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,
 250                 int lock_type,
 251                 struct lookup_intent *it,
 252                 int lock_mode,
 253                 struct md_op_data *op_data,
 254                 struct lustre_handle *lockh,
 255                 void *lmm,
 256                 int lmmsize,
 257                 ldlm_completion_callback cb_completion,
 258                 ldlm_blocking_callback cb_blocking,
 259                 void *cb_data, int extra_lock_flags)
 260 {
 261         struct ptlrpc_request *req;
 262         struct obd_device *obddev = class_exp2obd(exp);
 263         struct ldlm_res_id res_id =
 264                 { .name = {fid_seq(&op_data->op_fid1),
 265                            fid_oid(&op_data->op_fid1),
 266                            fid_ver(&op_data->op_fid1)} };
 267         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 268         struct ldlm_request *lockreq;
 269         struct ldlm_intent *lit;
 270         struct ldlm_reply *lockrep;
 271         int size[9] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 272                         [DLM_LOCKREQ_OFF]     = sizeof(*lockreq),
 273                         [DLM_INTENT_IT_OFF]   = sizeof(*lit),
 274                         0, 0, 0, 0, 0, 0 };
 275         int repsize[7] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 276                            [DLM_LOCKREPLY_OFF]   = sizeof(*lockrep),
 277                            [DLM_REPLY_REC_OFF]   = sizeof(struct mdt_body),
 278                            [DLM_REPLY_REC_OFF+1] = obddev->u.cli.
 279                                                    cl_max_mds_easize };
 280         int flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 281         int repbufcnt = 4, rc;
 282         ENTRY;
 283
 284         LASSERTF(lock_type == LDLM_IBITS, "lock type %d\n", lock_type);
 285 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 286 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 287
 288         if (it->it_op & IT_OPEN) {
 289                 int do_join = !!(it->it_flags & O_JOIN_FILE);
 290                 CFS_LIST_HEAD(cancels);
 291                 int count = 0;
 292                 int mode;
 293
 294                 it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 295
 296                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_create);
 297                 /* parent capability */
 298                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 299                                                sizeof(struct lustre_capa) : 0;
 300                 /* child capability, used for replay only */
 301                 size[DLM_INTENT_REC_OFF + 2] = sizeof(struct lustre_capa);
 302                 size[DLM_INTENT_REC_OFF + 3] = op_data->op_namelen + 1;
 303                 /* As an optimization, we allocate an RPC request buffer for
 304                  * at least a default-sized LOV EA even if we aren't sending
 305                  * one.
 306                  */
 307                 size[DLM_INTENT_REC_OFF + 4] = max(lmmsize,
 308                                            obddev->u.cli.cl_default_mds_easize);
 309
 310                 /* XXX: openlock is not cancelled for cross-refs. */
 311                 /* If inode is known, cancel conflicting OPEN locks. */
 312                 if (fid_is_sane(&op_data->op_fid2)) {
 313                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 314                                 mode = LCK_CW;
 315 #ifdef FMODE_EXEC
 316                         else if (it->it_flags & FMODE_EXEC)
 317                                 mode = LCK_PR;
 318 #endif
 319                         else
 320                                 mode = LCK_CR;
 321                         count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 322                                                         &cancels, mode,
 323                                                         MDS_INODELOCK_OPEN);
 324                 }
 325
 326                 /* If CREATE or JOIN_FILE, cancel parent's UPDATE lock. */
 327                 if (it->it_op & IT_CREAT || it->it_flags & O_JOIN_FILE)
 328                         mode = LCK_EX;
 329                 else
 330                         mode = LCK_CR;
 331                 count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 332                                                  &cancels, mode,
 333                                                  MDS_INODELOCK_UPDATE);
 334
 335                 if (do_join)
 336                         size[DLM_INTENT_REC_OFF + 5] =
 337                                                 sizeof(struct mdt_rec_join);
 338
 339                 req = ldlm_prep_enqueue_req(exp, 8 + do_join, size, &cancels,
 340                                             count);
 341                 if (!req)
 342                         RETURN(-ENOMEM);
 343
 344                 if (do_join) {
 345                         /* join is like an unlink of the tail */
 346                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 347                         mdc_join_pack(req, DLM_INTENT_REC_OFF + 5, op_data,
 348                                       (*(__u64 *)op_data->op_data));
 349                 }
 350
 351                 spin_lock(&req->rq_lock);
 352                 req->rq_replay = 1;
 353                 spin_unlock(&req->rq_lock);
 354
 355                 /* pack the intent */
 356                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 357                                      sizeof(*lit));
 358                 lit->opc = (__u64)it->it_op;
 359
 360                 /* pack the intended request */
 361                 mdc_open_pack(req, DLM_INTENT_REC_OFF, op_data,
 362                               it->it_create_mode, 0, it->it_flags,
 363                               lmm, lmmsize);
 364
 365                 /* for remote client, fetch remote perm for current user */
 366                 repsize[repbufcnt++] = client_is_remote(exp) ?
