lustre/mdc/mdc_locks.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_MDC
  29
  30 #ifdef __KERNEL__
  31 # include <linux/module.h>
  32 # include <linux/pagemap.h>
  33 # include <linux/miscdevice.h>
  34 # include <linux/init.h>
  35 #else
  36 # include <liblustre.h>
  37 #endif
  38
  39 #include <linux/lustre_acl.h>
  40 #include <obd_class.h>
  41 #include <lustre_dlm.h>
  42 /* fid_res_name_eq() */
  43 #include <lustre_fid.h>
  44 #include <lprocfs_status.h>
  45 #include "mdc_internal.h"
  46
  47 int it_disposition(struct lookup_intent *it, int flag)
  48 {
  49         return it->d.lustre.it_disposition & flag;
  50 }
  51 EXPORT_SYMBOL(it_disposition);
  52
  53 void it_set_disposition(struct lookup_intent *it, int flag)
  54 {
  55         it->d.lustre.it_disposition |= flag;
  56 }
  57 EXPORT_SYMBOL(it_set_disposition);
  58
  59 void it_clear_disposition(struct lookup_intent *it, int flag)
  60 {
  61         it->d.lustre.it_disposition &= ~flag;
  62 }
  63 EXPORT_SYMBOL(it_clear_disposition);
  64
  65 static int it_to_lock_mode(struct lookup_intent *it)
  66 {
  67         ENTRY;
  68
  69         /* CREAT needs to be tested before open (both could be set) */
  70         if (it->it_op & IT_CREAT)
  71                 return LCK_PW;
  72         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP))
  73                 return LCK_PR;
  74
  75         LBUG();
  76         RETURN(-EINVAL);
  77 }
  78
  79 int it_open_error(int phase, struct lookup_intent *it)
  80 {
  81         if (it_disposition(it, DISP_OPEN_OPEN)) {
  82                 if (phase >= DISP_OPEN_OPEN)
  83                         return it->d.lustre.it_status;
  84                 else
  85                         return 0;
  86         }
  87
  88         if (it_disposition(it, DISP_OPEN_CREATE)) {
  89                 if (phase >= DISP_OPEN_CREATE)
  90                         return it->d.lustre.it_status;
  91                 else
  92                         return 0;
  93         }
  94
  95         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  96                 if (phase >= DISP_LOOKUP_EXECD)
  97                         return it->d.lustre.it_status;
  98                 else
  99                         return 0;
 100         }
 101
 102         if (it_disposition(it, DISP_IT_EXECD)) {
 103                 if (phase >= DISP_IT_EXECD)
 104                         return it->d.lustre.it_status;
 105                 else
 106                         return 0;
 107         }
 108         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 109                it->d.lustre.it_status);
 110         LBUG();
 111         return 0;
 112 }
 113 EXPORT_SYMBOL(it_open_error);
 114
 115 /* this must be called on a lockh that is known to have a referenced lock */
 116 int mdc_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
 117 {
 118         struct ldlm_lock *lock;
 119         ENTRY;
 120
 121         if (!*lockh) {
 122                 EXIT;
 123                 RETURN(0);
 124         }
 125
 126         lock = ldlm_handle2lock((struct lustre_handle *)lockh);
 127
 128         LASSERT(lock != NULL);
 129         lock_res_and_lock(lock);
 130 #ifdef __KERNEL__
 131         if (lock->l_ast_data && lock->l_ast_data != data) {
 132                 struct inode *new_inode = data;
 133                 struct inode *old_inode = lock->l_ast_data;
 134                 LASSERTF(old_inode->i_state & I_FREEING,
 135                          "Found existing inode %p/%lu/%u state %lu in lock: "
 136                          "setting data to %p/%lu/%u\n", old_inode,
 137                          old_inode->i_ino, old_inode->i_generation,
 138                          old_inode->i_state,
 139                          new_inode, new_inode->i_ino, new_inode->i_generation);
 140         }
 141 #endif
 142         lock->l_ast_data = data;
 143         unlock_res_and_lock(lock);
 144         LDLM_LOCK_PUT(lock);
 145
 146         RETURN(0);
 147 }
 148
 149 int mdc_lock_match(struct obd_export *exp, int flags,
 150                    const struct lu_fid *fid, ldlm_type_t type,
 151                    ldlm_policy_data_t *policy, ldlm_mode_t mode,
 152                    struct lustre_handle *lockh)
 153 {
 154         struct ldlm_res_id res_id =
 155                 { .name = {fid_seq(fid),
