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, 2014, 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 #include <obd.h>
  41 #include <obd_class.h>
  42 #include <lustre_dlm.h>
  43 #include <lustre_fid.h> /* fid_res_name_eq() */
  44 #include <lustre_intent.h>
  45 #include <lustre_mdc.h>
  46 #include <lustre_net.h>
  47 #include <lustre_req_layout.h>
  48 #include "mdc_internal.h"
  49
  50 struct mdc_getattr_args {
  51         struct obd_export           *ga_exp;
  52         struct md_enqueue_info      *ga_minfo;
  53         struct ldlm_enqueue_info    *ga_einfo;
  54 };
  55
  56 int it_open_error(int phase, struct lookup_intent *it)
  57 {
  58         if (it_disposition(it, DISP_OPEN_LEASE)) {
  59                 if (phase >= DISP_OPEN_LEASE)
  60                         return it->d.lustre.it_status;
  61                 else
  62                         return 0;
  63         }
  64         if (it_disposition(it, DISP_OPEN_OPEN)) {
  65                 if (phase >= DISP_OPEN_OPEN)
  66                         return it->d.lustre.it_status;
  67                 else
  68                         return 0;
  69         }
  70
  71         if (it_disposition(it, DISP_OPEN_CREATE)) {
  72                 if (phase >= DISP_OPEN_CREATE)
  73                         return it->d.lustre.it_status;
  74                 else
  75                         return 0;
  76         }
  77
  78         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  79                 if (phase >= DISP_LOOKUP_EXECD)
  80                         return it->d.lustre.it_status;
  81                 else
  82                         return 0;
  83         }
  84
  85         if (it_disposition(it, DISP_IT_EXECD)) {
  86                 if (phase >= DISP_IT_EXECD)
  87                         return it->d.lustre.it_status;
  88                 else
  89                         return 0;
  90         }
  91         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
  92                it->d.lustre.it_status);
  93         LBUG();
  94         return 0;
  95 }
  96 EXPORT_SYMBOL(it_open_error);
  97
  98 /* this must be called on a lockh that is known to have a referenced lock */
  99 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
 100                       __u64 *bits)
 101 {
 102         struct ldlm_lock *lock;
 103         struct inode *new_inode = data;
 104         ENTRY;
 105
 106         if(bits)
 107                 *bits = 0;
 108
 109         if (!*lockh)
 110                 RETURN(0);
 111
 112         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 113
 114         LASSERT(lock != NULL);
 115         lock_res_and_lock(lock);
 116         if (lock->l_resource->lr_lvb_inode &&
 117             lock->l_resource->lr_lvb_inode != data) {
 118                 struct inode *old_inode = lock->l_resource->lr_lvb_inode;
 119                 LASSERTF(old_inode->i_state & I_FREEING,
 120                          "Found existing inode %p/%lu/%u state %lu in lock: "
 121                          "setting data to %p/%lu/%u\n", old_inode,
 122                          old_inode->i_ino, old_inode->i_generation,
 123                          old_inode->i_state,
 124                          new_inode, new_inode->i_ino, new_inode->i_generation);
 125         }
 126         lock->l_resource->lr_lvb_inode = new_inode;
 127         if (bits)
 128                 *bits = lock->l_policy_data.l_inodebits.bits;
 129
 130         unlock_res_and_lock(lock);
 131         LDLM_LOCK_PUT(lock);
 132
 133         RETURN(0);
 134 }
 135
 136 ldlm_mode_t mdc_lock_match(struct obd_export *exp, __u64 flags,
 137                            const struct lu_fid *fid, ldlm_type_t type,
 138                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
 139                            struct lustre_handle *lockh)
 140 {
 141         struct ldlm_res_id res_id;
 142         ldlm_mode_t rc;
 143         ENTRY;
 144
 145         fid_build_reg_res_name(fid, &res_id);
 146         /* LU-4405: Clear bits not supported by server */
 147         policy->l_inodebits.bits &= exp_connect_ibits(exp);
 148         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 149                              &res_id, type, policy, mode, lockh, 0);
 150         RETURN(rc);
 151 }
 152
 153 int mdc_cancel_unused(struct obd_export *exp,
 154                       const struct lu_fid *fid,
 155                       ldlm_policy_data_t *policy,
 156                       ldlm_mode_t mode,
 157                       ldlm_cancel_flags_t flags,
 158                       void *opaque)
 159 {
 160         struct ldlm_res_id res_id;
 161         struct obd_device *obd = class_exp2obd(exp);
 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 /* find any ldlm lock of the inode in mdc
 197  * return 0    not find
 198  *        1    find one
 199  *      < 0    error */
 200 int mdc_find_cbdata(struct obd_export *exp,
 201                     const struct lu_fid *fid,
 202                     ldlm_iterator_t it, void *data)
 203 {
 204         struct ldlm_res_id res_id;
 205         int rc = 0;
 206         ENTRY;
 207
 208         fid_build_reg_res_name((struct lu_fid*)fid, &res_id);
 209         rc = ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace, &res_id,
 210                                    it, data);
 211         if (rc == LDLM_ITER_STOP)
 212                 RETURN(1);
 213         else if (rc == LDLM_ITER_CONTINUE)
 214                 RETURN(0);
 215         RETURN(rc);
 216 }
 217
 218 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 219 {
 220         /* Don't hold error requests for replay. */
 221         if (req->rq_replay) {
 222                 spin_lock(&req->rq_lock);
 223                 req->rq_replay = 0;
 224                 spin_unlock(&req->rq_lock);
 225         }
 226         if (rc && req->rq_transno != 0) {
 227                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 228                 LBUG();
 229         }
 230 }
 231
 232 /* Save a large LOV EA into the request buffer so that it is available
 233  * for replay.  We don't do this in the initial request because the
 234  * original request doesn't need this buffer (at most it sends just the
 235  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 236  * buffer and may also be difficult to allocate and save a very large
 237  * request buffer for each open. (bug 5707)
 238  *
 239  * OOM here may cause recovery failure if lmm is needed (only for the
 240  * original open if the MDS crashed just when this client also OOM'd)
 241  * but this is incredibly unlikely, and questionable whether the client
 242  * could do MDS recovery under OOM anyways... */
 243 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 244                                 struct mdt_body *body)
 245 {
 246         int     rc;
 247
 248         /* FIXME: remove this explicit offset. */
 249         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 250                                         body->mbo_eadatasize);
 251         if (rc) {
 252                 CERROR("Can't enlarge segment %d size to %d\n",
 253                        DLM_INTENT_REC_OFF + 4, body->mbo_eadatasize);
 254                 body->mbo_valid &= ~OBD_MD_FLEASIZE;
 255                 body->mbo_eadatasize = 0;
 256         }
 257 }
 258
 259 static struct ptlrpc_request *
 260 mdc_intent_open_pack(struct obd_export *exp, struct lookup_intent *it,
 261                      struct md_op_data *op_data)
 262 {
 263         struct ptlrpc_request   *req;
 264         struct obd_device       *obddev = class_exp2obd(exp);
 265         struct ldlm_intent      *lit;
 266         const void              *lmm = op_data->op_data;
 267         __u32                    lmmsize = op_data->op_data_size;
 268         struct list_head         cancels = LIST_HEAD_INIT(cancels);
