lustre/mdc/mdc_locks.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MDC
  38
  39 #ifdef __KERNEL__
  40 # include <linux/module.h>
  41 # include <linux/lustre_intent.h>
  42 #else
  43 # include <liblustre.h>
  44 #endif
  45
  46 #include <obd.h>
  47 #include <obd_class.h>
  48 #include <lustre_dlm.h>
  49 #include <lustre_fid.h> /* fid_res_name_eq() */
  50 #include <lustre_mdc.h>
  51 #include <lustre_net.h>
  52 #include <lustre_req_layout.h>
  53 #include "mdc_internal.h"
  54
  55 struct mdc_getattr_args {
  56         struct obd_export           *ga_exp;
  57         struct md_enqueue_info      *ga_minfo;
  58         struct ldlm_enqueue_info    *ga_einfo;
  59 };
  60
  61 int it_open_error(int phase, struct lookup_intent *it)
  62 {
  63         if (it_disposition(it, DISP_OPEN_LEASE)) {
  64                 if (phase >= DISP_OPEN_LEASE)
  65                         return it->d.lustre.it_status;
  66                 else
  67                         return 0;
  68         }
  69         if (it_disposition(it, DISP_OPEN_OPEN)) {
  70                 if (phase >= DISP_OPEN_OPEN)
  71                         return it->d.lustre.it_status;
  72                 else
  73                         return 0;
  74         }
  75
  76         if (it_disposition(it, DISP_OPEN_CREATE)) {
  77                 if (phase >= DISP_OPEN_CREATE)
  78                         return it->d.lustre.it_status;
  79                 else
  80                         return 0;
  81         }
  82
  83         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  84                 if (phase >= DISP_LOOKUP_EXECD)
  85                         return it->d.lustre.it_status;
  86                 else
  87                         return 0;
  88         }
  89
  90         if (it_disposition(it, DISP_IT_EXECD)) {
  91                 if (phase >= DISP_IT_EXECD)
  92                         return it->d.lustre.it_status;
  93                 else
  94                         return 0;
  95         }
  96         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
  97                it->d.lustre.it_status);
  98         LBUG();
  99         return 0;
 100 }
 101 EXPORT_SYMBOL(it_open_error);
 102
 103 /* this must be called on a lockh that is known to have a referenced lock */
 104 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
 105                       __u64 *bits)
 106 {
 107         struct ldlm_lock *lock;
 108         struct inode *new_inode = data;
 109         ENTRY;
 110
 111         if(bits)
 112                 *bits = 0;
 113
 114         if (!*lockh)
 115                 RETURN(0);
 116
 117         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 118
 119         LASSERT(lock != NULL);
 120         lock_res_and_lock(lock);
 121 #ifdef __KERNEL__
 122         if (lock->l_resource->lr_lvb_inode &&
 123             lock->l_resource->lr_lvb_inode != data) {
 124                 struct inode *old_inode = lock->l_resource->lr_lvb_inode;
 125                 LASSERTF(old_inode->i_state & I_FREEING,
 126                          "Found existing inode %p/%lu/%u state %lu in lock: "
 127                          "setting data to %p/%lu/%u\n", old_inode,
 128                          old_inode->i_ino, old_inode->i_generation,
 129                          old_inode->i_state,
 130                          new_inode, new_inode->i_ino, new_inode->i_generation);
 131         }
 132 #endif
 133         lock->l_resource->lr_lvb_inode = new_inode;
 134         if (bits)
 135                 *bits = lock->l_policy_data.l_inodebits.bits;
 136
 137         unlock_res_and_lock(lock);
 138         LDLM_LOCK_PUT(lock);
 139
 140         RETURN(0);
 141 }
 142
 143 ldlm_mode_t mdc_lock_match(struct obd_export *exp, __u64 flags,
 144                            const struct lu_fid *fid, ldlm_type_t type,
 145                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
 146                            struct lustre_handle *lockh)
 147 {
 148         struct ldlm_res_id res_id;
 149         ldlm_mode_t rc;
 150         ENTRY;
 151
 152         fid_build_reg_res_name(fid, &res_id);
 153         /* LU-4405: Clear bits not supported by server */
 154         policy->l_inodebits.bits &= exp_connect_ibits(exp);
 155         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 156                              &res_id, type, policy, mode, lockh, 0);
 157         RETURN(rc);
 158 }
 159
 160 int mdc_cancel_unused(struct obd_export *exp,
 161                       const struct lu_fid *fid,
 162                       ldlm_policy_data_t *policy,
 163                       ldlm_mode_t mode,
 164                       ldlm_cancel_flags_t flags,
 165                       void *opaque)
 166 {
 167         struct ldlm_res_id res_id;
 168         struct obd_device *obd = class_exp2obd(exp);
 169         int rc;
 170
 171         ENTRY;
 172
 173         fid_build_reg_res_name(fid, &res_id);
 174         rc = ldlm_cli_cancel_unused_resource(obd->obd_namespace, &res_id,
 175                                              policy, mode, flags, opaque);
 176         RETURN(rc);
 177 }
 178
 179 int mdc_null_inode(struct obd_export *exp,
 180                    const struct lu_fid *fid)
 181 {
 182         struct ldlm_res_id res_id;
 183         struct ldlm_resource *res;
 184         struct ldlm_namespace *ns = class_exp2obd(exp)->obd_namespace;
 185         ENTRY;
 186
 187         LASSERTF(ns != NULL, "no namespace passed\n");
 188
 189         fid_build_reg_res_name(fid, &res_id);
 190
 191         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 192         if(res == NULL)
 193                 RETURN(0);
 194
 195         lock_res(res);
 196         res->lr_lvb_inode = NULL;
 197         unlock_res(res);
 198
 199         ldlm_resource_putref(res);
 200         RETURN(0);
 201 }
 202
 203 /* find any ldlm lock of the inode in mdc
 204  * return 0    not find
 205  *        1    find one
 206  *      < 0    error */
 207 int mdc_find_cbdata(struct obd_export *exp,
 208                     const struct lu_fid *fid,
 209                     ldlm_iterator_t it, void *data)
 210 {
 211         struct ldlm_res_id res_id;
 212         int rc = 0;
 213         ENTRY;
 214
 215         fid_build_reg_res_name((struct lu_fid*)fid, &res_id);
 216         rc = ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace, &res_id,
 217                                    it, data);
 218         if (rc == LDLM_ITER_STOP)
 219                 RETURN(1);
 220         else if (rc == LDLM_ITER_CONTINUE)
 221                 RETURN(0);
 222         RETURN(rc);
 223 }
 224
 225 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 226 {
 227         /* Don't hold error requests for replay. */
