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, int size[9])
 230 {
 231         int     rc;
 232         ENTRY;
 233
 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         EXIT;
 243 }
 244
 245 /* We always reserve enough space in the reply packet for a stripe MD, because
 246  * we don't know in advance the file type. */
 247 int mdc_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
 248                 struct lookup_intent *it, struct md_op_data *op_data,
 249                 struct lustre_handle *lockh, void *lmm, int lmmsize,
 250                 int extra_lock_flags)
 251 {
 252         struct ptlrpc_request *req;
 253         struct obd_device *obddev = class_exp2obd(exp);
 254         struct ldlm_res_id res_id =
 255                 { .name = {fid_seq(&op_data->op_fid1),
 256                            fid_oid(&op_data->op_fid1),
 257                            fid_ver(&op_data->op_fid1)} };
 258         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 259         struct ldlm_request *lockreq;
 260         struct ldlm_intent *lit;
 261         struct ldlm_reply *lockrep;
 262         int size[9] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 263                         [DLM_LOCKREQ_OFF]     = sizeof(*lockreq),
 264                         [DLM_INTENT_IT_OFF]   = sizeof(*lit),
 265                         0, 0, 0, 0, 0, 0 };
 266         int repsize[7] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 267                            [DLM_LOCKREPLY_OFF]   = sizeof(*lockrep),
 268                            [DLM_REPLY_REC_OFF]   = sizeof(struct mdt_body),
 269                            [DLM_REPLY_REC_OFF+1] = obddev->u.cli.
 270                                                    cl_max_mds_easize,
 271                            0, 0, 0 };
 272         int flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 273         int repbufcnt = 4, rc;
 274         ENTRY;
 275
 276         LASSERTF(einfo->ei_type == LDLM_IBITS,"lock type %d\n", einfo->ei_type);
 277
 278         if (it->it_op & IT_OPEN) {
 279                 int do_join = !!(it->it_flags & O_JOIN_FILE);
 280                 CFS_LIST_HEAD(cancels);
 281                 int count = 0;
 282                 int mode;
 283
 284                 it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 285
 286                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_create);
 287                 /* parent capability */
 288                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 289                                                sizeof(struct lustre_capa) : 0;
 290                 /* child capability, used for replay only */
 291                 size[DLM_INTENT_REC_OFF + 2] = sizeof(struct lustre_capa);
 292                 size[DLM_INTENT_REC_OFF + 3] = op_data->op_namelen + 1;
 293                 /* As an optimization, we allocate an RPC request buffer for
 294                  * at least a default-sized LOV EA even if we aren't sending
 295                  * one.
 296                  */
 297                 size[DLM_INTENT_REC_OFF + 4] = max(lmmsize,
 298                                            obddev->u.cli.cl_default_mds_easize);
 299
 300                 /* XXX: openlock is not cancelled for cross-refs. */
 301                 /* If inode is known, cancel conflicting OPEN locks. */
 302                 if (fid_is_sane(&op_data->op_fid2)) {
 303                         if (it->it_flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
 304                                 mode = LCK_CW;
 305 #ifdef FMODE_EXEC
 306                         else if (it->it_flags & FMODE_EXEC)
 307                                 mode = LCK_PR;
 308 #endif
 309                         else
 310                                 mode = LCK_CR;
 311                         count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 312                                                         &cancels, mode,
 313                                                         MDS_INODELOCK_OPEN);
 314                 }
 315
 316                 /* If CREATE or JOIN_FILE, cancel parent's UPDATE lock. */
 317                 if (it->it_op & IT_CREAT || it->it_flags & O_JOIN_FILE)
 318                         mode = LCK_EX;
 319                 else
 320                         mode = LCK_CR;
 321                 count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 322                                                  &cancels, mode,
 323                                                  MDS_INODELOCK_UPDATE);
 324
 325                 if (do_join)
 326                         size[DLM_INTENT_REC_OFF + 5] =
 327                                                 sizeof(struct mdt_rec_join);
 328
 329                 req = ldlm_prep_enqueue_req(exp, 8 + do_join, size, &cancels,
 330                                             count);
 331                 if (!req)
 332                         RETURN(-ENOMEM);
 333
 334                 if (do_join) {
 335                         /* join is like an unlink of the tail */
 336                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 337                         mdc_join_pack(req, DLM_INTENT_REC_OFF + 5, op_data,
 338                                       (*(__u64 *)op_data->op_data));
 339                 }
 340
 341                 spin_lock(&req->rq_lock);
 342                 req->rq_replay = 1;
 343                 spin_unlock(&req->rq_lock);
 344
 345                 /* pack the intent */
 346                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 347                                      sizeof(*lit));
 348                 lit->opc = (__u64)it->it_op;
 349
 350                 /* pack the intended request */
 351                 mdc_open_pack(req, DLM_INTENT_REC_OFF, op_data,
 352                               it->it_create_mode, 0, it->it_flags,
 353                               lmm, lmmsize);
 354
 355                 /* for remote client, fetch remote perm for current user */
 356                 repsize[repbufcnt++] = client_is_remote(exp) ?