 156                            fid_oid(fid),
 157                            fid_ver(fid)} };
 158         struct obd_device *obd = class_exp2obd(exp);
 159         int rc;
 160         ENTRY;
 161
 162         rc = ldlm_lock_match(obd->obd_namespace, flags,
 163                              &res_id, type, policy, mode, lockh);
 164
 165         RETURN(rc);
 166 }
 167
 168 int mdc_cancel_unused(struct obd_export *exp,
 169                       const struct lu_fid *fid,
 170                       int flags, void *opaque)
 171 {
 172         struct ldlm_res_id res_id =
 173                 { .name = {fid_seq(fid),
 174                            fid_oid(fid),
 175                            fid_ver(fid)} };
 176         struct obd_device *obd = class_exp2obd(exp);
 177         int rc;
 178
 179         ENTRY;
 180
 181         rc = ldlm_cli_cancel_unused(obd->obd_namespace, &res_id,
 182                                     flags, opaque);
 183         RETURN(rc);
 184 }
 185
 186 int mdc_change_cbdata(struct obd_export *exp,
 187                       const struct lu_fid *fid,
 188                       ldlm_iterator_t it, void *data)
 189 {
 190         struct ldlm_res_id res_id = { .name = {0} };
 191         ENTRY;
 192
 193         res_id.name[0] = fid_seq(fid);
 194         res_id.name[1] = fid_oid(fid);
 195         res_id.name[2] = fid_ver(fid);
 196
 197         ldlm_resource_iterate(class_exp2obd(exp)->obd_namespace,
 198                               &res_id, it, data);
 199
 200         EXIT;
 201         return 0;
 202 }
 203
 204 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 205 {
 206         /* Don't hold error requests for replay. */
 207         if (req->rq_replay) {
 208                 spin_lock(&req->rq_lock);
 209                 req->rq_replay = 0;
 210                 spin_unlock(&req->rq_lock);
 211         }
 212         if (rc && req->rq_transno != 0) {
 213                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 214                 LBUG();
 215         }
 216 }
 217
 218 /* Save a large LOV EA into the request buffer so that it is available
 219  * for replay.  We don't do this in the initial request because the
 220  * original request doesn't need this buffer (at most it sends just the
 221  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 222  * buffer and may also be difficult to allocate and save a very large
 223  * request buffer for each open. (bug 5707)
 224  *
 225  * OOM here may cause recovery failure if lmm is needed (only for the
 226  * original open if the MDS crashed just when this client also OOM'd)
 227  * but this is incredibly unlikely, and questionable whether the client
 228  * could do MDS recovery under OOM anyways... */
 229 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 230                                 struct mdt_body *body, int size[9])
 231 {
 232         int     rc;
 233         ENTRY;
 234
 235         rc = sptlrpc_cli_enlarge_reqbuf(req, DLM_INTENT_REC_OFF + 4,
 236                                         body->eadatasize);
 237         if (rc) {
 238                 CERROR("Can't enlarge segment %d size to %d\n",
 239                        DLM_INTENT_REC_OFF + 4, body->eadatasize);
 240                 body->valid &= ~OBD_MD_FLEASIZE;
 241                 body->eadatasize = 0;
 242         }
 243         EXIT;
 244 }
 245
 246 /* We always reserve enough space in the reply packet for a stripe MD, because
 247  * we don't know in advance the file type. */
 248 int mdc_enqueue(struct obd_export *exp,
 249                 int lock_type,
 250                 struct lookup_intent *it,
 251                 int lock_mode,
 252                 struct md_op_data *op_data,
 253                 struct lustre_handle *lockh,
 254                 void *lmm,
 255                 int lmmsize,
 256                 ldlm_completion_callback cb_completion,
 257                 ldlm_blocking_callback cb_blocking,
 258                 void *cb_data, int extra_lock_flags)
 259 {
 260         struct ptlrpc_request *req;
 261         struct obd_device *obddev = class_exp2obd(exp);
 262         struct ldlm_res_id res_id =
 263                 { .name = {fid_seq(&op_data->op_fid1),
 264                            fid_oid(&op_data->op_fid1),
 265                            fid_ver(&op_data->op_fid1)} };
