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 ldlm_mode_t 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         ldlm_mode_t rc;
 159         ENTRY;
 160
 161         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 162                              &res_id, type, policy, mode, lockh);
 163         RETURN(rc);
 164 }
 165
 166 int mdc_cancel_unused(struct obd_export *exp,
 167                       const struct lu_fid *fid,
 168                       ldlm_policy_data_t *policy,
 169                       ldlm_mode_t mode, int flags, void *opaque)
 170 {
 171         struct ldlm_res_id res_id =
 172                 { .name = {fid_seq(fid),
 173                            fid_oid(fid),
 174                            fid_ver(fid)} };
 175         struct obd_device *obd = class_exp2obd(exp);
 176         int rc;
 177
 178         ENTRY;
 179
 180         rc = ldlm_cli_cancel_unused_resource(obd->obd_namespace, &res_id,
 181                                              policy, mode, flags, opaque);
 182         RETURN(rc);
 183 }
 184
 185 int mdc_change_cbdata(struct obd_export *exp,
 186                       const struct lu_fid *fid,
 187                       ldlm_iterator_t it, void *data)
 188 {
 189         struct ldlm_res_id res_id = { .name = {0} };
 190         ENTRY;
 191
 192         res_id.name[0] = fid_seq(fid);
 193         res_id.name[1] = fid_oid(fid);
 194         res_id.name[2] = fid_ver(fid);
 195
 196         ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace,
 197                               &res_id, it, data);
 198
 199         EXIT;
 200         return 0;
 201 }
 202
 203 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 204 {
 205         /* Don't hold error requests for replay. */
 206         if (req->rq_replay) {
 207                 spin_lock(&req->rq_lock);
 208                 req->rq_replay = 0;
 209                 spin_unlock(&req->rq_lock);
 210         }
 211         if (rc && req->rq_transno != 0) {
 212                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 213                 LBUG();
 214         }
 215 }
 216
 217 /* Save a large LOV EA into the request buffer so that it is available
 218  * for replay.  We don't do this in the initial request because the
 219  * original request doesn't need this buffer (at most it sends just the
 220  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 221  * buffer and may also be difficult to allocate and save a very large
 222  * request buffer for each open. (bug 5707)
 223  *
 224  * OOM here may cause recovery failure if lmm is needed (only for the
 225  * original open if the MDS crashed just when this client also OOM'd)
 226  * but this is incredibly unlikely, and questionable whether the client
 227  * could do MDS recovery under OOM anyways... */
 228 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 229                                struct mdt_body *body)
 230 {
 231         int     rc;
 232
 233         /* FIXME: remove this explicit offset. */
 234         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 235                                         body->eadatasize);
 236         if (rc) {
 237                 CERROR("Can't enlarge segment %d size to %d\n",
 238                        DLM_INTENT_REC_OFF + 4, body->eadatasize);
 239                 body->valid &= ~OBD_MD_FLEASIZE;
 240                 body->eadatasize = 0;
 241         }
 242 }
 243
 244 static struct ptlrpc_request *mdc_intent_open_pack(struct obd_export *exp,
 245                                                    struct lookup_intent *it,
 246                                                    struct md_op_data *op_data,
 247                                                    void *lmm, int lmmsize,
 248                                                    void *cb_data)
 249 {
 250         struct ptlrpc_request *req;
 251         struct obd_device     *obddev = class_exp2obd(exp);
 252         struct ldlm_intent    *lit;
 253         int                    joinfile = !!((it->it_flags & O_JOIN_FILE) &&
 254                                               op_data->op_data);
 255         CFS_LIST_HEAD(cancels);
 256         int                    count = 0;
 257         int                    mode;
 258         int                    rc;
 259         ENTRY;
 260
 261         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 262
 263         /* XXX: openlock is not cancelled for cross-refs. */
 264         /* If inode is known, cancel conflicting OPEN locks. */
 265         if (fid_is_sane(&op_data->op_fid2)) {
 266                 if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 267                         mode = LCK_CW;
 268 #ifdef FMODE_EXEC
 269                 else if (it->it_flags & FMODE_EXEC)
 270                         mode = LCK_PR;
 271 #endif
 272                 else
 273                         mode = LCK_CR;
 274                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 275                                                 &cancels, mode,