 357                                        sizeof(struct mdt_remote_perm) :
 358                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 359                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 360                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 361         } else if (it->it_op & IT_UNLINK) {
 362                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_unlink);
 363                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 364                                                sizeof(struct lustre_capa) : 0;
 365                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 366                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 367                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 368                 if (!req)
 369                         RETURN(-ENOMEM);
 370
 371                 /* pack the intent */
 372                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 373                                      sizeof(*lit));
 374                 lit->opc = (__u64)it->it_op;
 375
 376                 /* pack the intended request */
 377                 mdc_unlink_pack(req, DLM_INTENT_REC_OFF, op_data);
 378
 379                 repsize[repbufcnt++] = obddev->u.cli.cl_max_mds_cookiesize;
 380         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 381                 obd_valid valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 382                         OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 383                         OBD_MD_FLMDSCAPA | OBD_MD_MEA;
 384                 valid |= client_is_remote(exp) ? OBD_MD_FLRMTPERM :
 385                                                  OBD_MD_FLACL;
 386                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_body);
 387                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 388                                                sizeof(struct lustre_capa) : 0;
 389                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 390
 391                 if (it->it_op & IT_GETATTR)
 392                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 393
 394                 req = ldlm_prep_enqueue_req(exp, 6, size, NULL, 0);
 395                 if (!req)
 396                         RETURN(-ENOMEM);
 397
 398                 /* pack the intent */
 399                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 400                                      sizeof(*lit));
 401                 lit->opc = (__u64)it->it_op;
 402
 403                 /* pack the intended request */
 404                 mdc_getattr_pack(req, DLM_INTENT_REC_OFF, valid,
 405                                  it->it_flags, op_data);
 406
 407                 repsize[repbufcnt++] = client_is_remote(exp) ?
