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, 2015, 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 #include <linux/module.h>
  40
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include <lustre_dlm.h>
  44 #include <lustre_fid.h>
  45 #include <lustre_intent.h>
  46 #include <lustre_mdc.h>
  47 #include <lustre_net.h>
  48 #include <lustre_req_layout.h>
  49 #include <lustre_swab.h>
  50
  51 #include "mdc_internal.h"
  52
  53 struct mdc_getattr_args {
  54         struct obd_export               *ga_exp;
  55         struct md_enqueue_info          *ga_minfo;
  56 };
  57
  58 int it_open_error(int phase, struct lookup_intent *it)
  59 {
  60         if (it_disposition(it, DISP_OPEN_LEASE)) {
  61                 if (phase >= DISP_OPEN_LEASE)
  62                         return it->it_status;
  63                 else
  64                         return 0;
  65         }
  66         if (it_disposition(it, DISP_OPEN_OPEN)) {
  67                 if (phase >= DISP_OPEN_OPEN)
  68                         return it->it_status;
  69                 else
  70                         return 0;
  71         }
  72
  73         if (it_disposition(it, DISP_OPEN_CREATE)) {
  74                 if (phase >= DISP_OPEN_CREATE)
  75                         return it->it_status;
  76                 else
  77                         return 0;
  78         }
  79
  80         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  81                 if (phase >= DISP_LOOKUP_EXECD)
  82                         return it->it_status;
  83                 else
  84                         return 0;
  85         }
  86
  87         if (it_disposition(it, DISP_IT_EXECD)) {
  88                 if (phase >= DISP_IT_EXECD)
  89                         return it->it_status;
  90                 else
  91                         return 0;
  92         }
  93
  94         CERROR("it disp: %X, status: %d\n", it->it_disposition, it->it_status);
  95         LBUG();
  96
  97         return 0;
  98 }
  99 EXPORT_SYMBOL(it_open_error);
 100
 101 /* this must be called on a lockh that is known to have a referenced lock */
 102 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
 103                       __u64 *bits)
 104 {
 105         struct ldlm_lock *lock;
 106         struct inode *new_inode = data;
 107         ENTRY;
 108
 109         if(bits)
 110                 *bits = 0;
 111
 112         if (!*lockh)
 113                 RETURN(0);
 114
 115         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 116
 117         LASSERT(lock != NULL);
 118         lock_res_and_lock(lock);
 119         if (lock->l_resource->lr_lvb_inode &&
 120             lock->l_resource->lr_lvb_inode != data) {
 121                 struct inode *old_inode = lock->l_resource->lr_lvb_inode;
 122                 LASSERTF(old_inode->i_state & I_FREEING,
 123                          "Found existing inode %p/%lu/%u state %lu in lock: "
 124                          "setting data to %p/%lu/%u\n", old_inode,
 125                          old_inode->i_ino, old_inode->i_generation,
 126                          old_inode->i_state,
 127                          new_inode, new_inode->i_ino, new_inode->i_generation);
 128         }
 129         lock->l_resource->lr_lvb_inode = new_inode;
 130         if (bits)
 131                 *bits = lock->l_policy_data.l_inodebits.bits;
 132
 133         unlock_res_and_lock(lock);
 134         LDLM_LOCK_PUT(lock);
 135
 136         RETURN(0);
 137 }
 138
 139 enum ldlm_mode mdc_lock_match(struct obd_export *exp, __u64 flags,
 140                               const struct lu_fid *fid, enum ldlm_type type,
 141                               union ldlm_policy_data *policy,
 142                               enum ldlm_mode mode, struct lustre_handle *lockh)
 143 {
 144         struct ldlm_res_id res_id;
 145         enum ldlm_mode rc;
 146         ENTRY;
 147
 148         fid_build_reg_res_name(fid, &res_id);
 149         /* LU-4405: Clear bits not supported by server */
 150         policy->l_inodebits.bits &= exp_connect_ibits(exp);
 151         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 152                              &res_id, type, policy, mode, lockh, 0);
 153         RETURN(rc);
 154 }
 155
 156 int mdc_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
 157                       union ldlm_policy_data *policy, enum ldlm_mode mode,
 158                       enum ldlm_cancel_flags flags, void *opaque)
 159 {
 160         struct obd_device *obd = class_exp2obd(exp);
 161         struct ldlm_res_id res_id;
 162         int rc;
 163
 164         ENTRY;
 165
 166         fid_build_reg_res_name(fid, &res_id);
 167         rc = ldlm_cli_cancel_unused_resource(obd->obd_namespace, &res_id,
 168                                              policy, mode, flags, opaque);
 169         RETURN(rc);
 170 }
 171
 172 int mdc_null_inode(struct obd_export *exp,
 173                    const struct lu_fid *fid)
 174 {
 175         struct ldlm_res_id res_id;
 176         struct ldlm_resource *res;
 177         struct ldlm_namespace *ns = class_exp2obd(exp)->obd_namespace;
 178         ENTRY;
 179
 180         LASSERTF(ns != NULL, "no namespace passed\n");
 181
 182         fid_build_reg_res_name(fid, &res_id);
 183
 184         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 185         if (IS_ERR(res))
 186                 RETURN(0);
 187
 188         lock_res(res);
 189         res->lr_lvb_inode = NULL;
 190         unlock_res(res);
 191
 192         ldlm_resource_putref(res);
 193         RETURN(0);
 194 }
 195
 196 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 197 {
 198         /* Don't hold error requests for replay. */
 199         if (req->rq_replay) {
 200                 spin_lock(&req->rq_lock);
 201                 req->rq_replay = 0;
 202                 spin_unlock(&req->rq_lock);
 203         }
 204         if (rc && req->rq_transno != 0) {
 205                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 206                 LBUG();
 207         }
 208 }
 209
 210 /* Save a large LOV EA into the request buffer so that it is available
 211  * for replay.  We don't do this in the initial request because the
 212  * original request doesn't need this buffer (at most it sends just the
 213  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 214  * buffer and may also be difficult to allocate and save a very large
 215  * request buffer for each open. (bug 5707)
 216  *
 217  * OOM here may cause recovery failure if lmm is needed (only for the
 218  * original open if the MDS crashed just when this client also OOM'd)
 219  * but this is incredibly unlikely, and questionable whether the client
 220  * could do MDS recovery under OOM anyways... */
 221 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 222                                 struct mdt_body *body)
 223 {
 224         int     rc;
 225
 226         /* FIXME: remove this explicit offset. */
 227         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 228                                         body->mbo_eadatasize);
 229         if (rc) {
 230                 CERROR("Can't enlarge segment %d size to %d\n",
 231                        DLM_INTENT_REC_OFF + 4, body->mbo_eadatasize);
 232                 body->mbo_valid &= ~OBD_MD_FLEASIZE;
 233                 body->mbo_eadatasize = 0;
 234         }
 235 }
 236
 237 static struct ptlrpc_request *
 238 mdc_intent_open_pack(struct obd_export *exp, struct lookup_intent *it,
 239                      struct md_op_data *op_data)
 240 {
 241         struct ptlrpc_request   *req;
 242         struct obd_device       *obddev = class_exp2obd(exp);