 276                                                 MDS_INODELOCK_OPEN);
 277         }
 278
 279         /* If CREATE or JOIN_FILE, cancel parent's UPDATE lock. */
 280         if (it->it_op & IT_CREAT || joinfile)
 281                 mode = LCK_EX;
 282         else
 283                 mode = LCK_CR;
 284         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 285                                          &cancels, mode,
 286                                          MDS_INODELOCK_UPDATE);
 287
 288         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 289                                    &RQF_LDLM_INTENT_OPEN);
 290         if (req == NULL) {
 291                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 292                 RETURN(ERR_PTR(-ENOMEM));
 293         }
 294
 295         /* parent capability */
 296         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 297         /* child capability, reserve the size according to parent capa, it will
 298          * be filled after we get the reply */
 299         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa1);
 300
 301         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 302                              op_data->op_namelen + 1);
 303         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 304                              max(lmmsize, obddev->u.cli.cl_default_mds_easize));
 305         if (!joinfile) {
 306                 req_capsule_set_size(&req->rq_pill, &RMF_REC_JOINFILE,
 307                                      RCL_CLIENT, 0);
 308         }
 309
 310         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 311         if (rc) {
 312                 ptlrpc_request_free(req);
 313                 return NULL;
 314         }
 315
 316         if (joinfile) {
 317                 __u64 head_size = *(__u64 *)op_data->op_data;
 318                 mdc_join_pack(req, op_data, head_size);
 319         }
 320
 321         spin_lock(&req->rq_lock);
 322         req->rq_replay = 1;
 323         spin_unlock(&req->rq_lock);
 324
 325         /* pack the intent */
 326         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 327         lit->opc = (__u64)it->it_op;
 328
 329         /* pack the intended request */
 330         mdc_open_pack(req, op_data, it->it_create_mode, 0, it->it_flags, lmm,
 331                       lmmsize);
 332
 333         /* for remote client, fetch remote perm for current user */
 334         if (client_is_remote(exp))
 335                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 336                                      sizeof(struct mdt_remote_perm));
 337         ptlrpc_request_set_replen(req);
 338         return req;
 339 }
 340
 341 static struct ptlrpc_request *mdc_intent_unlink_pack(struct obd_export *exp,
 342                                                      struct lookup_intent *it,
 343                                                      struct md_op_data *op_data)
 344 {
 345         struct ptlrpc_request *req;
 346         struct obd_device     *obddev = class_exp2obd(exp);
 347         struct ldlm_intent    *lit;
 348         int                    rc;
 349         ENTRY;
 350
 351         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 352                                    &RQF_LDLM_INTENT_UNLINK);
 353         if (req == NULL)
 354                 RETURN(ERR_PTR(-ENOMEM));
 355
 356         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 357         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 358                              op_data->op_namelen + 1);
 359
 360         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 361         if (rc) {
 362                 ptlrpc_request_free(req);
 363                 RETURN(ERR_PTR(rc));
 364         }
 365
 366         /* pack the intent */
 367         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 368         lit->opc = (__u64)it->it_op;
 369
 370         /* pack the intended request */
 371         mdc_unlink_pack(req, op_data);
 372
 373         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 374                              obddev->u.cli.cl_max_mds_easize);
 375         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 376                              obddev->u.cli.cl_max_mds_cookiesize);
 377         ptlrpc_request_set_replen(req);
 378         RETURN(req);
 379 }
 380
 381 static struct ptlrpc_request *mdc_intent_getattr_pack(struct obd_export *exp,
 382                                                       struct lookup_intent *it,
 383                                                      struct md_op_data *op_data)
 384 {
 385         struct ptlrpc_request *req;
 386         struct obd_device     *obddev = class_exp2obd(exp);
 387         obd_valid              valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 388                                        OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 389                                        OBD_MD_FLMDSCAPA | OBD_MD_MEA |
 390                                        (client_is_remote(exp) ?
