drivers/staging/lustre/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, 2012, 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 "../include/linux/lustre_intent.h"
  42 #include "../include/obd.h"
  43 #include "../include/obd_class.h"
  44 #include "../include/lustre_dlm.h"
  45 #include "../include/lustre_fid.h"      /* fid_res_name_eq() */
  46 #include "../include/lustre_mdc.h"
  47 #include "../include/lustre_net.h"
  48 #include "../include/lustre_req_layout.h"
  49 #include "mdc_internal.h"
  50
  51 struct mdc_getattr_args {
  52         struct obd_export          *ga_exp;
  53         struct md_enqueue_info      *ga_minfo;
  54         struct ldlm_enqueue_info    *ga_einfo;
  55 };
  56
  57 int it_disposition(struct lookup_intent *it, int flag)
  58 {
  59         return it->d.lustre.it_disposition & flag;
  60 }
  61 EXPORT_SYMBOL(it_disposition);
  62
  63 void it_set_disposition(struct lookup_intent *it, int flag)
  64 {
  65         it->d.lustre.it_disposition |= flag;
  66 }
  67 EXPORT_SYMBOL(it_set_disposition);
  68
  69 void it_clear_disposition(struct lookup_intent *it, int flag)
  70 {
  71         it->d.lustre.it_disposition &= ~flag;
  72 }
  73 EXPORT_SYMBOL(it_clear_disposition);
  74
  75 int it_open_error(int phase, struct lookup_intent *it)
  76 {
  77         if (it_disposition(it, DISP_OPEN_LEASE)) {
  78                 if (phase >= DISP_OPEN_LEASE)
  79                         return it->d.lustre.it_status;
  80                 else
  81                         return 0;
  82         }
  83         if (it_disposition(it, DISP_OPEN_OPEN)) {
  84                 if (phase >= DISP_OPEN_OPEN)
  85                         return it->d.lustre.it_status;
  86                 else
  87                         return 0;
  88         }
  89
  90         if (it_disposition(it, DISP_OPEN_CREATE)) {
  91                 if (phase >= DISP_OPEN_CREATE)
  92                         return it->d.lustre.it_status;
  93                 else
  94                         return 0;
  95         }
  96
  97         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  98                 if (phase >= DISP_LOOKUP_EXECD)
  99                         return it->d.lustre.it_status;
 100                 else
 101                         return 0;
 102         }
 103
 104         if (it_disposition(it, DISP_IT_EXECD)) {
 105                 if (phase >= DISP_IT_EXECD)
 106                         return it->d.lustre.it_status;
 107                 else
 108                         return 0;
 109         }
 110         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 111                it->d.lustre.it_status);
 112         LBUG();
 113         return 0;
 114 }
 115 EXPORT_SYMBOL(it_open_error);
 116
 117 /* this must be called on a lockh that is known to have a referenced lock */
 118 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
 119                       __u64 *bits)
 120 {
 121         struct ldlm_lock *lock;
 122         struct inode *new_inode = data;
 123
 124         if (bits)
 125                 *bits = 0;
 126
 127         if (!*lockh)
 128                 return 0;
 129
 130         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 131
 132         LASSERT(lock != NULL);
 133         lock_res_and_lock(lock);
 134         if (lock->l_resource->lr_lvb_inode &&
 135             lock->l_resource->lr_lvb_inode != data) {
 136                 struct inode *old_inode = lock->l_resource->lr_lvb_inode;
 137
 138                 LASSERTF(old_inode->i_state & I_FREEING,
 139                          "Found existing inode %p/%lu/%u state %lu in lock: setting data to %p/%lu/%u\n",
 140                          old_inode, old_inode->i_ino, old_inode->i_generation,
 141                          old_inode->i_state, new_inode, new_inode->i_ino,
 142                          new_inode->i_generation);
 143         }
 144         lock->l_resource->lr_lvb_inode = new_inode;
 145         if (bits)
 146                 *bits = lock->l_policy_data.l_inodebits.bits;
 147
 148         unlock_res_and_lock(lock);
 149         LDLM_LOCK_PUT(lock);
 150
 151         return 0;
 152 }
 153
 154 ldlm_mode_t mdc_lock_match(struct obd_export *exp, __u64 flags,
 155                            const struct lu_fid *fid, ldlm_type_t type,
 156                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
 157                            struct lustre_handle *lockh)
 158 {
 159         struct ldlm_res_id res_id;
 160         ldlm_mode_t rc;
 161
 162         fid_build_reg_res_name(fid, &res_id);
 163         /* LU-4405: Clear bits not supported by server */
 164         policy->l_inodebits.bits &= exp_connect_ibits(exp);
 165         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 166                              &res_id, type, policy, mode, lockh, 0);
 167         return rc;
 168 }
 169
 170 int mdc_cancel_unused(struct obd_export *exp,
 171                       const struct lu_fid *fid,
 172                       ldlm_policy_data_t *policy,
 173                       ldlm_mode_t mode,
 174                       ldlm_cancel_flags_t flags,
 175                       void *opaque)
 176 {
 177         struct ldlm_res_id res_id;
 178         struct obd_device *obd = class_exp2obd(exp);
 179         int rc;
 180
 181         fid_build_reg_res_name(fid, &res_id);
 182         rc = ldlm_cli_cancel_unused_resource(obd->obd_namespace, &res_id,
 183                                              policy, mode, flags, opaque);
 184         return rc;
 185 }
 186
 187 int mdc_null_inode(struct obd_export *exp,
 188                    const struct lu_fid *fid)
 189 {
 190         struct ldlm_res_id res_id;
 191         struct ldlm_resource *res;
 192         struct ldlm_namespace *ns = class_exp2obd(exp)->obd_namespace;
 193
 194         LASSERTF(ns != NULL, "no namespace passed\n");
 195
 196         fid_build_reg_res_name(fid, &res_id);
 197
 198         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 199         if (res == NULL)
 200                 return 0;
 201
 202         lock_res(res);
 203         res->lr_lvb_inode = NULL;
 204         unlock_res(res);
 205
 206         ldlm_resource_putref(res);
 207         return 0;
 208 }
 209
 210 /* find any ldlm lock of the inode in mdc
 211  * return 0    not find
 212  *      1    find one
 213  *      < 0    error */
 214 int mdc_find_cbdata(struct obd_export *exp,
 215                     const struct lu_fid *fid,
 216                     ldlm_iterator_t it, void *data)
 217 {
 218         struct ldlm_res_id res_id;
 219         int rc = 0;
 220
 221         fid_build_reg_res_name((struct lu_fid *)fid, &res_id);
 222         rc = ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace, &res_id,