 243         struct ldlm_intent      *lit;
 244         const void              *lmm = op_data->op_data;
 245         __u32                    lmmsize = op_data->op_data_size;
 246         struct list_head         cancels = LIST_HEAD_INIT(cancels);
 247         int                      count = 0;
 248         enum ldlm_mode           mode;
 249         int                      rc;
 250         ENTRY;
 251
 252         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 253
 254         /* XXX: openlock is not cancelled for cross-refs. */
 255         /* If inode is known, cancel conflicting OPEN locks. */
 256         if (fid_is_sane(&op_data->op_fid2)) {
 257                 if (it->it_flags & MDS_OPEN_LEASE) { /* try to get lease */
 258                         if (it->it_flags & FMODE_WRITE)
 259                                 mode = LCK_EX;
 260                         else
 261                                 mode = LCK_PR;
 262                 } else {
 263                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 264                                 mode = LCK_CW;
 265 #ifdef FMODE_EXEC
 266                         else if (it->it_flags & FMODE_EXEC)
 267                                 mode = LCK_PR;
 268 #endif
 269                         else
 270                                 mode = LCK_CR;
 271                 }
 272                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 273                                                 &cancels, mode,
 274                                                 MDS_INODELOCK_OPEN);
 275         }
 276
 277         /* If CREATE, cancel parent's UPDATE lock. */
 278         if (it->it_op & IT_CREAT)
 279                 mode = LCK_EX;
 280         else
 281                 mode = LCK_CR;
 282         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 283                                          &cancels, mode,
 284                                          MDS_INODELOCK_UPDATE);
 285
 286         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 287                                    &RQF_LDLM_INTENT_OPEN);
 288         if (req == NULL) {
 289                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 290                 RETURN(ERR_PTR(-ENOMEM));
 291         }
 292
 293         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 294                              op_data->op_namelen + 1);
 295         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 296                              max(lmmsize, obddev->u.cli.cl_default_mds_easize));
 297
 298         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 299         if (rc < 0) {
 300                 ptlrpc_request_free(req);
 301                 RETURN(ERR_PTR(rc));
 302         }
 303
 304         spin_lock(&req->rq_lock);
 305         req->rq_replay = req->rq_import->imp_replayable;
 306         spin_unlock(&req->rq_lock);
 307
 308         /* pack the intent */
 309         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 310         lit->opc = (__u64)it->it_op;
 311
 312         /* pack the intended request */
 313         mdc_open_pack(req, op_data, it->it_create_mode, 0, it->it_flags, lmm,
 314                       lmmsize);
 315
 316         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 317                              obddev->u.cli.cl_max_mds_easize);
 318
 319         /* for remote client, fetch remote perm for current user */
 320         if (client_is_remote(exp))
 321                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 322                                      sizeof(struct mdt_remote_perm));
 323         ptlrpc_request_set_replen(req);
 324         return req;
 325 }
 326
 327 static struct ptlrpc_request *
 328 mdc_intent_getxattr_pack(struct obd_export *exp,
 329                          struct lookup_intent *it,
 330                          struct md_op_data *op_data)
 331 {
 332         struct ptlrpc_request   *req;
 333         struct ldlm_intent      *lit;
 334         int                     rc, count = 0;
 335         __u32                   maxdata;
 336         struct list_head        cancels = LIST_HEAD_INIT(cancels);
 337
 338         ENTRY;
 339
 340         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 341                                         &RQF_LDLM_INTENT_GETXATTR);
 342         if (req == NULL)
 343                 RETURN(ERR_PTR(-ENOMEM));
 344
 345         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 346         if (rc) {
 347                 ptlrpc_request_free(req);
 348                 RETURN(ERR_PTR(rc));
 349         }
 350
 351         /* pack the intent */
 352         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 353         lit->opc = IT_GETXATTR;
 354
 355         maxdata = class_exp2cliimp(exp)->imp_connect_data.ocd_max_easize;
 356
 357         /* pack the intended request */
 358         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid, maxdata, -1,
 359                       0);
 360
 361         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 362                                 RCL_SERVER, maxdata);
 363
 364         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS,
 365                                 RCL_SERVER, maxdata);
 366
 367         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS_LENS,
 368                                 RCL_SERVER, maxdata);
 369
 370         ptlrpc_request_set_replen(req);
 371
 372         RETURN(req);
 373 }
 374
 375 static struct ptlrpc_request *mdc_intent_unlink_pack(struct obd_export *exp,
 376                                                      struct lookup_intent *it,
 377                                                      struct md_op_data *op_data)
 378 {
 379         struct ptlrpc_request *req;
 380         struct obd_device     *obddev = class_exp2obd(exp);
 381         struct ldlm_intent    *lit;
 382         int                    rc;
 383         ENTRY;
 384
 385         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 386                                    &RQF_LDLM_INTENT_UNLINK);
 387         if (req == NULL)
 388                 RETURN(ERR_PTR(-ENOMEM));
 389
 390         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 391                              op_data->op_namelen + 1);
 392
 393         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 394         if (rc) {
 395                 ptlrpc_request_free(req);
 396                 RETURN(ERR_PTR(rc));
 397         }
 398
 399         /* pack the intent */
 400         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 401         lit->opc = (__u64)it->it_op;
 402
 403         /* pack the intended request */
 404         mdc_unlink_pack(req, op_data);
 405
 406         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 407                              obddev->u.cli.cl_default_mds_easize);
 408         ptlrpc_request_set_replen(req);
 409         RETURN(req);
 410 }
 411
 412 static struct ptlrpc_request *mdc_intent_getattr_pack(struct obd_export *exp,
 413                                                       struct lookup_intent *it,
 414                                                       struct md_op_data *op_data)
 415 {
 416         struct ptlrpc_request   *req;
 417         struct obd_device       *obddev = class_exp2obd(exp);
 418         u64                      valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 419                                          OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 420                                          OBD_MD_MEA |
 421                                          (client_is_remote(exp) ?