 228         if (req->rq_replay) {
 229                 spin_lock(&req->rq_lock);
 230                 req->rq_replay = 0;
 231                 spin_unlock(&req->rq_lock);
 232         }
 233         if (rc && req->rq_transno != 0) {
 234                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 235                 LBUG();
 236         }
 237 }
 238
 239 /* Save a large LOV EA into the request buffer so that it is available
 240  * for replay.  We don't do this in the initial request because the
 241  * original request doesn't need this buffer (at most it sends just the
 242  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 243  * buffer and may also be difficult to allocate and save a very large
 244  * request buffer for each open. (bug 5707)
 245  *
 246  * OOM here may cause recovery failure if lmm is needed (only for the
 247  * original open if the MDS crashed just when this client also OOM'd)
 248  * but this is incredibly unlikely, and questionable whether the client
 249  * could do MDS recovery under OOM anyways... */
 250 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 251                                 struct mdt_body *body)
 252 {
 253         int     rc;
 254
 255         /* FIXME: remove this explicit offset. */
 256         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 257                                         body->eadatasize);
 258         if (rc) {
 259                 CERROR("Can't enlarge segment %d size to %d\n",
 260                        DLM_INTENT_REC_OFF + 4, body->eadatasize);
 261                 body->valid &= ~OBD_MD_FLEASIZE;
 262                 body->eadatasize = 0;
 263         }
 264 }
 265
 266 static struct ptlrpc_request *
 267 mdc_intent_open_pack(struct obd_export *exp, struct lookup_intent *it,
 268                      struct md_op_data *op_data)
 269 {
 270         struct ptlrpc_request   *req;
 271         struct obd_device       *obddev = class_exp2obd(exp);
 272         struct ldlm_intent      *lit;
 273         const void              *lmm = op_data->op_data;
 274         int                      lmmsize = op_data->op_data_size;
 275         struct list_head         cancels = LIST_HEAD_INIT(cancels);
 276         int                      count = 0;
 277         int                      mode;
 278         int                      rc;
 279         ENTRY;
 280
 281         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 282
 283         /* XXX: openlock is not cancelled for cross-refs. */
 284         /* If inode is known, cancel conflicting OPEN locks. */
 285         if (fid_is_sane(&op_data->op_fid2)) {
 286                 if (it->it_flags & MDS_OPEN_LEASE) { /* try to get lease */
 287                         if (it->it_flags & FMODE_WRITE)
 288                                 mode = LCK_EX;
 289                         else
 290                                 mode = LCK_PR;
 291                 } else {
 292                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 293                                 mode = LCK_CW;
 294 #ifdef FMODE_EXEC
 295                         else if (it->it_flags & FMODE_EXEC)
 296                                 mode = LCK_PR;
 297 #endif
 298                         else
 299                                 mode = LCK_CR;
 300                 }
 301                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 302                                                 &cancels, mode,
 303                                                 MDS_INODELOCK_OPEN);
 304         }
 305
 306         /* If CREATE, cancel parent's UPDATE lock. */
 307         if (it->it_op & IT_CREAT)
 308                 mode = LCK_EX;
 309         else
 310                 mode = LCK_CR;
 311         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 312                                          &cancels, mode,
 313                                          MDS_INODELOCK_UPDATE);
 314
 315         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 316                                    &RQF_LDLM_INTENT_OPEN);
 317         if (req == NULL) {
 318                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 319                 RETURN(ERR_PTR(-ENOMEM));
 320         }
 321
 322         /* parent capability */
 323         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 324         /* child capability, reserve the size according to parent capa, it will
 325          * be filled after we get the reply */
 326         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa1);
 327
 328         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 329                              op_data->op_namelen + 1);
 330         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 331                              max(lmmsize, obddev->u.cli.cl_default_mds_easize));
 332
 333         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 334         if (rc < 0) {
 335                 ptlrpc_request_free(req);
 336                 RETURN(ERR_PTR(rc));
 337         }
 338
 339         spin_lock(&req->rq_lock);
 340         req->rq_replay = req->rq_import->imp_replayable;
 341         spin_unlock(&req->rq_lock);
 342
 343         /* pack the intent */
 344         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 345         lit->opc = (__u64)it->it_op;
 346
 347         /* pack the intended request */
 348         mdc_open_pack(req, op_data, it->it_create_mode, 0, it->it_flags, lmm,
 349                       lmmsize);
 350
 351         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 352                              obddev->u.cli.cl_max_mds_easize);
 353
 354         /* for remote client, fetch remote perm for current user */
 355         if (client_is_remote(exp))
 356                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 357                                      sizeof(struct mdt_remote_perm));
 358         ptlrpc_request_set_replen(req);
 359         return req;
 360 }
 361
 362 static struct ptlrpc_request *
 363 mdc_intent_getxattr_pack(struct obd_export *exp,
 364                          struct lookup_intent *it,
 365                          struct md_op_data *op_data)
 366 {
 367         struct ptlrpc_request   *req;
 368         struct ldlm_intent      *lit;
 369         int                     rc, count = 0, maxdata;
 370         struct list_head        cancels = LIST_HEAD_INIT(cancels);
 371
 372         ENTRY;
 373
 374         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 375                                         &RQF_LDLM_INTENT_GETXATTR);
 376         if (req == NULL)
 377                 RETURN(ERR_PTR(-ENOMEM));
 378