 223                                    it, data);
 224         if (rc == LDLM_ITER_STOP)
 225                 return 1;
 226         else if (rc == LDLM_ITER_CONTINUE)
 227                 return 0;
 228         return rc;
 229 }
 230
 231 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 232 {
 233         /* Don't hold error requests for replay. */
 234         if (req->rq_replay) {
 235                 spin_lock(&req->rq_lock);
 236                 req->rq_replay = 0;
 237                 spin_unlock(&req->rq_lock);
 238         }
 239         if (rc && req->rq_transno != 0) {
 240                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 241                 LBUG();
 242         }
 243 }
 244
 245 /* Save a large LOV EA into the request buffer so that it is available
 246  * for replay.  We don't do this in the initial request because the
 247  * original request doesn't need this buffer (at most it sends just the
 248  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 249  * buffer and may also be difficult to allocate and save a very large
 250  * request buffer for each open. (bug 5707)
 251  *
 252  * OOM here may cause recovery failure if lmm is needed (only for the
 253  * original open if the MDS crashed just when this client also OOM'd)
 254  * but this is incredibly unlikely, and questionable whether the client
 255  * could do MDS recovery under OOM anyways... */
 256 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 257                                 struct mdt_body *body)
 258 {
 259         int     rc;
 260
 261         /* FIXME: remove this explicit offset. */
 262         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 263                                         body->eadatasize);
 264         if (rc) {
 265                 CERROR("Can't enlarge segment %d size to %d\n",
 266                        DLM_INTENT_REC_OFF + 4, body->eadatasize);
 267                 body->valid &= ~OBD_MD_FLEASIZE;
 268                 body->eadatasize = 0;
 269         }
 270 }
 271
 272 static struct ptlrpc_request *mdc_intent_open_pack(struct obd_export *exp,
 273                                                    struct lookup_intent *it,
 274                                                    struct md_op_data *op_data,
 275                                                    void *lmm, int lmmsize,
 276                                                    void *cb_data)
 277 {
 278         struct ptlrpc_request *req;
 279         struct obd_device     *obddev = class_exp2obd(exp);
 280         struct ldlm_intent    *lit;
 281         LIST_HEAD(cancels);
 282         int                 count = 0;
 283         int                 mode;
 284         int                 rc;
 285
 286         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 287
 288         /* XXX: openlock is not cancelled for cross-refs. */
 289         /* If inode is known, cancel conflicting OPEN locks. */
 290         if (fid_is_sane(&op_data->op_fid2)) {
 291                 if (it->it_flags & MDS_OPEN_LEASE) { /* try to get lease */
 292                         if (it->it_flags & FMODE_WRITE)
 293                                 mode = LCK_EX;
 294                         else
 295                                 mode = LCK_PR;
 296                 } else {
 297                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 298                                 mode = LCK_CW;
 299 #ifdef FMODE_EXEC
 300                         else if (it->it_flags & FMODE_EXEC)
 301                                 mode = LCK_PR;
 302 #endif
 303                         else
 304                                 mode = LCK_CR;
 305                 }
 306                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 307                                                 &cancels, mode,
 308                                                 MDS_INODELOCK_OPEN);
 309         }
 310
 311         /* If CREATE, cancel parent's UPDATE lock. */
 312         if (it->it_op & IT_CREAT)
 313                 mode = LCK_EX;
 314         else
 315                 mode = LCK_CR;
 316         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 317                                          &cancels, mode,
 318                                          MDS_INODELOCK_UPDATE);
 319
 320         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 321                                    &RQF_LDLM_INTENT_OPEN);
 322         if (req == NULL) {
 323                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 324                 return ERR_PTR(-ENOMEM);
 325         }
 326
 327         /* parent capability */
 328         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 329         /* child capability, reserve the size according to parent capa, it will
 330          * be filled after we get the reply */
 331         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa1);
 332
 333         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 334                              op_data->op_namelen + 1);
 335         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 336                              max(lmmsize, obddev->u.cli.cl_default_mds_easize));
 337
 338         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 339         if (rc < 0) {
 340                 ptlrpc_request_free(req);
 341                 return ERR_PTR(rc);
 342         }
 343
 344         spin_lock(&req->rq_lock);
 345         req->rq_replay = req->rq_import->imp_replayable;
 346         spin_unlock(&req->rq_lock);
 347
 348         /* pack the intent */
 349         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 350         lit->opc = (__u64)it->it_op;
 351
 352         /* pack the intended request */
 353         mdc_open_pack(req, op_data, it->it_create_mode, 0, it->it_flags, lmm,
 354                       lmmsize);
 355
 356         /* for remote client, fetch remote perm for current user */
 357         if (client_is_remote(exp))
 358                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 359                                      sizeof(struct mdt_remote_perm));
 360         ptlrpc_request_set_replen(req);
 361         return req;
 362 }
 363
 364 static struct ptlrpc_request *
 365 mdc_intent_getxattr_pack(struct obd_export *exp,
 366                          struct lookup_intent *it,
 367                          struct md_op_data *op_data)
 368 {
 369         struct ptlrpc_request   *req;
 370         struct ldlm_intent      *lit;
 371         int                     rc, count = 0, maxdata;
 372         LIST_HEAD(cancels);
 373
 374
 375
 376         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 377                                         &RQF_LDLM_INTENT_GETXATTR);