 422                                           OBD_MD_FLRMTPERM : OBD_MD_FLACL);
 423         struct ldlm_intent      *lit;
 424         int                      rc;
 425         __u32                    easize;
 426         ENTRY;
 427
 428         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 429                                    &RQF_LDLM_INTENT_GETATTR);
 430         if (req == NULL)
 431                 RETURN(ERR_PTR(-ENOMEM));
 432
 433         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 434                              op_data->op_namelen + 1);
 435
 436         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 437         if (rc) {
 438                 ptlrpc_request_free(req);
 439                 RETURN(ERR_PTR(rc));
 440         }
 441
 442         /* pack the intent */
 443         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 444         lit->opc = (__u64)it->it_op;
 445
 446         if (obddev->u.cli.cl_default_mds_easize > 0)
 447                 easize = obddev->u.cli.cl_default_mds_easize;
 448         else
 449                 easize = obddev->u.cli.cl_max_mds_easize;
 450
 451         /* pack the intended request */
 452         mdc_getattr_pack(req, valid, it->it_flags, op_data, easize);
 453
 454         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, easize);
 455         if (client_is_remote(exp))
 456                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 457                                      sizeof(struct mdt_remote_perm));
 458         ptlrpc_request_set_replen(req);
 459         RETURN(req);
 460 }
 461
 462 static struct ptlrpc_request *mdc_intent_layout_pack(struct obd_export *exp,
 463                                                      struct lookup_intent *it,
 464                                                      struct md_op_data *unused)
 465 {
 466         struct obd_device     *obd = class_exp2obd(exp);
 467         struct ptlrpc_request *req;
 468         struct ldlm_intent    *lit;
 469         struct layout_intent  *layout;
 470         int rc;
 471         ENTRY;
 472
 473         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 474                                 &RQF_LDLM_INTENT_LAYOUT);
 475         if (req == NULL)
 476                 RETURN(ERR_PTR(-ENOMEM));
 477
 478         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 479         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 480         if (rc) {
 481                 ptlrpc_request_free(req);
 482                 RETURN(ERR_PTR(rc));
 483         }
 484
 485         /* pack the intent */
 486         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 487         lit->opc = (__u64)it->it_op;
 488
 489         /* pack the layout intent request */
 490         layout = req_capsule_client_get(&req->rq_pill, &RMF_LAYOUT_INTENT);
 491         /* LAYOUT_INTENT_ACCESS is generic, specific operation will be
 492          * set for replication */
 493         layout->li_opc = LAYOUT_INTENT_ACCESS;
 494
 495         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 496                              obd->u.cli.cl_default_mds_easize);
 497         ptlrpc_request_set_replen(req);
 498         RETURN(req);
 499 }
 500
 501 static struct ptlrpc_request *
 502 mdc_enqueue_pack(struct obd_export *exp, int lvb_len)
 503 {
 504         struct ptlrpc_request *req;
 505         int rc;
 506         ENTRY;
 507
 508         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 509         if (req == NULL)
 510                 RETURN(ERR_PTR(-ENOMEM));
 511
 512         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 513         if (rc) {
 514                 ptlrpc_request_free(req);
 515                 RETURN(ERR_PTR(rc));
 516         }
 517
 518         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 519         ptlrpc_request_set_replen(req);
 520         RETURN(req);
 521 }
 522
 523 static int mdc_finish_enqueue(struct obd_export *exp,
 524                               struct ptlrpc_request *req,
 525                               struct ldlm_enqueue_info *einfo,
 526                               struct lookup_intent *it,
 527                               struct lustre_handle *lockh,
 528                               int rc)
 529 {
 530         struct req_capsule  *pill = &req->rq_pill;
 531         struct ldlm_request *lockreq;
 532         struct ldlm_reply   *lockrep;
 533         struct ldlm_lock    *lock;
 534         void                *lvb_data = NULL;
 535         __u32                lvb_len = 0;
 536         ENTRY;
 537
 538         LASSERT(rc >= 0);
 539         /* Similarly, if we're going to replay this request, we don't want to
 540          * actually get a lock, just perform the intent. */
 541         if (req->rq_transno || req->rq_replay) {
 542                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 543                 lockreq->lock_flags |= ldlm_flags_to_wire(LDLM_FL_INTENT_ONLY);
 544         }
 545
 546         if (rc == ELDLM_LOCK_ABORTED) {
 547                 einfo->ei_mode = 0;
 548                 memset(lockh, 0, sizeof(*lockh));
 549                 rc = 0;
 550         } else { /* rc = 0 */
 551                 lock = ldlm_handle2lock(lockh);
 552                 LASSERT(lock != NULL);
 553
 554                 /* If the server gave us back a different lock mode, we should
 555                  * fix up our variables. */
 556                 if (lock->l_req_mode != einfo->ei_mode) {
 557                         ldlm_lock_addref(lockh, lock->l_req_mode);
 558                         ldlm_lock_decref(lockh, einfo->ei_mode);
 559                         einfo->ei_mode = lock->l_req_mode;
 560                 }
 561                 LDLM_LOCK_PUT(lock);
 562         }
 563
 564         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 565         LASSERT(lockrep != NULL); /* checked by ldlm_cli_enqueue() */
 566
 567         it->it_disposition = (int)lockrep->lock_policy_res1;
 568         it->it_status = (int)lockrep->lock_policy_res2;
 569         it->it_lock_mode = einfo->ei_mode;