 379         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 380
 381         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 382         if (rc) {
 383                 ptlrpc_request_free(req);
 384                 RETURN(ERR_PTR(rc));
 385         }
 386
 387         /* pack the intent */
 388         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 389         lit->opc = IT_GETXATTR;
 390
 391         maxdata = class_exp2cliimp(exp)->imp_connect_data.ocd_max_easize;
 392
 393         /* pack the intended request */
 394         mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
 395                         op_data->op_valid, maxdata, -1, 0);
 396
 397         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 398                                 RCL_SERVER, maxdata);
 399
 400         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS,
 401                                 RCL_SERVER, maxdata);
 402
 403         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS_LENS,
 404                                 RCL_SERVER, maxdata);
 405
 406         ptlrpc_request_set_replen(req);
 407
 408         RETURN(req);
 409 }
 410
 411 static struct ptlrpc_request *mdc_intent_unlink_pack(struct obd_export *exp,
 412                                                      struct lookup_intent *it,
 413                                                      struct md_op_data *op_data)
 414 {
 415         struct ptlrpc_request *req;
 416         struct obd_device     *obddev = class_exp2obd(exp);
 417         struct ldlm_intent    *lit;
 418         int                    rc;
 419         ENTRY;
 420
 421         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 422                                    &RQF_LDLM_INTENT_UNLINK);
 423         if (req == NULL)
 424                 RETURN(ERR_PTR(-ENOMEM));
 425
 426         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 427         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 428                              op_data->op_namelen + 1);
 429
 430         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 431         if (rc) {
 432                 ptlrpc_request_free(req);
 433                 RETURN(ERR_PTR(rc));
 434         }
 435
 436         /* pack the intent */
 437         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 438         lit->opc = (__u64)it->it_op;
 439
 440         /* pack the intended request */
 441         mdc_unlink_pack(req, op_data);
 442
 443         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 444                              obddev->u.cli.cl_default_mds_easize);
 445         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 446                              obddev->u.cli.cl_default_mds_cookiesize);
 447         ptlrpc_request_set_replen(req);
 448         RETURN(req);
 449 }
 450
 451 static struct ptlrpc_request *mdc_intent_getattr_pack(struct obd_export *exp,
 452                                                       struct lookup_intent *it,
 453                                                       struct md_op_data *op_data)
 454 {
 455         struct ptlrpc_request *req;
 456         struct obd_device     *obddev = class_exp2obd(exp);
 457         obd_valid              valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 458                                        OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 459                                        OBD_MD_FLMDSCAPA | OBD_MD_MEA |
 460                                        (client_is_remote(exp) ?
 461                                                OBD_MD_FLRMTPERM : OBD_MD_FLACL);
 462         struct ldlm_intent    *lit;
 463         int                    rc;
 464         int                     easize;
 465         ENTRY;
 466
 467         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 468                                    &RQF_LDLM_INTENT_GETATTR);
 469         if (req == NULL)
 470                 RETURN(ERR_PTR(-ENOMEM));
 471
 472         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 473         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 474                              op_data->op_namelen + 1);
 475
 476         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 477         if (rc) {
 478                 ptlrpc_request_free(req);
 479                 RETURN(ERR_PTR(rc));
 480         }
 481
 482         /* pack the intent */
 483         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 484         lit->opc = (__u64)it->it_op;
 485
 486         if (obddev->u.cli.cl_default_mds_easize > 0)
 487                 easize = obddev->u.cli.cl_default_mds_easize;
 488         else
 489                 easize = obddev->u.cli.cl_max_mds_easize;
 490
 491         /* pack the intended request */
 492         mdc_getattr_pack(req, valid, it->it_flags, op_data, easize);
 493
 494         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, easize);
 495         if (client_is_remote(exp))
 496                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 497                                      sizeof(struct mdt_remote_perm));
 498         ptlrpc_request_set_replen(req);
 499         RETURN(req);
 500 }
 501
 502 static struct ptlrpc_request *mdc_intent_layout_pack(struct obd_export *exp,
 503                                                      struct lookup_intent *it,
 504                                                      struct md_op_data *unused)
 505 {
 506         struct obd_device     *obd = class_exp2obd(exp);
 507         struct ptlrpc_request *req;
 508         struct ldlm_intent    *lit;
 509         struct layout_intent  *layout;
 510         int rc;
 511         ENTRY;
 512
 513         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 514                                 &RQF_LDLM_INTENT_LAYOUT);
 515         if (req == NULL)
 516                 RETURN(ERR_PTR(-ENOMEM));
 517
 518         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 519         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 520         if (rc) {
 521                 ptlrpc_request_free(req);
 522                 RETURN(ERR_PTR(rc));
 523         }
 524
 525         /* pack the intent */
 526         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 527         lit->opc = (__u64)it->it_op;
 528
 529         /* pack the layout intent request */
 530         layout = req_capsule_client_get(&req->rq_pill, &RMF_LAYOUT_INTENT);
 531         /* LAYOUT_INTENT_ACCESS is generic, specific operation will be
 532          * set for replication */
 533         layout->li_opc = LAYOUT_INTENT_ACCESS;
 534
 535         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 536                              obd->u.cli.cl_default_mds_easize);