 269         int                      count = 0;
 270         int                      mode;
 271         int                      rc;
 272         ENTRY;
 273
 274         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 275
 276         /* XXX: openlock is not cancelled for cross-refs. */
 277         /* If inode is known, cancel conflicting OPEN locks. */
 278         if (fid_is_sane(&op_data->op_fid2)) {
 279                 if (it->it_flags & MDS_OPEN_LEASE) { /* try to get lease */
 280                         if (it->it_flags & FMODE_WRITE)
 281                                 mode = LCK_EX;
 282                         else
 283                                 mode = LCK_PR;
 284                 } else {
 285                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 286                                 mode = LCK_CW;
 287 #ifdef FMODE_EXEC
 288                         else if (it->it_flags & FMODE_EXEC)
 289                                 mode = LCK_PR;
 290 #endif
 291                         else
 292                                 mode = LCK_CR;
 293                 }
 294                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 295                                                 &cancels, mode,
 296                                                 MDS_INODELOCK_OPEN);
 297         }
 298
 299         /* If CREATE, cancel parent's UPDATE lock. */
 300         if (it->it_op & IT_CREAT)
 301                 mode = LCK_EX;
 302         else
 303                 mode = LCK_CR;
 304         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 305                                          &cancels, mode,
 306                                          MDS_INODELOCK_UPDATE);
 307
 308         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 309                                    &RQF_LDLM_INTENT_OPEN);
 310         if (req == NULL) {
 311                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 312                 RETURN(ERR_PTR(-ENOMEM));
 313         }
 314
 315         /* parent capability */
 316         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 317         /* child capability, reserve the size according to parent capa, it will
 318          * be filled after we get the reply */
 319         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa1);
 320
 321         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 322                              op_data->op_namelen + 1);
 323         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 324                              max(lmmsize, obddev->u.cli.cl_default_mds_easize));
 325
 326         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 327         if (rc < 0) {
 328                 ptlrpc_request_free(req);
 329                 RETURN(ERR_PTR(rc));
 330         }
 331
 332         spin_lock(&req->rq_lock);
 333         req->rq_replay = req->rq_import->imp_replayable;
 334         spin_unlock(&req->rq_lock);
 335
 336         /* pack the intent */
 337         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 338         lit->opc = (__u64)it->it_op;
 339
 340         /* pack the intended request */
 341         mdc_open_pack(req, op_data, it->it_create_mode, 0, it->it_flags, lmm,
 342                       lmmsize);
 343
 344         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 345                              obddev->u.cli.cl_max_mds_easize);
 346
 347         /* for remote client, fetch remote perm for current user */
 348         if (client_is_remote(exp))
 349                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 350                                      sizeof(struct mdt_remote_perm));
 351         ptlrpc_request_set_replen(req);
 352         return req;
 353 }
 354
 355 static struct ptlrpc_request *
 356 mdc_intent_getxattr_pack(struct obd_export *exp,
 357                          struct lookup_intent *it,
 358                          struct md_op_data *op_data)
 359 {
 360         struct ptlrpc_request   *req;
 361         struct ldlm_intent      *lit;
 362         int                     rc, count = 0;
 363         __u32                   maxdata;
 364         struct list_head        cancels = LIST_HEAD_INIT(cancels);
 365
 366         ENTRY;
 367
 368         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 369                                         &RQF_LDLM_INTENT_GETXATTR);
 370         if (req == NULL)
 371                 RETURN(ERR_PTR(-ENOMEM));
 372
 373         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 374
 375         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 376         if (rc) {
 377                 ptlrpc_request_free(req);
 378                 RETURN(ERR_PTR(rc));
 379         }
 380
 381         /* pack the intent */
 382         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 383         lit->opc = IT_GETXATTR;
 384
 385         maxdata = class_exp2cliimp(exp)->imp_connect_data.ocd_max_easize;
 386
 387         /* pack the intended request */
 388         mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
 389                         op_data->op_valid, maxdata, -1, 0);
 390
 391         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 392                                 RCL_SERVER, maxdata);
 393
 394         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS,
 395                                 RCL_SERVER, maxdata);
 396
 397         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS_LENS,
 398                                 RCL_SERVER, maxdata);
 399
 400         ptlrpc_request_set_replen(req);
 401
 402         RETURN(req);
 403 }
 404
 405 static struct ptlrpc_request *mdc_intent_unlink_pack(struct obd_export *exp,
 406                                                      struct lookup_intent *it,
 407                                                      struct md_op_data *op_data)
 408 {
 409         struct ptlrpc_request *req;
 410         struct obd_device     *obddev = class_exp2obd(exp);
 411         struct ldlm_intent    *lit;
 412         int                    rc;
 413         ENTRY;
 414
 415         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 416                                    &RQF_LDLM_INTENT_UNLINK);
 417         if (req == NULL)
 418                 RETURN(ERR_PTR(-ENOMEM));
 419
 420         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 421         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 422                              op_data->op_namelen + 1);
 423
 424         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 425         if (rc) {
 426                 ptlrpc_request_free(req);
 427                 RETURN(ERR_PTR(rc));
 428         }
 429
 430         /* pack the intent */
 431         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 432         lit->opc = (__u64)it->it_op;
 433
 434         /* pack the intended request */
 435         mdc_unlink_pack(req, op_data);
 436
 437         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 438                              obddev->u.cli.cl_default_mds_easize);
 439         ptlrpc_request_set_replen(req);
 440         RETURN(req);
 441 }
 442
 443 static struct ptlrpc_request *mdc_intent_getattr_pack(struct obd_export *exp,
 444                                                       struct lookup_intent *it,
 445                                                       struct md_op_data *op_data)
 446 {
 447         struct ptlrpc_request   *req;
 448         struct obd_device       *obddev = class_exp2obd(exp);
 449         u64                      valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 450                                          OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 451                                          OBD_MD_FLMDSCAPA | OBD_MD_MEA |
 452                                          (client_is_remote(exp) ?