 266         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 267         struct ldlm_request *lockreq;
 268         struct ldlm_intent *lit;
 269         struct ldlm_reply *lockrep;
 270         int size[9] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 271                         [DLM_LOCKREQ_OFF]     = sizeof(*lockreq),
 272                         [DLM_INTENT_IT_OFF]   = sizeof(*lit) };
 273         int repsize[7] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 274                            [DLM_LOCKREPLY_OFF]   = sizeof(*lockrep),
 275                            [DLM_REPLY_REC_OFF]   = sizeof(struct mdt_body),
 276                            [DLM_REPLY_REC_OFF+1] = obddev->u.cli.
 277                                                    cl_max_mds_easize };
 278         int flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 279         int repbufcnt = 4, ea_off, rc;
 280         ENTRY;
 281
 282         LASSERTF(lock_type == LDLM_IBITS, "lock type %d\n", lock_type);
 283 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 284 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 285
 286         if (it->it_op & IT_OPEN) {
 287                 int do_join = !!(it->it_flags & O_JOIN_FILE);
 288
 289                 it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 290
 291                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_create);
 292                 /* parent capability */
 293                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 294                                                sizeof(struct lustre_capa) : 0;
 295                 /* child capability, used for replay only */
 296                 size[DLM_INTENT_REC_OFF + 2] = sizeof(struct lustre_capa);
 297                 size[DLM_INTENT_REC_OFF + 3] = op_data->op_namelen + 1;
 298                 /* As an optimization, we allocate an RPC request buffer for
 299                  * at least a default-sized LOV EA even if we aren't sending
 300                  * one.  We grow the whole request to the next power-of-two
 301                  * size since we get that much from a slab allocation anyways.
 302                  * This avoids an allocation below in the common case where
 303                  * we need to save a default-sized LOV EA for open replay. */
 304                 ea_off = DLM_INTENT_REC_OFF + 4;
 305                 size[ea_off] = max(lmmsize,
 306                                    obddev->u.cli.cl_default_mds_easize);
 307                 if (do_join)
 308                         size[DLM_INTENT_REC_OFF + 5] =
 309                                                 sizeof(struct mdt_rec_join);
 310
 311                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 312                                       LDLM_ENQUEUE, 8 + do_join, size, NULL);
 313                 if (!req)
 314                         RETURN(-ENOMEM);
 315
 316                 if (do_join) {
 317                         __u64 head_size = *(__u32*)cb_data;
 318                         __u32 tsize = *(__u32*)lmm;
 319
 320                         /* join is like an unlink of the tail */
 321                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 322                         /* when joining file, cb_data and lmm args together
 323                          * indicate the head file size*/
 324                         mdc_join_pack(req, DLM_INTENT_REC_OFF + 5, op_data,
 325                                       (head_size << 32) | tsize);
 326                         cb_data = NULL;
 327                         lmm = NULL;
 328                 }
 329
 330                 spin_lock(&req->rq_lock);
 331                 req->rq_replay = 1;
 332                 spin_unlock(&req->rq_lock);
 333
 334                 /* pack the intent */
 335                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 336                                      sizeof(*lit));
 337                 lit->opc = (__u64)it->it_op;
 338
 339                 /* pack the intended request */
 340                 mdc_open_pack(req, DLM_INTENT_REC_OFF, op_data,
 341                               it->it_create_mode, 0, it->it_flags,
 342                               lmm, lmmsize);
 343
 344                 /* for remote client, fetch remote perm for current user */
 345                 repsize[repbufcnt++] = client_is_remote(exp) ?