 537         ptlrpc_request_set_replen(req);
 538         RETURN(req);
 539 }
 540
 541 static struct ptlrpc_request *
 542 mdc_enqueue_pack(struct obd_export *exp, int lvb_len)
 543 {
 544         struct ptlrpc_request *req;
 545         int rc;
 546         ENTRY;
 547
 548         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 549         if (req == NULL)
 550                 RETURN(ERR_PTR(-ENOMEM));
 551
 552         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 553         if (rc) {
 554                 ptlrpc_request_free(req);
 555                 RETURN(ERR_PTR(rc));
 556         }
 557
 558         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 559         ptlrpc_request_set_replen(req);
 560         RETURN(req);
 561 }
 562
 563 static int mdc_finish_enqueue(struct obd_export *exp,
 564                               struct ptlrpc_request *req,
 565                               struct ldlm_enqueue_info *einfo,
 566                               struct lookup_intent *it,
 567                               struct lustre_handle *lockh,
 568                               int rc)
 569 {
 570         struct req_capsule  *pill = &req->rq_pill;
 571         struct ldlm_request *lockreq;
 572         struct ldlm_reply   *lockrep;
 573         struct lustre_intent_data *intent = &it->d.lustre;
 574         struct ldlm_lock    *lock;
 575         void                *lvb_data = NULL;
 576         int                  lvb_len = 0;
 577         ENTRY;
 578
 579         LASSERT(rc >= 0);
 580         /* Similarly, if we're going to replay this request, we don't want to
 581          * actually get a lock, just perform the intent. */
 582         if (req->rq_transno || req->rq_replay) {
 583                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 584                 lockreq->lock_flags |= ldlm_flags_to_wire(LDLM_FL_INTENT_ONLY);
 585         }
 586
 587         if (rc == ELDLM_LOCK_ABORTED) {
 588                 einfo->ei_mode = 0;
 589                 memset(lockh, 0, sizeof(*lockh));
 590                 rc = 0;
 591         } else { /* rc = 0 */
 592                 lock = ldlm_handle2lock(lockh);
 593                 LASSERT(lock != NULL);
 594
 595                 /* If the server gave us back a different lock mode, we should
 596                  * fix up our variables. */
 597                 if (lock->l_req_mode != einfo->ei_mode) {
 598                         ldlm_lock_addref(lockh, lock->l_req_mode);
 599                         ldlm_lock_decref(lockh, einfo->ei_mode);
 600                         einfo->ei_mode = lock->l_req_mode;
 601                 }
 602                 LDLM_LOCK_PUT(lock);
 603         }
 604
 605         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 606         LASSERT(lockrep != NULL); /* checked by ldlm_cli_enqueue() */
 607
 608         intent->it_disposition = (int)lockrep->lock_policy_res1;
 609         intent->it_status = (int)lockrep->lock_policy_res2;
 610         intent->it_lock_mode = einfo->ei_mode;
 611         intent->it_lock_handle = lockh->cookie;
 612         intent->it_data = req;
 613
 614         /* Technically speaking rq_transno must already be zero if
 615          * it_status is in error, so the check is a bit redundant */
 616         if ((!req->rq_transno || intent->it_status < 0) && req->rq_replay)
 617                 mdc_clear_replay_flag(req, intent->it_status);
 618
 619         /* If we're doing an IT_OPEN which did not result in an actual
 620          * successful open, then we need to remove the bit which saves
 621          * this request for unconditional replay.
 622          *
 623          * It's important that we do this first!  Otherwise we might exit the
 624          * function without doing so, and try to replay a failed create
 625          * (bug 3440) */
 626         if (it->it_op & IT_OPEN && req->rq_replay &&
 627             (!it_disposition(it, DISP_OPEN_OPEN) ||intent->it_status != 0))
 628                 mdc_clear_replay_flag(req, intent->it_status);
 629
 630         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 631                   it->it_op, intent->it_disposition, intent->it_status);
 632
 633         /* We know what to expect, so we do any byte flipping required here */
 634         if (it->it_op & (IT_OPEN | IT_UNLINK | IT_LOOKUP | IT_GETATTR)) {
 635                 struct mdt_body *body;
 636
 637                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 638                 if (body == NULL) {
 639                         CERROR ("Can't swab mdt_body\n");
 640                         RETURN (-EPROTO);
 641                 }
 642
 643                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 644                     !it_open_error(DISP_OPEN_OPEN, it)) {
 645                         /*
 646                          * If this is a successful OPEN request, we need to set
 647                          * replay handler and data early, so that if replay
 648                          * happens immediately after swabbing below, new reply
 649                          * is swabbed by that handler correctly.
 650                          */
 651                         mdc_set_open_replay_data(NULL, NULL, it);
 652                 }
 653
 654                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 655                         void *eadata;
 656
 657                         mdc_update_max_ea_from_body(exp, body);
 658
 659                         /*
 660                          * The eadata is opaque; just check that it is there.
 661                          * Eventually, obd_unpackmd() will check the contents.
 662                          */
 663                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 664                                                               body->eadatasize);
 665                         if (eadata == NULL)
 666                                 RETURN(-EPROTO);
 667
 668                         /* save lvb data and length in case this is for layout
 669                          * lock */
 670                         lvb_data = eadata;
 671                         lvb_len = body->eadatasize;
 672
 673                         /*
 674                          * We save the reply LOV EA in case we have to replay a
 675                          * create for recovery.  If we didn't allocate a large
 676                          * enough request buffer above we need to reallocate it
 677                          * here to hold the actual LOV EA.
 678                          *
 679                          * To not save LOV EA if request is not going to replay
 680                          * (for example error one).