 453                                           OBD_MD_FLRMTPERM : OBD_MD_FLACL);
 454         struct ldlm_intent      *lit;
 455         int                      rc;
 456         __u32                    easize;
 457         ENTRY;
 458
 459         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 460                                    &RQF_LDLM_INTENT_GETATTR);
 461         if (req == NULL)
 462                 RETURN(ERR_PTR(-ENOMEM));
 463
 464         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 465         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 466                              op_data->op_namelen + 1);
 467
 468         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 469         if (rc) {
 470                 ptlrpc_request_free(req);
 471                 RETURN(ERR_PTR(rc));
 472         }
 473
 474         /* pack the intent */
 475         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 476         lit->opc = (__u64)it->it_op;
 477
 478         if (obddev->u.cli.cl_default_mds_easize > 0)
 479                 easize = obddev->u.cli.cl_default_mds_easize;
 480         else
 481                 easize = obddev->u.cli.cl_max_mds_easize;
 482
 483         /* pack the intended request */
 484         mdc_getattr_pack(req, valid, it->it_flags, op_data, easize);
 485
 486         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, easize);
 487         if (client_is_remote(exp))
 488                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 489                                      sizeof(struct mdt_remote_perm));
 490         ptlrpc_request_set_replen(req);
 491         RETURN(req);
 492 }
 493
 494 static struct ptlrpc_request *mdc_intent_layout_pack(struct obd_export *exp,
 495                                                      struct lookup_intent *it,
 496                                                      struct md_op_data *unused)
 497 {
 498         struct obd_device     *obd = class_exp2obd(exp);
 499         struct ptlrpc_request *req;
 500         struct ldlm_intent    *lit;
 501         struct layout_intent  *layout;
 502         int rc;
 503         ENTRY;
 504
 505         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 506                                 &RQF_LDLM_INTENT_LAYOUT);
 507         if (req == NULL)
 508                 RETURN(ERR_PTR(-ENOMEM));
 509
 510         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 511         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 512         if (rc) {
 513                 ptlrpc_request_free(req);
 514                 RETURN(ERR_PTR(rc));
 515         }
 516
 517         /* pack the intent */
 518         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 519         lit->opc = (__u64)it->it_op;
 520
 521         /* pack the layout intent request */
 522         layout = req_capsule_client_get(&req->rq_pill, &RMF_LAYOUT_INTENT);
 523         /* LAYOUT_INTENT_ACCESS is generic, specific operation will be
 524          * set for replication */
 525         layout->li_opc = LAYOUT_INTENT_ACCESS;
 526
 527         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 528                              obd->u.cli.cl_default_mds_easize);
 529         ptlrpc_request_set_replen(req);
 530         RETURN(req);
 531 }
 532
 533 static struct ptlrpc_request *
 534 mdc_enqueue_pack(struct obd_export *exp, int lvb_len)
 535 {
 536         struct ptlrpc_request *req;
 537         int rc;
 538         ENTRY;
 539
 540         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 541         if (req == NULL)
 542                 RETURN(ERR_PTR(-ENOMEM));
 543
 544         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 545         if (rc) {
 546                 ptlrpc_request_free(req);
 547                 RETURN(ERR_PTR(rc));
 548         }
 549
 550         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 551         ptlrpc_request_set_replen(req);
 552         RETURN(req);
 553 }
 554
 555 static int mdc_finish_enqueue(struct obd_export *exp,
 556                               struct ptlrpc_request *req,
 557                               struct ldlm_enqueue_info *einfo,
 558                               struct lookup_intent *it,
 559                               struct lustre_handle *lockh,
 560                               int rc)
 561 {
 562         struct req_capsule  *pill = &req->rq_pill;
 563         struct ldlm_request *lockreq;
 564         struct ldlm_reply   *lockrep;
 565         struct lustre_intent_data *intent = &it->d.lustre;
 566         struct ldlm_lock    *lock;
 567         void                *lvb_data = NULL;
 568         __u32                lvb_len = 0;
 569         ENTRY;
 570
 571         LASSERT(rc >= 0);
 572         /* Similarly, if we're going to replay this request, we don't want to
 573          * actually get a lock, just perform the intent. */
 574         if (req->rq_transno || req->rq_replay) {
 575                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 576                 lockreq->lock_flags |= ldlm_flags_to_wire(LDLM_FL_INTENT_ONLY);
 577         }
 578
 579         if (rc == ELDLM_LOCK_ABORTED) {
 580                 einfo->ei_mode = 0;
 581                 memset(lockh, 0, sizeof(*lockh));
 582                 rc = 0;
 583         } else { /* rc = 0 */
 584                 lock = ldlm_handle2lock(lockh);
 585                 LASSERT(lock != NULL);
 586
 587                 /* If the server gave us back a different lock mode, we should
 588                  * fix up our variables. */
 589                 if (lock->l_req_mode != einfo->ei_mode) {
 590                         ldlm_lock_addref(lockh, lock->l_req_mode);
 591                         ldlm_lock_decref(lockh, einfo->ei_mode);
 592                         einfo->ei_mode = lock->l_req_mode;
 593                 }
 594                 LDLM_LOCK_PUT(lock);
 595         }
 596
 597         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 598         LASSERT(lockrep != NULL); /* checked by ldlm_cli_enqueue() */
 599
 600         intent->it_disposition = (int)lockrep->lock_policy_res1;
 601         intent->it_status = (int)lockrep->lock_policy_res2;
 602         intent->it_lock_mode = einfo->ei_mode;
 603         intent->it_lock_handle = lockh->cookie;
 604         intent->it_data = req;
 605
 606         /* Technically speaking rq_transno must already be zero if
 607          * it_status is in error, so the check is a bit redundant */
 608         if ((!req->rq_transno || intent->it_status < 0) && req->rq_replay)
 609                 mdc_clear_replay_flag(req, intent->it_status);
 610
 611         /* If we're doing an IT_OPEN which did not result in an actual
 612          * successful open, then we need to remove the bit which saves
 613          * this request for unconditional replay.