 346                                        sizeof(struct mdt_remote_perm) :
 347                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 348                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 349                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 350         } else if (it->it_op & IT_UNLINK) {
 351                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_rec_unlink);
 352                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 353                                                sizeof(struct lustre_capa) : 0;
 354                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 355                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 356                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 357                                       LDLM_ENQUEUE, 6, size, NULL);
 358                 if (!req)
 359                         RETURN(-ENOMEM);
 360
 361                 /* pack the intent */
 362                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 363                                      sizeof(*lit));
 364                 lit->opc = (__u64)it->it_op;
 365
 366                 /* pack the intended request */
 367                 mdc_unlink_pack(req, DLM_INTENT_REC_OFF, op_data);
 368
 369                 repsize[repbufcnt++] = obddev->u.cli.cl_max_mds_cookiesize;
 370         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 371                 obd_valid valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 372                                   OBD_MD_FLMODEASIZE | OBD_MD_FLDIREA |
 373                                   OBD_MD_FLMDSCAPA;
 374                 valid |= client_is_remote(exp) ? OBD_MD_FLRMTPERM :
 375                                                  OBD_MD_FLACL;
 376                 size[DLM_INTENT_REC_OFF] = sizeof(struct mdt_body);
 377                 size[DLM_INTENT_REC_OFF + 1] = op_data->op_capa1 ?
 378                                                sizeof(struct lustre_capa) : 0;
 379                 size[DLM_INTENT_REC_OFF + 2] = op_data->op_namelen + 1;
 380
 381                 if (it->it_op & IT_GETATTR)
 382                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 383
 384                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 385                                       LDLM_ENQUEUE, 6, size, NULL);
 386                 if (!req)
 387                         RETURN(-ENOMEM);
 388
 389                 /* pack the intent */
 390                 lit = lustre_msg_buf(req->rq_reqmsg, DLM_INTENT_IT_OFF,
 391                                      sizeof(*lit));
 392                 lit->opc = (__u64)it->it_op;
 393
 394                 /* pack the intended request */
 395                 mdc_getattr_pack(req, DLM_INTENT_REC_OFF, valid,
 396                                  it->it_flags, op_data);
 397
 398                 repsize[repbufcnt++] = client_is_remote(exp) ?
 399                                        sizeof(struct mdt_remote_perm) :
 400                                        LUSTRE_POSIX_ACL_MAX_SIZE;
 401                 repsize[repbufcnt++] = sizeof(struct lustre_capa);
 402         } else if (it->it_op == IT_READDIR) {
 403                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 404                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 405                                       LDLM_ENQUEUE, 2, size, NULL);
 406                 if (!req)
 407                         RETURN(-ENOMEM);
 408
 409                 repbufcnt = 2;
 410         } else {
 411                 LBUG();
 412                 RETURN(-EINVAL);
 413         }
 414
 415         /* get ready for the reply */
 416         ptlrpc_req_set_repsize(req, repbufcnt, repsize);
 417
 418          /* It is important to obtain rpc_lock first (if applicable), so that
 419           * threads that are serialised with rpc_lock are not polluting our
 420           * rpcs in flight counter */
 421         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 422         mdc_enter_request(&obddev->u.cli);
 423         rc = ldlm_cli_enqueue(exp, &req, &res_id, lock_type, &policy,
 424                               lock_mode, &flags, cb_blocking, cb_completion,
 425                               NULL, cb_data, NULL, 0, NULL, lockh, 0);
 426         mdc_exit_request(&obddev->u.cli);
 427         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 428
 429         /* Similarly, if we're going to replay this request, we don't want to
 430          * actually get a lock, just perform the intent. */
 431         if (req->rq_transno || req->rq_replay) {
 432                 lockreq = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
 433                                          sizeof(*lockreq));
 434                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 435         }
 436
 437         /* This can go when we're sure that this can never happen */
 438         LASSERT(rc != -ENOENT);
 439         if (rc == ELDLM_LOCK_ABORTED) {
 440                 lock_mode = 0;
 441                 memset(lockh, 0, sizeof(*lockh));
 442                 rc = 0;
