lustre/mdc/mdc_locks.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_MDC
  29
  30 #ifdef __KERNEL__
  31 # include <linux/module.h>
  32 # include <linux/pagemap.h>
  33 # include <linux/miscdevice.h>
  34 # include <linux/init.h>
  35 #else
  36 # include <liblustre.h>
  37 #endif
  38
  39 #include <linux/lustre_acl.h>
  40 #include <obd_class.h>
  41 #include <lustre_dlm.h>
  42 /* fid_res_name_eq() */
  43 #include <lustre_fid.h>
  44 #include <lprocfs_status.h>
  45 #include "mdc_internal.h"
  46
  47 int it_disposition(struct lookup_intent *it, int flag)
  48 {
  49         return it->d.lustre.it_disposition & flag;
  50 }
  51 EXPORT_SYMBOL(it_disposition);
  52
  53 void it_set_disposition(struct lookup_intent *it, int flag)
  54 {
  55         it->d.lustre.it_disposition |= flag;
  56 }
  57 EXPORT_SYMBOL(it_set_disposition);
  58
  59 void it_clear_disposition(struct lookup_intent *it, int flag)
  60 {
  61         it->d.lustre.it_disposition &= ~flag;
  62 }
  63 EXPORT_SYMBOL(it_clear_disposition);
  64
  65 static int it_to_lock_mode(struct lookup_intent *it)
  66 {
  67         ENTRY;
  68
  69         /* CREAT needs to be tested before open (both could be set) */
  70         if (it->it_op & IT_CREAT)
  71                 return LCK_PW;
  72         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP))
  73                 return LCK_PR;
  74
  75         LBUG();
  76         RETURN(-EINVAL);
  77 }
  78
  79 int it_open_error(int phase, struct lookup_intent *it)
  80 {
  81         if (it_disposition(it, DISP_OPEN_OPEN)) {
  82                 if (phase >= DISP_OPEN_OPEN)
  83                         return it->d.lustre.it_status;
  84                 else
  85                         return 0;
  86         }
  87
  88         if (it_disposition(it, DISP_OPEN_CREATE)) {
  89                 if (phase >= DISP_OPEN_CREATE)
  90                         return it->d.lustre.it_status;
  91                 else
  92                         return 0;
  93         }
  94
  95         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  96                 if (phase >= DISP_LOOKUP_EXECD)
  97                         return it->d.lustre.it_status;
  98                 else
  99                         return 0;
 100         }
 101
 102         if (it_disposition(it, DISP_IT_EXECD)) {
 103                 if (phase >= DISP_IT_EXECD)
 104                         return it->d.lustre.it_status;
 105                 else
 106                         return 0;
 107         }
 108         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 109                it->d.lustre.it_status);
 110         LBUG();
 111         return 0;
 112 }
 113 EXPORT_SYMBOL(it_open_error);
 114
 115 /* this must be called on a lockh that is known to have a referenced lock */
 116 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
 117 {
 118         struct ldlm_lock *lock;
 119         ENTRY;
 120
 121         if (!*lockh) {
 122                 EXIT;
 123                 RETURN(0);
 124         }
 125
 126         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 127
 128         LASSERT(lock != NULL);
 129         lock_res_and_lock(lock);
 130 #ifdef __KERNEL__
 131         if (lock->l_ast_data && lock->l_ast_data != data) {
 132                 struct inode *new_inode = data;
 133                 struct inode *old_inode = lock->l_ast_data;
 134                 LASSERTF(old_inode->i_state & I_FREEING,
 135                          "Found existing inode %p/%lu/%u state %lu in lock: "
 136                          "setting data to %p/%lu/%u\n", old_inode,
 137                          old_inode->i_ino, old_inode->i_generation,
 138                          old_inode->i_state,
 139                          new_inode, new_inode->i_ino, new_inode->i_generation);
 140         }
 141 #endif
 142         lock->l_ast_data = data;
 143         unlock_res_and_lock(lock);
 144         LDLM_LOCK_PUT(lock);
 145
 146         RETURN(0);
 147 }
 148
 149 int mdc_lock_match(struct obd_export *exp, int flags,
 150                    const struct lu_fid *fid, ldlm_type_t type,
 151                    ldlm_policy_data_t *policy, ldlm_mode_t mode,
 152                    struct lustre_handle *lockh)
 153 {
 154         struct ldlm_res_id res_id =
 155                 { .name = {fid_seq(fid),
 156                            fid_oid(fid),
 157                            fid_ver(fid)} };
 158         struct obd_device *obd = class_exp2obd(exp);
 159         int rc;
 160         ENTRY;
 161