 391                                                OBD_MD_FLRMTPERM : OBD_MD_FLACL);
 392         struct ldlm_intent    *lit;
 393         int                    rc;
 394         ENTRY;
 395
 396         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 397                                    &RQF_LDLM_INTENT_GETATTR);
 398         if (req == NULL)
 399                 RETURN(ERR_PTR(-ENOMEM));
 400
 401         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 402         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 403                              op_data->op_namelen + 1);
 404
 405         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 406         if (rc) {
 407                 ptlrpc_request_free(req);
 408                 RETURN(ERR_PTR(rc));
 409         }
 410
 411         /* pack the intent */
 412         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 413         lit->opc = (__u64)it->it_op;
 414
 415         /* pack the intended request */
 416         mdc_getattr_pack(req, valid, it->it_flags, op_data);
 417
 418         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 419                              obddev->u.cli.cl_max_mds_easize);
 420         if (client_is_remote(exp))
 421                 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
 422                                      sizeof(struct mdt_remote_perm));
 423         ptlrpc_request_set_replen(req);
 424         RETURN(req);
 425 }
 426
 427 static struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp)
 428 {
 429         struct ptlrpc_request *req;
 430         int rc;
 431         ENTRY;
 432
 433         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 434         if (req == NULL)
 435                 RETURN(ERR_PTR(-ENOMEM));
 436
 437         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 438         if (rc) {
 439                 ptlrpc_request_free(req);
 440                 RETURN(ERR_PTR(rc));
 441         }
 442
 443         ptlrpc_request_set_replen(req);
 444         RETURN(req);
 445 }
 446
 447 /* We always reserve enough space in the reply packet for a stripe MD, because
 448  * we don't know in advance the file type. */
 449 int mdc_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
 450                 struct lookup_intent *it, struct md_op_data *op_data,
 451                 struct lustre_handle *lockh, void *lmm, int lmmsize,
 452                 int extra_lock_flags)
 453 {
 454         struct obd_device     *obddev = class_exp2obd(exp);
 455         struct ptlrpc_request *req;
 456         struct req_capsule    *pill;
 457         struct ldlm_request   *lockreq;
 458         struct ldlm_reply     *lockrep;
 459         int                    flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 460         int                    rc;
 461         struct ldlm_res_id res_id =
 462                 { .name = {fid_seq(&op_data->op_fid1),
 463                            fid_oid(&op_data->op_fid1),
 464                            fid_ver(&op_data->op_fid1)} };
 465         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 466         ENTRY;
 467
 468         LASSERTF(einfo->ei_type == LDLM_IBITS,"lock type %d\n", einfo->ei_type);
 469
 470         if (it->it_op & (IT_UNLINK | IT_GETATTR | IT_READDIR))
 471                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 472
 473         if (it->it_op & IT_OPEN) {
 474                 int joinfile = !!((it->it_flags & O_JOIN_FILE) &&
 475                                               op_data->op_data);
 476
 477                 req = mdc_intent_open_pack(exp, it, op_data, lmm, lmmsize,
 478                                            einfo->ei_cbdata);
 479                 if (!joinfile) {
 480                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 481                         einfo->ei_cbdata = NULL;
 482                         lmm = NULL;
 483                 } else
 484                         it->it_flags &= ~O_JOIN_FILE;
 485         } else if (it->it_op & IT_UNLINK)
 486                 req = mdc_intent_unlink_pack(exp, it, op_data);
 487         else if (it->it_op & (IT_GETATTR | IT_LOOKUP))
 488                 req = mdc_intent_getattr_pack(exp, it, op_data);
 489         else if (it->it_op == IT_READDIR)
 490                 req = ldlm_enqueue_pack(exp);
 491         else {
 492                 LBUG();
 493                 RETURN(-EINVAL);
 494         }
 495
 496         if (IS_ERR(req))
 497                 RETURN(PTR_ERR(req));
 498         pill = &req->rq_pill;
 499
 500         /* It is important to obtain rpc_lock first (if applicable), so that