 408                                        sizeof(struct mdt_remote_perm) :
 409                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 410                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 411         } else if (it->it_op == IT_READDIR) {
 412                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 413                 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
 414                 if (!req)
 415                         RETURN(-ENOMEM);
 416
 417                 repbufcnt = 2;
 418         } else {
 419                 LBUG();
 420                 RETURN(-EINVAL);
 421         }
 422
 423         /* get ready for the reply */
 424         ptlrpc_req_set_repsize(req, repbufcnt, repsize);
 425
 426          /* It is important to obtain rpc_lock first (if applicable), so that
 427           * threads that are serialised with rpc_lock are not polluting our
 428           * rpcs in flight counter */
 429         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 430         mdc_enter_request(&obddev->u.cli);
 431         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, &policy, &flags, NULL,
 432                               0, NULL, lockh, 0);
 433         mdc_exit_request(&obddev->u.cli);
 434         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 435
 436         /* Similarly, if we're going to replay this request, we don't want to
 437          * actually get a lock, just perform the intent. */
 438         if (req->rq_transno || req->rq_replay) {
 439                 lockreq = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 440                                          sizeof(*lockreq));
 441                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 442         }
 443
 444         if (rc == ELDLM_LOCK_ABORTED) {
 445                 einfo->ei_mode = 0;
 446                 memset(lockh, 0, sizeof(*lockh));
 447                 rc = 0;
 448         } else if (rc != 0) {
 449                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 450                 LASSERTF(rc < 0, "rc %d\n", rc);
 451                 mdc_clear_replay_flag(req, rc);
 452                 ptlrpc_req_finished(req);
 453                 RETURN(rc);
 454         } else { /* rc = 0 */
 455                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 456                 LASSERT(lock);
 457
 458                 /* If the server gave us back a different lock mode, we should
 459                  * fix up our variables. */
 460                 if (lock->l_req_mode != einfo->ei_mode) {
 461                         ldlm_lock_addref(lockh, lock->l_req_mode);
 462                         ldlm_lock_decref(lockh, einfo->ei_mode);
 463                         einfo->ei_mode = lock->l_req_mode;
 464                 }
 465                 LDLM_LOCK_PUT(lock);
 466         }
 467
 468         lockrep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 469                                  sizeof(*lockrep));
 470         LASSERT(lockrep != NULL);                 /* checked by ldlm_cli_enqueue() */
 471         /* swabbed by ldlm_cli_enqueue() */
 472         LASSERT(lustre_rep_swabbed(req, DLM_LOCKREPLY_OFF));
 473
 474         it->d.lustre.it_disposition = (int)lockrep->lock_policy_res1;
 475         it->d.lustre.it_status = (int)lockrep->lock_policy_res2;
 476         it->d.lustre.it_lock_mode = einfo->ei_mode;
 477         it->d.lustre.it_data = req;
 478
 479         if (it->d.lustre.it_status < 0 && req->rq_replay)
 480                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 481
 482         /* If we're doing an IT_OPEN which did not result in an actual
 483          * successful open, then we need to remove the bit which saves
 484          * this request for unconditional replay.
 485          *
 486          * It's important that we do this first!  Otherwise we might exit the
 487          * function without doing so, and try to replay a failed create
 488          * (bug 3440) */
 489         if (it->it_op & IT_OPEN && req->rq_replay &&
 490             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 491                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 492
 493         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 494                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 495
 496         /* We know what to expect, so we do any byte flipping required here */
 497         LASSERT(repbufcnt == 7 || repbufcnt == 6 || repbufcnt == 2);
 498         if (repbufcnt >= 6) {
 499                 int reply_off = DLM_REPLY_REC_OFF;
 500                 struct mdt_body *body;
 501
 502                 body = lustre_swab_repbuf(req, reply_off++, sizeof(*body),
 503                                          lustre_swab_mdt_body);
 504                 if (body == NULL) {
 505                         CERROR ("Can't swab mdt_body\n");
 506                         RETURN (-EPROTO);
 507                 }
 508
 509                 if (req->rq_replay && it_disposition(it, DISP_OPEN_OPEN) &&
 510                     !it_open_error(DISP_OPEN_OPEN, it)) {
 511                         /*
 512                          * If this is a successful OPEN request, we need to set
 513                          * replay handler and data early, so that if replay
 514                          * happens immediately after swabbing below, new reply
 515                          * is swabbed by that handler correctly.
 516                          */
 517                         mdc_set_open_replay_data(NULL, NULL, req);
 518                 }
 519
 520                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 521                         void *eadata;
 522
 523                         /*
 524                          * The eadata is opaque; just check that it is there.
 525                          * Eventually, obd_unpackmd() will check the contents.