 614          *
 615          * It's important that we do this first!  Otherwise we might exit the
 616          * function without doing so, and try to replay a failed create
 617          * (bug 3440) */
 618         if (it->it_op & IT_OPEN && req->rq_replay &&
 619             (!it_disposition(it, DISP_OPEN_OPEN) ||intent->it_status != 0))
 620                 mdc_clear_replay_flag(req, intent->it_status);
 621
 622         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 623                   it->it_op, intent->it_disposition, intent->it_status);
 624
 625         /* We know what to expect, so we do any byte flipping required here */
 626         if (it->it_op & (IT_OPEN | IT_UNLINK | IT_LOOKUP | IT_GETATTR)) {
 627                 struct mdt_body *body;
 628
 629                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 630                 if (body == NULL) {
 631                         CERROR ("Can't swab mdt_body\n");
 632                         RETURN (-EPROTO);
 633                 }
 634
 635                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 636                     !it_open_error(DISP_OPEN_OPEN, it)) {
 637                         /*
 638                          * If this is a successful OPEN request, we need to set
 639                          * replay handler and data early, so that if replay
 640                          * happens immediately after swabbing below, new reply
 641                          * is swabbed by that handler correctly.
 642                          */
 643                         mdc_set_open_replay_data(NULL, NULL, it);
 644                 }
 645
 646                 if (body->mbo_valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) {
 647                         void *eadata;
 648
 649                         mdc_update_max_ea_from_body(exp, body);
 650
 651                         /*
 652                          * The eadata is opaque; just check that it is there.
 653                          * Eventually, obd_unpackmd() will check the contents.
 654                          */
 655                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 656                                                         body->mbo_eadatasize);
 657                         if (eadata == NULL)
 658                                 RETURN(-EPROTO);
 659
 660                         /* save lvb data and length in case this is for layout
 661                          * lock */
 662                         lvb_data = eadata;
 663                         lvb_len = body->mbo_eadatasize;
 664
 665                         /*
 666                          * We save the reply LOV EA in case we have to replay a
 667                          * create for recovery.  If we didn't allocate a large
 668                          * enough request buffer above we need to reallocate it
 669                          * here to hold the actual LOV EA.
 670                          *
 671                          * To not save LOV EA if request is not going to replay
 672                          * (for example error one).
 673                          */
 674                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 675                                 void *lmm;
 676                                 if (req_capsule_get_size(pill, &RMF_EADATA,
 677                                                          RCL_CLIENT) <
 678                                     body->mbo_eadatasize)
 679                                         mdc_realloc_openmsg(req, body);
 680                                 else
 681                                         req_capsule_shrink(pill, &RMF_EADATA,
 682                                                            body->mbo_eadatasize,
 683                                                            RCL_CLIENT);
 684
 685                                 req_capsule_set_size(pill, &RMF_EADATA,
 686                                                      RCL_CLIENT,
 687                                                      body->mbo_eadatasize);
 688
 689                                 lmm = req_capsule_client_get(pill, &RMF_EADATA);
 690                                 if (lmm)
 691                                         memcpy(lmm, eadata,
 692                                                body->mbo_eadatasize);
 693                         }
 694                 }
 695
 696                 if (body->mbo_valid & OBD_MD_FLRMTPERM) {
 697                         struct mdt_remote_perm *perm;
 698
 699                         LASSERT(client_is_remote(exp));
 700                         perm = req_capsule_server_swab_get(pill, &RMF_ACL,
 701                                                 lustre_swab_mdt_remote_perm);
 702                         if (perm == NULL)
 703                                 RETURN(-EPROTO);
 704                 }
 705                 if (body->mbo_valid & OBD_MD_FLMDSCAPA) {
 706                         struct lustre_capa *capa, *p;
 707
 708                         capa = req_capsule_server_get(pill, &RMF_CAPA1);
 709                         if (capa == NULL)
 710                                 RETURN(-EPROTO);
 711
 712                         if (it->it_op & IT_OPEN) {
 713                                 /* client fid capa will be checked in replay */
 714                                 p = req_capsule_client_get(pill, &RMF_CAPA2);
 715                                 LASSERT(p);
 716                                 *p = *capa;
 717                         }
 718                 }
 719                 if (body->mbo_valid & OBD_MD_FLOSSCAPA) {
 720                         struct lustre_capa *capa;
 721
 722                         capa = req_capsule_server_get(pill, &RMF_CAPA2);
 723                         if (capa == NULL)
 724                                 RETURN(-EPROTO);
 725                 }
 726         } else if (it->it_op & IT_LAYOUT) {
 727                 /* maybe the lock was granted right away and layout
 728                  * is packed into RMF_DLM_LVB of req */
 729                 lvb_len = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
 730                 if (lvb_len > 0) {
 731                         lvb_data = req_capsule_server_sized_get(pill,