 367                                        sizeof(struct mdt_remote_perm) :
 368                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 369                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 370                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 371         } else if (it->it_op & IT_UNLINK) {
 372                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_unlink);
 373                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 374                                                sizeof(struct lustre_capa) : 0;
 375                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 376                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 377                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 378                 if (!req)
 379                         RETURN(-ENOMEM);
 380
 381                 /* pack the intent */
 382                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 383                                      sizeof(*lit));
 384                 lit->opc = (__u64)it->it_op;
 385
 386                 /* pack the intended request */
 387                 mdc_unlink_pack(req, DLM_INTENT_REC_OFF, op_data);
 388
 389                 repsize[repbufcnt++] = obddev->u.cli.cl_max_mds_cookiesize;
 390         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 391                 obd_valid valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 392                         OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 393                         OBD_MD_FLMDSCAPA | OBD_MD_MEA;
 394                 valid |= client_is_remote(exp) ? OBD_MD_FLRMTPERM :
 395                                                  OBD_MD_FLACL;
 396                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_body);
 397                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 398                                                sizeof(struct lustre_capa) : 0;
 399                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 400
 401                 if (it->it_op & IT_GETATTR)
 402                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 403
 404                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 405                 if (!req)
 406                         RETURN(-ENOMEM);
 407
 408                 /* pack the intent */
 409                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 410                                      sizeof(*lit));
 411                 lit->opc = (__u64)it->it_op;
 412
 413                 /* pack the intended request */
 414                 mdc_getattr_pack(req, DLM_INTENT_REC_OFF, valid,
 415                                  it->it_flags, op_data);
 416
 417                 repsize[repbufcnt++] = client_is_remote(exp) ?
 418                                        sizeof(struct mdt_remote_perm) :
 419                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 420                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 421         } else if (it->it_op == IT_READDIR) {
 422                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 423                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 424                 if (!req)
 425                         RETURN(-ENOMEM);
 426
 427                 repbufcnt = 2;
 428         } else {
 429                 LBUG();
 430                 RETURN(-EINVAL);
 431         }
 432
 433         /* get ready for the reply */
 434         ptlrpc_req_set_repsize(req, repbufcnt, repsize);
 435
 436          /* It is important to obtain rpc_lock first (if applicable), so that
 437           * threads that are serialised with rpc_lock are not polluting our
 438           * rpcs in flight counter */
 439         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 440         mdc_enter_request(&obddev->u.cli);
 441         rc = ldlm_cli_enqueue(exp, &req, &res_id, lock_type, &policy,
 442                               lock_mode, &flags, cb_blocking, cb_completion,
 443                               NULL, cb_data, NULL, 0, NULL, lockh, 0);
 444         mdc_exit_request(&obddev->u.cli);
 445         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 446
 447         /* Similarly, if we're going to replay this request, we don't want to
 448          * actually get a lock, just perform the intent. */
 449         if (req->rq_transno || req->rq_replay) {