 501          * threads that are serialised with rpc_lock are not polluting our
 502          * rpcs in flight counter */
 503         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 504         mdc_enter_request(&obddev->u.cli);
 505         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
 506                               0, NULL, lockh, 0);
 507         mdc_exit_request(&obddev->u.cli);
 508         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 509
 510         /* Similarly, if we're going to replay this request, we don't want to
 511          * actually get a lock, just perform the intent. */
 512         if (req->rq_transno || req->rq_replay) {
 513                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 514                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 515         }
 516
 517         if (rc == ELDLM_LOCK_ABORTED) {
 518                 einfo->ei_mode = 0;
 519                 memset(lockh, 0, sizeof(*lockh));
 520                 rc = 0;
 521         } else if (rc != 0) {
 522                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 523                 LASSERTF(rc < 0, "rc %d\n", rc);
 524                 mdc_clear_replay_flag(req, rc);
 525                 ptlrpc_req_finished(req);
 526                 RETURN(rc);
 527         } else { /* rc = 0 */
 528                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 529                 LASSERT(lock);
 530
 531                 /* If the server gave us back a different lock mode, we should
 532                  * fix up our variables. */
 533                 if (lock->l_req_mode != einfo->ei_mode) {
 534                         ldlm_lock_addref(lockh, lock->l_req_mode);
 535                         ldlm_lock_decref(lockh, einfo->ei_mode);
 536                         einfo->ei_mode = lock->l_req_mode;
 537                 }
 538                 LDLM_LOCK_PUT(lock);
 539         }
 540
 541         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 542         LASSERT(lockrep != NULL);                 /* checked by ldlm_cli_enqueue() */
 543
 544         it->d.lustre.it_disposition = (int)lockrep->lock_policy_res1;
 545         it->d.lustre.it_status = (int)lockrep->lock_policy_res2;
 546         it->d.lustre.it_lock_mode = einfo->ei_mode;
 547         it->d.lustre.it_data = req;
 548
 549         if (it->d.lustre.it_status < 0 && req->rq_replay)
 550                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 551
 552         /* If we're doing an IT_OPEN which did not result in an actual
 553          * successful open, then we need to remove the bit which saves
 554          * this request for unconditional replay.
 555          *
 556          * It's important that we do this first!  Otherwise we might exit the
 557          * function without doing so, and try to replay a failed create
 558          * (bug 3440) */
 559         if (it->it_op & IT_OPEN && req->rq_replay &&
 560             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 561                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 562
 563         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 564                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 565
 566         /* We know what to expect, so we do any byte flipping required here */
 567         if (it->it_op & (IT_OPEN | IT_UNLINK | IT_LOOKUP | IT_GETATTR)) {
 568                 struct mdt_body *body;
 569
 570                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 571                 if (body == NULL) {
 572                         CERROR ("Can't swab mdt_body\n");
 573                         RETURN (-EPROTO);
 574                 }
 575
 576                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 577                     !it_open_error(DISP_OPEN_OPEN, it)) {
 578                         /*
 579                          * If this is a successful OPEN request, we need to set
 580                          * replay handler and data early, so that if replay
 581                          * happens immediately after swabbing below, new reply
 582                          * is swabbed by that handler correctly.
 583                          */
 584                         mdc_set_open_replay_data(NULL, NULL, req);
 585                 }
 586
 587                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 588                         void *eadata;
 589
 590                         /*
 591                          * The eadata is opaque; just check that it is there.
 592                          * Eventually, obd_unpackmd() will check the contents.