 443         } else if (rc != 0) {
 444                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 445                 LASSERTF(rc < 0, "rc %d\n", rc);
 446                 mdc_clear_replay_flag(req, rc);
 447                 ptlrpc_req_finished(req);
 448                 RETURN(rc);
 449         } else { /* rc = 0 */
 450                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 451                 LASSERT(lock);
 452
 453                 /* If the server gave us back a different lock mode, we should
 454                  * fix up our variables. */
 455                 if (lock->l_req_mode != lock_mode) {
 456                         ldlm_lock_addref(lockh, lock->l_req_mode);
 457                         ldlm_lock_decref(lockh, lock_mode);
 458                         lock_mode = lock->l_req_mode;
 459                 }
 460
 461                 ldlm_lock_allow_match(lock);
 462                 LDLM_LOCK_PUT(lock);
 463         }
 464
 465         lockrep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 466                                  sizeof(*lockrep));
 467         LASSERT(lockrep != NULL);                 /* checked by ldlm_cli_enqueue() */
 468         LASSERT_REPSWABBED(req, DLM_LOCKREPLY_OFF); /* swabbed by ldlm_cli_enqueue() */
 469
 470         it->d.lustre.it_disposition = (int)lockrep->lock_policy_res1;
 471         it->d.lustre.it_status = (int)lockrep->lock_policy_res2;
 472         it->d.lustre.it_lock_mode = lock_mode;
 473         it->d.lustre.it_data = req;
 474
 475         if (it->d.lustre.it_status < 0 && req->rq_replay)
 476                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 477
 478         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 479                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 480
 481         /* We know what to expect, so we do any byte flipping required here */
 482         LASSERT(repbufcnt == 7 || repbufcnt == 6 || repbufcnt == 2);
 483         if (repbufcnt >= 6) {
 484                 int reply_off = DLM_REPLY_REC_OFF;
 485                 struct mdt_body *body;
 486
 487                 body = lustre_swab_repbuf(req, reply_off++, sizeof(*body),
 488                                          lustre_swab_mdt_body);
 489                 if (body == NULL) {
 490                         CERROR ("Can't swab mdt_body\n");
 491                         RETURN (-EPROTO);
 492                 }
 493
 494                 if (req->rq_replay && it_disposition(it, DISP_OPEN_OPEN) &&
 495                     !it_open_error(DISP_OPEN_OPEN, it)) {
 496                         /*
 497                          * If this is a successful OPEN request, we need to set
 498                          * replay handler and data early, so that if replay
 499                          * happens immediately after swabbing below, new reply
 500                          * is swabbed by that handler correctly.
 501                          */
 502                         mdc_set_open_replay_data(NULL, NULL, req);
 503                 }
 504
 505                 if ((body->valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) != 0) {
 506                         void *eadata;
 507
 508                         /*
 509                          * The eadata is opaque; just check that it is there.
 510                          * Eventually, obd_unpackmd() will check the contents.
 511                          */
 512                         eadata = lustre_swab_repbuf(req, reply_off++,
 513                                                     body->eadatasize, NULL);
 514                         if (eadata == NULL) {
 515                                 CERROR("Missing/short eadata\n");
 516                                 RETURN(-EPROTO);
 517                         }
 518                         if (body->valid & OBD_MD_FLMODEASIZE) {
 519                                 if (obddev->u.cli.cl_max_mds_easize <
 520                                     body->max_mdsize) {
 521                                         obddev->u.cli.cl_max_mds_easize =
 522                                                 body->max_mdsize;
 523                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 524                                                body->max_mdsize);
 525                                 }
 526                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 527                                     body->max_cookiesize) {
 528                                         obddev->u.cli.cl_max_mds_cookiesize =
 529                                                 body->max_cookiesize;
 530                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 531                                                body->max_cookiesize);
 532                                 }
 533                         }
 534
 535                         /*
 536                          * We save the reply LOV EA in case we have to replay a
 537                          * create for recovery.  If we didn't allocate a large
 538                          * enough request buffer above we need to reallocate it
 539                          * here to hold the actual LOV EA.
