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