 593                          */
 594                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 595                                                               body->eadatasize);
 596                         if (eadata == NULL)
 597                                 RETURN(-EPROTO);
 598
 599                         if (body->valid & OBD_MD_FLMODEASIZE) {
 600                                 if (obddev->u.cli.cl_max_mds_easize <
 601                                     body->max_mdsize) {
 602                                         obddev->u.cli.cl_max_mds_easize =
 603                                                 body->max_mdsize;
 604                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 605                                                body->max_mdsize);
 606                                 }
 607                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 608                                     body->max_cookiesize) {
 609                                         obddev->u.cli.cl_max_mds_cookiesize =
 610                                                 body->max_cookiesize;
 611                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 612                                                body->max_cookiesize);
 613                                 }
 614                         }
 615
 616                         /*
 617                          * We save the reply LOV EA in case we have to replay a
 618                          * create for recovery.  If we didn't allocate a large
 619                          * enough request buffer above we need to reallocate it
 620                          * here to hold the actual LOV EA.
 621                          *
 622                          * To not save LOV EA if request is not going to replay
 623                          * (for example error one).
 624                          */
 625                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 626                                 if (req_capsule_get_size(pill, &RMF_EADATA,
 627                                                          RCL_CLIENT) <
 628                                     body->eadatasize) {
 629                                         mdc_realloc_openmsg(req, body);
 630                                         req_capsule_set_size(pill, &RMF_EADATA,
 631                                                              RCL_CLIENT,
 632                                                              body->eadatasize);
 633                                 }
 634                                 lmm = req_capsule_client_get(pill, &RMF_EADATA);
 635                                 if (lmm)
 636                                         memcpy(lmm, eadata, body->eadatasize);
 637                         }
 638                 }
 639
 640                 if (body->valid & OBD_MD_FLRMTPERM) {
 641                         struct mdt_remote_perm *perm;
 642
 643                         LASSERT(client_is_remote(exp));
 644                         perm = req_capsule_server_swab_get(pill, &RMF_ACL,
 645                                                 lustre_swab_mdt_remote_perm);
 646                         if (perm == NULL)
 647                                 RETURN(-EPROTO);
 648                 }
 649                 if (body->valid & OBD_MD_FLMDSCAPA) {
 650                         struct lustre_capa *capa, *p;
 651
 652                         capa = req_capsule_server_get(pill, &RMF_CAPA1);
 653                         if (capa == NULL)
 654                                 RETURN(-EPROTO);
 655
 656                         if (it->it_op & IT_OPEN) {
 657                                 /* client fid capa will be checked in replay */
 658                                 p = req_capsule_client_get(pill, &RMF_CAPA2);
 659                                 LASSERT(p);
 660                                 *p = *capa;
 661                         }
 662                 }
 663                 if (body->valid & OBD_MD_FLOSSCAPA) {
 664                         struct lustre_capa *capa;
 665
 666                         capa = req_capsule_server_get(pill, &RMF_CAPA2);
 667                         if (capa == NULL)
 668                                 RETURN(-EPROTO);
 669                 }
 670         }
 671
 672         RETURN(rc);
 673 }
 674 /*
 675  * This long block is all about fixing up the lock and request state
 676  * so that it is correct as of the moment _before_ the operation was
 677  * applied; that way, the VFS will think that everything is normal and
 678  * call Lustre's regular VFS methods.
 679  *
 680  * If we're performing a creation, that means that unless the creation
 681  * failed with EEXIST, we should fake up a negative dentry.
 682  *
 683  * For everything else, we want to lookup to succeed.
 684  *
 685  * One additional note: if CREATE or OPEN succeeded, we add an extra
 686  * reference to the request because we need to keep it around until
 687  * ll_create/ll_open gets called.
 688  *
 689  * The server will return to us, in it_disposition, an indication of
 690  * exactly what d.lustre.it_status refers to.
 691  *
 692  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 693  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 694  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 695  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 696  * was successful.
 697  *
 698  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 699  * child lookup.
