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_sec.h>
  40 #include <linux/lprocfs_status.h>
  41 #include <linux/lustre_acl.h>
  42 #include <linux/lustre_gs.h>
  43 #include <linux/lustre_lite.h>
  44 #include "mdc_internal.h"
  45
  46 int it_disposition(struct lookup_intent *it, int flag)
  47 {
  48         return LUSTRE_IT(it)->it_disposition & flag;
  49 }
  50 EXPORT_SYMBOL(it_disposition);
  51
  52 void it_set_disposition(struct lookup_intent *it, int flag)
  53 {
  54         LUSTRE_IT(it)->it_disposition |= flag;
  55 }
  56 EXPORT_SYMBOL(it_set_disposition);
  57
  58 static void mdc_id2mdc_data(struct mdc_op_data *data,
  59                             struct lustre_id *f1,
  60                             struct lustre_id *f2,
  61                             const char *name,
  62                             int namelen, int mode)
  63 {
  64         LASSERT(data);
  65         LASSERT(f1);
  66
  67         data->id1 = *f1;
  68         if (f2)
  69                 data->id2 = *f2;
  70
  71         data->valid = 0;
  72         data->name = name;
  73         data->namelen = namelen;
  74         data->create_mode = mode;
  75         data->mod_time = LTIME_S(CURRENT_TIME);
  76 }
  77
  78 static int it_to_lock_mode(struct lookup_intent *it)
  79 {
  80         /* CREAT needs to be tested before open (both could be set) */
  81         if (it->it_op & IT_CREAT)
  82                 return LCK_PW;
  83         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
  84                               IT_CHDIR))
  85                 return LCK_PR;
  86
  87         LBUG();
  88         RETURN(-EINVAL);
  89 }
  90
  91 int it_open_error(int phase, struct lookup_intent *it)
  92 {
  93         if (it_disposition(it, DISP_OPEN_OPEN)) {
  94                 if (phase == DISP_OPEN_OPEN)
  95                         return LUSTRE_IT(it)->it_status;
  96                 else
  97                         return 0;
  98         }
  99
 100         if (it_disposition(it, DISP_OPEN_CREATE)) {
 101                 if (phase == DISP_OPEN_CREATE)
 102                         return LUSTRE_IT(it)->it_status;
 103                 else
 104                         return 0;
 105         }
 106
 107         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
 108                 if (phase == DISP_LOOKUP_EXECD)
 109                         return LUSTRE_IT(it)->it_status;
 110                 else
 111                         return 0;
 112         }
 113
 114         if (it_disposition(it, DISP_IT_EXECD)) {
 115                 if (phase == DISP_IT_EXECD)
 116                         return LUSTRE_IT(it)->it_status;
 117                 else
 118                         return 0;
 119         }
 120         CERROR("it disp: %X, status: %d\n", LUSTRE_IT(it)->it_disposition,
 121                LUSTRE_IT(it)->it_status);
 122         LBUG();
 123         return 0;
 124 }
 125 EXPORT_SYMBOL(it_open_error);
 126
 127 /* this must be called on a lockh that is known to have a referenced lock */
 128 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
 129 {
 130         struct ldlm_lock *lock;
 131         struct lustre_handle *lockh = (struct lustre_handle *)l;
 132         ENTRY;
 133
 134         if (!*l) {
 135                 EXIT;
 136                 return 0;
 137         }
 138
 139         lock = ldlm_handle2lock(lockh);
 140
 141         LASSERT(lock != NULL);
 142         lock_res_and_lock(lock);
 143 #ifdef __KERNEL__
 144         if (lock->l_ast_data && lock->l_ast_data != data) {
 145                 struct ll_inode_info *lli = ll_i2info(data);
 146                 struct inode *new_inode = data;
 147                 LASSERT(lli->lli_inode_magic == LLI_INODE_MAGIC);
 148                 struct inode *old_inode = lock->l_ast_data;
 149                 if (!(old_inode->i_state & I_FREEING)) {
 150                         CERROR("Found existing inode %p/%lu/%u state %lu "
 151                                "in lock: setting data to %p/%lu/%u\n",
 152                                old_inode, old_inode->i_ino,
 153                                old_inode->i_generation, old_inode->i_state,
 154                                new_inode, new_inode->i_ino,
 155                                new_inode->i_generation);
 156                         unlock_res_and_lock(lock);
 157                         LBUG();
 158                 }
 159         }
 160 #endif
 161         lock->l_ast_data = data;
 162         unlock_res_and_lock(lock);
 163         LDLM_LOCK_PUT(lock);
 164
 165         EXIT;
 166         return 0;
 167 }
 168 EXPORT_SYMBOL(mdc_set_lock_data);
 169
 170 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id,
 171                       ldlm_iterator_t it, void *data)
 172 {
 173         struct ldlm_res_id res_id = { .name = {0} };
 174         ENTRY;
 175