 450                 lockreq = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 451                                          sizeof(*lockreq));
 452                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 453         }
 454
 455         /* This can go when we're sure that this can never happen */
 456         LASSERT(rc != -ENOENT);
 457         if (rc == ELDLM_LOCK_ABORTED) {
 458                 lock_mode = 0;
 459                 memset(lockh, 0, sizeof(*lockh));
 460                 rc = 0;
 461         } else if (rc != 0) {
 462                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 463                 LASSERTF(rc < 0, "rc %d\n", rc);
 464                 mdc_clear_replay_flag(req, rc);
 465                 ptlrpc_req_finished(req);
 466                 RETURN(rc);
 467         } else { /* rc = 0 */
 468                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 469                 LASSERT(lock);
 470
 471                 /* If the server gave us back a different lock mode, we should
 472                  * fix up our variables. */
 473                 if (lock->l_req_mode != lock_mode) {
 474                         ldlm_lock_addref(lockh, lock->l_req_mode);
 475                         ldlm_lock_decref(lockh, lock_mode);
 476                         lock_mode = lock->l_req_mode;
 477                 }
 478                 LDLM_LOCK_PUT(lock);
 479         }
 480
 481         lockrep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 482                                  sizeof(*lockrep));
 483         LASSERT(lockrep != NULL);                 /* checked by ldlm_cli_enqueue() */
 484         LASSERT_REPSWABBED(req, DLM_LOCKREPLY_OFF); /* swabbed by ldlm_cli_enqueue() */
 485
 486         it->d.lustre.it_disposition = (int)lockrep->lock_policy_res1;
 487         it->d.lustre.it_status = (int)lockrep->lock_policy_res2;
 488         it->d.lustre.it_lock_mode = lock_mode;
 489         it->d.lustre.it_data = req;
 490
 491         if (it->d.lustre.it_status < 0 && req->rq_replay)
 492                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 493
 494         /* If we're doing an IT_OPEN which did not result in an actual
 495          * successful open, then we need to remove the bit which saves
 496          * this request for unconditional replay.
 497          *
 498          * It's important that we do this first!  Otherwise we might exit the
 499          * function without doing so, and try to replay a failed create
 500          * (bug 3440) */
 501         if (it->it_op & IT_OPEN && req->rq_replay &&
 502             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 503                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 504
 505         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 506                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 507
 508         /* We know what to expect, so we do any byte flipping required here */
 509         LASSERT(repbufcnt == 7 || repbufcnt == 6 || repbufcnt == 2);
 510         if (repbufcnt >= 6) {
 511                 int reply_off = DLM_REPLY_REC_OFF;
 512                 struct mdt_body *body;
 513
 514                 body = lustre_swab_repbuf(req, reply_off++, sizeof(*body),
 515                                          lustre_swab_mdt_body);
 516                 if (body == NULL) {
 517                         CERROR ("Can't swab mdt_body\n");
 518                         RETURN (-EPROTO);
 519                 }
 520
 521                 if (req->rq_replay && it_disposition(it, DISP_OPEN_OPEN) &&
 522                     !it_open_error(DISP_OPEN_OPEN, it)) {
 523                         /*
 524                          * If this is a successful OPEN request, we need to set
 525                          * replay handler and data early, so that if replay
 526                          * happens immediately after swabbing below, new reply
 527                          * is swabbed by that handler correctly.
 528                          */
 529                         mdc_set_open_replay_data(NULL, NULL, req);
 530                 }
 531
 532                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 533                         void *eadata;
 534
 535                         /*
 536                          * The eadata is opaque; just check that it is there.
 537                          * Eventually, obd_unpackmd() will check the contents.
