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 "mdc_internal.h"
  41
  42 int it_disposition(struct lookup_intent *it, int flag)
  43 {
  44         return it->d.lustre.it_disposition & flag;
  45 }
  46 EXPORT_SYMBOL(it_disposition);
  47
  48 void it_set_disposition(struct lookup_intent *it, int flag)
  49 {
  50         it->d.lustre.it_disposition |= flag;
  51 }
  52 EXPORT_SYMBOL(it_set_disposition);
  53
  54 static void mdc_id2mdc_op_data(struct mdc_op_data *data,
  55                                struct lustre_id *f1,
  56                                struct lustre_id *f2,
  57                                const char *name,
  58                                int namelen, int mode)
  59 {
  60         LASSERT(data);
  61         LASSERT(f1);
  62
  63         data->id1 = *f1;
  64         if (f2)
  65                 data->id2 = *f2;
  66
  67         data->valid = 0;
  68         data->name = name;
  69         data->namelen = namelen;
  70         data->create_mode = mode;
  71         data->mod_time = LTIME_S(CURRENT_TIME);
  72 }
  73
  74 static int it_to_lock_mode(struct lookup_intent *it)
  75 {
  76         /* CREAT needs to be tested before open (both could be set) */
  77         if (it->it_op & IT_CREAT)
  78                 return LCK_PW;
  79         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
  80                               IT_CHDIR))
  81                 return LCK_PR;
  82
  83         LBUG();
  84         RETURN(-EINVAL);
  85 }
  86
  87 int it_open_error(int phase, struct lookup_intent *it)
  88 {
  89         if (it_disposition(it, DISP_OPEN_OPEN)) {
  90                 if (phase == DISP_OPEN_OPEN)
  91                         return it->d.lustre.it_status;
  92                 else
  93                         return 0;
  94         }
  95
  96         if (it_disposition(it, DISP_OPEN_CREATE)) {
  97                 if (phase == DISP_OPEN_CREATE)
  98                         return it->d.lustre.it_status;
  99                 else
 100                         return 0;
 101         }
 102
 103         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
 104                 if (phase == DISP_LOOKUP_EXECD)
 105                         return it->d.lustre.it_status;
 106                 else
 107                         return 0;
 108         }
 109
 110         if (it_disposition(it, DISP_IT_EXECD)) {
 111                 if (phase == DISP_IT_EXECD)
 112                         return it->d.lustre.it_status;
 113                 else
 114                         return 0;
 115         }
 116         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 117                it->d.lustre.it_status);
 118         LBUG();
 119         return 0;
 120 }
 121 EXPORT_SYMBOL(it_open_error);
 122
 123 /* this must be called on a lockh that is known to have a referenced lock */
 124 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
 125 {
 126         struct ldlm_lock *lock;
 127         struct lustre_handle *lockh = (struct lustre_handle *)l;
 128         ENTRY;
 129
 130         if (!*l) {
 131                 EXIT;
 132                 return 0;
 133         }
 134
 135         lock = ldlm_handle2lock(lockh);
 136
 137         LASSERT(lock != NULL);
 138         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 139 #ifdef __KERNEL__
 140         if (lock->l_ast_data && lock->l_ast_data != data) {
 141                 struct inode *new_inode = data;
 142                 struct inode *old_inode = lock->l_ast_data;
 143                 LASSERTF(old_inode->i_state & I_FREEING,
 144                          "Found existing inode %p/%lu/%u state %lu in lock: "
 145                          "setting data to %p/%lu/%u\n", old_inode,
 146                          old_inode->i_ino, old_inode->i_generation,
 147                          old_inode->i_state, new_inode, new_inode->i_ino,
 148                          new_inode->i_generation);
 149         }
 150 #endif
 151         lock->l_ast_data = data;
 152         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 153         LDLM_LOCK_PUT(lock);
 154
 155         EXIT;
 156         return 0;
 157 }
 158 EXPORT_SYMBOL(mdc_set_lock_data);
 159
 160 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id,
 161                       ldlm_iterator_t it, void *data)
 162 {
 163         struct ldlm_res_id res_id = { .name = {0} };
 164         ENTRY;
 165
 166         res_id.name[0] = id_fid(id);
 167         res_id.name[1] = id_group(id);
 168
 169         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
 170                            &res_id, it, data);
 171
 172         EXIT;
 173         return 0;
 174 }
 175
 176 /* We always reserve enough space in the reply packet for a stripe MD, because
 177  * we don't know in advance the file type. */
 178 int mdc_enqueue(struct obd_export *exp,
 179                 int lock_type,
 180                 struct lookup_intent *it,
 181                 int lock_mode,
 182                 struct mdc_op_data *data,
 183                 struct lustre_handle *lockh,
 184                 void *lmm,
 185                 int lmmsize,
 186                 ldlm_completion_callback cb_completion,
 187                 ldlm_blocking_callback cb_blocking,
 188                 void *cb_data)
 189 {
 190         struct ptlrpc_request *req;
 191         struct ldlm_res_id res_id = {
 192                 .name = {id_fid(&data->id1), id_group(&data->id1)}