 176         res_id.name[0] = id_fid(id);
 177         res_id.name[1] = id_group(id);
 178
 179         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
 180                            &res_id, it, data);
 181
 182         EXIT;
 183         return 0;
 184 }
 185
 186 static inline void
 187 mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 188 {
 189         /* Don't hold error requests for replay. */
 190         if (req->rq_replay) {
 191                 unsigned long irqflags;
 192                 spin_lock_irqsave(&req->rq_lock, irqflags);
 193                 req->rq_replay = 0;
 194                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
 195         }
 196         if (rc && req->rq_transno != 0) {
 197                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 198                 LBUG();
 199         }
 200 }
 201
 202 /* We always reserve enough space in the reply packet for a stripe MD, because
 203  * we don't know in advance the file type. */
 204 int mdc_enqueue(struct obd_export *exp,
 205                 int lock_type,
 206                 struct lookup_intent *it,
 207                 int lock_mode,
 208                 struct mdc_op_data *data,
 209                 struct lustre_handle *lockh,
 210                 void *lmm,
 211                 int lmmsize,
 212                 ldlm_completion_callback cb_completion,
 213                 ldlm_blocking_callback cb_blocking,
 214                 void *cb_data)
 215 {
 216         struct ptlrpc_request *req;
 217         struct ldlm_res_id res_id = {
 218                 .name = {id_fid(&data->id1), id_group(&data->id1)}
 219         };
 220         struct obd_device *obddev = class_exp2obd(exp);
 221         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 222         struct ldlm_intent *lit;
 223         struct ldlm_request *lockreq;
 224         int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
 225                           [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
 226                           [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
 227         int repsize[5] = {sizeof(struct ldlm_reply),
 228                           sizeof(struct mds_body),
 229                           obddev->u.cli.cl_max_mds_easize};
 230         int req_buffers = 3, reply_buffers = 0;
 231         int rc, flags = LDLM_FL_HAS_INTENT;
 232         struct ldlm_reply *dlm_rep = NULL;
 233         void *eadata;
 234         unsigned long irqflags;
 235         ENTRY;
 236
 237 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 238 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 239
 240         reqsize[0] = lustre_secdesc_size();
 241
 242         if (it->it_op & IT_OPEN) {
 243                 struct lustre_intent_data *lustre_data =
 244                                    (struct lustre_intent_data *) it->d.fs_data;
 245                 it->it_create_mode |= S_IFREG;
 246                 it->it_create_mode &= ~current->fs->umask;
 247
 248                 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
 249                 reqsize[req_buffers++] = data->namelen + 1;
 250                 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
 251                 /*pack the lustre key*/
 252                 if (lustre_data->it_key_size > 0 && lustre_data->it_key)
 253                         reqsize[req_buffers++] = lustre_data->it_key_size;
 254
 255                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 256                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 257                 if (!req)
 258                         RETURN(-ENOMEM);
 259
 260                 spin_lock_irqsave (&req->rq_lock, irqflags);
 261                 req->rq_replay = 1;
 262                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 263
 264                 /* pack the intent */
 265                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 266                                      sizeof (*lit));
 267                 lit->opc = (__u64)it->it_op;
 268
 269                 /* pack the intended request */
 270                 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
 271                               it->it_create_mode, 0, it->it_flags, lmm, lmmsize,
 272                               lustre_data->it_key, lustre_data->it_key_size);
 273                 /* get ready for the reply */
 274                 reply_buffers = 3;
 275                 repsize[reply_buffers++] = sizeof(int);
 276                 repsize[reply_buffers++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 277                 repsize[reply_buffers++] = sizeof(int);
 278                 repsize[reply_buffers++] = sizeof(struct crypto_key);
 279                 req->rq_replen = lustre_msg_size(reply_buffers, repsize);