 162         rc = ldlm_lock_match(obd->obd_namespace, flags,
 163                              &res_id, type, policy, mode, lockh);
 164
 165         RETURN(rc);
 166 }
 167
 168 int mdc_cancel_unused(struct obd_export *exp,
 169                       const struct lu_fid *fid,
 170                       ldlm_policy_data_t *policy,
 171                       ldlm_mode_t mode, int flags, void *opaque)
 172 {
 173         struct ldlm_res_id res_id =
 174                 { .name = {fid_seq(fid),
 175                            fid_oid(fid),
 176                            fid_ver(fid)} };
 177         struct obd_device *obd = class_exp2obd(exp);
 178         int rc;
 179
 180         ENTRY;
 181
 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_change_cbdata(struct obd_export *exp,
 188                       const struct lu_fid *fid,
 189                       ldlm_iterator_t it, void *data)
 190 {
 191         struct ldlm_res_id res_id = { .name = {0} };
 192         ENTRY;
 193
 194         res_id.name[0] = fid_seq(fid);
 195         res_id.name[1] = fid_oid(fid);
 196         res_id.name[2] = fid_ver(fid);
 197
 198         ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace,
 199                               &res_id, it, data);
 200
 201         EXIT;
 202         return 0;
 203 }
 204
 205 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 206 {
 207         /* Don't hold error requests for replay. */
 208         if (req->rq_replay) {
 209                 spin_lock(&req->rq_lock);
 210                 req->rq_replay = 0;
 211                 spin_unlock(&req->rq_lock);
 212         }
 213         if (rc && req->rq_transno != 0) {
 214                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 215                 LBUG();
 216         }
 217 }
 218
 219 /* Save a large LOV EA into the request buffer so that it is available
 220  * for replay.  We don't do this in the initial request because the
 221  * original request doesn't need this buffer (at most it sends just the
 222  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 223  * buffer and may also be difficult to allocate and save a very large
 224  * request buffer for each open. (bug 5707)
 225  *
 226  * OOM here may cause recovery failure if lmm is needed (only for the
 227  * original open if the MDS crashed just when this client also OOM'd)
 228  * but this is incredibly unlikely, and questionable whether the client
 229  * could do MDS recovery under OOM anyways... */
 230 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 231                                 struct mdt_body *body, int size[9])
 232 {
 233         int     rc;
 234         ENTRY;
 235
 236         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 237                                         body->eadatasize);
 238         if (rc) {
 239                 CERROR("Can't enlarge segment %d size to %d\n",
 240                        DLM_INTENT_REC_OFF + 4, body->eadatasize);
 241                 body->valid &= ~OBD_MD_FLEASIZE;
 242                 body->eadatasize = 0;
 243         }
 244         EXIT;
 245 }
 246
 247 /* We always reserve enough space in the reply packet for a stripe MD, because
 248  * we don't know in advance the file type. */
 249 int mdc_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
 250                 struct lookup_intent *it, struct md_op_data *op_data,
 251                 struct lustre_handle *lockh, void *lmm, int lmmsize,
 252                 int extra_lock_flags)
 253 {
 254         struct ptlrpc_request *req;
 255         struct obd_device *obddev = class_exp2obd(exp);
 256         struct ldlm_res_id res_id =
 257                 { .name = {fid_seq(&op_data->op_fid1),
 258                            fid_oid(&op_data->op_fid1),
 259                            fid_ver(&op_data->op_fid1)} };
 260         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 261         struct ldlm_request *lockreq;
 262         struct ldlm_intent *lit;
 263         struct ldlm_reply *lockrep;
 264         int size[9] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 265                         [DLM_LOCKREQ_OFF]     = sizeof(*lockreq),
 266                         [DLM_INTENT_IT_OFF]   = sizeof(*lit),
 267                         0, 0, 0, 0, 0, 0 };
 268         int repsize[7] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 269                            [DLM_LOCKREPLY_OFF]   = sizeof(*lockrep),
 270                            [DLM_REPLY_REC_OFF]   = sizeof(struct mdt_body),
 271                            [DLM_REPLY_REC_OFF+1] = obddev->u.cli.
