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/obd_class.h>
  40 #include <linux/lustre_mds.h>
  41 #include <linux/lustre_dlm.h>
  42 #include <linux/lprocfs_status.h>
  43 #include "mdc_internal.h"
  44
  45 int it_disposition(struct lookup_intent *it, int flag)
  46 {
  47         return it->d.lustre.it_disposition & flag;
  48 }
  49 EXPORT_SYMBOL(it_disposition);
  50
  51 void it_set_disposition(struct lookup_intent *it, int flag)
  52 {
  53         it->d.lustre.it_disposition |= flag;
  54 }
  55 EXPORT_SYMBOL(it_set_disposition);
  56
  57 void it_clear_disposition(struct lookup_intent *it, int flag)
  58 {
  59         it->d.lustre.it_disposition &= ~flag;
  60 }
  61
  62 static int it_to_lock_mode(struct lookup_intent *it)
  63 {
  64         /* CREAT needs to be tested before open (both could be set) */
  65         if (it->it_op & IT_CREAT)
  66                 return LCK_CW;
  67         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP))
  68                 return LCK_CR;
  69
  70         LBUG();
  71         RETURN(-EINVAL);
  72 }
  73
  74 int it_open_error(int phase, struct lookup_intent *it)
  75 {
  76         if (it_disposition(it, DISP_OPEN_OPEN)) {
  77                 if (phase >= DISP_OPEN_OPEN)
  78                         return it->d.lustre.it_status;
  79                 else
  80                         return 0;
  81         }
  82
  83         if (it_disposition(it, DISP_OPEN_CREATE)) {
  84                 if (phase >= DISP_OPEN_CREATE)
  85                         return it->d.lustre.it_status;
  86                 else
  87                         return 0;
  88         }
  89
  90         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  91                 if (phase >= DISP_LOOKUP_EXECD)
  92                         return it->d.lustre.it_status;
  93                 else
  94                         return 0;
  95         }
  96
  97         if (it_disposition(it, DISP_IT_EXECD)) {
  98                 if (phase >= DISP_IT_EXECD)
  99                         return it->d.lustre.it_status;
 100                 else
 101                         return 0;
 102         }
 103         CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
 104                it->d.lustre.it_status);
 105         LBUG();
 106         return 0;
 107 }
 108 EXPORT_SYMBOL(it_open_error);
 109
 110 /* this must be called on a lockh that is known to have a referenced lock */
 111 void mdc_set_lock_data(__u64 *l, void *data)
 112 {
 113         struct ldlm_lock *lock;
 114         struct lustre_handle *lockh = (struct lustre_handle *)l;
 115         ENTRY;
 116
 117         if (!*l) {
 118                 EXIT;
 119                 return;
 120         }
 121
 122         lock = ldlm_handle2lock(lockh);
 123
 124         LASSERT(lock != NULL);
 125         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 126 #ifdef __KERNEL__
 127         if (lock->l_ast_data && lock->l_ast_data != data) {
 128                 struct inode *new_inode = data;
 129                 struct inode *old_inode = lock->l_ast_data;
 130                 LASSERTF(old_inode->i_state & I_FREEING,
 131                          "Found existing inode %p/%lu/%u state %lu in lock: "
 132                          "setting data to %p/%lu/%u\n", old_inode,
 133                          old_inode->i_ino, old_inode->i_generation,
 134                          old_inode->i_state,
 135                          new_inode, new_inode->i_ino, new_inode->i_generation);
 136         }
 137 #endif
 138         lock->l_ast_data = data;
 139         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 140         LDLM_LOCK_PUT(lock);
 141
 142         EXIT;
 143 }
 144 EXPORT_SYMBOL(mdc_set_lock_data);
 145
 146 int mdc_change_cbdata(struct obd_export *exp, struct ll_fid *fid,
 147                       ldlm_iterator_t it, void *data)
 148 {
 149         struct ldlm_res_id res_id = { .name = {0} };
 150         ENTRY;
 151
 152         res_id.name[0] = fid->id;
 153         res_id.name[1] = fid->generation;
 154
 155         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace, &res_id, it,
 156                            data);
 157
 158         EXIT;
 159         return 0;
 160 }
 161
 162 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 163 {
 164         /* Don't hold error requests for replay. */
 165         if (req->rq_replay) {
 166                 unsigned long irqflags;
 167                 spin_lock_irqsave(&req->rq_lock, irqflags);
 168                 req->rq_replay = 0;
 169                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
 170         }
 171         if (rc && req->rq_transno != 0) {