 700  */
 701 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 702                     void *lmm, int lmmsize, struct lookup_intent *it,
 703                     int lookup_flags, struct ptlrpc_request **reqp,
 704                     ldlm_blocking_callback cb_blocking,
 705                     int extra_lock_flags)
 706 {
 707         struct ptlrpc_request *request;
 708         struct lustre_handle old_lock;
 709         struct lustre_handle lockh;
 710         struct mdt_body *mdt_body;
 711         struct ldlm_lock *lock;
 712         int rc = 0;
 713         ENTRY;
 714         LASSERT(it);
 715
 716         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
 717                ", intent: %s flags %#o\n", op_data->op_namelen,
 718                op_data->op_name, PFID(&op_data->op_fid2),
 719                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
 720                it->it_flags);
 721
 722         if (fid_is_sane(&op_data->op_fid2) &&
 723             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
 724                 /* We could just return 1 immediately, but since we should only
 725                  * be called in revalidate_it if we already have a lock, let's
 726                  * verify that. */
 727                 ldlm_policy_data_t policy;
 728                 ldlm_mode_t mode;
 729
 730                 /* As not all attributes are kept under update lock, e.g.
 731                    owner/group/acls are under lookup lock, we need both
 732                    ibits for GETATTR. */
 733
 734                 /* For CMD, UPDATE lock and LOOKUP lock can not be got
 735                  * at the same for cross-object, so we can not match
 736                  * the 2 lock at the same time FIXME: but how to handle
 737                  * the above situation */
 738                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 739                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 740
 741                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED,
 742                                       &op_data->op_fid2, LDLM_IBITS, &policy,
 743                                       LCK_CR|LCK_CW|LCK_PR|LCK_PW, &lockh);
 744                 if (mode) {
 745                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 746                                sizeof(lockh));
 747                         it->d.lustre.it_lock_mode = mode;
 748                 }
 749
 750                 /* Only return failure if it was not GETATTR by cfid
 751                    (from inode_revalidate) */
 752                 if (mode || op_data->op_namelen != 0)
 753                         RETURN(!!mode);
 754         }
 755
 756         /* lookup_it may be called only after revalidate_it has run, because
 757          * revalidate_it cannot return errors, only zero.  Returning zero causes
 758          * this call to lookup, which *can* return an error.
 759          *
 760          * We only want to execute the request associated with the intent one
 761          * time, however, so don't send the request again.  Instead, skip past
 762          * this and use the request from revalidate.  In this case, revalidate
 763          * never dropped its reference, so the refcounts are all OK */
 764         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 765                 struct ldlm_enqueue_info einfo =
 766                         { LDLM_IBITS, it_to_lock_mode(it), cb_blocking,
 767                           ldlm_completion_ast, NULL, NULL };
 768
 769                 /* For case if upper layer did not alloc fid, do it now. */
 770                 if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
 771                         rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
 772                         if (rc < 0) {
 773                                 CERROR("Can't alloc new fid, rc %d\n", rc);
 774                                 RETURN(rc);
 775                         }
 776                 }
 777                 rc = mdc_enqueue(exp, &einfo, it, op_data, &lockh,
 778                                  lmm, lmmsize, extra_lock_flags);
 779                 if (rc < 0)
 780                         RETURN(rc);
 781                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 782         } else if (!fid_is_sane(&op_data->op_fid2) ||
 783                    !(it->it_flags & O_CHECK_STALE)) {
 784                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 785                  * request referenced from this intent, saved for subsequent
 786                  * lookup.  This path is executed when we proceed to this
 787                  * lookup, so we clear DISP_ENQ_COMPLETE */
 788                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 789         }
 790         request = *reqp = it->d.lustre.it_data;
 791         LASSERT(request != NULL);
 792         LASSERT(request != LP_POISON);
 793         LASSERT(request->rq_repmsg != LP_POISON);
 794
 795         if (!it_disposition(it, DISP_IT_EXECD)) {
 796                 /* The server failed before it even started executing the
 797                  * intent, i.e. because it couldn't unpack the request. */
 798                 LASSERT(it->d.lustre.it_status != 0);
 799                 RETURN(it->d.lustre.it_status);
 800         }
 801         rc = it_open_error(DISP_IT_EXECD, it);
 802         if (rc)
 803                 RETURN(rc);
 804
 805         mdt_body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 806         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 807
 808         /* If we were revalidating a fid/name pair, mark the intent in
 809          * case we fail and get called again from lookup */
 810         if (fid_is_sane(&op_data->op_fid2) &&
 811             (it->it_flags & O_CHECK_STALE) &&
 812             it->it_op != IT_GETATTR) {
 813                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 814
 815                 /* Also: did we find the same inode? */
 816                 /* sever can return one of two fids:
 817                  * op_fid2 - new allocated fid - if file is created.