 272                                                    cl_max_mds_easize,
 273                            0, 0, 0 };
 274         int flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 275         int repbufcnt = 4, rc;
 276         ENTRY;
 277
 278         LASSERTF(einfo->ei_type == LDLM_IBITS,"lock type %d\n", einfo->ei_type);
 279
 280         if (it->it_op & IT_OPEN) {
 281                 int do_join = !!(it->it_flags & O_JOIN_FILE);
 282                 CFS_LIST_HEAD(cancels);
 283                 int count = 0;
 284                 int mode;
 285
 286                 it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 287
 288                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_create);
 289                 /* parent capability */
 290                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 291                                                sizeof(struct lustre_capa) : 0;
 292                 /* child capability, used for replay only */
 293                 size[DLM_INTENT_REC_OFF + 2] = sizeof(struct lustre_capa);
 294                 size[DLM_INTENT_REC_OFF + 3] = op_data->op_namelen + 1;
 295                 /* As an optimization, we allocate an RPC request buffer for
 296                  * at least a default-sized LOV EA even if we aren't sending
 297                  * one.
 298                  */
 299                 size[DLM_INTENT_REC_OFF + 4] = max(lmmsize,
 300                                            obddev->u.cli.cl_default_mds_easize);
 301
 302                 /* XXX: openlock is not cancelled for cross-refs. */
 303                 /* If inode is known, cancel conflicting OPEN locks. */
 304                 if (fid_is_sane(&op_data->op_fid2)) {
 305                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 306                                 mode = LCK_CW;
 307 #ifdef FMODE_EXEC
 308                         else if (it->it_flags & FMODE_EXEC)
 309                                 mode = LCK_PR;
 310 #endif
 311                         else
 312                                 mode = LCK_CR;
 313                         count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 314                                                         &cancels, mode,
 315                                                         MDS_INODELOCK_OPEN);
 316                 }
 317
 318                 /* If CREATE or JOIN_FILE, cancel parent's UPDATE lock. */
 319                 if (it->it_op & IT_CREAT || it->it_flags & O_JOIN_FILE)
 320                         mode = LCK_EX;
 321                 else
 322                         mode = LCK_CR;
 323                 count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 324                                                  &cancels, mode,
 325                                                  MDS_INODELOCK_UPDATE);
 326
 327                 if (do_join)
 328                         size[DLM_INTENT_REC_OFF + 5] =
 329                                                 sizeof(struct mdt_rec_join);
 330
 331                 req = ldlm_prep_enqueue_req(exp, 8 + do_join, size, &cancels,
 332                                             count);
 333                 if (!req)
 334                         RETURN(-ENOMEM);
 335
 336                 if (do_join) {
 337                         /* join is like an unlink of the tail */
 338                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 339                         mdc_join_pack(req, DLM_INTENT_REC_OFF + 5, op_data,
 340                                       (*(__u64 *)op_data->op_data));
 341                 }
 342
 343                 spin_lock(&req->rq_lock);
 344                 req->rq_replay = 1;
 345                 spin_unlock(&req->rq_lock);
 346
 347                 /* pack the intent */
 348                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 349                                      sizeof(*lit));
 350                 lit->opc = (__u64)it->it_op;
 351
 352                 /* pack the intended request */
 353                 mdc_open_pack(req, DLM_INTENT_REC_OFF, op_data,
 354                               it->it_create_mode, 0, it->it_flags,
 355                               lmm, lmmsize);
 356
 357                 /* for remote client, fetch remote perm for current user */
 358                 repsize[repbufcnt++] = client_is_remote(exp) ?