 193         };
 194         struct obd_device *obddev = class_exp2obd(exp);
 195         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 196         struct ldlm_intent *lit;
 197         struct ldlm_request *lockreq;
 198         struct ldlm_reply *dlm_rep;
 199         int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
 200                           [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
 201                           [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
 202         int repsize[4] = {sizeof(struct ldlm_reply),
 203                           sizeof(struct mds_body),
 204                           obddev->u.cli.cl_max_mds_easize,
 205                           obddev->u.cli.cl_max_mds_cookiesize};
 206         int req_buffers = 3, reply_buffers = 0;
 207         int rc, flags = LDLM_FL_HAS_INTENT;
 208         void *eadata;
 209         unsigned long irqflags;
 210         ENTRY;
 211
 212 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 213 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 214
 215         reqsize[0] = mdc_get_secdesc_size();
 216
 217         if (it->it_op & IT_OPEN) {
 218                 it->it_create_mode |= S_IFREG;
 219                 it->it_create_mode &= ~current->fs->umask;
 220
 221                 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
 222                 reqsize[req_buffers++] = data->namelen + 1;
 223                 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
 224                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 225                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 226                 if (!req)
 227                         RETURN(-ENOMEM);
 228
 229                 spin_lock_irqsave (&req->rq_lock, irqflags);
 230                 req->rq_replay = 1;
 231                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 232
 233                 /* pack the intent */
 234                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 235                                      sizeof (*lit));
 236                 lit->opc = (__u64)it->it_op;
 237
 238                 /* pack the intended request */
 239                 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
 240                               it->it_create_mode, 0, it->it_flags,
 241                               lmm, lmmsize);
 242                 /* get ready for the reply */
 243                 reply_buffers = 3;
 244                 req->rq_replen = lustre_msg_size(3, repsize);
 245         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 246                 int valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
 247
 248                 reqsize[req_buffers++] = sizeof(struct mds_body);
 249                 reqsize[req_buffers++] = data->namelen + 1;
 250
 251                 if (it->it_op & IT_GETATTR)
 252                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 253
 254                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 255                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 256
 257                 if (!req)
 258                         RETURN(-ENOMEM);
 259
 260                 /* pack the intent */
 261                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 262                                      sizeof (*lit));
 263                 lit->opc = (__u64)it->it_op;
 264
 265                 /* pack the intended request */
 266                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 267                                  valid, it->it_flags, data);
 268
 269                 /* get ready for the reply */
 270                 reply_buffers = 3;
 271                 req->rq_replen = lustre_msg_size(3, repsize);
 272         } else if (it->it_op == IT_READDIR) {
 273                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 274                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 275                                       LDLM_ENQUEUE, 2, reqsize, NULL);
 276
 277                 if (!req)
 278                         RETURN(-ENOMEM);
 279                 /* get ready for the reply */
 280                 reply_buffers = 1;
 281                 req->rq_replen = lustre_msg_size(1, repsize);
 282         } else if (it->it_op == IT_UNLINK) {
 283                 reqsize[req_buffers++] = sizeof(struct mds_body);
 284                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 285                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 286                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 287                 if (!req)
 288                         RETURN(-ENOMEM);
 289
 290                 /* pack the intended request */
 291                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 292                                  0, 0, data);
 293
 294                 /* pack the intent */
 295                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 296                                      sizeof (*lit));
 297                 lit->opc = (__u64)it->it_op;
 298
 299                 /* get ready for the reply */
 300                 reply_buffers = 3;
 301                 req->rq_replen = lustre_msg_size(3, repsize);
 302         } else {
 303                 LBUG();
 304                 RETURN(-EINVAL);
 305         }
 306