 570         it->it_lock_handle = lockh->cookie;
 571         it->it_request = req;
 572
 573         /* Technically speaking rq_transno must already be zero if
 574          * it_status is in error, so the check is a bit redundant */
 575         if ((!req->rq_transno || it->it_status < 0) && req->rq_replay)
 576                 mdc_clear_replay_flag(req, it->it_status);
 577
 578         /* If we're doing an IT_OPEN which did not result in an actual
 579          * successful open, then we need to remove the bit which saves
 580          * this request for unconditional replay.
 581          *
 582          * It's important that we do this first!  Otherwise we might exit the
 583          * function without doing so, and try to replay a failed create
 584          * (bug 3440) */
 585         if (it->it_op & IT_OPEN && req->rq_replay &&
 586             (!it_disposition(it, DISP_OPEN_OPEN) || it->it_status != 0))
 587                 mdc_clear_replay_flag(req, it->it_status);
 588
 589         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 590                   it->it_op, it->it_disposition, it->it_status);
 591
 592         /* We know what to expect, so we do any byte flipping required here */
 593         if (it->it_op & (IT_OPEN | IT_UNLINK | IT_LOOKUP | IT_GETATTR)) {
 594                 struct mdt_body *body;
 595
 596                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 597                 if (body == NULL) {
 598                         CERROR ("Can't swab mdt_body\n");
 599                         RETURN (-EPROTO);
 600                 }
 601
 602                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 603                     !it_open_error(DISP_OPEN_OPEN, it)) {
 604                         /*
 605                          * If this is a successful OPEN request, we need to set
 606                          * replay handler and data early, so that if replay
 607                          * happens immediately after swabbing below, new reply
 608                          * is swabbed by that handler correctly.
 609                          */
 610                         mdc_set_open_replay_data(NULL, NULL, it);
 611                 }
 612
 613                 if (body->mbo_valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) {
 614                         void *eadata;
 615
 616                         mdc_update_max_ea_from_body(exp, body);
 617
 618                         /*
 619                          * The eadata is opaque; just check that it is there.
 620                          * Eventually, obd_unpackmd() will check the contents.
 621                          */
 622                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 623                                                         body->mbo_eadatasize);
 624                         if (eadata == NULL)
 625                                 RETURN(-EPROTO);
 626
 627                         /* save lvb data and length in case this is for layout
 628                          * lock */
 629                         lvb_data = eadata;
 630                         lvb_len = body->mbo_eadatasize;
 631
 632                         /*
 633                          * We save the reply LOV EA in case we have to replay a
 634                          * create for recovery.  If we didn't allocate a large
 635                          * enough request buffer above we need to reallocate it
 636                          * here to hold the actual LOV EA.
 637                          *
 638                          * To not save LOV EA if request is not going to replay
 639                          * (for example error one).
 640                          */
 641                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 642                                 void *lmm;
 643                                 if (req_capsule_get_size(pill, &RMF_EADATA,
 644                                                          RCL_CLIENT) <
 645                                     body->mbo_eadatasize)
 646                                         mdc_realloc_openmsg(req, body);
 647                                 else
 648                                         req_capsule_shrink(pill, &RMF_EADATA,
 649                                                            body->mbo_eadatasize,
 650                                                            RCL_CLIENT);
 651
 652                                 req_capsule_set_size(pill, &RMF_EADATA,
 653                                                      RCL_CLIENT,
 654                                                      body->mbo_eadatasize);
 655
 656                                 lmm = req_capsule_client_get(pill, &RMF_EADATA);
 657                                 if (lmm)
 658                                         memcpy(lmm, eadata,
 659                                                body->mbo_eadatasize);
 660                         }
 661                 }
 662
 663                 if (body->mbo_valid & OBD_MD_FLRMTPERM) {
 664                         struct mdt_remote_perm *perm;
 665
 666                         LASSERT(client_is_remote(exp));
 667                         perm = req_capsule_server_swab_get(pill, &RMF_ACL,
 668                                                 lustre_swab_mdt_remote_perm);
 669                         if (perm == NULL)
 670                                 RETURN(-EPROTO);
 671                 }
 672         } else if (it->it_op & IT_LAYOUT) {
 673                 /* maybe the lock was granted right away and layout
 674                  * is packed into RMF_DLM_LVB of req */
 675                 lvb_len = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
 676                 if (lvb_len > 0) {
 677                         lvb_data = req_capsule_server_sized_get(pill,
 678                                                         &RMF_DLM_LVB, lvb_len);
 679                         if (lvb_data == NULL)
 680                                 RETURN(-EPROTO);
 681                 }
 682         }
 683
 684         /* fill in stripe data for layout lock.