 359                                        sizeof(struct mdt_remote_perm) :
 360                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 361                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 362                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 363         } else if (it->it_op & IT_UNLINK) {
 364                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_unlink);
 365                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 366                                                sizeof(struct lustre_capa) : 0;
 367                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 368                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 369                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 370                 if (!req)
 371                         RETURN(-ENOMEM);
 372
 373                 /* pack the intent */
 374                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 375                                      sizeof(*lit));
 376                 lit->opc = (__u64)it->it_op;
 377
 378                 /* pack the intended request */
 379                 mdc_unlink_pack(req, DLM_INTENT_REC_OFF, op_data);
 380
 381                 repsize[repbufcnt++] = obddev->u.cli.cl_max_mds_cookiesize;
 382         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 383                 obd_valid valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 384                         OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 385                         OBD_MD_FLMDSCAPA | OBD_MD_MEA;
 386                 valid |= client_is_remote(exp) ? OBD_MD_FLRMTPERM :
 387                                                  OBD_MD_FLACL;
 388                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_body);
 389                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 390                                                sizeof(struct lustre_capa) : 0;
 391                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 392
 393                 if (it->it_op & IT_GETATTR)
 394                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 395
 396                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 397                 if (!req)
 398                         RETURN(-ENOMEM);
 399
 400                 /* pack the intent */
 401                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 402                                      sizeof(*lit));
 403                 lit->opc = (__u64)it->it_op;
 404
 405                 /* pack the intended request */
 406                 mdc_getattr_pack(req, DLM_INTENT_REC_OFF, valid,
 407                                  it->it_flags, op_data);
 408
 409                 repsize[repbufcnt++] = client_is_remote(exp) ?
 410                                        sizeof(struct mdt_remote_perm) :
 411                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 412                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 413         } else if (it->it_op == IT_READDIR) {
 414                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 415                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 416                 if (!req)
 417                         RETURN(-ENOMEM);
 418
 419                 repbufcnt = 2;
 420         } else {
 421                 LBUG();
 422                 RETURN(-EINVAL);
 423         }
 424
 425         /* get ready for the reply */
 426         ptlrpc_req_set_repsize(req, repbufcnt, repsize);
 427
 428          /* It is important to obtain rpc_lock first (if applicable), so that
 429           * threads that are serialised with rpc_lock are not polluting our
 430           * rpcs in flight counter */
 431         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 432         mdc_enter_request(&obddev->u.cli);
 433         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
 434                               0, NULL, lockh, 0);
 435         mdc_exit_request(&obddev->u.cli);
 436         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 437
 438         /* Similarly, if we're going to replay this request, we don't want to
 439          * actually get a lock, just perform the intent. */
 440         if (req->rq_transno || req->rq_replay) {
 441                 lockreq = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 442                                          sizeof(*lockreq));
 443                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 444         }
 445
 446         if (rc == ELDLM_LOCK_ABORTED) {
 447                 einfo->ei_mode = 0;
 448                 memset(lockh, 0, sizeof(*lockh));
 449                 rc = 0;
 450         } else if (rc != 0) {
 451                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 452                 LASSERTF(rc < 0, "rc %d\n", rc);
 453                 mdc_clear_replay_flag(req, rc);
 454                 ptlrpc_req_finished(req);
 455                 RETURN(rc);
 456         } else { /* rc = 0 */
 457                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 458                 LASSERT(lock);
 459
 460                 /* If the server gave us back a different lock mode, we should