 172                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 173                 LBUG();
 174         }
 175 }
 176
 177 static int round_up(int val)
 178 {
 179         int ret = 1;
 180         while (val) {
 181                 val >>= 1;
 182                 ret <<= 1;
 183         }
 184         return ret;
 185 }
 186
 187 /* Save a large LOV EA into the request buffer so that it is available
 188  * for replay.  We don't do this in the initial request because the
 189  * original request doesn't need this buffer (at most it sends just the
 190  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 191  * buffer and may also be difficult to allocate and save a very large
 192  * request buffer for each open. (bug 5707)
 193  *
 194  * OOM here may cause recovery failure if lmm is needed (only for the
 195  * original open if the MDS crashed just when this client also OOM'd)
 196  * but this is incredibly unlikely, and questionable whether the client
 197  * could do MDS recovery under OOM anyways... */
 198 static void mdc_realloc_openmsg(struct ptlrpc_request *req,
 199                                 struct mds_body *body, int size[5])
 200 {
 201         int new_size, old_size;
 202         struct lustre_msg *new_msg;
 203
 204         /* save old size */
 205         old_size = lustre_msg_size(5, size);
 206
 207         size[4] = body->eadatasize;
 208         new_size = lustre_msg_size(5, size);
 209         OBD_ALLOC(new_msg, new_size);
 210         if (new_msg != NULL) {
 211                 struct lustre_msg *old_msg = req->rq_reqmsg;
 212                 unsigned long irqflags;
 213
 214                 DEBUG_REQ(D_INFO, req, "replace reqmsg for larger EA %u\n",
 215                           body->eadatasize);
 216                 memcpy(new_msg, old_msg, old_size);
 217                 new_msg->buflens[4] = body->eadatasize;
 218
 219                 spin_lock_irqsave(&req->rq_lock, irqflags);
 220                 req->rq_reqmsg = new_msg;
 221                 req->rq_reqlen = new_size;
 222                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
 223
 224                 OBD_FREE(old_msg, old_size);
 225         } else {
 226                 body->valid &= ~OBD_MD_FLEASIZE;
 227                 body->eadatasize = 0;
 228         }
 229 }
 230
 231 /* We always reserve enough space in the reply packet for a stripe MD, because
 232  * we don't know in advance the file type. */
 233 int mdc_enqueue(struct obd_export *exp,
 234                 int lock_type,
 235                 struct lookup_intent *it,
 236                 int lock_mode,
 237                 struct mdc_op_data *data,
 238                 struct lustre_handle *lockh,
 239                 void *lmm,
 240                 int lmmsize,
 241                 ldlm_completion_callback cb_completion,
 242                 ldlm_blocking_callback cb_blocking,
 243                 void *cb_data, int extra_lock_flags)
 244 {
 245         struct ptlrpc_request *req;
 246         struct obd_device *obddev = class_exp2obd(exp);
 247         struct ldlm_res_id res_id =
 248                 { .name = {data->fid1.id, data->fid1.generation} };
 249         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 250         struct ldlm_request *lockreq;
 251         struct ldlm_intent *lit;
 252         int size[6] = {[MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
 253                        [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit) };
 254         struct ldlm_reply *dlm_rep;
 255         int repsize[4] = {sizeof(*dlm_rep),
 256                           sizeof(struct mds_body),
 257                           obddev->u.cli.cl_max_mds_easize};
 258         void *eadata;
 259         unsigned long irqflags;
 260         int repbufcnt = 3, req_buffers = 2;
 261         int rc, flags = extra_lock_flags | LDLM_FL_HAS_INTENT;
 262         ENTRY;
 263
 264         LASSERTF(lock_type == LDLM_IBITS, "lock type %d\n", lock_type);
 265 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 266 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 267
 268         if (it->it_op & IT_OPEN) {
 269                 it->it_create_mode |= S_IFREG;
 270
 271                 size[req_buffers++] = sizeof(struct mds_rec_create);
 272                 size[req_buffers++] = data->namelen + 1;
 273                 /* As an optimization, we allocate an RPC request buffer for
 274                  * at least a default-sized LOV EA even if we aren't sending
 275                  * one.  We grow the whole request to the next power-of-two
 276                  * size since we get that much from a slab allocation anyways.