 685          * LU-6581: trust layout data only if layout lock is granted. The MDT
 686          * has stopped sending layout unless the layout lock is granted. The
 687          * client still does this checking in case it's talking with an old
 688          * server. - Jinshan */
 689         lock = ldlm_handle2lock(lockh);
 690         if (lock != NULL && ldlm_has_layout(lock) && lvb_data != NULL &&
 691             !(lockrep->lock_flags & LDLM_FL_BLOCKED_MASK)) {
 692                 void *lmm;
 693
 694                 LDLM_DEBUG(lock, "layout lock returned by: %s, lvb_len: %d",
 695                         ldlm_it2str(it->it_op), lvb_len);
 696
 697                 OBD_ALLOC_LARGE(lmm, lvb_len);
 698                 if (lmm == NULL) {
 699                         LDLM_LOCK_PUT(lock);
 700                         RETURN(-ENOMEM);
 701                 }
 702                 memcpy(lmm, lvb_data, lvb_len);
 703
 704                 /* install lvb_data */
 705                 lock_res_and_lock(lock);
 706                 if (lock->l_lvb_data == NULL) {
 707                         lock->l_lvb_type = LVB_T_LAYOUT;
 708                         lock->l_lvb_data = lmm;
 709                         lock->l_lvb_len = lvb_len;
 710                         lmm = NULL;
 711                 }
 712                 unlock_res_and_lock(lock);
 713                 if (lmm != NULL)
 714                         OBD_FREE_LARGE(lmm, lvb_len);
 715         }
 716         if (lock != NULL)
 717                 LDLM_LOCK_PUT(lock);
 718
 719         RETURN(rc);
 720 }
 721
 722 /* We always reserve enough space in the reply packet for a stripe MD, because
 723  * we don't know in advance the file type. */
 724 int mdc_enqueue(struct obd_export *exp,
 725                 struct ldlm_enqueue_info *einfo,
 726                 const union ldlm_policy_data *policy,
 727                 struct lookup_intent *it, struct md_op_data *op_data,
 728                 struct lustre_handle *lockh, __u64 extra_lock_flags)
 729 {
 730         struct obd_device *obddev = class_exp2obd(exp);
 731         struct ptlrpc_request *req = NULL;
 732         __u64 flags, saved_flags = extra_lock_flags;
 733         struct ldlm_res_id res_id;
 734         static const union ldlm_policy_data lookup_policy = {
 735                                   .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 736         static const union ldlm_policy_data update_policy = {
 737                                   .l_inodebits = { MDS_INODELOCK_UPDATE } };
 738         static const union ldlm_policy_data layout_policy = {
 739                                   .l_inodebits = { MDS_INODELOCK_LAYOUT } };
 740         static const union ldlm_policy_data getxattr_policy = {
 741                                   .l_inodebits = { MDS_INODELOCK_XATTR } };
 742         int generation, resends = 0;
 743         struct ldlm_reply *lockrep;
 744         enum lvb_type lvb_type = 0;
 745         int rc;
 746         ENTRY;
 747
 748         LASSERTF(!it || einfo->ei_type == LDLM_IBITS, "lock type %d\n",
 749                  einfo->ei_type);
 750         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
 751
 752         if (it != NULL) {
 753                 LASSERT(policy == NULL);
 754
 755                 saved_flags |= LDLM_FL_HAS_INTENT;
 756                 if (it->it_op & (IT_OPEN | IT_UNLINK | IT_GETATTR | IT_READDIR))
 757                         policy = &update_policy;
 758                 else if (it->it_op & IT_LAYOUT)
 759                         policy = &layout_policy;
 760                 else if (it->it_op & (IT_GETXATTR | IT_SETXATTR))
 761                         policy = &getxattr_policy;
 762                 else
 763                         policy = &lookup_policy;
 764         }
 765
 766         generation = obddev->u.cli.cl_import->imp_generation;
 767 resend:
 768         flags = saved_flags;
 769         if (it == NULL) {
 770                 /* The only way right now is FLOCK. */
 771                 LASSERTF(einfo->ei_type == LDLM_FLOCK, "lock type %d\n",
 772                          einfo->ei_type);
 773                 res_id.name[3] = LDLM_FLOCK;
 774         } else if (it->it_op & IT_OPEN) {
 775                 req = mdc_intent_open_pack(exp, it, op_data);
 776         } else if (it->it_op & IT_UNLINK) {
 777                 req = mdc_intent_unlink_pack(exp, it, op_data);
 778         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 779                 req = mdc_intent_getattr_pack(exp, it, op_data);
 780         } else if (it->it_op & IT_READDIR) {
 781                 req = mdc_enqueue_pack(exp, 0);
 782         } else if (it->it_op & IT_LAYOUT) {
 783                 if (!imp_connect_lvb_type(class_exp2cliimp(exp)))
 784                         RETURN(-EOPNOTSUPP);
 785                 req = mdc_intent_layout_pack(exp, it, op_data);
 786                 lvb_type = LVB_T_LAYOUT;
 787         } else if (it->it_op & IT_GETXATTR) {
 788                 req = mdc_intent_getxattr_pack(exp, it, op_data);
 789         } else {
 790                 LBUG();
 791                 RETURN(-EINVAL);
 792         }
 793
 794         if (IS_ERR(req))
 795                 RETURN(PTR_ERR(req));
 796
 797         if (resends) {
 798                 req->rq_generation_set = 1;
 799                 req->rq_import_generation = generation;
 800                 req->rq_sent = cfs_time_current_sec() + resends;
 801         }
 802
 803         /* It is important to obtain modify RPC slot first (if applicable), so
 804          * that threads that are waiting for a modify RPC slot are not polluting
 805          * our rpcs in flight counter.