 378         if (req == NULL)
 379                 return ERR_PTR(-ENOMEM);
 380
 381         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 382
 383         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 384         if (rc) {
 385                 ptlrpc_request_free(req);
 386                 return ERR_PTR(rc);
 387         }
 388
 389         /* pack the intent */
 390         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 391         lit->opc = IT_GETXATTR;
 392
 393         maxdata = class_exp2cliimp(exp)->imp_connect_data.ocd_max_easize;
 394
 395         /* pack the intended request */
 396         mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
 397                         op_data->op_valid, maxdata, -1, 0);
 398
 399         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 400                                 RCL_SERVER, maxdata);
 401
 402         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS,
 403                                 RCL_SERVER, maxdata);
 404
 405         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS_LENS,
 406                                 RCL_SERVER, maxdata);
 407
 408         ptlrpc_request_set_replen(req);
 409
 410         return req;
 411 }
 412
 413 static struct ptlrpc_request *mdc_intent_unlink_pack(struct obd_export *exp,
 414                                                      struct lookup_intent *it,
 415                                                      struct md_op_data *op_data)
 416 {
 417         struct ptlrpc_request *req;
 418         struct obd_device     *obddev = class_exp2obd(exp);
 419         struct ldlm_intent    *lit;
 420         int                 rc;
 421
 422         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 423                                    &RQF_LDLM_INTENT_UNLINK);
 424         if (req == NULL)
 425                 return ERR_PTR(-ENOMEM);
 426
 427         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 428         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 429                              op_data->op_namelen + 1);
 430
 431         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 432         if (rc) {
 433                 ptlrpc_request_free(req);
 434                 return ERR_PTR(rc);
 435         }
 436
 437         /* pack the intent */
 438         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 439         lit->opc = (__u64)it->it_op;
 440
 441         /* pack the intended request */
 442         mdc_unlink_pack(req, op_data);
 443
 444         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 445                              obddev->u.cli.cl_default_mds_easize);
 446         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 447                              obddev->u.cli.cl_default_mds_cookiesize);
 448         ptlrpc_request_set_replen(req);
 449         return req;
 450 }
 451
 452 static struct ptlrpc_request *mdc_intent_getattr_pack(struct obd_export *exp,
 453                                                     struct lookup_intent *it,
 454                                                     struct md_op_data *op_data)
 455 {
 456         struct ptlrpc_request *req;
 457         struct obd_device     *obddev = class_exp2obd(exp);
 458         obd_valid             valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 459                                        OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 460                                        OBD_MD_FLMDSCAPA | OBD_MD_MEA |
 461                                        (client_is_remote(exp) ?
 462                                                OBD_MD_FLRMTPERM : OBD_MD_FLACL);
 463         struct ldlm_intent    *lit;
 464         int                 rc;
 465         int                 easize;
 466
 467         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 468                                    &RQF_LDLM_INTENT_GETATTR);
 469         if (req == NULL)
 470                 return ERR_PTR(-ENOMEM);
 471
 472         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 473         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 474                              op_data->op_namelen + 1);
 475
 476         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 477         if (rc) {
 478                 ptlrpc_request_free(req);
 479                 return ERR_PTR(rc);
 480         }
 481
 482         /* pack the intent */
 483         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 484         lit->opc = (__u64)it->it_op;
 485
 486         if (obddev->u.cli.cl_default_mds_easize > 0)
 487                 easize = obddev->u.cli.cl_default_mds_easize;
 488         else
 489                 easize = obddev->u.cli.cl_max_mds_easize;
 490
 491         /* pack the intended request */
 492         mdc_getattr_pack(req, valid, it->it_flags, op_data, easize);
 493
 494         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, easize);
 495         if (client_is_remote(exp))
 496                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 497                                      sizeof(struct mdt_remote_perm));
 498         ptlrpc_request_set_replen(req);
 499         return req;
 500 }
 501
 502 static struct ptlrpc_request *mdc_intent_layout_pack(struct obd_export *exp,
 503                                                      struct lookup_intent *it,
 504                                                      struct md_op_data *unused)
 505 {
 506         struct obd_device     *obd = class_exp2obd(exp);
 507         struct ptlrpc_request *req;
 508         struct ldlm_intent    *lit;
 509         struct layout_intent  *layout;
 510         int rc;
 511
 512         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 513                                 &RQF_LDLM_INTENT_LAYOUT);
 514         if (req == NULL)
 515                 return ERR_PTR(-ENOMEM);
 516
 517         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 518         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 519         if (rc) {
 520                 ptlrpc_request_free(req);
 521                 return ERR_PTR(rc);
 522         }
 523
 524         /* pack the intent */
 525         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 526         lit->opc = (__u64)it->it_op;
 527
 528         /* pack the layout intent request */
 529         layout = req_capsule_client_get(&req->rq_pill, &RMF_LAYOUT_INTENT);
 530         /* LAYOUT_INTENT_ACCESS is generic, specific operation will be
 531          * set for replication */
 532         layout->li_opc = LAYOUT_INTENT_ACCESS;
 533