 277                  * This avoids an allocation below in the common case where
 278                  * we need to save a default-sized LOV EA for open replay. */
 279                 size[req_buffers++] = max(lmmsize,
 280                                           obddev->u.cli.cl_default_mds_easize);
 281                 rc = lustre_msg_size(req_buffers, size);
 282                 if (rc & (rc - 1))
 283                         size[req_buffers - 1] = min(size[req_buffers - 1] +
 284                                                     round_up(rc) - rc,
 285                                         obddev->u.cli.cl_max_mds_easize);
 286
 287                 if (it->it_flags & O_JOIN_FILE) {
 288                         __u64 head_size = *(__u32*)cb_data;
 289                         __u32 tsize = *(__u32*)lmm;
 290
 291                         /* join is like an unlink of the tail */
 292                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 293                         size[req_buffers++] = sizeof(struct mds_rec_join);
 294                         req = ptlrpc_prep_req(class_exp2cliimp(exp),
 295                                               LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 296                                               req_buffers, size, NULL);
 297                         /* when joining file, cb_data and lmm args together
 298                          * indicate the head file size*/
 299                         mdc_join_pack(req, req_buffers - 1, data,
 300                                       (head_size << 32) | tsize);
 301                         cb_data = NULL;
 302                         lmm = NULL;
 303                 } else
 304                         req = ptlrpc_prep_req(class_exp2cliimp(exp),
 305                                               LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
 306                                               req_buffers, size, NULL);
 307                 if (!req)
 308                         RETURN(-ENOMEM);
 309
 310                 spin_lock_irqsave (&req->rq_lock, irqflags);
 311                 req->rq_replay = 1;
 312                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 313
 314                 /* pack the intent */
 315                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 316                                      sizeof (*lit));
 317                 lit->opc = (__u64)it->it_op;
 318
 319                 /* pack the intended request */
 320                 mdc_open_pack(req, MDS_REQ_INTENT_REC_OFF, data,
 321                               it->it_create_mode, 0,
 322                               it->it_flags, lmm, lmmsize);
 323
 324                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
 325         } else if (it->it_op & IT_UNLINK) {
 326                 size[req_buffers++] = sizeof(struct mds_rec_unlink);
 327                 size[req_buffers++] = data->namelen + 1;
 328                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 329                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 330                                       LDLM_ENQUEUE, req_buffers, size, NULL);
 331                 if (!req)
 332                         RETURN(-ENOMEM);
 333
 334                 /* pack the intent */
 335                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 336                                      sizeof (*lit));
 337                 lit->opc = (__u64)it->it_op;
 338
 339                 /* pack the intended request */
 340                 mdc_unlink_pack(req, MDS_REQ_INTENT_REC_OFF, data);
 341                 /* get ready for the reply */
 342                 repsize[repbufcnt++] = obddev->u.cli.cl_max_mds_cookiesize;
 343         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 344                 obd_valid valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE |
 345                                   OBD_MD_FLACL | OBD_MD_FLMODEASIZE;
 346                 size[req_buffers++] = sizeof(struct mds_body);
 347                 size[req_buffers++] = data->namelen + 1;
 348
 349                 if (it->it_op & IT_GETATTR)
 350                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 351
 352                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 353                                       LDLM_ENQUEUE, req_buffers, size, NULL);
 354                 if (!req)
 355                         RETURN(-ENOMEM);
 356
 357                 /* pack the intent */
 358                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 359                                      sizeof (*lit));
 360                 lit->opc = (__u64)it->it_op;
 361
 362                 /* pack the intended request */
 363                 mdc_getattr_pack(req, MDS_REQ_INTENT_REC_OFF, valid,
 364                                  it->it_flags, data);
 365                 /* get ready for the reply */
 366                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
 367         } else if (it->it_op == IT_READDIR) {
 368                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 369                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 370                                       LDLM_ENQUEUE, 1, size, NULL);
 371                 if (!req)
 372                         RETURN(-ENOMEM);
 373
 374                 /* get ready for the reply */
 375                 repbufcnt = 1;
 376         } else {
 377                 LBUG();
 378                 RETURN(-EINVAL);
 379         }
 380
 381         /* get ready for the reply */
 382         req->rq_replen = lustre_msg_size(repbufcnt, repsize);
 383
 384         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 385         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 386                               lock_type, &policy,lock_mode, &flags, cb_blocking,
 387                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 388                               lockh);
 389         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 390