 806          * We do not do flock request limiting, though */
 807         if (it) {
 808                 mdc_get_mod_rpc_slot(req, it);
 809                 rc = obd_get_request_slot(&obddev->u.cli);
 810                 if (rc != 0) {
 811                         mdc_put_mod_rpc_slot(req, it);
 812                         mdc_clear_replay_flag(req, 0);
 813                         ptlrpc_req_finished(req);
 814                         RETURN(rc);
 815                 }
 816         }
 817
 818         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, policy, &flags, NULL,
 819                               0, lvb_type, lockh, 0);
 820         if (!it) {
 821                 /* For flock requests we immediatelly return without further
 822                    delay and let caller deal with the rest, since rest of
 823                    this function metadata processing makes no sense for flock
 824                    requests anyway. But in case of problem during comms with
 825                    Server (ETIMEDOUT) or any signal/kill attempt (EINTR), we
 826                    can not rely on caller and this mainly for F_UNLCKs
 827                    (explicits or automatically generated by Kernel to clean
 828                    current FLocks upon exit) that can't be trashed */
 829                 if (((rc == -EINTR) || (rc == -ETIMEDOUT)) &&
 830                     (einfo->ei_type == LDLM_FLOCK) &&
 831                     (einfo->ei_mode == LCK_NL))
 832                         goto resend;
 833                 RETURN(rc);
 834         }
 835
 836         obd_put_request_slot(&obddev->u.cli);
 837         mdc_put_mod_rpc_slot(req, it);
 838
 839         if (rc < 0) {
 840                 CDEBUG(D_INFO, "%s: ldlm_cli_enqueue failed: rc = %d\n",
 841                        obddev->obd_name, rc);
 842
 843                 mdc_clear_replay_flag(req, rc);
 844                 ptlrpc_req_finished(req);
 845                 RETURN(rc);
 846         }
 847
 848         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 849         LASSERT(lockrep != NULL);
 850
 851         lockrep->lock_policy_res2 =
 852                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
 853
 854         /* Retry infinitely when the server returns -EINPROGRESS for the
 855          * intent operation, when server returns -EINPROGRESS for acquiring
 856          * intent lock, we'll retry in after_reply(). */
 857         if (it && (int)lockrep->lock_policy_res2 == -EINPROGRESS) {
 858                 mdc_clear_replay_flag(req, rc);
 859                 ptlrpc_req_finished(req);
 860                 resends++;
 861
 862                 CDEBUG(D_HA, "%s: resend:%d op:%d "DFID"/"DFID"\n",
 863                        obddev->obd_name, resends, it->it_op,
 864                        PFID(&op_data->op_fid1), PFID(&op_data->op_fid2));
 865
 866                 if (generation == obddev->u.cli.cl_import->imp_generation) {
 867                         goto resend;
 868                 } else {
 869                         CDEBUG(D_HA, "resend cross eviction\n");
 870                         RETURN(-EIO);
 871                 }
 872         }
 873
 874         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
 875         if (rc < 0) {
 876                 if (lustre_handle_is_used(lockh)) {
 877                         ldlm_lock_decref(lockh, einfo->ei_mode);
 878                         memset(lockh, 0, sizeof(*lockh));
 879                 }
 880                 ptlrpc_req_finished(req);
 881
 882                 it->it_lock_handle = 0;
 883                 it->it_lock_mode = 0;
 884                 it->it_request = NULL;
 885         }
 886
 887         RETURN(rc);
 888 }
 889
 890 static int mdc_finish_intent_lock(struct obd_export *exp,
 891                                   struct ptlrpc_request *request,
 892                                   struct md_op_data *op_data,
 893                                   struct lookup_intent *it,
 894                                   struct lustre_handle *lockh)
 895 {
 896         struct lustre_handle old_lock;
 897         struct mdt_body *mdt_body;
 898         struct ldlm_lock *lock;
 899         int rc;
 900         ENTRY;
 901
 902         LASSERT(request != NULL);
 903         LASSERT(request != LP_POISON);
 904         LASSERT(request->rq_repmsg != LP_POISON);
 905
 906         if (it->it_op & IT_READDIR)
 907                 RETURN(0);
 908
 909         if (!it_disposition(it, DISP_IT_EXECD)) {
 910                 /* The server failed before it even started executing the
 911                  * intent, i.e. because it couldn't unpack the request. */
 912                 LASSERT(it->it_status != 0);
 913                 RETURN(it->it_status);
 914         }
 915         rc = it_open_error(DISP_IT_EXECD, it);
 916         if (rc)
 917                 RETURN(rc);
 918
 919         mdt_body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 920         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 921
 922         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 923         if (rc)
 924                 RETURN(rc);
 925
 926         /* keep requests around for the multiple phases of the call
 927          * this shows the DISP_XX must guarantee we make it into the call
 928          */
 929         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 930             it_disposition(it, DISP_OPEN_CREATE) &&
 931             !it_open_error(DISP_OPEN_CREATE, it)) {
 932                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 933                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 934         }
 935         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 936             it_disposition(it, DISP_OPEN_OPEN) &&
 937             !it_open_error(DISP_OPEN_OPEN, it)) {
 938                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 939                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 940                 /* BUG 11546 - eviction in the middle of open rpc processing */
 941                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 942         }
 943
 944         if (it->it_op & IT_CREAT) {
 945                 /* XXX this belongs in ll_create_it */
 946         } else if (it->it_op == IT_OPEN) {