 534         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 535                              obd->u.cli.cl_default_mds_easize);
 536         ptlrpc_request_set_replen(req);
 537         return req;
 538 }
 539
 540 static struct ptlrpc_request *
 541 mdc_enqueue_pack(struct obd_export *exp, int lvb_len)
 542 {
 543         struct ptlrpc_request *req;
 544         int rc;
 545
 546         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 547         if (req == NULL)
 548                 return ERR_PTR(-ENOMEM);
 549
 550         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 551         if (rc) {
 552                 ptlrpc_request_free(req);
 553                 return ERR_PTR(rc);
 554         }
 555
 556         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 557         ptlrpc_request_set_replen(req);
 558         return req;
 559 }
 560
 561 static int mdc_finish_enqueue(struct obd_export *exp,
 562                               struct ptlrpc_request *req,
 563                               struct ldlm_enqueue_info *einfo,
 564                               struct lookup_intent *it,
 565                               struct lustre_handle *lockh,
 566                               int rc)
 567 {
 568         struct req_capsule  *pill = &req->rq_pill;
 569         struct ldlm_request *lockreq;
 570         struct ldlm_reply   *lockrep;
 571         struct lustre_intent_data *intent = &it->d.lustre;
 572         struct ldlm_lock    *lock;
 573         void            *lvb_data = NULL;
 574         int               lvb_len = 0;
 575
 576         LASSERT(rc >= 0);
 577         /* Similarly, if we're going to replay this request, we don't want to
 578          * actually get a lock, just perform the intent. */
 579         if (req->rq_transno || req->rq_replay) {
 580                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 581                 lockreq->lock_flags |= ldlm_flags_to_wire(LDLM_FL_INTENT_ONLY);
 582         }
 583
 584         if (rc == ELDLM_LOCK_ABORTED) {
 585                 einfo->ei_mode = 0;
 586                 memset(lockh, 0, sizeof(*lockh));
 587                 rc = 0;
 588         } else { /* rc = 0 */
 589                 lock = ldlm_handle2lock(lockh);
 590                 LASSERT(lock != NULL);
 591
 592                 /* If the server gave us back a different lock mode, we should
 593                  * fix up our variables. */
 594                 if (lock->l_req_mode != einfo->ei_mode) {
 595                         ldlm_lock_addref(lockh, lock->l_req_mode);
 596                         ldlm_lock_decref(lockh, einfo->ei_mode);
 597                         einfo->ei_mode = lock->l_req_mode;
 598                 }
 599                 LDLM_LOCK_PUT(lock);
 600         }
 601
 602         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 603         LASSERT(lockrep != NULL); /* checked by ldlm_cli_enqueue() */
 604
 605         intent->it_disposition = (int)lockrep->lock_policy_res1;
 606         intent->it_status = (int)lockrep->lock_policy_res2;
 607         intent->it_lock_mode = einfo->ei_mode;
 608         intent->it_lock_handle = lockh->cookie;
 609         intent->it_data = req;
 610
 611         /* Technically speaking rq_transno must already be zero if
 612          * it_status is in error, so the check is a bit redundant */
 613         if ((!req->rq_transno || intent->it_status < 0) && req->rq_replay)
 614                 mdc_clear_replay_flag(req, intent->it_status);
 615
 616         /* If we're doing an IT_OPEN which did not result in an actual
 617          * successful open, then we need to remove the bit which saves
 618          * this request for unconditional replay.
 619          *
 620          * It's important that we do this first!  Otherwise we might exit the
 621          * function without doing so, and try to replay a failed create
 622          * (bug 3440) */
 623         if (it->it_op & IT_OPEN && req->rq_replay &&
 624             (!it_disposition(it, DISP_OPEN_OPEN) || intent->it_status != 0))
 625                 mdc_clear_replay_flag(req, intent->it_status);
 626
 627         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 628                   it->it_op, intent->it_disposition, intent->it_status);
 629
 630         /* We know what to expect, so we do any byte flipping required here */
 631         if (it->it_op & (IT_OPEN | IT_UNLINK | IT_LOOKUP | IT_GETATTR)) {
 632                 struct mdt_body *body;
 633
 634                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 635                 if (body == NULL) {
 636                         CERROR("Can't swab mdt_body\n");
 637                         return -EPROTO;
 638                 }
 639
 640                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 641                     !it_open_error(DISP_OPEN_OPEN, it)) {
 642                         /*
 643                          * If this is a successful OPEN request, we need to set
 644                          * replay handler and data early, so that if replay
 645                          * happens immediately after swabbing below, new reply
 646                          * is swabbed by that handler correctly.
 647                          */
 648                         mdc_set_open_replay_data(NULL, NULL, it);
 649                 }
 650
 651                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 652                         void *eadata;
 653
 654                         mdc_update_max_ea_from_body(exp, body);
 655
 656                         /*
 657                          * The eadata is opaque; just check that it is there.
 658                          * Eventually, obd_unpackmd() will check the contents.
 659                          */
 660                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 661                                                               body->eadatasize);
 662                         if (eadata == NULL)
 663                                 return -EPROTO;
 664
 665                         /* save lvb data and length in case this is for layout
 666                          * lock */
 667                         lvb_data = eadata;
 668                         lvb_len = body->eadatasize;
 669
 670                         /*
 671                          * We save the reply LOV EA in case we have to replay a
 672                          * create for recovery.  If we didn't allocate a large
 673                          * enough request buffer above we need to reallocate it
 674                          * here to hold the actual LOV EA.
 675                          *
 676                          * To not save LOV EA if request is not going to replay
 677                          * (for example error one).