 391         /* Similarly, if we're going to replay this request, we don't want to
 392          * actually get a lock, just perform the intent. */
 393         if (req->rq_transno || req->rq_replay) {
 394                 lockreq = lustre_msg_buf(req->rq_reqmsg,
 395                                          MDS_REQ_INTENT_LOCKREQ_OFF,
 396                                          sizeof(*lockreq));
 397                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 398         }
 399
 400         /* This can go when we're sure that this can never happen */
 401         LASSERT(rc != -ENOENT);
 402         if (rc == ELDLM_LOCK_ABORTED) {
 403                 lock_mode = 0;
 404                 memset(lockh, 0, sizeof(*lockh));
 405                 rc = 0;
 406         } else if (rc != 0) {
 407                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 408                 LASSERTF(rc < 0, "rc %d\n", rc);
 409                 mdc_clear_replay_flag(req, rc);
 410                 ptlrpc_req_finished(req);
 411                 RETURN(rc);
 412         } else { /* rc = 0 */
 413                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 414                 LASSERT(lock);
 415
 416                 /* If the server gave us back a different lock mode, we should
 417                  * fix up our variables. */
 418                 if (lock->l_req_mode != lock_mode) {
 419                         ldlm_lock_addref(lockh, lock->l_req_mode);
 420                         ldlm_lock_decref(lockh, lock_mode);
 421                         lock_mode = lock->l_req_mode;
 422                 }
 423
 424                 ldlm_lock_allow_match(lock);
 425                 LDLM_LOCK_PUT(lock);
 426         }
 427
 428         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 429         LASSERT(dlm_rep != NULL);           /* checked by ldlm_cli_enqueue() */
 430         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 431
 432         it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
 433         it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
 434         it->d.lustre.it_lock_mode = lock_mode;
 435         it->d.lustre.it_data = req;
 436
 437         if (it->d.lustre.it_status < 0 && req->rq_replay)
 438                 mdc_clear_replay_flag(req, it->d.lustre.it_status);
 439
 440         DEBUG_REQ(D_RPCTRACE, req, "op: %d disposition: %x, status: %d",
 441                   it->it_op,it->d.lustre.it_disposition,it->d.lustre.it_status);
 442
 443         /* We know what to expect, so we do any byte flipping required here */
 444         LASSERT(repbufcnt == 4 || repbufcnt == 1);
 445         if (repbufcnt == 4) {
 446                 struct mds_body *body;
 447
 448                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 449                                           lustre_swab_mds_body);
 450                 if (body == NULL) {
 451                         CERROR ("Can't swab mds_body\n");
 452                         RETURN (-EPROTO);
 453                 }
 454
 455                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 456                         /* The eadata is opaque; just check that it is there.
 457                          * Eventually, obd_unpackmd() will check the contents */
 458                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 459                                                     NULL);
 460                         if (eadata == NULL) {
 461                                 CERROR ("Missing/short eadata\n");
 462                                 RETURN (-EPROTO);
 463                         }
 464                         if (body->valid & OBD_MD_FLMODEASIZE) {
 465                                 if (obddev->u.cli.cl_max_mds_easize <
 466                                                         body->max_mdsize) {
 467                                         obddev->u.cli.cl_max_mds_easize =
 468                                                 body->max_mdsize;
 469                                         CDEBUG(D_INFO, "maxeasize become %d\n",
 470                                                body->max_mdsize);
 471                                 }
 472                                 if (obddev->u.cli.cl_max_mds_cookiesize <
 473                                                         body->max_cookiesize) {
 474                                         obddev->u.cli.cl_max_mds_cookiesize =
 475                                                 body->max_cookiesize;
 476                                         CDEBUG(D_INFO, "cookiesize become %d\n",
 477                                                body->max_cookiesize);
 478                                 }
 479                         }
 480                         /* We save the reply LOV EA in case we have to replay
 481                          * a create for recovery.  If we didn't allocate a
 482                          * large enough request buffer above we need to
 483                          * reallocate it here to hold the actual LOV EA. */
 484                         if (it->it_op & IT_OPEN) {
 485                                 int pos = MDS_REQ_INTENT_REC_OFF + 2;
 486
 487                                 if (req->rq_reqmsg->buflens[pos] <
 488                                     body->eadatasize)
 489                                         mdc_realloc_openmsg(req, body, size);
 490
 491                                 lmm = lustre_msg_buf(req->rq_reqmsg, pos,
 492                                                      body->eadatasize);
 493                                 if (lmm)
 494                                         memcpy(lmm, eadata, body->eadatasize);
 495                         }
 496                 }
 497         }
 498
 499         RETURN(rc);
 500 }
 501 EXPORT_SYMBOL(mdc_enqueue);
 502
 503 /*
 504  * This long block is all about fixing up the lock and request state
 505  * so that it is correct as of the moment _before_ the operation was
 506  * applied; that way, the VFS will think that everything is normal and
 507  * call Lustre's regular VFS methods.
