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_export *exp, struct lustre_msg *msg,
 179                        int offset)
 180 {
 181         struct client_obd *cli_obd = &exp->exp_obd->u.cli;
 182         struct clonefs_info *cl_info;
 183         ENTRY;
 184
 185         if (cli_obd->cl_clone_info) {
 186                 cl_info = (struct clonefs_info *)lustre_msg_buf(msg, offset,
 187                                                                  sizeof (*cl_info));
 188                 memcpy(cl_info, cli_obd->cl_clone_info, sizeof(*cl_info));
 189         }
 190         RETURN(0);
 191 }
 192 #endif
 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 obd_device *obddev = class_exp2obd(exp);
 210         struct ldlm_res_id res_id =
 211                 { .name = {data->fid1.id, data->fid1.generation} };
 212         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 213         int size[6] = {sizeof(struct ldlm_request), sizeof(struct ldlm_intent)};
 214         int rc, flags = LDLM_FL_HAS_INTENT;
 215         int repsize[4] = {sizeof(struct ldlm_reply),
 216                           sizeof(struct mds_body),
 217                           obddev->u.cli.cl_max_mds_easize,
 218                           obddev->u.cli.cl_max_mds_cookiesize};
 219         struct ldlm_reply *dlm_rep;
 220         struct ldlm_intent *lit;
 221         struct ldlm_request *lockreq;
 222         void *eadata;
 223         unsigned long irqflags;
 224         int   reply_buffers = 0;
 225         ENTRY;
 226
 227 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 228 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 229
 230         if (it->it_op & IT_OPEN) {
 231                 it->it_create_mode |= S_IFREG;
 232                 it->it_create_mode &= ~current->fs->umask;
 233
 234                 size[2] = sizeof(struct mds_rec_create);
 235                 size[3] = data->namelen + 1;
 236                 size[4] = obddev->u.cli.cl_max_mds_easize;
 237                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE,
 238                                       5, size, NULL);
 239                 if (!req)
 240                         RETURN(-ENOMEM);
 241
 242                 spin_lock_irqsave (&req->rq_lock, irqflags);
 243                 req->rq_replay = 1;
 244                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 245
 246                 /* pack the intent */
 247                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 248                 lit->opc = (__u64)it->it_op;
 249
 250                 /* pack the intended request */
 251                 mdc_open_pack(req->rq_reqmsg, 2, data, it->it_create_mode, 0,
 252                               it->it_flags, lmm, lmmsize);
 253                 /* get ready for the reply */
 254                 reply_buffers = 3;
 255                 req->rq_replen = lustre_msg_size(3, repsize);
 256         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 257                 int valid = OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
 258                 size[2] = sizeof(struct mds_body);
 259                 size[3] = data->namelen + 1;
 260
 261                 if (it->it_op & IT_GETATTR)
 262                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 263 #if CONFIG_SNAPFS
 264                 size[4] = sizeof(struct clonefs_info);
 265                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 5,
 266                                       size, NULL);
 267 #else
 268                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 4,
 269                                       size, NULL);
 270
 271 #endif
 272                 if (!req)
 273                         RETURN(-ENOMEM);
 274
 275                 /* pack the intent */
 276                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 277                 lit->opc = (__u64)it->it_op;
 278
 279                 /* pack the intended request */
 280                 mdc_getattr_pack(req->rq_reqmsg, valid, 2, it->it_flags, data);
 281 #if CONFIG_SNAPFS
 282                 mdc_set_clone_info(exp, req->rq_reqmsg, 4);
 283 #endif
 284                 /* get ready for the reply */
 285                 reply_buffers = 3;
 286                 req->rq_replen = lustre_msg_size(3, repsize);
 287         } else if (it->it_op == IT_READDIR) {
 288                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 289                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
 290                                       size, NULL);
 291
 292                 if (!req)
 293                         RETURN(-ENOMEM);
 294                 /* get ready for the reply */
 295                 reply_buffers = 1;
 296                 req->rq_replen = lustre_msg_size(1, repsize);
 297         } else if (it->it_op == IT_UNLINK) {
 298                 size[2] = sizeof(struct mds_body);
 299                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 300                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 3,
 301                                       size, NULL);
 302                 if (!req)
 303                         RETURN(-ENOMEM);
 304
 305                 /* pack the intended request */
 306                 mdc_getattr_pack(req->rq_reqmsg, 0,  2, 0, data);
 307
 308                 /* pack the intent */
 309                 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
 310                 lit->opc = (__u64)it->it_op;