 947                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 948         } else {
 949                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_LAYOUT));
 950         }
 951
 952         /* If we already have a matching lock, then cancel the new
 953          * one.  We have to set the data here instead of in
 954          * mdc_enqueue, because we need to use the child's inode as
 955          * the l_ast_data to match, and that's not available until
 956          * intent_finish has performed the iget().) */
 957         lock = ldlm_handle2lock(lockh);
 958         if (lock) {
 959                 union ldlm_policy_data policy = lock->l_policy_data;
 960                 LDLM_DEBUG(lock, "matching against this");
 961
 962                 LASSERTF(fid_res_name_eq(&mdt_body->mbo_fid1,
 963                                          &lock->l_resource->lr_name),
 964                          "Lock res_id: "DLDLMRES", fid: "DFID"\n",
 965                          PLDLMRES(lock->l_resource), PFID(&mdt_body->mbo_fid1));
 966                 LDLM_LOCK_PUT(lock);
 967
 968                 memcpy(&old_lock, lockh, sizeof(*lockh));
 969                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 970                                     LDLM_IBITS, &policy, LCK_NL, &old_lock, 0)) {
 971                         ldlm_lock_decref_and_cancel(lockh, it->it_lock_mode);
 972                         memcpy(lockh, &old_lock, sizeof(old_lock));
 973                         it->it_lock_handle = lockh->cookie;
 974                 }
 975         }
 976
 977         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 978                 (int)op_data->op_namelen, op_data->op_name,
 979                 ldlm_it2str(it->it_op), it->it_status,
 980                 it->it_disposition, rc);
 981
 982         RETURN(rc);
 983 }
 984
 985 int mdc_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
 986                         struct lu_fid *fid, __u64 *bits)
 987 {
 988         /* We could just return 1 immediately, but since we should only
 989          * be called in revalidate_it if we already have a lock, let's
 990          * verify that. */
 991         struct ldlm_res_id res_id;
 992         struct lustre_handle lockh;
 993         union ldlm_policy_data policy;
 994         enum ldlm_mode mode;
 995         ENTRY;
 996
 997         if (it->it_lock_handle) {
 998                 lockh.cookie = it->it_lock_handle;
 999                 mode = ldlm_revalidate_lock_handle(&lockh, bits);
1000         } else {
1001                 fid_build_reg_res_name(fid, &res_id);
1002                 switch (it->it_op) {
1003                 case IT_GETATTR:
1004                         /* File attributes are held under multiple bits:
1005                          * nlink is under lookup lock, size and times are
1006                          * under UPDATE lock and recently we've also got
1007                          * a separate permissions lock for owner/group/acl that
1008                          * were protected by lookup lock before.
1009                          * Getattr must provide all of that information,
1010                          * so we need to ensure we have all of those locks.
1011                          * Unfortunately, if the bits are split across multiple
1012                          * locks, there's no easy way to match all of them here,
1013                          * so an extra RPC would be performed to fetch all
1014                          * of those bits at once for now. */
1015                         /* For new MDTs(> 2.4), UPDATE|PERM should be enough,
1016                          * but for old MDTs (< 2.4), permission is covered
1017                          * by LOOKUP lock, so it needs to match all bits here.*/
1018                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE |
1019                                                   MDS_INODELOCK_LOOKUP |
1020                                                   MDS_INODELOCK_PERM;
1021                         break;
1022                 case IT_READDIR:
1023                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1024                         break;
1025                 case IT_LAYOUT:
1026                         policy.l_inodebits.bits = MDS_INODELOCK_LAYOUT;
1027                         break;
1028                 default:
1029                         policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1030                         break;
1031                 }
1032
1033                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED, fid,
1034                                       LDLM_IBITS, &policy,
1035                                       LCK_CR | LCK_CW | LCK_PR | LCK_PW,
1036                                       &lockh);
1037         }
1038
1039         if (mode) {
1040                 it->it_lock_handle = lockh.cookie;
1041                 it->it_lock_mode = mode;
1042         } else {
1043                 it->it_lock_handle = 0;
1044                 it->it_lock_mode = 0;
1045         }
1046
1047         RETURN(!!mode);
1048 }
1049
1050 /*
1051  * This long block is all about fixing up the lock and request state
1052  * so that it is correct as of the moment _before_ the operation was
1053  * applied; that way, the VFS will think that everything is normal and
1054  * call Lustre's regular VFS methods.
1055  *
1056  * If we're performing a creation, that means that unless the creation
1057  * failed with EEXIST, we should fake up a negative dentry.
1058  *
1059  * For everything else, we want to lookup to succeed.
1060  *
1061  * One additional note: if CREATE or OPEN succeeded, we add an extra
1062  * reference to the request because we need to keep it around until
1063  * ll_create/ll_open gets called.
1064  *
1065  * The server will return to us, in it_disposition, an indication of
1066  * exactly what it_status refers to.
1067  *
1068  * If DISP_OPEN_OPEN is set, then it_status refers to the open() call,
1069  * otherwise if DISP_OPEN_CREATE is set, then it status is the
1070  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
1071  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
1072  * was successful.
1073  *
1074  * Else, if DISP_LOOKUP_EXECD then it_status is the rc of the
1075  * child lookup.