 280         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 281                 __u64 valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE |
 282                             OBD_MD_FLACL | OBD_MD_FLKEY;
 283
 284                 /* we don't expect xattr retrieve could reach here */
 285                 LASSERT(!(valid & (OBD_MD_FLXATTR | OBD_MD_FLXATTRLIST)));
 286
 287                 reqsize[req_buffers++] = sizeof(struct mds_body);
 288                 reqsize[req_buffers++] = data->namelen + 1;
 289
 290                 if (it->it_op & IT_GETATTR)
 291                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 292
 293                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 294                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 295
 296                 if (!req)
 297                         RETURN(-ENOMEM);
 298
 299                 /* pack the intent */
 300                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 301                                      sizeof (*lit));
 302                 lit->opc = (__u64)it->it_op;
 303
 304                 /* pack the intended request */
 305                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 306                                  valid, it->it_flags, data);
 307
 308                 /* get ready for the reply */
 309                 reply_buffers = 3;
 310                 repsize[reply_buffers++] = sizeof(int);
 311                 repsize[reply_buffers++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 312                 repsize[reply_buffers++] = sizeof(int);
 313                 repsize[reply_buffers++] = sizeof(struct crypto_key);
 314                 req->rq_replen = lustre_msg_size(reply_buffers, repsize);
 315         } else if (it->it_op == IT_READDIR) {
 316                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 317                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 318                                       LDLM_ENQUEUE, 2, reqsize, NULL);
 319
 320                 if (!req)
 321                         RETURN(-ENOMEM);
 322                 /* get ready for the reply */
 323                 reply_buffers = 1;
 324                 req->rq_replen = lustre_msg_size(1, repsize);
 325         } else if (it->it_op == IT_UNLINK) {
 326                 reqsize[req_buffers++] = sizeof(struct mds_body);
 327                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 328                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 329                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 330                 if (!req)
 331                         RETURN(-ENOMEM);
 332
 333                 /* pack the intended request */
 334                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 335                                  0, 0, data);
 336
 337                 /* pack the intent */
 338                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 339                                      sizeof (*lit));
 340                 lit->opc = (__u64)it->it_op;
 341
 342                 /* get ready for the reply */
 343                 reply_buffers = 3;
 344                 req->rq_replen = lustre_msg_size(3, repsize);
 345         } else {
 346                 LBUG();
 347                 RETURN(-EINVAL);
 348         }
 349
 350         lustre_pack_secdesc(req, reqsize[0]);
 351
 352         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 353         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 354                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 355                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 356                               lockh);
 357         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 358
 359         /* Similarly, if we're going to replay this request, we don't want to
 360          * actually get a lock, just perform the intent. */
 361         if (req->rq_transno || req->rq_replay) {
 362                 lockreq = lustre_msg_buf(req->rq_reqmsg,
 363                                          MDS_REQ_INTENT_LOCKREQ_OFF,
 364                                          sizeof (*lockreq));
 365                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 366         }
 367
 368         /* This can go when we're sure that this can never happen */
 369         LASSERT(rc != -ENOENT);
 370         /* We need dlm_rep to be assigned this early, to check lock mode of
 371            returned lock from request to avoid possible race with lock
 372            conversion */
 373         if (rc == ELDLM_LOCK_ABORTED || !rc) {
 374                 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 375                 LASSERT(dlm_rep != NULL);   /* checked by ldlm_cli_enqueue() */
 376         }
 377         if (rc == ELDLM_LOCK_ABORTED) {
 378                 lock_mode = 0;