 307         mdc_pack_secdesc(req, reqsize[0]);
 308
 309         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 310         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 311                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 312                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 313                               lockh);
 314         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 315
 316         /* Similarly, if we're going to replay this request, we don't want to
 317          * actually get a lock, just perform the intent. */
 318         if (req->rq_transno || req->rq_replay) {
 319                 lockreq = lustre_msg_buf(req->rq_reqmsg,
 320                                          MDS_REQ_INTENT_LOCKREQ_OFF,
 321                                          sizeof (*lockreq));
 322                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 323         }
 324
 325         /* This can go when we're sure that this can never happen */
 326         LASSERT(rc != -ENOENT);
 327         if (rc == ELDLM_LOCK_ABORTED) {
 328                 lock_mode = 0;
 329                 memset(lockh, 0, sizeof(*lockh));
 330                 rc = 0;
 331         } else if (rc != 0) {
 332                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 333                 LASSERT (rc < 0);
 334                 ptlrpc_req_finished(req);
 335                 RETURN(rc);
 336         } else { /* rc = 0 */
 337                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 338                 LASSERT(lock);
 339
 340                 /* If the server gave us back a different lock mode, we should
 341                  * fix up our variables. */
 342                 if (lock->l_req_mode != lock_mode) {
 343                         ldlm_lock_addref(lockh, lock->l_req_mode);
 344                         ldlm_lock_decref(lockh, lock_mode);
 345                         lock_mode = lock->l_req_mode;
 346                 }
 347
 348                 ldlm_lock_allow_match(lock);
 349                 LDLM_LOCK_PUT(lock);
 350         }
 351
 352         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 353         LASSERT(dlm_rep != NULL);           /* checked by ldlm_cli_enqueue() */
 354         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 355
 356         it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
 357         it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
 358         it->d.lustre.it_lock_mode = lock_mode;
 359         it->d.lustre.it_data = req;
 360
 361         if (it->d.lustre.it_status < 0 && req->rq_replay) {
 362                 LASSERT(req->rq_transno == 0);
 363                 /* Don't hold error requests for replay. */
 364                 spin_lock(&req->rq_lock);
 365                 req->rq_replay = 0;
 366                 spin_unlock(&req->rq_lock);
 367         }
 368
 369         DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
 370                   it->d.lustre.it_disposition, it->d.lustre.it_status);
 371
 372         /* We know what to expect, so we do any byte flipping required here */
 373         LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
 374         if (reply_buffers >= 3) {
 375                 struct mds_body *body;
 376
 377                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 378                                            lustre_swab_mds_body);
 379                 if (body == NULL) {
 380                         CERROR ("Can't swab mds_body\n");
 381                         RETURN (-EPROTO);
 382                 }
 383
 384                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 385                         /* The eadata is opaque; just check that it is there.
 386                          * Eventually, obd_unpackmd() will check the contents */
 387                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 388                                                     NULL);
 389                         if (eadata == NULL) {
 390                                 CERROR ("Missing/short eadata\n");
 391                                 RETURN (-EPROTO);
 392                         }
 393                         if (it->it_op & IT_OPEN) {
 394                                 void *replayea;
 395
 396                                 replayea = lustre_msg_buf(req->rq_reqmsg,
 397                                                           MDS_REQ_INTENT_REC_OFF + 2,
 398                                                           body->eadatasize);
 399                                 LASSERT(replayea);
 400                                 memcpy(replayea, eadata, body->eadatasize);
 401
 402                                 LASSERT(req->rq_reqmsg->bufcount == 6);
 403                                 req->rq_reqmsg->buflens[5] = body->eadatasize;
 404                                 /* If this isn't the last buffer, we might
 405                                  * have to shift other data around. */
 406                         }
 407                 }
 408         }
 409
 410         RETURN(rc);
 411 }
 412 EXPORT_SYMBOL(mdc_enqueue);
 413
 414 /*
 415  * This long block is all about fixing up the lock and request state
 416  * so that it is correct as of the moment _before_ the operation was
 417  * applied; that way, the VFS will think that everything is normal and
 418  * call Lustre's regular VFS methods.