 526                          */
 527                         eadata = lustre_swab_repbuf(req, reply_off++,
 528                                                     body->eadatasize, NULL);
 529                         if (eadata == NULL) {
 530                                 CERROR("Missing/short eadata\n");
 531                                 RETURN(-EPROTO);
 532                         }
 533                         if (body->valid & OBD_MD_FLMODEASIZE) {
 534                                 if (obddev->u.cli.cl_max_mds_easize <
 535                                     body->max_mdsize) {
 536                                         obddev->u.cli.cl_max_mds_easize =
 537                                                 body->max_mdsize;
 538                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 539                                                body->max_mdsize);
 540                                 }
 541                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 542                                     body->max_cookiesize) {
 543                                         obddev->u.cli.cl_max_mds_cookiesize =
 544                                                 body->max_cookiesize;
 545                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 546                                                body->max_cookiesize);
 547                                 }
 548                         }
 549
 550                         /*
 551                          * We save the reply LOV EA in case we have to replay a
 552                          * create for recovery.  If we didn't allocate a large
 553                          * enough request buffer above we need to reallocate it
 554                          * here to hold the actual LOV EA.
 555                          *
 556                          * To not save LOV EA if request is not going to replay
 557                          * (for example error one).
 558                          */
 559                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 560                                 if (lustre_msg_buflen(req->rq_reqmsg,
 561                                                       DLM_INTENT_REC_OFF + 4) <
 562                                     body->eadatasize)
 563                                         mdc_realloc_openmsg(req, body, size);
 564
 565                                 lmm = lustre_msg_buf(req->rq_reqmsg,
 566                                                      DLM_INTENT_REC_OFF + 4,
 567                                                      body->eadatasize);
 568                                 if (lmm)
 569                                         memcpy(lmm, eadata, body->eadatasize);
 570                         }
 571                 }
 572                 if (body->valid & OBD_MD_FLRMTPERM) {
 573                         struct mdt_remote_perm *perm;
 574
 575                         LASSERT(client_is_remote(exp));
 576                         perm = lustre_swab_repbuf(req, reply_off++,
 577                                                   sizeof(*perm),
 578                                                   lustre_swab_mdt_remote_perm);
 579                         if (perm == NULL) {
 580                                 CERROR("missing remote permission!\n");
 581                                 RETURN(-EPROTO);
 582                         }
 583                 } else if ((body->valid & OBD_MD_FLACL) && body->aclsize) {
 584                         reply_off++;
 585                 }
 586                 if (body->valid & OBD_MD_FLMDSCAPA) {
 587                         struct lustre_capa *capa, *p;
 588
 589                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 590                         if (capa == NULL) {
 591                                 CERROR("Missing/short MDS capability\n");
 592                                 RETURN(-EPROTO);
 593                         }
 594
 595                         if (it->it_op & IT_OPEN) {
 596                                 /* client fid capa will be checked in replay */
 597                                 p = lustre_msg_buf(req->rq_reqmsg,
 598                                                    DLM_INTENT_REC_OFF + 2,
 599                                                    sizeof(*p));
 600                                 LASSERT(p);
 601                                 *p = *capa;
 602                         }
 603                 }
 604                 if (body->valid & OBD_MD_FLOSSCAPA) {
 605                         struct lustre_capa *capa;
 606
 607                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 608                         if (capa == NULL) {
 609                                 CERROR("Missing/short OSS capability\n");
 610                                 RETURN(-EPROTO);
 611                         }
 612                 }
 613         }
 614
 615         RETURN(rc);
 616 }
 617 /*
 618  * This long block is all about fixing up the lock and request state
 619  * so that it is correct as of the moment _before_ the operation was
 620  * applied; that way, the VFS will think that everything is normal and
 621  * call Lustre's regular VFS methods.
 622  *
 623  * If we're performing a creation, that means that unless the creation
 624  * failed with EEXIST, we should fake up a negative dentry.
 625  *
 626  * For everything else, we want to lookup to succeed.
 627  *
 628  * One additional note: if CREATE or OPEN succeeded, we add an extra
 629  * reference to the request because we need to keep it around until
 630  * ll_create/ll_open gets called.