 461                  * fix up our variables. */
 462                 if (lock->l_req_mode != einfo->ei_mode) {
 463                         ldlm_lock_addref(lockh, lock->l_req_mode);
 464                         ldlm_lock_decref(lockh, einfo->ei_mode);
 465                         einfo->ei_mode = lock->l_req_mode;
 466                 }
 467                 LDLM_LOCK_PUT(lock);
 468         }
 469
 470         lockrep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 471                                  sizeof(*lockrep));
 472         LASSERT(lockrep != NULL);                 /* checked by ldlm_cli_enqueue() */
 473         /* swabbed by ldlm_cli_enqueue() */
 474         LASSERT(lustre_rep_swabbed(req, DLM_LOCKREPLY_OFF));
 475
 476         it->d.lustre.it_disposition = (int)lockrep->lock_policy_res1;
 477         it->d.lustre.it_status = (int)lockrep->lock_policy_res2;
 478         it->d.lustre.it_lock_mode = einfo->ei_mode;
 479         it->d.lustre.it_data = req;
 480
 481         if (it->d.lustre.it_status < 0 && req->rq_replay)
 482                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 483
 484         /* If we're doing an IT_OPEN which did not result in an actual
 485          * successful open, then we need to remove the bit which saves
 486          * this request for unconditional replay.
 487          *
 488          * It's important that we do this first!  Otherwise we might exit the
 489          * function without doing so, and try to replay a failed create
 490          * (bug 3440) */
 491         if (it->it_op & IT_OPEN && req->rq_replay &&
 492             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 493                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 494
 495         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 496                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 497
 498         /* We know what to expect, so we do any byte flipping required here */
 499         LASSERT(repbufcnt == 7 || repbufcnt == 6 || repbufcnt == 2);
 500         if (repbufcnt >= 6) {
 501                 int reply_off = DLM_REPLY_REC_OFF;
 502                 struct mdt_body *body;
 503
 504                 body = lustre_swab_repbuf(req, reply_off++, sizeof(*body),
 505                                          lustre_swab_mdt_body);
 506                 if (body == NULL) {
 507                         CERROR ("Can't swab mdt_body\n");
 508                         RETURN (-EPROTO);
 509                 }
 510
 511                 if (req->rq_replay && it_disposition(it, DISP_OPEN_OPEN) &&
 512                     !it_open_error(DISP_OPEN_OPEN, it)) {
 513                         /*
 514                          * If this is a successful OPEN request, we need to set
 515                          * replay handler and data early, so that if replay
 516                          * happens immediately after swabbing below, new reply
 517                          * is swabbed by that handler correctly.
 518                          */
 519                         mdc_set_open_replay_data(NULL, NULL, req);
 520                 }
 521
 522                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 523                         void *eadata;
 524
 525                         /*
 526                          * The eadata is opaque; just check that it is there.
 527                          * Eventually, obd_unpackmd() will check the contents.
 528                          */
 529                         eadata = lustre_swab_repbuf(req, reply_off++,
 530                                                     body->eadatasize, NULL);
 531                         if (eadata == NULL) {
 532                                 CERROR("Missing/short eadata\n");
 533                                 RETURN(-EPROTO);
 534                         }
 535                         if (body->valid & OBD_MD_FLMODEASIZE) {
 536                                 if (obddev->u.cli.cl_max_mds_easize <
 537                                     body->max_mdsize) {
 538                                         obddev->u.cli.cl_max_mds_easize =
 539                                                 body->max_mdsize;
 540                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 541                                                body->max_mdsize);
 542                                 }
 543                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 544                                     body->max_cookiesize) {
 545                                         obddev->u.cli.cl_max_mds_cookiesize =
 546                                                 body->max_cookiesize;
 547                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 548                                                body->max_cookiesize);
 549                                 }
 550                         }
 551
 552                         /*
 553                          * We save the reply LOV EA in case we have to replay a
 554                          * create for recovery.  If we didn't allocate a large
 555                          * enough request buffer above we need to reallocate it
 556                          * here to hold the actual LOV EA.