 732                                                         &RMF_DLM_LVB, lvb_len);
 733                         if (lvb_data == NULL)
 734                                 RETURN(-EPROTO);
 735                 }
 736         }
 737
 738         /* fill in stripe data for layout lock */
 739         lock = ldlm_handle2lock(lockh);
 740         if (lock != NULL && ldlm_has_layout(lock) && lvb_data != NULL) {
 741                 void *lmm;
 742
 743                 LDLM_DEBUG(lock, "layout lock returned by: %s, lvb_len: %d\n",
 744                         ldlm_it2str(it->it_op), lvb_len);
 745
 746                 OBD_ALLOC_LARGE(lmm, lvb_len);
 747                 if (lmm == NULL) {
 748                         LDLM_LOCK_PUT(lock);
 749                         RETURN(-ENOMEM);
 750                 }
 751                 memcpy(lmm, lvb_data, lvb_len);
 752
 753                 /* install lvb_data */
 754                 lock_res_and_lock(lock);
 755                 if (lock->l_lvb_data == NULL) {
 756                         lock->l_lvb_type = LVB_T_LAYOUT;
 757                         lock->l_lvb_data = lmm;
 758                         lock->l_lvb_len = lvb_len;
 759                         lmm = NULL;
 760                 }
 761                 unlock_res_and_lock(lock);
 762                 if (lmm != NULL)
 763                         OBD_FREE_LARGE(lmm, lvb_len);
 764         }
 765         if (lock != NULL)
 766                 LDLM_LOCK_PUT(lock);
 767
 768         RETURN(rc);
 769 }
 770
 771 /* We always reserve enough space in the reply packet for a stripe MD, because
 772  * we don't know in advance the file type. */
 773 int mdc_enqueue(struct obd_export *exp,
 774                 struct ldlm_enqueue_info *einfo,
 775                 const union ldlm_policy_data *policy,
 776                 struct lookup_intent *it, struct md_op_data *op_data,
 777                 struct lustre_handle *lockh, __u64 extra_lock_flags)
 778 {
 779         struct obd_device     *obddev = class_exp2obd(exp);
 780         struct ptlrpc_request *req = NULL;
 781         __u64                  flags, saved_flags = extra_lock_flags;
 782         int                    rc;
 783         struct ldlm_res_id res_id;
 784         static const ldlm_policy_data_t lookup_policy =
 785                             { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 786         static const ldlm_policy_data_t update_policy =
 787                             { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 788         static const ldlm_policy_data_t layout_policy =
 789                             { .l_inodebits = { MDS_INODELOCK_LAYOUT } };
 790         static const ldlm_policy_data_t getxattr_policy = {
 791                               .l_inodebits = { MDS_INODELOCK_XATTR } };
 792         int                    generation, resends = 0;
 793         struct ldlm_reply     *lockrep;
 794         enum lvb_type          lvb_type = 0;
 795         ENTRY;
 796
 797         LASSERTF(!it || einfo->ei_type == LDLM_IBITS, "lock type %d\n",
 798                  einfo->ei_type);
 799         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
 800
 801         if (it != NULL) {
 802                 LASSERT(policy == NULL);
 803
 804                 saved_flags |= LDLM_FL_HAS_INTENT;
 805                 if (it->it_op & (IT_OPEN | IT_UNLINK | IT_GETATTR | IT_READDIR))
 806                         policy = &update_policy;
 807                 else if (it->it_op & IT_LAYOUT)
 808                         policy = &layout_policy;
 809                 else if (it->it_op & (IT_GETXATTR | IT_SETXATTR))
 810                         policy = &getxattr_policy;
 811                 else
 812                         policy = &lookup_policy;
 813         }
 814
 815         generation = obddev->u.cli.cl_import->imp_generation;
 816 resend:
 817         flags = saved_flags;
 818         if (it == NULL) {
 819                 /* The only way right now is FLOCK. */
 820                 LASSERTF(einfo->ei_type == LDLM_FLOCK, "lock type %d\n",
 821                          einfo->ei_type);
 822                 res_id.name[3] = LDLM_FLOCK;
 823         } else if (it->it_op & IT_OPEN) {
 824                 req = mdc_intent_open_pack(exp, it, op_data);
 825         } else if (it->it_op & IT_UNLINK) {
 826                 req = mdc_intent_unlink_pack(exp, it, op_data);
 827         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 828                 req = mdc_intent_getattr_pack(exp, it, op_data);
 829         } else if (it->it_op & IT_READDIR) {
 830                 req = mdc_enqueue_pack(exp, 0);
 831         } else if (it->it_op & IT_LAYOUT) {
 832                 if (!imp_connect_lvb_type(class_exp2cliimp(exp)))
 833                         RETURN(-EOPNOTSUPP);
 834                 req = mdc_intent_layout_pack(exp, it, op_data);
 835                 lvb_type = LVB_T_LAYOUT;
 836         } else if (it->it_op & IT_GETXATTR) {
 837                 req = mdc_intent_getxattr_pack(exp, it, op_data);
 838         } else {
 839                 LBUG();
 840                 RETURN(-EINVAL);
 841         }
 842
 843         if (IS_ERR(req))
 844                 RETURN(PTR_ERR(req));
 845
 846         if (req != NULL && it && it->it_op & IT_CREAT)
 847                 /* ask ptlrpc not to resend on EINPROGRESS since we have our own
 848                  * retry logic */
 849                 req->rq_no_retry_einprogress = 1;
 850
 851         if (resends) {
 852                 req->rq_generation_set = 1;
 853                 req->rq_import_generation = generation;
 854                 req->rq_sent = cfs_time_current_sec() + resends;
 855         }
 856
 857         /* It is important to obtain rpc_lock first (if applicable), so that
 858          * threads that are serialised with rpc_lock are not polluting our
 859          * rpcs in flight counter. We do not do flock request limiting, though*/
 860         if (it) {
 861                 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 862                 rc = obd_get_request_slot(&obddev->u.cli);
 863                 if (rc != 0) {
 864                         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 865                         mdc_clear_replay_flag(req, 0);