 678                          */
 679                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 680                                 void *lmm;
 681
 682                                 if (req_capsule_get_size(pill, &RMF_EADATA,
 683                                                          RCL_CLIENT) <
 684                                     body->eadatasize)
 685                                         mdc_realloc_openmsg(req, body);
 686                                 else
 687                                         req_capsule_shrink(pill, &RMF_EADATA,
 688                                                            body->eadatasize,
 689                                                            RCL_CLIENT);
 690
 691                                 req_capsule_set_size(pill, &RMF_EADATA,
 692                                                      RCL_CLIENT,
 693                                                      body->eadatasize);
 694
 695                                 lmm = req_capsule_client_get(pill, &RMF_EADATA);
 696                                 if (lmm)
 697                                         memcpy(lmm, eadata, body->eadatasize);
 698                         }
 699                 }
 700
 701                 if (body->valid & OBD_MD_FLRMTPERM) {
 702                         struct mdt_remote_perm *perm;
 703
 704                         LASSERT(client_is_remote(exp));
 705                         perm = req_capsule_server_swab_get(pill, &RMF_ACL,
 706                                                 lustre_swab_mdt_remote_perm);
 707                         if (perm == NULL)
 708                                 return -EPROTO;
 709                 }
 710                 if (body->valid & OBD_MD_FLMDSCAPA) {
 711                         struct lustre_capa *capa, *p;
 712
 713                         capa = req_capsule_server_get(pill, &RMF_CAPA1);
 714                         if (capa == NULL)
 715                                 return -EPROTO;
 716
 717                         if (it->it_op & IT_OPEN) {
 718                                 /* client fid capa will be checked in replay */
 719                                 p = req_capsule_client_get(pill, &RMF_CAPA2);
 720                                 LASSERT(p);
 721                                 *p = *capa;
 722                         }
 723                 }
 724                 if (body->valid & OBD_MD_FLOSSCAPA) {
 725                         struct lustre_capa *capa;
 726
 727                         capa = req_capsule_server_get(pill, &RMF_CAPA2);
 728                         if (capa == NULL)
 729                                 return -EPROTO;
 730                 }
 731         } else if (it->it_op & IT_LAYOUT) {
 732                 /* maybe the lock was granted right away and layout
 733                  * is packed into RMF_DLM_LVB of req */
 734                 lvb_len = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
 735                 if (lvb_len > 0) {
 736                         lvb_data = req_capsule_server_sized_get(pill,
 737                                                         &RMF_DLM_LVB, lvb_len);
 738                         if (lvb_data == NULL)
 739                                 return -EPROTO;
 740                 }
 741         }
 742
 743         /* fill in stripe data for layout lock */
 744         lock = ldlm_handle2lock(lockh);
 745         if (lock != NULL && ldlm_has_layout(lock) && lvb_data != NULL) {
 746                 void *lmm;
 747
 748                 LDLM_DEBUG(lock, "layout lock returned by: %s, lvb_len: %d\n",
 749                         ldlm_it2str(it->it_op), lvb_len);
 750
 751                 OBD_ALLOC_LARGE(lmm, lvb_len);
 752                 if (lmm == NULL) {
 753                         LDLM_LOCK_PUT(lock);
 754                         return -ENOMEM;
 755                 }
 756                 memcpy(lmm, lvb_data, lvb_len);
 757
 758                 /* install lvb_data */
 759                 lock_res_and_lock(lock);
 760                 if (lock->l_lvb_data == NULL) {
 761                         lock->l_lvb_type = LVB_T_LAYOUT;
 762                         lock->l_lvb_data = lmm;
 763                         lock->l_lvb_len = lvb_len;
 764                         lmm = NULL;
 765                 }
 766                 unlock_res_and_lock(lock);
 767                 if (lmm != NULL)
 768                         OBD_FREE_LARGE(lmm, lvb_len);
 769         }
 770         if (lock != NULL)
 771                 LDLM_LOCK_PUT(lock);
 772
 773         return rc;
 774 }
 775
 776 /* We always reserve enough space in the reply packet for a stripe MD, because
 777  * we don't know in advance the file type. */
 778 int mdc_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
 779                 struct lookup_intent *it, struct md_op_data *op_data,
 780                 struct lustre_handle *lockh, void *lmm, int lmmsize,
 781                 struct ptlrpc_request **reqp, __u64 extra_lock_flags)
 782 {
 783         struct obd_device     *obddev = class_exp2obd(exp);
 784         struct ptlrpc_request *req = NULL;
 785         __u64             flags, saved_flags = extra_lock_flags;
 786         int                 rc;
 787         struct ldlm_res_id res_id;
 788         static const ldlm_policy_data_t lookup_policy =
 789                             { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 790         static const ldlm_policy_data_t update_policy =
 791                             { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 792         static const ldlm_policy_data_t layout_policy =
 793                             { .l_inodebits = { MDS_INODELOCK_LAYOUT } };
 794         static const ldlm_policy_data_t getxattr_policy = {
 795                               .l_inodebits = { MDS_INODELOCK_XATTR } };
 796         ldlm_policy_data_t const *policy = &lookup_policy;
 797         int                 generation, resends = 0;
 798         struct ldlm_reply     *lockrep;
 799         enum lvb_type          lvb_type = 0;
 800
 801         LASSERTF(!it || einfo->ei_type == LDLM_IBITS, "lock type %d\n",
 802                  einfo->ei_type);
 803
 804         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
 805
 806         if (it) {
 807                 saved_flags |= LDLM_FL_HAS_INTENT;
 808                 if (it->it_op & (IT_UNLINK | IT_GETATTR | IT_READDIR))
 809                         policy = &update_policy;
 810                 else if (it->it_op & IT_LAYOUT)
 811                         policy = &layout_policy;
 812                 else if (it->it_op & (IT_GETXATTR | IT_SETXATTR))
 813                         policy = &getxattr_policy;
 814         }
 815
 816         LASSERT(reqp == NULL);
 817
 818         generation = obddev->u.cli.cl_import->imp_generation;
 819 resend:
 820         flags = saved_flags;
 821         if (!it) {
 822                 /* The only way right now is FLOCK, in this case we hide flock