 311
 312                 /* get ready for the reply */
 313                 reply_buffers = 3;
 314                 req->rq_replen = lustre_msg_size(3, repsize);
 315         } else {
 316                 LBUG();
 317                 RETURN(-EINVAL);
 318         }
 319         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 320         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 321                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 322                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 323                               lockh);
 324         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 325
 326         /* Similarly, if we're going to replay this request, we don't want to
 327          * actually get a lock, just perform the intent. */
 328         if (req->rq_transno || req->rq_replay) {
 329                 lockreq = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*lockreq));
 330                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 331         }
 332
 333         /* This can go when we're sure that this can never happen */
 334         LASSERT(rc != -ENOENT);
 335         if (rc == ELDLM_LOCK_ABORTED) {
 336                 lock_mode = 0;
 337                 memset(lockh, 0, sizeof(*lockh));
 338                 rc = 0;
 339         } else if (rc != 0) {
 340                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 341                 LASSERT (rc < 0);
 342                 ptlrpc_req_finished(req);
 343                 RETURN(rc);
 344         } else { /* rc = 0 */
 345                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 346                 LASSERT(lock);
 347
 348                 /* If the server gave us back a different lock mode, we should
 349                  * fix up our variables. */
 350                 if (lock->l_req_mode != lock_mode) {
 351                         ldlm_lock_addref(lockh, lock->l_req_mode);
 352                         ldlm_lock_decref(lockh, lock_mode);
 353                         lock_mode = lock->l_req_mode;
 354                 }
 355
 356                 ldlm_lock_allow_match(lock);
 357                 LDLM_LOCK_PUT(lock);
 358         }
 359
 360         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 361         LASSERT(dlm_rep != NULL);           /* checked by ldlm_cli_enqueue() */
 362         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 363
 364         it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
 365         it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
 366         it->d.lustre.it_lock_mode = lock_mode;
 367         it->d.lustre.it_data = req;
 368
 369         if (it->d.lustre.it_status < 0 && req->rq_replay) {
 370                 LASSERT(req->rq_transno == 0);
 371                 /* Don't hold error requests for replay. */
 372                 spin_lock(&req->rq_lock);
 373                 req->rq_replay = 0;
 374                 spin_unlock(&req->rq_lock);
 375         }
 376
 377         DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
 378                   it->d.lustre.it_disposition, it->d.lustre.it_status);
 379
 380         /* We know what to expect, so we do any byte flipping required here */
 381         LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
 382         if (reply_buffers >= 3) {
 383                 struct mds_body *body;
 384
 385                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 386                                            lustre_swab_mds_body);
 387                 if (body == NULL) {
 388                         CERROR ("Can't swab mds_body\n");
 389                         RETURN (-EPROTO);
 390                 }
 391
 392                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 393                         /* The eadata is opaque; just check that it is there.
 394                          * Eventually, obd_unpackmd() will check the contents */
 395                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 396                                                     NULL);
 397                         if (eadata == NULL) {
 398                                 CERROR ("Missing/short eadata\n");
 399                                 RETURN (-EPROTO);
 400                         }
 401                         if (it->it_op & IT_OPEN) {
 402                                 void *replayea;
 403
 404                                 replayea = lustre_msg_buf(req->rq_reqmsg, 4,
 405                                                           body->eadatasize);
 406                                 LASSERT(replayea);
 407                                 memcpy(replayea, eadata, body->eadatasize);
 408
 409                                 LASSERT(req->rq_reqmsg->bufcount == 5);
 410                                 req->rq_reqmsg->buflens[4] = body->eadatasize;
 411                                 /* If this isn't the last buffer, we might
 412                                  * have to shift other data around. */
 413                         }
 414                 }
 415         }
 416
 417         RETURN(rc);
 418 }
 419 EXPORT_SYMBOL(mdc_enqueue);
 420
 421 /*
 422  * This long block is all about fixing up the lock and request state
 423  * so that it is correct as of the moment _before_ the operation was
 424  * applied; that way, the VFS will think that everything is normal and
 425  * call Lustre's regular VFS methods.