 866                         ptlrpc_req_finished(req);
 867                         RETURN(rc);
 868                 }
 869         }
 870
 871         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, policy, &flags, NULL,
 872                               0, lvb_type, lockh, 0);
 873         if (!it) {
 874                 /* For flock requests we immediatelly return without further
 875                    delay and let caller deal with the rest, since rest of
 876                    this function metadata processing makes no sense for flock
 877                    requests anyway. But in case of problem during comms with
 878                    Server (ETIMEDOUT) or any signal/kill attempt (EINTR), we
 879                    can not rely on caller and this mainly for F_UNLCKs
 880                    (explicits or automatically generated by Kernel to clean
 881                    current FLocks upon exit) that can't be trashed */
 882                 if (((rc == -EINTR) || (rc == -ETIMEDOUT)) &&
 883                     (einfo->ei_type == LDLM_FLOCK) &&
 884                     (einfo->ei_mode == LCK_NL))
 885                         goto resend;
 886                 RETURN(rc);
 887         }
 888
 889         obd_put_request_slot(&obddev->u.cli);
 890         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 891
 892         if (rc < 0) {
 893                 CDEBUG(D_INFO, "%s: ldlm_cli_enqueue failed: rc = %d\n",
 894                        obddev->obd_name, rc);
 895
 896                 mdc_clear_replay_flag(req, rc);
 897                 ptlrpc_req_finished(req);
 898                 RETURN(rc);
 899         }
 900
 901         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 902         LASSERT(lockrep != NULL);
 903
 904         lockrep->lock_policy_res2 =
 905                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
 906
 907         /* Retry the create infinitely when we get -EINPROGRESS from
 908          * server. This is required by the new quota design. */
 909         if (it && it->it_op & IT_CREAT &&
 910             (int)lockrep->lock_policy_res2 == -EINPROGRESS) {
 911                 mdc_clear_replay_flag(req, rc);
 912                 ptlrpc_req_finished(req);
 913                 resends++;
 914
 915                 CDEBUG(D_HA, "%s: resend:%d op:%d "DFID"/"DFID"\n",
 916                        obddev->obd_name, resends, it->it_op,
 917                        PFID(&op_data->op_fid1), PFID(&op_data->op_fid2));
 918
 919                 if (generation == obddev->u.cli.cl_import->imp_generation) {
 920                         goto resend;
 921                 } else {
 922                         CDEBUG(D_HA, "resend cross eviction\n");
 923                         RETURN(-EIO);
 924                 }
 925         }
 926
 927         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
 928         if (rc < 0) {
 929                 if (lustre_handle_is_used(lockh)) {
 930                         ldlm_lock_decref(lockh, einfo->ei_mode);
 931                         memset(lockh, 0, sizeof(*lockh));
 932                 }
 933                 ptlrpc_req_finished(req);
 934
 935                 it->d.lustre.it_lock_handle = 0;
 936                 it->d.lustre.it_lock_mode = 0;
 937                 it->d.lustre.it_data = NULL;
 938         }
 939
 940         RETURN(rc);
 941 }
 942
 943 static int mdc_finish_intent_lock(struct obd_export *exp,
 944                                   struct ptlrpc_request *request,
 945                                   struct md_op_data *op_data,
 946                                   struct lookup_intent *it,
 947                                   struct lustre_handle *lockh)
 948 {
 949         struct lustre_handle old_lock;
 950         struct mdt_body *mdt_body;
 951         struct ldlm_lock *lock;
 952         int rc;
 953         ENTRY;
 954
 955         LASSERT(request != NULL);
 956         LASSERT(request != LP_POISON);
 957         LASSERT(request->rq_repmsg != LP_POISON);
 958
 959         if (it->it_op & IT_READDIR)
 960                 RETURN(0);
 961
 962         if (!it_disposition(it, DISP_IT_EXECD)) {
 963                 /* The server failed before it even started executing the
 964                  * intent, i.e. because it couldn't unpack the request. */
 965                 LASSERT(it->d.lustre.it_status != 0);
 966                 RETURN(it->d.lustre.it_status);
 967         }
 968         rc = it_open_error(DISP_IT_EXECD, it);
 969         if (rc)
 970                 RETURN(rc);
 971
 972         mdt_body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 973         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 974
 975         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 976         if (rc)
 977                 RETURN(rc);
 978
 979         /* keep requests around for the multiple phases of the call
 980          * this shows the DISP_XX must guarantee we make it into the call
 981          */
 982         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 983             it_disposition(it, DISP_OPEN_CREATE) &&
 984             !it_open_error(DISP_OPEN_CREATE, it)) {
 985                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 986                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 987         }
 988         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 989             it_disposition(it, DISP_OPEN_OPEN) &&
 990             !it_open_error(DISP_OPEN_OPEN, it)) {
 991                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 992                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 993                 /* BUG 11546 - eviction in the middle of open rpc processing */
 994                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 995         }
 996
 997         if (it->it_op & IT_CREAT) {
 998                 /* XXX this belongs in ll_create_it */
 999         } else if (it->it_op == IT_OPEN) {
1000                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
1001         } else {
1002                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_LAYOUT));
1003         }
1004