 823                    policy as lmm, but lmmsize is 0 */
 824                 LASSERT(lmm && lmmsize == 0);
 825                 LASSERTF(einfo->ei_type == LDLM_FLOCK, "lock type %d\n",
 826                          einfo->ei_type);
 827                 policy = (ldlm_policy_data_t *)lmm;
 828                 res_id.name[3] = LDLM_FLOCK;
 829         } else if (it->it_op & IT_OPEN) {
 830                 req = mdc_intent_open_pack(exp, it, op_data, lmm, lmmsize,
 831                                            einfo->ei_cbdata);
 832                 policy = &update_policy;
 833                 einfo->ei_cbdata = NULL;
 834                 lmm = NULL;
 835         } else if (it->it_op & IT_UNLINK) {
 836                 req = mdc_intent_unlink_pack(exp, it, op_data);
 837         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 838                 req = mdc_intent_getattr_pack(exp, it, op_data);
 839         } else if (it->it_op & IT_READDIR) {
 840                 req = mdc_enqueue_pack(exp, 0);
 841         } else if (it->it_op & IT_LAYOUT) {
 842                 if (!imp_connect_lvb_type(class_exp2cliimp(exp)))
 843                         return -EOPNOTSUPP;
 844                 req = mdc_intent_layout_pack(exp, it, op_data);
 845                 lvb_type = LVB_T_LAYOUT;
 846         } else if (it->it_op & IT_GETXATTR) {
 847                 req = mdc_intent_getxattr_pack(exp, it, op_data);
 848         } else {
 849                 LBUG();
 850                 return -EINVAL;
 851         }
 852
 853         if (IS_ERR(req))
 854                 return PTR_ERR(req);
 855
 856         if (req != NULL && it && it->it_op & IT_CREAT)
 857                 /* ask ptlrpc not to resend on EINPROGRESS since we have our own
 858                  * retry logic */
 859                 req->rq_no_retry_einprogress = 1;
 860
 861         if (resends) {
 862                 req->rq_generation_set = 1;
 863                 req->rq_import_generation = generation;
 864                 req->rq_sent = get_seconds() + resends;
 865         }
 866
 867         /* It is important to obtain rpc_lock first (if applicable), so that
 868          * threads that are serialised with rpc_lock are not polluting our
 869          * rpcs in flight counter. We do not do flock request limiting, though*/
 870         if (it) {
 871                 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 872                 rc = mdc_enter_request(&obddev->u.cli);
 873                 if (rc != 0) {
 874                         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 875                         mdc_clear_replay_flag(req, 0);
 876                         ptlrpc_req_finished(req);
 877                         return rc;
 878                 }
 879         }
 880
 881         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, policy, &flags, NULL,
 882                               0, lvb_type, lockh, 0);
 883         if (!it) {
 884                 /* For flock requests we immediately return without further
 885                    delay and let caller deal with the rest, since rest of
 886                    this function metadata processing makes no sense for flock
 887                    requests anyway. But in case of problem during comms with
 888                    Server (ETIMEDOUT) or any signal/kill attempt (EINTR), we
 889                    can not rely on caller and this mainly for F_UNLCKs
 890                    (explicits or automatically generated by Kernel to clean
 891                    current FLocks upon exit) that can't be trashed */
 892                 if ((rc == -EINTR) || (rc == -ETIMEDOUT))
 893                         goto resend;
 894                 return rc;
 895         }
 896
 897         mdc_exit_request(&obddev->u.cli);
 898         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 899
 900         if (rc < 0) {
 901                 CDEBUG_LIMIT((rc == -EACCES || rc == -EIDRM) ? D_INFO : D_ERROR,
 902                              "%s: ldlm_cli_enqueue failed: rc = %d\n",
 903                              obddev->obd_name, rc);
 904
 905                 mdc_clear_replay_flag(req, rc);
 906                 ptlrpc_req_finished(req);
 907                 return rc;
 908         }
 909
 910         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
 911         LASSERT(lockrep != NULL);
 912
 913         lockrep->lock_policy_res2 =
 914                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
 915
 916         /* Retry the create infinitely when we get -EINPROGRESS from
 917          * server. This is required by the new quota design. */
 918         if (it && it->it_op & IT_CREAT &&
 919             (int)lockrep->lock_policy_res2 == -EINPROGRESS) {
 920                 mdc_clear_replay_flag(req, rc);
 921                 ptlrpc_req_finished(req);
 922                 resends++;
 923
 924                 CDEBUG(D_HA, "%s: resend:%d op:%d "DFID"/"DFID"\n",
 925                        obddev->obd_name, resends, it->it_op,
 926                        PFID(&op_data->op_fid1), PFID(&op_data->op_fid2));
 927
 928                 if (generation == obddev->u.cli.cl_import->imp_generation) {
 929                         goto resend;
 930                 } else {
 931                         CDEBUG(D_HA, "resend cross eviction\n");
 932                         return -EIO;
 933                 }
 934         }
 935
 936         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
 937         if (rc < 0) {
 938                 if (lustre_handle_is_used(lockh)) {
 939                         ldlm_lock_decref(lockh, einfo->ei_mode);
 940                         memset(lockh, 0, sizeof(*lockh));
 941                 }
 942                 ptlrpc_req_finished(req);
 943         }
 944         return rc;
 945 }
 946
 947 static int mdc_finish_intent_lock(struct obd_export *exp,
 948                                   struct ptlrpc_request *request,
 949                                   struct md_op_data *op_data,
 950                                   struct lookup_intent *it,
 951                                   struct lustre_handle *lockh)
 952 {
 953         struct lustre_handle old_lock;
 954         struct mdt_body *mdt_body;
 955         struct ldlm_lock *lock;
 956         int rc;
 957
 958         LASSERT(request != NULL);
 959         LASSERT(request != LP_POISON);
 960         LASSERT(request->rq_repmsg != LP_POISON);
 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         /* If we were revalidating a fid/name pair, mark the intent in
 976          * case we fail and get called again from lookup */
 977         if (fid_is_sane(&op_data->op_fid2) &&
 978             it->it_create_mode & M_CHECK_STALE &&
 979             it->it_op != IT_GETATTR) {
 980
 981                 /* Also: did we find the same inode? */
 982                 /* sever can return one of two fids:
 983                  * op_fid2 - new allocated fid - if file is created.
 984                  * op_fid3 - existent fid - if file only open.
