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