 538                          */
 539                         eadata = lustre_swab_repbuf(req, reply_off++,
 540                                                     body->eadatasize, NULL);
 541                         if (eadata == NULL) {
 542                                 CERROR("Missing/short eadata\n");
 543                                 RETURN(-EPROTO);
 544                         }
 545                         if (body->valid & OBD_MD_FLMODEASIZE) {
 546                                 if (obddev->u.cli.cl_max_mds_easize <
 547                                     body->max_mdsize) {
 548                                         obddev->u.cli.cl_max_mds_easize =
 549                                                 body->max_mdsize;
 550                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 551                                                body->max_mdsize);
 552                                 }
 553                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 554                                     body->max_cookiesize) {
 555                                         obddev->u.cli.cl_max_mds_cookiesize =
 556                                                 body->max_cookiesize;
 557                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 558                                                body->max_cookiesize);
 559                                 }
 560                         }
 561
 562                         /*
 563                          * We save the reply LOV EA in case we have to replay a
 564                          * create for recovery.  If we didn't allocate a large
 565                          * enough request buffer above we need to reallocate it
 566                          * here to hold the actual LOV EA.
 567                          *
 568                          * To not save LOV EA if request is not going to replay
 569                          * (for example error one).
 570                          */
 571                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 572                                 if (lustre_msg_buflen(req->rq_reqmsg,
 573                                                       DLM_INTENT_REC_OFF + 4) <
 574                                     body->eadatasize)
 575                                         mdc_realloc_openmsg(req, body, size);
 576
 577                                 lmm = lustre_msg_buf(req->rq_reqmsg,
 578                                                      DLM_INTENT_REC_OFF + 4,
 579                                                      body->eadatasize);
 580                                 if (lmm)
 581                                         memcpy(lmm, eadata, body->eadatasize);
 582                         }
 583                 }
 584                 if (body->valid & OBD_MD_FLRMTPERM) {
 585                         struct mdt_remote_perm *perm;
 586
 587                         LASSERT(client_is_remote(exp));
 588                         perm = lustre_swab_repbuf(req, reply_off++,
 589                                                   sizeof(*perm),
 590                                                   lustre_swab_mdt_remote_perm);
 591                         if (perm == NULL) {
 592                                 CERROR("missing remote permission!\n");
 593                                 RETURN(-EPROTO);
 594                         }
 595                 } else if ((body->valid & OBD_MD_FLACL) && body->aclsize) {
 596                         reply_off++;
 597                 }
 598                 if (body->valid & OBD_MD_FLMDSCAPA) {
 599                         struct lustre_capa *capa, *p;
 600
 601                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 602                         if (capa == NULL) {
 603                                 CERROR("Missing/short MDS capability\n");
 604                                 RETURN(-EPROTO);
 605                         }
 606
 607                         if (it->it_op & IT_OPEN) {
 608                                 /* client fid capa will be checked in replay */
 609                                 p = lustre_msg_buf(req->rq_reqmsg,
 610                                                    DLM_INTENT_REC_OFF + 2,
 611                                                    sizeof(*p));
 612                                 LASSERT(p);
 613                                 *p = *capa;
 614                         }
 615                 }
 616                 if (body->valid & OBD_MD_FLOSSCAPA) {
 617                         struct lustre_capa *capa;
 618
 619                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 620                         if (capa == NULL) {
 621                                 CERROR("Missing/short OSS capability\n");
 622                                 RETURN(-EPROTO);
 623                         }
 624                 }
 625         }
 626
 627         RETURN(rc);
 628 }
 629 /*
 630  * This long block is all about fixing up the lock and request state
 631  * so that it is correct as of the moment _before_ the operation was
 632  * applied; that way, the VFS will think that everything is normal and
 633  * call Lustre's regular VFS methods.
 634  *
 635  * If we're performing a creation, that means that unless the creation
 636  * failed with EEXIST, we should fake up a negative dentry.
 637  *
 638  * For everything else, we want to lookup to succeed.
 639  *
 640  * One additional note: if CREATE or OPEN succeeded, we add an extra
 641  * reference to the request because we need to keep it around until
 642  * ll_create/ll_open gets called.
 643  *
 644  * The server will return to us, in it_disposition, an indication of
 645  * exactly what d.lustre.it_status refers to.
 646  *
 647  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 648  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 649  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 650  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 651  * was successful.