1076  */
1077 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
1078                     struct lookup_intent *it, struct ptlrpc_request **reqp,
1079                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
1080 {
1081         struct ldlm_enqueue_info einfo = {
1082                 .ei_type        = LDLM_IBITS,
1083                 .ei_mode        = it_to_lock_mode(it),
1084                 .ei_cb_bl       = cb_blocking,
1085                 .ei_cb_cp       = ldlm_completion_ast,
1086         };
1087         struct lustre_handle lockh;
1088         int rc = 0;
1089         ENTRY;
1090         LASSERT(it);
1091
1092         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
1093                 ", intent: %s flags %#"LPF64"o\n", (int)op_data->op_namelen,
1094                 op_data->op_name, PFID(&op_data->op_fid2),
1095                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
1096                 it->it_flags);
1097
1098         lockh.cookie = 0;
1099         if (fid_is_sane(&op_data->op_fid2) &&
1100             (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_READDIR))) {
1101                 /* We could just return 1 immediately, but since we should only
1102                  * be called in revalidate_it if we already have a lock, let's
1103                  * verify that. */
1104                 it->it_lock_handle = 0;
1105                 rc = mdc_revalidate_lock(exp, it, &op_data->op_fid2, NULL);
1106                 /* Only return failure if it was not GETATTR by cfid
1107                    (from inode_revalidate) */
1108                 if (rc || op_data->op_namelen != 0)
1109                         RETURN(rc);
1110         }
1111
1112         /* For case if upper layer did not alloc fid, do it now. */
1113         if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
1114                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1115                 if (rc < 0) {
1116                         CERROR("Can't alloc new fid, rc %d\n", rc);
1117                         RETURN(rc);
1118                 }
1119         }
1120
1121         rc = mdc_enqueue(exp, &einfo, NULL, it, op_data, &lockh,
1122                          extra_lock_flags);
1123         if (rc < 0)
1124                 RETURN(rc);
1125
1126         *reqp = it->it_request;
1127         rc = mdc_finish_intent_lock(exp, *reqp, op_data, it, &lockh);
1128         RETURN(rc);
1129 }
1130
1131 static int mdc_intent_getattr_async_interpret(const struct lu_env *env,
1132                                               struct ptlrpc_request *req,
1133                                               void *args, int rc)
1134 {
1135         struct mdc_getattr_args  *ga = args;
1136         struct obd_export        *exp = ga->ga_exp;
1137         struct md_enqueue_info   *minfo = ga->ga_minfo;
1138         struct ldlm_enqueue_info *einfo = &minfo->mi_einfo;
1139         struct lookup_intent     *it;
1140         struct lustre_handle     *lockh;
1141         struct obd_device        *obddev;
1142         struct ldlm_reply        *lockrep;
1143         __u64                     flags = LDLM_FL_HAS_INTENT;
1144         ENTRY;
1145
1146         it    = &minfo->mi_it;
1147         lockh = &minfo->mi_lockh;
1148
1149         obddev = class_exp2obd(exp);
1150
1151         obd_put_request_slot(&obddev->u.cli);
1152         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_GETATTR_ENQUEUE))
1153                 rc = -ETIMEDOUT;
1154
1155         rc = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, 1, einfo->ei_mode,
1156                                    &flags, NULL, 0, lockh, rc);
1157         if (rc < 0) {
1158                 CERROR("ldlm_cli_enqueue_fini: %d\n", rc);
1159                 mdc_clear_replay_flag(req, rc);
1160                 GOTO(out, rc);
1161         }
1162
1163         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1164         LASSERT(lockrep != NULL);
1165
1166         lockrep->lock_policy_res2 =
1167                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1168
1169         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
1170         if (rc)
1171                 GOTO(out, rc);
1172
1173         rc = mdc_finish_intent_lock(exp, req, &minfo->mi_data, it, lockh);
1174         EXIT;
1175
1176 out:
1177         minfo->mi_cb(req, minfo, rc);
1178         return 0;
1179 }
1180
1181 int mdc_intent_getattr_async(struct obd_export *exp,
1182                              struct md_enqueue_info *minfo)
1183 {
1184         struct md_op_data       *op_data = &minfo->mi_data;
1185         struct lookup_intent    *it = &minfo->mi_it;
1186         struct ptlrpc_request   *req;
1187         struct mdc_getattr_args *ga;
1188         struct obd_device       *obddev = class_exp2obd(exp);
1189         struct ldlm_res_id       res_id;
1190         union ldlm_policy_data policy = {
1191                                 .l_inodebits = { MDS_INODELOCK_LOOKUP |
1192                                                  MDS_INODELOCK_UPDATE } };
1193         int                      rc = 0;
1194         __u64                    flags = LDLM_FL_HAS_INTENT;
1195         ENTRY;
1196
1197         CDEBUG(D_DLMTRACE, "name: %.*s in inode "DFID", intent: %s flags %#"
1198                 LPF64"o\n",
1199                 (int)op_data->op_namelen, op_data->op_name,
1200                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op), it->it_flags);
1201
1202         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
1203         req = mdc_intent_getattr_pack(exp, it, op_data);
1204         if (IS_ERR(req))
1205                 RETURN(PTR_ERR(req));
1206
1207         rc = obd_get_request_slot(&obddev->u.cli);
1208         if (rc != 0) {
1209                 ptlrpc_req_finished(req);
1210                 RETURN(rc);
1211         }
1212
1213         rc = ldlm_cli_enqueue(exp, &req, &minfo->mi_einfo, &res_id, &policy,
1214                               &flags, NULL, 0, LVB_T_NONE, &minfo->mi_lockh, 1);
1215         if (rc < 0) {
1216                 obd_put_request_slot(&obddev->u.cli);
1217                 ptlrpc_req_finished(req);
1218                 RETURN(rc);
1219         }
1220
1221         CLASSERT(sizeof(*ga) <= sizeof(req->rq_async_args));
1222         ga = ptlrpc_req_async_args(req);
1223         ga->ga_exp = exp;
1224         ga->ga_minfo = minfo;
1225
1226         req->rq_interpret_reply = mdc_intent_getattr_async_interpret;
1227         ptlrpcd_add_req(req);
1228
1229         RETURN(0);
1230 }