 681                          */
 682                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 683                                 void *lmm;
 684                                 if (req_capsule_get_size(pill, &RMF_EADATA,
 685                                                          RCL_CLIENT) <
 686                                     body->eadatasize)
 687                                         mdc_realloc_openmsg(req, body);
 688                                 else
 689                                         req_capsule_shrink(pill, &RMF_EADATA,
 690                                                            body->eadatasize,
 691                                                            RCL_CLIENT);
 692
 693                                 req_capsule_set_size(pill, &RMF_EADATA,
 694                                                      RCL_CLIENT,
 695                                                      body->eadatasize);
 696
 697                                 lmm = req_capsule_client_get(pill, &RMF_EADATA);
 698                                 if (lmm)
 699                                         memcpy(lmm, eadata, body->eadatasize);
 700                         }
 701                 }
 702
 703                 if (body->valid & OBD_MD_FLRMTPERM) {
 704                         struct mdt_remote_perm *perm;
 705
 706                         LASSERT(client_is_remote(exp));
 707                         perm = req_capsule_server_swab_get(pill, &RMF_ACL,
 708                                                 lustre_swab_mdt_remote_perm);
 709                         if (perm == NULL)
 710                                 RETURN(-EPROTO);
 711                 }
 712                 if (body->valid & OBD_MD_FLMDSCAPA) {
 713                         struct lustre_capa *capa, *p;
 714
 715                         capa = req_capsule_server_get(pill, &RMF_CAPA1);
 716                         if (capa == NULL)
 717                                 RETURN(-EPROTO);
 718
 719                         if (it->it_op & IT_OPEN) {
 720                                 /* client fid capa will be checked in replay */
 721                                 p = req_capsule_client_get(pill, &RMF_CAPA2);
 722                                 LASSERT(p);
 723                                 *p = *capa;
 724                         }
 725                 }
 726                 if (body->valid & OBD_MD_FLOSSCAPA) {
 727                         struct lustre_capa *capa;
 728
 729                         capa = req_capsule_server_get(pill, &RMF_CAPA2);
 730                         if (capa == NULL)
 731                                 RETURN(-EPROTO);
 732                 }
 733         } else if (it->it_op & IT_LAYOUT) {
 734                 /* maybe the lock was granted right away and layout
 735                  * is packed into RMF_DLM_LVB of req */
 736                 lvb_len = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
 737                 if (lvb_len > 0) {
 738                         lvb_data = req_capsule_server_sized_get(pill,
 739                                                         &RMF_DLM_LVB, lvb_len);
 740                         if (lvb_data == NULL)
 741                                 RETURN(-EPROTO);
 742                 }
 743         }
 744
 745         /* fill in stripe data for layout lock */
 746         lock = ldlm_handle2lock(lockh);
 747         if (lock != NULL && ldlm_has_layout(lock) && lvb_data != NULL) {
 748                 void *lmm;
 749
 750                 LDLM_DEBUG(lock, "layout lock returned by: %s, lvb_len: %d\n",
 751                         ldlm_it2str(it->it_op), lvb_len);
 752
 753                 OBD_ALLOC_LARGE(lmm, lvb_len);
 754                 if (lmm == NULL) {
 755                         LDLM_LOCK_PUT(lock);
 756                         RETURN(-ENOMEM);
 757                 }
 758                 memcpy(lmm, lvb_data, lvb_len);
 759
 760                 /* install lvb_data */
 761                 lock_res_and_lock(lock);
 762                 if (lock->l_lvb_data == NULL) {
 763                         lock->l_lvb_type = LVB_T_LAYOUT;
 764                         lock->l_lvb_data = lmm;
 765                         lock->l_lvb_len = lvb_len;
 766                         lmm = NULL;
 767                 }
 768                 unlock_res_and_lock(lock);
 769                 if (lmm != NULL)
 770                         OBD_FREE_LARGE(lmm, lvb_len);
 771         }
 772         if (lock != NULL)
 773                 LDLM_LOCK_PUT(lock);
 774
 775         RETURN(rc);
 776 }
 777
 778 /* We always reserve enough space in the reply packet for a stripe MD, because
 779  * we don't know in advance the file type. */
 780 int mdc_enqueue(struct obd_export *exp,
 781                 struct ldlm_enqueue_info *einfo,
 782                 const union ldlm_policy_data *policy,
 783                 struct lookup_intent *it, struct md_op_data *op_data,
 784                 struct lustre_handle *lockh, __u64 extra_lock_flags)
 785 {
 786         struct obd_device     *obddev = class_exp2obd(exp);
 787         struct ptlrpc_request *req = NULL;
 788         __u64                  flags, saved_flags = extra_lock_flags;
 789         int                    rc;
 790         struct ldlm_res_id res_id;
 791         static const ldlm_policy_data_t lookup_policy =
 792                             { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 793         static const ldlm_policy_data_t update_policy =
 794                             { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 795         static const ldlm_policy_data_t layout_policy =
 796                             { .l_inodebits = { MDS_INODELOCK_LAYOUT } };
 797         static const ldlm_policy_data_t getxattr_policy = {
 798                               .l_inodebits = { MDS_INODELOCK_XATTR } };
 799         int                    generation, resends = 0;
 800         struct ldlm_reply     *lockrep;
 801         enum lvb_type          lvb_type = 0;
 802         ENTRY;
 803
 804         LASSERTF(!it || einfo->ei_type == LDLM_IBITS, "lock type %d\n",
 805                  einfo->ei_type);
 806         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
 807
 808         if (it != NULL) {
 809                 LASSERT(policy == NULL);
 810
 811                 saved_flags |= LDLM_FL_HAS_INTENT;
 812                 if (it->it_op & (IT_OPEN | IT_UNLINK | IT_GETATTR | IT_READDIR))
 813                         policy = &update_policy;
 814                 else if (it->it_op & IT_LAYOUT)
 815                         policy = &layout_policy;
 816                 else if (it->it_op & (IT_GETXATTR | IT_SETXATTR))
 817                         policy = &getxattr_policy;
 818                 else
 819                         policy = &lookup_policy;
 820         }
 821
 822         generation = obddev->u.cli.cl_import->imp_generation;
 823 resend:
 824         flags = saved_flags;
 825         if (it == NULL) {
 826                 /* The only way right now is FLOCK. */
 827                 LASSERTF(einfo->ei_type == LDLM_FLOCK, "lock type %d\n",
 828                          einfo->ei_type);
 829                 res_id.name[3] = LDLM_FLOCK;