 379                 memset(lockh, 0, sizeof(*lockh));
 380                 rc = 0;
 381         } else if (rc != 0) {
 382                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 383                 LASSERTF(rc < 0, "rc = %d\n", rc);
 384                 mdc_clear_replay_flag(req, rc);
 385                 ptlrpc_req_finished(req);
 386                 RETURN(rc);
 387         } else { /* rc = 0 */
 388                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 389                 LASSERT(lock);
 390
 391                 /* If the server gave us back a different lock mode, we should
 392                  * fix up our variables. */
 393                 if (dlm_rep->lock_desc.l_req_mode != lock_mode) {
 394                         ldlm_lock_addref(lockh, dlm_rep->lock_desc.l_req_mode);
 395                         ldlm_lock_decref(lockh, lock_mode);
 396                         lock_mode = dlm_rep->lock_desc.l_req_mode;
 397                 }
 398
 399                 ldlm_lock_allow_match(lock);
 400                 LDLM_LOCK_PUT(lock);
 401         }
 402
 403         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 404
 405         LUSTRE_IT(it)->it_disposition = (int) dlm_rep->lock_policy_res1;
 406         LUSTRE_IT(it)->it_status = (int) dlm_rep->lock_policy_res2;
 407         LUSTRE_IT(it)->it_lock_mode = lock_mode;
 408         LUSTRE_IT(it)->it_data = req;
 409
 410         if (LUSTRE_IT(it)->it_status < 0 && req->rq_replay)
 411                 mdc_clear_replay_flag(req, LUSTRE_IT(it)->it_status);
 412
 413         DEBUG_REQ(D_DLMTRACE, req, "disposition: %x, status: %d",
 414                   LUSTRE_IT(it)->it_disposition, LUSTRE_IT(it)->it_status);
 415
 416         /* We know what to expect, so we do any byte flipping required here */
 417         LASSERT(reply_buffers == 5 || reply_buffers == 4 ||
 418                 reply_buffers == 3 || reply_buffers == 1 ||
 419                 reply_buffers == 6 || reply_buffers == 7);
 420         if (reply_buffers >= 3) {
 421                 struct mds_body *body;
 422
 423                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 424                                           lustre_swab_mds_body);
 425                 if (body == NULL) {
 426                         CERROR ("Can't swab mds_body\n");
 427                         RETURN (-EPROTO);
 428                 }
 429
 430                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 431                         /* The eadata is opaque; just check that it is there.
 432                          * Eventually, obd_unpackmd() will check the contents */
 433                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 434                                                     NULL);
 435                         if (eadata == NULL) {
 436                                 CERROR ("Missing/short eadata\n");
 437                                 RETURN (-EPROTO);
 438                         }
 439                         if (it->it_op & IT_OPEN) {
 440                                 void *replayea;
 441
 442                                 replayea = lustre_msg_buf(req->rq_reqmsg,
 443                                                           MDS_REQ_INTENT_REC_OFF + 2,
 444                                                           body->eadatasize);
 445                                 LASSERT(replayea);
 446                                 memcpy(replayea, eadata, body->eadatasize);
 447
 448                                 LASSERT(req->rq_reqmsg->bufcount == 6 ||
 449                                         req->rq_reqmsg->bufcount == 7);
 450                                 req->rq_reqmsg->buflens[5] = body->eadatasize;
 451                                 /* If this isn't the last buffer, we might
 452                                  * have to shift other data around. */
 453                         }
 454                 }
 455         }
 456
 457         RETURN(rc);
 458 }
 459 EXPORT_SYMBOL(mdc_enqueue);
 460
 461 /*
 462  * This long block is all about fixing up the lock and request state
 463  * so that it is correct as of the moment _before_ the operation was
 464  * applied; that way, the VFS will think that everything is normal and
 465  * call Lustre's regular VFS methods.
 466  *
 467  * If we're performing a creation, that means that unless the creation
 468  * failed with EEXIST, we should fake up a negative dentry.
 469  *
 470  * For everything else, we want to lookup to succeed.
 471  *
 472  * One additional note: if CREATE or OPEN succeeded, we add an extra
 473  * reference to the request because we need to keep it around until
 474  * ll_create/ll_open gets called.
 475  *
 476  * The server will return to us, in it_disposition, an indication of
 477  * exactly what d.lustre->it_status refers to.