 540                          *
 541                          * To not save LOV EA if request is not going to replay
 542                          * (for example error one).
 543                          */
 544                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 545                                 if (lustre_msg_buflen(req->rq_reqmsg,
 546                                                       DLM_INTENT_REC_OFF + 4) <
 547                                     body->eadatasize)
 548                                         mdc_realloc_openmsg(req, body, size);
 549
 550                                 lmm = lustre_msg_buf(req->rq_reqmsg,
 551                                                      DLM_INTENT_REC_OFF + 4,
 552                                                      body->eadatasize);
 553                                 if (lmm)
 554                                         memcpy(lmm, eadata, body->eadatasize);
 555                         }
 556                 }
 557                 if (body->valid & OBD_MD_FLRMTPERM) {
 558                         struct mdt_remote_perm *perm;
 559
 560                         LASSERT(client_is_remote(exp));
 561                         perm = lustre_swab_repbuf(req, reply_off++,
 562                                                   sizeof(*perm),
 563                                                   lustre_swab_mdt_remote_perm);
 564                         if (perm == NULL) {
 565                                 CERROR("missing remote permission!\n");
 566                                 RETURN(-EPROTO);
 567                         }
 568                 } else if ((body->valid & OBD_MD_FLACL) && body->aclsize) {
 569                         reply_off++;
 570                 }
 571                 if (body->valid & OBD_MD_FLMDSCAPA) {
 572                         struct lustre_capa *capa, *p;
 573
 574                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 575                         if (capa == NULL) {
 576                                 CERROR("Missing/short MDS capability\n");
 577                                 RETURN(-EPROTO);
 578                         }
 579
 580                         if (it->it_op & IT_OPEN) {
 581                                 /* client fid capa will be checked in replay */
 582                                 p = lustre_msg_buf(req->rq_reqmsg,
 583                                                    DLM_INTENT_REC_OFF + 2,
 584                                                    sizeof(*p));
 585                                 LASSERT(p);
 586                                 *p = *capa;
 587                         }
 588                 }
 589                 if (body->valid & OBD_MD_FLOSSCAPA) {
 590                         struct lustre_capa *capa;
 591
 592                         capa = lustre_unpack_capa(req->rq_repmsg, reply_off++);
 593                         if (capa == NULL) {
 594                                 CERROR("Missing/short OSS capability\n");
 595                                 RETURN(-EPROTO);
 596                         }
 597                 }
 598         }
 599
 600         RETURN(rc);
 601 }
 602 /*
 603  * This long block is all about fixing up the lock and request state
 604  * so that it is correct as of the moment _before_ the operation was
 605  * applied; that way, the VFS will think that everything is normal and
 606  * call Lustre's regular VFS methods.
 607  *
 608  * If we're performing a creation, that means that unless the creation
 609  * failed with EEXIST, we should fake up a negative dentry.
 610  *
 611  * For everything else, we want to lookup to succeed.
 612  *
 613  * One additional note: if CREATE or OPEN succeeded, we add an extra
 614  * reference to the request because we need to keep it around until
 615  * ll_create/ll_open gets called.
 616  *
 617  * The server will return to us, in it_disposition, an indication of
 618  * exactly what d.lustre.it_status refers to.
 619  *
 620  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 621  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 622  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 623  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 624  * was successful.
 625  *
 626  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 627  * child lookup.