 508  *
 509  * If we're performing a creation, that means that unless the creation
 510  * failed with EEXIST, we should fake up a negative dentry.
 511  *
 512  * For everything else, we want to lookup to succeed.
 513  *
 514  * One additional note: if CREATE or OPEN succeeded, we add an extra
 515  * reference to the request because we need to keep it around until
 516  * ll_create/ll_open gets called.
 517  *
 518  * The server will return to us, in it_disposition, an indication of
 519  * exactly what d.lustre.it_status refers to.
 520  *
 521  * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
 522  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 523  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 524  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 525  * was successful.
 526  *
 527  * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
 528  * child lookup.
 529  */
 530 int mdc_intent_lock(struct obd_export *exp, struct mdc_op_data *op_data,
 531                     void *lmm, int lmmsize, struct lookup_intent *it,
 532                     int lookup_flags, struct ptlrpc_request **reqp,
 533                     ldlm_blocking_callback cb_blocking, int extra_lock_flags)
 534 {
 535         struct lustre_handle lockh;
 536         struct ptlrpc_request *request;
 537         int rc = 0;
 538         struct mds_body *mds_body;
 539         struct lustre_handle old_lock;
 540         struct ldlm_lock *lock;
 541         ENTRY;
 542         LASSERT(it);
 543
 544         CDEBUG(D_DLMTRACE,"name: %.*s in inode "LPU64", intent: %s flags %#o\n",
 545                op_data->namelen, op_data->name, op_data->fid1.id,
 546                ldlm_it2str(it->it_op), it->it_flags);
 547
 548         if (op_data->fid2.id &&
 549             (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR)) {
 550                 /* We could just return 1 immediately, but since we should only
 551                  * be called in revalidate_it if we already have a lock, let's
 552                  * verify that. */
 553                 struct ldlm_res_id res_id = {.name ={op_data->fid2.id,
 554                                                      op_data->fid2.generation}};
 555                 struct lustre_handle lockh;
 556                 ldlm_policy_data_t policy;
 557                 int mode = LCK_CR;
 558
 559                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 560                                 MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 561                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 562                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 563                                      LDLM_IBITS, &policy, LCK_CR, &lockh);
 564                 if (!rc) {
 565                         mode = LCK_CW;
 566                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 567                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 568                                              LDLM_IBITS, &policy, LCK_CW, &lockh);
 569                 }
 570                 if (!rc) {
 571                         mode = LCK_PR;
 572                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 573                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 574                                              LDLM_IBITS, &policy, LCK_PR, &lockh);
 575                 }
 576                 if (rc) {
 577                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 578                                sizeof(lockh));
 579                         it->d.lustre.it_lock_mode = mode;
 580                 }
 581                 RETURN(rc);
 582         }
 583
 584         /* lookup_it may be called only after revalidate_it has run, because
 585          * revalidate_it cannot return errors, only zero.  Returning zero causes
 586          * this call to lookup, which *can* return an error.