 557                          *
 558                          * To not save LOV EA if request is not going to replay
 559                          * (for example error one).
 560                          */
 561                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 562                                 if (lustre_msg_buflen(req->rq_reqmsg,
 563                                                       DLM_INTENT_REC_OFF + 4) <
 564                                     body->eadatasize)
 565                                         mdc_realloc_openmsg(req, body, size);
 566
 567                                 lmm = lustre_msg_buf(req->rq_reqmsg,
 568                                                      DLM_INTENT_REC_OFF + 4,
 569                                                      body->eadatasize);
 570                                 if (lmm)
 571                                         memcpy(lmm, eadata, body->eadatasize);
 572                         }
 573                 }
 574                 if (body->valid & OBD_MD_FLRMTPERM) {
 575                         struct mdt_remote_perm *perm;
 576
 577                         LASSERT(client_is_remote(exp));
 578                         perm = lustre_swab_repbuf(req, reply_off++,
 579                                                   sizeof(*perm),
 580                                                   lustre_swab_mdt_remote_perm);
 581                         if (perm == NULL) {
 582                                 CERROR("missing remote permission!\n");
 583                                 RETURN(-EPROTO);
 584                         }
 585                 } else if ((body->valid & OBD_MD_FLACL) && body->aclsize) {
 586                         reply_off++;
 587                 }
 588                 if (body->valid & OBD_MD_FLMDSCAPA) {
 589                         struct lustre_capa *capa, *p;
 590
 591                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 592                         if (capa == NULL) {
 593                                 CERROR("Missing/short MDS capability\n");
 594                                 RETURN(-EPROTO);
 595                         }
 596
 597                         if (it->it_op & IT_OPEN) {
 598                                 /* client fid capa will be checked in replay */
 599                                 p = lustre_msg_buf(req->rq_reqmsg,
 600                                                    DLM_INTENT_REC_OFF + 2,
 601                                                    sizeof(*p));
 602                                 LASSERT(p);
 603                                 *p = *capa;
 604                         }
 605                 }
 606                 if (body->valid & OBD_MD_FLOSSCAPA) {
 607                         struct lustre_capa *capa;
 608
 609                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 610                         if (capa == NULL) {
 611                                 CERROR("Missing/short OSS capability\n");
 612                                 RETURN(-EPROTO);
 613                         }
 614                 }
 615         }
 616
 617         RETURN(rc);
 618 }
 619 /*
 620  * This long block is all about fixing up the lock and request state
 621  * so that it is correct as of the moment _before_ the operation was
 622  * applied; that way, the VFS will think that everything is normal and
 623  * call Lustre's regular VFS methods.
 624  *
 625  * If we're performing a creation, that means that unless the creation
 626  * failed with EEXIST, we should fake up a negative dentry.
 627  *
 628  * For everything else, we want to lookup to succeed.
 629  *
 630  * One additional note: if CREATE or OPEN succeeded, we add an extra
 631  * reference to the request because we need to keep it around until
 632  * ll_create/ll_open gets called.
 633  *
 634  * The server will return to us, in it_disposition, an indication of
 635  * exactly what d.lustre.it_status refers to.
 636  *
 637  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 638  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 639  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 640  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 641  * was successful.
 642  *
 643  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 644  * child lookup.