 426  *
 427  * If we're performing a creation, that means that unless the creation
 428  * failed with EEXIST, we should fake up a negative dentry.
 429  *
 430  * For everything else, we want to lookup to succeed.
 431  *
 432  * One additional note: if CREATE or OPEN succeeded, we add an extra
 433  * reference to the request because we need to keep it around until
 434  * ll_create/ll_open gets called.
 435  *
 436  * The server will return to us, in it_disposition, an indication of
 437  * exactly what d.lustre.it_status refers to.
 438  *
 439  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 440  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 441  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 442  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 443  * was successful.
 444  *
 445  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 446  * child lookup.
 447  */
 448 int mdc_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
 449                     struct ll_fid *pfid, const char *name, int len,
 450                     void *lmm, int lmmsize, struct ll_fid *cfid,
 451                     struct lookup_intent *it, int lookup_flags,
 452                     struct ptlrpc_request **reqp,
 453                     ldlm_blocking_callback cb_blocking)
 454 {
 455         struct lustre_handle lockh;
 456         struct ptlrpc_request *request;
 457         int rc = 0;
 458         struct mds_body *mds_body;
 459         struct lustre_handle old_lock;
 460         struct ldlm_lock *lock;
 461         ENTRY;
 462         LASSERT(it);
 463
 464         CDEBUG(D_DLMTRACE, "name: %*s in %ld, intent: %s\n", len, name,
 465                pfid ? (unsigned long) pfid->id : 0 , ldlm_it2str(it->it_op));
 466
 467         if (cfid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 468                      it->it_op == IT_CHDIR)) {
 469                 /* We could just return 1 immediately, but since we should only
 470                  * be called in revalidate_it if we already have a lock, let's
 471                  * verify that. */
 472                 struct ldlm_res_id res_id ={.name = {cfid->id,
 473                                                      cfid->generation}};
 474                 struct lustre_handle lockh;
 475                 ldlm_policy_data_t policy;
 476                 int mode = LCK_PR;
 477
 478                 /* For the GETATTR case, ll_revalidate_it issues two separate
 479                    queries - for LOOKUP and for UPDATE lock because if cannot
 480                    check them together - we might have those two bits to be
 481                    present in two separate granted locks */
 482                 policy.l_inodebits.bits =
 483                                  (it->it_op == IT_GETATTR)?MDS_INODELOCK_UPDATE:
 484                                                            MDS_INODELOCK_LOOKUP;
 485                 mode = LCK_PR;
 486                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 487                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 488                                      LDLM_IBITS, &policy, LCK_PR, &lockh);
 489                 if (!rc) {
 490                         mode = LCK_PW;
 491                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 492                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 493                                              LDLM_IBITS, &policy, LCK_PW,
 494                                              &lockh);
 495                 }
 496                 if (rc) {
 497                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 498                                sizeof(lockh));
 499                         it->d.lustre.it_lock_mode = mode;
 500                 }
 501                 RETURN(rc);
 502         }
 503
 504         /* lookup_it may be called only after revalidate_it has run, because
 505          * revalidate_it cannot return errors, only zero.  Returning zero causes
 506          * this call to lookup, which *can* return an error.