 419  *
 420  * If we're performing a creation, that means that unless the creation
 421  * failed with EEXIST, we should fake up a negative dentry.
 422  *
 423  * For everything else, we want to lookup to succeed.
 424  *
 425  * One additional note: if CREATE or OPEN succeeded, we add an extra
 426  * reference to the request because we need to keep it around until
 427  * ll_create/ll_open gets called.
 428  *
 429  * The server will return to us, in it_disposition, an indication of
 430  * exactly what d.lustre.it_status refers to.
 431  *
 432  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 433  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 434  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 435  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 436  * was successful.
 437  *
 438  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 439  * child lookup.
 440  */
 441 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid,
 442                     const char *name, int len, void *lmm, int lmmsize,
 443                     struct lustre_id *cid, struct lookup_intent *it,
 444                     int lookup_flags, struct ptlrpc_request **reqp,
 445                     ldlm_blocking_callback cb_blocking)
 446 {
 447         struct lustre_handle lockh;
 448         struct ptlrpc_request *request;
 449         struct mds_body *mds_body;
 450         struct lustre_handle old_lock;
 451         struct ldlm_lock *lock;
 452         int rc = 0;
 453         ENTRY;
 454         LASSERT(it);
 455
 456         CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
 457                len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
 458
 459         if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 460                     it->it_op == IT_CHDIR)) {
 461                 /* We could just return 1 immediately, but since we should only
 462                  * be called in revalidate_it if we already have a lock, let's
 463                  * verify that. */
 464                 struct ldlm_res_id res_id = {.name = {id_fid(cid),
 465                                                       id_group(cid)}};
 466                 struct lustre_handle lockh;
 467                 ldlm_policy_data_t policy;
 468                 int mode = LCK_PR;
 469
 470                 /* For the GETATTR case, ll_revalidate_it issues two separate
 471                    queries - for LOOKUP and for UPDATE lock because it cannot
 472                    check them together - we might have those two bits to be
 473                    present in two separate granted locks */
 474                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 475                         MDS_INODELOCK_UPDATE: MDS_INODELOCK_LOOKUP;
 476
 477                 mode = LCK_PR;
 478                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 479                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 480                                      LDLM_IBITS, &policy, LCK_PR, &lockh);
 481                 if (!rc) {
 482                         mode = LCK_PW;
 483                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 484                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 485                                              LDLM_IBITS, &policy, LCK_PW,
 486                                              &lockh);
 487                 }
 488                 if (rc) {
 489                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 490                                sizeof(lockh));
 491                         it->d.lustre.it_lock_mode = mode;
 492                 }
 493
 494                 /* Only return failure if it was not GETATTR by cfid
 495                    (from inode_revalidate) */
 496                 if (rc || name)
 497                         RETURN(rc);
 498         }
 499
 500         /* lookup_it may be called only after revalidate_it has run, because
 501          * revalidate_it cannot return errors, only zero.  Returning zero causes
 502          * this call to lookup, which *can* return an error.
 503          *
 504          * We only want to execute the request associated with the intent one
 505          * time, however, so don't send the request again.  Instead, skip past
 506          * this and use the request from revalidate.  In this case, revalidate
 507          * never dropped its reference, so the refcounts are all OK */
 508         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 509                 struct mdc_op_data op_data;
 510
 511                 mdc_id2mdc_op_data(&op_data, pid, cid, name, len, 0);
 512
 513                 /*
 514                  * if we getting inode by name (ll_lookup_it() case), we always
 515                  * should ask for fid, as later we will not be able to take
 516                  * locks, revalidate dentry, etc. with invalid fid in inode.
 517                  */
 518                 if (cid == NULL && name != NULL)
 519                         op_data.valid |= OBD_MD_FID;
 520
 521                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 522                                  &op_data, &lockh, lmm, lmmsize,
 523                                  ldlm_completion_ast, cb_blocking, NULL);
 524                 if (rc < 0)
 525                         RETURN(rc);
 526
 527                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 528         }
 529         request = *reqp = it->d.lustre.it_data;
 530         LASSERT(request != NULL);
 531
 532         /* If we're doing an IT_OPEN which did not result in an actual
 533          * successful open, then we need to remove the bit which saves
 534          * this request for unconditional replay.