1005         /* If we already have a matching lock, then cancel the new
1006          * one.  We have to set the data here instead of in
1007          * mdc_enqueue, because we need to use the child's inode as
1008          * the l_ast_data to match, and that's not available until
1009          * intent_finish has performed the iget().) */
1010         lock = ldlm_handle2lock(lockh);
1011         if (lock) {
1012                 ldlm_policy_data_t policy = lock->l_policy_data;
1013                 LDLM_DEBUG(lock, "matching against this");
1014
1015                 LASSERTF(fid_res_name_eq(&mdt_body->mbo_fid1,
1016                                          &lock->l_resource->lr_name),
1017                          "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1018                          PLDLMRES(lock->l_resource), PFID(&mdt_body->mbo_fid1));
1019                 LDLM_LOCK_PUT(lock);
1020
1021                 memcpy(&old_lock, lockh, sizeof(*lockh));
1022                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
1023                                     LDLM_IBITS, &policy, LCK_NL, &old_lock, 0)) {
1024                         ldlm_lock_decref_and_cancel(lockh,
1025                                                     it->d.lustre.it_lock_mode);
1026                         memcpy(lockh, &old_lock, sizeof(old_lock));
1027                         it->d.lustre.it_lock_handle = lockh->cookie;
1028                 }
1029         }
1030         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
1031                 (int)op_data->op_namelen, op_data->op_name,
1032                 ldlm_it2str(it->it_op), it->d.lustre.it_status,
1033                 it->d.lustre.it_disposition, rc);
1034         RETURN(rc);
1035 }
1036
1037 int mdc_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
1038                         struct lu_fid *fid, __u64 *bits)
1039 {
1040         /* We could just return 1 immediately, but since we should only
1041          * be called in revalidate_it if we already have a lock, let's
1042          * verify that. */
1043         struct ldlm_res_id res_id;
1044         struct lustre_handle lockh;
1045         ldlm_policy_data_t policy;
1046         ldlm_mode_t mode;
1047         ENTRY;
1048
1049         if (it->d.lustre.it_lock_handle) {
1050                 lockh.cookie = it->d.lustre.it_lock_handle;
1051                 mode = ldlm_revalidate_lock_handle(&lockh, bits);
1052         } else {
1053                 fid_build_reg_res_name(fid, &res_id);
1054                 switch (it->it_op) {
1055                 case IT_GETATTR:
1056                         /* File attributes are held under multiple bits:
1057                          * nlink is under lookup lock, size and times are
1058                          * under UPDATE lock and recently we've also got
1059                          * a separate permissions lock for owner/group/acl that
1060                          * were protected by lookup lock before.
1061                          * Getattr must provide all of that information,
1062                          * so we need to ensure we have all of those locks.
1063                          * Unfortunately, if the bits are split across multiple
1064                          * locks, there's no easy way to match all of them here,
1065                          * so an extra RPC would be performed to fetch all
1066                          * of those bits at once for now. */
1067                         /* For new MDTs(> 2.4), UPDATE|PERM should be enough,
1068                          * but for old MDTs (< 2.4), permission is covered
1069                          * by LOOKUP lock, so it needs to match all bits here.*/
1070                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE |
1071                                                   MDS_INODELOCK_LOOKUP |
1072                                                   MDS_INODELOCK_PERM;
1073                         break;
1074                 case IT_READDIR:
1075                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1076                         break;
1077                 case IT_LAYOUT:
1078                         policy.l_inodebits.bits = MDS_INODELOCK_LAYOUT;
1079                         break;
1080                 default:
1081                         policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1082                         break;
1083                 }
1084
1085                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED, fid,
1086                                       LDLM_IBITS, &policy,
1087                                       LCK_CR | LCK_CW | LCK_PR | LCK_PW,
1088                                       &lockh);
1089         }
1090
1091         if (mode) {
1092                 it->d.lustre.it_lock_handle = lockh.cookie;
1093                 it->d.lustre.it_lock_mode = mode;
1094         } else {
1095                 it->d.lustre.it_lock_handle = 0;
1096                 it->d.lustre.it_lock_mode = 0;
1097         }
1098
1099         RETURN(!!mode);
1100 }
1101
1102 /*
1103  * This long block is all about fixing up the lock and request state
1104  * so that it is correct as of the moment _before_ the operation was
1105  * applied; that way, the VFS will think that everything is normal and
1106  * call Lustre's regular VFS methods.
1107  *
1108  * If we're performing a creation, that means that unless the creation
1109  * failed with EEXIST, we should fake up a negative dentry.
1110  *
1111  * For everything else, we want to lookup to succeed.
1112  *
1113  * One additional note: if CREATE or OPEN succeeded, we add an extra
1114  * reference to the request because we need to keep it around until
1115  * ll_create/ll_open gets called.
1116  *
1117  * The server will return to us, in it_disposition, an indication of
1118  * exactly what d.lustre.it_status refers to.
1119  *
1120  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
1121  * otherwise if DISP_OPEN_CREATE is set, then it status is the
1122  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
1123  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
1124  * was successful.
1125  *
1126  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
1127  * child lookup.