 830         } else if (it->it_op & IT_OPEN) {
 831                 req = mdc_intent_open_pack(exp, it, op_data);
 832         } else if (it->it_op & IT_UNLINK) {
 833                 req = mdc_intent_unlink_pack(exp, it, op_data);
 834         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 835                 req = mdc_intent_getattr_pack(exp, it, op_data);
 836         } else if (it->it_op & IT_READDIR) {
 837                 req = mdc_enqueue_pack(exp, 0);
 838         } else if (it->it_op & IT_LAYOUT) {
 839                 if (!imp_connect_lvb_type(class_exp2cliimp(exp)))
 840                         RETURN(-EOPNOTSUPP);
 841                 req = mdc_intent_layout_pack(exp, it, op_data);
 842                 lvb_type = LVB_T_LAYOUT;
 843         } else if (it->it_op & IT_GETXATTR) {
 844                 req = mdc_intent_getxattr_pack(exp, it, op_data);
 845         } else {
 846                 LBUG();
 847                 RETURN(-EINVAL);
 848         }
 849
 850         if (IS_ERR(req))
 851                 RETURN(PTR_ERR(req));
 852
 853         if (req != NULL && it && it->it_op & IT_CREAT)
 854                 /* ask ptlrpc not to resend on EINPROGRESS since we have our own
 855                  * retry logic */
 856                 req->rq_no_retry_einprogress = 1;
 857
 858         if (resends) {
 859                 req->rq_generation_set = 1;
 860                 req->rq_import_generation = generation;
 861                 req->rq_sent = cfs_time_current_sec() + resends;
 862         }
 863
 864         /* It is important to obtain rpc_lock first (if applicable), so that
 865          * threads that are serialised with rpc_lock are not polluting our
 866          * rpcs in flight counter. We do not do flock request limiting, though*/
 867         if (it) {
 868                 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 869                 rc = obd_get_request_slot(&obddev->u.cli);
 870                 if (rc != 0) {
 871                         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 872                         mdc_clear_replay_flag(req, 0);
 873                         ptlrpc_req_finished(req);
 874                         RETURN(rc);
 875                 }
 876         }
 877
 878         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, policy, &flags, NULL,
 879                               0, lvb_type, lockh, 0);
 880         if (!it) {
 881                 /* For flock requests we immediatelly return without further
 882                    delay and let caller deal with the rest, since rest of
 883                    this function metadata processing makes no sense for flock
 884                    requests anyway. But in case of problem during comms with
 885                    Server (ETIMEDOUT) or any signal/kill attempt (EINTR), we
 886                    can not rely on caller and this mainly for F_UNLCKs
 887                    (explicits or automatically generated by Kernel to clean
 888                    current FLocks upon exit) that can't be trashed */
 889                 if (((rc == -EINTR) || (rc == -ETIMEDOUT)) &&
 890                     (einfo->ei_type == LDLM_FLOCK) &&
 891                     (einfo->ei_mode == LCK_NL))
 892                         goto resend;
 893                 RETURN(rc);
 894         }
 895
 896         obd_put_request_slot(&obddev->u.cli);
 897         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 898
 899         if (rc < 0) {
 900                 CDEBUG_LIMIT((rc == -EACCES || rc == -EIDRM) ? D_INFO : D_ERROR,
 901                              "%s: ldlm_cli_enqueue failed: rc = %d\n",
 902                              obddev->obd_name, rc);
 903
 904                 mdc_clear_replay_flag(req, rc);
 905                 ptlrpc_req_finished(req);
 906                 RETURN(rc);
 907         }
 908
 909         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 910         LASSERT(lockrep != NULL);
 911
 912         lockrep->lock_policy_res2 =
 913                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
 914
 915         /* Retry the create infinitely when we get -EINPROGRESS from
 916          * server. This is required by the new quota design. */
 917         if (it && it->it_op & IT_CREAT &&
 918             (int)lockrep->lock_policy_res2 == -EINPROGRESS) {
 919                 mdc_clear_replay_flag(req, rc);
 920                 ptlrpc_req_finished(req);
 921                 resends++;
 922
 923                 CDEBUG(D_HA, "%s: resend:%d op:%d "DFID"/"DFID"\n",
 924                        obddev->obd_name, resends, it->it_op,
 925                        PFID(&op_data->op_fid1), PFID(&op_data->op_fid2));
 926
 927                 if (generation == obddev->u.cli.cl_import->imp_generation) {
 928                         goto resend;
 929                 } else {
 930                         CDEBUG(D_HA, "resend cross eviction\n");
 931                         RETURN(-EIO);
 932                 }
 933         }
 934
 935         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
 936         if (rc < 0) {
 937                 if (lustre_handle_is_used(lockh)) {
 938                         ldlm_lock_decref(lockh, einfo->ei_mode);
 939                         memset(lockh, 0, sizeof(*lockh));
 940                 }
 941                 ptlrpc_req_finished(req);
 942         }
 943         RETURN(rc);
 944 }
 945
 946 static int mdc_finish_intent_lock(struct obd_export *exp,
 947                                   struct ptlrpc_request *request,
 948                                   struct md_op_data *op_data,
 949                                   struct lookup_intent *it,
 950                                   struct lustre_handle *lockh)
 951 {
 952         struct lustre_handle old_lock;
 953         struct mdt_body *mdt_body;
 954         struct ldlm_lock *lock;
 955         int rc;
 956         ENTRY;
 957
 958         LASSERT(request != NULL);
 959         LASSERT(request != LP_POISON);
 960         LASSERT(request->rq_repmsg != LP_POISON);
 961
 962         if (it->it_op & IT_READDIR)
 963                 RETURN(0);
 964
 965         if (!it_disposition(it, DISP_IT_EXECD)) {
 966                 /* The server failed before it even started executing the
 967                  * intent, i.e. because it couldn't unpack the request. */
 968                 LASSERT(it->d.lustre.it_status != 0);
 969                 RETURN(it->d.lustre.it_status);
 970         }
 971         rc = it_open_error(DISP_IT_EXECD, it);
 972         if (rc)
 973                 RETURN(rc);
 974
 975         mdt_body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 976         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 977
 978         /* If we were revalidating a fid/name pair, mark the intent in
 979          * case we fail and get called again from lookup */
 980         if (fid_is_sane(&op_data->op_fid2) &&
 981             it->it_create_mode & M_CHECK_STALE &&
 982             it->it_op != IT_GETATTR) {
 983                 /* Also: did we find the same inode? */
 984                 /* sever can return one of two fids:
 985                  * op_fid2 - new allocated fid - if file is created.
 986                  * op_fid3 - existent fid - if file only open.