 985                  * op_fid3 is saved in lmv_intent_open */
 986                 if ((!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1)) &&
 987                     (!lu_fid_eq(&op_data->op_fid3, &mdt_body->fid1))) {
 988                         CDEBUG(D_DENTRY, "Found stale data "DFID"("DFID")/"DFID
 989                                "\n", PFID(&op_data->op_fid2),
 990                                PFID(&op_data->op_fid2), PFID(&mdt_body->fid1));
 991                         return -ESTALE;
 992                 }
 993         }
 994
 995         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 996         if (rc)
 997                 return rc;
 998
 999         /* keep requests around for the multiple phases of the call
1000          * this shows the DISP_XX must guarantee we make it into the call
1001          */
1002         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
1003             it_disposition(it, DISP_OPEN_CREATE) &&
1004             !it_open_error(DISP_OPEN_CREATE, it)) {
1005                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
1006                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
1007         }
1008         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
1009             it_disposition(it, DISP_OPEN_OPEN) &&
1010             !it_open_error(DISP_OPEN_OPEN, it)) {
1011                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
1012                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
1013                 /* BUG 11546 - eviction in the middle of open rpc processing */
1014                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
1015         }
1016
1017         if (it->it_op & IT_CREAT) {
1018                 /* XXX this belongs in ll_create_it */
1019         } else if (it->it_op == IT_OPEN) {
1020                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
1021         } else {
1022                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_LAYOUT));
1023         }
1024
1025         /* If we already have a matching lock, then cancel the new
1026          * one.  We have to set the data here instead of in
1027          * mdc_enqueue, because we need to use the child's inode as
1028          * the l_ast_data to match, and that's not available until
1029          * intent_finish has performed the iget().) */
1030         lock = ldlm_handle2lock(lockh);
1031         if (lock) {
1032                 ldlm_policy_data_t policy = lock->l_policy_data;
1033
1034                 LDLM_DEBUG(lock, "matching against this");
1035
1036                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
1037                                          &lock->l_resource->lr_name),
1038                          "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1039                          PLDLMRES(lock->l_resource), PFID(&mdt_body->fid1));
1040                 LDLM_LOCK_PUT(lock);
1041
1042                 memcpy(&old_lock, lockh, sizeof(*lockh));
1043                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
1044                                     LDLM_IBITS, &policy, LCK_NL,
1045                                     &old_lock, 0)) {
1046                         ldlm_lock_decref_and_cancel(lockh,
1047                                                     it->d.lustre.it_lock_mode);
1048                         memcpy(lockh, &old_lock, sizeof(old_lock));
1049                         it->d.lustre.it_lock_handle = lockh->cookie;
1050                 }
1051         }
1052         CDEBUG(D_DENTRY,
1053                "D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
1054                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
1055                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
1056         return rc;
1057 }
1058
1059 int mdc_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
1060                         struct lu_fid *fid, __u64 *bits)
1061 {
1062         /* We could just return 1 immediately, but since we should only
1063          * be called in revalidate_it if we already have a lock, let's
1064          * verify that. */
1065         struct ldlm_res_id res_id;
1066         struct lustre_handle lockh;
1067         ldlm_policy_data_t policy;
1068         ldlm_mode_t mode;
1069
1070         if (it->d.lustre.it_lock_handle) {
1071                 lockh.cookie = it->d.lustre.it_lock_handle;
1072                 mode = ldlm_revalidate_lock_handle(&lockh, bits);
1073         } else {
1074                 fid_build_reg_res_name(fid, &res_id);
1075                 switch (it->it_op) {
1076                 case IT_GETATTR:
1077                         /* File attributes are held under multiple bits:
1078                          * nlink is under lookup lock, size and times are
1079                          * under UPDATE lock and recently we've also got
1080                          * a separate permissions lock for owner/group/acl that
1081                          * were protected by lookup lock before.
1082                          * Getattr must provide all of that information,
1083                          * so we need to ensure we have all of those locks.
1084                          * Unfortunately, if the bits are split across multiple
1085                          * locks, there's no easy way to match all of them here,
1086                          * so an extra RPC would be performed to fetch all
1087                          * of those bits at once for now. */
1088                         /* For new MDTs(> 2.4), UPDATE|PERM should be enough,
1089                          * but for old MDTs (< 2.4), permission is covered
1090                          * by LOOKUP lock, so it needs to match all bits here.*/
1091                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE |
1092                                                   MDS_INODELOCK_LOOKUP |
1093                                                   MDS_INODELOCK_PERM;
1094                         break;
1095                 case IT_LAYOUT:
1096                         policy.l_inodebits.bits = MDS_INODELOCK_LAYOUT;
1097                         break;
1098                 default:
1099                         policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1100                         break;
1101                 }
1102
1103                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED, fid,
1104                                        LDLM_IBITS, &policy,
1105                                       LCK_CR | LCK_CW | LCK_PR | LCK_PW,
1106                                       &lockh);
1107         }
1108
1109         if (mode) {
1110                 it->d.lustre.it_lock_handle = lockh.cookie;
1111                 it->d.lustre.it_lock_mode = mode;
1112         } else {
1113                 it->d.lustre.it_lock_handle = 0;
1114                 it->d.lustre.it_lock_mode = 0;
1115         }
1116
1117         return !!mode;
1118 }
1119
1120 /*
1121  * This long block is all about fixing up the lock and request state
1122  * so that it is correct as of the moment _before_ the operation was
1123  * applied; that way, the VFS will think that everything is normal and
1124  * call Lustre's regular VFS methods.
1125  *
1126  * If we're performing a creation, that means that unless the creation
1127  * failed with EEXIST, we should fake up a negative dentry.
1128  *
1129  * For everything else, we want to lookup to succeed.
1130  *
1131  * One additional note: if CREATE or OPEN succeeded, we add an extra
1132  * reference to the request because we need to keep it around until
1133  * ll_create/ll_open gets called.
1134  *
1135  * The server will return to us, in it_disposition, an indication of
1136  * exactly what d.lustre.it_status refers to.
1137  *
1138  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
1139  * otherwise if DISP_OPEN_CREATE is set, then it status is the
1140  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
1141  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
1142  * was successful.
1143  *
1144  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
1145  * child lookup.