 645  */
 646 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 647                     void *lmm, int lmmsize, struct lookup_intent *it,
 648                     int lookup_flags, struct ptlrpc_request **reqp,
 649                     ldlm_blocking_callback cb_blocking,
 650                     int extra_lock_flags)
 651 {
 652         struct ptlrpc_request *request;
 653         struct lustre_handle old_lock;
 654         struct lustre_handle lockh;
 655         struct mdt_body *mdt_body;
 656         struct ldlm_lock *lock;
 657         int rc = 0;
 658         ENTRY;
 659         LASSERT(it);
 660
 661         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
 662                ", intent: %s flags %#o\n", op_data->op_namelen,
 663                op_data->op_name, PFID(&op_data->op_fid2),
 664                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
 665                it->it_flags);
 666
 667         if (fid_is_sane(&op_data->op_fid2) &&
 668             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
 669                 /* We could just return 1 immediately, but since we should only
 670                  * be called in revalidate_it if we already have a lock, let's
 671                  * verify that. */
 672                 struct ldlm_res_id res_id = { .name = { fid_seq(&op_data->op_fid2),
 673                                                         fid_oid(&op_data->op_fid2),
 674                                                         fid_ver(&op_data->op_fid2) } };
 675                 ldlm_policy_data_t policy;
 676                 ldlm_mode_t mode = LCK_CR;
 677
 678                 /* As not all attributes are kept under update lock, e.g.
 679                    owner/group/acls are under lookup lock, we need both
 680                    ibits for GETATTR. */
 681
 682                 /* For CMD, UPDATE lock and LOOKUP lock can not be got
 683                  * at the same for cross-object, so we can not match
 684                  * the 2 lock at the same time FIXME: but how to handle
 685                  * the above situation */
 686                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 687                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 688
 689                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 690                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 691                                      LDLM_IBITS, &policy, mode, &lockh);
 692                 if (!rc) {
 693                         mode = LCK_CW;
 694                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 695                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 696                                              LDLM_IBITS, &policy, mode, &lockh);
 697                 }
 698                 if (!rc) {
 699                         mode = LCK_PR;
 700                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 701                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 702                                              LDLM_IBITS, &policy, mode, &lockh);
 703                 }
 704
 705                 if (!rc) {
 706                         mode = LCK_PW;
 707                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 708                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 709                                              LDLM_IBITS, &policy, mode, &lockh);
 710                 }
 711
 712                 if (rc) {
 713                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 714                                sizeof(lockh));
 715                         it->d.lustre.it_lock_mode = mode;
 716                 }
 717
 718                 /* Only return failure if it was not GETATTR by cfid
 719                    (from inode_revalidate) */
 720                 if (rc || op_data->op_namelen != 0)
 721                         RETURN(rc);
 722         }
 723
 724         /* lookup_it may be called only after revalidate_it has run, because
 725          * revalidate_it cannot return errors, only zero.  Returning zero causes
 726          * this call to lookup, which *can* return an error.
 727          *
 728          * We only want to execute the request associated with the intent one
 729          * time, however, so don't send the request again.  Instead, skip past
 730          * this and use the request from revalidate.  In this case, revalidate
 731          * never dropped its reference, so the refcounts are all OK */
 732         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 733                 struct ldlm_enqueue_info einfo =
 734                         { LDLM_IBITS, it_to_lock_mode(it), cb_blocking,
 735                           ldlm_completion_ast, NULL, NULL };
 736
 737                 /* For case if upper layer did not alloc fid, do it now. */
 738                 if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
 739                         rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
 740                         if (rc < 0) {
 741                                 CERROR("Can't alloc new fid, rc %d\n", rc);
 742                                 RETURN(rc);
 743                         }
 744                 }
 745                 rc = mdc_enqueue(exp, &einfo, it, op_data, &lockh,
 746                                  lmm, lmmsize, extra_lock_flags);
 747                 if (rc < 0)
 748                         RETURN(rc);
 749                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 750         } else if (!fid_is_sane(&op_data->op_fid2) ||