 478  *
 479  * If DISP_OPEN_OPEN is set, then d.lustre->it_status refers to the open() call,
 480  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 481  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 482  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 483  * was successful.
 484  *
 485  * Else, if DISP_LOOKUP_EXECD then d.lustre->it_status is the rc of the
 486  * child lookup.
 487  */
 488 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid,
 489                     const char *name, int len, void *lmm, int lmmsize,
 490                     struct lustre_id *cid, struct lookup_intent *it,
 491                     int lookup_flags, struct ptlrpc_request **reqp,
 492                     ldlm_blocking_callback cb_blocking)
 493 {
 494         struct lustre_handle lockh;
 495         struct ptlrpc_request *request;
 496         struct mds_body *mds_body;
 497         struct lustre_handle old_lock;
 498         struct ldlm_lock *lock;
 499         int rc = 0;
 500         ENTRY;
 501         LASSERT(it);
 502
 503         CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
 504                len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
 505
 506         if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 507                     it->it_op == IT_CHDIR)) {
 508                 /* We could just return 1 immediately, but since we should only
 509                  * be called in revalidate_it if we already have a lock, let's
 510                  * verify that. */
 511                 struct ldlm_res_id res_id = {.name = {id_fid(cid),
 512                                                       id_group(cid)}};
 513                 struct lustre_handle lockh;
 514                 ldlm_policy_data_t policy;
 515                 int mode;
 516
 517                 /* For the GETATTR case, ll_revalidate_it issues two separate
 518                    queries - for LOOKUP and for UPDATE lock because it cannot
 519                    check them together - we might have those two bits to be
 520                    present in two separate granted locks */
 521                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 522                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 523
 524                 mode = LCK_PR;
 525                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 526                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 527                                      LDLM_IBITS, &policy, mode,
 528                                      &lockh);
 529
 530                 if (!rc) {
 531                         mode = LCK_CR;
 532                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 533                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 534                                              LDLM_IBITS, &policy, mode,
 535                                              &lockh);
 536                 }
 537                 if (!rc) {
 538                         mode = LCK_PW;
 539                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 540                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 541                                              LDLM_IBITS, &policy, mode,
 542                                              &lockh);
 543                 }
 544                 if (!rc) {
 545                         mode = LCK_CW;
 546                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 547                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 548                                              LDLM_IBITS, &policy, mode,
 549                                              &lockh);
 550                 }
 551                 if (rc) {
 552                         if (ptlrpcs_check_cred(exp->exp_obd->u.cli.cl_import)) {
 553                                 /* return immediately if no credential held */
 554                                 ldlm_lock_decref(&lockh, mode);
 555                                 RETURN(-EACCES);
 556                         }
 557                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 558                                sizeof(lockh));
 559                         LUSTRE_IT(it)->it_lock_mode = mode;
 560                 }
 561
 562                 /* Only return failure if it was not GETATTR by cid (from
 563                    inode_revalidate) */
 564                 if (rc || name)
 565                         RETURN(rc);
 566         }
 567
 568         /* lookup_it may be called only after revalidate_it has run, because
 569          * revalidate_it cannot return errors, only zero.  Returning zero causes
 570          * this call to lookup, which *can* return an error.
 571          *
 572          * We only want to execute the request associated with the intent one
 573          * time, however, so don't send the request again.  Instead, skip past
 574          * this and use the request from revalidate.  In this case, revalidate
 575          * never dropped its reference, so the refcounts are all OK */
 576         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 577                 struct mdc_op_data *op_data;
 578
 579                 OBD_ALLOC(op_data, sizeof(*op_data));
 580                 if (op_data == NULL)
 581                         RETURN(-ENOMEM);
 582
 583                 mdc_id2mdc_data(op_data, pid, cid, name, len, 0);
 584
 585                 if (name != NULL)
 586                         op_data->valid |= OBD_MD_FID;
 587
 588                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 589                                  op_data, &lockh, lmm, lmmsize,
 590                                  ldlm_completion_ast, cb_blocking, NULL);
 591                 OBD_FREE(op_data, sizeof(*op_data));
 592                 if (rc < 0)
 593                         RETURN(rc);
 594
 595                 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, sizeof(lockh));
 596         }
 597         request = *reqp = LUSTRE_IT(it)->it_data;
 598         LASSERT(request != NULL);
 599
 600         /* If we're doing an IT_OPEN which did not result in an actual
 601          * successful open, then we need to remove the bit which saves this
 602          * request for unconditional replay.