 987                  * op_fid3 is saved in lmv_intent_open */
 988                 if ((!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1)) &&
 989                     (!lu_fid_eq(&op_data->op_fid3, &mdt_body->fid1))) {
 990                         CDEBUG(D_DENTRY, "Found stale data "DFID"("DFID")/"DFID
 991                                "\n", PFID(&op_data->op_fid2),
 992                                PFID(&op_data->op_fid2), PFID(&mdt_body->fid1));
 993                         RETURN(-ESTALE);
 994                 }
 995         }
 996
 997         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 998         if (rc)
 999                 RETURN(rc);
1000
1001         /* keep requests around for the multiple phases of the call
1002          * this shows the DISP_XX must guarantee we make it into the call
1003          */
1004         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
1005             it_disposition(it, DISP_OPEN_CREATE) &&
1006             !it_open_error(DISP_OPEN_CREATE, it)) {
1007                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
1008                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
1009         }
1010         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
1011             it_disposition(it, DISP_OPEN_OPEN) &&
1012             !it_open_error(DISP_OPEN_OPEN, it)) {
1013                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
1014                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
1015                 /* BUG 11546 - eviction in the middle of open rpc processing */
1016                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
1017         }
1018
1019         if (it->it_op & IT_CREAT) {
1020                 /* XXX this belongs in ll_create_it */
1021         } else if (it->it_op == IT_OPEN) {
1022                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
1023         } else {
1024                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_LAYOUT));
1025         }
1026
1027         /* If we already have a matching lock, then cancel the new
1028          * one.  We have to set the data here instead of in
1029          * mdc_enqueue, because we need to use the child's inode as
1030          * the l_ast_data to match, and that's not available until
1031          * intent_finish has performed the iget().) */
1032         lock = ldlm_handle2lock(lockh);
1033         if (lock) {
1034                 ldlm_policy_data_t policy = lock->l_policy_data;
1035                 LDLM_DEBUG(lock, "matching against this");
1036
1037                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
1038                                          &lock->l_resource->lr_name),
1039                          "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1040                          PLDLMRES(lock->l_resource), PFID(&mdt_body->fid1));
1041                 LDLM_LOCK_PUT(lock);
1042
1043                 memcpy(&old_lock, lockh, sizeof(*lockh));
1044                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
1045                                     LDLM_IBITS, &policy, LCK_NL, &old_lock, 0)) {
1046                         ldlm_lock_decref_and_cancel(lockh,
1047                                                     it->d.lustre.it_lock_mode);
1048                         memcpy(lockh, &old_lock, sizeof(old_lock));
1049                         it->d.lustre.it_lock_handle = lockh->cookie;
1050                 }
1051         }
1052         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
1053                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
1054                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
1055         RETURN(rc);
1056 }
1057
1058 int mdc_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
1059                         struct lu_fid *fid, __u64 *bits)
1060 {
1061         /* We could just return 1 immediately, but since we should only
1062          * be called in revalidate_it if we already have a lock, let's
1063          * verify that. */
1064         struct ldlm_res_id res_id;
1065         struct lustre_handle lockh;
1066         ldlm_policy_data_t policy;
1067         ldlm_mode_t mode;
1068         ENTRY;
1069
1070         if (it->d.lustre.it_lock_handle) {
1071                 lockh.cookie = it->d.lustre.it_lock_handle;
1072                 mode = ldlm_revalidate_lock_handle(&lockh, bits);
1073         } else {
1074                 fid_build_reg_res_name(fid, &res_id);
1075                 switch (it->it_op) {
1076                 case IT_GETATTR:
1077                         /* File attributes are held under multiple bits:
1078                          * nlink is under lookup lock, size and times are
1079                          * under UPDATE lock and recently we've also got
1080                          * a separate permissions lock for owner/group/acl that
1081                          * were protected by lookup lock before.
1082                          * Getattr must provide all of that information,
1083                          * so we need to ensure we have all of those locks.
1084                          * Unfortunately, if the bits are split across multiple
1085                          * locks, there's no easy way to match all of them here,
1086                          * so an extra RPC would be performed to fetch all
1087                          * of those bits at once for now. */
1088                         /* For new MDTs(> 2.4), UPDATE|PERM should be enough,
1089                          * but for old MDTs (< 2.4), permission is covered
1090                          * by LOOKUP lock, so it needs to match all bits here.*/
1091                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE |
1092                                                   MDS_INODELOCK_LOOKUP |
1093                                                   MDS_INODELOCK_PERM;
1094                         break;
1095                 case IT_READDIR:
1096                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1097                         break;
1098                 case IT_LAYOUT:
1099                         policy.l_inodebits.bits = MDS_INODELOCK_LAYOUT;
1100                         break;
1101                 default:
1102                         policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1103                         break;
1104                 }
1105
1106                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED, fid,
1107                                       LDLM_IBITS, &policy,
1108                                       LCK_CR | LCK_CW | LCK_PR | LCK_PW,
1109                                       &lockh);
1110         }
1111
1112         if (mode) {
1113                 it->d.lustre.it_lock_handle = lockh.cookie;
1114                 it->d.lustre.it_lock_mode = mode;
1115         } else {
1116                 it->d.lustre.it_lock_handle = 0;
1117                 it->d.lustre.it_lock_mode = 0;
1118         }
1119
1120         RETURN(!!mode);
1121 }
1122
1123 /*
1124  * This long block is all about fixing up the lock and request state
1125  * so that it is correct as of the moment _before_ the operation was
1126  * applied; that way, the VFS will think that everything is normal and
1127  * call Lustre's regular VFS methods.
1128  *
1129  * If we're performing a creation, that means that unless the creation
1130  * failed with EEXIST, we should fake up a negative dentry.
1131  *
1132  * For everything else, we want to lookup to succeed.
1133  *
1134  * One additional note: if CREATE or OPEN succeeded, we add an extra
1135  * reference to the request because we need to keep it around until
1136  * ll_create/ll_open gets called.
1137  *
1138  * The server will return to us, in it_disposition, an indication of
1139  * exactly what d.lustre.it_status refers to.
1140  *
1141  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
1142  * otherwise if DISP_OPEN_CREATE is set, then it status is the
1143  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
1144  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
1145  * was successful.
1146  *
1147  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
1148  * child lookup.