 631  *
 632  * The server will return to us, in it_disposition, an indication of
 633  * exactly what d.lustre.it_status refers to.
 634  *
 635  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 636  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 637  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 638  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 639  * was successful.
 640  *
 641  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 642  * child lookup.
 643  */
 644 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 645                     void *lmm, int lmmsize, struct lookup_intent *it,
 646                     int lookup_flags, struct ptlrpc_request **reqp,
 647                     ldlm_blocking_callback cb_blocking,
 648                     int extra_lock_flags)
 649 {
 650         struct ptlrpc_request *request;
 651         struct lustre_handle old_lock;
 652         struct lustre_handle lockh;
 653         struct mdt_body *mdt_body;
 654         struct ldlm_lock *lock;
 655         int rc = 0;
 656         ENTRY;
 657         LASSERT(it);
 658
 659         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
 660                ", intent: %s flags %#o\n", op_data->op_namelen,
 661                op_data->op_name, PFID(&op_data->op_fid2),
 662                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
 663                it->it_flags);
 664
 665         if (fid_is_sane(&op_data->op_fid2) &&
 666             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
 667                 /* We could just return 1 immediately, but since we should only
 668                  * be called in revalidate_it if we already have a lock, let's
 669                  * verify that. */
 670                 ldlm_policy_data_t policy;
 671                 ldlm_mode_t mode;
 672
 673                 /* As not all attributes are kept under update lock, e.g.
 674                    owner/group/acls are under lookup lock, we need both
 675                    ibits for GETATTR. */
 676
 677                 /* For CMD, UPDATE lock and LOOKUP lock can not be got
 678                  * at the same for cross-object, so we can not match
 679                  * the 2 lock at the same time FIXME: but how to handle
 680                  * the above situation */
 681                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 682                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 683
 684                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED,
 685                                       &op_data->op_fid2, LDLM_IBITS, &policy,
 686                                       LCK_CR|LCK_CW|LCK_PR|LCK_PW, &lockh);
 687                 if (mode) {
 688                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 689                                sizeof(lockh));
 690                         it->d.lustre.it_lock_mode = mode;
 691                 }
 692
 693                 /* Only return failure if it was not GETATTR by cfid
 694                    (from inode_revalidate) */
 695                 if (mode || op_data->op_namelen != 0)
 696                         RETURN(!!mode);
 697         }
 698
 699         /* lookup_it may be called only after revalidate_it has run, because
 700          * revalidate_it cannot return errors, only zero.  Returning zero causes
 701          * this call to lookup, which *can* return an error.
 702          *
 703          * We only want to execute the request associated with the intent one
 704          * time, however, so don't send the request again.  Instead, skip past
 705          * this and use the request from revalidate.  In this case, revalidate
 706          * never dropped its reference, so the refcounts are all OK */
 707         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 708                 struct ldlm_enqueue_info einfo =
 709                         { LDLM_IBITS, it_to_lock_mode(it), cb_blocking,
 710                           ldlm_completion_ast, NULL, NULL };
 711
 712                 /* For case if upper layer did not alloc fid, do it now. */
 713                 if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
 714                         rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
 715                         if (rc < 0) {
 716                                 CERROR("Can't alloc new fid, rc %d\n", rc);
 717                                 RETURN(rc);
 718                         }
 719                 }
 720                 rc = mdc_enqueue(exp, &einfo, it, op_data, &lockh,
 721                                  lmm, lmmsize, extra_lock_flags);
 722                 if (rc < 0)
 723                         RETURN(rc);
 724                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 725         } else if (!fid_is_sane(&op_data->op_fid2) ||
 726                    !(it->it_flags & O_CHECK_STALE)) {
 727                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 728                  * request referenced from this intent, saved for subsequent
 729                  * lookup.  This path is executed when we proceed to this
 730                  * lookup, so we clear DISP_ENQ_COMPLETE */
