /* keep requests around for the multiple phases of the call
 575          * this shows the DISP_XX must guarantee we make it into the call
 576          */
 577         if (it_disposition(it, DISP_OPEN_CREATE) &&
 578             !it_open_error(DISP_OPEN_CREATE, it))
 579                 ptlrpc_request_addref(request);
 580         if (it_disposition(it, DISP_OPEN_OPEN) &&
 581             !it_open_error(DISP_OPEN_OPEN, it))
 582                 ptlrpc_request_addref(request);
 583
 584         if (it->it_op & IT_CREAT) {
 585                 /* XXX this belongs in ll_create_iit */
 586         } else if (it->it_op == IT_OPEN) {
 587                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 588         } else {
 589                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 590         }
 591
 592         /* If we already have a matching lock, then cancel the new
 593          * one.  We have to set the data here instead of in
 594          * mdc_enqueue, because we need to use the child's inode as
 595          * the l_ast_data to match, and that's not available until
 596          * intent_finish has performed the iget().) */
 597         lock = ldlm_handle2lock(&lockh);
 598         if (lock) {
 599                 ldlm_policy_data_t policy = lock->l_policy_data;
 600                 LDLM_DEBUG(lock, "matching against this");
 601                 LDLM_LOCK_PUT(lock);
 602                 memcpy(&old_lock, &lockh, sizeof(lockh));
 603                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 604                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 605                         ldlm_lock_decref_and_cancel(&lockh,
 606                                                     it->d.lustre.it_lock_mode);
 607                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 608                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 609                                sizeof(lockh));
 610                 }
 611         }
 612         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 613                len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
 614                it->d.lustre.it_disposition, rc);
 615
 616         RETURN(rc);
 617 }
 618 EXPORT_SYMBOL(mdc_intent_lock);