 628  */
 629 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
 630                     void *lmm, int lmmsize, struct lookup_intent *it,
 631                     int lookup_flags, struct ptlrpc_request **reqp,
 632                     ldlm_blocking_callback cb_blocking,
 633                     int extra_lock_flags)
 634 {
 635         struct ptlrpc_request *request;
 636         struct lustre_handle old_lock;
 637         struct lustre_handle lockh;
 638         struct mdt_body *mdt_body;
 639         struct ldlm_lock *lock;
 640         int rc = 0;
 641         ENTRY;
 642         LASSERT(it);
 643
 644         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
 645                ", intent: %s flags %#o\n", op_data->op_namelen,
 646                op_data->op_name, PFID(&op_data->op_fid2),
 647                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
 648                it->it_flags);
 649
 650         if (fid_is_sane((struct lu_fid *)&op_data->op_fid2) &&
 651             (it->it_op & (IT_LOOKUP | IT_GETATTR))) {
 652                 /* We could just return 1 immediately, but since we should only
 653                  * be called in revalidate_it if we already have a lock, let's
 654                  * verify that. */
 655                 struct ldlm_res_id res_id = { .name = { fid_seq(&op_data->op_fid2),
 656                                                         fid_oid(&op_data->op_fid2),
 657                                                         fid_ver(&op_data->op_fid2) } };
 658                 struct lustre_handle lockh;
 659                 ldlm_policy_data_t policy;
 660                 ldlm_mode_t mode = LCK_CR;
 661
 662                 /* As not all attributes are kept under update lock, e.g.
 663                    owner/group/acls are under lookup lock, we need both
 664                    ibits for GETATTR. */
 665
 666                 /* For CMD, UPDATE lock and LOOKUP lock can not be got
 667                  * at the same for cross-object, so we can not match
 668                  * the 2 lock at the same time FIXME: but how to handle
 669                  * the above situation */
 670                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 671                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 672
 673                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 674                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 675                                      LDLM_IBITS, &policy, mode, &lockh);
 676                 if (!rc) {
 677                         mode = LCK_CW;
 678                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 679                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 680                                              LDLM_IBITS, &policy, mode, &lockh);
 681                 }
 682                 if (!rc) {
 683                         mode = LCK_PR;
 684                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 685                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 686                                              LDLM_IBITS, &policy, mode, &lockh);
 687                 }
 688
 689                 if (!rc) {
 690                         mode = LCK_PW;
 691                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 692                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 693                                              LDLM_IBITS, &policy, mode, &lockh);
 694                 }
 695
 696                 if (rc) {
 697                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 698                                sizeof(lockh));
 699                         it->d.lustre.it_lock_mode = mode;
 700                 }
 701
 702                 /* Only return failure if it was not GETATTR by cfid
 703                    (from inode_revalidate) */
 704                 if (rc || op_data->op_namelen != 0)
 705                         RETURN(rc);
 706         }
 707
 708         /* lookup_it may be called only after revalidate_it has run, because
 709          * revalidate_it cannot return errors, only zero.  Returning zero causes
 710          * this call to lookup, which *can* return an error.
 711          *
 712          * We only want to execute the request associated with the intent one
 713          * time, however, so don't send the request again.  Instead, skip past
 714          * this and use the request from revalidate.  In this case, revalidate
 715          * never dropped its reference, so the refcounts are all OK */
 716         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 717                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 718                                  op_data, &lockh, lmm, lmmsize,
 719                                  ldlm_completion_ast, cb_blocking, NULL,
 720                                  extra_lock_flags);
 721                 if (rc < 0)
 722                         RETURN(rc);
 723                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 724         } else if (!fid_is_sane(&op_data->op_fid2) ||
 725                         !(it->it_flags & O_CHECK_STALE)) {
 726                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 727                  * request referenced from this intent, saved for subsequent
 728                  * lookup.  This path is executed when we proceed to this
 729                  * lookup, so we clear DISP_ENQ_COMPLETE */
 730                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 731         }
 732         request = *reqp = it->d.lustre.it_data;
 733         LASSERT(request != NULL);
 734         LASSERT(request != LP_POISON);
 735         LASSERT(request->rq_repmsg != LP_POISON);
 736
 737         /* If we're doing an IT_OPEN which did not result in an actual
 738          * successful open, then we need to remove the bit which saves
 739          * this request for unconditional replay.