1149  */
1150 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
1151                     struct lookup_intent *it, struct ptlrpc_request **reqp,
1152                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
1153 {
1154         struct ldlm_enqueue_info einfo = {
1155                 .ei_type        = LDLM_IBITS,
1156                 .ei_mode        = it_to_lock_mode(it),
1157                 .ei_cb_bl       = cb_blocking,
1158                 .ei_cb_cp       = ldlm_completion_ast,
1159         };
1160         struct lustre_handle lockh;
1161         int rc = 0;
1162         ENTRY;
1163         LASSERT(it);
1164
1165         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
1166                 ", intent: %s flags %#"LPF64"o\n", op_data->op_namelen,
1167                 op_data->op_name, PFID(&op_data->op_fid2),
1168                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
1169                 it->it_flags);
1170
1171         lockh.cookie = 0;
1172         if (fid_is_sane(&op_data->op_fid2) &&
1173             (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_READDIR))) {
1174                 /* We could just return 1 immediately, but since we should only
1175                  * be called in revalidate_it if we already have a lock, let's
1176                  * verify that. */
1177                 it->d.lustre.it_lock_handle = 0;
1178                 rc = mdc_revalidate_lock(exp, it, &op_data->op_fid2, NULL);
1179                 /* Only return failure if it was not GETATTR by cfid
1180                    (from inode_revalidate) */
1181                 if (rc || op_data->op_namelen != 0)
1182                         RETURN(rc);
1183         }
1184
1185         /* For case if upper layer did not alloc fid, do it now. */
1186         if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
1187                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1188                 if (rc < 0) {
1189                         CERROR("Can't alloc new fid, rc %d\n", rc);
1190                         RETURN(rc);
1191                 }
1192         }
1193
1194         rc = mdc_enqueue(exp, &einfo, NULL, it, op_data, &lockh,
1195                          extra_lock_flags);
1196         if (rc < 0)
1197                 RETURN(rc);
1198
1199         *reqp = it->d.lustre.it_data;
1200         rc = mdc_finish_intent_lock(exp, *reqp, op_data, it, &lockh);
1201         RETURN(rc);
1202 }
1203
1204 static int mdc_intent_getattr_async_interpret(const struct lu_env *env,
1205                                               struct ptlrpc_request *req,
1206                                               void *args, int rc)
1207 {
1208         struct mdc_getattr_args  *ga = args;
1209         struct obd_export        *exp = ga->ga_exp;
1210         struct md_enqueue_info   *minfo = ga->ga_minfo;
1211         struct ldlm_enqueue_info *einfo = ga->ga_einfo;
1212         struct lookup_intent     *it;
1213         struct lustre_handle     *lockh;
1214         struct obd_device        *obddev;
1215         struct ldlm_reply        *lockrep;
1216         __u64                     flags = LDLM_FL_HAS_INTENT;
1217         ENTRY;
1218
1219         it    = &minfo->mi_it;
1220         lockh = &minfo->mi_lockh;
1221
1222         obddev = class_exp2obd(exp);
1223
1224         obd_put_request_slot(&obddev->u.cli);
1225         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_GETATTR_ENQUEUE))
1226                 rc = -ETIMEDOUT;
1227
1228         rc = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, 1, einfo->ei_mode,
1229                                    &flags, NULL, 0, lockh, rc);
1230         if (rc < 0) {
1231                 CERROR("ldlm_cli_enqueue_fini: %d\n", rc);
1232                 mdc_clear_replay_flag(req, rc);
1233                 GOTO(out, rc);
1234         }
1235
1236         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1237         LASSERT(lockrep != NULL);
1238
1239         lockrep->lock_policy_res2 =
1240                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1241
1242         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
1243         if (rc)
1244                 GOTO(out, rc);
1245
1246         rc = mdc_finish_intent_lock(exp, req, &minfo->mi_data, it, lockh);
1247         EXIT;
1248
1249 out:
1250         OBD_FREE_PTR(einfo);
1251         minfo->mi_cb(req, minfo, rc);
1252         return 0;
1253 }
1254
1255 int mdc_intent_getattr_async(struct obd_export *exp,
1256                              struct md_enqueue_info *minfo,
1257                              struct ldlm_enqueue_info *einfo)
1258 {
1259         struct md_op_data       *op_data = &minfo->mi_data;
1260         struct lookup_intent    *it = &minfo->mi_it;
1261         struct ptlrpc_request   *req;
1262         struct mdc_getattr_args *ga;
1263         struct obd_device       *obddev = class_exp2obd(exp);
1264         struct ldlm_res_id       res_id;
1265         /*XXX: Both MDS_INODELOCK_LOOKUP and MDS_INODELOCK_UPDATE are needed
1266          *     for statahead currently. Consider CMD in future, such two bits
1267          *     maybe managed by different MDS, should be adjusted then. */
1268         ldlm_policy_data_t       policy = {
1269                                         .l_inodebits = { MDS_INODELOCK_LOOKUP |
1270                                                          MDS_INODELOCK_UPDATE }
1271                                  };
1272         int                      rc = 0;
1273         __u64                    flags = LDLM_FL_HAS_INTENT;
1274         ENTRY;
1275
1276         CDEBUG(D_DLMTRACE, "name: %.*s in inode "DFID", intent: %s flags %#"
1277                 LPF64"o\n",
1278                 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1279                 ldlm_it2str(it->it_op), it->it_flags);
1280
1281         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
1282         req = mdc_intent_getattr_pack(exp, it, op_data);
1283         if (IS_ERR(req))
1284                 RETURN(PTR_ERR(req));
1285
1286         rc = obd_get_request_slot(&obddev->u.cli);
1287         if (rc != 0) {
1288                 ptlrpc_req_finished(req);
1289                 RETURN(rc);
1290         }
1291
1292         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
1293                               0, LVB_T_NONE, &minfo->mi_lockh, 1);
1294         if (rc < 0) {
1295                 obd_put_request_slot(&obddev->u.cli);
1296                 ptlrpc_req_finished(req);
1297                 RETURN(rc);
1298         }
1299
1300         CLASSERT(sizeof(*ga) <= sizeof(req->rq_async_args));
1301         ga = ptlrpc_req_async_args(req);
1302         ga->ga_exp = exp;
1303         ga->ga_minfo = minfo;
1304         ga->ga_einfo = einfo;
1305
1306         req->rq_interpret_reply = mdc_intent_getattr_async_interpret;
1307         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1308
1309         RETURN(0);
1310 }