 818                  * op_fid3 - existent fid - if file only open.
 819                  * op_fid3 is saved in lmv_intent_open */
 820                 if ((!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1)) &&
 821                     (!lu_fid_eq(&op_data->op_fid3, &mdt_body->fid1))) {
 822                         CDEBUG(D_DENTRY, "Found stale data "DFID"("DFID")/"DFID
 823                                "\n", PFID(&op_data->op_fid2),
 824                                PFID(&op_data->op_fid2), PFID(&mdt_body->fid1));
 825                         RETURN(-ESTALE);
 826                 }
 827         }
 828
 829         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 830         if (rc)
 831                 RETURN(rc);
 832
 833         /* keep requests around for the multiple phases of the call
 834          * this shows the DISP_XX must guarantee we make it into the call
 835          */
 836         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 837             it_disposition(it, DISP_OPEN_CREATE) &&
 838             !it_open_error(DISP_OPEN_CREATE, it)) {
 839                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 840                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 841         }
 842         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 843             it_disposition(it, DISP_OPEN_OPEN) &&
 844             !it_open_error(DISP_OPEN_OPEN, it)) {
 845                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 846                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 847                 /* BUG 11546 - eviction in the middle of open rpc processing */
 848                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 849         }
 850
 851         if (it->it_op & IT_CREAT) {
 852                 /* XXX this belongs in ll_create_it */
 853         } else if (it->it_op == IT_OPEN) {
 854                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 855         } else {
 856                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 857         }
 858
 859         /* If we already have a matching lock, then cancel the new
 860          * one.  We have to set the data here instead of in
 861          * mdc_enqueue, because we need to use the child's inode as
 862          * the l_ast_data to match, and that's not available until
 863          * intent_finish has performed the iget().) */
 864         lock = ldlm_handle2lock(&lockh);
 865         if (lock) {
 866                 ldlm_policy_data_t policy = lock->l_policy_data;
 867                 LDLM_DEBUG(lock, "matching against this");
 868
 869                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
 870                                          &lock->l_resource->lr_name),
 871                          "Lock res_id: %lu/%lu/%lu, fid: %lu/%lu/%lu.\n",
 872                          (unsigned long)lock->l_resource->lr_name.name[0],
 873                          (unsigned long)lock->l_resource->lr_name.name[1],
 874                          (unsigned long)lock->l_resource->lr_name.name[2],
 875                          (unsigned long)fid_seq(&mdt_body->fid1),
 876                          (unsigned long)fid_oid(&mdt_body->fid1),
 877                          (unsigned long)fid_ver(&mdt_body->fid1));
 878                 LDLM_LOCK_PUT(lock);
 879
 880                 memcpy(&old_lock, &lockh, sizeof(lockh));
 881                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 882                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 883                         ldlm_lock_decref_and_cancel(&lockh,
 884                                                     it->d.lustre.it_lock_mode);
 885                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 886                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 887                                sizeof(lockh));
 888                 }
 889         }
 890         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 891                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
 892                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 893
 894         RETURN(rc);
 895 }