 740          *
 741          * It's important that we do this first!  Otherwise we might exit the
 742          * function without doing so, and try to replay a failed create
 743          * (bug 3440) */
 744         if (it->it_op & IT_OPEN && request->rq_replay &&
 745             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 746                 mdc_clear_replay_flag(request, it->d.lustre.it_status);
 747
 748         if (!it_disposition(it, DISP_IT_EXECD)) {
 749                 /* The server failed before it even started executing the
 750                  * intent, i.e. because it couldn't unpack the request. */
 751                 LASSERT(it->d.lustre.it_status != 0);
 752                 RETURN(it->d.lustre.it_status);
 753         }
 754         rc = it_open_error(DISP_IT_EXECD, it);
 755         if (rc)
 756                 RETURN(rc);
 757
 758         mdt_body = lustre_msg_buf(request->rq_repmsg, DLM_REPLY_REC_OFF,
 759                                   sizeof(*mdt_body));
 760         LASSERT(mdt_body != NULL);      /* mdc_enqueue checked */
 761         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 762
 763         /* If we were revalidating a fid/name pair, mark the intent in
 764          * case we fail and get called again from lookup */
 765         if (fid_is_sane(&op_data->op_fid2) && it->it_flags & O_CHECK_STALE
 766                         && it->it_op != IT_GETATTR) {
 767                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 768
 769                 /* Also: did we find the same inode? */
 770                 if (!lu_fid_eq(&op_data->op_fid2, &mdt_body->fid1))
 771                         RETURN(-ESTALE);
 772         }
 773
 774         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 775         if (rc)
 776                 RETURN(rc);
 777
 778         /* keep requests around for the multiple phases of the call
 779          * this shows the DISP_XX must guarantee we make it into the call
 780          */
 781         if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
 782             it_disposition(it, DISP_OPEN_CREATE) &&
 783             !it_open_error(DISP_OPEN_CREATE, it)) {
 784                 it_set_disposition(it, DISP_ENQ_CREATE_REF);
 785                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 786         }
 787         if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
 788             it_disposition(it, DISP_OPEN_OPEN) &&
 789             !it_open_error(DISP_OPEN_OPEN, it)) {
 790                 it_set_disposition(it, DISP_ENQ_OPEN_REF);
 791                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 792         }
 793
 794         if (it->it_op & IT_CREAT) {
 795                 /* XXX this belongs in ll_create_it */
 796         } else if (it->it_op == IT_OPEN) {
 797                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 798         } else {
 799                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 800         }
 801
 802         /* If we already have a matching lock, then cancel the new
 803          * one.  We have to set the data here instead of in
 804          * mdc_enqueue, because we need to use the child's inode as
 805          * the l_ast_data to match, and that's not available until
 806          * intent_finish has performed the iget().) */
 807         lock = ldlm_handle2lock(&lockh);
 808         if (lock) {
 809                 ldlm_policy_data_t policy = lock->l_policy_data;
 810                 LDLM_DEBUG(lock, "matching against this");
 811
 812                 LASSERTF(fid_res_name_eq(&mdt_body->fid1,
 813                                          &lock->l_resource->lr_name),
 814                          "Lock res_id: %lu/%lu/%lu, fid: %lu/%lu/%lu.\n",
 815                          (unsigned long)lock->l_resource->lr_name.name[0],
 816                          (unsigned long)lock->l_resource->lr_name.name[1],
 817                          (unsigned long)lock->l_resource->lr_name.name[2],
 818                          (unsigned long)fid_seq(&mdt_body->fid1),
 819                          (unsigned long)fid_oid(&mdt_body->fid1),
 820                          (unsigned long)fid_ver(&mdt_body->fid1));
 821                 LDLM_LOCK_PUT(lock);
 822
 823                 memcpy(&old_lock, &lockh, sizeof(lockh));
 824                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 825                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 826                         ldlm_lock_decref_and_cancel(&lockh,
 827                                                     it->d.lustre.it_lock_mode);
 828                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 829                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 830                                sizeof(lockh));
 831                 }
 832         }
 833         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 834                op_data->op_namelen, op_data->op_name, ldlm_it2str(it->it_op),
 835                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 836
 837         RETURN(rc);
 838 }