 652  *
 653  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 654  * child lookup.
 655  */
 656 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 657                     void *lmm, int lmmsize, struct lookup_intent *it,
 658                     int lookup_flags, struct ptlrpc_request **reqp,
 659                     ldlm_blocking_callback cb_blocking,
 660                     int extra_lock_flags)
 661 {
 662         struct ptlrpc_request *request;
 663         struct lustre_handle old_lock;
 664         struct lustre_handle lockh;
 665         struct mdt_body *mdt_body;
 666         struct ldlm_lock *lock;
 667         int rc = 0;
 668         ENTRY;
 669         LASSERT(it);
 670
 671         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
 672                ", intent: %s flags %#o\n", op_data->op_namelen,
 673                op_data->op_name, PFID(&op_data->op_fid2),
 674                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
 675                it->it_flags);
 676
 677         if (fid_is_sane(&op_data->op_fid2) &&
 678             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
 679                 /* We could just return 1 immediately, but since we should only
 680                  * be called in revalidate_it if we already have a lock, let's
 681                  * verify that. */
 682                 struct ldlm_res_id res_id = { .name = { fid_seq(&op_data->op_fid2),
 683                                                         fid_oid(&op_data->op_fid2),
 684                                                         fid_ver(&op_data->op_fid2) } };
 685                 ldlm_policy_data_t policy;
 686                 ldlm_mode_t mode = LCK_CR;
 687
 688                 /* As not all attributes are kept under update lock, e.g.
 689                    owner/group/acls are under lookup lock, we need both
 690                    ibits for GETATTR. */
 691
 692                 /* For CMD, UPDATE lock and LOOKUP lock can not be got
 693                  * at the same for cross-object, so we can not match
 694                  * the 2 lock at the same time FIXME: but how to handle
 695                  * the above situation */
 696                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 697                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 698
 699                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 700                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 701                                      LDLM_IBITS, &policy, mode, &lockh);
 702                 if (!rc) {
 703                         mode = LCK_CW;
 704                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 705                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 706                                              LDLM_IBITS, &policy, mode, &lockh);
 707                 }
 708                 if (!rc) {
 709                         mode = LCK_PR;
 710                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 711                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 712                                              LDLM_IBITS, &policy, mode, &lockh);
 713                 }
 714
 715                 if (!rc) {
 716                         mode = LCK_PW;
 717                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 718                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 719                                              LDLM_IBITS, &policy, mode, &lockh);
 720                 }
 721
 722                 if (rc) {
 723                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 724                                sizeof(lockh));
 725                         it->d.lustre.it_lock_mode = mode;
 726                 }
 727
 728                 /* Only return failure if it was not GETATTR by cfid
 729                    (from inode_revalidate) */
 730                 if (rc || op_data->op_namelen != 0)
 731                         RETURN(rc);
 732         }
 733
 734         /* lookup_it may be called only after revalidate_it has run, because
 735          * revalidate_it cannot return errors, only zero.  Returning zero causes
 736          * this call to lookup, which *can* return an error.