 731                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 732         }
 733         request = *reqp = it->d.lustre.it_data;
 734         LASSERT(request != NULL);
 735         LASSERT(request != LP_POISON);
 736         LASSERT(request->rq_repmsg != LP_POISON);
 737
 738         if (!it_disposition(it, DISP_IT_EXECD)) {
 739                 /* The server failed before it even started executing the
 740                  * intent, i.e. because it couldn't unpack the request. */
 741                 LASSERT(it->d.lustre.it_status != 0);
 742                 RETURN(it->d.lustre.it_status);
 743         }
 744         rc = it_open_error(DISP_IT_EXECD, it);
 745         if (rc)
 746                 RETURN(rc);
 747
 748         mdt_body = lustre_msg_buf(request->rq_repmsg, DLM_REPLY_REC_OFF,
 749                                   sizeof(*mdt_body));
 750         /* mdc_enqueue checked */
 751         LASSERT(mdt_body != NULL);
 752         /* mdc_enqueue swabbed */
 753         LASSERT(lustre_rep_swabbed(request, 1));
 754
 755         /* If we were revalidating a fid/name pair, mark the intent in
 756          * case we fail and get called again from lookup */
 757         if (fid_is_sane(&op_data->op_fid2) && (it->it_flags & O_CHECK_STALE) &&
 758             (it->it_op != IT_GETATTR)) {
 759                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 760
 761                 /* Also: did we find the same inode? */
 762                 if (!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1))
 763                         RETURN(-ESTALE);
 764         }
 765
 766         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 767         if (rc)
 768                 RETURN(rc);
 769
 770         /* keep requests around for the multiple phases of the call
 771          * this shows the DISP_XX must guarantee we make it into the call
 772          */
 773         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 774             it_disposition(it, DISP_OPEN_CREATE) &&
 775             !it_open_error(DISP_OPEN_CREATE, it)) {
 776                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 777                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 778         }
 779         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 780             it_disposition(it, DISP_OPEN_OPEN) &&
 781             !it_open_error(DISP_OPEN_OPEN, it)) {
 782                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 783                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 784                 /* BUG 11546 - eviction in the middle of open rpc processing */
 785                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE, obd_timeout);
 786         }
 787
 788         if (it->it_op & IT_CREAT) {
 789                 /* XXX this belongs in ll_create_it */
 790         } else if (it->it_op == IT_OPEN) {
 791                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 792         } else {
 793                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 794         }
 795
 796         /* If we already have a matching lock, then cancel the new
 797          * one.  We have to set the data here instead of in
 798          * mdc_enqueue, because we need to use the child's inode as
 799          * the l_ast_data to match, and that's not available until
 800          * intent_finish has performed the iget().) */
 801         lock = ldlm_handle2lock(&lockh);
 802         if (lock) {
 803                 ldlm_policy_data_t policy = lock->l_policy_data;
 804                 LDLM_DEBUG(lock, "matching against this");
 805
 806                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
 807                                          &lock->l_resource->lr_name),
 808                          "Lock res_id: %lu/%lu/%lu, fid: %lu/%lu/%lu.\n",
 809                          (unsigned long)lock->l_resource->lr_name.name[0],
 810                          (unsigned long)lock->l_resource->lr_name.name[1],
 811                          (unsigned long)lock->l_resource->lr_name.name[2],
 812                          (unsigned long)fid_seq(&mdt_body->fid1),
 813                          (unsigned long)fid_oid(&mdt_body->fid1),
 814                          (unsigned long)fid_ver(&mdt_body->fid1));
 815                 LDLM_LOCK_PUT(lock);
 816
 817                 memcpy(&old_lock, &lockh, sizeof(lockh));
 818                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 819                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 820                         ldlm_lock_decref_and_cancel(&lockh,
 821                                                     it->d.lustre.it_lock_mode);
 822                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 823                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 824                                sizeof(lockh));
 825                 }
 826         }
 827         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 828                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
 829                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 830
 831         RETURN(rc);
 832 }