 603          *
 604          * It's important that we do this first!  Otherwise we might exit the
 605          * function without doing so, and try to replay a failed create (bug
 606          * 3440) */
 607         if (it->it_op & IT_OPEN && request->rq_replay &&
 608             (!it_disposition(it, DISP_OPEN_OPEN) || LUSTRE_IT(it)->it_status != 0))
 609                 mdc_clear_replay_flag(request, LUSTRE_IT(it)->it_status);
 610
 611         if (!it_disposition(it, DISP_IT_EXECD)) {
 612                 /* The server failed before it even started executing the
 613                  * intent, i.e. because it couldn't unpack the request. */
 614                 LASSERT(LUSTRE_IT(it)->it_status != 0);
 615                 RETURN(LUSTRE_IT(it)->it_status);
 616         }
 617         rc = it_open_error(DISP_IT_EXECD, it);
 618         if (rc)
 619                 RETURN(rc);
 620
 621         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 622         LASSERT(mds_body != NULL);      /* mdc_enqueue checked */
 623         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 624
 625         /* If we were revalidating a fid/name pair, mark the intent in case we
 626          * fail and get called again from lookup */
 627         if (cid != NULL) {
 628                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 629                 /* Also: did we find the same inode? */
 630
 631                 /* we have to compare all the fields but type, because MDS can
 632                  * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
 633                 if (!(lookup_flags & LOOKUP_COBD) && !id_equal(cid, &mds_body->id1))
 634                         RETURN(-ESTALE);
 635         }
 636
 637         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 638         if (rc)
 639                 RETURN(rc);
 640
 641         /*
 642          * keep requests around for the multiple phases of the call this shows
 643          * the DISP_XX must guarantee we make it into the call.
 644          */
 645         if (it_disposition(it, DISP_OPEN_CREATE) &&
 646             !it_open_error(DISP_OPEN_CREATE, it))
 647                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 648         if (it_disposition(it, DISP_OPEN_OPEN) &&
 649             !it_open_error(DISP_OPEN_OPEN, it))
 650                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 651
 652         if (it->it_op & IT_CREAT) {
 653                 /* XXX this belongs in ll_create_it */
 654         } else if (it->it_op == IT_OPEN) {
 655                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 656         } else {
 657                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 658         }
 659
 660         /*
 661          * if we already have a matching lock, then cancel the new one. We have
 662          * to set the data here instead of in mdc_enqueue, because we need to
 663          * use the child's inode as the l_ast_data to match, and that's not
 664          * available until intent_finish has performed the iget().)
 665          */
 666         lock = ldlm_handle2lock(&lockh);
 667         if (lock) {
 668                 ldlm_policy_data_t policy = lock->l_policy_data;
 669                 LDLM_DEBUG(lock, "matching against this");
 670                 LDLM_LOCK_PUT(lock);
 671
 672                 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
 673                          id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
 674                          "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
 675                          "response res_id: %lu/%lu.\n",
 676                          (unsigned long)lock->l_resource->lr_name.name[0],
 677                          (unsigned long)lock->l_resource->lr_name.name[1],
 678                          (unsigned long)id_fid(&mds_body->id1),
 679                          (unsigned long)id_group(&mds_body->id1));
 680
 681                 memcpy(&old_lock, &lockh, sizeof(lockh));
 682                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 683                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 684                         ldlm_lock_decref_and_cancel(&lockh,
 685                                                     LUSTRE_IT(it)->it_lock_mode);
 686                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 687                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 688                                sizeof(lockh));
 689                 }
 690         }
 691         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 692                len, name, ldlm_it2str(it->it_op), LUSTRE_IT(it)->it_status,
 693                LUSTRE_IT(it)->it_disposition, rc);
 694
 695         RETURN(rc);
 696 }
 697 EXPORT_SYMBOL(mdc_intent_lock);