 737          *
 738          * We only want to execute the request associated with the intent one
 739          * time, however, so don't send the request again.  Instead, skip past
 740          * this and use the request from revalidate.  In this case, revalidate
 741          * never dropped its reference, so the refcounts are all OK */
 742         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 743                 /* For case if upper layer did not alloc fid, do it now. */
 744                 if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
 745                         rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
 746                         if (rc < 0) {
 747                                 CERROR("Can't alloc new fid, rc %d\n", rc);
 748                                 RETURN(rc);
 749                         }
 750                 }
 751
 752                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 753                                  op_data, &lockh, lmm, lmmsize,
 754                                  ldlm_completion_ast, cb_blocking, NULL,
 755                                  extra_lock_flags);
 756                 if (rc < 0)
 757                         RETURN(rc);
 758                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 759         } else if (!fid_is_sane(&op_data->op_fid2) ||
 760                    !(it->it_flags & O_CHECK_STALE)) {
 761                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 762                  * request referenced from this intent, saved for subsequent
 763                  * lookup.  This path is executed when we proceed to this
 764                  * lookup, so we clear DISP_ENQ_COMPLETE */
 765                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 766         }
 767         request = *reqp = it->d.lustre.it_data;
 768         LASSERT(request != NULL);
 769         LASSERT(request != LP_POISON);
 770         LASSERT(request->rq_repmsg != LP_POISON);
 771
 772         if (!it_disposition(it, DISP_IT_EXECD)) {
 773                 /* The server failed before it even started executing the
 774                  * intent, i.e. because it couldn't unpack the request. */
 775                 LASSERT(it->d.lustre.it_status != 0);
 776                 RETURN(it->d.lustre.it_status);
 777         }
 778         rc = it_open_error(DISP_IT_EXECD, it);
 779         if (rc)
 780                 RETURN(rc);
 781
 782         mdt_body = lustre_msg_buf(request->rq_repmsg, DLM_REPLY_REC_OFF,
 783                                   sizeof(*mdt_body));
 784         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 785         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 786
 787         /* If we were revalidating a fid/name pair, mark the intent in
 788          * case we fail and get called again from lookup */
 789         if (fid_is_sane(&op_data->op_fid2) && (it->it_flags & O_CHECK_STALE) &&
 790             (it->it_op != IT_GETATTR)) {
 791                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 792
 793                 /* Also: did we find the same inode? */
 794                 if (!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1))
 795                         RETURN(-ESTALE);
 796         }
 797
 798         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 799         if (rc)
 800                 RETURN(rc);
 801
 802         /* keep requests around for the multiple phases of the call
 803          * this shows the DISP_XX must guarantee we make it into the call
 804          */
 805         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 806             it_disposition(it, DISP_OPEN_CREATE) &&
 807             !it_open_error(DISP_OPEN_CREATE, it)) {
 808                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 809                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 810         }
 811         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 812             it_disposition(it, DISP_OPEN_OPEN) &&
 813             !it_open_error(DISP_OPEN_OPEN, it)) {
 814                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 815                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 816                 /* BUG 11546 - eviction in the middle of open rpc processing */
 817                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 818         }
 819
 820         if (it->it_op & IT_CREAT) {
 821                 /* XXX this belongs in ll_create_it */
 822         } else if (it->it_op == IT_OPEN) {
 823                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 824         } else {
 825                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 826         }
 827
 828         /* If we already have a matching lock, then cancel the new
 829          * one.  We have to set the data here instead of in
 830          * mdc_enqueue, because we need to use the child's inode as
 831          * the l_ast_data to match, and that's not available until
 832          * intent_finish has performed the iget().) */
 833         lock = ldlm_handle2lock(&lockh);
 834         if (lock) {
 835                 ldlm_policy_data_t policy = lock->l_policy_data;
 836                 LDLM_DEBUG(lock, "matching against this");
 837
 838                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
 839                                          &lock->l_resource->lr_name),
 840                          "Lock res_id: %lu/%lu/%lu, fid: %lu/%lu/%lu.\n",
 841                          (unsigned long)lock->l_resource->lr_name.name[0],
 842                          (unsigned long)lock->l_resource->lr_name.name[1],
 843                          (unsigned long)lock->l_resource->lr_name.name[2],
 844                          (unsigned long)fid_seq(&mdt_body->fid1),
 845                          (unsigned long)fid_oid(&mdt_body->fid1),
 846                          (unsigned long)fid_ver(&mdt_body->fid1));
 847                 LDLM_LOCK_PUT(lock);
 848
 849                 memcpy(&old_lock, &lockh, sizeof(lockh));
 850                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 851                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 852                         ldlm_lock_decref_and_cancel(&lockh,
 853                                                     it->d.lustre.it_lock_mode);
 854                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 855                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 856                                sizeof(lockh));
 857                 }
 858         }
 859         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 860                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
 861                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 862
 863         RETURN(rc);
 864 }