1146  */
1147 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
1148                     void *lmm, int lmmsize, struct lookup_intent *it,
1149                     int lookup_flags, struct ptlrpc_request **reqp,
1150                     ldlm_blocking_callback cb_blocking,
1151                     __u64 extra_lock_flags)
1152 {
1153         struct ldlm_enqueue_info einfo = {
1154                 .ei_type        = LDLM_IBITS,
1155                 .ei_mode        = it_to_lock_mode(it),
1156                 .ei_cb_bl       = cb_blocking,
1157                 .ei_cb_cp       = ldlm_completion_ast,
1158         };
1159         struct lustre_handle lockh;
1160         int rc = 0;
1161
1162         LASSERT(it);
1163
1164         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
1165                 ", intent: %s flags %#Lo\n", op_data->op_namelen,
1166                 op_data->op_name, PFID(&op_data->op_fid2),
1167                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
1168                 it->it_flags);
1169
1170         lockh.cookie = 0;
1171         if (fid_is_sane(&op_data->op_fid2) &&
1172             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
1173                 /* We could just return 1 immediately, but since we should only
1174                  * be called in revalidate_it if we already have a lock, let's
1175                  * verify that. */
1176                 it->d.lustre.it_lock_handle = 0;
1177                 rc = mdc_revalidate_lock(exp, it, &op_data->op_fid2, NULL);
1178                 /* Only return failure if it was not GETATTR by cfid
1179                    (from inode_revalidate) */
1180                 if (rc || op_data->op_namelen != 0)
1181                         return rc;
1182         }
1183
1184         /* For case if upper layer did not alloc fid, do it now. */
1185         if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
1186                 rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
1187                 if (rc < 0) {
1188                         CERROR("Can't alloc new fid, rc %d\n", rc);
1189                         return rc;
1190                 }
1191         }
1192         rc = mdc_enqueue(exp, &einfo, it, op_data, &lockh, lmm, lmmsize, NULL,
1193                          extra_lock_flags);
1194         if (rc < 0)
1195                 return rc;
1196
1197         *reqp = it->d.lustre.it_data;
1198         rc = mdc_finish_intent_lock(exp, *reqp, op_data, it, &lockh);
1199         return rc;
1200 }
1201
1202 static int mdc_intent_getattr_async_interpret(const struct lu_env *env,
1203                                               struct ptlrpc_request *req,
1204                                               void *args, int rc)
1205 {
1206         struct mdc_getattr_args  *ga = args;
1207         struct obd_export       *exp = ga->ga_exp;
1208         struct md_enqueue_info   *minfo = ga->ga_minfo;
1209         struct ldlm_enqueue_info *einfo = ga->ga_einfo;
1210         struct lookup_intent     *it;
1211         struct lustre_handle     *lockh;
1212         struct obd_device       *obddev;
1213         struct ldlm_reply        *lockrep;
1214         __u64                flags = LDLM_FL_HAS_INTENT;
1215
1216         it    = &minfo->mi_it;
1217         lockh = &minfo->mi_lockh;
1218
1219         obddev = class_exp2obd(exp);
1220
1221         mdc_exit_request(&obddev->u.cli);
1222         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_GETATTR_ENQUEUE))
1223                 rc = -ETIMEDOUT;
1224
1225         rc = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, 1, einfo->ei_mode,
1226                                    &flags, NULL, 0, lockh, rc);
1227         if (rc < 0) {
1228                 CERROR("ldlm_cli_enqueue_fini: %d\n", rc);
1229                 mdc_clear_replay_flag(req, rc);
1230                 GOTO(out, rc);
1231         }
1232
1233         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1234         LASSERT(lockrep != NULL);
1235
1236         lockrep->lock_policy_res2 =
1237                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1238
1239         rc = mdc_finish_enqueue(exp, req, einfo, it, lockh, rc);
1240         if (rc)
1241                 GOTO(out, rc);
1242
1243         rc = mdc_finish_intent_lock(exp, req, &minfo->mi_data, it, lockh);
1244
1245 out:
1246         OBD_FREE_PTR(einfo);
1247         minfo->mi_cb(req, minfo, rc);
1248         return 0;
1249 }
1250
1251 int mdc_intent_getattr_async(struct obd_export *exp,
1252                              struct md_enqueue_info *minfo,
1253                              struct ldlm_enqueue_info *einfo)
1254 {
1255         struct md_op_data       *op_data = &minfo->mi_data;
1256         struct lookup_intent    *it = &minfo->mi_it;
1257         struct ptlrpc_request   *req;
1258         struct mdc_getattr_args *ga;
1259         struct obd_device       *obddev = class_exp2obd(exp);
1260         struct ldlm_res_id       res_id;
1261         /*XXX: Both MDS_INODELOCK_LOOKUP and MDS_INODELOCK_UPDATE are needed
1262          *     for statahead currently. Consider CMD in future, such two bits
1263          *     maybe managed by different MDS, should be adjusted then. */
1264         ldlm_policy_data_t       policy = {
1265                                         .l_inodebits = { MDS_INODELOCK_LOOKUP |
1266                                                          MDS_INODELOCK_UPDATE }
1267                                  };
1268         int                   rc = 0;
1269         __u64               flags = LDLM_FL_HAS_INTENT;
1270
1271         CDEBUG(D_DLMTRACE,
1272                 "name: %.*s in inode "DFID", intent: %s flags %#Lo\n",
1273                 op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1274                 ldlm_it2str(it->it_op), it->it_flags);
1275
1276         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
1277         req = mdc_intent_getattr_pack(exp, it, op_data);
1278         if (IS_ERR(req))
1279                 return PTR_ERR(req);
1280
1281         rc = mdc_enter_request(&obddev->u.cli);
1282         if (rc != 0) {
1283                 ptlrpc_req_finished(req);
1284                 return rc;
1285         }
1286
1287         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
1288                               0, LVB_T_NONE, &minfo->mi_lockh, 1);
1289         if (rc < 0) {
1290                 mdc_exit_request(&obddev->u.cli);
1291                 ptlrpc_req_finished(req);
1292                 return rc;
1293         }
1294
1295         CLASSERT(sizeof(*ga) <= sizeof(req->rq_async_args));
1296         ga = ptlrpc_req_async_args(req);
1297         ga->ga_exp = exp;
1298         ga->ga_minfo = minfo;
1299         ga->ga_einfo = einfo;
1300
1301         req->rq_interpret_reply = mdc_intent_getattr_async_interpret;
1302         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1303
1304         return 0;
1305 }