 587          *
 588          * We only want to execute the request associated with the intent one
 589          * time, however, so don't send the request again.  Instead, skip past
 590          * this and use the request from revalidate.  In this case, revalidate
 591          * never dropped its reference, so the refcounts are all OK */
 592         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 593
 594                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 595                                  op_data, &lockh, lmm, lmmsize,
 596                                  ldlm_completion_ast, cb_blocking, NULL,
 597                                  extra_lock_flags);
 598                 if (rc < 0)
 599                         RETURN(rc);
 600                 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
 601         } else if (!op_data->fid2.id) {
 602                 /* DISP_ENQ_COMPLETE set means there is extra reference on
 603                  * request referenced from this intent, saved for subsequent
 604                  * lookup.  This path is executed when we proceed to this
 605                  * lookup, so we clear DISP_ENQ_COMPLETE */
 606                 it_clear_disposition(it, DISP_ENQ_COMPLETE);
 607         }
 608         request = *reqp = it->d.lustre.it_data;
 609         LASSERT(request != NULL);
 610         LASSERT(request != LP_POISON);
 611         LASSERT(request->rq_repmsg != LP_POISON);
 612
 613         /* If we're doing an IT_OPEN which did not result in an actual
 614          * successful open, then we need to remove the bit which saves
 615          * this request for unconditional replay.
 616          *
 617          * It's important that we do this first!  Otherwise we might exit the
 618          * function without doing so, and try to replay a failed create
 619          * (bug 3440) */
 620         if (it->it_op & IT_OPEN && request->rq_replay &&
 621             (!it_disposition(it, DISP_OPEN_OPEN) ||it->d.lustre.it_status != 0))
 622                 mdc_clear_replay_flag(request, it->d.lustre.it_status);
 623
 624         if (!it_disposition(it, DISP_IT_EXECD)) {
 625                 /* The server failed before it even started executing the
 626                  * intent, i.e. because it couldn't unpack the request. */
 627                 LASSERT(it->d.lustre.it_status != 0);
 628                 RETURN(it->d.lustre.it_status);
 629         }
 630         rc = it_open_error(DISP_IT_EXECD, it);
 631         if (rc)
 632                 RETURN(rc);
 633
 634         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 635         LASSERT(mds_body != NULL);           /* mdc_enqueue checked */
 636         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 637
 638         /* If we were revalidating a fid/name pair, mark the intent in
 639          * case we fail and get called again from lookup */
 640         if (op_data->fid2.id) {
 641                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 642                 /* Also: did we find the same inode? */
 643                 if (memcmp(&op_data->fid2, &mds_body->fid1, sizeof(op_data->fid2)))
 644                         RETURN(-ESTALE);
 645         }
 646
 647         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 648         if (rc)
 649                 RETURN(rc);
 650
 651         /* keep requests around for the multiple phases of the call
 652          * this shows the DISP_XX must guarantee we make it into the call
 653          */
 654         if (it_disposition(it, DISP_OPEN_CREATE) &&
 655             !it_open_error(DISP_OPEN_CREATE, it))
 656                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 657         if (it_disposition(it, DISP_OPEN_OPEN) &&
 658             !it_open_error(DISP_OPEN_OPEN, it))
 659                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 660
 661         if (it->it_op & IT_CREAT) {
 662                 /* XXX this belongs in ll_create_it */
 663         } else if (it->it_op == IT_OPEN) {
 664                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 665         } else {
 666                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
 667         }
 668
 669         /* If we already have a matching lock, then cancel the new
 670          * one.  We have to set the data here instead of in
 671          * mdc_enqueue, because we need to use the child's inode as
 672          * the l_ast_data to match, and that's not available until
 673          * intent_finish has performed the iget().) */
 674         lock = ldlm_handle2lock(&lockh);
 675         if (lock) {
 676                 ldlm_policy_data_t policy = lock->l_policy_data;
 677                 LDLM_DEBUG(lock, "matching against this");
 678                 LDLM_LOCK_PUT(lock);
 679                 memcpy(&old_lock, &lockh, sizeof(lockh));
 680                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 681                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 682                         ldlm_lock_decref_and_cancel(&lockh,
 683                                                     it->d.lustre.it_lock_mode);
 684                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 685                         memcpy(&it->d.lustre.it_lock_handle, &lockh,
 686                                sizeof(lockh));
 687                 }
 688         }
 689         CDEBUG(D_DENTRY,"D_IT dentry %.*s intent: %s status %d disp %x rc %d\n",
 690                op_data->namelen, op_data->name, ldlm_it2str(it->it_op),
 691                it->d.lustre.it_status, it->d.lustre.it_disposition, rc);
 692
 693         RETURN(rc);
 694 }
 695 EXPORT_SYMBOL(mdc_intent_lock);