 507          *
 508          * We only want to execute the request associated with the intent one
 509          * time, however, so don't send the request again.  Instead, skip past
 510          * this and use the request from revalidate.  In this case, revalidate
 511          * never dropped its reference, so the refcounts are all OK */
 512         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 513                 struct mdc_op_data op_data;
 514                 mdc_fid2mdc_op_data(&op_data, uctxt, pfid, cfid, name, len, 0);
 515
 516                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 517                                  &op_data, &lockh, lmm, lmmsize,
 518                                  ldlm_completion_ast, cb_blocking, NULL);
 519                 if (rc < 0)
 520                         RETURN(rc);
 521                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 522         }
 523         request = *reqp = it->d.lustre.it_data;
 524         LASSERT(request != NULL);
 525
 526         if (!it_disposition(it, DISP_IT_EXECD)) {
 527                 /* The server failed before it even started executing the
 528                  * intent, i.e. because it couldn't unpack the request. */
 529                 LASSERT(it->d.lustre.it_status != 0);
 530                 RETURN(it->d.lustre.it_status);
 531         }
 532         rc = it_open_error(DISP_IT_EXECD, it);
 533         if (rc)
 534                 RETURN(rc);
 535
 536         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 537         LASSERT(mds_body != NULL);           /* mdc_enqueue checked */
 538         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 539
 540         /* If we were revalidating a fid/name pair, mark the intent in
 541          * case we fail and get called again from lookup */
 542         if (cfid != NULL) {
 543                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 544                 /* Also: did we find the same inode? */
 545                 /* we have to compare all the fields but type, because
 546                  * MDS can return mds/ino/generation triple if inode
 547                  * lives on another MDS -bzzz */
 548                 if (cfid->generation != mds_body->fid1.generation ||
 549                                 cfid->id != mds_body->fid1.id ||
 550                                 cfid->mds != mds_body->fid1.mds)
 551                         RETURN(-ESTALE);
 552         }
 553
 554         /* If we're doing an IT_OPEN which did not result in an actual
 555          * successful open, then we need to remove the bit which saves
 556          * this request for unconditional replay. */
 557         if (it->it_op & IT_OPEN) {
 558                 if (!it_disposition(it, DISP_OPEN_OPEN) ||
 559                     it->d.lustre.it_status != 0) {
 560                         unsigned long irqflags;
 561
 562                         spin_lock_irqsave(&request->rq_lock, irqflags);
 563                         request->rq_replay = 0;
 564                         spin_unlock_irqrestore(&request->rq_lock, irqflags);
 565                 }
 566         }
 567
 568         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 569         if (rc)
 570                 RETURN(rc);
 571
 572         /* keep requests around for the multiple phases of the call
 573          * this shows the DISP_XX must guarantee we make it into the call
 574          */
 575         if (it_disposition(it, DISP_OPEN_CREATE) &&
 576             !it_open_error(DISP_OPEN_CREATE, it))
 577                 ptlrpc_request_addref(request);
 578         if (it_disposition(it, DISP_OPEN_OPEN) &&
 579             !it_open_error(DISP_OPEN_OPEN, it))
 580                 ptlrpc_request_addref(request);
 581
 582         if (it->it_op & IT_CREAT) {
 583                 /* XXX this belongs in ll_create_iit */
 584         } else if (it->it_op == IT_OPEN) {
 585                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 586         } else {
 587                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 588         }
 589
 590         /* If we already have a matching lock, then cancel the new
 591          * one.  We have to set the data here instead of in
 592          * mdc_enqueue, because we need to use the child's inode as
 593          * the l_ast_data to match, and that's not available until
 594          * intent_finish has performed the iget().) */
 595         lock = ldlm_handle2lock(&lockh);
 596         if (lock) {
 597                 ldlm_policy_data_t policy = lock->l_policy_data;
 598                 LDLM_DEBUG(lock, "matching against this");
 599                 LDLM_LOCK_PUT(lock);
 600                 memcpy(&old_lock, &lockh, sizeof(lockh));
 601                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 602                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 603                         ldlm_lock_decref_and_cancel(&lockh,
 604                                                     it->d.lustre.it_lock_mode);
 605                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 606                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 607                                sizeof(lockh));
 608                 }
 609         }
 610         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 611                len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
 612                it->d.lustre.it_disposition, rc);
 613
 614         RETURN(rc);
 615 }
 616 EXPORT_SYMBOL(mdc_intent_lock);