 535          *
 536          * It's important that we do this first!  Otherwise we might exit the
 537          * function without doing so, and try to replay a failed create
 538          * (bug 3440) */
 539         if (it->it_op & IT_OPEN) {
 540                 if (!it_disposition(it, DISP_OPEN_OPEN) ||
 541                     it->d.lustre.it_status != 0) {
 542                         unsigned long irqflags;
 543
 544                         spin_lock_irqsave(&request->rq_lock, irqflags);
 545                         request->rq_replay = 0;
 546                         spin_unlock_irqrestore(&request->rq_lock, irqflags);
 547                 }
 548         }
 549         if (!it_disposition(it, DISP_IT_EXECD)) {
 550                 /* The server failed before it even started executing the
 551                  * intent, i.e. because it couldn't unpack the request. */
 552                 LASSERT(it->d.lustre.it_status != 0);
 553                 RETURN(it->d.lustre.it_status);
 554         }
 555         rc = it_open_error(DISP_IT_EXECD, it);
 556         if (rc)
 557                 RETURN(rc);
 558
 559         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 560         LASSERT(mds_body != NULL);      /* mdc_enqueue checked */
 561         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 562
 563         /* If we were revalidating a fid/name pair, mark the intent in case we
 564          * fail and get called again from lookup */
 565         if (cid != NULL) {
 566                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 567                 /* Also: did we find the same inode? */
 568
 569                 /* we have to compare all the fields but type, because MDS can
 570                  * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
 571                 if (id_gen(cid) != id_gen(&mds_body->id1) ||
 572                     id_ino(cid) != id_ino(&mds_body->id1) ||
 573                     id_fid(cid) != id_fid(&mds_body->id1) ||
 574                     id_group(cid) != id_group(&mds_body->id1))
 575                         RETURN(-ESTALE);
 576         }
 577
 578         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 579         if (rc)
 580                 RETURN(rc);
 581
 582         /*
 583          * keep requests around for the multiple phases of the call this shows
 584          * the DISP_XX must guarantee we make it into the call.
 585          */
 586         if (it_disposition(it, DISP_OPEN_CREATE) &&
 587             !it_open_error(DISP_OPEN_CREATE, it))
 588                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 589         if (it_disposition(it, DISP_OPEN_OPEN) &&
 590             !it_open_error(DISP_OPEN_OPEN, it))
 591                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 592
 593         if (it->it_op & IT_CREAT) {
 594                 /* XXX this belongs in ll_create_it */
 595         } else if (it->it_op == IT_OPEN) {
 596                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 597         } else {
 598                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 599         }
 600
 601         /*
 602          * if we already have a matching lock, then cancel the new one. We have
 603          * to set the data here instead of in mdc_enqueue, because we need to
 604          * use the child's inode as the l_ast_data to match, and that's not
 605          * available until intent_finish has performed the iget().)
 606          */
 607         lock = ldlm_handle2lock(&lockh);
 608         if (lock) {
 609                 ldlm_policy_data_t policy = lock->l_policy_data;
 610                 LDLM_DEBUG(lock, "matching against this");
 611                 LDLM_LOCK_PUT(lock);
 612
 613                 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
 614                          id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
 615                          "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
 616                          "response res_id: %lu/%lu.\n",
 617                          (unsigned long)lock->l_resource->lr_name.name[0],
 618                          (unsigned long)lock->l_resource->lr_name.name[1],
 619                          (unsigned long)id_fid(&mds_body->id1),
 620                          (unsigned long)id_group(&mds_body->id1));
 621
 622                 memcpy(&old_lock, &lockh, sizeof(lockh));
 623                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 624                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 625                         ldlm_lock_decref_and_cancel(&lockh,
 626                                                     it->d.lustre.it_lock_mode);
 627                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 628                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 629                                sizeof(lockh));
 630                 }
 631         }
 632         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 633                len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
 634                it->d.lustre.it_disposition, rc);
 635
 636         RETURN(rc);
 637 }
 638 EXPORT_SYMBOL(mdc_intent_lock);