 751                    !(it->it_flags & O_CHECK_STALE)) {
 752                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 753                  * request referenced from this intent, saved for subsequent
 754                  * lookup.  This path is executed when we proceed to this
 755                  * lookup, so we clear DISP_ENQ_COMPLETE */
 756                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 757         }
 758         request = *reqp = it->d.lustre.it_data;
 759         LASSERT(request != NULL);
 760         LASSERT(request != LP_POISON);
 761         LASSERT(request->rq_repmsg != LP_POISON);
 762
 763         if (!it_disposition(it, DISP_IT_EXECD)) {
 764                 /* The server failed before it even started executing the
 765                  * intent, i.e. because it couldn't unpack the request. */
 766                 LASSERT(it->d.lustre.it_status != 0);
 767                 RETURN(it->d.lustre.it_status);
 768         }
 769         rc = it_open_error(DISP_IT_EXECD, it);
 770         if (rc)
 771                 RETURN(rc);
 772
 773         mdt_body = lustre_msg_buf(request->rq_repmsg, DLM_REPLY_REC_OFF,
 774                                   sizeof(*mdt_body));
 775         /* mdc_enqueue checked */
 776         LASSERT(mdt_body != NULL);
 777         /* mdc_enqueue swabbed */
 778         LASSERT(lustre_rep_swabbed(request, 1));
 779
 780         /* If we were revalidating a fid/name pair, mark the intent in
 781          * case we fail and get called again from lookup */
 782         if (fid_is_sane(&op_data->op_fid2) && (it->it_flags & O_CHECK_STALE) &&
 783             (it->it_op != IT_GETATTR)) {
 784                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 785
 786                 /* Also: did we find the same inode? */
 787                 if (!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1))
 788                         RETURN(-ESTALE);
 789         }
 790
 791         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 792         if (rc)
 793                 RETURN(rc);
 794
 795         /* keep requests around for the multiple phases of the call
 796          * this shows the DISP_XX must guarantee we make it into the call
 797          */
 798         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 799             it_disposition(it, DISP_OPEN_CREATE) &&
 800             !it_open_error(DISP_OPEN_CREATE, it)) {
 801                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 802                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 803         }
 804         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 805             it_disposition(it, DISP_OPEN_OPEN) &&
 806             !it_open_error(DISP_OPEN_OPEN, it)) {
 807                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 808                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 809                 /* BUG 11546 - eviction in the middle of open rpc processing */
 810                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 811         }
 812
 813         if (it->it_op & IT_CREAT) {
 814                 /* XXX this belongs in ll_create_it */
 815         } else if (it->it_op == IT_OPEN) {
 816                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 817         } else {
 818                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 819         }
 820
 821         /* If we already have a matching lock, then cancel the new
 822          * one.  We have to set the data here instead of in
 823          * mdc_enqueue, because we need to use the child's inode as
 824          * the l_ast_data to match, and that's not available until
 825          * intent_finish has performed the iget().) */
 826         lock = ldlm_handle2lock(&lockh);
 827         if (lock) {
 828                 ldlm_policy_data_t policy = lock->l_policy_data;
 829                 LDLM_DEBUG(lock, "matching against this");
 830
 831                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
 832                                          &lock->l_resource->lr_name),
 833                          "Lock res_id: %lu/%lu/%lu, fid: %lu/%lu/%lu.\n",
 834                          (unsigned long)lock->l_resource->lr_name.name[0],
 835                          (unsigned long)lock->l_resource->lr_name.name[1],
 836                          (unsigned long)lock->l_resource->lr_name.name[2],
 837                          (unsigned long)fid_seq(&mdt_body->fid1),
 838                          (unsigned long)fid_oid(&mdt_body->fid1),
 839                          (unsigned long)fid_ver(&mdt_body->fid1));
 840                 LDLM_LOCK_PUT(lock);
 841
 842                 memcpy(&old_lock, &lockh, sizeof(lockh));
 843                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 844                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 845                         ldlm_lock_decref_and_cancel(&lockh,
 846                                                     it->d.lustre.it_lock_mode);
 847                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 848                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 849                                sizeof(lockh));
 850                 }
 851         }
 852         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 853                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
 854                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 855
 856         RETURN(rc);
 857 }