1128  */
1129 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
1130                     struct lookup_intent *it, struct ptlrpc_request **reqp,
1131                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
1132 {
1133         struct ldlm_enqueue_info einfo = {
1134                 .ei_type        = LDLM_IBITS,
1135                 .ei_mode        = it_to_lock_mode(it),
1136                 .ei_cb_bl       = cb_blocking,
1137                 .ei_cb_cp       = ldlm_completion_ast,
1138         };
1139         struct lustre_handle lockh;
1140         int rc = 0;
1141         ENTRY;
1142         LASSERT(it);
1143
1144         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
1145                 ", intent: %s flags %#"LPF64"o\n", (int)op_data->op_namelen,
1146                 op_data->op_name, PFID(&op_data->op_fid2),
1147                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
1148                 it->it_flags);
1149
1150         lockh.cookie = 0;
1151         if (fid_is_sane(&op_data->op_fid2) &&
1152             (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_READDIR))) {
1153                 /* We could just return 1 immediately, but since we should only
1154                  * be called in revalidate_it if we already have a lock, let's
1155                  * verify that. */
1156                 it->d.lustre.it_lock_handle = 0;
1157                 rc = mdc_revalidate_lock(exp, it, &op_data->op_fid2, NULL);
1158                 /* Only return failure if it was not GETATTR by cfid
1159                    (from inode_revalidate) */
1160                 if (rc || op_data->op_namelen != 0)
1161                         RETURN(rc);
1162         }
1163
1164         /* For case if upper layer did not alloc fid, do it now. */
1165         if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
1166                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1167                 if (rc < 0) {
1168                         CERROR("Can't alloc new fid, rc %d\n", rc);
1169                         RETURN(rc);
1170                 }
1171         }
1172
1173         rc = mdc_enqueue(exp, &einfo, NULL, it, op_data, &lockh,
1174                          extra_lock_flags);
1175         if (rc < 0)
1176                 RETURN(rc);
1177
1178         *reqp = it->d.lustre.it_data;
1179         rc = mdc_finish_intent_lock(exp, *reqp, op_data, it, &lockh);
1180         RETURN(rc);
1181 }
1182
1183 static int mdc_intent_getattr_async_interpret(const struct lu_env *env,
1184                                               struct ptlrpc_request *req,
1185                                               void *args, int rc)
1186 {
1187         struct mdc_getattr_args  *ga = args;
1188         struct obd_export        *exp = ga->ga_exp;
1189         struct md_enqueue_info   *minfo = ga->ga_minfo;
1190         struct ldlm_enqueue_info *einfo = ga->ga_einfo;
1191         struct lookup_intent     *it;
1192         struct lustre_handle     *lockh;
1193         struct obd_device        *obddev;
1194         struct ldlm_reply        *lockrep;
1195         __u64                     flags = LDLM_FL_HAS_INTENT;
1196         ENTRY;
1197
1198         it    = &minfo->mi_it;
1199         lockh = &minfo->mi_lockh;
1200
1201         obddev = class_exp2obd(exp);
1202
1203         obd_put_request_slot(&obddev->u.cli);
1204         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_GETATTR_ENQUEUE))
1205                 rc = -ETIMEDOUT;
1206
1207         rc = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, 1, einfo->ei_mode,
1208                                    &flags, NULL, 0, lockh, rc);
1209         if (rc < 0) {
1210                 CERROR("ldlm_cli_enqueue_fini: %d\n", rc);
1211                 mdc_clear_replay_flag(req, rc);
1212                 GOTO(out, rc);
1213         }
1214
1215         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1216         LASSERT(lockrep != NULL);
1217
1218         lockrep->lock_policy_res2 =
1219                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1220
1221         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
1222         if (rc)
1223                 GOTO(out, rc);
1224
1225         rc = mdc_finish_intent_lock(exp, req, &minfo->mi_data, it, lockh);
1226         EXIT;
1227
1228 out:
1229         OBD_FREE_PTR(einfo);
1230         minfo->mi_cb(req, minfo, rc);
1231         return 0;
1232 }
1233
1234 int mdc_intent_getattr_async(struct obd_export *exp,
1235                              struct md_enqueue_info *minfo,
1236                              struct ldlm_enqueue_info *einfo)
1237 {
1238         struct md_op_data       *op_data = &minfo->mi_data;
1239         struct lookup_intent    *it = &minfo->mi_it;
1240         struct ptlrpc_request   *req;
1241         struct mdc_getattr_args *ga;
1242         struct obd_device       *obddev = class_exp2obd(exp);
1243         struct ldlm_res_id       res_id;
1244         /*XXX: Both MDS_INODELOCK_LOOKUP and MDS_INODELOCK_UPDATE are needed
1245          *     for statahead currently. Consider CMD in future, such two bits
1246          *     maybe managed by different MDS, should be adjusted then. */
1247         ldlm_policy_data_t       policy = {
1248                                         .l_inodebits = { MDS_INODELOCK_LOOKUP |
1249                                                          MDS_INODELOCK_UPDATE }
1250                                  };
1251         int                      rc = 0;
1252         __u64                    flags = LDLM_FL_HAS_INTENT;
1253         ENTRY;
1254
1255         CDEBUG(D_DLMTRACE, "name: %.*s in inode "DFID", intent: %s flags %#"
1256                 LPF64"o\n",
1257                 (int)op_data->op_namelen, op_data->op_name,
1258                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op), it->it_flags);
1259
1260         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
1261         req = mdc_intent_getattr_pack(exp, it, op_data);
1262         if (IS_ERR(req))
1263                 RETURN(PTR_ERR(req));
1264
1265         rc = obd_get_request_slot(&obddev->u.cli);
1266         if (rc != 0) {
1267                 ptlrpc_req_finished(req);
1268                 RETURN(rc);
1269         }
1270
1271         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
1272                               0, LVB_T_NONE, &minfo->mi_lockh, 1);
1273         if (rc < 0) {
1274                 obd_put_request_slot(&obddev->u.cli);
1275                 ptlrpc_req_finished(req);
1276                 RETURN(rc);
1277         }
1278
1279         CLASSERT(sizeof(*ga) <= sizeof(req->rq_async_args));
1280         ga = ptlrpc_req_async_args(req);
1281         ga->ga_exp = exp;
1282         ga->ga_minfo = minfo;
1283         ga->ga_einfo = einfo;
1284
1285         req->rq_interpret_reply = mdc_intent_getattr_async_interpret;
1286         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1287
1288         RETURN(0);
1289 }